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 "benode_t.h"
  50 #include "beutil.h"
  51 #include "bespillremat.h"
  52 #include "bespill.h"
  53 #include "bepressurestat.h"
  54
  55 #include "bechordal_t.h"
  56
  57 #ifdef WITH_LIBCORE
  58 #include <libcore/lc_opts.h>
  59 #include <libcore/lc_opts_enum.h>
  60 #endif /* WITH_LIBCORE */
  61
  62 #define DUMP_PROBLEM       1
  63 #define DUMP_MPS           2
  64 #define DUMP_SOLUTION      4
  65
  66 #define KEEPALIVE_REMATS   1
  67 #define KEEPALIVE_SPILLS   2
  68 #define KEEPALIVE_RELOADS  4
  69
  70 #define VERIFY_MEMINTERF   1
  71 #define VERIFY_DOMINANCE   2
  72
  73 #define REMATS_NONE        0
  74 #define REMATS_BRIGGS      1
  75 #define REMATS_NOINVERSE   2
  76 #define REMATS_ALL         3
  77
  78 static int opt_dump_flags   = 0;
  79 static int opt_log = 0;
  80 static int opt_keep_alive   = 0;
  81 static int opt_goodwin = 1;
  82 static int opt_memcopies = 1;
  83 static int opt_memoperands = 1;
  84 static int opt_verify = VERIFY_MEMINTERF;
  85 static int opt_remats = REMATS_ALL;
  86 static int opt_repair_schedule = 0;
  87 static int opt_no_enlarge_liveness = 0;
  88 static int opt_remat_while_live = 1;
  89 static int opt_timeout = 300;
  90 static double opt_cost_reload = 8.0;
  91 static double opt_cost_memoperand =  7.0;
  92 static double opt_cost_spill =  50.0;
  93 static double opt_cost_remat =  1.0;
  94
  95
  96 #ifdef WITH_LIBCORE
  97 static const lc_opt_enum_mask_items_t dump_items[] = {
  98         { "problem",  DUMP_PROBLEM  },
  99         { "mps",      DUMP_MPS      },
 100         { "solution", DUMP_SOLUTION },
 101         { NULL,       0 }
 102 };
 103
 104 static lc_opt_enum_mask_var_t dump_var = {
 105         &opt_dump_flags, dump_items
 106 };
 107
 108 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 109         { "remats",  KEEPALIVE_REMATS  },
 110         { "spills",  KEEPALIVE_SPILLS  },
 111         { "reloads", KEEPALIVE_RELOADS },
 112         { NULL,      0 }
 113 };
 114
 115 static lc_opt_enum_mask_var_t keep_alive_var = {
 116         &opt_keep_alive, keepalive_items
 117 };
 118
 119 static const lc_opt_enum_mask_items_t remats_items[] = {
 120         { "none",      REMATS_NONE      },
 121         { "briggs",    REMATS_BRIGGS    },
 122         { "noinverse", REMATS_NOINVERSE },
 123         { "all",       REMATS_ALL       },
 124         { NULL,        0 }
 125 };
 126
 127 static lc_opt_enum_mask_var_t remats_var = {
 128         &opt_remats, remats_items
 129 };
 130
 131 static const lc_opt_table_entry_t options[] = {
 132         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads",                   &keep_alive_var),
 133
 134         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 135         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 136         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 137         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
 138         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 139         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 140         LC_OPT_ENT_BOOL     ("remat_while_live",  "remat only values that can be used by real ops", &opt_remat_while_live),
 141
 142         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution",                               &dump_var),
 143         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 144         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 145
 146         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 147         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 148         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 149         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 150         { NULL }
 151 };
 152
 153 void be_spill_remat_register_options(lc_opt_entry_t *grp)
 154 {
 155         lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
 156         lc_opt_add_table(my_grp, options);
 157 }
 158 #endif
 159
 160
 161 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 162 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 163
 164 #define  SOLVE
 165 //#define  SOLVE_LOCAL
 166 #define LPP_SERVER "i44pc52"
 167 #define LPP_SOLVER "cplex"
 168
 169
 170 #define MAX_PATHS      INT_MAX
 171 #define ILP_UNDEF               -1
 172
 173 typedef struct _spill_ilp_t {
 174         const arch_register_class_t  *cls;
 175         int                           n_regs;
 176         const be_chordal_env_t       *chordal_env;
 177         be_lv_t                      *lv;
 178         lpp_t                        *lpp;
 179         struct obstack               *obst;
 180         set                          *remat_info;
 181         pset                         *all_possible_remats;
 182         pset                         *inverse_ops;
 183         ir_node                      *keep;
 184         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 185         pset                         *spills;
 186         set                          *interferences;
 187         ir_node                      *m_unknown;
 188         set                          *memoperands;
 189         DEBUG_ONLY(firm_dbg_module_t * dbg);
 190 } spill_ilp_t;
 191
 192 typedef int ilp_var_t;
 193 typedef int ilp_cst_t;
 194
 195 typedef struct _spill_bb_t {
 196         set      *ilp;
 197         set      *reloads;
 198 } spill_bb_t;
 199
 200 typedef struct _remat_t {
 201         const ir_node        *op;      /**< for copy_irn */
 202         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 203         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 204         int                   cost;    /**< cost of this remat */
 205         int                   inverse; /**< nonzero if this is an inverse remat */
 206 } remat_t;
 207
 208 /**
 209  * Data to be attached to each IR node. For remats this contains the ilp_var
 210  * for this remat and for normal ops this contains the ilp_vars for
 211  * reloading each operand
 212  */
 213 typedef struct _op_t {
 214         int             is_remat;
 215         union {
 216                 struct {
 217                         ilp_var_t       ilp;
 218                         const remat_t  *remat; /** the remat this op belongs to */
 219                         int             pre; /** 1, if this is a pressure-increasing remat */
 220                 } remat;
 221                 struct {
 222                         ilp_var_t       ilp;
 223                         ir_node        *op; /** the operation this live range belongs to */
 224                         union {
 225                                 ilp_var_t      *reloads;
 226                                 ilp_var_t      *copies;
 227                         } args;
 228                 } live_range;
 229         } attr;
 230 } op_t;
 231
 232 typedef struct _defs_t {
 233         const ir_node   *value;
 234         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 235         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 236 } defs_t;
 237
 238 typedef struct _remat_info_t {
 239         const ir_node       *irn; /**< the irn to which these remats belong */
 240         pset                *remats; /**< possible remats for this value */
 241         pset                *remats_by_operand; /**< remats with this value as operand */
 242 } remat_info_t;
 243
 244 typedef struct _keyval_t {
 245         const void          *key;
 246         const void          *val;
 247 } keyval_t;
 248
 249 typedef struct _spill_t {
 250         ir_node            *irn;
 251         ilp_var_t           reg_in;
 252         ilp_var_t           mem_in;
 253         ilp_var_t           reg_out;
 254         ilp_var_t           mem_out;
 255         ilp_var_t           spill;
 256 } spill_t;
 257
 258 typedef struct _memoperand_t {
 259         ir_node             *irn; /**< the irn */
 260         unsigned int         pos; /**< the position of the argument */
 261         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 262 } memoperand_t;
 263
 264 static INLINE int
 265 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 266 {
 267         return chordal_has_class(si->chordal_env, irn);
 268 }
 269
 270 #if 0
 271 static int
 272 cmp_remat(const void *a, const void *b)
 273 {
 274         const keyval_t *p = a;
 275         const keyval_t *q = b;
 276         const remat_t  *r = p->val;
 277         const remat_t  *s = q->val;
 278
 279         assert(r && s);
 280
 281         return !(r == s || r->op == s->op);
 282 }
 283 #endif
 284 static int
 285 cmp_remat(const void *a, const void *b)
 286 {
 287         const remat_t  *r = a;
 288         const remat_t  *s = a;
 289
 290         return !(r == s || r->op == s->op);
 291 }
 292
 293 static int
 294 cmp_spill(const void *a, const void *b, size_t size)
 295 {
 296         const spill_t *p = a;
 297         const spill_t *q = b;
 298
 299 //      return !(p->irn == q->irn && p->bb == q->bb);
 300         return !(p->irn == q->irn);
 301 }
 302
 303 static int
 304 cmp_memoperands(const void *a, const void *b, size_t size)
 305 {
 306         const memoperand_t *p = a;
 307         const memoperand_t *q = b;
 308
 309         return !(p->irn == q->irn && p->pos == q->pos);
 310 }
 311
 312 static keyval_t *
 313 set_find_keyval(set * set, const void * key)
 314 {
 315         keyval_t     query;
 316
 317         query.key = key;
 318         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 319 }
 320
 321 static keyval_t *
 322 set_insert_keyval(set * set, void * key, void * val)
 323 {
 324         keyval_t     query;
 325
 326         query.key = key;
 327         query.val = val;
 328         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 329 }
 330
 331 static defs_t *
 332 set_find_def(set * set, const ir_node * value)
 333 {
 334         defs_t     query;
 335
 336         query.value = value;
 337         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 338 }
 339
 340 static defs_t *
 341 set_insert_def(set * set, const ir_node * value)
 342 {
 343         defs_t     query;
 344
 345         query.value = value;
 346         query.spills = NULL;
 347         query.remats = NULL;
 348         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 349 }
 350
 351 static memoperand_t *
 352 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 353 {
 354         memoperand_t     query;
 355
 356         query.irn = irn;
 357         query.pos = pos;
 358         query.ilp = ilp;
 359         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 360 }
 361
 362 static memoperand_t *
 363 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 364 {
 365         memoperand_t     query;
 366
 367         query.irn = (ir_node*)irn;
 368         query.pos = pos;
 369         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 370 }
 371
 372
 373 static spill_t *
 374 set_find_spill(set * set, const ir_node * value)
 375 {
 376         spill_t     query;
 377
 378         query.irn = (ir_node*)value;
 379         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 380 }
 381
 382 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 383 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 384 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 385 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 386 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 387
 388 static int
 389 cmp_remat_info(const void *a, const void *b, size_t size)
 390 {
 391         const remat_info_t *p = a;
 392         const remat_info_t *q = b;
 393
 394         return !(p->irn == q->irn);
 395 }
 396
 397 static int
 398 cmp_defs(const void *a, const void *b, size_t size)
 399 {
 400         const defs_t *p = a;
 401         const defs_t *q = b;
 402
 403         return !(p->value == q->value);
 404 }
 405
 406 static int
 407 cmp_keyval(const void *a, const void *b, size_t size)
 408 {
 409         const keyval_t *p = a;
 410         const keyval_t *q = b;
 411
 412         return !(p->key == q->key);
 413 }
 414
 415 static double
 416 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 417 {
 418 #define FUDGE 0.001
 419 #ifndef EXECFREQ_LOOPDEPH
 420         return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
 421 #else
 422         if(is_Block(irn))
 423                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 424         else
 425                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 426 #endif
 427 }
 428
 429 static double
 430 get_cost(const spill_ilp_t * si, const ir_node * irn)
 431 {
 432         if(be_is_Spill(irn)) {
 433                 return opt_cost_spill;
 434         } else if(be_is_Reload(irn)){
 435                 return opt_cost_reload;
 436         } else {
 437                 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
 438         }
 439 }
 440
 441 /**
 442  * Checks, whether node and its operands have suitable reg classes
 443  */
 444 static INLINE int
 445 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 446 {
 447         int               n;
 448         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 449         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 450
 451 #if 0
 452         if(!remat)
 453                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 454 #endif
 455
 456         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 457                 ir_node        *op = get_irn_n(irn, n);
 458                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 459
 460 //              if(!remat)
 461 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 462         }
 463
 464         return remat;
 465 }
 466
 467 /**
 468  * Try to create a remat from @p op with destination value @p dest_value
 469  */
 470 static INLINE remat_t *
 471 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 472 {
 473         remat_t  *remat = NULL;
 474
 475 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 476 //              return NULL;
 477
 478         if(dest_value == op) {
 479                 const ir_node *proj = NULL;
 480
 481                 if(is_Proj(dest_value)) {
 482                         op = get_Proj_pred(op);
 483                         proj = dest_value;
 484                 }
 485
 486                 if(!is_rematerializable(si, op))
 487                         return NULL;
 488
 489                 remat = obstack_alloc(si->obst, sizeof(*remat));
 490                 remat->op = op;
 491                 remat->cost = get_cost(si, op);
 492                 remat->value = dest_value;
 493                 remat->proj = proj;
 494                 remat->inverse = 0;
 495         } else {
 496                 arch_inverse_t     inverse;
 497                 int                n;
 498
 499                 /* get the index of the operand we want to retrieve by the inverse op */
 500                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 501                         ir_node        *arg = get_irn_n(op, n);
 502
 503                         if(arg == dest_value) break;
 504                 }
 505                 if(n<0) return NULL;
 506
 507                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 508
 509                 /* else ask the backend to give an inverse op */
 510                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 511                         int   i;
 512
 513                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 514
 515                         assert(inverse.n > 0 && "inverse op should have at least one node");
 516
 517                         for(i=inverse.n-1; i>=0; --i) {
 518                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 519                         }
 520
 521                         if(inverse.n <= 2) {
 522                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 523                                 remat->op = inverse.nodes[0];
 524                                 remat->cost = inverse.costs;
 525                                 remat->value = dest_value;
 526                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 527                                 remat->inverse = 1;
 528
 529                                 assert(is_Proj(remat->proj));
 530                         } else {
 531                                 assert(0 && "I can not handle remats with more than 2 nodes");
 532                         }
 533                 }
 534         }
 535
 536         if(remat) {
 537                 if(remat->proj) {
 538                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 539                 } else {
 540                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 541                 }
 542         }
 543         return remat;
 544 }
 545
 546
 547 static INLINE void
 548 add_remat(const spill_ilp_t * si, const remat_t * remat)
 549 {
 550         remat_info_t    *remat_info,
 551                      query;
 552         int              n;
 553
 554         assert(remat->op);
 555         assert(remat->value);
 556
 557         query.irn = remat->value;
 558         query.remats = NULL;
 559         query.remats_by_operand = NULL;
 560         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 561
 562         if(remat_info->remats == NULL) {
 563                 remat_info->remats = new_pset(cmp_remat, 4096);
 564         }
 565         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 566
 567         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 568         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 569                 ir_node        *arg = get_irn_n(remat->op, n);
 570
 571                 query.irn = arg;
 572                 query.remats = NULL;
 573                 query.remats_by_operand = NULL;
 574                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 575
 576                 if(remat_info->remats_by_operand == NULL) {
 577                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 578                 }
 579                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 580         }
 581 }
 582
 583 static int
 584 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 585 {
 586         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 587         int                i = 0;
 588
 589         while(edge) {
 590                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 591                         ++i;
 592                 }
 593                 edge = get_irn_out_edge_next(irn, edge);
 594         }
 595
 596         return i;
 597 }
 598
 599 static int
 600 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 601 {
 602         int n;
 603         int ret = 0;
 604
 605 //      if(is_Proj(irn))
 606 //              irn = get_Proj_pred(irn);
 607
 608         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 609                 const ir_node  *arg = get_irn_n(irn, n);
 610
 611                 if(has_reg_class(si, arg)) ++ret;
 612         }
 613
 614         return ret;
 615 }
 616
 617 static INLINE void
 618 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 619 {
 620         int      n;
 621         remat_t *remat;
 622
 623         if( has_reg_class(si, op)
 624         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 625         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 626         ) {
 627                 remat = get_remat_from_op(si, op, op);
 628                 if(remat) {
 629                         add_remat(si, remat);
 630                 }
 631         }
 632
 633         if(opt_remats == REMATS_ALL) {
 634                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 635                    for each arg */
 636                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 637                         ir_node        *arg = get_irn_n(op, n);
 638
 639                         if(has_reg_class(si, arg)) {
 640                                 /* try to get an inverse remat */
 641                                 remat = get_remat_from_op(si, arg, op);
 642                                 if(remat) {
 643                                         add_remat(si, remat);
 644                                 }
 645                         }
 646                 }
 647         }
 648 }
 649
 650 static INLINE int
 651 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 652 {
 653         ir_node *block;
 654         ir_node *def_block = get_nodes_block(val);
 655         int      ret;
 656
 657         if(val == pos)
 658                 return 0;
 659
 660         /* if pos is at end of a basic block */
 661         if(is_Block(pos)) {
 662                 ret = (pos == def_block || block_dominates(def_block, pos));
 663 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 664                 return ret;
 665         }
 666
 667         /* else if this is a normal operation */
 668         block = get_nodes_block(pos);
 669         if(block == def_block) {
 670                 if(!sched_is_scheduled(val)) return 1;
 671
 672                 ret = sched_comes_after(val, pos);
 673 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 674                 return ret;
 675         }
 676
 677         ret = block_dominates(def_block, block);
 678 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 679         return ret;
 680 }
 681
 682 static INLINE ir_node *
 683 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 684 {
 685     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 686 }
