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
  49 typedef struct _ilpsched_options_t {
  50         unsigned limit_dead;
  51         unsigned time_limit;
  52         char     log_file[1024];
  53 } ilpsched_options_t;
  54
  55 typedef struct _unit_type_info_t {
  56         int                            n_units;
  57         const be_execution_unit_type_t *tp;
  58 } unit_type_info_t;
  59
  60 /**
  61  * holding the ILP variables of the different types
  62  */
  63 typedef struct _ilp_var_types_t {
  64         int *x;   /* x_{nt}^k variables */
  65         int *a;   /* a_{nt}^k variables */
  66         int *d;   /* d_{nt}^k variables */
  67         int *y;   /* y_{nt}^k variables */
  68 } ilp_var_types_t;
  69
  70 /* attributes for a node */
  71 typedef struct _ilpsched_node_attr_t {
  72         unsigned asap;                     /**< The ASAP scheduling control step */
  73         unsigned alap;                     /**< The ALAP scheduling control step */
  74         unsigned sched_point;              /**< the step in which the node is finally scheduled */
  75         unsigned visit_idx;                /**< Index of the node having visited this node last */
  76         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
  77         unsigned n_consumer;               /**< Number of consumers */
  78         ir_node  **block_consumer;         /**< List of consumer being in the same block */
  79         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
  80         unsigned block_idx     : 30;       /**< A unique per block index */
  81         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
  82         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
  83         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
  84                                                                                         for ASAP, successors for ALAP */
  85         unsigned n_unit_types;             /**< number of allowed execution unit types */
  86         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
  87         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
  88 } ilpsched_node_attr_t;
  89
  90 /* attributes for a block */
  91 typedef struct _ilpsched_block_attr_t {
  92         unsigned block_last_idx;        /**< The highest node index in block so far */
  93         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
  94         unsigned max_steps;             /**< Upper bound for block execution */
  95         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
  96         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
  97 } ilpsched_block_attr_t;
  98
  99 typedef union _ilpsched_attr_ {
 100         ilpsched_node_attr_t  node_attr;
 101         ilpsched_block_attr_t block_attr;
 102 } ilpsched_attr_t;
 103
 104 /* A irn for the phase and it's attributes (either node or block) */
 105 typedef struct {
 106         ir_node         *irn;
 107         ilpsched_attr_t attr;
 108 } be_ilpsched_irn_t;
 109
 110 /* The ILP scheduling environment */
 111 typedef struct {
 112         phase_t              ph;            /**< The phase */
 113         ir_graph             *irg;          /**< The current irg */
 114         heights_t            *height;       /**< The heights object of the irg */
 115         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 116         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 117         const arch_env_t     *arch_env;
 118         const arch_isa_t     *isa;          /**< The ISA */
 119         const be_main_env_t  *main_env;
 120         const be_machine_t   *cpu;          /**< the current abstract machine */
 121         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 122         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 123         DEBUG_ONLY(firm_dbg_module_t *dbg);
 124 } be_ilpsched_env_t;
 125
 126 /* convenience macros to handle phase irn data */
 127 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 128 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 129 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 130 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 131
 132 /* iterate over a list of ir_nodes linked by link field */
 133 #define foreach_linked_irns(head, iter) for ((iter) = (head); (iter); (iter) = get_irn_link((iter)))
 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_Jmp(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         70,    /* if we have more than 70 nodes: use alive nodes constraint */
 175         300,   /* 300 sec per block time limit */
 176         ""     /* no log file */
 177 };
 178
 179 #ifdef WITH_LIBCORE
 180 /* ILP options */
 181 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 182         LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead),
 183         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 184         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 185         { NULL }
 186 };
 187 #endif /* WITH_LIBCORE */
 188
 189 /*
 190         We need this global variable as we compare nodes dependent on heights,
 191         but we cannot pass any information to the qsort compare function.
 192 */
 193 static heights_t *glob_heights;
 194
 195 /**
 196  * Check if irn is a Proj, which has no execution units assigned.
 197  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 198  */
 199 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
 200         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
 201 }
 202
 203 /**
 204  * Skips normal Projs.
 205  * @return predecessor if irn is a normal Proj, otherwise irn.
 206  */
 207 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
 208         if (is_normal_Proj(isa, irn))
 209                 return get_Proj_pred(irn);
 210         return irn;
 211 }
 212
 213 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 214         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 215         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 216                 lat = 1;
 217         return lat;
 218 }
 219
 220
 221 /**
 222  * Compare scheduling time steps of two be_ilpsched_irn's.
 223  */
 224 static int cmp_ilpsched_irn(const void *a, const void *b) {
 225         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 226         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 227         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 228         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 229
 230         if (n1_a->sched_point == n2_a->sched_point) {
 231                 ir_node *irn_a = n1->irn;
 232                 ir_node *irn_b = n2->irn;
 233
 234                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 235                         return 1;
 236                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 237                         return -1;
 238                 return 0;
 239         }
 240         else
 241                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 242 }
 243
 244 /**
 245  * In case there is no phase information for irn, initialize it.
 246  */
 247 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
 248         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 249
 250         if (res == old) {
 251                 /* if we have already some data: check for reinitialization */
 252
 253                 if (! is_Block(irn)) {
 254                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 255
 256                         if (! na->transitive_block_nodes) {
 257                                 ir_node               *block      = get_nodes_block(irn);
 258                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 259                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 260
 261                                 /* we are called after the block indices have been build: create bitset */
 262                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 263                         }
 264                         else {
 265                                 /* we are called from reinit block data: clear the bitset */
 266                                 bitset_clear_all(na->transitive_block_nodes);
 267                                 na->visit_idx    = 0;
 268                                 na->alap_changed = 1;
 269                         }
 270                 }
 271                 return old;
 272         }
 273
 274         res->irn = irn;
 275
 276         /* set ilpsched irn attributes (either block or irn) */
 277         if (is_Block(irn)) {
 278                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 279
 280                 ba->n_interesting_nodes = 0;
 281                 ba->block_last_idx      = 0;
 282                 ba->root_nodes          = plist_new();
 283                 ba->head_ilp_nodes      = NULL;
 284                 ba->max_steps           = 0;
 285         }
 286         else {
 287                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 288                 memset(na, 0, sizeof(*na));
 289         }
 290
 291         return res;
 292 }
 293
 294 /**
 295  * Assign a per block unique number to each node.
