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

   1 /**
   2  * Scheduling algorithms.
   3  * An ILP scheduler based on
   4  * "ILP-based Instruction Scheduling for IA-64"
   5  * by Daniel Kaestner and Sebastian Winkel
   6  *
   7  * @date   22.10.2005
   8  * @author Christian Wuerdig
   9  * @cvs-id $Id$
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #ifndef _WIN32
  20 #include <strings.h>
  21 #endif /* _WIN32 */
  22
  23 #include "irnode_t.h"
  24 #include "irgwalk.h"
  25 #include "irbitset.h"
  26 #include "irphase_t.h"
  27 #include "height.h"
  28 #include "iredges.h"
  29 #include "pdeq.h"
  30 #include "debug.h"
  31 #include "irtools.h"
  32 #include "irdump.h"
  33 #include "plist.h"
  34
  35 #include <lpp/lpp.h>
  36 #include <lpp/lpp_net.h>
  37
  38 #include <libcore/lc_opts.h>
  39 #include <libcore/lc_opts_enum.h>
  40 #include <libcore/lc_timing.h>
  41
  42 #include "be.h"
  43 #include "benode_t.h"
  44 #include "besched_t.h"
  45 #include "beilpsched.h"
  46 #include "beutil.h"
  47 #include "bestat.h"
  48
  49 typedef struct _ilpsched_options_t {
  50         unsigned regpress;
  51         unsigned time_limit;
  52         char     log_file[1024];
  53 } ilpsched_options_t;
  54
  55 typedef struct _unit_type_info_t {
  56         int                            n_units;
  57         const be_execution_unit_type_t *tp;
  58 } unit_type_info_t;
  59
  60 /**
  61  * holding the ILP variables of the different types
  62  */
  63 typedef struct _ilp_var_types_t {
  64         int *x;   /* x_{nt}^k variables */
  65         int *a;   /* a_{nt}^k variables */
  66         int *y;   /* y_{nt}^k variables */
  67 } ilp_var_types_t;
  68
  69 /**
  70  * Holds alive variables for a node live-in to a block.
  71  */
  72 typedef struct _ilp_livein_node_t {
  73         ir_node  *irn;
  74         unsigned max_alive_steps;
  75         int      *a;
  76 } ilp_livein_node_t;
  77
  78 /* attributes for a node */
  79 typedef struct _ilpsched_node_attr_t {
  80         unsigned asap;                     /**< The ASAP scheduling control step */
  81         unsigned alap;                     /**< The ALAP scheduling control step */
  82         unsigned latency;                  /**< Latency of this node (needed for sorting) */
  83         unsigned sched_point;              /**< the step in which the node is finally scheduled */
  84         unsigned visit_idx;                /**< Index of the node having visited this node last */
  85         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
  86         unsigned n_consumer;               /**< Number of consumers */
  87         ir_node  **block_consumer;         /**< List of consumer being in the same block */
  88         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
  89         unsigned block_idx     : 30;       /**< A unique per block index */
  90         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
  91         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
  92         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
  93                                                                                         for ASAP, successors for ALAP */
  94         unsigned         n_unit_types;     /**< number of allowed execution unit types */
  95         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
  96         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
  97 } ilpsched_node_attr_t;
  98
  99 /* attributes for a block */
 100 typedef struct _ilpsched_block_attr_t {
 101         unsigned block_last_idx;        /**< The highest node index in block so far */
 102         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 103         unsigned max_steps;             /**< Upper bound for block execution */
 104         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 105         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 106         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 107 } ilpsched_block_attr_t;
 108
 109 typedef union _ilpsched_attr_ {
 110         ilpsched_node_attr_t  node_attr;
 111         ilpsched_block_attr_t block_attr;
 112 } ilpsched_attr_t;
 113
 114 /* A irn for the phase and it's attributes (either node or block) */
 115 typedef struct {
 116         ir_node         *irn;
 117         ilpsched_attr_t attr;
 118 } be_ilpsched_irn_t;
 119
 120 /* The ILP scheduling environment */
 121 typedef struct {
 122         phase_t              ph;            /**< The phase */
 123         ir_graph             *irg;          /**< The current irg */
 124         heights_t            *height;       /**< The heights object of the irg */
 125         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 126         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 127         const arch_env_t     *arch_env;
 128         const arch_isa_t     *isa;          /**< The ISA */
 129         const be_main_env_t  *main_env;
 130         const be_machine_t   *cpu;          /**< the current abstract machine */
 131         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 132         const be_irg_t       *birg;         /**< The birg object */
 133         be_options_t         *be_opts;      /**< backend options */
 134         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 135         DEBUG_ONLY(firm_dbg_module_t *dbg);
 136 } be_ilpsched_env_t;
 137
 138 /* convenience macros to handle phase irn data */
 139 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 140 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 141 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 142 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 143
 144 /* check if node is considered for ILP scheduling */
 145 #define consider_for_sched(isa, irn) \
 146         (! (is_Block(irn)            ||  \
 147                 is_normal_Proj(isa, irn) ||  \
 148                 is_Phi(irn)              ||  \
 149                 is_NoMem(irn)            ||  \
 150                 is_Unknown(irn)          ||  \
 151                 is_End(irn)                  \
 152                 ))
 153
 154 /* gives the valid scheduling time step interval for a node */
 155 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 156
 157 /* gives the valid interval where a node can die */
 158 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 159
 160 /* gives the corresponding ILP variable for given node, unit and time step */
 161 #define ILPVAR_IDX(na, unit, control_step) \
 162         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 163
 164 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 165 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 166         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 167
 168 /* check if a double value is within an epsilon environment of 0 */
 169 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 170
 171 #define ilp_timer_push(t)         lc_timer_push((t))
 172 #define ilp_timer_pop()           lc_timer_pop()
 173 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
 174
 175 /* option variable */
 176 static ilpsched_options_t ilp_opts = {
 177         1,     /* default is with register pressure constraints */
 178         300,   /* 300 sec per block time limit */
 179         ""     /* no log file */
 180 };
 181
 182 /* ILP options */
 183 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 184         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 185         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 186         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 187         { NULL }
 188 };
 189
 190 /*
 191         We need this global variable as we compare nodes dependent on heights,
 192         but we cannot pass any information to the qsort compare function.
 193 */
 194 static heights_t *glob_heights;
 195
 196 /**
 197  * Check if irn is a Proj, which has no execution units assigned.
 198  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 199  */
 200 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
 201         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
 202 }
 203
 204 /**
 205  * Skips normal Projs.
 206  * @return predecessor if irn is a normal Proj, otherwise irn.
 207  */
 208 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
 209         if (is_normal_Proj(isa, irn))
 210                 return get_Proj_pred(irn);
 211         return irn;
 212 }
 213
 214 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
 215         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 216         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 217                 lat = 1;
 218         return lat;
 219 }
 220
 221 static int cmp_live_in_nodes(const void *a, const void *b) {
 222         const ilp_livein_node_t *n1 = a;
 223         const ilp_livein_node_t *n2 = b;
 224
 225         return n1->irn != n2->irn;
 226 }
 227
 228 /**
 229  * Compare scheduling time steps of two be_ilpsched_irn's.
