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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass_t.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35 #include "irvrfy.h"
  36
  37 #include <lpp/lpp.h>
  38 #include <lpp/mps.h>
  39 #include <lpp/lpp_net.h>
  40 #include <lpp/lpp_cplex.h>
  41 //#include <lc_pset.h>
  42 #include <libcore/lc_bitset.h>
  43
  44 #include "be_t.h"
  45 #include "belive_t.h"
  46 #include "besched_t.h"
  47 #include "beirgmod.h"
  48 #include "bearch.h"
  49 #include "benode_t.h"
  50 #include "beutil.h"
  51 #include "bespillremat.h"
  52 #include "bespill.h"
  53 #include "bepressurestat.h"
  54
  55 #include "bechordal_t.h"
  56
  57 #ifdef WITH_LIBCORE
  58 #include <libcore/lc_opts.h>
  59 #include <libcore/lc_opts_enum.h>
  60 #endif /* WITH_LIBCORE */
  61
  62 #define DUMP_PROBLEM  1
  63 #define DUMP_MPS      2
  64 #define DUMP_SOLUTION 4
  65
  66 static int dump_flags   = 0;
  67 static int activate_log = 0;
  68
  69 #ifdef WITH_LIBCORE
  70 static const lc_opt_enum_mask_items_t dump_items[] = {
  71         { "problem",  DUMP_PROBLEM  },
  72         { "mps",      DUMP_MPS      },
  73         { "solution", DUMP_SOLUTION },
  74         { NULL,      0 }
  75 };
  76
  77 static lc_opt_enum_mask_var_t dump_var = {
  78         &dump_flags, dump_items
  79 };
  80
  81 static const lc_opt_table_entry_t options[] = {
  82         LC_OPT_ENT_ENUM_MASK("dump", "dump ifg before, after or after each cloud",  &dump_var),
  83         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                        &activate_log),
  84         { NULL }
  85 };
  86
  87 void be_spill_remat_register_options(lc_opt_entry_t *grp)
  88 {
  89         lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
  90         lc_opt_add_table(my_grp, options);
  91 }
  92 #endif
  93
  94 //#define DUMP_SOLUTION
  95 //#define DUMP_ILP
  96 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
  97 //#define COLLECT_REMATS /* enable rematerialization */
  98 //#define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
  99 //#define ONLY_BRIGGS_REMATS /* only remats without parameters (or only with ignored params) */
 100 #define REMAT_WHILE_LIVE /* only remat values that are live */
 101 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
 102 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
 103 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
 104 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
 105 //#define KEEPALIVE_SPILLS
 106 //#define KEEPALIVE_RELOADS
 107 #define GOODWIN_REDUCTION
 108 //#define NO_MEMCOPIES
 109 //#define VERIFY_DOMINANCE
 110 #define WITH_MEMOPERANDS
 111
 112 #define  SOLVE
 113 //#define  SOLVE_LOCAL
 114 #define LPP_SERVER "i44pc52"
 115 #define LPP_SOLVER "cplex"
 116
 117 #define COST_LOAD        8
 118 #define COST_MEMOPERAND  7
 119 #define COST_STORE       50
 120 #define COST_REMAT       1
 121
 122 #define ILP_TIMEOUT    300
 123 #define MAX_PATHS      16
 124 #define ILP_UNDEF               -1
 125
 126 typedef struct _spill_ilp_t {
 127         const arch_register_class_t  *cls;
 128         int                           n_regs;
 129         const be_chordal_env_t       *chordal_env;
 130         be_lv_t                      *lv;
 131         lpp_t                        *lpp;
 132         struct obstack               *obst;
 133         set                          *remat_info;
 134         pset                         *all_possible_remats;
 135         pset                         *inverse_ops;
 136 #ifdef KEEPALIVE
 137         ir_node                      *keep;
 138 #endif
 139         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 140         pset                         *spills;
 141         set                          *interferences;
 142         ir_node                      *m_unknown;
 143 #ifdef WITH_MEMOPERANDS
 144         set                          *memoperands;
 145 #endif
 146         DEBUG_ONLY(firm_dbg_module_t * dbg);
 147 } spill_ilp_t;
 148
 149 typedef int ilp_var_t;
 150 typedef int ilp_cst_t;
 151
 152 typedef struct _spill_bb_t {
 153         set      *ilp;
