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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       ILP based instruction scheduling.
  23  * @author      Christian Wuerdig
  24  * @date        22.10.2006
  25  * @version     $Id$
  26  *
  27  * An ILP scheduler based on
  28  * "ILP-based Instruction Scheduling for IA-64"
  29  * by Daniel Kaestner and Sebastian Winkel
  30  * extended with register pressure constraints by Christian Wuerdig
  31  */
  32 #include "config.h"
  33
  34 #ifdef WITH_ILP
  35
  36 #include <math.h>
  37
  38 #ifndef _WIN32
  39 #include <strings.h>
  40 #endif /* _WIN32 */
  41
  42 #include "irnode_t.h"
  43 #include "irgwalk.h"
  44 #include "irbitset.h"
  45 #include "irphase_t.h"
  46 #include "height.h"
  47 #include "iredges.h"
  48 #include "pdeq.h"
  49 #include "debug.h"
  50 #include "irtools.h"
  51 #include "irdump.h"
  52 #include "irprintf.h"
  53 #include "plist.h"
  54 #include "irprintf.h"
  55 #include "timing.h"
  56
  57 #include <lpp/lpp.h>
  58 #include <lpp/lpp_net.h>
  59
  60 #include "lc_opts.h"
  61 #include "lc_opts_enum.h"
  62
  63 #include "be.h"
  64 #include "benode.h"
  65 #include "besched.h"
  66 #include "beilpsched.h"
  67 #include "beutil.h"
  68 #include "bestat.h"
  69 #include "beirg.h"
  70 #include "bemachine.h"
  71 #include "belistsched.h"
  72
  73 typedef struct _ilpsched_options_t {
  74         unsigned regpress;
  75         unsigned time_limit;
  76         char     log_file[1024];
  77 } ilpsched_options_t;
  78
  79 typedef struct _unit_type_info_t {
  80         int                            n_units;
  81         const be_execution_unit_type_t *tp;
  82 } unit_type_info_t;
  83
  84 /**
  85  * holding the ILP variables of the different types
  86  */
  87 typedef struct _ilp_var_types_t {
  88         int *x;   /* x_{nt}^k variables */
  89         int *a;   /* a_{nt}^k variables */
  90         int *y;   /* y_{nt}^k variables */
  91 } ilp_var_types_t;
  92
  93 /**
  94  * Holds alive variables for a node live-in to a block.
  95  */
  96 typedef struct _ilp_livein_node_t {
  97         ir_node  *irn;
  98         unsigned max_alive_steps;
  99         int      *a;
 100 } ilp_livein_node_t;
 101
 102 /* attributes for a node */
 103 typedef struct _ilpsched_node_attr_t {
 104         unsigned asap;                     /**< The ASAP scheduling control step */
 105         unsigned alap;                     /**< The ALAP scheduling control step */
 106         unsigned latency;                  /**< Latency of this node (needed for sorting) */
 107         unsigned sched_point;              /**< the step in which the node is finally scheduled */
 108         unsigned visit_idx;                /**< Index of the node having visited this node last */
 109         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
 110         unsigned n_consumer;               /**< Number of consumers */
 111         ir_node  **block_consumer;         /**< List of consumer being in the same block */
 112         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
 113         unsigned block_idx     : 30;       /**< A unique per block index */
 114         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
 115         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
 116         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
 117                                                                                         for ASAP, successors for ALAP */
 118         unsigned         n_unit_types;     /**< number of allowed execution unit types */
 119         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
 120         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
 121 } ilpsched_node_attr_t;
 122
 123 /* attributes for a block */
 124 typedef struct _ilpsched_block_attr_t {
 125         unsigned block_last_idx;        /**< The highest node index in block so far */
 126         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 127         unsigned max_steps;             /**< Upper bound for block execution */
 128         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 129         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 130         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 131 } ilpsched_block_attr_t;
 132
 133 typedef union _ilpsched_attr_ {
 134         ilpsched_node_attr_t  node_attr;
 135         ilpsched_block_attr_t block_attr;
 136 } ilpsched_attr_t;
 137
 138 /* A irn for the phase and it's attributes (either node or block) */
 139 typedef struct {
 140         const ir_node   *irn;
 141         ilpsched_attr_t attr;
 142 } be_ilpsched_irn_t;
 143
 144 /* The ILP scheduling environment */
 145 typedef struct {
 146         ir_phase             ph;            /**< The phase */
 147         ir_graph             *irg;          /**< The current irg */
 148         heights_t            *height;       /**< The heights object of the irg */
 149         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 150         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 151         const arch_env_t     *arch_env;
 152         const be_main_env_t  *main_env;
 153         const be_machine_t   *cpu;          /**< the current abstract machine */
 154         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 155         const be_irg_t       *birg;         /**< The birg object */
 156         be_options_t         *be_opts;      /**< backend options */
 157         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 158         DEBUG_ONLY(firm_dbg_module_t *dbg);
 159 } be_ilpsched_env_t;
 160
 161 /* convenience macros to handle phase irn data */
 162 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 163 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 164 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 165 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 166
 167 /* check if node is considered for ILP scheduling */
 168 #define consider_for_sched(env, irn) \
 169         (! (is_Block(irn)            ||  \
 170                 is_normal_Proj(env, irn) ||  \
 171                 is_Phi(irn)              ||  \
 172                 is_NoMem(irn)            ||  \
 173                 is_Unknown(irn)          ||  \
 174                 is_End(irn)                  \
 175                 ))
 176
 177 /* gives the valid scheduling time step interval for a node */
 178 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 179
 180 /* gives the valid interval where a node can die */
 181 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 182
 183 /* gives the corresponding ILP variable for given node, unit and time step */
 184 #define ILPVAR_IDX(na, unit, control_step) \
 185         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 186
 187 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 188 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 189         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 190
 191 /* check if a double value is within an epsilon environment of 0 */
 192 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 193
 194 /* option variable */
 195 static ilpsched_options_t ilp_opts = {
 196         1,     /* default is with register pressure constraints */
 197         300,   /* 300 sec per block time limit */
 198         ""     /* no log file */
 199 };
 200
 201 /* ILP options */
 202 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 203         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 204         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 205         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 206         LC_OPT_LAST
 207 };
 208
 209 /*
 210         We need this global variable as we compare nodes dependent on heights,
 211         but we cannot pass any information to the qsort compare function.
 212 */
 213 static heights_t *glob_heights;
 214
 215 /**
 216  * Check if irn is a Proj, which has no execution units assigned.
 217  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 218  */
 219 static inline int is_normal_Proj(const arch_env_t *env, const ir_node *irn)
 220 {
 221         return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
 222 }
 223
 224 /**
 225  * Skips normal Projs.
 226  * @return predecessor if irn is a normal Proj, otherwise irn.
 227  */
 228 static inline ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn)
 229 {
 230         if (is_normal_Proj(env, irn))
 231                 return get_Proj_pred(irn);
 232         return irn;
 233 }
 234
 235 static inline int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env)
 236 {
 237         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 238         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 239                 lat = 1;
 240         return lat;
 241 }
 242
 243 static int cmp_live_in_nodes(const void *a, const void *b)
 244 {
 245         const ilp_livein_node_t *n1 = a;
 246         const ilp_livein_node_t *n2 = b;
 247
 248         return n1->irn != n2->irn;
 249 }
 250
 251 /**
 252  * Compare scheduling time steps of two be_ilpsched_irn's.
