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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       ILP based instruction scheduling.
  23  * @author      Christian Wuerdig
  24  * @date        22.10.2006
  25  * @version     $Id$
  26  *
  27  * An ILP scheduler based on
  28  * "ILP-based Instruction Scheduling for IA-64"
  29  * by Daniel Kaestner and Sebastian Winkel
  30  * extended with register pressure constraints by Christian Wuerdig
  31  */
  32 #include "config.h"
  33
  34 #ifdef WITH_ILP
  35
  36 #include <math.h>
  37
  38 #ifndef _WIN32
  39 #include <strings.h>
  40 #endif /* _WIN32 */
  41
  42 #include "irnode_t.h"
  43 #include "irgwalk.h"
  44 #include "irbitset.h"
  45 #include "irphase_t.h"
  46 #include "height.h"
  47 #include "iredges.h"
  48 #include "pdeq.h"
  49 #include "debug.h"
  50 #include "irtools.h"
  51 #include "irdump.h"
  52 #include "irprintf.h"
  53 #include "plist.h"
  54 #include "irprintf.h"
  55 #include "timing.h"
  56
  57 #include <lpp/lpp.h>
  58 #include <lpp/lpp_net.h>
  59
  60 #include "lc_opts.h"
  61 #include "lc_opts_enum.h"
  62
  63 #include "be.h"
  64 #include "benode.h"
  65 #include "besched.h"
  66 #include "beilpsched.h"
  67 #include "beutil.h"
  68 #include "bestat.h"
  69 #include "beirg.h"
  70 #include "bemachine.h"
  71 #include "belistsched.h"
  72 #include "bemodule.h"
  73
  74 typedef struct _ilpsched_options_t {
  75         unsigned regpress;
  76         unsigned time_limit;
  77         char     log_file[1024];
  78 } ilpsched_options_t;
  79
  80 typedef struct _unit_type_info_t {
  81         int                            n_units;
  82         const be_execution_unit_type_t *tp;
  83 } unit_type_info_t;
  84
  85 /**
  86  * holding the ILP variables of the different types
  87  */
  88 typedef struct _ilp_var_types_t {
  89         int *x;   /* x_{nt}^k variables */
  90         int *a;   /* a_{nt}^k variables */
  91         int *y;   /* y_{nt}^k variables */
  92 } ilp_var_types_t;
  93
  94 /**
  95  * Holds alive variables for a node live-in to a block.
  96  */
  97 typedef struct _ilp_livein_node_t {
  98         ir_node  *irn;
  99         unsigned max_alive_steps;
 100         int      *a;
 101 } ilp_livein_node_t;
 102
 103 /* attributes for a node */
 104 typedef struct _ilpsched_node_attr_t {
 105         unsigned asap;                     /**< The ASAP scheduling control step */
 106         unsigned alap;                     /**< The ALAP scheduling control step */
 107         unsigned latency;                  /**< Latency of this node (needed for sorting) */
 108         unsigned sched_point;              /**< the step in which the node is finally scheduled */
 109         unsigned visit_idx;                /**< Index of the node having visited this node last */
 110         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
 111         unsigned n_consumer;               /**< Number of consumers */
 112         ir_node  **block_consumer;         /**< List of consumer being in the same block */
 113         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
 114         unsigned block_idx     : 30;       /**< A unique per block index */
 115         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
 116         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
 117         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
 118                                                                                         for ASAP, successors for ALAP */
 119         unsigned         n_unit_types;     /**< number of allowed execution unit types */
 120         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
 121         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
 122 } ilpsched_node_attr_t;
 123
 124 /* attributes for a block */
 125 typedef struct _ilpsched_block_attr_t {
 126         unsigned block_last_idx;        /**< The highest node index in block so far */
 127         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
 128         unsigned max_steps;             /**< Upper bound for block execution */
 129         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
 130         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
 131         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 132 } ilpsched_block_attr_t;
 133
 134 typedef union _ilpsched_attr_ {
 135         ilpsched_node_attr_t  node_attr;
 136         ilpsched_block_attr_t block_attr;
 137 } ilpsched_attr_t;
 138
 139 /* A irn for the phase and it's attributes (either node or block) */
 140 typedef struct {
 141         const ir_node   *irn;
 142         ilpsched_attr_t attr;
 143 } be_ilpsched_irn_t;
 144
 145 /* The ILP scheduling environment */
 146 typedef struct {
 147         ir_phase             ph;            /**< The phase */
 148         ir_graph             *irg;          /**< The current irg */
 149         heights_t            *height;       /**< The heights object of the irg */
 150         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
 151         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
 152         const arch_env_t     *arch_env;
 153         const be_main_env_t  *main_env;
 154         const be_machine_t   *cpu;          /**< the current abstract machine */
 155         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
 156         const be_irg_t       *birg;         /**< The birg object */
 157         be_options_t         *be_opts;      /**< backend options */
 158         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
 159         DEBUG_ONLY(firm_dbg_module_t *dbg);
 160 } be_ilpsched_env_t;
 161
 162 /* convenience macros to handle phase irn data */
 163 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
 164 #define is_ilpsched_block(node)             (is_Block((node)->irn))
 165 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 166 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 167
 168 /* check if node is considered for ILP scheduling */
 169 #define consider_for_sched(env, irn) \
 170         (! (is_Block(irn)            ||  \
 171                 is_normal_Proj(env, irn) ||  \
 172                 is_Phi(irn)              ||  \
 173                 is_NoMem(irn)            ||  \
 174                 is_Unknown(irn)          ||  \
 175                 is_End(irn)                  \
 176                 ))
 177
 178 /* gives the valid scheduling time step interval for a node */
 179 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
 180
 181 /* gives the valid interval where a node can die */
 182 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
 183
 184 /* gives the corresponding ILP variable for given node, unit and time step */
 185 #define ILPVAR_IDX(na, unit, control_step) \
 186         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
 187
 188 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
 189 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
 190         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
 191
 192 /* check if a double value is within an epsilon environment of 0 */
 193 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 194
 195 /* option variable */
 196 static ilpsched_options_t ilp_opts = {
 197         1,     /* default is with register pressure constraints */
 198         300,   /* 300 sec per block time limit */
 199         ""     /* no log file */
 200 };
 201
 202 /* ILP options */
 203 static const lc_opt_table_entry_t ilpsched_option_table[] = {
 204         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
 205         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
 206         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
 207         LC_OPT_LAST
 208 };
 209
 210 /*
 211         We need this global variable as we compare nodes dependent on heights,
 212         but we cannot pass any information to the qsort compare function.
 213 */
 214 static heights_t *glob_heights;
 215
 216 /**
 217  * Check if irn is a Proj, which has no execution units assigned.
 218  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 219  */
 220 static inline int is_normal_Proj(const arch_env_t *env, const ir_node *irn)
 221 {
 222         return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
 223 }
 224
 225 /**
 226  * Skips normal Projs.
 227  * @return predecessor if irn is a normal Proj, otherwise irn.
 228  */
 229 static inline ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn)
 230 {
 231         if (is_normal_Proj(env, irn))
 232                 return get_Proj_pred(irn);
 233         return irn;
 234 }
 235
 236 static inline int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env)
 237 {
 238         unsigned lat = be_ilp_sched_latency(sel, irn, env);
 239         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
 240                 lat = 1;
 241         return lat;
 242 }
 243
 244 static int cmp_live_in_nodes(const void *a, const void *b)
 245 {
 246         const ilp_livein_node_t *n1 = a;
 247         const ilp_livein_node_t *n2 = b;
 248
 249         return n1->irn != n2->irn;
 250 }
 251
 252 /**
 253  * Compare scheduling time steps of two be_ilpsched_irn's.
