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