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

   1 /**
   2  * Scheduling algorithms.
   3  * An ILP scheduler based on
   4  * "ILP-based Instruction Scheduling for IA-64"
   5  * by Daniel Kaestner and Sebastian Winkel
   6  *
   7  * @date   22.10.2005
   8  * @author Christian Wuerdig
   9  * @cvs-id $Id$
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #ifndef _WIN32
  20 #include <strings.h>
  21 #endif /* _WIN32 */
  22
  23 #include "irnode_t.h"
  24 #include "irgwalk.h"
  25 #include "irbitset.h"
  26 #include "irphase_t.h"
  27 #include "height.h"
  28 #include "iredges.h"
  29 #include "pdeq.h"
  30 #include "debug.h"
  31 #include "irtools.h"
  32 #include "irdump.h"
  33 #include "plist.h"
  34
  35 #include <lpp/lpp.h>
  36 #include <lpp/lpp_net.h>
  37
  38 #ifdef WITH_LIBCORE
  39 #include <libcore/lc_opts.h>
  40 #include <libcore/lc_opts_enum.h>
  41 #include <libcore/lc_timing.h>
  42 #endif /* WITH_LIBCORE */
  43
  44 #include "be.h"
  45 #include "benode_t.h"
  46 #include "besched_t.h"
  47 #include "beilpsched.h"
  48
  49 typedef struct _ilpsched_options_t {
  50         unsigned limit_dead;
  51         unsigned time_limit;
  52         char     log_file[1024];
  53 } ilpsched_options_t;
  54
  55 typedef struct _unit_type_info_t {
  56         int                            n_units;
  57         const be_execution_unit_type_t *tp;
  58 } unit_type_info_t;
  59
  60 /**
  61  * holding the ILP variables of the different types
  62  */
  63 typedef struct _ilp_var_types_t {
  64         int *x;   /* x_{nt}^k variables */
  65         int *a;   /* a_{nt}^k variables */
  66         int *d;   /* d_{nt}^k variables */
  67         int *y;   /* y_{nt}^k variables */
  68 } ilp_var_types_t;
  69
  70 /* attributes for a node */
  71 typedef struct _ilpsched_node_attr_t {
  72         unsigned asap;                     /**< The ASAP scheduling control step */
  73         unsigned alap;                     /**< The ALAP scheduling control step */
  74         unsigned latency;                  /**< Latency of this node (needed for sorting) */
  75         unsigned sched_point;              /**< the step in which the node is finally scheduled */
  76         unsigned visit_idx;                /**< Index of the node having visited this node last */
  77         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
  78         unsigned n_consumer;               /**< Number of consumers */
  79         ir_node  **block_consumer;         /**< List of consumer being in the same block */
  80         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
  81         unsigned block_idx     : 30;       /**< A unique per block index */
  82         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
  83         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
  84         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
  85                                                                                         for ASAP, successors for ALAP */
  86         unsigned n_unit_types;             /**< number of allowed execution unit types */
  87         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
  88         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
  89 } ilpsched_node_attr_t;
  90
  91 /* attributes for a block */
  92 typedef struct _ilpsched_block_attr_t {
  93         unsigned block_last_idx;        /**< The highest node index in block so far */
  94         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
  95         unsigned max_steps;             /**< Upper bound for block execution */
  96         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
  97         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
  98 } ilpsched_block_attr_t;
  99
 100 typedef union _ilpsched_attr_ {
 101         ilpsched_node_attr_t  node_attr;
 102         ilpsched_block_attr_t block_attr;
 103 } ilpsched_attr_t;
 104
 105 /* A irn for the phase and it's attributes (either node or block) */
 106 typedef struct {
 107         ir_node         *irn;
 108         ilpsched_attr_t attr;
 109 } be_ilpsched_irn_t;
 110
 111 /* The ILP scheduling environment */
 112 typedef struct {
 113         phase_t              ph;            /**< The phase */
 114         ir_graph             *irg;          /**< The current irg */
 115         heights_t            *height;       /**< The heights object of the irg */
 116         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 117         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 118         const arch_env_t     *arch_env;
 119         const arch_isa_t     *isa;          /**< The ISA */
 120         const be_main_env_t  *main_env;
 121         const be_machine_t   *cpu;          /**< the current abstract machine */
 122         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 123         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 124         DEBUG_ONLY(firm_dbg_module_t *dbg);
 125 } be_ilpsched_env_t;
 126
 127 /* convenience macros to handle phase irn data */
 128 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 129 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 130 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 131 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 132
 133 /* iterate over a list of ir_nodes linked by link field */
 134 #define foreach_linked_irns(head, iter) for ((iter) = (head); (iter); (iter) = get_irn_link((iter)))
 135
 136 /* check if node is considered for ILP scheduling */
 137 #define consider_for_sched(isa, irn) \
 138         (! (is_Block(irn)            ||  \
 139                 is_normal_Proj(isa, irn) ||  \
 140                 is_Phi(irn)              ||  \
 141                 is_NoMem(irn)            ||  \
 142                 is_Unknown(irn)          ||  \
 143                 is_End(irn)                  \
 144                 ))
 145
 146 /* gives the valid scheduling time step interval for a node */
 147 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 148
 149 /* gives the valid interval where a node can die */
 150 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 151
 152 /* gives the corresponding ILP variable for given node, unit and time step */
 153 #define ILPVAR_IDX(na, unit, control_step) \
 154         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 155
 156 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 157 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 158         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 159
 160 /* check if a double value is within an epsilon environment of 0 */
 161 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 162
 163 #ifdef WITH_LIBCORE
 164         #define ilp_timer_push(t)         lc_timer_push((t))
 165         #define ilp_timer_pop()           lc_timer_pop()
 166         #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
 167 #else /* WITH_LIBCORE */
 168         #define ilp_timer_push(t)
 169         #define ilp_timer_pop()
 170         #define ilp_timer_elapsed_usec(t) 0.0
 171 #endif /* WITH_LIBCORE */
 172
 173 /* option variable */
 174 static ilpsched_options_t ilp_opts = {
 175         70,    /* if we have more than 70 nodes: use alive nodes constraint */
 176         300,   /* 300 sec per block time limit */
 177         ""     /* no log file */
 178 };
 179
 180 #ifdef WITH_LIBCORE
 181 /* ILP options */
 182 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 183         LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead),
 184         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 185         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 186         { NULL }
 187 };
 188 #endif /* WITH_LIBCORE */
 189
 190 /*
 191         We need this global variable as we compare nodes dependent on heights,
 192         but we cannot pass any information to the qsort compare function.
 193 */
 194 static heights_t *glob_heights;
 195
 196 /**
 197  * Check if irn is a Proj, which has no execution units assigned.
 198  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 199  */
 200 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
 201         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
 202 }
 203
 204 /**
 205  * Skips normal Projs.
