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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass_t.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35 #include "irvrfy.h"
  36
  37 #include <lpp/lpp.h>
  38 #include <lpp/mps.h>
  39 #include <lpp/lpp_net.h>
  40 #include <lpp/lpp_cplex.h>
  41 //#include <lc_pset.h>
  42 //#include <libcore/lc_bitset.h>
  43
  44 #include "be_t.h"
  45 #include "belive_t.h"
  46 #include "besched_t.h"
  47 #include "beirgmod.h"
  48 #include "bearch.h"
  49 #include "beabi.h"
  50 #include "benode_t.h"
  51 #include "beutil.h"
  52 #include "bespillremat.h"
  53 #include "bespill.h"
  54 #include "bepressurestat.h"
  55
  56 #include "bechordal_t.h"
  57
  58 #ifdef WITH_LIBCORE
  59 #include <libcore/lc_opts.h>
  60 #include <libcore/lc_opts_enum.h>
  61 #endif /* WITH_LIBCORE */
  62
  63 #define DUMP_PROBLEM       1
  64 #define DUMP_MPS           2
  65 #define DUMP_SOLUTION      4
  66
  67 #define KEEPALIVE_REMATS   1
  68 #define KEEPALIVE_SPILLS   2
  69 #define KEEPALIVE_RELOADS  4
  70
  71 #define VERIFY_MEMINTERF   1
  72 #define VERIFY_DOMINANCE   2
  73
  74 #define REMATS_NONE        0
  75 #define REMATS_BRIGGS      1
  76 #define REMATS_NOINVERSE   2
  77 #define REMATS_ALL         3
  78
  79 static int opt_dump_flags   = 0;
  80 static int opt_log = 0;
  81 static int opt_keep_alive   = 0;
  82 static int opt_goodwin = 1;
  83 static int opt_memcopies = 1;
  84 static int opt_memoperands = 1;
  85 static int opt_verify = VERIFY_MEMINTERF;
  86 static int opt_remats = REMATS_ALL;
  87 static int opt_repair_schedule = 0;
  88 static int opt_no_enlarge_liveness = 0;
  89 static int opt_remat_while_live = 1;
  90 static int opt_timeout = 300;
  91 static double opt_cost_reload = 8.0;
  92 static double opt_cost_memoperand =  7.0;
  93 static double opt_cost_spill =  50.0;
  94 static double opt_cost_remat =  1.0;
  95
  96
  97 #ifdef WITH_LIBCORE
  98 static const lc_opt_enum_mask_items_t dump_items[] = {
  99         { "problem",  DUMP_PROBLEM  },
 100         { "mps",      DUMP_MPS      },
 101         { "solution", DUMP_SOLUTION },
 102         { NULL,       0 }
 103 };
 104
 105 static lc_opt_enum_mask_var_t dump_var = {
 106         &opt_dump_flags, dump_items
 107 };
 108
 109 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 110         { "remats",  KEEPALIVE_REMATS  },
 111         { "spills",  KEEPALIVE_SPILLS  },
 112         { "reloads", KEEPALIVE_RELOADS },
 113         { NULL,      0 }
 114 };
 115
 116 static lc_opt_enum_mask_var_t keep_alive_var = {
 117         &opt_keep_alive, keepalive_items
 118 };
 119
 120 static const lc_opt_enum_mask_items_t remats_items[] = {
 121         { "none",      REMATS_NONE      },
 122         { "briggs",    REMATS_BRIGGS    },
 123         { "noinverse", REMATS_NOINVERSE },
 124         { "all",       REMATS_ALL       },
 125         { NULL,        0 }
 126 };
 127
 128 static lc_opt_enum_mask_var_t remats_var = {
 129         &opt_remats, remats_items
 130 };
 131
 132 static const lc_opt_table_entry_t options[] = {
 133         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads",                   &keep_alive_var),
 134
 135         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 136         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 137         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 138         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
 139         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 140         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 141         LC_OPT_ENT_BOOL     ("remat_while_live",  "remat only values that can be used by real ops", &opt_remat_while_live),
 142
 143         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution",                               &dump_var),
 144         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 145         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 146
 147         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 148         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 149         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 150         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 151         { NULL }
 152 };
 153
 154 void be_spill_remat_register_options(lc_opt_entry_t *grp)
 155 {
 156         lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
 157         lc_opt_add_table(my_grp, options);
 158 }
 159 #endif
 160
 161
 162 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 163 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 164
 165 #define  SOLVE
 166 //#define  SOLVE_LOCAL
 167 #define LPP_SERVER "i44pc52"
 168 #define LPP_SOLVER "cplex"
 169
 170
 171 #define MAX_PATHS      INT_MAX
 172 #define ILP_UNDEF               -1
 173
 174 typedef struct _spill_ilp_t {
 175         const arch_register_class_t  *cls;
 176         int                           n_regs;
 177         const be_chordal_env_t       *chordal_env;
 178         be_lv_t                      *lv;
 179         lpp_t                        *lpp;
 180         struct obstack               *obst;
 181         set                          *remat_info;
 182         pset                         *all_possible_remats;
 183         pset                         *inverse_ops;
 184         ir_node                      *keep;
 185         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 186         pset                         *spills;
 187         set                          *interferences;
 188         ir_node                      *m_unknown;
 189         set                          *memoperands;
 190         DEBUG_ONLY(firm_dbg_module_t * dbg);
 191 } spill_ilp_t;
 192
 193 typedef int ilp_var_t;
 194 typedef int ilp_cst_t;
 195
 196 typedef struct _spill_bb_t {
 197         set      *ilp;
 198         set      *reloads;
 199 } spill_bb_t;
 200
 201 typedef struct _remat_t {
 202         const ir_node        *op;      /**< for copy_irn */
 203         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 204         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 205         int                   cost;    /**< cost of this remat */
 206         int                   inverse; /**< nonzero if this is an inverse remat */
 207 } remat_t;
 208
 209 /**
 210  * Data to be attached to each IR node. For remats this contains the ilp_var
 211  * for this remat and for normal ops this contains the ilp_vars for
 212  * reloading each operand
 213  */
 214 typedef struct _op_t {
 215         int             is_remat;
 216         union {
 217                 struct {
 218                         ilp_var_t       ilp;
 219                         const remat_t  *remat; /** the remat this op belongs to */
 220                         int             pre; /** 1, if this is a pressure-increasing remat */
 221                 } remat;
 222                 struct {
 223                         ilp_var_t       ilp;
 224                         ir_node        *op; /** the operation this live range belongs to */
 225                         union {
 226                                 ilp_var_t      *reloads;
 227                                 ilp_var_t      *copies;
 228                         } args;
 229                 } live_range;
 230         } attr;
 231 } op_t;
 232
 233 typedef struct _defs_t {
 234         const ir_node   *value;
 235         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 236         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 237 } defs_t;
 238
 239 typedef struct _remat_info_t {
 240         const ir_node       *irn; /**< the irn to which these remats belong */
 241         pset                *remats; /**< possible remats for this value */
 242         pset                *remats_by_operand; /**< remats with this value as operand */
 243 } remat_info_t;
 244
 245 typedef struct _keyval_t {
 246         const void          *key;
 247         const void          *val;
 248 } keyval_t;
 249
 250 typedef struct _spill_t {
 251         ir_node            *irn;
 252         ilp_var_t           reg_in;
 253         ilp_var_t           mem_in;
 254         ilp_var_t           reg_out;
 255         ilp_var_t           mem_out;
 256         ilp_var_t           spill;
 257 } spill_t;
 258
 259 typedef struct _memoperand_t {
 260         ir_node             *irn; /**< the irn */
 261         unsigned int         pos; /**< the position of the argument */
 262         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 263 } memoperand_t;
 264
 265 static INLINE int
 266 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 267 {
 268         return chordal_has_class(si->chordal_env, irn);
 269 }
 270
 271 #if 0
 272 static int
 273 cmp_remat(const void *a, const void *b)
 274 {
 275         const keyval_t *p = a;
 276         const keyval_t *q = b;
 277         const remat_t  *r = p->val;
 278         const remat_t  *s = q->val;
 279
 280         assert(r && s);
 281
 282         return !(r == s || r->op == s->op);
 283 }
 284 #endif
 285 static int
 286 cmp_remat(const void *a, const void *b)
 287 {
 288         const remat_t  *r = a;
 289         const remat_t  *s = a;
 290
 291         return !(r == s || r->op == s->op);
 292 }
 293
 294 static int
 295 cmp_spill(const void *a, const void *b, size_t size)
 296 {
 297         const spill_t *p = a;
 298         const spill_t *q = b;
 299
 300 //      return !(p->irn == q->irn && p->bb == q->bb);
 301         return !(p->irn == q->irn);
 302 }
 303
 304 static int
 305 cmp_memoperands(const void *a, const void *b, size_t size)
 306 {
 307         const memoperand_t *p = a;
 308         const memoperand_t *q = b;
 309
 310         return !(p->irn == q->irn && p->pos == q->pos);
 311 }
 312
 313 static keyval_t *
 314 set_find_keyval(set * set, const void * key)
 315 {
 316         keyval_t     query;
 317
 318         query.key = key;
 319         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 320 }
 321
 322 static keyval_t *
 323 set_insert_keyval(set * set, void * key, void * val)
 324 {
 325         keyval_t     query;
 326
 327         query.key = key;
 328         query.val = val;
 329         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 330 }
 331
 332 static defs_t *
 333 set_find_def(set * set, const ir_node * value)
 334 {
 335         defs_t     query;
 336
 337         query.value = value;
 338         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 339 }
 340
 341 static defs_t *
 342 set_insert_def(set * set, const ir_node * value)
 343 {
 344         defs_t     query;
 345
 346         query.value = value;
 347         query.spills = NULL;
 348         query.remats = NULL;
 349         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 350 }
 351
 352 static memoperand_t *
 353 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 354 {
 355         memoperand_t     query;
 356
 357         query.irn = irn;
 358         query.pos = pos;
 359         query.ilp = ilp;
 360         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 361 }
 362
 363 static memoperand_t *
 364 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 365 {
 366         memoperand_t     query;
 367
 368         query.irn = (ir_node*)irn;
 369         query.pos = pos;
 370         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 371 }
 372
 373
 374 static spill_t *
 375 set_find_spill(set * set, const ir_node * value)
 376 {
 377         spill_t     query;
 378
 379         query.irn = (ir_node*)value;
 380         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 381 }
 382
 383 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 384 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 385 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 386 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 387 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 388
 389 static int
 390 cmp_remat_info(const void *a, const void *b, size_t size)
 391 {
 392         const remat_info_t *p = a;
 393         const remat_info_t *q = b;
 394
 395         return !(p->irn == q->irn);
 396 }
 397
 398 static int
 399 cmp_defs(const void *a, const void *b, size_t size)
 400 {
 401         const defs_t *p = a;
 402         const defs_t *q = b;
 403
 404         return !(p->value == q->value);
 405 }
 406
 407 static int
 408 cmp_keyval(const void *a, const void *b, size_t size)
 409 {
 410         const keyval_t *p = a;
 411         const keyval_t *q = b;
 412
 413         return !(p->key == q->key);
 414 }
 415
 416 static double
 417 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 418 {
 419 #define FUDGE 0.001
 420 #ifndef EXECFREQ_LOOPDEPH
 421         return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
 422 #else
 423         if(is_Block(irn))
 424                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 425         else
 426                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 427 #endif
 428 }
 429
 430 static double
 431 get_cost(const spill_ilp_t * si, const ir_node * irn)
 432 {
 433         if(be_is_Spill(irn)) {
 434                 return opt_cost_spill;
 435         } else if(be_is_Reload(irn)){
 436                 return opt_cost_reload;
 437         } else {
 438                 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
 439         }
 440 }
 441
 442 /**
 443  * Checks, whether node and its operands have suitable reg classes
 444  */
 445 static INLINE int
 446 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 447 {
 448         int               n;
 449         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 450         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 451
 452 #if 0
 453         if(!remat)
 454                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 455 #endif
 456
 457         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 458                 ir_node        *op = get_irn_n(irn, n);
 459                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 460
 461 //              if(!remat)
 462 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 463         }
 464
 465         return remat;
 466 }
 467
 468 /**
 469  * Try to create a remat from @p op with destination value @p dest_value
 470  */
 471 static INLINE remat_t *
 472 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 473 {
 474         remat_t  *remat = NULL;
 475
 476 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 477 //              return NULL;
 478
 479         if(dest_value == op) {
 480                 const ir_node *proj = NULL;
 481
 482                 if(is_Proj(dest_value)) {
 483                         op = get_Proj_pred(op);
 484                         proj = dest_value;
 485                 }
 486
 487                 if(!is_rematerializable(si, op))
 488                         return NULL;
 489
 490                 remat = obstack_alloc(si->obst, sizeof(*remat));
 491                 remat->op = op;
 492                 remat->cost = get_cost(si, op);
 493                 remat->value = dest_value;
 494                 remat->proj = proj;
 495                 remat->inverse = 0;
 496         } else {
 497                 arch_inverse_t     inverse;
 498                 int                n;
 499
 500                 /* get the index of the operand we want to retrieve by the inverse op */
 501                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 502                         ir_node        *arg = get_irn_n(op, n);
 503
 504                         if(arg == dest_value) break;
 505                 }
 506                 if(n<0) return NULL;
 507
 508                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 509
 510                 /* else ask the backend to give an inverse op */
 511                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 512                         int   i;
 513
 514                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 515
 516                         assert(inverse.n > 0 && "inverse op should have at least one node");
 517
 518                         for(i=inverse.n-1; i>=0; --i) {
 519                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 520                         }
 521
 522                         if(inverse.n <= 2) {
 523                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 524                                 remat->op = inverse.nodes[0];
 525                                 remat->cost = inverse.costs;
 526                                 remat->value = dest_value;
 527                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 528                                 remat->inverse = 1;
 529
 530                                 assert(is_Proj(remat->proj));
 531                         } else {
 532                                 assert(0 && "I can not handle remats with more than 2 nodes");
 533                         }
 534                 }
 535         }
 536
 537         if(remat) {
 538                 if(remat->proj) {
 539                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 540                 } else {
 541                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 542                 }
 543         }
 544         return remat;
 545 }
 546
 547
 548 static INLINE void
 549 add_remat(const spill_ilp_t * si, const remat_t * remat)
 550 {
 551         remat_info_t    *remat_info,
 552                      query;
 553         int              n;
 554
 555         assert(remat->op);
 556         assert(remat->value);
 557
 558         query.irn = remat->value;
 559         query.remats = NULL;
 560         query.remats_by_operand = NULL;
 561         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 562
 563         if(remat_info->remats == NULL) {
 564                 remat_info->remats = new_pset(cmp_remat, 4096);
 565         }
 566         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 567
 568         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 569         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 570                 ir_node        *arg = get_irn_n(remat->op, n);
 571
 572                 query.irn = arg;
 573                 query.remats = NULL;
 574                 query.remats_by_operand = NULL;
 575                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 576
 577                 if(remat_info->remats_by_operand == NULL) {
 578                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 579                 }
 580                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 581         }
 582 }
 583
 584 static int
 585 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 586 {
 587         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 588         int                i = 0;
 589
 590         while(edge) {
 591                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 592                         ++i;
 593                 }
 594                 edge = get_irn_out_edge_next(irn, edge);
 595         }
 596
 597         return i;
 598 }
 599
 600 static int
 601 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 602 {
 603         int n;
 604         int ret = 0;
 605
 606         if(is_Proj(irn))
 607                 irn = get_Proj_pred(irn);
 608
 609         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 610                 const ir_node  *arg = get_irn_n(irn, n);
 611
 612                 if(has_reg_class(si, arg)) ++ret;
 613         }
 614
 615         return ret;
 616 }
 617
 618 static INLINE void
 619 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 620 {
 621         int      n;
 622         remat_t *remat;
 623
 624         if( has_reg_class(si, op)
 625         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 626         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 627         ) {
 628                 remat = get_remat_from_op(si, op, op);
 629                 if(remat) {
 630                         add_remat(si, remat);
 631                 }
 632         }
 633
 634         if(opt_remats == REMATS_ALL) {
 635                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 636                    for each arg */
 637                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 638                         ir_node        *arg = get_irn_n(op, n);
 639
 640                         if(has_reg_class(si, arg)) {
 641                                 /* try to get an inverse remat */
 642                                 remat = get_remat_from_op(si, arg, op);
 643                                 if(remat) {
 644                                         add_remat(si, remat);
 645                                 }
 646                         }
 647                 }
 648         }
 649 }
 650
 651 static INLINE int
 652 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 653 {
 654         ir_node *block;
 655         ir_node *def_block = get_nodes_block(val);
 656         int      ret;
 657
 658         if(val == pos)
 659                 return 0;
 660
 661         /* if pos is at end of a basic block */
 662         if(is_Block(pos)) {
 663                 ret = (pos == def_block || block_dominates(def_block, pos));
 664 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 665                 return ret;
 666         }
 667
 668         /* else if this is a normal operation */
 669         block = get_nodes_block(pos);
 670         if(block == def_block) {
 671                 if(!sched_is_scheduled(val)) return 1;
 672
 673                 ret = sched_comes_after(val, pos);
 674 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 675                 return ret;
 676         }
 677
 678         ret = block_dominates(def_block, block);
 679 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 680         return ret;
 681 }
 682
 683 static INLINE ir_node *
 684 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 685 {
 686     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 687 }
