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 =  50.0;
  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                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2223 #if 1
2224                                 ilp_var_t  delete;
2225                                 assert(spill);
2226
2227                                 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2228                                 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2229
2230                                 /* op may not be killed if its first live_range is 1 */
2231                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2232                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2233                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2234                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2235
2236                                 /* op may not be killed if it is spilled after the definition */
2237                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2238                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2239                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2240                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2241 #else
2242                                 ilp_var_t  keep;
2243                                 assert(spill);
2244
2245                                 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2246                                 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2247
2248                                 /* op may not be killed if its first live_range is 1 */
2249                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2250                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2251                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2252                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2253
2254                                 /* op may not be killed if it is spilled after the definition */
2255                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2256                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2257                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2258                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2259 #endif
2260                         }
2261                 } else {
2262                         pset_foreach(defs, tmp) {
2263                                 op_t      *tmp_op = get_irn_link(tmp);
2264                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2265                                 assert(spill);
2266
2267                                 /* live_range or spill should be 1
2268                                    TODO: lr should be live until first use */
2269                                 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2270                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2271                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2272                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2273                         }
2274                 }
2275
2276
2277                 /******************
2278                  *   P R O L O G
2279                  ******************/
2280
2281                 /* check the register pressure in the prolog */
2282                 /* sum_{L\U} lr <= k - |U| */
2283                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2284                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2285
2286                 /* for the prolog remove defined values from the live set */
2287                 pset_foreach(defs, tmp) {
2288                         pset_remove_ptr(live, tmp);
2289                 }
2290
2291                 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2292                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2293                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2294                 }
2295
2296                 /***********************************************************
2297                  *  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
2298                  **********************************************************/
2299
2300
2301                 set_foreach(args, keyval) {
2302                         spill_t          *spill;
2303                         const ir_node    *arg = keyval->key;
2304                         int               i = PTR_TO_INT(keyval->val);
2305                         op_t             *arg_op = get_irn_link(arg);
2306                         ilp_cst_t         requirements;
2307                         int               n_memoperands;
2308
2309                         spill = set_find_spill(spill_bb->ilp, arg);
2310                         assert(spill);
2311
2312                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2313                         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);
2314
2315                         /* reload <= mem_out */
2316                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2317                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2318                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2319                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2320
2321                         /* requirement: arg must be in register for use */
2322                         /* reload + remat + live_range == 1 */
2323                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2324                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2325
2326                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2327                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2328                         foreach_pre_remat(si, irn, tmp) {
2329                                 op_t     *remat_op = get_irn_link(tmp);
2330                                 if(remat_op->attr.remat.remat->value == arg) {
2331                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2332                                 }
2333                         }
2334
2335                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2336                                 n_memoperands = 0;
2337                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2338                                         if(get_irn_n(irn, n) == arg) {
2339                                                 n_memoperands++;
2340                                         }
2341                                 }
2342                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2343                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2344                                                 memoperand_t  *memoperand;
2345                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2346
2347                                                 /* memoperand <= mem_out */
2348                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2349                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2350                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2351                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2352
2353                                                 /* the memoperand is only sufficient if it is used once by the op */
2354                                                 if(n_memoperands == 1)
2355                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2356
2357                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2358
2359                                                 /* we have one more free register if we use a memory operand */
2360                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2361                                         }
2362                                 }
2363                         }
2364                 }
2365
2366                 /* iterate over L\U */
2367                 pset_foreach(live, tmp) {
2368                         if(!set_find_keyval(args, tmp)) {
2369                                 /* if a live value is not used by irn */
2370                                 tmp_op = get_irn_link(tmp);
2371                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2372                         }
2373                 }
2374
2375                 /* requirements for remats */
2376                 foreach_pre_remat(si, irn, tmp) {
2377                         op_t        *remat_op = get_irn_link(tmp);
2378                         int          n;
2379
2380                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2381                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2382                                 op_t           *arg_op = get_irn_link(remat_arg);
2383
2384                                 if(!has_reg_class(si, remat_arg)) continue;
2385
2386                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2387                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2388                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2389
2390                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2391                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2392
2393                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2394                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2395                                         int    index = (int)keyval->val;
2396
2397                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2398                                 }
2399                         }
2400                 }
2401
2402
2403
2404
2405                 /*************************
2406                  *  D O N E  W I T H  O P
2407                  *************************/
2408
2409                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2410
2411                 pset_foreach(live, tmp) {
2412                         assert(has_reg_class(si, tmp));
2413                 }
2414
2415                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2416                         ir_node        *arg = get_irn_n(irn, n);
2417
2418                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2419                 }
2420
2421                 del_pset(remat_defs);
2422                 del_pset(used);
2423                 del_set(args);
2424                 del_pset(defs);
2425                 defs = pset_new_ptr_default();
2426
2427                 /* skip everything above barrier in start block */
2428                 if(is_start_block(bb) && be_is_Barrier(irn)) {
2429                         assert(pset_count(live) == 0);
2430                         break;
2431                 }
2432
2433         }
2434         del_pset(defs);
2435
2436
2437
2438         /***************************************
2439          *   B E G I N N I N G  O F  B L O C K
2440          ***************************************/
2441
2442
2443         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2444         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2445
2446         pset_foreach(live, irn) {
2447                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2448         }
2449
2450         /* construct mem_outs for all values */
2451         set_foreach(spill_bb->ilp, spill) {
2452                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2453                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2454
2455                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2456                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2457
2458                 if(pset_find_ptr(live, spill->irn)) {
2459                         int default_spilled;
2460                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2461
2462                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2463                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2464                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2465                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2466
2467                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2468                                 int   n;
2469                                 op_t *op = get_irn_link(spill->irn);
2470
2471                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2472                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2473                                         double          freq=0.0;
2474                                         int             m;