 206  * @return predecessor if irn is a normal Proj, otherwise irn.
 207  */
 208 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
 209         if (is_normal_Proj(isa, irn))
 210                 return get_Proj_pred(irn);
 211         return irn;
 212 }
 213
 214 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 215         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 216         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 217                 lat = 1;
 218         return lat;
 219 }
 220
 221
 222 /**
 223  * Compare scheduling time steps of two be_ilpsched_irn's.
 224  */
 225 static int cmp_ilpsched_irn(const void *a, const void *b) {
 226         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 227         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 228         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 229         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 230
 231         if (n1_a->sched_point == n2_a->sched_point) {
 232                 ir_node *irn_a = n1->irn;
 233                 ir_node *irn_b = n2->irn;
 234
 235                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 236                         return 1;
 237                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 238                         return -1;
 239
 240                 /*
 241                         Ok, timestep is equal and the nodes are parallel,
 242                         so check latency and schedule high latency first.
 243                 */
 244                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 245         }
 246         else
 247                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 248 }
 249
 250 /**
 251  * In case there is no phase information for irn, initialize it.
 252  */
 253 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
 254         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 255
 256         if (res == old) {
 257                 /* if we have already some data: check for reinitialization */
 258
 259                 if (! is_Block(irn)) {
 260                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 261
 262                         if (! na->transitive_block_nodes) {
 263                                 ir_node               *block      = get_nodes_block(irn);
 264                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 265                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 266
 267                                 /* we are called after the block indices have been build: create bitset */
 268                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 269                         }
 270                         else {
 271                                 /* we are called from reinit block data: clear the bitset */
 272                                 bitset_clear_all(na->transitive_block_nodes);
 273                                 na->visit_idx    = 0;
 274                                 na->alap_changed = 1;
 275                         }
 276                 }
 277                 return old;
 278         }
 279
 280         res->irn = irn;
 281
 282         /* set ilpsched irn attributes (either block or irn) */
 283         if (is_Block(irn)) {
 284                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 285
 286                 ba->n_interesting_nodes = 0;
 287                 ba->block_last_idx      = 0;
 288                 ba->root_nodes          = plist_new();
 289                 ba->head_ilp_nodes      = NULL;
 290                 ba->max_steps           = 0;
 291         }
 292         else {
 293                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 294                 memset(na, 0, sizeof(*na));
 295         }
 296
 297         return res;
 298 }
 299
 300 /**
 301  * Assign a per block unique number to each node.
 302  */
 303 static void build_block_idx(ir_node *irn, void *walk_env) {
 304         be_ilpsched_env_t     *env = walk_env;
 305         be_ilpsched_irn_t     *node, *block_node;
 306         ilpsched_node_attr_t  *na;
 307         ilpsched_block_attr_t *ba;
 308
 309         if (! consider_for_sched(env->arch_env->isa, irn))
 310                 return;
 311
 312         node       = get_ilpsched_irn(env, irn);
 313         na         = get_ilpsched_node_attr(node);
 314         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 315         ba         = get_ilpsched_block_attr(block_node);
 316
 317         na->block_idx = ba->block_last_idx++;
 318 }
 319
 320 /********************************************************
 321  *                              __        _
 322  *                             / /       | |
 323  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 324  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 325  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 326  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 327  *                 | |                           | |
 328  *                 |_|                           |_|
 329  ********************************************************/
 330
 331 /**
 332  * Add all nodes having no user in current block to last_nodes list.
 333  */
 334 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 335         ir_node               *block;
 336         const ir_edge_t       *edge;
 337         be_ilpsched_irn_t     *block_node, *node;
 338         ilpsched_block_attr_t *ba;
 339         ilpsched_node_attr_t  *na;
 340         int                   i, j;
 341         be_ilpsched_env_t     *env           = walk_env;
 342         int                   has_block_user = 0;
 343         unsigned              n_consumer     = 0;
 344         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 345         ir_node               **consumer;
 346         int                   idx;
 347
 348         if (! consider_for_sched(env->arch_env->isa, irn))
 349                 return;
 350
 351         block    = get_nodes_block(irn);
 352     idx      = get_irn_idx(irn);
 353         consumer = NEW_ARR_F(ir_node *, 0);
 354
 355         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 356
 357         /* check data and dependency out edges */
 358         for (i = 0; i < 2 && ! has_block_user; ++i) {
 359                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 360                         ir_node *user = get_edge_src_irn(edge);
 361
 362                         if (is_normal_Proj(env->arch_env->isa, user)) {
 363                                 const ir_edge_t *user_edge;
 364
 365                                 if (get_irn_mode(user) == mode_X)
 366                                         continue;
 367
 368                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 369                                 for (j = 0; j < 2; ++j) {
 370                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 371                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 372
 373                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 374                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 375                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 376
 377                                                         /* skip already visited nodes */
 378                                                         if (ua->consumer_idx == idx)
 379                                                                 continue;
 380
 381                                                         /* check if node has user in this block and collect the user if it's a data user */
 382                                                         if (get_nodes_block(real_user) == block) {
 383                                                                 if (i == 0 && j == 0)
 384                                                                         ARR_APP1(ir_node *, consumer, real_user);
 385                                                                 has_block_user = 1;
 386                                                         }
 387
 388                                                         /* only count data consumer */
 389                                                         if (i == 0)
 390                                                                 n_consumer++;
 391
 392                                                         /* mark user as visited by this node */
 393                                                         ua->consumer_idx = idx;
 394                                                 }
 395                                         }
 396                                 }
 397                         }
 398                         else if (is_Block(user)) {
 399                                 continue;
 400                         }
 401                         else if (! is_Phi(user)) {
 402                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 403                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 404
 405                                 /* skip already visited nodes */
 406                                 if (ua->consumer_idx == idx)
 407                                         continue;
 408
 409                                 /* check if node has user in this block and collect the user if it's a data user */
 410                                 if (get_nodes_block(user) == block) {
 411                                         if (i == 0)
 412                                                 ARR_APP1(ir_node *, consumer, user);
 413                                         has_block_user = 1;
 414                                 }
 415
 416                                 /* only count data consumer */
 417                                 if (i == 0)
 418                                         n_consumer++;
 419
 420                                 /* mark user visited by this node */
 421                                 ua->consumer_idx = idx;
 422                         }
 423                 }
 424         }
 425
 426         block_node = get_ilpsched_irn(env, block);
 427         ba         = get_ilpsched_block_attr(block_node);
 428
 429         ba->n_interesting_nodes++;
 430
 431         /* current irn has no user inside this block, add to queue */
 432         if (! has_block_user) {
 433                 DB((env->dbg, LEVEL_3, "root node\n"));
 434                 plist_insert_back(ba->root_nodes, irn);
 435         }
 436         else {
 437                 DB((env->dbg, LEVEL_3, "normal node\n"));
 438         }
 439
 440         /* record number of all consumer and the consumer within the same block */
 441         node = get_ilpsched_irn(env, irn);
 442         na   = get_ilpsched_node_attr(node);
 443         na->n_consumer     = n_consumer;
 444         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 445         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 446         DEL_ARR_F(consumer);
 447 }
 448
 449 /**
 450  * Calculate the ASAP scheduling step for current irn.
