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

   1 /**
   2  * Scheduling algorithms.
   3  * An ILP scheduler based on
   4  * "ILP-based Instruction Scheduling for IA-64"
   5  * by Daniel Kaestner and Sebastian Winkel
   6  *
   7  * @date   22.10.2005
   8  * @author Christian Wuerdig
   9  * @cvs-id $Id$
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #ifndef _WIN32
  20 #include <strings.h>
  21 #endif /* _WIN32 */
  22
  23 #include "irnode_t.h"
  24 #include "irgwalk.h"
  25 #include "irbitset.h"
  26 #include "irphase_t.h"
  27 #include "height.h"
  28 #include "iredges.h"
  29 #include "pdeq.h"
  30 #include "debug.h"
  31 #include "irtools.h"
  32 #include "irdump.h"
  33 #include "plist.h"
  34
  35 #include <lpp/lpp.h>
  36 #include <lpp/lpp_net.h>
  37
  38 #include <libcore/lc_opts.h>
  39 #include <libcore/lc_opts_enum.h>
  40 #include <libcore/lc_timing.h>
  41
  42 #include "be.h"
  43 #include "benode_t.h"
  44 #include "besched_t.h"
  45 #include "beilpsched.h"
  46 #include "beutil.h"
  47
  48 typedef struct _ilpsched_options_t {
  49         unsigned regpress;
  50         unsigned limit_dead;
  51         unsigned time_limit;
  52         char     log_file[1024];
  53 } ilpsched_options_t;
  54
  55 typedef struct _unit_type_info_t {
  56         int                            n_units;
  57         const be_execution_unit_type_t *tp;
  58 } unit_type_info_t;
  59
  60 /**
  61  * holding the ILP variables of the different types
  62  */
  63 typedef struct _ilp_var_types_t {
  64         int *x;   /* x_{nt}^k variables */
  65         int *a;   /* a_{nt}^k variables */
  66         int *d;   /* d_{nt}^k variables */
  67         int *y;   /* y_{nt}^k variables */
  68 } ilp_var_types_t;
  69
  70 /**
  71  * Holds alive variables for a node live-in to a block.
  72  */
  73 typedef struct _ilp_livein_node_t {
  74         ir_node  *irn;
  75         unsigned max_alive_steps;
  76         int      *a;
  77 } ilp_livein_node_t;
  78
  79 /* attributes for a node */
  80 typedef struct _ilpsched_node_attr_t {
  81         unsigned asap;                     /**< The ASAP scheduling control step */
  82         unsigned alap;                     /**< The ALAP scheduling control step */
  83         unsigned latency;                  /**< Latency of this node (needed for sorting) */
  84         unsigned sched_point;              /**< the step in which the node is finally scheduled */
  85         unsigned visit_idx;                /**< Index of the node having visited this node last */
  86         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
  87         unsigned n_consumer;               /**< Number of consumers */
  88         ir_node  **block_consumer;         /**< List of consumer being in the same block */
  89         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
  90         unsigned block_idx     : 30;       /**< A unique per block index */
  91         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
  92         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
  93         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
  94                                                                                         for ASAP, successors for ALAP */
  95         unsigned n_unit_types;             /**< number of allowed execution unit types */
  96         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
  97         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
  98 } ilpsched_node_attr_t;
  99
 100 /* attributes for a block */
 101 typedef struct _ilpsched_block_attr_t {
 102         unsigned block_last_idx;        /**< The highest node index in block so far */
 103         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 104         unsigned max_steps;             /**< Upper bound for block execution */
 105         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 106         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 107         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 108 } ilpsched_block_attr_t;
 109
 110 typedef union _ilpsched_attr_ {
 111         ilpsched_node_attr_t  node_attr;
 112         ilpsched_block_attr_t block_attr;
 113 } ilpsched_attr_t;
 114
 115 /* A irn for the phase and it's attributes (either node or block) */
 116 typedef struct {
 117         ir_node         *irn;
 118         ilpsched_attr_t attr;
 119 } be_ilpsched_irn_t;
 120
 121 /* The ILP scheduling environment */
 122 typedef struct {
 123         phase_t              ph;            /**< The phase */
 124         ir_graph             *irg;          /**< The current irg */
 125         heights_t            *height;       /**< The heights object of the irg */
 126         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 127         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 128         const arch_env_t     *arch_env;
 129         const arch_isa_t     *isa;          /**< The ISA */
 130         const be_main_env_t  *main_env;
 131         const be_machine_t   *cpu;          /**< the current abstract machine */
 132         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 133         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 134         DEBUG_ONLY(firm_dbg_module_t *dbg);
 135 } be_ilpsched_env_t;
 136
 137 /* convenience macros to handle phase irn data */
 138 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 139 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 140 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 141 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 142
 143 /* check if node is considered for ILP scheduling */
 144 #define consider_for_sched(isa, irn) \
 145         (! (is_Block(irn)            ||  \
 146                 is_normal_Proj(isa, irn) ||  \
 147                 is_Phi(irn)              ||  \
 148                 is_NoMem(irn)            ||  \
 149                 is_Unknown(irn)          ||  \
 150                 is_End(irn)                  \
 151                 ))
 152
 153 /* gives the valid scheduling time step interval for a node */
 154 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 155
 156 /* gives the valid interval where a node can die */
 157 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 158
 159 /* gives the corresponding ILP variable for given node, unit and time step */
 160 #define ILPVAR_IDX(na, unit, control_step) \
 161         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 162
 163 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 164 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 165         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 166
 167 /* check if a double value is within an epsilon environment of 0 */
 168 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 169
 170 #define ilp_timer_push(t)         lc_timer_push((t))
 171 #define ilp_timer_pop()           lc_timer_pop()
 172 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
 173
 174 /* option variable */
 175 static ilpsched_options_t ilp_opts = {
 176         1,     /* default is with register pressure constraints */
 177         120,   /* if we have more than 70 nodes: use alive nodes constraint */
 178         300,   /* 300 sec per block time limit */
 179         ""     /* no log file */
 180 };
 181
 182 /* ILP options */
 183 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 184         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 185         LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead),
 186         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 187         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 188         { NULL }
 189 };
 190
 191 /*
 192         We need this global variable as we compare nodes dependent on heights,
 193         but we cannot pass any information to the qsort compare function.
 194 */
 195 static heights_t *glob_heights;
 196
 197 /**
 198  * Check if irn is a Proj, which has no execution units assigned.
 199  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 200  */
 201 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
 202         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
 203 }
 204
 205 /**
 206  * Skips normal Projs.
 207  * @return predecessor if irn is a normal Proj, otherwise irn.
 208  */
 209 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
 210         if (is_normal_Proj(isa, irn))
 211                 return get_Proj_pred(irn);
 212         return irn;
 213 }
 214
 215 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 216         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 217         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 218                 lat = 1;
 219         return lat;
 220 }
 221
 222 static int cmp_live_in_nodes(const void *a, const void *b) {
 223         const ilp_livein_node_t *n1 = a;
 224         const ilp_livein_node_t *n2 = b;
 225
 226         return n1->irn != n2->irn;
 227 }
 228
 229 /**
 230  * Compare scheduling time steps of two be_ilpsched_irn's.
 231  */
 232 static int cmp_ilpsched_irn(const void *a, const void *b) {
 233         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 234         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 235         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 236         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 237
 238         if (n1_a->sched_point == n2_a->sched_point) {
 239                 ir_node *irn_a = n1->irn;
 240                 ir_node *irn_b = n2->irn;
 241
 242                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 243                         return 1;
 244                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 245                         return -1;
 246
 247                 /*
 248                         Ok, timestep is equal and the nodes are parallel,
 249                         so check latency and schedule high latency first.