2475                                         ilp_var_t       var;
2476
2477
2478                                         /* argument already done? */
2479                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2480
2481                                         /* get sum of execution frequencies of blocks with the same phi argument */
2482                                         for(m=n; m>=0; --m) {
2483                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2484
2485                                                 if(arg==arg2) {
2486                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2487                                                 }
2488                                         }
2489
2490                                         /* copies are not for free */
2491                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2492                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2493
2494                                         for(m=n; m>=0; --m) {
2495                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2496
2497                                                 if(arg==arg2) {
2498                                                         op->attr.live_range.args.copies[m] = var;
2499                                                 }
2500                                         }
2501
2502 #if 0
2503                                         /* copy <= mem_in */
2504                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2505                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2506                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2507                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2508 #endif
2509                                 }
2510                         }
2511                 }
2512         }
2513
2514         foreach_post_remat(bb, tmp) {
2515                 int         n;
2516                 op_t       *remat_op = get_irn_link(tmp);
2517                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2518                 ir_node    *remat_arg;
2519
2520                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2521                         remat_arg = get_irn_n(tmp, n);
2522
2523                         if(has_reg_class(si, remat_arg)) {
2524                                 pset_insert_ptr(remat_args, remat_arg);
2525                         }
2526                 }
2527
2528                 /* remat + \sum live_range(remat_arg) <= |args| */
2529                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2530                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2531                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2532
2533                 pset_foreach(remat_args, remat_arg) {
2534                         if(pset_find_ptr(live, remat_arg)) {
2535                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2536                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2537                         }
2538                 }
2539                 del_pset(remat_args);
2540         }
2541
2542         foreach_post_remat(bb, tmp) {
2543                 int  n;
2544
2545                 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2546                         ir_node  *remat_arg = get_irn_n(tmp, n);
2547
2548                         /* if value is becoming live through use by remat2 */
2549                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2550                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2551                                 ilp_cst_t   nomem;
2552
2553                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2554
2555                                 pset_insert_ptr(live, remat_arg);
2556                                 spill = add_to_spill_bb(si, bb, remat_arg);
2557                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2558
2559                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2560                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2561                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2562                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2563                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2564
2565
2566                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2567                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2568                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2569                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2570                         }
2571                 }
2572         }
2573
2574         /* L\U is empty at bb start */
2575         /* arg is live throughout epilog if it is reg_in into this block */
2576
2577         /* check the register pressure at the beginning of the block
2578          * including remats
2579          */
2580         /* reg_in entspricht post_use */
2581
2582         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2583         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2584
2585         pset_foreach(live, irn) {
2586         ilp_cst_t  nospill;
2587
2588                 spill = set_find_spill(spill_bb->ilp, irn);
2589                 assert(spill);
2590
2591                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2592                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2593
2594                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2595
2596                 /* spill + mem_in <= 1 */
2597                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2598                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2599
2600                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2601                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2602
2603         } /* post_remats are NOT included in register pressure check because
2604            they do not increase regpressure */
2605
2606         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2607         pset_foreach(live, irn) {
2608                 int          p = 0,
2609                                          n;
2610
2611                 spill = set_find_spill(spill_bb->ilp, irn);
2612                 assert(spill && spill->irn == irn);
2613
2614                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2615                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2616                                 ilp_cst_t       mem_in,
2617                                                                 reg_in;
2618                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2619                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2620                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2621                                 spill_t        *spill_p;
2622                                 op_t           *op = get_irn_link(irn);
2623
2624                                 /* although the phi is in the right regclass one or more of
2625                                  * its arguments can be in a different one or at least to
2626                                  * ignore
2627                                  */
2628                                 if(has_reg_class(si, phi_arg)) {
2629                                         /* mem_in < mem_out_arg + copy */
2630                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2631                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2632
2633                                         /* reg_in < reg_out_arg */
2634                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2635                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2636
2637                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2638                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2639
2640                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2641                                         assert(spill_p);
2642
2643                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2644                                         if(opt_memcopies)
2645                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2646
2647                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2648                                 }
2649                         }
2650                 } else {
2651                         /* else assure the value arrives on all paths in the same resource */
2652
2653                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2654                                 ilp_cst_t       mem_in,
2655                                                                 reg_in;
2656                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2657                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2658                                 spill_t        *spill_p;
2659
2660                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2661                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2662                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2663                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2664
2665                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2666                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2667
2668                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2669                                 assert(spill_p);
2670
2671                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2672                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2673                         }
2674                 }
2675         }
2676
2677         foreach_post_remat(bb, tmp) {
2678                 int         n;
2679
2680                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2681                         ir_node    *remat_arg = get_irn_n(tmp, n);
2682                         op_t       *remat_op = get_irn_link(tmp);
2683
2684                         if(!has_reg_class(si, remat_arg)) continue;
2685
2686                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2687                         assert(spill);
2688
2689                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2690                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2691                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2692                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2693                 }
2694         }
2695
2696         pset_foreach(live, irn) {
2697                 const op_t      *op = get_irn_link(irn);
2698                 const ir_node   *remat;
2699                 int              n_remats = 0;
2700
2701                 cst = ILP_UNDEF;
2702
2703                 foreach_post_remat(bb, remat) {
2704                         int   n;
2705
2706                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2707                                 const ir_node  *arg = get_irn_n(remat, n);
2708
2709                                 if(arg == irn) {
2710                                         const op_t   *remat_op = get_irn_link(remat);
2711
2712                                         if(cst == ILP_UNDEF) {
2713                                                 /* sum remat2s <= 1 + n_remats*live_range */
2714                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2715                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2716                                         }
2717                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2718                                         ++n_remats;
2719                                         break;
2720                                 }
2721                         }
2722                 }
2723                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2724                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2725                 }
2726         }
2727
2728         /* first live ranges from reg_ins */
2729         pset_foreach(live, irn) {
2730                 op_t      *op = get_irn_link(irn);
2731
2732                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2733
2734                         spill = set_find_spill(spill_bb->ilp, irn);
2735                         assert(spill && spill->irn == irn);
2736
2737                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2738                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2739                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2740                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2741
