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

   1 /**
   2  * Scheduling algorithms.
   3  * An ILP scheduler based on
   4  * "ILP-based Instruction Scheduling for IA-64"
   5  * by Daniel Kaestner and Sebastian Winkel
   6  *
   7  * @date   22.10.2005
   8  * @author Christian Wuerdig
   9  * @cvs-id $Id$
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #ifndef _WIN32
  20 #include <strings.h>
  21 #endif /* _WIN32 */
  22
  23 #include "irnode_t.h"
  24 #include "irgwalk.h"
  25 #include "irbitset.h"
  26 #include "irphase_t.h"
  27 #include "height.h"
  28 #include "iredges.h"
  29 #include "pdeq.h"
  30 #include "debug.h"
  31 #include "irtools.h"
  32 #include "irdump.h"
  33 #include "plist.h"
  34
  35 #include <lpp/lpp.h>
  36 #include <lpp/lpp_net.h>
  37
  38 #ifdef WITH_LIBCORE
  39 #include <libcore/lc_opts.h>
  40 #include <libcore/lc_opts_enum.h>
  41 #include <libcore/lc_timing.h>
  42 #endif /* WITH_LIBCORE */
  43
  44 #include "be.h"
  45 #include "benode_t.h"
  46 #include "besched_t.h"
  47 #include "beilpsched.h"
  48 #include "beutil.h"
  49
  50 typedef struct _ilpsched_options_t {
  51         unsigned limit_dead;
  52         unsigned time_limit;
  53         char     log_file[1024];
  54 } ilpsched_options_t;
  55
  56 typedef struct _unit_type_info_t {
  57         int                            n_units;
  58         const be_execution_unit_type_t *tp;
  59 } unit_type_info_t;
  60
  61 /**
  62  * holding the ILP variables of the different types
  63  */
  64 typedef struct _ilp_var_types_t {
  65         int *x;   /* x_{nt}^k variables */
  66         int *a;   /* a_{nt}^k variables */
  67         int *d;   /* d_{nt}^k variables */
  68         int *y;   /* y_{nt}^k variables */
  69 } ilp_var_types_t;
  70
  71 /* attributes for a node */
  72 typedef struct _ilpsched_node_attr_t {
  73         unsigned asap;                     /**< The ASAP scheduling control step */
  74         unsigned alap;                     /**< The ALAP scheduling control step */
  75         unsigned latency;                  /**< Latency of this node (needed for sorting) */
  76         unsigned sched_point;              /**< the step in which the node is finally scheduled */
  77         unsigned visit_idx;                /**< Index of the node having visited this node last */
  78         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
  79         unsigned n_consumer;               /**< Number of consumers */
  80         ir_node  **block_consumer;         /**< List of consumer being in the same block */
  81         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
  82         unsigned block_idx     : 30;       /**< A unique per block index */
  83         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
  84         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
  85         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
  86                                                                                         for ASAP, successors for ALAP */
  87         unsigned n_unit_types;             /**< number of allowed execution unit types */
  88         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
  89         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
  90 } ilpsched_node_attr_t;
  91
  92 /* attributes for a block */
  93 typedef struct _ilpsched_block_attr_t {
  94         unsigned block_last_idx;        /**< The highest node index in block so far */
  95         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
  96         unsigned max_steps;             /**< Upper bound for block execution */
  97         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
  98         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
  99 } ilpsched_block_attr_t;
 100
 101 typedef union _ilpsched_attr_ {
 102         ilpsched_node_attr_t  node_attr;
 103         ilpsched_block_attr_t block_attr;
 104 } ilpsched_attr_t;
 105
 106 /* A irn for the phase and it's attributes (either node or block) */
 107 typedef struct {
 108         ir_node         *irn;
 109         ilpsched_attr_t attr;
 110 } be_ilpsched_irn_t;
 111
 112 /* The ILP scheduling environment */
 113 typedef struct {
 114         phase_t              ph;            /**< The phase */
 115         ir_graph             *irg;          /**< The current irg */
 116         heights_t            *height;       /**< The heights object of the irg */
 117         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 118         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 119         const arch_env_t     *arch_env;
 120         const arch_isa_t     *isa;          /**< The ISA */
 121         const be_main_env_t  *main_env;
 122         const be_machine_t   *cpu;          /**< the current abstract machine */
 123         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 124         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 125         DEBUG_ONLY(firm_dbg_module_t *dbg);
 126 } be_ilpsched_env_t;
 127
 128 /* convenience macros to handle phase irn data */
 129 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 130 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 131 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 132 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 133
 134 /* check if node is considered for ILP scheduling */
 135 #define consider_for_sched(isa, irn) \
 136         (! (is_Block(irn)            ||  \
 137                 is_normal_Proj(isa, irn) ||  \
 138                 is_Phi(irn)              ||  \
 139                 is_NoMem(irn)            ||  \
 140                 is_Unknown(irn)          ||  \
 141                 is_End(irn)                  \
 142                 ))
 143
 144 /* gives the valid scheduling time step interval for a node */
 145 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 146
 147 /* gives the valid interval where a node can die */
 148 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 149
 150 /* gives the corresponding ILP variable for given node, unit and time step */
 151 #define ILPVAR_IDX(na, unit, control_step) \
 152         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 153
 154 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 155 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 156         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 157
 158 /* check if a double value is within an epsilon environment of 0 */
 159 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 160
 161 #ifdef WITH_LIBCORE
 162         #define ilp_timer_push(t)         lc_timer_push((t))
 163         #define ilp_timer_pop()           lc_timer_pop()
 164         #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
 165 #else /* WITH_LIBCORE */
 166         #define ilp_timer_push(t)
 167         #define ilp_timer_pop()
 168         #define ilp_timer_elapsed_usec(t) 0.0
 169 #endif /* WITH_LIBCORE */
 170
 171 /* option variable */
 172 static ilpsched_options_t ilp_opts = {
 173         70,    /* if we have more than 70 nodes: use alive nodes constraint */
 174         300,   /* 300 sec per block time limit */
 175         ""     /* no log file */
 176 };
 177
 178 #ifdef WITH_LIBCORE
 179 /* ILP options */
 180 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 181         LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead),
 182         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 183         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 184         { NULL }
 185 };
 186 #endif /* WITH_LIBCORE */
 187
 188 /*
 189         We need this global variable as we compare nodes dependent on heights,
 190         but we cannot pass any information to the qsort compare function.
 191 */
 192 static heights_t *glob_heights;
 193
 194 /**
 195  * Check if irn is a Proj, which has no execution units assigned.
 196  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 197  */
 198 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
 199         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
 200 }
 201
 202 /**
 203  * Skips normal Projs.
 204  * @return predecessor if irn is a normal Proj, otherwise irn.