 230  */
 231 static int cmp_ilpsched_irn(const void *a, const void *b) {
 232         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 233         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 234         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 235         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 236
 237         if (n1_a->sched_point == n2_a->sched_point) {
 238                 ir_node *irn_a = n1->irn;
 239                 ir_node *irn_b = n2->irn;
 240
 241                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 242                         return 1;
 243                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 244                         return -1;
 245
 246                 /*
 247                         Ok, timestep is equal and the nodes are parallel,
 248                         so check latency and schedule high latency first.
 249                 */
 250                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 251         }
 252         else
 253                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 254 }
 255
 256 /**
 257  * In case there is no phase information for irn, initialize it.
 258  */
 259 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
 260         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 261
 262         if (res == old) {
 263                 /* if we have already some data: check for reinitialization */
 264
 265                 if (! is_Block(irn)) {
 266                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 267
 268                         if (! na->transitive_block_nodes) {
 269                                 ir_node               *block      = get_nodes_block(irn);
 270                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 271                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 272
 273                                 /* we are called after the block indices have been build: create bitset */
 274                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 275                         }
 276                         else {
 277                                 /* we are called from reinit block data: clear the bitset */
 278                                 bitset_clear_all(na->transitive_block_nodes);
 279                                 na->visit_idx    = 0;
 280                                 na->alap_changed = 1;
 281                         }
 282                 }
 283                 return old;
 284         }
 285
 286         res->irn = irn;
 287
 288         /* set ilpsched irn attributes (either block or irn) */
 289         if (is_Block(irn)) {
 290                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 291
 292                 ba->n_interesting_nodes = 0;
 293                 ba->block_last_idx      = 0;
 294                 ba->root_nodes          = plist_new();
 295                 ba->head_ilp_nodes      = NULL;
 296                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 297                 ba->max_steps           = 0;
 298         }
 299         else {
 300                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 301                 memset(na, 0, sizeof(*na));
 302         }
 303
 304         return res;
 305 }
 306
 307 /**
 308  * Assign a per block unique number to each node.
 309  */
 310 static void build_block_idx(ir_node *irn, void *walk_env) {
 311         be_ilpsched_env_t     *env = walk_env;
 312         be_ilpsched_irn_t     *node, *block_node;
 313         ilpsched_node_attr_t  *na;
 314         ilpsched_block_attr_t *ba;
 315
 316         set_irn_link(irn, NULL);
 317         if (! consider_for_sched(env->arch_env->isa, irn))
 318                 return;
 319
 320         node       = get_ilpsched_irn(env, irn);
 321         na         = get_ilpsched_node_attr(node);
 322         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 323         ba         = get_ilpsched_block_attr(block_node);
 324
 325         na->block_idx = ba->block_last_idx++;
 326 }
 327
 328 /********************************************************
 329  *                              __        _
 330  *                             / /       | |
 331  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 332  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 333  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 334  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 335  *                 | |                           | |
 336  *                 |_|                           |_|
 337  ********************************************************/
 338
 339 /**
 340  * Add all nodes having no user in current block to last_nodes list.
 341  */
 342 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
 343         ir_node               *block;
 344         const ir_edge_t       *edge;
 345         be_ilpsched_irn_t     *block_node, *node;
 346         ilpsched_block_attr_t *ba;
 347         ilpsched_node_attr_t  *na;
 348         int                   i, j;
 349         be_ilpsched_env_t     *env           = walk_env;
 350         int                   has_block_user = 0;
 351         unsigned              n_consumer     = 0;
 352         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 353         ir_node               **consumer;
 354         int                   idx;
 355
 356         if (! consider_for_sched(env->arch_env->isa, irn))
 357                 return;
 358
 359         block    = get_nodes_block(irn);
 360     idx      = get_irn_idx(irn);
 361         consumer = NEW_ARR_F(ir_node *, 0);
 362
 363         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 364
 365         /* check data and dependency out edges */
 366         for (i = 0; i < 2 && ! has_block_user; ++i) {
 367                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 368                         ir_node *user = get_edge_src_irn(edge);
 369
 370                         if (is_normal_Proj(env->arch_env->isa, user)) {
 371                                 const ir_edge_t *user_edge;
 372
 373                                 if (get_irn_mode(user) == mode_X)
 374                                         continue;
 375
 376                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 377                                 for (j = 0; j < 2; ++j) {
 378                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 379                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 380
 381                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 382                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 383                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 384
 385                                                         /* skip already visited nodes */
 386                                                         if (ua->consumer_idx == idx)
 387                                                                 continue;
 388
 389                                                         /* check if node has user in this block and collect the user if it's a data user */
 390                                                         if (get_nodes_block(real_user) == block) {
 391                                                                 if (i == 0 && j == 0)
 392                                                                         ARR_APP1(ir_node *, consumer, real_user);
 393                                                                 has_block_user = 1;
 394                                                         }
 395
 396                                                         /* only count data consumer */
 397                                                         if (i == 0)
 398                                                                 n_consumer++;
 399
 400                                                         /* mark user as visited by this node */
 401                                                         ua->consumer_idx = idx;
 402                                                 }
 403                                         }
 404                                 }
 405                         }
 406                         else if (is_Block(user)) {
 407                                 continue;
 408                         }
 409                         else if (! is_Phi(user)) {
 410                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 411                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 412
 413                                 /* skip already visited nodes */
 414                                 if (ua->consumer_idx == idx)
 415                                         continue;
 416
 417                                 /* check if node has user in this block and collect the user if it's a data user */
 418                                 if (get_nodes_block(user) == block) {
 419                                         if (i == 0)
 420                                                 ARR_APP1(ir_node *, consumer, user);
 421                                         has_block_user = 1;
 422                                 }
 423
 424                                 /* only count data consumer */
 425                                 if (i == 0)
 426                                         n_consumer++;
 427
 428                                 /* mark user visited by this node */
 429                                 ua->consumer_idx = idx;
 430                         }
 431                         else if (get_nodes_block(user) != block) {
 432                                 n_consumer++;
 433                         }
 434                 }
 435         }
 436
 437         block_node = get_ilpsched_irn(env, block);
 438         ba         = get_ilpsched_block_attr(block_node);
 439
 440         ba->n_interesting_nodes++;
 441
 442         /* current irn has no user inside this block, add to queue */
 443         if (! has_block_user) {
 444                 DB((env->dbg, LEVEL_3, "root node\n"));
 445                 plist_insert_back(ba->root_nodes, irn);
 446         }
 447         else {
 448                 DB((env->dbg, LEVEL_3, "normal node\n"));
 449         }
 450
 451         /* record number of all consumer and the consumer within the same block */
 452         node = get_ilpsched_irn(env, irn);
 453         na   = get_ilpsched_node_attr(node);
 454         na->n_consumer     = n_consumer;
 455         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 456         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 457         DEL_ARR_F(consumer);
 458 }
 459
 460 /**
 461  * Calculate the ASAP scheduling step for current irn.