 688
 689 /**
 690  * Returns first non-Phi node of block @p bb
 691  */
 692 static INLINE ir_node *
 693 sched_block_first_nonphi(const ir_node * bb)
 694 {
 695         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 696 }
 697
 698 static int
 699 sched_skip_proj_predicator(const ir_node * irn, void * data)
 700 {
 701         return (is_Proj(irn));
 702 }
 703
 704 static INLINE ir_node *
 705 sched_next_nonproj(const ir_node * irn, int forward)
 706 {
 707         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 708 }
 709
 710 /**
 711  * Returns next operation node (non-Proj) after @p irn
 712  * or the basic block of this node
 713  */
 714 static INLINE ir_node *
 715 sched_next_op(const ir_node * irn)
 716 {
 717         ir_node *next = sched_next(irn);
 718
 719         if(is_Block(next))
 720                 return next;
 721
 722         return sched_next_nonproj(next, 1);
 723 }
 724
 725 /**
 726  * Returns previous operation node (non-Proj) before @p irn
 727  * or the basic block of this node
 728  */
 729 static INLINE ir_node *
 730 sched_prev_op(const ir_node * irn)
 731 {
 732         ir_node *prev = sched_prev(irn);
 733
 734         if(is_Block(prev))
 735                 return prev;
 736
 737         return sched_next_nonproj(prev, 0);
 738 }
 739
 740 static void
 741 sched_put_after(ir_node * insert, ir_node * irn)
 742 {
 743         if(is_Block(insert)) {
 744                 insert = sched_block_first_nonphi(insert);
 745         } else {
 746                 insert = sched_next_op(insert);
 747         }
 748         sched_add_before(insert, irn);
 749 }
 750
 751 static void
 752 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 753 {
 754   if(is_Block(insert)) {
 755           insert = sched_block_last_noncf(si, insert);
 756   } else {
 757           insert = sched_next_nonproj(insert, 0);
 758           insert = sched_prev(insert);
 759   }
 760   sched_add_after(insert, irn);
 761 }
 762
 763 /**
 764  * Tells you whether a @p remat can be placed before the irn @p pos
 765  */
 766 static INLINE int
 767 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 768 {
 769         const ir_node   *op = remat->op;
 770         const ir_node   *prev;
 771         int        n,
 772                            res = 1;
 773
 774         if(is_Block(pos)) {
 775                 prev = sched_block_last_noncf(si, pos);
 776                 prev = sched_next_nonproj(prev, 0);
 777         } else {
 778                 prev = sched_prev_op(pos);
 779         }
 780         /* do not remat if the rematted value is defined immediately before this op */
 781         if(prev == remat->op) {
 782                 return 0;
 783         }
 784
 785 #if 0
 786         /* this should be just fine, the following OP will be using this value, right? */
 787
 788         /* only remat AFTER the real definition of a value (?) */
 789         if(!value_is_defined_before(si, pos, remat->value)) {
 790 //              ir_fprintf(stderr, "error(not defined)");
 791                 return 0;
 792         }
 793 #endif
 794
 795         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 796                 const ir_node   *arg = get_irn_n(op, n);
 797
 798                 if(opt_no_enlarge_liveness) {
 799                         if(has_reg_class(si, arg) && live) {
 800                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 801                         } else {
 802                                 res &= value_is_defined_before(si, pos, arg);
 803                         }
 804                 } else {
 805                         res &= value_is_defined_before(si, pos, arg);
 806                 }
 807         }
 808
 809         return res;
 810 }
 811
 812 /**
 813  * Tells you whether a @p remat can be placed after the irn @p pos
 814  */
 815 static INLINE int
 816 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 817 {
 818         if(is_Block(pos)) {
 819                 pos = sched_block_first_nonphi(pos);
 820         } else {
 821                 pos = sched_next_op(pos);
 822         }
 823
 824         /* only remat AFTER the real definition of a value (?) */
 825         if(!value_is_defined_before(si, pos, remat->value)) {
 826                 return 0;
 827         }
 828
 829         return can_remat_before(si, remat, pos, live);
 830 }
 831
 832 /**
 833  * Collect potetially rematerializable OPs
 834  */
 835 static void
 836 walker_remat_collector(ir_node * irn, void * data)
 837 {
 838         spill_ilp_t    *si = data;
 839
 840         if(!is_Block(irn) && !is_Phi(irn)) {
 841                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 842                 get_remats_from_op(si, irn);
 843         }
 844 }
 845
 846 /**
 847  * Inserts a copy of @p irn before @p pos
 848  */
 849 static ir_node *
 850 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 851 {
 852         ir_node     *bb;
 853         ir_node     *copy;
 854
 855         bb = is_Block(pos)?pos:get_nodes_block(pos);
 856         copy = exact_copy(irn);
 857
 858         _set_phi_class(copy, NULL);
 859         set_nodes_block(copy, bb);
 860         sched_put_before(si, pos, copy);
 861
 862         return copy;
 863 }
 864
 865 /**
 866  * Inserts a copy of @p irn after @p pos
 867  */
 868 static ir_node *
 869 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 870 {
 871         ir_node     *bb;
 872         ir_node     *copy;
 873
 874         bb = is_Block(pos)?pos:get_nodes_block(pos);
 875         copy = exact_copy(irn);
 876
 877         _set_phi_class(copy, NULL);
 878         set_nodes_block(copy, bb);
 879         sched_put_after(pos, copy);
 880
 881         return copy;
 882 }
 883
 884 static ir_node *
 885 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 886 {
 887         char     buf[256];
 888
 889         if(can_remat_after(si, remat, pos, live)) {
 890                 ir_node         *copy,
 891                                                 *proj_copy;
 892                 op_t            *op;
 893
 894                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 895
 896                 copy = insert_copy_after(si, remat->op, pos);
 897
 898                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 899                 op = obstack_alloc(si->obst, sizeof(*op));
 900                 op->is_remat = 1;
 901                 op->attr.remat.remat = remat;
 902                 op->attr.remat.pre = 0;
 903                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 904
 905                 set_irn_link(copy, op);
 906                 pset_insert_ptr(si->all_possible_remats, copy);
 907                 if(remat->proj) {
 908                         proj_copy = insert_copy_after(si, remat->proj, copy);
 909                         set_irn_n(proj_copy, 0, copy);
 910                         set_irn_link(proj_copy, op);
 911                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 912                 } else {
 913                         proj_copy = NULL;
 914                 }
 915
 916                 return copy;
 917         }
 918
 919         return NULL;
 920 }
 921
 922 static ir_node *
 923 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 924 {
 925         char     buf[256];
 926
 927         if(can_remat_before(si, remat, pos, live)) {
 928                 ir_node         *copy,
 929                                                 *proj_copy;
 930                 op_t            *op;
 931
 932                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 933
 934                 copy = insert_copy_before(si, remat->op, pos);
 935
 936                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 937                 op = obstack_alloc(si->obst, sizeof(*op));
 938                 op->is_remat = 1;
 939                 op->attr.remat.remat = remat;
 940                 op->attr.remat.pre = 1;
 941                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 942
 943                 set_irn_link(copy, op);
 944                 pset_insert_ptr(si->all_possible_remats, copy);
 945                 if(remat->proj) {
 946                         proj_copy = insert_copy_after(si, remat->proj, copy);
 947                         set_irn_n(proj_copy, 0, copy);
 948                         set_irn_link(proj_copy, op);
 949                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 950                 } else {
 951                         proj_copy = NULL;
 952                 }
 953
 954                 return copy;
 955         }
 956
 957         return NULL;
 958 }
 959
 960 static int
 961 get_block_n_succs(const ir_node *block) {
 962         const ir_edge_t *edge;
 963
 964         assert(edges_activated(current_ir_graph));
 965
 966         edge = get_block_succ_first(block);
 967         if (! edge)
 968                 return 0;
 969
 970         edge = get_block_succ_next(block, edge);
 971         return edge ? 2 : 1;
 972 }
 973
 974 static int
 975 is_start_block(const ir_node * bb)
 976 {
 977         return get_irg_start_block(get_irn_irg(bb)) == bb;
 978 }
 979
 980 static int
 981 is_before_frame(const ir_node * bb, const ir_node * irn)
 982 {
 983         const ir_node  *frame  = get_irg_frame(get_irn_irg(bb));
 984
 985         if(is_start_block(bb) && sched_get_time_step(frame) >= sched_get_time_step(irn))
 986                 return 1;
 987         else
 988                 return 0;
 989 }
 990
 991 static int
 992 is_merge_edge(const ir_node * bb)
 993 {
 994         if(is_start_block(bb))
 995                 return 0;
 996
 997         if(opt_goodwin)
 998                 return get_block_n_succs(bb) == 1;
 999         else
1000                 return 1;
1001 }
1002
1003 static int
1004 is_diverge_edge(const ir_node * bb)
1005 {
1006         if(is_start_block(bb))
1007                 return 0;
1008
1009         if(opt_goodwin)
1010                 return get_Block_n_cfgpreds(bb) == 1;
1011         else
1012                 return 1;
1013 }
1014
1015 static void
1016 walker_regclass_copy_insertor(ir_node * irn, void * data)
1017 {
1018         spill_ilp_t    *si = data;
1019
1020         if(is_Phi(irn) && has_reg_class(si, irn)) {
1021                 int n;
1022
1023                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1024                         ir_node  *phi_arg = get_irn_n(irn, n);
1025                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1026
1027                         if(!has_reg_class(si, phi_arg)) {
1028                                 ir_node   *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1029                                 ir_node   *pos = sched_block_last_noncf(si, bb);
1030                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1031
1032                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1033                                 sched_add_after(pos, copy);
1034                                 set_irn_n(irn, n, copy);
1035
1036                                 op->is_remat = 0;
1037                                 op->attr.live_range.args.reloads = NULL;
1038                                 op->attr.live_range.ilp = ILP_UNDEF;
1039                                 set_irn_link(copy, op);
1040                         }
1041                 }
1042         }
1043 }
1044
1045
1046 /**
1047  * Insert (so far unused) remats into the irg to
1048  * recompute the potential liveness of all values
1049  */
1050 static void
1051 walker_remat_insertor(ir_node * bb, void * data)
1052 {
1053         spill_ilp_t    *si = data;
1054         spill_bb_t     *spill_bb;
1055         ir_node        *irn;
1056         int             n, i;
1057         pset           *live = pset_new_ptr_default();
1058
1059         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1060
1061         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1062                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1063
1064                 /* add remats at end of block */
1065                 if (has_reg_class(si, value)) {
1066                         pset_insert_ptr(live, value);
1067                 }
1068         }
1069
1070         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1071         set_irn_link(bb, spill_bb);
1072
1073         irn = sched_last(bb);
1074         while(!sched_is_end(irn)) {
1075                 ir_node   *next;
1076                 op_t      *op;
1077                 pset      *args;
1078                 ir_node   *arg;
1079                 pset      *remat_args;
1080
1081                 next = sched_prev(irn);
1082
1083                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1084
1085                 if(is_Phi(irn) || is_Proj(irn)) {
1086                         op_t      *op;
1087
1088                         if(has_reg_class(si, irn)) {
1089                                 pset_remove_ptr(live, irn);
1090                         }
1091
1092                         op = obstack_alloc(si->obst, sizeof(*op));
1093                         op->is_remat = 0;
1094                         op->attr.live_range.args.reloads = NULL;
1095                         op->attr.live_range.ilp = ILP_UNDEF;
1096                         set_irn_link(irn, op);
1097
1098                         irn = next;
1099                         continue;
1100                 }
1101
1102                 op = obstack_alloc(si->obst, sizeof(*op));
1103                 op->is_remat = 0;
1104                 op->attr.live_range.ilp = ILP_UNDEF;
1105                 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1106                 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1107                 set_irn_link(irn, op);
1108
1109                 args = pset_new_ptr_default();
1110
1111                 /* collect arguments of op */
1112                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1113                         ir_node        *arg = get_irn_n(irn, n);
1114
1115                         pset_insert_ptr(args, arg);
1116                 }
1117
1118                 /* set args of op already live in epilog */
1119                 pset_foreach(args, arg) {
1120                         if(has_reg_class(si, arg)) {
1121                                 pset_insert_ptr(live, arg);
1122                         }
1123                 }
1124                 /* delete defined value from live set */
1125                 if(has_reg_class(si, irn)) {
1126                         pset_remove_ptr(live, irn);
1127                 }
1128
1129
1130                 remat_args = pset_new_ptr_default();
1131
1132                 /* insert all possible remats before irn */
1133                 pset_foreach(args, arg) {
1134                         remat_info_t   *remat_info,
1135                                                     query;
1136                         remat_t        *remat;
1137
1138                         /* continue if the operand has the wrong reg class
1139                          */
1140                         if(!has_reg_class(si, arg))
1141                                 continue;
1142
1143                         query.irn = arg;
1144                         query.remats = NULL;
1145                         query.remats_by_operand = NULL;
1146                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1147
1148                         if(!remat_info) {
1149                                 continue;
1150                         }
1151
1152                         if(remat_info->remats) {
1153                                 pset_foreach(remat_info->remats, remat) {
1154                                         ir_node  *remat_irn = NULL;
1155
1156                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1157                                         if(opt_remat_while_live) {
1158                                                 if(pset_find_ptr(live, remat->value)) {
1159                                                         remat_irn = insert_remat_before(si, remat, irn, live);
1160                                                 }
1161                                         } else {
1162                                                 remat_irn = insert_remat_before(si, remat, irn, live);
1163                                         }
1164                                         if(remat_irn) {
1165                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1166                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1167
1168                                                         if(!has_reg_class(si, remat_arg)) continue;
1169
1170                                                         pset_insert_ptr(remat_args, remat_arg);
1171                                                 }
1172                                         }
1173                                 }
1174                         }
1175                 }
1176
1177                 /* now we add remat args to op's args because they could also die at this op */
1178                 pset_foreach(args,arg) {
1179                         if(pset_find_ptr(remat_args, arg)) {
1180                                 pset_remove_ptr(remat_args, arg);
1181                         }
1182                 }
1183                 pset_foreach(remat_args,arg) {
1184                         pset_insert_ptr(args, arg);
1185                 }
1186
1187                 /* insert all possible remats after irn */
1188                 pset_foreach(args, arg) {
1189                         remat_info_t   *remat_info,
1190                                                     query;
1191                         remat_t        *remat;
1192
1193                         /* continue if the operand has the wrong reg class */
1194                         if(!has_reg_class(si, arg))
1195                                 continue;
1196
1197                         query.irn = arg;
1198                         query.remats = NULL;
1199                         query.remats_by_operand = NULL;
1200                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1201
1202                         if(!remat_info) {
1203                                 continue;
1204                         }
1205
1206                         /* do not place post remats after jumps */
1207                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1208
1209                         if(remat_info->remats_by_operand) {
1210                                 pset_foreach(remat_info->remats_by_operand, remat) {
1211                                         /* do not insert remats producing the same value as one of the operands */
1212                                         if(!pset_find_ptr(args, remat->value)) {
1213                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1214                                                 if(opt_remat_while_live) {
1215                                                         if(pset_find_ptr(live, remat->value)) {
1216                                                                 insert_remat_after(si, remat, irn, live);
1217                                                         }
1218                                                 } else {
1219                                                         insert_remat_after(si, remat, irn, live);
1220                                                 }
1221                                         }
1222                                 }
1223                         }
1224                 }
1225
1226                 del_pset(remat_args);
1227                 del_pset(args);
1228                 irn = next;
1229         }
1230
1231         /* add remats at end if successor has multiple predecessors */
1232         if(is_merge_edge(bb)) {
1233                 pset     *live_out = pset_new_ptr_default();
1234                 ir_node  *value;
1235
1236                 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1237                         value = be_lv_get_irn(si->lv, bb, i);
1238
1239                         if (has_reg_class(si, value)) {
1240                                 pset_insert_ptr(live_out, value);
1241                         }
1242                 }
1243
1244                 /* add remats at end of block */
1245                 pset_foreach(live_out, value) {
1246                         remat_info_t   *remat_info,
1247                                                    query;
1248                         remat_t        *remat;
1249
1250                         query.irn = value;
1251                         query.remats = NULL;
1252                         query.remats_by_operand = NULL;
1253                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1254
1255                         if(remat_info && remat_info->remats) {
1256                                 pset_foreach(remat_info->remats, remat) {
1257                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1258
1259                                         insert_remat_before(si, remat, bb, live_out);
1260                                 }
1261                         }
1262                 }
1263                 del_pset(live_out);
1264         }
1265
1266         if(is_diverge_edge(bb)) {
1267                 pset     *live_in = pset_new_ptr_default();
1268                 ir_node  *value;
1269
1270                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1271                         value = be_lv_get_irn(si->lv, bb, i);
1272
1273                         if(has_reg_class(si, value)) {
1274                                 pset_insert_ptr(live_in, value);
1275                         }
1276                 }
1277                 sched_foreach(bb, value) {
1278                         if(!is_Phi(value)) break;
1279
1280                         if(has_reg_class(si, value)) {
1281                                 pset_insert_ptr(live_in, value);
1282                         }
1283                 }
1284