2742                         foreach_post_remat(bb, tmp) {
2743                                 op_t     *remat_op = get_irn_link(tmp);
2744
2745                                 if(remat_op->attr.remat.remat->value == irn) {
2746                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2747                                 }
2748                         }
2749                 }
2750         }
2751
2752         /* walk forward now and compute constraints for placing spills */
2753         /* this must only be done for values that are not defined in this block */
2754         pset_foreach(live, irn) {
2755                 /*
2756                  * if value is defined in this block we can anways place the spill directly after the def
2757                  *    -> no constraint necessary
2758                  */
2759                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2760                         assert(0);
2761                 }
2762
2763
2764                 spill = set_find_spill(spill_bb->ilp, irn);
2765                 assert(spill);
2766
2767                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2768                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2769
2770                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2771                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2772
2773                 if(!is_Phi(irn)) {
2774                         sched_foreach_op(bb, tmp) {
2775                                 op_t   *op = get_irn_link(tmp);
2776
2777                                 if(is_Phi(tmp)) continue;
2778                                 assert(!is_Proj(tmp));
2779
2780                                 if(op->is_remat) {
2781                                         const ir_node   *value = op->attr.remat.remat->value;
2782
2783                                         if(value == irn) {
2784                                                 /* only collect remats up to the first real use of a value */
2785                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2786                                         }
2787                                 } else {
2788                                         int   n;
2789
2790                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2791                                                 ir_node    *arg = get_irn_n(tmp, n);
2792
2793                                                 if(arg == irn) {
2794                                                         /* if a value is used stop collecting remats */
2795                             goto next_live;
2796                                                 }
2797                                         }
2798                                 }
2799                         }
2800                 }
2801 next_live: ;
2802         }
2803
2804         del_pset(live);
2805 }
2806
2807 typedef struct _irnlist_t {
2808         struct list_head   list;
2809         ir_node           *irn;
2810 } irnlist_t;
2811
2812 typedef struct _interference_t {
2813         struct list_head    blocklist;
2814         ir_node            *a;
2815         ir_node            *b;
2816 } interference_t;
2817
2818 static int
2819 cmp_interference(const void *a, const void *b, size_t size)
2820 {
2821         const interference_t *p = a;
2822         const interference_t *q = b;
2823
2824         return !(p->a == q->a && p->b == q->b);
2825 }
2826
2827 static interference_t *
2828 set_find_interference(set * set, ir_node * a, ir_node * b)
2829 {
2830         interference_t     query;
2831
2832         query.a = (a>b)?a:b;
2833         query.b = (a>b)?b:a;
2834
2835         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2836 }
2837
2838 static interference_t *
2839 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2840 {
2841         interference_t     query,
2842                                           *result;
2843         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2844
2845         list->irn = bb;
2846
2847         result = set_find_interference(set, a, b);
2848         if(result) {
2849
2850                 list_add(&list->list, &result->blocklist);
2851                 return result;
2852         }
2853
2854         query.a = (a>b)?a:b;
2855         query.b = (a>b)?b:a;
2856
2857         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2858
2859         INIT_LIST_HEAD(&result->blocklist);
2860         list_add(&list->list, &result->blocklist);
2861
2862         return result;
2863 }
2864
2865 static int
2866 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2867 {
2868         const ir_edge_t *edge;
2869
2870         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2871                 /* both values are live in, so they interfere */
2872                 return 1;
2873         }
2874
2875         /* ensure a dominates b */
2876         if(value_dominates(b,a)) {
2877                 const ir_node * t;
2878                 t = b;
2879                 b = a;
2880                 a = t;
2881         }
2882         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2883
2884
2885         /* the following code is stolen from bera.c */
2886         if(be_is_live_end(si->lv, bb, a))
2887                 return 1;
2888
2889         foreach_out_edge(a, edge) {
2890                 const ir_node *user = edge->src;
2891                 if(get_nodes_block(user) == bb
2892                                 && !is_Phi(user)
2893                                 && b != user
2894                                 && !pset_find_ptr(si->inverse_ops, user)
2895                                 && value_dominates(b, user))
2896                         return 1;
2897         }
2898
2899         return 0;
2900 }
2901
2902 /**
2903  * Walk all irg blocks and collect interfering values inside of phi classes
2904  */
2905 static void
2906 luke_interferencewalker(ir_node * bb, void * data)
2907 {
2908         spill_ilp_t    *si = (spill_ilp_t*)data;
2909         int             l1, l2;
2910
2911         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2912                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2913                 op_t           *a_op = get_irn_link(a);
2914
2915
2916                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2917                 if (has_reg_class(si, a) && get_phi_class(a)) {
2918                         if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2919                                 continue;
2920
2921                         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)) {
2922                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
2923                                 op_t           *b_op = get_irn_link(b);
2924
2925
2926                                 /* a and b are only interesting if they are in the same phi class */
2927                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2928                                         if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2929                                                 continue;
2930
2931                                         if(values_interfere_in_block(si, bb, a, b)) {
2932                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2933                                                 set_insert_interference(si, si->interferences, a, b, bb);
2934                                         }
2935                                 }
2936                         }
2937                 }
2938         }
2939 }
2940
2941 static unsigned int copy_path_id = 0;
2942
2943 static void
2944 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2945 {
2946         ilp_cst_t  cst;
2947         ilp_var_t  copy;
2948         char       buf[256];
2949         void      *ptr;
2950
2951         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2952         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2953
2954         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2955
2956         pset_foreach(copies, ptr) {
2957                 copy = PTR_TO_INT(ptr);
2958                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2959         }
2960 }
2961
2962 /**
2963  * @parameter copies   contains a path of copies which lead us to irn
2964  * @parameter visited  contains a set of nodes already visited on this path
2965  */
2966 static int
2967 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2968 {
2969         const ir_edge_t *edge;
2970         op_t            *op = get_irn_link(irn);
2971     pset            *visited_users = pset_new_ptr_default();
2972         int              paths = 0;
2973
2974         if(op->is_remat) return 0;
2975
2976         pset_insert_ptr(visited, irn);
2977
2978         if(is_Phi(irn)) {
2979                 int    n;
2980         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2981
2982                 /* visit all operands */
2983                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2984                         ir_node  *arg = get_irn_n(irn, n);
2985                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2986
2987                         if(!has_reg_class(si, arg)) continue;
2988             if(pset_find_ptr(visited_operands, arg)) continue;
2989             pset_insert_ptr(visited_operands, arg);
2990
2991                         if(arg == target) {
2992                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2993                                         del_pset(visited_operands);
2994                                         del_pset(visited_users);
2995                                         pset_remove_ptr(visited, irn);
2996                                         return paths;
2997                                 }
2998                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2999                                 write_copy_path_cst(si, copies, any_interfere);
3000                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3001                         } else if(!pset_find_ptr(visited, arg)) {
3002                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3003                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3004                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3005
3006                 if(paths > MAX_PATHS) {
3007                     if(pset_count(copies) == 0) {
3008                         ilp_cst_t  cst;
3009                         char       buf[256];
3010
3011                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3012                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3013                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3014                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3015                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3016
3017                         paths = 0;
3018                     } else {
3019                         del_pset(visited_operands);
3020                         del_pset(visited_users);
3021                         pset_remove_ptr(visited, irn);
3022                         return paths;
3023                     }
3024                 } else if(pset_count(copies) == 0) {
3025                                         paths = 0;
3026                                 }
3027                         }
3028                 }
3029
3030         del_pset(visited_operands);
3031         }
3032
3033         /* visit all uses which are phis */
3034         foreach_out_edge(irn, edge) {
3035                 ir_node  *user = edge->src;
3036                 int       pos  = edge->pos;
3037                 op_t     *op = get_irn_link(user);
3038                 ilp_var_t copy;
3039
3040                 if(!is_Phi(user)) continue;
3041                 if(!has_reg_class(si, user)) continue;
3042         if(pset_find_ptr(visited_users, user)) continue;
3043         pset_insert_ptr(visited_users, user);
3044
3045                 copy = op->attr.live_range.args.copies[pos];
3046
3047                 if(user == target) {
3048                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3049                                 del_pset(visited_users);
3050                                 pset_remove_ptr(visited, irn);
3051                                 return paths;
3052                         }
3053                         pset_insert(copies, INT_TO_PTR(copy), copy);
3054                         write_copy_path_cst(si, copies, any_interfere);
3055                         pset_remove(copies, INT_TO_PTR(copy), copy);
3056                 } else if(!pset_find_ptr(visited, user)) {
3057                         pset_insert(copies, INT_TO_PTR(copy), copy);