 462  */
 463 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
 464         be_ilpsched_env_t     *env = walk_env;
 465         int                   i;
 466         ir_node               *block;
 467         be_ilpsched_irn_t     *node, *block_node;
 468         ilpsched_node_attr_t  *na;
 469         ilpsched_block_attr_t *ba;
 470
 471         /* These nodes are handled separate */
 472         if (! consider_for_sched(env->arch_env->isa, irn))
 473                 return;
 474
 475         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 476
 477         block    = get_nodes_block(irn);
 478         node     = get_ilpsched_irn(env, irn);
 479         na       = get_ilpsched_node_attr(node);
 480         na->asap = 1;
 481
 482         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 483                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
 484
 485                 /* check for greatest distance to top */
 486                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 487                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 488                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 489                         unsigned             lat;
 490
 491                         lat         = fixed_latency(env->sel, pred, env->block_env);
 492                         na->latency = lat;
 493                         na->asap    = MAX(na->asap, pna->asap + lat);
 494                 }
 495         }
 496
 497         /* add node to ILP node list and update max_steps */
 498         block_node = get_ilpsched_irn(env, block);
 499         ba         = get_ilpsched_block_attr(block_node);
 500
 501         set_irn_link(irn, ba->head_ilp_nodes);
 502         ba->head_ilp_nodes = irn;
 503         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 504
 505         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 506 }
 507
 508 /**
 509  * Calculate the ALAP scheduling step of all irns in current block.
 510  * Depends on max_steps being calculated.
 511  */
 512 static void calculate_block_alap(ir_node *block, void *walk_env) {
 513         be_ilpsched_env_t     *env        = walk_env;
 514         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 515         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 516         waitq                 *cur_queue  = new_waitq();
 517         plist_element_t       *el;
 518
 519         assert(is_Block(block));
 520
 521         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 522                 block, ba->n_interesting_nodes, ba->max_steps));
 523
 524         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 525         //phase_reinit_block_irn_data(&env->ph, block);
 526
 527         /* init start queue */
 528         foreach_plist(ba->root_nodes, el) {
 529                 waitq_put(cur_queue, plist_element_get_value(el));
 530         }
 531
 532         /* repeat until all nodes are processed */
 533         while (! waitq_empty(cur_queue)) {
 534                 waitq *next_queue = new_waitq();
 535
 536                 /* process all nodes in current step */
 537                 while (! waitq_empty(cur_queue)) {
 538                         ir_node              *cur_irn = waitq_get(cur_queue);
 539                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 540                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 541                         int                  i;
 542
 543                         /* cur_node has no alap set -> it's a root node, set to max alap */
 544                         if (na->alap == 0) {
 545                                 na->alap = ba->max_steps;
 546                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 547                                         cur_irn, na->alap));
 548                         }
 549                         else {
 550                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 551                                         cur_irn, na->alap));
 552                         }
 553
 554                         /* set the alap's of all predecessors */
 555                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 556                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
 557
 558                                 /* check for greatest distance to bottom */
 559                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 560                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 561                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 562                                         unsigned             lat;
 563
 564                                         /* mark the predecessor as visited by current irn */
 565                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 566                                                 continue;
 567                                         pna->visit_idx = get_irn_idx(cur_irn);
 568
 569                                         lat = fixed_latency(env->sel, pred, env->block_env);
 570
 571                                         /* set ALAP of current pred */
 572                                         if (pna->alap == 0) {
 573                                                 /* current ALAP is 0: node has not yet been visited */
 574                                                 pna->alap_changed = 1;
 575                                                 pna->alap         = na->alap - lat;
 576                                         }
 577                                         else if (pna->alap > na->alap - lat) {
 578                                                 /* we found a longer path to root node: change ALAP */
 579                                                 pna->alap         = na->alap - lat;
 580                                                 pna->alap_changed = 1;
 581                                         }
 582                                         else {
 583                                                 /* current ALAP is best found so far: keep it */
 584                                                 pna->alap_changed = 0;
 585                                         }
 586
 587                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 588
 589                                         /* enqueue node for next iteration */
 590                                         if (get_irn_ins_or_deps(pred) > 0)
 591                                                 waitq_put(next_queue, pred);
 592                                 }
 593                         }
 594                 }
 595
 596                 /* prepare for next iteration */
 597                 del_waitq(cur_queue);
 598                 cur_queue = next_queue;
 599         }
 600 }
 601
 602 /**
 603  * Free list of root nodes and the set of live-in nodes.
 604  */
 605 static void clear_unwanted_data(ir_node *block, void *walk_env) {
 606         be_ilpsched_env_t     *env        = walk_env;
 607         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 608         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 609
 610         plist_free(ba->root_nodes);
 611         ba->root_nodes = NULL;
 612         del_pset(ba->livein_nodes);
 613         ba->livein_nodes = NULL;
 614 }
 615
 616 /**
 617  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 618  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 619  */
 620 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
 621         be_ilpsched_env_t    *env  = walk_env;
 622         ir_node              *pred = irn;
 623         be_ilpsched_irn_t    *node, *pred_node;
 624         ilpsched_node_attr_t *na, *pna;
 625
 626         if (! consider_for_sched(env->arch_env->isa, irn))
 627                 return;
 628
 629         if (! is_Proj(irn) && ! be_is_Keep(irn))
 630                 return;
 631
 632         /* go to the ancestor */
 633         if (be_is_Keep(irn))
 634                 pred = get_irn_n(irn, 0);
 635         pred = skip_Proj(pred);
 636
 637         node      = get_ilpsched_irn(env, irn);
 638         pred_node = get_ilpsched_irn(env, pred);
 639         na        = get_ilpsched_node_attr(node);
 640         pna       = get_ilpsched_node_attr(pred_node);
 641
 642         na->asap = pna->asap;
 643         na->alap = pna->alap;
 644
 645         /* record all Projs and Keeps for this node */
 646         if (! pna->projkeeps)
 647                 pna->projkeeps = new_waitq();
 648         waitq_put(pna->projkeeps, irn);
 649
 650         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 651 }
 652
 653 /*******************************************
 654  *           _              _       _
 655  *          | |            | |     | |
 656  *  ___  ___| |__   ___  __| |_   _| | ___
 657  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 658  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 659  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 660  *
 661  *******************************************/
 662
 663 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
 664         const ir_edge_t *edge;
 665
 666         foreach_out_edge(irn, edge) {
 667                 ir_node *user = get_edge_src_irn(edge);
 668
 669                 if (be_is_Keep(user)) {
 670                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 671                         waitq_put(keeps, user);
 672                 }
 673         }
 674 }
 675
 676 /**
 677  * Inserts @p irn before @p before into schedule and notifies backend.
 678  */
 679 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
 680         ir_node *before, ir_node *irn, unsigned cycle)
 681 {
 682         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 683         sched_add_before(before, irn);
 684 }
 685
 686 /**
 687  * Adds a node, it's Projs (in case of mode_T nodes) and
 688  * it's Keeps to schedule.
