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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass_t.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35 #include "irvrfy.h"
  36
  37 #include <lpp/lpp.h>
  38 #include <lpp/mps.h>
  39 #include <lpp/lpp_net.h>
  40 #include <lpp/lpp_cplex.h>
  41 //#include <lc_pset.h>
  42 //#include <libcore/lc_bitset.h>
  43
  44 #include "be_t.h"
  45 #include "belive_t.h"
  46 #include "besched_t.h"
  47 #include "beirgmod.h"
  48 #include "bearch.h"
  49 #include "beabi.h"
  50 #include "benode_t.h"
  51 #include "beutil.h"
  52 #include "bespillremat.h"
  53 #include "bespill.h"
  54 #include "bepressurestat.h"
  55 #include "beprofile.h"
  56
  57 #include "bechordal_t.h"
  58
  59 #ifdef WITH_LIBCORE
  60 #include <libcore/lc_opts.h>
  61 #include <libcore/lc_opts_enum.h>
  62 #endif /* WITH_LIBCORE */
  63
  64 #define DUMP_PROBLEM       1
  65 #define DUMP_MPS           2
  66 #define DUMP_SOLUTION      4
  67 #define DUMP_STATS         8
  68 #define DUMP_PRESSURE      16
  69
  70 #define KEEPALIVE_REMATS   1
  71 #define KEEPALIVE_SPILLS   2
  72 #define KEEPALIVE_RELOADS  4
  73
  74 #define VERIFY_MEMINTERF   1
  75 #define VERIFY_DOMINANCE   2
  76
  77 #define REMATS_NONE        0
  78 #define REMATS_BRIGGS      1
  79 #define REMATS_NOINVERSE   2
  80 #define REMATS_ALL         3
  81
  82 static int opt_dump_flags   = 0;
  83 static int opt_log = 0;
  84 static int opt_keep_alive   = 0;
  85 static int opt_goodwin = 1;
  86 static int opt_memcopies = 1;
  87 static int opt_memoperands = 1;
  88 static int opt_verify = VERIFY_MEMINTERF;
  89 static int opt_remats = REMATS_ALL;
  90 static int opt_repair_schedule = 0;
  91 static int opt_no_enlarge_liveness = 0;
  92 static int opt_timeout = 300;
  93 static double opt_cost_reload = 8.0;
  94 static double opt_cost_memoperand =  7.0;
  95 static double opt_cost_spill =  30.0;
  96 static double opt_cost_remat =  1.0;
  97
  98
  99 #ifdef WITH_LIBCORE
 100 static const lc_opt_enum_mask_items_t dump_items[] = {
 101         { "problem",  DUMP_PROBLEM  },
 102         { "mps",      DUMP_MPS      },
 103         { "solution", DUMP_SOLUTION },
 104         { "stats",    DUMP_STATS },
 105         { "pressure", DUMP_PRESSURE },
 106         { NULL,       0 }
 107 };
 108
 109 static lc_opt_enum_mask_var_t dump_var = {
 110         &opt_dump_flags, dump_items
 111 };
 112
 113 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 114         { "remats",  KEEPALIVE_REMATS  },
 115         { "spills",  KEEPALIVE_SPILLS  },
 116         { "reloads", KEEPALIVE_RELOADS },
 117         { NULL,      0 }
 118 };
 119
 120 static lc_opt_enum_mask_var_t keep_alive_var = {
 121         &opt_keep_alive, keepalive_items
 122 };
 123
 124 static const lc_opt_enum_mask_items_t remats_items[] = {
 125         { "none",      REMATS_NONE      },
 126         { "briggs",    REMATS_BRIGGS    },
 127         { "noinverse", REMATS_NOINVERSE },
 128         { "all",       REMATS_ALL       },
 129         { NULL,        0 }
 130 };
 131
 132 static lc_opt_enum_mask_var_t remats_var = {
 133         &opt_remats, remats_items
 134 };
 135
 136 static const lc_opt_table_entry_t options[] = {
 137         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads",                   &keep_alive_var),
 138
 139         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 140         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 141         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 142         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
 143         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 144         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 145
 146         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps, solution, stats or pressure",              &dump_var),
 147         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 148         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 149
 150         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 151         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 152         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 153         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 154         { NULL }
 155 };
 156
 157 void be_spill_remat_register_options(lc_opt_entry_t *grp)
 158 {
 159         lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
 160         lc_opt_add_table(my_grp, options);
 161 }
 162 #endif
 163
 164
 165 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 166 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 167
 168 #define  SOLVE
 169 //#define  SOLVE_LOCAL
 170 #define LPP_SERVER "i44pc52"
 171 #define LPP_SOLVER "cplex"
 172
 173
 174 #define MAX_PATHS      INT_MAX
 175 #define ILP_UNDEF               -1
 176
 177 typedef struct _spill_ilp_t {
 178         const arch_register_class_t  *cls;
 179         int                           n_regs;
 180         const be_chordal_env_t       *chordal_env;
 181         be_lv_t                      *lv;
 182         lpp_t                        *lpp;
 183         struct obstack               *obst;
 184         set                          *remat_info;
 185         pset                         *all_possible_remats;
 186         pset                         *inverse_ops;
 187         ir_node                      *keep;
 188         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 189         pset                         *spills;
 190         set                          *interferences;
 191         ir_node                      *m_unknown;
 192         set                          *memoperands;
 193         DEBUG_ONLY(firm_dbg_module_t * dbg);
 194 } spill_ilp_t;
 195
 196 typedef int ilp_var_t;
 197 typedef int ilp_cst_t;
 198
 199 typedef struct _spill_bb_t {
 200         set      *ilp;
 201         set      *reloads;
 202 } spill_bb_t;
 203
 204 typedef struct _remat_t {
 205         const ir_node        *op;      /**< for copy_irn */
 206         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 207         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 208         int                   cost;    /**< cost of this remat */
 209         int                   inverse; /**< nonzero if this is an inverse remat */
 210 } remat_t;
 211
 212 /**
 213  * Data to be attached to each IR node. For remats this contains the ilp_var
 214  * for this remat and for normal ops this contains the ilp_vars for
 215  * reloading each operand
 216  */
 217 typedef struct _op_t {
 218         int             is_remat;
 219         union {
 220                 struct {
 221                         ilp_var_t       ilp;
 222                         const remat_t  *remat; /** the remat this op belongs to */
 223                         int             pre; /** 1, if this is a pressure-increasing remat */
 224                 } remat;
 225                 struct {
 226                         ilp_var_t       ilp;
 227                         ir_node        *op; /** the operation this live range belongs to */
 228                         union {
 229                                 ilp_var_t      *reloads;
 230                                 ilp_var_t      *copies;
 231                         } args;
 232                 } live_range;
 233         } attr;
 234 } op_t;
 235
 236 typedef struct _defs_t {
 237         const ir_node   *value;
 238         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 239         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 240 } defs_t;
 241
 242 typedef struct _remat_info_t {
 243         const ir_node       *irn; /**< the irn to which these remats belong */
 244         pset                *remats; /**< possible remats for this value */
 245         pset                *remats_by_operand; /**< remats with this value as operand */
 246 } remat_info_t;
 247
 248 typedef struct _keyval_t {
 249         const void          *key;
 250         const void          *val;
 251 } keyval_t;
 252
 253 typedef struct _spill_t {
 254         ir_node            *irn;
 255         ilp_var_t           reg_in;
 256         ilp_var_t           mem_in;
 257         ilp_var_t           reg_out;
 258         ilp_var_t           mem_out;
 259         ilp_var_t           spill;
 260 } spill_t;
 261
 262 typedef struct _memoperand_t {
 263         ir_node             *irn; /**< the irn */
 264         unsigned int         pos; /**< the position of the argument */
 265         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 266 } memoperand_t;
 267
 268 static INLINE int
 269 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 270 {
 271         return chordal_has_class(si->chordal_env, irn);
 272 }
 273
 274 #if 0
 275 static int
 276 cmp_remat(const void *a, const void *b)
 277 {
 278         const keyval_t *p = a;
 279         const keyval_t *q = b;
 280         const remat_t  *r = p->val;
 281         const remat_t  *s = q->val;
 282
 283         assert(r && s);
 284
 285         return !(r == s || r->op == s->op);
 286 }
 287 #endif
 288 static int
 289 cmp_remat(const void *a, const void *b)
 290 {
 291         const remat_t  *r = a;
 292         const remat_t  *s = a;
 293
 294         return !(r == s || r->op == s->op);
 295 }
 296
 297 static int
 298 cmp_spill(const void *a, const void *b, size_t size)
 299 {
 300         const spill_t *p = a;
 301         const spill_t *q = b;
 302
 303 //      return !(p->irn == q->irn && p->bb == q->bb);
 304         return !(p->irn == q->irn);
 305 }
 306
 307 static int
 308 cmp_memoperands(const void *a, const void *b, size_t size)
 309 {
 310         const memoperand_t *p = a;
 311         const memoperand_t *q = b;
 312
 313         return !(p->irn == q->irn && p->pos == q->pos);
 314 }
 315
 316 static keyval_t *
 317 set_find_keyval(set * set, const void * key)
 318 {
 319         keyval_t     query;
 320
 321         query.key = key;
 322         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 323 }
 324
 325 static keyval_t *
 326 set_insert_keyval(set * set, void * key, void * val)
 327 {
 328         keyval_t     query;
 329
 330         query.key = key;
 331         query.val = val;
 332         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 333 }
 334
 335 static defs_t *
 336 set_find_def(set * set, const ir_node * value)
 337 {
 338         defs_t     query;
 339
 340         query.value = value;
 341         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 342 }
 343
 344 static defs_t *
 345 set_insert_def(set * set, const ir_node * value)
 346 {
 347         defs_t     query;
 348
 349         query.value = value;
 350         query.spills = NULL;
 351         query.remats = NULL;
 352         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 353 }
 354
 355 static memoperand_t *
 356 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 357 {
 358         memoperand_t     query;
 359
 360         query.irn = irn;
 361         query.pos = pos;
 362         query.ilp = ilp;
 363         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 364 }
 365
 366 static memoperand_t *
 367 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 368 {
 369         memoperand_t     query;
 370
 371         query.irn = (ir_node*)irn;
 372         query.pos = pos;
 373         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 374 }
 375
 376
 377 static spill_t *
 378 set_find_spill(set * set, const ir_node * value)
 379 {
 380         spill_t     query;
 381
 382         query.irn = (ir_node*)value;
 383         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 384 }
 385
 386 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 387 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 388 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 389 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 390 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 391
 392 static int
 393 cmp_remat_info(const void *a, const void *b, size_t size)
 394 {
 395         const remat_info_t *p = a;
 396         const remat_info_t *q = b;
 397
 398         return !(p->irn == q->irn);
 399 }
 400
 401 static int
 402 cmp_defs(const void *a, const void *b, size_t size)
 403 {
 404         const defs_t *p = a;
 405         const defs_t *q = b;
 406
 407         return !(p->value == q->value);
 408 }
 409
 410 static int
 411 cmp_keyval(const void *a, const void *b, size_t size)
 412 {
 413         const keyval_t *p = a;
 414         const keyval_t *q = b;
 415
 416         return !(p->key == q->key);
 417 }
 418
 419 static double
 420 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 421 {
 422 #define FUDGE 0.001
 423         if(be_profile_has_data())
 424                 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
 425
 426 #ifndef EXECFREQ_LOOPDEPH
 427         return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
 428 #else
 429         if(is_Block(irn))
 430                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 431         else
 432                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 433 #endif
 434 }
 435
 436 static double
 437 get_cost(const spill_ilp_t * si, const ir_node * irn)
 438 {
 439         if(be_is_Spill(irn)) {
 440                 return opt_cost_spill;
 441         } else if(be_is_Reload(irn)){
 442                 return opt_cost_reload;
 443         } else {
 444                 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
 445         }
 446 }
 447
 448 /**
 449  * Checks, whether node and its operands have suitable reg classes
 450  */
 451 static INLINE int
 452 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 453 {
 454         int               n;
 455         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 456         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 457
 458 #if 0
 459         if(!remat)
 460                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 461 #endif
 462
 463         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 464                 ir_node        *op = get_irn_n(irn, n);
 465                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 466
 467 //              if(!remat)
 468 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 469         }
 470
 471         return remat;
 472 }
 473
 474 /**
 475  * Try to create a remat from @p op with destination value @p dest_value
 476  */
 477 static INLINE remat_t *
 478 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 479 {
 480         remat_t  *remat = NULL;
 481
 482 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 483 //              return NULL;
 484
 485         if(dest_value == op) {
 486                 const ir_node *proj = NULL;
 487
 488                 if(is_Proj(dest_value)) {
 489                         op = get_Proj_pred(op);
 490                         proj = dest_value;
 491                 }
 492
 493                 if(!is_rematerializable(si, op))
 494                         return NULL;
 495
 496                 remat          = obstack_alloc(si->obst, sizeof(*remat));
 497                 remat->op      = op;
 498                 remat->cost    = get_cost(si, op);
 499                 remat->value   = dest_value;
 500                 remat->proj    = proj;
 501                 remat->inverse = 0;
 502         } else {
 503                 arch_inverse_t     inverse;
 504                 int                n;
 505
 506                 /* get the index of the operand we want to retrieve by the inverse op */
 507                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 508                         ir_node        *arg = get_irn_n(op, n);
 509
 510                         if(arg == dest_value) break;
 511                 }
 512                 if(n<0) return NULL;
 513
 514                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 515
 516                 /* else ask the backend to give an inverse op */
 517                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 518                         int   i;
 519
 520                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 521
 522                         assert(inverse.n > 0 && "inverse op should have at least one node");
 523
 524                         for(i=inverse.n-1; i>=0; --i) {
 525                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 526                         }
 527
 528                         if(inverse.n <= 2) {
 529                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 530                                 remat->op = inverse.nodes[0];
 531                                 remat->cost = inverse.costs;
 532                                 remat->value = dest_value;
 533                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 534                                 remat->inverse = 1;
 535
 536                                 assert(is_Proj(remat->proj));
 537                         } else {
 538                                 assert(0 && "I can not handle remats with more than 2 nodes");
 539                         }
 540                 }
 541         }
 542
 543         if(remat) {
 544                 if(remat->proj) {
 545                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 546                 } else {
 547                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 548                 }
 549         }
 550         return remat;
 551 }
 552
 553
 554 static INLINE void
 555 add_remat(const spill_ilp_t * si, const remat_t * remat)
 556 {
 557         remat_info_t    *remat_info,
 558                      query;
 559         int              n;
 560
 561         assert(remat->op);
 562         assert(remat->value);
 563
 564         query.irn = remat->value;
 565         query.remats = NULL;
 566         query.remats_by_operand = NULL;
 567         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 568
 569         if(remat_info->remats == NULL) {
 570                 remat_info->remats = new_pset(cmp_remat, 4096);
 571         }
 572         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 573
 574         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 575         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 576                 ir_node        *arg = get_irn_n(remat->op, n);
 577
 578                 query.irn = arg;
 579                 query.remats = NULL;
 580                 query.remats_by_operand = NULL;
 581                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 582
 583                 if(remat_info->remats_by_operand == NULL) {
 584                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 585                 }
 586                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 587         }
 588 }
 589
 590 static int
 591 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 592 {
 593         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 594         int                i = 0;
 595
 596         while(edge) {
 597                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 598                         ++i;
 599                 }
 600                 edge = get_irn_out_edge_next(irn, edge);
 601         }
 602
 603         return i;
 604 }
 605
 606 static int
 607 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 608 {
 609         int n;
 610         int ret = 0;
 611
 612         if(is_Proj(irn))
 613                 irn = get_Proj_pred(irn);
 614
 615         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 616                 const ir_node  *arg = get_irn_n(irn, n);
 617
 618                 if(has_reg_class(si, arg)) ++ret;
 619         }
 620
 621         return ret;
 622 }
 623
 624 static INLINE void
 625 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 626 {
 627         int      n;
 628         remat_t *remat;
 629
 630         if( has_reg_class(si, op)
 631         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 632         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 633         ) {
 634                 remat = get_remat_from_op(si, op, op);
 635                 if(remat) {
 636                         add_remat(si, remat);
 637                 }
 638         }
 639
 640         if(opt_remats == REMATS_ALL) {
 641                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 642                    for each arg */
 643                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 644                         ir_node        *arg = get_irn_n(op, n);
 645
 646                         if(has_reg_class(si, arg)) {
 647                                 /* try to get an inverse remat */
 648                                 remat = get_remat_from_op(si, arg, op);
 649                                 if(remat) {
 650                                         add_remat(si, remat);
 651                                 }
 652                         }
 653                 }
 654         }
 655 }
 656
 657 static INLINE int
 658 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 659 {
 660         ir_node *block;
 661         ir_node *def_block = get_nodes_block(val);
 662         int      ret;
 663
 664         if(val == pos)
 665                 return 0;
 666
 667         /* if pos is at end of a basic block */
 668         if(is_Block(pos)) {
 669                 ret = (pos == def_block || block_dominates(def_block, pos));
 670 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 671                 return ret;
 672         }
 673
 674         /* else if this is a normal operation */
 675         block = get_nodes_block(pos);
 676         if(block == def_block) {
 677                 if(!sched_is_scheduled(val)) return 1;
 678
 679                 ret = sched_comes_after(val, pos);
 680 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 681                 return ret;
 682         }
 683
 684         ret = block_dominates(def_block, block);
 685 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 686         return ret;
 687 }
 688
 689 static INLINE ir_node *
 690 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 691 {
 692     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 693 }