 451  */
 452 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 453         be_ilpsched_env_t     *env = walk_env;
 454         int                   i;
 455         ir_node               *block;
 456         be_ilpsched_irn_t     *node, *block_node;
 457         ilpsched_node_attr_t  *na;
 458         ilpsched_block_attr_t *ba;
 459
 460         /* These nodes are handled separate */
 461         if (! consider_for_sched(env->arch_env->isa, irn))
 462                 return;
 463
 464         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 465
 466         block    = get_nodes_block(irn);
 467         node     = get_ilpsched_irn(env, irn);
 468         na       = get_ilpsched_node_attr(node);
 469         na->asap = 1;
 470
 471         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 472                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
 473
 474                 /* check for greatest distance to top */
 475                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 476                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 477                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 478                         unsigned             lat;
 479
 480                         lat         = fixed_latency(env->sel, pred, env->block_env);
 481                         na->latency = lat;
 482                         na->asap    = MAX(na->asap, pna->asap + lat);
 483                 }
 484         }
 485
 486         /* add node to ILP node list and update max_steps */
 487         block_node = get_ilpsched_irn(env, block);
 488         ba         = get_ilpsched_block_attr(block_node);
 489
 490         set_irn_link(irn, ba->head_ilp_nodes);
 491         ba->head_ilp_nodes = irn;
 492         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 493
 494         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 495 }
 496
 497 /**
 498  * Calculate the ALAP scheduling step of all irns in current block.
 499  * Depends on max_steps being calculated.
 500  */
 501 static void calculate_block_alap(ir_node *block, void *walk_env) {
 502         be_ilpsched_env_t     *env        = walk_env;
 503         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 504         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 505         waitq                 *cur_queue  = new_waitq();
 506         plist_element_t       *el;
 507
 508         assert(is_Block(block));
 509
 510         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 511                 block, ba->n_interesting_nodes, ba->max_steps));
 512
 513         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 514         //phase_reinit_block_irn_data(&env->ph, block);
 515
 516         /* init start queue */
 517         foreach_plist(ba->root_nodes, el) {
 518                 waitq_put(cur_queue, plist_element_get_value(el));
 519         }
 520
 521         /* repeat until all nodes are processed */
 522         while (! waitq_empty(cur_queue)) {
 523                 waitq *next_queue = new_waitq();
 524
 525                 /* process all nodes in current step */
 526                 while (! waitq_empty(cur_queue)) {
 527                         ir_node              *cur_irn = waitq_get(cur_queue);
 528                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 529                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 530                         int                  i;
 531
 532                         /* cur_node has no alap set -> it's a root node, set to max alap */
 533                         if (na->alap == 0) {
 534                                 na->alap = ba->max_steps;
 535                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 536                                         cur_irn, na->alap));
 537                         }
 538                         else {
 539                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 540                                         cur_irn, na->alap));
 541                         }
 542
 543                         /* set the alap's of all predecessors */
 544                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 545                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
 546
 547                                 /* check for greatest distance to bottom */
 548                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 549                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 550                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 551                                         unsigned             lat;
 552
 553                                         /* mark the predecessor as visited by current irn */
 554                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 555                                                 continue;
 556                                         pna->visit_idx = get_irn_idx(cur_irn);
 557
 558                                         lat = fixed_latency(env->sel, pred, env->block_env);
 559
 560                                         /* set ALAP of current pred */
 561                                         if (pna->alap == 0) {
 562                                                 /* current ALAP is 0: node has not yet been visited */
 563                                                 pna->alap_changed = 1;
 564                                                 pna->alap         = na->alap - lat;
 565                                         }
 566                                         else if (pna->alap > na->alap - lat) {
 567                                                 /* we found a longer path to root node: change ALAP */
 568                                                 pna->alap         = na->alap - lat;
 569                                                 pna->alap_changed = 1;
 570                                         }
 571                                         else {
 572                                                 /* current ALAP is best found so far: keep it */
 573                                                 pna->alap_changed = 0;
 574                                         }
 575
 576                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 577
 578                                         /* enqueue node for next iteration */
 579                                         if (get_irn_ins_or_deps(pred) > 0)
 580                                                 waitq_put(next_queue, pred);
 581                                 }
 582                         }
 583                 }
 584
 585                 /* prepare for next iteration */
 586                 del_waitq(cur_queue);
 587                 cur_queue = next_queue;
 588         }
 589 }
 590
 591 /**
 592  * We can free the list of root nodes here.
 593  */
 594 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 595         be_ilpsched_env_t     *env        = walk_env;
 596         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 597         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 598
 599         plist_free(ba->root_nodes);
 600         ba->root_nodes = NULL;
 601 }
 602
 603 /**
 604  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 605  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 606  */
 607 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 608         be_ilpsched_env_t    *env  = walk_env;
 609         ir_node              *pred = irn;
 610         be_ilpsched_irn_t    *node, *pred_node;
 611         ilpsched_node_attr_t *na, *pna;
 612
 613         if (! consider_for_sched(env->arch_env->isa, irn))
 614                 return;
 615
 616         if (! is_Proj(irn) && ! be_is_Keep(irn))
 617                 return;
 618
 619         /* go to the ancestor */
 620         if (be_is_Keep(irn))
 621                 pred = get_irn_n(irn, 0);
 622         pred = skip_Proj(pred);
 623
 624         node      = get_ilpsched_irn(env, irn);
 625         pred_node = get_ilpsched_irn(env, pred);
 626         na        = get_ilpsched_node_attr(node);
 627         pna       = get_ilpsched_node_attr(pred_node);
 628
 629         na->asap = pna->asap;
 630         na->alap = pna->alap;
 631
 632         /* record all Projs and Keeps for this node */
 633         if (! pna->projkeeps)
 634                 pna->projkeeps = new_waitq();
 635         waitq_put(pna->projkeeps, irn);
 636
 637         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 638 }
 639
 640 /*******************************************
 641  *           _              _       _
 642  *          | |            | |     | |
 643  *  ___  ___| |__   ___  __| |_   _| | ___
 644  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 645  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 646  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 647  *
 648  *******************************************/
 649
 650 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
 651         const ir_edge_t *edge;
 652
 653         foreach_out_edge(irn, edge) {
 654                 ir_node *user = get_edge_src_irn(edge);
 655
 656                 if (be_is_Keep(user)) {
 657                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 658                         waitq_put(keeps, user);
 659                 }
 660         }
 661 }
 662
 663 /**
 664  * Inserts @p irn before @p before into schedule and notifies backend.
