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 #ifdef HAVE_CONFIG_H
  33 #include "config.h"
  34 #endif
  35
  36 #include "firm_config.h"
  37
  38 #ifdef WITH_ILP
  39
  40 #include <math.h>
  41
  42 #ifndef _WIN32
  43 #include <strings.h>
  44 #endif /* _WIN32 */
  45
  46 #include "irnode_t.h"
  47 #include "irgwalk.h"
  48 #include "irbitset.h"
  49 #include "irphase_t.h"
  50 #include "height.h"
  51 #include "iredges.h"
  52 #include "pdeq.h"
  53 #include "debug.h"
  54 #include "irtools.h"
  55 #include "irdump.h"
  56 #include "irprintf.h"
  57 #include "plist.h"
  58 #include "irprintf.h"
  59 #include "timing.h"
  60
  61 #include <lpp/lpp.h>
  62 #include <lpp/lpp_net.h>
  63
  64 #include "lc_opts.h"
  65 #include "lc_opts_enum.h"
  66
  67 #include "be.h"
  68 #include "benode_t.h"
  69 #include "besched_t.h"
  70 #include "beilpsched.h"
  71 #include "beutil.h"
  72 #include "bestat.h"
  73 #include "beirg_t.h"
  74
  75 typedef struct _ilpsched_options_t {
  76         unsigned regpress;
  77         unsigned time_limit;
  78         char     log_file[1024];
  79 } ilpsched_options_t;
  80
  81 typedef struct _unit_type_info_t {
  82         int                            n_units;
  83         const be_execution_unit_type_t *tp;
  84 } unit_type_info_t;
  85
  86 /**
  87  * holding the ILP variables of the different types
  88  */
  89 typedef struct _ilp_var_types_t {
  90         int *x;   /* x_{nt}^k variables */
  91         int *a;   /* a_{nt}^k variables */
  92         int *y;   /* y_{nt}^k variables */
  93 } ilp_var_types_t;
  94
  95 /**
  96  * Holds alive variables for a node live-in to a block.
  97  */
  98 typedef struct _ilp_livein_node_t {
  99         ir_node  *irn;
 100         unsigned max_alive_steps;
 101         int      *a;
 102 } ilp_livein_node_t;
 103
 104 /* attributes for a node */
 105 typedef struct _ilpsched_node_attr_t {
 106         unsigned asap;                     /**< The ASAP scheduling control step */
 107         unsigned alap;                     /**< The ALAP scheduling control step */
 108         unsigned latency;                  /**< Latency of this node (needed for sorting) */
 109         unsigned sched_point;              /**< the step in which the node is finally scheduled */
 110         unsigned visit_idx;                /**< Index of the node having visited this node last */
 111         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
 112         unsigned n_consumer;               /**< Number of consumers */
 113         ir_node  **block_consumer;         /**< List of consumer being in the same block */
 114         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
 115         unsigned block_idx     : 30;       /**< A unique per block index */
 116         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
 117         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
 118         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
 119                                                                                         for ASAP, successors for ALAP */
 120         unsigned         n_unit_types;     /**< number of allowed execution unit types */
 121         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
 122         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
 123 } ilpsched_node_attr_t;
 124
 125 /* attributes for a block */
 126 typedef struct _ilpsched_block_attr_t {
 127         unsigned block_last_idx;        /**< The highest node index in block so far */
 128         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 129         unsigned max_steps;             /**< Upper bound for block execution */
 130         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 131         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 132         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 133 } ilpsched_block_attr_t;
 134
 135 typedef union _ilpsched_attr_ {
 136         ilpsched_node_attr_t  node_attr;
 137         ilpsched_block_attr_t block_attr;
 138 } ilpsched_attr_t;
 139
 140 /* A irn for the phase and it's attributes (either node or block) */
 141 typedef struct {
 142         const ir_node   *irn;
 143         ilpsched_attr_t attr;
 144 } be_ilpsched_irn_t;
 145
 146 /* The ILP scheduling environment */
 147 typedef struct {
 148         ir_phase             ph;            /**< The phase */
 149         ir_graph             *irg;          /**< The current irg */
 150         heights_t            *height;       /**< The heights object of the irg */
 151         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 152         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 153         const arch_env_t     *arch_env;
 154         const be_main_env_t  *main_env;
 155         const be_machine_t   *cpu;          /**< the current abstract machine */
 156         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 157         const be_irg_t       *birg;         /**< The birg object */
 158         be_options_t         *be_opts;      /**< backend options */
 159         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 160         DEBUG_ONLY(firm_dbg_module_t *dbg);
 161 } be_ilpsched_env_t;
 162
 163 /* convenience macros to handle phase irn data */
 164 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 165 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 166 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 167 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 168
 169 /* check if node is considered for ILP scheduling */
 170 #define consider_for_sched(env, irn) \
 171         (! (is_Block(irn)            ||  \
 172                 is_normal_Proj(env, irn) ||  \
 173                 is_Phi(irn)              ||  \
 174                 is_NoMem(irn)            ||  \
 175                 is_Unknown(irn)          ||  \
 176                 is_End(irn)                  \
 177                 ))
 178
 179 /* gives the valid scheduling time step interval for a node */
 180 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 181
 182 /* gives the valid interval where a node can die */
 183 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 184
 185 /* gives the corresponding ILP variable for given node, unit and time step */
 186 #define ILPVAR_IDX(na, unit, control_step) \
 187         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 188
 189 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 190 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 191         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 192
 193 /* check if a double value is within an epsilon environment of 0 */
 194 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 195
 196 #define ilp_timer_push(t)         ir_timer_push((t))
 197 #define ilp_timer_pop()           ir_timer_pop()
 198 #define ilp_timer_elapsed_usec(t) ir_timer_elapsed_usec((t))
 199
 200 /* option variable */
 201 static ilpsched_options_t ilp_opts = {
 202         1,     /* default is with register pressure constraints */
 203         300,   /* 300 sec per block time limit */
 204         ""     /* no log file */
 205 };
 206
 207 /* ILP options */
 208 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 209         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 210         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 211         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 212         LC_OPT_LAST
 213 };
 214
 215 /*
 216         We need this global variable as we compare nodes dependent on heights,
 217         but we cannot pass any information to the qsort compare function.
 218 */
 219 static heights_t *glob_heights;
 220
 221 /**
 222  * Check if irn is a Proj, which has no execution units assigned.
 223  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 224  */
 225 static INLINE int is_normal_Proj(const arch_env_t *env, const ir_node *irn) {
 226         return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
 227 }
 228
 229 /**
 230  * Skips normal Projs.
 231  * @return predecessor if irn is a normal Proj, otherwise irn.
 232  */
 233 static INLINE ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn) {
 234         if (is_normal_Proj(env, irn))
 235                 return get_Proj_pred(irn);
 236         return irn;
 237 }
 238
 239 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 240         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 241         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 242                 lat = 1;
 243         return lat;
 244 }
 245
 246 static int cmp_live_in_nodes(const void *a, const void *b) {
 247         const ilp_livein_node_t *n1 = a;
 248         const ilp_livein_node_t *n2 = b;
 249
 250         return n1->irn != n2->irn;
 251 }
 252
 253 /**
 254  * Compare scheduling time steps of two be_ilpsched_irn's.