 296  */
 297 static void build_block_idx(ir_node *irn, void *walk_env) {
 298         be_ilpsched_env_t     *env = walk_env;
 299         be_ilpsched_irn_t     *node, *block_node;
 300         ilpsched_node_attr_t  *na;
 301         ilpsched_block_attr_t *ba;
 302
 303         if (! consider_for_sched(env->arch_env->isa, irn))
 304                 return;
 305
 306         node       = get_ilpsched_irn(env, irn);
 307         na         = get_ilpsched_node_attr(node);
 308         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 309         ba         = get_ilpsched_block_attr(block_node);
 310
 311         na->block_idx = ba->block_last_idx++;
 312 }
 313
 314 /********************************************************
 315  *                              __        _
 316  *                             / /       | |
 317  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 318  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 319  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 320  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 321  *                 | |                           | |
 322  *                 |_|                           |_|
 323  ********************************************************/
 324
 325 /**
 326  * Add all nodes having no user in current block to last_nodes list.
 327  */
 328 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 329         ir_node               *block;
 330         const ir_edge_t       *edge;
 331         be_ilpsched_irn_t     *block_node, *node;
 332         ilpsched_block_attr_t *ba;
 333         ilpsched_node_attr_t  *na;
 334         int                   i, j;
 335         be_ilpsched_env_t     *env           = walk_env;
 336         int                   has_block_user = 0;
 337         unsigned              n_consumer     = 0;
 338         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 339         ir_node               **consumer;
 340         int                   idx;
 341
 342         if (! consider_for_sched(env->arch_env->isa, irn))
 343                 return;
 344
 345         block    = get_nodes_block(irn);
 346     idx      = get_irn_idx(irn);
 347         consumer = NEW_ARR_F(ir_node *, 0);
 348
 349         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 350
 351         /* check data and dependency out edges */
 352         for (i = 0; i < 2 && ! has_block_user; ++i) {
 353                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 354                         ir_node *user = get_edge_src_irn(edge);
 355
 356                         if (is_normal_Proj(env->arch_env->isa, user)) {
 357                                 const ir_edge_t *user_edge;
 358
 359                                 if (get_irn_mode(user) == mode_X)
 360                                         continue;
 361
 362                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 363                                 for (j = 0; j < 2; ++j) {
 364                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 365                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 366
 367                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 368                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 369                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 370
 371                                                         /* skip already visited nodes */
 372                                                         if (ua->consumer_idx == idx)
 373                                                                 continue;
 374
 375                                                         /* check if node has user in this block and collect the user if it's a data user */
 376                                                         if (get_nodes_block(real_user) == block) {
 377                                                                 if (i == 0 && j == 0)
 378                                                                         ARR_APP1(ir_node *, consumer, real_user);
 379                                                                 has_block_user = 1;
 380                                                         }
 381
 382                                                         /* only count data consumer */
 383                                                         if (i == 0)
 384                                                                 n_consumer++;
 385
 386                                                         /* mark user as visited by this node */
 387                                                         ua->consumer_idx = idx;
 388                                                 }
 389                                         }
 390                                 }
 391                         }
 392                         else if (is_Block(user)) {
 393                                 continue;
 394                         }
 395                         else if (! is_Phi(user)) {
 396                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 397                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 398
 399                                 /* skip already visited nodes */
 400                                 if (ua->consumer_idx == idx)
 401                                         continue;
 402
 403                                 /* check if node has user in this block and collect the user if it's a data user */
 404                                 if (get_nodes_block(user) == block) {
 405                                         if (i == 0)
 406                                                 ARR_APP1(ir_node *, consumer, user);
 407                                         has_block_user = 1;
 408                                 }
 409
 410                                 /* only count data consumer */
 411                                 if (i == 0)
 412                                         n_consumer++;
 413
 414                                 /* mark user visited by this node */
 415                                 ua->consumer_idx = idx;
 416                         }
 417                 }
 418         }
 419
 420         block_node = get_ilpsched_irn(env, block);
 421         ba         = get_ilpsched_block_attr(block_node);
 422
 423         ba->n_interesting_nodes++;
 424
 425         /* current irn has no user inside this block, add to queue */
 426         if (! has_block_user) {
 427                 DB((env->dbg, LEVEL_3, "root node\n"));
 428                 plist_insert_back(ba->root_nodes, irn);
 429         }
 430         else {
 431                 DB((env->dbg, LEVEL_3, "normal node\n"));
 432         }
 433
 434         /* record number of all consumer and the consumer within the same block */
 435         node = get_ilpsched_irn(env, irn);
 436         na   = get_ilpsched_node_attr(node);
 437         na->n_consumer     = n_consumer;
 438         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 439         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 440         DEL_ARR_F(consumer);
 441 }
 442
 443 /**
 444  * Calculate the ASAP scheduling step for current irn.
 445  */
 446 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 447         be_ilpsched_env_t     *env = walk_env;
 448         int                   i;
 449         ir_node               *block;
 450         be_ilpsched_irn_t     *node, *block_node;
 451         ilpsched_node_attr_t  *na;
 452         ilpsched_block_attr_t *ba;
 453
 454         /* These nodes are handled separate */
 455         if (! consider_for_sched(env->arch_env->isa, irn))
 456                 return;
 457
 458         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 459
 460         block    = get_nodes_block(irn);
 461         node     = get_ilpsched_irn(env, irn);
 462         na       = get_ilpsched_node_attr(node);
 463         na->asap = 1;
 464
 465         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 466                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
 467
 468                 /* check for greatest distance to top */
 469                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 470                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 471                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 472                         unsigned             lat;
 473
 474                         lat      = fixed_latency(env->sel, pred, env->block_env);
 475                         na->asap = MAX(na->asap, pna->asap + lat);
 476                 }
 477         }
 478
 479         /* add node to ILP node list and update max_steps */
 480         block_node = get_ilpsched_irn(env, block);
 481         ba         = get_ilpsched_block_attr(block_node);
 482
 483         set_irn_link(irn, ba->head_ilp_nodes);
 484         ba->head_ilp_nodes = irn;
 485         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 486
 487         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 488 }
 489
 490 /**
 491  * Calculate the ALAP scheduling step of all irns in current block.
 492  * Depends on max_steps being calculated.