1285                 /* add remat2s at beginning of block */
1286                 pset_foreach(live_in, value) {
1287                         remat_info_t   *remat_info,
1288                                                    query;
1289                         remat_t        *remat;
1290
1291                         query.irn = value;
1292                         query.remats = NULL;
1293                         query.remats_by_operand = NULL;
1294                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1295
1296                         if(remat_info && remat_info->remats_by_operand) {
1297                                 pset_foreach(remat_info->remats_by_operand, remat) {
1298                                         DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1299
1300                                         /* put the remat here if all its args are available */
1301                                         insert_remat_after(si, remat, bb, live_in);
1302
1303                                 }
1304                         }
1305                 }
1306                 del_pset(live_in);
1307         }
1308 }
1309
1310 int
1311 can_be_copied(const ir_node * bb, const ir_node * irn)
1312 {
1313         assert(is_merge_edge(bb));
1314
1315         const ir_edge_t *edge = get_block_succ_first(bb);
1316         const ir_node   *next_bb = edge->src;
1317         int              pos = edge->pos;
1318         const ir_node   *phi;
1319
1320         sched_foreach(next_bb, phi) {
1321                 const ir_node  *phi_arg;
1322
1323                 if(!is_Phi(phi)) break;
1324
1325                 phi_arg = get_irn_n(phi, pos);
1326
1327                 if(phi_arg == irn) {
1328                         return 1;
1329                 }
1330         }
1331         return 0;
1332 }
1333
1334 /**
1335  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1336  */
1337 static void
1338 luke_endwalker(ir_node * bb, void * data)
1339 {
1340         spill_ilp_t    *si = (spill_ilp_t*)data;
1341         pset           *live;
1342         pset           *use_end;
1343         char            buf[256];
1344         ilp_cst_t       cst;
1345         ir_node        *irn;
1346         spill_bb_t     *spill_bb = get_irn_link(bb);
1347         int             i;
1348
1349
1350         live = pset_new_ptr_default();
1351         use_end = pset_new_ptr_default();
1352
1353         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1354                 irn = be_lv_get_irn(si->lv, bb, i);
1355                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1356                         op_t      *op;
1357
1358                         pset_insert_ptr(live, irn);
1359                         op = get_irn_link(irn);
1360                         assert(!op->is_remat);
1361                 }
1362         }
1363
1364         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1365         /* their reg_out must always be set */
1366         sched_foreach_reverse(bb, irn) {
1367                 int   n;
1368
1369                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1370
1371                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1372                         ir_node        *irn_arg = get_irn_n(irn, n);
1373
1374                         if(has_reg_class(si, irn_arg)) {
1375                                 pset_insert_ptr(use_end, irn_arg);
1376                         }
1377                 }
1378         }
1379
1380         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1381         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1382         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1383
1384         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1385
1386         /* if this is a merge edge we can reload at the end of this block */
1387         if(is_merge_edge(bb)) {
1388                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1389         } else if(pset_count(use_end)){
1390                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1391         } else {
1392                 spill_bb->reloads = NULL;
1393         }
1394
1395         pset_foreach(live,irn) {
1396                 spill_t     query,
1397                                         *spill;
1398                 double      spill_cost;
1399                 int         default_spilled;
1400
1401
1402                 /* handle values used by control flow nodes later separately */
1403                 if(pset_find_ptr(use_end, irn)) continue;
1404
1405                 query.irn = irn;
1406                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1407
1408                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1409
1410                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1411                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1412                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1413
1414                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1415                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1416
1417                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1418                 /* by default spill value right after definition */
1419                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1420                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1421
1422                 if(is_merge_edge(bb)) {
1423                         ilp_var_t   reload;
1424                         ilp_cst_t   rel_cst;
1425
1426                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1427                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1428                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1429
1430                         /* reload <= mem_out */
1431                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1432                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1433                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1434                 }
1435
1436                 spill->reg_in = ILP_UNDEF;
1437                 spill->mem_in = ILP_UNDEF;
1438         }
1439
1440         pset_foreach(use_end,irn) {
1441                 spill_t     query,
1442                                         *spill;
1443                 double      spill_cost;
1444                 ilp_cst_t   end_use_req,
1445                                         rel_cst;
1446                 ilp_var_t   reload;
1447                 int         default_spilled;
1448
1449                 query.irn = irn;
1450                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1451
1452                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1453
1454                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1455                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1456                 /* if irn is used at the end of the block, then it is live anyway */
1457                 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1458
1459                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1460                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1461
1462                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1463                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1464                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1465
1466                 /* reload for use be control flow op */
1467                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1468                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1469                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1470
1471                 /* reload <= mem_out */
1472                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1473                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1474                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1475
1476                 spill->reg_in = ILP_UNDEF;
1477                 spill->mem_in = ILP_UNDEF;
1478
1479                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1480                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1481                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1482         }
1483
1484         del_pset(live);
1485         del_pset(use_end);
1486 }
1487
1488 static ir_node *
1489 next_post_remat(const ir_node * irn)
1490 {
1491         op_t      *op;
1492     ir_node   *next;
1493
1494         if(is_Block(irn)) {
1495                 next = sched_block_first_nonphi(irn);
1496         } else {
1497                 next = sched_next_op(irn);
1498         }
1499
1500         if(sched_is_end(next))
1501                 return NULL;
1502
1503         op = get_irn_link(next);
1504         if(op->is_remat && !op->attr.remat.pre) {
1505                 return next;
1506         }
1507
1508         return NULL;
1509 }
1510
1511
1512 static ir_node *
1513 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1514 {
1515         op_t      *op;
1516         ir_node   *ret;
1517
1518         if(is_Block(irn)) {
1519                 ret = sched_block_last_noncf(si, irn);
1520                 ret = sched_next(ret);
1521                 ret = sched_prev_op(ret);
1522         } else {
1523                 ret = sched_prev_op(irn);
1524         }
1525
1526         if(sched_is_end(ret) || is_Phi(ret))
1527                 return NULL;
1528
1529         op = (op_t*)get_irn_link(ret);
1530         if(op->is_remat && op->attr.remat.pre) {
1531                 return ret;
1532         }
1533
1534         return NULL;
1535 }
1536
1537 /**
1538  * Find a remat of value @p value in the epilog of @p pos
1539  */
1540 static ir_node *
1541 find_post_remat(const ir_node * value, const ir_node * pos)
1542 {
1543         while((pos = next_post_remat(pos)) != NULL) {
1544                 op_t   *op;
1545
1546                 op = get_irn_link(pos);
1547                 assert(op->is_remat && !op->attr.remat.pre);
1548
1549                 if(op->attr.remat.remat->value == value)
1550                         return (ir_node*)pos;
1551
1552 #if 0
1553         const ir_edge_t *edge;
1554                 foreach_out_edge(pos, edge) {
1555                         ir_node   *proj = get_edge_src_irn(edge);
1556                         assert(is_Proj(proj));
1557                 }
1558 #endif
1559
1560         }
1561
1562         return NULL;
1563 }
1564
1565 static spill_t *
1566 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1567 {
1568         spill_bb_t  *spill_bb = get_irn_link(bb);
1569         spill_t     *spill,
1570                                  query;
1571         char         buf[256];
1572         int          default_spilled;
1573
1574         query.irn = irn;
1575         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1576         if(!spill) {
1577                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1578
1579                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1580
1581                 spill->reg_out = ILP_UNDEF;
1582                 spill->reg_in  = ILP_UNDEF;
1583                 spill->mem_in  = ILP_UNDEF;
1584
1585                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1586                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1587
1588                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1589                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1590                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1591         }
1592
1593         return spill;
1594 }
1595
1596 static void
1597 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1598 {
1599         ir_node        *irn;
1600         int i;
1601
1602         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1603                 irn = be_lv_get_irn(si->lv, bb, i);
1604
1605                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1606                         pset_insert_ptr(live, irn);
1607                 }
1608         }
1609
1610         irn = sched_last(bb);
1611
1612         /* all values eaten by control flow operations are also live until the end of the block */
1613         sched_foreach_reverse(bb, irn) {
1614                 int  i;
1615
1616                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1617
1618                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1619                         ir_node *arg = get_irn_n(irn,i);
1620
1621                         if(has_reg_class(si, arg)) {
1622                                 pset_insert_ptr(live, arg);
1623                         }
1624                 }
1625         }
1626 }
1627
1628 /**
1629  *  Inserts ILP-constraints and variables for memory copying before the given position
1630  */
1631 static void
1632 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1633 {
1634         const ir_node    *succ;
1635         const ir_edge_t  *edge;
1636         spill_bb_t       *spill_bb = get_irn_link(block);
1637         ir_node          *phi;
1638         int               pos;
1639         ilp_cst_t         cst;
1640         ilp_var_t         copyreg;
1641         char              buf[256];
1642         ir_node          *tmp;
1643
1644
1645         assert(edges_activated(current_ir_graph));
1646
1647         edge = get_block_succ_first(block);
1648         if(!edge) return;
1649
1650         succ = edge->src;
1651         pos = edge->pos;
1652
1653         edge = get_block_succ_next(block, edge);
1654         /* next block can only contain phis, if this is a merge edge */
1655         if(edge) return;
1656
1657         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1658         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1659
1660         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1661         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1662
1663         pset_foreach(live, tmp) {
1664                 spill_t  *spill;
1665 #if 0
1666                 op_t  *op = get_irn_link(irn);
1667                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1668 #endif
1669                 spill = set_find_spill(spill_bb->ilp, tmp);
1670                 assert(spill);
1671
1672                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1673         }
1674         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1675
1676         sched_foreach(succ, phi) {
1677                 const ir_node  *to_copy;
1678                 op_t           *to_copy_op;
1679                 spill_t        *to_copy_spill;
1680                 op_t           *phi_op = get_irn_link(phi);
1681                 ilp_var_t       reload = ILP_UNDEF;
1682
1683
1684                 if(!is_Phi(phi)) break;
1685                 if(!has_reg_class(si, phi)) continue;
1686
1687                 to_copy = get_irn_n(phi, pos);
1688
1689                 to_copy_op = get_irn_link(to_copy);
1690
1691                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1692                 assert(to_copy_spill);
1693
1694                 if(spill_bb->reloads) {
1695                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1696
1697                         if(keyval) {
1698                                 reload = PTR_TO_INT(keyval->val);
1699                         }
1700                 }
1701
1702                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1703                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1704
1705                 /* copy - reg_out - reload - remat - live_range <= 0 */
1706                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1707                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1708                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1709                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1710                 foreach_pre_remat(si, block, tmp) {
1711                         op_t     *remat_op = get_irn_link(tmp);
1712                         if(remat_op->attr.remat.remat->value == to_copy) {
1713                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1714                         }
1715                 }
1716
1717                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1718                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1719
1720                 /* copy - reg_out - copyreg <= 0 */
1721                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1722                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1723                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1724         }
1725 }
1726
1727
1728 /**
1729  * Walk all irg blocks and emit this ILP
1730  */
1731 static void
1732 luke_blockwalker(ir_node * bb, void * data)
1733 {
1734         spill_ilp_t    *si = (spill_ilp_t*)data;
1735         ir_node        *irn;
1736         pset           *live;
1737         char            buf[256];
1738         ilp_cst_t       cst;
1739         spill_bb_t     *spill_bb = get_irn_link(bb);
1740         ir_node        *tmp;
1741         spill_t        *spill;
1742         pset           *defs = pset_new_ptr_default();
1743         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1744
1745
1746         live = pset_new_ptr_default();
1747
1748         /****************************************
1749          *      B A S I C  B L O C K  E N D
1750          ***************************************/
1751
1752
1753         /* init live values at end of block */
1754         get_live_end(si, bb, live);
1755
1756         pset_foreach(live, irn) {
1757                 op_t           *op;
1758                 ilp_var_t       reload = ILP_UNDEF;
1759
1760                 spill = set_find_spill(spill_bb->ilp, irn);
1761                 assert(spill);
1762
1763                 if(spill_bb->reloads) {
1764                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1765
1766                         if(keyval) {
1767                                 reload = PTR_TO_INT(keyval->val);
1768                         }
1769                 }
1770
1771                 op = get_irn_link(irn);
1772                 assert(!op->is_remat);
1773
1774                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1775                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1776                 op->attr.live_range.op = bb;
1777
1778                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1779                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1780
1781                 /* reg_out - reload - remat - live_range <= 0 */
1782                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1783                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1784                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1785                 foreach_pre_remat(si, bb, tmp) {
1786                         op_t     *remat_op = get_irn_link(tmp);
1787                         if(remat_op->attr.remat.remat->value == irn) {
1788                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1789                         }
1790                 }
1791                 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1792                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1793
1794                 /* value may only die at bb end if it is used for a mem copy */
1795                 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1796                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1797                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1798                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1799                 foreach_pre_remat(si, bb, tmp) {
1800                         op_t     *remat_op = get_irn_link(tmp);
1801                         if(remat_op->attr.remat.remat->value == irn) {
1802                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1803                         }
1804                 }
1805                 if(is_merge_edge(bb)) {
1806                         const ir_edge_t *edge = get_block_succ_first(bb);
1807                         const ir_node   *next_bb = edge->src;
1808                         int              pos = edge->pos;
1809                         const ir_node   *phi;
1810
1811                         sched_foreach(next_bb, phi) {
1812                                 const ir_node  *phi_arg;
1813
1814                                 if(!is_Phi(phi)) break;
1815
1816                                 phi_arg = get_irn_n(phi, pos);
1817
1818                                 if(phi_arg == irn) {
1819                                         op_t      *phi_op = get_irn_link(phi);
1820                                         ilp_var_t  copy = phi_op->attr.live_range.args.copies[pos];
1821
1822                                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1823                                 }
1824                         }
1825                 }
1826         }
1827
1828         if(opt_memcopies)
1829                 insert_mem_copy_position(si, live, bb);
1830
1831         /*
1832          * start new live ranges for values used by remats at end of block
1833          * and assure the remat args are available
1834          */
1835         foreach_pre_remat(si, bb, tmp) {
1836                 op_t     *remat_op = get_irn_link(tmp);
1837                 int       n;
1838
1839                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1840                         ir_node        *remat_arg = get_irn_n(tmp, n);
1841                         op_t           *arg_op = get_irn_link(remat_arg);
1842                         ilp_var_t       prev_lr;
1843
1844                         if(!has_reg_class(si, remat_arg)) continue;
1845
1846                         /* if value is becoming live through use by remat */
1847                         if(!pset_find_ptr(live, remat_arg)) {
