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 #include "irprintf.h"
  35
  36 #include <lpp/lpp.h>
  37 #include <lpp/lpp_net.h>
  38
  39 #include <libcore/lc_opts.h>
  40 #include <libcore/lc_opts_enum.h>
  41 #include <libcore/lc_timing.h>
  42
  43 #include "be.h"
  44 #include "benode_t.h"
  45 #include "besched_t.h"
  46 #include "beilpsched.h"
  47 #include "beutil.h"
  48 #include "bestat.h"
  49
  50 typedef struct _ilpsched_options_t {
  51         unsigned regpress;
  52         unsigned time_limit;
  53         char     log_file[1024];
  54 } ilpsched_options_t;
  55
  56 typedef struct _unit_type_info_t {
  57         int                            n_units;
  58         const be_execution_unit_type_t *tp;
  59 } unit_type_info_t;
  60
  61 /**
  62  * holding the ILP variables of the different types
  63  */
  64 typedef struct _ilp_var_types_t {
  65         int *x;   /* x_{nt}^k variables */
  66         int *a;   /* a_{nt}^k variables */
  67         int *y;   /* y_{nt}^k variables */
  68 } ilp_var_types_t;
  69
  70 /**
  71  * Holds alive variables for a node live-in to a block.
  72  */
  73 typedef struct _ilp_livein_node_t {
  74         ir_node  *irn;
  75         unsigned max_alive_steps;
  76         int      *a;
  77 } ilp_livein_node_t;
  78
  79 /* attributes for a node */
  80 typedef struct _ilpsched_node_attr_t {
  81         unsigned asap;                     /**< The ASAP scheduling control step */
  82         unsigned alap;                     /**< The ALAP scheduling control step */
  83         unsigned latency;                  /**< Latency of this node (needed for sorting) */
  84         unsigned sched_point;              /**< the step in which the node is finally scheduled */
  85         unsigned visit_idx;                /**< Index of the node having visited this node last */
  86         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
  87         unsigned n_consumer;               /**< Number of consumers */
  88         ir_node  **block_consumer;         /**< List of consumer being in the same block */
  89         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
  90         unsigned block_idx     : 30;       /**< A unique per block index */
  91         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
  92         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
  93         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
  94                                                                                         for ASAP, successors for ALAP */
  95         unsigned         n_unit_types;     /**< number of allowed execution unit types */
  96         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
  97         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
  98 } ilpsched_node_attr_t;
  99
 100 /* attributes for a block */
 101 typedef struct _ilpsched_block_attr_t {
 102         unsigned block_last_idx;        /**< The highest node index in block so far */
 103         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 104         unsigned max_steps;             /**< Upper bound for block execution */
 105         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 106         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 107         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 108 } ilpsched_block_attr_t;
 109
 110 typedef union _ilpsched_attr_ {
 111         ilpsched_node_attr_t  node_attr;
 112         ilpsched_block_attr_t block_attr;
 113 } ilpsched_attr_t;
 114
 115 /* A irn for the phase and it's attributes (either node or block) */
 116 typedef struct {
 117         ir_node         *irn;
 118         ilpsched_attr_t attr;
 119 } be_ilpsched_irn_t;
 120
 121 /* The ILP scheduling environment */
 122 typedef struct {
 123         phase_t              ph;            /**< The phase */
 124         ir_graph             *irg;          /**< The current irg */
 125         heights_t            *height;       /**< The heights object of the irg */
 126         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 127         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 128         const arch_env_t     *arch_env;
 129         const arch_isa_t     *isa;          /**< The ISA */
 130         const be_main_env_t  *main_env;
 131         const be_machine_t   *cpu;          /**< the current abstract machine */
 132         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 133         const be_irg_t       *birg;         /**< The birg object */
 134         be_options_t         *be_opts;      /**< backend options */
 135         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 136         DEBUG_ONLY(firm_dbg_module_t *dbg);
 137 } be_ilpsched_env_t;
 138
 139 /* convenience macros to handle phase irn data */
 140 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 141 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 142 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 143 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 144
 145 /* check if node is considered for ILP scheduling */
 146 #define consider_for_sched(isa, irn) \
 147         (! (is_Block(irn)            ||  \
 148                 is_normal_Proj(isa, irn) ||  \
 149                 is_Phi(irn)              ||  \
 150                 is_NoMem(irn)            ||  \
 151                 is_Unknown(irn)          ||  \
 152                 is_End(irn)                  \
 153                 ))
 154
 155 /* gives the valid scheduling time step interval for a node */
 156 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 157
 158 /* gives the valid interval where a node can die */
 159 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 160
 161 /* gives the corresponding ILP variable for given node, unit and time step */
 162 #define ILPVAR_IDX(na, unit, control_step) \
 163         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 164
 165 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 166 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 167         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 168
 169 /* check if a double value is within an epsilon environment of 0 */
 170 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 171
 172 #define ilp_timer_push(t)         lc_timer_push((t))
 173 #define ilp_timer_pop()           lc_timer_pop()
 174 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
 175
 176 /* option variable */
 177 static ilpsched_options_t ilp_opts = {
 178         1,     /* default is with register pressure constraints */
 179         300,   /* 300 sec per block time limit */
 180         ""     /* no log file */
 181 };
 182
 183 /* ILP options */
 184 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 185         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 186         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 187         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 188         { NULL }
 189 };
 190
 191 /*
 192         We need this global variable as we compare nodes dependent on heights,
 193         but we cannot pass any information to the qsort compare function.
 194 */
 195 static heights_t *glob_heights;
 196
 197 /**
 198  * Check if irn is a Proj, which has no execution units assigned.
 199  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 200  */
 201 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
 202         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
 203 }
 204
 205 /**
 206  * Skips normal Projs.
 207  * @return predecessor if irn is a normal Proj, otherwise irn.
 208  */
 209 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
 210         if (is_normal_Proj(isa, irn))
 211                 return get_Proj_pred(irn);
 212         return irn;
 213 }
 214
 215 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 216         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 217         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 218                 lat = 1;
 219         return lat;
 220 }
 221
 222 static int cmp_live_in_nodes(const void *a, const void *b) {
 223         const ilp_livein_node_t *n1 = a;
 224         const ilp_livein_node_t *n2 = b;
 225
 226         return n1->irn != n2->irn;
 227 }
 228
 229 /**
 230  * Compare scheduling time steps of two be_ilpsched_irn's.
 231  */
 232 static int cmp_ilpsched_irn(const void *a, const void *b) {
 233         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 234         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 235         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 236         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 237
 238         if (n1_a->sched_point == n2_a->sched_point) {
 239                 ir_node *irn_a = n1->irn;
 240                 ir_node *irn_b = n2->irn;
 241
 242                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 243                         return 1;
 244                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 245                         return -1;
 246
 247                 /*
 248                         Ok, timestep is equal and the nodes are parallel,
 249                         so check latency and schedule high latency first.
 250                 */
 251                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 252         }
 253         else
 254                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 255 }
 256
 257 /**
 258  * In case there is no phase information for irn, initialize it.