 254  */
 255 static int cmp_ilpsched_irn(const void *a, const void *b)
 256 {
 257         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
 258         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
 259         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
 260         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
 261
 262         if (n1_a->sched_point == n2_a->sched_point) {
 263                 const ir_node *irn_a = n1->irn;
 264                 const ir_node *irn_b = n2->irn;
 265
 266                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
 267                         return 1;
 268                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
 269                         return -1;
 270
 271                 /*
 272                         Ok, timestep is equal and the nodes are parallel,
 273                         so check latency and schedule high latency first.
 274                 */
 275                 return QSORT_CMP(n2_a->latency, n1_a->latency);
 276         }
 277         else
 278                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
 279 }
 280
 281 /**
 282  * In case there is no phase information for irn, initialize it.
 283  */
 284 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old)
 285 {
 286         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 287
 288         if (res == old) {
 289                 /* if we have already some data: check for reinitialization */
 290
 291                 if (! is_Block(irn)) {
 292                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 293
 294                         if (! na->transitive_block_nodes) {
 295                                 ir_node               *block      = get_nodes_block(irn);
 296                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
 297                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 298
 299                                 /* we are called after the block indices have been build: create bitset */
 300                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
 301                         }
 302                         else {
 303                                 /* we are called from reinit block data: clear the bitset */
 304                                 bitset_clear_all(na->transitive_block_nodes);
 305                                 na->visit_idx    = 0;
 306                                 na->alap_changed = 1;
 307                         }
 308                 }
 309                 return old;
 310         }
 311
 312         res->irn = irn;
 313
 314         /* set ilpsched irn attributes (either block or irn) */
 315         if (is_Block(irn)) {
 316                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
 317
 318                 ba->n_interesting_nodes = 0;
 319                 ba->block_last_idx      = 0;
 320                 ba->root_nodes          = plist_new();
 321                 ba->head_ilp_nodes      = NULL;
 322                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
 323                 ba->max_steps           = 0;
 324         }
 325         else {
 326                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
 327                 memset(na, 0, sizeof(*na));
 328         }
 329
 330         return res;
 331 }
 332
 333 /**
 334  * Assign a per block unique number to each node.
 335  */
 336 static void build_block_idx(ir_node *irn, void *walk_env)
 337 {
 338         be_ilpsched_env_t     *env = walk_env;
 339         be_ilpsched_irn_t     *node, *block_node;
 340         ilpsched_node_attr_t  *na;
 341         ilpsched_block_attr_t *ba;
 342
 343         set_irn_link(irn, NULL);
 344         if (! consider_for_sched(env->arch_env, irn))
 345                 return;
 346
 347         node       = get_ilpsched_irn(env, irn);
 348         na         = get_ilpsched_node_attr(node);
 349         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
 350         ba         = get_ilpsched_block_attr(block_node);
 351
 352         na->block_idx = ba->block_last_idx++;
 353 }
 354
 355 /********************************************************
 356  *                              __        _
 357  *                             / /       | |
 358  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 359  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 360  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 361  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 362  *                 | |                           | |
 363  *                 |_|                           |_|
 364  ********************************************************/
 365
 366 /**
 367  * Add all nodes having no user in current block to last_nodes list.
 368  */
 369 static void collect_alap_root_nodes(ir_node *irn, void *walk_env)
 370 {
 371         ir_node               *block;
 372         const ir_edge_t       *edge;
 373         be_ilpsched_irn_t     *block_node, *node;
 374         ilpsched_block_attr_t *ba;
 375         ilpsched_node_attr_t  *na;
 376         int                   i, j;
 377         be_ilpsched_env_t     *env           = walk_env;
 378         int                   has_block_user = 0;
 379         unsigned              n_consumer     = 0;
 380         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
 381         ir_node               **consumer;
 382         unsigned              idx;
 383
 384         if (! consider_for_sched(env->arch_env, irn))
 385                 return;
 386
 387         block    = get_nodes_block(irn);
 388     idx      = get_irn_idx(irn);
 389         consumer = NEW_ARR_F(ir_node *, 0);
 390
 391         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
 392
 393         /* check data and dependency out edges */
 394         for (i = 0; i < 2 && ! has_block_user; ++i) {
 395                 foreach_out_edge_kind(irn, edge, ekind[i]) {
 396                         ir_node *user = get_edge_src_irn(edge);
 397
 398                         if (is_normal_Proj(env->arch_env, user)) {
 399                                 const ir_edge_t *user_edge;
 400
 401                                 if (get_irn_mode(user) == mode_X)
 402                                         continue;
 403
 404                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
 405                                 for (j = 0; j < 2; ++j) {
 406                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
 407                                                 ir_node *real_user = get_edge_src_irn(user_edge);
 408
 409                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
 410                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
 411                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 412
 413                                                         /* skip already visited nodes */
 414                                                         if (ua->consumer_idx == idx)
 415                                                                 continue;
 416
 417                                                         /* check if node has user in this block and collect the user if it's a data user */
 418                                                         if (get_nodes_block(real_user) == block) {
 419                                                                 if (i == 0 && j == 0)
 420                                                                         ARR_APP1(ir_node *, consumer, real_user);
 421                                                                 has_block_user = 1;
 422                                                         }
 423
 424                                                         /* only count data consumer */
 425                                                         if (i == 0)
 426                                                                 n_consumer++;
 427
 428                                                         /* mark user as visited by this node */
 429                                                         ua->consumer_idx = idx;
 430                                                 }
 431                                         }
 432                                 }
 433                         }
 434                         else if (is_Block(user)) {
 435                                 continue;
 436                         }
 437                         else if (! is_Phi(user)) {
 438                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 439                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
 440
 441                                 /* skip already visited nodes */
 442                                 if (ua->consumer_idx == idx)
 443                                         continue;
 444
 445                                 /* check if node has user in this block and collect the user if it's a data user */
 446                                 if (get_nodes_block(user) == block) {
 447                                         if (i == 0)
 448                                                 ARR_APP1(ir_node *, consumer, user);
 449                                         has_block_user = 1;
 450                                 }
 451
 452                                 /* only count data consumer */
 453                                 if (i == 0)
 454                                         n_consumer++;
 455
 456                                 /* mark user visited by this node */
 457                                 ua->consumer_idx = idx;
 458                         }
 459                         else if (get_nodes_block(user) != block) {
 460                                 n_consumer++;
 461                         }
 462                 }
 463         }
 464
 465         block_node = get_ilpsched_irn(env, block);
 466         ba         = get_ilpsched_block_attr(block_node);
 467
 468         ba->n_interesting_nodes++;
 469
 470         /* current irn has no user inside this block, add to queue */
 471         if (! has_block_user) {
 472                 DB((env->dbg, LEVEL_3, "root node\n"));
 473                 plist_insert_back(ba->root_nodes, irn);
 474         }
 475         else {
 476                 DB((env->dbg, LEVEL_3, "normal node\n"));
 477         }
 478
 479         /* record number of all consumer and the consumer within the same block */
 480         node = get_ilpsched_irn(env, irn);
 481         na   = get_ilpsched_node_attr(node);
 482         na->n_consumer     = n_consumer;
 483         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
 484         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
 485         DEL_ARR_F(consumer);
 486 }
 487
 488 /**
 489  * Calculate the ASAP scheduling step for current irn.