 493  */
 494 static void calculate_block_alap(ir_node *block, void *walk_env) {
 495         be_ilpsched_env_t     *env        = walk_env;
 496         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 497         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 498         waitq                 *cur_queue  = new_waitq();
 499         plist_element_t       *el;
 500
 501         assert(is_Block(block));
 502
 503         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 504                 block, ba->n_interesting_nodes, ba->max_steps));
 505
 506         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 507         //phase_reinit_block_irn_data(&env->ph, block);
 508
 509         /* init start queue */
 510         foreach_plist(ba->root_nodes, el) {
 511                 waitq_put(cur_queue, plist_element_get_value(el));
 512         }
 513
 514         /* repeat until all nodes are processed */
 515         while (! waitq_empty(cur_queue)) {
 516                 waitq *next_queue = new_waitq();
 517
 518                 /* process all nodes in current step */
 519                 while (! waitq_empty(cur_queue)) {
 520                         ir_node              *cur_irn = waitq_get(cur_queue);
 521                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 522                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 523                         int                  i;
 524
 525                         /* cur_node has no alap set -> it's a root node, set to max alap */
 526                         if (na->alap == 0) {
 527                                 na->alap = ba->max_steps;
 528                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 529                                         cur_irn, na->alap));
 530                         }
 531                         else {
 532                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 533                                         cur_irn, na->alap));
 534                         }
 535
 536                         /* set the alap's of all predecessors */
 537                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 538                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
 539
 540                                 /* check for greatest distance to bottom */
 541                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 542                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 543                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 544                                         unsigned             lat;
 545
 546                                         /* mark the predecessor as visited by current irn */
 547                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 548                                                 continue;
 549                                         pna->visit_idx = get_irn_idx(cur_irn);
 550
 551                                         lat = fixed_latency(env->sel, pred, env->block_env);
 552
 553                                         /* set ALAP of current pred */
 554                                         if (pna->alap == 0) {
 555                                                 /* current ALAP is 0: node has not yet been visited */
 556                                                 pna->alap_changed = 1;
 557                                                 pna->alap         = na->alap - lat;
 558                                         }
 559                                         else if (pna->alap > na->alap - lat) {
 560                                                 /* we found a longer path to root node: change ALAP */
 561                                                 pna->alap         = na->alap - lat;
 562                                                 pna->alap_changed = 1;
 563                                         }
 564                                         else {
 565                                                 /* current ALAP is best found so far: keep it */
 566                                                 pna->alap_changed = 0;
 567                                         }
 568
 569                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 570
 571                                         /* enqueue node for next iteration */
 572                                         if (get_irn_ins_or_deps(pred) > 0)
 573                                                 waitq_put(next_queue, pred);
 574                                 }
 575                         }
 576                 }
 577
 578                 /* prepare for next iteration */
 579                 del_waitq(cur_queue);
 580                 cur_queue = next_queue;
 581         }
 582 }
 583
 584 /**
 585  * We can free the list of root nodes here.
 586  */
 587 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 588         be_ilpsched_env_t     *env        = walk_env;
 589         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 590         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 591
 592         plist_free(ba->root_nodes);
 593         ba->root_nodes = NULL;
 594 }
 595
 596 /**
 597  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 598  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 599  */
 600 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 601         be_ilpsched_env_t    *env  = walk_env;
 602         ir_node              *pred = irn;
 603         be_ilpsched_irn_t    *node, *pred_node;
 604         ilpsched_node_attr_t *na, *pna;
 605
 606         if (! consider_for_sched(env->arch_env->isa, irn))
 607                 return;
 608
 609         if (! is_Proj(irn) && ! be_is_Keep(irn))
 610                 return;
 611
 612         /* go to the ancestor */
 613         if (be_is_Keep(irn))
 614                 pred = get_irn_n(irn, 0);
 615         pred = skip_Proj(pred);
 616
 617         node      = get_ilpsched_irn(env, irn);
 618         pred_node = get_ilpsched_irn(env, pred);
 619         na        = get_ilpsched_node_attr(node);
 620         pna       = get_ilpsched_node_attr(pred_node);
 621
 622         na->asap = pna->asap;
 623         na->alap = pna->alap;
 624
 625         /* record all Projs and Keeps for this node */
 626         if (! pna->projkeeps)
 627                 pna->projkeeps = new_waitq();
 628         waitq_put(pna->projkeeps, irn);
 629
 630         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 631 }
 632
 633 /*******************************************
 634  *           _              _       _
 635  *          | |            | |     | |
 636  *  ___  ___| |__   ___  __| |_   _| | ___
 637  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 638  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 639  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 640  *
 641  *******************************************/
 642
 643 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
 644         const ir_edge_t *edge;
 645
 646         foreach_out_edge(irn, edge) {
 647                 ir_node *user = get_edge_src_irn(edge);
 648
 649                 if (be_is_Keep(user)) {
 650                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 651                         waitq_put(keeps, user);
 652                 }
 653         }
 654 }
 655
 656 /**
 657  * Inserts @p irn before @p before into schedule and notifies backend.
 658  */
 659 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 660         ir_node *before, ir_node *irn, unsigned cycle)
 661 {
 662         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 663         sched_add_before(before, irn);
 664 }
 665
 666 /**
 667  * Adds a node, it's Projs (in case of mode_T nodes) and
 668  * it's Keeps to schedule.
 669  */
 670 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
 671         const ir_edge_t *edge;
 672         waitq           *keeps = new_waitq();
 673
 674         /* mode_M nodes are not scheduled */
 675         if (get_irn_mode(irn) == mode_M)
 676                 return;
 677
 678         if (! sched_is_scheduled(irn))
 679                 notified_sched_add_before(env, block, irn, cycle);
 680
 681         /* add Projs */
 682         if (get_irn_mode(irn) == mode_T) {
 683                 foreach_out_edge(irn, edge) {
 684                         ir_node *user = get_edge_src_irn(edge);
 685
 686                         if (to_appear_in_schedule(user) || get_irn_mode(user) == mode_b)
 687                                 notified_sched_add_before(env, block, user, cycle);
 688
 689                         check_for_keeps(keeps, block, user);
 690                 }
 691         }
 692         else {
 693                 check_for_keeps(keeps, block, irn);
 694         }
 695
 696         /* add Keeps */
 697         while (! waitq_empty(keeps)) {
 698                 ir_node *keep = waitq_get(keeps);
 699                 if (! sched_is_scheduled(keep))
 700                         notified_sched_add_before(env, block, keep, cycle);
 701         }
 702
 703         del_waitq(keeps);
 704 }
 705
 706 /**
 707  * Schedule all nodes in the given block, according to the ILP solution.