3058                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3059                         pset_remove(copies, INT_TO_PTR(copy), copy);
3060
3061             if(paths > MAX_PATHS) {
3062                 if(pset_count(copies) == 0) {
3063                     ilp_cst_t  cst;
3064                     char       buf[256];
3065
3066                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3067                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3068                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3069                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3070                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3071
3072                     paths = 0;
3073                 } else {
3074                     del_pset(visited_users);
3075                     pset_remove_ptr(visited, irn);
3076                     return paths;
3077                 }
3078             } else if(pset_count(copies) == 0) {
3079                                 paths = 0;
3080                         }
3081                 }
3082         }
3083
3084     del_pset(visited_users);
3085         pset_remove_ptr(visited, irn);
3086         return paths;
3087 }
3088
3089 static void
3090 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3091 {
3092         pset * copies = pset_new_ptr_default();
3093         pset * visited = pset_new_ptr_default();
3094
3095         find_copy_path(si, a, b, any_interfere, copies, visited);
3096
3097         del_pset(visited);
3098         del_pset(copies);
3099 }
3100
3101
3102 static void
3103 memcopyhandler(spill_ilp_t * si)
3104 {
3105         interference_t   *interference;
3106         char              buf[256];
3107         /* teste Speicherwerte auf Interferenz */
3108
3109         /* analyze phi classes */
3110         phi_class_compute(si->chordal_env->irg);
3111
3112         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3113         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3114
3115         /* now lets emit the ILP unequations for the crap */
3116         set_foreach(si->interferences, interference) {
3117                 irnlist_t      *irnlist;
3118                 ilp_var_t       interfere,
3119                                                 any_interfere;
3120                 ilp_cst_t       any_interfere_cst,
3121                                                 cst;
3122                 const ir_node  *a  = interference->a;
3123                 const ir_node  *b  = interference->b;
3124
3125                 /* any_interf <= \sum interf */
3126                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3127                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3128                 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3129
3130                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3131
3132                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3133                         const ir_node  *bb = irnlist->irn;
3134                         spill_bb_t     *spill_bb = get_irn_link(bb);
3135                         spill_t        *spilla,
3136                                                    *spillb;
3137                         char           buf[256];
3138
3139                         spilla = set_find_spill(spill_bb->ilp, a);
3140                         assert(spilla);
3141
3142                         spillb = set_find_spill(spill_bb->ilp, b);
3143                         assert(spillb);
3144
3145                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3146                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3147                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3148                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3149                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3150                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3151
3152                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3153                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3154
3155                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3156                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3157                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3158                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3159                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3160
3161                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3162                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3163
3164                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3165                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3166                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3167
3168                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3169                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3170
3171                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3172                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3173                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3174
3175
3176                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3177
3178                         /* any_interfere >= interf */
3179                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3180                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3181
3182                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3183                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3184                 }
3185
3186                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3187                 gen_copy_constraints(si,a,b,any_interfere);
3188         }
3189 }
3190
3191
3192 static INLINE int
3193 is_zero(double x)
3194 {
3195         return fabs(x) < 0.00001;
3196 }
3197
3198 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3199 {
3200         spill_ilp_t *si = get_irg_link(current_ir_graph);
3201
3202         if(pset_find_ptr(si->all_possible_remats, n)) {
3203                 op_t   *op = (op_t*)get_irn_link(n);
3204                 assert(op && op->is_remat);
3205
3206                 if(!op->attr.remat.remat->inverse) {
3207                         if(op->attr.remat.pre) {
3208                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3209                         } else {
3210                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3211                         }
3212
3213                         return 1;
3214                 } else {
3215                         op_t   *op = (op_t*)get_irn_link(n);
3216                         assert(op && op->is_remat);
3217
3218                         if(op->attr.remat.pre) {
3219                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3220                         } else {
3221                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3222                         }
3223
3224                         return 1;
3225                 }
3226         }
3227
3228         return 0;
3229 }
3230
3231 static void
3232 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3233 {
3234         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3235         be_dump(irg, suffix, dump_ir_block_graph_sched);
3236         set_dump_node_vcgattr_hook(NULL);
3237 }
3238
3239 /**
3240  * Edge hook to dump the schedule edges with annotated register pressure.
3241  */
3242 static int
3243 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3244 {
3245         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3246                 ir_node *prev = sched_prev(irn);
3247                 fprintf(F, "edge:{sourcename:\"");
3248                 PRINT_NODEID(irn);
3249                 fprintf(F, "\" targetname:\"");
3250                 PRINT_NODEID(prev);
3251                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3252                 fprintf(F, "\" color:magenta}\n");
3253         }
3254         return 1;
3255 }
3256
3257 static void
3258 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3259 {
3260         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3261
3262         dump_consts_local(0);
3263         set_dump_node_edge_hook(sched_pressure_edge_hook);
3264         dump_ir_block_graph(irg, suffix);
3265         set_dump_node_edge_hook(old_edge_hook);
3266 }
3267
3268 static void
3269 walker_pressure_annotator(ir_node * bb, void * data)
3270 {
3271         spill_ilp_t  *si = data;
3272         ir_node      *irn;
3273         int           n, i;
3274         pset         *live = pset_new_ptr_default();
3275         int           projs = 0;
3276
3277         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3278                 irn = be_lv_get_irn(si->lv, bb, i);
3279
3280                 if (has_reg_class(si, irn)) {
3281                         pset_insert_ptr(live, irn);
3282                 }
3283         }
3284
3285         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3286
3287         sched_foreach_reverse(bb, irn) {
3288                 if(is_Phi(irn)) {
3289                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3290                         continue;
3291                 }
3292
3293                 if(has_reg_class(si, irn)) {
3294                         pset_remove_ptr(live, irn);
3295                         if(is_Proj(irn)) ++projs;
3296                 }
3297
3298                 if(!is_Proj(irn)) projs = 0;
3299
3300                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3301                         ir_node    *arg = get_irn_n(irn, n);
3302
3303                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3304                 }
3305                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3306         }
3307
3308         del_pset(live);
3309 }
3310
3311 static void
3312 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3313 {
3314         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3315 }
3316
3317 static void
3318 connect_all_remats_with_keep(spill_ilp_t * si)
3319 {
3320         ir_node   *irn;
3321         ir_node  **ins,
3322                          **pos;
3323         int        n_remats;
3324
3325
3326         n_remats = pset_count(si->all_possible_remats);
3327         if(n_remats) {
3328                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3329
3330                 pos = ins;
3331                 pset_foreach(si->all_possible_remats, irn) {
3332                         *pos = irn;
3333                         ++pos;
3334                 }
3335
3336                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3337
3338                 obstack_free(si->obst, ins);
3339         }
3340 }
3341
3342 static void
3343 connect_all_spills_with_keep(spill_ilp_t * si)
3344 {
3345         ir_node   *irn;
3346         ir_node  **ins,
3347                          **pos;
3348         int        n_spills;
3349         ir_node   *keep;
3350
3351
3352         n_spills = pset_count(si->spills);
3353         if(n_spills) {
3354                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3355
3356                 pos = ins;
3357                 pset_foreach(si->spills, irn) {
3358                         *pos = irn;
3359                         ++pos;
3360                 }
3361
3362                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3363
3364                 obstack_free(si->obst, ins);
3365         }
3366 }
3367
3368 /** insert a spill at an arbitrary position */
3369 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3370 {
3371         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
3372         ir_graph *irg   = get_irn_irg(bl);
3373         ir_node *frame  = get_irg_frame(irg);
3374         ir_node *spill;
3375         ir_node *next;
3376
3377         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3378         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3379
3380         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3381
3382         /*
3383          * search the right insertion point. a spill of a phi cannot be put
3384          * directly after the phi, if there are some phis behind the one which
3385          * is spilled. Also, a spill of a Proj must be after all Projs of the
3386          * same tuple node.
3387          *
3388          * Here's one special case:
3389          * If the spill is in the start block, the spill must be after the frame
3390          * pointer is set up. This is done by setting insert to the end of the block
3391          * which is its default initialization (see above).
