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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       ILP based instruction scheduling.
  23  * @author      Christian Wuerdig
  24  * @date        22.10.2006
  25  * @version     $Id$
  26  *
  27  * An ILP scheduler based on
  28  * "ILP-based Instruction Scheduling for IA-64"
  29  * by Daniel Kaestner and Sebastian Winkel
  30  * extended with register pressure constraints by Christian Wuerdig
  31  */
  32 #include "config.h"
  33
  34 #ifdef WITH_ILP
  35
  36 #include <math.h>
  37
  38 #ifndef _WIN32
  39 #include <strings.h>
  40 #endif /* _WIN32 */
  41
  42 #include "irnode_t.h"
  43 #include "irgwalk.h"
  44 #include "irbitset.h"
  45 #include "irphase_t.h"
  46 #include "height.h"
  47 #include "iredges.h"
  48 #include "pdeq.h"
  49 #include "debug.h"
  50 #include "irtools.h"
  51 #include "irdump.h"
  52 #include "irprintf.h"
  53 #include "plist.h"
  54 #include "irprintf.h"
  55 #include "timing.h"
  56
  57 #include <lpp/lpp.h>
  58 #include <lpp/lpp_net.h>
  59
  60 #include "lc_opts.h"
  61 #include "lc_opts_enum.h"
  62
  63 #include "be.h"
  64 #include "benode_t.h"
  65 #include "besched_t.h"
  66 #include "beilpsched.h"
  67 #include "beutil.h"
  68 #include "bestat.h"
  69 #include "beirg_t.h"
  70
  71 typedef struct _ilpsched_options_t {
  72         unsigned regpress;
  73         unsigned time_limit;
  74         char     log_file[1024];
  75 } ilpsched_options_t;
  76
  77 typedef struct _unit_type_info_t {
  78         int                            n_units;
  79         const be_execution_unit_type_t *tp;
  80 } unit_type_info_t;
  81
  82 /**
  83  * holding the ILP variables of the different types
  84  */
  85 typedef struct _ilp_var_types_t {
  86         int *x;   /* x_{nt}^k variables */
  87         int *a;   /* a_{nt}^k variables */
  88         int *y;   /* y_{nt}^k variables */
  89 } ilp_var_types_t;
  90
  91 /**
  92  * Holds alive variables for a node live-in to a block.
  93  */
  94 typedef struct _ilp_livein_node_t {
  95         ir_node  *irn;
  96         unsigned max_alive_steps;
  97         int      *a;
  98 } ilp_livein_node_t;
  99
 100 /* attributes for a node */
 101 typedef struct _ilpsched_node_attr_t {
 102         unsigned asap;                     /**< The ASAP scheduling control step */
 103         unsigned alap;                     /**< The ALAP scheduling control step */
 104         unsigned latency;                  /**< Latency of this node (needed for sorting) */
 105         unsigned sched_point;              /**< the step in which the node is finally scheduled */
 106         unsigned visit_idx;                /**< Index of the node having visited this node last */
 107         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
 108         unsigned n_consumer;               /**< Number of consumers */
 109         ir_node  **block_consumer;         /**< List of consumer being in the same block */
 110         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
 111         unsigned block_idx     : 30;       /**< A unique per block index */
 112         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
 113         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
 114         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
 115                                                                                         for ASAP, successors for ALAP */
 116         unsigned         n_unit_types;     /**< number of allowed execution unit types */
 117         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
 118         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
 119 } ilpsched_node_attr_t;
 120
 121 /* attributes for a block */
 122 typedef struct _ilpsched_block_attr_t {
 123         unsigned block_last_idx;        /**< The highest node index in block so far */
 124         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 125         unsigned max_steps;             /**< Upper bound for block execution */
 126         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 127         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 128         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 129 } ilpsched_block_attr_t;
 130
 131 typedef union _ilpsched_attr_ {
 132         ilpsched_node_attr_t  node_attr;
 133         ilpsched_block_attr_t block_attr;
 134 } ilpsched_attr_t;
 135
 136 /* A irn for the phase and it's attributes (either node or block) */
 137 typedef struct {
 138         const ir_node   *irn;
 139         ilpsched_attr_t attr;
 140 } be_ilpsched_irn_t;
 141
 142 /* The ILP scheduling environment */
 143 typedef struct {
 144         ir_phase             ph;            /**< The phase */
 145         ir_graph             *irg;          /**< The current irg */
 146         heights_t            *height;       /**< The heights object of the irg */
 147         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 148         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 149         const arch_env_t     *arch_env;
 150         const be_main_env_t  *main_env;
 151         const be_machine_t   *cpu;          /**< the current abstract machine */
 152         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 153         const be_irg_t       *birg;         /**< The birg object */
 154         be_options_t         *be_opts;      /**< backend options */
 155         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 156         DEBUG_ONLY(firm_dbg_module_t *dbg);
 157 } be_ilpsched_env_t;
 158
 159 /* convenience macros to handle phase irn data */
 160 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 161 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 162 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 163 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 164
 165 /* check if node is considered for ILP scheduling */
 166 #define consider_for_sched(env, irn) \
 167         (! (is_Block(irn)            ||  \
 168                 is_normal_Proj(env, irn) ||  \
 169                 is_Phi(irn)              ||  \
 170                 is_NoMem(irn)            ||  \
 171                 is_Unknown(irn)          ||  \
 172                 is_End(irn)                  \
 173                 ))
 174
 175 /* gives the valid scheduling time step interval for a node */
 176 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 177
 178 /* gives the valid interval where a node can die */
 179 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 180
 181 /* gives the corresponding ILP variable for given node, unit and time step */
 182 #define ILPVAR_IDX(na, unit, control_step) \
 183         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 184
 185 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 186 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 187         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 188
 189 /* check if a double value is within an epsilon environment of 0 */
 190 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 191
 192 #define ilp_timer_push(t)         ir_timer_push((t))
 193 #define ilp_timer_pop()           ir_timer_pop()
 194 #define ilp_timer_elapsed_usec(t) ir_timer_elapsed_usec((t))
 195
 196 /* option variable */
 197 static ilpsched_options_t ilp_opts = {
 198         1,     /* default is with register pressure constraints */
 199         300,   /* 300 sec per block time limit */
 200         ""     /* no log file */
 201 };
 202
 203 /* ILP options */
 204 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 205         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 206         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 207         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 208         LC_OPT_LAST
 209 };
 210
 211 /*
 212         We need this global variable as we compare nodes dependent on heights,
 213         but we cannot pass any information to the qsort compare function.
 214 */
 215 static heights_t *glob_heights;
 216
 217 /**
 218  * Check if irn is a Proj, which has no execution units assigned.
 219  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 220  */
 221 static inline int is_normal_Proj(const arch_env_t *env, const ir_node *irn) {
 222         return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
 223 }
 224
 225 /**
 226  * Skips normal Projs.
 227  * @return predecessor if irn is a normal Proj, otherwise irn.
 228  */
 229 static inline ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn) {
 230         if (is_normal_Proj(env, irn))
 231                 return get_Proj_pred(irn);
 232         return irn;
 233 }
 234
 235 static inline int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 236         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 237         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 238                 lat = 1;
 239         return lat;
 240 }
 241
 242 static int cmp_live_in_nodes(const void *a, const void *b) {
 243         const ilp_livein_node_t *n1 = a;
 244         const ilp_livein_node_t *n2 = b;
 245
 246         return n1->irn != n2->irn;
 247 }
 248
 249 /**
 250  * Compare scheduling time steps of two be_ilpsched_irn's.