 255  */
 256 static int cmp_ilpsched_irn(const void *a, const void *b) {
 257         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 258         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 259         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 260         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 261
 262         if (n1_a->sched_point == n2_a->sched_point) {
 263                 const ir_node *irn_a = n1->irn;
 264                 const ir_node *irn_b = n2->irn;
 265
 266                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 267                         return 1;
 268                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 269                         return -1;
 270
 271                 /*
 272                         Ok, timestep is equal and the nodes are parallel,
 273                         so check latency and schedule high latency first.
 274                 */
 275                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 276         }
 277         else
 278                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 279 }
 280
 281 /**
 282  * In case there is no phase information for irn, initialize it.
 283  */
 284 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old) {
 285         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 286
 287         if (res == old) {
 288                 /* if we have already some data: check for reinitialization */
 289
 290                 if (! is_Block(irn)) {
 291                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 292
 293                         if (! na->transitive_block_nodes) {
 294                                 ir_node               *block      = get_nodes_block(irn);
 295                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 296                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 297
 298                                 /* we are called after the block indices have been build: create bitset */
 299                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 300                         }
 301                         else {
 302                                 /* we are called from reinit block data: clear the bitset */
 303                                 bitset_clear_all(na->transitive_block_nodes);
 304                                 na->visit_idx    = 0;
 305                                 na->alap_changed = 1;
 306                         }
 307                 }
 308                 return old;
 309         }
 310
 311         res->irn = irn;
 312
 313         /* set ilpsched irn attributes (either block or irn) */
 314         if (is_Block(irn)) {
 315                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 316
 317                 ba->n_interesting_nodes = 0;
 318                 ba->block_last_idx      = 0;
 319                 ba->root_nodes          = plist_new();
 320                 ba->head_ilp_nodes      = NULL;
 321                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 322                 ba->max_steps           = 0;
 323         }
 324         else {
 325                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 326                 memset(na, 0, sizeof(*na));
 327         }
 328
 329         return res;
 330 }
 331
 332 /**
 333  * Assign a per block unique number to each node.
 334  */
 335 static void build_block_idx(ir_node *irn, void *walk_env) {
 336         be_ilpsched_env_t     *env = walk_env;
 337         be_ilpsched_irn_t     *node, *block_node;
 338         ilpsched_node_attr_t  *na;
 339         ilpsched_block_attr_t *ba;
 340
 341         set_irn_link(irn, NULL);
 342         if (! consider_for_sched(env->arch_env, irn))
 343                 return;
 344
 345         node       = get_ilpsched_irn(env, irn);
 346         na         = get_ilpsched_node_attr(node);
 347         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 348         ba         = get_ilpsched_block_attr(block_node);
 349
 350         na->block_idx = ba->block_last_idx++;
 351 }
 352
 353 /********************************************************
 354  *                              __        _
 355  *                             / /       | |
 356  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 357  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 358  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 359  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 360  *                 | |                           | |
 361  *                 |_|                           |_|
 362  ********************************************************/
 363
 364 /**
 365  * Add all nodes having no user in current block to last_nodes list.
 366  */
 367 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 368         ir_node               *block;
 369         const ir_edge_t       *edge;
 370         be_ilpsched_irn_t     *block_node, *node;
 371         ilpsched_block_attr_t *ba;
 372         ilpsched_node_attr_t  *na;
 373         int                   i, j;
 374         be_ilpsched_env_t     *env           = walk_env;
 375         int                   has_block_user = 0;
 376         unsigned              n_consumer     = 0;
 377         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 378         ir_node               **consumer;
 379         unsigned              idx;
 380
 381         if (! consider_for_sched(env->arch_env, irn))
 382                 return;
 383
 384         block    = get_nodes_block(irn);
 385     idx      = get_irn_idx(irn);
 386         consumer = NEW_ARR_F(ir_node *, 0);
 387
 388         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 389
 390         /* check data and dependency out edges */
 391         for (i = 0; i < 2 && ! has_block_user; ++i) {
 392                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 393                         ir_node *user = get_edge_src_irn(edge);
 394
 395                         if (is_normal_Proj(env->arch_env, user)) {
 396                                 const ir_edge_t *user_edge;
 397
 398                                 if (get_irn_mode(user) == mode_X)
 399                                         continue;
 400
 401                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 402                                 for (j = 0; j < 2; ++j) {
 403                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 404                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 405
 406                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 407                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 408                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 409
 410                                                         /* skip already visited nodes */
 411                                                         if (ua->consumer_idx == idx)
 412                                                                 continue;
 413
 414                                                         /* check if node has user in this block and collect the user if it's a data user */
 415                                                         if (get_nodes_block(real_user) == block) {
 416                                                                 if (i == 0 && j == 0)
 417                                                                         ARR_APP1(ir_node *, consumer, real_user);
 418                                                                 has_block_user = 1;
 419                                                         }
 420
 421                                                         /* only count data consumer */
 422                                                         if (i == 0)
 423                                                                 n_consumer++;
 424
 425                                                         /* mark user as visited by this node */
 426                                                         ua->consumer_idx = idx;
 427                                                 }
 428                                         }
 429                                 }
 430                         }
 431                         else if (is_Block(user)) {
 432                                 continue;
 433                         }
 434                         else if (! is_Phi(user)) {
 435                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 436                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 437
 438                                 /* skip already visited nodes */
 439                                 if (ua->consumer_idx == idx)
 440                                         continue;
 441
 442                                 /* check if node has user in this block and collect the user if it's a data user */
 443                                 if (get_nodes_block(user) == block) {
 444                                         if (i == 0)
 445                                                 ARR_APP1(ir_node *, consumer, user);
 446                                         has_block_user = 1;
 447                                 }
 448
 449                                 /* only count data consumer */
 450                                 if (i == 0)
 451                                         n_consumer++;
 452
 453                                 /* mark user visited by this node */
 454                                 ua->consumer_idx = idx;
 455                         }
 456                         else if (get_nodes_block(user) != block) {
 457                                 n_consumer++;
 458                         }
 459                 }
 460         }
 461
 462         block_node = get_ilpsched_irn(env, block);
 463         ba         = get_ilpsched_block_attr(block_node);
 464
 465         ba->n_interesting_nodes++;
 466
 467         /* current irn has no user inside this block, add to queue */
 468         if (! has_block_user) {
 469                 DB((env->dbg, LEVEL_3, "root node\n"));
 470                 plist_insert_back(ba->root_nodes, irn);
 471         }
 472         else {
 473                 DB((env->dbg, LEVEL_3, "normal node\n"));
 474         }
 475
 476         /* record number of all consumer and the consumer within the same block */
 477         node = get_ilpsched_irn(env, irn);
 478         na   = get_ilpsched_node_attr(node);
 479         na->n_consumer     = n_consumer;
 480         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 481         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 482         DEL_ARR_F(consumer);
 483 }
 484
 485 /**
 486  * Calculate the ASAP scheduling step for current irn.