 205  */
 206 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
 207         if (is_normal_Proj(isa, irn))
 208                 return get_Proj_pred(irn);
 209         return irn;
 210 }
 211
 212 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 213         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 214         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 215                 lat = 1;
 216         return lat;
 217 }
 218
 219
 220 /**
 221  * Compare scheduling time steps of two be_ilpsched_irn's.
 222  */
 223 static int cmp_ilpsched_irn(const void *a, const void *b) {
 224         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 225         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 226         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 227         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 228
 229         if (n1_a->sched_point == n2_a->sched_point) {
 230                 ir_node *irn_a = n1->irn;
 231                 ir_node *irn_b = n2->irn;
 232
 233                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 234                         return 1;
 235                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 236                         return -1;
 237
 238                 /*
 239                         Ok, timestep is equal and the nodes are parallel,
 240                         so check latency and schedule high latency first.
 241                 */
 242                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 243         }
 244         else
 245                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 246 }
 247
 248 /**
 249  * In case there is no phase information for irn, initialize it.
 250  */
 251 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
 252         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 253
 254         if (res == old) {
 255                 /* if we have already some data: check for reinitialization */
 256
 257                 if (! is_Block(irn)) {
 258                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 259
 260                         if (! na->transitive_block_nodes) {
 261                                 ir_node               *block      = get_nodes_block(irn);
 262                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 263                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 264
 265                                 /* we are called after the block indices have been build: create bitset */
 266                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 267                         }
 268                         else {
 269                                 /* we are called from reinit block data: clear the bitset */
 270                                 bitset_clear_all(na->transitive_block_nodes);
 271                                 na->visit_idx    = 0;
 272                                 na->alap_changed = 1;
 273                         }
 274                 }
 275                 return old;
 276         }
 277
 278         res->irn = irn;
 279
 280         /* set ilpsched irn attributes (either block or irn) */
 281         if (is_Block(irn)) {
 282                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 283
 284                 ba->n_interesting_nodes = 0;
 285                 ba->block_last_idx      = 0;
 286                 ba->root_nodes          = plist_new();
 287                 ba->head_ilp_nodes      = NULL;
 288                 ba->max_steps           = 0;
 289         }
 290         else {
 291                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 292                 memset(na, 0, sizeof(*na));
 293         }
 294
 295         return res;
 296 }
 297
 298 /**
 299  * Assign a per block unique number to each node.
 300  */
 301 static void build_block_idx(ir_node *irn, void *walk_env) {
 302         be_ilpsched_env_t     *env = walk_env;
 303         be_ilpsched_irn_t     *node, *block_node;
 304         ilpsched_node_attr_t  *na;
 305         ilpsched_block_attr_t *ba;
 306
 307         if (! consider_for_sched(env->arch_env->isa, irn))
 308                 return;
 309
 310         node       = get_ilpsched_irn(env, irn);
 311         na         = get_ilpsched_node_attr(node);
 312         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 313         ba         = get_ilpsched_block_attr(block_node);
 314
 315         na->block_idx = ba->block_last_idx++;
 316 }
 317
 318 /********************************************************
 319  *                              __        _
 320  *                             / /       | |
 321  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 322  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 323  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 324  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 325  *                 | |                           | |
 326  *                 |_|                           |_|
 327  ********************************************************/
 328
 329 /**
 330  * Add all nodes having no user in current block to last_nodes list.
 331  */
 332 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 333         ir_node               *block;
 334         const ir_edge_t       *edge;
 335         be_ilpsched_irn_t     *block_node, *node;
 336         ilpsched_block_attr_t *ba;
 337         ilpsched_node_attr_t  *na;
 338         int                   i, j;
 339         be_ilpsched_env_t     *env           = walk_env;
 340         int                   has_block_user = 0;
 341         unsigned              n_consumer     = 0;
 342         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 343         ir_node               **consumer;
 344         int                   idx;
 345
 346         if (! consider_for_sched(env->arch_env->isa, irn))
 347                 return;
 348
 349         block    = get_nodes_block(irn);
 350     idx      = get_irn_idx(irn);
 351         consumer = NEW_ARR_F(ir_node *, 0);
 352
 353         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 354
 355         /* check data and dependency out edges */
 356         for (i = 0; i < 2 && ! has_block_user; ++i) {
 357                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 358                         ir_node *user = get_edge_src_irn(edge);
 359
 360                         if (is_normal_Proj(env->arch_env->isa, user)) {
 361                                 const ir_edge_t *user_edge;
 362
 363                                 if (get_irn_mode(user) == mode_X)
 364                                         continue;
 365
 366                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 367                                 for (j = 0; j < 2; ++j) {
 368                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 369                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 370
 371                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 372                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 373                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 374
 375                                                         /* skip already visited nodes */
 376                                                         if (ua->consumer_idx == idx)
 377                                                                 continue;
 378
 379                                                         /* check if node has user in this block and collect the user if it's a data user */
 380                                                         if (get_nodes_block(real_user) == block) {
 381                                                                 if (i == 0 && j == 0)
 382                                                                         ARR_APP1(ir_node *, consumer, real_user);
 383                                                                 has_block_user = 1;
 384                                                         }
 385
 386                                                         /* only count data consumer */
 387                                                         if (i == 0)
 388                                                                 n_consumer++;
 389
 390                                                         /* mark user as visited by this node */
 391                                                         ua->consumer_idx = idx;
 392                                                 }
 393                                         }
 394                                 }
 395                         }
 396                         else if (is_Block(user)) {
 397                                 continue;
 398                         }
 399                         else if (! is_Phi(user)) {
 400                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 401                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 402
 403                                 /* skip already visited nodes */
 404                                 if (ua->consumer_idx == idx)
 405                                         continue;
 406
 407                                 /* check if node has user in this block and collect the user if it's a data user */
 408                                 if (get_nodes_block(user) == block) {
 409                                         if (i == 0)
 410                                                 ARR_APP1(ir_node *, consumer, user);
 411                                         has_block_user = 1;
 412                                 }
 413
 414                                 /* only count data consumer */
 415                                 if (i == 0)
 416                                         n_consumer++;
 417
 418                                 /* mark user visited by this node */
 419                                 ua->consumer_idx = idx;
 420                         }
 421                 }
 422         }
 423
 424         block_node = get_ilpsched_irn(env, block);
 425         ba         = get_ilpsched_block_attr(block_node);
 426
 427         ba->n_interesting_nodes++;
 428
 429         /* current irn has no user inside this block, add to queue */
 430         if (! has_block_user) {
 431                 DB((env->dbg, LEVEL_3, "root node\n"));
 432                 plist_insert_back(ba->root_nodes, irn);
 433         }
 434         else {
 435                 DB((env->dbg, LEVEL_3, "normal node\n"));
 436         }
 437
 438         /* record number of all consumer and the consumer within the same block */
 439         node = get_ilpsched_irn(env, irn);
 440         na   = get_ilpsched_node_attr(node);
 441         na->n_consumer     = n_consumer;
 442         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 443         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 444         DEL_ARR_F(consumer);
 445 }
 446
 447 /**
 448  * Calculate the ASAP scheduling step for current irn.