 251  */
 252 static int cmp_ilpsched_irn(const void *a, const void *b) {
 253         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 254         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 255         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 256         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 257
 258         if (n1_a->sched_point == n2_a->sched_point) {
 259                 const ir_node *irn_a = n1->irn;
 260                 const ir_node *irn_b = n2->irn;
 261
 262                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 263                         return 1;
 264                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 265                         return -1;
 266
 267                 /*
 268                         Ok, timestep is equal and the nodes are parallel,
 269                         so check latency and schedule high latency first.
 270                 */
 271                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 272         }
 273         else
 274                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 275 }
 276
 277 /**
 278  * In case there is no phase information for irn, initialize it.
 279  */
 280 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old) {
 281         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 282
 283         if (res == old) {
 284                 /* if we have already some data: check for reinitialization */
 285
 286                 if (! is_Block(irn)) {
 287                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 288
 289                         if (! na->transitive_block_nodes) {
 290                                 ir_node               *block      = get_nodes_block(irn);
 291                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 292                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 293
 294                                 /* we are called after the block indices have been build: create bitset */
 295                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 296                         }
 297                         else {
 298                                 /* we are called from reinit block data: clear the bitset */
 299                                 bitset_clear_all(na->transitive_block_nodes);
 300                                 na->visit_idx    = 0;
 301                                 na->alap_changed = 1;
 302                         }
 303                 }
 304                 return old;
 305         }
 306
 307         res->irn = irn;
 308
 309         /* set ilpsched irn attributes (either block or irn) */
 310         if (is_Block(irn)) {
 311                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 312
 313                 ba->n_interesting_nodes = 0;
 314                 ba->block_last_idx      = 0;
 315                 ba->root_nodes          = plist_new();
 316                 ba->head_ilp_nodes      = NULL;
 317                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 318                 ba->max_steps           = 0;
 319         }
 320         else {
 321                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 322                 memset(na, 0, sizeof(*na));
 323         }
 324
 325         return res;
 326 }
 327
 328 /**
 329  * Assign a per block unique number to each node.
 330  */
 331 static void build_block_idx(ir_node *irn, void *walk_env) {
 332         be_ilpsched_env_t     *env = walk_env;
 333         be_ilpsched_irn_t     *node, *block_node;
 334         ilpsched_node_attr_t  *na;
 335         ilpsched_block_attr_t *ba;
 336
 337         set_irn_link(irn, NULL);
 338         if (! consider_for_sched(env->arch_env, irn))
 339                 return;
 340
 341         node       = get_ilpsched_irn(env, irn);
 342         na         = get_ilpsched_node_attr(node);
 343         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 344         ba         = get_ilpsched_block_attr(block_node);
 345
 346         na->block_idx = ba->block_last_idx++;
 347 }
 348
 349 /********************************************************
 350  *                              __        _
 351  *                             / /       | |
 352  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 353  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 354  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 355  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 356  *                 | |                           | |
 357  *                 |_|                           |_|
 358  ********************************************************/
 359
 360 /**
 361  * Add all nodes having no user in current block to last_nodes list.
 362  */
 363 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 364         ir_node               *block;
 365         const ir_edge_t       *edge;
 366         be_ilpsched_irn_t     *block_node, *node;
 367         ilpsched_block_attr_t *ba;
 368         ilpsched_node_attr_t  *na;
 369         int                   i, j;
 370         be_ilpsched_env_t     *env           = walk_env;
 371         int                   has_block_user = 0;
 372         unsigned              n_consumer     = 0;
 373         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 374         ir_node               **consumer;
 375         unsigned              idx;
 376
 377         if (! consider_for_sched(env->arch_env, irn))
 378                 return;
 379
 380         block    = get_nodes_block(irn);
 381     idx      = get_irn_idx(irn);
 382         consumer = NEW_ARR_F(ir_node *, 0);
 383
 384         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 385
 386         /* check data and dependency out edges */
 387         for (i = 0; i < 2 && ! has_block_user; ++i) {
 388                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 389                         ir_node *user = get_edge_src_irn(edge);
 390
 391                         if (is_normal_Proj(env->arch_env, user)) {
 392                                 const ir_edge_t *user_edge;
 393
 394                                 if (get_irn_mode(user) == mode_X)
 395                                         continue;
 396
 397                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 398                                 for (j = 0; j < 2; ++j) {
 399                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 400                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 401
 402                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 403                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 404                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 405
 406                                                         /* skip already visited nodes */
 407                                                         if (ua->consumer_idx == idx)
 408                                                                 continue;
 409
 410                                                         /* check if node has user in this block and collect the user if it's a data user */
 411                                                         if (get_nodes_block(real_user) == block) {
 412                                                                 if (i == 0 && j == 0)
 413                                                                         ARR_APP1(ir_node *, consumer, real_user);
 414                                                                 has_block_user = 1;
 415                                                         }
 416
 417                                                         /* only count data consumer */
 418                                                         if (i == 0)
 419                                                                 n_consumer++;
 420
 421                                                         /* mark user as visited by this node */
 422                                                         ua->consumer_idx = idx;
 423                                                 }
 424                                         }
 425                                 }
 426                         }
 427                         else if (is_Block(user)) {
 428                                 continue;
 429                         }
 430                         else if (! is_Phi(user)) {
 431                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 432                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 433
 434                                 /* skip already visited nodes */
 435                                 if (ua->consumer_idx == idx)
 436                                         continue;
 437
 438                                 /* check if node has user in this block and collect the user if it's a data user */
 439                                 if (get_nodes_block(user) == block) {
 440                                         if (i == 0)
 441                                                 ARR_APP1(ir_node *, consumer, user);
 442                                         has_block_user = 1;
 443                                 }
 444
 445                                 /* only count data consumer */
 446                                 if (i == 0)
 447                                         n_consumer++;
 448
 449                                 /* mark user visited by this node */
 450                                 ua->consumer_idx = idx;
 451                         }
 452                         else if (get_nodes_block(user) != block) {
 453                                 n_consumer++;
 454                         }
 455                 }
 456         }
 457
 458         block_node = get_ilpsched_irn(env, block);
 459         ba         = get_ilpsched_block_attr(block_node);
 460
 461         ba->n_interesting_nodes++;
 462
 463         /* current irn has no user inside this block, add to queue */
 464         if (! has_block_user) {
 465                 DB((env->dbg, LEVEL_3, "root node\n"));
 466                 plist_insert_back(ba->root_nodes, irn);
 467         }
 468         else {
 469                 DB((env->dbg, LEVEL_3, "normal node\n"));
 470         }
 471
 472         /* record number of all consumer and the consumer within the same block */
 473         node = get_ilpsched_irn(env, irn);
 474         na   = get_ilpsched_node_attr(node);
 475         na->n_consumer     = n_consumer;
 476         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 477         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 478         DEL_ARR_F(consumer);
 479 }
 480
 481 /**
 482  * Calculate the ASAP scheduling step for current irn.