 487  */
 488 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 489         be_ilpsched_env_t     *env = walk_env;
 490         int                   i;
 491         ir_node               *block;
 492         be_ilpsched_irn_t     *node, *block_node;
 493         ilpsched_node_attr_t  *na;
 494         ilpsched_block_attr_t *ba;
 495
 496         /* These nodes are handled separate */
 497         if (! consider_for_sched(env->arch_env, irn))
 498                 return;
 499
 500         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 501
 502         block    = get_nodes_block(irn);
 503         node     = get_ilpsched_irn(env, irn);
 504         na       = get_ilpsched_node_attr(node);
 505         na->asap = 1;
 506
 507         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 508                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
 509
 510                 /* check for greatest distance to top */
 511                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 512                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 513                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 514                         unsigned             lat;
 515
 516                         lat         = fixed_latency(env->sel, pred, env->block_env);
 517                         na->latency = lat;
 518                         na->asap    = MAX(na->asap, pna->asap + lat);
 519                 }
 520         }
 521
 522         /* add node to ILP node list and update max_steps */
 523         block_node = get_ilpsched_irn(env, block);
 524         ba         = get_ilpsched_block_attr(block_node);
 525
 526         set_irn_link(irn, ba->head_ilp_nodes);
 527         ba->head_ilp_nodes = irn;
 528         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 529
 530         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 531 }
 532
 533 /**
 534  * Calculate the ALAP scheduling step of all irns in current block.
 535  * Depends on max_steps being calculated.
 536  */
 537 static void calculate_block_alap(ir_node *block, void *walk_env) {
 538         be_ilpsched_env_t     *env        = walk_env;
 539         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 540         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 541         waitq                 *cur_queue  = new_waitq();
 542         plist_element_t       *el;
 543
 544         assert(is_Block(block));
 545
 546         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 547                 block, ba->n_interesting_nodes, ba->max_steps));
 548
 549         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 550         //phase_reinit_block_irn_data(&env->ph, block);
 551
 552         /* init start queue */
 553         foreach_plist(ba->root_nodes, el) {
 554                 waitq_put(cur_queue, plist_element_get_value(el));
 555         }
 556
 557         /* repeat until all nodes are processed */
 558         while (! waitq_empty(cur_queue)) {
 559                 waitq *next_queue = new_waitq();
 560
 561                 /* process all nodes in current step */
 562                 while (! waitq_empty(cur_queue)) {
 563                         ir_node              *cur_irn = waitq_get(cur_queue);
 564                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 565                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 566                         int                  i;
 567
 568                         /* cur_node has no alap set -> it's a root node, set to max alap */
 569                         if (na->alap == 0) {
 570                                 na->alap = ba->max_steps;
 571                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 572                                         cur_irn, na->alap));
 573                         }
 574                         else {
 575                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 576                                         cur_irn, na->alap));
 577                         }
 578
 579                         /* set the alap's of all predecessors */
 580                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 581                                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
 582
 583                                 /* check for greatest distance to bottom */
 584                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 585                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 586                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 587                                         unsigned             lat;
 588
 589                                         /* mark the predecessor as visited by current irn */
 590                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 591                                                 continue;
 592                                         pna->visit_idx = get_irn_idx(cur_irn);
 593
 594                                         lat = fixed_latency(env->sel, pred, env->block_env);
 595
 596                                         /* set ALAP of current pred */
 597                                         if (pna->alap == 0) {
 598                                                 /* current ALAP is 0: node has not yet been visited */
 599                                                 pna->alap_changed = 1;
 600                                                 pna->alap         = na->alap - lat;
 601                                         }
 602                                         else if (pna->alap > na->alap - lat) {
 603                                                 /* we found a longer path to root node: change ALAP */
 604                                                 pna->alap         = na->alap - lat;
 605                                                 pna->alap_changed = 1;
 606                                         }
 607                                         else {
 608                                                 /* current ALAP is best found so far: keep it */
 609                                                 pna->alap_changed = 0;
 610                                         }
 611
 612                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 613
 614                                         /* enqueue node for next iteration */
 615                                         if (get_irn_ins_or_deps(pred) > 0)
 616                                                 waitq_put(next_queue, pred);
 617                                 }
 618                         }
 619                 }
 620
 621                 /* prepare for next iteration */
 622                 del_waitq(cur_queue);
 623                 cur_queue = next_queue;
 624         }
 625 }
 626
 627 /**
 628  * Free list of root nodes and the set of live-in nodes.
 629  */
 630 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 631         be_ilpsched_env_t     *env        = walk_env;
 632         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 633         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 634
 635         plist_free(ba->root_nodes);
 636         ba->root_nodes = NULL;
 637         del_pset(ba->livein_nodes);
 638         ba->livein_nodes = NULL;
 639 }
 640
 641 /**
 642  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 643  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 644  */
 645 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 646         be_ilpsched_env_t    *env  = walk_env;
 647         ir_node              *pred = irn;
 648         be_ilpsched_irn_t    *node, *pred_node;
 649         ilpsched_node_attr_t *na, *pna;
 650
 651         if (! consider_for_sched(env->arch_env, irn))
 652                 return;
 653
 654         if (! is_Proj(irn) && ! be_is_Keep(irn))
 655                 return;
 656
 657         /* go to the ancestor */
 658         if (be_is_Keep(irn))
 659                 pred = get_irn_n(irn, 0);
 660         pred = skip_Proj(pred);
 661
 662         node      = get_ilpsched_irn(env, irn);
 663         pred_node = get_ilpsched_irn(env, pred);
 664         na        = get_ilpsched_node_attr(node);
 665         pna       = get_ilpsched_node_attr(pred_node);
 666
 667         na->asap = pna->asap;
 668         na->alap = pna->alap;
 669
 670         /* record all Projs and Keeps for this node */
 671         if (! pna->projkeeps)
 672                 pna->projkeeps = new_waitq();
 673         waitq_put(pna->projkeeps, irn);
 674
 675         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 676 }
 677
 678 /*******************************************
 679  *           _              _       _
 680  *          | |            | |     | |
 681  *  ___  ___| |__   ___  __| |_   _| | ___
 682  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 683  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 684  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 685  *
 686  *******************************************/
 687
 688 static INLINE void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn) {
 689         const ir_edge_t *edge;
 690         (void) block;
 691
 692         foreach_out_edge(irn, edge) {
 693                 ir_node *user = get_edge_src_irn(edge);
 694
 695                 if (be_is_Keep(user)) {
 696                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 697                         waitq_put(keeps, user);
 698                 }
 699         }
 700 }
 701
 702 /**
 703  * Inserts @p irn before @p before into schedule and notifies backend.
 704  */
 705 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 706         const ir_node *before, const ir_node *irn, unsigned cycle)
 707 {
 708         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 709         sched_add_before(before, irn);
 710 }
 711
 712 /**
 713  * Adds a node, it's Projs (in case of mode_T nodes) and
 714  * it's Keeps to schedule.