 694
 695 /**
 696  * Returns first non-Phi node of block @p bb
 697  */
 698 static INLINE ir_node *
 699 sched_block_first_nonphi(const ir_node * bb)
 700 {
 701         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 702 }
 703
 704 static int
 705 sched_skip_proj_predicator(const ir_node * irn, void * data)
 706 {
 707         return (is_Proj(irn));
 708 }
 709
 710 static INLINE ir_node *
 711 sched_next_nonproj(const ir_node * irn, int forward)
 712 {
 713         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 714 }
 715
 716 /**
 717  * Returns next operation node (non-Proj) after @p irn
 718  * or the basic block of this node
 719  */
 720 static INLINE ir_node *
 721 sched_next_op(const ir_node * irn)
 722 {
 723         ir_node *next = sched_next(irn);
 724
 725         if(is_Block(next))
 726                 return next;
 727
 728         return sched_next_nonproj(next, 1);
 729 }
 730
 731 /**
 732  * Returns previous operation node (non-Proj) before @p irn
 733  * or the basic block of this node
 734  */
 735 static INLINE ir_node *
 736 sched_prev_op(const ir_node * irn)
 737 {
 738         ir_node *prev = sched_prev(irn);
 739
 740         if(is_Block(prev))
 741                 return prev;
 742
 743         return sched_next_nonproj(prev, 0);
 744 }
 745
 746 static void
 747 sched_put_after(ir_node * insert, ir_node * irn)
 748 {
 749         if(is_Block(insert)) {
 750                 insert = sched_block_first_nonphi(insert);
 751         } else {
 752                 insert = sched_next_op(insert);
 753         }
 754         sched_add_before(insert, irn);
 755 }
 756
 757 static void
 758 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 759 {
 760   if(is_Block(insert)) {
 761           insert = sched_block_last_noncf(si, insert);
 762   } else {
 763           insert = sched_next_nonproj(insert, 0);
 764           insert = sched_prev(insert);
 765   }
 766   sched_add_after(insert, irn);
 767 }
 768
 769 /**
 770  * Tells you whether a @p remat can be placed before the irn @p pos
 771  */
 772 static INLINE int
 773 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 774 {
 775         const ir_node   *op = remat->op;
 776         const ir_node   *prev;
 777         int        n,
 778                            res = 1;
 779
 780         if(is_Block(pos)) {
 781                 prev = sched_block_last_noncf(si, pos);
 782                 prev = sched_next_nonproj(prev, 0);
 783         } else {
 784                 prev = sched_prev_op(pos);
 785         }
 786         /* do not remat if the rematted value is defined immediately before this op */
 787         if(prev == remat->op) {
 788                 return 0;
 789         }
 790
 791 #if 0
 792         /* this should be just fine, the following OP will be using this value, right? */
 793
 794         /* only remat AFTER the real definition of a value (?) */
 795         if(!value_is_defined_before(si, pos, remat->value)) {
 796 //              ir_fprintf(stderr, "error(not defined)");
 797                 return 0;
 798         }
 799 #endif
 800
 801         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 802                 const ir_node   *arg = get_irn_n(op, n);
 803
 804                 if(opt_no_enlarge_liveness) {
 805                         if(has_reg_class(si, arg) && live) {
 806                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 807                         } else {
 808                                 res &= value_is_defined_before(si, pos, arg);
 809                         }
 810                 } else {
 811                         res &= value_is_defined_before(si, pos, arg);
 812                 }
 813         }
 814
 815         return res;
 816 }
 817
 818 /**
 819  * Tells you whether a @p remat can be placed after the irn @p pos
 820  */
 821 static INLINE int
 822 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 823 {
 824         if(is_Block(pos)) {
 825                 pos = sched_block_first_nonphi(pos);
 826         } else {
 827                 pos = sched_next_op(pos);
 828         }
 829
 830         /* only remat AFTER the real definition of a value (?) */
 831         if(!value_is_defined_before(si, pos, remat->value)) {
 832                 return 0;
 833         }
 834
 835         return can_remat_before(si, remat, pos, live);
 836 }
 837
 838 /**
 839  * Collect potetially rematerializable OPs
 840  */
 841 static void
 842 walker_remat_collector(ir_node * irn, void * data)
 843 {
 844         spill_ilp_t    *si = data;
 845
 846         if(!is_Block(irn) && !is_Phi(irn)) {
 847                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 848                 get_remats_from_op(si, irn);
 849         }
 850 }
 851
 852 /**
 853  * Inserts a copy of @p irn before @p pos
 854  */
 855 static ir_node *
 856 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 857 {
 858         ir_node     *bb;
 859         ir_node     *copy;
 860
 861         bb = is_Block(pos)?pos:get_nodes_block(pos);
 862         copy = exact_copy(irn);
 863
 864         _set_phi_class(copy, NULL);
 865         set_nodes_block(copy, bb);
 866         sched_put_before(si, pos, copy);
 867
 868         return copy;
 869 }
 870
 871 /**
 872  * Inserts a copy of @p irn after @p pos
 873  */
 874 static ir_node *
 875 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 876 {
 877         ir_node     *bb;
 878         ir_node     *copy;
 879
 880         bb = is_Block(pos)?pos:get_nodes_block(pos);
 881         copy = exact_copy(irn);
 882
 883         _set_phi_class(copy, NULL);
 884         set_nodes_block(copy, bb);
 885         sched_put_after(pos, copy);
 886
 887         return copy;
 888 }
 889
 890 static ir_node *
 891 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 892 {
 893         char     buf[256];
 894
 895         if(can_remat_after(si, remat, pos, live)) {
 896                 ir_node         *copy,
 897                                                 *proj_copy;
 898                 op_t            *op;
 899
 900                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 901
 902                 copy = insert_copy_after(si, remat->op, pos);
 903
 904                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 905                 op = obstack_alloc(si->obst, sizeof(*op));
 906                 op->is_remat = 1;
 907                 op->attr.remat.remat = remat;
 908                 op->attr.remat.pre = 0;
 909                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 910
 911                 set_irn_link(copy, op);
 912                 pset_insert_ptr(si->all_possible_remats, copy);
 913                 if(remat->proj) {
 914                         proj_copy = insert_copy_after(si, remat->proj, copy);
 915                         set_irn_n(proj_copy, 0, copy);
 916                         set_irn_link(proj_copy, op);
 917                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 918                 } else {
 919                         proj_copy = NULL;
 920                 }
 921
 922                 return copy;
 923         }
 924
 925         return NULL;
 926 }
 927
 928 static ir_node *
 929 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 930 {
 931         char     buf[256];
 932
 933         if(can_remat_before(si, remat, pos, live)) {
 934                 ir_node         *copy,
 935                                                 *proj_copy;
 936                 op_t            *op;
 937
 938                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 939
 940                 copy = insert_copy_before(si, remat->op, pos);
 941
 942                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 943                 op = obstack_alloc(si->obst, sizeof(*op));
 944                 op->is_remat = 1;
 945                 op->attr.remat.remat = remat;
 946                 op->attr.remat.pre = 1;
 947                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 948
 949                 set_irn_link(copy, op);
 950                 pset_insert_ptr(si->all_possible_remats, copy);
 951                 if(remat->proj) {
 952                         proj_copy = insert_copy_after(si, remat->proj, copy);
 953                         set_irn_n(proj_copy, 0, copy);
 954                         set_irn_link(proj_copy, op);
 955                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 956                 } else {
 957                         proj_copy = NULL;
 958                 }
 959
 960                 return copy;
 961         }
 962
 963         return NULL;
 964 }
 965
 966 static int
 967 get_block_n_succs(const ir_node *block) {
 968         const ir_edge_t *edge;
 969
 970         assert(edges_activated(current_ir_graph));
 971
 972         edge = get_block_succ_first(block);
 973         if (! edge)
 974                 return 0;
 975
 976         edge = get_block_succ_next(block, edge);
 977         return edge ? 2 : 1;
 978 }
 979
 980 static int
 981 is_start_block(const ir_node * bb)
 982 {
 983         return get_irg_start_block(get_irn_irg(bb)) == bb;
 984 }
 985
 986 static int
 987 is_merge_edge(const ir_node * bb)
 988 {
 989         if(is_start_block(bb))
 990                 return 0;
 991
 992         if(opt_goodwin)
 993                 return get_block_n_succs(bb) == 1;
 994         else
 995                 return 1;
 996 }
 997
 998 static int
 999 is_diverge_edge(const ir_node * bb)
1000 {
1001         if(is_start_block(bb))
1002                 return 0;
1003
1004         if(opt_goodwin)
1005                 return get_Block_n_cfgpreds(bb) == 1;
1006         else
1007                 return 1;
1008 }
1009
1010 static void
1011 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1012 {
1013         ir_node        *irn;
1014         int i;
1015
1016         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1017                 irn = be_lv_get_irn(si->lv, bb, i);
1018
1019                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1020                         pset_insert_ptr(live, irn);
1021                 }
1022         }
1023
1024         irn = sched_last(bb);
1025
1026         /* all values eaten by control flow operations are also live until the end of the block */
1027         sched_foreach_reverse(bb, irn) {
1028                 int  i;
1029
1030                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1031
1032                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1033                         ir_node *arg = get_irn_n(irn,i);
1034
1035                         if(has_reg_class(si, arg)) {
1036                                 pset_insert_ptr(live, arg);
1037                         }
1038                 }
1039         }
1040 }
1041
1042 static void
1043 walker_regclass_copy_insertor(ir_node * irn, void * data)
1044 {
1045         spill_ilp_t    *si = data;
1046
1047         if(is_Phi(irn) && has_reg_class(si, irn)) {
1048                 int n;
1049
1050                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1051                         ir_node  *phi_arg = get_irn_n(irn, n);
1052                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1053
1054                         if(!has_reg_class(si, phi_arg)) {
1055                                 ir_node   *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1056                                 ir_node   *pos = sched_block_last_noncf(si, bb);
1057                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1058
1059                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1060                                 sched_add_after(pos, copy);
1061                                 set_irn_n(irn, n, copy);
1062
1063                                 op->is_remat = 0;
1064                                 op->attr.live_range.args.reloads = NULL;
1065                                 op->attr.live_range.ilp = ILP_UNDEF;
1066                                 set_irn_link(copy, op);
1067                         }
1068                 }
1069         }
1070 }
1071
1072 /**
1073  * Insert (so far unused) remats into the irg to
1074  * recompute the potential liveness of all values
1075  */
1076 static void
1077 walker_remat_insertor(ir_node * bb, void * data)
1078 {
1079         spill_ilp_t    *si = data;
1080         ir_node        *irn;
1081         int             n, i;
1082         pset           *live;
1083         pset           *post_remats;
1084         remat_t        *remat;
1085
1086         /* skip start block, no remats to do there */
1087         if(is_start_block(bb)) return;
1088
1089         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1090
1091         live = pset_new_ptr_default();
1092         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1093                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1094
1095                 /* add remats at end of block */
1096                 if (has_reg_class(si, value)) {
1097                         pset_insert_ptr(live, value);
1098                 }
1099         }
1100
1101         irn = sched_last(bb);
1102         while(!sched_is_end(irn)) {
1103                 ir_node   *next;
1104                 pset      *args;
1105                 ir_node   *arg;
1106                 pset      *used;
1107
1108                 next = sched_prev(irn);
1109
1110                 /* delete defined value from live set */
1111                 if(has_reg_class(si, irn)) {
1112                         pset_remove_ptr(live, irn);
1113                 }
1114
1115                 if(is_Phi(irn) || is_Proj(irn)) {
1116                         irn = next;
1117                         continue;
1118                 }
1119
1120                 args = pset_new_ptr_default();
1121                 used = pset_new_ptr_default();
1122
1123                 /* collect arguments of op and set args of op already live in epilog */
1124                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1125                         ir_node        *arg = get_irn_n(irn, n);
1126
1127                         pset_insert_ptr(args, arg);
1128                         if(has_reg_class(si, arg)) {
1129                                 pset_insert_ptr(live, arg);
1130                                 pset_insert_ptr(used, arg);
1131                         }
1132                 }
1133
1134                 /* insert all possible remats before irn */
1135                 pset_foreach(args, arg) {
1136                         remat_info_t   *remat_info,
1137                                                     query;
1138
1139                         /* continue if the operand has the wrong reg class */
1140                         if(!has_reg_class(si, arg))
1141                                 continue;
1142
1143                         query.irn = arg;
1144                         query.remats = NULL;
1145                         query.remats_by_operand = NULL;
1146                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1147
1148                         if(!remat_info) {
1149                                 continue;
1150                         }
1151
1152                         if(remat_info->remats) {
1153                                 pset_foreach(remat_info->remats, remat) {
1154                                         ir_node  *remat_irn = NULL;
1155
1156                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1157                                         remat_irn = insert_remat_before(si, remat, irn, live);
1158
1159                                         if(remat_irn) {
1160                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1161                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1162
1163                                                         /* collect args of remats which are not args of op */
1164                                                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1165                                                                 pset_insert_ptr(used, remat_arg);
1166                                                         }
1167                                                 }
1168                                         }
1169                                 }
1170                         }
1171                 }
1172
1173                 /* do not place post remats after jumps */
1174                 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) {
1175                         del_pset(used);
1176                         del_pset(args);
1177                         break;
1178                 }
1179
1180                 /* insert all possible remats after irn */
1181                 post_remats = pset_new_ptr_default();
1182                 pset_foreach(used, arg) {
1183                         remat_info_t   *remat_info,
1184                                                     query;
1185
1186                         /* continue if the operand has the wrong reg class */
1187                         if(!has_reg_class(si, arg))
1188                                 continue;
1189
1190                         query.irn = arg;
1191                         query.remats = NULL;
1192                         query.remats_by_operand = NULL;
1193                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1194
1195                         if(!remat_info) {
1196                                 continue;
1197                         }
1198
1199                         if(remat_info->remats_by_operand) {
1200                                 pset_foreach(remat_info->remats_by_operand, remat) {
1201                                         /* do not insert remats producing the same value as one of the operands */
1202                                         if(!pset_find_ptr(args, remat->value)) {
1203                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1204
1205                                                 /* only remat values that can be used be real ops */
1206                                                 if(pset_find_ptr(live, remat->value)) {
1207                                                         pset_insert_ptr(post_remats, remat);
1208                                                 }
1209                                         }
1210                                 }
1211                         }
1212                 }
1213                 pset_foreach(post_remats, remat) {
1214                         insert_remat_after(si, remat, irn, live);
1215                 }
1216                 del_pset(post_remats);
1217
1218                 del_pset(used);
1219                 del_pset(args);
1220                 irn = next;
1221         }
1222
1223         /* add remats at end if successor has multiple predecessors */
1224         if(is_merge_edge(bb)) {
1225                 pset     *live_out = pset_new_ptr_default();
1226                 ir_node  *value;
1227
1228                 get_live_end(si, bb, live_out);
1229
1230                 /* add remats at end of block */
1231                 pset_foreach(live_out, value) {
1232                         remat_info_t   *remat_info,
1233                                                    query;
1234
1235                         query.irn = value;
1236                         query.remats = NULL;
1237                         query.remats_by_operand = NULL;
1238                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1239
1240                         if(remat_info && remat_info->remats) {
1241                                 pset_foreach(remat_info->remats, remat) {
1242                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1243
1244                                         insert_remat_before(si, remat, bb, live_out);
1245                                 }
1246                         }
1247                 }
1248                 del_pset(live_out);
1249         }
1250
1251         if(is_diverge_edge(bb)) {
1252                 pset     *live_in = pset_new_ptr_default();
1253                 ir_node  *value;
1254
1255                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1256                         value = be_lv_get_irn(si->lv, bb, i);
1257
1258                         if(has_reg_class(si, value)) {
1259                                 pset_insert_ptr(live_in, value);
1260                         }
1261                 }
1262                 /* add phis to live_in */
1263                 sched_foreach(bb, value) {
1264                         if(!is_Phi(value)) break;
1265
1266                         if(has_reg_class(si, value)) {
1267                                 pset_insert_ptr(live_in, value);
1268                         }
1269                 }
1270
1271                 /* add remat2s at beginning of block */
1272                 post_remats = pset_new_ptr_default();
1273                 pset_foreach(live_in, value) {
1274                         remat_info_t   *remat_info,
1275                                                    query;
1276
1277                         query.irn = value;
1278                         query.remats = NULL;
1279                         query.remats_by_operand = NULL;
1280                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1281
1282                         if(remat_info && remat_info->remats_by_operand) {
1283                                 pset_foreach(remat_info->remats_by_operand, remat) {
1284                                         DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1285
1286                                         /* put the remat here if all its args are available and result is still live */
1287                                         if(pset_find_ptr(live_in, remat->value)) {
1288                                                 pset_insert_ptr(post_remats, remat);
1289                                         }
1290                                 }
1291                         }
1292                 }
1293                 pset_foreach(post_remats, remat) {
1294                         insert_remat_after(si, remat, bb, live_in);
1295                 }
1296                 del_pset(post_remats);