 483  */
 484 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 485         be_ilpsched_env_t     *env = walk_env;
 486         int                   i;
 487         ir_node               *block;
 488         be_ilpsched_irn_t     *node, *block_node;
 489         ilpsched_node_attr_t  *na;
 490         ilpsched_block_attr_t *ba;
 491
 492         /* These nodes are handled separate */
 493         if (! consider_for_sched(env->arch_env, irn))
 494                 return;
 495
 496         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 497
 498         block    = get_nodes_block(irn);
 499         node     = get_ilpsched_irn(env, irn);
 500         na       = get_ilpsched_node_attr(node);
 501         na->asap = 1;
 502
 503         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 504                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
 505
 506                 /* check for greatest distance to top */
 507                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 508                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 509                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 510                         unsigned             lat;
 511
 512                         lat         = fixed_latency(env->sel, pred, env->block_env);
 513                         na->latency = lat;
 514                         na->asap    = MAX(na->asap, pna->asap + lat);
 515                 }
 516         }
 517
 518         /* add node to ILP node list and update max_steps */
 519         block_node = get_ilpsched_irn(env, block);
 520         ba         = get_ilpsched_block_attr(block_node);
 521
 522         set_irn_link(irn, ba->head_ilp_nodes);
 523         ba->head_ilp_nodes = irn;
 524         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 525
 526         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 527 }
 528
 529 /**
 530  * Calculate the ALAP scheduling step of all irns in current block.
 531  * Depends on max_steps being calculated.
 532  */
 533 static void calculate_block_alap(ir_node *block, void *walk_env) {
 534         be_ilpsched_env_t     *env        = walk_env;
 535         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 536         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 537         waitq                 *cur_queue  = new_waitq();
 538         plist_element_t       *el;
 539
 540         assert(is_Block(block));
 541
 542         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 543                 block, ba->n_interesting_nodes, ba->max_steps));
 544
 545         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 546         //phase_reinit_block_irn_data(&env->ph, block);
 547
 548         /* init start queue */
 549         foreach_plist(ba->root_nodes, el) {
 550                 waitq_put(cur_queue, plist_element_get_value(el));
 551         }
 552
 553         /* repeat until all nodes are processed */
 554         while (! waitq_empty(cur_queue)) {
 555                 waitq *next_queue = new_waitq();
 556
 557                 /* process all nodes in current step */
 558                 while (! waitq_empty(cur_queue)) {
 559                         ir_node              *cur_irn = waitq_get(cur_queue);
 560                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 561                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 562                         int                  i;
 563
 564                         /* cur_node has no alap set -> it's a root node, set to max alap */
 565                         if (na->alap == 0) {
 566                                 na->alap = ba->max_steps;
 567                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 568                                         cur_irn, na->alap));
 569                         }
 570                         else {
 571                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 572                                         cur_irn, na->alap));
 573                         }
 574
 575                         /* set the alap's of all predecessors */
 576                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 577                                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
 578
 579                                 /* check for greatest distance to bottom */
 580                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 581                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 582                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 583                                         unsigned             lat;
 584
 585                                         /* mark the predecessor as visited by current irn */
 586                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 587                                                 continue;
 588                                         pna->visit_idx = get_irn_idx(cur_irn);
 589
 590                                         lat = fixed_latency(env->sel, pred, env->block_env);
 591
 592                                         /* set ALAP of current pred */
 593                                         if (pna->alap == 0) {
 594                                                 /* current ALAP is 0: node has not yet been visited */
 595                                                 pna->alap_changed = 1;
 596                                                 pna->alap         = na->alap - lat;
 597                                         }
 598                                         else if (pna->alap > na->alap - lat) {
 599                                                 /* we found a longer path to root node: change ALAP */
 600                                                 pna->alap         = na->alap - lat;
 601                                                 pna->alap_changed = 1;
 602                                         }
 603                                         else {
 604                                                 /* current ALAP is best found so far: keep it */
 605                                                 pna->alap_changed = 0;
 606                                         }
 607
 608                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 609
 610                                         /* enqueue node for next iteration */
 611                                         if (get_irn_ins_or_deps(pred) > 0)
 612                                                 waitq_put(next_queue, pred);
 613                                 }
 614                         }
 615                 }
 616
 617                 /* prepare for next iteration */
 618                 del_waitq(cur_queue);
 619                 cur_queue = next_queue;
 620         }
 621 }
 622
 623 /**
 624  * Free list of root nodes and the set of live-in nodes.
 625  */
 626 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 627         be_ilpsched_env_t     *env        = walk_env;
 628         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 629         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 630
 631         plist_free(ba->root_nodes);
 632         ba->root_nodes = NULL;
 633         del_pset(ba->livein_nodes);
 634         ba->livein_nodes = NULL;
 635 }
 636
 637 /**
 638  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 639  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 640  */
 641 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 642         be_ilpsched_env_t    *env  = walk_env;
 643         ir_node              *pred = irn;
 644         be_ilpsched_irn_t    *node, *pred_node;
 645         ilpsched_node_attr_t *na, *pna;
 646
 647         if (! consider_for_sched(env->arch_env, irn))
 648                 return;
 649
 650         if (! is_Proj(irn) && ! be_is_Keep(irn))
 651                 return;
 652
 653         /* go to the ancestor */
 654         if (be_is_Keep(irn))
 655                 pred = get_irn_n(irn, 0);
 656         pred = skip_Proj(pred);
 657
 658         node      = get_ilpsched_irn(env, irn);
 659         pred_node = get_ilpsched_irn(env, pred);
 660         na        = get_ilpsched_node_attr(node);
 661         pna       = get_ilpsched_node_attr(pred_node);
 662
 663         na->asap = pna->asap;
 664         na->alap = pna->alap;
 665
 666         /* record all Projs and Keeps for this node */
 667         if (! pna->projkeeps)
 668                 pna->projkeeps = new_waitq();
 669         waitq_put(pna->projkeeps, irn);
 670
 671         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 672 }
 673
 674 /*******************************************
 675  *           _              _       _
 676  *          | |            | |     | |
 677  *  ___  ___| |__   ___  __| |_   _| | ___
 678  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 679  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 680  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 681  *
 682  *******************************************/
 683
 684 static inline void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn) {
 685         const ir_edge_t *edge;
 686         (void) block;
 687
 688         foreach_out_edge(irn, edge) {
 689                 ir_node *user = get_edge_src_irn(edge);
 690
 691                 if (be_is_Keep(user)) {
 692                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 693                         waitq_put(keeps, user);
 694                 }
 695         }
 696 }
 697
 698 /**
 699  * Inserts @p irn before @p before into schedule and notifies backend.
 700  */
 701 static inline void notified_sched_add_before(be_ilpsched_env_t *env,
 702         const ir_node *before, const ir_node *irn, unsigned cycle)
 703 {
 704         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 705         sched_add_before(before, irn);
 706 }
 707
 708 /**
 709  * Adds a node, it's Projs (in case of mode_T nodes) and
 710  * it's Keeps to schedule.
