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