 253  */
 254 static int cmp_ilpsched_irn(const void *a, const void *b)
 255 {
 256         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 257         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 258         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 259         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 260
 261         if (n1_a->sched_point == n2_a->sched_point) {
 262                 const ir_node *irn_a = n1->irn;
 263                 const ir_node *irn_b = n2->irn;
 264
 265                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 266                         return 1;
 267                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 268                         return -1;
 269
 270                 /*
 271                         Ok, timestep is equal and the nodes are parallel,
 272                         so check latency and schedule high latency first.
 273                 */
 274                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 275         }
 276         else
 277                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 278 }
 279
 280 /**
 281  * In case there is no phase information for irn, initialize it.
 282  */
 283 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old)
 284 {
 285         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 286
 287         if (res == old) {
 288                 /* if we have already some data: check for reinitialization */
 289
 290                 if (! is_Block(irn)) {
 291                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 292
 293                         if (! na->transitive_block_nodes) {
 294                                 ir_node               *block      = get_nodes_block(irn);
 295                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 296                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 297
 298                                 /* we are called after the block indices have been build: create bitset */
 299                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 300                         }
 301                         else {
 302                                 /* we are called from reinit block data: clear the bitset */
 303                                 bitset_clear_all(na->transitive_block_nodes);
 304                                 na->visit_idx    = 0;
 305                                 na->alap_changed = 1;
 306                         }
 307                 }
 308                 return old;
 309         }
 310
 311         res->irn = irn;
 312
 313         /* set ilpsched irn attributes (either block or irn) */
 314         if (is_Block(irn)) {
 315                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 316
 317                 ba->n_interesting_nodes = 0;
 318                 ba->block_last_idx      = 0;
 319                 ba->root_nodes          = plist_new();
 320                 ba->head_ilp_nodes      = NULL;
 321                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 322                 ba->max_steps           = 0;
 323         }
 324         else {
 325                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 326                 memset(na, 0, sizeof(*na));
 327         }
 328
 329         return res;
 330 }
 331
 332 /**
 333  * Assign a per block unique number to each node.
 334  */
 335 static void build_block_idx(ir_node *irn, void *walk_env)
 336 {
 337         be_ilpsched_env_t     *env = walk_env;
 338         be_ilpsched_irn_t     *node, *block_node;
 339         ilpsched_node_attr_t  *na;
 340         ilpsched_block_attr_t *ba;
 341
 342         set_irn_link(irn, NULL);
 343         if (! consider_for_sched(env->arch_env, irn))
 344                 return;
 345
 346         node       = get_ilpsched_irn(env, irn);
 347         na         = get_ilpsched_node_attr(node);
 348         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 349         ba         = get_ilpsched_block_attr(block_node);
 350
 351         na->block_idx = ba->block_last_idx++;
 352 }
 353
 354 /********************************************************
 355  *                              __        _
 356  *                             / /       | |
 357  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 358  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 359  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 360  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 361  *                 | |                           | |
 362  *                 |_|                           |_|
 363  ********************************************************/
 364
 365 /**
 366  * Add all nodes having no user in current block to last_nodes list.
 367  */
 368 static void collect_alap_root_nodes(ir_node *irn, void *walk_env)
 369 {
 370         ir_node               *block;
 371         const ir_edge_t       *edge;
 372         be_ilpsched_irn_t     *block_node, *node;
 373         ilpsched_block_attr_t *ba;
 374         ilpsched_node_attr_t  *na;
 375         int                   i, j;
 376         be_ilpsched_env_t     *env           = walk_env;
 377         int                   has_block_user = 0;
 378         unsigned              n_consumer     = 0;
 379         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 380         ir_node               **consumer;
 381         unsigned              idx;
 382
 383         if (! consider_for_sched(env->arch_env, irn))
 384                 return;
 385
 386         block    = get_nodes_block(irn);
 387     idx      = get_irn_idx(irn);
 388         consumer = NEW_ARR_F(ir_node *, 0);
 389
 390         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 391
 392         /* check data and dependency out edges */
 393         for (i = 0; i < 2 && ! has_block_user; ++i) {
 394                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 395                         ir_node *user = get_edge_src_irn(edge);
 396
 397                         if (is_normal_Proj(env->arch_env, user)) {
 398                                 const ir_edge_t *user_edge;
 399
 400                                 if (get_irn_mode(user) == mode_X)
 401                                         continue;
 402
 403                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 404                                 for (j = 0; j < 2; ++j) {
 405                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 406                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 407
 408                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 409                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 410                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 411
 412                                                         /* skip already visited nodes */
 413                                                         if (ua->consumer_idx == idx)
 414                                                                 continue;
 415
 416                                                         /* check if node has user in this block and collect the user if it's a data user */
 417                                                         if (get_nodes_block(real_user) == block) {
 418                                                                 if (i == 0 && j == 0)
 419                                                                         ARR_APP1(ir_node *, consumer, real_user);
 420                                                                 has_block_user = 1;
 421                                                         }
 422
 423                                                         /* only count data consumer */
 424                                                         if (i == 0)
 425                                                                 n_consumer++;
 426
 427                                                         /* mark user as visited by this node */
 428                                                         ua->consumer_idx = idx;
 429                                                 }
 430                                         }
 431                                 }
 432                         }
 433                         else if (is_Block(user)) {
 434                                 continue;
 435                         }
 436                         else if (! is_Phi(user)) {
 437                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 438                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 439
 440                                 /* skip already visited nodes */
 441                                 if (ua->consumer_idx == idx)
 442                                         continue;
 443
 444                                 /* check if node has user in this block and collect the user if it's a data user */
 445                                 if (get_nodes_block(user) == block) {
 446                                         if (i == 0)
 447                                                 ARR_APP1(ir_node *, consumer, user);
 448                                         has_block_user = 1;
 449                                 }
 450
 451                                 /* only count data consumer */
 452                                 if (i == 0)
 453                                         n_consumer++;
 454
 455                                 /* mark user visited by this node */
 456                                 ua->consumer_idx = idx;
 457                         }
 458                         else if (get_nodes_block(user) != block) {
 459                                 n_consumer++;
 460                         }
 461                 }
 462         }
 463
 464         block_node = get_ilpsched_irn(env, block);
 465         ba         = get_ilpsched_block_attr(block_node);
 466
 467         ba->n_interesting_nodes++;
 468
 469         /* current irn has no user inside this block, add to queue */
 470         if (! has_block_user) {
 471                 DB((env->dbg, LEVEL_3, "root node\n"));
 472                 plist_insert_back(ba->root_nodes, irn);
 473         }
 474         else {
 475                 DB((env->dbg, LEVEL_3, "normal node\n"));
 476         }
 477
 478         /* record number of all consumer and the consumer within the same block */
 479         node = get_ilpsched_irn(env, irn);
 480         na   = get_ilpsched_node_attr(node);
 481         na->n_consumer     = n_consumer;
 482         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 483         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 484         DEL_ARR_F(consumer);
 485 }
 486
 487 /**
 488  * Calculate the ASAP scheduling step for current irn.
