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 "irnodeset.h"
  33 #include "phiclass.h"
  34 #include "iredges.h"
  35 #include "execfreq.h"
  36 #include "irvrfy.h"
  37 #include "irbackedge_t.h"
  38
  39 #include <lpp/lpp.h>
  40 #include <lpp/mps.h>
  41 #include <lpp/lpp_net.h>
  42 #include <lpp/lpp_cplex.h>
  43
  44 #include "be_t.h"
  45 #include "beirg_t.h"
  46 #include "belive_t.h"
  47 #include "besched_t.h"
  48 #include "bessaconstr.h"
  49 #include "bearch_t.h"
  50 #include "beabi.h"
  51 #include "benode_t.h"
  52 #include "beutil.h"
  53 #include "bespillremat.h"
  54 #include "bespill.h"
  55 #include "bepressurestat.h"
  56 #include "beprofile.h"
  57 #include "bespilloptions.h"
  58
  59 #include "bechordal_t.h"
  60
  61 #include <libcore/lc_opts.h>
  62 #include <libcore/lc_opts_enum.h>
  63
  64 #define DUMP_PROBLEM       1
  65 #define DUMP_MPS           2
  66 #define DUMP_SOLUTION      4
  67 #define DUMP_STATS         8
  68 #define DUMP_PRESSURE      16
  69
  70 #define KEEPALIVE_REMATS   1
  71 #define KEEPALIVE_SPILLS   2
  72 #define KEEPALIVE_RELOADS  4
  73
  74 #define VERIFY_MEMINTERF   1
  75 #define VERIFY_DOMINANCE   2
  76
  77 #define REMATS_NONE        0
  78 #define REMATS_BRIGGS      1
  79 #define REMATS_NOINVERSE   2
  80 #define REMATS_ALL         3
  81
  82 static unsigned opt_dump_flags = 0;
  83 static int opt_log = 0;
  84 static unsigned opt_keep_alive = 0;
  85 static int opt_goodwin = 1;
  86 static int opt_memcopies = 1;
  87 static int opt_memoperands = 1;
  88 static int opt_verify = VERIFY_MEMINTERF;
  89 static unsigned opt_remats = REMATS_ALL;
  90 static int opt_repair_schedule = 0;
  91 static int opt_no_enlarge_liveness = 0;
  92 static int opt_remat_while_live = 1;
  93 static int opt_timeout = 300;
  94 static double opt_cost_reload = 8.0;
  95 static double opt_cost_memoperand =  7.0;
  96 static double opt_cost_spill =  15.0;
  97 static double opt_cost_remat =  1.0;
  98
  99
 100 static const lc_opt_enum_mask_items_t dump_items[] = {
 101         { "problem",  DUMP_PROBLEM  },
 102         { "mps",      DUMP_MPS      },
 103         { "solution", DUMP_SOLUTION },
 104         { "stats",    DUMP_STATS },
 105         { "pressure", DUMP_PRESSURE },
 106         { NULL,       0 }
 107 };
 108
 109 static lc_opt_enum_mask_var_t dump_var = {
 110         &opt_dump_flags, dump_items
 111 };
 112
 113 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 114         { "remats",  KEEPALIVE_REMATS  },
 115         { "spills",  KEEPALIVE_SPILLS  },
 116         { "reloads", KEEPALIVE_RELOADS },
 117         { NULL,      0 }
 118 };
 119
 120 static lc_opt_enum_mask_var_t keep_alive_var = {
 121         &opt_keep_alive, keepalive_items
 122 };
 123
 124 static const lc_opt_enum_mask_items_t remats_items[] = {
 125         { "none",      REMATS_NONE      },
 126         { "briggs",    REMATS_BRIGGS    },
 127         { "noinverse", REMATS_NOINVERSE },
 128         { "all",       REMATS_ALL       },
 129         { NULL,        0 }
 130 };
 131
 132 static lc_opt_enum_mask_var_t remats_var = {
 133         &opt_remats, remats_items
 134 };
 135
 136 static const lc_opt_table_entry_t options[] = {
 137         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes",                              &keep_alive_var),
 138
 139         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 140         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 141         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 142         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert",                                 &remats_var),
 143         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 144         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 145         LC_OPT_ENT_BOOL     ("remat_while_live",  "only remat where rematted value was live",       &opt_remat_while_live),
 146
 147         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data",                  &dump_var),
 148         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 149         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 150
 151         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 152         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 153         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 154         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 155         { NULL }
 156 };
 157
 158 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 159 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 160
 161 #define  SOLVE
 162 //#define  SOLVE_LOCAL
 163 #define LPP_SERVER "i44pc52"
 164 #define LPP_SOLVER "cplex"
 165
 166
 167 #define MAX_PATHS      INT_MAX
 168 #define ILP_UNDEF               -1
 169
 170 typedef struct _spill_ilp_t {
 171         const arch_register_class_t  *cls;
 172         int                           n_regs;
 173         be_irg_t                     *birg;
 174         be_lv_t                      *lv;
 175         lpp_t                        *lpp;
 176         struct obstack               *obst;
 177         set                          *remat_info;
 178         pset                         *all_possible_remats;
 179         pset                         *inverse_ops;
 180         ir_node                      *keep;
 181         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 182         pset                         *spills;
 183         set                          *interferences;
 184         ir_node                      *m_unknown;
 185         set                          *memoperands;
 186         phi_classes_t                *pc;
 187 #ifndef SCHEDULE_PHIM
 188         pset                         *phims;
 189 #endif
 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 arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
 269                                               si->cls, irn);
 270 }
 271
 272 #if 0
 273 static int
 274 cmp_remat(const void *a, const void *b)
 275 {
 276         const keyval_t *p = a;
 277         const keyval_t *q = b;
 278         const remat_t  *r = p->val;
 279         const remat_t  *s = q->val;
 280
 281         assert(r && s);
 282
 283         return !(r == s || r->op == s->op);
 284 }
 285 #endif
 286 static int
 287 cmp_remat(const void *a, const void *b)
 288 {
 289         const remat_t  *r = a;
 290         const remat_t  *s = a;
 291
 292         return !(r == s || r->op == s->op);
 293 }
 294
 295 static int
 296 cmp_spill(const void *a, const void *b, size_t size)
 297 {
 298         const spill_t *p = a;
 299         const spill_t *q = b;
 300
 301 //      return !(p->irn == q->irn && p->bb == q->bb);
 302         return !(p->irn == q->irn);
 303 }
 304
 305 static int
 306 cmp_memoperands(const void *a, const void *b, size_t size)
 307 {
 308         const memoperand_t *p = a;
 309         const memoperand_t *q = b;
 310
 311         return !(p->irn == q->irn && p->pos == q->pos);
 312 }
 313
 314 static keyval_t *
 315 set_find_keyval(set * set, const void * key)
 316 {
 317         keyval_t     query;
 318
 319         query.key = key;
 320         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 321 }
 322
 323 static keyval_t *
 324 set_insert_keyval(set * set, void * key, void * val)
 325 {
 326         keyval_t     query;
 327
 328         query.key = key;
 329         query.val = val;
 330         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 331 }
 332
 333 static defs_t *
 334 set_find_def(set * set, const ir_node * value)
 335 {
 336         defs_t     query;
 337
 338         query.value = value;
 339         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 340 }
 341
 342 static defs_t *
 343 set_insert_def(set * set, const ir_node * value)
 344 {
 345         defs_t     query;
 346
 347         query.value = value;
 348         query.spills = NULL;
 349         query.remats = NULL;
 350         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 351 }
 352
 353 static memoperand_t *
 354 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 355 {
 356         memoperand_t     query;
 357
 358         query.irn = irn;
 359         query.pos = pos;
 360         query.ilp = ilp;
 361         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 362 }
 363
 364 static memoperand_t *
 365 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 366 {
 367         memoperand_t     query;
 368
 369         query.irn = (ir_node*)irn;
 370         query.pos = pos;
 371         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 372 }
 373
 374
 375 static spill_t *
 376 set_find_spill(set * set, const ir_node * value)
 377 {
 378         spill_t     query;
 379
 380         query.irn = (ir_node*)value;
 381         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 382 }
 383
 384 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 385 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 386 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 387 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 388 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 389
 390 static int
 391 cmp_remat_info(const void *a, const void *b, size_t size)
 392 {
 393         const remat_info_t *p = a;
 394         const remat_info_t *q = b;
 395
 396         return !(p->irn == q->irn);
 397 }
 398
 399 static int
 400 cmp_defs(const void *a, const void *b, size_t size)
 401 {
 402         const defs_t *p = a;
 403         const defs_t *q = b;
 404
 405         return !(p->value == q->value);
 406 }
 407
 408 static int
 409 cmp_keyval(const void *a, const void *b, size_t size)
 410 {
 411         const keyval_t *p = a;
 412         const keyval_t *q = b;
 413
 414         return !(p->key == q->key);
 415 }
 416
 417 static double
 418 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 419 {
 420 #define FUDGE 0.001
 421         if(be_profile_has_data())
 422                 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
 423
 424 #ifndef EXECFREQ_LOOPDEPH
 425         return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
 426 #else
 427         if(is_Block(irn))
 428                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 429         else
 430                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 431 #endif
 432 }
 433
 434 static double
 435 get_cost(const spill_ilp_t * si, const ir_node * irn)
 436 {
 437         if(be_is_Spill(irn)) {
 438                 return opt_cost_spill;
 439         } else if(be_is_Reload(irn)){
 440                 return opt_cost_reload;
 441         } else {
 442                 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
 443         }
 444 }
 445
 446 /**
 447  * Checks, whether node and its operands have suitable reg classes
 448  */
 449 static INLINE int
 450 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 451 {
 452         int               n;
 453         const arch_env_t *arch_env = si->birg->main_env->arch_env;
 454         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 455
 456 #if 0
 457         if(!remat)
 458                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 459 #endif
 460
 461         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 462                 ir_node        *op = get_irn_n(irn, n);
 463                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 464
 465 //              if(!remat)
 466 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 467         }
 468
 469         return remat;
 470 }
 471
 472 /**
 473  * Try to create a remat from @p op with destination value @p dest_value
 474  */
 475 static INLINE remat_t *
 476 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 477 {
 478         remat_t  *remat = NULL;
 479
 480 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 481 //              return NULL;
 482
 483         if(dest_value == op) {
 484                 const ir_node *proj = NULL;
 485
 486                 if(is_Proj(dest_value)) {
 487                         op = get_Proj_pred(op);
 488                         proj = dest_value;
 489                 }
 490
 491                 if(!is_rematerializable(si, op))
 492                         return NULL;
 493
 494                 remat          = obstack_alloc(si->obst, sizeof(*remat));
 495                 remat->op      = op;
 496                 remat->cost    = (int)get_cost(si, op);
 497                 remat->value   = dest_value;
 498                 remat->proj    = proj;
 499                 remat->inverse = 0;
 500         } else {
 501                 arch_inverse_t     inverse;
 502                 int                n;
 503
 504                 /* get the index of the operand we want to retrieve by the inverse op */
 505                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 506                         ir_node        *arg = get_irn_n(op, n);
 507
 508                         if(arg == dest_value) break;
 509                 }
 510                 if(n<0) return NULL;
 511
 512                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 513
 514                 /* else ask the backend to give an inverse op */
 515                 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 516                         int   i;
 517
 518                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 519
 520                         assert(inverse.n > 0 && "inverse op should have at least one node");
 521
 522                         for(i=inverse.n-1; i>=0; --i) {
 523                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 524                         }
 525
 526                         if(inverse.n <= 2) {
 527                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 528                                 remat->op = inverse.nodes[0];
 529                                 remat->cost = inverse.costs;
 530                                 remat->value = dest_value;
 531                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 532                                 remat->inverse = 1;
 533
 534                                 // Matze: commented this out, this doesn't seem to be true if
 535                                 // the inverse is a simple operation with only 1 result...