 711  */
 712 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle) {
 713         const ir_edge_t *edge;
 714         waitq           *keeps = new_waitq();
 715
 716         /* mode_M nodes are not scheduled */
 717         if (get_irn_mode(irn) == mode_M)
 718                 return;
 719
 720         if (! sched_is_scheduled(irn))
 721                 notified_sched_add_before(env, block, irn, cycle);
 722
 723         /* add Projs */
 724         if (get_irn_mode(irn) == mode_T) {
 725                 foreach_out_edge(irn, edge) {
 726                         ir_node *user = get_edge_src_irn(edge);
 727
 728                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 729                                 get_irn_n_edges(user) > 0)
 730                         {
 731                                 notified_sched_add_before(env, block, user, cycle);
 732                         }
 733
 734                         check_for_keeps(keeps, block, user);
 735                 }
 736         }
 737         else {
 738                 check_for_keeps(keeps, block, irn);
 739         }
 740
 741         /* add Keeps */
 742         while (! waitq_empty(keeps)) {
 743                 ir_node *keep = waitq_get(keeps);
 744                 if (! sched_is_scheduled(keep))
 745                         notified_sched_add_before(env, block, keep, cycle);
 746         }
 747
 748         del_waitq(keeps);
 749 }
 750
 751 /**
 752  * Schedule all nodes in the given block, according to the ILP solution.
 753  */
 754 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 755         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 756         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 757         sched_info_t          *info       = get_irn_sched_info(block);
 758         be_ilpsched_irn_t     **sched_nodes;
 759         unsigned              i, l;
 760         ir_node               *cfop, *irn;
 761         const ir_edge_t       *edge;
 762
 763         /* init block schedule list */
 764         INIT_LIST_HEAD(&info->list);
 765         info->scheduled = 1;
 766
 767         /* collect nodes and their scheduling time step */
 768         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 769         if (ba->n_interesting_nodes == 0) {
 770                 /* ignore */
 771         }
 772         else if (ba->n_interesting_nodes == 1) {
 773                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 774
 775                 /* add the single node */
 776                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 777         }
 778         else {
 779                 /* check all nodes for their positive solution */
 780                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 781                         be_ilpsched_irn_t    *node;
 782                         ilpsched_node_attr_t *na;
 783                         int                  tp_idx, found;
 784                         unsigned             cur_var, t;
 785
 786                         node    = get_ilpsched_irn(env, irn);
 787                         na      = get_ilpsched_node_attr(node);
 788                         cur_var = 0;
 789                         found   = 0;
 790
 791                         if (! na->is_dummy_node) {
 792                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 793                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 794                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 795
 796                                                 if (! LPP_VALUE_IS_0(cost)) {
 797                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
 798                                                         found = 1;
 799                                                 }
 800                                         }
 801                                 }
 802                         }
 803
 804                         found = 0;
 805                         /* go over all variables of a node until the non-zero one is found */
 806                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 807                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 808                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 809
 810                                         /* check, if variable is set to one (it's not zero then :) */
 811                                         if (! LPP_VALUE_IS_0(val)) {
 812                                                 na->sched_point = t;
 813                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 814                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 815                                                         irn, t, na->type_info[tp_idx].tp->name));
 816                                                 found = 1;
 817                                         }
 818                                 }
 819                         }
 820                 }
 821
 822                 glob_heights = env->height;
 823                 /* sort nodes ascending by scheduling time step */
 824                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 825         }
 826
 827         /* make all Phis ready and remember the single cf op */
 828         cfop = NULL;
 829         foreach_out_edge(block, edge) {
 830                 irn = get_edge_src_irn(edge);
 831
 832                 switch (get_irn_opcode(irn)) {
 833                         case iro_Phi:
 834                                 add_to_sched(env, block, irn, 0);
 835                                 break;
 836                         case iro_Start:
 837                         case iro_End:
 838                         case iro_Proj:
 839                         case iro_Bad:
 840                         case iro_Unknown:
 841                                 break;
 842                         default:
 843                                 if (is_cfop(irn)) {
 844                                         assert(cfop == NULL && "Highlander - there can be only one");
 845                                         cfop = irn;
 846                                 }
 847                         break;
 848                 }
 849         }
 850
 851         /* add all nodes from list */
 852         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 853                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 854                 if (sched_nodes[i]->irn != cfop)
 855                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 856         }
 857
 858         /* schedule control flow node if not already done */
 859         if (cfop && ! sched_is_scheduled(cfop))
 860                 add_to_sched(env, block, cfop, 0);
 861
 862         DEL_ARR_F(sched_nodes);
 863 }
 864
 865 /***************************************************************
 866  *   _____ _      _____     _____           _   _
 867  *  |_   _| |    |  __ \   / ____|         | | (_)
 868  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 869  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 870  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 871  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 872  *
 873  ***************************************************************/
 874
 875 /**
 876  * Check if node can be executed on given unit type.
 877  */
 878 static inline int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 879         int                  i;
 880         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 881
 882         for (i = na->n_unit_types - 1; i >= 0; --i) {
 883                 if (na->type_info[i].tp == tp)
 884                         return i;
 885         }
 886
 887         return -1;
 888 }
 889
 890 /************************************************
 891  *                   _       _     _
 892  *                  (_)     | |   | |
 893  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 894  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 895  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 896  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 897  *
 898  ************************************************/
 899
 900 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
 901         const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
 902         unsigned                  n_unit_types = 0;
 903         be_ilpsched_irn_t         *node;
 904         ilpsched_node_attr_t      *na;
 905         unsigned                  unit_idx, tp_idx;
 906
 907         /* count number of available unit types for this node */
 908         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 909                 /* just count */ ;
 910
 911         node = get_ilpsched_irn(env, irn);
 912         na   = get_ilpsched_node_attr(node);
 913
 914         if (! na->type_info) {
 915                 na->n_unit_types = n_unit_types;
 916                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 917
 918                 /* fill the type info array */
 919                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 920                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 921                                 /* beware: we also count number of available units here */
 922                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 923                                         na->is_dummy_node = 1;
 924                         }
 925
 926                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 927                         na->type_info[tp_idx].n_units = unit_idx;
 928                 }
 929         }
 930
 931         return n_unit_types;
 932 }
 933
 934 /**
 935  * Returns the largest alap time of a user of @p irn.
 936  * The user must be in block @p block.