 715  */
 716 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle) {
 717         const ir_edge_t *edge;
 718         waitq           *keeps = new_waitq();
 719
 720         /* mode_M nodes are not scheduled */
 721         if (get_irn_mode(irn) == mode_M)
 722                 return;
 723
 724         if (! sched_is_scheduled(irn))
 725                 notified_sched_add_before(env, block, irn, cycle);
 726
 727         /* add Projs */
 728         if (get_irn_mode(irn) == mode_T) {
 729                 foreach_out_edge(irn, edge) {
 730                         ir_node *user = get_edge_src_irn(edge);
 731
 732                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 733                                 get_irn_n_edges(user) > 0)
 734                         {
 735                                 notified_sched_add_before(env, block, user, cycle);
 736                         }
 737
 738                         check_for_keeps(keeps, block, user);
 739                 }
 740         }
 741         else {
 742                 check_for_keeps(keeps, block, irn);
 743         }
 744
 745         /* add Keeps */
 746         while (! waitq_empty(keeps)) {
 747                 ir_node *keep = waitq_get(keeps);
 748                 if (! sched_is_scheduled(keep))
 749                         notified_sched_add_before(env, block, keep, cycle);
 750         }
 751
 752         del_waitq(keeps);
 753 }
 754
 755 /**
 756  * Schedule all nodes in the given block, according to the ILP solution.
 757  */
 758 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 759         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 760         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 761         sched_info_t          *info       = get_irn_sched_info(block);
 762         be_ilpsched_irn_t     **sched_nodes;
 763         unsigned              i, l;
 764         ir_node               *cfop, *irn;
 765         const ir_edge_t       *edge;
 766
 767         /* init block schedule list */
 768         INIT_LIST_HEAD(&info->list);
 769         info->scheduled = 1;
 770
 771         /* collect nodes and their scheduling time step */
 772         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 773         if (ba->n_interesting_nodes == 0) {
 774                 /* ignore */
 775         }
 776         else if (ba->n_interesting_nodes == 1) {
 777                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 778
 779                 /* add the single node */
 780                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 781         }
 782         else {
 783                 /* check all nodes for their positive solution */
 784                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 785                         be_ilpsched_irn_t    *node;
 786                         ilpsched_node_attr_t *na;
 787                         int                  tp_idx, found;
 788                         unsigned             cur_var, t;
 789
 790                         node    = get_ilpsched_irn(env, irn);
 791                         na      = get_ilpsched_node_attr(node);
 792                         cur_var = 0;
 793                         found   = 0;
 794
 795                         if (! na->is_dummy_node) {
 796                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 797                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 798                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 799
 800                                                 if (! LPP_VALUE_IS_0(cost)) {
 801                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
 802                                                         found = 1;
 803                                                 }
 804                                         }
 805                                 }
 806                         }
 807
 808                         found = 0;
 809                         /* go over all variables of a node until the non-zero one is found */
 810                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 811                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 812                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 813
 814                                         /* check, if variable is set to one (it's not zero then :) */
 815                                         if (! LPP_VALUE_IS_0(val)) {
 816                                                 na->sched_point = t;
 817                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 818                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 819                                                         irn, t, na->type_info[tp_idx].tp->name));
 820                                                 found = 1;
 821                                         }
 822                                 }
 823                         }
 824                 }
 825
 826                 glob_heights = env->height;
 827                 /* sort nodes ascending by scheduling time step */
 828                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 829         }
 830
 831         /* make all Phis ready and remember the single cf op */
 832         cfop = NULL;
 833         foreach_out_edge(block, edge) {
 834                 irn = get_edge_src_irn(edge);
 835
 836                 switch (get_irn_opcode(irn)) {
 837                         case iro_Phi:
 838                                 add_to_sched(env, block, irn, 0);
 839                                 break;
 840                         case iro_Start:
 841                         case iro_End:
 842                         case iro_Proj:
 843                         case iro_Bad:
 844                         case iro_Unknown:
 845                                 break;
 846                         default:
 847                                 if (is_cfop(irn)) {
 848                                         assert(cfop == NULL && "Highlander - there can be only one");
 849                                         cfop = irn;
 850                                 }
 851                         break;
 852                 }
 853         }
 854
 855         /* add all nodes from list */
 856         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 857                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 858                 if (sched_nodes[i]->irn != cfop)
 859                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 860         }
 861
 862         /* schedule control flow node if not already done */
 863         if (cfop && ! sched_is_scheduled(cfop))
 864                 add_to_sched(env, block, cfop, 0);
 865
 866         DEL_ARR_F(sched_nodes);
 867 }
 868
 869 /***************************************************************
 870  *   _____ _      _____     _____           _   _
 871  *  |_   _| |    |  __ \   / ____|         | | (_)
 872  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 873  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 874  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 875  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 876  *
 877  ***************************************************************/
 878
 879 /**
 880  * Check if node can be executed on given unit type.
 881  */
 882 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 883         int                  i;
 884         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 885
 886         for (i = na->n_unit_types - 1; i >= 0; --i) {
 887                 if (na->type_info[i].tp == tp)
 888                         return i;
 889         }
 890
 891         return -1;
 892 }
 893
 894 /************************************************
 895  *                   _       _     _
 896  *                  (_)     | |   | |
 897  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 898  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 899  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 900  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 901  *
 902  ************************************************/
 903
 904 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
 905         const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
 906         unsigned                  n_unit_types = 0;
 907         be_ilpsched_irn_t         *node;
 908         ilpsched_node_attr_t      *na;
 909         unsigned                  unit_idx, tp_idx;
 910
 911         /* count number of available unit types for this node */
 912         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 913                 /* just count */ ;
 914
 915         node = get_ilpsched_irn(env, irn);
 916         na   = get_ilpsched_node_attr(node);
 917
 918         if (! na->type_info) {
 919                 na->n_unit_types = n_unit_types;
 920                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 921
 922                 /* fill the type info array */
 923                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 924                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 925                                 /* beware: we also count number of available units here */
 926                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 927                                         na->is_dummy_node = 1;
 928                         }
 929
 930                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 931                         na->type_info[tp_idx].n_units = unit_idx;
 932                 }
 933         }
 934
 935         return n_unit_types;
 936 }
 937
 938 /**
 939  * Returns the largest alap time of a user of @p irn.
 940  * The user must be in block @p block.