 687
 688 /**
 689  * Returns first non-Phi node of block @p bb
 690  */
 691 static INLINE ir_node *
 692 sched_block_first_nonphi(const ir_node * bb)
 693 {
 694         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 695 }
 696
 697 static int
 698 sched_skip_proj_predicator(const ir_node * irn, void * data)
 699 {
 700         return (is_Proj(irn));
 701 }
 702
 703 static INLINE ir_node *
 704 sched_next_nonproj(const ir_node * irn, int forward)
 705 {
 706         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 707 }
 708
 709 /**
 710  * Returns next operation node (non-Proj) after @p irn
 711  * or the basic block of this node
 712  */
 713 static INLINE ir_node *
 714 sched_next_op(const ir_node * irn)
 715 {
 716         ir_node *next = sched_next(irn);
 717
 718         if(is_Block(next))
 719                 return next;
 720
 721         return sched_next_nonproj(next, 1);
 722 }
 723
 724 /**
 725  * Returns previous operation node (non-Proj) before @p irn
 726  * or the basic block of this node
 727  */
 728 static INLINE ir_node *
 729 sched_prev_op(const ir_node * irn)
 730 {
 731         ir_node *prev = sched_prev(irn);
 732
 733         if(is_Block(prev))
 734                 return prev;
 735
 736         return sched_next_nonproj(prev, 0);
 737 }
 738
 739 static void
 740 sched_put_after(ir_node * insert, ir_node * irn)
 741 {
 742         if(is_Block(insert)) {
 743                 insert = sched_block_first_nonphi(insert);
 744         } else {
 745                 insert = sched_next_op(insert);
 746         }
 747         sched_add_before(insert, irn);
 748 }
 749
 750 static void
 751 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 752 {
 753   if(is_Block(insert)) {
 754           insert = sched_block_last_noncf(si, insert);
 755   } else {
 756           insert = sched_next_nonproj(insert, 0);
 757           insert = sched_prev(insert);
 758   }
 759   sched_add_after(insert, irn);
 760 }
 761
 762 /**
 763  * Tells you whether a @p remat can be placed before the irn @p pos
 764  */
 765 static INLINE int
 766 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 767 {
 768         const ir_node   *op = remat->op;
 769         const ir_node   *prev;
 770         int        n,
 771                            res = 1;
 772
 773         if(is_Block(pos)) {
 774                 prev = sched_block_last_noncf(si, pos);
 775                 prev = sched_next_nonproj(prev, 0);
 776         } else {
 777                 prev = sched_prev_op(pos);
 778         }
 779         /* do not remat if the rematted value is defined immediately before this op */
 780         if(prev == remat->op) {
 781                 return 0;
 782         }
 783
 784 #if 0
 785         /* this should be just fine, the following OP will be using this value, right? */
 786
 787         /* only remat AFTER the real definition of a value (?) */
 788         if(!value_is_defined_before(si, pos, remat->value)) {
 789 //              ir_fprintf(stderr, "error(not defined)");
 790                 return 0;
 791         }
 792 #endif
 793
 794         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 795                 const ir_node   *arg = get_irn_n(op, n);
 796
 797                 if(opt_no_enlarge_liveness) {
 798                         if(has_reg_class(si, arg) && live) {
 799                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 800                         } else {
 801                                 res &= value_is_defined_before(si, pos, arg);
 802                         }
 803                 } else {
 804                         res &= value_is_defined_before(si, pos, arg);
 805                 }
 806         }
 807
 808         return res;
 809 }
 810
 811 /**
 812  * Tells you whether a @p remat can be placed after the irn @p pos
 813  */
 814 static INLINE int
 815 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 816 {
 817         if(is_Block(pos)) {
 818                 pos = sched_block_first_nonphi(pos);
 819         } else {
 820                 pos = sched_next_op(pos);
 821         }
 822
 823         /* only remat AFTER the real definition of a value (?) */
 824         if(!value_is_defined_before(si, pos, remat->value)) {
 825                 return 0;
 826         }
 827
 828         return can_remat_before(si, remat, pos, live);
 829 }
 830
 831 /**
 832  * Collect potetially rematerializable OPs
 833  */
 834 static void
 835 walker_remat_collector(ir_node * irn, void * data)
 836 {
 837         spill_ilp_t    *si = data;
 838
 839         if(!is_Block(irn) && !is_Phi(irn)) {
 840                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 841                 get_remats_from_op(si, irn);
 842         }
 843 }
 844
 845 /**
 846  * Inserts a copy of @p irn before @p pos
 847  */
 848 static ir_node *
 849 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 850 {
 851         ir_node     *bb;
 852         ir_node     *copy;
 853
 854         bb = is_Block(pos)?pos:get_nodes_block(pos);
 855         copy = exact_copy(irn);
 856
 857         _set_phi_class(copy, NULL);
 858         set_nodes_block(copy, bb);
 859         sched_put_before(si, pos, copy);
 860
 861         return copy;
 862 }
 863
 864 /**
 865  * Inserts a copy of @p irn after @p pos
 866  */
 867 static ir_node *
 868 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 869 {
 870         ir_node     *bb;
 871         ir_node     *copy;
 872
 873         bb = is_Block(pos)?pos:get_nodes_block(pos);
 874         copy = exact_copy(irn);
 875
 876         _set_phi_class(copy, NULL);
 877         set_nodes_block(copy, bb);
 878         sched_put_after(pos, copy);
 879
 880         return copy;
 881 }
 882
 883 static ir_node *
 884 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 885 {
 886         char     buf[256];
 887
 888         if(can_remat_after(si, remat, pos, live)) {
 889                 ir_node         *copy,
 890                                                 *proj_copy;
 891                 op_t            *op;
 892
 893                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 894
 895                 copy = insert_copy_after(si, remat->op, pos);
 896
 897                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 898                 op = obstack_alloc(si->obst, sizeof(*op));
 899                 op->is_remat = 1;
 900                 op->attr.remat.remat = remat;
 901                 op->attr.remat.pre = 0;
 902                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 903
 904                 set_irn_link(copy, op);
 905                 pset_insert_ptr(si->all_possible_remats, copy);
 906                 if(remat->proj) {
 907                         proj_copy = insert_copy_after(si, remat->proj, copy);
 908                         set_irn_n(proj_copy, 0, copy);
 909                         set_irn_link(proj_copy, op);
 910                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 911                 } else {
 912                         proj_copy = NULL;
 913                 }
 914
 915                 return copy;
 916         }
 917
 918         return NULL;
 919 }
 920
 921 static ir_node *
 922 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 923 {
 924         char     buf[256];
 925
 926         if(can_remat_before(si, remat, pos, live)) {
 927                 ir_node         *copy,
 928                                                 *proj_copy;
 929                 op_t            *op;
 930
 931                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 932
 933                 copy = insert_copy_before(si, remat->op, pos);
 934
 935                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 936                 op = obstack_alloc(si->obst, sizeof(*op));
 937                 op->is_remat = 1;
 938                 op->attr.remat.remat = remat;
 939                 op->attr.remat.pre = 1;
 940                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 941
 942                 set_irn_link(copy, op);
 943                 pset_insert_ptr(si->all_possible_remats, copy);
 944                 if(remat->proj) {
 945                         proj_copy = insert_copy_after(si, remat->proj, copy);
 946                         set_irn_n(proj_copy, 0, copy);
 947                         set_irn_link(proj_copy, op);
 948                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 949                 } else {
 950                         proj_copy = NULL;
 951                 }
 952
 953                 return copy;
 954         }
 955
 956         return NULL;
 957 }
 958
 959 static int
 960 get_block_n_succs(const ir_node *block) {
 961         const ir_edge_t *edge;
 962
 963         assert(edges_activated(current_ir_graph));
 964
 965         edge = get_block_succ_first(block);
 966         if (! edge)
 967                 return 0;
 968
 969         edge = get_block_succ_next(block, edge);
 970         return edge ? 2 : 1;
 971 }
 972
 973 static int
 974 is_start_block(const ir_node * bb)
 975 {
 976         return get_irg_start_block(get_irn_irg(bb)) == bb;
 977 }
 978
 979 static int
 980 is_before_frame(const ir_node * bb, const ir_node * irn)
 981 {
 982         const ir_node  *frame  = get_irg_frame(get_irn_irg(bb));
 983
 984         if(is_start_block(bb) && sched_get_time_step(frame) >= sched_get_time_step(irn))
 985                 return 1;
 986         else
 987                 return 0;
 988 }
 989
 990 static int
 991 is_merge_edge(const ir_node * bb)
 992 {
 993         if(is_start_block(bb))
 994                 return 0;
 995
 996         if(opt_goodwin)
 997                 return get_block_n_succs(bb) == 1;
 998         else
 999                 return 1;
1000 }
1001
1002 static int
1003 is_diverge_edge(const ir_node * bb)
1004 {
1005         if(is_start_block(bb))
1006                 return 0;
1007
1008         if(opt_goodwin)
1009                 return get_Block_n_cfgpreds(bb) == 1;
1010         else
1011                 return 1;
1012 }
1013
1014 static void
1015 walker_regclass_copy_insertor(ir_node * irn, void * data)
1016 {
1017         spill_ilp_t    *si = data;
1018
1019         if(is_Phi(irn) && has_reg_class(si, irn)) {
1020                 int n;
1021
1022                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1023                         ir_node  *phi_arg = get_irn_n(irn, n);
1024                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1025
1026                         if(!has_reg_class(si, phi_arg)) {
1027                                 ir_node   *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1028                                 ir_node   *pos = sched_block_last_noncf(si, bb);
1029                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1030
1031                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1032                                 sched_add_after(pos, copy);
1033                                 set_irn_n(irn, n, copy);
1034
1035                                 op->is_remat = 0;
1036                                 op->attr.live_range.args.reloads = NULL;
1037                                 op->attr.live_range.ilp = ILP_UNDEF;
1038                                 set_irn_link(copy, op);
1039                         }
1040                 }
1041         }
1042 }
1043
1044
1045 /**
1046  * Insert (so far unused) remats into the irg to
1047  * recompute the potential liveness of all values
1048  */
1049 static void
1050 walker_remat_insertor(ir_node * bb, void * data)
1051 {
1052         spill_ilp_t    *si = data;
1053         spill_bb_t     *spill_bb;
1054         ir_node        *irn;
1055         int             n, i;
1056         pset           *live = pset_new_ptr_default();
1057
1058         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1059
1060         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1061                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1062
1063                 /* add remats at end of block */
1064                 if (has_reg_class(si, value)) {
1065                         pset_insert_ptr(live, value);
1066                 }
1067         }
1068
1069         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1070         set_irn_link(bb, spill_bb);
1071
1072         irn = sched_last(bb);
1073         while(!sched_is_end(irn)) {
1074                 ir_node   *next;
1075                 op_t      *op;
1076                 pset      *args;
1077                 ir_node   *arg;
1078                 pset      *remat_args;
1079
1080                 next = sched_prev(irn);
1081
1082                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1083
1084                 if(is_Phi(irn) || is_Proj(irn)) {
1085                         op_t      *op;
1086
1087                         if(has_reg_class(si, irn)) {
1088                                 pset_remove_ptr(live, irn);
1089                         }
1090
1091                         op = obstack_alloc(si->obst, sizeof(*op));
1092                         op->is_remat = 0;
1093                         op->attr.live_range.args.reloads = NULL;
1094                         op->attr.live_range.ilp = ILP_UNDEF;
1095                         set_irn_link(irn, op);
1096
1097                         irn = next;
1098                         continue;
1099                 }
1100
1101                 op = obstack_alloc(si->obst, sizeof(*op));
1102                 op->is_remat = 0;
1103                 op->attr.live_range.ilp = ILP_UNDEF;
1104                 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1105                 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1106                 set_irn_link(irn, op);
1107
1108                 args = pset_new_ptr_default();
1109
1110                 /* collect arguments of op */
1111                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1112                         ir_node        *arg = get_irn_n(irn, n);
1113
1114                         pset_insert_ptr(args, arg);
1115                 }
1116
1117                 /* set args of op already live in epilog */
1118                 pset_foreach(args, arg) {
1119                         if(has_reg_class(si, arg)) {
1120                                 pset_insert_ptr(live, arg);
1121                         }
1122                 }
1123                 /* delete defined value from live set */
1124                 if(has_reg_class(si, irn)) {
1125                         pset_remove_ptr(live, irn);
1126                 }
1127
1128
1129                 remat_args = pset_new_ptr_default();
1130
1131                 /* insert all possible remats before irn */
1132                 pset_foreach(args, arg) {
1133                         remat_info_t   *remat_info,
1134                                                     query;
1135                         remat_t        *remat;
1136
1137                         /* continue if the operand has the wrong reg class
1138                          */
1139                         if(!has_reg_class(si, arg))
1140                                 continue;
1141
1142                         query.irn = arg;
1143                         query.remats = NULL;
1144                         query.remats_by_operand = NULL;
1145                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1146
1147                         if(!remat_info) {
1148                                 continue;
1149                         }
1150
1151                         if(remat_info->remats) {
1152                                 pset_foreach(remat_info->remats, remat) {
1153                                         ir_node  *remat_irn = NULL;
1154
1155                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1156                                         if(opt_remat_while_live) {
1157                                                 if(pset_find_ptr(live, remat->value)) {
1158                                                         remat_irn = insert_remat_before(si, remat, irn, live);
1159                                                 }
1160                                         } else {
1161                                                 remat_irn = insert_remat_before(si, remat, irn, live);
1162                                         }
1163                                         if(remat_irn) {
1164                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1165                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1166
1167                                                         if(!has_reg_class(si, remat_arg)) continue;
1168
1169                                                         pset_insert_ptr(remat_args, remat_arg);
1170                                                 }
1171                                         }
1172                                 }
1173                         }
1174                 }
1175
1176                 /* now we add remat args to op's args because they could also die at this op */
1177                 pset_foreach(args,arg) {
1178                         if(pset_find_ptr(remat_args, arg)) {
1179                                 pset_remove_ptr(remat_args, arg);
1180                         }
1181                 }
1182                 pset_foreach(remat_args,arg) {
1183                         pset_insert_ptr(args, arg);
1184                 }
1185
1186                 /* insert all possible remats after irn */
1187                 pset_foreach(args, arg) {
1188                         remat_info_t   *remat_info,
1189                                                     query;
1190                         remat_t        *remat;
1191
1192                         /* continue if the operand has the wrong reg class */
1193                         if(!has_reg_class(si, arg))
1194                                 continue;
1195
1196                         query.irn = arg;
1197                         query.remats = NULL;
1198                         query.remats_by_operand = NULL;
1199                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1200
1201                         if(!remat_info) {
1202                                 continue;
1203                         }
1204
1205                         /* do not place post remats after jumps */
1206                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1207
1208                         if(remat_info->remats_by_operand) {
1209                                 pset_foreach(remat_info->remats_by_operand, remat) {
1210                                         /* do not insert remats producing the same value as one of the operands */
1211                                         if(!pset_find_ptr(args, remat->value)) {
1212                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1213                                                 if(opt_remat_while_live) {
1214                                                         if(pset_find_ptr(live, remat->value)) {
1215                                                                 insert_remat_after(si, remat, irn, live);
1216                                                         }
1217                                                 } else {
1218                                                         insert_remat_after(si, remat, irn, live);
1219                                                 }
1220                                         }
1221                                 }
1222                         }
1223                 }
1224
1225                 del_pset(remat_args);
1226                 del_pset(args);
1227                 irn = next;
1228         }
1229
1230         /* add remats at end if successor has multiple predecessors */
1231         if(is_merge_edge(bb)) {
1232                 pset     *live_out = pset_new_ptr_default();
1233                 ir_node  *value;
1234
1235                 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1236                         value = be_lv_get_irn(si->lv, bb, i);
1237
1238                         if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1239                                 pset_insert_ptr(live_out, value);
1240                         }
1241                 }
1242
1243                 /* add remats at end of block */
1244                 pset_foreach(live_out, value) {
1245                         remat_info_t   *remat_info,
1246                                                    query;
1247                         remat_t        *remat;
1248
1249                         query.irn = value;
1250                         query.remats = NULL;
1251                         query.remats_by_operand = NULL;
1252                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1253
1254                         if(remat_info && remat_info->remats) {
1255                                 pset_foreach(remat_info->remats, remat) {
1256                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1257
1258                                         insert_remat_before(si, remat, bb, live_out);
1259                                 }
1260                         }
1261                 }
1262                 del_pset(live_out);
1263         }
1264
1265         if(is_diverge_edge(bb)) {
1266                 pset     *live_in = pset_new_ptr_default();
1267                 ir_node  *value;
1268
1269                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1270                         value = be_lv_get_irn(si->lv, bb, i);
1271