1297                 del_pset(live_in);
1298         }
1299 }
1300
1301 static int
1302 can_be_copied(const ir_node * bb, const ir_node * irn)
1303 {
1304         const ir_edge_t *edge    = get_block_succ_first(bb);
1305         const ir_node   *next_bb = edge->src;
1306         int             pos      = edge->pos;
1307         const ir_node   *phi;
1308
1309         assert(is_merge_edge(bb));
1310
1311         sched_foreach(next_bb, phi) {
1312                 const ir_node  *phi_arg;
1313
1314                 if(!is_Phi(phi)) break;
1315
1316                 phi_arg = get_irn_n(phi, pos);
1317
1318                 if(phi_arg == irn) {
1319                         return 1;
1320                 }
1321         }
1322         return 0;
1323 }
1324
1325 /**
1326  * Initialize additional node info
1327  */
1328 static void
1329 luke_initializer(ir_node * bb, void * data)
1330 {
1331         spill_ilp_t    *si = (spill_ilp_t*)data;
1332         spill_bb_t     *spill_bb;
1333         ir_node        *irn;
1334
1335         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1336         set_irn_link(bb, spill_bb);
1337
1338         sched_foreach(bb, irn) {
1339                 op_t      *op;
1340
1341                 op = obstack_alloc(si->obst, sizeof(*op));
1342                 op->is_remat = 0;
1343                 op->attr.live_range.ilp = ILP_UNDEF;
1344                 if(is_Phi(irn)) {
1345                         if(opt_memcopies) {
1346                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1347                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1348                         }
1349                 } else if(!is_Proj(irn)) {
1350                         op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1351                         memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1352                 } else {
1353                         op->attr.live_range.args.reloads = NULL;
1354                 }
1355                 set_irn_link(irn, op);
1356         }
1357 }
1358
1359
1360 /**
1361  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1362  */
1363 static void
1364 luke_endwalker(ir_node * bb, void * data)
1365 {
1366         spill_ilp_t    *si = (spill_ilp_t*)data;
1367         pset           *live;
1368         pset           *use_end;
1369         char            buf[256];
1370         ilp_cst_t       cst;
1371         ir_node        *irn;
1372         spill_bb_t     *spill_bb = get_irn_link(bb);
1373         int             i;
1374
1375         live = pset_new_ptr_default();
1376         use_end = pset_new_ptr_default();
1377
1378         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1379                 irn = be_lv_get_irn(si->lv, bb, i);
1380                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1381                         pset_insert_ptr(live, irn);
1382                 }
1383         }
1384
1385         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1386         /* their reg_out must always be set */
1387         sched_foreach_reverse(bb, irn) {
1388                 int   n;
1389
1390                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1391
1392                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1393                         ir_node        *irn_arg = get_irn_n(irn, n);
1394
1395                         if(has_reg_class(si, irn_arg)) {
1396                                 pset_insert_ptr(use_end, irn_arg);
1397                         }
1398                 }
1399         }
1400
1401         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1402         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1403         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1404
1405         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1406
1407         /* if this is a merge edge we can reload at the end of this block */
1408         if(is_merge_edge(bb)) {
1409                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1410         } else if(pset_count(use_end)){
1411                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1412         } else {
1413                 spill_bb->reloads = NULL;
1414         }
1415
1416         pset_foreach(live,irn) {
1417                 spill_t     query,
1418                                         *spill;
1419                 double      spill_cost;
1420                 int         default_spilled;
1421
1422
1423                 /* handle values used by control flow nodes later separately */
1424                 if(pset_find_ptr(use_end, irn)) continue;
1425
1426                 query.irn = irn;
1427                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1428
1429                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1430
1431                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1432                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1433                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1434
1435                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1436                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1437
1438                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1439                 /* by default spill value right after definition */
1440                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1441                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1442
1443                 if(is_merge_edge(bb)) {
1444                         ilp_var_t   reload;
1445                         ilp_cst_t   rel_cst;
1446
1447                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1448                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1449                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1450
1451                         /* reload <= mem_out */
1452                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1453                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1454                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1455                 }
1456
1457                 spill->reg_in = ILP_UNDEF;
1458                 spill->mem_in = ILP_UNDEF;
1459         }
1460
1461         pset_foreach(use_end,irn) {
1462                 spill_t     query,
1463                                         *spill;
1464                 double      spill_cost;
1465                 ilp_cst_t   end_use_req,
1466                                         rel_cst;
1467                 ilp_var_t   reload;
1468                 int         default_spilled;
1469
1470                 query.irn = irn;
1471                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1472
1473                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1474
1475                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1476                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1477                 /* if irn is used at the end of the block, then it is live anyway */
1478                 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1479
1480                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1481                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1482
1483                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1484                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1485                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1486
1487                 /* reload for use be control flow op */
1488                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1489                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1490                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1491
1492                 /* reload <= mem_out */
1493                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1494                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1495                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1496
1497                 spill->reg_in = ILP_UNDEF;
1498                 spill->mem_in = ILP_UNDEF;
1499
1500                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1501                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1502                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1503         }
1504
1505         del_pset(live);
1506         del_pset(use_end);
1507 }
1508
1509 static ir_node *
1510 next_post_remat(const ir_node * irn)
1511 {
1512         op_t      *op;
1513     ir_node   *next;
1514
1515         if(is_Block(irn)) {
1516                 next = sched_block_first_nonphi(irn);
1517         } else {
1518                 next = sched_next_op(irn);
1519         }
1520
1521         if(sched_is_end(next))
1522                 return NULL;
1523
1524         op = get_irn_link(next);
1525         if(op->is_remat && !op->attr.remat.pre) {
1526                 return next;
1527         }
1528
1529         return NULL;
1530 }
1531
1532
1533 static ir_node *
1534 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1535 {
1536         op_t      *op;
1537         ir_node   *ret;
1538
1539         if(is_Block(irn)) {
1540                 ret = sched_block_last_noncf(si, irn);
1541                 ret = sched_next(ret);
1542                 ret = sched_prev_op(ret);
1543         } else {
1544                 ret = sched_prev_op(irn);
1545         }
1546
1547         if(sched_is_end(ret) || is_Phi(ret))
1548                 return NULL;
1549
1550         op = (op_t*)get_irn_link(ret);
1551         if(op->is_remat && op->attr.remat.pre) {
1552                 return ret;
1553         }
1554
1555         return NULL;
1556 }
1557
1558 /**
1559  * Find a remat of value @p value in the epilog of @p pos
1560  */
1561 static ir_node *
1562 find_post_remat(const ir_node * value, const ir_node * pos)
1563 {
1564         while((pos = next_post_remat(pos)) != NULL) {
1565                 op_t   *op;
1566
1567                 op = get_irn_link(pos);
1568                 assert(op->is_remat && !op->attr.remat.pre);
1569
1570                 if(op->attr.remat.remat->value == value)
1571                         return (ir_node*)pos;
1572
1573 #if 0
1574         const ir_edge_t *edge;
1575                 foreach_out_edge(pos, edge) {
1576                         ir_node   *proj = get_edge_src_irn(edge);
1577                         assert(is_Proj(proj));
1578                 }
1579 #endif
1580
1581         }
1582
1583         return NULL;
1584 }
1585
1586 static spill_t *
1587 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1588 {
1589         spill_bb_t  *spill_bb = get_irn_link(bb);
1590         spill_t     *spill,
1591                                  query;
1592         char         buf[256];
1593         int          default_spilled;
1594
1595         query.irn = irn;
1596         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1597         if(!spill) {
1598                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1599
1600                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1601
1602                 spill->reg_out = ILP_UNDEF;
1603                 spill->reg_in  = ILP_UNDEF;
1604                 spill->mem_in  = ILP_UNDEF;
1605
1606                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1607                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1608
1609                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1610                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1611                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1612         }
1613
1614         return spill;
1615 }
1616
1617 /**
1618  *  Inserts ILP-constraints and variables for memory copying before the given position
1619  */
1620 static void
1621 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1622 {
1623         const ir_node    *succ;
1624         const ir_edge_t  *edge;
1625         spill_bb_t       *spill_bb = get_irn_link(block);
1626         ir_node          *phi;
1627         int               pos;
1628         ilp_cst_t         cst;
1629         ilp_var_t         copyreg;
1630         char              buf[256];
1631         ir_node          *tmp;
1632
1633
1634         assert(edges_activated(current_ir_graph));
1635
1636         edge = get_block_succ_first(block);
1637         if(!edge) return;
1638
1639         succ = edge->src;
1640         pos = edge->pos;
1641
1642         edge = get_block_succ_next(block, edge);
1643         /* next block can only contain phis, if this is a merge edge */
1644         if(edge) return;
1645
1646         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1647         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1648
1649         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1650         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1651
1652         pset_foreach(live, tmp) {
1653                 spill_t  *spill;
1654 #if 0
1655                 op_t  *op = get_irn_link(irn);
1656                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1657 #endif
1658                 spill = set_find_spill(spill_bb->ilp, tmp);
1659                 assert(spill);
1660
1661                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1662         }
1663         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1664
1665         sched_foreach(succ, phi) {
1666                 const ir_node  *to_copy;
1667                 op_t           *to_copy_op;
1668                 spill_t        *to_copy_spill;
1669                 op_t           *phi_op = get_irn_link(phi);
1670                 ilp_var_t       reload = ILP_UNDEF;
1671
1672
1673                 if(!is_Phi(phi)) break;
1674                 if(!has_reg_class(si, phi)) continue;
1675
1676                 to_copy = get_irn_n(phi, pos);
1677                 to_copy_op = get_irn_link(to_copy);
1678
1679                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1680                 assert(to_copy_spill);
1681
1682                 if(spill_bb->reloads) {
1683                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1684
1685                         if(keyval) {
1686                                 reload = PTR_TO_INT(keyval->val);
1687                         }
1688                 }
1689
1690                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1691                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1692
1693                 /* copy - reg_out - reload - remat - live_range <= 0 */
1694                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1695                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1696                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1697                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1698                 foreach_pre_remat(si, block, tmp) {
1699                         op_t     *remat_op = get_irn_link(tmp);
1700                         if(remat_op->attr.remat.remat->value == to_copy) {
1701                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1702                         }
1703                 }
1704
1705                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1706                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1707
1708                 /* copy - reg_out - copyreg <= 0 */
1709                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1710                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1711                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1712         }
1713 }
1714
1715
1716 /**
1717  * Walk all irg blocks and emit this ILP
1718  */
1719 static void
1720 luke_blockwalker(ir_node * bb, void * data)
1721 {
1722         spill_ilp_t    *si = (spill_ilp_t*)data;
1723         ir_node        *irn;
1724         pset           *live;
1725         char            buf[256];
1726         ilp_cst_t       cst;
1727         spill_bb_t     *spill_bb = get_irn_link(bb);
1728         ir_node        *tmp;
1729         spill_t        *spill;
1730         pset           *defs = pset_new_ptr_default();
1731         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1732
1733         live = pset_new_ptr_default();
1734
1735         /****************************************
1736          *      B A S I C  B L O C K  E N D
1737          ***************************************/
1738
1739
1740         /* init live values at end of block */
1741         get_live_end(si, bb, live);
1742
1743         pset_foreach(live, irn) {
1744                 op_t           *op;
1745                 ilp_var_t       reload = ILP_UNDEF;
1746
1747                 spill = set_find_spill(spill_bb->ilp, irn);
1748                 assert(spill);
1749
1750                 if(spill_bb->reloads) {
1751                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1752
1753                         if(keyval) {
1754                                 reload = PTR_TO_INT(keyval->val);
1755                         }
1756                 }
1757
1758                 op = get_irn_link(irn);
1759                 assert(!op->is_remat);
1760
1761                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1762                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1763                 op->attr.live_range.op = bb;
1764
1765                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1766                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1767
1768                 /* reg_out - reload - remat - live_range <= 0 */
1769                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1770                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1771                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1772                 foreach_pre_remat(si, bb, tmp) {
1773                         op_t     *remat_op = get_irn_link(tmp);
1774                         if(remat_op->attr.remat.remat->value == irn) {
1775                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1776                         }
1777                 }
1778                 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1779                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1780
1781                 /* value may only die at bb end if it is used for a mem copy */
1782                 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1783                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1784                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1785                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1786                 foreach_pre_remat(si, bb, tmp) {
1787                         op_t     *remat_op = get_irn_link(tmp);
1788                         if(remat_op->attr.remat.remat->value == irn) {
1789                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1790                         }
1791                 }
1792                 if(is_merge_edge(bb)) {
1793                         const ir_edge_t *edge = get_block_succ_first(bb);
1794                         const ir_node   *next_bb = edge->src;
1795                         int              pos = edge->pos;
1796                         const ir_node   *phi;
1797
1798                         sched_foreach(next_bb, phi) {
1799                                 const ir_node  *phi_arg;
1800
1801                                 if(!is_Phi(phi)) break;
1802
1803                                 phi_arg = get_irn_n(phi, pos);
1804
1805                                 if(phi_arg == irn) {
1806                                         op_t      *phi_op = get_irn_link(phi);
1807                                         ilp_var_t  copy = phi_op->attr.live_range.args.copies[pos];
1808
1809                                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1810                                 }
1811                         }
1812                 }
1813         }
1814
1815         if(opt_memcopies)
1816                 insert_mem_copy_position(si, live, bb);
1817
1818         /*
1819          * start new live ranges for values used by remats at end of block
1820          * and assure the remat args are available
1821          */
1822         foreach_pre_remat(si, bb, tmp) {
1823                 op_t     *remat_op = get_irn_link(tmp);
1824                 int       n;
1825
1826                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1827                         ir_node        *remat_arg = get_irn_n(tmp, n);
1828                         op_t           *arg_op = get_irn_link(remat_arg);
1829                         ilp_var_t       prev_lr;
1830
1831                         if(!has_reg_class(si, remat_arg)) continue;
1832
1833                         /* if value is becoming live through use by remat */
1834                         if(!pset_find_ptr(live, remat_arg)) {
1835                                 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1836                                 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1837
1838                                 arg_op->attr.live_range.ilp = prev_lr;
1839                                 arg_op->attr.live_range.op = bb;
1840
1841                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1842
1843                                 pset_insert_ptr(live, remat_arg);
1844                                 add_to_spill_bb(si, bb, remat_arg);
1845                         }
1846
1847                         /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1848                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1849                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1850
1851                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1852                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1853
1854                         /* use reload placed for this argument */
1855                         if(spill_bb->reloads) {
1856                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1857
1858                                 if(keyval) {