 937  */
 938 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block) {
 939         const ir_edge_t *edge;
 940         unsigned        max_alap = 0;
 941
 942         foreach_out_edge(irn, edge) {
 943                 ir_node *user = get_edge_src_irn(edge);
 944
 945                 if (get_nodes_block(user) == block) {
 946                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 947                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 948
 949                         max_alap = MAX(max_alap, na->alap);
 950                 }
 951         }
 952
 953         assert(max_alap > 0);
 954         return max_alap;
 955 }
 956
 957 /**
 958  * Create the following variables:
 959  * - x_{nt}^k    binary     weigthed with: t
 960  *      node n is scheduled at time step t to unit type k
 961  * ==>> These variables represent the schedule
 962  *
 963  * - a_{nt}^k    binary     weighted with num_nodes
 964  *      node n is alive at time step t on unit type k
 965  *
 966  * - y_{nt}^k    continuous  weighted with: num_nodes^2
 967  *      register pressure over limit for unit type k
 968  * ==>> These variables represent the register pressure
 969  *
 970  */
 971 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 972         char                  buf[1024];
 973         ir_node               *irn;
 974         unsigned              num_block_var, num_nodes;
 975         ilp_livein_node_t     *livein;
 976         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 977         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 978         ir_timer_t            *t_var   = ir_timer_register("beilpsched_var", "create ilp variables");
 979
 980         ilp_timer_push(t_var);
 981         num_block_var = num_nodes = 0;
 982         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 983                 be_ilpsched_irn_t    *node;
 984                 ilpsched_node_attr_t *na;
 985                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
 986                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
 987                 int                  i;
 988
 989                 node         = get_ilpsched_irn(env, irn);
 990                 na           = get_ilpsched_node_attr(node);
 991                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
 992
 993                 /* allocate space for ilp variables */
 994                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 995                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 996
 997                 /* we need these variables only for "real" nodes */
 998                 if (! na->is_dummy_node) {
 999                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1000                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1001
1002                         num_ad         = ba->max_steps - na->asap + 1;
1003                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1004                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1005                 }
1006
1007                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1008                         irn, na->asap, na->alap, na->n_unit_types));
1009
1010                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1011                 /* create variables */
1012                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1013                         unsigned t;
1014
1015                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1016                                 /* x_{nt}^k variables */
1017                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1018                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1019                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1020                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1021                                 /* variable counter */
1022                                 n_var++;
1023                                 num_block_var++;
1024
1025                                 if (! na->is_dummy_node) {
1026                                         /* y_{nt}^k variables */
1027                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1028                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1029                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1030                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1031
1032                                         /* variable counter */
1033                                         n_var++;
1034                                         num_block_var++;
1035                                 }
1036                         }
1037
1038                         /* a node can die at any step t: asap(n) <= t <= U */
1039                         if (! na->is_dummy_node) {
1040                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1041
1042                                         /* a_{nt}^k variables */
1043                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1044                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1045                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1046                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1047
1048                                         /* variable counter */
1049                                         n_var++;
1050                                         num_block_var++;
1051                                 }
1052                         }
1053
1054                         /* collect live-in nodes */
1055                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1056                                 ir_node *pred = get_irn_n(irn, i);
1057
1058                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1059                                         be_ilpsched_set_type_info(env, pred, var_obst);
1060                                         if (! na->is_dummy_node) {
1061                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1062                                                 entry->irn = pred;
1063                                                 entry->a   = NULL;
1064                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1065                                         }
1066                                 }
1067                         }
1068                 }
1069
1070                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1071                 num_nodes++;
1072         }
1073
1074         /* create alive variables a_{nt}^k for live-ins */
1075         foreach_pset(ba->livein_nodes, livein) {
1076                 be_ilpsched_irn_t    *node;
1077                 ilpsched_node_attr_t *na;
1078                 unsigned             tp_idx, var_idx;
1079                 ir_node              *irn;
1080
1081                 irn  = livein->irn;
1082                 node = get_ilpsched_irn(env, irn);
1083                 na   = get_ilpsched_node_attr(node);
1084
1085                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1086
1087                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1088                 var_idx   = 0;
1089
1090                 /* create variables */
1091                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1092                         unsigned t;
1093
1094                         for (t = 0; t < livein->max_alive_steps; ++t) {
1095                                 /* a_{nt}^k variables */
1096                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1097                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1098                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1099                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1100                                 num_block_var++;
1101                         }
1102                 }
1103         }
1104
1105         ilp_timer_pop();
1106         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1107                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1108 }
1109
1110 /*******************************************************
1111  *                      _             _       _
1112  *                     | |           (_)     | |
1113  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1114  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1115  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1116  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1117  *
1118  *******************************************************/
1119
1120 /**
1121  * Collect all operands and nodes @p irn depends on.
1122  * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1123  */
1124 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps) {
1125         int i;
1126
1127         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1128                 ir_node *p = get_irn_in_or_dep(irn, i);
1129
1130                 if (is_Proj(p)) {
1131                         const ir_edge_t *edge;
1132
1133                         p = get_Proj_pred(p);
1134                         foreach_out_edge(p, edge) {
1135                                 ir_node *src = get_edge_src_irn(edge);
1136                                 ir_nodeset_insert(deps, src);
1137                         }
1138                 }
1139                 else {
1140                         ir_nodeset_insert(deps, p);
1141                 }
1142         }
1143 }
1144
1145 /**
1146  * Create following ILP constraints:
1147  * - the assignment constraints:
1148  *     assure each node is executed once by exactly one (allowed) execution unit
1149  * - the dead node assignment constraints:
1150  *     assure a node can only die at most once
1151  * - the precedence constraints:
1152  *     assure that no data dependencies are violated
1153  */
1154 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1155         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1156         char                  buf[1024];
1157         ir_node               *irn;
1158         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1159         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1160         ir_timer_t            *t_cst_assign  = ir_timer_register("beilpsched_cst_assign", "create assignment constraints");
1161         ir_timer_t            *t_cst_prec    = ir_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1162
1163         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1164         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1165                 int                  cst, tp_idx;
1166                 unsigned             cur_var;
1167                 be_ilpsched_irn_t    *node;
1168                 ilpsched_node_attr_t *na;
1169                 ir_node              *pred;
1170                 ir_nodeset_t          deps;
1171                 ir_nodeset_iterator_t iter;
1172
1173                 ir_nodeset_init(&deps);
1174
1175                 node    = get_ilpsched_irn(env, irn);
1176                 na      = get_ilpsched_node_attr(node);
1177                 cur_var = 0;
1178
1179                 /* the assignment constraint */
1180                 ilp_timer_push(t_cst_assign);
1181                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1182                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1183                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1184                 num_cst_assign++;
1185
1186                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1187                 ilp_timer_pop();
1188
1189                 /* We have separate constraints for Projs and Keeps */
1190                 // ILP becomes infeasible ?!?
1191 //              if (is_Proj(irn) || be_is_Keep(irn))
1192 //                      continue;
1193
1194                 /* the precedence constraints */
1195                 ilp_timer_push(t_cst_prec);
1196                 bs_block_irns = bitset_clear_all(bs_block_irns);
1197
1198                 sta_collect_in_deps(irn, &deps);
1199                 foreach_ir_nodeset(&deps, pred, iter) {
1200                         unsigned             t_low, t_high, t;
1201                         be_ilpsched_irn_t    *pred_node;
1202                         ilpsched_node_attr_t *pna;
1203                         unsigned             delay;
1204
1205                         pred = skip_normal_Proj(env->arch_env, pred);
1206                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1207                                 continue;
1208
1209                         pred_node = get_ilpsched_irn(env, pred);
1210                         pna       = get_ilpsched_node_attr(pred_node);
1211
1212                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1213
1214                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1215                                 bitset_set(bs_block_irns, pna->block_idx);
1216                         else
1217                                 continue;
1218
1219                         /* irn = n, pred = m */
1220                         delay  = fixed_latency(env->sel, pred, env->block_env);
1221                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1222                         t_high = MIN(na->alap, pna->alap + delay - 1);
1223                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1224                                 unsigned tn, tm;
1225                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1226
1227                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1228                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1229                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1230                                 num_cst_prec++;
1231
1232                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1233                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1234                                         /* node variables have smaller index than pred variables */
1235                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1236                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1237                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1238                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1239                                                 }
1240                                         }
1241
1242                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1243                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1244                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1245                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1246                                                 }
1247                                         }
1248                                 }
1249                                 else {
1250                                         /* pred variables have smaller index than node variables */
1251                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1252                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1253                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1254                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1255                                                 }
1256                                         }
1257
1258                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1259                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1260                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1261                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1262                                                 }
1263                                         }
1264                                 }
1265
1266                                 if (ARR_LEN(tmp_var_idx) > 0)
1267                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1268
1269                                 DEL_ARR_F(tmp_var_idx);
1270                         }
1271                 }
1272                 ir_nodeset_destroy(&deps);
1273                 ilp_timer_pop();
1274         }
1275         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1276                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1277         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1278                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1279 }
1280
1281 /**
1282  * Create ILP resource constraints:
1283  * - assure that for each time step not more instructions are scheduled
1284  *   to the same unit types as units of this type are available
1285  */
1286 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1287         int                   glob_type_idx;
1288         char                  buf[1024];
1289         unsigned              num_cst_resrc = 0;
1290         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1291         ir_timer_t            *t_cst_rsrc   = ir_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1292
1293         ilp_timer_push(t_cst_rsrc);
1294         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1295                 unsigned                 t;
1296                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1297
1298                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1299
1300                 /* check each time step */
1301                 for (t = 0; t < ba->max_steps; ++t) {
1302                         ir_node *irn;
1303                         int     cst;
1304                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1305
1306                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1307                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1308                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1309                         num_cst_resrc++;
1310
1311                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1312                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1313                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1314                                 int                  tp_idx;
1315
1316                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1317
1318                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1319                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1320                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1321                                 }
1322                         }
1323
1324                         /* set constraints if we have some */
1325                         if (ARR_LEN(tmp_var_idx) > 0)
1326                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1327
1328                         DEL_ARR_F(tmp_var_idx);
1329                 }
1330         }
1331         ilp_timer_pop();
1332         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1333                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1334 }
1335
1336 /**
1337  * Create ILP bundle constraints:
1338  * - assure, at most bundle_size * bundles_per_cycle instructions
1339  *   can be started at a certain point.