 259  */
 260 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
 261         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 262
 263         if (res == old) {
 264                 /* if we have already some data: check for reinitialization */
 265
 266                 if (! is_Block(irn)) {
 267                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 268
 269                         if (! na->transitive_block_nodes) {
 270                                 ir_node               *block      = get_nodes_block(irn);
 271                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 272                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 273
 274                                 /* we are called after the block indices have been build: create bitset */
 275                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 276                         }
 277                         else {
 278                                 /* we are called from reinit block data: clear the bitset */
 279                                 bitset_clear_all(na->transitive_block_nodes);
 280                                 na->visit_idx    = 0;
 281                                 na->alap_changed = 1;
 282                         }
 283                 }
 284                 return old;
 285         }
 286
 287         res->irn = irn;
 288
 289         /* set ilpsched irn attributes (either block or irn) */
 290         if (is_Block(irn)) {
 291                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 292
 293                 ba->n_interesting_nodes = 0;
 294                 ba->block_last_idx      = 0;
 295                 ba->root_nodes          = plist_new();
 296                 ba->head_ilp_nodes      = NULL;
 297                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 298                 ba->max_steps           = 0;
 299         }
 300         else {
 301                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 302                 memset(na, 0, sizeof(*na));
 303         }
 304
 305         return res;
 306 }
 307
 308 /**
 309  * Assign a per block unique number to each node.
 310  */
 311 static void build_block_idx(ir_node *irn, void *walk_env) {
 312         be_ilpsched_env_t     *env = walk_env;
 313         be_ilpsched_irn_t     *node, *block_node;
 314         ilpsched_node_attr_t  *na;
 315         ilpsched_block_attr_t *ba;
 316
 317         set_irn_link(irn, NULL);
 318         if (! consider_for_sched(env->arch_env->isa, irn))
 319                 return;
 320
 321         node       = get_ilpsched_irn(env, irn);
 322         na         = get_ilpsched_node_attr(node);
 323         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 324         ba         = get_ilpsched_block_attr(block_node);
 325
 326         na->block_idx = ba->block_last_idx++;
 327 }
 328
 329 /********************************************************
 330  *                              __        _
 331  *                             / /       | |
 332  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 333  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 334  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 335  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 336  *                 | |                           | |
 337  *                 |_|                           |_|
 338  ********************************************************/
 339
 340 /**
 341  * Add all nodes having no user in current block to last_nodes list.
 342  */
 343 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 344         ir_node               *block;
 345         const ir_edge_t       *edge;
 346         be_ilpsched_irn_t     *block_node, *node;
 347         ilpsched_block_attr_t *ba;
 348         ilpsched_node_attr_t  *na;
 349         int                   i, j;
 350         be_ilpsched_env_t     *env           = walk_env;
 351         int                   has_block_user = 0;
 352         unsigned              n_consumer     = 0;
 353         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 354         ir_node               **consumer;
 355         int                   idx;
 356
 357         if (! consider_for_sched(env->arch_env->isa, irn))
 358                 return;
 359
 360         block    = get_nodes_block(irn);
 361     idx      = get_irn_idx(irn);
 362         consumer = NEW_ARR_F(ir_node *, 0);
 363
 364         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 365
 366         /* check data and dependency out edges */
 367         for (i = 0; i < 2 && ! has_block_user; ++i) {
 368                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 369                         ir_node *user = get_edge_src_irn(edge);
 370
 371                         if (is_normal_Proj(env->arch_env->isa, user)) {
 372                                 const ir_edge_t *user_edge;
 373
 374                                 if (get_irn_mode(user) == mode_X)
 375                                         continue;
 376
 377                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 378                                 for (j = 0; j < 2; ++j) {
 379                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 380                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 381
 382                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 383                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 384                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 385
 386                                                         /* skip already visited nodes */
 387                                                         if (ua->consumer_idx == idx)
 388                                                                 continue;
 389
 390                                                         /* check if node has user in this block and collect the user if it's a data user */
 391                                                         if (get_nodes_block(real_user) == block) {
 392                                                                 if (i == 0 && j == 0)
 393                                                                         ARR_APP1(ir_node *, consumer, real_user);
 394                                                                 has_block_user = 1;
 395                                                         }
 396
 397                                                         /* only count data consumer */
 398                                                         if (i == 0)
 399                                                                 n_consumer++;
 400
 401                                                         /* mark user as visited by this node */
 402                                                         ua->consumer_idx = idx;
 403                                                 }
 404                                         }
 405                                 }
 406                         }
 407                         else if (is_Block(user)) {
 408                                 continue;
 409                         }
 410                         else if (! is_Phi(user)) {
 411                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 412                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 413
 414                                 /* skip already visited nodes */
 415                                 if (ua->consumer_idx == idx)
 416                                         continue;
 417
 418                                 /* check if node has user in this block and collect the user if it's a data user */
 419                                 if (get_nodes_block(user) == block) {
 420                                         if (i == 0)
 421                                                 ARR_APP1(ir_node *, consumer, user);
 422                                         has_block_user = 1;
 423                                 }
 424
 425                                 /* only count data consumer */
 426                                 if (i == 0)
 427                                         n_consumer++;
 428
 429                                 /* mark user visited by this node */
 430                                 ua->consumer_idx = idx;
 431                         }
 432                         else if (get_nodes_block(user) != block) {
 433                                 n_consumer++;
 434                         }
 435                 }
 436         }
 437
 438         block_node = get_ilpsched_irn(env, block);
 439         ba         = get_ilpsched_block_attr(block_node);
 440
 441         ba->n_interesting_nodes++;
 442
 443         /* current irn has no user inside this block, add to queue */
 444         if (! has_block_user) {
 445                 DB((env->dbg, LEVEL_3, "root node\n"));
 446                 plist_insert_back(ba->root_nodes, irn);
 447         }
 448         else {
 449                 DB((env->dbg, LEVEL_3, "normal node\n"));
 450         }
 451
 452         /* record number of all consumer and the consumer within the same block */
 453         node = get_ilpsched_irn(env, irn);
 454         na   = get_ilpsched_node_attr(node);
 455         na->n_consumer     = n_consumer;
 456         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 457         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 458         DEL_ARR_F(consumer);
 459 }
 460
 461 /**
 462  * Calculate the ASAP scheduling step for current irn.