 941  */
 942 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block) {
 943         const ir_edge_t *edge;
 944         unsigned        max_alap = 0;
 945
 946         foreach_out_edge(irn, edge) {
 947                 ir_node *user = get_edge_src_irn(edge);
 948
 949                 if (get_nodes_block(user) == block) {
 950                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 951                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 952
 953                         max_alap = MAX(max_alap, na->alap);
 954                 }
 955         }
 956
 957         assert(max_alap > 0);
 958         return max_alap;
 959 }
 960
 961 /**
 962  * Create the following variables:
 963  * - x_{nt}^k    binary     weigthed with: t
 964  *      node n is scheduled at time step t to unit type k
 965  * ==>> These variables represent the schedule
 966  *
 967  * - a_{nt}^k    binary     weighted with num_nodes
 968  *      node n is alive at time step t on unit type k
 969  *
 970  * - y_{nt}^k    continuous  weighted with: num_nodes^2
 971  *      register pressure over limit for unit type k
 972  * ==>> These variables represent the register pressure
 973  *
 974  */
 975 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 976         char                  buf[1024];
 977         ir_node               *irn;
 978         unsigned              num_block_var, num_nodes;
 979         ilp_livein_node_t     *livein;
 980         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 981         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 982         ir_timer_t            *t_var   = ir_timer_register("beilpsched_var", "create ilp variables");
 983
 984         ilp_timer_push(t_var);
 985         num_block_var = num_nodes = 0;
 986         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 987                 be_ilpsched_irn_t    *node;
 988                 ilpsched_node_attr_t *na;
 989                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
 990                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
 991                 int                  i;
 992
 993                 node         = get_ilpsched_irn(env, irn);
 994                 na           = get_ilpsched_node_attr(node);
 995                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
 996
 997                 /* allocate space for ilp variables */
 998                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 999                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
1000
1001                 /* we need these variables only for "real" nodes */
1002                 if (! na->is_dummy_node) {
1003                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1004                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1005
1006                         num_ad         = ba->max_steps - na->asap + 1;
1007                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1008                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1009                 }
1010
1011                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1012                         irn, na->asap, na->alap, na->n_unit_types));
1013
1014                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1015                 /* create variables */
1016                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1017                         unsigned t;
1018
1019                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1020                                 /* x_{nt}^k variables */
1021                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1022                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1023                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1024                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1025                                 /* variable counter */
1026                                 n_var++;
1027                                 num_block_var++;
1028
1029                                 if (! na->is_dummy_node) {
1030                                         /* y_{nt}^k variables */
1031                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1032                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1033                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1034                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1035
1036                                         /* variable counter */
1037                                         n_var++;
1038                                         num_block_var++;
1039                                 }
1040                         }
1041
1042                         /* a node can die at any step t: asap(n) <= t <= U */
1043                         if (! na->is_dummy_node) {
1044                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1045
1046                                         /* a_{nt}^k variables */
1047                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1048                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1049                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1050                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1051
1052                                         /* variable counter */
1053                                         n_var++;
1054                                         num_block_var++;
1055                                 }
1056                         }
1057
1058                         /* collect live-in nodes */
1059                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1060                                 ir_node *pred = get_irn_n(irn, i);
1061
1062                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1063                                         be_ilpsched_set_type_info(env, pred, var_obst);
1064                                         if (! na->is_dummy_node) {
1065                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1066                                                 entry->irn = pred;
1067                                                 entry->a   = NULL;
1068                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1069                                         }
1070                                 }
1071                         }
1072                 }
1073
1074                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1075                 num_nodes++;
1076         }
1077
1078         /* create alive variables a_{nt}^k for live-ins */
1079         foreach_pset(ba->livein_nodes, livein) {
1080                 be_ilpsched_irn_t    *node;
1081                 ilpsched_node_attr_t *na;
1082                 unsigned             tp_idx, var_idx;
1083                 ir_node              *irn;
1084
1085                 irn  = livein->irn;
1086                 node = get_ilpsched_irn(env, irn);
1087                 na   = get_ilpsched_node_attr(node);
1088
1089                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1090
1091                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1092                 var_idx   = 0;
1093
1094                 /* create variables */
1095                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1096                         unsigned t;
1097
1098                         for (t = 0; t < livein->max_alive_steps; ++t) {
1099                                 /* a_{nt}^k variables */
1100                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1101                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1102                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1103                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1104                                 num_block_var++;
1105                         }
1106                 }
1107         }
1108
1109         ilp_timer_pop();
1110         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1111                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1112 }
1113
1114 /*******************************************************
1115  *                      _             _       _
1116  *                     | |           (_)     | |
1117  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1118  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1119  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1120  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1121  *
1122  *******************************************************/
1123
1124 /**
1125  * Collect all operands and nodes @p irn depends on.
1126  * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1127  */
1128 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps) {
1129         int i;
1130
1131         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1132                 ir_node *p = get_irn_in_or_dep(irn, i);
1133
1134                 if (is_Proj(p)) {
1135                         const ir_edge_t *edge;
1136
1137                         p = get_Proj_pred(p);
1138                         foreach_out_edge(p, edge) {
1139                                 ir_node *src = get_edge_src_irn(edge);
1140                                 ir_nodeset_insert(deps, src);
1141                         }
1142                 }
1143                 else {
1144                         ir_nodeset_insert(deps, p);
1145                 }
1146         }
1147 }
1148
1149 /**
1150  * Create following ILP constraints:
1151  * - the assignment constraints:
1152  *     assure each node is executed once by exactly one (allowed) execution unit
1153  * - the dead node assignment constraints:
1154  *     assure a node can only die at most once
1155  * - the precedence constraints:
1156  *     assure that no data dependencies are violated
1157  */
1158 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1159         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1160         char                  buf[1024];
1161         ir_node               *irn;
1162         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1163         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1164         ir_timer_t            *t_cst_assign  = ir_timer_register("beilpsched_cst_assign", "create assignment constraints");
1165         ir_timer_t            *t_cst_prec    = ir_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1166
1167         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1168         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1169                 int                  cst, tp_idx;
1170                 unsigned             cur_var;
1171                 be_ilpsched_irn_t    *node;
1172                 ilpsched_node_attr_t *na;
1173                 ir_node              *pred;
1174                 ir_nodeset_t          deps;
1175                 ir_nodeset_iterator_t iter;
1176
1177                 ir_nodeset_init(&deps);
1178
1179                 node    = get_ilpsched_irn(env, irn);
1180                 na      = get_ilpsched_node_attr(node);
1181                 cur_var = 0;
1182
1183                 /* the assignment constraint */
1184                 ilp_timer_push(t_cst_assign);
1185                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1186                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1187                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1188                 num_cst_assign++;
1189
1190                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1191                 ilp_timer_pop();
1192
1193                 /* We have separate constraints for Projs and Keeps */
1194                 // ILP becomes infeasible ?!?
1195 //              if (is_Proj(irn) || be_is_Keep(irn))
1196 //                      continue;
1197
1198                 /* the precedence constraints */
1199                 ilp_timer_push(t_cst_prec);
1200                 bs_block_irns = bitset_clear_all(bs_block_irns);
1201
1202                 sta_collect_in_deps(irn, &deps);
1203                 foreach_ir_nodeset(&deps, pred, iter) {
1204                         unsigned             t_low, t_high, t;
1205                         be_ilpsched_irn_t    *pred_node;
1206                         ilpsched_node_attr_t *pna;
1207                         unsigned             delay;
1208
1209                         pred = skip_normal_Proj(env->arch_env, pred);
1210                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1211                                 continue;
1212
1213                         pred_node = get_ilpsched_irn(env, pred);
1214                         pna       = get_ilpsched_node_attr(pred_node);
1215
1216                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1217
1218                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1219                                 bitset_set(bs_block_irns, pna->block_idx);
1220                         else
1221                                 continue;
1222
1223                         /* irn = n, pred = m */
1224                         delay  = fixed_latency(env->sel, pred, env->block_env);
1225                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1226                         t_high = MIN(na->alap, pna->alap + delay - 1);
1227                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1228                                 unsigned tn, tm;
1229                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1230
1231                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1232                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1233                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1234                                 num_cst_prec++;
1235
1236                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1237                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1238                                         /* node variables have smaller index than pred variables */
1239                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1240                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1241                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1242                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1243                                                 }
1244                                         }
1245
1246                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1247                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1248                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1249                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1250                                                 }
1251                                         }
1252                                 }
1253                                 else {
1254                                         /* pred variables have smaller index than node variables */
1255                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1256                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1257                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1258                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1259                                                 }
1260                                         }
1261
1262                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1263                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1264                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1265                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1266                                                 }
1267                                         }
1268                                 }
1269
1270                                 if (ARR_LEN(tmp_var_idx) > 0)
1271                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1272
1273                                 DEL_ARR_F(tmp_var_idx);
1274                         }
1275                 }
1276                 ir_nodeset_destroy(&deps);
1277                 ilp_timer_pop();
1278         }
1279         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1280                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1281         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1282                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1283 }
1284
1285 /**
1286  * Create ILP resource constraints:
1287  * - assure that for each time step not more instructions are scheduled
1288  *   to the same unit types as units of this type are available
1289  */
1290 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1291         int                   glob_type_idx;
1292         char                  buf[1024];
1293         unsigned              num_cst_resrc = 0;
1294         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1295         ir_timer_t            *t_cst_rsrc   = ir_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1296
1297         ilp_timer_push(t_cst_rsrc);
1298         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1299                 unsigned                 t;
1300                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1301
1302                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1303
1304                 /* check each time step */
1305                 for (t = 0; t < ba->max_steps; ++t) {
1306                         ir_node *irn;
1307                         int     cst;
1308                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1309
1310                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1311                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1312                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1313                         num_cst_resrc++;
1314
1315                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1316                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1317                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1318                                 int                  tp_idx;
1319
1320                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1321
1322                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1323                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1324                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1325                                 }
1326                         }
1327
1328                         /* set constraints if we have some */
1329                         if (ARR_LEN(tmp_var_idx) > 0)
1330                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1331
1332                         DEL_ARR_F(tmp_var_idx);
1333                 }
1334         }
1335         ilp_timer_pop();
1336         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1337                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1338 }
1339
1340 /**
1341  * Create ILP bundle constraints:
1342  * - assure, at most bundle_size * bundles_per_cycle instructions
1343  *   can be started at a certain point.