1272                         if (has_reg_class(si, value)) {
1273                                 pset_insert_ptr(live_in, value);
1274                         }
1275                 }
1276
1277                 /* add remat2s at beginning of block */
1278                 pset_foreach(live_in, value) {
1279                         if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1280                                 remat_info_t   *remat_info,
1281                                                            query;
1282                                 remat_t        *remat;
1283
1284                                 query.irn = value;
1285                                 query.remats = NULL;
1286                                 query.remats_by_operand = NULL;
1287                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1288
1289                                 if(remat_info && remat_info->remats_by_operand) {
1290                                         pset_foreach(remat_info->remats_by_operand, remat) {
1291                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1292
1293                                                 /* put the remat here if all its args are available */
1294                                                 insert_remat_after(si, remat, bb, live_in);
1295
1296                                         }
1297                                 }
1298                         }
1299                 }
1300                 del_pset(live_in);
1301         }
1302 }
1303
1304 /**
1305  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1306  */
1307 static void
1308 luke_endwalker(ir_node * bb, void * data)
1309 {
1310         spill_ilp_t    *si = (spill_ilp_t*)data;
1311         pset           *live;
1312         pset           *use_end;
1313         char            buf[256];
1314         ilp_cst_t       cst;
1315         ir_node        *irn;
1316         spill_bb_t     *spill_bb = get_irn_link(bb);
1317         int             i;
1318
1319
1320         live = pset_new_ptr_default();
1321         use_end = pset_new_ptr_default();
1322
1323         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1324                 irn = be_lv_get_irn(si->lv, bb, i);
1325                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1326                         op_t      *op;
1327
1328                         pset_insert_ptr(live, irn);
1329                         op = get_irn_link(irn);
1330                         assert(!op->is_remat);
1331                 }
1332         }
1333
1334         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1335         /* their reg_out must always be set */
1336         sched_foreach_reverse(bb, irn) {
1337                 int   n;
1338
1339                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1340
1341                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1342                         ir_node        *irn_arg = get_irn_n(irn, n);
1343
1344                         if(has_reg_class(si, irn_arg)) {
1345                                 pset_insert_ptr(use_end, irn_arg);
1346                         }
1347                 }
1348         }
1349
1350         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1351         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1352         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1353
1354         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1355
1356         /* if this is a merge edge we can reload at the end of this block */
1357         if(is_merge_edge(bb)) {
1358                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1359         } else if(pset_count(use_end)){
1360                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1361         } else {
1362                 spill_bb->reloads = NULL;
1363         }
1364
1365         pset_foreach(live,irn) {
1366                 spill_t     query,
1367                                         *spill;
1368                 double      spill_cost;
1369                 int         default_spilled;
1370
1371
1372                 /* handle values used by control flow nodes later separately */
1373                 if(pset_find_ptr(use_end, irn)) continue;
1374
1375                 query.irn = irn;
1376                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1377
1378                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1379
1380                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1381                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1382                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1383
1384                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1385                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1386
1387                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1388                 /* by default spill value right after definition */
1389                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1390                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1391
1392                 if(is_merge_edge(bb)) {
1393                         ilp_var_t   reload;
1394                         ilp_cst_t   rel_cst;
1395
1396                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1397                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1398                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1399
1400                         /* reload <= mem_out */
1401                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1402                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1403                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1404                 }
1405
1406                 spill->reg_in = ILP_UNDEF;
1407                 spill->mem_in = ILP_UNDEF;
1408         }
1409
1410         pset_foreach(use_end,irn) {
1411                 spill_t     query,
1412                                         *spill;
1413                 double      spill_cost;
1414                 ilp_cst_t   end_use_req,
1415                                         rel_cst;
1416                 ilp_var_t   reload;
1417                 int         default_spilled;
1418
1419                 query.irn = irn;
1420                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1421
1422                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1423
1424                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1425                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1426                 /* if irn is used at the end of the block, then it is live anyway */
1427                 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1428
1429                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1430                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1431
1432                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1433                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1434                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1435
1436                 /* reload for use be control flow op */
1437                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1438                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1439                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1440
1441                 /* reload <= mem_out */
1442                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1443                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1444                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1445
1446                 spill->reg_in = ILP_UNDEF;
1447                 spill->mem_in = ILP_UNDEF;
1448
1449                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1450                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1451                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1452         }
1453
1454         del_pset(live);
1455         del_pset(use_end);
1456 }
1457
1458 static ir_node *
1459 next_post_remat(const ir_node * irn)
1460 {
1461         op_t      *op;
1462     ir_node   *next;
1463
1464         if(is_Block(irn)) {
1465                 next = sched_block_first_nonphi(irn);
1466         } else {
1467                 next = sched_next_op(irn);
1468         }
1469
1470         if(sched_is_end(next))
1471                 return NULL;
1472
1473         op = get_irn_link(next);
1474         if(op->is_remat && !op->attr.remat.pre) {
1475                 return next;
1476         }
1477
1478         return NULL;
1479 }
1480
1481
1482 static ir_node *
1483 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1484 {
1485         op_t      *op;
1486         ir_node   *ret;
1487
1488         if(is_Block(irn)) {
1489                 ret = sched_block_last_noncf(si, irn);
1490                 ret = sched_next(ret);
1491                 ret = sched_prev_op(ret);
1492         } else {
1493                 ret = sched_prev_op(irn);
1494         }
1495
1496         if(sched_is_end(ret) || is_Phi(ret))
1497                 return NULL;
1498
1499         op = (op_t*)get_irn_link(ret);
1500         if(op->is_remat && op->attr.remat.pre) {
1501                 return ret;
1502         }
1503
1504         return NULL;
1505 }
1506
1507 /**
1508  * Find a remat of value @p value in the epilog of @p pos
1509  */
1510 static ir_node *
1511 find_post_remat(const ir_node * value, const ir_node * pos)
1512 {
1513         while((pos = next_post_remat(pos)) != NULL) {
1514                 op_t   *op;
1515
1516                 op = get_irn_link(pos);
1517                 assert(op->is_remat && !op->attr.remat.pre);
1518
1519                 if(op->attr.remat.remat->value == value)
1520                         return (ir_node*)pos;
1521
1522 #if 0
1523         const ir_edge_t *edge;
1524                 foreach_out_edge(pos, edge) {
1525                         ir_node   *proj = get_edge_src_irn(edge);
1526                         assert(is_Proj(proj));
1527                 }
1528 #endif
1529
1530         }
1531
1532         return NULL;
1533 }
1534
1535 static spill_t *
1536 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1537 {
1538         spill_bb_t  *spill_bb = get_irn_link(bb);
1539         spill_t     *spill,
1540                                  query;
1541         char         buf[256];
1542         int          default_spilled;
1543
1544         query.irn = irn;
1545         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1546         if(!spill) {
1547                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1548
1549                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1550
1551                 spill->reg_out = ILP_UNDEF;
1552                 spill->reg_in  = ILP_UNDEF;
1553                 spill->mem_in  = ILP_UNDEF;
1554
1555                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1556                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1557
1558                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1559                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1560                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1561         }
1562
1563         return spill;
1564 }
1565
1566 static void
1567 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1568 {
1569         ir_node        *irn;
1570         int i;
1571
1572         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1573                 irn = be_lv_get_irn(si->lv, bb, i);
1574
1575                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1576                         pset_insert_ptr(live, irn);
1577                 }
1578         }
1579
1580         irn = sched_last(bb);
1581
1582         /* all values eaten by control flow operations are also live until the end of the block */
1583         sched_foreach_reverse(bb, irn) {
1584                 int  i;
1585
1586                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1587
1588                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1589                         ir_node *arg = get_irn_n(irn,i);
1590
1591                         if(has_reg_class(si, arg)) {
1592                                 pset_insert_ptr(live, arg);
1593                         }
1594                 }
1595         }
1596 }
1597
1598 /**
1599  *  Inserts ILP-constraints and variables for memory copying before the given position
1600  */
1601 static void
1602 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1603 {
1604         const ir_node    *succ;
1605         const ir_edge_t  *edge;
1606         spill_bb_t       *spill_bb = get_irn_link(block);
1607         ir_node          *phi;
1608         int               pos;
1609         ilp_cst_t         cst;
1610         ilp_var_t         copyreg;
1611         char              buf[256];
1612         ir_node          *tmp;
1613
1614
1615         assert(edges_activated(current_ir_graph));
1616
1617         edge = get_block_succ_first(block);
1618         if(!edge) return;
1619
1620         succ = edge->src;
1621         pos = edge->pos;
1622
1623         edge = get_block_succ_next(block, edge);
1624         /* next block can only contain phis, if this is a merge edge */
1625         if(edge) return;
1626
1627         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1628         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1629
1630         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1631         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1632
1633         pset_foreach(live, tmp) {
1634                 spill_t  *spill;
1635 #if 0
1636                 op_t  *op = get_irn_link(irn);
1637                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1638 #endif
1639                 spill = set_find_spill(spill_bb->ilp, tmp);
1640                 assert(spill);
1641
1642                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1643         }
1644         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1645
1646         sched_foreach(succ, phi) {
1647                 const ir_node  *to_copy;
1648                 op_t           *to_copy_op;
1649                 spill_t        *to_copy_spill;
1650                 op_t           *phi_op = get_irn_link(phi);
1651                 ilp_var_t       reload = ILP_UNDEF;
1652
1653
1654                 if(!is_Phi(phi)) break;
1655                 if(!has_reg_class(si, phi)) continue;
1656
1657                 to_copy = get_irn_n(phi, pos);
1658
1659                 to_copy_op = get_irn_link(to_copy);
1660
1661                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1662                 assert(to_copy_spill);
1663
1664                 if(spill_bb->reloads) {
1665                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1666
1667                         if(keyval) {
1668                                 reload = PTR_TO_INT(keyval->val);
1669                         }
1670                 }
1671
1672                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1673                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1674
1675                 /* copy - reg_out - reload - remat - live_range <= 0 */
1676                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1677                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1678                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1679                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1680                 foreach_pre_remat(si, block, tmp) {
1681                         op_t     *remat_op = get_irn_link(tmp);
1682                         if(remat_op->attr.remat.remat->value == to_copy) {
1683                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1684                         }
1685                 }
1686
1687                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1688                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1689
1690                 /* copy - reg_out - copyreg <= 0 */
1691                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1692                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1693                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1694         }
1695 }
1696
1697
1698 /**
1699  * Walk all irg blocks and emit this ILP
1700  */
1701 static void
1702 luke_blockwalker(ir_node * bb, void * data)
1703 {
1704         spill_ilp_t    *si = (spill_ilp_t*)data;
1705         ir_node        *irn;
1706         pset           *live;
1707         char            buf[256];
1708         ilp_cst_t       cst;
1709         spill_bb_t     *spill_bb = get_irn_link(bb);
1710         ir_node        *tmp;
1711         spill_t        *spill;
1712         pset           *defs = pset_new_ptr_default();
1713         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1714
1715
1716         live = pset_new_ptr_default();
1717
1718         /****************************************
1719          *      B A S I C  B L O C K  E N D
1720          ***************************************/
1721
1722
1723         /* init live values at end of block */
1724         get_live_end(si, bb, live);
1725
1726         pset_foreach(live, irn) {
1727                 op_t           *op;
1728                 ilp_var_t       reload = ILP_UNDEF;
1729
1730                 spill = set_find_spill(spill_bb->ilp, irn);
1731                 assert(spill);
1732
1733                 if(spill_bb->reloads) {
1734                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1735
1736                         if(keyval) {
1737                                 reload = PTR_TO_INT(keyval->val);
1738                         }
1739                 }
1740
1741                 op = get_irn_link(irn);
1742                 assert(!op->is_remat);
1743
1744                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1745                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1746                 op->attr.live_range.op = bb;
1747
1748                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1749                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1750
1751                 /* reg_out - reload - remat - live_range <= 0 */
1752                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1753                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1754                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1755                 foreach_pre_remat(si, bb, tmp) {
1756                         op_t     *remat_op = get_irn_link(tmp);
1757                         if(remat_op->attr.remat.remat->value == irn) {
1758                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1759                         }
1760                 }
1761                 /* maybe we should also assure that reg_out >= live_range etc. */
1762         }
1763
1764         if(opt_memcopies)
1765                 insert_mem_copy_position(si, live, bb);
1766
1767         /*
1768          * start new live ranges for values used by remats at end of block
1769          * and assure the remat args are available
1770          */
1771         foreach_pre_remat(si, bb, tmp) {
1772                 op_t     *remat_op = get_irn_link(tmp);
1773                 int       n;
1774
1775                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1776                         ir_node        *remat_arg = get_irn_n(tmp, n);
1777                         op_t           *arg_op = get_irn_link(remat_arg);
1778                         ilp_var_t       prev_lr;
1779
1780                         if(!has_reg_class(si, remat_arg)) continue;
1781
1782                         /* if value is becoming live through use by remat */
1783                         if(!pset_find_ptr(live, remat_arg)) {
1784                                 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1785                                 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1786
1787                                 arg_op->attr.live_range.ilp = prev_lr;
1788                                 arg_op->attr.live_range.op = bb;
1789
1790                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1791
1792                                 pset_insert_ptr(live, remat_arg);
1793                                 add_to_spill_bb(si, bb, remat_arg);
1794                         }
1795
1796                         /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1797                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1798                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1799
1800                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1801                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1802
1803                         /* use reload placed for this argument */
1804                         if(spill_bb->reloads) {
1805                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1806
1807                                 if(keyval) {
1808                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1809
1810                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1811                                 }
1812                         }
1813                 }
1814         }
1815         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1816
1817
1818
1819
1820         /**************************************