 463  */
 464 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 465         be_ilpsched_env_t     *env = walk_env;
 466         int                   i;
 467         ir_node               *block;
 468         be_ilpsched_irn_t     *node, *block_node;
 469         ilpsched_node_attr_t  *na;
 470         ilpsched_block_attr_t *ba;
 471
 472         /* These nodes are handled separate */
 473         if (! consider_for_sched(env->arch_env->isa, irn))
 474                 return;
 475
 476         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 477
 478         block    = get_nodes_block(irn);
 479         node     = get_ilpsched_irn(env, irn);
 480         na       = get_ilpsched_node_attr(node);
 481         na->asap = 1;
 482
 483         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 484                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
 485
 486                 /* check for greatest distance to top */
 487                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 488                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 489                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 490                         unsigned             lat;
 491
 492                         lat         = fixed_latency(env->sel, pred, env->block_env);
 493                         na->latency = lat;
 494                         na->asap    = MAX(na->asap, pna->asap + lat);
 495                 }
 496         }
 497
 498         /* add node to ILP node list and update max_steps */
 499         block_node = get_ilpsched_irn(env, block);
 500         ba         = get_ilpsched_block_attr(block_node);
 501
 502         set_irn_link(irn, ba->head_ilp_nodes);
 503         ba->head_ilp_nodes = irn;
 504         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 505
 506         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 507 }
 508
 509 /**
 510  * Calculate the ALAP scheduling step of all irns in current block.
 511  * Depends on max_steps being calculated.
 512  */
 513 static void calculate_block_alap(ir_node *block, void *walk_env) {
 514         be_ilpsched_env_t     *env        = walk_env;
 515         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 516         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 517         waitq                 *cur_queue  = new_waitq();
 518         plist_element_t       *el;
 519
 520         assert(is_Block(block));
 521
 522         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 523                 block, ba->n_interesting_nodes, ba->max_steps));
 524
 525         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 526         //phase_reinit_block_irn_data(&env->ph, block);
 527
 528         /* init start queue */
 529         foreach_plist(ba->root_nodes, el) {
 530                 waitq_put(cur_queue, plist_element_get_value(el));
 531         }
 532
 533         /* repeat until all nodes are processed */
 534         while (! waitq_empty(cur_queue)) {
 535                 waitq *next_queue = new_waitq();
 536
 537                 /* process all nodes in current step */
 538                 while (! waitq_empty(cur_queue)) {
 539                         ir_node              *cur_irn = waitq_get(cur_queue);
 540                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 541                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 542                         int                  i;
 543
 544                         /* cur_node has no alap set -> it's a root node, set to max alap */
 545                         if (na->alap == 0) {
 546                                 na->alap = ba->max_steps;
 547                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 548                                         cur_irn, na->alap));
 549                         }
 550                         else {
 551                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 552                                         cur_irn, na->alap));
 553                         }
 554
 555                         /* set the alap's of all predecessors */
 556                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 557                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
 558
 559                                 /* check for greatest distance to bottom */
 560                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 561                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 562                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 563                                         unsigned             lat;
 564
 565                                         /* mark the predecessor as visited by current irn */
 566                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 567                                                 continue;
 568                                         pna->visit_idx = get_irn_idx(cur_irn);
 569
 570                                         lat = fixed_latency(env->sel, pred, env->block_env);
 571
 572                                         /* set ALAP of current pred */
 573                                         if (pna->alap == 0) {
 574                                                 /* current ALAP is 0: node has not yet been visited */
 575                                                 pna->alap_changed = 1;
 576                                                 pna->alap         = na->alap - lat;
 577                                         }
 578                                         else if (pna->alap > na->alap - lat) {
 579                                                 /* we found a longer path to root node: change ALAP */
 580                                                 pna->alap         = na->alap - lat;
 581                                                 pna->alap_changed = 1;
 582                                         }
 583                                         else {
 584                                                 /* current ALAP is best found so far: keep it */
 585                                                 pna->alap_changed = 0;
 586                                         }
 587
 588                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 589
 590                                         /* enqueue node for next iteration */
 591                                         if (get_irn_ins_or_deps(pred) > 0)
 592                                                 waitq_put(next_queue, pred);
 593                                 }
 594                         }
 595                 }
 596
 597                 /* prepare for next iteration */
 598                 del_waitq(cur_queue);
 599                 cur_queue = next_queue;
 600         }
 601 }
 602
 603 /**
 604  * Free list of root nodes and the set of live-in nodes.
 605  */
 606 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 607         be_ilpsched_env_t     *env        = walk_env;
 608         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 609         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 610
 611         plist_free(ba->root_nodes);
 612         ba->root_nodes = NULL;
 613         del_pset(ba->livein_nodes);
 614         ba->livein_nodes = NULL;
 615 }
 616
 617 /**
 618  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 619  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 620  */
 621 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 622         be_ilpsched_env_t    *env  = walk_env;
 623         ir_node              *pred = irn;
 624         be_ilpsched_irn_t    *node, *pred_node;
 625         ilpsched_node_attr_t *na, *pna;
 626
 627         if (! consider_for_sched(env->arch_env->isa, irn))
 628                 return;
 629
 630         if (! is_Proj(irn) && ! be_is_Keep(irn))
 631                 return;
 632
 633         /* go to the ancestor */
 634         if (be_is_Keep(irn))
 635                 pred = get_irn_n(irn, 0);
 636         pred = skip_Proj(pred);
 637
 638         node      = get_ilpsched_irn(env, irn);
 639         pred_node = get_ilpsched_irn(env, pred);
 640         na        = get_ilpsched_node_attr(node);
 641         pna       = get_ilpsched_node_attr(pred_node);
 642
 643         na->asap = pna->asap;
 644         na->alap = pna->alap;
 645
 646         /* record all Projs and Keeps for this node */
 647         if (! pna->projkeeps)
 648                 pna->projkeeps = new_waitq();
 649         waitq_put(pna->projkeeps, irn);
 650
 651         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 652 }
 653
 654 /*******************************************
 655  *           _              _       _
 656  *          | |            | |     | |
 657  *  ___  ___| |__   ___  __| |_   _| | ___
 658  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 659  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 660  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 661  *
 662  *******************************************/
 663
 664 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
 665         const ir_edge_t *edge;
 666
 667         foreach_out_edge(irn, edge) {
 668                 ir_node *user = get_edge_src_irn(edge);
 669
 670                 if (be_is_Keep(user)) {
 671                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 672                         waitq_put(keeps, user);
 673                 }
 674         }
 675 }
 676
 677 /**
 678  * Inserts @p irn before @p before into schedule and notifies backend.
 679  */
 680 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 681         ir_node *before, ir_node *irn, unsigned cycle)
 682 {
 683         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 684         sched_add_before(before, irn);
 685 }
 686
 687 /**
 688  * Adds a node, it's Projs (in case of mode_T nodes) and
 689  * it's Keeps to schedule.
 690  */
 691 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
 692         const ir_edge_t *edge;
 693         waitq           *keeps = new_waitq();
 694
 695         /* mode_M nodes are not scheduled */
 696         if (get_irn_mode(irn) == mode_M)
 697                 return;
 698
 699         if (! sched_is_scheduled(irn))
 700                 notified_sched_add_before(env, block, irn, cycle);
 701
 702         /* add Projs */
 703         if (get_irn_mode(irn) == mode_T) {
 704                 foreach_out_edge(irn, edge) {
 705                         ir_node *user = get_edge_src_irn(edge);
 706
 707                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 708                                 get_irn_n_edges(user) > 0)
 709                         {
 710                                 notified_sched_add_before(env, block, user, cycle);
 711                         }
 712
 713                         check_for_keeps(keeps, block, user);
 714                 }
 715         }
 716         else {
 717                 check_for_keeps(keeps, block, irn);
 718         }
 719
 720         /* add Keeps */
 721         while (! waitq_empty(keeps)) {
 722                 ir_node *keep = waitq_get(keeps);
 723                 if (! sched_is_scheduled(keep))
 724                         notified_sched_add_before(env, block, keep, cycle);
 725         }
 726
 727         del_waitq(keeps);
 728 }
 729
 730 /**
 731  * Schedule all nodes in the given block, according to the ILP solution.