1848                                 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1849                                 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1850
1851                                 arg_op->attr.live_range.ilp = prev_lr;
1852                                 arg_op->attr.live_range.op = bb;
1853
1854                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1855
1856                                 pset_insert_ptr(live, remat_arg);
1857                                 add_to_spill_bb(si, bb, remat_arg);
1858                         }
1859
1860                         /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1861                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1862                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1863
1864                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1865                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1866
1867                         /* use reload placed for this argument */
1868                         if(spill_bb->reloads) {
1869                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1870
1871                                 if(keyval) {
1872                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1873
1874                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1875                                 }
1876                         }
1877                 }
1878         }
1879         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1880
1881
1882
1883
1884         /**************************************
1885          *    B A S I C  B L O C K  B O D Y
1886          **************************************/
1887
1888         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1889                 op_t       *op;
1890                 op_t       *tmp_op;
1891                 int         n,
1892                                         u = 0,
1893                                         d = 0;
1894                 ilp_cst_t       check_pre,
1895                                         check_post;
1896                 set        *args;
1897                 pset       *used;
1898                 pset       *remat_defs;
1899                 keyval_t   *keyval;
1900                 ilp_cst_t   one_memoperand;
1901
1902                 /* iterate only until first phi */
1903                 if(is_Phi(irn))
1904                         break;
1905
1906                 op = get_irn_link(irn);
1907                 /* skip remats */
1908                 if(op->is_remat) continue;
1909                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1910
1911                 /* collect defined values */
1912                 if(has_reg_class(si, irn)) {
1913                         pset_insert_ptr(defs, irn);
1914                 }
1915
1916                 /* skip projs */
1917                 if(is_Proj(irn)) continue;
1918
1919                 /*
1920                  * init set of irn's arguments
1921                  * and all possibly used values around this op
1922                  * and values defined by post remats
1923                  */
1924                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1925                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1926                 remat_defs = pset_new_ptr(pset_count(live));
1927
1928                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1929                         ir_node        *irn_arg = get_irn_n(irn, n);
1930                         if(has_reg_class(si, irn_arg)) {
1931                                 set_insert_keyval(args, irn_arg, (void*)n);
1932                                 pset_insert_ptr(used, irn_arg);
1933                         }
1934                 }
1935                 foreach_post_remat(irn, tmp) {
1936                         op_t    *remat_op = get_irn_link(tmp);
1937
1938                         pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1939
1940                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1941                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1942                                 if(has_reg_class(si, remat_arg)) {
1943                                         pset_insert_ptr(used, remat_arg);
1944                                 }
1945                         }
1946                 }
1947                 foreach_pre_remat(si, irn, tmp) {
1948                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1949                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1950                                 if(has_reg_class(si, remat_arg)) {
1951                                         pset_insert_ptr(used, remat_arg);
1952                                 }
1953                         }
1954                 }
1955
1956                 /**********************************
1957                  *   I N  E P I L O G  O F  irn
1958                  **********************************/
1959
1960                 /* ensure each dying value is used by only one post remat */
1961                 pset_foreach(used, tmp) {
1962                         ir_node     *value = tmp;
1963                         op_t        *value_op = get_irn_link(value);
1964                         ir_node     *remat;
1965                         int          n_remats = 0;
1966
1967                         cst = ILP_UNDEF;
1968                         foreach_post_remat(irn, remat) {
1969                                 op_t  *remat_op = get_irn_link(remat);
1970
1971                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1972                                         ir_node   *remat_arg = get_irn_n(remat, n);
1973
1974                                         /* if value is used by this remat add it to constraint */
1975                                         if(remat_arg == value) {
1976                                                 if(n_remats == 0) {
1977                                                         /* sum remat2s <= 1 + n_remats*live_range */
1978                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1979                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1980                                                 }
1981
1982                                                 n_remats++;
1983                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1984                                                 break;
1985                                         }
1986                                 }
1987                         }
1988
1989                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1990                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1991                         }
1992                 }
1993
1994         /* ensure at least one value dies at post remat */
1995         foreach_post_remat(irn, tmp) {
1996             op_t     *remat_op = get_irn_link(tmp);
1997             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
1998             ir_node  *remat_arg;
1999
2000             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2001                 remat_arg = get_irn_n(tmp, n);
2002
2003                 if(has_reg_class(si, remat_arg)) {
2004
2005                     /* does arg always die at this op? */
2006                     if(!pset_find_ptr(live, remat_arg))
2007                         goto skip_one_must_die;
2008
2009                     pset_insert_ptr(remat_args, remat_arg);
2010                 }
2011             }
2012
2013             /* remat + \sum live_range(remat_arg) <= |args| */
2014             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2015             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2016             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2017
2018             pset_foreach(remat_args, remat_arg) {
2019                 op_t  *arg_op = get_irn_link(remat_arg);
2020
2021                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2022             }
2023
2024 skip_one_must_die:
2025             del_pset(remat_args);
2026         }
2027
2028                 /* new live ranges for values from L\U defined by post remats */
2029                 pset_foreach(live, tmp) {
2030                         ir_node     *value = tmp;
2031                         op_t        *value_op = get_irn_link(value);
2032
2033                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2034                                 ilp_var_t    prev_lr = ILP_UNDEF;
2035                                 ir_node     *remat;
2036
2037                                 if(pset_find_ptr(remat_defs, value)) {
2038
2039                                         /* next_live_range <= prev_live_range + sum remat2s */
2040                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2041                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2042
2043                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2044                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2045
2046                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2047                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2048
2049                                         foreach_post_remat(irn, remat) {
2050                                                 op_t        *remat_op = get_irn_link(remat);
2051
2052                                                 /* if value is being rematerialized by this remat */
2053                                                 if(value == remat_op->attr.remat.remat->value) {
2054                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2055                                                 }
2056                                         }
2057
2058                                         value_op->attr.live_range.ilp = prev_lr;
2059                                         value_op->attr.live_range.op = irn;
2060                                 }
2061                         }
2062                 }
2063
2064                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2065                 foreach_post_remat(irn, tmp) {
2066                         op_t        *remat_op = get_irn_link(tmp);
2067                         int          n;
2068
2069                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2070                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2071                                 op_t           *arg_op = get_irn_link(remat_arg);
2072
2073                                 if(!has_reg_class(si, remat_arg)) continue;
2074
2075                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2076                                 if(!pset_find_ptr(used, remat_arg)) {
2077                                         /* remat <= live_range(remat_arg) */
2078                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2079                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2080
2081                                         /* if value is becoming live through use by remat2 */
2082                                         if(!pset_find_ptr(live, remat_arg)) {
2083                                                 ilp_var_t     lr;
2084
2085                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2086                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2087
2088                                                 arg_op->attr.live_range.ilp = lr;
2089                                                 arg_op->attr.live_range.op = irn;
2090
2091                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2092
2093                                                 pset_insert_ptr(live, remat_arg);
2094                                                 add_to_spill_bb(si, bb, remat_arg);
2095                                         }
2096
2097                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2098                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2099                                 }
2100                         }
2101                 }
2102
2103                 d = pset_count(defs);
2104                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2105
2106                 /* count how many regs irn needs for arguments */
2107                 u = set_count(args);
2108
2109
2110                 /* check the register pressure in the epilog */
2111                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2112                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2113                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2114
2115                 /* add L\U' to check_post */
2116                 pset_foreach(live, tmp) {
2117                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2118                                 /* if a live value is not used by irn */
2119                                 tmp_op = get_irn_link(tmp);
2120                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2121                         }
2122                 }
2123
2124                 /***********************************************************
2125                  *  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
2126                  **********************************************************/
2127
2128
2129                 pset_foreach(used, tmp) {
2130                         ilp_var_t       prev_lr;
2131                         ilp_var_t       post_use;
2132                         int             p = 0;
2133                         spill_t        *spill;
2134                         ir_node        *arg = tmp;
2135                         op_t           *arg_op = get_irn_link(arg);
2136                         ir_node        *remat;
2137
2138                         spill = add_to_spill_bb(si, bb, arg);
2139
2140                         /* new live range for each used value */
2141                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2142                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0);
2143
2144                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2145                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2146                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2147
2148                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2149
2150                         /* arg is live throughout epilog if the next live_range is in a register */
2151                         if(pset_find_ptr(live, arg)) {
2152                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2153
2154                                 /* post_use >= next_lr + remat */
2155                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2156                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2157                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2158                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2159
2160                         }
2161
2162                         /* if value is not an arg of op and not possibly defined by post remat
2163                          * then it may only die and not become live
2164                          */
2165                         if(!set_find_keyval(args, arg)) {
2166                                 /* post_use <= prev_lr */
2167                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2168                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2169                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2170                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2171
2172                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2173                                         /* next_lr <= prev_lr */
2174                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2175                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2176                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2177                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2178                                 }
2179                         }
2180
2181
2182                         /* forall post remat which use arg add a similar cst */
2183                         foreach_post_remat(irn, remat) {
2184                                 int      n;
2185
2186                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2187                                         ir_node    *remat_arg = get_irn_n(remat, n);
2188                                         op_t       *remat_op = get_irn_link(remat);
2189
2190                                         if(remat_arg == arg) {
2191                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2192
2193                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2194                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2195                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2196                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2197                                         }
2198                                 }
2199                         }
2200
2201                         if(opt_memoperands) {
2202                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2203                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2204                                                 ilp_var_t       memoperand;
2205
2206                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2207                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2208                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2209
2210                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2211                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2212
2213                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2214                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2215                                         }
2216                                 }
2217                         }
2218
2219                         /* new live range begins for each used value */
2220                         arg_op->attr.live_range.ilp = prev_lr;
2221                         arg_op->attr.live_range.op = irn;
2222
2223                         pset_insert_ptr(live, arg);
2224                 }
2225
2226                 /* just to be sure */
2227                 check_post = ILP_UNDEF;
2228
2229
2230
2231
2232                 /******************
2233                  *   P R O L O G
2234                  ******************/
2235
2236                 /* check the register pressure in the prolog */
2237                 /* sum_{L\U} lr <= k - |U| */
2238                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2239                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2240
2241                 /* for the prolog remove defined values from the live set */
2242                 pset_foreach(defs, tmp) {
2243                         pset_remove_ptr(live, tmp);
2244                 }
2245
2246                 if(opt_memoperands) {
2247                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2248                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2249                 }
2250
2251                 /***********************************************************
2252                  *  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
2253                  **********************************************************/
2254
2255
2256                 set_foreach(args, keyval) {
2257                         spill_t          *spill;
2258                         const ir_node    *arg = keyval->key;
2259                         int               i = PTR_TO_INT(keyval->val);
2260                         op_t             *arg_op = get_irn_link(arg);
2261                         ilp_cst_t         requirements;
2262                         int               n_memoperands;
2263
2264                         spill = set_find_spill(spill_bb->ilp, arg);
2265                         assert(spill);
2266
2267                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2268                         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);
2269
2270                         /* reload <= mem_out */
2271                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2272                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2273                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2274                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2275
2276                         /* requirement: arg must be in register for use */
2277                         /* reload + remat + live_range == 1 */
2278                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2279                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2280
2281                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2282                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2283                         foreach_pre_remat(si, irn, tmp) {
2284                                 op_t     *remat_op = get_irn_link(tmp);
2285                                 if(remat_op->attr.remat.remat->value == arg) {
2286                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2287                                 }
2288                         }
2289
2290                         if(opt_memoperands) {
2291                                 n_memoperands = 0;
2292                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2293                                         if(get_irn_n(irn, n) == arg) {
2294                                                 n_memoperands++;
2295                                         }
2296                                 }
2297                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2298                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2299                                                 memoperand_t  *memoperand;