 250                 */
 251                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 252         }
 253         else
 254                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 255 }
 256
 257 /**
 258  * In case there is no phase information for irn, initialize it.
 259  */
 260 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
 261         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 262
 263         if (res == old) {
 264                 /* if we have already some data: check for reinitialization */
 265
 266                 if (! is_Block(irn)) {
 267                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 268
 269                         if (! na->transitive_block_nodes) {
 270                                 ir_node               *block      = get_nodes_block(irn);
 271                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 272                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 273
 274                                 /* we are called after the block indices have been build: create bitset */
 275                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 276                         }
 277                         else {
 278                                 /* we are called from reinit block data: clear the bitset */
 279                                 bitset_clear_all(na->transitive_block_nodes);
 280                                 na->visit_idx    = 0;
 281                                 na->alap_changed = 1;
 282                         }
 283                 }
 284                 return old;
 285         }
 286
 287         res->irn = irn;
 288
 289         /* set ilpsched irn attributes (either block or irn) */
 290         if (is_Block(irn)) {
 291                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 292
 293                 ba->n_interesting_nodes = 0;
 294                 ba->block_last_idx      = 0;
 295                 ba->root_nodes          = plist_new();
 296                 ba->head_ilp_nodes      = NULL;
 297                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 298                 ba->max_steps           = 0;
 299         }
 300         else {
 301                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 302                 memset(na, 0, sizeof(*na));
 303         }
 304
 305         return res;
 306 }
 307
 308 /**
 309  * Assign a per block unique number to each node.
 310  */
 311 static void build_block_idx(ir_node *irn, void *walk_env) {
 312         be_ilpsched_env_t     *env = walk_env;
 313         be_ilpsched_irn_t     *node, *block_node;
 314         ilpsched_node_attr_t  *na;
 315         ilpsched_block_attr_t *ba;
 316
 317         if (! consider_for_sched(env->arch_env->isa, irn))
 318                 return;
 319
 320         node       = get_ilpsched_irn(env, irn);
 321         na         = get_ilpsched_node_attr(node);
 322         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 323         ba         = get_ilpsched_block_attr(block_node);
 324
 325         na->block_idx = ba->block_last_idx++;
 326 }
 327
 328 /********************************************************
 329  *                              __        _
 330  *                             / /       | |
 331  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 332  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 333  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 334  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 335  *                 | |                           | |
 336  *                 |_|                           |_|
 337  ********************************************************/
 338
 339 /**
 340  * Add all nodes having no user in current block to last_nodes list.
 341  */
 342 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 343         ir_node               *block;
 344         const ir_edge_t       *edge;
 345         be_ilpsched_irn_t     *block_node, *node;
 346         ilpsched_block_attr_t *ba;
 347         ilpsched_node_attr_t  *na;
 348         int                   i, j;
 349         be_ilpsched_env_t     *env           = walk_env;
 350         int                   has_block_user = 0;
 351         unsigned              n_consumer     = 0;
 352         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 353         ir_node               **consumer;
 354         int                   idx;
 355
 356         if (! consider_for_sched(env->arch_env->isa, irn))
 357                 return;
 358
 359         block    = get_nodes_block(irn);
 360     idx      = get_irn_idx(irn);
 361         consumer = NEW_ARR_F(ir_node *, 0);
 362
 363         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 364
 365         /* check data and dependency out edges */
 366         for (i = 0; i < 2 && ! has_block_user; ++i) {
 367                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 368                         ir_node *user = get_edge_src_irn(edge);
 369
 370                         if (is_normal_Proj(env->arch_env->isa, user)) {
 371                                 const ir_edge_t *user_edge;
 372
 373                                 if (get_irn_mode(user) == mode_X)
 374                                         continue;
 375
 376                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 377                                 for (j = 0; j < 2; ++j) {
 378                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 379                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 380
 381                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 382                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 383                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 384
 385                                                         /* skip already visited nodes */
 386                                                         if (ua->consumer_idx == idx)
 387                                                                 continue;
 388
 389                                                         /* check if node has user in this block and collect the user if it's a data user */
 390                                                         if (get_nodes_block(real_user) == block) {
 391                                                                 if (i == 0 && j == 0)
 392                                                                         ARR_APP1(ir_node *, consumer, real_user);
 393                                                                 has_block_user = 1;
 394                                                         }
 395
 396                                                         /* only count data consumer */
 397                                                         if (i == 0)
 398                                                                 n_consumer++;
 399
 400                                                         /* mark user as visited by this node */
 401                                                         ua->consumer_idx = idx;
 402                                                 }
 403                                         }
 404                                 }
 405                         }
 406                         else if (is_Block(user)) {
 407                                 continue;
 408                         }
 409                         else if (! is_Phi(user)) {
 410                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 411                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 412
 413                                 /* skip already visited nodes */
 414                                 if (ua->consumer_idx == idx)
 415                                         continue;
 416
 417                                 /* check if node has user in this block and collect the user if it's a data user */
 418                                 if (get_nodes_block(user) == block) {
 419                                         if (i == 0)
 420                                                 ARR_APP1(ir_node *, consumer, user);
 421                                         has_block_user = 1;
 422                                 }
 423
 424                                 /* only count data consumer */
 425                                 if (i == 0)
 426                                         n_consumer++;
 427
 428                                 /* mark user visited by this node */
 429                                 ua->consumer_idx = idx;
 430                         }
 431                         else if (get_nodes_block(user) != block) {
 432                                 n_consumer++;
 433                         }
 434                 }
 435         }
 436
 437         block_node = get_ilpsched_irn(env, block);
 438         ba         = get_ilpsched_block_attr(block_node);
 439
 440         ba->n_interesting_nodes++;
 441
 442         /* current irn has no user inside this block, add to queue */
 443         if (! has_block_user) {
 444                 DB((env->dbg, LEVEL_3, "root node\n"));
 445                 plist_insert_back(ba->root_nodes, irn);
 446         }
 447         else {
 448                 DB((env->dbg, LEVEL_3, "normal node\n"));
 449         }
 450
 451         /* record number of all consumer and the consumer within the same block */
 452         node = get_ilpsched_irn(env, irn);
 453         na   = get_ilpsched_node_attr(node);
 454         na->n_consumer     = n_consumer;
 455         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 456         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 457         DEL_ARR_F(consumer);
 458 }
 459
 460 /**
 461  * Calculate the ASAP scheduling step for current irn.
 462  */
 463 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 464         be_ilpsched_env_t     *env = walk_env;
 465         int                   i;
 466         ir_node               *block;
 467         be_ilpsched_irn_t     *node, *block_node;
 468         ilpsched_node_attr_t  *na;
 469         ilpsched_block_attr_t *ba;
 470
 471         /* These nodes are handled separate */
 472         if (! consider_for_sched(env->arch_env->isa, irn))
 473                 return;
 474
 475         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 476
 477         block    = get_nodes_block(irn);
 478         node     = get_ilpsched_irn(env, irn);
 479         na       = get_ilpsched_node_attr(node);
 480         na->asap = 1;
 481
 482         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 483                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
 484
 485                 /* check for greatest distance to top */
 486                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 487                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 488                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 489                         unsigned             lat;
 490
 491                         lat         = fixed_latency(env->sel, pred, env->block_env);
 492                         na->latency = lat;
 493                         na->asap    = MAX(na->asap, pna->asap + lat);
 494                 }
 495         }
 496
 497         /* add node to ILP node list and update max_steps */
 498         block_node = get_ilpsched_irn(env, block);
 499         ba         = get_ilpsched_block_attr(block_node);
 500
 501         set_irn_link(irn, ba->head_ilp_nodes);
 502         ba->head_ilp_nodes = irn;
 503         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 504
 505         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 506 }
 507
 508 /**
 509  * Calculate the ALAP scheduling step of all irns in current block.