 732  */
 733 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 734         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 735         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 736         sched_info_t          *info       = get_irn_sched_info(block);
 737         be_ilpsched_irn_t     **sched_nodes;
 738         unsigned              i, l;
 739         ir_node               *cfop, *irn;
 740         const ir_edge_t       *edge;
 741
 742         /* init block schedule list */
 743         INIT_LIST_HEAD(&info->list);
 744         info->scheduled = 1;
 745
 746         /* collect nodes and their scheduling time step */
 747         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 748         if (ba->n_interesting_nodes == 0) {
 749                 /* ignore */
 750         }
 751         else if (ba->n_interesting_nodes == 1) {
 752                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 753
 754                 /* add the single node */
 755                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 756         }
 757         else {
 758                 /* check all nodes for their positive solution */
 759                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 760                         be_ilpsched_irn_t    *node;
 761                         ilpsched_node_attr_t *na;
 762                         int                  tp_idx, found;
 763                         unsigned             cur_var, t;
 764
 765                         node    = get_ilpsched_irn(env, irn);
 766                         na      = get_ilpsched_node_attr(node);
 767                         cur_var = 0;
 768                         found   = 0;
 769
 770                         if (! na->is_dummy_node) {
 771                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 772                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 773                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 774
 775                                                 if (! LPP_VALUE_IS_0(cost)) {
 776                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
 777                                                         found = 1;
 778                                                 }
 779                                         }
 780                                 }
 781                         }
 782
 783                         found = 0;
 784                         /* go over all variables of a node until the non-zero one is found */
 785                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 786                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 787                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 788
 789                                         /* check, if variable is set to one (it's not zero then :) */
 790                                         if (! LPP_VALUE_IS_0(val)) {
 791                                                 na->sched_point = t;
 792                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 793                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 794                                                         irn, t, na->type_info[tp_idx].tp->name));
 795                                                 found = 1;
 796                                         }
 797                                 }
 798                         }
 799                 }
 800
 801                 glob_heights = env->height;
 802                 /* sort nodes ascending by scheduling time step */
 803                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 804         }
 805
 806         /* make all Phis ready and remember the single cf op */
 807         cfop = NULL;
 808         foreach_out_edge(block, edge) {
 809                 irn = get_edge_src_irn(edge);
 810
 811                 switch (get_irn_opcode(irn)) {
 812                         case iro_Phi:
 813                                 add_to_sched(env, block, irn, 0);
 814                                 break;
 815                         case iro_Start:
 816                         case iro_End:
 817                         case iro_Proj:
 818                         case iro_Bad:
 819                         case iro_Unknown:
 820                                 break;
 821                         default:
 822                                 if (is_cfop(irn)) {
 823                                         assert(cfop == NULL && "Highlander - there can be only one");
 824                                         cfop = irn;
 825                                 }
 826                         break;
 827                 }
 828         }
 829
 830         /* add all nodes from list */
 831         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 832                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 833                 if (sched_nodes[i]->irn != cfop)
 834                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 835         }
 836
 837         /* schedule control flow node if not already done */
 838         if (cfop && ! sched_is_scheduled(cfop))
 839                 add_to_sched(env, block, cfop, 0);
 840
 841         DEL_ARR_F(sched_nodes);
 842 }
 843
 844 /***************************************************************
 845  *   _____ _      _____     _____           _   _
 846  *  |_   _| |    |  __ \   / ____|         | | (_)
 847  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 848  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 849  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 850  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 851  *
 852  ***************************************************************/
 853
 854 /**
 855  * Check if node can be executed on given unit type.
 856  */
 857 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 858         int                  i;
 859         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 860
 861         for (i = na->n_unit_types - 1; i >= 0; --i) {
 862                 if (na->type_info[i].tp == tp)
 863                         return i;
 864         }
 865
 866         return -1;
 867 }
 868
 869 /************************************************
 870  *                   _       _     _
 871  *                  (_)     | |   | |
 872  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 873  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 874  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 875  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 876  *
 877  ************************************************/
 878
 879 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
 880         const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
 881         unsigned                  n_unit_types = 0;
 882         be_ilpsched_irn_t         *node;
 883         ilpsched_node_attr_t      *na;
 884         unsigned                  unit_idx, tp_idx;
 885
 886         /* count number of available unit types for this node */
 887         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 888                 /* just count */ ;
 889
 890         node = get_ilpsched_irn(env, irn);
 891         na   = get_ilpsched_node_attr(node);
 892
 893         if (! na->type_info) {
 894                 na->n_unit_types = n_unit_types;
 895                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 896
 897                 /* fill the type info array */
 898                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 899                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 900                                 /* beware: we also count number of available units here */
 901                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 902                                         na->is_dummy_node = 1;
 903                         }
 904
 905                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 906                         na->type_info[tp_idx].n_units = unit_idx;
 907                 }
 908         }
 909
 910         return n_unit_types;
 911 }
 912
 913 /**
 914  * Returns the largest alap time of a user of @p irn.
 915  * The user must be in block @p block.