 490  */
 491 static void calculate_irn_asap(ir_node *irn, void *walk_env)
 492 {
 493         be_ilpsched_env_t     *env = walk_env;
 494         int                   i;
 495         ir_node               *block;
 496         be_ilpsched_irn_t     *node, *block_node;
 497         ilpsched_node_attr_t  *na;
 498         ilpsched_block_attr_t *ba;
 499
 500         /* These nodes are handled separate */
 501         if (! consider_for_sched(env->arch_env, irn))
 502                 return;
 503
 504         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
 505
 506         block    = get_nodes_block(irn);
 507         node     = get_ilpsched_irn(env, irn);
 508         na       = get_ilpsched_node_attr(node);
 509         na->asap = 1;
 510
 511         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
 512                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
 513
 514                 /* check for greatest distance to top */
 515                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 516                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 517                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 518                         unsigned             lat;
 519
 520                         lat         = fixed_latency(env->sel, pred, env->block_env);
 521                         na->latency = lat;
 522                         na->asap    = MAX(na->asap, pna->asap + lat);
 523                 }
 524         }
 525
 526         /* add node to ILP node list and update max_steps */
 527         block_node = get_ilpsched_irn(env, block);
 528         ba         = get_ilpsched_block_attr(block_node);
 529
 530         set_irn_link(irn, ba->head_ilp_nodes);
 531         ba->head_ilp_nodes = irn;
 532         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
 533
 534         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
 535 }
 536
 537 /**
 538  * Calculate the ALAP scheduling step of all irns in current block.
 539  * Depends on max_steps being calculated.
 540  */
 541 static void calculate_block_alap(ir_node *block, void *walk_env)
 542 {
 543         be_ilpsched_env_t     *env        = walk_env;
 544         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 545         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 546         waitq                 *cur_queue  = new_waitq();
 547         plist_element_t       *el;
 548
 549         assert(is_Block(block));
 550
 551         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
 552                 block, ba->n_interesting_nodes, ba->max_steps));
 553
 554         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
 555         //phase_reinit_block_irn_data(&env->ph, block);
 556
 557         /* init start queue */
 558         foreach_plist(ba->root_nodes, el) {
 559                 waitq_put(cur_queue, plist_element_get_value(el));
 560         }
 561
 562         /* repeat until all nodes are processed */
 563         while (! waitq_empty(cur_queue)) {
 564                 waitq *next_queue = new_waitq();
 565
 566                 /* process all nodes in current step */
 567                 while (! waitq_empty(cur_queue)) {
 568                         ir_node              *cur_irn = waitq_get(cur_queue);
 569                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
 570                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
 571                         int                  i;
 572
 573                         /* cur_node has no alap set -> it's a root node, set to max alap */
 574                         if (na->alap == 0) {
 575                                 na->alap = ba->max_steps;
 576                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
 577                                         cur_irn, na->alap));
 578                         }
 579                         else {
 580                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
 581                                         cur_irn, na->alap));
 582                         }
 583
 584                         /* set the alap's of all predecessors */
 585                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
 586                                 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
 587
 588                                 /* check for greatest distance to bottom */
 589                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
 590                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
 591                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
 592                                         unsigned             lat;
 593
 594                                         /* mark the predecessor as visited by current irn */
 595                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
 596                                                 continue;
 597                                         pna->visit_idx = get_irn_idx(cur_irn);
 598
 599                                         lat = fixed_latency(env->sel, pred, env->block_env);
 600
 601                                         /* set ALAP of current pred */
 602                                         if (pna->alap == 0) {
 603                                                 /* current ALAP is 0: node has not yet been visited */
 604                                                 pna->alap_changed = 1;
 605                                                 pna->alap         = na->alap - lat;
 606                                         }
 607                                         else if (pna->alap > na->alap - lat) {
 608                                                 /* we found a longer path to root node: change ALAP */
 609                                                 pna->alap         = na->alap - lat;
 610                                                 pna->alap_changed = 1;
 611                                         }
 612                                         else {
 613                                                 /* current ALAP is best found so far: keep it */
 614                                                 pna->alap_changed = 0;
 615                                         }
 616
 617                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
 618
 619                                         /* enqueue node for next iteration */
 620                                         if (get_irn_ins_or_deps(pred) > 0)
 621                                                 waitq_put(next_queue, pred);
 622                                 }
 623                         }
 624                 }
 625
 626                 /* prepare for next iteration */
 627                 del_waitq(cur_queue);
 628                 cur_queue = next_queue;
 629         }
 630 }
 631
 632 /**
 633  * Free list of root nodes and the set of live-in nodes.
 634  */
 635 static void clear_unwanted_data(ir_node *block, void *walk_env)
 636 {
 637         be_ilpsched_env_t     *env        = walk_env;
 638         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 639         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 640
 641         plist_free(ba->root_nodes);
 642         ba->root_nodes = NULL;
 643         del_pset(ba->livein_nodes);
 644         ba->livein_nodes = NULL;
 645 }
 646
 647 /**
 648  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 649  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 650  */
 651 static void refine_asap_alap_times(ir_node *irn, void *walk_env)
 652 {
 653         be_ilpsched_env_t    *env  = walk_env;
 654         ir_node              *pred = irn;
 655         be_ilpsched_irn_t    *node, *pred_node;
 656         ilpsched_node_attr_t *na, *pna;
 657
 658         if (! consider_for_sched(env->arch_env, irn))
 659                 return;
 660
 661         if (! is_Proj(irn) && ! be_is_Keep(irn))
 662                 return;
 663
 664         /* go to the ancestor */
 665         if (be_is_Keep(irn))
 666                 pred = get_irn_n(irn, 0);
 667         pred = skip_Proj(pred);
 668
 669         node      = get_ilpsched_irn(env, irn);
 670         pred_node = get_ilpsched_irn(env, pred);
 671         na        = get_ilpsched_node_attr(node);
 672         pna       = get_ilpsched_node_attr(pred_node);
 673
 674         na->asap = pna->asap;
 675         na->alap = pna->alap;
 676
 677         /* record all Projs and Keeps for this node */
 678         if (! pna->projkeeps)
 679                 pna->projkeeps = new_waitq();
 680         waitq_put(pna->projkeeps, irn);
 681
 682         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
 683 }
 684
 685 /*******************************************
 686  *           _              _       _
 687  *          | |            | |     | |
 688  *  ___  ___| |__   ___  __| |_   _| | ___
 689  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 690  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 691  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 692  *
 693  *******************************************/
 694
 695 static inline void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn)
 696 {
 697         const ir_edge_t *edge;
 698         (void) block;
 699
 700         foreach_out_edge(irn, edge) {
 701                 ir_node *user = get_edge_src_irn(edge);
 702
 703                 if (be_is_Keep(user)) {
 704                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
 705                         waitq_put(keeps, user);
 706                 }
 707         }
 708 }
 709
 710 /**
 711  * Inserts @p irn before @p before into schedule and notifies backend.