 510  * Depends on max_steps being calculated.
 511  */
 512 static void calculate_block_alap(ir_node *block, void *walk_env) {
 513         be_ilpsched_env_t     *env        = walk_env;
 514         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 515         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 516         waitq                 *cur_queue  = new_waitq();
 517         plist_element_t       *el;
 518
 519         assert(is_Block(block));
 520
 521         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 522                 block, ba->n_interesting_nodes, ba->max_steps));
 523
 524         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 525         //phase_reinit_block_irn_data(&env->ph, block);
 526
 527         /* init start queue */
 528         foreach_plist(ba->root_nodes, el) {
 529                 waitq_put(cur_queue, plist_element_get_value(el));
 530         }
 531
 532         /* repeat until all nodes are processed */
 533         while (! waitq_empty(cur_queue)) {
 534                 waitq *next_queue = new_waitq();
 535
 536                 /* process all nodes in current step */
 537                 while (! waitq_empty(cur_queue)) {
 538                         ir_node              *cur_irn = waitq_get(cur_queue);
 539                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 540                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 541                         int                  i;
 542
 543                         /* cur_node has no alap set -> it's a root node, set to max alap */
 544                         if (na->alap == 0) {
 545                                 na->alap = ba->max_steps;
 546                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 547                                         cur_irn, na->alap));
 548                         }
 549                         else {
 550                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 551                                         cur_irn, na->alap));
 552                         }
 553
 554                         /* set the alap's of all predecessors */
 555                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 556                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
 557
 558                                 /* check for greatest distance to bottom */
 559                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 560                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 561                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 562                                         unsigned             lat;
 563
 564                                         /* mark the predecessor as visited by current irn */
 565                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 566                                                 continue;
 567                                         pna->visit_idx = get_irn_idx(cur_irn);
 568
 569                                         lat = fixed_latency(env->sel, pred, env->block_env);
 570
 571                                         /* set ALAP of current pred */
 572                                         if (pna->alap == 0) {
 573                                                 /* current ALAP is 0: node has not yet been visited */
 574                                                 pna->alap_changed = 1;
 575                                                 pna->alap         = na->alap - lat;
 576                                         }
 577                                         else if (pna->alap > na->alap - lat) {
 578                                                 /* we found a longer path to root node: change ALAP */
 579                                                 pna->alap         = na->alap - lat;
 580                                                 pna->alap_changed = 1;
 581                                         }
 582                                         else {
 583                                                 /* current ALAP is best found so far: keep it */
 584                                                 pna->alap_changed = 0;
 585                                         }
 586
 587                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 588
 589                                         /* enqueue node for next iteration */
 590                                         if (get_irn_ins_or_deps(pred) > 0)
 591                                                 waitq_put(next_queue, pred);
 592                                 }
 593                         }
 594                 }
 595
 596                 /* prepare for next iteration */
 597                 del_waitq(cur_queue);
 598                 cur_queue = next_queue;
 599         }
 600 }
 601
 602 /**
 603  * Free list of root nodes and the set of live-in nodes.
 604  */
 605 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 606         be_ilpsched_env_t     *env        = walk_env;
 607         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 608         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 609
 610         plist_free(ba->root_nodes);
 611         ba->root_nodes = NULL;
 612         del_pset(ba->livein_nodes);
 613         ba->livein_nodes = NULL;
 614 }
 615
 616 /**
 617  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 618  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 619  */
 620 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 621         be_ilpsched_env_t    *env  = walk_env;
 622         ir_node              *pred = irn;
 623         be_ilpsched_irn_t    *node, *pred_node;
 624         ilpsched_node_attr_t *na, *pna;
 625
 626         if (! consider_for_sched(env->arch_env->isa, irn))
 627                 return;
 628
 629         if (! is_Proj(irn) && ! be_is_Keep(irn))
 630                 return;
 631
 632         /* go to the ancestor */
 633         if (be_is_Keep(irn))
 634                 pred = get_irn_n(irn, 0);
 635         pred = skip_Proj(pred);
 636
 637         node      = get_ilpsched_irn(env, irn);
 638         pred_node = get_ilpsched_irn(env, pred);
 639         na        = get_ilpsched_node_attr(node);
 640         pna       = get_ilpsched_node_attr(pred_node);
 641
 642         na->asap = pna->asap;
 643         na->alap = pna->alap;
 644
 645         /* record all Projs and Keeps for this node */
 646         if (! pna->projkeeps)
 647                 pna->projkeeps = new_waitq();
 648         waitq_put(pna->projkeeps, irn);
 649
 650         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 651 }
 652
 653 /*******************************************
 654  *           _              _       _
 655  *          | |            | |     | |
 656  *  ___  ___| |__   ___  __| |_   _| | ___
 657  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 658  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 659  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 660  *
 661  *******************************************/
 662
 663 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
 664         const ir_edge_t *edge;
 665
 666         foreach_out_edge(irn, edge) {
 667                 ir_node *user = get_edge_src_irn(edge);
 668
 669                 if (be_is_Keep(user)) {
 670                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 671                         waitq_put(keeps, user);
 672                 }
 673         }
 674 }
 675
 676 /**
 677  * Inserts @p irn before @p before into schedule and notifies backend.
 678  */
 679 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 680         ir_node *before, ir_node *irn, unsigned cycle)
 681 {
 682         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 683         sched_add_before(before, irn);
 684 }
 685
 686 /**
 687  * Adds a node, it's Projs (in case of mode_T nodes) and
 688  * it's Keeps to schedule.
 689  */
 690 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
 691         const ir_edge_t *edge;
 692         waitq           *keeps = new_waitq();
 693
 694         /* mode_M nodes are not scheduled */
 695         if (get_irn_mode(irn) == mode_M)
 696                 return;
 697
 698         if (! sched_is_scheduled(irn))
 699                 notified_sched_add_before(env, block, irn, cycle);
 700
 701         /* add Projs */
 702         if (get_irn_mode(irn) == mode_T) {
 703                 foreach_out_edge(irn, edge) {
 704                         ir_node *user = get_edge_src_irn(edge);
 705
 706                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 707                                 get_irn_n_edges(user) > 0)
 708                         {
 709                                 notified_sched_add_before(env, block, user, cycle);
 710                         }
 711
 712                         check_for_keeps(keeps, block, user);
 713                 }
 714         }
 715         else {
 716                 check_for_keeps(keeps, block, irn);
 717         }
 718
 719         /* add Keeps */
 720         while (! waitq_empty(keeps)) {
 721                 ir_node *keep = waitq_get(keeps);
 722                 if (! sched_is_scheduled(keep))
 723                         notified_sched_add_before(env, block, keep, cycle);
 724         }
 725
 726         del_waitq(keeps);
 727 }
 728
 729 /**
 730  * Schedule all nodes in the given block, according to the ILP solution.