 449  */
 450 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 451         be_ilpsched_env_t     *env = walk_env;
 452         int                   i;
 453         ir_node               *block;
 454         be_ilpsched_irn_t     *node, *block_node;
 455         ilpsched_node_attr_t  *na;
 456         ilpsched_block_attr_t *ba;
 457
 458         /* These nodes are handled separate */
 459         if (! consider_for_sched(env->arch_env->isa, irn))
 460                 return;
 461
 462         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 463
 464         block    = get_nodes_block(irn);
 465         node     = get_ilpsched_irn(env, irn);
 466         na       = get_ilpsched_node_attr(node);
 467         na->asap = 1;
 468
 469         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 470                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
 471
 472                 /* check for greatest distance to top */
 473                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 474                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 475                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 476                         unsigned             lat;
 477
 478                         lat         = fixed_latency(env->sel, pred, env->block_env);
 479                         na->latency = lat;
 480                         na->asap    = MAX(na->asap, pna->asap + lat);
 481                 }
 482         }
 483
 484         /* add node to ILP node list and update max_steps */
 485         block_node = get_ilpsched_irn(env, block);
 486         ba         = get_ilpsched_block_attr(block_node);
 487
 488         set_irn_link(irn, ba->head_ilp_nodes);
 489         ba->head_ilp_nodes = irn;
 490         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 491
 492         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 493 }
 494
 495 /**
 496  * Calculate the ALAP scheduling step of all irns in current block.
 497  * Depends on max_steps being calculated.
 498  */
 499 static void calculate_block_alap(ir_node *block, void *walk_env) {
 500         be_ilpsched_env_t     *env        = walk_env;
 501         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 502         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 503         waitq                 *cur_queue  = new_waitq();
 504         plist_element_t       *el;
 505
 506         assert(is_Block(block));
 507
 508         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 509                 block, ba->n_interesting_nodes, ba->max_steps));
 510
 511         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 512         //phase_reinit_block_irn_data(&env->ph, block);
 513
 514         /* init start queue */
 515         foreach_plist(ba->root_nodes, el) {
 516                 waitq_put(cur_queue, plist_element_get_value(el));
 517         }
 518
 519         /* repeat until all nodes are processed */
 520         while (! waitq_empty(cur_queue)) {
 521                 waitq *next_queue = new_waitq();
 522
 523                 /* process all nodes in current step */
 524                 while (! waitq_empty(cur_queue)) {
 525                         ir_node              *cur_irn = waitq_get(cur_queue);
 526                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 527                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 528                         int                  i;
 529
 530                         /* cur_node has no alap set -> it's a root node, set to max alap */
 531                         if (na->alap == 0) {
 532                                 na->alap = ba->max_steps;
 533                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 534                                         cur_irn, na->alap));
 535                         }
 536                         else {
 537                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 538                                         cur_irn, na->alap));
 539                         }
 540
 541                         /* set the alap's of all predecessors */
 542                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 543                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
 544
 545                                 /* check for greatest distance to bottom */
 546                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 547                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 548                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 549                                         unsigned             lat;
 550
 551                                         /* mark the predecessor as visited by current irn */
 552                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 553                                                 continue;
 554                                         pna->visit_idx = get_irn_idx(cur_irn);
 555
 556                                         lat = fixed_latency(env->sel, pred, env->block_env);
 557
 558                                         /* set ALAP of current pred */
 559                                         if (pna->alap == 0) {
 560                                                 /* current ALAP is 0: node has not yet been visited */
 561                                                 pna->alap_changed = 1;
 562                                                 pna->alap         = na->alap - lat;
 563                                         }
 564                                         else if (pna->alap > na->alap - lat) {
 565                                                 /* we found a longer path to root node: change ALAP */
 566                                                 pna->alap         = na->alap - lat;
 567                                                 pna->alap_changed = 1;
 568                                         }
 569                                         else {
 570                                                 /* current ALAP is best found so far: keep it */
 571                                                 pna->alap_changed = 0;
 572                                         }
 573
 574                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 575
 576                                         /* enqueue node for next iteration */
 577                                         if (get_irn_ins_or_deps(pred) > 0)
 578                                                 waitq_put(next_queue, pred);
 579                                 }
 580                         }
 581                 }
 582
 583                 /* prepare for next iteration */
 584                 del_waitq(cur_queue);
 585                 cur_queue = next_queue;
 586         }
 587 }
 588
 589 /**
 590  * We can free the list of root nodes here.
 591  */
 592 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 593         be_ilpsched_env_t     *env        = walk_env;
 594         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 595         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 596
 597         plist_free(ba->root_nodes);
 598         ba->root_nodes = NULL;
 599 }
 600
 601 /**
 602  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 603  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 604  */
 605 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 606         be_ilpsched_env_t    *env  = walk_env;
 607         ir_node              *pred = irn;
 608         be_ilpsched_irn_t    *node, *pred_node;
 609         ilpsched_node_attr_t *na, *pna;
 610
 611         if (! consider_for_sched(env->arch_env->isa, irn))
 612                 return;
 613
 614         if (! is_Proj(irn) && ! be_is_Keep(irn))
 615                 return;
 616
 617         /* go to the ancestor */
 618         if (be_is_Keep(irn))
 619                 pred = get_irn_n(irn, 0);
 620         pred = skip_Proj(pred);
 621
 622         node      = get_ilpsched_irn(env, irn);
 623         pred_node = get_ilpsched_irn(env, pred);
 624         na        = get_ilpsched_node_attr(node);
 625         pna       = get_ilpsched_node_attr(pred_node);
 626
 627         na->asap = pna->asap;
 628         na->alap = pna->alap;
 629
 630         /* record all Projs and Keeps for this node */
 631         if (! pna->projkeeps)
 632                 pna->projkeeps = new_waitq();
 633         waitq_put(pna->projkeeps, irn);
 634
 635         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 636 }
 637
 638 /*******************************************
 639  *           _              _       _
 640  *          | |            | |     | |
 641  *  ___  ___| |__   ___  __| |_   _| | ___
 642  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 643  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 644  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 645  *
 646  *******************************************/
 647
 648 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
 649         const ir_edge_t *edge;
 650
 651         foreach_out_edge(irn, edge) {
 652                 ir_node *user = get_edge_src_irn(edge);
 653
 654                 if (be_is_Keep(user)) {
 655                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 656                         waitq_put(keeps, user);
 657                 }
 658         }
 659 }
 660
 661 /**
 662  * Inserts @p irn before @p before into schedule and notifies backend.