 689  */
 690 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
 691         const ir_edge_t *edge;
 692         waitq           *keeps = new_waitq();
 693
 694         /* mode_M nodes are not scheduled */
 695         if (get_irn_mode(irn) == mode_M)
 696                 return;
 697
 698         if (! sched_is_scheduled(irn))
 699                 notified_sched_add_before(env, block, irn, cycle);
 700
 701         /* add Projs */
 702         if (get_irn_mode(irn) == mode_T) {
 703                 foreach_out_edge(irn, edge) {
 704                         ir_node *user = get_edge_src_irn(edge);
 705
 706                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 707                                 get_irn_n_edges(user) > 0)
 708                         {
 709                                 notified_sched_add_before(env, block, user, cycle);
 710                         }
 711
 712                         check_for_keeps(keeps, block, user);
 713                 }
 714         }
 715         else {
 716                 check_for_keeps(keeps, block, irn);
 717         }
 718
 719         /* add Keeps */
 720         while (! waitq_empty(keeps)) {
 721                 ir_node *keep = waitq_get(keeps);
 722                 if (! sched_is_scheduled(keep))
 723                         notified_sched_add_before(env, block, keep, cycle);
 724         }
 725
 726         del_waitq(keeps);
 727 }
 728
 729 /**
 730  * Schedule all nodes in the given block, according to the ILP solution.
 731  */
 732 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
 733         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 734         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 735         sched_info_t          *info       = get_irn_sched_info(block);
 736         be_ilpsched_irn_t     **sched_nodes;
 737         unsigned              i, l;
 738         ir_node               *cfop, *irn;
 739         const ir_edge_t       *edge;
 740
 741         /* init block schedule list */
 742         INIT_LIST_HEAD(&info->list);
 743         info->scheduled = 1;
 744
 745         /* collect nodes and their scheduling time step */
 746         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 747         if (ba->n_interesting_nodes == 0) {
 748                 /* ignore */
 749         }
 750         else if (ba->n_interesting_nodes == 1) {
 751                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 752
 753                 /* add the single node */
 754                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 755         }
 756         else {
 757                 /* check all nodes for their positive solution */
 758                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 759                         be_ilpsched_irn_t    *node;
 760                         ilpsched_node_attr_t *na;
 761                         int                  tp_idx, found;
 762                         unsigned             cur_var, t;
 763
 764                         node    = get_ilpsched_irn(env, irn);
 765                         na      = get_ilpsched_node_attr(node);
 766                         cur_var = 0;
 767                         found   = 0;
 768
 769                         if (! na->is_dummy_node) {
 770                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 771                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 772                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 773
 774                                                 if (! LPP_VALUE_IS_0(cost)) {
 775                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
 776                                                         found = 1;
 777                                                 }
 778                                         }
 779                                 }
 780                         }
 781
 782                         found = 0;
 783                         /* go over all variables of a node until the non-zero one is found */
 784                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 785                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 786                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 787
 788                                         /* check, if variable is set to one (it's not zero then :) */
 789                                         if (! LPP_VALUE_IS_0(val)) {
 790                                                 na->sched_point = t;
 791                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 792                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 793                                                         irn, t, na->type_info[tp_idx].tp->name));
 794                                                 found = 1;
 795                                         }
 796                                 }
 797                         }
 798                 }
 799
 800                 glob_heights = env->height;
 801                 /* sort nodes ascending by scheduling time step */
 802                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 803         }
 804
 805         /* make all Phis ready and remember the single cf op */
 806         cfop = NULL;
 807         foreach_out_edge(block, edge) {
 808                 irn = get_edge_src_irn(edge);
 809
 810                 switch (get_irn_opcode(irn)) {
 811                         case iro_Phi:
 812                                 add_to_sched(env, block, irn, 0);
 813                                 break;
 814                         case iro_Start:
 815                         case iro_End:
 816                         case iro_Proj:
 817                         case iro_Bad:
 818                         case iro_Unknown:
 819                                 break;
 820                         default:
 821                                 if (is_cfop(irn)) {
 822                                         assert(cfop == NULL && "Highlander - there can be only one");
 823                                         cfop = irn;
 824                                 }
 825                         break;
 826                 }
 827         }
 828
 829         /* add all nodes from list */
 830         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 831                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 832                 if (sched_nodes[i]->irn != cfop)
 833                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 834         }
 835
 836         /* schedule control flow node if not already done */
 837         if (cfop && ! sched_is_scheduled(cfop))
 838                 add_to_sched(env, block, cfop, 0);
 839
 840         DEL_ARR_F(sched_nodes);
 841 }
 842
 843 /***************************************************************
 844  *   _____ _      _____     _____           _   _
 845  *  |_   _| |    |  __ \   / ____|         | | (_)
 846  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 847  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 848  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 849  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 850  *
 851  ***************************************************************/
 852
 853 /**
 854  * Check if node can be executed on given unit type.
 855  */
 856 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
 857         int                  i;
 858         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 859
 860         for (i = na->n_unit_types - 1; i >= 0; --i) {
 861                 if (na->type_info[i].tp == tp)
 862                         return i;
 863         }
 864
 865         return -1;
 866 }
 867
 868 /************************************************
 869  *                   _       _     _
 870  *                  (_)     | |   | |
 871  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 872  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 873  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 874  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 875  *
 876  ************************************************/
 877
 878 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
 879         const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
 880         unsigned                  n_unit_types = 0;
 881         be_ilpsched_irn_t         *node;
 882         ilpsched_node_attr_t      *na;
 883         unsigned                  unit_idx, tp_idx;
 884
 885         /* count number of available unit types for this node */
 886         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 887                 /* just count */ ;
 888
 889         node = get_ilpsched_irn(env, irn);
 890         na   = get_ilpsched_node_attr(node);
 891
 892         if (! na->type_info) {
 893                 na->n_unit_types = n_unit_types;
 894                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 895
 896                 /* fill the type info array */
 897                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 898                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 899                                 /* beware: we also count number of available units here */
 900                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 901                                         na->is_dummy_node = 1;
 902                         }
 903
 904                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 905                         na->type_info[tp_idx].n_units = unit_idx;
 906                 }
 907         }
 908
 909         return n_unit_types;
 910 }
 911
 912 /**
 913  * Returns the largest alap time of a user of @p irn.
 914  * The user must be in block @p block.