1859                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1860
1861                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1862                                 }
1863                         }
1864                 }
1865         }
1866         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1867
1868
1869
1870
1871         /**************************************
1872          *    B A S I C  B L O C K  B O D Y
1873          **************************************/
1874
1875         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1876                 op_t       *op;
1877                 op_t       *tmp_op;
1878                 int         n,
1879                                         u = 0,
1880                                         d = 0;
1881                 ilp_cst_t       check_pre,
1882                                         check_post;
1883                 set        *args;
1884                 pset       *used;
1885                 pset       *remat_defs;
1886                 keyval_t   *keyval;
1887                 ilp_cst_t   one_memoperand;
1888
1889                 /* iterate only until first phi */
1890                 if(is_Phi(irn))
1891                         break;
1892
1893                 op = get_irn_link(irn);
1894                 /* skip remats */
1895                 if(op->is_remat) continue;
1896
1897                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1898
1899                 /* collect defined values */
1900                 if(has_reg_class(si, irn)) {
1901                         pset_insert_ptr(defs, irn);
1902                 }
1903
1904                 /* skip projs */
1905                 if(is_Proj(irn)) continue;
1906
1907                 /*
1908                  * init set of irn's arguments
1909                  * and all possibly used values around this op
1910                  * and values defined by post remats
1911                  */
1912                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1913                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1914                 remat_defs = pset_new_ptr(pset_count(live));
1915
1916                 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1917                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
1918                                 ir_node        *irn_arg = get_irn_n(irn, n);
1919                                 if(has_reg_class(si, irn_arg)) {
1920                                         set_insert_keyval(args, irn_arg, (void*)n);
1921                                         pset_insert_ptr(used, irn_arg);
1922                                 }
1923                         }
1924                         foreach_post_remat(irn, tmp) {
1925                                 op_t    *remat_op = get_irn_link(tmp);
1926
1927                                 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1928
1929                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1930                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1931                                         if(has_reg_class(si, remat_arg)) {
1932                                                 pset_insert_ptr(used, remat_arg);
1933                                         }
1934                                 }
1935                         }
1936                         foreach_pre_remat(si, irn, tmp) {
1937                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1938                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1939                                         if(has_reg_class(si, remat_arg)) {
1940                                                 pset_insert_ptr(used, remat_arg);
1941                                         }
1942                                 }
1943                         }
1944                 }
1945
1946                 /**********************************
1947                  *   I N  E P I L O G  O F  irn
1948                  **********************************/
1949
1950                 /* ensure each dying value is used by only one post remat */
1951                 pset_foreach(used, tmp) {
1952                         ir_node     *value = tmp;
1953                         op_t        *value_op = get_irn_link(value);
1954                         ir_node     *remat;
1955                         int          n_remats = 0;
1956
1957                         cst = ILP_UNDEF;
1958                         foreach_post_remat(irn, remat) {
1959                                 op_t  *remat_op = get_irn_link(remat);
1960
1961                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1962                                         ir_node   *remat_arg = get_irn_n(remat, n);
1963
1964                                         /* if value is used by this remat add it to constraint */
1965                                         if(remat_arg == value) {
1966                                                 if(n_remats == 0) {
1967                                                         /* sum remat2s <= 1 + n_remats*live_range */
1968                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1969                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1970                                                 }
1971
1972                                                 n_remats++;
1973                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1974                                                 break;
1975                                         }
1976                                 }
1977                         }
1978
1979                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1980                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1981                         }
1982                 }
1983
1984         /* ensure at least one value dies at post remat */
1985         foreach_post_remat(irn, tmp) {
1986             op_t     *remat_op = get_irn_link(tmp);
1987             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
1988             ir_node  *remat_arg;
1989
1990             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1991                 remat_arg = get_irn_n(tmp, n);
1992
1993                 if(has_reg_class(si, remat_arg)) {
1994
1995                     /* does arg always die at this op? */
1996                     if(!pset_find_ptr(live, remat_arg))
1997                         goto skip_one_must_die;
1998
1999                     pset_insert_ptr(remat_args, remat_arg);
2000                 }
2001             }
2002
2003             /* remat + \sum live_range(remat_arg) <= |args| */
2004             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2005             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2006             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2007
2008             pset_foreach(remat_args, remat_arg) {
2009                 op_t  *arg_op = get_irn_link(remat_arg);
2010
2011                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2012             }
2013
2014 skip_one_must_die:
2015             del_pset(remat_args);
2016         }
2017
2018                 /* new live ranges for values from L\U defined by post remats */
2019                 pset_foreach(live, tmp) {
2020                         ir_node     *value = tmp;
2021                         op_t        *value_op = get_irn_link(value);
2022
2023                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2024                                 ilp_var_t    prev_lr = ILP_UNDEF;
2025                                 ir_node     *remat;
2026
2027                                 if(pset_find_ptr(remat_defs, value)) {
2028
2029                                         /* next_live_range <= prev_live_range + sum remat2s */
2030                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2031                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2032
2033                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2034                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2035
2036                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2037                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2038
2039                                         foreach_post_remat(irn, remat) {
2040                                                 op_t        *remat_op = get_irn_link(remat);
2041
2042                                                 /* if value is being rematerialized by this remat */
2043                                                 if(value == remat_op->attr.remat.remat->value) {
2044                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2045                                                 }
2046                                         }
2047
2048                                         value_op->attr.live_range.ilp = prev_lr;
2049                                         value_op->attr.live_range.op = irn;
2050                                 }
2051                         }
2052                 }
2053
2054                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2055                 foreach_post_remat(irn, tmp) {
2056                         op_t        *remat_op = get_irn_link(tmp);
2057                         int          n;
2058
2059                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2060                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2061                                 op_t           *arg_op = get_irn_link(remat_arg);
2062
2063                                 if(!has_reg_class(si, remat_arg)) continue;
2064
2065                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2066                                 if(!pset_find_ptr(used, remat_arg)) {
2067                                         /* remat <= live_range(remat_arg) */
2068                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2069                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2070
2071                                         /* if value is becoming live through use by remat2 */
2072                                         if(!pset_find_ptr(live, remat_arg)) {
2073                                                 ilp_var_t     lr;
2074
2075                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2076                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2077
2078                                                 arg_op->attr.live_range.ilp = lr;
2079                                                 arg_op->attr.live_range.op = irn;
2080
2081                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2082
2083                                                 pset_insert_ptr(live, remat_arg);
2084                                                 add_to_spill_bb(si, bb, remat_arg);
2085                                         }
2086
2087                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2088                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2089                                 }
2090                         }
2091                 }
2092
2093                 d = pset_count(defs);
2094                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2095
2096                 /* count how many regs irn needs for arguments */
2097                 u = set_count(args);
2098
2099
2100                 /* check the register pressure in the epilog */
2101                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2102                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2103                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2104
2105                 /* add L\U' to check_post */
2106                 pset_foreach(live, tmp) {
2107                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2108                                 /* if a live value is not used by irn */
2109                                 tmp_op = get_irn_link(tmp);
2110                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2111                         }
2112                 }
2113
2114                 /***********************************************************
2115                  *  I T E R A T I O N  O V E R  U S E S  F O R  E P I L O G
2116                  **********************************************************/
2117
2118
2119                 pset_foreach(used, tmp) {
2120                         ilp_var_t       prev_lr;
2121                         ilp_var_t       post_use;
2122                         int             p = 0;
2123                         spill_t        *spill;
2124                         ir_node        *arg = tmp;
2125                         op_t           *arg_op = get_irn_link(arg);
2126                         ir_node        *remat;
2127
2128                         spill = add_to_spill_bb(si, bb, arg);
2129
2130                         /* new live range for each used value */
2131                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2132                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2133
2134                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2135                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2136                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2137
2138                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2139
2140                         /* arg is live throughout epilog if the next live_range is in a register */
2141                         if(pset_find_ptr(live, arg)) {
2142                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2143
2144                                 /* post_use >= next_lr + remat */
2145                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2146                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2147                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2148                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2149                         }
2150
2151                         /* forall post remat which use arg add a similar cst */
2152                         foreach_post_remat(irn, remat) {
2153                                 int      n;
2154
2155                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2156                                         ir_node    *remat_arg = get_irn_n(remat, n);
2157                                         op_t       *remat_op = get_irn_link(remat);
2158
2159                                         if(remat_arg == arg) {
2160                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2161
2162                                                 /* post_use >= remat */
2163                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2164                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2165                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2166                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2167                                         }
2168                                 }
2169                         }
2170
2171                         /* if value is not an arg of op and not possibly defined by post remat
2172                          * then it may only die and not become live
2173                          */
2174                         if(!set_find_keyval(args, arg)) {
2175                                 /* post_use <= prev_lr */
2176                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2177                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2178                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2179                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2180
2181                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2182                                         /* next_lr <= prev_lr */
2183                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2184                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2185                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2186                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2187                                 }
2188                         }
2189
2190                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2191                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2192                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2193                                                 ilp_var_t       memoperand;
2194
2195                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2196                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2197                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2198
2199                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2200                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2201
2202                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2203                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2204                                         }
2205                                 }
2206                         }
2207
2208                         /* new live range begins for each used value */
2209                         arg_op->attr.live_range.ilp = prev_lr;
2210                         arg_op->attr.live_range.op = irn;
2211
2212                         pset_insert_ptr(live, arg);
2213                 }
2214
2215                 /* just to be sure */
2216                 check_post = ILP_UNDEF;
2217
2218                 /* allow original defintions to be removed */
2219                 if(opt_repair_schedule) {
2220                         pset_foreach(defs, tmp) {
2221                                 op_t      *tmp_op = get_irn_link(tmp);
2222                                 ilp_var_t  keep;
2223                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2224                                 assert(spill);
2225
2226                                 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2227                                 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2228
2229                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2230                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2231                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2232                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2233
2234                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2235                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2236                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2237                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2238                         }
2239                 }
2240
2241
2242                 /******************
2243                  *   P R O L O G
2244                  ******************/
2245
2246                 /* check the register pressure in the prolog */
2247                 /* sum_{L\U} lr <= k - |U| */
2248                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2249                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2250
2251                 /* for the prolog remove defined values from the live set */
2252                 pset_foreach(defs, tmp) {
2253                         pset_remove_ptr(live, tmp);
2254                 }
2255
2256                 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2257                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2258                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2259                 }
2260
2261                 /***********************************************************
2262                  *  I T E R A T I O N  O V E R  A R G S  F O R  P R O L O G
2263                  **********************************************************/
2264
2265
2266                 set_foreach(args, keyval) {
2267                         spill_t          *spill;
2268                         const ir_node    *arg = keyval->key;
2269                         int               i = PTR_TO_INT(keyval->val);
2270                         op_t             *arg_op = get_irn_link(arg);
2271                         ilp_cst_t         requirements;
2272                         int               n_memoperands;
2273
2274                         spill = set_find_spill(spill_bb->ilp, arg);
2275                         assert(spill);
2276
2277                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2278                         op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2279
2280                         /* reload <= mem_out */
2281                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2282                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2283                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2284                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2285
2286                         /* requirement: arg must be in register for use */
2287                         /* reload + remat + live_range == 1 */
2288                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2289                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2290
2291                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2292                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2293                         foreach_pre_remat(si, irn, tmp) {
2294                                 op_t     *remat_op = get_irn_link(tmp);
2295                                 if(remat_op->attr.remat.remat->value == arg) {
2296                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2297                                 }
2298                         }
2299
2300                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2301                                 n_memoperands = 0;
2302                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2303                                         if(get_irn_n(irn, n) == arg) {
