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