3392          */
3393
3394         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3395                 insert = frame;
3396
3397         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3398                 insert = next;
3399
3400         sched_add_after(insert, spill);
3401         return spill;
3402 }
3403
3404 static void
3405 delete_remat(spill_ilp_t * si, ir_node * remat) {
3406         int       n;
3407         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3408
3409         sched_remove(remat);
3410
3411         /* kill links to operands */
3412         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3413                 set_irn_n(remat, n, bad);
3414         }
3415 }
3416
3417 static void
3418 clean_remat_info(spill_ilp_t * si)
3419 {
3420         int            n;
3421         remat_t       *remat;
3422         remat_info_t  *remat_info;
3423         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
3424
3425         set_foreach(si->remat_info, remat_info) {
3426                 if(!remat_info->remats) continue;
3427
3428                 pset_foreach(remat_info->remats, remat)
3429                 {
3430                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3431                                 if(sched_is_scheduled(remat->proj)) {
3432                                         sched_remove((ir_node*)remat->proj);
3433                                 }
3434                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3435                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3436                         }
3437
3438                         if(get_irn_n_edges(remat->op) == 0) {
3439                                 if(sched_is_scheduled(remat->op)) {
3440                                         sched_remove((ir_node*)remat->op);
3441                                 }
3442                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3443                                         set_irn_n((ir_node*)remat->op, n, bad);
3444                                 }
3445                         }
3446                 }
3447
3448                 if(remat_info->remats) del_pset(remat_info->remats);
3449                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3450         }
3451 }
3452
3453 static void
3454 delete_unnecessary_remats(spill_ilp_t * si)
3455 {
3456         if(opt_keep_alive & KEEPALIVE_REMATS) {
3457                 int       n;
3458                 ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3459
3460                 if(si->keep) {
3461                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3462                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3463                                 op_t           *arg_op = get_irn_link(keep_arg);
3464                                 lpp_name_t     *name;
3465
3466                                 assert(arg_op->is_remat);
3467
3468                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3469
3470                                 if(is_zero(name->value)) {
3471                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3472                                         /* TODO check whether reload is preferred over remat (could be bug) */
3473                                         delete_remat(si, keep_arg);
3474                                 } else {
3475                                         if(!arg_op->attr.remat.remat->inverse) {
3476                                                 if(arg_op->attr.remat.pre) {
3477                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3478                                                 } else {
3479                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3480                                                 }
3481                                         } else {
3482                                                 if(arg_op->attr.remat.pre) {
3483                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3484                                                 } else {
3485                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3486                                                 }
3487                                         }
3488                                 }
3489
3490                                 set_irn_n(si->keep, n, bad);
3491                         }
3492                 } else {
3493                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3494                 }
3495         } else {
3496                 ir_node  *remat;
3497
3498                 pset_foreach(si->all_possible_remats, remat) {
3499                         op_t           *remat_op = get_irn_link(remat);
3500                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3501
3502                         if(is_zero(name->value)) {
3503                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3504                                 /* TODO check whether reload is preferred over remat (could be bug) */
3505                                 delete_remat(si, remat);
3506                         } else {
3507                                 if(!remat_op->attr.remat.remat->inverse) {
3508                                         if(remat_op->attr.remat.pre) {
3509                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3510                                         } else {
3511                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3512                                         }
3513                                 } else {
3514                                         if(remat_op->attr.remat.pre) {
3515                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3516                                         } else {
3517                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3518                                         }
3519                                 }
3520                         }
3521                 }
3522         }
3523 }
3524
3525 static pset *
3526 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3527 {
3528         pset     *spills = pset_new_ptr_default();
3529
3530         const ir_node  *next;
3531         defs_t         *defs;
3532
3533         defs = set_find_def(si->values, value);
3534
3535         if(defs && defs->spills) {
3536                 for(next = defs->spills; next; next = get_irn_link(next)) {
3537                         pset_insert_ptr(spills, next);
3538                 }
3539         }
3540
3541         return spills;
3542 }
3543
3544 /**
3545  * @param before   The node after which the spill will be placed in the schedule
3546  */
3547 static ir_node *
3548 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3549 {
3550         defs_t   *defs;
3551         ir_node  *spill;
3552         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3553
3554         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3555
3556         spill = be_spill2(arch_env, irn, before);
3557
3558         defs = set_insert_def(si->values, value);
3559         assert(defs);
3560
3561         /* enter into the linked list */
3562         set_irn_link(spill, defs->spills);
3563         defs->spills = spill;
3564
3565         if(opt_keep_alive & KEEPALIVE_SPILLS)
3566                 pset_insert_ptr(si->spills, spill);
3567
3568         return spill;
3569 }
3570
3571 /**
3572  * @param before   The Phi node which has to be spilled
3573  */
3574 static ir_node *
3575 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3576 {
3577         ir_node   *mem_phi;
3578         ir_node  **ins;
3579         defs_t    *defs;
3580         int        n;
3581
3582         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3583
3584         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3585                 ins[n] = si->m_unknown;
3586         }
3587
3588         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3589
3590         defs = set_insert_def(si->values, phi);
3591         assert(defs);
3592
3593         /* enter into the linked list */
3594         set_irn_link(mem_phi, defs->spills);
3595         defs->spills = mem_phi;
3596
3597 #ifdef SCHEDULE_PHIM
3598         sched_add_after(phi, mem_phi);
3599 #endif
3600
3601         if(opt_keep_alive & KEEPALIVE_SPILLS)
3602                 pset_insert_ptr(si->spills, mem_phi);
3603
3604
3605         return mem_phi;
3606 }
3607
3608 /**
3609  * Add remat to list of defs, destroys link field!