 731  */
 732 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 733         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 734         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 735         sched_info_t          *info       = get_irn_sched_info(block);
 736         be_ilpsched_irn_t     **sched_nodes;
 737         unsigned              i, l;
 738         ir_node               *cfop, *irn;
 739         const ir_edge_t       *edge;
 740
 741         /* init block schedule list */
 742         INIT_LIST_HEAD(&info->list);
 743         info->scheduled = 1;
 744
 745         /* collect nodes and their scheduling time step */
 746         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 747         if (ba->n_interesting_nodes == 0) {
 748                 /* ignore */
 749         }
 750         else if (ba->n_interesting_nodes == 1) {
 751                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 752
 753                 /* add the single node */
 754                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 755         }
 756         else {
 757                 /* check all nodes for their positive solution */
 758                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 759                         be_ilpsched_irn_t    *node;
 760                         ilpsched_node_attr_t *na;
 761                         int                  tp_idx, found;
 762                         unsigned             cur_var, t;
 763
 764                         node    = get_ilpsched_irn(env, irn);
 765                         na      = get_ilpsched_node_attr(node);
 766                         cur_var = 0;
 767                         found   = 0;
 768
 769                         if (! na->is_dummy_node) {
 770                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 771                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 772                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 773
 774                                                 if (! LPP_VALUE_IS_0(cost)) {
 775                                                         ir_fprintf(stderr, "\t\t%+F has additional regpressure costs of %f\n", irn, cost);
 776                                                         found = 1;
 777                                                 }
 778                                         }
 779                                 }
 780                         }
 781
 782                         found = 0;
 783                         /* go over all variables of a node until the non-zero one is found */
 784                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 785                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 786                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 787
 788                                         /* check, if variable is set to one (it's not zero then :) */
 789                                         if (! LPP_VALUE_IS_0(val)) {
 790                                                 na->sched_point = t;
 791                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 792                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 793                                                         irn, t, na->type_info[tp_idx].tp->name));
 794                                                 found = 1;
 795                                         }
 796                                 }
 797                         }
 798                 }
 799
 800                 glob_heights = env->height;
 801                 /* sort nodes ascending by scheduling time step */
 802                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 803         }
 804
 805         /* make all Phis ready and remember the single cf op */
 806         cfop = NULL;
 807         foreach_out_edge(block, edge) {
 808                 irn = get_edge_src_irn(edge);
 809
 810                 switch (get_irn_opcode(irn)) {
 811                         case iro_Phi:
 812                                 add_to_sched(env, block, irn, 0);
 813                                 break;
 814                         case iro_Start:
 815                         case iro_End:
 816                         case iro_Proj:
 817                         case iro_Bad:
 818                         case iro_Unknown:
 819                                 break;
 820                         default:
 821                                 if (is_cfop(irn)) {
 822                                         assert(cfop == NULL && "Highlander - there can be only one");
 823                                         cfop = irn;
 824                                 }
 825                         break;
 826                 }
 827         }
 828
 829         /* add all nodes from list */
 830         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 831                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 832                 if (sched_nodes[i]->irn != cfop)
 833                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 834         }
 835
 836         /* schedule control flow node if not already done */
 837         if (cfop && ! sched_is_scheduled(cfop))
 838                 add_to_sched(env, block, cfop, 0);
 839
 840         DEL_ARR_F(sched_nodes);
 841 }
 842
 843 /***************************************************************
 844  *   _____ _      _____     _____           _   _
 845  *  |_   _| |    |  __ \   / ____|         | | (_)
 846  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 847  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 848  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 849  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 850  *
 851  ***************************************************************/
 852
 853 /**
 854  * Check if node can be executed on given unit type.
 855  */
 856 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 857         int                  i;
 858         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 859
 860         for (i = na->n_unit_types - 1; i >= 0; --i) {
 861                 if (na->type_info[i].tp == tp)
 862                         return i;
 863         }
 864
 865         return -1;
 866 }
 867
 868 /************************************************
 869  *                   _       _     _
 870  *                  (_)     | |   | |
 871  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 872  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 873  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 874  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 875  *
 876  ************************************************/
 877
 878 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
 879         const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
 880         unsigned                  n_unit_types = 0;
 881         be_ilpsched_irn_t         *node;
 882         ilpsched_node_attr_t      *na;
 883         unsigned                  unit_idx, tp_idx;
 884
 885         /* count number of available unit types for this node */
 886         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 887                 /* just count */ ;
 888
 889         node = get_ilpsched_irn(env, irn);
 890         na   = get_ilpsched_node_attr(node);
 891
 892         if (! na->type_info) {
 893                 na->n_unit_types = n_unit_types;
 894                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 895
 896                 /* fill the type info array */
 897                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 898                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 899                                 /* beware: we also count number of available units here */
 900                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 901                                         na->is_dummy_node = 1;
 902                         }
 903
 904                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 905                         na->type_info[tp_idx].n_units = unit_idx;
 906                 }
 907         }
 908
 909         return n_unit_types;
 910 }
 911
 912 /**
 913  * Returns the largest alap time of a user of @p irn.
 914  * The user must be in block @p block.
 915  */
 916 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
 917         const ir_edge_t *edge;
 918         unsigned        max_alap = 0;
 919
 920         foreach_out_edge(irn, edge) {
 921                 ir_node *user = get_edge_src_irn(edge);
 922
 923                 if (get_nodes_block(user) == block) {
 924                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 925                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 926
 927                         max_alap = MAX(max_alap, na->alap);
 928                 }
 929         }
 930
 931         assert(max_alap > 0);
 932         return max_alap;
 933 }
 934
 935 /**
 936  * Create the following variables:
 937  * - x_{nt}^k    binary     weigthed with: t
 938  *      node n is scheduled at time step t to unit type k
 939  * ==>> These variables represent the schedule
 940  *
 941  * - d_{nt}^k    binary     weighted with: t
 942  *      node n dies at time step t on unit type k
 943  * - a_{nt}^k    binary     weighted with num_nodes
 944  *      node n is alive at time step t on unit type k
 945  *
 946  * - y_{nt}^k    binary     weighted with: num_nodes^2
 947  *      node n is scheduled at time step t to unit type k
 948  *      although all units of this type are occupied
 949  * ==>> These variables represent the register pressure
 950  *
 951  */
 952 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 953         char                  buf[1024];
 954         ir_node               *irn;
 955         unsigned              num_block_var, num_nodes;
 956         ilp_livein_node_t     *livein;
 957         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 958         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 959         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
 960
 961         ilp_timer_push(t_var);
 962         num_block_var = num_nodes = 0;
 963         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 964                 be_ilpsched_irn_t    *node;
 965                 ilpsched_node_attr_t *na;
 966                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
 967                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
 968                 int                  i;
 969
 970                 node         = get_ilpsched_irn(env, irn);
 971                 na           = get_ilpsched_node_attr(node);
 972                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
 973
 974                 /* allocate space for ilp variables */
 975                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 976                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 977
 978                 /* we need these variables only for "real" nodes */
 979                 if (! na->is_dummy_node) {
 980                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 981                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 982
 983                         num_ad = ba->max_steps - na->asap + 1;
 984
 985                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 986                                 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 987                                 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
 988                         }
 989                         else {
 990                                 na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 991                                 memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0]));
 992                         }
 993                 }
 994
 995                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
 996                         irn, na->asap, na->alap, na->n_unit_types));
 997
 998                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
 999                 /* create variables */
1000                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1001                         unsigned t;
1002
1003                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1004                                 /* x_{nt}^k variables */
1005                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1006                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1007                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1008                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1009                                 /* variable counter */
1010                                 n_var++;
1011                                 num_block_var++;
1012
1013                                 if (! na->is_dummy_node) {
1014                                         /* y_{nt}^k variables */
1015                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1016                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1017                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1018                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1019
1020                                         /* variable counter */
1021                                         n_var++;
1022                                         num_block_var++;
1023                                 }
1024                         }
1025
1026                         /* a node can die at any step t: asap(n) <= t <= U */
1027                         if (! na->is_dummy_node) {
1028                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1029
1030                                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
1031                                                 /* a_{nt}^k variables */
1032                                                 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1033                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1034                                                 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1035                                         }