2304                                                 n_memoperands++;
2305                                         }
2306                                 }
2307                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2308                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2309                                                 memoperand_t  *memoperand;
2310                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2311
2312                                                 /* memoperand <= mem_out */
2313                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2314                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2315                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2316                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2317
2318                                                 /* the memoperand is only sufficient if it is used once by the op */
2319                                                 if(n_memoperands == 1)
2320                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2321
2322                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2323
2324                                                 /* we have one more free register if we use a memory operand */
2325                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2326                                         }
2327                                 }
2328                         }
2329                 }
2330
2331                 /* iterate over L\U */
2332                 pset_foreach(live, tmp) {
2333                         if(!set_find_keyval(args, tmp)) {
2334                                 /* if a live value is not used by irn */
2335                                 tmp_op = get_irn_link(tmp);
2336                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2337                         }
2338                 }
2339
2340                 /* requirements for remats */
2341                 foreach_pre_remat(si, irn, tmp) {
2342                         op_t        *remat_op = get_irn_link(tmp);
2343                         int          n;
2344
2345                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2346                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2347                                 op_t           *arg_op = get_irn_link(remat_arg);
2348
2349                                 if(!has_reg_class(si, remat_arg)) continue;
2350
2351                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2352                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2353                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2354
2355                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2356                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2357
2358                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2359                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2360                                         int    index = (int)keyval->val;
2361
2362                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2363                                 }
2364                         }
2365                 }
2366
2367
2368
2369
2370                 /*************************
2371                  *  D O N E  W I T H  O P
2372                  *************************/
2373
2374                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2375
2376                 pset_foreach(live, tmp) {
2377                         assert(has_reg_class(si, tmp));
2378                 }
2379
2380                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2381                         ir_node        *arg = get_irn_n(irn, n);
2382
2383                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2384                 }
2385
2386                 del_pset(remat_defs);
2387                 del_pset(used);
2388                 del_set(args);
2389                 del_pset(defs);
2390                 defs = pset_new_ptr_default();
2391
2392                 /* skip everything above barrier in start block */
2393                 if(is_start_block(bb) && be_is_Barrier(irn)) {
2394                         assert(pset_count(live) == 0);
2395                         break;
2396                 }
2397
2398         }
2399         del_pset(defs);
2400
2401
2402
2403         /***************************************
2404          *   B E G I N N I N G  O F  B L O C K
2405          ***************************************/
2406
2407
2408         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2409         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2410
2411         pset_foreach(live, irn) {
2412                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2413         }
2414
2415         /* construct mem_outs for all values */
2416         set_foreach(spill_bb->ilp, spill) {
2417                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2418                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2419
2420                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2421                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2422
2423                 if(pset_find_ptr(live, spill->irn)) {
2424                         int default_spilled;
2425                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2426
2427                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2428                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2429                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2430                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2431
2432                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2433                                 int   n;
2434                                 op_t *op = get_irn_link(spill->irn);
2435
2436                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2437                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2438                                         double          freq=0.0;
2439                                         int             m;
2440                                         ilp_var_t       var;
2441
2442
2443                                         /* argument already done? */
2444                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2445
2446                                         /* get sum of execution frequencies of blocks with the same phi argument */
2447                                         for(m=n; m>=0; --m) {
2448                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2449
2450                                                 if(arg==arg2) {
2451                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2452                                                 }
2453                                         }
2454
2455                                         /* copies are not for free */
2456                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2457                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2458
2459                                         for(m=n; m>=0; --m) {
2460                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2461
2462                                                 if(arg==arg2) {
2463                                                         op->attr.live_range.args.copies[m] = var;
2464                                                 }
2465                                         }
2466
2467 #if 0
2468                                         /* copy <= mem_in */
2469                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2470                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2471                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2472                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2473 #endif
2474                                 }
2475                         }
2476                 }
2477         }
2478
2479         foreach_post_remat(bb, tmp) {
2480                 int         n;
2481                 op_t       *remat_op = get_irn_link(tmp);
2482                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2483                 ir_node    *remat_arg;
2484
2485                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2486                         remat_arg = get_irn_n(tmp, n);
2487
2488                         if(has_reg_class(si, remat_arg)) {
2489                                 pset_insert_ptr(remat_args, remat_arg);
2490                         }
2491                 }
2492
2493                 /* remat + \sum live_range(remat_arg) <= |args| */
2494                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2495                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2496                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2497
2498                 pset_foreach(remat_args, remat_arg) {
2499                         if(pset_find_ptr(live, remat_arg)) {
2500                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2501                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2502                         }
2503                 }
2504                 del_pset(remat_args);
2505         }
2506
2507         foreach_post_remat(bb, tmp) {
2508                 int  n;
2509
2510                 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2511                         ir_node  *remat_arg = get_irn_n(tmp, n);
2512
2513                         /* if value is becoming live through use by remat2 */
2514                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2515                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2516                                 ilp_cst_t   nomem;
2517
2518                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2519
2520                                 pset_insert_ptr(live, remat_arg);
2521                                 spill = add_to_spill_bb(si, bb, remat_arg);
2522                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2523
2524                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2525                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2526                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2527                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2528                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2529
2530
2531                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2532                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2533                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2534                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2535                         }
2536                 }
2537         }
2538
2539         /* L\U is empty at bb start */
2540         /* arg is live throughout epilog if it is reg_in into this block */
2541
2542         /* check the register pressure at the beginning of the block
2543          * including remats
2544          */
2545         /* reg_in entspricht post_use */
2546
2547         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2548         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2549
2550         pset_foreach(live, irn) {
2551         ilp_cst_t  nospill;
2552
2553                 spill = set_find_spill(spill_bb->ilp, irn);
2554                 assert(spill);
2555
2556                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2557                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2558
2559                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2560
2561                 /* spill + mem_in <= 1 */
2562                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2563                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2564
2565                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2566                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2567
2568         } /* post_remats are NOT included in register pressure check because
2569            they do not increase regpressure */
2570
2571         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2572         pset_foreach(live, irn) {
2573                 int          p = 0,
2574                                          n;
2575
2576                 spill = set_find_spill(spill_bb->ilp, irn);
2577                 assert(spill && spill->irn == irn);
2578
2579                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2580                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2581                                 ilp_cst_t       mem_in,
2582                                                                 reg_in;
2583                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2584                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2585                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2586                                 spill_t        *spill_p;
2587                                 op_t           *op = get_irn_link(irn);
2588
2589                                 /* although the phi is in the right regclass one or more of
2590                                  * its arguments can be in a different one or at least to
2591                                  * ignore
2592                                  */
2593                                 if(has_reg_class(si, phi_arg)) {
2594                                         /* mem_in < mem_out_arg + copy */
2595                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2596                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2597
2598                                         /* reg_in < reg_out_arg */
2599                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2600                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2601
2602                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2603                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2604
2605                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2606                                         assert(spill_p);
2607
2608                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2609                                         if(opt_memcopies)
2610                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2611
2612                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2613                                 }
2614                         }
2615                 } else {
2616                         /* else assure the value arrives on all paths in the same resource */
2617
2618                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2619                                 ilp_cst_t       mem_in,
2620                                                                 reg_in;
2621                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2622                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2623                                 spill_t        *spill_p;
2624
2625                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2626                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2627                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2628                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2629
2630                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2631                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2632
2633                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2634                                 assert(spill_p);
2635
2636                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2637                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2638                         }
2639                 }
2640         }
2641
2642         foreach_post_remat(bb, tmp) {
2643                 int         n;
2644
2645                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2646                         ir_node    *remat_arg = get_irn_n(tmp, n);
2647                         op_t       *remat_op = get_irn_link(tmp);
2648
2649                         if(!has_reg_class(si, remat_arg)) continue;
2650
2651                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2652                         assert(spill);
2653
2654                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2655                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2656                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2657                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2658                 }
2659         }
2660
2661         pset_foreach(live, irn) {
2662                 const op_t      *op = get_irn_link(irn);
2663                 const ir_node   *remat;
2664                 int              n_remats = 0;
2665
2666                 cst = ILP_UNDEF;
2667
2668                 foreach_post_remat(bb, remat) {
2669                         int   n;
2670
2671                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2672                                 const ir_node  *arg = get_irn_n(remat, n);
2673
2674                                 if(arg == irn) {
2675                                         const op_t   *remat_op = get_irn_link(remat);
2676
2677                                         if(cst == ILP_UNDEF) {
2678                                                 /* sum remat2s <= 1 + n_remats*live_range */
2679                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2680                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2681                                         }
2682                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2683                                         ++n_remats;
2684                                         break;
2685                                 }
2686                         }
2687                 }
2688                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2689                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2690                 }
2691         }
2692
2693         /* first live ranges from reg_ins */
2694         pset_foreach(live, irn) {
2695                 op_t      *op = get_irn_link(irn);
2696
2697                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2698
2699                         spill = set_find_spill(spill_bb->ilp, irn);
2700                         assert(spill && spill->irn == irn);
2701
2702                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2703                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2704                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2705                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2706
2707                         foreach_post_remat(bb, tmp) {
2708                                 op_t     *remat_op = get_irn_link(tmp);
2709
2710                                 if(remat_op->attr.remat.remat->value == irn) {
2711                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2712                                 }
2713                         }
2714                 }
2715         }
2716
2717         /* walk forward now and compute constraints for placing spills */
2718         /* this must only be done for values that are not defined in this block */
2719         pset_foreach(live, irn) {
2720                 /*
2721                  * if value is defined in this block we can anways place the spill directly after the def
2722                  *    -> no constraint necessary
2723                  */
2724                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2725                         assert(0);
2726                 }
2727
2728
2729                 spill = set_find_spill(spill_bb->ilp, irn);
2730                 assert(spill);
2731
2732                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2733                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2734
2735                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2736                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2737
2738                 if(!is_Phi(irn)) {
2739                         sched_foreach_op(bb, tmp) {
2740                                 op_t   *op = get_irn_link(tmp);
2741
2742                                 if(is_Phi(tmp)) continue;
2743                                 assert(!is_Proj(tmp));
2744
2745                                 if(op->is_remat) {
2746                                         const ir_node   *value = op->attr.remat.remat->value;
2747
2748                                         if(value == irn) {
2749                                                 /* only collect remats up to the first real use of a value */
2750                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2751                                         }
2752                                 } else {
2753                                         int   n;
2754
2755                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2756                                                 ir_node    *arg = get_irn_n(tmp, n);
2757
2758                                                 if(arg == irn) {
2759                                                         /* if a value is used stop collecting remats */
2760                             goto next_live;
2761                                                 }
2762                                         }
2763                                 }
2764                         }
2765                 }
2766 next_live: ;
2767         }
2768
2769         del_pset(live);
2770 }
2771
2772 typedef struct _irnlist_t {
2773         struct list_head   list;
2774         ir_node           *irn;
2775 } irnlist_t;
2776
2777 typedef struct _interference_t {
2778         struct list_head    blocklist;
2779         ir_node            *a;
2780         ir_node            *b;
2781 } interference_t;
2782
2783 static int