1821          *    B A S I C  B L O C K  B O D Y
1822          **************************************/
1823
1824         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1825                 op_t       *op;
1826                 op_t       *tmp_op;
1827                 int         n,
1828                                         u = 0,
1829                                         d = 0;
1830                 ilp_cst_t       check_pre,
1831                                         check_post;
1832                 set        *args;
1833                 pset       *used;
1834                 pset       *remat_defs;
1835                 keyval_t   *keyval;
1836                 ilp_cst_t   one_memoperand;
1837
1838                 /* iterate only until first phi */
1839                 if(is_Phi(irn))
1840                         break;
1841
1842                 op = get_irn_link(irn);
1843                 /* skip remats */
1844                 if(op->is_remat) continue;
1845                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1846
1847                 /* collect defined values */
1848                 if(has_reg_class(si, irn)) {
1849                         pset_insert_ptr(defs, irn);
1850                 }
1851
1852                 /* skip projs */
1853                 if(is_Proj(irn)) continue;
1854
1855                 /*
1856                  * init set of irn's arguments
1857                  * and all possibly used values around this op
1858                  * and values defined by post remats
1859                  */
1860                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1861                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1862                 remat_defs = pset_new_ptr(pset_count(live));
1863
1864                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1865                         ir_node        *irn_arg = get_irn_n(irn, n);
1866                         if(has_reg_class(si, irn_arg)) {
1867                                 set_insert_keyval(args, irn_arg, (void*)n);
1868                                 pset_insert_ptr(used, irn_arg);
1869                         }
1870                 }
1871                 foreach_post_remat(irn, tmp) {
1872                         op_t    *remat_op = get_irn_link(tmp);
1873
1874                         pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1875
1876                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1877                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1878                                 if(has_reg_class(si, remat_arg)) {
1879                                         pset_insert_ptr(used, remat_arg);
1880                                 }
1881                         }
1882                 }
1883                 foreach_pre_remat(si, irn, tmp) {
1884                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1885                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1886                                 if(has_reg_class(si, remat_arg)) {
1887                                         pset_insert_ptr(used, remat_arg);
1888                                 }
1889                         }
1890                 }
1891
1892                 /**********************************
1893                  *   I N  E P I L O G  O F  irn
1894                  **********************************/
1895
1896                 /* ensure each dying value is used by only one post remat */
1897                 pset_foreach(used, tmp) {
1898                         ir_node     *value = tmp;
1899                         op_t        *value_op = get_irn_link(value);
1900                         ir_node     *remat;
1901                         int          n_remats = 0;
1902
1903                         cst = ILP_UNDEF;
1904                         foreach_post_remat(irn, remat) {
1905                                 op_t  *remat_op = get_irn_link(remat);
1906
1907                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1908                                         ir_node   *remat_arg = get_irn_n(remat, n);
1909
1910                                         /* if value is used by this remat add it to constraint */
1911                                         if(remat_arg == value) {
1912                                                 if(n_remats == 0) {
1913                                                         /* sum remat2s <= 1 + n_remats*live_range */
1914                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1915                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1916                                                 }
1917
1918                                                 n_remats++;
1919                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1920                                                 break;
1921                                         }
1922                                 }
1923                         }
1924
1925             // value_op->attr.live_range.ilp != ILP_UNDEF
1926                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1927                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1928                         }
1929                 }
1930
1931         /* ensure at least one value dies at post remat */
1932         foreach_post_remat(irn, tmp) {
1933             op_t     *remat_op = get_irn_link(tmp);
1934             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
1935             ir_node  *remat_arg;
1936
1937             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1938                 remat_arg = get_irn_n(tmp, n);
1939
1940                 if(has_reg_class(si, remat_arg)) {
1941
1942                     /* does arg always die at this op? */
1943                     if(!pset_find_ptr(live, remat_arg))
1944                         goto skip_one_must_die;
1945
1946                     pset_insert_ptr(remat_args, remat_arg);
1947                 }
1948             }
1949
1950             /* remat + \sum live_range(remat_arg) <= |args| */
1951             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1952             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1953             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1954
1955             pset_foreach(remat_args, remat_arg) {
1956                 op_t  *arg_op = get_irn_link(remat_arg);
1957
1958                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1959             }
1960
1961 skip_one_must_die:
1962             del_pset(remat_args);
1963         }
1964
1965                 /* new live ranges for values from L\U defined by post remats */
1966                 pset_foreach(live, tmp) {
1967                         ir_node     *value = tmp;
1968                         op_t        *value_op = get_irn_link(value);
1969
1970                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1971                                 ilp_var_t    prev_lr = ILP_UNDEF;
1972                                 ir_node     *remat;
1973
1974                                 if(pset_find_ptr(remat_defs, value)) {
1975
1976                                         /* next_live_range <= prev_live_range + sum remat2s */
1977                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1978                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1979
1980                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1981                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1982
1983                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1984                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1985
1986                                         foreach_post_remat(irn, remat) {
1987                                                 op_t        *remat_op = get_irn_link(remat);
1988
1989                                                 /* if value is being rematerialized by this remat */
1990                                                 if(value == remat_op->attr.remat.remat->value) {
1991                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1992                                                 }
1993                                         }
1994
1995                                         value_op->attr.live_range.ilp = prev_lr;
1996                                         value_op->attr.live_range.op = irn;
1997                                 }
1998                         }
1999                 }
2000
2001                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2002                 foreach_post_remat(irn, tmp) {
2003                         op_t        *remat_op = get_irn_link(tmp);
2004                         int          n;
2005
2006                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2007                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2008                                 op_t           *arg_op = get_irn_link(remat_arg);
2009
2010                                 if(!has_reg_class(si, remat_arg)) continue;
2011
2012                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2013                                 if(!pset_find_ptr(used, remat_arg)) {
2014                                         /* remat <= live_range(remat_arg) */
2015                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2016                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2017
2018                                         /* if value is becoming live through use by remat2 */
2019                                         if(!pset_find_ptr(live, remat_arg)) {
2020                                                 ilp_var_t     lr;
2021
2022                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2023                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2024
2025                                                 arg_op->attr.live_range.ilp = lr;
2026                                                 arg_op->attr.live_range.op = irn;
2027
2028                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2029
2030                                                 pset_insert_ptr(live, remat_arg);
2031                                                 add_to_spill_bb(si, bb, remat_arg);
2032                                         }
2033
2034                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2035                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2036                                 }
2037                         }
2038                 }
2039
2040                 d = pset_count(defs);
2041                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2042
2043                 /* count how many regs irn needs for arguments */
2044                 u = set_count(args);
2045
2046
2047                 /* check the register pressure in the epilog */
2048                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2049                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2050                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2051
2052                 /* add L\U' to check_post */
2053                 pset_foreach(live, tmp) {
2054                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2055                                 /* if a live value is not used by irn */
2056                                 tmp_op = get_irn_link(tmp);
2057                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2058                         }
2059                 }
2060
2061                 /***********************************************************
2062                  *  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
2063                  **********************************************************/
2064
2065
2066                 pset_foreach(used, tmp) {
2067                         ilp_var_t       prev_lr;
2068                         ilp_var_t       post_use;
2069                         int             p = 0;
2070                         spill_t        *spill;
2071                         ir_node        *arg = tmp;
2072                         op_t           *arg_op = get_irn_link(arg);
2073                         ir_node        *remat;
2074
2075                         spill = add_to_spill_bb(si, bb, arg);
2076
2077                         /* new live range for each used value */
2078                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2079                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0);
2080
2081                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2082                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2083                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2084
2085                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2086
2087                         /* arg is live throughout epilog if the next live_range is in a register */
2088                         if(pset_find_ptr(live, arg)) {
2089                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2090
2091                                 /* post_use >= next_lr + remat */
2092                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2093                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2094                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2095                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2096
2097                         }
2098
2099                         /* if value is not an arg of op and not possibly defined by post remat
2100                          * then it may only die and not become live
2101                          */
2102                         if(!set_find_keyval(args, arg)) {
2103                                 /* post_use <= prev_lr */
2104                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2105                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2106                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2107                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2108
2109                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2110                                         /* next_lr <= prev_lr */
2111                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2112                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2113                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2114                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2115                                 }
2116                         }
2117
2118
2119                         /* forall post remat which use arg add a similar cst */
2120                         foreach_post_remat(irn, remat) {
2121                                 int      n;
2122
2123                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2124                                         ir_node    *remat_arg = get_irn_n(remat, n);
2125                                         op_t       *remat_op = get_irn_link(remat);
2126
2127                                         if(remat_arg == arg) {
2128                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2129
2130                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2131                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2132                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2133                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2134                                         }
2135                                 }
2136                         }
2137
2138                         if(opt_memoperands) {
2139                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2140                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2141                                                 ilp_var_t       memoperand;
2142
2143                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2144                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2145                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2146
2147                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2148                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2149
2150                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2151                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2152 //                                              if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2153 //                                                      lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2154                                         }
2155                                 }
2156                         }
2157
2158                         /* new live range begins for each used value */
2159                         arg_op->attr.live_range.ilp = prev_lr;
2160                         arg_op->attr.live_range.op = irn;
2161
2162                         /*if(!pset_find_ptr(live, arg)) {
2163                                 pset_insert_ptr(live, arg);
2164                                 add_to_spill_bb(si, bb, arg);
2165                         }*/
2166                         pset_insert_ptr(live, arg);
2167
2168                 }
2169
2170                 /* just to be sure */
2171                 check_post = ILP_UNDEF;
2172
2173
2174
2175
2176                 /******************
2177                  *   P R O L O G
2178                  ******************/
2179
2180                 /* check the register pressure in the prolog */
2181                 /* sum_{L\U} lr <= k - |U| */
2182                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2183                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2184
2185                 /* for the prolog remove defined values from the live set */
2186                 pset_foreach(defs, tmp) {
2187                         pset_remove_ptr(live, tmp);
2188                 }
2189
2190                 if(opt_memoperands) {
2191                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2192                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2193                 }
2194
2195                 /***********************************************************
2196                  *  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
2197                  **********************************************************/
2198
2199
2200                 set_foreach(args, keyval) {
2201                         spill_t          *spill;
2202                         const ir_node    *arg = keyval->key;
2203                         int               i = PTR_TO_INT(keyval->val);
2204                         op_t             *arg_op = get_irn_link(arg);
2205                         ilp_cst_t         requirements;
2206                         int               n_memoperands;
2207
2208                         spill = set_find_spill(spill_bb->ilp, arg);
2209                         assert(spill);
2210
2211                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2212                         op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), is_before_frame(bb, irn)?0.0:1.0);
2213
2214                         /* reload <= mem_out */
2215                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2216                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2217                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2218                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2219
2220                         /* requirement: arg must be in register for use */
2221                         /* reload + remat + live_range == 1 */
2222                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2223                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2224
2225                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2226                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2227                         foreach_pre_remat(si, irn, tmp) {
2228                                 op_t     *remat_op = get_irn_link(tmp);
2229                                 if(remat_op->attr.remat.remat->value == arg) {
2230                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2231                                 }
2232                         }
2233
2234                         if(opt_memoperands) {
2235                                 n_memoperands = 0;
2236                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2237                                         if(get_irn_n(irn, n) == arg) {
2238                                                 n_memoperands++;
2239                                         }
2240                                 }
2241                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2242                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2243                                                 memoperand_t  *memoperand;
2244                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2245
2246                                                 /* memoperand <= mem_out */
2247                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2248                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2249                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2250                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2251