1036                                         else {
1037                                                 /* d_{nt}^k variables */
1038                                                 snprintf(buf, sizeof(buf), "d_n%u_%s_%u",
1039                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1040                                                 na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1041                                         }
1042                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1043
1044                                         /* variable counter */
1045                                         n_var++;
1046                                         num_block_var++;
1047                                 }
1048                         }
1049
1050                         /* collect live-in nodes */
1051                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1052                                 ir_node *pred = get_irn_n(irn, i);
1053
1054                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1055                                         be_ilpsched_set_type_info(env, pred, var_obst);
1056                                         if (! na->is_dummy_node) {
1057                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1058                                                 entry->irn = pred;
1059                                                 entry->a   = NULL;
1060                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1061                                         }
1062                                 }
1063                         }
1064                 }
1065
1066                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1067                 num_nodes++;
1068         }
1069
1070         /* create alive variables a_{nt}^k for live-ins */
1071         foreach_pset(ba->livein_nodes, livein) {
1072                 be_ilpsched_irn_t    *node;
1073                 ilpsched_node_attr_t *na;
1074                 unsigned             tp_idx, var_idx, max_alap;
1075                 ir_node              *irn;
1076
1077                 irn  = livein->irn;
1078                 node = get_ilpsched_irn(env, irn);
1079                 na   = get_ilpsched_node_attr(node);
1080
1081                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1082
1083                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1084                 var_idx   = 0;
1085
1086                 /* create variables */
1087                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1088                         unsigned t;
1089
1090                         for (t = 0; t < livein->max_alive_steps; ++t) {
1091                                 /* a_{nt}^k variables */
1092                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1093                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1094                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1095                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1096                                 num_block_var++;
1097                         }
1098                 }
1099         }
1100
1101         ilp_timer_pop();
1102         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1103                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1104 }
1105
1106 /*******************************************************
1107  *                      _             _       _
1108  *                     | |           (_)     | |
1109  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1110  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1111  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1112  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1113  *
1114  *******************************************************/
1115
1116 /**
1117  * Create following ILP constraints:
1118  * - the assignment constraints:
1119  *     assure each node is executed once by exactly one (allowed) execution unit
1120  * - the dead node assignment constraints:
1121  *     assure a node can only die at most once
1122  * - the precedence constraints:
1123  *     assure that no data dependencies are violated
1124  */
1125 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1126         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1127         char                  buf[1024];
1128         ir_node               *irn;
1129         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1130         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1131         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign",      "create assignment constraints");
1132         lc_timer_t            *t_cst_dead    = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints");
1133         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",        "create precedence constraints");
1134
1135         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1136         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1137                 int                  cst, tp_idx, i;
1138                 unsigned             cur_var;
1139                 be_ilpsched_irn_t    *node;
1140                 ilpsched_node_attr_t *na;
1141
1142                 node    = get_ilpsched_irn(env, irn);
1143                 na      = get_ilpsched_node_attr(node);
1144                 cur_var = 0;
1145
1146                 /* the assignment constraint */
1147                 ilp_timer_push(t_cst_assign);
1148                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1149                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1150                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1151                 num_cst_assign++;
1152
1153                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1154                 ilp_timer_pop();
1155
1156                 /* the dead node assignment constraint */
1157                 if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) {
1158                         ilp_timer_push(t_cst_dead);
1159                         snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn));
1160                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1161                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1162
1163                         lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0);
1164                         ilp_timer_pop();
1165                 }
1166
1167                 /* We have separate constraints for Projs and Keeps */
1168                 // ILP becomes infeasible ?!?
1169 //              if (is_Proj(irn) || be_is_Keep(irn))
1170 //                      continue;
1171
1172                 /* the precedence constraints */
1173                 ilp_timer_push(t_cst_prec);
1174                 bs_block_irns = bitset_clear_all(bs_block_irns);
1175                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1176                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1177                         unsigned             t_low, t_high, t;
1178                         be_ilpsched_irn_t    *pred_node;
1179                         ilpsched_node_attr_t *pna;
1180                         unsigned             delay;
1181
1182                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1183                                 continue;
1184
1185                         pred_node = get_ilpsched_irn(env, pred);
1186                         pna       = get_ilpsched_node_attr(pred_node);
1187
1188                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1189
1190                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1191                                 bitset_set(bs_block_irns, pna->block_idx);
1192                         else
1193                                 continue;
1194
1195                         /* irn = n, pred = m */
1196                         delay  = fixed_latency(env->sel, pred, env->block_env);
1197                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1198                         t_high = MIN(na->alap, pna->alap + delay - 1);
1199                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1200                                 unsigned tn, tm;
1201                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1202
1203                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1204                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1205                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1206                                 num_cst_prec++;
1207
1208                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1209                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1210                                         /* node variables have smaller index than pred variables */
1211                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1212                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1213                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1214                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1215                                                 }
1216                                         }
1217
1218                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1219                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1220                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1221                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1222                                                 }
1223                                         }
1224                                 }
1225                                 else {
1226                                         /* pred variables have smaller index than node variables */
1227                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1228                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1229                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1230                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1231                                                 }
1232                                         }
1233
1234                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1235                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1236                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1237                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1238                                                 }
1239                                         }
1240                                 }
1241
1242                                 if (ARR_LEN(tmp_var_idx) > 0)
1243                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1244
1245                                 DEL_ARR_F(tmp_var_idx);
1246                         }
1247                 }
1248                 ilp_timer_pop();
1249         }
1250         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1251                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1252         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1253                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1254 }
1255
1256 /**
1257  * Create ILP resource constraints:
1258  * - assure that for each time step not more instructions are scheduled
1259  *   to the same unit types as units of this type are available
1260  */
1261 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1262         int                   glob_type_idx;
1263         char                  buf[1024];
1264         unsigned              num_cst_resrc = 0;
1265         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1266         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1267
1268         ilp_timer_push(t_cst_rsrc);
1269         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1270                 unsigned                 t;
1271                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1272
1273                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1274
1275                 /* check each time step */
1276                 for (t = 0; t < ba->max_steps; ++t) {
1277                         ir_node *irn;
1278                         int     cst;
1279                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1280
1281                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1282                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1283                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1284                         num_cst_resrc++;
1285
1286                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1287                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1288                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1289                                 int                  tp_idx;
1290
1291                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1292
1293                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1294                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1295                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1296                                 }
1297                         }
1298
1299                         /* set constraints if we have some */
1300                         if (ARR_LEN(tmp_var_idx) > 0)
1301                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1302
1303                         DEL_ARR_F(tmp_var_idx);
1304                 }
1305         }
1306         ilp_timer_pop();
1307         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1308                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1309 }
1310
1311 /**
1312  * Create ILP bundle constraints:
1313  * - assure, at most bundle_size * bundles_per_cycle instructions
1314  *   can be started at a certain point.