3610  */
3611 static void
3612 insert_remat(spill_ilp_t * si, ir_node * remat)
3613 {
3614         defs_t   *defs;
3615         op_t     *remat_op = get_irn_link(remat);
3616
3617         assert(remat_op->is_remat);
3618
3619         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3620         assert(defs);
3621
3622         /* enter into the linked list */
3623         set_irn_link(remat, defs->remats);
3624         defs->remats = remat;
3625 }
3626
3627
3628 /**
3629  * Add reload before operation and add to list of defs
3630  */
3631 static ir_node *
3632 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3633 {
3634         defs_t   *defs;
3635         ir_node  *reload,
3636                          *spill;
3637         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3638
3639         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3640
3641         defs = set_find_def(si->values, value);
3642
3643         spill = defs->spills;
3644         assert(spill && "no spill placed before reload");
3645
3646         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3647
3648         /* enter into the linked list */
3649         set_irn_link(reload, defs->remats);
3650         defs->remats = reload;
3651
3652         return reload;
3653 }
3654
3655 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3656 {
3657         defs_t           *defs;
3658         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3659         ir_node          *spill;
3660         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3661
3662         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3663
3664         defs = set_find_def(si->values, value);
3665
3666         spill = defs->spills;
3667         assert(spill && "no spill placed before reload");
3668
3669         arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3670 }
3671
3672 void insert_memoperands(spill_ilp_t * si)
3673 {
3674         memoperand_t   *memoperand;
3675         lpp_name_t     *name;
3676
3677         set_foreach(si->memoperands, memoperand) {
3678                 name = si->lpp->vars[memoperand->ilp];
3679                 if(!is_zero(name->value)) {
3680                         perform_memory_operand(si, memoperand);
3681                 }
3682         }
3683 }
3684
3685 static void
3686 walker_spill_placer(ir_node * bb, void * data) {
3687         spill_ilp_t   *si = (spill_ilp_t*)data;
3688         ir_node       *irn;
3689         spill_bb_t    *spill_bb = get_irn_link(bb);
3690         pset          *spills_to_do = pset_new_ptr_default();
3691         spill_t       *spill;
3692
3693         set_foreach(spill_bb->ilp, spill) {
3694                 lpp_name_t    *name;
3695
3696                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3697                         name = si->lpp->vars[spill->mem_in];
3698                         if(!is_zero(name->value)) {
3699                                 ir_node   *mem_phi;
3700
3701                                 mem_phi = insert_mem_phi(si, spill->irn);
3702
3703                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3704                         }
3705                 }
3706
3707                 name = si->lpp->vars[spill->spill];
3708                 if(!is_zero(name->value)) {
3709                         /* place spill directly after definition */
3710                         if(get_nodes_block(spill->irn) == bb) {
3711                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3712                                 continue;
3713                         }
3714
3715                         /* place spill at bb start */
3716                         if(spill->reg_in > 0) {
3717                                 name = si->lpp->vars[spill->reg_in];
3718                                 if(!is_zero(name->value)) {
3719                                         insert_spill(si, spill->irn, spill->irn, bb);
3720                                         continue;
3721                                 }
3722                         }
3723                         /* place spill after a remat */
3724                         pset_insert_ptr(spills_to_do, spill->irn);
3725                 }
3726         }
3727         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3728
3729
3730         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3731                 op_t     *op = get_irn_link(irn);
3732
3733                 if(be_is_Spill(irn)) continue;
3734
3735                 if(op->is_remat) {
3736                         /* TODO fix this if we want to support remats with more than two nodes */
3737                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3738                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3739
3740                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3741                         }
3742                 } else {
3743                         if(pset_find_ptr(spills_to_do, irn)) {
3744                                 pset_remove_ptr(spills_to_do, irn);
3745
3746                                 insert_spill(si, irn, irn, irn);
3747                         }
3748                 }
3749
3750         }
3751
3752         assert(pset_count(spills_to_do) == 0);
3753
3754         /* afterwards free data in block */
3755         del_pset(spills_to_do);
3756 }
3757
3758 static ir_node *
3759 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3760 {
3761         ir_node          *insert_pos = bb;
3762         ir_node          *spill;
3763         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3764
3765         /* find last definition of arg value in block */
3766         ir_node  *next;
3767         defs_t   *defs;
3768         int       last = 0;
3769
3770         defs = set_find_def(si->values, value);
3771
3772         if(defs && defs->remats) {
3773                 for(next = defs->remats; next; next = get_irn_link(next)) {
3774                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3775                                 last = sched_get_time_step(next);
3776                                 insert_pos = next;
3777                         }
3778                 }
3779         }
3780
3781         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3782                 last = sched_get_time_step(value);
3783                 insert_pos = value;
3784         }
3785
3786         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3787
3788         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3789
3790         return spill;
3791 }
3792
3793 static void
3794 phim_fixer(spill_ilp_t *si) {
3795         defs_t  *defs;
3796
3797         set_foreach(si->values, defs) {
3798                 const ir_node  *phi = defs->value;
3799                 op_t           *op = get_irn_link(phi);
3800                 ir_node        *phi_m = NULL;
3801                 ir_node        *next = defs->spills;
3802                 int             n;
3803
3804                 if(!is_Phi(phi)) continue;
3805
3806                 while(next) {
3807                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3808                                 phi_m = next;
3809                                 break;
3810                         } else {
3811                                 next = get_irn_link(next);
3812                         }
3813                 }
3814                 if(!phi_m) continue;
3815
3816                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3817                         ir_node        *value = get_irn_n(phi, n);
3818                         defs_t         *val_defs = set_find_def(si->values, value);
3819
3820                         /* a spill of this value */
3821                         ir_node      *spill;
3822
3823
3824                         if(opt_memcopies) {
3825                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3826                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3827
3828                                 if(!is_zero(name->value)) {
3829                                         spill = insert_mem_copy(si, pred, value);
3830                                 } else {
3831                                         spill = val_defs->spills;
3832                                 }
3833                         } else {
3834                                 spill = val_defs->spills;
3835                         }
3836
3837                         assert(spill && "no spill placed before PhiM");
3838                         set_irn_n(phi_m, n, spill);
3839                 }
3840         }
3841 }
3842
3843 static void
3844 walker_reload_placer(ir_node * bb, void * data) {
3845         spill_ilp_t   *si = (spill_ilp_t*)data;
3846         ir_node       *irn;
3847         spill_bb_t    *spill_bb = get_irn_link(bb);
3848
3849         /* reloads at end of block */
3850         if(spill_bb->reloads) {
3851                 keyval_t    *keyval;
3852
3853                 set_foreach(spill_bb->reloads, keyval) {
3854                         ir_node        *irn = (ir_node*)keyval->key;
3855                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3856                         lpp_name_t     *name;
3857
3858                         name = si->lpp->vars[reload];
3859                         if(!is_zero(name->value)) {
3860                                 ir_node    *reload;
3861                                 ir_node    *insert_pos = bb;
3862                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3863                                 op_t       *prev_op = get_irn_link(prev);
3864
3865                                 while(be_is_Spill(prev)) {
3866                                         prev = sched_prev(prev);
3867                                 }
3868
3869                                 prev_op = get_irn_link(prev);
3870
3871                                 /* insert reload before pre-remats */
3872                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3873                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3874                                         insert_pos = prev;
3875
3876                                         do {
3877                                                 prev = sched_prev(prev);
3878                                         } while(be_is_Spill(prev));
3879
3880                                         prev_op = get_irn_link(prev);
3881
3882                                 }
3883
3884                                 reload = insert_reload(si, irn, insert_pos);
3885
3886                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3887                                         pset_insert_ptr(si->spills, reload);
3888                         }
3889                 }
3890         }
3891
3892         /* walk and insert more reloads and collect remats */