 665  */
 666 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 667         ir_node *before, ir_node *irn, unsigned cycle)
 668 {
 669         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 670         sched_add_before(before, irn);
 671 }
 672
 673 /**
 674  * Adds a node, it's Projs (in case of mode_T nodes) and
 675  * it's Keeps to schedule.
 676  */
 677 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
 678         const ir_edge_t *edge;
 679         waitq           *keeps = new_waitq();
 680
 681         /* mode_M nodes are not scheduled */
 682         if (get_irn_mode(irn) == mode_M)
 683                 return;
 684
 685         if (! sched_is_scheduled(irn))
 686                 notified_sched_add_before(env, block, irn, cycle);
 687
 688         /* add Projs */
 689         if (get_irn_mode(irn) == mode_T) {
 690                 foreach_out_edge(irn, edge) {
 691                         ir_node *user = get_edge_src_irn(edge);
 692
 693                         if (to_appear_in_schedule(user) || get_irn_mode(user) == mode_b)
 694                                 notified_sched_add_before(env, block, user, cycle);
 695
 696                         check_for_keeps(keeps, block, user);
 697                 }
 698         }
 699         else {
 700                 check_for_keeps(keeps, block, irn);
 701         }
 702
 703         /* add Keeps */
 704         while (! waitq_empty(keeps)) {
 705                 ir_node *keep = waitq_get(keeps);
 706                 if (! sched_is_scheduled(keep))
 707                         notified_sched_add_before(env, block, keep, cycle);
 708         }
 709
 710         del_waitq(keeps);
 711 }
 712
 713 /**
 714  * Schedule all nodes in the given block, according to the ILP solution.
 715  */
 716 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 717         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 718         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 719         sched_info_t          *info       = get_irn_sched_info(block);
 720         be_ilpsched_irn_t     **sched_nodes;
 721         unsigned              i, l;
 722         ir_node               *cfop, *irn;
 723         const ir_edge_t       *edge;
 724
 725         /* init block schedule list */
 726         INIT_LIST_HEAD(&info->list);
 727         info->scheduled = 1;
 728
 729         /* collect nodes and their scheduling time step */
 730         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 731         if (ba->n_interesting_nodes == 0) {
 732                 /* ignore */
 733         }
 734         else if (ba->n_interesting_nodes == 1) {
 735                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 736
 737                 /* add the single node */
 738                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 739         }
 740         else {
 741                 /* check all nodes for their positive solution */
 742                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 743                         be_ilpsched_irn_t    *node;
 744                         ilpsched_node_attr_t *na;
 745                         int                  tp_idx, found;
 746                         unsigned             cur_var, t;
 747
 748                         node    = get_ilpsched_irn(env, irn);
 749                         na      = get_ilpsched_node_attr(node);
 750                         cur_var = 0;
 751                         found   = 0;
 752
 753                         /* go over all variables of a node until the non-zero one is found */
 754                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 755                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 756                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 757
 758                                         /* check, if variable is set to one (it's not zero then :) */
 759                                         if (! LPP_VALUE_IS_0(val)) {
 760                                                 na->sched_point = t;
 761                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 762                                                 DBG((env->dbg, LEVEL_1, "Schedpoint of %+F is %u at unit type %s\n",
 763                                                         irn, t, na->type_info[tp_idx].tp->name));
 764                                                 found = 1;
 765                                         }
 766                                 }
 767                         }
 768                 }
 769
 770                 glob_heights = env->height;
 771                 /* sort nodes ascending by scheduling time step */
 772                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 773         }
 774
 775         /* make all Phis ready and remember the single cf op */
 776         cfop = NULL;
 777         foreach_out_edge(block, edge) {
 778                 irn = get_edge_src_irn(edge);
 779
 780                 switch (get_irn_opcode(irn)) {
 781                         case iro_Phi:
 782                                 add_to_sched(env, block, irn, 0);
 783                                 break;
 784                         case iro_Start:
 785                         case iro_End:
 786                         case iro_Proj:
 787                         case iro_Bad:
 788                         case iro_Unknown:
 789                                 break;
 790                         default:
 791                                 if (is_cfop(irn)) {
 792                                         assert(cfop == NULL && "Highlander - there can be only one");
 793                                         cfop = irn;
 794                                 }
 795                         break;
 796                 }
 797         }
 798
 799         /* add all nodes from list */
 800         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 801                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 802                 if (sched_nodes[i]->irn != cfop)
 803                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 804         }
 805
 806         /* schedule control flow node if not already done */
 807         if (cfop && ! sched_is_scheduled(cfop))
 808                 add_to_sched(env, block, cfop, 0);
 809
 810         DEL_ARR_F(sched_nodes);
 811 }
 812
 813 /***************************************************************
 814  *   _____ _      _____     _____           _   _
 815  *  |_   _| |    |  __ \   / ____|         | | (_)
 816  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 817  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 818  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 819  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 820  *
 821  ***************************************************************/
 822
 823 /**
 824  * Check if node can be executed on given unit type.