1344  */
1345 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1346         char                  buf[1024];
1347         unsigned              t;
1348         unsigned              num_cst_bundle = 0;
1349         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1350         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1351         ir_timer_t            *t_cst_bundle  = ir_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1352
1353         ilp_timer_push(t_cst_bundle);
1354         for (t = 0; t < ba->max_steps; ++t) {
1355                 ir_node *irn;
1356                 int     cst;
1357                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1358
1359                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1360                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1361                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1362                 num_cst_bundle++;
1363
1364                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1365                         be_ilpsched_irn_t    *node;
1366                         ilpsched_node_attr_t *na;
1367                         int                  tp_idx;
1368
1369                         /* Projs and Keeps do not contribute to bundle size */
1370                         if (is_Proj(irn) || be_is_Keep(irn))
1371                                 continue;
1372
1373                         node = get_ilpsched_irn(env, irn);
1374                         na   = get_ilpsched_node_attr(node);
1375
1376                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1377                         if (na->is_dummy_node)
1378                                 continue;
1379
1380                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1381                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1382                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1383                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1384                                 }
1385                         }
1386                 }
1387
1388                 if (ARR_LEN(tmp_var_idx) > 0)
1389                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1390
1391                 DEL_ARR_F(tmp_var_idx);
1392         }
1393         ilp_timer_pop();
1394         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1395                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1396 }
1397
1398 /**
1399  * Create ILP alive nodes constraints:
1400  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1401  */
1402 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1403         char                  buf[1024];
1404         ir_node               *irn;
1405         unsigned              num_cst = 0;
1406         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1407         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1408
1409         ilp_timer_push(t_cst);
1410         /* for each node */
1411         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1412                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1413                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1414                 unsigned             t;
1415
1416                 /* we ignore nodes assigned to dummy unit here */
1417                 if (na->is_dummy_node)
1418                         continue;
1419
1420                 /* check check all time steps: asap(n) <= t <= U */
1421                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1422                         int node_tp_idx;
1423
1424                         /* for all unit types available for this node */
1425                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1426                                 unsigned tn, tn_max, idx;
1427                                 int      cst, i;
1428                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1429                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1430
1431                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1432                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1433                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1434                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1435                                 num_cst++;
1436
1437                                 tn_max = MIN(na->alap - 1, t);
1438                                 /* check if the node has been scheduled so far */
1439                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1440                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1441                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1442                                 }
1443
1444                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1445                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1446                                 DEL_ARR_F(tmp_var_idx_n);
1447
1448                                 /* subtract the number of consumer scheduled so far */
1449                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1450                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1451                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1452                                         int                  tp_idx;
1453                                         unsigned             tm, tm_max;
1454
1455                                         tm_max = MIN(ca->alap - 1, t);
1456                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1457                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1458                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1459                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1460                                                 }
1461                                         }
1462                                 }
1463
1464                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1465                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1466                                 DEL_ARR_F(tmp_var_idx_m);
1467
1468                                 /* -c * a_{nt}^k */
1469                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1470                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1471
1472                         }
1473                 }
1474         }
1475         ilp_timer_pop();
1476         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1477                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1478 }
1479
1480 /**
1481  * Create ILP alive nodes constraints for live-in nodes:
1482  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1483  */
1484 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1485         char                  buf[1024];
1486         ilp_livein_node_t     *livein;
1487         unsigned              num_cst = 0;
1488         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1489         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1490
1491         ilp_timer_push(t_cst);
1492         /* for each node */
1493         foreach_pset(ba->livein_nodes, livein) {
1494                 ir_node              *irn  = livein->irn;
1495                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1496                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1497                 unsigned             t;
1498
1499                 /* check check all time steps: 0 <= t < max_alive_steps */
1500                 for (t = 0; t < livein->max_alive_steps; ++t) {
1501                         int node_tp_idx;
1502
1503                         /* for all unit types available for this node */
1504                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1505                                 const ir_edge_t *edge;
1506                                 unsigned idx;
1507                                 int      cst, num_block_user;
1508                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1509
1510                                 /* check the number of consumer scheduled so far */
1511                                 num_block_user = 0;
1512                                 foreach_out_edge(irn, edge) {
1513                                         ir_node              *user = get_edge_src_irn(edge);
1514                                         be_ilpsched_irn_t    *cons;
1515                                         ilpsched_node_attr_t *ca;
1516                                         int                  tp_idx;
1517                                         unsigned             tm, tm_max;
1518
1519                                         /* check only users within current block */
1520                                         if (get_nodes_block(user) != block_node->irn)
1521                                                 continue;
1522
1523                                         num_block_user++;
1524                                         cons = get_ilpsched_irn(env, user);
1525                                         ca   = get_ilpsched_node_attr(cons);
1526
1527                                         tm_max = MIN(ca->alap - 1, t);
1528                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1529                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1530                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1531                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1532                                                 }
1533                                         }
1534                                 }
1535
1536                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1537                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1538                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1539                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1540                                 num_cst++;
1541
1542                                 /* sum(scheduled users) */
1543                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1544                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1545                                 DEL_ARR_F(tmp_var_idx_m);
1546
1547                                 /* + c * a_{nt}^k */
1548                                 idx = node_tp_idx * livein->max_alive_steps + t;
1549                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1550                         }
1551                 }
1552         }
1553         ilp_timer_pop();
1554         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1555                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1556 }
1557
1558 /**
1559  * Create ILP pressure constraints, based on alive nodes:
1560  * - add additional costs to objective function if a node is scheduled
1561  *   on a unit although all units of this type are currently occupied
1562  */
1563 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1564         char                  buf[1024];
1565         ir_node               *cur_irn;
1566         unsigned              num_cst = 0;
1567         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1568         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1569
1570         ilp_timer_push(t_cst);
1571         /* y_{nt}^k is set for each node and timestep and unit type */
1572         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1573                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1574                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1575                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1576                 int                  glob_type_idx;
1577
1578                 /* we ignore nodes assigned to DUMMY unit here */
1579                 if (cur_na->is_dummy_node)
1580                         continue;
1581
1582                 /* for all types */