 708  */
 709 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 710         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 711         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 712         sched_info_t          *info       = get_irn_sched_info(block);
 713         be_ilpsched_irn_t     **sched_nodes;
 714         unsigned              i, l;
 715         ir_node               *cfop, *irn;
 716         const ir_edge_t       *edge;
 717
 718         /* init block schedule list */
 719         INIT_LIST_HEAD(&info->list);
 720         info->scheduled = 1;
 721
 722         /* collect nodes and their scheduling time step */
 723         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 724         if (ba->n_interesting_nodes == 0) {
 725                 /* ignore */
 726         }
 727         else if (ba->n_interesting_nodes == 1) {
 728                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 729
 730                 /* add the single node */
 731                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 732         }
 733         else {
 734                 /* check all nodes for their positive solution */
 735                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 736                         be_ilpsched_irn_t    *node;
 737                         ilpsched_node_attr_t *na;
 738                         int                  tp_idx, found;
 739                         unsigned             cur_var, t;
 740
 741                         node    = get_ilpsched_irn(env, irn);
 742                         na      = get_ilpsched_node_attr(node);
 743                         cur_var = 0;
 744                         found   = 0;
 745
 746                         /* go over all variables of a node until the non-zero one is found */
 747                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 748                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 749                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 750
 751                                         /* check, if variable is set to one (it's not zero then :) */
 752                                         if (! LPP_VALUE_IS_0(val)) {
 753                                                 na->sched_point = t;
 754                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 755                                                 DBG((env->dbg, LEVEL_1, "Schedpoint of %+F is %u at unit type %s\n",
 756                                                         irn, t, na->type_info[tp_idx].tp->name));
 757                                                 found = 1;
 758                                         }
 759                                 }
 760                         }
 761                 }
 762
 763                 glob_heights = heights_new(env->irg);
 764                 /* sort nodes ascending by scheduling time step */
 765                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 766                 heights_free(glob_heights);
 767         }
 768
 769         /* make all Phis ready and remember the single cf op */
 770         cfop = NULL;
 771         foreach_out_edge(block, edge) {
 772                 irn = get_edge_src_irn(edge);
 773
 774                 switch (get_irn_opcode(irn)) {
 775                         case iro_Phi:
 776                                 add_to_sched(env, block, irn, 0);
 777                                 break;
 778                         case iro_Start:
 779                         case iro_End:
 780                         case iro_Proj:
 781                         case iro_Bad:
 782                                 break;
 783                         default:
 784                                 if (is_cfop(irn)) {
 785                                         assert(cfop == NULL && "Highlander - there can be only one");
 786                                         cfop = irn;
 787                                 }
 788                         break;
 789                 }
 790         }
 791
 792         /* add all nodes from list */
 793         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 794                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 795                 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 796         }
 797
 798         /* schedule control flow node if not already done */
 799         if (cfop && ! sched_is_scheduled(cfop))
 800                 add_to_sched(env, block, cfop, 0);
 801
 802         DEL_ARR_F(sched_nodes);
 803 }
 804
 805 /***************************************************************
 806  *   _____ _      _____     _____           _   _
 807  *  |_   _| |    |  __ \   / ____|         | | (_)
 808  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 809  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 810  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 811  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 812  *
 813  ***************************************************************/
 814
 815 /**
 816  * Check if node can be executed on given unit type.
 817  */
 818 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 819         int                  i;
 820         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 821
 822         for (i = na->n_unit_types - 1; i >= 0; --i) {
 823                 if (na->type_info[i].tp == tp)
 824                         return i;
 825         }
 826
 827         return -1;
 828 }
 829
 830 /************************************************
 831  *                   _       _     _
 832  *                  (_)     | |   | |
 833  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 834  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 835  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 836  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 837  *
 838  ************************************************/
 839
 840 /**
 841  * Create the following variables:
 842  * - x_{nt}^k    binary     weigthed with: t
 843  *      node n is scheduled at time step t to unit type k
 844  * ==>> These variables represent the schedule
 845  *
 846  * - d_{nt}^k    binary     weighted with: t
 847  *      node n dies at time step t on unit type k
 848  * - a_{nt}^k    binary     weighted with num_nodes
 849  *      node n is alive at time step t on unit type k
 850  *
 851  * - y_{nt}^k    binary     weighted with: num_nodes^2
 852  *      node n is scheduled at time step t to unit type k
 853  *      although all units of this type are occupied
 854  * ==>> These variables represent the register pressure
 855  *
 856  */
 857 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 858         char                  buf[1024];
 859         ir_node               *irn;
 860         unsigned              num_block_var, num_nodes;
 861         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 862         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 863 #ifdef WITH_LIBCORE
 864         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
 865 #endif /* WITH_LIBCORE */
 866
 867         ilp_timer_push(t_var);
 868         num_block_var = num_nodes = 0;
 869         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 870                 const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
 871                 be_ilpsched_irn_t         *node;
 872                 ilpsched_node_attr_t      *na;
 873                 unsigned                  n_unit_types, tp_idx, unit_idx, n_var, cur_unit;
 874                 unsigned                  cur_var_ad, cur_var_x, cur_var_y, num_ad;
 875
 876                 /* count number of available unit types for this node */
 877                 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 878                         /* just count */ ;
 879
 880                 node = get_ilpsched_irn(env, irn);
 881                 na   = get_ilpsched_node_attr(node);
 882
 883                 na->n_unit_types = n_unit_types;
 884                 na->type_info    = NEW_ARR_D(unit_type_info_t, var_obst, n_unit_types);
 885
 886                 /* fill the type info array */
 887                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 888                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 889                                 /* beware: we also count number of available units here */
 890                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 891                                         na->is_dummy_node = 1;
 892                         }
 893
 894                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 895                         na->type_info[tp_idx].n_units = unit_idx;
 896                 }
 897
 898                 /* allocate space for ilp variables */
 899                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 900                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 901
 902                 /* we need these variables only for "real" nodes */
 903                 if (! na->is_dummy_node) {
 904                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 905                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 906
 907                         num_ad = ba->max_steps - na->asap + 1;
 908
 909                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 910                                 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 911                                 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
 912                         }
 913                         else {
 914                                 na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 915                                 memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0]));
 916                         }
 917                 }
 918
 919                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
 920                         irn, na->asap, na->alap, na->n_unit_types));
 921
 922                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
 923                 /* create variables */
 924                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 925                         unsigned t;
 926
 927                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
 928                                 /* x_{nt}^k variables */
 929                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
 930                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 931                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 932                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 933                                 /* variable counter */
 934                                 n_var++;
 935                                 num_block_var++;
 936
 937                                 if (! na->is_dummy_node) {
 938                                         /* y_{nt}^k variables */
 939                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
 940                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 941                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_binary, (double)(weigth_y));
 942                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 943
 944                                         /* variable counter */
 945                                         n_var++;
 946                                         num_block_var++;
 947                                 }
 948                         }
 949
 950                         /* a node can die at any step t: asap(n) <= t <= U */
 951                         if (! na->is_dummy_node) {
 952                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
 953
 954                                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 955                                                 /* a_{nt}^k variables */