 825  */
 826 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 827         int                  i;
 828         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 829
 830         for (i = na->n_unit_types - 1; i >= 0; --i) {
 831                 if (na->type_info[i].tp == tp)
 832                         return i;
 833         }
 834
 835         return -1;
 836 }
 837
 838 /************************************************
 839  *                   _       _     _
 840  *                  (_)     | |   | |
 841  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 842  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 843  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 844  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 845  *
 846  ************************************************/
 847
 848 /**
 849  * Create the following variables:
 850  * - x_{nt}^k    binary     weigthed with: t
 851  *      node n is scheduled at time step t to unit type k
 852  * ==>> These variables represent the schedule
 853  *
 854  * - d_{nt}^k    binary     weighted with: t
 855  *      node n dies at time step t on unit type k
 856  * - a_{nt}^k    binary     weighted with num_nodes
 857  *      node n is alive at time step t on unit type k
 858  *
 859  * - y_{nt}^k    binary     weighted with: num_nodes^2
 860  *      node n is scheduled at time step t to unit type k
 861  *      although all units of this type are occupied
 862  * ==>> These variables represent the register pressure
 863  *
 864  */
 865 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 866         char                  buf[1024];
 867         ir_node               *irn;
 868         unsigned              num_block_var, num_nodes;
 869         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 870         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 871 #ifdef WITH_LIBCORE
 872         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
 873 #endif /* WITH_LIBCORE */
 874
 875         ilp_timer_push(t_var);
 876         num_block_var = num_nodes = 0;
 877         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 878                 const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
 879                 be_ilpsched_irn_t         *node;
 880                 ilpsched_node_attr_t      *na;
 881                 unsigned                  n_unit_types, tp_idx, unit_idx, n_var, cur_unit;
 882                 unsigned                  cur_var_ad, cur_var_x, cur_var_y, num_ad;
 883
 884                 /* count number of available unit types for this node */
 885                 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 886                         /* just count */ ;
 887
 888                 node = get_ilpsched_irn(env, irn);
 889                 na   = get_ilpsched_node_attr(node);
 890
 891                 na->n_unit_types = n_unit_types;
 892                 na->type_info    = NEW_ARR_D(unit_type_info_t, var_obst, n_unit_types);
 893
 894                 /* fill the type info array */
 895                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 896                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 897                                 /* beware: we also count number of available units here */
 898                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 899                                         na->is_dummy_node = 1;
 900                         }
 901
 902                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 903                         na->type_info[tp_idx].n_units = unit_idx;
 904                 }
 905
 906                 /* allocate space for ilp variables */
 907                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 908                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 909
 910                 /* we need these variables only for "real" nodes */
 911                 if (! na->is_dummy_node) {
 912                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 913                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 914
 915                         num_ad = ba->max_steps - na->asap + 1;
 916
 917                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 918                                 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 919                                 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
 920                         }
 921                         else {
 922                                 na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 923                                 memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0]));
 924                         }
 925                 }
 926
 927                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
 928                         irn, na->asap, na->alap, na->n_unit_types));
 929
 930                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
 931                 /* create variables */
 932                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 933                         unsigned t;
 934
 935                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
 936                                 /* x_{nt}^k variables */
 937                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
 938                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 939                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 940                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 941                                 /* variable counter */
 942                                 n_var++;
 943                                 num_block_var++;
 944
 945                                 if (! na->is_dummy_node) {
 946                                         /* y_{nt}^k variables */
 947                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
 948                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 949                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_binary, (double)(weigth_y));
 950                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 951
 952                                         /* variable counter */
 953                                         n_var++;
 954                                         num_block_var++;
 955                                 }
 956                         }
 957
 958                         /* a node can die at any step t: asap(n) <= t <= U */
 959                         if (! na->is_dummy_node) {
 960                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
 961
 962                                         if (ba->n_interesting_nodes > env->opts->limit_dead) {
 963                                                 /* a_{nt}^k variables */
 964                                                 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
 965                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 966                                                 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
 967                                         }
 968                                         else {
 969                                                 /* d_{nt}^k variables */
 970                                                 snprintf(buf, sizeof(buf), "d_n%u_%s_%u",
 971                                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 972                                                 na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 973                                         }
 974                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 975
 976                                         /* variable counter */
 977                                         n_var++;
 978                                         num_block_var++;
 979                                 }
 980                         }
 981                 }
 982
 983                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
 984                 num_nodes++;
 985         }
 986         ilp_timer_pop();
 987         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
 988                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
 989 }
 990
 991 /*******************************************************
 992  *                      _             _       _
 993  *                     | |           (_)     | |
 994  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
 995  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
 996  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
 997  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
 998  *
 999  *******************************************************/
1000
1001 /**
1002  * Create following ILP constraints:
1003  * - the assignment constraints:
1004  *     assure each node is executed once by exactly one (allowed) execution unit
1005  * - the dead node assignment constraints:
1006  *     assure a node can only die at most once
1007  * - the precedence constraints:
1008  *     assure that no data dependencies are violated
1009  */
1010 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1011         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1012         char                  buf[1024];
1013         ir_node               *irn;
1014         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1015         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1016 #ifdef WITH_LIBCORE
1017         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign",      "create assignment constraints");
1018         lc_timer_t            *t_cst_dead    = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints");
1019         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",        "create precedence constraints");
1020 #endif /* WITH_LIBCORE */
1021
1022         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1023         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1024                 int                  cst, tp_idx, i;
1025                 unsigned             cur_var;
1026                 be_ilpsched_irn_t    *node;
1027                 ilpsched_node_attr_t *na;
1028
1029                 node    = get_ilpsched_irn(env, irn);
1030                 na      = get_ilpsched_node_attr(node);
1031                 cur_var = 0;
1032
1033                 /* the assignment constraint */
1034                 ilp_timer_push(t_cst_assign);
1035                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1036                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1037                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1038                 num_cst_assign++;
1039
1040                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1041                 ilp_timer_pop();
1042
1043                 /* the dead node assignment constraint */
1044                 if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) {
1045                         ilp_timer_push(t_cst_dead);
1046                         snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn));
1047                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1048                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1049
1050                         lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0);
1051                         ilp_timer_pop();
1052                 }
1053
1054                 /* We have separate constraints for Projs and Keeps */
1055                 // ILP becomes infeasible ?!?
1056 //              if (is_Proj(irn) || be_is_Keep(irn))
1057 //                      continue;
1058
1059                 /* the precedence constraints */
1060                 ilp_timer_push(t_cst_prec);
1061                 bs_block_irns = bitset_clear_all(bs_block_irns);
1062                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1063                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1064                         unsigned             t_low, t_high, t;
1065                         be_ilpsched_irn_t    *pred_node;
1066                         ilpsched_node_attr_t *pna;
1067                         unsigned             delay;
1068
1069                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1070                                 continue;
1071
1072                         pred_node = get_ilpsched_irn(env, pred);
1073                         pna       = get_ilpsched_node_attr(pred_node);
1074
1075                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1076
1077                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1078                                 bitset_set(bs_block_irns, pna->block_idx);
1079                         else
1080                                 continue;
1081
1082                         /* irn = n, pred = m */
1083                         delay  = fixed_latency(env->sel, pred, env->block_env);
1084                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1085                         t_high = MIN(na->alap, pna->alap + delay - 1);
1086                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1087                                 unsigned tn, tm;
1088                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1089
1090                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1091                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1092                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1093                                 num_cst_prec++;
1094
1095                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1096                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1097                                         /* node variables have smaller index than pred variables */
1098                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1099                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1100                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1101                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1102                                                 }
1103                                         }
1104
1105                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1106                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1107                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1108                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1109                                                 }
1110                                         }
1111                                 }
1112                                 else {
1113                                         /* pred variables have smaller index than node variables */
1114                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1115                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1116                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1117                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1118                                                 }
1119                                         }
1120
1121                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1122                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1123                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1124                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1125                                                 }
1126                                         }
1127                                 }
1128
1129                                 if (ARR_LEN(tmp_var_idx) > 0)
1130                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1131
1132                                 DEL_ARR_F(tmp_var_idx);
1133                         }
1134                 }
1135                 ilp_timer_pop();
1136         }
1137         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1138                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1139         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1140                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1141 }
1142
1143 /**
1144  * Create ILP resource constraints:
1145  * - assure that for each time step not more instructions are scheduled
1146  *   to the same unit types as units of this type are available
1147  */
1148 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1149         int                   glob_type_idx;
1150         char                  buf[1024];
1151         unsigned              num_cst_resrc = 0;
1152         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1153 #ifdef WITH_LIBCORE
1154         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1155 #endif /* WITH_LIBCORE */
1156
1157         ilp_timer_push(t_cst_rsrc);
1158         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1159                 unsigned                 t;
1160                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1161
1162                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1163
1164                 /* check each time step */
1165                 for (t = 0; t < ba->max_steps; ++t) {
1166                         ir_node *irn;
1167                         int     cst;
1168                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1169
1170                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1171                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1172                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1173                         num_cst_resrc++;
1174
1175                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1176                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1177                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1178                                 int                  tp_idx;
1179
1180                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1181
1182                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1183                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1184                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1185                                 }
1186                         }
1187
1188                         /* set constraints if we have some */
1189                         if (ARR_LEN(tmp_var_idx) > 0)
1190                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1191
1192                         DEL_ARR_F(tmp_var_idx);
1193                 }
1194         }
1195         ilp_timer_pop();
1196         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1197                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1198 }
1199
1200 /**
1201  * Create ILP bundle constraints:
1202  * - assure, at most bundle_size * bundles_per_cycle instructions
1203  *   can be started at a certain point.