 712  */
 713 static inline void notified_sched_add_before(be_ilpsched_env_t *env,
 714         const ir_node *before, const ir_node *irn, unsigned cycle)
 715 {
 716         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
 717         sched_add_before((ir_node*) before, (ir_node*) irn);
 718 }
 719
 720 /**
 721  * Adds a node, it's Projs (in case of mode_T nodes) and
 722  * it's Keeps to schedule.
 723  */
 724 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle)
 725 {
 726         const ir_edge_t *edge;
 727         waitq           *keeps = new_waitq();
 728
 729         /* mode_M nodes are not scheduled */
 730         if (get_irn_mode(irn) == mode_M)
 731                 return;
 732
 733         if (! sched_is_scheduled(irn))
 734                 notified_sched_add_before(env, block, irn, cycle);
 735
 736         /* add Projs */
 737         if (get_irn_mode(irn) == mode_T) {
 738                 foreach_out_edge(irn, edge) {
 739                         ir_node *user = get_edge_src_irn(edge);
 740
 741                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
 742                                 get_irn_n_edges(user) > 0)
 743                         {
 744                                 notified_sched_add_before(env, block, user, cycle);
 745                         }
 746
 747                         check_for_keeps(keeps, block, user);
 748                 }
 749         }
 750         else {
 751                 check_for_keeps(keeps, block, irn);
 752         }
 753
 754         /* add Keeps */
 755         while (! waitq_empty(keeps)) {
 756                 ir_node *keep = waitq_get(keeps);
 757                 if (! sched_is_scheduled(keep))
 758                         notified_sched_add_before(env, block, keep, cycle);
 759         }
 760
 761         del_waitq(keeps);
 762 }
 763
 764 /**
 765  * Schedule all nodes in the given block, according to the ILP solution.
 766  */
 767 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block)
 768 {
 769         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
 770         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
 771         be_ilpsched_irn_t     **sched_nodes;
 772         unsigned              i, l;
 773         ir_node               *cfop, *irn;
 774         const ir_edge_t       *edge;
 775
 776         /* init block schedule list */
 777         sched_init_block(block);
 778
 779         /* collect nodes and their scheduling time step */
 780         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
 781         if (ba->n_interesting_nodes == 0) {
 782                 /* ignore */
 783         }
 784         else if (ba->n_interesting_nodes == 1) {
 785                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
 786
 787                 /* add the single node */
 788                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 789         }
 790         else {
 791                 /* check all nodes for their positive solution */
 792                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
 793                         be_ilpsched_irn_t    *node;
 794                         ilpsched_node_attr_t *na;
 795                         int                  tp_idx, found;
 796                         unsigned             cur_var, t;
 797
 798                         node    = get_ilpsched_irn(env, irn);
 799                         na      = get_ilpsched_node_attr(node);
 800                         cur_var = 0;
 801                         found   = 0;
 802
 803                         if (! na->is_dummy_node) {
 804                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 805                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 806                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
 807
 808                                                 if (! LPP_VALUE_IS_0(cost)) {
 809                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
 810                                                         found = 1;
 811                                                 }
 812                                         }
 813                                 }
 814                         }
 815
 816                         found = 0;
 817                         /* go over all variables of a node until the non-zero one is found */
 818                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
 819                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
 820                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
 821
 822                                         /* check, if variable is set to one (it's not zero then :) */
 823                                         if (! LPP_VALUE_IS_0(val)) {
 824                                                 na->sched_point = t;
 825                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
 826                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
 827                                                         irn, t, na->type_info[tp_idx].tp->name));
 828                                                 found = 1;
 829                                         }
 830                                 }
 831                         }
 832                 }
 833
 834                 glob_heights = env->height;
 835                 /* sort nodes ascending by scheduling time step */
 836                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
 837         }
 838
 839         /* make all Phis ready and remember the single cf op */
 840         cfop = NULL;
 841         foreach_out_edge(block, edge) {
 842                 irn = get_edge_src_irn(edge);
 843
 844                 switch (get_irn_opcode(irn)) {
 845                         case iro_Phi:
 846                                 add_to_sched(env, block, irn, 0);
 847                                 break;
 848                         case beo_Start:
 849                         case iro_End:
 850                         case iro_Proj:
 851                         case iro_Bad:
 852                         case iro_Unknown:
 853                                 break;
 854                         default:
 855                                 if (is_cfop(irn)) {
 856                                         assert(cfop == NULL && "Highlander - there can be only one");
 857                                         cfop = irn;
 858                                 }
 859                         break;
 860                 }
 861         }
 862
 863         /* add all nodes from list */
 864         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
 865                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
 866                 if (sched_nodes[i]->irn != cfop)
 867                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
 868         }
 869
 870         /* schedule control flow node if not already done */
 871         if (cfop && ! sched_is_scheduled(cfop))
 872                 add_to_sched(env, block, cfop, 0);
 873
 874         DEL_ARR_F(sched_nodes);
 875 }
 876
 877 /***************************************************************
 878  *   _____ _      _____     _____           _   _
 879  *  |_   _| |    |  __ \   / ____|         | | (_)
 880  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 881  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 882  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 883  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 884  *
 885  ***************************************************************/
 886
 887 /**
 888  * Check if node can be executed on given unit type.
 889  */
 890 static inline int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node)
 891 {
 892         int                  i;
 893         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
 894
 895         for (i = na->n_unit_types - 1; i >= 0; --i) {
 896                 if (na->type_info[i].tp == tp)
 897                         return i;
 898         }
 899
 900         return -1;
 901 }
 902
 903 /************************************************
 904  *                   _       _     _
 905  *                  (_)     | |   | |
 906  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 907  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 908  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 909  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 910  *
 911  ************************************************/
 912
 913 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst)
 914 {
 915         const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
 916         unsigned                  n_unit_types = 0;
 917         be_ilpsched_irn_t         *node;
 918         ilpsched_node_attr_t      *na;
 919         unsigned                  unit_idx, tp_idx;
 920
 921         /* count number of available unit types for this node */
 922         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
 923                 /* just count */ ;
 924
 925         node = get_ilpsched_irn(env, irn);
 926         na   = get_ilpsched_node_attr(node);
 927
 928         if (! na->type_info) {
 929                 na->n_unit_types = n_unit_types;
 930                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
 931
 932                 /* fill the type info array */
 933                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
 934                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
 935                                 /* beware: we also count number of available units here */
 936                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
 937                                         na->is_dummy_node = 1;
 938                         }
 939
 940                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
 941                         na->type_info[tp_idx].n_units = unit_idx;
 942                 }
 943         }
 944
 945         return n_unit_types;
 946 }
 947
 948 /**
 949  * Returns the largest alap time of a user of @p irn.
 950  * The user must be in block @p block.