 916  */
 917 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
 918         const ir_edge_t *edge;
 919         unsigned        max_alap = 0;
 920
 921         foreach_out_edge(irn, edge) {
 922                 ir_node *user = get_edge_src_irn(edge);
 923
 924                 if (get_nodes_block(user) == block) {
 925                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 926                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 927
 928                         max_alap = MAX(max_alap, na->alap);
 929                 }
 930         }
 931
 932         assert(max_alap > 0);
 933         return max_alap;
 934 }
 935
 936 /**
 937  * Create the following variables:
 938  * - x_{nt}^k    binary     weigthed with: t
 939  *      node n is scheduled at time step t to unit type k
 940  * ==>> These variables represent the schedule
 941  *
 942  * - a_{nt}^k    binary     weighted with num_nodes
 943  *      node n is alive at time step t on unit type k
 944  *
 945  * - y_{nt}^k    continuous  weighted with: num_nodes^2
 946  *      register pressure over limit for unit type k
 947  * ==>> These variables represent the register pressure
 948  *
 949  */
 950 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 951         char                  buf[1024];
 952         ir_node               *irn;
 953         unsigned              num_block_var, num_nodes;
 954         ilp_livein_node_t     *livein;
 955         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 956         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 957         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
 958
 959         ilp_timer_push(t_var);
 960         num_block_var = num_nodes = 0;
 961         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 962                 be_ilpsched_irn_t    *node;
 963                 ilpsched_node_attr_t *na;
 964                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
 965                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
 966                 int                  i;
 967
 968                 node         = get_ilpsched_irn(env, irn);
 969                 na           = get_ilpsched_node_attr(node);
 970                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
 971
 972                 /* allocate space for ilp variables */
 973                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 974                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 975
 976                 /* we need these variables only for "real" nodes */
 977                 if (! na->is_dummy_node) {
 978                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 979                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 980
 981                         num_ad         = ba->max_steps - na->asap + 1;
 982                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 983                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
 984                 }
 985
 986                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
 987                         irn, na->asap, na->alap, na->n_unit_types));
 988
 989                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
 990                 /* create variables */
 991                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 992                         unsigned t;
 993
 994                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
 995                                 /* x_{nt}^k variables */
 996                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
 997                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 998                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 999                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1000                                 /* variable counter */
1001                                 n_var++;
1002                                 num_block_var++;
1003
1004                                 if (! na->is_dummy_node) {
1005                                         /* y_{nt}^k variables */
1006                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1007                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1008                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1009                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1010
1011                                         /* variable counter */
1012                                         n_var++;
1013                                         num_block_var++;
1014                                 }
1015                         }
1016
1017                         /* a node can die at any step t: asap(n) <= t <= U */
1018                         if (! na->is_dummy_node) {
1019                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1020
1021                                         /* a_{nt}^k variables */
1022                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1023                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1024                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1025                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1026
1027                                         /* variable counter */
1028                                         n_var++;
1029                                         num_block_var++;
1030                                 }
1031                         }
1032
1033                         /* collect live-in nodes */
1034                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1035                                 ir_node *pred = get_irn_n(irn, i);
1036
1037                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1038                                         be_ilpsched_set_type_info(env, pred, var_obst);
1039                                         if (! na->is_dummy_node) {
1040                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1041                                                 entry->irn = pred;
1042                                                 entry->a   = NULL;
1043                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1044                                         }
1045                                 }
1046                         }
1047                 }
1048
1049                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1050                 num_nodes++;
1051         }
1052
1053         /* create alive variables a_{nt}^k for live-ins */
1054         foreach_pset(ba->livein_nodes, livein) {
1055                 be_ilpsched_irn_t    *node;
1056                 ilpsched_node_attr_t *na;
1057                 unsigned             tp_idx, var_idx;
1058                 ir_node              *irn;
1059
1060                 irn  = livein->irn;
1061                 node = get_ilpsched_irn(env, irn);
1062                 na   = get_ilpsched_node_attr(node);
1063
1064                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1065
1066                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1067                 var_idx   = 0;
1068
1069                 /* create variables */
1070                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1071                         unsigned t;
1072
1073                         for (t = 0; t < livein->max_alive_steps; ++t) {
1074                                 /* a_{nt}^k variables */
1075                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1076                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1077                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1078                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1079                                 num_block_var++;
1080                         }
1081                 }
1082         }
1083
1084         ilp_timer_pop();
1085         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1086                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1087 }
1088
1089 /*******************************************************
1090  *                      _             _       _
1091  *                     | |           (_)     | |
1092  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1093  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1094  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1095  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1096  *
1097  *******************************************************/
1098
1099 /**
1100  * Create following ILP constraints:
1101  * - the assignment constraints:
1102  *     assure each node is executed once by exactly one (allowed) execution unit
1103  * - the dead node assignment constraints:
1104  *     assure a node can only die at most once
1105  * - the precedence constraints:
1106  *     assure that no data dependencies are violated
1107  */
1108 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1109         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1110         char                  buf[1024];
1111         ir_node               *irn;
1112         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1113         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1114         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1115         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1116
1117         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1118         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1119                 int                  cst, tp_idx, i;
1120                 unsigned             cur_var;
1121                 be_ilpsched_irn_t    *node;
1122                 ilpsched_node_attr_t *na;
1123
1124                 node    = get_ilpsched_irn(env, irn);
1125                 na      = get_ilpsched_node_attr(node);
1126                 cur_var = 0;
1127
1128                 /* the assignment constraint */
1129                 ilp_timer_push(t_cst_assign);
1130                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1131                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1132                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1133                 num_cst_assign++;
1134
1135                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1136                 ilp_timer_pop();
1137
1138                 /* We have separate constraints for Projs and Keeps */
1139                 // ILP becomes infeasible ?!?
1140 //              if (is_Proj(irn) || be_is_Keep(irn))
1141 //                      continue;
1142
1143                 /* the precedence constraints */
1144                 ilp_timer_push(t_cst_prec);
1145                 bs_block_irns = bitset_clear_all(bs_block_irns);
1146                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1147                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1148                         unsigned             t_low, t_high, t;
1149                         be_ilpsched_irn_t    *pred_node;
1150                         ilpsched_node_attr_t *pna;
1151                         unsigned             delay;
1152
1153                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1154                                 continue;
1155
1156                         pred_node = get_ilpsched_irn(env, pred);
1157                         pna       = get_ilpsched_node_attr(pred_node);
1158
1159                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1160
1161                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1162                                 bitset_set(bs_block_irns, pna->block_idx);
1163                         else
1164                                 continue;
1165
1166                         /* irn = n, pred = m */
1167                         delay  = fixed_latency(env->sel, pred, env->block_env);
1168                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1169                         t_high = MIN(na->alap, pna->alap + delay - 1);
1170                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1171                                 unsigned tn, tm;
1172                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1173
1174                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1175                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1176                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1177                                 num_cst_prec++;
1178
1179                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1180                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1181                                         /* node variables have smaller index than pred variables */
1182                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1183                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1184                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1185                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1186                                                 }
1187                                         }
1188
1189                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1190                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1191                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1192                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1193                                                 }
1194                                         }
1195                                 }
1196                                 else {
1197                                         /* pred variables have smaller index than node variables */
1198                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1199                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1200                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1201                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1202                                                 }
1203                                         }
1204
1205                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1206                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1207                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1208                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1209                                                 }
1210                                         }
1211                                 }
1212
1213                                 if (ARR_LEN(tmp_var_idx) > 0)
1214                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1215
1216                                 DEL_ARR_F(tmp_var_idx);
1217                         }
1218                 }
1219                 ilp_timer_pop();
1220         }
1221         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1222                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1223         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1224                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1225 }
1226
1227 /**
1228  * Create ILP resource constraints:
1229  * - assure that for each time step not more instructions are scheduled
1230  *   to the same unit types as units of this type are available
1231  */
1232 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1233         int                   glob_type_idx;
1234         char                  buf[1024];
1235         unsigned              num_cst_resrc = 0;
1236         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1237         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1238
1239         ilp_timer_push(t_cst_rsrc);
1240         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1241                 unsigned                 t;
1242                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1243
1244                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1245
1246                 /* check each time step */
1247                 for (t = 0; t < ba->max_steps; ++t) {
1248                         ir_node *irn;
1249                         int     cst;
1250                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1251
1252                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1253                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1254                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1255                         num_cst_resrc++;
1256
1257                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1258                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1259                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1260                                 int                  tp_idx;
1261
1262                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1263
1264                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1265                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1266                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1267                                 }
1268                         }
1269
1270                         /* set constraints if we have some */
1271                         if (ARR_LEN(tmp_var_idx) > 0)
1272                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1273
1274                         DEL_ARR_F(tmp_var_idx);
1275                 }
1276         }
1277         ilp_timer_pop();
1278         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1279                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1280 }
1281
1282 /**
1283  * Create ILP bundle constraints:
1284  * - assure, at most bundle_size * bundles_per_cycle instructions
1285  *   can be started at a certain point.