2300                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2301
2302                                                 /* memoperand <= mem_out */
2303                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2304                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2305                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2306                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2307
2308                                                 /* the memoperand is only sufficient if it is used once by the op */
2309                                                 if(n_memoperands == 1)
2310                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2311
2312                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2313
2314                                                 /* we have one more free register if we use a memory operand */
2315                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2316                                         }
2317                                 }
2318                         }
2319                 }
2320
2321                 /* iterate over L\U */
2322                 pset_foreach(live, tmp) {
2323                         if(!set_find_keyval(args, tmp)) {
2324                                 /* if a live value is not used by irn */
2325                                 tmp_op = get_irn_link(tmp);
2326                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2327                         }
2328                 }
2329
2330
2331                 /* requirements for remats */
2332                 foreach_pre_remat(si, irn, tmp) {
2333                         op_t        *remat_op = get_irn_link(tmp);
2334                         int          n;
2335
2336                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2337                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2338                                 op_t           *arg_op = get_irn_link(remat_arg);
2339
2340                                 if(!has_reg_class(si, remat_arg)) continue;
2341
2342                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2343                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2344                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2345
2346                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2347                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2348
2349                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2350                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2351                                         int    index = (int)keyval->val;
2352
2353                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2354                                 }
2355                         }
2356                 }
2357
2358
2359
2360
2361                 /*************************
2362                  *  D O N E  W I T H  O P
2363                  *************************/
2364
2365                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2366
2367                 pset_foreach(live, tmp) {
2368                         assert(has_reg_class(si, tmp));
2369                 }
2370
2371                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2372                         ir_node        *arg = get_irn_n(irn, n);
2373
2374                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2375                 }
2376
2377                 del_pset(remat_defs);
2378                 del_pset(used);
2379                 del_set(args);
2380                 del_pset(defs);
2381                 defs = pset_new_ptr_default();
2382         }
2383
2384
2385
2386         /***************************************
2387          *   B E G I N N I N G  O F  B L O C K
2388          ***************************************/
2389
2390
2391         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2392         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2393
2394         pset_foreach(live, irn) {
2395                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2396         }
2397
2398         /* construct mem_outs for all values */
2399
2400         set_foreach(spill_bb->ilp, spill) {
2401                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2402                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2403
2404                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2405                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2406
2407                 if(pset_find_ptr(live, spill->irn)) {
2408                         int default_spilled;
2409                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2410
2411                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2412                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2413                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2414                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2415
2416                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2417                                 int   n;
2418                                 op_t *op = get_irn_link(spill->irn);
2419
2420                                 /* do we have to copy a phi argument? */
2421                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2422                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2423
2424                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2425                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2426                                         double          freq=0.0;
2427                                         int             m;
2428                                         ilp_var_t       var;
2429
2430
2431                                         /* argument already done? */
2432                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2433
2434                                         /* get sum of execution frequencies of blocks with the same phi argument */
2435                                         for(m=n; m>=0; --m) {
2436                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2437
2438                                                 if(arg==arg2) {
2439                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2440                                                 }
2441                                         }
2442
2443                                         /* copies are not for free */
2444                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2445                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2446
2447                                         for(m=n; m>=0; --m) {
2448                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2449
2450                                                 if(arg==arg2) {
2451                                                         op->attr.live_range.args.copies[m] = var;
2452                                                 }
2453                                         }
2454
2455 #if 0
2456                                         /* copy <= mem_in */
2457                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2458                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2459                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2460                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2461 #endif
2462                                 }
2463                         }
2464                 }
2465         }
2466
2467         foreach_post_remat(bb, tmp) {
2468                 int         n;
2469                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2470                 op_t       *remat_op = get_irn_link(tmp);
2471                 ir_node    *remat_arg;
2472
2473                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2474                         remat_arg = get_irn_n(tmp, n);
2475                         if(has_reg_class(si, remat_arg)) {
2476                                 pset_insert_ptr(remat_args, remat_arg);
2477                         }
2478                 }
2479                 assert(pset_count(remat_args) > 0 && "post remats should have at least one arg");
2480
2481                 /* remat + \sum live_range(remat_arg) <= |args| */
2482                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2483                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2484                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2485
2486                 pset_foreach(remat_args, remat_arg) {
2487                         /* if value is becoming live through use by remat2 */
2488                         if(!pset_find_ptr(live, remat_arg)) {
2489                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2490                                 ilp_cst_t   nomem;
2491
2492                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2493
2494                                 pset_insert_ptr(live, remat_arg);
2495                                 spill = add_to_spill_bb(si, bb, remat_arg);
2496                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2497
2498                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2499                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2500                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2501                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2502                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2503
2504
2505                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2506                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2507                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2508
2509                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2510                                 if(spill_bb->reloads) {
2511                                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
2512
2513                                         if(keyval) {
2514                                                 ilp_var_t reload = PTR_TO_INT(keyval->val);
2515                                                 lpp_set_factor_fast(si->lpp, nomem, reload, 1.0);
2516                                         }
2517                                 }
2518                         } else {
2519                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2520                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2521                         }
2522                 }
2523                 del_pset(remat_args);
2524         }
2525
2526         /* L\U is empty at bb start */
2527         /* arg is live throughout epilog if it is reg_in into this block */
2528
2529         /* check the register pressure at the beginning of the block
2530          * including remats
2531          */
2532         /* reg_in entspricht post_use */
2533
2534         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2535         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2536
2537         pset_foreach(live, irn) {
2538         ilp_cst_t  nospill;
2539
2540                 spill = set_find_spill(spill_bb->ilp, irn);
2541                 assert(spill);
2542
2543                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2544                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2545
2546                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2547
2548                 /* spill + mem_in <= 1 */
2549                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2550                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2551
2552                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2553                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2554
2555         } /* post_remats are NOT included in register pressure check because
2556            they do not increase regpressure */
2557
2558         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2559         pset_foreach(live, irn) {
2560                 int          p = 0,
2561                                          n;
2562
2563                 spill = set_find_spill(spill_bb->ilp, irn);
2564                 assert(spill && spill->irn == irn);
2565
2566                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2567                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2568                                 ilp_cst_t       mem_in,
2569                                                                 reg_in;
2570                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2571                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2572                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2573                                 spill_t        *spill_p;
2574                                 op_t           *op = get_irn_link(irn);
2575
2576                                 /* although the phi is in the right regclass one or more of
2577                                  * its arguments can be in a different one or at least to
2578                                  * ignore
2579                                  */
2580                                 if(has_reg_class(si, phi_arg)) {
2581                                         /* mem_in < mem_out_arg + copy */
2582                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2583                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2584
2585                                         /* reg_in < reg_out_arg */
2586                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2587                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2588
2589                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2590                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2591
2592                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2593                                         assert(spill_p);
2594
2595                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2596                                         if(opt_memcopies)
2597                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2598
2599                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2600                                 }
2601                         }
2602                 } else {
2603                         /* else assure the value arrives on all paths in the same resource */
2604
2605                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2606                                 ilp_cst_t       mem_in,
2607                                                                 reg_in;
2608                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2609                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2610                                 spill_t        *spill_p;
2611
2612                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2613                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2614                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2615                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2616
2617                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2618                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2619
2620                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2621                                 assert(spill_p);
2622
2623                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2624                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2625                         }
2626                 }
2627         }
2628
2629         foreach_post_remat(bb, tmp) {
2630                 int         n;
2631
2632                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2633                         ir_node    *remat_arg = get_irn_n(tmp, n);
2634                         op_t       *remat_op = get_irn_link(tmp);
2635
2636                         if(!has_reg_class(si, remat_arg)) continue;
2637
2638                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2639                         assert(spill);
2640
2641                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2642                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2643                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2644                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2645                 }
2646         }
2647
2648         pset_foreach(live, irn) {
2649                 const op_t      *op = get_irn_link(irn);
2650                 const ir_node   *remat;
2651                 int              n_remats = 0;
2652
2653                 cst = ILP_UNDEF;
2654
2655                 foreach_post_remat(bb, remat) {
2656                         int   n;
2657
2658                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2659                                 const ir_node  *arg = get_irn_n(remat, n);
2660
2661                                 if(arg == irn) {
2662                                         const op_t   *remat_op = get_irn_link(remat);
2663
2664                                         if(cst == ILP_UNDEF) {
2665                                                 /* sum remat2s <= 1 + n_remats*live_range */
2666                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2667                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2668                                         }
2669                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2670                                         ++n_remats;
2671                                         break;
2672                                 }
2673                         }
2674                 }
2675                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2676                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2677                 }
2678         }
2679
2680         /* first live ranges from reg_ins */
2681         pset_foreach(live, irn) {
2682                 op_t      *op = get_irn_link(irn);
2683
2684                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2685
2686                         spill = set_find_spill(spill_bb->ilp, irn);
2687                         assert(spill && spill->irn == irn);
2688
2689                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2690                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2691                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2692                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2693
2694                         foreach_post_remat(bb, tmp) {
2695                                 op_t     *remat_op = get_irn_link(tmp);
2696
2697                                 if(remat_op->attr.remat.remat->value == irn) {
2698                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2699                                 }
2700                         }
2701                 }
2702         }
2703
2704         /* walk forward now and compute constraints for placing spills */
2705         /* this must only be done for values that are not defined in this block */
2706         /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2707         pset_foreach(live, irn) {
2708                 /*
2709                  * if value is defined in this block we can anways place the spill directly after the def
2710                  *    -> no constraint necessary
2711                  */
2712                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2713
2714
2715                 spill = set_find_spill(spill_bb->ilp, irn);
2716                 assert(spill);
2717
2718                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2719                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2720
2721                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2722                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2723
2724                 if(!is_Phi(irn)) {
2725                         sched_foreach_op(bb, tmp) {
2726                                 op_t   *op = get_irn_link(tmp);
2727
2728                                 if(is_Phi(tmp)) continue;
2729                                 assert(!is_Proj(tmp));
2730
2731                                 if(op->is_remat) {
2732                                         const ir_node   *value = op->attr.remat.remat->value;
2733
2734                                         if(value == irn) {
2735                                                 /* only collect remats up to the first real use of a value */
2736                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2737                                         }
2738                                 } else {
2739                                         int   n;
2740
2741                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2742                                                 ir_node    *arg = get_irn_n(tmp, n);
2743
2744                                                 if(arg == irn) {
2745                                                         /* if a value is used stop collecting remats */
2746                             goto next_live;
2747                                                 }
2748                                         }
2749                                 }
2750                         }
2751                 }
2752 next_live: ;
2753         }
2754
2755         del_pset(live);
2756 }
2757
2758 typedef struct _irnlist_t {
2759         struct list_head   list;
2760         ir_node           *irn;
2761 } irnlist_t;
2762
2763 typedef struct _interference_t {
2764         struct list_head    blocklist;