2252                                                 /* the memoperand is only sufficient if it is used once by the op */
2253                                                 if(n_memoperands == 1)
2254                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2255
2256                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2257
2258                                                 /* we have one more free register if we use a memory operand */
2259                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2260                                         }
2261                                 }
2262                         }
2263                 }
2264
2265                 /* iterate over L\U */
2266                 pset_foreach(live, tmp) {
2267                         if(!set_find_keyval(args, tmp)) {
2268                                 /* if a live value is not used by irn */
2269                                 tmp_op = get_irn_link(tmp);
2270                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2271                         }
2272                 }
2273
2274
2275                 /* requirements for remats */
2276                 /* start new live ranges for values used by remats */
2277                 foreach_pre_remat(si, irn, tmp) {
2278                         op_t        *remat_op = get_irn_link(tmp);
2279                         int          n;
2280
2281                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2282                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2283                                 op_t           *arg_op = get_irn_link(remat_arg);
2284
2285                                 if(!has_reg_class(si, remat_arg)) continue;
2286
2287                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2288                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2289                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2290
2291                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2292                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2293
2294                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2295                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2296                                         int    index = (int)keyval->val;
2297
2298                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2299                                 }
2300                         }
2301                 }
2302
2303
2304
2305
2306                 /*************************
2307                  *  D O N E  W I T H  O P
2308                  *************************/
2309
2310                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2311
2312                 pset_foreach(live, tmp) {
2313                         assert(has_reg_class(si, tmp));
2314                 }
2315
2316                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2317                         ir_node        *arg = get_irn_n(irn, n);
2318
2319                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2320                 }
2321
2322                 del_pset(remat_defs);
2323                 del_pset(used);
2324                 del_set(args);
2325                 del_pset(defs);
2326                 defs = pset_new_ptr_default();
2327         }
2328
2329
2330
2331         /***************************************
2332          *   B E G I N N I N G  O F  B L O C K
2333          ***************************************/
2334
2335
2336         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2337         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2338
2339         pset_foreach(live, irn) {
2340                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2341         }
2342
2343         /* construct mem_outs for all values */
2344
2345         set_foreach(spill_bb->ilp, spill) {
2346                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2347                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2348
2349                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2350                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2351
2352                 if(pset_find_ptr(live, spill->irn)) {
2353                         int default_spilled;
2354                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2355
2356                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2357                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2358                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2359                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2360
2361                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2362                                 int   n;
2363                                 op_t *op = get_irn_link(spill->irn);
2364
2365                                 /* do we have to copy a phi argument? */
2366                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2367                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2368
2369                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2370                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2371                                         double          freq=0.0;
2372                                         int             m;
2373                                         ilp_var_t       var;
2374
2375
2376                                         /* argument already done? */
2377                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2378
2379                                         /* get sum of execution frequencies of blocks with the same phi argument */
2380                                         for(m=n; m>=0; --m) {
2381                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2382
2383                                                 if(arg==arg2) {
2384                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2385                                                 }
2386                                         }
2387
2388                                         /* copies are not for free */
2389                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2390                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2391
2392                                         for(m=n; m>=0; --m) {
2393                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2394
2395                                                 if(arg==arg2) {
2396                                                         op->attr.live_range.args.copies[m] = var;
2397                                                 }
2398                                         }
2399
2400 #if 0
2401                                         /* copy <= mem_in */
2402                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2403                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2404                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2405                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2406 #endif
2407                                 }
2408                         }
2409                 }
2410         }
2411
2412         foreach_post_remat(bb, tmp) {
2413                 int         n;
2414
2415                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2416                         ir_node    *remat_arg = get_irn_n(tmp, n);
2417
2418                         if(!has_reg_class(si, remat_arg)) continue;
2419
2420                         /* if value is becoming live through use by remat2 */
2421                         if(!pset_find_ptr(live, remat_arg)) {
2422                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2423                                 ilp_cst_t   nomem;
2424
2425                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2426
2427                                 pset_insert_ptr(live, remat_arg);
2428                                 spill = add_to_spill_bb(si, bb, remat_arg);
2429                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2430
2431                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2432                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2433                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2434                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2435                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2436
2437
2438                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2439                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2440                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2441
2442                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2443                                 if(spill_bb->reloads) {
2444                                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
2445
2446                                         if(keyval) {
2447                                                 ilp_var_t reload = PTR_TO_INT(keyval->val);
2448                                                 lpp_set_factor_fast(si->lpp, nomem, reload, 1.0);
2449                                         }
2450                                 }
2451                         }
2452                 }
2453         }
2454
2455         /* L\U is empty at bb start */
2456         /* arg is live throughout epilog if it is reg_in into this block */
2457
2458         /* check the register pressure at the beginning of the block
2459          * including remats
2460          */
2461         /* reg_in entspricht post_use */
2462
2463         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2464         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2465
2466         pset_foreach(live, irn) {
2467         ilp_cst_t  nospill;
2468
2469                 spill = set_find_spill(spill_bb->ilp, irn);
2470                 assert(spill);
2471
2472                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2473                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2474
2475                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2476
2477                 /* spill + mem_in <= 1 */
2478                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2479                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2480
2481                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2482                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2483
2484         } /* post_remats are NOT included in register pressure check because
2485            they do not increase regpressure */
2486
2487         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2488         pset_foreach(live, irn) {
2489                 int          p = 0,
2490                                          n;
2491
2492                 spill = set_find_spill(spill_bb->ilp, irn);
2493                 assert(spill && spill->irn == irn);
2494
2495                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2496                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2497                                 ilp_cst_t       mem_in,
2498                                                                 reg_in;
2499                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2500                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2501                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2502                                 spill_t        *spill_p;
2503                                 op_t           *op = get_irn_link(irn);
2504
2505                                 /* although the phi is in the right regclass one or more of
2506                                  * its arguments can be in a different one or at least to
2507                                  * ignore
2508                                  */
2509                                 if(has_reg_class(si, phi_arg)) {
2510                                         /* mem_in < mem_out_arg + copy */
2511                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2512                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2513
2514                                         /* reg_in < reg_out_arg */
2515                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2516                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2517
2518                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2519                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2520
2521                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2522                                         assert(spill_p);
2523
2524                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2525                                         if(opt_memcopies)
2526                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2527
2528                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2529                                 }
2530                         }
2531                 } else {
2532                         /* else assure the value arrives on all paths in the same resource */
2533
2534                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2535                                 ilp_cst_t       mem_in,
2536                                                                 reg_in;
2537                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2538                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2539                                 spill_t        *spill_p;
2540
2541                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2542                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2543                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2544                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2545
2546                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2547                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2548
2549                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2550                                 assert(spill_p);
2551
2552                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2553                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2554                         }
2555                 }
2556         }
2557
2558         foreach_post_remat(bb, tmp) {
2559                 int         n;
2560
2561                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2562                         ir_node    *remat_arg = get_irn_n(tmp, n);
2563                         op_t       *remat_op = get_irn_link(tmp);
2564
2565                         if(!has_reg_class(si, remat_arg)) continue;
2566
2567                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2568                         assert(spill);
2569
2570                         /* remat <= reg_in_argument */
2571                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2572                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2573                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2574                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2575                 }
2576         }
2577
2578         pset_foreach(live, irn) {
2579                 const op_t      *op = get_irn_link(irn);
2580                 const ir_node   *remat;
2581                 int              n_remats = 0;
2582
2583                 if(op->attr.live_range.ilp == ILP_UNDEF) continue;
2584
2585                 cst = ILP_UNDEF;
2586
2587                 foreach_post_remat(bb, remat) {
2588                         int   n;
2589
2590                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2591                                 const ir_node  *arg = get_irn_n(remat, n);
2592
2593                                 if(arg == irn) {
2594                                         const op_t   *remat_op = get_irn_link(remat);
2595
2596                                         if(cst == ILP_UNDEF) {
2597                                                 /* \sum post_remat <= 1 + #post_remats * next(lr) */
2598                                                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N_arg_%N", remat, bb, irn);
2599                                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
2600                                         }
2601                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2602                                         ++n_remats;
2603                                         break;
2604                                 }
2605                         }
2606                 }
2607                 if(n_remats) {
2608                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, n_remats);
2609                 }
2610         }
2611
2612         /* first live ranges from reg_ins */
2613         pset_foreach(live, irn) {
2614                 op_t      *op = get_irn_link(irn);
2615
2616                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2617
2618                         spill = set_find_spill(spill_bb->ilp, irn);
2619                         assert(spill && spill->irn == irn);
2620
2621                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2622                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2623                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2624                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2625
2626                         foreach_post_remat(bb, tmp) {
2627                                 op_t     *remat_op = get_irn_link(tmp);
2628
2629                                 if(remat_op->attr.remat.remat->value == irn) {
2630                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2631                                 }
2632                         }
2633                 }
2634         }
2635
2636         /* walk forward now and compute constraints for placing spills */
2637         /* this must only be done for values that are not defined in this block */
2638         /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2639         pset_foreach(live, irn) {
2640                 /*
2641                  * if value is defined in this block we can anways place the spill directly after the def
2642                  *    -> no constraint necessary
2643                  */
2644                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2645
2646
2647                 spill = set_find_spill(spill_bb->ilp, irn);
2648                 assert(spill);
2649
2650                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2651                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2652
2653                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2654                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2655
2656                 if(!is_Phi(irn)) {
2657                         sched_foreach_op(bb, tmp) {
2658                                 op_t   *op = get_irn_link(tmp);
2659
2660                                 if(is_Phi(tmp)) continue;
2661                                 assert(!is_Proj(tmp));
2662
2663                                 if(op->is_remat) {
2664                                         const ir_node   *value = op->attr.remat.remat->value;
2665
2666                                         if(value == irn) {
2667                                                 /* only collect remats up to the first real use of a value */
2668                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2669                                         }
2670                                 } else {
2671                                         int   n;
2672
2673                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2674                                                 ir_node    *arg = get_irn_n(tmp, n);
2675
2676                                                 if(arg == irn) {
2677                                                         /* if a value is used stop collecting remats */
2678                             goto next_live;
2679                                                 }
2680                                         }
2681                                 }
2682                         }
2683                 }
2684 next_live: ;
2685         }
2686
2687         del_pset(live);
2688 }
2689
2690 typedef struct _irnlist_t {
2691         struct list_head   list;
2692         ir_node           *irn;
2693 } irnlist_t;
2694
2695 typedef struct _interference_t {
2696         struct list_head    blocklist;
2697         ir_node            *a;
2698         ir_node            *b;
2699 } interference_t;
2700
2701 static int
2702 cmp_interference(const void *a, const void *b, size_t size)
2703 {
2704         const interference_t *p = a;
2705         const interference_t *q = b;
2706
2707         return !(p->a == q->a && p->b == q->b);
2708 }
2709
2710 static interference_t *
2711 set_find_interference(set * set, ir_node * a, ir_node * b)
2712 {
2713         interference_t     query;
2714
2715         query.a = (a>b)?a:b;
2716         query.b = (a>b)?b:a;
2717
2718         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2719 }
2720
2721 static interference_t *
2722 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2723 {
2724         interference_t     query,