1286  */
1287 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1288         char                  buf[1024];
1289         unsigned              t;
1290         unsigned              num_cst_bundle = 0;
1291         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1292         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1293         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1294
1295         ilp_timer_push(t_cst_bundle);
1296         for (t = 0; t < ba->max_steps; ++t) {
1297                 ir_node *irn;
1298                 int     cst;
1299                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1300
1301                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1302                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1303                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1304                 num_cst_bundle++;
1305
1306                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1307                         be_ilpsched_irn_t    *node;
1308                         ilpsched_node_attr_t *na;
1309                         int                  tp_idx;
1310
1311                         /* Projs and Keeps do not contribute to bundle size */
1312                         if (is_Proj(irn) || be_is_Keep(irn))
1313                                 continue;
1314
1315                         node = get_ilpsched_irn(env, irn);
1316                         na   = get_ilpsched_node_attr(node);
1317
1318                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1319                         if (na->is_dummy_node)
1320                                 continue;
1321
1322                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1323                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1324                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1325                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1326                                 }
1327                         }
1328                 }
1329
1330                 if (ARR_LEN(tmp_var_idx) > 0)
1331                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1332
1333                 DEL_ARR_F(tmp_var_idx);
1334         }
1335         ilp_timer_pop();
1336         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1337                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1338 }
1339
1340 /**
1341  * Create ILP alive nodes constraints:
1342  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1343  */
1344 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1345         char                  buf[1024];
1346         ir_node               *irn;
1347         unsigned              num_cst = 0;
1348         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1349         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1350
1351         ilp_timer_push(t_cst);
1352         /* for each node */
1353         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1354                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1355                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1356                 unsigned             t;
1357
1358                 /* we ignore nodes assigned to dummy unit here */
1359                 if (na->is_dummy_node)
1360                         continue;
1361
1362                 /* check check all time steps: asap(n) <= t <= U */
1363                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1364                         int node_tp_idx;
1365
1366                         /* for all unit types available for this node */
1367                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1368                                 unsigned tn, tn_max, idx;
1369                                 int      cst, i;
1370                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1371                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1372
1373                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1374                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1375                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1376                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1377                                 num_cst++;
1378
1379                                 tn_max = MIN(na->alap - 1, t);
1380                                 /* check if the node has been scheduled so far */
1381                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1382                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1383                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1384                                 }
1385
1386                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1387                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1388                                 DEL_ARR_F(tmp_var_idx_n);
1389
1390                                 /* subtract the number of consumer scheduled so far */
1391                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1392                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1393                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1394                                         int                  tp_idx;
1395                                         unsigned             tm, tm_max;
1396
1397                                         tm_max = MIN(ca->alap - 1, t);
1398                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1399                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1400                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1401                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1402                                                 }
1403                                         }
1404                                 }
1405
1406                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1407                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1408                                 DEL_ARR_F(tmp_var_idx_m);
1409
1410                                 /* -c * a_{nt}^k */
1411                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1412                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1413
1414                         }
1415                 }
1416         }
1417         ilp_timer_pop();
1418         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1419                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1420 }
1421
1422 /**
1423  * Create ILP alive nodes constraints for live-in nodes:
1424  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1425  */
1426 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1427         char                  buf[1024];
1428         ilp_livein_node_t     *livein;
1429         unsigned              num_cst = 0;
1430         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1431         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1432
1433         ilp_timer_push(t_cst);
1434         /* for each node */
1435         foreach_pset(ba->livein_nodes, livein) {
1436                 ir_node              *irn  = livein->irn;
1437                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1438                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1439                 unsigned             t;
1440
1441                 /* check check all time steps: 0 <= t < max_alive_steps */
1442                 for (t = 0; t < livein->max_alive_steps; ++t) {
1443                         int node_tp_idx;
1444
1445                         /* for all unit types available for this node */
1446                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1447                                 const ir_edge_t *edge;
1448                                 unsigned idx;
1449                                 int      cst, num_block_user;
1450                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1451
1452                                 /* check the number of consumer scheduled so far */
1453                                 num_block_user = 0;
1454                                 foreach_out_edge(irn, edge) {
1455                                         ir_node              *user = get_edge_src_irn(edge);
1456                                         be_ilpsched_irn_t    *cons;
1457                                         ilpsched_node_attr_t *ca;
1458                                         int                  tp_idx;
1459                                         unsigned             tm, tm_max;
1460
1461                                         /* check only users within current block */
1462                                         if (get_nodes_block(user) != block_node->irn)
1463                                                 continue;
1464
1465                                         num_block_user++;
1466                                         cons = get_ilpsched_irn(env, user);
1467                                         ca   = get_ilpsched_node_attr(cons);
1468
1469                                         tm_max = MIN(ca->alap - 1, t);
1470                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1471                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1472                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1473                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1474                                                 }
1475                                         }
1476                                 }
1477
1478                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1479                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1480                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1481                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1482                                 num_cst++;
1483
1484                                 /* sum(scheduled users) */
1485                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1486                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1487                                 DEL_ARR_F(tmp_var_idx_m);
1488
1489                                 /* + c * a_{nt}^k */
1490                                 idx = node_tp_idx * livein->max_alive_steps + t;
1491                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1492                         }
1493                 }
1494         }
1495         ilp_timer_pop();
1496         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1497                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1498 }
1499
1500 /**
1501  * Create ILP pressure constraints, based on alive nodes:
1502  * - add additional costs to objective function if a node is scheduled
1503  *   on a unit although all units of this type are currently occupied
1504  */
1505 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1506         char                  buf[1024];
1507         ir_node               *cur_irn;
1508         unsigned              num_cst = 0;
1509         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1510         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1511
1512         ilp_timer_push(t_cst);
1513         /* y_{nt}^k is set for each node and timestep and unit type */
1514         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1515                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1516                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1517                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1518                 int                  glob_type_idx;
1519
1520                 /* we ignore nodes assigned to DUMMY unit here */
1521                 if (cur_na->is_dummy_node)
1522                         continue;
1523
1524                 /* for all types */
1525                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1526                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1527                         int                      cur_tp_idx;
1528                         unsigned                 t;
1529
1530                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1531
1532                         /* check if node can be executed on this unit type */
1533                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1534                         if (cur_tp_idx < 0)
1535                                 continue;
1536
1537                         /* check all time_steps at which the current node can be scheduled */
1538                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1539                                 int     cst, y_idx;
1540                                 ir_node *irn;
1541                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1542                                 ilp_livein_node_t *livein;
1543
1544                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1545                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1546                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1547                                 num_cst++;
1548
1549                                 /* - accumulate all nodes alive at point t on unit type k */
1550                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1551                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1552                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1553                                         int                  a_idx, tp_idx;
1554
1555                                         /* check if node can be alive here */
1556                                         if (t < na->asap - 1)
1557                                                 continue;
1558
1559                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1560
1561                                         /* current type is not suitable */
1562                                         if (tp_idx < 0)
1563                                                 continue;
1564
1565                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1566                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1567                                 }
1568                                 /* do the same for livein nodes */
1569                                 foreach_pset(ba->livein_nodes, livein) {
1570                                         ir_node              *irn  = livein->irn;
1571                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1572                                         int                  a_idx, tp_idx;
1573
1574                                         /* check if node can be alive here */
1575                                         if (t >= livein->max_alive_steps)
1576                                                 continue;
1577
1578                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1579
1580                                         /* current type is not suitable */
1581                                         if (tp_idx < 0)
1582                                                 continue;
1583
1584                                         a_idx = tp_idx * livein->max_alive_steps + t;
1585                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1586                                 }
1587
1588                                 if (ARR_LEN(tmp_var_idx) > 0)
1589                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1590                                 DEL_ARR_F(tmp_var_idx);
1591
1592                                 /* - num_nodes * y_{nt}^k */
1593                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1594                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1595                         }
1596                 }
1597         }
1598         ilp_timer_pop();
1599         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1600                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1601 }
1602
1603 /**
1604  * Create ILP branch constraints:
1605  * Assure, alle nodes are scheduled prior to cfg op.