2765         ir_node            *a;
2766         ir_node            *b;
2767 } interference_t;
2768
2769 static int
2770 cmp_interference(const void *a, const void *b, size_t size)
2771 {
2772         const interference_t *p = a;
2773         const interference_t *q = b;
2774
2775         return !(p->a == q->a && p->b == q->b);
2776 }
2777
2778 static interference_t *
2779 set_find_interference(set * set, ir_node * a, ir_node * b)
2780 {
2781         interference_t     query;
2782
2783         query.a = (a>b)?a:b;
2784         query.b = (a>b)?b:a;
2785
2786         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2787 }
2788
2789 static interference_t *
2790 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2791 {
2792         interference_t     query,
2793                                           *result;
2794         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2795
2796         list->irn = bb;
2797
2798         result = set_find_interference(set, a, b);
2799         if(result) {
2800
2801                 list_add(&list->list, &result->blocklist);
2802                 return result;
2803         }
2804
2805         query.a = (a>b)?a:b;
2806         query.b = (a>b)?b:a;
2807
2808         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2809
2810         INIT_LIST_HEAD(&result->blocklist);
2811         list_add(&list->list, &result->blocklist);
2812
2813         return result;
2814 }
2815
2816 static int
2817 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2818 {
2819         const ir_edge_t *edge;
2820
2821         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2822                 /* both values are live in, so they interfere */
2823                 return 1;
2824         }
2825
2826         /* ensure a dominates b */
2827         if(value_dominates(b,a)) {
2828                 const ir_node * t;
2829                 t = b;
2830                 b = a;
2831                 a = t;
2832         }
2833         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2834
2835
2836         /* the following code is stolen from bera.c */
2837         if(be_is_live_end(si->lv, bb, a))
2838                 return 1;
2839
2840         foreach_out_edge(a, edge) {
2841                 const ir_node *user = edge->src;
2842                 if(get_nodes_block(user) == bb
2843                                 && !is_Phi(user)
2844                                 && b != user
2845                                 && !pset_find_ptr(si->inverse_ops, user)
2846                                 && value_dominates(b, user))
2847                         return 1;
2848         }
2849
2850         return 0;
2851 }
2852
2853 /**
2854  * Walk all irg blocks and collect interfering values inside of phi classes
2855  */
2856 static void
2857 luke_interferencewalker(ir_node * bb, void * data)
2858 {
2859         spill_ilp_t    *si = (spill_ilp_t*)data;
2860         int             l1, l2;
2861
2862         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2863                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2864                 op_t           *a_op = get_irn_link(a);
2865
2866
2867                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2868                 if (has_reg_class(si, a) && get_phi_class(a)) {
2869                         if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2870                                 continue;
2871
2872                         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)) {
2873                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
2874                                 op_t           *b_op = get_irn_link(b);
2875
2876
2877                                 /* a and b are only interesting if they are in the same phi class */
2878                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2879                                         if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2880                                                 continue;
2881
2882                                         if(values_interfere_in_block(si, bb, a, b)) {
2883                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2884                                                 set_insert_interference(si, si->interferences, a, b, bb);
2885                                         }
2886                                 }
2887                         }
2888                 }
2889         }
2890 }
2891
2892 static unsigned int copy_path_id = 0;
2893
2894 static void
2895 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2896 {
2897         ilp_cst_t  cst;
2898         ilp_var_t  copy;
2899         char       buf[256];
2900         void      *ptr;
2901
2902         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2903         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2904
2905         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2906
2907         pset_foreach(copies, ptr) {
2908                 copy = PTR_TO_INT(ptr);
2909                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2910         }
2911 }
2912
2913 /**
2914  * @parameter copies   contains a path of copies which lead us to irn
2915  * @parameter visited  contains a set of nodes already visited on this path
2916  */
2917 static int
2918 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2919 {
2920         const ir_edge_t *edge;
2921         op_t            *op = get_irn_link(irn);
2922     pset            *visited_users = pset_new_ptr_default();
2923         int              paths = 0;
2924
2925         if(op->is_remat) return 0;
2926
2927         pset_insert_ptr(visited, irn);
2928
2929         if(is_Phi(irn)) {
2930                 int    n;
2931         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2932
2933                 /* visit all operands */
2934                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2935                         ir_node  *arg = get_irn_n(irn, n);
2936                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2937
2938                         if(!has_reg_class(si, arg)) continue;
2939             if(pset_find_ptr(visited_operands, arg)) continue;
2940             pset_insert_ptr(visited_operands, arg);
2941
2942                         if(arg == target) {
2943                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2944                                         del_pset(visited_operands);
2945                                         del_pset(visited_users);
2946                                         pset_remove_ptr(visited, irn);
2947                                         return paths;
2948                                 }
2949                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2950                                 write_copy_path_cst(si, copies, any_interfere);
2951                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2952                         } else if(!pset_find_ptr(visited, arg)) {
2953                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2954                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2955                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2956
2957                 if(paths > MAX_PATHS) {
2958                     if(pset_count(copies) == 0) {
2959                         ilp_cst_t  cst;
2960                         char       buf[256];
2961
2962                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2963                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2964                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2965                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2966                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2967
2968                         paths = 0;
2969                     } else {
2970                         del_pset(visited_operands);
2971                         del_pset(visited_users);
2972                         pset_remove_ptr(visited, irn);
2973                         return paths;
2974                     }
2975                 } else if(pset_count(copies) == 0) {
2976                                         paths = 0;
2977                                 }
2978                         }
2979                 }
2980
2981         del_pset(visited_operands);
2982         }
2983
2984         /* visit all uses which are phis */
2985         foreach_out_edge(irn, edge) {
2986                 ir_node  *user = edge->src;
2987                 int       pos  = edge->pos;
2988                 op_t     *op = get_irn_link(user);
2989                 ilp_var_t copy;
2990
2991                 if(!is_Phi(user)) continue;
2992                 if(!has_reg_class(si, user)) continue;
2993         if(pset_find_ptr(visited_users, user)) continue;
2994         pset_insert_ptr(visited_users, user);
2995
2996                 copy = op->attr.live_range.args.copies[pos];
2997
2998                 if(user == target) {
2999                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3000                                 del_pset(visited_users);
3001                                 pset_remove_ptr(visited, irn);
3002                                 return paths;
3003                         }
3004                         pset_insert(copies, INT_TO_PTR(copy), copy);
3005                         write_copy_path_cst(si, copies, any_interfere);
3006                         pset_remove(copies, INT_TO_PTR(copy), copy);
3007                 } else if(!pset_find_ptr(visited, user)) {
3008                         pset_insert(copies, INT_TO_PTR(copy), copy);
3009                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3010                         pset_remove(copies, INT_TO_PTR(copy), copy);
3011
3012             if(paths > MAX_PATHS) {
3013                 if(pset_count(copies) == 0) {
3014                     ilp_cst_t  cst;
3015                     char       buf[256];
3016
3017                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3018                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3019                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3020                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3021                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3022
3023                     paths = 0;
3024                 } else {
3025                     del_pset(visited_users);
3026                     pset_remove_ptr(visited, irn);
3027                     return paths;
3028                 }
3029             } else if(pset_count(copies) == 0) {
3030                                 paths = 0;
3031                         }
3032                 }
3033         }
3034
3035     del_pset(visited_users);
3036         pset_remove_ptr(visited, irn);
3037         return paths;
3038 }
3039
3040 static void
3041 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3042 {
3043         pset * copies = pset_new_ptr_default();
3044         pset * visited = pset_new_ptr_default();
3045
3046         find_copy_path(si, a, b, any_interfere, copies, visited);
3047
3048         del_pset(visited);
3049         del_pset(copies);
3050 }
3051
3052
3053 static void
3054 memcopyhandler(spill_ilp_t * si)
3055 {
3056         interference_t   *interference;
3057         char              buf[256];
3058         /* teste Speicherwerte auf Interferenz */
3059
3060         /* analyze phi classes */
3061         phi_class_compute(si->chordal_env->irg);
3062
3063         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3064         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3065
3066         /* now lets emit the ILP unequations for the crap */
3067         set_foreach(si->interferences, interference) {
3068                 irnlist_t      *irnlist;
3069                 ilp_var_t       interfere,
3070                                                 any_interfere;
3071                 ilp_cst_t       any_interfere_cst,
3072                                                 cst;
3073                 const ir_node  *a  = interference->a;
3074                 const ir_node  *b  = interference->b;
3075
3076                 /* any_interf <= \sum interf */
3077                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3078                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3079                 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3080
3081                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3082
3083                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3084                         const ir_node  *bb = irnlist->irn;
3085                         spill_bb_t     *spill_bb = get_irn_link(bb);
3086                         spill_t        *spilla,
3087                                                    *spillb;
3088                         char           buf[256];
3089
3090                         spilla = set_find_spill(spill_bb->ilp, a);
3091                         assert(spilla);
3092
3093                         spillb = set_find_spill(spill_bb->ilp, b);
3094                         assert(spillb);
3095
3096                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3097                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3098                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3099                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3100                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3101                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3102
3103                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3104                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3105
3106                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3107                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3108                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3109                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3110                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3111
3112                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3113                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3114
3115                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3116                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3117                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3118
3119                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3120                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3121
3122                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3123                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3124                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3125
3126
3127                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3128
3129                         /* any_interfere >= interf */
3130                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3131                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3132
3133                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3134                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3135                 }
3136
3137                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3138                 gen_copy_constraints(si,a,b,any_interfere);
3139         }
3140 }
3141
3142
3143 static INLINE int
3144 is_zero(double x)
3145 {
3146         return fabs(x) < 0.00001;
3147 }
3148
3149 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3150 {
3151         spill_ilp_t *si = get_irg_link(current_ir_graph);
3152
3153         if(pset_find_ptr(si->all_possible_remats, n)) {
3154                 op_t   *op = (op_t*)get_irn_link(n);
3155                 assert(op && op->is_remat);
3156
3157                 if(!op->attr.remat.remat->inverse) {
3158                         if(op->attr.remat.pre) {
3159                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3160                         } else {
3161                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3162                         }
3163
3164                         return 1;
3165                 } else {
3166                         op_t   *op = (op_t*)get_irn_link(n);
3167                         assert(op && op->is_remat);
3168
3169                         if(op->attr.remat.pre) {
3170                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3171                         } else {
3172                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3173                         }
3174
3175                         return 1;
3176                 }
3177         }
3178
3179         return 0;
3180 }
3181
3182 static void
3183 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3184 {
3185         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3186         be_dump(irg, suffix, dump_ir_block_graph_sched);
3187         set_dump_node_vcgattr_hook(NULL);
3188 }
3189
3190 /**
3191  * Edge hook to dump the schedule edges with annotated register pressure.
3192  */
3193 static int
3194 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3195 {
3196         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3197                 ir_node *prev = sched_prev(irn);
3198                 fprintf(F, "edge:{sourcename:\"");
3199                 PRINT_NODEID(irn);
3200                 fprintf(F, "\" targetname:\"");
3201                 PRINT_NODEID(prev);
3202                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3203                 fprintf(F, "\" color:magenta}\n");
3204         }
3205         return 1;
3206 }
3207
3208 static void
3209 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3210 {
3211         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3212
3213         dump_consts_local(0);
3214         set_dump_node_edge_hook(sched_pressure_edge_hook);
3215         dump_ir_block_graph(irg, suffix);
3216         set_dump_node_edge_hook(old_edge_hook);
3217 }
3218
3219 static void
3220 walker_pressure_annotator(ir_node * bb, void * data)
3221 {
3222         spill_ilp_t  *si = data;
3223         ir_node      *irn;
3224         int           n, i;
3225         pset         *live = pset_new_ptr_default();
3226         int           projs = 0;
3227
3228         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3229                 irn = be_lv_get_irn(si->lv, bb, i);
3230
3231                 if (has_reg_class(si, irn)) {
3232                         pset_insert_ptr(live, irn);
3233                 }
3234         }
3235
3236         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3237
3238         sched_foreach_reverse(bb, irn) {
3239                 if(is_Phi(irn)) {
3240                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3241                         continue;
3242                 }
3243
3244                 if(has_reg_class(si, irn)) {
3245                         pset_remove_ptr(live, irn);
3246                         if(is_Proj(irn)) ++projs;
3247                 }
3248
3249                 if(!is_Proj(irn)) projs = 0;
3250
3251                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3252                         ir_node    *arg = get_irn_n(irn, n);
3253
3254                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3255                 }
3256                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3257         }
3258
3259         del_pset(live);
3260 }
3261
3262 static void
3263 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3264 {
3265         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3266 }
3267
3268 static void
3269 connect_all_remats_with_keep(spill_ilp_t * si)
3270 {
3271         ir_node   *irn;
3272         ir_node  **ins,
3273                          **pos;
3274         int        n_remats;
3275
3276
3277         n_remats = pset_count(si->all_possible_remats);
3278         if(n_remats) {
3279                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3280
3281                 pos = ins;
3282                 pset_foreach(si->all_possible_remats, irn) {
3283                         *pos = irn;
3284                         ++pos;
3285                 }
3286
3287                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3288
3289                 obstack_free(si->obst, ins);
3290         }
3291 }
3292
3293 static void
3294 connect_all_spills_with_keep(spill_ilp_t * si)
3295 {
3296         ir_node   *irn;
3297         ir_node  **ins,
3298                          **pos;
3299         int        n_spills;
3300         ir_node   *keep;
3301
3302
3303         n_spills = pset_count(si->spills);
3304         if(n_spills) {
3305                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3306
3307                 pos = ins;
3308                 pset_foreach(si->spills, irn) {
3309                         *pos = irn;
3310                         ++pos;
3311                 }
3312
3313                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3314
3315                 obstack_free(si->obst, ins);
3316         }
3317 }
3318
3319 /** insert a spill at an arbitrary position */
3320 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3321 {
3322         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
3323         ir_graph *irg   = get_irn_irg(bl);
3324         ir_node *frame  = get_irg_frame(irg);
3325         ir_node *spill;
3326         ir_node *next;
3327
3328         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3329         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3330
3331         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3332
3333         /*
3334          * search the right insertion point. a spill of a phi cannot be put
3335          * directly after the phi, if there are some phis behind the one which
3336          * is spilled. Also, a spill of a Proj must be after all Projs of the
3337          * same tuple node.