 951  */
 952 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block)
 953 {
 954         const ir_edge_t *edge;
 955         unsigned        max_alap = 0;
 956
 957         foreach_out_edge(irn, edge) {
 958                 ir_node *user = get_edge_src_irn(edge);
 959
 960                 if (get_nodes_block(user) == block) {
 961                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
 962                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
 963
 964                         max_alap = MAX(max_alap, na->alap);
 965                 }
 966         }
 967
 968         assert(max_alap > 0);
 969         return max_alap;
 970 }
 971
 972 /**
 973  * Create the following variables:
 974  * - x_{nt}^k    binary     weigthed with: t
 975  *      node n is scheduled at time step t to unit type k
 976  * ==>> These variables represent the schedule
 977  *
 978  * - a_{nt}^k    binary     weighted with num_nodes
 979  *      node n is alive at time step t on unit type k
 980  *
 981  * - y_{nt}^k    continuous  weighted with: num_nodes^2
 982  *      register pressure over limit for unit type k
 983  * ==>> These variables represent the register pressure
 984  *
 985  */
 986 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst)
 987 {
 988         char                  buf[1024];
 989         ir_node               *irn;
 990         unsigned              num_block_var, num_nodes;
 991         ilp_livein_node_t     *livein;
 992         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
 993         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
 994
 995         stat_ev_tim_push();
 996
 997         num_block_var = num_nodes = 0;
 998         foreach_linked_irns(ba->head_ilp_nodes, irn) {
 999                 be_ilpsched_irn_t    *node;
1000                 ilpsched_node_attr_t *na;
1001                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
1002                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
1003                 int                  i;
1004
1005                 node         = get_ilpsched_irn(env, irn);
1006                 na           = get_ilpsched_node_attr(node);
1007                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
1008
1009                 /* allocate space for ilp variables */
1010                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1011                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
1012
1013                 /* we need these variables only for "real" nodes */
1014                 if (! na->is_dummy_node) {
1015                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1016                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1017
1018                         num_ad         = ba->max_steps - na->asap + 1;
1019                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1020                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1021                 }
1022
1023                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1024                         irn, na->asap, na->alap, na->n_unit_types));
1025
1026                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1027                 /* create variables */
1028                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1029                         unsigned t;
1030
1031                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1032                                 /* x_{nt}^k variables */
1033                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1034                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1035                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1036                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1037                                 /* variable counter */
1038                                 n_var++;
1039                                 num_block_var++;
1040
1041                                 if (! na->is_dummy_node) {
1042                                         /* y_{nt}^k variables */
1043                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1044                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1045                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1046                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1047
1048                                         /* variable counter */
1049                                         n_var++;
1050                                         num_block_var++;
1051                                 }
1052                         }
1053
1054                         /* a node can die at any step t: asap(n) <= t <= U */
1055                         if (! na->is_dummy_node) {
1056                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1057
1058                                         /* a_{nt}^k variables */
1059                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1060                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1061                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1062                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1063
1064                                         /* variable counter */
1065                                         n_var++;
1066                                         num_block_var++;
1067                                 }
1068                         }
1069
1070                         /* collect live-in nodes */
1071                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1072                                 ir_node *pred = get_irn_n(irn, i);
1073
1074                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1075                                         be_ilpsched_set_type_info(env, pred, var_obst);
1076                                         if (! na->is_dummy_node) {
1077                                                 ilp_livein_node_t *entry = OALLOC(var_obst, ilp_livein_node_t);
1078                                                 entry->irn = pred;
1079                                                 entry->a   = NULL;
1080                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1081                                         }
1082                                 }
1083                         }
1084                 }
1085
1086                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1087                 num_nodes++;
1088         }
1089
1090         /* create alive variables a_{nt}^k for live-ins */
1091         foreach_pset(ba->livein_nodes, livein) {
1092                 be_ilpsched_irn_t    *node;
1093                 ilpsched_node_attr_t *na;
1094                 unsigned             tp_idx, var_idx;
1095                 ir_node              *irn;
1096
1097                 irn  = livein->irn;
1098                 node = get_ilpsched_irn(env, irn);
1099                 na   = get_ilpsched_node_attr(node);
1100
1101                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1102
1103                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1104                 var_idx   = 0;
1105
1106                 /* create variables */
1107                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1108                         unsigned t;
1109
1110                         for (t = 0; t < livein->max_alive_steps; ++t) {
1111                                 /* a_{nt}^k variables */
1112                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1113                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1114                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1115                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1116                                 num_block_var++;
1117                         }
1118                 }
1119         }
1120
1121         stat_ev_tim_pop("beilpsched_var");
1122 }
1123
1124 /*******************************************************
1125  *                      _             _       _
1126  *                     | |           (_)     | |
1127  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1128  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1129  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1130  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1131  *
1132  *******************************************************/
1133
1134 /**
1135  * Collect all operands and nodes @p irn depends on.
1136  * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1137  */
1138 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps)
1139 {
1140         int i;
1141
1142         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1143                 ir_node *p = get_irn_in_or_dep(irn, i);
1144
1145                 if (is_Proj(p)) {
1146                         const ir_edge_t *edge;
1147
1148                         p = get_Proj_pred(p);
1149                         foreach_out_edge(p, edge) {
1150                                 ir_node *src = get_edge_src_irn(edge);
1151                                 ir_nodeset_insert(deps, src);
1152                         }
1153                 }
1154                 else {
1155                         ir_nodeset_insert(deps, p);
1156                 }
1157         }
1158 }
1159
1160 /**
1161  * Create following ILP constraints:
1162  * - the assignment constraints:
1163  *     assure each node is executed once by exactly one (allowed) execution unit
1164  * - the dead node assignment constraints:
1165  *     assure a node can only die at most once
1166  * - the precedence constraints:
1167  *     assure that no data dependencies are violated
1168  */
1169 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1170 {
1171         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1172         char                  buf[1024];
1173         ir_node               *irn;
1174         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1175         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1176
1177         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1178         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1179                 int                  cst, tp_idx;
1180                 unsigned             cur_var;
1181                 be_ilpsched_irn_t    *node;
1182                 ilpsched_node_attr_t *na;
1183                 ir_node              *pred;
1184                 ir_nodeset_t          deps;
1185                 ir_nodeset_iterator_t iter;
1186
1187                 ir_nodeset_init(&deps);
1188
1189                 node    = get_ilpsched_irn(env, irn);
1190                 na      = get_ilpsched_node_attr(node);
1191                 cur_var = 0;
1192
1193                 /* the assignment constraint */
1194                 stat_ev_tim_push();
1195                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1196                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1197                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1198                 num_cst_assign++;
1199
1200                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1201                 stat_ev_tim_pop("beilpsched_cst_assign");
1202
1203                 /* We have separate constraints for Projs and Keeps */
1204                 // ILP becomes infeasible ?!?