 489  */
 490 static void calculate_irn_asap(ir_node *irn, void *walk_env)
 491 {
 492         be_ilpsched_env_t     *env = walk_env;
 493         int                   i;
 494         ir_node               *block;
 495         be_ilpsched_irn_t     *node, *block_node;
 496         ilpsched_node_attr_t  *na;
 497         ilpsched_block_attr_t *ba;
 498
 499         /* These nodes are handled separate */
 500         if (! consider_for_sched(env->arch_env, irn))
 501                 return;
 502
 503         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 504
 505         block    = get_nodes_block(irn);
 506         node     = get_ilpsched_irn(env, irn);
 507         na       = get_ilpsched_node_attr(node);
 508         na->asap = 1;
 509
 510         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 511                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
 512
 513                 /* check for greatest distance to top */
 514                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 515                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 516                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 517                         unsigned             lat;
 518
 519                         lat         = fixed_latency(env->sel, pred, env->block_env);
 520                         na->latency = lat;
 521                         na->asap    = MAX(na->asap, pna->asap + lat);
 522                 }
 523         }
 524
 525         /* add node to ILP node list and update max_steps */
 526         block_node = get_ilpsched_irn(env, block);
 527         ba         = get_ilpsched_block_attr(block_node);
 528
 529         set_irn_link(irn, ba->head_ilp_nodes);
 530         ba->head_ilp_nodes = irn;
 531         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 532
 533         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 534 }
 535
 536 /**
 537  * Calculate the ALAP scheduling step of all irns in current block.
 538  * Depends on max_steps being calculated.
 539  */
 540 static void calculate_block_alap(ir_node *block, void *walk_env)
 541 {
 542         be_ilpsched_env_t     *env        = walk_env;
 543         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 544         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 545         waitq                 *cur_queue  = new_waitq();
 546         plist_element_t       *el;
 547
 548         assert(is_Block(block));
 549
 550         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 551                 block, ba->n_interesting_nodes, ba->max_steps));
 552
 553         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 554         //phase_reinit_block_irn_data(&env->ph, block);
 555
 556         /* init start queue */
 557         foreach_plist(ba->root_nodes, el) {
 558                 waitq_put(cur_queue, plist_element_get_value(el));
 559         }
 560
 561         /* repeat until all nodes are processed */
 562         while (! waitq_empty(cur_queue)) {
 563                 waitq *next_queue = new_waitq();
 564
 565                 /* process all nodes in current step */
 566                 while (! waitq_empty(cur_queue)) {
 567                         ir_node              *cur_irn = waitq_get(cur_queue);
 568                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 569                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 570                         int                  i;
 571
 572                         /* cur_node has no alap set -> it's a root node, set to max alap */
 573                         if (na->alap == 0) {
 574                                 na->alap = ba->max_steps;
 575                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 576                                         cur_irn, na->alap));
 577                         }
 578                         else {
 579                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 580                                         cur_irn, na->alap));
 581                         }
 582
 583                         /* set the alap's of all predecessors */
 584                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 585                                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
 586
 587                                 /* check for greatest distance to bottom */
 588                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 589                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 590                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 591                                         unsigned             lat;
 592
 593                                         /* mark the predecessor as visited by current irn */
 594                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 595                                                 continue;
 596                                         pna->visit_idx = get_irn_idx(cur_irn);
 597
 598                                         lat = fixed_latency(env->sel, pred, env->block_env);
 599
 600                                         /* set ALAP of current pred */
 601                                         if (pna->alap == 0) {
 602                                                 /* current ALAP is 0: node has not yet been visited */
 603                                                 pna->alap_changed = 1;
 604                                                 pna->alap         = na->alap - lat;
 605                                         }
 606                                         else if (pna->alap > na->alap - lat) {
 607                                                 /* we found a longer path to root node: change ALAP */
 608                                                 pna->alap         = na->alap - lat;
 609                                                 pna->alap_changed = 1;
 610                                         }
 611                                         else {
 612                                                 /* current ALAP is best found so far: keep it */
 613                                                 pna->alap_changed = 0;
 614                                         }
 615
 616                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 617
 618                                         /* enqueue node for next iteration */
 619                                         if (get_irn_ins_or_deps(pred) > 0)
 620                                                 waitq_put(next_queue, pred);
 621                                 }
 622                         }
 623                 }
 624
 625                 /* prepare for next iteration */
 626                 del_waitq(cur_queue);
 627                 cur_queue = next_queue;
 628         }
 629 }
 630
 631 /**
 632  * Free list of root nodes and the set of live-in nodes.
 633  */
 634 static void clear_unwanted_data(ir_node *block, void *walk_env)
 635 {
 636         be_ilpsched_env_t     *env        = walk_env;
 637         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 638         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 639
 640         plist_free(ba->root_nodes);
 641         ba->root_nodes = NULL;
 642         del_pset(ba->livein_nodes);
 643         ba->livein_nodes = NULL;
 644 }
 645
 646 /**
 647  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 648  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 649  */
 650 static void refine_asap_alap_times(ir_node *irn, void *walk_env)
 651 {
 652         be_ilpsched_env_t    *env  = walk_env;
 653         ir_node              *pred = irn;
 654         be_ilpsched_irn_t    *node, *pred_node;
 655         ilpsched_node_attr_t *na, *pna;
 656
 657         if (! consider_for_sched(env->arch_env, irn))
 658                 return;
 659
 660         if (! is_Proj(irn) && ! be_is_Keep(irn))
 661                 return;
 662
 663         /* go to the ancestor */
 664         if (be_is_Keep(irn))
 665                 pred = get_irn_n(irn, 0);
 666         pred = skip_Proj(pred);
 667
 668         node      = get_ilpsched_irn(env, irn);
 669         pred_node = get_ilpsched_irn(env, pred);
 670         na        = get_ilpsched_node_attr(node);
 671         pna       = get_ilpsched_node_attr(pred_node);
 672
 673         na->asap = pna->asap;
 674         na->alap = pna->alap;
 675
 676         /* record all Projs and Keeps for this node */
 677         if (! pna->projkeeps)
 678                 pna->projkeeps = new_waitq();
 679         waitq_put(pna->projkeeps, irn);
 680
 681         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 682 }
 683
 684 /*******************************************
 685  *           _              _       _
 686  *          | |            | |     | |
 687  *  ___  ___| |__   ___  __| |_   _| | ___
 688  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 689  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 690  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 691  *
 692  *******************************************/
 693
 694 static inline void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn)
 695 {
 696         const ir_edge_t *edge;
 697         (void) block;
 698
 699         foreach_out_edge(irn, edge) {
 700                 ir_node *user = get_edge_src_irn(edge);
 701
 702                 if (be_is_Keep(user)) {
 703                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 704                         waitq_put(keeps, user);
 705                 }
 706         }
 707 }
 708
 709 /**
 710  * Inserts @p irn before @p before into schedule and notifies backend.
 711  */
 712 static inline void notified_sched_add_before(be_ilpsched_env_t *env,
 713         const ir_node *before, const ir_node *irn, unsigned cycle)
 714 {
 715         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 716         sched_add_before((ir_node*) before, (ir_node*) irn);
 717 }
 718
 719 /**
 720  * Adds a node, it's Projs (in case of mode_T nodes) and
 721  * it's Keeps to schedule.