3338          *
3339          * Here's one special case:
3340          * If the spill is in the start block, the spill must be after the frame
3341          * pointer is set up. This is done by setting insert to the end of the block
3342          * which is its default initialization (see above).
3343          */
3344
3345         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3346                 insert = frame;
3347
3348         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3349                 insert = next;
3350
3351         sched_add_after(insert, spill);
3352         return spill;
3353 }
3354
3355 static void
3356 delete_remat(spill_ilp_t * si, ir_node * remat) {
3357         int       n;
3358         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3359
3360         sched_remove(remat);
3361
3362         /* kill links to operands */
3363         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3364                 set_irn_n(remat, n, bad);
3365         }
3366 }
3367
3368 static void
3369 clean_remat_info(spill_ilp_t * si)
3370 {
3371         int            n;
3372         remat_t       *remat;
3373         remat_info_t  *remat_info;
3374         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
3375
3376         set_foreach(si->remat_info, remat_info) {
3377                 if(!remat_info->remats) continue;
3378
3379                 pset_foreach(remat_info->remats, remat)
3380                 {
3381                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3382                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3383                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3384                         }
3385
3386                         if(get_irn_n_edges(remat->op) == 0) {
3387                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3388                                         set_irn_n((ir_node*)remat->op, n, bad);
3389                                 }
3390                         }
3391                 }
3392
3393                 if(remat_info->remats) del_pset(remat_info->remats);
3394                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3395         }
3396 }
3397
3398 static void
3399 delete_unnecessary_remats(spill_ilp_t * si)
3400 {
3401         if(opt_keep_alive & KEEPALIVE_REMATS) {
3402                 int       n;
3403                 ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3404
3405                 if(si->keep) {
3406 //                      ir_node   *end = get_irg_end(si->chordal_env->irg);
3407 //                      ir_node  **keeps;
3408
3409                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3410                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3411                                 op_t           *arg_op = get_irn_link(keep_arg);
3412                                 lpp_name_t     *name;
3413
3414                                 assert(arg_op->is_remat);
3415
3416                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3417
3418                                 if(is_zero(name->value)) {
3419                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3420                                         /* TODO check whether reload is preferred over remat (could be bug) */
3421                                         delete_remat(si, keep_arg);
3422                                 } else {
3423                                         if(!arg_op->attr.remat.remat->inverse) {
3424                                                 if(arg_op->attr.remat.pre) {
3425                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3426                                                 } else {
3427                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3428                                                 }
3429                                         } else {
3430                                                 if(arg_op->attr.remat.pre) {
3431                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3432                                                 } else {
3433                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3434                                                 }
3435                                         }
3436                                 }
3437
3438                                 set_irn_n(si->keep, n, bad);
3439                         }
3440 #if 0
3441                         for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3442                                 ir_node        *end_arg = get_End_keepalive(end, i);
3443
3444                                 if(end_arg != si->keep) {
3445                                         obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3446                                 }
3447                         }
3448                         keeps = obstack_finish(si->obst);
3449                         set_End_keepalives(end, n-1, keeps);
3450                         obstack_free(si->obst, keeps);
3451 #endif
3452                 } else {
3453                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3454                 }
3455         } else {
3456                 ir_node  *remat;
3457
3458                 pset_foreach(si->all_possible_remats, remat) {
3459                         op_t           *remat_op = get_irn_link(remat);
3460                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3461
3462                         if(is_zero(name->value)) {
3463                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3464                                 /* TODO check whether reload is preferred over remat (could be bug) */
3465                                 delete_remat(si, remat);
3466                         } else {
3467                                 if(!remat_op->attr.remat.remat->inverse) {
3468                                         if(remat_op->attr.remat.pre) {
3469                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3470                                         } else {
3471                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3472                                         }
3473                                 } else {
3474                                         if(remat_op->attr.remat.pre) {
3475                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3476                                         } else {
3477                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3478                                         }
3479                                 }
3480                         }
3481                 }
3482         }
3483 }
3484
3485 static pset *
3486 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3487 {
3488         pset     *spills = pset_new_ptr_default();
3489
3490         const ir_node  *next;
3491         defs_t         *defs;
3492
3493         defs = set_find_def(si->values, value);
3494
3495         if(defs && defs->spills) {
3496                 for(next = defs->spills; next; next = get_irn_link(next)) {
3497                         pset_insert_ptr(spills, next);
3498                 }
3499         }
3500
3501         return spills;
3502 }
3503
3504 /**
3505  * @param before   The node after which the spill will be placed in the schedule
3506  */
3507 static ir_node *
3508 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3509 {
3510         defs_t   *defs;
3511         ir_node  *spill;
3512         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3513
3514         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3515
3516         spill = be_spill2(arch_env, irn, before);
3517
3518         defs = set_insert_def(si->values, value);
3519         assert(defs);
3520
3521         /* enter into the linked list */
3522         set_irn_link(spill, defs->spills);
3523         defs->spills = spill;
3524
3525         if(opt_keep_alive & KEEPALIVE_SPILLS)
3526                 pset_insert_ptr(si->spills, spill);
3527
3528         return spill;
3529 }
3530
3531 /**
3532  * @param before   The Phi node which has to be spilled
3533  */
3534 static ir_node *
3535 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3536 {
3537         ir_node   *mem_phi;
3538         ir_node  **ins;
3539         defs_t    *defs;
3540         int        n;
3541
3542         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3543
3544         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3545                 ins[n] = si->m_unknown;
3546         }
3547
3548         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3549
3550         defs = set_insert_def(si->values, phi);
3551         assert(defs);
3552
3553         /* enter into the linked list */
3554         set_irn_link(mem_phi, defs->spills);
3555         defs->spills = mem_phi;
3556
3557 #ifdef SCHEDULE_PHIM
3558         sched_add_after(phi, mem_phi);
3559 #endif
3560
3561         if(opt_keep_alive & KEEPALIVE_SPILLS)
3562                 pset_insert_ptr(si->spills, mem_phi);
3563
3564
3565         return mem_phi;
3566 }
3567
3568 /**
3569  * Add remat to list of defs, destroys link field!
3570  */
3571 static void
3572 insert_remat(spill_ilp_t * si, ir_node * remat)
3573 {
3574         defs_t   *defs;
3575         op_t     *remat_op = get_irn_link(remat);
3576
3577         assert(remat_op->is_remat);
3578
3579         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3580         assert(defs);
3581
3582         /* enter into the linked list */
3583         set_irn_link(remat, defs->remats);
3584         defs->remats = remat;
3585 }
3586
3587
3588 /**
3589  * Add reload before operation and add to list of defs
3590  */
3591 static ir_node *
3592 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3593 {
3594         defs_t   *defs;
3595         ir_node  *reload,
3596                          *spill;
3597         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3598
3599         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3600
3601         defs = set_find_def(si->values, value);
3602
3603         spill = defs->spills;
3604         assert(spill && "no spill placed before reload");
3605
3606         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3607
3608         /* enter into the linked list */
3609         set_irn_link(reload, defs->remats);
3610         defs->remats = reload;
3611
3612         return reload;
3613 }
3614
3615 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3616 {
3617         defs_t           *defs;
3618         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3619         ir_node          *spill;
3620         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3621
3622         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3623
3624         defs = set_find_def(si->values, value);
3625
3626         spill = defs->spills;
3627         assert(spill && "no spill placed before reload");
3628
3629         arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3630 }
3631
3632 void insert_memoperands(spill_ilp_t * si)
3633 {
3634         memoperand_t   *memoperand;
3635         lpp_name_t     *name;
3636
3637         set_foreach(si->memoperands, memoperand) {
3638                 name = si->lpp->vars[memoperand->ilp];
3639                 if(!is_zero(name->value)) {
3640                         perform_memory_operand(si, memoperand);
3641                 }
3642         }
3643 }
3644
3645 static void
3646 walker_spill_placer(ir_node * bb, void * data) {
3647         spill_ilp_t   *si = (spill_ilp_t*)data;
3648         ir_node       *irn;
3649         spill_bb_t    *spill_bb = get_irn_link(bb);
3650         pset          *spills_to_do = pset_new_ptr_default();
3651         spill_t       *spill;
3652
3653         set_foreach(spill_bb->ilp, spill) {
3654                 lpp_name_t    *name;
3655
3656                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3657                         name = si->lpp->vars[spill->mem_in];
3658                         if(!is_zero(name->value)) {
3659                                 ir_node   *mem_phi;
3660
3661                                 mem_phi = insert_mem_phi(si, spill->irn);
3662
3663                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3664                         }
3665                 }
3666
3667                 name = si->lpp->vars[spill->spill];
3668                 if(!is_zero(name->value)) {
3669                         /* place spill directly after definition */
3670                         if(get_nodes_block(spill->irn) == bb) {
3671                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3672                                 continue;
3673                         }
3674
3675                         /* place spill at bb start */
3676                         if(spill->reg_in > 0) {
3677                                 name = si->lpp->vars[spill->reg_in];
3678                                 if(!is_zero(name->value)) {
3679                                         insert_spill(si, spill->irn, spill->irn, bb);
3680                                         continue;
3681                                 }
3682                         }
3683                         /* place spill after a remat */
3684                         pset_insert_ptr(spills_to_do, spill->irn);
3685                 }
3686         }
3687         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3688
3689
3690         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3691                 op_t     *op = get_irn_link(irn);
3692
3693                 if(be_is_Spill(irn)) continue;
3694
3695                 if(op->is_remat) {
3696                         /* TODO fix this if we want to support remats with more than two nodes */
3697                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3698                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3699
3700                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3701                         }
3702                 } else {
3703                         if(pset_find_ptr(spills_to_do, irn)) {
3704                                 pset_remove_ptr(spills_to_do, irn);
3705
3706                                 insert_spill(si, irn, irn, irn);
3707                         }
3708                 }
3709
3710         }
3711
3712         assert(pset_count(spills_to_do) == 0);
3713
3714         /* afterwards free data in block */
3715         del_pset(spills_to_do);
3716 }
3717
3718 static ir_node *
3719 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3720 {
3721         ir_node          *insert_pos = bb;
3722         ir_node          *spill;
3723         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3724
3725         /* find last definition of arg value in block */
3726         ir_node  *next;
3727         defs_t   *defs;
3728         int       last = 0;
3729
3730         defs = set_find_def(si->values, value);
3731
3732         if(defs && defs->remats) {
3733                 for(next = defs->remats; next; next = get_irn_link(next)) {
3734                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3735                                 last = sched_get_time_step(next);
3736                                 insert_pos = next;
3737                         }
3738                 }
3739         }
3740
3741         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3742                 last = sched_get_time_step(value);
3743                 insert_pos = value;
3744         }
3745
3746         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3747
3748         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3749
3750         return spill;
3751 }
3752
3753 static void
3754 phim_fixer(spill_ilp_t *si) {
3755         defs_t  *defs;
3756
3757         set_foreach(si->values, defs) {
3758                 const ir_node  *phi = defs->value;
3759                 op_t           *op = get_irn_link(phi);
3760                 ir_node        *phi_m = NULL;
3761                 ir_node        *next = defs->spills;
3762                 int             n;
3763
3764                 if(!is_Phi(phi)) continue;
3765
3766                 while(next) {
3767                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3768                                 phi_m = next;
3769                                 break;
3770                         } else {
3771                                 next = get_irn_link(next);
3772                         }
3773                 }
3774                 if(!phi_m) continue;
3775
3776                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3777                         ir_node        *value = get_irn_n(phi, n);
3778                         defs_t         *val_defs = set_find_def(si->values, value);
3779
3780                         /* a spill of this value */
3781                         ir_node      *spill;
3782
3783
3784                         if(opt_memcopies) {
3785                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3786                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3787
3788                                 if(!is_zero(name->value)) {
3789                                         spill = insert_mem_copy(si, pred, value);
3790                                 } else {
3791                                         spill = val_defs->spills;
3792                                 }
3793                         } else {
3794                                 spill = val_defs->spills;
3795                         }
3796
3797                         assert(spill && "no spill placed before PhiM");
3798                         set_irn_n(phi_m, n, spill);
3799                 }
3800         }
3801 }
3802
3803 static void
3804 walker_reload_placer(ir_node * bb, void * data) {
3805         spill_ilp_t   *si = (spill_ilp_t*)data;
3806         ir_node       *irn;
3807         spill_bb_t    *spill_bb = get_irn_link(bb);
3808
3809         /* reloads at end of block */
3810         if(spill_bb->reloads) {
3811                 keyval_t    *keyval;
3812
3813                 set_foreach(spill_bb->reloads, keyval) {
3814                         ir_node        *irn = (ir_node*)keyval->key;
3815                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3816                         lpp_name_t     *name;
3817
3818                         name = si->lpp->vars[reload];
3819                         if(!is_zero(name->value)) {
3820                                 ir_node    *reload;
3821                                 ir_node    *insert_pos = bb;
3822                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3823                                 op_t       *prev_op = get_irn_link(prev);
3824
3825                                 while(be_is_Spill(prev)) {
3826                                         prev = sched_prev(prev);
3827                                 }
3828
3829                                 prev_op = get_irn_link(prev);
3830
3831                                 /* insert reload before pre-remats */
3832                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3833                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3834                                         insert_pos = prev;
3835
3836                                         do {
3837                                                 prev = sched_prev(prev);
3838                                         } while(be_is_Spill(prev));
3839
3840                                         prev_op = get_irn_link(prev);
3841
3842                                 }
3843
3844                                 reload = insert_reload(si, irn, insert_pos);
3845
3846                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3847                                         pset_insert_ptr(si->spills, reload);
3848                         }
3849                 }
3850         }
3851
3852         /* walk and insert more reloads and collect remats */
3853         sched_foreach_reverse(bb, irn) {
3854                 op_t     *op = get_irn_link(irn);
3855
3856                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3857                 if(is_Phi(irn)) break;
3858
3859                 if(op->is_remat) {
3860                         if(get_irn_mode(irn) != mode_T) {
3861                                 insert_remat(si, irn);
3862                         }
3863                 } else {
3864                         int    n;
3865
3866                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3867                                 ir_node    *arg = get_irn_n(irn, n);