1315  */
1316 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1317         char                  buf[1024];
1318         unsigned              t;
1319         unsigned              num_cst_bundle = 0;
1320         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1321         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1322         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1323
1324         ilp_timer_push(t_cst_bundle);
1325         for (t = 0; t < ba->max_steps; ++t) {
1326                 ir_node *irn;
1327                 int     cst;
1328                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1329
1330                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1331                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1332                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1333                 num_cst_bundle++;
1334
1335                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1336                         be_ilpsched_irn_t    *node;
1337                         ilpsched_node_attr_t *na;
1338                         int                  tp_idx;
1339
1340                         /* Projs and Keeps do not contribute to bundle size */
1341                         if (is_Proj(irn) || be_is_Keep(irn))
1342                                 continue;
1343
1344                         node = get_ilpsched_irn(env, irn);
1345                         na   = get_ilpsched_node_attr(node);
1346
1347                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1348                         if (na->is_dummy_node)
1349                                 continue;
1350
1351                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1352                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1353                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1354                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1355                                 }
1356                         }
1357                 }
1358
1359                 if (ARR_LEN(tmp_var_idx) > 0)
1360                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1361
1362                 DEL_ARR_F(tmp_var_idx);
1363         }
1364         ilp_timer_pop();
1365         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1366                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1367 }
1368
1369 /**
1370  * Create ILP dying nodes constraints:
1371  * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k
1372  */
1373 static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1374         char                  buf[1024];
1375         unsigned              t;
1376         unsigned              num_cst = 0;
1377         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1378         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints");
1379
1380         ilp_timer_push(t_cst);
1381         /* check all time_steps */
1382         for (t = 0; t < ba->max_steps; ++t) {
1383                 ir_node *irn;
1384
1385                 /* for all nodes */
1386                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1387                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1388                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1389
1390                         /* if node has no consumer within current block, it cannot die here */
1391                         /* we also ignore nodes assigned to dummy unit */
1392                         if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node)
1393                                 continue;
1394
1395                         /* node can only die here if t at least asap(n) */
1396                         if (t >= na->asap - 1) {
1397                                 int node_tp_idx;
1398
1399                                 /* for all unit types */
1400                                 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1401                                         int tp_idx, i, cst;
1402                                         int *tmp_var_idx = NEW_ARR_F(int, 0);
1403
1404                                         snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn));
1405                                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1));
1406                                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1407                                         num_cst++;
1408
1409                                         /* number of consumer scheduled till t */
1410                                         for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1411                                                 be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1412                                                 ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1413
1414                                                 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1415                                                         unsigned tm;
1416
1417                                                         for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) {
1418                                                                 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1419                                                                 ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]);
1420                                                         }
1421                                                 }
1422                                         }
1423
1424                                         /* could be that no consumer can be scheduled at this point */
1425                                         if (ARR_LEN(tmp_var_idx)) {
1426                                                 int      idx;
1427                                                 unsigned tn;
1428
1429                                                 /* subtract possible prior kill points */
1430                                                 for (tn = na->asap - 1; tn < t; ++tn) {
1431                                                         idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn);
1432                                                         lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0);
1433                                                 }
1434
1435                                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1436                                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer));
1437                                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1438                                         }
1439
1440                                         DEL_ARR_F(tmp_var_idx);
1441                                 }
1442                         }
1443
1444                 }
1445         }
1446         ilp_timer_pop();
1447         DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n",
1448                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1449 }
1450
1451 /**
1452  * Create ILP alive nodes constraints:
1453  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1454  */
1455 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1456         char                  buf[1024];
1457         ir_node               *irn;
1458         unsigned              num_cst = 0;
1459         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1460         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1461
1462         ilp_timer_push(t_cst);
1463         /* for each node */
1464         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1465                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1466                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1467                 unsigned             t;
1468
1469                 /* we ignore nodes assigned to dummy unit here */
1470                 if (na->is_dummy_node)
1471                         continue;
1472
1473                 /* check check all time steps: asap(n) <= t <= U */
1474                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1475                         int node_tp_idx;
1476
1477                         /* for all unit types available for this node */
1478                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1479                                 unsigned tn, tn_max, idx;
1480                                 int      cst, i;
1481                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1482                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1483
1484                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1485                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1486                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1487                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1488                                 num_cst++;
1489
1490                                 tn_max = MIN(na->alap - 1, t);
1491                                 /* check if the node has been scheduled so far */
1492                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1493                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1494                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1495                                 }
1496
1497                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1498                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1499                                 DEL_ARR_F(tmp_var_idx_n);
1500
1501                                 /* subtract the number of consumer scheduled so far */
1502                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1503                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1504                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1505                                         int                  tp_idx;
1506                                         unsigned             tm, tm_max;
1507
1508                                         tm_max = MIN(ca->alap - 1, t);
1509                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1510                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1511                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1512                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1513                                                 }
1514                                         }
1515                                 }
1516
1517                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1518                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1519                                 DEL_ARR_F(tmp_var_idx_m);
1520
1521                                 /* -c * a_{nt}^k */
1522                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1523                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1524
1525                         }
1526                 }
1527         }
1528         ilp_timer_pop();
1529         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1530                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1531 }
1532
1533 /**
1534  * Create ILP alive nodes constraints for live-in nodes:
1535  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1536  */
1537 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1538         char                  buf[1024];
1539         ilp_livein_node_t     *livein;
1540         unsigned              num_cst = 0;
1541         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1542         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1543
1544         ilp_timer_push(t_cst);
1545         /* for each node */
1546         foreach_pset(ba->livein_nodes, livein) {
1547                 ir_node              *irn  = livein->irn;
1548                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1549                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1550                 unsigned             t;
1551
1552                 /* check check all time steps: 0 <= t < max_alive_steps */
1553                 for (t = 0; t < livein->max_alive_steps; ++t) {
1554                         int node_tp_idx;
1555
1556                         /* for all unit types available for this node */
1557                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1558                                 const ir_edge_t *edge;
1559                                 unsigned tn, tn_max, idx;
1560                                 int      cst, i, num_block_user;
1561                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1562
1563                                 /* check the number of consumer scheduled so far */
1564                                 num_block_user = 0;
1565                                 foreach_out_edge(irn, edge) {
1566                                         ir_node              *user = get_edge_src_irn(edge);
1567                                         be_ilpsched_irn_t    *cons;
1568                                         ilpsched_node_attr_t *ca;
1569                                         int                  tp_idx;
1570                                         unsigned             tm, tm_max;
1571
1572                                         /* check only users within current block */
1573                                         if (get_nodes_block(user) != block_node->irn)
1574                                                 continue;
1575
1576                                         num_block_user++;
1577                                         cons = get_ilpsched_irn(env, user);
1578                                         ca   = get_ilpsched_node_attr(cons);
1579
1580                                         tm_max = MIN(ca->alap - 1, t);
1581                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1582                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1583                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1584                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1585                                                 }
1586                                         }
1587                                 }
1588
1589                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1590                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1591                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1592                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1593                                 num_cst++;
1594
1595                                 /* sum(scheduled users) */
1596                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1597                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1598                                 DEL_ARR_F(tmp_var_idx_m);
1599