 722  */
 723 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle)
 724 {
 725         const ir_edge_t *edge;
 726         waitq           *keeps = new_waitq();
 727
 728         /* mode_M nodes are not scheduled */
 729         if (get_irn_mode(irn) == mode_M)
 730                 return;
 731
 732         if (! sched_is_scheduled(irn))
 733                 notified_sched_add_before(env, block, irn, cycle);
 734
 735         /* add Projs */
 736         if (get_irn_mode(irn) == mode_T) {
 737                 foreach_out_edge(irn, edge) {
 738                         ir_node *user = get_edge_src_irn(edge);
 739
 740                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 741                                 get_irn_n_edges(user) > 0)
 742                         {
 743                                 notified_sched_add_before(env, block, user, cycle);
 744                         }
 745
 746                         check_for_keeps(keeps, block, user);
 747                 }
 748         }
 749         else {
 750                 check_for_keeps(keeps, block, irn);
 751         }
 752
 753         /* add Keeps */
 754         while (! waitq_empty(keeps)) {
 755                 ir_node *keep = waitq_get(keeps);
 756                 if (! sched_is_scheduled(keep))
 757                         notified_sched_add_before(env, block, keep, cycle);
 758         }
 759
 760         del_waitq(keeps);
 761 }
 762
 763 /**
 764  * Schedule all nodes in the given block, according to the ILP solution.
 765  */
 766 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block)
 767 {
 768         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 769         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 770         be_ilpsched_irn_t     **sched_nodes;
 771         unsigned              i, l;
 772         ir_node               *cfop, *irn;
 773         const ir_edge_t       *edge;
 774
 775         /* init block schedule list */
 776         sched_init_block(block);
 777
 778         /* collect nodes and their scheduling time step */
 779         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 780         if (ba->n_interesting_nodes == 0) {
 781                 /* ignore */
 782         }
 783         else if (ba->n_interesting_nodes == 1) {
 784                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 785
 786                 /* add the single node */
 787                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 788         }
 789         else {
 790                 /* check all nodes for their positive solution */
 791                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 792                         be_ilpsched_irn_t    *node;
 793                         ilpsched_node_attr_t *na;
 794                         int                  tp_idx, found;
 795                         unsigned             cur_var, t;
 796
 797                         node    = get_ilpsched_irn(env, irn);
 798                         na      = get_ilpsched_node_attr(node);
 799                         cur_var = 0;
 800                         found   = 0;
 801
 802                         if (! na->is_dummy_node) {
 803                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 804                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 805                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 806
 807                                                 if (! LPP_VALUE_IS_0(cost)) {
 808                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
 809                                                         found = 1;
 810                                                 }
 811                                         }
 812                                 }
 813                         }
 814
 815                         found = 0;
 816                         /* go over all variables of a node until the non-zero one is found */
 817                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 818                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 819                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 820
 821                                         /* check, if variable is set to one (it's not zero then :) */
 822                                         if (! LPP_VALUE_IS_0(val)) {
 823                                                 na->sched_point = t;
 824                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 825                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 826                                                         irn, t, na->type_info[tp_idx].tp->name));
 827                                                 found = 1;
 828                                         }
 829                                 }
 830                         }
 831                 }
 832
 833                 glob_heights = env->height;
 834                 /* sort nodes ascending by scheduling time step */
 835                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 836         }
 837
 838         /* make all Phis ready and remember the single cf op */
 839         cfop = NULL;
 840         foreach_out_edge(block, edge) {
 841                 irn = get_edge_src_irn(edge);
 842
 843                 switch (get_irn_opcode(irn)) {
 844                         case iro_Phi:
 845                                 add_to_sched(env, block, irn, 0);
 846                                 break;
 847                         case beo_Start:
 848                         case iro_End:
 849                         case iro_Proj:
 850                         case iro_Bad:
 851                         case iro_Unknown:
 852                                 break;
 853                         default:
 854                                 if (is_cfop(irn)) {
 855                                         assert(cfop == NULL && "Highlander - there can be only one");
 856                                         cfop = irn;
 857                                 }
 858                         break;
 859                 }
 860         }
 861
 862         /* add all nodes from list */
 863         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 864                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 865                 if (sched_nodes[i]->irn != cfop)
 866                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 867         }
 868
 869         /* schedule control flow node if not already done */
 870         if (cfop && ! sched_is_scheduled(cfop))
 871                 add_to_sched(env, block, cfop, 0);
 872
 873         DEL_ARR_F(sched_nodes);
 874 }
 875
 876 /***************************************************************
 877  *   _____ _      _____     _____           _   _
 878  *  |_   _| |    |  __ \   / ____|         | | (_)
 879  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 880  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 881  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 882  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 883  *
 884  ***************************************************************/
 885
 886 /**
 887  * Check if node can be executed on given unit type.
 888  */
 889 static inline int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node)
 890 {
 891         int                  i;
 892         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 893
 894         for (i = na->n_unit_types - 1; i >= 0; --i) {
 895                 if (na->type_info[i].tp == tp)
 896                         return i;
 897         }
 898
 899         return -1;
 900 }
 901
 902 /************************************************
 903  *                   _       _     _
 904  *                  (_)     | |   | |
 905  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 906  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 907  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 908  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 909  *
 910  ************************************************/
 911
 912 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst)
 913 {
 914         const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
 915         unsigned                  n_unit_types = 0;
 916         be_ilpsched_irn_t         *node;
 917         ilpsched_node_attr_t      *na;
 918         unsigned                  unit_idx, tp_idx;
 919
 920         /* count number of available unit types for this node */
 921         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 922                 /* just count */ ;
 923
 924         node = get_ilpsched_irn(env, irn);
 925         na   = get_ilpsched_node_attr(node);
 926
 927         if (! na->type_info) {
 928                 na->n_unit_types = n_unit_types;
 929                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 930
 931                 /* fill the type info array */
 932                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 933                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 934                                 /* beware: we also count number of available units here */
 935                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 936                                         na->is_dummy_node = 1;
 937                         }
 938
 939                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 940                         na->type_info[tp_idx].n_units = unit_idx;
 941                 }
 942         }
 943
 944         return n_unit_types;
 945 }
 946
 947 /**
 948  * Returns the largest alap time of a user of @p irn.
 949  * The user must be in block @p block.