 956                                                 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
 957                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 958                                                 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
 959                                         }
 960                                         else {
 961                                                 /* d_{nt}^k variables */
 962                                                 snprintf(buf, sizeof(buf), "d_n%u_%s_%u",
 963                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 964                                                 na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 965                                         }
 966                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 967
 968                                         /* variable counter */
 969                                         n_var++;
 970                                         num_block_var++;
 971                                 }
 972                         }
 973                 }
 974
 975                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
 976                 num_nodes++;
 977         }
 978         ilp_timer_pop();
 979         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
 980                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
 981 }
 982
 983 /*******************************************************
 984  *                      _             _       _
 985  *                     | |           (_)     | |
 986  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
 987  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
 988  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
 989  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
 990  *
 991  *******************************************************/
 992
 993 /**
 994  * Create following ILP constraints:
 995  * - the assignment constraints:
 996  *     assure each node is executed once by exactly one (allowed) execution unit
 997  * - the dead node assignment constraints:
 998  *     assure a node can only die at most once
 999  * - the precedence constraints:
1000  *     assure that no data dependencies are violated
1001  */
1002 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1003         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1004         char                  buf[1024];
1005         ir_node               *irn;
1006         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1007         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1008 #ifdef WITH_LIBCORE
1009         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign",      "create assignment constraints");
1010         lc_timer_t            *t_cst_dead    = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints");
1011         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",        "create precedence constraints");
1012 #endif /* WITH_LIBCORE */
1013
1014         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1015         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1016                 int                  cst, tp_idx, i;
1017                 unsigned             cur_var;
1018                 be_ilpsched_irn_t    *node;
1019                 ilpsched_node_attr_t *na;
1020
1021                 node    = get_ilpsched_irn(env, irn);
1022                 na      = get_ilpsched_node_attr(node);
1023                 cur_var = 0;
1024
1025                 /* the assignment constraint */
1026                 ilp_timer_push(t_cst_assign);
1027                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1028                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1029                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1030                 num_cst_assign++;
1031
1032                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1033                 ilp_timer_pop();
1034
1035                 /* the dead node assignment constraint */
1036                 if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) {
1037                         ilp_timer_push(t_cst_dead);
1038                         snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn));
1039                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1040                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1041
1042                         lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0);
1043                         ilp_timer_pop();
1044                 }
1045
1046                 /* We have separate constraints for Projs and Keeps */
1047                 // ILP becomes infeasible ?!?
1048 //              if (is_Proj(irn) || be_is_Keep(irn))
1049 //                      continue;
1050
1051                 /* the precedence constraints */
1052                 ilp_timer_push(t_cst_prec);
1053                 bs_block_irns = bitset_clear_all(bs_block_irns);
1054                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1055                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1056                         unsigned             t_low, t_high, t;
1057                         be_ilpsched_irn_t    *pred_node;
1058                         ilpsched_node_attr_t *pna;
1059                         unsigned             delay;
1060
1061                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1062                                 continue;
1063
1064                         pred_node = get_ilpsched_irn(env, pred);
1065                         pna       = get_ilpsched_node_attr(pred_node);
1066
1067                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1068
1069                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1070                                 bitset_set(bs_block_irns, pna->block_idx);
1071                         else
1072                                 continue;
1073
1074                         /* irn = n, pred = m */
1075                         delay  = fixed_latency(env->sel, pred, env->block_env);
1076                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1077                         t_high = MIN(na->alap, pna->alap + delay - 1);
1078                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1079                                 unsigned tn, tm;
1080                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1081
1082                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1083                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1084                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1085                                 num_cst_prec++;
1086
1087                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1088                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1089                                         /* node variables have smaller index than pred variables */
1090                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1091                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1092                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1093                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1094                                                 }
1095                                         }
1096
1097                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1098                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1099                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1100                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1101                                                 }
1102                                         }
1103                                 }
1104                                 else {
1105                                         /* pred variables have smaller index than node variables */
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                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1114                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1115                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1116                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1117                                                 }
1118                                         }
1119                                 }
1120
1121                                 if (ARR_LEN(tmp_var_idx) > 0)
1122                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1123
1124                                 DEL_ARR_F(tmp_var_idx);
1125                         }
1126                 }
1127                 ilp_timer_pop();
1128         }
1129         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1130                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1131         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1132                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1133 }
1134
1135 /**
1136  * Create ILP resource constraints:
1137  * - assure that for each time step not more instructions are scheduled
1138  *   to the same unit types as units of this type are available
1139  */
1140 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1141         int                   glob_type_idx;
1142         char                  buf[1024];
1143         unsigned              num_cst_resrc = 0;
1144         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1145 #ifdef WITH_LIBCORE
1146         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1147 #endif /* WITH_LIBCORE */
1148
1149         ilp_timer_push(t_cst_rsrc);
1150         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1151                 unsigned                 t;
1152                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1153
1154                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1155
1156                 /* check each time step */
1157                 for (t = 0; t < ba->max_steps; ++t) {
1158                         ir_node *irn;
1159                         int     cst;
1160                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1161
1162                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1163                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1164                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1165                         num_cst_resrc++;
1166
1167                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1168                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1169                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1170                                 int                  tp_idx;
1171
1172                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1173
1174                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1175                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1176                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1177                                 }
1178                         }
1179
1180                         /* set constraints if we have some */
1181                         if (ARR_LEN(tmp_var_idx) > 0)
1182                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1183
1184                         DEL_ARR_F(tmp_var_idx);
1185                 }
1186         }
1187         ilp_timer_pop();
1188         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1189                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1190 }
1191
1192 /**
1193  * Create ILP bundle constraints:
1194  * - assure, at most bundle_size * bundles_per_cycle instructions
1195  *   can be started at a certain point.