3868
3869                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3870                                         lpp_name_t    *name;
3871
3872                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3873                                         if(!is_zero(name->value)) {
3874                                                 ir_node    *reload;
3875                                                 ir_node    *insert_pos = irn;
3876                                                 ir_node    *prev = sched_prev(insert_pos);
3877                                                 op_t       *prev_op;
3878
3879                                                 while(be_is_Spill(prev)) {
3880                                                         prev = sched_prev(prev);
3881                                                 }
3882
3883                                                 prev_op = get_irn_link(prev);
3884
3885                                                 /* insert reload before pre-remats */
3886                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3887                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3888                                                         insert_pos = prev;
3889
3890                                                         do {
3891                                                                 prev = sched_prev(prev);
3892                                                         } while(be_is_Spill(prev));
3893
3894                                                         prev_op = get_irn_link(prev);
3895
3896                                                 }
3897
3898                                                 reload = insert_reload(si, arg, insert_pos);
3899
3900                                                 set_irn_n(irn, n, reload);
3901
3902                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3903                                                         pset_insert_ptr(si->spills, reload);
3904                                         }
3905                                 }
3906                         }
3907                 }
3908         }
3909
3910         del_set(spill_bb->ilp);
3911         if(spill_bb->reloads) del_set(spill_bb->reloads);
3912 }
3913
3914 static void
3915 walker_collect_used(ir_node * irn, void * data)
3916 {
3917         bitset_t   *used = data;
3918
3919         bitset_set(used, get_irn_idx(irn));
3920 }
3921
3922 struct kill_helper {
3923         bitset_t  *used;
3924         spill_ilp_t  *si;
3925 };
3926
3927 static void
3928 walker_kill_unused(ir_node * bb, void * data)
3929 {
3930         struct kill_helper *kh = data;
3931         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
3932         ir_node            *irn;
3933
3934
3935         for(irn=sched_first(bb); !sched_is_end(irn);) {
3936                 ir_node     *next = sched_next(irn);
3937                 int          n;
3938
3939                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3940                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3941                                 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)));
3942 #if 0
3943                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3944 #endif
3945                         }
3946
3947                         sched_remove(irn);
3948
3949                         set_nodes_block(irn, bad);
3950                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3951                                 set_irn_n(irn, n, bad);
3952                         }
3953                 }
3954                 irn = next;
3955         }
3956 }
3957
3958 static void
3959 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3960 {
3961         struct kill_helper kh;
3962
3963         kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3964         kh.si = si;
3965
3966         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3967         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3968
3969         bitset_free(kh.used);
3970 }
3971
3972 void
3973 print_irn_pset(pset * p)
3974 {
3975         ir_node   *irn;
3976
3977         pset_foreach(p, irn) {
3978                 ir_printf("%+F\n", irn);
3979         }
3980 }
3981
3982 void
3983 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3984 {
3985     FILE           *f = fopen(file, "w");
3986     ir_node        *irn;
3987     interference_t *interference;
3988
3989     pset_break(phiclass);
3990     set_break(si->interferences);
3991
3992     ir_fprintf(f, "digraph phiclass {\n");
3993
3994     pset_foreach(phiclass, irn) {
3995         if(is_Phi(irn))
3996             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
3997     }
3998
3999     pset_foreach(phiclass, irn) {
4000         int n;
4001
4002         if(!is_Phi(irn)) continue;
4003
4004         for(n=get_irn_arity(irn)-1; n>=0; --n) {
4005             ir_node  *arg = get_irn_n(irn, n);
4006
4007             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
4008         }
4009     }
4010
4011     set_foreach(si->interferences, interference) {
4012         const ir_node  *a  = interference->a;
4013         const ir_node  *b  = interference->b;
4014         if(get_phi_class(a) == phiclass) {
4015             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4016         }
4017     }
4018
4019     ir_fprintf(f, "}");
4020     fclose(f);
4021 }
4022
4023 static void
4024 rewire_uses(spill_ilp_t * si)
4025 {
4026         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4027         defs_t               *defs;
4028         pset                 *ignore = pset_new_ptr(1);
4029
4030         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4031
4032         /* then fix uses of spills */
4033         set_foreach(si->values, defs) {
4034                 pset           *reloads;
4035                 pset           *spills;
4036                 const ir_node  *next = defs->remats;
4037                 int remats = 0;
4038
4039                 reloads = pset_new_ptr_default();
4040
4041                 while(next) {
4042                         if(be_is_Reload(next)) {
4043                                 pset_insert_ptr(reloads, next);
4044                         } else {
4045                                 ++remats;
4046                         }
4047                         next = get_irn_link(next);
4048                 }
4049
4050                 spills = get_spills_for_value(si, defs->value);
4051                 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));
4052                 if(pset_count(spills) > 1) {
4053                         //assert(pset_count(reloads) > 0);
4054                         //                              print_irn_pset(spills);
4055                         //                              print_irn_pset(reloads);
4056
4057                         be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4058                 }
4059
4060                 del_pset(reloads);
4061                 del_pset(spills);
4062         }
4063
4064         /* first fix uses of remats and reloads */
4065         set_foreach(si->values, defs) {
4066                 pset           *nodes;
4067                 const ir_node  *next = defs->remats;
4068
4069                 if(next) {
4070                         nodes = pset_new_ptr_default();
4071                         pset_insert_ptr(nodes, defs->value);
4072
4073                         while(next) {
4074                                 pset_insert_ptr(nodes, next);
4075                                 next = get_irn_link(next);
4076                         }
4077
4078                         if(pset_count(nodes) > 1) {
4079                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
4080                                 be_ssa_constr_set(dfi, si->lv, nodes);
4081                         }
4082
4083                         del_pset(nodes);
4084                 }
4085         }
4086
4087 //      remove_unused_defs(si);
4088
4089         be_free_dominance_frontiers(dfi);
4090 }
4091
4092
4093 static void
4094 writeback_results(spill_ilp_t * si)
4095 {
4096         /* walk through the graph and collect all spills, reloads and remats for a value */
4097
4098         si->values = new_set(cmp_defs, 4096);
4099
4100         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4101         delete_unnecessary_remats(si);
4102         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4103         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4104         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4105         if(opt_memoperands)
4106                 insert_memoperands(si);
4107         phim_fixer(si);
4108
4109         /* clean the remat info! there are still back-edges leading there! */
4110         clean_remat_info(si);
4111
4112         rewire_uses(si);
4113
4114         connect_all_spills_with_keep(si);
4115
4116         del_set(si->values);
4117 }
4118
4119 static int
4120 get_n_regs(spill_ilp_t * si)
4121 {
4122         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4123
4124         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4125         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4126
4127         arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4128     be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4129
4130         bitset_andnot(arch_regs, abi_regs);
4131         arch_n_regs = bitset_popcnt(arch_regs);
4132
4133         bitset_free(arch_regs);
4134         bitset_free(abi_regs);
4135
4136         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4137         return arch_n_regs;
4138 }
4139
4140 static void
4141 walker_reload_mover(ir_node * bb, void * data)
4142 {
4143         spill_ilp_t   *si = data;
4144         ir_node           *tmp;
4145
4146         sched_foreach(bb, tmp) {
4147                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4148                         ir_node       *reload = tmp;
4149                         ir_node       *irn = tmp;
4150
4151                         /* move reload upwards */
4152
4153                         int pressure = (int)get_irn_link(reload);
4154                         if(pressure < si->n_regs) {
4155                                 irn = sched_prev(reload);
4156                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4157                                 sched_remove(reload);
4158                                 pressure = (int)get_irn_link(irn);
4159
4160                                 while(pressure < si->n_regs) {
4161                                         if( sched_is_end(irn) ||
4162                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4163                                            /* do not move reload before its spill */
4164                                            (irn == be_get_Reload_mem(reload)) ||
4165                                            /* do not move before phi */
4166                                            is_Phi(irn)) break;
4167
4168                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4169                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4170                                         irn = sched_prev(irn);
4171
4172                                         pressure = (int)get_irn_link(irn);
4173                                 }
4174
4175                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4176                                 sched_put_after(irn, reload);
4177                         }
4178                 }
4179         }
4180 }
4181
4182 static void
4183 move_reloads_upward(spill_ilp_t * si)
4184 {
4185         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4186 }
4187
4188
4189 /**
4190  * Walk all irg blocks and check for interfering spills inside of phi classes
4191  */
4192 static void
4193 luke_meminterferencechecker(ir_node * bb, void * data)
4194 {
4195         spill_ilp_t    *si = (spill_ilp_t*)data;
4196         int             l1, l2;
4197
4198         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4199                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4200
4201                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4202
4203                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4204                 if (has_reg_class(si, a) && get_phi_class(a)) {
4205                         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)) {
4206                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
4207
4208                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4209
4210                                 /* a and b are only interesting if they are in the same phi class */
4211                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4212                                         if(values_interfere_in_block(si, bb, a, b)) {
4213                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4214                                         }
4215                                 }
4216                         }
4217                 }
4218         }
4219 }
4220
4221 static void
4222 verify_phiclasses(spill_ilp_t * si)
4223 {
4224         /* analyze phi classes */
4225         phi_class_compute(si->chordal_env->irg);
4226
4227         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4228         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4229 }
4230
4231 void
4232 be_spill_remat(const be_chordal_env_t * chordal_env)
4233 {
4234         char            buf[256];
4235         char            problem_name[256];
4236         char            dump_suffix[256];
4237         char            dump_suffix2[256];
4238         struct obstack  obst;
4239         spill_ilp_t     si;
4240
4241         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4242         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4243         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4244
4245         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4246         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4247
4248         if(opt_verify & VERIFY_DOMINANCE)
4249                 be_check_dominance(chordal_env->irg);
4250
4251         obstack_init(&obst);
4252         si.chordal_env = chordal_env;
4253         si.obst = &obst;
4254         si.cls = chordal_env->cls;
4255         si.lpp = new_lpp(problem_name, lpp_minimize);
4256         si.remat_info = new_set(cmp_remat_info, 4096);
4257         si.interferences = new_set(cmp_interference, 32);
4258         si.memoperands = new_set(cmp_memoperands, 128);
4259         si.all_possible_remats = pset_new_ptr_default();
4260         si.spills = pset_new_ptr_default();
4261         si.inverse_ops = pset_new_ptr_default();
4262         si.lv = chordal_env->lv;
4263         si.keep = NULL;
4264         si.n_regs = get_n_regs(&si);
4265
4266         set_irg_link(chordal_env->irg, &si);
4267         compute_doms(chordal_env->irg);
4268
4269         /* compute phi classes */
4270 //      phi_class_compute(chordal_env->irg);
4271
4272         be_analyze_regpressure(chordal_env, "-pre");
4273
4274         if(opt_remats) {
4275                 /* collect remats */
4276                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4277                 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4278         }
4279
4280         /* insert possible remats */
4281         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4282         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4283         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4284
4285         if(opt_keep_alive & KEEPALIVE_REMATS) {
4286                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4287                 connect_all_remats_with_keep(&si);
4288                 /* dump graph with inserted remats */
4289                 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4290         }
4291
4292         /* insert copies for phi arguments not in my regclass */
4293         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4294
4295         /* recompute liveness */
4296         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4297         be_liveness_recompute(si.lv);
4298
4299         /* build the ILP */
4300
4301         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4302         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4303         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4304
4305         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4306         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4307
4308         if(opt_memcopies) {
4309                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4310                 memcopyhandler(&si);
4311         }
4312
4313         if(opt_dump_flags & DUMP_PROBLEM) {
4314                 FILE           *f;
4315                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4316                 if ((f = fopen(buf, "wt")) != NULL) {
4317                         lpp_dump_plain(si.lpp, f);
4318                         fclose(f);
4319                 }
4320         }
4321
4322         if(opt_dump_flags & DUMP_MPS) {
4323                 FILE *f;
4324
4325                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4326                 if((f = fopen(buf, "wt")) != NULL) {
4327                         mps_write_mps(si.lpp, s_mps_fixed, f);
4328                         fclose(f);
4329                 }
4330
4331                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4332                 if((f = fopen(buf, "wt")) != NULL) {
4333                         mps_write_mst(si.lpp, s_mps_fixed, f);
4334                         fclose(f);
4335                 }
4336         }
4337
4338         lpp_check_startvals(si.lpp);
4339
4340 #ifdef SOLVE
4341         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4342         lpp_set_time_limit(si.lpp, opt_timeout);
4343
4344         if(opt_log)
4345                 lpp_set_log(si.lpp, stdout);
4346
4347 #ifdef SOLVE_LOCAL
4348         lpp_solve_cplex(si.lpp);
4349 #else
4350         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4351 #endif
4352         assert(lpp_is_sol_valid(si.lpp)
4353                && "solution of ILP must be valid");
4354
4355         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));
4356
4357         if(opt_dump_flags & DUMP_SOLUTION) {
4358                 FILE           *f;
4359                 char            buf[256];
4360
4361                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4362                 if ((f = fopen(buf, "wt")) != NULL) {
4363                         int             i;
4364                         for (i = 0; i < si.lpp->var_next; ++i) {
4365                                 lpp_name_t     *name = si.lpp->vars[i];
4366                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4367                         }
4368                         fclose(f);
4369                 }
4370         }
4371
4372         writeback_results(&si);
4373
4374 #endif                          /* SOLVE */
4375
4376         kill_all_unused_values_in_schedule(&si);
4377
4378         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4379                 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4380
4381         // move reloads upwards
4382         be_liveness_recompute(si.lv);
4383         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4384         move_reloads_upward(&si);
4385
4386         if(opt_memcopies) {
4387                 verify_phiclasses(&si);
4388         }
4389
4390         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4391
4392         dump_pressure_graph(&si, dump_suffix2);
4393
4394         be_analyze_regpressure(chordal_env, "-post");
4395
4396         if(opt_verify & VERIFY_DOMINANCE)
4397                 be_check_dominance(chordal_env->irg);
4398
4399         free_dom(chordal_env->irg);
4400         del_set(si.interferences);
4401         del_pset(si.inverse_ops);
4402         del_pset(si.all_possible_remats);
4403         del_set(si.memoperands);
4404         del_pset(si.spills);
4405         free_lpp(si.lpp);
4406         obstack_free(&obst, NULL);
4407         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4408 }
4409
4410 #else                           /* WITH_ILP */
4411
4412 static void
4413 only_that_you_can_compile_without_WITH_ILP_defined(void)
4414 {
4415 }
4416
4417 #endif                          /* WITH_ILP */