 915  */
 916 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
 917         const ir_edge_t *edge;
 918         unsigned        max_alap = 0;
 919
 920         foreach_out_edge(irn, edge) {
 921                 ir_node *user = get_edge_src_irn(edge);
 922
 923                 if (get_nodes_block(user) == block) {
 924                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 925                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 926
 927                         max_alap = MAX(max_alap, na->alap);
 928                 }
 929         }
 930
 931         assert(max_alap > 0);
 932         return max_alap;
 933 }
 934
 935 /**
 936  * Create the following variables:
 937  * - x_{nt}^k    binary     weigthed with: t
 938  *      node n is scheduled at time step t to unit type k
 939  * ==>> These variables represent the schedule
 940  *
 941  * - a_{nt}^k    binary     weighted with num_nodes
 942  *      node n is alive at time step t on unit type k
 943  *
 944  * - y_{nt}^k    continuous  weighted with: num_nodes^2
 945  *      register pressure over limit for unit type k
 946  * ==>> These variables represent the register pressure
 947  *
 948  */
 949 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
 950         char                  buf[1024];
 951         ir_node               *irn;
 952         unsigned              num_block_var, num_nodes;
 953         ilp_livein_node_t     *livein;
 954         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 955         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 956         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
 957
 958         ilp_timer_push(t_var);
 959         num_block_var = num_nodes = 0;
 960         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 961                 be_ilpsched_irn_t    *node;
 962                 ilpsched_node_attr_t *na;
 963                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
 964                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
 965                 int                  i;
 966
 967                 node         = get_ilpsched_irn(env, irn);
 968                 na           = get_ilpsched_node_attr(node);
 969                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
 970
 971                 /* allocate space for ilp variables */
 972                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 973                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
 974
 975                 /* we need these variables only for "real" nodes */
 976                 if (! na->is_dummy_node) {
 977                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
 978                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 979
 980                         num_ad         = ba->max_steps - na->asap + 1;
 981                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
 982                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
 983                 }
 984
 985                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
 986                         irn, na->asap, na->alap, na->n_unit_types));
 987
 988                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
 989                 /* create variables */
 990                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 991                         unsigned t;
 992
 993                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
 994                                 /* x_{nt}^k variables */
 995                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
 996                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
 997                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
 998                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 999                                 /* variable counter */
1000                                 n_var++;
1001                                 num_block_var++;
1002
1003                                 if (! na->is_dummy_node) {
1004                                         /* y_{nt}^k variables */
1005                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1006                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1007                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1008                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1009
1010                                         /* variable counter */
1011                                         n_var++;
1012                                         num_block_var++;
1013                                 }
1014                         }
1015
1016                         /* a node can die at any step t: asap(n) <= t <= U */
1017                         if (! na->is_dummy_node) {
1018                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1019
1020                                         /* a_{nt}^k variables */
1021                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1022                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1023                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1024                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1025
1026                                         /* variable counter */
1027                                         n_var++;
1028                                         num_block_var++;
1029                                 }
1030                         }
1031
1032                         /* collect live-in nodes */
1033                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1034                                 ir_node *pred = get_irn_n(irn, i);
1035
1036                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1037                                         be_ilpsched_set_type_info(env, pred, var_obst);
1038                                         if (! na->is_dummy_node) {
1039                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1040                                                 entry->irn = pred;
1041                                                 entry->a   = NULL;
1042                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1043                                         }
1044                                 }
1045                         }
1046                 }
1047
1048                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1049                 num_nodes++;
1050         }
1051
1052         /* create alive variables a_{nt}^k for live-ins */
1053         foreach_pset(ba->livein_nodes, livein) {
1054                 be_ilpsched_irn_t    *node;
1055                 ilpsched_node_attr_t *na;
1056                 unsigned             tp_idx, var_idx;
1057                 ir_node              *irn;
1058
1059                 irn  = livein->irn;
1060                 node = get_ilpsched_irn(env, irn);
1061                 na   = get_ilpsched_node_attr(node);
1062
1063                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1064
1065                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1066                 var_idx   = 0;
1067
1068                 /* create variables */
1069                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1070                         unsigned t;
1071
1072                         for (t = 0; t < livein->max_alive_steps; ++t) {
1073                                 /* a_{nt}^k variables */
1074                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1075                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1076                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1077                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1078                                 num_block_var++;
1079                         }
1080                 }
1081         }
1082
1083         ilp_timer_pop();
1084         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1085                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1086 }
1087
1088 /*******************************************************
1089  *                      _             _       _
1090  *                     | |           (_)     | |
1091  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1092  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1093  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1094  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1095  *
1096  *******************************************************/
1097
1098 /**
1099  * Create following ILP constraints:
1100  * - the assignment constraints:
1101  *     assure each node is executed once by exactly one (allowed) execution unit
1102  * - the dead node assignment constraints:
1103  *     assure a node can only die at most once
1104  * - the precedence constraints:
1105  *     assure that no data dependencies are violated
1106  */
1107 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1108         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1109         char                  buf[1024];
1110         ir_node               *irn;
1111         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1112         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1113         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1114         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1115
1116         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1117         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1118                 int                  cst, tp_idx, i;
1119                 unsigned             cur_var;
1120                 be_ilpsched_irn_t    *node;
1121                 ilpsched_node_attr_t *na;
1122
1123                 node    = get_ilpsched_irn(env, irn);
1124                 na      = get_ilpsched_node_attr(node);
1125                 cur_var = 0;
1126
1127                 /* the assignment constraint */
1128                 ilp_timer_push(t_cst_assign);
1129                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1130                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1131                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1132                 num_cst_assign++;
1133
1134                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1135                 ilp_timer_pop();
1136
1137                 /* We have separate constraints for Projs and Keeps */
1138                 // ILP becomes infeasible ?!?
1139 //              if (is_Proj(irn) || be_is_Keep(irn))
1140 //                      continue;
1141
1142                 /* the precedence constraints */
1143                 ilp_timer_push(t_cst_prec);
1144                 bs_block_irns = bitset_clear_all(bs_block_irns);
1145                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1146                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1147                         unsigned             t_low, t_high, t;
1148                         be_ilpsched_irn_t    *pred_node;
1149                         ilpsched_node_attr_t *pna;
1150                         unsigned             delay;
1151
1152                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1153                                 continue;
1154
1155                         pred_node = get_ilpsched_irn(env, pred);
1156                         pna       = get_ilpsched_node_attr(pred_node);
1157
1158                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1159
1160                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1161                                 bitset_set(bs_block_irns, pna->block_idx);
1162                         else
1163                                 continue;
1164
1165                         /* irn = n, pred = m */
1166                         delay  = fixed_latency(env->sel, pred, env->block_env);
1167                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1168                         t_high = MIN(na->alap, pna->alap + delay - 1);
1169                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1170                                 unsigned tn, tm;
1171                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1172
1173                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1174                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1175                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1176                                 num_cst_prec++;
1177
1178                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1179                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1180                                         /* node variables have smaller index than pred variables */
1181                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1182                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1183                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1184                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1185                                                 }
1186                                         }
1187
1188                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1189                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1190                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1191                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1192                                                 }
1193                                         }
1194                                 }
1195                                 else {
1196                                         /* pred variables have smaller index than node variables */
1197                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1198                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1199                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1200                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1201                                                 }
1202                                         }
1203
1204                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1205                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1206                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1207                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1208                                                 }
1209                                         }
1210                                 }
1211
1212                                 if (ARR_LEN(tmp_var_idx) > 0)
1213                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1214
1215                                 DEL_ARR_F(tmp_var_idx);
1216                         }
1217                 }
1218                 ilp_timer_pop();
1219         }
1220         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1221                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1222         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1223                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1224 }
1225
1226 /**
1227  * Create ILP resource constraints:
1228  * - assure that for each time step not more instructions are scheduled
1229  *   to the same unit types as units of this type are available
1230  */
1231 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1232         int                   glob_type_idx;
1233         char                  buf[1024];
1234         unsigned              num_cst_resrc = 0;
1235         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1236         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1237
1238         ilp_timer_push(t_cst_rsrc);
1239         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1240                 unsigned                 t;
1241                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1242
1243                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1244
1245                 /* check each time step */
1246                 for (t = 0; t < ba->max_steps; ++t) {
1247                         ir_node *irn;
1248                         int     cst;
1249                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1250
1251                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1252                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1253                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1254                         num_cst_resrc++;
1255
1256                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1257                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1258                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1259                                 int                  tp_idx;
1260
1261                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1262
1263                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1264                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1265                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1266                                 }
1267                         }
1268
1269                         /* set constraints if we have some */
1270                         if (ARR_LEN(tmp_var_idx) > 0)
1271                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1272
1273                         DEL_ARR_F(tmp_var_idx);
1274                 }
1275         }
1276         ilp_timer_pop();
1277         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1278                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1279 }
1280
1281 /**
1282  * Create ILP bundle constraints:
1283  * - assure, at most bundle_size * bundles_per_cycle instructions
1284  *   can be started at a certain point.