1583                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1584                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1585                         int                      cur_tp_idx;
1586                         unsigned                 t;
1587
1588                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1589
1590                         /* check if node can be executed on this unit type */
1591                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1592                         if (cur_tp_idx < 0)
1593                                 continue;
1594
1595                         /* check all time_steps at which the current node can be scheduled */
1596                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1597                                 int     cst, y_idx;
1598                                 ir_node *irn;
1599                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1600                                 ilp_livein_node_t *livein;
1601
1602                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1603                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1604                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1605                                 num_cst++;
1606
1607                                 /* - accumulate all nodes alive at point t on unit type k */
1608                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1609                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1610                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1611                                         int                  a_idx, tp_idx;
1612
1613                                         /* check if node can be alive here */
1614                                         if (t < na->asap - 1)
1615                                                 continue;
1616
1617                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1618
1619                                         /* current type is not suitable */
1620                                         if (tp_idx < 0)
1621                                                 continue;
1622
1623                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1624                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1625                                 }
1626                                 /* do the same for livein nodes */
1627                                 foreach_pset(ba->livein_nodes, livein) {
1628                                         ir_node              *irn  = livein->irn;
1629                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1630                                         int                  a_idx, tp_idx;
1631
1632                                         /* check if node can be alive here */
1633                                         if (t >= livein->max_alive_steps)
1634                                                 continue;
1635
1636                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1637
1638                                         /* current type is not suitable */
1639                                         if (tp_idx < 0)
1640                                                 continue;
1641
1642                                         a_idx = tp_idx * livein->max_alive_steps + t;
1643                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1644                                 }
1645
1646                                 if (ARR_LEN(tmp_var_idx) > 0)
1647                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1648                                 DEL_ARR_F(tmp_var_idx);
1649
1650                                 /* - num_nodes * y_{nt}^k */
1651                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1652                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1653                         }
1654                 }
1655         }
1656         ilp_timer_pop();
1657         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1658                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1659 }
1660
1661 /**
1662  * Create ILP branch constraints:
1663  * Assure, alle nodes are scheduled prior to cfg op.
1664  */
1665 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1666         char                  buf[1024];
1667         ir_node               *cur_irn, *cfop;
1668         unsigned              num_cst          = 0;
1669         unsigned              num_non_branches = 0;
1670         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1671         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_branch", "create branch constraints");
1672
1673         ilp_timer_push(t_cst);
1674         cfop = NULL;
1675         /* determine number of non-branch nodes and the one and only branch node */
1676         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1677                 switch (get_irn_opcode(cur_irn)) {
1678                         case iro_Phi:
1679                         case iro_Start:
1680                         case iro_End:
1681                         case iro_Proj:
1682                         case iro_Bad:
1683                         case iro_Unknown:
1684                                 num_non_branches++;
1685                                 break;
1686                         default:
1687                                 if (is_cfop(cur_irn)) {
1688                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1689                                         cfop = cur_irn;
1690                                 }
1691                                 else {
1692                                         num_non_branches++;
1693                                 }
1694                                 break;
1695                 }
1696         }
1697
1698         if (cfop) {
1699                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1700                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1701                 unsigned t;
1702
1703                 /* for each time step */
1704                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1705                         int *non_branch_vars, *branch_vars;
1706                         int cst;
1707
1708                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1709                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1710                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1711                         num_cst++;
1712
1713                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1714                         non_branch_vars = NEW_ARR_F(int, 0);
1715                         branch_vars     = NEW_ARR_F(int, 0);
1716                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1717                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1718                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1719                                 int                  tp_idx;
1720
1721                                 if (cur_irn == cfop) {
1722                                         /* for all unit types available for this node */
1723                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1724                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1725                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1726                                         }
1727                                 }
1728                                 else {
1729                                         /* sum up all possible schedule points for this node upto current timestep */
1730                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1731                                                 unsigned tn;
1732                                                 unsigned tmax = MIN(t, na->alap - 1);
1733
1734                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1735                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1736                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1737                                                 }
1738                                         }
1739                                 }
1740
1741                         }
1742
1743                         if (ARR_LEN(non_branch_vars) > 0)
1744                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1745                         if (ARR_LEN(branch_vars) > 0)
1746                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1747
1748                         DEL_ARR_F(branch_vars);
1749                         DEL_ARR_F(non_branch_vars);
1750                 }
1751         }
1752         ilp_timer_pop();
1753         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1754                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1755 }
1756
1757 #if 0
1758 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1759         char                  buf[1024];
1760         ir_node               *irn;
1761         unsigned              num_cst = 0;
1762         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1763         ir_timer_t            *t_cst  = ir_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1764
1765         ilp_timer_push(t_cst);
1766         /* check all nodes */
1767         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1768                 be_ilpsched_irn_t    *node;
1769                 ilpsched_node_attr_t *na;
1770                 unsigned             t;
1771                 ir_node              **pk;
1772
1773                 /* only mode_T nodes can have Projs and Keeps assigned */
1774                 if (get_irn_mode(irn) != mode_T)
1775                         continue;
1776
1777                 node = get_ilpsched_irn(env, irn);
1778                 na   = get_ilpsched_node_attr(node);
1779
1780                 /* check if has some Projs and Keeps assigned */
1781                 if (! na->projkeeps)
1782                         continue;
1783
1784                 /* we can run only once over the queue, so preserve the nodes */
1785                 pk = NEW_ARR_F(ir_node *, 0);
1786                 while (! waitq_empty(na->projkeeps))
1787                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1788                 del_waitq(na->projkeeps);
1789                 na->projkeeps = NULL;
1790
1791                 /* for all time steps at which this node can be scheduled */
1792                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1793                         int cst, tp_idx, i;
1794                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1795
1796                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1797                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1798                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1799                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1800                         num_cst++;
1801
1802                         /* sum up scheduling variables for this time step */
1803                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1804                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1805                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1806                         }
1807
1808                         if (ARR_LEN(tmp_var_idx_n) > 0)
1809                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1810                         DEL_ARR_F(tmp_var_idx_n);
1811
1812                         /* subtract all Proj and Keep variables for this step */
1813                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1814                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1815                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1816                                 int                  pk_tp_idx;
1817
1818                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1819                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1820                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1821                                 }