2725                                           *result;
2726         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2727
2728         list->irn = bb;
2729
2730         result = set_find_interference(set, a, b);
2731         if(result) {
2732
2733                 list_add(&list->list, &result->blocklist);
2734                 return result;
2735         }
2736
2737         query.a = (a>b)?a:b;
2738         query.b = (a>b)?b:a;
2739
2740         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2741
2742         INIT_LIST_HEAD(&result->blocklist);
2743         list_add(&list->list, &result->blocklist);
2744
2745         return result;
2746 }
2747
2748 static int
2749 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2750 {
2751         const ir_edge_t *edge;
2752
2753         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2754                 /* both values are live in, so they interfere */
2755                 return 1;
2756         }
2757
2758         /* ensure a dominates b */
2759         if(value_dominates(b,a)) {
2760                 const ir_node * t;
2761                 t = b;
2762                 b = a;
2763                 a = t;
2764         }
2765         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2766
2767
2768         /* the following code is stolen from bera.c */
2769         if(be_is_live_end(si->lv, bb, a))
2770                 return 1;
2771
2772         foreach_out_edge(a, edge) {
2773                 const ir_node *user = edge->src;
2774                 if(get_nodes_block(user) == bb
2775                                 && !is_Phi(user)
2776                                 && b != user
2777                                 && !pset_find_ptr(si->inverse_ops, user)
2778                                 && value_dominates(b, user))
2779                         return 1;
2780         }
2781
2782         return 0;
2783 }
2784
2785 /**
2786  * Walk all irg blocks and collect interfering values inside of phi classes
2787  */
2788 static void
2789 luke_interferencewalker(ir_node * bb, void * data)
2790 {
2791         spill_ilp_t    *si = (spill_ilp_t*)data;
2792         int             l1, l2;
2793
2794         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2795                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2796                 op_t           *a_op = get_irn_link(a);
2797
2798
2799                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2800                 if (has_reg_class(si, a) && get_phi_class(a)) {
2801                         if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2802                                 continue;
2803
2804                         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)) {
2805                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
2806                                 op_t           *b_op = get_irn_link(b);
2807
2808
2809                                 /* a and b are only interesting if they are in the same phi class */
2810                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2811                                         if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2812                                                 continue;
2813
2814                                         if(values_interfere_in_block(si, bb, a, b)) {
2815                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2816                                                 set_insert_interference(si, si->interferences, a, b, bb);
2817                                         }
2818                                 }
2819                         }
2820                 }
2821         }
2822 }
2823
2824 static unsigned int copy_path_id = 0;
2825
2826 static void
2827 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2828 {
2829         ilp_cst_t  cst;
2830         ilp_var_t  copy;
2831         char       buf[256];
2832         void      *ptr;
2833
2834         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2835         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2836
2837         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2838
2839         pset_foreach(copies, ptr) {
2840                 copy = PTR_TO_INT(ptr);
2841                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2842         }
2843 }
2844
2845 /**
2846  * @parameter copies   contains a path of copies which lead us to irn
2847  * @parameter visited  contains a set of nodes already visited on this path
2848  */
2849 static int
2850 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2851 {
2852         const ir_edge_t *edge;
2853         op_t            *op = get_irn_link(irn);
2854     pset            *visited_users = pset_new_ptr_default();
2855         int              paths = 0;
2856
2857         if(op->is_remat) return 0;
2858
2859         pset_insert_ptr(visited, irn);
2860
2861         if(is_Phi(irn)) {
2862                 int    n;
2863         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2864
2865                 /* visit all operands */
2866                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2867                         ir_node  *arg = get_irn_n(irn, n);
2868                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2869
2870                         if(!has_reg_class(si, arg)) continue;
2871             if(pset_find_ptr(visited_operands, arg)) continue;
2872             pset_insert_ptr(visited_operands, arg);
2873
2874                         if(arg == target) {
2875                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2876                                         del_pset(visited_operands);
2877                                         del_pset(visited_users);
2878                                         pset_remove_ptr(visited, irn);
2879                                         return paths;
2880                                 }
2881                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2882                                 write_copy_path_cst(si, copies, any_interfere);
2883                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2884                         } else if(!pset_find_ptr(visited, arg)) {
2885                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2886                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2887                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2888
2889                 if(paths > MAX_PATHS) {
2890                     if(pset_count(copies) == 0) {
2891                         ilp_cst_t  cst;
2892                         char       buf[256];
2893
2894                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2895                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2896                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2897                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2898                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2899
2900                         paths = 0;
2901                     } else {
2902                         del_pset(visited_operands);
2903                         del_pset(visited_users);
2904                         pset_remove_ptr(visited, irn);
2905                         return paths;
2906                     }
2907                 } else if(pset_count(copies) == 0) {
2908                                         paths = 0;
2909                                 }
2910                         }
2911                 }
2912
2913         del_pset(visited_operands);
2914         }
2915
2916         /* visit all uses which are phis */
2917         foreach_out_edge(irn, edge) {
2918                 ir_node  *user = edge->src;
2919                 int       pos  = edge->pos;
2920                 op_t     *op = get_irn_link(user);
2921                 ilp_var_t copy;
2922
2923                 if(!is_Phi(user)) continue;
2924                 if(!has_reg_class(si, user)) continue;
2925         if(pset_find_ptr(visited_users, user)) continue;
2926         pset_insert_ptr(visited_users, user);
2927
2928                 copy = op->attr.live_range.args.copies[pos];
2929
2930                 if(user == target) {
2931                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2932                                 del_pset(visited_users);
2933                                 pset_remove_ptr(visited, irn);
2934                                 return paths;
2935                         }
2936                         pset_insert(copies, INT_TO_PTR(copy), copy);
2937                         write_copy_path_cst(si, copies, any_interfere);
2938                         pset_remove(copies, INT_TO_PTR(copy), copy);
2939                 } else if(!pset_find_ptr(visited, user)) {
2940                         pset_insert(copies, INT_TO_PTR(copy), copy);
2941                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2942                         pset_remove(copies, INT_TO_PTR(copy), copy);
2943
2944             if(paths > MAX_PATHS) {
2945                 if(pset_count(copies) == 0) {
2946                     ilp_cst_t  cst;
2947                     char       buf[256];
2948
2949                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2950                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2951                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2952                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2953                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2954
2955                     paths = 0;
2956                 } else {
2957                     del_pset(visited_users);
2958                     pset_remove_ptr(visited, irn);
2959                     return paths;
2960                 }
2961             } else if(pset_count(copies) == 0) {
2962                                 paths = 0;
2963                         }
2964                 }
2965         }
2966
2967     del_pset(visited_users);
2968         pset_remove_ptr(visited, irn);
2969         return paths;
2970 }
2971
2972 static void
2973 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2974 {
2975         pset * copies = pset_new_ptr_default();
2976         pset * visited = pset_new_ptr_default();
2977
2978         find_copy_path(si, a, b, any_interfere, copies, visited);
2979
2980         del_pset(visited);
2981         del_pset(copies);
2982 }
2983
2984
2985 static void
2986 memcopyhandler(spill_ilp_t * si)
2987 {
2988         interference_t   *interference;
2989         char              buf[256];
2990         /* teste Speicherwerte auf Interferenz */
2991
2992         /* analyze phi classes */
2993         phi_class_compute(si->chordal_env->irg);
2994
2995         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2996         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2997
2998         /* now lets emit the ILP unequations for the crap */
2999         set_foreach(si->interferences, interference) {
3000                 irnlist_t      *irnlist;
3001                 ilp_var_t       interfere,
3002                                                 any_interfere;
3003                 ilp_cst_t       any_interfere_cst,
3004                                                 cst;
3005                 const ir_node  *a  = interference->a;
3006                 const ir_node  *b  = interference->b;
3007
3008                 /* any_interf <= \sum interf */
3009                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3010                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3011                 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3012
3013                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3014
3015                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3016                         const ir_node  *bb = irnlist->irn;
3017                         spill_bb_t     *spill_bb = get_irn_link(bb);
3018                         spill_t        *spilla,
3019                                                    *spillb;
3020                         char           buf[256];
3021
3022                         spilla = set_find_spill(spill_bb->ilp, a);
3023                         assert(spilla);
3024
3025                         spillb = set_find_spill(spill_bb->ilp, b);
3026                         assert(spillb);
3027
3028                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3029                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3030                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3031                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3032                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3033                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3034
3035                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3036                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3037
3038                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3039                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3040                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3041                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3042                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3043
3044                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3045                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3046
3047                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3048                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3049                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3050
3051                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3052                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3053
3054                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3055                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3056                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3057
3058
3059                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3060
3061                         /* any_interfere >= interf */
3062                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3063                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3064
3065                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3066                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3067                 }
3068
3069                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3070                 gen_copy_constraints(si,a,b,any_interfere);
3071         }
3072 }
3073
3074
3075 static INLINE int
3076 is_zero(double x)
3077 {
3078         return fabs(x) < 0.00001;
3079 }
3080
3081 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3082 {
3083         spill_ilp_t *si = get_irg_link(current_ir_graph);
3084
3085         if(pset_find_ptr(si->all_possible_remats, n)) {
3086                 op_t   *op = (op_t*)get_irn_link(n);
3087                 assert(op && op->is_remat);
3088
3089                 if(!op->attr.remat.remat->inverse) {
3090                         if(op->attr.remat.pre) {
3091                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3092                         } else {
3093                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3094                         }
3095
3096                         return 1;
3097                 } else {
3098                         op_t   *op = (op_t*)get_irn_link(n);
3099                         assert(op && op->is_remat);
3100
3101                         if(op->attr.remat.pre) {
3102                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3103                         } else {
3104                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3105                         }
3106
3107                         return 1;
3108                 }
3109         }
3110
3111         return 0;
3112 }
3113
3114 static void
3115 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3116 {
3117         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3118         be_dump(irg, suffix, dump_ir_block_graph_sched);
3119         set_dump_node_vcgattr_hook(NULL);
3120 }
3121
3122 /**
3123  * Edge hook to dump the schedule edges with annotated register pressure.
3124  */
3125 static int
3126 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3127 {
3128         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3129                 ir_node *prev = sched_prev(irn);
3130                 fprintf(F, "edge:{sourcename:\"");
3131                 PRINT_NODEID(irn);
3132                 fprintf(F, "\" targetname:\"");
3133                 PRINT_NODEID(prev);
3134                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3135                 fprintf(F, "\" color:magenta}\n");
3136         }
3137         return 1;
3138 }
3139
3140 static void
3141 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3142 {
3143         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3144
3145         dump_consts_local(0);
3146         set_dump_node_edge_hook(sched_pressure_edge_hook);
3147         dump_ir_block_graph(irg, suffix);
3148         set_dump_node_edge_hook(old_edge_hook);
3149 }
3150
3151 static void
3152 walker_pressure_annotator(ir_node * bb, void * data)
3153 {
3154         spill_ilp_t  *si = data;
3155         ir_node      *irn;
3156         int           n, i;
3157         pset         *live = pset_new_ptr_default();
3158         int           projs = 0;
3159
3160         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3161                 irn = be_lv_get_irn(si->lv, bb, i);
3162
3163                 if (has_reg_class(si, irn)) {
3164                         pset_insert_ptr(live, irn);
3165                 }
3166         }
3167
3168         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3169
3170         sched_foreach_reverse(bb, irn) {
3171                 if(is_Phi(irn)) {
3172                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3173                         continue;
3174                 }
3175
3176                 if(has_reg_class(si, irn)) {
3177                         pset_remove_ptr(live, irn);
3178                         if(is_Proj(irn)) ++projs;
3179                 }
3180
3181                 if(!is_Proj(irn)) projs = 0;
3182
3183                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3184                         ir_node    *arg = get_irn_n(irn, n);
3185
3186                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3187                 }
3188                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3189         }
3190
3191         del_pset(live);
3192 }
3193
3194 static void
3195 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3196 {
3197         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3198 }
3199
3200 static void
3201 connect_all_remats_with_keep(spill_ilp_t * si)
3202 {
3203         ir_node   *irn;
3204         ir_node  **ins,
3205                          **pos;
3206         int        n_remats;
3207
3208
3209         n_remats = pset_count(si->all_possible_remats);
3210         if(n_remats) {
3211                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3212
3213                 pos = ins;
3214                 pset_foreach(si->all_possible_remats, irn) {
3215                         *pos = irn;
3216                         ++pos;
3217                 }
3218
3219                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3220
3221                 obstack_free(si->obst, ins);
3222         }
3223 }
3224
3225 static void
3226 connect_all_spills_with_keep(spill_ilp_t * si)
3227 {
3228         ir_node   *irn;
3229         ir_node  **ins,
3230                          **pos;
3231         int        n_spills;
3232         ir_node   *keep;
3233
3234
3235         n_spills = pset_count(si->spills);
3236         if(n_spills) {
3237                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3238
3239                 pos = ins;
3240                 pset_foreach(si->spills, irn) {
3241                         *pos = irn;
3242                         ++pos;
3243                 }
3244
3245                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3246
3247                 obstack_free(si->obst, ins);
3248         }
3249 }
3250
3251 /** insert a spill at an arbitrary position */
3252 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3253 {
3254         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
3255         ir_graph *irg   = get_irn_irg(bl);
3256         ir_node *frame  = get_irg_frame(irg);
3257         ir_node *spill;
3258         ir_node *next;
3259
3260         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3261         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3262
3263         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3264
3265         /*
3266          * search the right insertion point. a spill of a phi cannot be put
3267          * directly after the phi, if there are some phis behind the one which
3268          * is spilled. Also, a spill of a Proj must be after all Projs of the
3269          * same tuple node.
3270          *
3271          * Here's one special case:
3272          * If the spill is in the start block, the spill must be after the frame
3273          * pointer is set up. This is done by setting insert to the end of the block
3274          * which is its default initialization (see above).