1285  */
1286 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1287         char                  buf[1024];
1288         unsigned              t;
1289         unsigned              num_cst_bundle = 0;
1290         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1291         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1292         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1293
1294         ilp_timer_push(t_cst_bundle);
1295         for (t = 0; t < ba->max_steps; ++t) {
1296                 ir_node *irn;
1297                 int     cst;
1298                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1299
1300                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1301                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1302                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1303                 num_cst_bundle++;
1304
1305                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1306                         be_ilpsched_irn_t    *node;
1307                         ilpsched_node_attr_t *na;
1308                         int                  tp_idx;
1309
1310                         /* Projs and Keeps do not contribute to bundle size */
1311                         if (is_Proj(irn) || be_is_Keep(irn))
1312                                 continue;
1313
1314                         node = get_ilpsched_irn(env, irn);
1315                         na   = get_ilpsched_node_attr(node);
1316
1317                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1318                         if (na->is_dummy_node)
1319                                 continue;
1320
1321                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1322                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1323                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1324                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1325                                 }
1326                         }
1327                 }
1328
1329                 if (ARR_LEN(tmp_var_idx) > 0)
1330                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1331
1332                 DEL_ARR_F(tmp_var_idx);
1333         }
1334         ilp_timer_pop();
1335         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1336                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1337 }
1338
1339 /**
1340  * Create ILP alive nodes constraints:
1341  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1342  */
1343 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1344         char                  buf[1024];
1345         ir_node               *irn;
1346         unsigned              num_cst = 0;
1347         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1348         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1349
1350         ilp_timer_push(t_cst);
1351         /* for each node */
1352         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1353                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1354                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1355                 unsigned             t;
1356
1357                 /* we ignore nodes assigned to dummy unit here */
1358                 if (na->is_dummy_node)
1359                         continue;
1360
1361                 /* check check all time steps: asap(n) <= t <= U */
1362                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1363                         int node_tp_idx;
1364
1365                         /* for all unit types available for this node */
1366                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1367                                 unsigned tn, tn_max, idx;
1368                                 int      cst, i;
1369                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1370                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1371
1372                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1373                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1374                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1375                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1376                                 num_cst++;
1377
1378                                 tn_max = MIN(na->alap - 1, t);
1379                                 /* check if the node has been scheduled so far */
1380                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1381                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1382                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1383                                 }
1384
1385                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1386                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1387                                 DEL_ARR_F(tmp_var_idx_n);
1388
1389                                 /* subtract the number of consumer scheduled so far */
1390                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1391                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1392                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1393                                         int                  tp_idx;
1394                                         unsigned             tm, tm_max;
1395
1396                                         tm_max = MIN(ca->alap - 1, t);
1397                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1398                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1399                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1400                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1401                                                 }
1402                                         }
1403                                 }
1404
1405                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1406                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1407                                 DEL_ARR_F(tmp_var_idx_m);
1408
1409                                 /* -c * a_{nt}^k */
1410                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1411                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1412
1413                         }
1414                 }
1415         }
1416         ilp_timer_pop();
1417         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1418                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1419 }
1420
1421 /**
1422  * Create ILP alive nodes constraints for live-in nodes:
1423  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1424  */
1425 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1426         char                  buf[1024];
1427         ilp_livein_node_t     *livein;
1428         unsigned              num_cst = 0;
1429         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1430         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1431
1432         ilp_timer_push(t_cst);
1433         /* for each node */
1434         foreach_pset(ba->livein_nodes, livein) {
1435                 ir_node              *irn  = livein->irn;
1436                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1437                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1438                 unsigned             t;
1439
1440                 /* check check all time steps: 0 <= t < max_alive_steps */
1441                 for (t = 0; t < livein->max_alive_steps; ++t) {
1442                         int node_tp_idx;
1443
1444                         /* for all unit types available for this node */
1445                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1446                                 const ir_edge_t *edge;
1447                                 unsigned idx;
1448                                 int      cst, num_block_user;
1449                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1450
1451                                 /* check the number of consumer scheduled so far */
1452                                 num_block_user = 0;
1453                                 foreach_out_edge(irn, edge) {
1454                                         ir_node              *user = get_edge_src_irn(edge);
1455                                         be_ilpsched_irn_t    *cons;
1456                                         ilpsched_node_attr_t *ca;
1457                                         int                  tp_idx;
1458                                         unsigned             tm, tm_max;
1459
1460                                         /* check only users within current block */
1461                                         if (get_nodes_block(user) != block_node->irn)
1462                                                 continue;
1463
1464                                         num_block_user++;
1465                                         cons = get_ilpsched_irn(env, user);
1466                                         ca   = get_ilpsched_node_attr(cons);
1467
1468                                         tm_max = MIN(ca->alap - 1, t);
1469                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1470                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1471                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1472                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1473                                                 }
1474                                         }
1475                                 }
1476
1477                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1478                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1479                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1480                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1481                                 num_cst++;
1482
1483                                 /* sum(scheduled users) */
1484                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1485                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1486                                 DEL_ARR_F(tmp_var_idx_m);
1487
1488                                 /* + c * a_{nt}^k */
1489                                 idx = node_tp_idx * livein->max_alive_steps + t;
1490                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1491                         }
1492                 }
1493         }
1494         ilp_timer_pop();
1495         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1496                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1497 }
1498
1499 /**
1500  * Create ILP pressure constraints, based on alive nodes:
1501  * - add additional costs to objective function if a node is scheduled
1502  *   on a unit although all units of this type are currently occupied
1503  */
1504 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1505         char                  buf[1024];
1506         ir_node               *cur_irn;
1507         unsigned              num_cst = 0;
1508         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1509         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1510
1511         ilp_timer_push(t_cst);
1512         /* y_{nt}^k is set for each node and timestep and unit type */
1513         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1514                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1515                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1516                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1517                 int                  glob_type_idx;
1518
1519                 /* we ignore nodes assigned to DUMMY unit here */
1520                 if (cur_na->is_dummy_node)
1521                         continue;
1522
1523                 /* for all types */
1524                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1525                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1526                         int                      cur_tp_idx;
1527                         unsigned                 t;
1528
1529                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1530
1531                         /* check if node can be executed on this unit type */
1532                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1533                         if (cur_tp_idx < 0)
1534                                 continue;
1535
1536                         /* check all time_steps at which the current node can be scheduled */
1537                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1538                                 int     cst, y_idx;
1539                                 ir_node *irn;
1540                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1541                                 ilp_livein_node_t *livein;
1542
1543                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1544                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1545                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1546                                 num_cst++;
1547
1548                                 /* - accumulate all nodes alive at point t on unit type k */
1549                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1550                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1551                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1552                                         int                  a_idx, tp_idx;
1553
1554                                         /* check if node can be alive here */
1555                                         if (t < na->asap - 1)
1556                                                 continue;
1557
1558                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1559
1560                                         /* current type is not suitable */
1561                                         if (tp_idx < 0)
1562                                                 continue;
1563
1564                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1565                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1566                                 }
1567                                 /* do the same for livein nodes */
1568                                 foreach_pset(ba->livein_nodes, livein) {
1569                                         ir_node              *irn  = livein->irn;
1570                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1571                                         int                  a_idx, tp_idx;
1572
1573                                         /* check if node can be alive here */
1574                                         if (t >= livein->max_alive_steps)
1575                                                 continue;
1576
1577                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1578
1579                                         /* current type is not suitable */
1580                                         if (tp_idx < 0)
1581                                                 continue;
1582
1583                                         a_idx = tp_idx * livein->max_alive_steps + t;
1584                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1585                                 }
1586
1587                                 if (ARR_LEN(tmp_var_idx) > 0)
1588                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1589                                 DEL_ARR_F(tmp_var_idx);
1590
1591                                 /* - num_nodes * y_{nt}^k */
1592                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1593                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1594                         }
1595                 }
1596         }
1597         ilp_timer_pop();
1598         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1599                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1600 }
1601
1602 /**
1603  * Create ILP branch constraints:
1604  * Assure, alle nodes are scheduled prior to cfg op.