1204  */
1205 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1206         char                  buf[1024];
1207         unsigned              t;
1208         unsigned              num_cst_bundle = 0;
1209         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1210         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1211 #ifdef WITH_LIBCORE
1212         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1213 #endif /* WITH_LIBCORE */
1214
1215         ilp_timer_push(t_cst_bundle);
1216         for (t = 0; t < ba->max_steps; ++t) {
1217                 ir_node *irn;
1218                 int     cst;
1219                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1220
1221                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1222                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1223                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1224                 num_cst_bundle++;
1225
1226                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1227                         be_ilpsched_irn_t    *node;
1228                         ilpsched_node_attr_t *na;
1229                         int                  tp_idx;
1230
1231                         /* Projs and Keeps do not contribute to bundle size */
1232                         if (is_Proj(irn) || be_is_Keep(irn))
1233                                 continue;
1234
1235                         node = get_ilpsched_irn(env, irn);
1236                         na   = get_ilpsched_node_attr(node);
1237
1238                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1239                         if (na->is_dummy_node)
1240                                 continue;
1241
1242                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1243                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1244                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1245                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1246                                 }
1247                         }
1248                 }
1249
1250                 if (ARR_LEN(tmp_var_idx) > 0)
1251                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1252
1253                 DEL_ARR_F(tmp_var_idx);
1254         }
1255         ilp_timer_pop();
1256         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1257                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1258 }
1259
1260 /**
1261  * Create ILP dying nodes constraints:
1262  * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k
1263  */
1264 static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1265         char                  buf[1024];
1266         unsigned              t;
1267         unsigned              num_cst = 0;
1268         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1269 #ifdef WITH_LIBCORE
1270         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints");
1271 #endif /* WITH_LIBCORE */
1272
1273         ilp_timer_push(t_cst);
1274         /* check all time_steps */
1275         for (t = 0; t < ba->max_steps; ++t) {
1276                 ir_node *irn;
1277
1278                 /* for all nodes */
1279                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1280                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1281                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1282
1283                         /* if node has no consumer within current block, it cannot die here */
1284                         /* we also ignore nodes assigned to dummy unit */
1285                         if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node)
1286                                 continue;
1287
1288                         /* node can only die here if t at least asap(n) */
1289                         if (t >= na->asap - 1) {
1290                                 int node_tp_idx;
1291
1292                                 /* for all unit types */
1293                                 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1294                                         int tp_idx, i, cst;
1295                                         int *tmp_var_idx = NEW_ARR_F(int, 0);
1296
1297                                         snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn));
1298                                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1));
1299                                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1300                                         num_cst++;
1301
1302                                         /* number of consumer scheduled till t */
1303                                         for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1304                                                 be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1305                                                 ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1306
1307                                                 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1308                                                         unsigned tm;
1309
1310                                                         for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) {
1311                                                                 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1312                                                                 ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]);
1313                                                         }
1314                                                 }
1315                                         }
1316
1317                                         /* could be that no consumer can be scheduled at this point */
1318                                         if (ARR_LEN(tmp_var_idx)) {
1319                                                 int      idx;
1320                                                 unsigned tn;
1321
1322                                                 /* subtract possible prior kill points */
1323                                                 for (tn = na->asap - 1; tn < t; ++tn) {
1324                                                         idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn);
1325                                                         lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0);
1326                                                 }
1327
1328                                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1329                                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer));
1330                                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1331                                         }
1332
1333                                         DEL_ARR_F(tmp_var_idx);
1334                                 }
1335                         }
1336
1337                 }
1338         }
1339         ilp_timer_pop();
1340         DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n",
1341                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1342 }
1343
1344 /**
1345 * Create ILP alive nodes constraints:
1346 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1347 */
1348 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1349         char                  buf[1024];
1350         ir_node               *irn;
1351         unsigned              num_cst = 0;
1352         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1353 #ifdef WITH_LIBCORE
1354         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1355 #endif /* WITH_LIBCORE */
1356
1357         ilp_timer_push(t_cst);
1358         /* for each node */
1359         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1360                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1361                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1362                 unsigned             t;
1363
1364                 /* we ignore nodes assigned to dummy unit here */
1365                 if (na->is_dummy_node)
1366                         continue;
1367
1368                 /* check check all time steps: asap(n) <= t <= U */
1369                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1370                         int node_tp_idx;
1371
1372                         /* for all unit types available for this node */
1373                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1374                                 unsigned tn, tn_max, idx;
1375                                 int      cst, i;
1376                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1377                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1378
1379                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1380                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1381                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1382                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1383                                 num_cst++;
1384
1385                                 tn_max = MIN(na->alap - 1, t);
1386                                 /* check if the node has been scheduled so far */
1387                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1388                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1389                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1390                                 }
1391
1392                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1393                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1394                                 DEL_ARR_F(tmp_var_idx_n);
1395
1396                                 /* subtract the number of consumer scheduled so far */
1397                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1398                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1399                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1400                                         int                  tp_idx;
1401                                         unsigned             tm, tm_max;
1402
1403                                         tm_max = MIN(ca->alap - 1, t);
1404                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1405                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1406                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1407                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1408                                                 }
1409                                         }
1410                                 }
1411
1412                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1413                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1414                                 DEL_ARR_F(tmp_var_idx_m);
1415
1416                                 /* -c * a_{nt}^k */
1417                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1418                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1419
1420                         }
1421                 }
1422         }
1423         ilp_timer_pop();
1424         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1425                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1426 }
1427
1428 /**
1429  * Create ILP pressure constraints, based on dead nodes:
1430  * - add additional costs to objective function if a node is scheduled