1205 //              if (is_Proj(irn) || be_is_Keep(irn))
1206 //                      continue;
1207
1208                 /* the precedence constraints */
1209                 stat_ev_tim_push();
1210                 bitset_clear_all(bs_block_irns);
1211
1212                 sta_collect_in_deps(irn, &deps);
1213                 foreach_ir_nodeset(&deps, pred, iter) {
1214                         unsigned             t_low, t_high, t;
1215                         be_ilpsched_irn_t    *pred_node;
1216                         ilpsched_node_attr_t *pna;
1217                         unsigned             delay;
1218
1219                         pred = skip_normal_Proj(env->arch_env, pred);
1220                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1221                                 continue;
1222
1223                         pred_node = get_ilpsched_irn(env, pred);
1224                         pna       = get_ilpsched_node_attr(pred_node);
1225
1226                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1227
1228                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1229                                 bitset_set(bs_block_irns, pna->block_idx);
1230                         else
1231                                 continue;
1232
1233                         /* irn = n, pred = m */
1234                         delay  = fixed_latency(env->sel, pred, env->block_env);
1235                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1236                         t_high = MIN(na->alap, pna->alap + delay - 1);
1237                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1238                                 unsigned tn, tm;
1239                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1240
1241                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1242                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1243                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1244                                 num_cst_prec++;
1245
1246                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1247                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1248                                         /* node variables have smaller index than pred variables */
1249                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1250                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1251                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1252                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1253                                                 }
1254                                         }
1255
1256                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1257                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1258                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1259                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1260                                                 }
1261                                         }
1262                                 }
1263                                 else {
1264                                         /* pred variables have smaller index than node variables */
1265                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1266                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1267                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1268                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1269                                                 }
1270                                         }
1271
1272                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1273                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1274                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1275                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1276                                                 }
1277                                         }
1278                                 }
1279
1280                                 if (ARR_LEN(tmp_var_idx) > 0)
1281                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1282
1283                                 DEL_ARR_F(tmp_var_idx);
1284                         }
1285                 }
1286                 ir_nodeset_destroy(&deps);
1287                 stat_ev_tim_pop("beilpsched_cst_prec");
1288         }
1289 }
1290
1291 /**
1292  * Create ILP resource constraints:
1293  * - assure that for each time step not more instructions are scheduled
1294  *   to the same unit types as units of this type are available
1295  */
1296 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1297 {
1298         int                   glob_type_idx;
1299         char                  buf[1024];
1300         unsigned              num_cst_resrc = 0;
1301         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1302
1303         stat_ev_tim_push();
1304         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1305                 unsigned                 t;
1306                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1307
1308                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1309
1310                 /* check each time step */
1311                 for (t = 0; t < ba->max_steps; ++t) {
1312                         ir_node *irn;
1313                         int     cst;
1314                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1315
1316                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1317                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1318                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1319                         num_cst_resrc++;
1320
1321                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1322                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1323                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1324                                 int                  tp_idx;
1325
1326                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1327
1328                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1329                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1330                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1331                                 }
1332                         }
1333
1334                         /* set constraints if we have some */
1335                         if (ARR_LEN(tmp_var_idx) > 0)
1336                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1337
1338                         DEL_ARR_F(tmp_var_idx);
1339                 }
1340         }
1341         stat_ev_tim_pop("beilpsched_cst_rsrc");
1342 }
1343
1344 /**
1345  * Create ILP bundle constraints:
1346  * - assure, at most bundle_size * bundles_per_cycle instructions
1347  *   can be started at a certain point.
1348  */
1349 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1350 {
1351         char                  buf[1024];
1352         unsigned              t;
1353         unsigned              num_cst_bundle = 0;
1354         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1355         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1356
1357         stat_ev_tim_push();
1358         for (t = 0; t < ba->max_steps; ++t) {
1359                 ir_node *irn;
1360                 int     cst;
1361                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1362
1363                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1364                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1365                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1366                 num_cst_bundle++;
1367
1368                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1369                         be_ilpsched_irn_t    *node;
1370                         ilpsched_node_attr_t *na;
1371                         int                  tp_idx;
1372
1373                         /* Projs and Keeps do not contribute to bundle size */
1374                         if (is_Proj(irn) || be_is_Keep(irn))
1375                                 continue;
1376
1377                         node = get_ilpsched_irn(env, irn);
1378                         na   = get_ilpsched_node_attr(node);
1379
1380                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1381                         if (na->is_dummy_node)
1382                                 continue;
1383
1384                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1385                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1386                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1387                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1388                                 }
1389                         }
1390                 }
1391
1392                 if (ARR_LEN(tmp_var_idx) > 0)
1393                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1394
1395                 DEL_ARR_F(tmp_var_idx);
1396         }
1397         stat_ev_tim_pop("beilpsched_cst_bundle");
1398 }
1399
1400 /**
1401  * Create ILP alive nodes constraints:
1402  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1403  */
1404 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1405 {
1406         char                  buf[1024];
1407         ir_node               *irn;
1408         unsigned              num_cst = 0;
1409         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1410
1411         stat_ev_tim_push();
1412         /* for each node */
1413         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1414                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1415                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1416                 unsigned             t;
1417
1418                 /* we ignore nodes assigned to dummy unit here */
1419                 if (na->is_dummy_node)
1420                         continue;
1421
1422                 /* check check all time steps: asap(n) <= t <= U */
1423                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1424                         int node_tp_idx;
1425
1426                         /* for all unit types available for this node */
1427                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1428                                 unsigned tn, tn_max, idx;
1429                                 int      cst, i;
1430                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1431                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1432
1433                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1434                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1435                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1436                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1437                                 num_cst++;
1438
1439                                 tn_max = MIN(na->alap - 1, t);
1440                                 /* check if the node has been scheduled so far */
1441                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1442                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1443                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1444                                 }
1445
1446                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1447                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1448                                 DEL_ARR_F(tmp_var_idx_n);
1449
1450                                 /* subtract the number of consumer scheduled so far */
1451                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1452                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1453                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1454                                         int                  tp_idx;
1455                                         unsigned             tm, tm_max;
1456
1457                                         tm_max = MIN(ca->alap - 1, t);
1458                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1459                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1460                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1461                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1462                                                 }
1463                                         }
1464                                 }
1465
1466                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1467                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1468                                 DEL_ARR_F(tmp_var_idx_m);
1469
1470                                 /* -c * a_{nt}^k */
1471                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1472                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1473
1474                         }
1475                 }
1476         }
1477         stat_ev_tim_pop("beilpsched_cst_alive_nodes");
1478 }
1479
1480 /**
1481  * Create ILP alive nodes constraints for live-in nodes:
1482  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1483  */
1484 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1485 {
1486         char                  buf[1024];
1487         ilp_livein_node_t     *livein;
1488         unsigned              num_cst = 0;
1489         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1490
1491         stat_ev_tim_push();
1492         /* for each node */
1493         foreach_pset(ba->livein_nodes, livein) {
1494                 ir_node              *irn  = livein->irn;
1495                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1496                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1497                 unsigned             t;