2784 cmp_interference(const void *a, const void *b, size_t size)
2785 {
2786         const interference_t *p = a;
2787         const interference_t *q = b;
2788
2789         return !(p->a == q->a && p->b == q->b);
2790 }
2791
2792 static interference_t *
2793 set_find_interference(set * set, ir_node * a, ir_node * b)
2794 {
2795         interference_t     query;
2796
2797         query.a = (a>b)?a:b;
2798         query.b = (a>b)?b:a;
2799
2800         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2801 }
2802
2803 static interference_t *
2804 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2805 {
2806         interference_t     query,
2807                                           *result;
2808         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2809
2810         list->irn = bb;
2811
2812         result = set_find_interference(set, a, b);
2813         if(result) {
2814
2815                 list_add(&list->list, &result->blocklist);
2816                 return result;
2817         }
2818
2819         query.a = (a>b)?a:b;
2820         query.b = (a>b)?b:a;
2821
2822         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2823
2824         INIT_LIST_HEAD(&result->blocklist);
2825         list_add(&list->list, &result->blocklist);
2826
2827         return result;
2828 }
2829
2830 static int
2831 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2832 {
2833         const ir_edge_t *edge;
2834
2835         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2836                 /* both values are live in, so they interfere */
2837                 return 1;
2838         }
2839
2840         /* ensure a dominates b */
2841         if(value_dominates(b,a)) {
2842                 const ir_node * t;
2843                 t = b;
2844                 b = a;
2845                 a = t;
2846         }
2847         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2848
2849
2850         /* the following code is stolen from bera.c */
2851         if(be_is_live_end(si->lv, bb, a))
2852                 return 1;
2853
2854         foreach_out_edge(a, edge) {
2855                 const ir_node *user = edge->src;
2856                 if(get_nodes_block(user) == bb
2857                                 && !is_Phi(user)
2858                                 && b != user
2859                                 && !pset_find_ptr(si->inverse_ops, user)
2860                                 && value_dominates(b, user))
2861                         return 1;
2862         }
2863
2864         return 0;
2865 }
2866
2867 /**
2868  * Walk all irg blocks and collect interfering values inside of phi classes
2869  */
2870 static void
2871 luke_interferencewalker(ir_node * bb, void * data)
2872 {
2873         spill_ilp_t    *si = (spill_ilp_t*)data;
2874         int             l1, l2;
2875
2876         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2877                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2878                 op_t           *a_op = get_irn_link(a);
2879
2880
2881                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2882                 if (has_reg_class(si, a) && get_phi_class(a)) {
2883                         if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2884                                 continue;
2885
2886                         for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
2887                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
2888                                 op_t           *b_op = get_irn_link(b);
2889
2890
2891                                 /* a and b are only interesting if they are in the same phi class */
2892                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2893                                         if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2894                                                 continue;
2895
2896                                         if(values_interfere_in_block(si, bb, a, b)) {
2897                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2898                                                 set_insert_interference(si, si->interferences, a, b, bb);
2899                                         }
2900                                 }
2901                         }
2902                 }
2903         }
2904 }
2905
2906 static unsigned int copy_path_id = 0;
2907
2908 static void
2909 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2910 {
2911         ilp_cst_t  cst;
2912         ilp_var_t  copy;
2913         char       buf[256];
2914         void      *ptr;
2915
2916         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2917         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2918
2919         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2920
2921         pset_foreach(copies, ptr) {
2922                 copy = PTR_TO_INT(ptr);
2923                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2924         }
2925 }
2926
2927 /**
2928  * @parameter copies   contains a path of copies which lead us to irn
2929  * @parameter visited  contains a set of nodes already visited on this path
2930  */
2931 static int
2932 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2933 {
2934         const ir_edge_t *edge;
2935         op_t            *op = get_irn_link(irn);
2936     pset            *visited_users = pset_new_ptr_default();
2937         int              paths = 0;
2938
2939         if(op->is_remat) return 0;
2940
2941         pset_insert_ptr(visited, irn);
2942
2943         if(is_Phi(irn)) {
2944                 int    n;
2945         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2946
2947                 /* visit all operands */
2948                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2949                         ir_node  *arg = get_irn_n(irn, n);
2950                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2951
2952                         if(!has_reg_class(si, arg)) continue;
2953             if(pset_find_ptr(visited_operands, arg)) continue;
2954             pset_insert_ptr(visited_operands, arg);
2955
2956                         if(arg == target) {
2957                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2958                                         del_pset(visited_operands);
2959                                         del_pset(visited_users);
2960                                         pset_remove_ptr(visited, irn);
2961                                         return paths;
2962                                 }
2963                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2964                                 write_copy_path_cst(si, copies, any_interfere);
2965                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2966                         } else if(!pset_find_ptr(visited, arg)) {
2967                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2968                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2969                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2970
2971                 if(paths > MAX_PATHS) {
2972                     if(pset_count(copies) == 0) {
2973                         ilp_cst_t  cst;
2974                         char       buf[256];
2975
2976                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2977                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2978                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2979                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2980                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2981
2982                         paths = 0;
2983                     } else {
2984                         del_pset(visited_operands);
2985                         del_pset(visited_users);
2986                         pset_remove_ptr(visited, irn);
2987                         return paths;
2988                     }
2989                 } else if(pset_count(copies) == 0) {
2990                                         paths = 0;
2991                                 }
2992                         }
2993                 }
2994
2995         del_pset(visited_operands);
2996         }
2997
2998         /* visit all uses which are phis */
2999         foreach_out_edge(irn, edge) {
3000                 ir_node  *user = edge->src;
3001                 int       pos  = edge->pos;
3002                 op_t     *op = get_irn_link(user);
3003                 ilp_var_t copy;
3004
3005                 if(!is_Phi(user)) continue;
3006                 if(!has_reg_class(si, user)) continue;
3007         if(pset_find_ptr(visited_users, user)) continue;
3008         pset_insert_ptr(visited_users, user);
3009
3010                 copy = op->attr.live_range.args.copies[pos];
3011
3012                 if(user == target) {
3013                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3014                                 del_pset(visited_users);
3015                                 pset_remove_ptr(visited, irn);
3016                                 return paths;
3017                         }
3018                         pset_insert(copies, INT_TO_PTR(copy), copy);
3019                         write_copy_path_cst(si, copies, any_interfere);
3020                         pset_remove(copies, INT_TO_PTR(copy), copy);
3021                 } else if(!pset_find_ptr(visited, user)) {
3022                         pset_insert(copies, INT_TO_PTR(copy), copy);
3023                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3024                         pset_remove(copies, INT_TO_PTR(copy), copy);
3025
3026             if(paths > MAX_PATHS) {
3027                 if(pset_count(copies) == 0) {
3028                     ilp_cst_t  cst;
3029                     char       buf[256];
3030
3031                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3032                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3033                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3034                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3035                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3036
3037                     paths = 0;
3038                 } else {
3039                     del_pset(visited_users);
3040                     pset_remove_ptr(visited, irn);
3041                     return paths;
3042                 }
3043             } else if(pset_count(copies) == 0) {
3044                                 paths = 0;
3045                         }
3046                 }
3047         }
3048
3049     del_pset(visited_users);
3050         pset_remove_ptr(visited, irn);
3051         return paths;
3052 }
3053
3054 static void
3055 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3056 {
3057         pset * copies = pset_new_ptr_default();
3058         pset * visited = pset_new_ptr_default();
3059
3060         find_copy_path(si, a, b, any_interfere, copies, visited);
3061
3062         del_pset(visited);
3063         del_pset(copies);
3064 }
3065
3066
3067 static void
3068 memcopyhandler(spill_ilp_t * si)
3069 {
3070         interference_t   *interference;
3071         char              buf[256];
3072         /* teste Speicherwerte auf Interferenz */
3073
3074         /* analyze phi classes */
3075         phi_class_compute(si->chordal_env->irg);
3076
3077         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3078         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3079
3080         /* now lets emit the ILP unequations for the crap */
3081         set_foreach(si->interferences, interference) {
3082                 irnlist_t      *irnlist;
3083                 ilp_var_t       interfere,
3084                                                 any_interfere;
3085                 ilp_cst_t       any_interfere_cst,
3086                                                 cst;
3087                 const ir_node  *a  = interference->a;
3088                 const ir_node  *b  = interference->b;
3089
3090                 /* any_interf <= \sum interf */
3091                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3092                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3093                 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3094
3095                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3096
3097                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3098                         const ir_node  *bb = irnlist->irn;
3099                         spill_bb_t     *spill_bb = get_irn_link(bb);
3100                         spill_t        *spilla,
3101                                                    *spillb;
3102                         char           buf[256];
3103
3104                         spilla = set_find_spill(spill_bb->ilp, a);
3105                         assert(spilla);
3106
3107                         spillb = set_find_spill(spill_bb->ilp, b);
3108                         assert(spillb);
3109
3110                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3111                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3112                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3113                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3114                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3115                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3116
3117                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3118                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3119
3120                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3121                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3122                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3123                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3124                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3125
3126                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3127                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3128
3129                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3130                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3131                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3132
3133                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3134                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3135
3136                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3137                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3138                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3139
3140
3141                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3142
3143                         /* any_interfere >= interf */
3144                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3145                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3146
3147                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3148                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3149                 }
3150
3151                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3152                 gen_copy_constraints(si,a,b,any_interfere);
3153         }
3154 }
3155
3156
3157 static INLINE int
3158 is_zero(double x)
3159 {
3160         return fabs(x) < 0.00001;
3161 }
3162
3163 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3164 {
3165         spill_ilp_t *si = get_irg_link(current_ir_graph);
3166
3167         if(pset_find_ptr(si->all_possible_remats, n)) {
3168                 op_t   *op = (op_t*)get_irn_link(n);
3169                 assert(op && op->is_remat);
3170
3171                 if(!op->attr.remat.remat->inverse) {
3172                         if(op->attr.remat.pre) {
3173                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3174                         } else {
3175                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3176                         }
3177
3178                         return 1;
3179                 } else {
3180                         op_t   *op = (op_t*)get_irn_link(n);
3181                         assert(op && op->is_remat);
3182
3183                         if(op->attr.remat.pre) {
3184                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3185                         } else {
3186                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3187                         }
3188
3189                         return 1;
3190                 }
3191         }
3192
3193         return 0;
3194 }
3195
3196 static void
3197 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3198 {
3199         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3200         be_dump(irg, suffix, dump_ir_block_graph_sched);
3201         set_dump_node_vcgattr_hook(NULL);
3202 }
3203
3204 /**
3205  * Edge hook to dump the schedule edges with annotated register pressure.
3206  */
3207 static int
3208 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3209 {
3210         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3211                 ir_node *prev = sched_prev(irn);
3212                 fprintf(F, "edge:{sourcename:\"");
3213                 PRINT_NODEID(irn);
3214                 fprintf(F, "\" targetname:\"");
3215                 PRINT_NODEID(prev);
3216                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3217                 fprintf(F, "\" color:magenta}\n");
3218         }
3219         return 1;
3220 }
3221
3222 static void
3223 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3224 {
3225         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3226
3227         dump_consts_local(0);
3228         set_dump_node_edge_hook(sched_pressure_edge_hook);
3229         dump_ir_block_graph(irg, suffix);
3230         set_dump_node_edge_hook(old_edge_hook);
3231 }
3232
3233 static void
3234 walker_pressure_annotator(ir_node * bb, void * data)
3235 {
3236         spill_ilp_t  *si = data;
3237         ir_node      *irn;
3238         int           n, i;
3239         pset         *live = pset_new_ptr_default();
3240         int           projs = 0;
3241
3242         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3243                 irn = be_lv_get_irn(si->lv, bb, i);
3244
3245                 if (has_reg_class(si, irn)) {
3246                         pset_insert_ptr(live, irn);
3247                 }
3248         }
3249
3250         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3251
3252         sched_foreach_reverse(bb, irn) {
3253                 if(is_Phi(irn)) {
3254                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3255                         continue;
3256                 }
3257
3258                 if(has_reg_class(si, irn)) {
3259                         pset_remove_ptr(live, irn);
3260                         if(is_Proj(irn)) ++projs;
3261                 }
3262
3263                 if(!is_Proj(irn)) projs = 0;
3264
3265                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3266                         ir_node    *arg = get_irn_n(irn, n);
3267
3268                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3269                 }
3270                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3271         }
3272
3273         del_pset(live);
3274 }
3275
3276 static void
3277 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3278 {
3279         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3280 }
3281
3282 static void
3283 connect_all_remats_with_keep(spill_ilp_t * si)
3284 {
3285         ir_node   *irn;
3286         ir_node  **ins,
3287                          **pos;
3288         int        n_remats;
3289
3290
3291         n_remats = pset_count(si->all_possible_remats);
3292         if(n_remats) {
3293                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3294
3295                 pos = ins;
3296                 pset_foreach(si->all_possible_remats, irn) {
3297                         *pos = irn;
3298                         ++pos;
3299                 }
3300
3301                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3302
3303                 obstack_free(si->obst, ins);
3304         }
3305 }
3306
3307 static void
3308 connect_all_spills_with_keep(spill_ilp_t * si)
3309 {
3310         ir_node   *irn;
3311         ir_node  **ins,
3312                          **pos;
3313         int        n_spills;
3314         ir_node   *keep;
3315
3316
3317         n_spills = pset_count(si->spills);
3318         if(n_spills) {
3319                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3320
3321                 pos = ins;
3322                 pset_foreach(si->spills, irn) {
3323                         *pos = irn;
3324                         ++pos;
3325                 }
3326
3327                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3328
3329                 obstack_free(si->obst, ins);
3330         }
3331 }
3332
3333 /** insert a spill at an arbitrary position */
3334 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3335 {
3336         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
3337         ir_graph *irg   = get_irn_irg(bl);
3338         ir_node *frame  = get_irg_frame(irg);
3339         ir_node *spill;
3340         ir_node *next;
3341
3342         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3343         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3344
3345         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3346
3347         /*
3348          * search the right insertion point. a spill of a phi cannot be put
3349          * directly after the phi, if there are some phis behind the one which
3350          * is spilled. Also, a spill of a Proj must be after all Projs of the
3351          * same tuple node.
3352          *
3353          * Here's one special case:
3354          * If the spill is in the start block, the spill must be after the frame
3355          * pointer is set up. This is done by setting insert to the end of the block
3356          * which is its default initialization (see above).