1340  */
1341 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1342         char                  buf[1024];
1343         unsigned              t;
1344         unsigned              num_cst_bundle = 0;
1345         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1346         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1347         ir_timer_t            *t_cst_bundle  = ir_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1348
1349         ilp_timer_push(t_cst_bundle);
1350         for (t = 0; t < ba->max_steps; ++t) {
1351                 ir_node *irn;
1352                 int     cst;
1353                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1354
1355                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1356                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1357                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1358                 num_cst_bundle++;
1359
1360                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1361                         be_ilpsched_irn_t    *node;
1362                         ilpsched_node_attr_t *na;
1363                         int                  tp_idx;
1364
1365                         /* Projs and Keeps do not contribute to bundle size */
1366                         if (is_Proj(irn) || be_is_Keep(irn))
1367                                 continue;
1368
1369                         node = get_ilpsched_irn(env, irn);
1370                         na   = get_ilpsched_node_attr(node);
1371
1372                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1373                         if (na->is_dummy_node)
1374                                 continue;
1375
1376                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1377                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1378                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1379                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1380                                 }
1381                         }
1382                 }
1383
1384                 if (ARR_LEN(tmp_var_idx) > 0)
1385                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1386
1387                 DEL_ARR_F(tmp_var_idx);
1388         }
1389         ilp_timer_pop();
1390         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1391                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1392 }
1393
1394 /**
1395  * Create ILP alive nodes constraints:
1396  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1397  */
1398 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1399         char                  buf[1024];
1400         ir_node               *irn;
1401         unsigned              num_cst = 0;
1402         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1403         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1404
1405         ilp_timer_push(t_cst);
1406         /* for each node */
1407         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1408                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1409                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1410                 unsigned             t;
1411
1412                 /* we ignore nodes assigned to dummy unit here */
1413                 if (na->is_dummy_node)
1414                         continue;
1415
1416                 /* check check all time steps: asap(n) <= t <= U */
1417                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1418                         int node_tp_idx;
1419
1420                         /* for all unit types available for this node */
1421                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1422                                 unsigned tn, tn_max, idx;
1423                                 int      cst, i;
1424                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1425                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1426
1427                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1428                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1429                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1430                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1431                                 num_cst++;
1432
1433                                 tn_max = MIN(na->alap - 1, t);
1434                                 /* check if the node has been scheduled so far */
1435                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1436                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1437                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1438                                 }
1439
1440                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1441                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1442                                 DEL_ARR_F(tmp_var_idx_n);
1443
1444                                 /* subtract the number of consumer scheduled so far */
1445                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1446                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1447                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1448                                         int                  tp_idx;
1449                                         unsigned             tm, tm_max;
1450
1451                                         tm_max = MIN(ca->alap - 1, t);
1452                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1453                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1454                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1455                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1456                                                 }
1457                                         }
1458                                 }
1459
1460                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1461                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1462                                 DEL_ARR_F(tmp_var_idx_m);
1463
1464                                 /* -c * a_{nt}^k */
1465                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1466                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1467
1468                         }
1469                 }
1470         }
1471         ilp_timer_pop();
1472         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1473                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1474 }
1475
1476 /**
1477  * Create ILP alive nodes constraints for live-in nodes:
1478  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1479  */
1480 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1481         char                  buf[1024];
1482         ilp_livein_node_t     *livein;
1483         unsigned              num_cst = 0;
1484         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1485         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1486
1487         ilp_timer_push(t_cst);
1488         /* for each node */
1489         foreach_pset(ba->livein_nodes, livein) {
1490                 ir_node              *irn  = livein->irn;
1491                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1492                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1493                 unsigned             t;
1494
1495                 /* check check all time steps: 0 <= t < max_alive_steps */
1496                 for (t = 0; t < livein->max_alive_steps; ++t) {
1497                         int node_tp_idx;
1498
1499                         /* for all unit types available for this node */
1500                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1501                                 const ir_edge_t *edge;
1502                                 unsigned idx;
1503                                 int      cst, num_block_user;
1504                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1505
1506                                 /* check the number of consumer scheduled so far */
1507                                 num_block_user = 0;
1508                                 foreach_out_edge(irn, edge) {
1509                                         ir_node              *user = get_edge_src_irn(edge);
1510                                         be_ilpsched_irn_t    *cons;
1511                                         ilpsched_node_attr_t *ca;
1512                                         int                  tp_idx;
1513                                         unsigned             tm, tm_max;
1514
1515                                         /* check only users within current block */
1516                                         if (get_nodes_block(user) != block_node->irn)
1517                                                 continue;
1518
1519                                         num_block_user++;
1520                                         cons = get_ilpsched_irn(env, user);
1521                                         ca   = get_ilpsched_node_attr(cons);
1522
1523                                         tm_max = MIN(ca->alap - 1, t);
1524                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1525                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1526                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1527                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1528                                                 }
1529                                         }
1530                                 }
1531
1532                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1533                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1534                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1535                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1536                                 num_cst++;
1537
1538                                 /* sum(scheduled users) */
1539                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1540                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1541                                 DEL_ARR_F(tmp_var_idx_m);
1542
1543                                 /* + c * a_{nt}^k */
1544                                 idx = node_tp_idx * livein->max_alive_steps + t;
1545                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1546                         }
1547                 }
1548         }
1549         ilp_timer_pop();
1550         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1551                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1552 }
1553
1554 /**
1555  * Create ILP pressure constraints, based on alive nodes:
1556  * - add additional costs to objective function if a node is scheduled
1557  *   on a unit although all units of this type are currently occupied
1558  */
1559 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1560         char                  buf[1024];
1561         ir_node               *cur_irn;
1562         unsigned              num_cst = 0;
1563         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1564         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1565
1566         ilp_timer_push(t_cst);
1567         /* y_{nt}^k is set for each node and timestep and unit type */
1568         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1569                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1570                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1571                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1572                 int                  glob_type_idx;
1573
1574                 /* we ignore nodes assigned to DUMMY unit here */
1575                 if (cur_na->is_dummy_node)
1576                         continue;
1577
1578                 /* for all types */