3275          */
3276
3277         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3278                 insert = frame;
3279
3280         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3281                 insert = next;
3282
3283         sched_add_after(insert, spill);
3284         return spill;
3285 }
3286
3287 static void
3288 delete_remat(spill_ilp_t * si, ir_node * remat) {
3289         int       n;
3290         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3291
3292         sched_remove(remat);
3293
3294         /* kill links to operands */
3295         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3296                 set_irn_n(remat, n, bad);
3297         }
3298 }
3299
3300 static void
3301 clean_remat_info(spill_ilp_t * si)
3302 {
3303         int            n;
3304         remat_t       *remat;
3305         remat_info_t  *remat_info;
3306         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
3307
3308         set_foreach(si->remat_info, remat_info) {
3309                 if(!remat_info->remats) continue;
3310
3311                 pset_foreach(remat_info->remats, remat)
3312                 {
3313                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3314                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3315                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3316                         }
3317
3318                         if(get_irn_n_edges(remat->op) == 0) {
3319                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3320                                         set_irn_n((ir_node*)remat->op, n, bad);
3321                                 }
3322                         }
3323                 }
3324
3325                 if(remat_info->remats) del_pset(remat_info->remats);
3326                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3327         }
3328 }
3329
3330 static void
3331 delete_unnecessary_remats(spill_ilp_t * si)
3332 {
3333         if(opt_keep_alive & KEEPALIVE_REMATS) {
3334                 int       n;
3335                 ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3336
3337                 if(si->keep) {
3338 //                      ir_node   *end = get_irg_end(si->chordal_env->irg);
3339 //                      ir_node  **keeps;
3340
3341                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3342                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3343                                 op_t           *arg_op = get_irn_link(keep_arg);
3344                                 lpp_name_t     *name;
3345
3346                                 assert(arg_op->is_remat);
3347
3348                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3349
3350                                 if(is_zero(name->value)) {
3351                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3352                                         /* TODO check whether reload is preferred over remat (could be bug) */
3353                                         delete_remat(si, keep_arg);
3354                                 } else {
3355                                         if(!arg_op->attr.remat.remat->inverse) {
3356                                                 if(arg_op->attr.remat.pre) {
3357                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3358                                                 } else {
3359                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3360                                                 }
3361                                         } else {
3362                                                 if(arg_op->attr.remat.pre) {
3363                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3364                                                 } else {
3365                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3366                                                 }
3367                                         }
3368                                 }
3369
3370                                 set_irn_n(si->keep, n, bad);
3371                         }
3372 #if 0
3373                         for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3374                                 ir_node        *end_arg = get_End_keepalive(end, i);
3375
3376                                 if(end_arg != si->keep) {
3377                                         obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3378                                 }
3379                         }
3380                         keeps = obstack_finish(si->obst);
3381                         set_End_keepalives(end, n-1, keeps);
3382                         obstack_free(si->obst, keeps);
3383 #endif
3384                 } else {
3385                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3386                 }
3387         } else {
3388                 ir_node  *remat;
3389
3390                 pset_foreach(si->all_possible_remats, remat) {
3391                         op_t           *remat_op = get_irn_link(remat);
3392                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3393
3394                         if(is_zero(name->value)) {
3395                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3396                                 /* TODO check whether reload is preferred over remat (could be bug) */
3397                                 delete_remat(si, remat);
3398                         } else {
3399                                 if(!remat_op->attr.remat.remat->inverse) {
3400                                         if(remat_op->attr.remat.pre) {
3401                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3402                                         } else {
3403                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3404                                         }
3405                                 } else {
3406                                         if(remat_op->attr.remat.pre) {
3407                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3408                                         } else {
3409                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3410                                         }
3411                                 }
3412                         }
3413                 }
3414         }
3415 }
3416
3417 static pset *
3418 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3419 {
3420         pset     *spills = pset_new_ptr_default();
3421
3422         const ir_node  *next;
3423         defs_t         *defs;
3424
3425         defs = set_find_def(si->values, value);
3426
3427         if(defs && defs->spills) {
3428                 for(next = defs->spills; next; next = get_irn_link(next)) {
3429                         pset_insert_ptr(spills, next);
3430                 }
3431         }
3432
3433         return spills;
3434 }
3435
3436 /**
3437  * @param before   The node after which the spill will be placed in the schedule
3438  */
3439 static ir_node *
3440 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3441 {
3442         defs_t   *defs;
3443         ir_node  *spill;
3444         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3445
3446         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3447
3448         spill = be_spill2(arch_env, irn, before);
3449
3450         defs = set_insert_def(si->values, value);
3451         assert(defs);
3452
3453         /* enter into the linked list */
3454         set_irn_link(spill, defs->spills);
3455         defs->spills = spill;
3456
3457         if(opt_keep_alive & KEEPALIVE_SPILLS)
3458                 pset_insert_ptr(si->spills, spill);
3459
3460         return spill;
3461 }
3462
3463 /**
3464  * @param before   The Phi node which has to be spilled
3465  */
3466 static ir_node *
3467 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3468 {
3469         ir_node   *mem_phi;
3470         ir_node  **ins;
3471         defs_t    *defs;
3472         int        n;
3473
3474         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3475
3476         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3477                 ins[n] = si->m_unknown;
3478         }
3479
3480         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3481
3482         defs = set_insert_def(si->values, phi);
3483         assert(defs);
3484
3485         /* enter into the linked list */
3486         set_irn_link(mem_phi, defs->spills);
3487         defs->spills = mem_phi;
3488
3489 #ifdef SCHEDULE_PHIM
3490         sched_add_after(phi, mem_phi);
3491 #endif
3492
3493         if(opt_keep_alive & KEEPALIVE_SPILLS)
3494                 pset_insert_ptr(si->spills, mem_phi);
3495
3496
3497         return mem_phi;
3498 }
3499
3500 /**
3501  * Add remat to list of defs, destroys link field!
3502  */
3503 static void
3504 insert_remat(spill_ilp_t * si, ir_node * remat)
3505 {
3506         defs_t   *defs;
3507         op_t     *remat_op = get_irn_link(remat);
3508
3509         assert(remat_op->is_remat);
3510
3511         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3512         assert(defs);
3513
3514         /* enter into the linked list */
3515         set_irn_link(remat, defs->remats);
3516         defs->remats = remat;
3517 }
3518
3519
3520 /**
3521  * Add reload before operation and add to list of defs
3522  */
3523 static ir_node *
3524 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3525 {
3526         defs_t   *defs;
3527         ir_node  *reload,
3528                          *spill;
3529         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3530
3531         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3532
3533         defs = set_find_def(si->values, value);
3534
3535         spill = defs->spills;
3536         assert(spill && "no spill placed before reload");
3537
3538         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3539
3540         /* enter into the linked list */
3541         set_irn_link(reload, defs->remats);
3542         defs->remats = reload;
3543
3544         return reload;
3545 }
3546
3547 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3548 {
3549         defs_t           *defs;
3550         ir_node          *reload;
3551         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3552         ir_node          *spill;
3553         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3554
3555         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3556
3557         defs = set_find_def(si->values, value);
3558
3559         spill = defs->spills;
3560         assert(spill && "no spill placed before reload");
3561
3562         reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3563
3564         arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3565         sched_remove(reload);
3566 }
3567
3568 void insert_memoperands(spill_ilp_t * si)
3569 {
3570         memoperand_t   *memoperand;
3571         lpp_name_t     *name;
3572
3573         set_foreach(si->memoperands, memoperand) {
3574                 name = si->lpp->vars[memoperand->ilp];
3575                 if(!is_zero(name->value)) {
3576                         perform_memory_operand(si, memoperand);
3577                 }
3578         }
3579 }
3580
3581 static void
3582 walker_spill_placer(ir_node * bb, void * data) {
3583         spill_ilp_t   *si = (spill_ilp_t*)data;
3584         ir_node       *irn;
3585         spill_bb_t    *spill_bb = get_irn_link(bb);
3586         pset          *spills_to_do = pset_new_ptr_default();
3587         spill_t       *spill;
3588
3589         set_foreach(spill_bb->ilp, spill) {
3590                 lpp_name_t    *name;
3591
3592                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3593                         name = si->lpp->vars[spill->mem_in];
3594                         if(!is_zero(name->value)) {
3595                                 ir_node   *mem_phi;
3596
3597                                 mem_phi = insert_mem_phi(si, spill->irn);
3598
3599                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3600                         }
3601                 }
3602
3603                 name = si->lpp->vars[spill->spill];
3604                 if(!is_zero(name->value)) {
3605                         /* place spill directly after definition */
3606                         if(get_nodes_block(spill->irn) == bb) {
3607                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3608                                 continue;
3609                         }
3610
3611                         /* place spill at bb start */
3612                         if(spill->reg_in > 0) {
3613                                 name = si->lpp->vars[spill->reg_in];
3614                                 if(!is_zero(name->value)) {
3615                                         insert_spill(si, spill->irn, spill->irn, bb);
3616                                         continue;
3617                                 }
3618                         }
3619                         /* place spill after a remat */
3620                         pset_insert_ptr(spills_to_do, spill->irn);
3621                 }
3622         }
3623         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3624
3625
3626         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3627                 op_t     *op = get_irn_link(irn);
3628
3629                 if(be_is_Spill(irn)) continue;
3630
3631                 if(op->is_remat) {
3632                         /* TODO fix this if we want to support remats with more than two nodes */
3633                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3634                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3635
3636                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3637                         }
3638                 } else {
3639                         if(pset_find_ptr(spills_to_do, irn)) {
3640                                 pset_remove_ptr(spills_to_do, irn);
3641
3642                                 insert_spill(si, irn, irn, irn);
3643                         }
3644                 }
3645
3646         }
3647
3648         assert(pset_count(spills_to_do) == 0);
3649
3650         /* afterwards free data in block */
3651         del_pset(spills_to_do);
3652 }
3653
3654 static ir_node *
3655 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3656 {
3657         ir_node          *insert_pos = bb;
3658         ir_node          *spill;
3659         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3660
3661         /* find last definition of arg value in block */
3662         ir_node  *next;
3663         defs_t   *defs;
3664         int       last = 0;
3665
3666         defs = set_find_def(si->values, value);
3667
3668         if(defs && defs->remats) {
3669                 for(next = defs->remats; next; next = get_irn_link(next)) {
3670                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3671                                 last = sched_get_time_step(next);
3672                                 insert_pos = next;
3673                         }
3674                 }
3675         }
3676
3677         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3678                 last = sched_get_time_step(value);
3679                 insert_pos = value;
3680         }
3681
3682         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3683
3684         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3685
3686         return spill;
3687 }
3688
3689 static void
3690 phim_fixer(spill_ilp_t *si) {
3691         defs_t  *defs;
3692
3693         set_foreach(si->values, defs) {
3694                 const ir_node  *phi = defs->value;
3695                 op_t           *op = get_irn_link(phi);
3696                 ir_node        *phi_m = NULL;
3697                 ir_node        *next = defs->spills;
3698                 int             n;
3699
3700                 if(!is_Phi(phi)) continue;
3701
3702                 while(next) {
3703                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3704                                 phi_m = next;
3705                                 break;
3706                         } else {
3707                                 next = get_irn_link(next);
3708                         }
3709                 }
3710                 if(!phi_m) continue;
3711
3712                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3713                         ir_node        *value = get_irn_n(phi, n);
3714                         defs_t         *val_defs = set_find_def(si->values, value);
3715
3716                         /* a spill of this value */
3717                         ir_node      *spill;
3718
3719
3720                         if(opt_memcopies) {
3721                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3722                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3723
3724                                 if(!is_zero(name->value)) {
3725                                         spill = insert_mem_copy(si, pred, value);
3726                                 } else {
3727                                         spill = val_defs->spills;
3728                                 }
3729                         } else {
3730                                 spill = val_defs->spills;
3731                         }
3732
3733                         assert(spill && "no spill placed before PhiM");
3734                         set_irn_n(phi_m, n, spill);
3735                 }
3736         }
3737 }
3738
3739 static void
3740 walker_reload_placer(ir_node * bb, void * data) {
3741         spill_ilp_t   *si = (spill_ilp_t*)data;
3742         ir_node       *irn;
3743         spill_bb_t    *spill_bb = get_irn_link(bb);
3744
3745         /* reloads at end of block */
3746         if(spill_bb->reloads) {
3747                 keyval_t    *keyval;
3748
3749                 set_foreach(spill_bb->reloads, keyval) {
3750                         ir_node        *irn = (ir_node*)keyval->key;
3751                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3752                         lpp_name_t     *name;
3753
3754                         name = si->lpp->vars[reload];
3755                         if(!is_zero(name->value)) {
3756                                 ir_node    *reload;
3757                                 ir_node    *insert_pos = bb;
3758                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3759                                 op_t       *prev_op = get_irn_link(prev);
3760
3761                                 while(be_is_Spill(prev)) {
3762                                         prev = sched_prev(prev);
3763                                 }
3764
3765                                 prev_op = get_irn_link(prev);
3766
3767                                 /* insert reload before pre-remats */
3768                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3769                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3770                                         insert_pos = prev;
3771
3772                                         do {
3773                                                 prev = sched_prev(prev);
3774                                         } while(be_is_Spill(prev));
3775
3776                                         prev_op = get_irn_link(prev);
3777
3778                                 }
3779
3780                                 reload = insert_reload(si, irn, insert_pos);
3781
3782                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3783                                         pset_insert_ptr(si->spills, reload);
3784                         }
3785                 }
3786         }
3787
3788         /* walk and insert more reloads and collect remats */
3789         sched_foreach_reverse(bb, irn) {
3790                 op_t     *op = get_irn_link(irn);
3791
3792                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3793                 if(is_Phi(irn)) break;
3794
3795                 if(op->is_remat) {
3796                         if(get_irn_mode(irn) != mode_T) {
3797                                 insert_remat(si, irn);
3798                         }
3799                 } else {
3800                         int    n;
3801
3802                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3803                                 ir_node    *arg = get_irn_n(irn, n);