 950  */
 951 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block)
 952 {
 953         const ir_edge_t *edge;
 954         unsigned        max_alap = 0;
 955
 956         foreach_out_edge(irn, edge) {
 957                 ir_node *user = get_edge_src_irn(edge);
 958
 959                 if (get_nodes_block(user) == block) {
 960                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 961                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 962
 963                         max_alap = MAX(max_alap, na->alap);
 964                 }
 965         }
 966
 967         assert(max_alap > 0);
 968         return max_alap;
 969 }
 970
 971 /**
 972  * Create the following variables:
 973  * - x_{nt}^k    binary     weigthed with: t
 974  *      node n is scheduled at time step t to unit type k
 975  * ==>> These variables represent the schedule
 976  *
 977  * - a_{nt}^k    binary     weighted with num_nodes
 978  *      node n is alive at time step t on unit type k
 979  *
 980  * - y_{nt}^k    continuous  weighted with: num_nodes^2
 981  *      register pressure over limit for unit type k
 982  * ==>> These variables represent the register pressure
 983  *
 984  */
 985 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst)
 986 {
 987         char                  buf[1024];
 988         ir_node               *irn;
 989         unsigned              num_block_var, num_nodes;
 990         ilp_livein_node_t     *livein;
 991         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 992         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 993
 994         stat_ev_tim_push();
 995
 996         num_block_var = num_nodes = 0;
 997         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 998                 be_ilpsched_irn_t    *node;
 999                 ilpsched_node_attr_t *na;
1000                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
1001                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
1002                 int                  i;
1003
1004                 node         = get_ilpsched_irn(env, irn);
1005                 na           = get_ilpsched_node_attr(node);
1006                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
1007
1008                 /* allocate space for ilp variables */
1009                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1010                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
1011
1012                 /* we need these variables only for "real" nodes */
1013                 if (! na->is_dummy_node) {
1014                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1015                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1016
1017                         num_ad         = ba->max_steps - na->asap + 1;
1018                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1019                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1020                 }
1021
1022                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1023                         irn, na->asap, na->alap, na->n_unit_types));
1024
1025                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1026                 /* create variables */
1027                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1028                         unsigned t;
1029
1030                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1031                                 /* x_{nt}^k variables */
1032                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1033                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1034                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1035                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1036                                 /* variable counter */
1037                                 n_var++;
1038                                 num_block_var++;
1039
1040                                 if (! na->is_dummy_node) {
1041                                         /* y_{nt}^k variables */
1042                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1043                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1044                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1045                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1046
1047                                         /* variable counter */
1048                                         n_var++;
1049                                         num_block_var++;
1050                                 }
1051                         }
1052
1053                         /* a node can die at any step t: asap(n) <= t <= U */
1054                         if (! na->is_dummy_node) {
1055                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1056
1057                                         /* a_{nt}^k variables */
1058                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1059                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1060                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1061                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1062
1063                                         /* variable counter */
1064                                         n_var++;
1065                                         num_block_var++;
1066                                 }
1067                         }
1068
1069                         /* collect live-in nodes */
1070                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1071                                 ir_node *pred = get_irn_n(irn, i);
1072
1073                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1074                                         be_ilpsched_set_type_info(env, pred, var_obst);
1075                                         if (! na->is_dummy_node) {
1076                                                 ilp_livein_node_t *entry = OALLOC(var_obst, ilp_livein_node_t);
1077                                                 entry->irn = pred;
1078                                                 entry->a   = NULL;
1079                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1080                                         }
1081                                 }
1082                         }
1083                 }
1084
1085                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1086                 num_nodes++;
1087         }
1088
1089         /* create alive variables a_{nt}^k for live-ins */
1090         foreach_pset(ba->livein_nodes, livein) {
1091                 be_ilpsched_irn_t    *node;
1092                 ilpsched_node_attr_t *na;
1093                 unsigned             tp_idx, var_idx;
1094                 ir_node              *irn;
1095
1096                 irn  = livein->irn;
1097                 node = get_ilpsched_irn(env, irn);
1098                 na   = get_ilpsched_node_attr(node);
1099
1100                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1101
1102                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1103                 var_idx   = 0;
1104
1105                 /* create variables */
1106                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1107                         unsigned t;
1108
1109                         for (t = 0; t < livein->max_alive_steps; ++t) {
1110                                 /* a_{nt}^k variables */
1111                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1112                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1113                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1114                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1115                                 num_block_var++;
1116                         }
1117                 }
1118         }
1119
1120         stat_ev_tim_pop("beilpsched_var");
1121 }
1122
1123 /*******************************************************
1124  *                      _             _       _
1125  *                     | |           (_)     | |
1126  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1127  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1128  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1129  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1130  *
1131  *******************************************************/
1132
1133 /**
1134  * Collect all operands and nodes @p irn depends on.
1135  * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1136  */
1137 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps)
1138 {
1139         int i;
1140
1141         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1142                 ir_node *p = get_irn_in_or_dep(irn, i);
1143
1144                 if (is_Proj(p)) {
1145                         const ir_edge_t *edge;
1146
1147                         p = get_Proj_pred(p);
1148                         foreach_out_edge(p, edge) {
1149                                 ir_node *src = get_edge_src_irn(edge);
1150                                 ir_nodeset_insert(deps, src);
1151                         }
1152                 }
1153                 else {
1154                         ir_nodeset_insert(deps, p);
1155                 }
1156         }
1157 }
1158
1159 /**
1160  * Create following ILP constraints:
1161  * - the assignment constraints:
1162  *     assure each node is executed once by exactly one (allowed) execution unit
1163  * - the dead node assignment constraints:
1164  *     assure a node can only die at most once
1165  * - the precedence constraints:
1166  *     assure that no data dependencies are violated
1167  */
1168 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1169 {
1170         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1171         char                  buf[1024];
1172         ir_node               *irn;
1173         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1174         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1175
1176         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1177         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1178                 int                  cst, tp_idx;
1179                 unsigned             cur_var;
1180                 be_ilpsched_irn_t    *node;
1181                 ilpsched_node_attr_t *na;
1182                 ir_node              *pred;
1183                 ir_nodeset_t          deps;
1184                 ir_nodeset_iterator_t iter;
1185
1186                 ir_nodeset_init(&deps);
1187
1188                 node    = get_ilpsched_irn(env, irn);
1189                 na      = get_ilpsched_node_attr(node);
1190                 cur_var = 0;
1191
1192                 /* the assignment constraint */
1193                 stat_ev_tim_push();
1194                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1195                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1196                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1197                 num_cst_assign++;
1198
1199                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1200                 stat_ev_tim_pop("beilpsched_cst_assign");
1201
1202                 /* We have separate constraints for Projs and Keeps */
1203                 // ILP becomes infeasible ?!?