1196  */
1197 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1198         char                  buf[1024];
1199         unsigned              t;
1200         unsigned              num_cst_bundle = 0;
1201         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1202         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1203 #ifdef WITH_LIBCORE
1204         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1205 #endif /* WITH_LIBCORE */
1206
1207         ilp_timer_push(t_cst_bundle);
1208         for (t = 0; t < ba->max_steps; ++t) {
1209                 ir_node *irn;
1210                 int     cst;
1211                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1212
1213                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1214                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1215                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1216                 num_cst_bundle++;
1217
1218                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1219                         be_ilpsched_irn_t    *node;
1220                         ilpsched_node_attr_t *na;
1221                         int                  tp_idx;
1222
1223                         /* Projs and Keeps do not contribute to bundle size */
1224                         if (is_Proj(irn) || be_is_Keep(irn))
1225                                 continue;
1226
1227                         node = get_ilpsched_irn(env, irn);
1228                         na   = get_ilpsched_node_attr(node);
1229
1230                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1231                         if (na->is_dummy_node)
1232                                 continue;
1233
1234                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1235                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1236                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1237                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1238                                 }
1239                         }
1240                 }
1241
1242                 if (ARR_LEN(tmp_var_idx) > 0)
1243                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1244
1245                 DEL_ARR_F(tmp_var_idx);
1246         }
1247         ilp_timer_pop();
1248         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1249                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1250 }
1251
1252 /**
1253  * Create ILP dying nodes constraints:
1254  * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k
1255  */
1256 static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1257         char                  buf[1024];
1258         unsigned              t;
1259         unsigned              num_cst = 0;
1260         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1261 #ifdef WITH_LIBCORE
1262         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints");
1263 #endif /* WITH_LIBCORE */
1264
1265         ilp_timer_push(t_cst);
1266         /* check all time_steps */
1267         for (t = 0; t < ba->max_steps; ++t) {
1268                 ir_node *irn;
1269
1270                 /* for all nodes */
1271                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1272                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1273                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1274
1275                         /* if node has no consumer within current block, it cannot die here */
1276                         /* we also ignore nodes assigned to dummy unit */
1277                         if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node)
1278                                 continue;
1279
1280                         /* node can only die here if t at least asap(n) */
1281                         if (t >= na->asap - 1) {
1282                                 int node_tp_idx;
1283
1284                                 /* for all unit types */
1285                                 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1286                                         int tp_idx, i, cst;
1287                                         int *tmp_var_idx = NEW_ARR_F(int, 0);
1288
1289                                         snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn));
1290                                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1));
1291                                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1292                                         num_cst++;
1293
1294                                         /* number of consumer scheduled till t */
1295                                         for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1296                                                 be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1297                                                 ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1298
1299                                                 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1300                                                         unsigned tm;
1301
1302                                                         for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) {
1303                                                                 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1304                                                                 ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]);
1305                                                         }
1306                                                 }
1307                                         }
1308
1309                                         /* could be that no consumer can be scheduled at this point */
1310                                         if (ARR_LEN(tmp_var_idx)) {
1311                                                 int      idx;
1312                                                 unsigned tn;
1313
1314                                                 /* subtract possible prior kill points */
1315                                                 for (tn = na->asap - 1; tn < t; ++tn) {
1316                                                         idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn);
1317                                                         lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0);
1318                                                 }
1319
1320                                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1321                                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer));
1322                                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1323                                         }
1324
1325                                         DEL_ARR_F(tmp_var_idx);
1326                                 }
1327                         }
1328
1329                 }
1330         }
1331         ilp_timer_pop();
1332         DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n",
1333                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1334 }
1335
1336 /**
1337 * Create ILP alive nodes constraints:
1338 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1339 */
1340 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1341         char                  buf[1024];
1342         ir_node               *irn;
1343         unsigned              num_cst = 0;
1344         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1345 #ifdef WITH_LIBCORE
1346         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1347 #endif /* WITH_LIBCORE */
1348
1349         ilp_timer_push(t_cst);
1350         /* for each node */
1351         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1352                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1353                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1354                 unsigned             t;
1355
1356                 /* we ignore nodes assigned to dummy unit here */
1357                 if (na->is_dummy_node)
1358                         continue;
1359
1360                 /* check check all time steps: asap(n) <= t <= U */
1361                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1362                         int node_tp_idx;
1363
1364                         /* for all unit types available for this node */
1365                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1366                                 unsigned tn, tn_max, idx;
1367                                 int      cst, i;
1368                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1369                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1370
1371                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1372                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1373                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1374                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1375                                 num_cst++;
1376
1377                                 tn_max = MIN(na->alap - 1, t);
1378                                 /* check if the node has been scheduled so far */
1379                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1380                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1381                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1382                                 }
1383
1384                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1385                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1386                                 DEL_ARR_F(tmp_var_idx_n);
1387
1388                                 /* subtract the number of consumer scheduled so far */
1389                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1390                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1391                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1392                                         int                  tp_idx;
1393                                         unsigned             tm, tm_max;
1394
1395                                         tm_max = MIN(ca->alap - 1, t);
1396                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1397                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1398                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1399                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1400                                                 }
1401                                         }
1402                                 }
1403
1404                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1405                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1406                                 DEL_ARR_F(tmp_var_idx_m);
1407
1408                                 /* -c * a_{nt}^k */
1409                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1410                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1411
1412                         }
1413                 }
1414         }
1415         ilp_timer_pop();
1416         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1417                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1418 }
1419
1420 /**
1421  * Create ILP pressure constraints, based on dead nodes:
1422  * - add additional costs to objective function if a node is scheduled
1423  *   on a unit although all units of this type are currently occupied