 536                                 //assert(is_Proj(remat->proj));
 537                         } else {
 538                                 assert(0 && "I can not handle remats with more than 2 nodes");
 539                         }
 540                 }
 541         }
 542
 543         if(remat) {
 544                 if(remat->proj) {
 545                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 546                 } else {
 547                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 548                 }
 549         }
 550         return remat;
 551 }
 552
 553
 554 static INLINE void
 555 add_remat(const spill_ilp_t * si, const remat_t * remat)
 556 {
 557         remat_info_t    *remat_info,
 558                      query;
 559         int              n;
 560
 561         assert(remat->op);
 562         assert(remat->value);
 563
 564         query.irn = remat->value;
 565         query.remats = NULL;
 566         query.remats_by_operand = NULL;
 567         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 568
 569         if(remat_info->remats == NULL) {
 570                 remat_info->remats = new_pset(cmp_remat, 4096);
 571         }
 572         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 573
 574         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 575         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 576                 ir_node        *arg = get_irn_n(remat->op, n);
 577
 578                 query.irn = arg;
 579                 query.remats = NULL;
 580                 query.remats_by_operand = NULL;
 581                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 582
 583                 if(remat_info->remats_by_operand == NULL) {
 584                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 585                 }
 586                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 587         }
 588 }
 589
 590 static int
 591 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 592 {
 593         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 594         int                i = 0;
 595
 596         while(edge) {
 597                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 598                         ++i;
 599                 }
 600                 edge = get_irn_out_edge_next(irn, edge);
 601         }
 602
 603         return i;
 604 }
 605
 606 static int
 607 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 608 {
 609         int n;
 610         int ret = 0;
 611
 612         if(is_Proj(irn))
 613                 irn = get_Proj_pred(irn);
 614
 615         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 616                 const ir_node  *arg = get_irn_n(irn, n);
 617
 618                 if(has_reg_class(si, arg)) ++ret;
 619         }
 620
 621         return ret;
 622 }
 623
 624 static INLINE void
 625 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 626 {
 627         int      n;
 628         remat_t *remat;
 629
 630         if( has_reg_class(si, op)
 631         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 632         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 633         ) {
 634                 remat = get_remat_from_op(si, op, op);
 635                 if(remat) {
 636                         add_remat(si, remat);
 637                 }
 638         }
 639
 640         if(opt_remats == REMATS_ALL) {
 641                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 642                    for each arg */
 643                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 644                         ir_node        *arg = get_irn_n(op, n);
 645
 646                         if(has_reg_class(si, arg)) {
 647                                 /* try to get an inverse remat */
 648                                 remat = get_remat_from_op(si, arg, op);
 649                                 if(remat) {
 650                                         add_remat(si, remat);
 651                                 }
 652                         }
 653                 }
 654         }
 655 }
 656
 657 static INLINE int
 658 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 659 {
 660         ir_node *block;
 661         ir_node *def_block = get_nodes_block(val);
 662         int      ret;
 663
 664         if(val == pos)
 665                 return 0;
 666
 667         /* if pos is at end of a basic block */
 668         if(is_Block(pos)) {
 669                 ret = (pos == def_block || block_dominates(def_block, pos));
 670 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 671                 return ret;
 672         }
 673
 674         /* else if this is a normal operation */
 675         block = get_nodes_block(pos);
 676         if(block == def_block) {
 677                 if(!sched_is_scheduled(val)) return 1;
 678
 679                 ret = sched_comes_after(val, pos);
 680 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 681                 return ret;
 682         }
 683
 684         ret = block_dominates(def_block, block);
 685 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 686         return ret;
 687 }
 688
 689 static INLINE ir_node *
 690 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 691 {
 692     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
 693 }
 694
 695 /**
 696  * Returns first non-Phi node of block @p bb
 697  */
 698 static INLINE ir_node *
 699 sched_block_first_nonphi(const ir_node * bb)
 700 {
 701         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 702 }
 703
 704 static int
 705 sched_skip_proj_predicator(const ir_node * irn, void * data)
 706 {
 707         return (is_Proj(irn));
 708 }
 709
 710 static INLINE ir_node *
 711 sched_next_nonproj(const ir_node * irn, int forward)
 712 {
 713         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 714 }
 715
 716 /**
 717  * Returns next operation node (non-Proj) after @p irn
 718  * or the basic block of this node
 719  */
 720 static INLINE ir_node *
 721 sched_next_op(const ir_node * irn)
 722 {
 723         ir_node *next = sched_next(irn);
 724
 725         if(is_Block(next))
 726                 return next;
 727
 728         return sched_next_nonproj(next, 1);
 729 }
 730
 731 /**
 732  * Returns previous operation node (non-Proj) before @p irn
 733  * or the basic block of this node
 734  */
 735 static INLINE ir_node *
 736 sched_prev_op(const ir_node * irn)
 737 {
 738         ir_node *prev = sched_prev(irn);
 739
 740         if(is_Block(prev))
 741                 return prev;
 742
 743         return sched_next_nonproj(prev, 0);
 744 }
 745
 746 static void
 747 sched_put_after(ir_node * insert, ir_node * irn)
 748 {
 749         if(is_Block(insert)) {
 750                 insert = sched_block_first_nonphi(insert);
 751         } else {
 752                 insert = sched_next_op(insert);
 753         }
 754         sched_reset(irn);
 755         sched_add_before(insert, irn);
 756 }
 757
 758 static void
 759 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 760 {
 761   if(is_Block(insert)) {
 762           insert = sched_block_last_noncf(si, insert);
 763   } else {
 764           insert = sched_next_nonproj(insert, 0);
 765           insert = sched_prev(insert);
 766   }
 767   sched_reset(irn);
 768   sched_add_after(insert, irn);
 769 }
 770
 771 static ir_node *
 772 next_post_remat(const ir_node * irn)
 773 {
 774         op_t      *op;
 775     ir_node   *next;
 776
 777         if(is_Block(irn)) {
 778                 next = sched_block_first_nonphi(irn);
 779         } else {
 780                 next = sched_next_op(irn);
 781         }
 782
 783         if(sched_is_end(next))
 784                 return NULL;
 785
 786         op = get_irn_link(next);
 787         if(op->is_remat && !op->attr.remat.pre) {
 788                 return next;
 789         }
 790
 791         return NULL;
 792 }
 793
 794
 795 static ir_node *
 796 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
 797 {
 798         op_t      *op;
 799         ir_node   *ret;
 800
 801         if(is_Block(irn)) {
 802                 ret = sched_block_last_noncf(si, irn);
 803                 ret = sched_next(ret);
 804                 ret = sched_prev_op(ret);
 805         } else {
 806                 ret = sched_prev_op(irn);
 807         }
 808
 809         if(sched_is_end(ret) || is_Phi(ret))
 810                 return NULL;
 811
 812         op = (op_t*)get_irn_link(ret);
 813         if(op->is_remat && op->attr.remat.pre) {
 814                 return ret;
 815         }
 816
 817         return NULL;
 818 }
 819
 820 /**
 821  * Tells you whether a @p remat can be placed before the irn @p pos
 822  */
 823 static INLINE int
 824 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 825 {
 826         const ir_node   *op = remat->op;
 827         const ir_node   *prev;
 828         int        n,
 829                            res = 1;
 830
 831         if(is_Block(pos)) {
 832                 prev = sched_block_last_noncf(si, pos);
 833                 prev = sched_next_nonproj(prev, 0);
 834         } else {
 835                 prev = sched_prev_op(pos);
 836         }
 837         /* do not remat if the rematted value is defined immediately before this op */
 838         if(prev == remat->op) {
 839                 return 0;
 840         }
 841
 842 #if 0
 843         /* this should be just fine, the following OP will be using this value, right? */
 844
 845         /* only remat AFTER the real definition of a value (?) */
 846         if(!value_is_defined_before(si, pos, remat->value)) {
 847 //              ir_fprintf(stderr, "error(not defined)");
 848                 return 0;
 849         }
 850 #endif
 851
 852         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 853                 const ir_node   *arg = get_irn_n(op, n);
 854
 855                 if(opt_no_enlarge_liveness) {
 856                         if(has_reg_class(si, arg) && live) {
 857                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 858                         } else {
 859                                 res &= value_is_defined_before(si, pos, arg);
 860                         }
 861                 } else {
 862                         res &= value_is_defined_before(si, pos, arg);
 863                 }
 864         }
 865
 866         return res;
 867 }
 868
 869 /**
 870  * Tells you whether a @p remat can be placed after the irn @p pos
 871  */
 872 static INLINE int
 873 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 874 {
 875         if(is_Block(pos)) {
 876                 pos = sched_block_first_nonphi(pos);
 877         } else {
 878                 pos = sched_next_op(pos);
 879         }
 880
 881         /* only remat AFTER the real definition of a value (?) */
 882         if(!value_is_defined_before(si, pos, remat->value)) {
 883                 return 0;
 884         }
 885
 886         return can_remat_before(si, remat, pos, live);
 887 }
 888
 889 /**
 890  * Collect potetially rematerializable OPs
 891  */
 892 static void
 893 walker_remat_collector(ir_node * irn, void * data)
 894 {
 895         spill_ilp_t    *si = data;
 896
 897         if(!is_Block(irn) && !is_Phi(irn)) {
 898                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 899                 get_remats_from_op(si, irn);
 900         }
 901 }
 902
 903 /**
 904  * Inserts a copy of @p irn before @p pos
 905  */
 906 static ir_node *
 907 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 908 {
 909         ir_node     *bb;
 910         ir_node     *copy;
 911
 912         bb = is_Block(pos)?pos:get_nodes_block(pos);
 913         copy = exact_copy(irn);
 914
 915         set_phi_class(si->pc, copy, NULL);
 916         set_nodes_block(copy, bb);
 917         sched_put_before(si, pos, copy);
 918
 919         return copy;
 920 }
 921
 922 /**
 923  * Inserts a copy of @p irn after @p pos
 924  */
 925 static ir_node *
 926 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 927 {
 928         ir_node     *bb;
 929         ir_node     *copy;
 930
 931         bb = is_Block(pos)?pos:get_nodes_block(pos);
 932         copy = exact_copy(irn);
 933
 934         set_phi_class(si->pc, copy, NULL);
 935         set_nodes_block(copy, bb);
 936         sched_put_after(pos, copy);
 937
 938         return copy;
 939 }
 940
 941 static ir_node *
 942 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 943 {
 944         char     buf[256];
 945
 946         if(can_remat_after(si, remat, pos, live)) {
 947                 ir_node         *copy,
 948                                                 *proj_copy;
 949                 op_t            *op;
 950
 951                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 952
 953                 copy = insert_copy_after(si, remat->op, pos);
 954
 955                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 956                 op = obstack_alloc(si->obst, sizeof(*op));
 957                 op->is_remat = 1;
 958                 op->attr.remat.remat = remat;
 959                 op->attr.remat.pre = 0;
 960                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 961
 962                 set_irn_link(copy, op);
 963                 pset_insert_ptr(si->all_possible_remats, copy);
 964                 if(remat->proj) {
 965                         proj_copy = insert_copy_after(si, remat->proj, copy);
 966                         set_irn_n(proj_copy, 0, copy);
 967                         set_irn_link(proj_copy, op);
 968                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 969                 } else {
 970                         proj_copy = NULL;
 971                 }
 972
 973                 return copy;
 974         }
 975
 976         return NULL;
 977 }
 978
 979 static ir_node *
 980 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 981 {
 982         char     buf[256];
 983
 984         if(can_remat_before(si, remat, pos, live)) {
 985                 ir_node         *copy,
 986                                                 *proj_copy;
 987                 op_t            *op;
 988
 989                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 990
 991                 copy = insert_copy_before(si, remat->op, pos);
 992
 993                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 994                 op = obstack_alloc(si->obst, sizeof(*op));
 995                 op->is_remat = 1;
 996                 op->attr.remat.remat = remat;
 997                 op->attr.remat.pre = 1;
 998                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 999
1000                 set_irn_link(copy, op);
1001                 pset_insert_ptr(si->all_possible_remats, copy);
1002                 if(remat->proj) {
1003                         proj_copy = insert_copy_after(si, remat->proj, copy);
1004                         set_irn_n(proj_copy, 0, copy);
1005                         set_irn_link(proj_copy, op);
1006                         pset_insert_ptr(si->all_possible_remats, proj_copy);
1007                 } else {
1008                         proj_copy = NULL;
1009                 }
1010
1011                 return copy;
1012         }
1013
1014         return NULL;
1015 }
1016
1017 static int
1018 get_block_n_succs(const ir_node *block) {
1019         const ir_edge_t *edge;
1020
1021         assert(edges_activated(current_ir_graph));
1022
1023         edge = get_block_succ_first(block);
1024         if (! edge)
1025                 return 0;
1026
1027         edge = get_block_succ_next(block, edge);
1028         return edge ? 2 : 1;
1029 }
1030
1031 static int
1032 is_start_block(const ir_node * bb)
1033 {
1034         return get_irg_start_block(get_irn_irg(bb)) == bb;
1035 }
1036
1037 static int
1038 is_merge_edge(const ir_node * bb)
1039 {
1040         if(is_start_block(bb))
1041                 return 0;
1042
1043         if(opt_goodwin)
1044                 return get_block_n_succs(bb) == 1;
1045         else
1046                 return 1;
1047 }
1048
1049 static int
1050 is_diverge_edge(const ir_node * bb)
1051 {
1052         if(is_start_block(bb))
1053                 return 0;
1054
1055         if(opt_goodwin)
1056                 return get_Block_n_cfgpreds(bb) == 1;
1057         else
1058                 return 1;
1059 }
1060
1061 static void
1062 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1063 {
1064         ir_node        *irn;
1065         int i;
1066
1067         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1068                 irn = be_lv_get_irn(si->lv, bb, i);
1069
1070                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1071                         pset_insert_ptr(live, irn);
1072                 }
1073         }
1074
1075         irn = sched_last(bb);
1076
1077         /* all values eaten by control flow operations are also live until the end of the block */
1078         sched_foreach_reverse(bb, irn) {
1079                 int  i;
1080
1081                 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1082
1083                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1084                         ir_node *arg = get_irn_n(irn,i);
1085
1086                         if(has_reg_class(si, arg)) {
1087                                 pset_insert_ptr(live, arg);
1088                         }
1089                 }
1090         }
1091         /*
1092          * find values that are used by remats at end of block
1093          * and insert them into live set
1094          */
1095         foreach_pre_remat(si, bb, irn) {
1096                 int       n;
1097
1098                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1099                         ir_node        *remat_arg = get_irn_n(irn, n);
1100
1101                         if(!has_reg_class(si, remat_arg)) continue;
1102
1103                         /* if value is becoming live through use by remat */
1104                         if(!pset_find_ptr(live, remat_arg)) {
1105                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1106
1107                                 pset_insert_ptr(live, remat_arg);
1108                         }
1109                 }
1110         }
1111 }
1112
1113 static void
1114 walker_regclass_copy_insertor(ir_node * irn, void * data)
1115 {
1116         spill_ilp_t    *si = data;
1117
1118         if(is_Phi(irn) && has_reg_class(si, irn)) {
1119                 int n;
1120
1121                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1122                         ir_node  *phi_arg = get_irn_n(irn, n);
1123                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1124
1125                         if(!has_reg_class(si, phi_arg)) {
1126                                 ir_node   *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1127                                 ir_node   *pos = sched_block_last_noncf(si, bb);
1128                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1129
1130                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1131                                 sched_add_after(pos, copy);
1132                                 set_irn_n(irn, n, copy);
1133
1134                                 op->is_remat = 0;
1135                                 op->attr.live_range.args.reloads = NULL;
1136                                 op->attr.live_range.ilp = ILP_UNDEF;
1137                                 set_irn_link(copy, op);
1138                         }
1139                 }
1140         }
1141 }
1142
1143 /**
1144  * Insert (so far unused) remats into the irg to
1145  * recompute the potential liveness of all values
1146  */
1147 static void
1148 walker_remat_insertor(ir_node * bb, void * data)
1149 {
1150         spill_ilp_t    *si = data;
1151         ir_node        *irn;
1152         int             n, i;
1153         pset           *live;
1154         pset           *post_remats;
1155         remat_t        *remat;
1156
1157         /* skip start block, no remats to do there */
1158         if(is_start_block(bb)) return;
1159
1160         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1161
1162         live = pset_new_ptr_default();
1163         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1164                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1165
1166                 /* add remats at end of block */
1167                 if (has_reg_class(si, value)) {
1168                         pset_insert_ptr(live, value);
1169                 }
1170         }
1171
1172         irn = sched_last(bb);
1173         while(!sched_is_end(irn)) {
1174                 ir_node   *next;
1175                 pset      *args;
1176                 ir_node   *arg;
1177                 pset      *used;
1178
1179                 next = sched_prev(irn);
1180
1181                 /* delete defined value from live set */
1182                 if(has_reg_class(si, irn)) {
1183                         pset_remove_ptr(live, irn);
1184                 }
1185
1186                 if(is_Phi(irn) || is_Proj(irn)) {
1187                         irn = next;
1188                         continue;
1189                 }
1190
1191                 args = pset_new_ptr_default();
1192                 used = pset_new_ptr_default();
1193
1194                 /* collect arguments of op and set args of op already live in epilog */
1195                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1196                         ir_node        *arg = get_irn_n(irn, n);
1197
1198                         pset_insert_ptr(args, arg);
1199                         if(has_reg_class(si, arg)) {
1200                                 pset_insert_ptr(live, arg);
1201                                 pset_insert_ptr(used, arg);
1202                         }
1203                 }
1204
1205                 /* insert all possible remats before irn */
1206                 pset_foreach(args, arg) {
1207                         remat_info_t   *remat_info,
1208                                                     query;
1209
1210                         /* continue if the operand has the wrong reg class */
1211                         if(!has_reg_class(si, arg))
1212                                 continue;
1213
1214                         query.irn = arg;
1215                         query.remats = NULL;
1216                         query.remats_by_operand = NULL;
1217                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1218
1219                         if(!remat_info) {
1220                                 continue;
1221                         }
1222
1223                         if(remat_info->remats) {
1224                                 pset_foreach(remat_info->remats, remat) {
1225                                         ir_node  *remat_irn = NULL;
1226
1227                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1228                                         remat_irn = insert_remat_before(si, remat, irn, live);
1229
1230                                         if(remat_irn) {
1231                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1232                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1233
1234                                                         /* collect args of remats which are not args of op */
1235                                                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1236                                                                 pset_insert_ptr(used, remat_arg);
1237                                                         }
1238                                                 }
1239                                         }
1240                                 }
1241                         }