1431  *   on a unit although all units of this type are currently occupied
1432  */
1433 static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1434         char                  buf[1024];
1435         ir_node               *cur_irn;
1436         unsigned              num_cst = 0;
1437         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1438 #ifdef WITH_LIBCORE
1439         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1440 #endif /* WITH_LIBCORE */
1441
1442         ilp_timer_push(t_cst);
1443         /* y_{nt}^k is set for each node and timestep and unit type */
1444         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1445                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1446                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1447                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1448                 int                  glob_type_idx;
1449
1450                 /* we ignore nodes assigned to DUMMY unit here */
1451                 if (cur_na->is_dummy_node)
1452                         continue;
1453
1454                 /* for all types */
1455                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1456                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1457                         int                      cur_tp_idx;
1458                         unsigned                 t;
1459
1460                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1461
1462                         /* check if node can be executed on this unit type */
1463                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1464                         if (cur_tp_idx < 0)
1465                                 continue;
1466
1467                         /* check all time_steps */
1468                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1469                                 int     cst, y_idx;
1470                                 ir_node *irn;
1471                                 int     *tmp_idx_1  = NEW_ARR_F(int, 0);
1472                                 int     *tmp_idx_m1 = NEW_ARR_F(int, 0);
1473
1474                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1475                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1476                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1477                                 num_cst++;
1478
1479                                 /*
1480                                         - accumulate all nodes scheduled on unit type k till t
1481                                         - subtract all nodes died on unit type k till t
1482                                 */
1483                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1484                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1485                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1486                                         unsigned             tn, tmax;
1487                                         int                  tp_idx;
1488
1489                                         tmax   = MIN(t, na->alap - 1);
1490                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1491
1492                                         /* current unit type is not suitable for current node */
1493                                         if (tp_idx < 0)
1494                                                 continue;
1495
1496                                         for (tn = na->asap - 1; tn <= tmax; ++tn) {
1497                                                 int idx;
1498
1499                                                 /* node scheduled */
1500                                                 idx = ILPVAR_IDX(na, tp_idx, tn);
1501                                                 ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]);
1502
1503                                                 /* node dead */
1504                                                 idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn);
1505                                                 ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]);
1506                                         }
1507                                 }
1508
1509                                 if (ARR_LEN(tmp_idx_1) > 0)
1510                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0);
1511
1512                                 if (ARR_LEN(tmp_idx_m1) > 0)
1513                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0);
1514
1515                                 /* BEWARE: t is unsigned, so (double)(-t) won't work */
1516                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1517                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(t));
1518
1519                                 DEL_ARR_F(tmp_idx_1);
1520                                 DEL_ARR_F(tmp_idx_m1);
1521                         }
1522                 }
1523         }
1524         ilp_timer_pop();
1525         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1526                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1527 }
1528
1529 /**
1530  * Create ILP pressure constraints, based on alive nodes:
1531  * - add additional costs to objective function if a node is scheduled
1532  *   on a unit although all units of this type are currently occupied
1533  */
1534 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1535         char                  buf[1024];
1536         ir_node               *cur_irn;
1537         unsigned              num_cst = 0;
1538         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1539 #ifdef WITH_LIBCORE
1540         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1541 #endif /* WITH_LIBCORE */
1542
1543         ilp_timer_push(t_cst);
1544         /* y_{nt}^k is set for each node and timestep and unit type */
1545         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1546                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1547                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1548                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1549                 int                  glob_type_idx;
1550
1551                 /* we ignore nodes assigned to DUMMY unit here */
1552                 if (cur_na->is_dummy_node)
1553                         continue;
1554
1555                 /* for all types */
1556                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1557                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1558                         int                      cur_tp_idx;
1559                         unsigned                 t;
1560
1561                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1562
1563                         /* check if node can be executed on this unit type */
1564                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1565                         if (cur_tp_idx < 0)
1566                                 continue;
1567
1568                         /* check all time_steps at which the current node can be scheduled */
1569                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1570                                 int     cst, y_idx;
1571                                 ir_node *irn;
1572                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1573
1574                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1575                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1576                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1577                                 num_cst++;
1578
1579                                 /* - accumulate all nodes alive at point t on unit type k */
1580                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1581                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1582                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1583                                         int                  a_idx, tp_idx;
1584
1585                                         /* check if node can be alive here */
1586                                         if (t < na->asap - 1)
1587                                                 continue;
1588
1589                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1590
1591                                         /* current type is not suitable */
1592                                         if (tp_idx < 0)
1593                                                 continue;
1594
1595                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1596                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1597                                 }
1598
1599                                 if (ARR_LEN(tmp_var_idx) > 0)
1600                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1601                                 DEL_ARR_F(tmp_var_idx);
1602
1603                                 /* - num_nodes * y_{nt}^k */
1604                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1605                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(ba->n_interesting_nodes));
1606                         }
1607                 }
1608         }
1609         ilp_timer_pop();
1610         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1611                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1612 }
1613
1614 #if 0
1615 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1616         char                  buf[1024];
1617         ir_node               *irn;
1618         unsigned              num_cst = 0;
1619         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1620 #ifdef WITH_LIBCORE
1621         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1622 #endif /* WITH_LIBCORE */
1623
1624         ilp_timer_push(t_cst);
1625         /* check all nodes */
1626         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1627                 be_ilpsched_irn_t    *node;
1628                 ilpsched_node_attr_t *na;
1629                 unsigned             t;
1630                 ir_node              **pk;
1631
1632                 /* only mode_T nodes can have Projs and Keeps assigned */
1633                 if (get_irn_mode(irn) != mode_T)
1634                         continue;
1635
1636                 node = get_ilpsched_irn(env, irn);
1637                 na   = get_ilpsched_node_attr(node);
1638
1639                 /* check if has some Projs and Keeps assigned */
1640                 if (! na->projkeeps)
1641                         continue;
1642
1643                 /* we can run only once over the queue, so preserve the nodes */
1644                 pk = NEW_ARR_F(ir_node *, 0);
1645                 while (! waitq_empty(na->projkeeps))
1646                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1647                 del_waitq(na->projkeeps);
1648                 na->projkeeps = NULL;
1649
1650                 /* for all time steps at which this node can be scheduled */
1651                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1652                         int cst, tp_idx, i;
1653                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1654
1655                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1656                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1657                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1658                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1659                         num_cst++;
1660