1579                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1580                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1581                         int                      cur_tp_idx;
1582                         unsigned                 t;
1583
1584                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1585
1586                         /* check if node can be executed on this unit type */
1587                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1588                         if (cur_tp_idx < 0)
1589                                 continue;
1590
1591                         /* check all time_steps at which the current node can be scheduled */
1592                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1593                                 int     cst, y_idx;
1594                                 ir_node *irn;
1595                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1596                                 ilp_livein_node_t *livein;
1597
1598                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1599                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1600                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1601                                 num_cst++;
1602
1603                                 /* - accumulate all nodes alive at point t on unit type k */
1604                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1605                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1606                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1607                                         int                  a_idx, tp_idx;
1608
1609                                         /* check if node can be alive here */
1610                                         if (t < na->asap - 1)
1611                                                 continue;
1612
1613                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1614
1615                                         /* current type is not suitable */
1616                                         if (tp_idx < 0)
1617                                                 continue;
1618
1619                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1620                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1621                                 }
1622                                 /* do the same for livein nodes */
1623                                 foreach_pset(ba->livein_nodes, livein) {
1624                                         ir_node              *irn  = livein->irn;
1625                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1626                                         int                  a_idx, tp_idx;
1627
1628                                         /* check if node can be alive here */
1629                                         if (t >= livein->max_alive_steps)
1630                                                 continue;
1631
1632                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1633
1634                                         /* current type is not suitable */
1635                                         if (tp_idx < 0)
1636                                                 continue;
1637
1638                                         a_idx = tp_idx * livein->max_alive_steps + t;
1639                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1640                                 }
1641
1642                                 if (ARR_LEN(tmp_var_idx) > 0)
1643                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1644                                 DEL_ARR_F(tmp_var_idx);
1645
1646                                 /* - num_nodes * y_{nt}^k */
1647                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1648                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1649                         }
1650                 }
1651         }
1652         ilp_timer_pop();
1653         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1654                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1655 }
1656
1657 /**
1658  * Create ILP branch constraints:
1659  * Assure, alle nodes are scheduled prior to cfg op.
1660  */
1661 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1662         char                  buf[1024];
1663         ir_node               *cur_irn, *cfop;
1664         unsigned              num_cst          = 0;
1665         unsigned              num_non_branches = 0;
1666         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1667         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_branch", "create branch constraints");
1668
1669         ilp_timer_push(t_cst);
1670         cfop = NULL;
1671         /* determine number of non-branch nodes and the one and only branch node */
1672         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1673                 switch (get_irn_opcode(cur_irn)) {
1674                         case iro_Phi:
1675                         case iro_Start:
1676                         case iro_End:
1677                         case iro_Proj:
1678                         case iro_Bad:
1679                         case iro_Unknown:
1680                                 num_non_branches++;
1681                                 break;
1682                         default:
1683                                 if (is_cfop(cur_irn)) {
1684                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1685                                         cfop = cur_irn;
1686                                 }
1687                                 else {
1688                                         num_non_branches++;
1689                                 }
1690                                 break;
1691                 }
1692         }
1693
1694         if (cfop) {
1695                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1696                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1697                 unsigned t;
1698
1699                 /* for each time step */
1700                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1701                         int *non_branch_vars, *branch_vars;
1702                         int cst;
1703
1704                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1705                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1706                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1707                         num_cst++;
1708
1709                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1710                         non_branch_vars = NEW_ARR_F(int, 0);
1711                         branch_vars     = NEW_ARR_F(int, 0);
1712                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1713                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1714                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1715                                 int                  tp_idx;
1716
1717                                 if (cur_irn == cfop) {
1718                                         /* for all unit types available for this node */
1719                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1720                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1721                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1722                                         }
1723                                 }
1724                                 else {
1725                                         /* sum up all possible schedule points for this node upto current timestep */
1726                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1727                                                 unsigned tn;
1728                                                 unsigned tmax = MIN(t, na->alap - 1);
1729
1730                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1731                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1732                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1733                                                 }
1734                                         }
1735                                 }
1736
1737                         }
1738
1739                         if (ARR_LEN(non_branch_vars) > 0)
1740                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1741                         if (ARR_LEN(branch_vars) > 0)
1742                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1743
1744                         DEL_ARR_F(branch_vars);
1745                         DEL_ARR_F(non_branch_vars);
1746                 }
1747         }
1748         ilp_timer_pop();
1749         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1750                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1751 }
1752
1753 #if 0
1754 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1755         char                  buf[1024];
1756         ir_node               *irn;
1757         unsigned              num_cst = 0;
1758         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1759         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1760
1761         ilp_timer_push(t_cst);
1762         /* check all nodes */
1763         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1764                 be_ilpsched_irn_t    *node;
1765                 ilpsched_node_attr_t *na;
1766                 unsigned             t;
1767                 ir_node              **pk;
1768
1769                 /* only mode_T nodes can have Projs and Keeps assigned */
1770                 if (get_irn_mode(irn) != mode_T)
1771                         continue;
1772
1773                 node = get_ilpsched_irn(env, irn);
1774                 na   = get_ilpsched_node_attr(node);
1775
1776                 /* check if has some Projs and Keeps assigned */
1777                 if (! na->projkeeps)
1778                         continue;
1779
1780                 /* we can run only once over the queue, so preserve the nodes */
1781                 pk = NEW_ARR_F(ir_node *, 0);
1782                 while (! waitq_empty(na->projkeeps))
1783                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1784                 del_waitq(na->projkeeps);
1785                 na->projkeeps = NULL;
1786
1787                 /* for all time steps at which this node can be scheduled */
1788                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1789                         int cst, tp_idx, i;
1790                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1791
1792                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1793                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1794                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1795                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1796                         num_cst++;
1797
1798                         /* sum up scheduling variables for this time step */
1799                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1800                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1801                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1802                         }
1803
1804                         if (ARR_LEN(tmp_var_idx_n) > 0)
1805                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1806                         DEL_ARR_F(tmp_var_idx_n);
1807
1808                         /* subtract all Proj and Keep variables for this step */
1809                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1810                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1811                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1812                                 int                  pk_tp_idx;
1813
1814                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1815                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1816                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1817                                 }