1242                 }
1243
1244                 /* do not place post remats after jumps */
1245                 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1246                         del_pset(used);
1247                         del_pset(args);
1248                         break;
1249                 }
1250
1251                 /* insert all possible remats after irn */
1252                 post_remats = pset_new_ptr_default();
1253                 pset_foreach(used, arg) {
1254                         remat_info_t   *remat_info,
1255                                                     query;
1256
1257                         /* continue if the operand has the wrong reg class */
1258                         if(!has_reg_class(si, arg))
1259                                 continue;
1260
1261                         query.irn = arg;
1262                         query.remats = NULL;
1263                         query.remats_by_operand = NULL;
1264                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1265
1266                         if(!remat_info) {
1267                                 continue;
1268                         }
1269
1270                         if(remat_info->remats_by_operand) {
1271                                 pset_foreach(remat_info->remats_by_operand, remat) {
1272                                         /* do not insert remats producing the same value as one of the operands */
1273                                         if(!pset_find_ptr(args, remat->value)) {
1274                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1275
1276                                                 /* only remat values that can be used by real ops */
1277                                                 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1278                                                         pset_insert_ptr(post_remats, remat);
1279                                                 }
1280                                         }
1281                                 }
1282                         }
1283                 }
1284                 pset_foreach(post_remats, remat) {
1285                         insert_remat_after(si, remat, irn, live);
1286                 }
1287                 del_pset(post_remats);
1288
1289                 del_pset(used);
1290                 del_pset(args);
1291                 irn = next;
1292         }
1293
1294         /* add remats at end if successor has multiple predecessors */
1295         if(is_merge_edge(bb)) {
1296                 pset     *live_out = pset_new_ptr_default();
1297                 ir_node  *value;
1298
1299                 get_live_end(si, bb, live_out);
1300
1301                 /* add remats at end of block */
1302                 pset_foreach(live_out, value) {
1303                         remat_info_t   *remat_info,
1304                                                    query;
1305
1306                         query.irn = value;
1307                         query.remats = NULL;
1308                         query.remats_by_operand = NULL;
1309                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1310
1311                         if(remat_info && remat_info->remats) {
1312                                 pset_foreach(remat_info->remats, remat) {
1313                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1314
1315                                         insert_remat_before(si, remat, bb, live_out);
1316                                 }
1317                         }
1318                 }
1319                 del_pset(live_out);
1320         }
1321
1322         if(is_diverge_edge(bb)) {
1323                 pset     *live_in = pset_new_ptr_default();
1324                 ir_node  *value;
1325
1326                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1327                         value = be_lv_get_irn(si->lv, bb, i);
1328
1329                         if(has_reg_class(si, value)) {
1330                                 pset_insert_ptr(live_in, value);
1331                         }
1332                 }
1333                 /* add phis to live_in */
1334                 sched_foreach(bb, value) {
1335                         if(!is_Phi(value)) break;
1336
1337                         if(has_reg_class(si, value)) {
1338                                 pset_insert_ptr(live_in, value);
1339                         }
1340                 }
1341
1342                 /* add remat2s at beginning of block */
1343                 post_remats = pset_new_ptr_default();
1344                 pset_foreach(live_in, value) {
1345                         remat_info_t   *remat_info,
1346                                                    query;
1347
1348                         query.irn = value;
1349                         query.remats = NULL;
1350                         query.remats_by_operand = NULL;
1351                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1352
1353                         if(remat_info && remat_info->remats_by_operand) {
1354                                 pset_foreach(remat_info->remats_by_operand, remat) {
1355                                         DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1356
1357                                         /* put the remat here if all its args are available and result is still live */
1358                                         if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1359                                                 pset_insert_ptr(post_remats, remat);
1360                                         }
1361                                 }
1362                         }
1363                 }
1364                 pset_foreach(post_remats, remat) {
1365                         insert_remat_after(si, remat, bb, live_in);
1366                 }
1367                 del_pset(post_remats);
1368                 del_pset(live_in);
1369         }
1370 }
1371
1372 static int
1373 can_be_copied(const ir_node * bb, const ir_node * irn)
1374 {
1375         const ir_edge_t *edge    = get_block_succ_first(bb);
1376         const ir_node   *next_bb = edge->src;
1377         int             pos      = edge->pos;
1378         const ir_node   *phi;
1379
1380         assert(is_merge_edge(bb));
1381
1382         sched_foreach(next_bb, phi) {
1383                 const ir_node  *phi_arg;
1384
1385                 if(!is_Phi(phi)) break;
1386
1387                 phi_arg = get_irn_n(phi, pos);
1388
1389                 if(phi_arg == irn) {
1390                         return 1;
1391                 }
1392         }
1393         return 0;
1394 }
1395
1396 /**
1397  * Initialize additional node info
1398  */
1399 static void
1400 luke_initializer(ir_node * bb, void * data)
1401 {
1402         spill_ilp_t    *si = (spill_ilp_t*)data;
1403         spill_bb_t     *spill_bb;
1404         ir_node        *irn;
1405
1406         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1407         set_irn_link(bb, spill_bb);
1408
1409         sched_foreach(bb, irn) {
1410                 op_t      *op;
1411
1412                 op = obstack_alloc(si->obst, sizeof(*op));
1413                 op->is_remat = 0;
1414                 op->attr.live_range.ilp = ILP_UNDEF;
1415                 if(is_Phi(irn)) {
1416                         if(opt_memcopies) {
1417                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1418                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1419                         }
1420                 } else if(!is_Proj(irn)) {
1421                         op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1422                         memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1423                 } else {
1424                         op->attr.live_range.args.reloads = NULL;
1425                 }
1426                 set_irn_link(irn, op);
1427         }
1428 }
1429
1430
1431 /**
1432  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1433  */
1434 static void
1435 luke_endwalker(ir_node * bb, void * data)
1436 {
1437         spill_ilp_t    *si = (spill_ilp_t*)data;
1438         pset           *live;
1439         pset           *use_end;
1440         char            buf[256];
1441         ilp_cst_t       cst;
1442         ir_node        *irn;
1443         spill_bb_t     *spill_bb = get_irn_link(bb);
1444         int             i;
1445
1446         live = pset_new_ptr_default();
1447         use_end = pset_new_ptr_default();
1448
1449         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1450                 irn = be_lv_get_irn(si->lv, bb, i);
1451                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1452                         pset_insert_ptr(live, irn);
1453                 }
1454         }
1455         /*
1456          * find values that are used by remats at end of block
1457          * and insert them into live set
1458          */
1459         foreach_pre_remat(si, bb, irn) {
1460                 int       n;
1461
1462                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1463                         ir_node        *remat_arg = get_irn_n(irn, n);
1464
1465                         if(has_reg_class(si, remat_arg)) {
1466                                 pset_insert_ptr(live, remat_arg);
1467                         }
1468                 }
1469         }
1470
1471         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1472         /* their reg_out must always be set */
1473         sched_foreach_reverse(bb, irn) {
1474                 int   n;
1475
1476                 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1477
1478                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1479                         ir_node        *irn_arg = get_irn_n(irn, n);
1480
1481                         if(has_reg_class(si, irn_arg)) {
1482                                 pset_insert_ptr(use_end, irn_arg);
1483                         }
1484                 }
1485         }
1486
1487         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1488         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1489         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1490
1491         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1492
1493         /* if this is a merge edge we can reload at the end of this block */
1494         if(is_merge_edge(bb)) {
1495                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1496         } else if(pset_count(use_end)){
1497                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1498         } else {
1499                 spill_bb->reloads = NULL;
1500         }
1501
1502         pset_foreach(live,irn) {
1503                 spill_t     query,
1504                                         *spill;
1505                 double      spill_cost;
1506                 int         default_spilled;
1507
1508
1509                 /* handle values used by control flow nodes later separately */
1510                 if(pset_find_ptr(use_end, irn)) continue;
1511
1512                 query.irn = irn;
1513                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1514
1515                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1516
1517                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1518                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1519                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1520
1521                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1522                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1523
1524                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1525                 /* by default spill value right after definition */
1526                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1527                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1528
1529                 if(is_merge_edge(bb)) {
1530                         ilp_var_t   reload;
1531                         ilp_cst_t   rel_cst;
1532
1533                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1534                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1535                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1536
1537                         /* reload <= mem_out */
1538                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1539                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1540                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1541                 }
1542
1543                 spill->reg_in = ILP_UNDEF;
1544                 spill->mem_in = ILP_UNDEF;
1545         }
1546
1547         pset_foreach(use_end,irn) {
1548                 spill_t     query,
1549                                         *spill;
1550                 double      spill_cost;
1551                 ilp_cst_t   end_use_req,
1552                                         rel_cst;
1553                 ilp_var_t   reload;
1554                 int         default_spilled;
1555
1556                 query.irn = irn;
1557                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1558
1559                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1560
1561                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1562                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1563
1564                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1565                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1566
1567                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1568                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1569                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1570
1571                 /* reload for use be control flow op */
1572                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1573                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1574                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1575
1576                 /* reload <= mem_out */
1577                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1578                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1579                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1580
1581                 spill->reg_in = ILP_UNDEF;
1582                 spill->mem_in = ILP_UNDEF;
1583
1584                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1585                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1586                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1587         }
1588
1589         del_pset(live);
1590         del_pset(use_end);
1591 }
1592
1593 #ifndef NDEBUG
1594 /**
1595  * Find a remat of value @p value in the epilog of @p pos
1596  */
1597 static ir_node *
1598 find_post_remat(const ir_node * value, const ir_node * pos)
1599 {
1600         while((pos = next_post_remat(pos)) != NULL) {
1601                 op_t   *op;
1602
1603                 op = get_irn_link(pos);
1604                 assert(op->is_remat && !op->attr.remat.pre);
1605
1606                 if(op->attr.remat.remat->value == value)
1607                         return (ir_node*)pos;
1608
1609 #if 0
1610         const ir_edge_t *edge;
1611                 foreach_out_edge(pos, edge) {
1612                         ir_node   *proj = get_edge_src_irn(edge);
1613                         assert(is_Proj(proj));
1614                 }
1615 #endif
1616
1617         }
1618
1619         return NULL;
1620 }
1621 #endif
1622
1623 static spill_t *
1624 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1625 {
1626         spill_bb_t  *spill_bb = get_irn_link(bb);
1627         spill_t     *spill,
1628                                  query;
1629         char         buf[256];
1630         int          default_spilled;
1631
1632         query.irn = irn;
1633         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1634         if(!spill) {
1635                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1636
1637                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1638
1639                 spill->reg_out = ILP_UNDEF;
1640                 spill->reg_in  = ILP_UNDEF;
1641                 spill->mem_in  = ILP_UNDEF;
1642
1643                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1644                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1645
1646                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1647                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1648                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1649         }
1650
1651         return spill;
1652 }
1653
1654 /**
1655  *  Inserts ILP-constraints and variables for memory copying before the given position
1656  */
1657 static void
1658 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1659 {
1660         const ir_node    *succ;
1661         const ir_edge_t  *edge;
1662         spill_bb_t       *spill_bb = get_irn_link(block);
1663         ir_node          *phi;
1664         int               pos;
1665         ilp_cst_t         cst;
1666         ilp_var_t         copyreg;
1667         char              buf[256];
1668         ir_node          *tmp;
1669
1670
1671         assert(edges_activated(current_ir_graph));
1672
1673         edge = get_block_succ_first(block);
1674         if(!edge) return;
1675
1676         succ = edge->src;
1677         pos = edge->pos;
1678
1679         edge = get_block_succ_next(block, edge);
1680         /* next block can only contain phis, if this is a merge edge */
1681         if(edge) return;
1682
1683         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1684         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1685
1686         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1687         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1688
1689         pset_foreach(live, tmp) {
1690                 spill_t  *spill;
1691 #if 0
1692                 op_t  *op = get_irn_link(irn);
1693                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1694 #endif
1695                 spill = set_find_spill(spill_bb->ilp, tmp);
1696                 assert(spill);
1697
1698                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1699         }
1700         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1701
1702         sched_foreach(succ, phi) {
1703                 const ir_node  *to_copy;
1704                 op_t           *to_copy_op;
1705                 spill_t        *to_copy_spill;
1706                 op_t           *phi_op = get_irn_link(phi);
1707                 ilp_var_t       reload = ILP_UNDEF;
1708
1709
1710                 if(!is_Phi(phi)) break;
1711                 if(!has_reg_class(si, phi)) continue;
1712
1713                 to_copy = get_irn_n(phi, pos);
1714                 to_copy_op = get_irn_link(to_copy);
1715
1716                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1717                 assert(to_copy_spill);
1718
1719                 if(spill_bb->reloads) {
1720                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1721
1722                         if(keyval) {
1723                                 reload = PTR_TO_INT(keyval->val);
1724                         }
1725                 }
1726
1727                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1728                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1729
1730                 /* copy - reg_out - reload - remat - live_range <= 0 */
1731                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1732                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1733                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1734                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1735                 foreach_pre_remat(si, block, tmp) {
1736                         op_t     *remat_op = get_irn_link(tmp);
1737                         if(remat_op->attr.remat.remat->value == to_copy) {
1738                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1739                         }
1740                 }
1741
1742                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1743                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1744
1745                 /* copy - reg_out - copyreg <= 0 */
1746                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1747                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1748                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1749         }
1750 }
1751
1752
1753 /**
1754  * Walk all irg blocks and emit this ILP
1755  */
1756 static void
1757 luke_blockwalker(ir_node * bb, void * data)
1758 {
1759         spill_ilp_t    *si = (spill_ilp_t*)data;
1760         ir_node        *irn;
1761         pset           *live;
1762         char            buf[256];
1763         ilp_cst_t       cst;
1764         spill_bb_t     *spill_bb = get_irn_link(bb);
1765         ir_node        *tmp;
1766         spill_t        *spill;
1767         pset           *defs = pset_new_ptr_default();
1768         const arch_env_t *arch_env = si->birg->main_env->arch_env;
1769
1770         live = pset_new_ptr_default();
1771
1772         /****************************************
1773          *      B A S I C  B L O C K  E N D
1774          ***************************************/
1775
1776
1777         /* init live values at end of block */
1778         get_live_end(si, bb, live);
1779
1780         pset_foreach(live, irn) {
1781                 op_t           *op;
1782                 ilp_var_t       reload = ILP_UNDEF;
1783
1784                 spill = set_find_spill(spill_bb->ilp, irn);
1785                 assert(spill);
1786
1787                 if(spill_bb->reloads) {
1788                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1789
1790                         if(keyval) {
1791                                 reload = PTR_TO_INT(keyval->val);
1792                         }
1793                 }
1794
1795                 op = get_irn_link(irn);
1796                 assert(!op->is_remat);
1797
1798                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1799                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1800                 op->attr.live_range.op = bb;
1801
1802                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1803                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1804
1805                 /* reg_out - reload - remat - live_range <= 0 */
1806                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1807                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1808                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1809                 foreach_pre_remat(si, bb, tmp) {
1810                         op_t     *remat_op = get_irn_link(tmp);
1811                         if(remat_op->attr.remat.remat->value == irn) {
1812                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1813                         }
1814                 }
1815                 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1816                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1817
1818                 /* value may only die at bb end if it is used for a mem copy */
1819                 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1820                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1821                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1822                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1823                 foreach_pre_remat(si, bb, tmp) {