1498
1499                 /* check check all time steps: 0 <= t < max_alive_steps */
1500                 for (t = 0; t < livein->max_alive_steps; ++t) {
1501                         int node_tp_idx;
1502
1503                         /* for all unit types available for this node */
1504                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1505                                 const ir_edge_t *edge;
1506                                 unsigned idx;
1507                                 int      cst, num_block_user;
1508                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1509
1510                                 /* check the number of consumer scheduled so far */
1511                                 num_block_user = 0;
1512                                 foreach_out_edge(irn, edge) {
1513                                         ir_node              *user = get_edge_src_irn(edge);
1514                                         be_ilpsched_irn_t    *cons;
1515                                         ilpsched_node_attr_t *ca;
1516                                         int                  tp_idx;
1517                                         unsigned             tm, tm_max;
1518
1519                                         /* check only users within current block */
1520                                         if (get_nodes_block(user) != block_node->irn)
1521                                                 continue;
1522
1523                                         num_block_user++;
1524                                         cons = get_ilpsched_irn(env, user);
1525                                         ca   = get_ilpsched_node_attr(cons);
1526
1527                                         tm_max = MIN(ca->alap - 1, t);
1528                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1529                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1530                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1531                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1532                                                 }
1533                                         }
1534                                 }
1535
1536                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1537                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1538                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1539                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1540                                 num_cst++;
1541
1542                                 /* sum(scheduled users) */
1543                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1544                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1545                                 DEL_ARR_F(tmp_var_idx_m);
1546
1547                                 /* + c * a_{nt}^k */
1548                                 idx = node_tp_idx * livein->max_alive_steps + t;
1549                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1550                         }
1551                 }
1552         }
1553         stat_ev_tim_pop("beilpsched_cst_alive_livein_nodes");
1554 }
1555
1556 /**
1557  * Create ILP pressure constraints, based on alive nodes:
1558  * - add additional costs to objective function if a node is scheduled
1559  *   on a unit although all units of this type are currently occupied
1560  */
1561 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1562 {
1563         char                  buf[1024];
1564         ir_node               *cur_irn;
1565         unsigned              num_cst = 0;
1566         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1567
1568         stat_ev_tim_push();
1569         /* y_{nt}^k is set for each node and timestep and unit type */
1570         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1571                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1572                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1573                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1574                 int                  glob_type_idx;
1575
1576                 /* we ignore nodes assigned to DUMMY unit here */
1577                 if (cur_na->is_dummy_node)
1578                         continue;
1579
1580                 /* for all types */
1581                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1582                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1583                         int                      cur_tp_idx;
1584                         unsigned                 t;
1585
1586                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1587
1588                         /* check if node can be executed on this unit type */
1589                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1590                         if (cur_tp_idx < 0)
1591                                 continue;
1592
1593                         /* check all time_steps at which the current node can be scheduled */
1594                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1595                                 int     cst, y_idx;
1596                                 ir_node *irn;
1597                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1598                                 ilp_livein_node_t *livein;
1599
1600                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1601                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1602                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1603                                 num_cst++;
1604
1605                                 /* - accumulate all nodes alive at point t on unit type k */
1606                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1607                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1608                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1609                                         int                  a_idx, tp_idx;
1610
1611                                         /* check if node can be alive here */
1612                                         if (t < na->asap - 1)
1613                                                 continue;
1614
1615                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1616
1617                                         /* current type is not suitable */
1618                                         if (tp_idx < 0)
1619                                                 continue;
1620
1621                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1622                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1623                                 }
1624                                 /* do the same for livein nodes */
1625                                 foreach_pset(ba->livein_nodes, livein) {
1626                                         ir_node              *irn  = livein->irn;
1627                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1628                                         int                  a_idx, tp_idx;
1629
1630                                         /* check if node can be alive here */
1631                                         if (t >= livein->max_alive_steps)
1632                                                 continue;
1633
1634                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1635
1636                                         /* current type is not suitable */
1637                                         if (tp_idx < 0)
1638                                                 continue;
1639
1640                                         a_idx = tp_idx * livein->max_alive_steps + t;
1641                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1642                                 }
1643
1644                                 if (ARR_LEN(tmp_var_idx) > 0)
1645                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1646                                 DEL_ARR_F(tmp_var_idx);
1647
1648                                 /* - num_nodes * y_{nt}^k */
1649                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1650                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1651                         }
1652                 }
1653         }
1654         stat_ev_tim_pop("beilpsched_cst_pressure");
1655 }
1656
1657 /**
1658  * Create ILP branch constraints:
1659  * Assure, alle nodes are scheduled prior to cfg op.
1660  */
1661 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1662 {
1663         char                  buf[1024];
1664         ir_node               *cur_irn, *cfop;
1665         unsigned              num_cst          = 0;
1666         unsigned              num_non_branches = 0;
1667         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1668
1669         stat_ev_tim_push();
1670         cfop = NULL;
1671         /* determine number of non-branch nodes and the one and only branch node */
1672         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1673                 switch (get_irn_opcode(cur_irn)) {
1674                         case iro_Phi:
1675                         case beo_Start:
1676                         case iro_End:
1677                         case iro_Proj:
1678                         case iro_Bad:
1679                         case iro_Unknown:
1680                                 num_non_branches++;
1681                                 break;
1682                         default:
1683                                 if (is_cfop(cur_irn)) {
1684                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1685                                         cfop = cur_irn;
1686                                 }
1687                                 else {
1688                                         num_non_branches++;
1689                                 }
1690                                 break;
1691                 }
1692         }
1693
1694         if (cfop) {
1695                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1696                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1697                 unsigned t;
1698
1699                 /* for each time step */
1700                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1701                         int *non_branch_vars, *branch_vars;
1702                         int cst;
1703
1704                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1705                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1706                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1707                         num_cst++;
1708
1709                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1710                         non_branch_vars = NEW_ARR_F(int, 0);
1711                         branch_vars     = NEW_ARR_F(int, 0);
1712                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1713                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1714                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1715                                 int                  tp_idx;
1716
1717                                 if (cur_irn == cfop) {
1718                                         /* for all unit types available for this node */
1719                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1720                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1721                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1722                                         }
1723                                 }
1724                                 else {
1725                                         /* sum up all possible schedule points for this node upto current timestep */
1726                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1727                                                 unsigned tn;
1728                                                 unsigned tmax = MIN(t, na->alap - 1);
1729
1730                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1731                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1732                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1733                                                 }
1734                                         }
1735                                 }
1736
1737                         }
1738
1739                         if (ARR_LEN(non_branch_vars) > 0)
1740                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1741                         if (ARR_LEN(branch_vars) > 0)
1742                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1743
1744                         DEL_ARR_F(branch_vars);
1745                         DEL_ARR_F(non_branch_vars);
1746                 }
1747         }
1748         stat_ev_tim_pop("beilpsched_cst_branch");
1749 }
1750
1751 #if 0
1752 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
1753 {
1754         char                  buf[1024];
1755         ir_node               *irn;
1756         unsigned              num_cst = 0;
1757         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1758
1759         stat_ev_tim_push();
1760         /* check all nodes */
1761         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1762                 be_ilpsched_irn_t    *node;
1763                 ilpsched_node_attr_t *na;
1764                 unsigned             t;
1765                 ir_node              **pk;
1766
1767                 /* only mode_T nodes can have Projs and Keeps assigned */
1768                 if (get_irn_mode(irn) != mode_T)
1769                         continue;
1770
1771                 node = get_ilpsched_irn(env, irn);
1772                 na   = get_ilpsched_node_attr(node);
1773
1774                 /* check if has some Projs and Keeps assigned */
1775                 if (! na->projkeeps)
1776                         continue;