1605  */
1606 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1607         char                  buf[1024];
1608         ir_node               *cur_irn, *cfop;
1609         unsigned              num_cst          = 0;
1610         unsigned              num_non_branches = 0;
1611         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1612         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1613
1614         ilp_timer_push(t_cst);
1615         cfop = NULL;
1616         /* determine number of non-branch nodes and the one and only branch node */
1617         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1618                 switch (get_irn_opcode(cur_irn)) {
1619                         case iro_Phi:
1620                         case iro_Start:
1621                         case iro_End:
1622                         case iro_Proj:
1623                         case iro_Bad:
1624                         case iro_Unknown:
1625                                 num_non_branches++;
1626                                 break;
1627                         default:
1628                                 if (is_cfop(cur_irn)) {
1629                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1630                                         cfop = cur_irn;
1631                                 }
1632                                 else {
1633                                         num_non_branches++;
1634                                 }
1635                                 break;
1636                 }
1637         }
1638
1639         if (cfop) {
1640                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1641                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1642                 unsigned t;
1643
1644                 /* for each time step */
1645                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1646                         int *non_branch_vars, *branch_vars;
1647                         int cst;
1648
1649                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1650                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1651                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1652                         num_cst++;
1653
1654                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1655                         non_branch_vars = NEW_ARR_F(int, 0);
1656                         branch_vars     = NEW_ARR_F(int, 0);
1657                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1658                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1659                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1660                                 int                  tp_idx;
1661
1662                                 if (cur_irn == cfop) {
1663                                         /* for all unit types available for this node */
1664                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1665                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1666                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1667                                         }
1668                                 }
1669                                 else {
1670                                         /* sum up all possible schedule points for this node upto current timestep */
1671                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1672                                                 unsigned tn;
1673                                                 unsigned tmax = MIN(t, na->alap - 1);
1674
1675                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1676                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1677                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1678                                                 }
1679                                         }
1680                                 }
1681
1682                         }
1683
1684                         if (ARR_LEN(non_branch_vars) > 0)
1685                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1686                         if (ARR_LEN(branch_vars) > 0)
1687                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1688
1689                         DEL_ARR_F(branch_vars);
1690                         DEL_ARR_F(non_branch_vars);
1691                 }
1692         }
1693         ilp_timer_pop();
1694         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1695                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1696 }
1697
1698 #if 0
1699 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1700         char                  buf[1024];
1701         ir_node               *irn;
1702         unsigned              num_cst = 0;
1703         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1704         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1705
1706         ilp_timer_push(t_cst);
1707         /* check all nodes */
1708         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1709                 be_ilpsched_irn_t    *node;
1710                 ilpsched_node_attr_t *na;
1711                 unsigned             t;
1712                 ir_node              **pk;
1713
1714                 /* only mode_T nodes can have Projs and Keeps assigned */
1715                 if (get_irn_mode(irn) != mode_T)
1716                         continue;
1717
1718                 node = get_ilpsched_irn(env, irn);
1719                 na   = get_ilpsched_node_attr(node);
1720
1721                 /* check if has some Projs and Keeps assigned */
1722                 if (! na->projkeeps)
1723                         continue;
1724
1725                 /* we can run only once over the queue, so preserve the nodes */
1726                 pk = NEW_ARR_F(ir_node *, 0);
1727                 while (! waitq_empty(na->projkeeps))
1728                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1729                 del_waitq(na->projkeeps);
1730                 na->projkeeps = NULL;
1731
1732                 /* for all time steps at which this node can be scheduled */
1733                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1734                         int cst, tp_idx, i;
1735                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1736
1737                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1738                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1739                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1740                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1741                         num_cst++;
1742
1743                         /* sum up scheduling variables for this time step */
1744                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1745                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1746                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1747                         }
1748
1749                         if (ARR_LEN(tmp_var_idx_n) > 0)
1750                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1751                         DEL_ARR_F(tmp_var_idx_n);
1752
1753                         /* subtract all Proj and Keep variables for this step */
1754                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1755                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1756                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1757                                 int                  pk_tp_idx;
1758
1759                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1760                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1761                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1762                                 }
1763                         }
1764                 }
1765         }
1766         ilp_timer_pop();
1767         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1768                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1769 }
1770 #endif /* if 0 */
1771
1772 /***************************************************
1773  *  _____ _      _____                    _
1774  * |_   _| |    |  __ \                  (_)
1775  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1776  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1777  *  _| |_| |____| |      | | | | | | (_| | | | | |
1778  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1779  *
1780  ***************************************************/
1781
1782 /**
1783  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1784  */
1785 static void create_ilp(ir_node *block, void *walk_env) {
1786         be_ilpsched_env_t     *env           = walk_env;
1787         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1788         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1789         FILE                  *logfile       = NULL;
1790         lpp_t                 *lpp           = NULL;
1791         int                   need_heur      = 0;
1792         struct obstack        var_obst;
1793         char                  name[1024];
1794
1795         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1796         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1797         DBG((env->dbg, 255, "=========================================\n\n"));
1798
1799         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1800                 block, ba->n_interesting_nodes, ba->max_steps));
1801
1802         /* notify backend and get block environment */
1803         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1804
1805         /* if we have less than two interesting nodes, there is no need to create the ILP */
1806         if (ba->n_interesting_nodes > 1) {
1807                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1808                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1809                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1810                 int    estimated_n_var = (int)((double)base_num * fact_var);