 663  */
 664 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 665         ir_node *before, ir_node *irn, unsigned cycle)
 666 {
 667         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 668         sched_add_before(before, irn);
 669 }
 670
 671 /**
 672  * Adds a node, it's Projs (in case of mode_T nodes) and
 673  * it's Keeps to schedule.
 674  */
 675 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
 676         const ir_edge_t *edge;
 677         waitq           *keeps = new_waitq();
 678
 679         /* mode_M nodes are not scheduled */
 680         if (get_irn_mode(irn) == mode_M)
 681                 return;
 682
 683         if (! sched_is_scheduled(irn))
 684                 notified_sched_add_before(env, block, irn, cycle);
 685
 686         /* add Projs */
 687         if (get_irn_mode(irn) == mode_T) {
 688                 foreach_out_edge(irn, edge) {
 689                         ir_node *user = get_edge_src_irn(edge);
 690
 691                         if (to_appear_in_schedule(user) || get_irn_mode(user) == mode_b)
 692                                 notified_sched_add_before(env, block, user, cycle);
 693
 694                         check_for_keeps(keeps, block, user);
 695                 }
 696         }
 697         else {
 698                 check_for_keeps(keeps, block, irn);
 699         }
 700
 701         /* add Keeps */
 702         while (! waitq_empty(keeps)) {
 703                 ir_node *keep = waitq_get(keeps);
 704                 if (! sched_is_scheduled(keep))
 705                         notified_sched_add_before(env, block, keep, cycle);
 706         }
 707
 708         del_waitq(keeps);
 709 }
 710
 711 /**
 712  * Schedule all nodes in the given block, according to the ILP solution.
 713  */
 714 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 715         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 716         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 717         sched_info_t          *info       = get_irn_sched_info(block);
 718         be_ilpsched_irn_t     **sched_nodes;
 719         unsigned              i, l;
 720         ir_node               *cfop, *irn;
 721         const ir_edge_t       *edge;
 722
 723         /* init block schedule list */
 724         INIT_LIST_HEAD(&info->list);
 725         info->scheduled = 1;
 726
 727         /* collect nodes and their scheduling time step */
 728         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 729         if (ba->n_interesting_nodes == 0) {
 730                 /* ignore */
 731         }
 732         else if (ba->n_interesting_nodes == 1) {
 733                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 734
 735                 /* add the single node */
 736                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 737         }
 738         else {
 739                 /* check all nodes for their positive solution */
 740                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 741                         be_ilpsched_irn_t    *node;
 742                         ilpsched_node_attr_t *na;
 743                         int                  tp_idx, found;
 744                         unsigned             cur_var, t;
 745
 746                         node    = get_ilpsched_irn(env, irn);
 747                         na      = get_ilpsched_node_attr(node);
 748                         cur_var = 0;
 749                         found   = 0;
 750
 751                         /* go over all variables of a node until the non-zero one is found */
 752                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 753                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 754                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 755
 756                                         /* check, if variable is set to one (it's not zero then :) */
 757                                         if (! LPP_VALUE_IS_0(val)) {
 758                                                 na->sched_point = t;
 759                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 760                                                 DBG((env->dbg, LEVEL_1, "Schedpoint of %+F is %u at unit type %s\n",
 761                                                         irn, t, na->type_info[tp_idx].tp->name));
 762                                                 found = 1;
 763                                         }
 764                                 }
 765                         }
 766                 }
 767
 768                 glob_heights = env->height;
 769                 /* sort nodes ascending by scheduling time step */
 770                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 771         }
 772
 773         /* make all Phis ready and remember the single cf op */
 774         cfop = NULL;
 775         foreach_out_edge(block, edge) {
 776                 irn = get_edge_src_irn(edge);
 777
 778                 switch (get_irn_opcode(irn)) {
 779                         case iro_Phi:
 780                                 add_to_sched(env, block, irn, 0);
 781                                 break;
 782                         case iro_Start:
 783                         case iro_End:
 784                         case iro_Proj:
 785                         case iro_Bad:
 786                         case iro_Unknown:
 787                                 break;
 788                         default:
 789                                 if (is_cfop(irn)) {
 790                                         assert(cfop == NULL && "Highlander - there can be only one");
 791                                         cfop = irn;
 792                                 }
 793                         break;
 794                 }
 795         }
 796
 797         /* add all nodes from list */
 798         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 799                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 800                 if (sched_nodes[i]->irn != cfop)
 801                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 802         }
 803
 804         /* schedule control flow node if not already done */
 805         if (cfop && ! sched_is_scheduled(cfop))
 806                 add_to_sched(env, block, cfop, 0);
 807
 808         DEL_ARR_F(sched_nodes);
 809 }
 810
 811 /***************************************************************
 812  *   _____ _      _____     _____           _   _
 813  *  |_   _| |    |  __ \   / ____|         | | (_)
 814  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 815  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 816  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 817  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 818  *
 819  ***************************************************************/
 820
 821 /**
 822  * Check if node can be executed on given unit type.