1204 //              if (is_Proj(irn) || be_is_Keep(irn))
1205 //                      continue;
1206
1207                 /* the precedence constraints */
1208                 stat_ev_tim_push();
1209                 bs_block_irns = bitset_clear_all(bs_block_irns);
1210
1211                 sta_collect_in_deps(irn, &deps);
1212                 foreach_ir_nodeset(&deps, pred, iter) {
1213                         unsigned             t_low, t_high, t;
1214                         be_ilpsched_irn_t    *pred_node;
1215                         ilpsched_node_attr_t *pna;
1216                         unsigned             delay;
1217
1218                         pred = skip_normal_Proj(env->arch_env, pred);
1219                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1220                                 continue;
1221
1222                         pred_node = get_ilpsched_irn(env, pred);
1223                         pna       = get_ilpsched_node_attr(pred_node);
1224
1225                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1226
1227                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1228                                 bitset_set(bs_block_irns, pna->block_idx);
1229                         else
1230                                 continue;
1231
1232                         /* irn = n, pred = m */
1233                         delay  = fixed_latency(env->sel, pred, env->block_env);
1234                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1235                         t_high = MIN(na->alap, pna->alap + delay - 1);
1236                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1237                                 unsigned tn, tm;
1238                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1239
1240                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1241                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1242                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1243                                 num_cst_prec++;
1244
1245                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1246                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1247                                         /* node variables have smaller index than pred variables */
1248                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1249                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1250                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1251                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1252                                                 }
1253                                         }
1254
1255                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1256                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1257                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1258                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1259                                                 }
1260                                         }
1261                                 }
1262                                 else {
1263                                         /* pred variables have smaller index than node variables */
1264                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1265                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1266                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1267                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1268                                                 }
1269                                         }
1270
1271                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1272                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1273                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1274                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1275                                                 }
1276                                         }
1277                                 }
1278
1279                                 if (ARR_LEN(tmp_var_idx) > 0)
1280                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1281
1282                                 DEL_ARR_F(tmp_var_idx);
1283                         }
1284                 }
1285                 ir_nodeset_destroy(&deps);
1286                 stat_ev_tim_pop("beilpsched_cst_prec");
1287         }
1288 }
1289
1290 /**
1291  * Create ILP resource constraints:
1292  * - assure that for each time step not more instructions are scheduled
1293  *   to the same unit types as units of this type are available
1294  */
1295 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1296 {
1297         int                   glob_type_idx;
1298         char                  buf[1024];
1299         unsigned              num_cst_resrc = 0;
1300         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1301
1302         stat_ev_tim_push();
1303         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1304                 unsigned                 t;
1305                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1306
1307                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1308
1309                 /* check each time step */
1310                 for (t = 0; t < ba->max_steps; ++t) {
1311                         ir_node *irn;
1312                         int     cst;
1313                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1314
1315                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1316                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1317                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1318                         num_cst_resrc++;
1319
1320                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1321                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1322                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1323                                 int                  tp_idx;
1324
1325                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1326
1327                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1328                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1329                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1330                                 }
1331                         }
1332
1333                         /* set constraints if we have some */
1334                         if (ARR_LEN(tmp_var_idx) > 0)
1335                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1336
1337                         DEL_ARR_F(tmp_var_idx);
1338                 }
1339         }
1340         stat_ev_tim_pop("beilpsched_cst_rsrc");
1341 }
1342
1343 /**
1344  * Create ILP bundle constraints:
1345  * - assure, at most bundle_size * bundles_per_cycle instructions
1346  *   can be started at a certain point.
1347  */
1348 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1349 {
1350         char                  buf[1024];
1351         unsigned              t;
1352         unsigned              num_cst_bundle = 0;
1353         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1354         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1355
1356         stat_ev_tim_push();
1357         for (t = 0; t < ba->max_steps; ++t) {
1358                 ir_node *irn;
1359                 int     cst;
1360                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1361
1362                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1363                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1364                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1365                 num_cst_bundle++;
1366
1367                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1368                         be_ilpsched_irn_t    *node;
1369                         ilpsched_node_attr_t *na;
1370                         int                  tp_idx;
1371
1372                         /* Projs and Keeps do not contribute to bundle size */
1373                         if (is_Proj(irn) || be_is_Keep(irn))
1374                                 continue;
1375
1376                         node = get_ilpsched_irn(env, irn);
1377                         na   = get_ilpsched_node_attr(node);
1378
1379                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1380                         if (na->is_dummy_node)
1381                                 continue;
1382
1383                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1384                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1385                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1386                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1387                                 }
1388                         }
1389                 }
1390
1391                 if (ARR_LEN(tmp_var_idx) > 0)
1392                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1393
1394                 DEL_ARR_F(tmp_var_idx);
1395         }
1396         stat_ev_tim_pop("beilpsched_cst_bundle");
1397 }
1398
1399 /**
1400  * Create ILP alive nodes constraints:
1401  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1402  */
1403 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1404 {
1405         char                  buf[1024];
1406         ir_node               *irn;
1407         unsigned              num_cst = 0;
1408         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1409
1410         stat_ev_tim_push();
1411         /* for each node */
1412         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1413                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1414                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1415                 unsigned             t;
1416
1417                 /* we ignore nodes assigned to dummy unit here */
1418                 if (na->is_dummy_node)
1419                         continue;
1420
1421                 /* check check all time steps: asap(n) <= t <= U */
1422                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1423                         int node_tp_idx;
1424
1425                         /* for all unit types available for this node */
1426                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1427                                 unsigned tn, tn_max, idx;
1428                                 int      cst, i;
1429                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1430                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1431
1432                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1433                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1434                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1435                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1436                                 num_cst++;
1437
1438                                 tn_max = MIN(na->alap - 1, t);
1439                                 /* check if the node has been scheduled so far */
1440                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1441                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1442                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1443                                 }
1444
1445                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1446                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1447                                 DEL_ARR_F(tmp_var_idx_n);
1448
1449                                 /* subtract the number of consumer scheduled so far */
1450                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1451                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1452                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1453                                         int                  tp_idx;
1454                                         unsigned             tm, tm_max;
1455
1456                                         tm_max = MIN(ca->alap - 1, t);
1457                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1458                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1459                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1460                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1461                                                 }
1462                                         }
1463                                 }
1464
1465                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1466                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1467                                 DEL_ARR_F(tmp_var_idx_m);
1468
1469                                 /* -c * a_{nt}^k */
1470                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1471                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1472
1473                         }
1474                 }
1475         }
1476         stat_ev_tim_pop("beilpsched_cst_alive_nodes");
1477 }
1478
1479 /**
1480  * Create ILP alive nodes constraints for live-in nodes:
1481  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1482  */
1483 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1484 {
1485         char                  buf[1024];
1486         ilp_livein_node_t     *livein;
1487         unsigned              num_cst = 0;
1488         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1489
1490         stat_ev_tim_push();
1491         /* for each node */
1492         foreach_pset(ba->livein_nodes, livein) {
1493                 ir_node              *irn  = livein->irn;
1494                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1495                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1496                 unsigned             t;