1824                         op_t     *remat_op = get_irn_link(tmp);
1825                         if(remat_op->attr.remat.remat->value == irn) {
1826                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1827                         }
1828                 }
1829                 if(is_merge_edge(bb)) {
1830                         const ir_edge_t *edge = get_block_succ_first(bb);
1831                         const ir_node   *next_bb = edge->src;
1832                         int              pos = edge->pos;
1833                         const ir_node   *phi;
1834
1835                         sched_foreach(next_bb, phi) {
1836                                 const ir_node  *phi_arg;
1837
1838                                 if(!is_Phi(phi)) break;
1839
1840                                 phi_arg = get_irn_n(phi, pos);
1841
1842                                 if(phi_arg == irn) {
1843                                         op_t      *phi_op = get_irn_link(phi);
1844                                         ilp_var_t  copy = phi_op->attr.live_range.args.copies[pos];
1845
1846                                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1847                                 }
1848                         }
1849                 }
1850         }
1851
1852         if(opt_memcopies)
1853                 insert_mem_copy_position(si, live, bb);
1854
1855         /*
1856          * assure the remat args are available
1857          */
1858         foreach_pre_remat(si, bb, tmp) {
1859                 op_t     *remat_op = get_irn_link(tmp);
1860                 int       n;
1861
1862                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1863                         ir_node        *remat_arg = get_irn_n(tmp, n);
1864                         op_t           *arg_op = get_irn_link(remat_arg);
1865
1866                         if(!has_reg_class(si, remat_arg)) continue;
1867
1868                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1869                         assert(spill);
1870
1871                         /* arguments of remats have to be live until the very end of the block
1872                          * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1873                          * no remats, they could be in wrong order
1874                          */
1875
1876                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1877                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1878
1879                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1880                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1881                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1882
1883                         /* use reload placed for this argument */
1884                         if(spill_bb->reloads) {
1885                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1886
1887                                 if(keyval) {
1888                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1889
1890                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1891                                 }
1892                         }
1893                 }
1894         }
1895         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1896
1897
1898
1899
1900         /**************************************
1901          *    B A S I C  B L O C K  B O D Y
1902          **************************************/
1903
1904         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1905                 op_t       *op;
1906                 op_t       *tmp_op;
1907                 int         n,
1908                                         u = 0,
1909                                         d = 0;
1910                 ilp_cst_t       check_pre,
1911                                         check_post;
1912                 set        *args;
1913                 pset       *used;
1914                 pset       *remat_defs;
1915                 keyval_t   *keyval;
1916                 ilp_cst_t   one_memoperand = -1;
1917
1918                 /* iterate only until first phi */
1919                 if(is_Phi(irn))
1920                         break;
1921
1922                 op = get_irn_link(irn);
1923                 /* skip remats */
1924                 if(op->is_remat) continue;
1925
1926                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1927
1928                 /* collect defined values */
1929                 if(has_reg_class(si, irn)) {
1930                         pset_insert_ptr(defs, irn);
1931                 }
1932
1933                 /* skip projs */
1934                 if(is_Proj(irn)) continue;
1935
1936                 /*
1937                  * init set of irn's arguments
1938                  * and all possibly used values around this op
1939                  * and values defined by post remats
1940                  */
1941                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1942                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1943                 remat_defs = pset_new_ptr(pset_count(live));
1944
1945                 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1946                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
1947                                 ir_node        *irn_arg = get_irn_n(irn, n);
1948                                 if(has_reg_class(si, irn_arg)) {
1949                                         set_insert_keyval(args, irn_arg, (void*)n);
1950                                         pset_insert_ptr(used, irn_arg);
1951                                 }
1952                         }
1953                         foreach_post_remat(irn, tmp) {
1954                                 op_t    *remat_op = get_irn_link(tmp);
1955
1956                                 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1957
1958                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1959                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1960                                         if(has_reg_class(si, remat_arg)) {
1961                                                 pset_insert_ptr(used, remat_arg);
1962                                         }
1963                                 }
1964                         }
1965                         foreach_pre_remat(si, irn, tmp) {
1966                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1967                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1968                                         if(has_reg_class(si, remat_arg)) {
1969                                                 pset_insert_ptr(used, remat_arg);
1970                                         }
1971                                 }
1972                         }
1973                 }
1974
1975                 /**********************************
1976                  *   I N  E P I L O G  O F  irn
1977                  **********************************/
1978
1979                 /* ensure each dying value is used by only one post remat */
1980                 pset_foreach(used, tmp) {
1981                         ir_node     *value = tmp;
1982                         op_t        *value_op = get_irn_link(value);
1983                         ir_node     *remat;
1984                         int          n_remats = 0;
1985
1986                         cst = ILP_UNDEF;
1987                         foreach_post_remat(irn, remat) {
1988                                 op_t  *remat_op = get_irn_link(remat);
1989
1990                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1991                                         ir_node   *remat_arg = get_irn_n(remat, n);
1992
1993                                         /* if value is used by this remat add it to constraint */
1994                                         if(remat_arg == value) {
1995                                                 if(n_remats == 0) {
1996                                                         /* sum remat2s <= 1 + n_remats*live_range */
1997                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1998                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1999                                                 }
2000
2001                                                 n_remats++;
2002                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2003                                                 break;
2004                                         }
2005                                 }
2006                         }
2007
2008                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2009                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2010                         }
2011                 }
2012
2013         /* ensure at least one value dies at post remat */
2014         foreach_post_remat(irn, tmp) {
2015             op_t     *remat_op = get_irn_link(tmp);
2016             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
2017             ir_node  *remat_arg;
2018
2019             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2020                 remat_arg = get_irn_n(tmp, n);
2021
2022                 if(has_reg_class(si, remat_arg)) {
2023
2024                     /* does arg always die at this op? */
2025                     if(!pset_find_ptr(live, remat_arg))
2026                         goto skip_one_must_die;
2027
2028                     pset_insert_ptr(remat_args, remat_arg);
2029                 }
2030             }
2031
2032             /* remat + \sum live_range(remat_arg) <= |args| */
2033             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2034             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2035             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2036
2037             pset_foreach(remat_args, remat_arg) {
2038                 op_t  *arg_op = get_irn_link(remat_arg);
2039
2040                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2041             }
2042
2043 skip_one_must_die:
2044             del_pset(remat_args);
2045         }
2046
2047                 /* new live ranges for values from L\U defined by post remats */
2048                 pset_foreach(live, tmp) {
2049                         ir_node     *value = tmp;
2050                         op_t        *value_op = get_irn_link(value);
2051
2052                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2053                                 ilp_var_t    prev_lr = ILP_UNDEF;
2054                                 ir_node     *remat;
2055
2056                                 if(pset_find_ptr(remat_defs, value)) {
2057
2058                                         /* next_live_range <= prev_live_range + sum remat2s */
2059                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2060                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2061
2062                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2063                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2064
2065                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2066                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2067
2068                                         foreach_post_remat(irn, remat) {
2069                                                 op_t        *remat_op = get_irn_link(remat);
2070
2071                                                 /* if value is being rematerialized by this remat */
2072                                                 if(value == remat_op->attr.remat.remat->value) {
2073                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2074                                                 }
2075                                         }
2076
2077                                         value_op->attr.live_range.ilp = prev_lr;
2078                                         value_op->attr.live_range.op = irn;
2079                                 }
2080                         }
2081                 }
2082
2083                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2084                 foreach_post_remat(irn, tmp) {
2085                         op_t        *remat_op = get_irn_link(tmp);
2086                         int          n;
2087
2088                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2089                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2090                                 op_t           *arg_op = get_irn_link(remat_arg);
2091
2092                                 if(!has_reg_class(si, remat_arg)) continue;
2093
2094                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2095                                 if(!pset_find_ptr(used, remat_arg)) {
2096                                         /* remat <= live_range(remat_arg) */
2097                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2098                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2099
2100                                         /* if value is becoming live through use by remat2 */
2101                                         if(!pset_find_ptr(live, remat_arg)) {
2102                                                 ilp_var_t     lr;
2103
2104                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2105                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2106
2107                                                 arg_op->attr.live_range.ilp = lr;
2108                                                 arg_op->attr.live_range.op = irn;
2109
2110                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2111
2112                                                 pset_insert_ptr(live, remat_arg);
2113                                                 add_to_spill_bb(si, bb, remat_arg);
2114                                         }
2115
2116                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2117                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2118                                 }
2119                         }
2120                 }
2121
2122                 d = pset_count(defs);
2123                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2124
2125                 /* count how many regs irn needs for arguments */
2126                 u = set_count(args);
2127
2128
2129                 /* check the register pressure in the epilog */
2130                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2131                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2132                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2133
2134                 /* add L\U' to check_post */
2135                 pset_foreach(live, tmp) {
2136                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2137                                 /* if a live value is not used by irn */
2138                                 tmp_op = get_irn_link(tmp);
2139                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2140                         }
2141                 }
2142
2143                 /***********************************************************
2144                  *  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
2145                  **********************************************************/
2146
2147
2148                 pset_foreach(used, tmp) {
2149                         ilp_var_t       prev_lr;
2150                         ilp_var_t       post_use;
2151                         int             p = 0;
2152                         spill_t        *spill;
2153                         ir_node        *arg = tmp;
2154                         op_t           *arg_op = get_irn_link(arg);
2155                         ir_node        *remat;
2156
2157                         spill = add_to_spill_bb(si, bb, arg);
2158
2159                         /* new live range for each used value */
2160                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2161                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2162
2163                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2164                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2165                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2166
2167                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2168
2169                         /* arg is live throughout epilog if the next live_range is in a register */
2170                         if(pset_find_ptr(live, arg)) {
2171                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2172
2173                                 /* post_use >= next_lr + remat */
2174                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2175                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2176                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2177                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2178                         }
2179
2180                         /* forall post remat which use arg add a similar cst */
2181                         foreach_post_remat(irn, remat) {
2182                                 int      n;
2183
2184                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2185                                         ir_node    *remat_arg = get_irn_n(remat, n);
2186                                         op_t       *remat_op = get_irn_link(remat);
2187
2188                                         if(remat_arg == arg) {
2189                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2190
2191                                                 /* post_use >= remat */
2192                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2193                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2194                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2195                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2196                                         }
2197                                 }
2198                         }
2199
2200                         /* if value is not an arg of op and not possibly defined by post remat
2201                          * then it may only die and not become live
2202                          */
2203                         if(!set_find_keyval(args, arg)) {
2204                                 /* post_use <= prev_lr */
2205                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2206                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2207                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2208                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2209
2210                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2211                                         /* next_lr <= prev_lr */
2212                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2213                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2214                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2215                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2216                                 }
2217                         }
2218
2219                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2220                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2221                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2222                                                 ilp_var_t       memoperand;
2223
2224                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2225                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2226                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2227
2228                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2229                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2230
2231                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2232                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2233                                         }
2234                                 }
2235                         }
2236
2237                         /* new live range begins for each used value */
2238                         arg_op->attr.live_range.ilp = prev_lr;
2239                         arg_op->attr.live_range.op = irn;
2240
2241                         pset_insert_ptr(live, arg);
2242                 }
2243
2244                 /* just to be sure */
2245                 check_post = ILP_UNDEF;
2246
2247                 /* allow original defintions to be removed */
2248                 if(opt_repair_schedule) {
2249                         pset_foreach(defs, tmp) {
2250                                 op_t      *tmp_op = get_irn_link(tmp);
2251                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2252 #if 1
2253                                 ilp_var_t  delete;
2254                                 assert(spill);
2255
2256                                 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2257                                 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2258
2259                                 /* op may not be killed if its first live_range is 1 */
2260                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2261                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2262                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2263                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2264
2265                                 /* op may not be killed if it is spilled after the definition */
2266                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2267                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2268                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2269                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2270 #else
2271                                 ilp_var_t  keep;
2272                                 assert(spill);
2273
2274                                 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2275                                 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2276
2277                                 /* op may not be killed if its first live_range is 1 */
2278                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2279                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2280                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2281                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2282
2283                                 /* op may not be killed if it is spilled after the definition */
2284                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2285                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2286                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2287                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2288 #endif
2289                         }
2290                 } else {
2291 #if 0
2292                         pset_foreach(defs, tmp) {
2293                                 op_t      *tmp_op = get_irn_link(tmp);
2294                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2295                                 assert(spill);
2296
2297                                 /* live_range or spill should be 1
2298                                    TODO: lr should be live until first use */
2299                                 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2300                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2301                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2302                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2303                         }