 823  */
 824 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 825         int                  i;
 826         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 827
 828         for (i = na->n_unit_types - 1; i >= 0; --i) {
 829                 if (na->type_info[i].tp == tp)
 830                         return i;
 831         }
 832
 833         return -1;
 834 }
 835
 836 /************************************************
 837  *                   _       _     _
 838  *                  (_)     | |   | |
 839  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 840  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 841  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 842  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 843  *
 844  ************************************************/
 845
 846 /**
 847  * Create the following variables:
 848  * - x_{nt}^k    binary     weigthed with: t
 849  *      node n is scheduled at time step t to unit type k
 850  * ==>> These variables represent the schedule
 851  *
 852  * - d_{nt}^k    binary     weighted with: t
 853  *      node n dies at time step t on unit type k
 854  * - a_{nt}^k    binary     weighted with num_nodes
 855  *      node n is alive at time step t on unit type k
 856  *
 857  * - y_{nt}^k    binary     weighted with: num_nodes^2
 858  *      node n is scheduled at time step t to unit type k
 859  *      although all units of this type are occupied
 860  * ==>> These variables represent the register pressure
 861  *
 862  */
 863 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 864         char                  buf[1024];
 865         ir_node               *irn;
 866         unsigned              num_block_var, num_nodes;
 867         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 868         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 869 #ifdef WITH_LIBCORE
 870         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
 871 #endif /* WITH_LIBCORE */
 872
 873         ilp_timer_push(t_var);
 874         num_block_var = num_nodes = 0;
 875         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 876                 const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
 877                 be_ilpsched_irn_t         *node;
 878                 ilpsched_node_attr_t      *na;
 879                 unsigned                  n_unit_types, tp_idx, unit_idx, n_var, cur_unit;
 880                 unsigned                  cur_var_ad, cur_var_x, cur_var_y, num_ad;
 881
 882                 /* count number of available unit types for this node */
 883                 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 884                         /* just count */ ;
 885
 886                 node = get_ilpsched_irn(env, irn);
 887                 na   = get_ilpsched_node_attr(node);
 888
 889                 na->n_unit_types = n_unit_types;
 890                 na->type_info    = NEW_ARR_D(unit_type_info_t, var_obst, n_unit_types);
 891
 892                 /* fill the type info array */
 893                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 894                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 895                                 /* beware: we also count number of available units here */
 896                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 897                                         na->is_dummy_node = 1;
 898                         }
 899
 900                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 901                         na->type_info[tp_idx].n_units = unit_idx;
 902                 }
 903
 904                 /* allocate space for ilp variables */
 905                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 906                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 907
 908                 /* we need these variables only for "real" nodes */
 909                 if (! na->is_dummy_node) {
 910                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 911                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 912
 913                         num_ad = ba->max_steps - na->asap + 1;
 914
 915                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 916                                 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 917                                 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
 918                         }
 919                         else {
 920                                 na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 921                                 memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0]));
 922                         }
 923                 }
 924
 925                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
 926                         irn, na->asap, na->alap, na->n_unit_types));
 927
 928                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
 929                 /* create variables */
 930                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 931                         unsigned t;
 932
 933                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
 934                                 /* x_{nt}^k variables */
 935                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
 936                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 937                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 938                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 939                                 /* variable counter */
 940                                 n_var++;
 941                                 num_block_var++;
 942
 943                                 if (! na->is_dummy_node) {
 944                                         /* y_{nt}^k variables */
 945                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
 946                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 947                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_binary, (double)(weigth_y));
 948                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 949
 950                                         /* variable counter */
 951                                         n_var++;
 952                                         num_block_var++;
 953                                 }
 954                         }
 955
 956                         /* a node can die at any step t: asap(n) <= t <= U */
 957                         if (! na->is_dummy_node) {
 958                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
 959
 960                                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 961                                                 /* a_{nt}^k variables */
 962                                                 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
 963                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 964                                                 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
 965                                         }
 966                                         else {
 967                                                 /* d_{nt}^k variables */
 968                                                 snprintf(buf, sizeof(buf), "d_n%u_%s_%u",
 969                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 970                                                 na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 971                                         }
 972                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 973
 974                                         /* variable counter */
 975                                         n_var++;
 976                                         num_block_var++;
 977                                 }
 978                         }
 979                 }
 980
 981                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
 982                 num_nodes++;
 983         }
 984         ilp_timer_pop();
 985         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
 986                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
 987 }
 988
 989 /*******************************************************
 990  *                      _             _       _
 991  *                     | |           (_)     | |
 992  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
 993  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
 994  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
 995  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
 996  *
 997  *******************************************************/
 998
 999 /**
1000  * Create following ILP constraints:
1001  * - the assignment constraints:
1002  *     assure each node is executed once by exactly one (allowed) execution unit
1003  * - the dead node assignment constraints:
1004  *     assure a node can only die at most once
1005  * - the precedence constraints:
1006  *     assure that no data dependencies are violated
1007  */
1008 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1009         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1010         char                  buf[1024];
1011         ir_node               *irn;
1012         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1013         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1014 #ifdef WITH_LIBCORE
1015         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign",      "create assignment constraints");
1016         lc_timer_t            *t_cst_dead    = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints");
1017         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",        "create precedence constraints");
1018 #endif /* WITH_LIBCORE */
1019
1020         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1021         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1022                 int                  cst, tp_idx, i;
1023                 unsigned             cur_var;
1024                 be_ilpsched_irn_t    *node;
1025                 ilpsched_node_attr_t *na;
1026
1027                 node    = get_ilpsched_irn(env, irn);
1028                 na      = get_ilpsched_node_attr(node);
1029                 cur_var = 0;
1030
1031                 /* the assignment constraint */
1032                 ilp_timer_push(t_cst_assign);
1033                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1034                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1035                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1036                 num_cst_assign++;
1037
1038                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1039                 ilp_timer_pop();
1040
1041                 /* the dead node assignment constraint */
1042                 if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) {
1043                         ilp_timer_push(t_cst_dead);
1044                         snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn));
1045                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1046                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1047
1048                         lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0);
1049                         ilp_timer_pop();
1050                 }
1051
1052                 /* We have separate constraints for Projs and Keeps */
1053                 // ILP becomes infeasible ?!?
1054 //              if (is_Proj(irn) || be_is_Keep(irn))
1055 //                      continue;
1056
1057                 /* the precedence constraints */
1058                 ilp_timer_push(t_cst_prec);
1059                 bs_block_irns = bitset_clear_all(bs_block_irns);
1060                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1061                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1062                         unsigned             t_low, t_high, t;
1063                         be_ilpsched_irn_t    *pred_node;
1064                         ilpsched_node_attr_t *pna;
1065                         unsigned             delay;
1066
1067                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1068                                 continue;
1069
1070                         pred_node = get_ilpsched_irn(env, pred);
1071                         pna       = get_ilpsched_node_attr(pred_node);
1072
1073                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1074
1075                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1076                                 bitset_set(bs_block_irns, pna->block_idx);
1077                         else
1078                                 continue;
1079
1080                         /* irn = n, pred = m */
1081                         delay  = fixed_latency(env->sel, pred, env->block_env);
1082                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1083                         t_high = MIN(na->alap, pna->alap + delay - 1);
1084                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1085                                 unsigned tn, tm;
1086                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1087
1088                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1089                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1090                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1091                                 num_cst_prec++;
1092
1093                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1094                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1095                                         /* node variables have smaller index than pred variables */
1096                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1097                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1098                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1099                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1100                                                 }
1101                                         }
1102
1103                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1104                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1105                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1106                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1107                                                 }
1108                                         }
1109                                 }
1110                                 else {
1111                                         /* pred variables have smaller index than node variables */
1112                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1113                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1114                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1115                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1116                                                 }
1117                                         }
1118
1119                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1120                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1121                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1122                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1123                                                 }
1124                                         }
1125                                 }
1126
1127                                 if (ARR_LEN(tmp_var_idx) > 0)
1128                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1129
1130                                 DEL_ARR_F(tmp_var_idx);
1131                         }
1132                 }
1133                 ilp_timer_pop();
1134         }
1135         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1136                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1137         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1138                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1139 }
1140
1141 /**
1142  * Create ILP resource constraints:
1143  * - assure that for each time step not more instructions are scheduled
1144  *   to the same unit types as units of this type are available
1145  */
1146 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1147         int                   glob_type_idx;
1148         char                  buf[1024];
1149         unsigned              num_cst_resrc = 0;
1150         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1151 #ifdef WITH_LIBCORE
1152         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1153 #endif /* WITH_LIBCORE */
1154
1155         ilp_timer_push(t_cst_rsrc);
1156         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1157                 unsigned                 t;
1158                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1159
1160                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1161
1162                 /* check each time step */
1163                 for (t = 0; t < ba->max_steps; ++t) {
1164                         ir_node *irn;
1165                         int     cst;
1166                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1167
1168                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1169                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1170                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1171                         num_cst_resrc++;
1172
1173                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1174                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1175                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1176                                 int                  tp_idx;
1177
1178                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1179
1180                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1181                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1182                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1183                                 }
1184                         }
1185
1186                         /* set constraints if we have some */
1187                         if (ARR_LEN(tmp_var_idx) > 0)
1188                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1189
1190                         DEL_ARR_F(tmp_var_idx);
1191                 }
1192         }
1193         ilp_timer_pop();
1194         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1195                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1196 }
1197
1198 /**
1199  * Create ILP bundle constraints:
1200  * - assure, at most bundle_size * bundles_per_cycle instructions
1201  *   can be started at a certain point.