1497
1498                 /* check check all time steps: 0 <= t < max_alive_steps */
1499                 for (t = 0; t < livein->max_alive_steps; ++t) {
1500                         int node_tp_idx;
1501
1502                         /* for all unit types available for this node */
1503                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1504                                 const ir_edge_t *edge;
1505                                 unsigned idx;
1506                                 int      cst, num_block_user;
1507                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1508
1509                                 /* check the number of consumer scheduled so far */
1510                                 num_block_user = 0;
1511                                 foreach_out_edge(irn, edge) {
1512                                         ir_node              *user = get_edge_src_irn(edge);
1513                                         be_ilpsched_irn_t    *cons;
1514                                         ilpsched_node_attr_t *ca;
1515                                         int                  tp_idx;
1516                                         unsigned             tm, tm_max;
1517
1518                                         /* check only users within current block */
1519                                         if (get_nodes_block(user) != block_node->irn)
1520                                                 continue;
1521
1522                                         num_block_user++;
1523                                         cons = get_ilpsched_irn(env, user);
1524                                         ca   = get_ilpsched_node_attr(cons);
1525
1526                                         tm_max = MIN(ca->alap - 1, t);
1527                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1528                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1529                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1530                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1531                                                 }
1532                                         }
1533                                 }
1534
1535                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1536                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1537                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1538                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1539                                 num_cst++;
1540
1541                                 /* sum(scheduled users) */
1542                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1543                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1544                                 DEL_ARR_F(tmp_var_idx_m);
1545
1546                                 /* + c * a_{nt}^k */
1547                                 idx = node_tp_idx * livein->max_alive_steps + t;
1548                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1549                         }
1550                 }
1551         }
1552         stat_ev_tim_pop("beilpsched_cst_alive_livein_nodes");
1553 }
1554
1555 /**
1556  * Create ILP pressure constraints, based on alive nodes:
1557  * - add additional costs to objective function if a node is scheduled
1558  *   on a unit although all units of this type are currently occupied
1559  */
1560 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1561 {
1562         char                  buf[1024];
1563         ir_node               *cur_irn;
1564         unsigned              num_cst = 0;
1565         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1566
1567         stat_ev_tim_push();
1568         /* y_{nt}^k is set for each node and timestep and unit type */
1569         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1570                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1571                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1572                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1573                 int                  glob_type_idx;
1574
1575                 /* we ignore nodes assigned to DUMMY unit here */
1576                 if (cur_na->is_dummy_node)
1577                         continue;
1578
1579                 /* for all types */
1580                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1581                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1582                         int                      cur_tp_idx;
1583                         unsigned                 t;
1584
1585                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1586
1587                         /* check if node can be executed on this unit type */
1588                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1589                         if (cur_tp_idx < 0)
1590                                 continue;
1591
1592                         /* check all time_steps at which the current node can be scheduled */
1593                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1594                                 int     cst, y_idx;
1595                                 ir_node *irn;
1596                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1597                                 ilp_livein_node_t *livein;
1598
1599                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1600                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1601                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1602                                 num_cst++;
1603
1604                                 /* - accumulate all nodes alive at point t on unit type k */
1605                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1606                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1607                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1608                                         int                  a_idx, tp_idx;
1609
1610                                         /* check if node can be alive here */
1611                                         if (t < na->asap - 1)
1612                                                 continue;
1613
1614                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1615
1616                                         /* current type is not suitable */
1617                                         if (tp_idx < 0)
1618                                                 continue;
1619
1620                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1621                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1622                                 }
1623                                 /* do the same for livein nodes */
1624                                 foreach_pset(ba->livein_nodes, livein) {
1625                                         ir_node              *irn  = livein->irn;
1626                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1627                                         int                  a_idx, tp_idx;
1628
1629                                         /* check if node can be alive here */
1630                                         if (t >= livein->max_alive_steps)
1631                                                 continue;
1632
1633                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1634
1635                                         /* current type is not suitable */
1636                                         if (tp_idx < 0)
1637                                                 continue;
1638
1639                                         a_idx = tp_idx * livein->max_alive_steps + t;
1640                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1641                                 }
1642
1643                                 if (ARR_LEN(tmp_var_idx) > 0)
1644                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1645                                 DEL_ARR_F(tmp_var_idx);
1646
1647                                 /* - num_nodes * y_{nt}^k */
1648                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1649                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1650                         }
1651                 }
1652         }
1653         stat_ev_tim_pop("beilpsched_cst_pressure");
1654 }
1655
1656 /**
1657  * Create ILP branch constraints:
1658  * Assure, alle nodes are scheduled prior to cfg op.
1659  */
1660 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1661 {
1662         char                  buf[1024];
1663         ir_node               *cur_irn, *cfop;
1664         unsigned              num_cst          = 0;
1665         unsigned              num_non_branches = 0;
1666         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1667
1668         stat_ev_tim_push();
1669         cfop = NULL;
1670         /* determine number of non-branch nodes and the one and only branch node */
1671         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1672                 switch (get_irn_opcode(cur_irn)) {
1673                         case iro_Phi:
1674                         case beo_Start:
1675                         case iro_End:
1676                         case iro_Proj:
1677                         case iro_Bad:
1678                         case iro_Unknown:
1679                                 num_non_branches++;
1680                                 break;
1681                         default:
1682                                 if (is_cfop(cur_irn)) {
1683                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1684                                         cfop = cur_irn;
1685                                 }
1686                                 else {
1687                                         num_non_branches++;
1688                                 }
1689                                 break;
1690                 }
1691         }
1692
1693         if (cfop) {
1694                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1695                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1696                 unsigned t;
1697
1698                 /* for each time step */
1699                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1700                         int *non_branch_vars, *branch_vars;
1701                         int cst;
1702
1703                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1704                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1705                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1706                         num_cst++;
1707
1708                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1709                         non_branch_vars = NEW_ARR_F(int, 0);
1710                         branch_vars     = NEW_ARR_F(int, 0);
1711                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1712                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1713                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1714                                 int                  tp_idx;
1715
1716                                 if (cur_irn == cfop) {
1717                                         /* for all unit types available for this node */
1718                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1719                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1720                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1721                                         }
1722                                 }
1723                                 else {
1724                                         /* sum up all possible schedule points for this node upto current timestep */
1725                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1726                                                 unsigned tn;
1727                                                 unsigned tmax = MIN(t, na->alap - 1);
1728
1729                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1730                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1731                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1732                                                 }
1733                                         }
1734                                 }
1735
1736                         }
1737
1738                         if (ARR_LEN(non_branch_vars) > 0)
1739                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1740                         if (ARR_LEN(branch_vars) > 0)
1741                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1742
1743                         DEL_ARR_F(branch_vars);
1744                         DEL_ARR_F(non_branch_vars);
1745                 }
1746         }
1747         stat_ev_tim_pop("beilpsched_cst_branch");
1748 }
1749
1750 #if 0
1751 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1752 {
1753         char                  buf[1024];
1754         ir_node               *irn;
1755         unsigned              num_cst = 0;
1756         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1757
1758         stat_ev_tim_push();
1759         /* check all nodes */
1760         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1761                 be_ilpsched_irn_t    *node;
1762                 ilpsched_node_attr_t *na;
1763                 unsigned             t;
1764                 ir_node              **pk;
1765
1766                 /* only mode_T nodes can have Projs and Keeps assigned */
1767                 if (get_irn_mode(irn) != mode_T)
1768                         continue;
1769
1770                 node = get_ilpsched_irn(env, irn);
1771                 na   = get_ilpsched_node_attr(node);
1772
1773                 /* check if has some Projs and Keeps assigned */
1774                 if (! na->projkeeps)
1775                         continue;