2304 #endif
2305                 }
2306
2307
2308                 /******************
2309                  *   P R O L O G
2310                  ******************/
2311
2312                 /* check the register pressure in the prolog */
2313                 /* sum_{L\U} lr <= k - |U| */
2314                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2315                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2316
2317                 /* for the prolog remove defined values from the live set */
2318                 pset_foreach(defs, tmp) {
2319                         pset_remove_ptr(live, tmp);
2320                 }
2321
2322                 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2323                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2324                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2325                 }
2326
2327                 /***********************************************************
2328                  *  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
2329                  **********************************************************/
2330
2331
2332                 set_foreach(args, keyval) {
2333                         spill_t          *spill;
2334                         const ir_node    *arg = keyval->key;
2335                         int               i = PTR_TO_INT(keyval->val);
2336                         op_t             *arg_op = get_irn_link(arg);
2337                         ilp_cst_t         requirements;
2338                         int               n_memoperands;
2339
2340                         spill = set_find_spill(spill_bb->ilp, arg);
2341                         assert(spill);
2342
2343                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2344                         op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2345
2346                         /* reload <= mem_out */
2347                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2348                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2349                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2350                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2351
2352                         /* requirement: arg must be in register for use */
2353                         /* reload + remat + live_range == 1 */
2354                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2355                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2356
2357                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2358                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2359                         foreach_pre_remat(si, irn, tmp) {
2360                                 op_t     *remat_op = get_irn_link(tmp);
2361                                 if(remat_op->attr.remat.remat->value == arg) {
2362                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2363                                 }
2364                         }
2365
2366                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2367                                 n_memoperands = 0;
2368                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2369                                         if(get_irn_n(irn, n) == arg) {
2370                                                 n_memoperands++;
2371                                         }
2372                                 }
2373                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2374                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2375                                                 memoperand_t  *memoperand;
2376                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2377
2378                                                 /* memoperand <= mem_out */
2379                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2380                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2381                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2382                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2383
2384                                                 /* the memoperand is only sufficient if it is used once by the op */
2385                                                 if(n_memoperands == 1)
2386                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2387
2388                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2389
2390                                                 /* we have one more free register if we use a memory operand */
2391                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2392                                         }
2393                                 }
2394                         }
2395                 }
2396
2397                 /* iterate over L\U */
2398                 pset_foreach(live, tmp) {
2399                         if(!set_find_keyval(args, tmp)) {
2400                                 /* if a live value is not used by irn */
2401                                 tmp_op = get_irn_link(tmp);
2402                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2403                         }
2404                 }
2405
2406                 /* requirements for remats */
2407                 foreach_pre_remat(si, irn, tmp) {
2408                         op_t        *remat_op = get_irn_link(tmp);
2409                         int          n;
2410
2411                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2412                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2413                                 op_t           *arg_op = get_irn_link(remat_arg);
2414
2415                                 if(!has_reg_class(si, remat_arg)) continue;
2416
2417                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2418                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2419                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2420
2421                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2422                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2423
2424                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2425                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2426                                         int    index = (int)keyval->val;
2427
2428                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2429                                 }
2430                         }
2431                 }
2432
2433
2434
2435
2436                 /*************************
2437                  *  D O N E  W I T H  O P
2438                  *************************/
2439
2440                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2441
2442                 pset_foreach(live, tmp) {
2443                         assert(has_reg_class(si, tmp));
2444                 }
2445
2446 #ifndef NDEBUG
2447                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2448                         ir_node        *arg = get_irn_n(irn, n);
2449
2450                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2451                 }
2452 #endif
2453
2454                 del_pset(remat_defs);
2455                 del_pset(used);
2456                 del_set(args);
2457                 del_pset(defs);
2458                 defs = pset_new_ptr_default();
2459
2460                 /* skip everything above barrier in start block */
2461                 if(is_start_block(bb) && be_is_Barrier(irn)) {
2462                         assert(pset_count(live) == 0);
2463                         break;
2464                 }
2465
2466         }
2467         del_pset(defs);
2468
2469
2470
2471         /***************************************
2472          *   B E G I N N I N G  O F  B L O C K
2473          ***************************************/
2474
2475
2476         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2477         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2478
2479         pset_foreach(live, irn) {
2480                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2481         }
2482
2483         /* construct mem_outs for all values */
2484         set_foreach(spill_bb->ilp, spill) {
2485                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2486                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2487
2488                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2489                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2490
2491                 if(pset_find_ptr(live, spill->irn)) {
2492                         int default_spilled;
2493                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2494
2495                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2496                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2497                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2498                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2499
2500                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2501                                 int   n;
2502                                 op_t *op = get_irn_link(spill->irn);
2503
2504                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2505                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2506                                         double          freq=0.0;
2507                                         int             m;
2508                                         ilp_var_t       var;
2509
2510
2511                                         /* argument already done? */
2512                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2513
2514                                         /* get sum of execution frequencies of blocks with the same phi argument */
2515                                         for(m=n; m>=0; --m) {
2516                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2517
2518                                                 if(arg==arg2) {
2519                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2520                                                 }
2521                                         }
2522
2523                                         /* copies are not for free */
2524                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2525                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2526
2527                                         for(m=n; m>=0; --m) {
2528                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2529
2530                                                 if(arg==arg2) {
2531                                                         op->attr.live_range.args.copies[m] = var;
2532                                                 }
2533                                         }
2534
2535 #if 0
2536                                         /* copy <= mem_in */
2537                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2538                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2539                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2540                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2541 #endif
2542                                 }
2543                         }
2544                 }
2545         }
2546
2547         foreach_post_remat(bb, tmp) {
2548                 int         n;
2549                 op_t       *remat_op = get_irn_link(tmp);
2550                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2551                 ir_node    *remat_arg;
2552
2553                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2554                         remat_arg = get_irn_n(tmp, n);
2555
2556                         if(has_reg_class(si, remat_arg)) {
2557                                 pset_insert_ptr(remat_args, remat_arg);
2558                         }
2559                 }
2560
2561                 /* remat + \sum live_range(remat_arg) <= |args| */
2562                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2563                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2564                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2565
2566                 pset_foreach(remat_args, remat_arg) {
2567                         if(pset_find_ptr(live, remat_arg)) {
2568                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2569                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2570                         }
2571                 }
2572                 del_pset(remat_args);
2573         }
2574
2575         foreach_post_remat(bb, tmp) {
2576                 int  n;
2577
2578                 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2579                         ir_node  *remat_arg = get_irn_n(tmp, n);
2580
2581                         /* if value is becoming live through use by remat2 */
2582                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2583                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2584                                 ilp_cst_t   nomem;
2585
2586                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2587
2588                                 pset_insert_ptr(live, remat_arg);
2589                                 spill = add_to_spill_bb(si, bb, remat_arg);
2590                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2591
2592                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2593                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2594                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2595                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2596                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2597
2598
2599                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2600                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2601                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2602                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2603                         }
2604                 }
2605         }
2606
2607         /* L\U is empty at bb start */
2608         /* arg is live throughout epilog if it is reg_in into this block */
2609
2610         /* check the register pressure at the beginning of the block
2611          * including remats
2612          */
2613         /* reg_in entspricht post_use */
2614
2615         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2616         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2617
2618         pset_foreach(live, irn) {
2619         ilp_cst_t  nospill;
2620
2621                 spill = set_find_spill(spill_bb->ilp, irn);
2622                 assert(spill);
2623
2624                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2625                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2626
2627                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2628
2629                 /* spill + mem_in <= 1 */
2630                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2631                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2632
2633                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2634                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2635
2636         } /* post_remats are NOT included in register pressure check because
2637            they do not increase regpressure */
2638
2639         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2640         pset_foreach(live, irn) {
2641                 int          p = 0,
2642                                          n;
2643
2644                 spill = set_find_spill(spill_bb->ilp, irn);
2645                 assert(spill && spill->irn == irn);
2646
2647                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2648                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2649                                 ilp_cst_t       mem_in,
2650                                                                 reg_in;
2651                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2652                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2653                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2654                                 spill_t        *spill_p;
2655                                 op_t           *op = get_irn_link(irn);
2656
2657                                 /* although the phi is in the right regclass one or more of
2658                                  * its arguments can be in a different one or at least to
2659                                  * ignore
2660                                  */
2661                                 if(has_reg_class(si, phi_arg)) {
2662                                         /* mem_in < mem_out_arg + copy */
2663                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2664                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2665
2666                                         /* reg_in < reg_out_arg */
2667                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2668                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2669
2670                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2671                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2672
2673                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2674                                         assert(spill_p);
2675
2676                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2677                                         if(opt_memcopies)
2678                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2679
2680                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2681                                 }
2682                         }
2683                 } else {
2684                         /* else assure the value arrives on all paths in the same resource */
2685
2686                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2687                                 ilp_cst_t       mem_in,
2688                                                                 reg_in;
2689                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2690                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2691                                 spill_t        *spill_p;
2692
2693                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2694                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2695                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2696                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2697
2698                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2699                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2700
2701                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2702                                 assert(spill_p);
2703
2704                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2705                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2706                         }
2707                 }
2708         }
2709
2710         foreach_post_remat(bb, tmp) {
2711                 int         n;
2712
2713                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2714                         ir_node    *remat_arg = get_irn_n(tmp, n);
2715                         op_t       *remat_op = get_irn_link(tmp);
2716
2717                         if(!has_reg_class(si, remat_arg)) continue;
2718
2719                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2720                         assert(spill);
2721
2722                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2723                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2724                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2725                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2726                 }
2727         }
2728
2729         pset_foreach(live, irn) {
2730                 const op_t      *op = get_irn_link(irn);
2731                 const ir_node   *remat;
2732                 int              n_remats = 0;
2733
2734                 cst = ILP_UNDEF;
2735
2736                 foreach_post_remat(bb, remat) {
2737                         int   n;
2738
2739                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2740                                 const ir_node  *arg = get_irn_n(remat, n);
2741
2742                                 if(arg == irn) {
2743                                         const op_t   *remat_op = get_irn_link(remat);
2744
2745                                         if(cst == ILP_UNDEF) {
2746                                                 /* sum remat2s <= 1 + n_remats*live_range */
2747                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2748                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2749                                         }
2750                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2751                                         ++n_remats;
2752                                         break;
2753                                 }
2754                         }
2755                 }
2756                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2757                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2758                 }
2759         }
2760
2761         /* first live ranges from reg_ins */
2762         pset_foreach(live, irn) {
2763                 op_t      *op = get_irn_link(irn);
2764
2765                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2766
2767                         spill = set_find_spill(spill_bb->ilp, irn);
2768                         assert(spill && spill->irn == irn);
2769
2770                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2771                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2772                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2773                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2774
2775                         foreach_post_remat(bb, tmp) {
2776                                 op_t     *remat_op = get_irn_link(tmp);
2777
2778                                 if(remat_op->attr.remat.remat->value == irn) {
2779                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2780                                 }
2781                         }
2782                 }
2783         }
2784
2785         /* walk forward now and compute constraints for placing spills */
2786         /* this must only be done for values that are not defined in this block */
2787         pset_foreach(live, irn) {
2788                 /*
2789                  * if value is defined in this block we can anways place the spill directly after the def
2790                  *    -> no constraint necessary
2791                  */
2792                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2793                         assert(0);
2794                 }
2795
2796
2797                 spill = set_find_spill(spill_bb->ilp, irn);
2798                 assert(spill);
2799
2800                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2801                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2802
2803                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2804                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2805
2806                 if(!is_Phi(irn)) {