1600                                 /* + c * a_{nt}^k */
1601                                 idx = node_tp_idx * livein->max_alive_steps + t;
1602                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1603                         }
1604                 }
1605         }
1606         ilp_timer_pop();
1607         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1608                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1609 }
1610
1611 /**
1612  * Create ILP pressure constraints, based on dead nodes:
1613  * - add additional costs to objective function if a node is scheduled
1614  *   on a unit although all units of this type are currently occupied
1615  */
1616 static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1617         char                  buf[1024];
1618         ir_node               *cur_irn;
1619         unsigned              num_cst = 0;
1620         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1621         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1622
1623         ilp_timer_push(t_cst);
1624         /* y_{nt}^k is set for each node and timestep and unit type */
1625         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1626                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1627                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1628                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1629                 int                  glob_type_idx;
1630
1631                 /* we ignore nodes assigned to DUMMY unit here */
1632                 if (cur_na->is_dummy_node)
1633                         continue;
1634
1635                 /* for all types */
1636                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1637                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1638                         int                      cur_tp_idx;
1639                         unsigned                 t;
1640
1641                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1642
1643                         /* check if node can be executed on this unit type */
1644                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1645                         if (cur_tp_idx < 0)
1646                                 continue;
1647
1648                         /* check all time_steps */
1649                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1650                                 int     cst, y_idx;
1651                                 ir_node *irn;
1652                                 int     *tmp_idx_1  = NEW_ARR_F(int, 0);
1653                                 int     *tmp_idx_m1 = NEW_ARR_F(int, 0);
1654
1655                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1656                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1657                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1658                                 num_cst++;
1659
1660                                 /*
1661                                         - accumulate all nodes scheduled on unit type k till t
1662                                         - subtract all nodes died on unit type k till t
1663                                 */
1664                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1665                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1666                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1667                                         unsigned             tn, tmax;
1668                                         int                  tp_idx;
1669
1670                                         tmax   = MIN(t, na->alap - 1);
1671                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1672
1673                                         /* current unit type is not suitable for current node */
1674                                         if (tp_idx < 0)
1675                                                 continue;
1676
1677                                         for (tn = na->asap - 1; tn <= tmax; ++tn) {
1678                                                 int idx;
1679
1680                                                 /* node scheduled */
1681                                                 idx = ILPVAR_IDX(na, tp_idx, tn);
1682                                                 ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]);
1683
1684                                                 /* node dead */
1685                                                 idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn);
1686                                                 ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]);
1687                                         }
1688                                 }
1689
1690                                 if (ARR_LEN(tmp_idx_1) > 0)
1691                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0);
1692
1693                                 if (ARR_LEN(tmp_idx_m1) > 0)
1694                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0);
1695
1696                                 /* BEWARE: t is unsigned, so (double)(-t) won't work */
1697                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1698                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1699
1700                                 DEL_ARR_F(tmp_idx_1);
1701                                 DEL_ARR_F(tmp_idx_m1);
1702                         }
1703                 }
1704         }
1705         ilp_timer_pop();
1706         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1707                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1708 }
1709
1710 /**
1711  * Create ILP pressure constraints, based on alive nodes:
1712  * - add additional costs to objective function if a node is scheduled
1713  *   on a unit although all units of this type are currently occupied
1714  */
1715 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1716         char                  buf[1024];
1717         ir_node               *cur_irn;
1718         unsigned              num_cst = 0;
1719         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1720         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1721
1722         ilp_timer_push(t_cst);
1723         /* y_{nt}^k is set for each node and timestep and unit type */
1724         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1725                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1726                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1727                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1728                 int                  glob_type_idx;
1729
1730                 /* we ignore nodes assigned to DUMMY unit here */
1731                 if (cur_na->is_dummy_node)
1732                         continue;
1733
1734                 /* for all types */
1735                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1736                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1737                         int                      cur_tp_idx;
1738                         unsigned                 t;
1739
1740                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1741
1742                         /* check if node can be executed on this unit type */
1743                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1744                         if (cur_tp_idx < 0)
1745                                 continue;
1746
1747                         /* check all time_steps at which the current node can be scheduled */
1748                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1749                                 int     cst, y_idx;
1750                                 ir_node *irn;
1751                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1752                                 ilp_livein_node_t *livein;
1753
1754                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1755                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1756                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1757                                 num_cst++;
1758
1759                                 /* - accumulate all nodes alive at point t on unit type k */
1760                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1761                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1762                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1763                                         int                  a_idx, tp_idx;
1764
1765                                         /* check if node can be alive here */
1766                                         if (t < na->asap - 1)
1767                                                 continue;
1768
1769                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1770
1771                                         /* current type is not suitable */
1772                                         if (tp_idx < 0)
1773                                                 continue;
1774
1775                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1776                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1777                                 }
1778                                 /* do the same for livein nodes */
1779                                 foreach_pset(ba->livein_nodes, livein) {
1780                                         ir_node              *irn  = livein->irn;
1781                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1782                                         int                  a_idx, tp_idx;
1783
1784                                         /* check if node can be alive here */
1785                                         if (t >= livein->max_alive_steps)
1786                                                 continue;
1787
1788                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1789
1790                                         /* current type is not suitable */
1791                                         if (tp_idx < 0)
1792                                                 continue;
1793
1794                                         a_idx = tp_idx * livein->max_alive_steps + t;
1795                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1796                                 }
1797
1798                                 if (ARR_LEN(tmp_var_idx) > 0)
1799                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1800                                 DEL_ARR_F(tmp_var_idx);
1801
1802                                 /* - num_nodes * y_{nt}^k */
1803                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1804                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1805                         }
1806                 }
1807         }
1808         ilp_timer_pop();
1809         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1810                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1811 }
1812
1813 /**
1814  * Create ILP branch constraints:
1815  * Assure, alle nodes are scheduled prior to cfg op.
1816  */
1817 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1818         char                  buf[1024];
1819         ir_node               *cur_irn, *cfop;
1820         unsigned              num_cst          = 0;
1821         unsigned              num_non_branches = 0;
1822         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1823         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1824
1825         ilp_timer_push(t_cst);
1826         cfop = NULL;
1827         /* determine number of non-branch nodes and the one and only branch node */
1828         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1829                 switch (get_irn_opcode(cur_irn)) {
1830                         case iro_Phi:
1831                         case iro_Start:
1832                         case iro_End:
1833                         case iro_Proj:
1834                         case iro_Bad:
1835                         case iro_Unknown:
1836                                 num_non_branches++;
1837                                 break;
1838                         default:
1839                                 if (is_cfop(cur_irn)) {
1840                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1841                                         cfop = cur_irn;
1842                                 }
1843                                 else {
1844                                         num_non_branches++;
1845                                 }
1846                                 break;
1847                 }
1848         }
1849
1850         if (cfop) {
1851                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1852                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1853                 unsigned t;
1854
1855                 /* for each time step */
1856                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1857                         int *non_branch_vars, *branch_vars;
1858                         int cst;
1859
1860                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1861                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1862                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1863                         num_cst++;
1864
1865                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1866                         non_branch_vars = NEW_ARR_F(int, 0);
1867                         branch_vars     = NEW_ARR_F(int, 0);
1868                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1869                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1870                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1871                                 int                  tp_idx;
1872
1873                                 if (cur_irn == cfop) {
1874                                         /* for all unit types available for this node */
1875                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1876                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1877                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1878                                         }
1879                                 }
1880                                 else {
1881                                         /* sum up all possible schedule points for this node upto current timestep */
1882                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1883                                                 unsigned tn;
1884                                                 unsigned tmax = MIN(t, na->alap - 1);
1885
1886                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1887                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1888                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1889                                                 }