1606  */
1607 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1608         char                  buf[1024];
1609         ir_node               *cur_irn, *cfop;
1610         unsigned              num_cst          = 0;
1611         unsigned              num_non_branches = 0;
1612         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1613         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1614
1615         ilp_timer_push(t_cst);
1616         cfop = NULL;
1617         /* determine number of non-branch nodes and the one and only branch node */
1618         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1619                 switch (get_irn_opcode(cur_irn)) {
1620                         case iro_Phi:
1621                         case iro_Start:
1622                         case iro_End:
1623                         case iro_Proj:
1624                         case iro_Bad:
1625                         case iro_Unknown:
1626                                 num_non_branches++;
1627                                 break;
1628                         default:
1629                                 if (is_cfop(cur_irn)) {
1630                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1631                                         cfop = cur_irn;
1632                                 }
1633                                 else {
1634                                         num_non_branches++;
1635                                 }
1636                                 break;
1637                 }
1638         }
1639
1640         if (cfop) {
1641                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1642                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1643                 unsigned t;
1644
1645                 /* for each time step */
1646                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1647                         int *non_branch_vars, *branch_vars;
1648                         int cst;
1649
1650                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1651                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1652                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1653                         num_cst++;
1654
1655                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1656                         non_branch_vars = NEW_ARR_F(int, 0);
1657                         branch_vars     = NEW_ARR_F(int, 0);
1658                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1659                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1660                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1661                                 int                  tp_idx;
1662
1663                                 if (cur_irn == cfop) {
1664                                         /* for all unit types available for this node */
1665                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1666                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1667                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1668                                         }
1669                                 }
1670                                 else {
1671                                         /* sum up all possible schedule points for this node upto current timestep */
1672                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1673                                                 unsigned tn;
1674                                                 unsigned tmax = MIN(t, na->alap - 1);
1675
1676                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1677                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1678                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1679                                                 }
1680                                         }
1681                                 }
1682
1683                         }
1684
1685                         if (ARR_LEN(non_branch_vars) > 0)
1686                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1687                         if (ARR_LEN(branch_vars) > 0)
1688                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1689
1690                         DEL_ARR_F(branch_vars);
1691                         DEL_ARR_F(non_branch_vars);
1692                 }
1693         }
1694         ilp_timer_pop();
1695         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1696                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1697 }
1698
1699 #if 0
1700 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1701         char                  buf[1024];
1702         ir_node               *irn;
1703         unsigned              num_cst = 0;
1704         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1705         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1706
1707         ilp_timer_push(t_cst);
1708         /* check all nodes */
1709         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1710                 be_ilpsched_irn_t    *node;
1711                 ilpsched_node_attr_t *na;
1712                 unsigned             t;
1713                 ir_node              **pk;
1714
1715                 /* only mode_T nodes can have Projs and Keeps assigned */
1716                 if (get_irn_mode(irn) != mode_T)
1717                         continue;
1718
1719                 node = get_ilpsched_irn(env, irn);
1720                 na   = get_ilpsched_node_attr(node);
1721
1722                 /* check if has some Projs and Keeps assigned */
1723                 if (! na->projkeeps)
1724                         continue;
1725
1726                 /* we can run only once over the queue, so preserve the nodes */
1727                 pk = NEW_ARR_F(ir_node *, 0);
1728                 while (! waitq_empty(na->projkeeps))
1729                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1730                 del_waitq(na->projkeeps);
1731                 na->projkeeps = NULL;
1732
1733                 /* for all time steps at which this node can be scheduled */
1734                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1735                         int cst, tp_idx, i;
1736                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1737
1738                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1739                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1740                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1741                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1742                         num_cst++;
1743
1744                         /* sum up scheduling variables for this time step */
1745                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1746                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1747                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1748                         }
1749
1750                         if (ARR_LEN(tmp_var_idx_n) > 0)
1751                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1752                         DEL_ARR_F(tmp_var_idx_n);
1753
1754                         /* subtract all Proj and Keep variables for this step */
1755                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1756                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1757                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1758                                 int                  pk_tp_idx;
1759
1760                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1761                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1762                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1763                                 }
1764                         }
1765                 }
1766         }
1767         ilp_timer_pop();
1768         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1769                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1770 }
1771 #endif /* if 0 */
1772
1773 /***************************************************
1774  *  _____ _      _____                    _
1775  * |_   _| |    |  __ \                  (_)
1776  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1777  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1778  *  _| |_| |____| |      | | | | | | (_| | | | | |
1779  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1780  *
1781  ***************************************************/
1782
1783 /**
1784  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1785  */
1786 static void create_ilp(ir_node *block, void *walk_env) {
1787         be_ilpsched_env_t     *env           = walk_env;
1788         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1789         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1790         FILE                  *logfile       = NULL;
1791         lpp_t                 *lpp           = NULL;
1792         int                   need_heur      = 0;
1793         struct obstack        var_obst;
1794         char                  name[1024];
1795
1796         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1797         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1798         DBG((env->dbg, 255, "=========================================\n\n"));
1799
1800         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1801                 block, ba->n_interesting_nodes, ba->max_steps));
1802
1803         /* notify backend and get block environment */
1804         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1805
1806         /* if we have less than two interesting nodes, there is no need to create the ILP */
1807         if (ba->n_interesting_nodes > 1) {
1808                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1809                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1810                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1811                 int    estimated_n_var = (int)((double)base_num * fact_var);