2807                         sched_foreach_op(bb, tmp) {
2808                                 op_t   *op = get_irn_link(tmp);
2809
2810                                 if(is_Phi(tmp)) continue;
2811                                 assert(!is_Proj(tmp));
2812
2813                                 if(op->is_remat) {
2814                                         const ir_node   *value = op->attr.remat.remat->value;
2815
2816                                         if(value == irn) {
2817                                                 /* only collect remats up to the first real use of a value */
2818                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2819                                         }
2820                                 } else {
2821                                         int   n;
2822
2823                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2824                                                 ir_node    *arg = get_irn_n(tmp, n);
2825
2826                                                 if(arg == irn) {
2827                                                         /* if a value is used stop collecting remats */
2828                             goto next_live;
2829                                                 }
2830                                         }
2831                                 }
2832                         }
2833                 }
2834 next_live: ;
2835         }
2836
2837         del_pset(live);
2838 }
2839
2840 typedef struct _irnlist_t {
2841         struct list_head   list;
2842         ir_node           *irn;
2843 } irnlist_t;
2844
2845 typedef struct _interference_t {
2846         struct list_head    blocklist;
2847         ir_node            *a;
2848         ir_node            *b;
2849 } interference_t;
2850
2851 static int
2852 cmp_interference(const void *a, const void *b, size_t size)
2853 {
2854         const interference_t *p = a;
2855         const interference_t *q = b;
2856
2857         return !(p->a == q->a && p->b == q->b);
2858 }
2859
2860 static interference_t *
2861 set_find_interference(set * set, ir_node * a, ir_node * b)
2862 {
2863         interference_t     query;
2864
2865         query.a = (a>b)?a:b;
2866         query.b = (a>b)?b:a;
2867
2868         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2869 }
2870
2871 static interference_t *
2872 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2873 {
2874         interference_t     query,
2875                                           *result;
2876         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2877
2878         list->irn = bb;
2879
2880         result = set_find_interference(set, a, b);
2881         if(result) {
2882
2883                 list_add(&list->list, &result->blocklist);
2884                 return result;
2885         }
2886
2887         query.a = (a>b)?a:b;
2888         query.b = (a>b)?b:a;
2889
2890         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2891
2892         INIT_LIST_HEAD(&result->blocklist);
2893         list_add(&list->list, &result->blocklist);
2894
2895         return result;
2896 }
2897
2898 static int
2899 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2900 {
2901         const ir_edge_t *edge;
2902
2903         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2904                 /* both values are live in, so they interfere */
2905                 return 1;
2906         }
2907
2908         /* ensure a dominates b */
2909         if(value_dominates(b,a)) {
2910                 const ir_node * t;
2911                 t = b;
2912                 b = a;
2913                 a = t;
2914         }
2915         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2916
2917
2918         /* the following code is stolen from bera.c */
2919         if(be_is_live_end(si->lv, bb, a))
2920                 return 1;
2921
2922         foreach_out_edge(a, edge) {
2923                 const ir_node *user = edge->src;
2924                 if(get_nodes_block(user) == bb
2925                                 && !is_Phi(user)
2926                                 && b != user
2927                                 && !pset_find_ptr(si->inverse_ops, user)
2928                                 && value_dominates(b, user))
2929                         return 1;
2930         }
2931
2932         return 0;
2933 }
2934
2935 /**
2936  * Walk all irg blocks and collect interfering values inside of phi classes
2937  */
2938 static void
2939 luke_interferencewalker(ir_node * bb, void * data)
2940 {
2941         spill_ilp_t    *si = (spill_ilp_t*)data;
2942         int             l1, l2;
2943
2944         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2945                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2946                 op_t           *a_op = get_irn_link(a);
2947
2948
2949                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2950                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2951                         if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2952                                 continue;
2953
2954                         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)) {
2955                                 ir_node *b    = be_lv_get_irn(si->lv, bb, l2);
2956                                 op_t    *b_op = get_irn_link(b);
2957
2958                                 /* a and b are only interesting if they are in the same phi class */
2959                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2960                                         if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2961                                                 continue;
2962
2963                                         if (values_interfere_in_block(si, bb, a, b)) {
2964                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2965                                                 set_insert_interference(si, si->interferences, a, b, bb);
2966                                         }
2967                                 }
2968                         }
2969                 }
2970         }
2971 }
2972
2973 static unsigned int copy_path_id = 0;
2974
2975 static void
2976 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2977 {
2978         ilp_cst_t  cst;
2979         ilp_var_t  copy;
2980         char       buf[256];
2981         void      *ptr;
2982
2983         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2984         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2985
2986         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2987
2988         pset_foreach(copies, ptr) {
2989                 copy = PTR_TO_INT(ptr);
2990                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2991         }
2992 }
2993
2994 /**
2995  * @parameter copies   contains a path of copies which lead us to irn
2996  * @parameter visited  contains a set of nodes already visited on this path
2997  */
2998 static int
2999 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3000 {
3001         const ir_edge_t *edge;
3002         op_t            *op = get_irn_link(irn);
3003     pset            *visited_users = pset_new_ptr_default();
3004         int              paths = 0;
3005
3006         if(op->is_remat) return 0;
3007
3008         pset_insert_ptr(visited, irn);
3009
3010         if(is_Phi(irn)) {
3011                 int    n;
3012         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
3013
3014                 /* visit all operands */
3015                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3016                         ir_node  *arg = get_irn_n(irn, n);
3017                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
3018
3019                         if(!has_reg_class(si, arg)) continue;
3020             if(pset_find_ptr(visited_operands, arg)) continue;
3021             pset_insert_ptr(visited_operands, arg);
3022
3023                         if(arg == target) {
3024                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3025                                         del_pset(visited_operands);
3026                                         del_pset(visited_users);
3027                                         pset_remove_ptr(visited, irn);
3028                                         return paths;
3029                                 }
3030                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3031                                 write_copy_path_cst(si, copies, any_interfere);
3032                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3033                         } else if(!pset_find_ptr(visited, arg)) {
3034                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3035                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3036                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3037
3038                 if(paths > MAX_PATHS) {
3039                     if(pset_count(copies) == 0) {
3040                         ilp_cst_t  cst;
3041                         char       buf[256];
3042
3043                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3044                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3045                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3046                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3047                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3048
3049                         paths = 0;
3050                     } else {
3051                         del_pset(visited_operands);
3052                         del_pset(visited_users);
3053                         pset_remove_ptr(visited, irn);
3054                         return paths;
3055                     }
3056                 } else if(pset_count(copies) == 0) {
3057                                         paths = 0;
3058                                 }
3059                         }
3060                 }
3061
3062         del_pset(visited_operands);
3063         }
3064
3065         /* visit all uses which are phis */
3066         foreach_out_edge(irn, edge) {
3067                 ir_node  *user = edge->src;
3068                 int       pos  = edge->pos;
3069                 op_t     *op = get_irn_link(user);
3070                 ilp_var_t copy;
3071
3072                 if(!is_Phi(user)) continue;
3073                 if(!has_reg_class(si, user)) continue;
3074         if(pset_find_ptr(visited_users, user)) continue;
3075         pset_insert_ptr(visited_users, user);
3076
3077                 copy = op->attr.live_range.args.copies[pos];
3078
3079                 if(user == target) {
3080                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3081                                 del_pset(visited_users);
3082                                 pset_remove_ptr(visited, irn);
3083                                 return paths;
3084                         }
3085                         pset_insert(copies, INT_TO_PTR(copy), copy);
3086                         write_copy_path_cst(si, copies, any_interfere);
3087                         pset_remove(copies, INT_TO_PTR(copy), copy);
3088                 } else if(!pset_find_ptr(visited, user)) {
3089                         pset_insert(copies, INT_TO_PTR(copy), copy);
3090                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3091                         pset_remove(copies, INT_TO_PTR(copy), copy);
3092
3093             if(paths > MAX_PATHS) {
3094                 if(pset_count(copies) == 0) {
3095                     ilp_cst_t  cst;
3096                     char       buf[256];
3097
3098                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3099                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3100                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3101                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3102                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3103
3104                     paths = 0;
3105                 } else {
3106                     del_pset(visited_users);
3107                     pset_remove_ptr(visited, irn);
3108                     return paths;
3109                 }
3110             } else if(pset_count(copies) == 0) {
3111                                 paths = 0;
3112                         }
3113                 }
3114         }
3115
3116     del_pset(visited_users);
3117         pset_remove_ptr(visited, irn);
3118         return paths;
3119 }
3120
3121 static void
3122 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3123 {
3124         pset * copies = pset_new_ptr_default();
3125         pset * visited = pset_new_ptr_default();
3126
3127         find_copy_path(si, a, b, any_interfere, copies, visited);
3128
3129         del_pset(visited);
3130         del_pset(copies);
3131 }
3132
3133
3134 static void
3135 memcopyhandler(spill_ilp_t * si)
3136 {
3137         interference_t   *interference;
3138         char              buf[256];
3139         /* teste Speicherwerte auf Interferenz */
3140
3141         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3142         irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3143
3144         /* now lets emit the ILP unequations for the crap */
3145         set_foreach(si->interferences, interference) {
3146                 irnlist_t      *irnlist;
3147                 ilp_var_t      interfere, any_interfere;
3148                 ilp_cst_t      any_interfere_cst, cst;
3149                 const ir_node  *a  = interference->a;
3150                 const ir_node  *b  = interference->b;
3151
3152                 /* any_interf <= \sum interf */
3153                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3154                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3155                 any_interfere     = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3156
3157                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3158
3159                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3160                         const ir_node  *bb = irnlist->irn;
3161                         spill_bb_t     *spill_bb = get_irn_link(bb);
3162                         spill_t        *spilla,
3163                                                    *spillb;
3164                         char           buf[256];
3165
3166                         spilla = set_find_spill(spill_bb->ilp, a);
3167                         assert(spilla);
3168
3169                         spillb = set_find_spill(spill_bb->ilp, b);
3170                         assert(spillb);
3171
3172                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3173                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3174                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3175                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3176                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3177                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3178
3179                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3180                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3181
3182                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3183                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3184                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3185                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3186                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3187
3188                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3189                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3190
3191                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3192                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3193                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3194
3195                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3196                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3197
3198                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3199                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3200                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3201
3202
3203                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3204
3205                         /* any_interfere >= interf */
3206                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3207                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3208
3209                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3210                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3211                 }
3212
3213                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3214                 gen_copy_constraints(si,a,b,any_interfere);
3215         }
3216 }
3217
3218
3219 static INLINE int
3220 is_zero(double x)
3221 {
3222         return fabs(x) < 0.00001;
3223 }
3224
3225 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3226 {
3227         spill_ilp_t *si = get_irg_link(current_ir_graph);
3228
3229         if(pset_find_ptr(si->all_possible_remats, n)) {
3230                 op_t   *op = (op_t*)get_irn_link(n);
3231                 assert(op && op->is_remat);
3232
3233                 if(!op->attr.remat.remat->inverse) {
3234                         if(op->attr.remat.pre) {
3235                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3236                         } else {
3237                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3238                         }
3239
3240                         return 1;
3241                 } else {
3242                         op_t   *op = (op_t*)get_irn_link(n);
3243                         assert(op && op->is_remat);
3244
3245                         if(op->attr.remat.pre) {
3246                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3247                         } else {
3248                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3249                         }
3250
3251                         return 1;
3252                 }
3253         }
3254
3255         return 0;
3256 }
3257
3258 static void
3259 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3260 {
3261         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3262         be_dump(irg, suffix, dump_ir_block_graph_sched);
3263         set_dump_node_vcgattr_hook(NULL);
3264 }
3265
3266 /**
3267  * Edge hook to dump the schedule edges with annotated register pressure.
3268  */
3269 static int
3270 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3271 {
3272         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3273                 ir_node *prev = sched_prev(irn);
3274                 fprintf(F, "edge:{sourcename:\"");
3275                 PRINT_NODEID(irn);
3276                 fprintf(F, "\" targetname:\"");
3277                 PRINT_NODEID(prev);
3278                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3279                 fprintf(F, "\" color:magenta}\n");
3280         }
3281         return 1;
3282 }
3283
3284 static void
3285 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3286 {
3287         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3288
3289         dump_consts_local(0);
3290         set_dump_node_edge_hook(sched_pressure_edge_hook);
3291         dump_ir_block_graph(irg, suffix);
3292         set_dump_node_edge_hook(old_edge_hook);
3293 }
3294
3295 static void
3296 walker_pressure_annotator(ir_node * bb, void * data)
3297 {
3298         spill_ilp_t  *si = data;
3299         ir_node      *irn;
3300         int           n, i;
3301         pset         *live = pset_new_ptr_default();
3302         int           projs = 0;
3303
3304         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3305                 irn = be_lv_get_irn(si->lv, bb, i);
3306
3307                 if (has_reg_class(si, irn)) {
3308                         pset_insert_ptr(live, irn);
3309                 }
3310         }
3311
3312         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3313
3314         sched_foreach_reverse(bb, irn) {
3315                 if(is_Phi(irn)) {
3316                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3317                         continue;
3318                 }
3319
3320                 if(has_reg_class(si, irn)) {
3321                         pset_remove_ptr(live, irn);
3322                         if(is_Proj(irn)) ++projs;
3323                 }
3324
3325                 if(!is_Proj(irn)) projs = 0;
3326
3327                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3328                         ir_node    *arg = get_irn_n(irn, n);
3329
3330                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3331                 }
3332                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3333         }
3334
3335         del_pset(live);
3336 }
3337
3338 static void
3339 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3340 {
3341         be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3342 }
3343
3344 static void
3345 connect_all_remats_with_keep(spill_ilp_t * si)
3346 {
3347         ir_node   *irn;
3348         ir_node  **ins,
3349                          **pos;
3350         int        n_remats;
3351
3352
3353         n_remats = pset_count(si->all_possible_remats);
3354         if(n_remats) {
3355                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3356
3357                 pos = ins;
3358                 pset_foreach(si->all_possible_remats, irn) {
3359                         *pos = irn;
3360                         ++pos;
3361                 }
3362
3363                 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3364
3365                 obstack_free(si->obst, ins);
3366         }
3367 }
3368
3369 static void
3370 connect_all_spills_with_keep(spill_ilp_t * si)
3371 {
3372         ir_node   *irn;
3373         ir_node  **ins,
3374                          **pos;
3375         int        n_spills;
3376         ir_node   *keep;
3377
3378
3379         n_spills = pset_count(si->spills);
3380         if(n_spills) {
3381                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3382
3383                 pos = ins;
3384                 pset_foreach(si->spills, irn) {
3385                         *pos = irn;
3386                         ++pos;
3387                 }
3388
3389                 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3390
3391                 obstack_free(si->obst, ins);
3392         }
3393 }
3394
3395 /** insert a spill at an arbitrary position */
3396 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3397 {
3398         ir_node  *bl    = is_Block(insert) ? insert : get_nodes_block(insert);
3399         ir_graph *irg   = get_irn_irg(bl);
3400         ir_node  *frame = get_irg_frame(irg);
3401         ir_node  *spill;
3402         ir_node  *next;
3403         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3404         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3405
3406         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3407
3408         /*
3409          * search the right insertion point. a spill of a phi cannot be put
3410          * directly after the phi, if there are some phis behind the one which
3411          * is spilled. Also, a spill of a Proj must be after all Projs of the
3412          * same tuple node.
3413          *
3414          * Here's one special case:
3415          * If the spill is in the start block, the spill must be after the frame
3416          * pointer is set up. This is done by setting insert to the end of the block
3417          * which is its default initialization (see above).