1776
1777                 /* we can run only once over the queue, so preserve the nodes */
1778                 pk = NEW_ARR_F(ir_node *, 0);
1779                 while (! waitq_empty(na->projkeeps))
1780                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1781                 del_waitq(na->projkeeps);
1782                 na->projkeeps = NULL;
1783
1784                 /* for all time steps at which this node can be scheduled */
1785                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1786                         int cst, tp_idx, i;
1787                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1788
1789                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1790                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1791                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1792                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1793                         num_cst++;
1794
1795                         /* sum up scheduling variables for this time step */
1796                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1797                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1798                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1799                         }
1800
1801                         if (ARR_LEN(tmp_var_idx_n) > 0)
1802                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1803                         DEL_ARR_F(tmp_var_idx_n);
1804
1805                         /* subtract all Proj and Keep variables for this step */
1806                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1807                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1808                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1809                                 int                  pk_tp_idx;
1810
1811                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1812                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1813                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1814                                 }
1815                         }
1816                 }
1817         }
1818         stat_ev_tim_pop("beilpsched_cst_projkeep");
1819 }
1820 #endif /* if 0 */
1821
1822 /***************************************************
1823  *  _____ _      _____                    _
1824  * |_   _| |    |  __ \                  (_)
1825  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1826  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1827  *  _| |_| |____| |      | | | | | | (_| | | | | |
1828  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1829  *
1830  ***************************************************/
1831
1832 /**
1833  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1834  */
1835 static void create_ilp(ir_node *block, void *walk_env)
1836 {
1837         be_ilpsched_env_t     *env           = walk_env;
1838         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1839         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1840         FILE                  *logfile       = NULL;
1841         lpp_t                 *lpp           = NULL;
1842         int                   need_heur      = 0;
1843         struct obstack        var_obst;
1844         char                  name[1024];
1845
1846         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1847         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1848         DBG((env->dbg, 255, "=========================================\n\n"));
1849
1850         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1851                 block, ba->n_interesting_nodes, ba->max_steps));
1852
1853         /* notify backend and get block environment */
1854         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1855
1856         /* if we have less than two interesting nodes, there is no need to create the ILP */
1857         if (ba->n_interesting_nodes > 1) {
1858                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1859                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1860                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1861                 int    estimated_n_var = (int)((double)base_num * fact_var);
1862                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1863
1864                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1865                         estimated_n_var, estimated_n_cst));
1866                 (void) estimated_n_var;
1867
1868                 /* set up the LPP object */
1869                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1870
1871                 lpp = new_lpp_userdef(
1872                         (const char *)name,
1873                         lpp_minimize,
1874                         estimated_n_cst,     /* num vars */
1875                         estimated_n_cst + 1, /* num cst */
1876                         1.3);                /* grow factor */
1877                 obstack_init(&var_obst);
1878
1879                 /* create ILP variables */
1880                 create_variables(env, lpp, block_node, &var_obst);
1881
1882                 /* create ILP constraints */
1883                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1884                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1885                 create_ressource_constraints(env, lpp, block_node);
1886                 create_bundle_constraints(env, lpp, block_node);
1887                 create_branch_constraint(env, lpp, block_node);
1888                 //create_proj_keep_constraints(env, lpp, block_node);
1889
1890                 if (env->opts->regpress) {
1891                         create_alive_nodes_constraint(env, lpp, block_node);
1892                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1893                         create_pressure_alive_constraint(env, lpp, block_node);
1894                 }
1895
1896                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1897
1898                 /* debug stuff, dump lpp when debugging is on  */
1899                 DEBUG_ONLY({
1900                         if (firm_dbg_get_mask(env->dbg) > 1) {
1901                                 char buf[1024];
1902                                 FILE *f;
1903
1904                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1905                                 f = fopen(buf, "w");
1906                                 lpp_dump_plain(lpp, f);
1907                                 fclose(f);
1908                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1909                                 lpp_dump(lpp, buf);
1910                         }
1911                 })
1912
1913                 /* set solve time limit */
1914                 lpp_set_time_limit(lpp, env->opts->time_limit);
1915
1916                 /* set logfile if requested */
1917                 if (strlen(env->opts->log_file) > 0) {
1918                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1919                                 lpp_set_log(lpp, stdout);
1920                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1921                                 lpp_set_log(lpp, stderr);
1922                         else {
1923                                 logfile = fopen(env->opts->log_file, "w");
1924                                 if (! logfile)
1925                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1926                                 else
1927                                         lpp_set_log(lpp, logfile);
1928                         }
1929                 }
1930
1931                 /* solve the ILP */
1932                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1933
1934                 if (logfile)
1935                         fclose(logfile);
1936
1937                 /* check for valid solution */
1938                 if (! lpp_is_sol_valid(lpp)) {
1939                         DEBUG_ONLY({
1940                             char buf[1024];
1941                             FILE *f;
1942
1943                             if (firm_dbg_get_mask(env->dbg) >= 2) {
1944                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1945                               f = fopen(buf, "w");
1946                               lpp_dump_plain(lpp, f);
1947                               fclose(f);
1948                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1949                               lpp_dump(lpp, buf);
1950                               dump_ir_block_graph(env->irg, "-infeasible");
1951                            }
1952                         })
1953
1954                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1955                         need_heur = 1;
1956                 }
1957
1958                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1959                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1960                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1961                 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));
1962                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1963                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1964                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1965                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1966
1967                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1968         }
1969
1970         /* apply solution */
1971         be_stat_ev("nodes", ba->block_last_idx);
1972         be_stat_ev("vars", lpp ? lpp->var_next : 0);
1973         be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1974         if (need_heur) {
1975                         be_stat_ev("time", -1);
1976                         be_stat_ev_dbl("opt", 0.0);
1977                 list_sched_single_block(env->birg, block, env->be_opts);
1978         }
1979         else {
1980                 if (lpp) {
1981                         double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1982                         be_stat_ev_dbl("time", lpp->sol_time);
1983                         be_stat_ev_dbl("opt", opt);
1984                 }
1985                 else {
1986                         be_stat_ev_dbl("time", 0.0);
1987                         be_stat_ev_dbl("opt", 100.0);
1988                 }
1989                 apply_solution(env, lpp, block);
1990         }
1991
1992         if (lpp)
1993                 free_lpp(lpp);
1994
1995         /* notify backend */
1996         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1997 }
1998
1999 /**
2000  * Perform ILP scheduling on the given irg.
2001  */
2002 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts)
2003 {
2004         be_ilpsched_env_t          env;
2005         const char                 *name     = "be ilp scheduling";
2006         ir_graph                   *irg      = be_get_birg_irg(birg);
2007         const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
2008         const ilp_sched_selector_t *sel      = arch_env->impl->get_ilp_sched_selector(arch_env);
2009
2010         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2011
2012         stat_ev_ctx_push("ilpsched");
2013
2014 //      firm_dbg_set_mask(env.dbg, 1);
2015
2016         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2017         env.sel        = sel;
2018         env.irg        = irg;
2019         env.height     = heights_new(irg);
2020         env.main_env   = birg->main_env;
2021         env.arch_env   = arch_env;
2022         env.cpu        = arch_env_get_machine(arch_env);
2023         env.opts       = &ilp_opts;
2024         env.birg       = birg;
2025         env.be_opts    = be_opts;
2026         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2027
2028         /* assign a unique per block number to all interesting nodes */
2029         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2030
2031         /*
2032                 The block indices are completely build after the walk,
2033                 now we can allocate the bitsets (size depends on block indices)
2034                 for all nodes.
2035         */
2036         phase_reinit_irn_data(&env.ph);
2037
2038         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2039         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2040
2041         /* Calculate ALAP of all irns */
2042         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2043
2044         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2045         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2046
2047         /* perform ILP scheduling */
2048         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2049
2050         DEBUG_ONLY(
2051                 if (firm_dbg_get_mask(env.dbg)) {
2052                         phase_stat_t stat;
2053                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2054
2055                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2056                 }
2057         );
2058
2059         /* free data allocated dynamically */
2060         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2061
2062         /* free all allocated object */
2063         phase_free(&env.ph);
2064         heights_free(env.height);
2065
2066         /* notify backend */
2067         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2068
2069         stat_ev_ctx_pop("ilpsched");
2070 }
2071
2072 /**
2073  * Register ILP scheduler options.
2074  */
2075 void be_init_ilpsched(void)
2076 {
2077         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2078         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2079
2080         lc_opt_add_table(sched_grp, ilpsched_option_table);
2081 }
2082
2083 #else /* WITH_ILP */
2084
2085 static inline void some_picky_compiler_do_not_allow_empty_files(void)
2086 {}
2087
2088 #endif /* WITH_ILP */