1812                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1813
1814                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1815                         estimated_n_var, estimated_n_cst));
1816
1817                 /* set up the LPP object */
1818                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1819
1820                 lpp = new_lpp_userdef(
1821                         (const char *)name,
1822                         lpp_minimize,
1823                         estimated_n_cst,     /* num vars */
1824                         estimated_n_cst + 1, /* num cst */
1825                         1.3);                /* grow factor */
1826                 obstack_init(&var_obst);
1827
1828                 /* create ILP variables */
1829                 create_variables(env, lpp, block_node, &var_obst);
1830
1831                 /* create ILP constraints */
1832                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1833                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1834                 create_ressource_constraints(env, lpp, block_node);
1835                 create_bundle_constraints(env, lpp, block_node);
1836                 create_branch_constraint(env, lpp, block_node);
1837                 //create_proj_keep_constraints(env, lpp, block_node);
1838
1839                 if (env->opts->regpress) {
1840                         create_alive_nodes_constraint(env, lpp, block_node);
1841                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1842                         create_pressure_alive_constraint(env, lpp, block_node);
1843                 }
1844
1845                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1846
1847                 /* debug stuff, dump lpp when debugging is on  */
1848                 DEBUG_ONLY(
1849                         if (firm_dbg_get_mask(env->dbg) > 1) {
1850                                 char buf[1024];
1851                                 FILE *f;
1852
1853                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1854                                 f = fopen(buf, "w");
1855                                 lpp_dump_plain(lpp, f);
1856                                 fclose(f);
1857                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1858                                 lpp_dump(lpp, buf);
1859                         }
1860                 );
1861
1862                 /* set solve time limit */
1863                 lpp_set_time_limit(lpp, env->opts->time_limit);
1864
1865                 /* set logfile if requested */
1866                 if (strlen(env->opts->log_file) > 0) {
1867                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1868                                 lpp_set_log(lpp, stdout);
1869                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1870                                 lpp_set_log(lpp, stderr);
1871                         else {
1872                                 logfile = fopen(env->opts->log_file, "w");
1873                                 if (! logfile)
1874                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1875                                 else
1876                                         lpp_set_log(lpp, logfile);
1877                         }
1878                 }
1879
1880                 /* solve the ILP */
1881                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1882
1883                 if (logfile)
1884                         fclose(logfile);
1885
1886                 /* check for valid solution */
1887                 if (! lpp_is_sol_valid(lpp)) {
1888                         char buf[1024];
1889                         FILE *f;
1890
1891                         DEBUG_ONLY(
1892                                 if (firm_dbg_get_mask(env->dbg) >= 2) {
1893                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1894                                         f = fopen(buf, "w");
1895                                         lpp_dump_plain(lpp, f);
1896                                         fclose(f);
1897                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1898                                         lpp_dump(lpp, buf);
1899                                         dump_ir_block_graph(env->irg, "-infeasible");
1900                                 }
1901                         )
1902
1903                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1904                         need_heur = 1;
1905                 }
1906
1907                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1908                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1909                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1910                 DBG((env->dbg, LEVEL_1, "\tmatrix: %u elements, density %.2f%%, size %.2fMB\n", lpp->n_elems, lpp->density, (double)lpp->matrix_mem / 1024.0 / 1024.0));
1911                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1912                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1913                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1914                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1915
1916                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1917         }
1918
1919         /* apply solution */
1920 #ifdef FIRM_STATISTICS
1921         if (be_stat_ev_is_active()) {
1922                 be_stat_ev("nodes", ba->block_last_idx);
1923         }
1924 #endif /* FIRM_STATISTICS */
1925         if (need_heur) {
1926 #ifdef FIRM_STATISTICS
1927                 if (be_stat_ev_is_active()) {
1928                         be_stat_ev("time", -1);
1929                 }
1930 #endif /* FIRM_STATISTICS */
1931                 list_sched_single_block(env->birg, block, env->be_opts);
1932         }
1933         else {
1934 #ifdef FIRM_STATISTICS
1935                 if (be_stat_ev_is_active()) {
1936                         if (lpp)
1937                                 be_stat_ev_dbl("time", lpp->sol_time);
1938                         else
1939                                 be_stat_ev("time", 0);
1940                 }
1941 #endif /* FIRM_STATISTICS */
1942                 apply_solution(env, lpp, block);
1943         }
1944
1945         if (lpp)
1946                 free_lpp(lpp);
1947
1948         /* notify backend */
1949         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1950 }
1951
1952 /**
1953  * Perform ILP scheduling on the given irg.
1954  */
1955 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
1956         be_ilpsched_env_t          env;
1957         const char                 *name = "be ilp scheduling";
1958         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
1959         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
1960
1961         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
1962
1963 #ifdef FIRM_STATISTICS
1964         if (be_stat_ev_is_active()) {
1965                 be_stat_tags[STAT_TAG_CLS] = "ilpsched";
1966                 be_stat_ev_push(be_stat_tags, STAT_TAG_LAST, be_stat_file);
1967         }
1968 #endif /* FIRM_STATISTICS */
1969
1970 //      firm_dbg_set_mask(env.dbg, 1);
1971
1972         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
1973         env.sel        = sel;
1974         env.irg        = birg->irg;
1975         env.height     = heights_new(birg->irg);
1976         env.main_env   = birg->main_env;
1977         env.arch_env   = birg->main_env->arch_env;
1978         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
1979         env.opts       = &ilp_opts;
1980         env.birg       = birg;
1981         env.be_opts    = be_opts;
1982         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
1983
1984         /* assign a unique per block number to all interesting nodes */
1985         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
1986
1987         /*
1988                 The block indices are completely build after the walk,
1989                 now we can allocate the bitsets (size depends on block indices)
1990                 for all nodes.
1991         */
1992         phase_reinit_irn_data(&env.ph);
1993
1994         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
1995         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
1996
1997         /* Calculate ALAP of all irns */
1998         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
1999
2000         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2001         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2002
2003         /* perform ILP scheduling */
2004         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2005
2006         DEBUG_ONLY(
2007                 if (firm_dbg_get_mask(env.dbg)) {
2008                         phase_stat_t stat;
2009                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2010
2011                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2012                 }
2013         );
2014
2015         /* free data allocated dynamically */
2016         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2017
2018         /* free all allocated object */
2019         phase_free(&env.ph);
2020         heights_free(env.height);
2021
2022         /* notify backend */
2023         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2024
2025 #ifdef FIRM_STATISTICS
2026         if (be_stat_ev_is_active()) {
2027                 be_stat_ev_pop();
2028         }
2029 #endif /* FIRM_STATISTICS */
2030 }
2031
2032 /**
2033  * Register ILP scheduler options.
2034  */
2035 void be_init_ilpsched(void)
2036 {
2037         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2038         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2039
2040         lc_opt_add_table(sched_grp, ilpsched_option_table);
2041 }
2042
2043 #else /* WITH_ILP */
2044
2045 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2046 {}
2047
2048 #endif /* WITH_ILP */