3357          */
3358
3359         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3360                 insert = frame;
3361
3362         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3363                 insert = next;
3364
3365         sched_add_after(insert, spill);
3366         return spill;
3367 }
3368
3369 static void
3370 delete_remat(spill_ilp_t * si, ir_node * remat) {
3371         int       n;
3372         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3373
3374         sched_remove(remat);
3375
3376         /* kill links to operands */
3377         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3378                 set_irn_n(remat, n, bad);
3379         }
3380 }
3381
3382 static void
3383 clean_remat_info(spill_ilp_t * si)
3384 {
3385         int            n;
3386         remat_t       *remat;
3387         remat_info_t  *remat_info;
3388         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
3389
3390         set_foreach(si->remat_info, remat_info) {
3391                 if(!remat_info->remats) continue;
3392
3393                 pset_foreach(remat_info->remats, remat)
3394                 {
3395                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3396                                 if(sched_is_scheduled(remat->proj)) {
3397                                         sched_remove((ir_node*)remat->proj);
3398                                 }
3399                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3400                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3401                         }
3402
3403                         if(get_irn_n_edges(remat->op) == 0) {
3404                                 if(sched_is_scheduled(remat->op)) {
3405                                         sched_remove((ir_node*)remat->op);
3406                                 }
3407                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3408                                         set_irn_n((ir_node*)remat->op, n, bad);
3409                                 }
3410                         }
3411                 }
3412
3413                 if(remat_info->remats) del_pset(remat_info->remats);
3414                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3415         }
3416 }
3417
3418 static void
3419 delete_unnecessary_remats(spill_ilp_t * si)
3420 {
3421         if(opt_keep_alive & KEEPALIVE_REMATS) {
3422                 int       n;
3423                 ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3424
3425                 if(si->keep) {
3426                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3427                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3428                                 op_t           *arg_op = get_irn_link(keep_arg);
3429                                 lpp_name_t     *name;
3430
3431                                 assert(arg_op->is_remat);
3432
3433                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3434
3435                                 if(is_zero(name->value)) {
3436                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3437                                         /* TODO check whether reload is preferred over remat (could be bug) */
3438                                         delete_remat(si, keep_arg);
3439                                 } else {
3440                                         if(!arg_op->attr.remat.remat->inverse) {
3441                                                 if(arg_op->attr.remat.pre) {
3442                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3443                                                 } else {
3444                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3445                                                 }
3446                                         } else {
3447                                                 if(arg_op->attr.remat.pre) {
3448                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3449                                                 } else {
3450                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3451                                                 }
3452                                         }
3453                                 }
3454
3455                                 set_irn_n(si->keep, n, bad);
3456                         }
3457                 } else {
3458                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3459                 }
3460         } else {
3461                 ir_node  *remat;
3462
3463                 pset_foreach(si->all_possible_remats, remat) {
3464                         op_t           *remat_op = get_irn_link(remat);
3465                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3466
3467                         if(is_zero(name->value)) {
3468                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3469                                 /* TODO check whether reload is preferred over remat (could be bug) */
3470                                 delete_remat(si, remat);
3471                         } else {
3472                                 if(!remat_op->attr.remat.remat->inverse) {
3473                                         if(remat_op->attr.remat.pre) {
3474                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3475                                         } else {
3476                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3477                                         }
3478                                 } else {
3479                                         if(remat_op->attr.remat.pre) {
3480                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3481                                         } else {
3482                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3483                                         }
3484                                 }
3485                         }
3486                 }
3487         }
3488 }
3489
3490 static pset *
3491 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3492 {
3493         pset     *spills = pset_new_ptr_default();
3494
3495         const ir_node  *next;
3496         defs_t         *defs;
3497
3498         defs = set_find_def(si->values, value);
3499
3500         if(defs && defs->spills) {
3501                 for(next = defs->spills; next; next = get_irn_link(next)) {
3502                         pset_insert_ptr(spills, next);
3503                 }
3504         }
3505
3506         return spills;
3507 }
3508
3509 /**
3510  * @param before   The node after which the spill will be placed in the schedule
3511  */
3512 static ir_node *
3513 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3514 {
3515         defs_t   *defs;
3516         ir_node  *spill;
3517         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3518
3519         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3520
3521         spill = be_spill2(arch_env, irn, before);
3522
3523         defs = set_insert_def(si->values, value);
3524         assert(defs);
3525
3526         /* enter into the linked list */
3527         set_irn_link(spill, defs->spills);
3528         defs->spills = spill;
3529
3530         if(opt_keep_alive & KEEPALIVE_SPILLS)
3531                 pset_insert_ptr(si->spills, spill);
3532
3533         return spill;
3534 }
3535
3536 /**
3537  * @param before   The Phi node which has to be spilled
3538  */
3539 static ir_node *
3540 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3541 {
3542         ir_node   *mem_phi;
3543         ir_node  **ins;
3544         defs_t    *defs;
3545         int        n;
3546
3547         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3548
3549         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3550                 ins[n] = si->m_unknown;
3551         }
3552
3553         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3554
3555         defs = set_insert_def(si->values, phi);
3556         assert(defs);
3557
3558         /* enter into the linked list */
3559         set_irn_link(mem_phi, defs->spills);
3560         defs->spills = mem_phi;
3561
3562 #ifdef SCHEDULE_PHIM
3563         sched_add_after(phi, mem_phi);
3564 #endif
3565
3566         if(opt_keep_alive & KEEPALIVE_SPILLS)
3567                 pset_insert_ptr(si->spills, mem_phi);
3568
3569
3570         return mem_phi;
3571 }
3572
3573 /**
3574  * Add remat to list of defs, destroys link field!
3575  */
3576 static void
3577 insert_remat(spill_ilp_t * si, ir_node * remat)
3578 {
3579         defs_t   *defs;
3580         op_t     *remat_op = get_irn_link(remat);
3581
3582         assert(remat_op->is_remat);
3583
3584         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3585         assert(defs);
3586
3587         /* enter into the linked list */
3588         set_irn_link(remat, defs->remats);
3589         defs->remats = remat;
3590 }
3591
3592
3593 /**
3594  * Add reload before operation and add to list of defs
3595  */
3596 static ir_node *
3597 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3598 {
3599         defs_t   *defs;
3600         ir_node  *reload,
3601                          *spill;
3602         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3603
3604         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3605
3606         defs = set_find_def(si->values, value);
3607
3608         spill = defs->spills;
3609         assert(spill && "no spill placed before reload");
3610
3611         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3612
3613         /* enter into the linked list */
3614         set_irn_link(reload, defs->remats);
3615         defs->remats = reload;
3616
3617         return reload;
3618 }
3619
3620 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3621 {
3622         defs_t           *defs;
3623         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3624         ir_node          *spill;
3625         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3626
3627         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3628
3629         defs = set_find_def(si->values, value);
3630
3631         spill = defs->spills;
3632         assert(spill && "no spill placed before reload");
3633
3634         arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3635 }
3636
3637 void insert_memoperands(spill_ilp_t * si)
3638 {
3639         memoperand_t   *memoperand;
3640         lpp_name_t     *name;
3641
3642         set_foreach(si->memoperands, memoperand) {
3643                 name = si->lpp->vars[memoperand->ilp];
3644                 if(!is_zero(name->value)) {
3645                         perform_memory_operand(si, memoperand);
3646                 }
3647         }
3648 }
3649
3650 static void
3651 walker_spill_placer(ir_node * bb, void * data) {
3652         spill_ilp_t   *si = (spill_ilp_t*)data;
3653         ir_node       *irn;
3654         spill_bb_t    *spill_bb = get_irn_link(bb);
3655         pset          *spills_to_do = pset_new_ptr_default();
3656         spill_t       *spill;
3657
3658         set_foreach(spill_bb->ilp, spill) {
3659                 lpp_name_t    *name;
3660
3661                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3662                         name = si->lpp->vars[spill->mem_in];
3663                         if(!is_zero(name->value)) {
3664                                 ir_node   *mem_phi;
3665
3666                                 mem_phi = insert_mem_phi(si, spill->irn);
3667
3668                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3669                         }
3670                 }
3671
3672                 name = si->lpp->vars[spill->spill];
3673                 if(!is_zero(name->value)) {
3674                         /* place spill directly after definition */
3675                         if(get_nodes_block(spill->irn) == bb) {
3676                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3677                                 continue;
3678                         }
3679
3680                         /* place spill at bb start */
3681                         if(spill->reg_in > 0) {
3682                                 name = si->lpp->vars[spill->reg_in];
3683                                 if(!is_zero(name->value)) {
3684                                         insert_spill(si, spill->irn, spill->irn, bb);
3685                                         continue;
3686                                 }
3687                         }
3688                         /* place spill after a remat */
3689                         pset_insert_ptr(spills_to_do, spill->irn);
3690                 }
3691         }
3692         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3693
3694
3695         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3696                 op_t     *op = get_irn_link(irn);
3697
3698                 if(be_is_Spill(irn)) continue;
3699
3700                 if(op->is_remat) {
3701                         /* TODO fix this if we want to support remats with more than two nodes */
3702                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3703                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3704
3705                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3706                         }
3707                 } else {
3708                         if(pset_find_ptr(spills_to_do, irn)) {
3709                                 pset_remove_ptr(spills_to_do, irn);
3710
3711                                 insert_spill(si, irn, irn, irn);
3712                         }
3713                 }
3714
3715         }
3716
3717         assert(pset_count(spills_to_do) == 0);
3718
3719         /* afterwards free data in block */
3720         del_pset(spills_to_do);
3721 }
3722
3723 static ir_node *
3724 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3725 {
3726         ir_node          *insert_pos = bb;
3727         ir_node          *spill;
3728         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3729
3730         /* find last definition of arg value in block */
3731         ir_node  *next;
3732         defs_t   *defs;
3733         int       last = 0;
3734
3735         defs = set_find_def(si->values, value);
3736
3737         if(defs && defs->remats) {
3738                 for(next = defs->remats; next; next = get_irn_link(next)) {
3739                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3740                                 last = sched_get_time_step(next);
3741                                 insert_pos = next;
3742                         }
3743                 }
3744         }
3745
3746         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3747                 last = sched_get_time_step(value);
3748                 insert_pos = value;
3749         }
3750
3751         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3752
3753         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3754
3755         return spill;
3756 }
3757
3758 static void
3759 phim_fixer(spill_ilp_t *si) {
3760         defs_t  *defs;
3761
3762         set_foreach(si->values, defs) {
3763                 const ir_node  *phi = defs->value;
3764                 op_t           *op = get_irn_link(phi);
3765                 ir_node        *phi_m = NULL;
3766                 ir_node        *next = defs->spills;
3767                 int             n;
3768
3769                 if(!is_Phi(phi)) continue;
3770
3771                 while(next) {
3772                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3773                                 phi_m = next;
3774                                 break;
3775                         } else {
3776                                 next = get_irn_link(next);
3777                         }
3778                 }
3779                 if(!phi_m) continue;
3780
3781                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3782                         ir_node        *value = get_irn_n(phi, n);
3783                         defs_t         *val_defs = set_find_def(si->values, value);
3784
3785                         /* a spill of this value */
3786                         ir_node      *spill;
3787
3788
3789                         if(opt_memcopies) {
3790                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3791                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3792
3793                                 if(!is_zero(name->value)) {
3794                                         spill = insert_mem_copy(si, pred, value);
3795                                 } else {
3796                                         spill = val_defs->spills;
3797                                 }
3798                         } else {
3799                                 spill = val_defs->spills;
3800                         }
3801
3802                         assert(spill && "no spill placed before PhiM");
3803                         set_irn_n(phi_m, n, spill);
3804                 }
3805         }
3806 }
3807
3808 static void
3809 walker_reload_placer(ir_node * bb, void * data) {
3810         spill_ilp_t   *si = (spill_ilp_t*)data;
3811         ir_node       *irn;
3812         spill_bb_t    *spill_bb = get_irn_link(bb);
3813
3814         /* reloads at end of block */
3815         if(spill_bb->reloads) {
3816                 keyval_t    *keyval;
3817
3818                 set_foreach(spill_bb->reloads, keyval) {
3819                         ir_node        *irn = (ir_node*)keyval->key;
3820                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3821                         lpp_name_t     *name;
3822
3823                         name = si->lpp->vars[reload];
3824                         if(!is_zero(name->value)) {
3825                                 ir_node    *reload;
3826                                 ir_node    *insert_pos = bb;
3827                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3828                                 op_t       *prev_op = get_irn_link(prev);
3829
3830                                 while(be_is_Spill(prev)) {
3831                                         prev = sched_prev(prev);
3832                                 }
3833
3834                                 prev_op = get_irn_link(prev);
3835
3836                                 /* insert reload before pre-remats */
3837                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3838                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3839                                         insert_pos = prev;
3840
3841                                         do {
3842                                                 prev = sched_prev(prev);
3843                                         } while(be_is_Spill(prev));
3844
3845                                         prev_op = get_irn_link(prev);
3846
3847                                 }
3848
3849                                 reload = insert_reload(si, irn, insert_pos);
3850
3851                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3852                                         pset_insert_ptr(si->spills, reload);
3853                         }
3854                 }
3855         }
3856
3857         /* walk and insert more reloads and collect remats */
3858         sched_foreach_reverse(bb, irn) {
3859                 op_t     *op = get_irn_link(irn);
3860
3861                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3862                 if(is_Phi(irn)) break;
3863
3864                 if(op->is_remat) {
3865                         if(get_irn_mode(irn) != mode_T) {
3866                                 insert_remat(si, irn);
3867                         }
3868                 } else {
3869                         int    n;
3870
3871                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3872                                 ir_node    *arg = get_irn_n(irn, n);
3873
3874                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3875                                         lpp_name_t    *name;
3876
3877                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3878                                         if(!is_zero(name->value)) {
3879                                                 ir_node    *reload;
3880                                                 ir_node    *insert_pos = irn;
3881                                                 ir_node    *prev = sched_prev(insert_pos);
3882                                                 op_t       *prev_op;