3418          */
3419
3420         if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3421                 insert = frame;
3422
3423         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3424                 insert = next;
3425
3426         sched_add_after(insert, spill);
3427         return spill;
3428 }
3429
3430 static void
3431 delete_remat(spill_ilp_t * si, ir_node * remat) {
3432         int       n;
3433         ir_node  *bad = get_irg_bad(si->birg->irg);
3434
3435         sched_remove(remat);
3436
3437         /* kill links to operands */
3438         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3439                 set_irn_n(remat, n, bad);
3440         }
3441 }
3442
3443 static void
3444 clean_remat_info(spill_ilp_t * si)
3445 {
3446         int            n;
3447         remat_t       *remat;
3448         remat_info_t  *remat_info;
3449         ir_node       *bad = get_irg_bad(si->birg->irg);
3450
3451         set_foreach(si->remat_info, remat_info) {
3452                 if(!remat_info->remats) continue;
3453
3454                 pset_foreach(remat_info->remats, remat)
3455                 {
3456                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3457                                 if(sched_is_scheduled(remat->proj)) {
3458                                         sched_remove((ir_node*)remat->proj);
3459                                 }
3460                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3461                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3462                         }
3463
3464                         if(get_irn_n_edges(remat->op) == 0) {
3465                                 if(sched_is_scheduled(remat->op)) {
3466                                         sched_remove((ir_node*)remat->op);
3467                                 }
3468                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3469                                         set_irn_n((ir_node*)remat->op, n, bad);
3470                                 }
3471                         }
3472                 }
3473
3474                 if(remat_info->remats) del_pset(remat_info->remats);
3475                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3476         }
3477 }
3478
3479 static void
3480 delete_unnecessary_remats(spill_ilp_t * si)
3481 {
3482         if(opt_keep_alive & KEEPALIVE_REMATS) {
3483                 int       n;
3484                 ir_node  *bad = get_irg_bad(si->birg->irg);
3485
3486                 if(si->keep) {
3487                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3488                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3489                                 op_t           *arg_op = get_irn_link(keep_arg);
3490                                 lpp_name_t     *name;
3491
3492                                 assert(arg_op->is_remat);
3493
3494                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3495
3496                                 if(is_zero(name->value)) {
3497                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3498                                         /* TODO check whether reload is preferred over remat (could be bug) */
3499                                         delete_remat(si, keep_arg);
3500                                 } else {
3501                                         if(!arg_op->attr.remat.remat->inverse) {
3502                                                 if(arg_op->attr.remat.pre) {
3503                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3504                                                 } else {
3505                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3506                                                 }
3507                                         } else {
3508                                                 if(arg_op->attr.remat.pre) {
3509                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3510                                                 } else {
3511                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3512                                                 }
3513                                         }
3514                                 }
3515
3516                                 set_irn_n(si->keep, n, bad);
3517                         }
3518                 } else {
3519                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3520                 }
3521         } else {
3522                 ir_node  *remat;
3523
3524                 pset_foreach(si->all_possible_remats, remat) {
3525                         op_t           *remat_op = get_irn_link(remat);
3526                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3527
3528                         if(is_zero(name->value)) {
3529                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3530                                 /* TODO check whether reload is preferred over remat (could be bug) */
3531                                 delete_remat(si, remat);
3532                         } else {
3533                                 if(!remat_op->attr.remat.remat->inverse) {
3534                                         if(remat_op->attr.remat.pre) {
3535                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3536                                         } else {
3537                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3538                                         }
3539                                 } else {
3540                                         if(remat_op->attr.remat.pre) {
3541                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3542                                         } else {
3543                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3544                                         }
3545                                 }
3546                         }
3547                 }
3548         }
3549 }
3550
3551 static pset *
3552 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3553 {
3554         pset     *spills = pset_new_ptr_default();
3555
3556         const ir_node  *next;
3557         defs_t         *defs;
3558
3559         defs = set_find_def(si->values, value);
3560
3561         if(defs && defs->spills) {
3562                 for(next = defs->spills; next; next = get_irn_link(next)) {
3563                         pset_insert_ptr(spills, next);
3564                 }
3565         }
3566
3567         return spills;
3568 }
3569
3570 static ir_node *
3571 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3572 {
3573         ir_node  *res;
3574
3575         assert( get_irn_arity(block) == arity );
3576
3577         res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3578         res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3579
3580         return res;
3581 }
3582
3583 /**
3584  * @param before   The node after which the spill will be placed in the schedule
3585  */
3586 static ir_node *
3587 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3588 {
3589         defs_t   *defs;
3590         ir_node  *spill;
3591         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3592
3593         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3594
3595         spill = be_spill2(arch_env, irn, before);
3596
3597         defs = set_insert_def(si->values, value);
3598         assert(defs);
3599
3600         /* enter into the linked list */
3601         set_irn_link(spill, defs->spills);
3602         defs->spills = spill;
3603
3604         if(opt_keep_alive & KEEPALIVE_SPILLS)
3605                 pset_insert_ptr(si->spills, spill);
3606
3607         return spill;
3608 }
3609
3610 /**
3611  * @param before   The Phi node which has to be spilled
3612  */
3613 static ir_node *
3614 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3615 {
3616         ir_node   *mem_phi;
3617         ir_node  **ins;
3618         defs_t    *defs;
3619         int        n;
3620
3621         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3622
3623         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3624                 ins[n] = si->m_unknown;
3625         }
3626
3627         mem_phi =  new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3628
3629         defs = set_insert_def(si->values, phi);
3630         assert(defs);
3631
3632         /* enter into the linked list */
3633         set_irn_link(mem_phi, defs->spills);
3634         defs->spills = mem_phi;
3635
3636 #ifdef SCHEDULE_PHIM
3637         sched_add_after(phi, mem_phi);
3638 #else
3639         pset_insert_ptr(si->phims, mem_phi);
3640 #endif
3641
3642         if(opt_keep_alive & KEEPALIVE_SPILLS)
3643                 pset_insert_ptr(si->spills, mem_phi);
3644
3645
3646         return mem_phi;
3647 }
3648
3649 /**
3650  * Add remat to list of defs, destroys link field!
3651  */
3652 static void
3653 insert_remat(spill_ilp_t * si, ir_node * remat)
3654 {
3655         defs_t   *defs;
3656         op_t     *remat_op = get_irn_link(remat);
3657
3658         assert(remat_op->is_remat);
3659
3660         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3661         assert(defs);
3662
3663         /* enter into the linked list */
3664         set_irn_link(remat, defs->remats);
3665         defs->remats = remat;
3666 }
3667
3668
3669 /**
3670  * Add reload before operation and add to list of defs
3671  */
3672 static ir_node *
3673 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3674 {
3675         defs_t   *defs;
3676         ir_node  *reload,
3677                          *spill;
3678         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3679
3680         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3681
3682         defs = set_find_def(si->values, value);
3683
3684         spill = defs->spills;
3685         assert(spill && "no spill placed before reload");
3686
3687         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3688
3689         /* enter into the linked list */
3690         set_irn_link(reload, defs->remats);
3691         defs->remats = reload;
3692
3693         return reload;
3694 }
3695
3696 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3697 {
3698         defs_t           *defs;
3699         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3700         ir_node          *spill;
3701         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3702
3703         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3704
3705         defs = set_find_def(si->values, value);
3706
3707         spill = defs->spills;
3708         assert(spill && "no spill placed before reload");
3709
3710         arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3711 }
3712
3713 void insert_memoperands(spill_ilp_t * si)
3714 {
3715         memoperand_t   *memoperand;
3716         lpp_name_t     *name;
3717
3718         set_foreach(si->memoperands, memoperand) {
3719                 name = si->lpp->vars[memoperand->ilp];
3720                 if(!is_zero(name->value)) {
3721                         perform_memory_operand(si, memoperand);
3722                 }
3723         }
3724 }
3725
3726 static void
3727 walker_spill_placer(ir_node * bb, void * data) {
3728         spill_ilp_t   *si = (spill_ilp_t*)data;
3729         ir_node       *irn;
3730         spill_bb_t    *spill_bb = get_irn_link(bb);
3731         pset          *spills_to_do = pset_new_ptr_default();
3732         spill_t       *spill;
3733
3734         set_foreach(spill_bb->ilp, spill) {
3735                 lpp_name_t    *name;
3736
3737                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3738                         name = si->lpp->vars[spill->mem_in];
3739                         if(!is_zero(name->value)) {
3740                                 ir_node   *mem_phi;
3741
3742                                 mem_phi = insert_mem_phi(si, spill->irn);
3743
3744                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3745                         }
3746                 }
3747
3748                 name = si->lpp->vars[spill->spill];
3749                 if(!is_zero(name->value)) {
3750                         /* place spill directly after definition */
3751                         if(get_nodes_block(spill->irn) == bb) {
3752                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3753                                 continue;
3754                         }
3755
3756                         /* place spill at bb start */
3757                         if(spill->reg_in > 0) {
3758                                 name = si->lpp->vars[spill->reg_in];
3759                                 if(!is_zero(name->value)) {
3760                                         insert_spill(si, spill->irn, spill->irn, bb);
3761                                         continue;
3762                                 }
3763                         }
3764                         /* place spill after a remat */
3765                         pset_insert_ptr(spills_to_do, spill->irn);
3766                 }
3767         }
3768         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3769
3770
3771         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3772                 op_t     *op = get_irn_link(irn);
3773
3774                 if(be_is_Spill(irn)) continue;
3775
3776                 if(op->is_remat) {
3777                         /* TODO fix this if we want to support remats with more than two nodes */
3778                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3779                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3780
3781                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3782                         }
3783                 } else {
3784                         if(pset_find_ptr(spills_to_do, irn)) {
3785                                 pset_remove_ptr(spills_to_do, irn);
3786
3787                                 insert_spill(si, irn, irn, irn);
3788                         }
3789                 }
3790
3791         }
3792
3793         assert(pset_count(spills_to_do) == 0);
3794
3795         /* afterwards free data in block */
3796         del_pset(spills_to_do);
3797 }
3798
3799 static ir_node *
3800 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3801 {
3802         ir_node          *insert_pos = bb;
3803         ir_node          *spill;
3804         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3805
3806         /* find last definition of arg value in block */
3807         ir_node  *next;
3808         defs_t   *defs;
3809         int       last = 0;
3810
3811         defs = set_find_def(si->values, value);
3812
3813         if(defs && defs->remats) {
3814                 for(next = defs->remats; next; next = get_irn_link(next)) {
3815                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3816                                 last = sched_get_time_step(next);
3817                                 insert_pos = next;
3818                         }
3819                 }
3820         }
3821
3822         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3823                 last = sched_get_time_step(value);
3824                 insert_pos = value;
3825         }
3826
3827         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3828
3829         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3830
3831         return spill;
3832 }
3833
3834 static void
3835 phim_fixer(spill_ilp_t *si) {
3836         defs_t  *defs;
3837
3838         set_foreach(si->values, defs) {
3839                 const ir_node  *phi = defs->value;
3840                 op_t           *op = get_irn_link(phi);
3841                 ir_node        *phi_m = NULL;
3842                 ir_node        *next = defs->spills;
3843                 int             n;
3844
3845                 if(!is_Phi(phi)) continue;
3846
3847                 while(next) {
3848                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3849                                 phi_m = next;
3850                                 break;
3851                         } else {
3852                                 next = get_irn_link(next);
3853                         }
3854                 }
3855                 if(!phi_m) continue;
3856
3857                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3858                         ir_node        *value = get_irn_n(phi, n);
3859                         defs_t         *val_defs = set_find_def(si->values, value);
3860
3861                         /* a spill of this value */
3862                         ir_node      *spill;
3863
3864
3865                         if(opt_memcopies) {
3866                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3867                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3868
3869                                 if(!is_zero(name->value)) {
3870                                         spill = insert_mem_copy(si, pred, value);
3871                                 } else {
3872                                         spill = val_defs->spills;
3873                                 }
3874                         } else {
3875                                 spill = val_defs->spills;
3876                         }
3877
3878                         assert(spill && "no spill placed before PhiM");
3879                         set_irn_n(phi_m, n, spill);
3880                 }
3881         }
3882 }
3883
3884 static void
3885 walker_reload_placer(ir_node * bb, void * data) {
3886         spill_ilp_t   *si = (spill_ilp_t*)data;
3887         ir_node       *irn;
3888         spill_bb_t    *spill_bb = get_irn_link(bb);
3889
3890         /* reloads at end of block */
3891         if(spill_bb->reloads) {
3892                 keyval_t    *keyval;
3893
3894                 set_foreach(spill_bb->reloads, keyval) {
3895                         ir_node        *irn = (ir_node*)keyval->key;
3896                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3897                         lpp_name_t     *name;
3898
3899                         name = si->lpp->vars[reload];
3900                         if(!is_zero(name->value)) {
3901                                 ir_node    *reload;
3902                                 ir_node    *insert_pos = bb;
3903                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3904                                 op_t       *prev_op = get_irn_link(prev);
3905
3906                                 while(be_is_Spill(prev)) {
3907                                         prev = sched_prev(prev);
3908                                 }
3909
3910                                 prev_op = get_irn_link(prev);
3911
3912                                 /* insert reload before pre-remats */
3913                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3914                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3915                                         insert_pos = prev;
3916
3917                                         do {
3918                                                 prev = sched_prev(prev);
3919                                         } while(be_is_Spill(prev));
3920
3921                                         prev_op = get_irn_link(prev);
3922
3923                                 }
3924
3925                                 reload = insert_reload(si, irn, insert_pos);
3926
3927                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3928                                         pset_insert_ptr(si->spills, reload);
3929                         }
3930                 }
3931         }
3932
3933         /* walk and insert more reloads and collect remats */
3934         sched_foreach_reverse(bb, irn) {
3935                 op_t     *op = get_irn_link(irn);
3936
3937                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3938                 if(is_Phi(irn)) break;
3939
3940                 if(op->is_remat) {
3941                         if(get_irn_mode(irn) != mode_T) {
3942                                 insert_remat(si, irn);
3943                         }
3944                 } else {
3945                         int    n;
3946
3947                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3948                                 ir_node    *arg = get_irn_n(irn, n);
3949
3950                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3951                                         lpp_name_t    *name;
3952
3953                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3954                                         if(!is_zero(name->value)) {
3955                                                 ir_node    *reload;
3956                                                 ir_node    *insert_pos = irn;
3957                                                 ir_node    *prev = sched_prev(insert_pos);
3958                                                 op_t       *prev_op;
3959
3960                                                 while(be_is_Spill(prev)) {
3961                                                         prev = sched_prev(prev);
3962                                                 }
3963
3964                                                 prev_op = get_irn_link(prev);
3965
3966                                                 /* insert reload before pre-remats */
3967                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3968                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3969                                                         insert_pos = prev;
3970
3971                                                         do {
3972                                                                 prev = sched_prev(prev);
3973                                                         } while(be_is_Spill(prev));
3974
3975                                                         prev_op = get_irn_link(prev);
3976
3977                                                 }
3978
3979                                                 reload = insert_reload(si, arg, insert_pos);
3980
3981                                                 assert(reload && "no reload returned");
3982                                                 set_irn_n(irn, n, reload);
3983
3984                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3985                                                         pset_insert_ptr(si->spills, reload);
3986                                         }
3987                                 }
3988                         }
3989                 }
3990         }
3991
3992         del_set(spill_bb->ilp);