1811                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1812
1813                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1814                         estimated_n_var, estimated_n_cst));
1815
1816                 /* set up the LPP object */
1817                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1818
1819                 lpp = new_lpp_userdef(
1820                         (const char *)name,
1821                         lpp_minimize,
1822                         estimated_n_cst,     /* num vars */
1823                         estimated_n_cst + 1, /* num cst */
1824                         1.3);                /* grow factor */
1825                 obstack_init(&var_obst);
1826
1827                 /* create ILP variables */
1828                 create_variables(env, lpp, block_node, &var_obst);
1829
1830                 /* create ILP constraints */
1831                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1832                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1833                 create_ressource_constraints(env, lpp, block_node);
1834                 create_bundle_constraints(env, lpp, block_node);
1835                 create_branch_constraint(env, lpp, block_node);
1836                 //create_proj_keep_constraints(env, lpp, block_node);
1837
1838                 if (env->opts->regpress) {
1839                         create_alive_nodes_constraint(env, lpp, block_node);
1840                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1841                         create_pressure_alive_constraint(env, lpp, block_node);
1842                 }
1843
1844                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1845
1846                 /* debug stuff, dump lpp when debugging is on  */
1847                 DEBUG_ONLY(
1848                         if (firm_dbg_get_mask(env->dbg) > 1) {
1849                                 char buf[1024];
1850                                 FILE *f;
1851
1852                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1853                                 f = fopen(buf, "w");
1854                                 lpp_dump_plain(lpp, f);
1855                                 fclose(f);
1856                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1857                                 lpp_dump(lpp, buf);
1858                         }
1859                 );
1860
1861                 /* set solve time limit */
1862                 lpp_set_time_limit(lpp, env->opts->time_limit);
1863
1864                 /* set logfile if requested */
1865                 if (strlen(env->opts->log_file) > 0) {
1866                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1867                                 lpp_set_log(lpp, stdout);
1868                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1869                                 lpp_set_log(lpp, stderr);
1870                         else {
1871                                 logfile = fopen(env->opts->log_file, "w");
1872                                 if (! logfile)
1873                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1874                                 else
1875                                         lpp_set_log(lpp, logfile);
1876                         }
1877                 }
1878
1879                 /* solve the ILP */
1880                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1881
1882                 if (logfile)
1883                         fclose(logfile);
1884
1885                 /* check for valid solution */
1886                 if (! lpp_is_sol_valid(lpp)) {
1887                         char buf[1024];
1888                         FILE *f;
1889
1890                         DEBUG_ONLY(
1891                                 if (firm_dbg_get_mask(env->dbg) >= 2) {
1892                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1893                                         f = fopen(buf, "w");
1894                                         lpp_dump_plain(lpp, f);
1895                                         fclose(f);
1896                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1897                                         lpp_dump(lpp, buf);
1898                                         dump_ir_block_graph(env->irg, "-infeasible");
1899                                 }
1900                         )
1901
1902                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1903                         need_heur = 1;
1904                 }
1905
1906                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1907                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1908                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1909                 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));
1910                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1911                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1912                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1913                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1914
1915                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1916         }
1917
1918         /* apply solution */
1919 #ifdef FIRM_STATISTICS
1920         if (be_stat_ev_is_active()) {
1921                 be_stat_ev("nodes", ba->block_last_idx);
1922         }
1923 #endif /* FIRM_STATISTICS */
1924         if (need_heur) {
1925 #ifdef FIRM_STATISTICS
1926                 if (be_stat_ev_is_active()) {
1927                         be_stat_ev("time", -1);
1928                 }
1929 #endif /* FIRM_STATISTICS */
1930                 list_sched_single_block(env->birg, block, env->be_opts);
1931         }
1932         else {
1933 #ifdef FIRM_STATISTICS
1934                 if (be_stat_ev_is_active()) {
1935                         if (lpp)
1936                                 be_stat_ev_dbl("time", lpp->sol_time);
1937                         else
1938                                 be_stat_ev("time", 0);
1939                 }
1940 #endif /* FIRM_STATISTICS */
1941                 apply_solution(env, lpp, block);
1942         }
1943
1944         if (lpp)
1945                 free_lpp(lpp);
1946
1947         /* notify backend */
1948         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1949 }
1950
1951 /**
1952  * Perform ILP scheduling on the given irg.
1953  */
1954 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
1955         be_ilpsched_env_t          env;
1956         const char                 *name = "be ilp scheduling";
1957         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
1958         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
1959
1960         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
1961
1962 #ifdef FIRM_STATISTICS
1963         if (be_stat_ev_is_active()) {
1964                 be_stat_tags[STAT_TAG_CLS] = "ilpsched";
1965                 be_stat_ev_push(be_stat_tags, STAT_TAG_LAST, be_stat_file);
1966         }
1967 #endif /* FIRM_STATISTICS */
1968
1969 //      firm_dbg_set_mask(env.dbg, 1);
1970
1971         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
1972         env.sel        = sel;
1973         env.irg        = birg->irg;
1974         env.height     = heights_new(birg->irg);
1975         env.main_env   = birg->main_env;
1976         env.arch_env   = birg->main_env->arch_env;
1977         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
1978         env.opts       = &ilp_opts;
1979         env.birg       = birg;
1980         env.be_opts    = be_opts;
1981         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
1982
1983         /* assign a unique per block number to all interesting nodes */
1984         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
1985
1986         /*
1987                 The block indices are completely build after the walk,
1988                 now we can allocate the bitsets (size depends on block indices)
1989                 for all nodes.
1990         */
1991         phase_reinit_irn_data(&env.ph);
1992
1993         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
1994         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
1995
1996         /* Calculate ALAP of all irns */
1997         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
1998
1999         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2000         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2001
2002         /* perform ILP scheduling */
2003         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2004
2005         DEBUG_ONLY(
2006                 if (firm_dbg_get_mask(env.dbg)) {
2007                         phase_stat_t stat;
2008                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2009
2010                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2011                 }
2012         );
2013
2014         /* free data allocated dynamically */
2015         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2016
2017         /* free all allocated object */
2018         phase_free(&env.ph);
2019         heights_free(env.height);
2020
2021         /* notify backend */
2022         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2023
2024 #ifdef FIRM_STATISTICS
2025         if (be_stat_ev_is_active()) {
2026                 be_stat_ev_pop();
2027         }
2028 #endif /* FIRM_STATISTICS */
2029 }
2030
2031 /**
2032  * Register ILP scheduler options.
2033  */
2034 void be_init_ilpsched(void)
2035 {
2036         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2037         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2038
2039         lc_opt_add_table(sched_grp, ilpsched_option_table);
2040 }
2041
2042 #else /* WITH_ILP */
2043
2044 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2045 {}
2046
2047 #endif /* WITH_ILP */