1202  */
1203 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1204         char                  buf[1024];
1205         unsigned              t;
1206         unsigned              num_cst_bundle = 0;
1207         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1208         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1209 #ifdef WITH_LIBCORE
1210         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1211 #endif /* WITH_LIBCORE */
1212
1213         ilp_timer_push(t_cst_bundle);
1214         for (t = 0; t < ba->max_steps; ++t) {
1215                 ir_node *irn;
1216                 int     cst;
1217                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1218
1219                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1220                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1221                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1222                 num_cst_bundle++;
1223
1224                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1225                         be_ilpsched_irn_t    *node;
1226                         ilpsched_node_attr_t *na;
1227                         int                  tp_idx;
1228
1229                         /* Projs and Keeps do not contribute to bundle size */
1230                         if (is_Proj(irn) || be_is_Keep(irn))
1231                                 continue;
1232
1233                         node = get_ilpsched_irn(env, irn);
1234                         na   = get_ilpsched_node_attr(node);
1235
1236                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1237                         if (na->is_dummy_node)
1238                                 continue;
1239
1240                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1241                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1242                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1243                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1244                                 }
1245                         }
1246                 }
1247
1248                 if (ARR_LEN(tmp_var_idx) > 0)
1249                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1250
1251                 DEL_ARR_F(tmp_var_idx);
1252         }
1253         ilp_timer_pop();
1254         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1255                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1256 }
1257
1258 /**
1259  * Create ILP dying nodes constraints:
1260  * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k
1261  */
1262 static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1263         char                  buf[1024];
1264         unsigned              t;
1265         unsigned              num_cst = 0;
1266         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1267 #ifdef WITH_LIBCORE
1268         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints");
1269 #endif /* WITH_LIBCORE */
1270
1271         ilp_timer_push(t_cst);
1272         /* check all time_steps */
1273         for (t = 0; t < ba->max_steps; ++t) {
1274                 ir_node *irn;
1275
1276                 /* for all nodes */
1277                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1278                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1279                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1280
1281                         /* if node has no consumer within current block, it cannot die here */
1282                         /* we also ignore nodes assigned to dummy unit */
1283                         if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node)
1284                                 continue;
1285
1286                         /* node can only die here if t at least asap(n) */
1287                         if (t >= na->asap - 1) {
1288                                 int node_tp_idx;
1289
1290                                 /* for all unit types */
1291                                 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1292                                         int tp_idx, i, cst;
1293                                         int *tmp_var_idx = NEW_ARR_F(int, 0);
1294
1295                                         snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn));
1296                                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1));
1297                                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1298                                         num_cst++;
1299
1300                                         /* number of consumer scheduled till t */
1301                                         for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1302                                                 be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1303                                                 ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1304
1305                                                 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1306                                                         unsigned tm;
1307
1308                                                         for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) {
1309                                                                 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1310                                                                 ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]);
1311                                                         }
1312                                                 }
1313                                         }
1314
1315                                         /* could be that no consumer can be scheduled at this point */
1316                                         if (ARR_LEN(tmp_var_idx)) {
1317                                                 int      idx;
1318                                                 unsigned tn;
1319
1320                                                 /* subtract possible prior kill points */
1321                                                 for (tn = na->asap - 1; tn < t; ++tn) {
1322                                                         idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn);
1323                                                         lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0);
1324                                                 }
1325
1326                                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1327                                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer));
1328                                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1329                                         }
1330
1331                                         DEL_ARR_F(tmp_var_idx);
1332                                 }
1333                         }
1334
1335                 }
1336         }
1337         ilp_timer_pop();
1338         DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n",
1339                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1340 }
1341
1342 /**
1343 * Create ILP alive nodes constraints:
1344 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1345 */
1346 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1347         char                  buf[1024];
1348         ir_node               *irn;
1349         unsigned              num_cst = 0;
1350         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1351 #ifdef WITH_LIBCORE
1352         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1353 #endif /* WITH_LIBCORE */
1354
1355         ilp_timer_push(t_cst);
1356         /* for each node */
1357         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1358                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1359                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1360                 unsigned             t;
1361
1362                 /* we ignore nodes assigned to dummy unit here */
1363                 if (na->is_dummy_node)
1364                         continue;
1365
1366                 /* check check all time steps: asap(n) <= t <= U */
1367                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1368                         int node_tp_idx;
1369
1370                         /* for all unit types available for this node */
1371                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1372                                 unsigned tn, tn_max, idx;
1373                                 int      cst, i;
1374                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1375                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1376
1377                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1378                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1379                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1380                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1381                                 num_cst++;
1382
1383                                 tn_max = MIN(na->alap - 1, t);
1384                                 /* check if the node has been scheduled so far */
1385                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1386                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1387                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1388                                 }
1389
1390                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1391                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1392                                 DEL_ARR_F(tmp_var_idx_n);
1393
1394                                 /* subtract the number of consumer scheduled so far */
1395                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1396                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1397                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1398                                         int                  tp_idx;
1399                                         unsigned             tm, tm_max;
1400
1401                                         tm_max = MIN(ca->alap - 1, t);
1402                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1403                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1404                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1405                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1406                                                 }
1407                                         }
1408                                 }
1409
1410                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1411                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1412                                 DEL_ARR_F(tmp_var_idx_m);
1413
1414                                 /* -c * a_{nt}^k */
1415                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1416                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1417
1418                         }
1419                 }
1420         }
1421         ilp_timer_pop();
1422         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1423                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1424 }
1425
1426 /**
1427  * Create ILP pressure constraints, based on dead nodes:
1428  * - add additional costs to objective function if a node is scheduled
