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

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