1890                                         }
1891                                 }
1892
1893                         }
1894
1895                         if (ARR_LEN(non_branch_vars) > 0)
1896                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1897                         if (ARR_LEN(branch_vars) > 0)
1898                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1899
1900                         DEL_ARR_F(branch_vars);
1901                         DEL_ARR_F(non_branch_vars);
1902                 }
1903         }
1904         ilp_timer_pop();
1905         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1906                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1907 }
1908
1909 #if 0
1910 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1911         char                  buf[1024];
1912         ir_node               *irn;
1913         unsigned              num_cst = 0;
1914         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1915         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1916
1917         ilp_timer_push(t_cst);
1918         /* check all nodes */
1919         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1920                 be_ilpsched_irn_t    *node;
1921                 ilpsched_node_attr_t *na;
1922                 unsigned             t;
1923                 ir_node              **pk;
1924
1925                 /* only mode_T nodes can have Projs and Keeps assigned */
1926                 if (get_irn_mode(irn) != mode_T)
1927                         continue;
1928
1929                 node = get_ilpsched_irn(env, irn);
1930                 na   = get_ilpsched_node_attr(node);
1931
1932                 /* check if has some Projs and Keeps assigned */
1933                 if (! na->projkeeps)
1934                         continue;
1935
1936                 /* we can run only once over the queue, so preserve the nodes */
1937                 pk = NEW_ARR_F(ir_node *, 0);
1938                 while (! waitq_empty(na->projkeeps))
1939                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1940                 del_waitq(na->projkeeps);
1941                 na->projkeeps = NULL;
1942
1943                 /* for all time steps at which this node can be scheduled */
1944                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1945                         int cst, tp_idx, i;
1946                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1947
1948                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1949                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1950                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1951                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1952                         num_cst++;
1953
1954                         /* sum up scheduling variables for this time step */
1955                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1956                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1957                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1958                         }
1959
1960                         if (ARR_LEN(tmp_var_idx_n) > 0)
1961                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1962                         DEL_ARR_F(tmp_var_idx_n);
1963
1964                         /* subtract all Proj and Keep variables for this step */
1965                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1966                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1967                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1968                                 int                  pk_tp_idx;
1969
1970                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1971                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1972                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1973                                 }
1974                         }
1975                 }
1976         }
1977         ilp_timer_pop();
1978         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1979                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1980 }
1981 #endif /* if 0 */
1982
1983 /***************************************************
1984  *  _____ _      _____                    _
1985  * |_   _| |    |  __ \                  (_)
1986  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1987  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1988  *  _| |_| |____| |      | | | | | | (_| | | | | |
1989  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1990  *
1991  ***************************************************/
1992
1993 /**
1994  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1995  */
1996 static void create_ilp(ir_node *block, void *walk_env) {
1997         be_ilpsched_env_t     *env           = walk_env;
1998         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1999         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
2000         FILE                  *logfile       = NULL;
2001         lpp_t                 *lpp           = NULL;
2002         struct obstack        var_obst;
2003         char                  name[1024];
2004
2005         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
2006         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
2007         DBG((env->dbg, 255, "=========================================\n\n"));
2008
2009         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
2010                 block, ba->n_interesting_nodes, ba->max_steps));
2011
2012         /* notify backend and get block environment */
2013         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
2014
2015         /* if we have less than two interesting nodes, there is no need to create the ILP */
2016         if (ba->n_interesting_nodes > 1) {
2017                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
2018                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
2019                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
2020                 int    estimated_n_var = (int)((double)base_num * fact_var);
2021                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
2022
2023                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
2024                         estimated_n_var, estimated_n_cst));
2025
2026                 /* set up the LPP object */
2027                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
2028
2029                 lpp = new_lpp_userdef(
2030                         (const char *)name,
2031                         lpp_minimize,
2032                         estimated_n_cst,     /* num vars */
2033                         estimated_n_cst + 1, /* num cst */
2034                         1.3);                /* grow factor */
2035                 obstack_init(&var_obst);
2036
2037                 /* create ILP variables */
2038                 create_variables(env, lpp, block_node, &var_obst);
2039
2040                 /* create ILP constraints */
2041                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
2042                 create_assignment_and_precedence_constraints(env, lpp, block_node);
2043                 create_ressource_constraints(env, lpp, block_node);
2044                 create_bundle_constraints(env, lpp, block_node);
2045                 create_branch_constraint(env, lpp, block_node);
2046                 //create_proj_keep_constraints(env, lpp, block_node);
2047
2048                 if (env->opts->regpress) {
2049                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
2050                                 create_alive_nodes_constraint(env, lpp, block_node);
2051                                 create_alive_livein_nodes_constraint(env, lpp, block_node);
2052                                 create_pressure_alive_constraint(env, lpp, block_node);
2053                         } else {
2054                                 create_dying_nodes_constraint(env, lpp, block_node);
2055                                 create_pressure_dead_constraint(env, lpp, block_node);
2056                         }
2057                 }
2058
2059                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
2060
2061                 /* debug stuff, dump lpp when debugging is on  */
2062                 DEBUG_ONLY(
2063                         if (firm_dbg_get_mask(env->dbg) > 1) {
2064                                 char buf[1024];
2065                                 FILE *f;
2066
2067                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
2068                                 f = fopen(buf, "w");
2069                                 lpp_dump_plain(lpp, f);
2070                                 fclose(f);
2071                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
2072                                 lpp_dump(lpp, buf);
2073                         }
2074                 );
2075
2076                 /* set solve time limit */
2077                 lpp_set_time_limit(lpp, env->opts->time_limit);
2078
2079                 /* set logfile if requested */
2080                 if (strlen(env->opts->log_file) > 0) {
2081                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
2082                                 lpp_set_log(lpp, stdout);
2083                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
2084                                 lpp_set_log(lpp, stderr);
2085                         else {
2086                                 logfile = fopen(env->opts->log_file, "w");
2087                                 if (! logfile)
2088                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
2089                                 else
2090                                         lpp_set_log(lpp, logfile);
2091                         }
2092                 }
2093
2094                 /* solve the ILP */
2095                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
2096
2097                 if (logfile)
2098                         fclose(logfile);
2099
2100                 /* check for valid solution */
2101                 if (! lpp_is_sol_valid(lpp)) {
2102                         char buf[1024];
2103                         FILE *f;
2104
2105                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block));
2106                         f = fopen(buf, "w");
2107                         lpp_dump_plain(lpp, f);
2108                         fclose(f);
2109                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block));
2110                         lpp_dump(lpp, buf);
2111                         dump_ir_block_graph(env->irg, "-assert");
2112
2113                         assert(0 && "ILP solution is not feasible!");
2114                 }
2115
2116                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
2117                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
2118                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
2119                 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));
2120                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
2121                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
2122                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
2123                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
2124
2125                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
2126         }
2127
2128         /* apply solution */
2129         apply_solution(env, lpp, block);
2130
2131         if (lpp)
2132                 free_lpp(lpp);
2133
2134         /* notify backend */
2135         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2136 }
2137
2138 /**
2139  * Perform ILP scheduling on the given irg.
2140  */
2141 void be_ilp_sched(const be_irg_t *birg) {
2142         be_ilpsched_env_t          env;
2143         const char                 *name = "be ilp scheduling";
2144         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
2145         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
2146
2147         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2148
2149 //      firm_dbg_set_mask(env.dbg, 1);
2150
2151         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
2152         env.sel        = sel;
2153         env.irg        = birg->irg;
2154         env.height     = heights_new(birg->irg);
2155         env.main_env   = birg->main_env;
2156         env.arch_env   = birg->main_env->arch_env;
2157         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
2158         env.opts       = &ilp_opts;
2159         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
2160
2161         /* assign a unique per block number to all interesting nodes */
2162         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2163
2164         /*
2165                 The block indices are completely build after the walk,
2166                 now we can allocate the bitsets (size depends on block indices)
2167                 for all nodes.
2168         */
2169         phase_reinit_irn_data(&env.ph);
2170
2171         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2172         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2173
2174         /* Calculate ALAP of all irns */
2175         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2176
2177         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2178         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2179
2180         /* perform ILP scheduling */
2181         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2182
2183         DEBUG_ONLY(
2184                 if (firm_dbg_get_mask(env.dbg)) {
2185                         phase_stat_t stat;
2186                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2187
2188                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2189                 }
2190         );
2191
2192         /* free data allocated dynamically */
2193         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2194
2195         /* free all allocated object */
2196         phase_free(&env.ph);
2197         heights_free(env.height);
2198
2199         /* notify backend */
2200         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2201 }
2202
2203 /**
2204  * Register ILP scheduler options.
2205  */
2206 void be_init_ilpsched(void)
2207 {
2208         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2209         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2210
2211         lc_opt_add_table(sched_grp, ilpsched_option_table);
2212 }
2213
2214 #else /* WITH_ILP */
2215
2216 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2217 {}
2218
2219 #endif /* WITH_ILP */