1429  *   on a unit although all units of this type are currently occupied
1430  */
1431 static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1432         char                  buf[1024];
1433         ir_node               *cur_irn;
1434         unsigned              num_cst = 0;
1435         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1436 #ifdef WITH_LIBCORE
1437         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1438 #endif /* WITH_LIBCORE */
1439
1440         ilp_timer_push(t_cst);
1441         /* y_{nt}^k is set for each node and timestep and unit type */
1442         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1443                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1444                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1445                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1446                 int                  glob_type_idx;
1447
1448                 /* we ignore nodes assigned to DUMMY unit here */
1449                 if (cur_na->is_dummy_node)
1450                         continue;
1451
1452                 /* for all types */
1453                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1454                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1455                         int                      cur_tp_idx;
1456                         unsigned                 t;
1457
1458                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1459
1460                         /* check if node can be executed on this unit type */
1461                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1462                         if (cur_tp_idx < 0)
1463                                 continue;
1464
1465                         /* check all time_steps */
1466                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1467                                 int     cst, y_idx;
1468                                 ir_node *irn;
1469                                 int     *tmp_idx_1  = NEW_ARR_F(int, 0);
1470                                 int     *tmp_idx_m1 = NEW_ARR_F(int, 0);
1471
1472                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1473                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1474                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1475                                 num_cst++;
1476
1477                                 /*
1478                                         - accumulate all nodes scheduled on unit type k till t
1479                                         - subtract all nodes died on unit type k till t
1480                                 */
1481                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1482                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1483                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1484                                         unsigned             tn, tmax;
1485                                         int                  tp_idx;
1486
1487                                         tmax   = MIN(t, na->alap - 1);
1488                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1489
1490                                         /* current unit type is not suitable for current node */
1491                                         if (tp_idx < 0)
1492                                                 continue;
1493
1494                                         for (tn = na->asap - 1; tn <= tmax; ++tn) {
1495                                                 int idx;
1496
1497                                                 /* node scheduled */
1498                                                 idx = ILPVAR_IDX(na, tp_idx, tn);
1499                                                 ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]);
1500
1501                                                 /* node dead */
1502                                                 idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn);
1503                                                 ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]);
1504                                         }
1505                                 }
1506
1507                                 if (ARR_LEN(tmp_idx_1) > 0)
1508                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0);
1509
1510                                 if (ARR_LEN(tmp_idx_m1) > 0)
1511                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0);
1512
1513                                 /* BEWARE: t is unsigned, so (double)(-t) won't work */
1514                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1515                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(t));
1516
1517                                 DEL_ARR_F(tmp_idx_1);
1518                                 DEL_ARR_F(tmp_idx_m1);
1519                         }
1520                 }
1521         }
1522         ilp_timer_pop();
1523         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1524                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1525 }
1526
1527 /**
1528  * Create ILP pressure constraints, based on alive nodes:
1529  * - add additional costs to objective function if a node is scheduled
1530  *   on a unit although all units of this type are currently occupied
1531  */
1532 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1533         char                  buf[1024];
1534         ir_node               *cur_irn;
1535         unsigned              num_cst = 0;
1536         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1537 #ifdef WITH_LIBCORE
1538         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1539 #endif /* WITH_LIBCORE */
1540
1541         ilp_timer_push(t_cst);
1542         /* y_{nt}^k is set for each node and timestep and unit type */
1543         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1544                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1545                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1546                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1547                 int                  glob_type_idx;
1548
1549                 /* we ignore nodes assigned to DUMMY unit here */
1550                 if (cur_na->is_dummy_node)
1551                         continue;
1552
1553                 /* for all types */
1554                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1555                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1556                         int                      cur_tp_idx;
1557                         unsigned                 t;
1558
1559                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1560
1561                         /* check if node can be executed on this unit type */
1562                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1563                         if (cur_tp_idx < 0)
1564                                 continue;
1565
1566                         /* check all time_steps at which the current node can be scheduled */
1567                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1568                                 int     cst, y_idx;
1569                                 ir_node *irn;
1570                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1571
1572                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1573                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1574                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1575                                 num_cst++;
1576
1577                                 /* - accumulate all nodes alive at point t on unit type k */
1578                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1579                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1580                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1581                                         int                  a_idx, tp_idx;
1582
1583                                         /* check if node can be alive here */
1584                                         if (t < na->asap - 1)
1585                                                 continue;
1586
1587                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1588
1589                                         /* current type is not suitable */
1590                                         if (tp_idx < 0)
1591                                                 continue;
1592
1593                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1594                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1595                                 }
1596
1597                                 if (ARR_LEN(tmp_var_idx) > 0)
1598                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1599                                 DEL_ARR_F(tmp_var_idx);
1600
1601                                 /* - num_nodes * y_{nt}^k */
1602                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1603                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(ba->n_interesting_nodes));
1604                         }
1605                 }
1606         }
1607         ilp_timer_pop();
1608         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1609                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1610 }
1611
1612 #if 0
1613 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1614         char                  buf[1024];
1615         ir_node               *irn;
1616         unsigned              num_cst = 0;
1617         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1618 #ifdef WITH_LIBCORE
1619         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1620 #endif /* WITH_LIBCORE */
1621
1622         ilp_timer_push(t_cst);
1623         /* check all nodes */
1624         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1625                 be_ilpsched_irn_t    *node;
1626                 ilpsched_node_attr_t *na;
1627                 unsigned             t;
1628                 ir_node              **pk;
1629
1630                 /* only mode_T nodes can have Projs and Keeps assigned */
1631                 if (get_irn_mode(irn) != mode_T)
1632                         continue;
1633
1634                 node = get_ilpsched_irn(env, irn);
1635                 na   = get_ilpsched_node_attr(node);
1636
1637                 /* check if has some Projs and Keeps assigned */
1638                 if (! na->projkeeps)
1639                         continue;
1640
1641                 /* we can run only once over the queue, so preserve the nodes */
1642                 pk = NEW_ARR_F(ir_node *, 0);
1643                 while (! waitq_empty(na->projkeeps))
1644                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1645                 del_waitq(na->projkeeps);
1646                 na->projkeeps = NULL;
1647
1648                 /* for all time steps at which this node can be scheduled */
1649                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1650                         int cst, tp_idx, i;
1651                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1652
1653                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1654                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1655                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1656                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1657                         num_cst++;
1658