1424  */
1425 static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1426         char                  buf[1024];
1427         ir_node               *cur_irn;
1428         unsigned              num_cst = 0;
1429         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1430 #ifdef WITH_LIBCORE
1431         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1432 #endif /* WITH_LIBCORE */
1433
1434         ilp_timer_push(t_cst);
1435         /* y_{nt}^k is set for each node and timestep and unit type */
1436         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1437                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1438                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1439                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1440                 int                  glob_type_idx;
1441
1442                 /* we ignore nodes assigned to DUMMY unit here */
1443                 if (cur_na->is_dummy_node)
1444                         continue;
1445
1446                 /* for all types */
1447                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1448                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1449                         int                      cur_tp_idx;
1450                         unsigned                 t;
1451
1452                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1453
1454                         /* check if node can be executed on this unit type */
1455                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1456                         if (cur_tp_idx < 0)
1457                                 continue;
1458
1459                         /* check all time_steps */
1460                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1461                                 int     cst, y_idx;
1462                                 ir_node *irn;
1463                                 int     *tmp_idx_1  = NEW_ARR_F(int, 0);
1464                                 int     *tmp_idx_m1 = NEW_ARR_F(int, 0);
1465
1466                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1467                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1468                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1469                                 num_cst++;
1470
1471                                 /*
1472                                         - accumulate all nodes scheduled on unit type k till t
1473                                         - subtract all nodes died on unit type k till t
1474                                 */
1475                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1476                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1477                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1478                                         unsigned             tn, tmax;
1479                                         int                  tp_idx;
1480
1481                                         tmax   = MIN(t, na->alap - 1);
1482                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1483
1484                                         /* current unit type is not suitable for current node */
1485                                         if (tp_idx < 0)
1486                                                 continue;
1487
1488                                         for (tn = na->asap - 1; tn <= tmax; ++tn) {
1489                                                 int idx;
1490
1491                                                 /* node scheduled */
1492                                                 idx = ILPVAR_IDX(na, tp_idx, tn);
1493                                                 ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]);
1494
1495                                                 /* node dead */
1496                                                 idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn);
1497                                                 ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]);
1498                                         }
1499                                 }
1500
1501                                 if (ARR_LEN(tmp_idx_1) > 0)
1502                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0);
1503
1504                                 if (ARR_LEN(tmp_idx_m1) > 0)
1505                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0);
1506
1507                                 /* BEWARE: t is unsigned, so (double)(-t) won't work */
1508                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1509                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(t));
1510
1511                                 DEL_ARR_F(tmp_idx_1);
1512                                 DEL_ARR_F(tmp_idx_m1);
1513                         }
1514                 }
1515         }
1516         ilp_timer_pop();
1517         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1518                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1519 }
1520
1521 /**
1522  * Create ILP pressure constraints, based on alive nodes:
1523  * - add additional costs to objective function if a node is scheduled
1524  *   on a unit although all units of this type are currently occupied
1525  */
1526 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1527         char                  buf[1024];
1528         ir_node               *cur_irn;
1529         unsigned              num_cst = 0;
1530         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1531 #ifdef WITH_LIBCORE
1532         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1533 #endif /* WITH_LIBCORE */
1534
1535         ilp_timer_push(t_cst);
1536         /* y_{nt}^k is set for each node and timestep and unit type */
1537         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1538                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1539                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1540                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1541                 int                  glob_type_idx;
1542
1543                 /* we ignore nodes assigned to DUMMY unit here */
1544                 if (cur_na->is_dummy_node)
1545                         continue;
1546
1547                 /* for all types */
1548                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1549                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1550                         int                      cur_tp_idx;
1551                         unsigned                 t;
1552
1553                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1554
1555                         /* check if node can be executed on this unit type */
1556                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1557                         if (cur_tp_idx < 0)
1558                                 continue;
1559
1560                         /* check all time_steps at which the current node can be scheduled */
1561                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1562                                 int     cst, y_idx;
1563                                 ir_node *irn;
1564                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1565
1566                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1567                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1568                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1569                                 num_cst++;
1570
1571                                 /* - accumulate all nodes alive at point t on unit type k */
1572                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1573                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1574                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1575                                         int                  a_idx, tp_idx;
1576
1577                                         /* check if node can be alive here */
1578                                         if (t < na->asap - 1)
1579                                                 continue;
1580
1581                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1582
1583                                         /* current type is not suitable */
1584                                         if (tp_idx < 0)
1585                                                 continue;
1586
1587                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1588                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1589                                 }
1590
1591                                 if (ARR_LEN(tmp_var_idx) > 0)
1592                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1593                                 DEL_ARR_F(tmp_var_idx);
1594
1595                                 /* - num_nodes * y_{nt}^k */
1596                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1597                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(ba->n_interesting_nodes));
1598                         }
1599                 }
1600         }
1601         ilp_timer_pop();
1602         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1603                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1604 }
1605
1606 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1607         char                  buf[1024];
1608         ir_node               *irn;
1609         unsigned              num_cst = 0;
1610         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1611 #ifdef WITH_LIBCORE
1612         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1613 #endif /* WITH_LIBCORE */
1614
1615         ilp_timer_push(t_cst);
1616         /* check all nodes */
1617         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1618                 be_ilpsched_irn_t    *node;
1619                 ilpsched_node_attr_t *na;
1620                 unsigned             t;
1621                 ir_node              **pk;
1622
1623                 /* only mode_T nodes can have Projs and Keeps assigned */
1624                 if (get_irn_mode(irn) != mode_T)
1625                         continue;
1626
1627                 node = get_ilpsched_irn(env, irn);
1628                 na   = get_ilpsched_node_attr(node);
1629
1630                 /* check if has some Projs and Keeps assigned */
1631                 if (! na->projkeeps)
1632                         continue;
1633
1634                 /* we can run only once over the queue, so preserve the nodes */
1635                 pk = NEW_ARR_F(ir_node *, 0);
1636                 while (! waitq_empty(na->projkeeps))
1637                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1638                 del_waitq(na->projkeeps);
1639                 na->projkeeps = NULL;
1640
1641                 /* for all time steps at which this node can be scheduled */
1642                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1643                         int cst, tp_idx, i;
1644                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1645
1646                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1647                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1648                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1649                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1650                         num_cst++;
1651
1652                         /* sum up scheduling variables for this time step */