3804
3805                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3806                                         lpp_name_t    *name;
3807
3808                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3809                                         if(!is_zero(name->value)) {
3810                                                 ir_node    *reload;
3811                                                 ir_node    *insert_pos = irn;
3812                                                 ir_node    *prev = sched_prev(insert_pos);
3813                                                 op_t       *prev_op;
3814
3815                                                 while(be_is_Spill(prev)) {
3816                                                         prev = sched_prev(prev);
3817                                                 }
3818
3819                                                 prev_op = get_irn_link(prev);
3820
3821                                                 /* insert reload before pre-remats */
3822                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3823                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3824                                                         insert_pos = prev;
3825
3826                                                         do {
3827                                                                 prev = sched_prev(prev);
3828                                                         } while(be_is_Spill(prev));
3829
3830                                                         prev_op = get_irn_link(prev);
3831
3832                                                 }
3833
3834                                                 reload = insert_reload(si, arg, insert_pos);
3835
3836                                                 set_irn_n(irn, n, reload);
3837
3838                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3839                                                         pset_insert_ptr(si->spills, reload);
3840                                         }
3841                                 }
3842                         }
3843                 }
3844         }
3845
3846         del_set(spill_bb->ilp);
3847         if(spill_bb->reloads) del_set(spill_bb->reloads);
3848 }
3849
3850 static void
3851 walker_collect_used(ir_node * irn, void * data)
3852 {
3853         lc_bitset_t   *used = data;
3854
3855         lc_bitset_set(used, get_irn_idx(irn));
3856 }
3857
3858 struct kill_helper {
3859         lc_bitset_t  *used;
3860         spill_ilp_t  *si;
3861 };
3862
3863 static void
3864 walker_kill_unused(ir_node * bb, void * data)
3865 {
3866         struct kill_helper *kh = data;
3867         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
3868         ir_node            *irn;
3869
3870
3871         for(irn=sched_first(bb); !sched_is_end(irn);) {
3872                 ir_node     *next = sched_next(irn);
3873                 int          n;
3874
3875                 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3876                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3877                                 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)));
3878 #if 0
3879                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3880 #endif
3881                         }
3882
3883                         sched_remove(irn);
3884
3885                         set_nodes_block(irn, bad);
3886                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3887                                 set_irn_n(irn, n, bad);
3888                         }
3889                 }
3890                 irn = next;
3891         }
3892 }
3893
3894 static void
3895 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3896 {
3897         struct kill_helper kh;
3898
3899         kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3900         kh.si = si;
3901
3902         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3903         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3904
3905         lc_bitset_free(kh.used);
3906 }
3907
3908 void
3909 print_irn_pset(pset * p)
3910 {
3911         ir_node   *irn;
3912
3913         pset_foreach(p, irn) {
3914                 ir_printf("%+F\n", irn);
3915         }
3916 }
3917
3918 void
3919 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3920 {
3921     FILE           *f = fopen(file, "w");
3922     ir_node        *irn;
3923     interference_t *interference;
3924
3925     pset_break(phiclass);
3926     set_break(si->interferences);
3927
3928     ir_fprintf(f, "digraph phiclass {\n");
3929
3930     pset_foreach(phiclass, irn) {
3931         if(is_Phi(irn))
3932             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
3933     }
3934
3935     pset_foreach(phiclass, irn) {
3936         int n;
3937
3938         if(!is_Phi(irn)) continue;
3939
3940         for(n=get_irn_arity(irn)-1; n>=0; --n) {
3941             ir_node  *arg = get_irn_n(irn, n);
3942
3943             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
3944         }
3945     }
3946
3947     set_foreach(si->interferences, interference) {
3948         const ir_node  *a  = interference->a;
3949         const ir_node  *b  = interference->b;
3950         if(get_phi_class(a) == phiclass) {
3951             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3952         }
3953     }
3954
3955     ir_fprintf(f, "}");
3956     fclose(f);
3957 }
3958
3959 static void
3960 rewire_uses(spill_ilp_t * si)
3961 {
3962         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3963         defs_t               *defs;
3964         pset                 *ignore = pset_new_ptr(1);
3965
3966         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3967
3968         /* then fix uses of spills */
3969         set_foreach(si->values, defs) {
3970                 pset           *reloads;
3971                 pset           *spills;
3972                 const ir_node  *next = defs->remats;
3973                 int remats = 0;
3974
3975                 reloads = pset_new_ptr_default();
3976
3977                 while(next) {
3978                         if(be_is_Reload(next)) {
3979                                 pset_insert_ptr(reloads, next);
3980                         } else {
3981                                 ++remats;
3982                         }
3983                         next = get_irn_link(next);
3984                 }
3985
3986                 spills = get_spills_for_value(si, defs->value);
3987                 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));
3988                 if(pset_count(spills) > 1) {
3989                         //assert(pset_count(reloads) > 0);
3990                         //                              print_irn_pset(spills);
3991                         //                              print_irn_pset(reloads);
3992
3993                         be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3994                 }
3995
3996                 del_pset(reloads);
3997                 del_pset(spills);
3998         }
3999
4000         /* first fix uses of remats and reloads */
4001         set_foreach(si->values, defs) {
4002                 pset           *nodes;
4003                 const ir_node  *next = defs->remats;
4004
4005                 if(next) {
4006                         nodes = pset_new_ptr_default();
4007                         pset_insert_ptr(nodes, defs->value);
4008
4009                         while(next) {
4010                                 pset_insert_ptr(nodes, next);
4011                                 next = get_irn_link(next);
4012                         }
4013
4014                         if(pset_count(nodes) > 1) {
4015                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
4016                                 be_ssa_constr_set(dfi, si->lv, nodes);
4017                         }
4018
4019                         del_pset(nodes);
4020                 }
4021         }
4022
4023 //      remove_unused_defs(si);
4024
4025         be_free_dominance_frontiers(dfi);
4026 }
4027
4028
4029 static void
4030 writeback_results(spill_ilp_t * si)
4031 {
4032         /* walk through the graph and collect all spills, reloads and remats for a value */
4033
4034         si->values = new_set(cmp_defs, 4096);
4035
4036         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4037         delete_unnecessary_remats(si);
4038         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4039         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4040         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4041         if(opt_memoperands)
4042                 insert_memoperands(si);
4043         phim_fixer(si);
4044
4045         /* clean the remat info! there are still back-edges leading there! */
4046         clean_remat_info(si);
4047
4048         rewire_uses(si);
4049
4050         connect_all_spills_with_keep(si);
4051
4052         del_set(si->values);
4053 }
4054
4055 static int
4056 get_n_regs(spill_ilp_t * si)
4057 {
4058         int     arch_n_regs = arch_register_class_n_regs(si->cls);
4059         int     free = 0;
4060         int     i;
4061
4062         for(i=0; i<arch_n_regs; i++) {
4063                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
4064                         free++;
4065                 }
4066         }
4067
4068         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
4069         return free;
4070 }
4071
4072 static void
4073 walker_reload_mover(ir_node * bb, void * data)
4074 {
4075         spill_ilp_t   *si = data;
4076         ir_node           *tmp;
4077
4078         sched_foreach(bb, tmp) {
4079                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4080                         ir_node       *reload = tmp;
4081                         ir_node       *irn = tmp;
4082
4083                         /* move reload upwards */
4084
4085                         int pressure = (int)get_irn_link(reload);
4086                         if(pressure < si->n_regs) {
4087                                 irn = sched_prev(reload);
4088                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4089                                 sched_remove(reload);
4090                                 pressure = (int)get_irn_link(irn);
4091
4092                                 while(pressure < si->n_regs) {
4093                                         if( sched_is_end(irn) ||
4094                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4095                                            /* do not move reload before its spill */
4096                                            (irn == be_get_Reload_mem(reload)) ||
4097                                            /* do not move before phi */
4098                                            is_Phi(irn)) break;
4099
4100                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4101                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4102                                         irn = sched_prev(irn);
4103
4104                                         pressure = (int)get_irn_link(irn);
4105                                 }
4106
4107                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4108                                 sched_put_after(irn, reload);
4109                         }
4110                 }
4111         }
4112 }
4113
4114 static void
4115 move_reloads_upward(spill_ilp_t * si)
4116 {
4117         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4118 }
4119
4120
4121 /**
4122  * Walk all irg blocks and check for interfering spills inside of phi classes
4123  */
4124 static void
4125 luke_meminterferencechecker(ir_node * bb, void * data)
4126 {
4127         spill_ilp_t    *si = (spill_ilp_t*)data;
4128         int             l1, l2;
4129
4130         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4131                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4132
4133                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4134
4135                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4136                 if (has_reg_class(si, a) && get_phi_class(a)) {
4137                         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)) {
4138                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
4139
4140                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4141
4142                                 /* a and b are only interesting if they are in the same phi class */
4143                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4144                                         if(values_interfere_in_block(si, bb, a, b)) {
4145                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4146                                         }
4147                                 }
4148                         }
4149                 }
4150         }
4151 }
4152
4153 static void
4154 verify_phiclasses(spill_ilp_t * si)
4155 {
4156         /* analyze phi classes */
4157         phi_class_compute(si->chordal_env->irg);
4158
4159         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4160         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4161 }
4162
4163 void
4164 be_spill_remat(const be_chordal_env_t * chordal_env)
4165 {
4166         char            buf[256];
4167         char            problem_name[256];
4168         char            dump_suffix[256];
4169         char            dump_suffix2[256];
4170         struct obstack  obst;
4171         spill_ilp_t     si;
4172
4173         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4174         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4175         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4176
4177         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4178         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4179
4180         if(opt_verify & VERIFY_DOMINANCE)
4181                 be_check_dominance(chordal_env->irg);
4182
4183         obstack_init(&obst);
4184         si.chordal_env = chordal_env;
4185         si.obst = &obst;
4186         si.cls = chordal_env->cls;
4187         si.lpp = new_lpp(problem_name, lpp_minimize);
4188         si.remat_info = new_set(cmp_remat_info, 4096);
4189         si.interferences = new_set(cmp_interference, 32);
4190         si.memoperands = new_set(cmp_memoperands, 128);
4191         si.all_possible_remats = pset_new_ptr_default();
4192         si.spills = pset_new_ptr_default();
4193         si.inverse_ops = pset_new_ptr_default();
4194         si.lv = chordal_env->lv;
4195         si.keep = NULL;
4196         si.n_regs = get_n_regs(&si);
4197
4198         set_irg_link(chordal_env->irg, &si);
4199         compute_doms(chordal_env->irg);
4200
4201         /* compute phi classes */
4202 //      phi_class_compute(chordal_env->irg);
4203
4204         be_analyze_regpressure(chordal_env, "-pre");
4205
4206         if(opt_remats) {
4207                 /* collect remats */
4208                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4209                 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4210         }
4211
4212         /* insert possible remats */
4213         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4214         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4215         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4216
4217         if(opt_keep_alive & KEEPALIVE_REMATS) {
4218                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4219                 connect_all_remats_with_keep(&si);
4220                 /* dump graph with inserted remats */
4221                 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4222         }
4223
4224         /* insert copies for phi arguments not in my regclass */
4225         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4226
4227         /* recompute liveness */
4228         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4229         be_liveness_recompute(si.lv);
4230
4231         /* build the ILP */
4232
4233         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4234         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4235         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4236
4237         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4238         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4239
4240         if(opt_memcopies) {
4241                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4242                 memcopyhandler(&si);
4243         }
4244
4245         if(opt_dump_flags & DUMP_PROBLEM) {
4246                 FILE           *f;
4247                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4248                 if ((f = fopen(buf, "wt")) != NULL) {
4249                         lpp_dump_plain(si.lpp, f);
4250                         fclose(f);
4251                 }
4252         }
4253
4254         if(opt_dump_flags & DUMP_MPS) {
4255                 FILE *f;
4256
4257                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4258                 if((f = fopen(buf, "wt")) != NULL) {
4259                         mps_write_mps(si.lpp, s_mps_fixed, f);
4260                         fclose(f);
4261                 }
4262
4263                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4264                 if((f = fopen(buf, "wt")) != NULL) {
4265                         mps_write_mst(si.lpp, s_mps_fixed, f);
4266                         fclose(f);
4267                 }
4268         }
4269
4270         lpp_check_startvals(si.lpp);
4271
4272 #ifdef SOLVE
4273         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4274         lpp_set_time_limit(si.lpp, opt_timeout);
4275
4276         if(opt_log)
4277                 lpp_set_log(si.lpp, stdout);
4278
4279 #ifdef SOLVE_LOCAL
4280         lpp_solve_cplex(si.lpp);
4281 #else
4282         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4283 #endif
4284         assert(lpp_is_sol_valid(si.lpp)
4285                && "solution of ILP must be valid");
4286
4287         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));
4288
4289         if(opt_dump_flags & DUMP_SOLUTION) {
4290                 FILE           *f;
4291                 char            buf[256];
4292
4293                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4294                 if ((f = fopen(buf, "wt")) != NULL) {
4295                         int             i;
4296                         for (i = 0; i < si.lpp->var_next; ++i) {
4297                                 lpp_name_t     *name = si.lpp->vars[i];
4298                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4299                         }
4300                         fclose(f);
4301                 }
4302         }
4303
4304         writeback_results(&si);
4305
4306 #endif                          /* SOLVE */
4307
4308         kill_all_unused_values_in_schedule(&si);
4309
4310         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4311                 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4312
4313         // move reloads upwards
4314         be_liveness_recompute(si.lv);
4315         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4316         move_reloads_upward(&si);
4317
4318         if(opt_memcopies) {
4319                 verify_phiclasses(&si);
4320         }
4321
4322         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4323
4324         dump_pressure_graph(&si, dump_suffix2);
4325
4326         be_analyze_regpressure(chordal_env, "-post");
4327
4328         if(opt_verify & VERIFY_DOMINANCE)
4329                 be_check_dominance(chordal_env->irg);
4330
4331         free_dom(chordal_env->irg);
4332         del_set(si.interferences);
4333         del_pset(si.inverse_ops);
4334         del_pset(si.all_possible_remats);
4335         del_set(si.memoperands);
4336         del_pset(si.spills);
4337         free_lpp(si.lpp);
4338         obstack_free(&obst, NULL);
4339         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4340 }
4341
4342 #else                           /* WITH_ILP */
4343
4344 static void
4345 only_that_you_can_compile_without_WITH_ILP_defined(void)
4346 {
4347 }
4348
4349 #endif                          /* WITH_ILP */