3893         sched_foreach_reverse(bb, irn) {
3894                 op_t     *op = get_irn_link(irn);
3895
3896                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3897                 if(is_Phi(irn)) break;
3898
3899                 if(op->is_remat) {
3900                         if(get_irn_mode(irn) != mode_T) {
3901                                 insert_remat(si, irn);
3902                         }
3903                 } else {
3904                         int    n;
3905
3906                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3907                                 ir_node    *arg = get_irn_n(irn, n);
3908
3909                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3910                                         lpp_name_t    *name;
3911
3912                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3913                                         if(!is_zero(name->value)) {
3914                                                 ir_node    *reload;
3915                                                 ir_node    *insert_pos = irn;
3916                                                 ir_node    *prev = sched_prev(insert_pos);
3917                                                 op_t       *prev_op;
3918
3919                                                 while(be_is_Spill(prev)) {
3920                                                         prev = sched_prev(prev);
3921                                                 }
3922
3923                                                 prev_op = get_irn_link(prev);
3924
3925                                                 /* insert reload before pre-remats */
3926                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3927                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3928                                                         insert_pos = prev;
3929
3930                                                         do {
3931                                                                 prev = sched_prev(prev);
3932                                                         } while(be_is_Spill(prev));
3933
3934                                                         prev_op = get_irn_link(prev);
3935
3936                                                 }
3937
3938                                                 reload = insert_reload(si, arg, insert_pos);
3939
3940                                                 set_irn_n(irn, n, reload);
3941
3942                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3943                                                         pset_insert_ptr(si->spills, reload);
3944                                         }
3945                                 }
3946                         }
3947                 }
3948         }
3949
3950         del_set(spill_bb->ilp);
3951         if(spill_bb->reloads) del_set(spill_bb->reloads);
3952 }
3953
3954 static void
3955 walker_collect_used(ir_node * irn, void * data)
3956 {
3957         bitset_t   *used = data;
3958
3959         bitset_set(used, get_irn_idx(irn));
3960 }
3961
3962 struct kill_helper {
3963         bitset_t  *used;
3964         spill_ilp_t  *si;
3965 };
3966
3967 static void
3968 walker_kill_unused(ir_node * bb, void * data)
3969 {
3970         struct kill_helper *kh = data;
3971         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
3972         ir_node            *irn;
3973
3974
3975         for(irn=sched_first(bb); !sched_is_end(irn);) {
3976                 ir_node     *next = sched_next(irn);
3977                 int          n;
3978
3979                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3980                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3981                                 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)));
3982 #if 0
3983                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3984 #endif
3985                         }
3986
3987                         sched_remove(irn);
3988
3989                         set_nodes_block(irn, bad);
3990                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3991                                 set_irn_n(irn, n, bad);
3992                         }
3993                 }
3994                 irn = next;
3995         }
3996 }
3997
3998 static void
3999 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4000 {
4001         struct kill_helper kh;
4002
4003         kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
4004         kh.si = si;
4005
4006         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
4007         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
4008
4009         bitset_free(kh.used);
4010 }
4011
4012 void
4013 print_irn_pset(pset * p)
4014 {
4015         ir_node   *irn;
4016
4017         pset_foreach(p, irn) {
4018                 ir_printf("%+F\n", irn);
4019         }
4020 }
4021
4022 void
4023 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
4024 {
4025     FILE           *f = fopen(file, "w");
4026     ir_node        *irn;
4027     interference_t *interference;
4028
4029     pset_break(phiclass);
4030     set_break(si->interferences);
4031
4032     ir_fprintf(f, "digraph phiclass {\n");
4033
4034     pset_foreach(phiclass, irn) {
4035         if(is_Phi(irn))
4036             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
4037     }
4038
4039     pset_foreach(phiclass, irn) {
4040         int n;
4041
4042         if(!is_Phi(irn)) continue;
4043
4044         for(n=get_irn_arity(irn)-1; n>=0; --n) {
4045             ir_node  *arg = get_irn_n(irn, n);
4046
4047             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
4048         }
4049     }
4050
4051     set_foreach(si->interferences, interference) {
4052         const ir_node  *a  = interference->a;
4053         const ir_node  *b  = interference->b;
4054         if(get_phi_class(a) == phiclass) {
4055             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4056         }
4057     }
4058
4059     ir_fprintf(f, "}");
4060     fclose(f);
4061 }
4062
4063 static void
4064 rewire_uses(spill_ilp_t * si)
4065 {
4066         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4067         defs_t               *defs;
4068         pset                 *ignore = pset_new_ptr(1);
4069
4070         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4071
4072         /* then fix uses of spills */
4073         set_foreach(si->values, defs) {
4074                 pset           *reloads;
4075                 pset           *spills;
4076                 const ir_node  *next = defs->remats;
4077                 int remats = 0;
4078
4079                 reloads = pset_new_ptr_default();
4080
4081                 while(next) {
4082                         if(be_is_Reload(next)) {
4083                                 pset_insert_ptr(reloads, next);
4084                         } else {
4085                                 ++remats;
4086                         }
4087                         next = get_irn_link(next);
4088                 }
4089
4090                 spills = get_spills_for_value(si, defs->value);
4091                 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));
4092                 if(pset_count(spills) > 1) {
4093                         //assert(pset_count(reloads) > 0);
4094                         //                              print_irn_pset(spills);
4095                         //                              print_irn_pset(reloads);
4096
4097                         be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4098                 }
4099
4100                 del_pset(reloads);
4101                 del_pset(spills);
4102         }
4103
4104         /* first fix uses of remats and reloads */
4105         set_foreach(si->values, defs) {
4106                 pset           *nodes;
4107                 const ir_node  *next = defs->remats;
4108                 int             orig_kept = 0;
4109
4110                 if(next) {
4111                         nodes = pset_new_ptr_default();
4112                         if(sched_is_scheduled(defs->value)) {
4113                                 pset_insert_ptr(nodes, defs->value);
4114                                 orig_kept = 1;
4115                         }
4116
4117                         while(next) {
4118                                 pset_insert_ptr(nodes, next);
4119                                 next = get_irn_link(next);
4120                         }
4121
4122                         DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4123                         be_ssa_constr_set(dfi, si->lv, nodes);
4124
4125                         del_pset(nodes);
4126                 }
4127         }
4128
4129 //      remove_unused_defs(si);
4130
4131         be_free_dominance_frontiers(dfi);
4132 }
4133
4134
4135 static void
4136 writeback_results(spill_ilp_t * si)
4137 {
4138         /* walk through the graph and collect all spills, reloads and remats for a value */
4139
4140         si->values = new_set(cmp_defs, 4096);
4141
4142         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4143         delete_unnecessary_remats(si);
4144         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4145         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4146         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4147         if(opt_memoperands)
4148                 insert_memoperands(si);
4149         phim_fixer(si);
4150
4151         /* clean the remat info! there are still back-edges leading there! */
4152         clean_remat_info(si);
4153
4154         rewire_uses(si);
4155
4156         connect_all_spills_with_keep(si);
4157
4158         del_set(si->values);
4159 }
4160
4161 static int
4162 get_n_regs(spill_ilp_t * si)
4163 {
4164         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4165
4166         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4167         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4168
4169         arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4170     be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4171
4172         bitset_andnot(arch_regs, abi_regs);
4173         arch_n_regs = bitset_popcnt(arch_regs);
4174
4175         bitset_free(arch_regs);
4176         bitset_free(abi_regs);
4177
4178         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4179         return arch_n_regs;
4180 }
4181
4182 static void
4183 walker_reload_mover(ir_node * bb, void * data)
4184 {
4185         spill_ilp_t   *si = data;
4186         ir_node           *tmp;
4187
4188         sched_foreach(bb, tmp) {
4189                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4190                         ir_node       *reload = tmp;