1659                         /* sum up scheduling variables for this time step */
1660                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1661                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1662                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1663                         }
1664
1665                         if (ARR_LEN(tmp_var_idx_n) > 0)
1666                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1667                         DEL_ARR_F(tmp_var_idx_n);
1668
1669                         /* subtract all Proj and Keep variables for this step */
1670                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1671                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1672                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1673                                 int                  pk_tp_idx;
1674
1675                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1676                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1677                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1678                                 }
1679                         }
1680                 }
1681         }
1682         ilp_timer_pop();
1683         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1684                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1685 }
1686 #endif
1687
1688 /***************************************************
1689  *  _____ _      _____                    _
1690  * |_   _| |    |  __ \                  (_)
1691  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1692  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1693  *  _| |_| |____| |      | | | | | | (_| | | | | |
1694  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1695  *
1696  ***************************************************/
1697
1698 /**
1699  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1700  */
1701 static void create_ilp(ir_node *block, void *walk_env) {
1702         be_ilpsched_env_t     *env           = walk_env;
1703         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1704         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1705         FILE                  *logfile       = NULL;
1706         lpp_t                 *lpp           = NULL;
1707         struct obstack        var_obst;
1708         char                  name[1024];
1709
1710         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1711         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1712         DBG((env->dbg, 255, "=========================================\n\n"));
1713
1714         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1715                 block, ba->n_interesting_nodes, ba->max_steps));
1716
1717         /* notify backend and get block environment */
1718         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1719
1720         /* if we have less than two interesting nodes, there is no need to create the ILP */
1721         if (ba->n_interesting_nodes > 1) {
1722                 double fact_var        = ba->n_interesting_nodes > 25 ? 1.1 : 1.2;
1723                 double fact_cst        = ba->n_interesting_nodes > 25 ? 0.7 : 1.5;
1724                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1725                 int    estimated_n_var = (int)((double)base_num * fact_var);
1726                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1727
1728                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimed numbers: %d vars, %d cst\n",
1729                         estimated_n_var, estimated_n_cst));
1730
1731                 /* set up the LPP object */
1732                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1733
1734                 lpp = new_lpp_userdef(
1735                         (const char *)name,
1736                         lpp_minimize,
1737                         estimated_n_cst + 1,  /* num vars */
1738                         estimated_n_cst + 20, /* num cst */
1739                         1.2);                 /* grow factor */
1740                 obstack_init(&var_obst);
1741
1742                 /* create ILP variables */
1743                 create_variables(env, lpp, block_node, &var_obst);
1744
1745                 /* create ILP constraints */
1746                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1747                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1748                 create_ressource_constraints(env, lpp, block_node);
1749                 create_bundle_constraints(env, lpp, block_node);
1750                 //create_proj_keep_constraints(env, lpp, block_node);
1751 #if 0
1752                 if (ba->n_interesting_nodes > env->opts->limit_dead) {
1753                         create_alive_nodes_constraint(env, lpp, block_node);
1754                         create_pressure_alive_constraint(env, lpp, block_node);
1755                 } else {
1756                         create_dying_nodes_constraint(env, lpp, block_node);
1757                         create_pressure_dead_constraint(env, lpp, block_node);
1758                 }
1759 #endif
1760
1761                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1762
1763                 /* debug stuff, dump lpp when debugging is on  */
1764                 DEBUG_ONLY(
1765                         if (firm_dbg_get_mask(env->dbg) > 0) {
1766                                 char buf[1024];
1767                                 FILE *f;
1768
1769                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1770                                 f = fopen(buf, "w");
1771                                 lpp_dump_plain(lpp, f);
1772                                 fclose(f);
1773                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1774                                 lpp_dump(lpp, buf);
1775                         }
1776                 );
1777
1778                 /* set solve time limit */
1779                 lpp_set_time_limit(lpp, env->opts->time_limit);
1780
1781                 /* set logfile if requested */
1782                 if (strlen(env->opts->log_file) > 0) {
1783                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1784                                 lpp_set_log(lpp, stdout);
1785                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1786                                 lpp_set_log(lpp, stderr);
1787                         else {
1788                                 logfile = fopen(env->opts->log_file, "w");
1789                                 if (! logfile)
1790                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1791                                 else
1792                                         lpp_set_log(lpp, logfile);
1793                         }
1794                 }
1795
1796                 /* solve the ILP */
1797                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1798
1799                 if (logfile)
1800                         fclose(logfile);
1801
1802                 /* check for valid solution */
1803                 if (! lpp_is_sol_valid(lpp)) {
1804                         char buf[1024];
1805                         FILE *f;
1806
1807                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block));
1808                         f = fopen(buf, "w");
1809                         lpp_dump_plain(lpp, f);
1810                         fclose(f);
1811                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block));
1812                         lpp_dump(lpp, buf);
1813                         dump_ir_block_graph(env->irg, "-assert");
1814
1815                         assert(0 && "ILP solution is not feasible!");
1816                 }
1817
1818                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1819                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1820                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1821                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1822                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1823                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1824                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1825
1826                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1827         }
1828
1829         /* apply solution */
1830         apply_solution(env, lpp, block);
1831
1832         if (lpp)
1833                 free_lpp(lpp);
1834
1835         /* notify backend */
1836         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1837 }
1838
1839 /**
1840  * Perform ILP scheduling on the given irg.
1841  */
1842 void be_ilp_sched(const be_irg_t *birg) {
1843         be_ilpsched_env_t          env;
1844         const char                 *name = "be ilp scheduling";
1845         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
1846         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
1847
1848         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
1849
1850         //firm_dbg_set_mask(env.dbg, 31);
1851
1852         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
1853         env.sel        = sel;
1854         env.irg        = birg->irg;
1855         env.height     = heights_new(birg->irg);
1856         env.main_env   = birg->main_env;
1857         env.arch_env   = birg->main_env->arch_env;
1858         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
1859         env.opts       = &ilp_opts;
1860         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
1861
1862         /* assign a unique per block number to all interesting nodes */
1863         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
1864
1865         /*
1866                 The block indices are completely build after the walk,
1867                 now we can allocate the bitsets (size depends on block indices)
1868                 for all nodes.
1869         */
1870         phase_reinit_irn_data(&env.ph);
1871
1872         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
1873         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
1874
1875         /* Calculate ALAP of all irns */
1876         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
1877
1878         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
1879         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
1880
1881         /* perform ILP scheduling */
1882         irg_block_walk_graph(env.irg, clear_unwanted_data, create_ilp, &env);
1883
1884         DEBUG_ONLY(
1885                 if (firm_dbg_get_mask(env.dbg)) {
1886                         phase_stat_t stat;
1887                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
1888
1889                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
1890                 }
1891         );
1892
1893         /* free all allocated object */
1894         phase_free(&env.ph);
1895         heights_free(env.height);
1896
1897         /* notify backend */
1898         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
1899 }
1900
1901 #ifdef WITH_LIBCORE
1902 /**
1903  * Register ILP scheduler options.
1904  */
1905 void be_init_ilpsched(void)
1906 {
1907         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1908         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
1909
1910         lc_opt_add_table(sched_grp, ilpsched_option_table);
1911 }
1912 #endif /* WITH_LIBCORE */
1913
1914 #else /* WITH_ILP */
1915
1916 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
1917 {}
1918
1919 #endif /* WITH_ILP */