1822                         }
1823                 }
1824         }
1825         ilp_timer_pop();
1826         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1827                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1828 }
1829 #endif /* if 0 */
1830
1831 /***************************************************
1832  *  _____ _      _____                    _
1833  * |_   _| |    |  __ \                  (_)
1834  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1835  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1836  *  _| |_| |____| |      | | | | | | (_| | | | | |
1837  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1838  *
1839  ***************************************************/
1840
1841 /**
1842  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1843  */
1844 static void create_ilp(ir_node *block, void *walk_env) {
1845         be_ilpsched_env_t     *env           = walk_env;
1846         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1847         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1848         FILE                  *logfile       = NULL;
1849         lpp_t                 *lpp           = NULL;
1850         int                   need_heur      = 0;
1851         struct obstack        var_obst;
1852         char                  name[1024];
1853
1854         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1855         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1856         DBG((env->dbg, 255, "=========================================\n\n"));
1857
1858         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1859                 block, ba->n_interesting_nodes, ba->max_steps));
1860
1861         /* notify backend and get block environment */
1862         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1863
1864         /* if we have less than two interesting nodes, there is no need to create the ILP */
1865         if (ba->n_interesting_nodes > 1) {
1866                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1867                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1868                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1869                 int    estimated_n_var = (int)((double)base_num * fact_var);
1870                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1871
1872                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1873                         estimated_n_var, estimated_n_cst));
1874                 (void) estimated_n_var;
1875
1876                 /* set up the LPP object */
1877                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1878
1879                 lpp = new_lpp_userdef(
1880                         (const char *)name,
1881                         lpp_minimize,
1882                         estimated_n_cst,     /* num vars */
1883                         estimated_n_cst + 1, /* num cst */
1884                         1.3);                /* grow factor */
1885                 obstack_init(&var_obst);
1886
1887                 /* create ILP variables */
1888                 create_variables(env, lpp, block_node, &var_obst);
1889
1890                 /* create ILP constraints */
1891                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1892                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1893                 create_ressource_constraints(env, lpp, block_node);
1894                 create_bundle_constraints(env, lpp, block_node);
1895                 create_branch_constraint(env, lpp, block_node);
1896                 //create_proj_keep_constraints(env, lpp, block_node);
1897
1898                 if (env->opts->regpress) {
1899                         create_alive_nodes_constraint(env, lpp, block_node);
1900                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1901                         create_pressure_alive_constraint(env, lpp, block_node);
1902                 }
1903
1904                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1905
1906                 /* debug stuff, dump lpp when debugging is on  */
1907                 DEBUG_ONLY({
1908                         if (firm_dbg_get_mask(env->dbg) > 1) {
1909                                 char buf[1024];
1910                                 FILE *f;
1911
1912                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1913                                 f = fopen(buf, "w");
1914                                 lpp_dump_plain(lpp, f);
1915                                 fclose(f);
1916                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1917                                 lpp_dump(lpp, buf);
1918                         }
1919                 })
1920
1921                 /* set solve time limit */
1922                 lpp_set_time_limit(lpp, env->opts->time_limit);
1923
1924                 /* set logfile if requested */
1925                 if (strlen(env->opts->log_file) > 0) {
1926                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1927                                 lpp_set_log(lpp, stdout);
1928                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1929                                 lpp_set_log(lpp, stderr);
1930                         else {
1931                                 logfile = fopen(env->opts->log_file, "w");
1932                                 if (! logfile)
1933                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1934                                 else
1935                                         lpp_set_log(lpp, logfile);
1936                         }
1937                 }
1938
1939                 /* solve the ILP */
1940                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1941
1942                 if (logfile)
1943                         fclose(logfile);
1944
1945                 /* check for valid solution */
1946                 if (! lpp_is_sol_valid(lpp)) {
1947                         DEBUG_ONLY({
1948                             char buf[1024];
1949                             FILE *f;
1950
1951                             if (firm_dbg_get_mask(env->dbg) >= 2) {
1952                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1953                               f = fopen(buf, "w");
1954                               lpp_dump_plain(lpp, f);
1955                               fclose(f);
1956                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1957                               lpp_dump(lpp, buf);
1958                               dump_ir_block_graph(env->irg, "-infeasible");
1959                            }
1960                         })
1961
1962                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1963                         need_heur = 1;
1964                 }
1965
1966                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1967                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1968                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1969                 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));
1970                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1971                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1972                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1973                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1974
1975                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1976         }
1977
1978         /* apply solution */
1979         be_stat_ev("nodes", ba->block_last_idx);
1980         be_stat_ev("vars", lpp ? lpp->var_next : 0);
1981         be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1982         if (need_heur) {
1983                         be_stat_ev("time", -1);
1984                         be_stat_ev_dbl("opt", 0.0);
1985                 list_sched_single_block(env->birg, block, env->be_opts);
1986         }
1987         else {
1988                 if (lpp) {
1989                         double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1990                         be_stat_ev_dbl("time", lpp->sol_time);
1991                         be_stat_ev_dbl("opt", opt);
1992                 }
1993                 else {
1994                         be_stat_ev_dbl("time", 0.0);
1995                         be_stat_ev_dbl("opt", 100.0);
1996                 }
1997                 apply_solution(env, lpp, block);
1998         }
1999
2000         if (lpp)
2001                 free_lpp(lpp);
2002
2003         /* notify backend */
2004         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2005 }
2006
2007 /**
2008  * Perform ILP scheduling on the given irg.
2009  */
2010 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2011         be_ilpsched_env_t          env;
2012         const char                 *name     = "be ilp scheduling";
2013         ir_graph                   *irg      = be_get_birg_irg(birg);
2014         const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
2015         const ilp_sched_selector_t *sel      = arch_env->impl->get_ilp_sched_selector(arch_env);
2016
2017         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2018
2019         stat_ev_ctx_push("ilpsched");
2020
2021 //      firm_dbg_set_mask(env.dbg, 1);
2022
2023         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2024         env.sel        = sel;
2025         env.irg        = irg;
2026         env.height     = heights_new(irg);
2027         env.main_env   = birg->main_env;
2028         env.arch_env   = arch_env;
2029         env.cpu        = arch_env_get_machine(arch_env);
2030         env.opts       = &ilp_opts;
2031         env.birg       = birg;
2032         env.be_opts    = be_opts;
2033         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2034
2035         /* assign a unique per block number to all interesting nodes */
2036         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2037
2038         /*
2039                 The block indices are completely build after the walk,
2040                 now we can allocate the bitsets (size depends on block indices)
2041                 for all nodes.
2042         */
2043         phase_reinit_irn_data(&env.ph);
2044
2045         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2046         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2047
2048         /* Calculate ALAP of all irns */
2049         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2050
2051         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2052         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2053
2054         /* perform ILP scheduling */
2055         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2056
2057         DEBUG_ONLY(
2058                 if (firm_dbg_get_mask(env.dbg)) {
2059                         phase_stat_t stat;
2060                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2061
2062                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2063                 }
2064         );
2065
2066         /* free data allocated dynamically */
2067         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2068
2069         /* free all allocated object */
2070         phase_free(&env.ph);
2071         heights_free(env.height);
2072
2073         /* notify backend */
2074         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2075
2076         stat_ev_ctx_pop("ilpsched");
2077 }
2078
2079 /**
2080  * Register ILP scheduler options.
2081  */
2082 void be_init_ilpsched(void)
2083 {
2084         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2085         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2086
2087         lc_opt_add_table(sched_grp, ilpsched_option_table);
2088 }
2089
2090 #else /* WITH_ILP */
2091
2092 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2093 {}
2094
2095 #endif /* WITH_ILP */