3993         if(spill_bb->reloads) del_set(spill_bb->reloads);
3994 }
3995
3996 static void
3997 walker_collect_used(ir_node * irn, void * data)
3998 {
3999         bitset_t   *used = data;
4000
4001         bitset_set(used, get_irn_idx(irn));
4002 }
4003
4004 struct kill_helper {
4005         bitset_t  *used;
4006         spill_ilp_t  *si;
4007 };
4008
4009 static void
4010 walker_kill_unused(ir_node * bb, void * data)
4011 {
4012         struct kill_helper *kh = data;
4013         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
4014         ir_node            *irn;
4015
4016
4017         for(irn=sched_first(bb); !sched_is_end(irn);) {
4018                 ir_node     *next = sched_next(irn);
4019                 int          n;
4020
4021                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4022                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
4023                                 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)));
4024 #if 0
4025                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4026 #endif
4027                         }
4028
4029                         sched_remove(irn);
4030
4031                         set_nodes_block(irn, bad);
4032                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
4033                                 set_irn_n(irn, n, bad);
4034                         }
4035                 }
4036                 irn = next;
4037         }
4038 }
4039
4040 #ifndef SCHEDULE_PHIM
4041 static void
4042 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4043 {
4044         ir_node  *phi;
4045         ir_node  *bad = get_irg_bad(si->birg->irg);
4046         int       n;
4047
4048         pset_foreach(si->phims, phi) {
4049                 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4050
4051                         set_nodes_block(phi, bad);
4052                         for (n=get_irn_arity(phi)-1; n>=0; --n) {
4053                                 set_irn_n(phi, n, bad);
4054                         }
4055                 }
4056         }
4057 }
4058 #endif
4059
4060 static void
4061 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4062 {
4063         struct kill_helper  kh;
4064
4065         kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4066         kh.si = si;
4067
4068         irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4069 #ifndef SCHEDULE_PHIM
4070         kill_unused_phims(si, &kh);
4071 #endif
4072         irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4073
4074         bitset_free(kh.used);
4075 }
4076
4077 void
4078 print_irn_pset(pset * p)
4079 {
4080         ir_node   *irn;
4081
4082         pset_foreach(p, irn) {
4083                 ir_printf("%+F\n", irn);
4084         }
4085 }
4086
4087 void
4088 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4089 {
4090     FILE           *f = fopen(file, "w");
4091     ir_node        *irn;
4092     interference_t *interference;
4093     int            i;
4094
4095     set_break(si->interferences);
4096
4097     ir_fprintf(f, "digraph phiclass {\n");
4098
4099     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4100         irn = phiclass[i];
4101         if (is_Phi(irn))
4102             ir_fprintf(f, "  %F%N [shape=box]\n", irn, irn);
4103     }
4104
4105     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4106         int n;
4107
4108         irn = phiclass[i];
4109         if (! is_Phi(irn))
4110             continue;
4111
4112         for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4113             ir_node  *arg = get_irn_n(irn, n);
4114
4115             ir_fprintf(f, "  %F%N -> %F%N\n", irn, irn, arg, arg);
4116         }
4117     }
4118
4119     set_foreach(si->interferences, interference) {
4120         const ir_node *a = interference->a;
4121         const ir_node *b = interference->b;
4122         if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4123             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4124         }
4125     }
4126
4127     ir_fprintf(f, "}");
4128     fclose(f);
4129 }
4130
4131 static void
4132 rewire_uses(spill_ilp_t * si)
4133 {
4134         defs_t               *defs;
4135         ir_nodeset_t         ignore;
4136
4137         ir_nodeset_init(&ignore);
4138         ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg));
4139
4140         /* then fix uses of spills */
4141         set_foreach(si->values, defs) {
4142                 pset           *reloads;
4143                 pset           *spills;
4144                 const ir_node  *next = defs->remats;
4145                 int remats = 0;
4146
4147                 reloads = pset_new_ptr_default();
4148
4149                 while(next) {
4150                         if(be_is_Reload(next)) {
4151                                 pset_insert_ptr(reloads, next);
4152                         } else {
4153                                 ++remats;
4154                         }
4155                         next = get_irn_link(next);
4156                 }
4157
4158                 spills = get_spills_for_value(si, defs->value);
4159                 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));
4160                 if(pset_count(spills) > 1) {
4161                         be_ssa_construction_env_t senv;
4162                         ir_node *node;
4163                         //assert(pset_count(reloads) > 0);
4164                         //                              print_irn_pset(spills);
4165                         //                              print_irn_pset(reloads);
4166
4167                         be_ssa_construction_init(&senv, si->birg);
4168                         be_ssa_construction_set_ignore_uses(&senv, &ignore);
4169                         pset_foreach(spills, node) {
4170                                 be_ssa_construction_add_copy(&senv, node);
4171                         }
4172                         pset_foreach(spills, node) {
4173                                 be_ssa_construction_fix_users(&senv, node);
4174                         }
4175                         be_ssa_construction_update_liveness_phis(&senv, si->lv);
4176                         pset_foreach(spills, node) {
4177                                 be_liveness_update(si->lv, node);
4178                         }
4179                         be_ssa_construction_destroy(&senv);
4180                 }
4181
4182                 del_pset(reloads);
4183                 del_pset(spills);
4184         }
4185
4186         /* first fix uses of remats and reloads */
4187         set_foreach(si->values, defs) {
4188                 const ir_node  *next = defs->remats;
4189                 int             orig_kept = 0;
4190
4191                 if(next) {
4192                         be_ssa_construction_env_t senv;
4193
4194                         be_ssa_construction_init(&senv, si->birg);
4195
4196                         if(sched_is_scheduled(defs->value)) {
4197                                 be_ssa_construction_add_copy(&senv, (ir_node*) defs->value);
4198                                 orig_kept = 1;
4199                         }
4200
4201                         next = defs->remats;
4202                         while(next) {
4203                                 be_ssa_construction_add_copy(&senv, (ir_node*) next);
4204                                 next = get_irn_link(next);
4205                         }
4206
4207                         if(sched_is_scheduled(defs->value)) {
4208                                 be_ssa_construction_fix_users(&senv, (ir_node*) defs->value);
4209                         }
4210
4211                         next = defs->remats;
4212                         while(next) {
4213                                 be_ssa_construction_fix_users(&senv, (ir_node*) next);
4214                                 next = get_irn_link(next);
4215                         }
4216
4217                         be_ssa_construction_update_liveness_phis(&senv, si->lv);
4218                         if(sched_is_scheduled(defs->value)) {
4219                                 be_liveness_update(si->lv, (ir_node*) defs->value);
4220                         }
4221
4222                         next = defs->remats;
4223                         while(next) {
4224                                 be_liveness_update(si->lv, (ir_node*) next);
4225                                 next = get_irn_link(next);
4226                         }
4227
4228                         be_ssa_construction_destroy(&senv);
4229                 }
4230         }
4231
4232         ir_nodeset_destroy(&ignore);
4233 //      remove_unused_defs(si);
4234 }
4235
4236
4237 static void
4238 writeback_results(spill_ilp_t * si)
4239 {
4240         /* walk through the graph and collect all spills, reloads and remats for a value */
4241
4242         si->values = new_set(cmp_defs, 4096);
4243
4244         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4245         delete_unnecessary_remats(si);
4246         si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4247         irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4248         irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4249         if(opt_memoperands)
4250                 insert_memoperands(si);
4251         phim_fixer(si);
4252
4253         /* clean the remat info! there are still back-edges leading there! */
4254         clean_remat_info(si);
4255
4256         rewire_uses(si);
4257
4258         connect_all_spills_with_keep(si);
4259
4260         del_set(si->values);
4261 }
4262
4263 static int
4264 get_n_regs(spill_ilp_t * si)
4265 {
4266         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4267
4268         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4269         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4270
4271         arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4272     be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4273
4274         bitset_andnot(arch_regs, abi_regs);
4275         arch_n_regs = bitset_popcnt(arch_regs);
4276
4277         bitset_free(arch_regs);
4278         bitset_free(abi_regs);
4279
4280         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4281         return arch_n_regs;
4282 }
4283
4284 static void
4285 walker_reload_mover(ir_node * bb, void * data)
4286 {
4287         spill_ilp_t   *si = data;
4288         ir_node           *tmp;
4289
4290         sched_foreach(bb, tmp) {
4291                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4292                         ir_node       *reload = tmp;
4293                         ir_node       *irn = tmp;
4294
4295                         /* move reload upwards */
4296
4297                         int pressure = (int)get_irn_link(reload);
4298                         if(pressure < si->n_regs) {
4299                                 irn = sched_prev(reload);
4300                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4301                                 sched_remove(reload);
4302                                 pressure = (int)get_irn_link(irn);
4303
4304                                 while(pressure < si->n_regs) {
4305                                         if( sched_is_end(irn) ||
4306                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4307                                            /* do not move reload before its spill */
4308                                            (irn == be_get_Reload_mem(reload)) ||
4309                                            /* do not move before phi */
4310                                            is_Phi(irn)) break;
4311
4312                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4313                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4314                                         irn = sched_prev(irn);
4315
4316                                         pressure = (int)get_irn_link(irn);
4317                                 }
4318
4319                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4320                                 sched_put_after(irn, reload);
4321                         }
4322                 }
4323         }
4324 }
4325
4326 static void
4327 move_reloads_upward(spill_ilp_t * si)
4328 {
4329         irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4330 }
4331
4332
4333 /**
4334  * Walk all irg blocks and check for interfering spills inside of phi classes
4335  */
4336 static void
4337 luke_meminterferencechecker(ir_node * bb, void * data)
4338 {
4339         spill_ilp_t *si = (spill_ilp_t*)data;
4340         int         l1, l2;
4341
4342         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4343                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4344
4345                 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4346                         continue;
4347
4348                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4349                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4350                         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)) {
4351                                 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4352
4353                                 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4354                                         continue;
4355
4356                                 /* a and b are only interesting if they are in the same phi class */
4357                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4358                                         if (values_interfere_in_block(si, bb, a, b)) {
4359                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4360                                         }
4361                                 }
4362                         }
4363                 }
4364         }
4365 }
4366
4367 static void
4368 verify_phiclasses(spill_ilp_t * si)
4369 {
4370         /* analyze phi classes */
4371         phi_class_free(si->pc);
4372         si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4373
4374         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4375         irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4376 }
4377
4378 void
4379 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4380 {
4381         char            buf[256];
4382         char            problem_name[256];
4383         char            dump_suffix[256];
4384         char            dump_suffix2[256];
4385         struct obstack  obst;
4386         spill_ilp_t     si;
4387         ir_graph       *irg = be_get_birg_irg(birg);
4388
4389         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4390         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4391         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4392
4393         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4394         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4395
4396         if(opt_verify & VERIFY_DOMINANCE)
4397                 be_check_dominance(irg);
4398
4399         be_assure_dom_front(birg);
4400         be_assure_liveness(birg);
4401
4402         obstack_init(&obst);
4403         si.obst                = &obst;
4404         si.birg                = birg;
4405         si.cls                 = cls;
4406         si.lpp                 = new_lpp(problem_name, lpp_minimize);
4407         si.remat_info          = new_set(cmp_remat_info, 4096);
4408         si.interferences       = new_set(cmp_interference, 32);
4409         si.memoperands         = new_set(cmp_memoperands, 128);
4410         si.all_possible_remats = pset_new_ptr_default();
4411         si.spills              = pset_new_ptr_default();
4412         si.inverse_ops         = pset_new_ptr_default();
4413         si.lv                  = birg->lv;
4414         si.keep                = NULL;
4415         si.n_regs              = get_n_regs(&si);
4416
4417         set_irg_link(irg, &si);
4418         compute_doms(irg);
4419
4420         /* compute phi classes */
4421         // phi_class_compute(irg);
4422
4423         if(opt_dump_flags & DUMP_STATS)
4424                 be_analyze_regpressure(birg, cls, "-pre");
4425
4426         DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4427         irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4428
4429         if(opt_remats) {
4430                 /* collect remats */
4431                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4432                 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4433         }
4434
4435         /* insert possible remats */
4436         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4437         irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4438         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4439
4440         if(opt_keep_alive & KEEPALIVE_REMATS) {
4441                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4442                 connect_all_remats_with_keep(&si);
4443                 /* dump graph with inserted remats */
4444                 dump_graph_with_remats(irg, dump_suffix);
4445         }
4446
4447         /* insert copies for phi arguments not in my regclass */
4448         irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4449
4450         /* recompute liveness */
4451         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4452         be_liveness_recompute(si.lv);
4453
4454         /* build the ILP */
4455         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4456         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4457         irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4458
4459         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4460         irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4461
4462         si.pc = phi_class_new_from_irg(birg->irg, 0);
4463         if (opt_memcopies) {
4464                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4465                 memcopyhandler(&si);
4466         }
4467
4468         if (opt_dump_flags & DUMP_PROBLEM) {
4469                 FILE           *f;
4470                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4471                 if ((f = fopen(buf, "wt")) != NULL) {
4472                         lpp_dump_plain(si.lpp, f);
4473                         fclose(f);
4474                 }
4475         }
4476
4477         if (opt_dump_flags & DUMP_MPS) {
4478                 FILE *f;
4479
4480                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4481                 if ((f = fopen(buf, "wt")) != NULL) {
4482                         mps_write_mps(si.lpp, s_mps_fixed, f);
4483                         fclose(f);
4484                 }
4485
4486                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4487                 if ((f = fopen(buf, "wt")) != NULL) {
4488                         mps_write_mst(si.lpp, s_mps_fixed, f);
4489                         fclose(f);
4490                 }
4491         }
4492
4493         lpp_check_startvals(si.lpp);
4494
4495 #ifdef SOLVE
4496         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4497         lpp_set_time_limit(si.lpp, opt_timeout);
4498
4499         if(opt_log)
4500                 lpp_set_log(si.lpp, stdout);
4501
4502 #ifdef SOLVE_LOCAL
4503         lpp_solve_cplex(si.lpp);
4504 #else
4505         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4506 #endif
4507         assert(lpp_is_sol_valid(si.lpp)
4508                && "solution of ILP must be valid");
4509
4510         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4511
4512         if(opt_dump_flags & DUMP_SOLUTION) {
4513                 FILE           *f;
4514                 char            buf[256];
4515
4516                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4517                 if ((f = fopen(buf, "wt")) != NULL) {
4518                         int             i;
4519                         for (i = 0; i < si.lpp->var_next; ++i) {
4520                                 lpp_name_t     *name = si.lpp->vars[i];
4521                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4522                         }
4523                         fclose(f);
4524                 }
4525         }
4526
4527 #ifndef SCHEDULE_PHIM
4528         si.phims = pset_new_ptr_default();
4529 #endif
4530         writeback_results(&si);
4531
4532
4533 #endif                          /* SOLVE */
4534
4535         kill_all_unused_values_in_schedule(&si);
4536
4537 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4538         del_pset(si.phims);
4539 #endif
4540
4541         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4542                 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4543
4544         // move reloads upwards
4545         be_liveness_recompute(si.lv);
4546         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4547         move_reloads_upward(&si);
4548
4549         if(opt_memcopies) {
4550                 verify_phiclasses(&si);
4551         }
4552
4553         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4554
4555         if(opt_dump_flags & DUMP_PRESSURE)
4556                 dump_pressure_graph(&si, dump_suffix2);
4557
4558         if(opt_dump_flags & DUMP_STATS)
4559                 be_analyze_regpressure(birg, cls, "-post");
4560
4561         if(opt_verify & VERIFY_DOMINANCE)
4562                 be_check_dominance(irg);
4563
4564         free_dom(irg);
4565         del_set(si.interferences);
4566         del_pset(si.inverse_ops);
4567         del_pset(si.all_possible_remats);
4568         del_set(si.memoperands);
4569         del_pset(si.spills);
4570         free_lpp(si.lpp);
4571         phi_class_free(si.pc);
4572         obstack_free(&obst, NULL);
4573         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4574 }
4575
4576 void be_init_spillremat(void)
4577 {
4578         static be_spiller_t remat_spiller = {
4579                 be_spill_remat
4580         };
4581         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4582         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4583         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4584         lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4585
4586         be_register_spiller("remat", &remat_spiller);
4587         lc_opt_add_table(remat_grp, options);
4588 }
4589
4590 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4591
4592 #else                           /* WITH_ILP */
4593
4594 static void INLINE
4595 only_that_you_can_compile_without_WITH_ILP_defined(void)
4596 {
4597 }
4598
4599 #endif                          /* WITH_ILP */