1777
1778                 /* we can run only once over the queue, so preserve the nodes */
1779                 pk = NEW_ARR_F(ir_node *, 0);
1780                 while (! waitq_empty(na->projkeeps))
1781                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1782                 del_waitq(na->projkeeps);
1783                 na->projkeeps = NULL;
1784
1785                 /* for all time steps at which this node can be scheduled */
1786                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1787                         int cst, tp_idx, i;
1788                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1789
1790                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1791                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1792                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1793                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1794                         num_cst++;
1795
1796                         /* sum up scheduling variables for this time step */
1797                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1798                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1799                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1800                         }
1801
1802                         if (ARR_LEN(tmp_var_idx_n) > 0)
1803                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1804                         DEL_ARR_F(tmp_var_idx_n);
1805
1806                         /* subtract all Proj and Keep variables for this step */
1807                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1808                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1809                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1810                                 int                  pk_tp_idx;
1811
1812                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1813                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1814                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1815                                 }
1816                         }
1817                 }
1818         }
1819         stat_ev_tim_pop("beilpsched_cst_projkeep");
1820 }
1821 #endif /* if 0 */
1822
1823 /***************************************************
1824  *  _____ _      _____                    _
1825  * |_   _| |    |  __ \                  (_)
1826  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1827  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1828  *  _| |_| |____| |      | | | | | | (_| | | | | |
1829  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1830  *
1831  ***************************************************/
1832
1833 /**
1834  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1835  */
1836 static void create_ilp(ir_node *block, void *walk_env)
1837 {
1838         be_ilpsched_env_t     *env           = walk_env;
1839         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1840         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1841         FILE                  *logfile       = NULL;
1842         lpp_t                 *lpp           = NULL;
1843         int                   need_heur      = 0;
1844         struct obstack        var_obst;
1845         char                  name[1024];
1846
1847         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1848         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1849         DBG((env->dbg, 255, "=========================================\n\n"));
1850
1851         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1852                 block, ba->n_interesting_nodes, ba->max_steps));
1853
1854         /* notify backend and get block environment */
1855         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1856
1857         /* if we have less than two interesting nodes, there is no need to create the ILP */
1858         if (ba->n_interesting_nodes > 1) {
1859                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1860                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1861                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1862                 int    estimated_n_var = (int)((double)base_num * fact_var);
1863                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1864
1865                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1866                         estimated_n_var, estimated_n_cst));
1867                 (void) estimated_n_var;
1868
1869                 /* set up the LPP object */
1870                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1871
1872                 lpp = new_lpp_userdef(
1873                         (const char *)name,
1874                         lpp_minimize,
1875                         estimated_n_cst,     /* num vars */
1876                         estimated_n_cst + 1, /* num cst */
1877                         1.3);                /* grow factor */
1878                 obstack_init(&var_obst);
1879
1880                 /* create ILP variables */
1881                 create_variables(env, lpp, block_node, &var_obst);
1882
1883                 /* create ILP constraints */
1884                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1885                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1886                 create_ressource_constraints(env, lpp, block_node);
1887                 create_bundle_constraints(env, lpp, block_node);
1888                 create_branch_constraint(env, lpp, block_node);
1889                 //create_proj_keep_constraints(env, lpp, block_node);
1890
1891                 if (env->opts->regpress) {
1892                         create_alive_nodes_constraint(env, lpp, block_node);
1893                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1894                         create_pressure_alive_constraint(env, lpp, block_node);
1895                 }
1896
1897                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1898
1899                 /* debug stuff, dump lpp when debugging is on  */
1900                 DEBUG_ONLY({
1901                         if (firm_dbg_get_mask(env->dbg) > 1) {
1902                                 char buf[1024];
1903                                 FILE *f;
1904
1905                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1906                                 f = fopen(buf, "w");
1907                                 lpp_dump_plain(lpp, f);
1908                                 fclose(f);
1909                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1910                                 lpp_dump(lpp, buf);
1911                         }
1912                 })
1913
1914                 /* set solve time limit */
1915                 lpp_set_time_limit(lpp, env->opts->time_limit);
1916
1917                 /* set logfile if requested */
1918                 if (strlen(env->opts->log_file) > 0) {
1919                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1920                                 lpp_set_log(lpp, stdout);
1921                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1922                                 lpp_set_log(lpp, stderr);
1923                         else {
1924                                 logfile = fopen(env->opts->log_file, "w");
1925                                 if (! logfile)
1926                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1927                                 else
1928                                         lpp_set_log(lpp, logfile);
1929                         }
1930                 }
1931
1932                 /* solve the ILP */
1933                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1934
1935                 if (logfile)
1936                         fclose(logfile);
1937
1938                 /* check for valid solution */
1939                 if (! lpp_is_sol_valid(lpp)) {
1940                         DEBUG_ONLY({
1941                             char buf[1024];
1942                             FILE *f;
1943
1944                             if (firm_dbg_get_mask(env->dbg) >= 2) {
1945                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1946                               f = fopen(buf, "w");
1947                               lpp_dump_plain(lpp, f);
1948                               fclose(f);
1949                               snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1950                               lpp_dump(lpp, buf);
1951                               dump_ir_graph(env->irg, "infeasible");
1952                            }
1953                         })
1954
1955                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1956                         need_heur = 1;
1957                 }
1958
1959                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1960                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1961                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1962                 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));
1963                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1964                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1965                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1966                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1967
1968                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1969         }
1970
1971         /* apply solution */
1972         be_stat_ev("nodes", ba->block_last_idx);
1973         be_stat_ev("vars", lpp ? lpp->var_next : 0);
1974         be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1975         if (need_heur) {
1976                         be_stat_ev("time", -1);
1977                         be_stat_ev_dbl("opt", 0.0);
1978                 list_sched_single_block(env->birg, block, env->be_opts);
1979         }
1980         else {
1981                 if (lpp) {
1982                         double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1983                         be_stat_ev_dbl("time", lpp->sol_time);
1984                         be_stat_ev_dbl("opt", opt);
1985                 }
1986                 else {
1987                         be_stat_ev_dbl("time", 0.0);
1988                         be_stat_ev_dbl("opt", 100.0);
1989                 }
1990                 apply_solution(env, lpp, block);
1991         }
1992
1993         if (lpp)
1994                 free_lpp(lpp);
1995
1996         /* notify backend */
1997         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1998 }
1999
2000 /**
2001  * Perform ILP scheduling on the given irg.
2002  */
2003 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts)
2004 {
2005         be_ilpsched_env_t          env;
2006         ir_graph                   *irg      = be_get_birg_irg(birg);
2007         const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
2008         const ilp_sched_selector_t *sel      = arch_env->impl->get_ilp_sched_selector(arch_env);
2009
2010         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2011
2012         stat_ev_ctx_push("ilpsched");
2013
2014 //      firm_dbg_set_mask(env.dbg, 1);
2015
2016         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2017         env.sel        = sel;
2018         env.irg        = irg;
2019         env.height     = heights_new(irg);
2020         env.main_env   = birg->main_env;
2021         env.arch_env   = arch_env;
2022         env.cpu        = arch_env_get_machine(arch_env);
2023         env.opts       = &ilp_opts;
2024         env.birg       = birg;
2025         env.be_opts    = be_opts;
2026         phase_init(&env.ph, env.irg, init_ilpsched_irn);
2027
2028         /* assign a unique per block number to all interesting nodes */
2029         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2030
2031         /*
2032                 The block indices are completely build after the walk,
2033                 now we can allocate the bitsets (size depends on block indices)
2034                 for all nodes.
2035         */
2036         phase_reinit_irn_data(&env.ph);
2037
2038         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2039         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2040
2041         /* Calculate ALAP of all irns */
2042         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2043
2044         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2045         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2046
2047         /* perform ILP scheduling */
2048         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2049
2050         DEBUG_ONLY(
2051                 if (firm_dbg_get_mask(env.dbg)) {
2052                         phase_stat_t stat;
2053                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2054
2055                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2056                 }
2057         );
2058
2059         /* free data allocated dynamically */
2060         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2061
2062         /* free all allocated object */
2063         phase_deinit(&env.ph);
2064         heights_free(env.height);
2065
2066         /* notify backend */
2067         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2068
2069         stat_ev_ctx_pop("ilpsched");
2070 }
2071
2072 /**
2073  * Register ILP scheduler options.
2074  */
2075 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_ilpsched);
2076 void be_init_ilpsched(void)
2077 {
2078         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2079         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2080
2081         lc_opt_add_table(sched_grp, ilpsched_option_table);
2082 }
2083
2084 #else /* WITH_ILP */
2085
2086 static inline void some_picky_compiler_do_not_allow_empty_files(void)
2087 {}
2088
2089 #endif /* WITH_ILP */