1653                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1654                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1655                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1656                         }
1657
1658                         if (ARR_LEN(tmp_var_idx_n) > 0)
1659                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1660                         DEL_ARR_F(tmp_var_idx_n);
1661
1662                         /* subtract all Proj and Keep variables for this step */
1663                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1664                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1665                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1666                                 int                  pk_tp_idx;
1667
1668                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1669                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1670                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1671                                 }
1672                         }
1673                 }
1674         }
1675         ilp_timer_pop();
1676         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1677                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1678 }
1679
1680 /***************************************************
1681  *  _____ _      _____                    _
1682  * |_   _| |    |  __ \                  (_)
1683  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1684  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1685  *  _| |_| |____| |      | | | | | | (_| | | | | |
1686  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1687  *
1688  ***************************************************/
1689
1690 /**
1691  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1692  */
1693 static void create_ilp(ir_node *block, void *walk_env) {
1694         be_ilpsched_env_t     *env           = walk_env;
1695         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1696         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1697         FILE                  *logfile       = NULL;
1698         lpp_t                 *lpp           = NULL;
1699         struct obstack        var_obst;
1700
1701         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1702         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1703         DBG((env->dbg, 255, "=========================================\n\n"));
1704
1705         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1706                 block, ba->n_interesting_nodes, ba->max_steps));
1707
1708         /* notify backend and get block environment */
1709         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1710
1711         /* if we have less than two interesting nodes, there is no need to create the ILP */
1712         if (ba->n_interesting_nodes > 1) {
1713                 double fact_var        = ba->n_interesting_nodes > 25 ? 1.1 : 1.2;
1714                 double fact_cst        = ba->n_interesting_nodes > 25 ? 0.7 : 1.5;
1715                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1716                 int    estimated_n_var = (int)((double)base_num * fact_var);
1717                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1718
1719                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimed numbers: %d vars, %d cst\n",
1720                         estimated_n_var, estimated_n_cst));
1721
1722                 /* set up the LPP object */
1723                 lpp = new_lpp_userdef(
1724                         "be ilp scheduling",
1725                         lpp_minimize,
1726                         estimated_n_cst + 1,  /* num vars */
1727                         estimated_n_cst + 20, /* num cst */
1728                         1.2);                 /* grow factor */
1729                 obstack_init(&var_obst);
1730
1731                 /* create ILP variables */
1732                 create_variables(env, lpp, block_node, &var_obst);
1733
1734                 /* create ILP constraints */
1735                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1736                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1737                 create_ressource_constraints(env, lpp, block_node);
1738                 create_bundle_constraints(env, lpp, block_node);
1739                 //create_proj_keep_constraints(env, lpp, block_node);
1740                 if (ba->n_interesting_nodes > env->opts->limit_dead) {
1741                         create_alive_nodes_constraint(env, lpp, block_node);
1742                         create_pressure_alive_constraint(env, lpp, block_node);
1743                 } else {
1744                         create_dying_nodes_constraint(env, lpp, block_node);
1745                         create_pressure_dead_constraint(env, lpp, block_node);
1746                 }
1747
1748                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1749
1750                 /* debug stuff, dump lpp when debugging is on  */
1751                 DEBUG_ONLY(
1752                         if (firm_dbg_get_mask(env->dbg) > 0) {
1753                                 char buf[1024];
1754                                 FILE *f;
1755
1756                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1757                                 f = fopen(buf, "w");
1758                                 lpp_dump_plain(lpp, f);
1759                                 fclose(f);
1760                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1761                                 lpp_dump(lpp, buf);
1762                         }
1763                 );
1764
1765                 /* set solve time limit */
1766                 lpp_set_time_limit(lpp, env->opts->time_limit);
1767
1768                 /* set logfile if requested */
1769                 if (strlen(env->opts->log_file) > 0) {
1770                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1771                                 lpp_set_log(lpp, stdout);
1772                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1773                                 lpp_set_log(lpp, stderr);
1774                         else {
1775                                 logfile = fopen(env->opts->log_file, "w");
1776                                 if (! logfile)
1777                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1778                                 else
1779                                         lpp_set_log(lpp, logfile);
1780                         }
1781                 }
1782
1783                 /* solve the ILP */
1784                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1785
1786                 if (logfile)
1787                         fclose(logfile);
1788
1789                 /* check for valid solution */
1790                 if (! lpp_is_sol_valid(lpp)) {
1791                         char buf[1024];
1792                         FILE *f;
1793
1794                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block));
1795                         f = fopen(buf, "w");
1796                         lpp_dump_plain(lpp, f);
1797                         fclose(f);
1798                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block));
1799                         lpp_dump(lpp, buf);
1800                         dump_ir_block_graph(env->irg, "-assert");
1801
1802                         assert(0 && "ILP solution is not feasible!");
1803                 }
1804
1805                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1806                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1807                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1808                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1809                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1810                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1811                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1812
1813                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1814         }
1815
1816         /* apply solution */
1817         apply_solution(env, lpp, block);
1818
1819         if (lpp)
1820                 free_lpp(lpp);
1821
1822         /* notify backend */
1823         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1824 }
1825
1826 /**
1827  * Perform ILP scheduling on the given irg.
1828  */
1829 void be_ilp_sched(const be_irg_t *birg) {
1830         be_ilpsched_env_t          env;
1831         const char                 *name = "be ilp scheduling";
1832         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
1833         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
1834
1835         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
1836
1837         //firm_dbg_set_mask(env.dbg, 31);
1838
1839         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
1840         env.sel        = sel;
1841         env.irg        = birg->irg;
1842         env.height     = heights_new(birg->irg);
1843         env.main_env   = birg->main_env;
1844         env.arch_env   = birg->main_env->arch_env;
1845         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
1846         env.opts       = &ilp_opts;
1847         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
1848
1849         /* assign a unique per block number to all interesting nodes */
1850         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
1851
1852         /*
1853                 The block indices are completely build after the walk,
1854                 now we can allocate the bitsets (size depends on block indices)
1855                 for all nodes.
1856         */
1857         phase_reinit_irn_data(&env.ph);
1858
1859         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
1860         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
1861
1862         /* Calculate ALAP of all irns */
1863         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
1864
1865         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
1866         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
1867
1868         /* we don't need this information any longer */
1869         heights_free(env.height);
1870
1871         /* perform ILP scheduling */
1872         irg_block_walk_graph(env.irg, clear_unwanted_data, create_ilp, &env);
1873
1874         DEBUG_ONLY(
1875                 if (firm_dbg_get_mask(env.dbg)) {
1876                         phase_stat_t stat;
1877                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
1878
1879                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
1880                 }
1881         );
1882
1883         /* free all allocated object */
1884         phase_free(&env.ph);
1885
1886         /* notify backend */
1887         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
1888 }
1889
1890 #ifdef WITH_LIBCORE
1891 /**
1892  * Register ILP scheduler options.
1893  */
1894 void ilpsched_register_options(lc_opt_entry_t *grp) {
1895         static int     run_once = 0;
1896         lc_opt_entry_t *sched_grp;
1897
1898         if (! run_once) {
1899                 run_once  = 1;
1900                 sched_grp = lc_opt_get_grp(grp, "ilpsched");
1901
1902                 lc_opt_add_table(sched_grp, ilpsched_option_table);
1903         }
1904 }
1905 #endif /* WITH_LIBCORE */
1906
1907 #else /* WITH_ILP */
1908
1909 static int some_picky_compiler_do_not_allow_empty_files;
1910
1911 #endif /* WITH_ILP */