1661                         /* sum up scheduling variables for this time step */
1662                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1663                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1664                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1665                         }
1666
1667                         if (ARR_LEN(tmp_var_idx_n) > 0)
1668                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1669                         DEL_ARR_F(tmp_var_idx_n);
1670
1671                         /* subtract all Proj and Keep variables for this step */
1672                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1673                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1674                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1675                                 int                  pk_tp_idx;
1676
1677                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1678                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1679                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1680                                 }
1681                         }
1682                 }
1683         }
1684         ilp_timer_pop();
1685         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1686                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1687 }
1688 #endif
1689
1690 /***************************************************
1691  *  _____ _      _____                    _
1692  * |_   _| |    |  __ \                  (_)
1693  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1694  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1695  *  _| |_| |____| |      | | | | | | (_| | | | | |
1696  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1697  *
1698  ***************************************************/
1699
1700 /**
1701  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1702  */
1703 static void create_ilp(ir_node *block, void *walk_env) {
1704         be_ilpsched_env_t     *env           = walk_env;
1705         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1706         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1707         FILE                  *logfile       = NULL;
1708         lpp_t                 *lpp           = NULL;
1709         struct obstack        var_obst;
1710         char                  name[1024];
1711
1712         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1713         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1714         DBG((env->dbg, 255, "=========================================\n\n"));
1715
1716         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1717                 block, ba->n_interesting_nodes, ba->max_steps));
1718
1719         /* notify backend and get block environment */
1720         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1721
1722         /* if we have less than two interesting nodes, there is no need to create the ILP */
1723         if (ba->n_interesting_nodes > 1) {
1724                 double fact_var        = ba->n_interesting_nodes > 25 ? 1.1 : 1.2;
1725                 double fact_cst        = ba->n_interesting_nodes > 25 ? 0.7 : 1.5;
1726                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1727                 int    estimated_n_var = (int)((double)base_num * fact_var);
1728                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1729
1730                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimed numbers: %d vars, %d cst\n",
1731                         estimated_n_var, estimated_n_cst));
1732
1733                 /* set up the LPP object */
1734                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1735
1736                 lpp = new_lpp_userdef(
1737                         (const char *)name,
1738                         lpp_minimize,
1739                         estimated_n_cst + 1,  /* num vars */
1740                         estimated_n_cst + 20, /* num cst */
1741                         1.2);                 /* grow factor */
1742                 obstack_init(&var_obst);
1743
1744                 /* create ILP variables */
1745                 create_variables(env, lpp, block_node, &var_obst);
1746
1747                 /* create ILP constraints */
1748                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1749                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1750                 create_ressource_constraints(env, lpp, block_node);
1751                 create_bundle_constraints(env, lpp, block_node);
1752                 //create_proj_keep_constraints(env, lpp, block_node);
1753 #if 0
1754                 if (ba->n_interesting_nodes > env->opts->limit_dead) {
1755                         create_alive_nodes_constraint(env, lpp, block_node);
1756                         create_pressure_alive_constraint(env, lpp, block_node);
1757                 } else {
1758                         create_dying_nodes_constraint(env, lpp, block_node);
1759                         create_pressure_dead_constraint(env, lpp, block_node);
1760                 }
1761 #endif
1762
1763                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1764
1765                 /* debug stuff, dump lpp when debugging is on  */
1766                 DEBUG_ONLY(
1767                         if (firm_dbg_get_mask(env->dbg) > 0) {
1768                                 char buf[1024];
1769                                 FILE *f;
1770
1771                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1772                                 f = fopen(buf, "w");
1773                                 lpp_dump_plain(lpp, f);
1774                                 fclose(f);
1775                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1776                                 lpp_dump(lpp, buf);
1777                         }
1778                 );
1779
1780                 /* set solve time limit */
1781                 lpp_set_time_limit(lpp, env->opts->time_limit);
1782
1783                 /* set logfile if requested */
1784                 if (strlen(env->opts->log_file) > 0) {
1785                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1786                                 lpp_set_log(lpp, stdout);
1787                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1788                                 lpp_set_log(lpp, stderr);
1789                         else {
1790                                 logfile = fopen(env->opts->log_file, "w");
1791                                 if (! logfile)
1792                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1793                                 else
1794                                         lpp_set_log(lpp, logfile);
1795                         }
1796                 }
1797
1798                 /* solve the ILP */
1799                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1800
1801                 if (logfile)
1802                         fclose(logfile);
1803
1804                 /* check for valid solution */
1805                 if (! lpp_is_sol_valid(lpp)) {
1806                         char buf[1024];
1807                         FILE *f;
1808
1809                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block));
1810                         f = fopen(buf, "w");
1811                         lpp_dump_plain(lpp, f);
1812                         fclose(f);
1813                         snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block));
1814                         lpp_dump(lpp, buf);
1815                         dump_ir_block_graph(env->irg, "-assert");
1816
1817                         assert(0 && "ILP solution is not feasible!");
1818                 }
1819
1820                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1821                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1822                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1823                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1824                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1825                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1826                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1827
1828                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1829         }
1830
1831         /* apply solution */
1832         apply_solution(env, lpp, block);
1833
1834         if (lpp)
1835                 free_lpp(lpp);
1836
1837         /* notify backend */
1838         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1839 }
1840
1841 /**
1842  * Perform ILP scheduling on the given irg.
1843  */
1844 void be_ilp_sched(const be_irg_t *birg) {
1845         be_ilpsched_env_t          env;
1846         const char                 *name = "be ilp scheduling";
1847         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
1848         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
1849
1850         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
1851
1852         //firm_dbg_set_mask(env.dbg, 31);
1853
1854         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
1855         env.sel        = sel;
1856         env.irg        = birg->irg;
1857         env.height     = heights_new(birg->irg);
1858         env.main_env   = birg->main_env;
1859         env.arch_env   = birg->main_env->arch_env;
1860         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
1861         env.opts       = &ilp_opts;
1862         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
1863
1864         /* assign a unique per block number to all interesting nodes */
1865         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
1866
1867         /*
1868                 The block indices are completely build after the walk,
1869                 now we can allocate the bitsets (size depends on block indices)
1870                 for all nodes.
1871         */
1872         phase_reinit_irn_data(&env.ph);
1873
1874         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
1875         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
1876
1877         /* Calculate ALAP of all irns */
1878         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
1879
1880         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
1881         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
1882
1883         /* perform ILP scheduling */
1884         irg_block_walk_graph(env.irg, clear_unwanted_data, create_ilp, &env);
1885
1886         DEBUG_ONLY(
1887                 if (firm_dbg_get_mask(env.dbg)) {
1888                         phase_stat_t stat;
1889                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
1890
1891                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
1892                 }
1893         );
1894
1895         /* free all allocated object */
1896         phase_free(&env.ph);
1897         heights_free(env.height);
1898
1899         /* notify backend */
1900         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
1901 }
1902
1903 #ifdef WITH_LIBCORE
1904 /**
1905  * Register ILP scheduler options.
1906  */
1907 void be_init_ilpsched(void)
1908 {
1909         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1910         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
1911
1912         lc_opt_add_table(sched_grp, ilpsched_option_table);
1913 }
1914 #endif /* WITH_LIBCORE */
1915
1916 #else /* WITH_ILP */
1917
1918 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
1919 {}
1920
1921 #endif /* WITH_ILP */