4191                         ir_node       *irn = tmp;
4192
4193                         /* move reload upwards */
4194
4195                         int pressure = (int)get_irn_link(reload);
4196                         if(pressure < si->n_regs) {
4197                                 irn = sched_prev(reload);
4198                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4199                                 sched_remove(reload);
4200                                 pressure = (int)get_irn_link(irn);
4201
4202                                 while(pressure < si->n_regs) {
4203                                         if( sched_is_end(irn) ||
4204                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4205                                            /* do not move reload before its spill */
4206                                            (irn == be_get_Reload_mem(reload)) ||
4207                                            /* do not move before phi */
4208                                            is_Phi(irn)) break;
4209
4210                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4211                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4212                                         irn = sched_prev(irn);
4213
4214                                         pressure = (int)get_irn_link(irn);
4215                                 }
4216
4217                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4218                                 sched_put_after(irn, reload);
4219                         }
4220                 }
4221         }
4222 }
4223
4224 static void
4225 move_reloads_upward(spill_ilp_t * si)
4226 {
4227         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4228 }
4229
4230
4231 /**
4232  * Walk all irg blocks and check for interfering spills inside of phi classes
4233  */
4234 static void
4235 luke_meminterferencechecker(ir_node * bb, void * data)
4236 {
4237         spill_ilp_t    *si = (spill_ilp_t*)data;
4238         int             l1, l2;
4239
4240         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4241                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4242
4243                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4244
4245                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4246                 if (has_reg_class(si, a) && get_phi_class(a)) {
4247                         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)) {
4248                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
4249
4250                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4251
4252                                 /* a and b are only interesting if they are in the same phi class */
4253                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4254                                         if(values_interfere_in_block(si, bb, a, b)) {
4255                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4256                                         }
4257                                 }
4258                         }
4259                 }
4260         }
4261 }
4262
4263 static void
4264 verify_phiclasses(spill_ilp_t * si)
4265 {
4266         /* analyze phi classes */
4267         phi_class_compute(si->chordal_env->irg);
4268
4269         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4270         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4271 }
4272
4273 void
4274 be_spill_remat(const be_chordal_env_t * chordal_env)
4275 {
4276         char            buf[256];
4277         char            problem_name[256];
4278         char            dump_suffix[256];
4279         char            dump_suffix2[256];
4280         struct obstack  obst;
4281         spill_ilp_t     si;
4282
4283         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4284         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4285         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4286
4287         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4288         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4289
4290         if(opt_verify & VERIFY_DOMINANCE)
4291                 be_check_dominance(chordal_env->irg);
4292
4293         obstack_init(&obst);
4294         si.chordal_env = chordal_env;
4295         si.obst = &obst;
4296         si.cls = chordal_env->cls;
4297         si.lpp = new_lpp(problem_name, lpp_minimize);
4298         si.remat_info = new_set(cmp_remat_info, 4096);
4299         si.interferences = new_set(cmp_interference, 32);
4300         si.memoperands = new_set(cmp_memoperands, 128);
4301         si.all_possible_remats = pset_new_ptr_default();
4302         si.spills = pset_new_ptr_default();
4303         si.inverse_ops = pset_new_ptr_default();
4304         si.lv = chordal_env->lv;
4305         si.keep = NULL;
4306         si.n_regs = get_n_regs(&si);
4307
4308         set_irg_link(chordal_env->irg, &si);
4309         compute_doms(chordal_env->irg);
4310
4311         /* compute phi classes */
4312 //      phi_class_compute(chordal_env->irg);
4313
4314         if(opt_dump_flags & DUMP_STATS)
4315                 be_analyze_regpressure(chordal_env, "-pre");
4316
4317         DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4318         irg_block_walk_graph(chordal_env->irg, luke_initializer, NULL, &si);
4319
4320         if(opt_remats) {
4321                 /* collect remats */
4322                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4323                 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4324         }
4325
4326         /* insert possible remats */
4327         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4328         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4329         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4330
4331         if(opt_keep_alive & KEEPALIVE_REMATS) {
4332                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4333                 connect_all_remats_with_keep(&si);
4334                 /* dump graph with inserted remats */
4335                 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4336         }
4337
4338         /* insert copies for phi arguments not in my regclass */
4339         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4340
4341         /* recompute liveness */
4342         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4343         be_liveness_recompute(si.lv);
4344
4345         /* build the ILP */
4346         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4347         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4348         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4349
4350         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4351         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4352
4353         if(opt_memcopies) {
4354                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4355                 memcopyhandler(&si);
4356         }
4357
4358         if(opt_dump_flags & DUMP_PROBLEM) {
4359                 FILE           *f;
4360                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4361                 if ((f = fopen(buf, "wt")) != NULL) {
4362                         lpp_dump_plain(si.lpp, f);
4363                         fclose(f);
4364                 }
4365         }
4366
4367         if(opt_dump_flags & DUMP_MPS) {
4368                 FILE *f;
4369
4370                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4371                 if((f = fopen(buf, "wt")) != NULL) {
4372                         mps_write_mps(si.lpp, s_mps_fixed, f);
4373                         fclose(f);
4374                 }
4375
4376                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4377                 if((f = fopen(buf, "wt")) != NULL) {
4378                         mps_write_mst(si.lpp, s_mps_fixed, f);
4379                         fclose(f);
4380                 }
4381         }
4382
4383         lpp_check_startvals(si.lpp);
4384
4385 #ifdef SOLVE
4386         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4387         lpp_set_time_limit(si.lpp, opt_timeout);
4388
4389         if(opt_log)
4390                 lpp_set_log(si.lpp, stdout);
4391
4392 #ifdef SOLVE_LOCAL
4393         lpp_solve_cplex(si.lpp);
4394 #else
4395         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4396 #endif
4397         assert(lpp_is_sol_valid(si.lpp)
4398                && "solution of ILP must be valid");
4399
4400         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));
4401
4402         if(opt_dump_flags & DUMP_SOLUTION) {
4403                 FILE           *f;
4404                 char            buf[256];
4405
4406                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4407                 if ((f = fopen(buf, "wt")) != NULL) {
4408                         int             i;
4409                         for (i = 0; i < si.lpp->var_next; ++i) {
4410                                 lpp_name_t     *name = si.lpp->vars[i];
4411                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4412                         }
4413                         fclose(f);
4414                 }
4415         }
4416
4417         writeback_results(&si);
4418
4419 #endif                          /* SOLVE */
4420
4421         kill_all_unused_values_in_schedule(&si);
4422
4423         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4424                 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4425
4426         // move reloads upwards
4427         be_liveness_recompute(si.lv);
4428         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4429         move_reloads_upward(&si);
4430
4431         if(opt_memcopies) {
4432                 verify_phiclasses(&si);
4433         }
4434
4435         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4436
4437         if(opt_dump_flags & DUMP_PRESSURE)
4438                 dump_pressure_graph(&si, dump_suffix2);
4439
4440         if(opt_dump_flags & DUMP_STATS)
4441                 be_analyze_regpressure(chordal_env, "-post");
4442
4443         if(opt_verify & VERIFY_DOMINANCE)
4444                 be_check_dominance(chordal_env->irg);
4445
4446         free_dom(chordal_env->irg);
4447         del_set(si.interferences);
4448         del_pset(si.inverse_ops);
4449         del_pset(si.all_possible_remats);
4450         del_set(si.memoperands);
4451         del_pset(si.spills);
4452         free_lpp(si.lpp);
4453         obstack_free(&obst, NULL);
4454         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4455 }
4456
4457 #else                           /* WITH_ILP */
4458
4459 static void
4460 only_that_you_can_compile_without_WITH_ILP_defined(void)
4461 {
4462 }
4463
4464 #endif                          /* WITH_ILP */