1818                         }
1819                 }
1820         }
1821         ilp_timer_pop();
1822         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1823                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1824 }
1825 #endif /* if 0 */
1826
1827 /***************************************************
1828  *  _____ _      _____                    _
1829  * |_   _| |    |  __ \                  (_)
1830  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1831  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1832  *  _| |_| |____| |      | | | | | | (_| | | | | |
1833  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1834  *
1835  ***************************************************/
1836
1837 /**
1838  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1839  */
1840 static void create_ilp(ir_node *block, void *walk_env) {
1841         be_ilpsched_env_t     *env           = walk_env;
1842         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1843         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1844         FILE                  *logfile       = NULL;
1845         lpp_t                 *lpp           = NULL;
1846         int                   need_heur      = 0;
1847         struct obstack        var_obst;
1848         char                  name[1024];
1849
1850         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1851         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1852         DBG((env->dbg, 255, "=========================================\n\n"));
1853
1854         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1855                 block, ba->n_interesting_nodes, ba->max_steps));
1856
1857         /* notify backend and get block environment */
1858         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1859
1860         /* if we have less than two interesting nodes, there is no need to create the ILP */
1861         if (ba->n_interesting_nodes > 1) {
1862                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1863                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1864                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1865                 int    estimated_n_var = (int)((double)base_num * fact_var);
1866                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1867
1868                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1869                         estimated_n_var, estimated_n_cst));
1870                 (void) estimated_n_var;
1871
1872                 /* set up the LPP object */
1873                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1874
1875                 lpp = new_lpp_userdef(
1876                         (const char *)name,
1877                         lpp_minimize,
1878                         estimated_n_cst,     /* num vars */
1879                         estimated_n_cst + 1, /* num cst */
1880                         1.3);                /* grow factor */
1881                 obstack_init(&var_obst);
1882
1883                 /* create ILP variables */
1884                 create_variables(env, lpp, block_node, &var_obst);
1885
1886                 /* create ILP constraints */
1887                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1888                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1889                 create_ressource_constraints(env, lpp, block_node);
1890                 create_bundle_constraints(env, lpp, block_node);
1891                 create_branch_constraint(env, lpp, block_node);
1892                 //create_proj_keep_constraints(env, lpp, block_node);
1893
1894                 if (env->opts->regpress) {
1895                         create_alive_nodes_constraint(env, lpp, block_node);
1896                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1897                         create_pressure_alive_constraint(env, lpp, block_node);
1898                 }
1899
1900                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1901
1902                 /* debug stuff, dump lpp when debugging is on  */
1903                 DEBUG_ONLY({
1904                         if (firm_dbg_get_mask(env->dbg) > 1) {
1905                                 char buf[1024];
1906                                 FILE *f;
1907
1908                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1909                                 f = fopen(buf, "w");
1910                                 lpp_dump_plain(lpp, f);
1911                                 fclose(f);
1912                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1913                                 lpp_dump(lpp, buf);
1914                         }
1915                 })
1916
1917                 /* set solve time limit */
1918                 lpp_set_time_limit(lpp, env->opts->time_limit);
1919
1920                 /* set logfile if requested */
1921                 if (strlen(env->opts->log_file) > 0) {
1922                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1923                                 lpp_set_log(lpp, stdout);
1924                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1925                                 lpp_set_log(lpp, stderr);
1926                         else {
1927                                 logfile = fopen(env->opts->log_file, "w");
1928                                 if (! logfile)
1929                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1930                                 else
1931                                         lpp_set_log(lpp, logfile);
1932                         }
1933                 }
1934
1935                 /* solve the ILP */
1936                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1937
1938                 if (logfile)
1939                         fclose(logfile);
1940
1941                 /* check for valid solution */
1942                 if (! lpp_is_sol_valid(lpp)) {
1943                         DEBUG_ONLY({
1944                             char buf[1024];
1945                             FILE *f;
1946
1947                             if (firm_dbg_get_mask(env->dbg) >= 2) {
1948                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1949                               f = fopen(buf, "w");
1950                               lpp_dump_plain(lpp, f);
1951                               fclose(f);
1952                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1953                               lpp_dump(lpp, buf);
1954                               dump_ir_block_graph(env->irg, "-infeasible");
1955                            }
1956                         })
1957
1958                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1959                         need_heur = 1;
1960                 }
1961
1962                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1963                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1964                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1965                 DBG((env->dbg, LEVEL_1, "\tmatrix: %u elements, density %.2f%%, size %.2fMB\n", lpp->n_elems, lpp->density, (double)lpp->matrix_mem / 1024.0 / 1024.0));
1966                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1967                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1968                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1969                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1970
1971                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1972         }
1973
1974         /* apply solution */
1975         be_stat_ev("nodes", ba->block_last_idx);
1976         be_stat_ev("vars", lpp ? lpp->var_next : 0);
1977         be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1978         if (need_heur) {
1979                         be_stat_ev("time", -1);
1980                         be_stat_ev_dbl("opt", 0.0);
1981                 list_sched_single_block(env->birg, block, env->be_opts);
1982         }
1983         else {
1984                 if (lpp) {
1985                         double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1986                         be_stat_ev_dbl("time", lpp->sol_time);
1987                         be_stat_ev_dbl("opt", opt);
1988                 }
1989                 else {
1990                         be_stat_ev_dbl("time", 0.0);
1991                         be_stat_ev_dbl("opt", 100.0);
1992                 }
1993                 apply_solution(env, lpp, block);
1994         }
1995
1996         if (lpp)
1997                 free_lpp(lpp);
1998
1999         /* notify backend */
2000         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2001 }
2002
2003 /**
2004  * Perform ILP scheduling on the given irg.
2005  */
2006 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2007         be_ilpsched_env_t          env;
2008         const char                 *name     = "be ilp scheduling";
2009         ir_graph                   *irg      = be_get_birg_irg(birg);
2010         const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
2011         const ilp_sched_selector_t *sel      = arch_env->impl->get_ilp_sched_selector(arch_env);
2012
2013         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2014
2015         stat_ev_ctx_push("ilpsched");
2016
2017 //      firm_dbg_set_mask(env.dbg, 1);
2018
2019         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2020         env.sel        = sel;
2021         env.irg        = irg;
2022         env.height     = heights_new(irg);
2023         env.main_env   = birg->main_env;
2024         env.arch_env   = arch_env;
2025         env.cpu        = arch_env_get_machine(arch_env);
2026         env.opts       = &ilp_opts;
2027         env.birg       = birg;
2028         env.be_opts    = be_opts;
2029         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2030
2031         /* assign a unique per block number to all interesting nodes */
2032         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2033
2034         /*
2035                 The block indices are completely build after the walk,
2036                 now we can allocate the bitsets (size depends on block indices)
2037                 for all nodes.
2038         */
2039         phase_reinit_irn_data(&env.ph);
2040
2041         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2042         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2043
2044         /* Calculate ALAP of all irns */
2045         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2046
2047         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2048         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2049
2050         /* perform ILP scheduling */
2051         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2052
2053         DEBUG_ONLY(
2054                 if (firm_dbg_get_mask(env.dbg)) {
2055                         phase_stat_t stat;
2056                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2057
2058                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2059                 }
2060         );
2061
2062         /* free data allocated dynamically */
2063         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2064
2065         /* free all allocated object */
2066         phase_free(&env.ph);
2067         heights_free(env.height);
2068
2069         /* notify backend */
2070         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2071
2072         stat_ev_ctx_pop("ilpsched");
2073 }
2074
2075 /**
2076  * Register ILP scheduler options.
2077  */
2078 void be_init_ilpsched(void)
2079 {
2080         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2081         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2082
2083         lc_opt_add_table(sched_grp, ilpsched_option_table);
2084 }
2085
2086 #else /* WITH_ILP */
2087
2088 static inline void some_picky_compiler_do_not_allow_empty_files(void)
2089 {}
2090
2091 #endif /* WITH_ILP */