3883
3884                                                 while(be_is_Spill(prev)) {
3885                                                         prev = sched_prev(prev);
3886                                                 }
3887
3888                                                 prev_op = get_irn_link(prev);
3889
3890                                                 /* insert reload before pre-remats */
3891                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3892                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3893                                                         insert_pos = prev;
3894
3895                                                         do {
3896                                                                 prev = sched_prev(prev);
3897                                                         } while(be_is_Spill(prev));
3898
3899                                                         prev_op = get_irn_link(prev);
3900
3901                                                 }
3902
3903                                                 reload = insert_reload(si, arg, insert_pos);
3904
3905                                                 set_irn_n(irn, n, reload);
3906
3907                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3908                                                         pset_insert_ptr(si->spills, reload);
3909                                         }
3910                                 }
3911                         }
3912                 }
3913         }
3914
3915         del_set(spill_bb->ilp);
3916         if(spill_bb->reloads) del_set(spill_bb->reloads);
3917 }
3918
3919 static void
3920 walker_collect_used(ir_node * irn, void * data)
3921 {
3922         bitset_t   *used = data;
3923
3924         bitset_set(used, get_irn_idx(irn));
3925 }
3926
3927 struct kill_helper {
3928         bitset_t  *used;
3929         spill_ilp_t  *si;
3930 };
3931
3932 static void
3933 walker_kill_unused(ir_node * bb, void * data)
3934 {
3935         struct kill_helper *kh = data;
3936         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
3937         ir_node            *irn;
3938
3939
3940         for(irn=sched_first(bb); !sched_is_end(irn);) {
3941                 ir_node     *next = sched_next(irn);
3942                 int          n;
3943
3944                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3945                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3946                                 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(kh->si, irn)*execution_frequency(kh->si, bb)));
3947 #if 0
3948                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3949 #endif
3950                         }
3951
3952                         sched_remove(irn);
3953
3954                         set_nodes_block(irn, bad);
3955                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3956                                 set_irn_n(irn, n, bad);
3957                         }
3958                 }
3959                 irn = next;
3960         }
3961 }
3962
3963 static void
3964 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3965 {
3966         struct kill_helper kh;
3967
3968         kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3969         kh.si = si;
3970
3971         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3972         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3973
3974         bitset_free(kh.used);
3975 }
3976
3977 void
3978 print_irn_pset(pset * p)
3979 {
3980         ir_node   *irn;
3981
3982         pset_foreach(p, irn) {
3983                 ir_printf("%+F\n", irn);
3984         }
3985 }
3986
3987 void
3988 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3989 {
3990     FILE           *f = fopen(file, "w");
3991     ir_node        *irn;
3992     interference_t *interference;
3993
3994     pset_break(phiclass);
3995     set_break(si->interferences);
3996
3997     ir_fprintf(f, "digraph phiclass {\n");
3998
3999     pset_foreach(phiclass, irn) {
4000         if(is_Phi(irn))
4001             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
4002     }
4003
4004     pset_foreach(phiclass, irn) {
4005         int n;
4006
4007         if(!is_Phi(irn)) continue;
4008
4009         for(n=get_irn_arity(irn)-1; n>=0; --n) {
4010             ir_node  *arg = get_irn_n(irn, n);
4011
4012             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
4013         }
4014     }
4015
4016     set_foreach(si->interferences, interference) {
4017         const ir_node  *a  = interference->a;
4018         const ir_node  *b  = interference->b;
4019         if(get_phi_class(a) == phiclass) {
4020             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4021         }
4022     }
4023
4024     ir_fprintf(f, "}");
4025     fclose(f);
4026 }
4027
4028 static void
4029 rewire_uses(spill_ilp_t * si)
4030 {
4031         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4032         defs_t               *defs;
4033         pset                 *ignore = pset_new_ptr(1);
4034
4035         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4036
4037         /* then fix uses of spills */
4038         set_foreach(si->values, defs) {
4039                 pset           *reloads;
4040                 pset           *spills;
4041                 const ir_node  *next = defs->remats;
4042                 int remats = 0;
4043
4044                 reloads = pset_new_ptr_default();
4045
4046                 while(next) {
4047                         if(be_is_Reload(next)) {
4048                                 pset_insert_ptr(reloads, next);
4049                         } else {
4050                                 ++remats;
4051                         }
4052                         next = get_irn_link(next);
4053                 }
4054
4055                 spills = get_spills_for_value(si, defs->value);
4056                 DBG((si->dbg, LEVEL_2, "\t  %d remats, %d reloads, and %d spills for value %+F\n", remats, pset_count(reloads), pset_count(spills), defs->value));
4057                 if(pset_count(spills) > 1) {
4058                         //assert(pset_count(reloads) > 0);
4059                         //                              print_irn_pset(spills);
4060                         //                              print_irn_pset(reloads);
4061
4062                         be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4063                 }
4064
4065                 del_pset(reloads);
4066                 del_pset(spills);
4067         }
4068
4069         /* first fix uses of remats and reloads */
4070         set_foreach(si->values, defs) {
4071                 pset           *nodes;
4072                 const ir_node  *next = defs->remats;
4073                 int             orig_kept = 0;
4074
4075                 if(next) {
4076                         nodes = pset_new_ptr_default();
4077                         if(sched_is_scheduled(defs->value)) {
4078                                 pset_insert_ptr(nodes, defs->value);
4079                                 orig_kept = 1;
4080                         }
4081
4082                         while(next) {
4083                                 pset_insert_ptr(nodes, next);
4084                                 next = get_irn_link(next);
4085                         }
4086
4087                         DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4088                         be_ssa_constr_set(dfi, si->lv, nodes);
4089
4090                         del_pset(nodes);
4091                 }
4092         }
4093
4094 //      remove_unused_defs(si);
4095
4096         be_free_dominance_frontiers(dfi);
4097 }
4098
4099
4100 static void
4101 writeback_results(spill_ilp_t * si)
4102 {
4103         /* walk through the graph and collect all spills, reloads and remats for a value */
4104
4105         si->values = new_set(cmp_defs, 4096);
4106
4107         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4108         delete_unnecessary_remats(si);
4109         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4110         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4111         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4112         if(opt_memoperands)
4113                 insert_memoperands(si);
4114         phim_fixer(si);
4115
4116         /* clean the remat info! there are still back-edges leading there! */
4117         clean_remat_info(si);
4118
4119         rewire_uses(si);
4120
4121         connect_all_spills_with_keep(si);
4122
4123         del_set(si->values);
4124 }
4125
4126 static int
4127 get_n_regs(spill_ilp_t * si)
4128 {
4129         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4130
4131         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4132         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4133
4134         arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4135     be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4136
4137         bitset_andnot(arch_regs, abi_regs);
4138         arch_n_regs = bitset_popcnt(arch_regs);
4139
4140         bitset_free(arch_regs);
4141         bitset_free(abi_regs);
4142
4143         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4144         return arch_n_regs;
4145 }
4146
4147 static void
4148 walker_reload_mover(ir_node * bb, void * data)
4149 {
4150         spill_ilp_t   *si = data;
4151         ir_node           *tmp;
4152
4153         sched_foreach(bb, tmp) {
4154                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4155                         ir_node       *reload = tmp;
4156                         ir_node       *irn = tmp;
4157
4158                         /* move reload upwards */
4159
4160                         int pressure = (int)get_irn_link(reload);
4161                         if(pressure < si->n_regs) {
4162                                 irn = sched_prev(reload);
4163                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4164                                 sched_remove(reload);
4165                                 pressure = (int)get_irn_link(irn);
4166
4167                                 while(pressure < si->n_regs) {
4168                                         if( sched_is_end(irn) ||
4169                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4170                                            /* do not move reload before its spill */
4171                                            (irn == be_get_Reload_mem(reload)) ||
4172                                            /* do not move before phi */
4173                                            is_Phi(irn)) break;
4174
4175                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4176                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4177                                         irn = sched_prev(irn);
4178
4179                                         pressure = (int)get_irn_link(irn);
4180                                 }
4181
4182                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4183                                 sched_put_after(irn, reload);
4184                         }
4185                 }
4186         }
4187 }
4188
4189 static void
4190 move_reloads_upward(spill_ilp_t * si)
4191 {
4192         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4193 }
4194
4195
4196 /**
4197  * Walk all irg blocks and check for interfering spills inside of phi classes
4198  */
4199 static void
4200 luke_meminterferencechecker(ir_node * bb, void * data)
4201 {
4202         spill_ilp_t    *si = (spill_ilp_t*)data;
4203         int             l1, l2;
4204
4205         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4206                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4207
4208                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4209
4210                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4211                 if (has_reg_class(si, a) && get_phi_class(a)) {
4212                         for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
4213                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
4214
4215                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4216
4217                                 /* a and b are only interesting if they are in the same phi class */
4218                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4219                                         if(values_interfere_in_block(si, bb, a, b)) {
4220                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4221                                         }
4222                                 }
4223                         }
4224                 }
4225         }
4226 }
4227
4228 static void
4229 verify_phiclasses(spill_ilp_t * si)
4230 {
4231         /* analyze phi classes */
4232         phi_class_compute(si->chordal_env->irg);
4233
4234         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4235         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4236 }
4237
4238 void
4239 be_spill_remat(const be_chordal_env_t * chordal_env)
4240 {
4241         char            buf[256];
4242         char            problem_name[256];
4243         char            dump_suffix[256];
4244         char            dump_suffix2[256];
4245         struct obstack  obst;
4246         spill_ilp_t     si;
4247
4248         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4249         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4250         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4251
4252         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4253         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4254
4255         if(opt_verify & VERIFY_DOMINANCE)
4256                 be_check_dominance(chordal_env->irg);
4257
4258         obstack_init(&obst);
4259         si.chordal_env = chordal_env;
4260         si.obst = &obst;
4261         si.cls = chordal_env->cls;
4262         si.lpp = new_lpp(problem_name, lpp_minimize);
4263         si.remat_info = new_set(cmp_remat_info, 4096);
4264         si.interferences = new_set(cmp_interference, 32);
4265         si.memoperands = new_set(cmp_memoperands, 128);
4266         si.all_possible_remats = pset_new_ptr_default();
4267         si.spills = pset_new_ptr_default();
4268         si.inverse_ops = pset_new_ptr_default();
4269         si.lv = chordal_env->lv;
4270         si.keep = NULL;
4271         si.n_regs = get_n_regs(&si);
4272
4273         set_irg_link(chordal_env->irg, &si);
4274         compute_doms(chordal_env->irg);
4275
4276         /* compute phi classes */
4277 //      phi_class_compute(chordal_env->irg);
4278
4279         if(opt_dump_flags & DUMP_STATS)
4280                 be_analyze_regpressure(chordal_env, "-pre");
4281
4282         DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4283         irg_block_walk_graph(chordal_env->irg, luke_initializer, NULL, &si);
4284
4285         if(opt_remats) {
4286                 /* collect remats */
4287                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4288                 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4289         }
4290
4291         /* insert possible remats */
4292         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4293         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4294         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4295
4296         if(opt_keep_alive & KEEPALIVE_REMATS) {
4297                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4298                 connect_all_remats_with_keep(&si);
4299                 /* dump graph with inserted remats */
4300                 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4301         }
4302
4303         /* insert copies for phi arguments not in my regclass */
4304         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4305
4306         /* recompute liveness */
4307         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4308         be_liveness_recompute(si.lv);
4309
4310         /* build the ILP */
4311         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4312         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4313         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4314
4315         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4316         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4317
4318         if(opt_memcopies) {
4319                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4320                 memcopyhandler(&si);
4321         }
4322
4323         if(opt_dump_flags & DUMP_PROBLEM) {
4324                 FILE           *f;
4325                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4326                 if ((f = fopen(buf, "wt")) != NULL) {
4327                         lpp_dump_plain(si.lpp, f);
4328                         fclose(f);
4329                 }
4330         }
4331
4332         if(opt_dump_flags & DUMP_MPS) {
4333                 FILE *f;
4334
4335                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4336                 if((f = fopen(buf, "wt")) != NULL) {
4337                         mps_write_mps(si.lpp, s_mps_fixed, f);
4338                         fclose(f);
4339                 }
4340
4341                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4342                 if((f = fopen(buf, "wt")) != NULL) {
4343                         mps_write_mst(si.lpp, s_mps_fixed, f);
4344                         fclose(f);
4345                 }
4346         }
4347
4348         lpp_check_startvals(si.lpp);
4349
4350 #ifdef SOLVE
4351         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4352         lpp_set_time_limit(si.lpp, opt_timeout);
4353
4354         if(opt_log)
4355                 lpp_set_log(si.lpp, stdout);
4356
4357 #ifdef SOLVE_LOCAL
4358         lpp_solve_cplex(si.lpp);
4359 #else
4360         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4361 #endif
4362         assert(lpp_is_sol_valid(si.lpp)
4363                && "solution of ILP must be valid");
4364
4365         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
4366
4367         if(opt_dump_flags & DUMP_SOLUTION) {
4368                 FILE           *f;
4369                 char            buf[256];
4370
4371                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4372                 if ((f = fopen(buf, "wt")) != NULL) {
4373                         int             i;
4374                         for (i = 0; i < si.lpp->var_next; ++i) {
4375                                 lpp_name_t     *name = si.lpp->vars[i];
4376                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4377                         }
4378                         fclose(f);
4379                 }
4380         }
4381
4382         writeback_results(&si);
4383
4384 #endif                          /* SOLVE */
4385
4386         kill_all_unused_values_in_schedule(&si);
4387
4388         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4389                 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4390
4391         // move reloads upwards
4392         be_liveness_recompute(si.lv);
4393         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4394         move_reloads_upward(&si);
4395
4396         if(opt_memcopies) {
4397                 verify_phiclasses(&si);
4398         }
4399
4400         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4401
4402         if(opt_dump_flags & DUMP_PRESSURE)
4403                 dump_pressure_graph(&si, dump_suffix2);
4404
4405         if(opt_dump_flags & DUMP_STATS)
4406                 be_analyze_regpressure(chordal_env, "-post");
4407
4408         if(opt_verify & VERIFY_DOMINANCE)
4409                 be_check_dominance(chordal_env->irg);
4410
4411         free_dom(chordal_env->irg);
4412         del_set(si.interferences);
4413         del_pset(si.inverse_ops);
4414         del_pset(si.all_possible_remats);
4415         del_set(si.memoperands);
4416         del_pset(si.spills);
4417         free_lpp(si.lpp);
4418         obstack_free(&obst, NULL);
4419         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4420 }
4421
4422 #else                           /* WITH_ILP */
4423
4424 static void
4425 only_that_you_can_compile_without_WITH_ILP_defined(void)
4426 {
4427 }
4428
4429 #endif                          /* WITH_ILP */