 154         set      *reloads;
 155 } spill_bb_t;
 156
 157 typedef struct _remat_t {
 158         const ir_node        *op;      /**< for copy_irn */
 159         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 160         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 161         int                   cost;    /**< cost of this remat */
 162         int                   inverse; /**< nonzero if this is an inverse remat */
 163 } remat_t;
 164
 165 /**
 166  * Data to be attached to each IR node. For remats this contains the ilp_var
 167  * for this remat and for normal ops this contains the ilp_vars for
 168  * reloading each operand
 169  */
 170 typedef struct _op_t {
 171         int             is_remat;
 172         union {
 173                 struct {
 174                         ilp_var_t       ilp;
 175                         const remat_t  *remat; /** the remat this op belongs to */
 176                         int             pre; /** 1, if this is a pressure-increasing remat */
 177                 } remat;
 178                 struct {
 179                         ilp_var_t       ilp;
 180                         ir_node        *op; /** the operation this live range belongs to */
 181                         union {
 182                                 ilp_var_t      *reloads;
 183                                 ilp_var_t      *copies;
 184                         } args;
 185                 } live_range;
 186         } attr;
 187 } op_t;
 188
 189 typedef struct _defs_t {
 190         const ir_node   *value;
 191         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 192         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 193 } defs_t;
 194
 195 typedef struct _remat_info_t {
 196         const ir_node       *irn; /**< the irn to which these remats belong */
 197         pset                *remats; /**< possible remats for this value */
 198         pset                *remats_by_operand; /**< remats with this value as operand */
 199 } remat_info_t;
 200
 201 typedef struct _keyval_t {
 202         const void          *key;
 203         const void          *val;
 204 } keyval_t;
 205
 206 typedef struct _spill_t {
 207         ir_node            *irn;
 208         ilp_var_t           reg_in;
 209         ilp_var_t           mem_in;
 210         ilp_var_t           reg_out;
 211         ilp_var_t           mem_out;
 212         ilp_var_t           spill;
 213 } spill_t;
 214
 215 #ifdef WITH_MEMOPERANDS
 216 typedef struct _memoperand_t {
 217         ir_node             *irn; /**< the irn */
 218         unsigned int         pos; /**< the position of the argument */
 219         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 220 } memoperand_t;
 221 #endif
 222
 223 static INLINE int
 224 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 225 {
 226         return chordal_has_class(si->chordal_env, irn);
 227 }
 228
 229 #if 0
 230 static int
 231 cmp_remat(const void *a, const void *b)
 232 {
 233         const keyval_t *p = a;
 234         const keyval_t *q = b;
 235         const remat_t  *r = p->val;
 236         const remat_t  *s = q->val;
 237
 238         assert(r && s);
 239
 240         return !(r == s || r->op == s->op);
 241 }
 242 #endif
 243 static int
 244 cmp_remat(const void *a, const void *b)
 245 {
 246         const remat_t  *r = a;
 247         const remat_t  *s = a;
 248
 249         return !(r == s || r->op == s->op);
 250 }
 251
 252 static int
 253 cmp_spill(const void *a, const void *b, size_t size)
 254 {
 255         const spill_t *p = a;
 256         const spill_t *q = b;
 257
 258 //      return !(p->irn == q->irn && p->bb == q->bb);
 259         return !(p->irn == q->irn);
 260 }
 261
 262 #ifdef WITH_MEMOPERANDS
 263 static int
 264 cmp_memoperands(const void *a, const void *b, size_t size)
 265 {
 266         const memoperand_t *p = a;
 267         const memoperand_t *q = b;
 268
 269         return !(p->irn == q->irn && p->pos == q->pos);
 270 }
 271 #endif
 272
 273 static keyval_t *
 274 set_find_keyval(set * set, const void * key)
 275 {
 276         keyval_t     query;
 277
 278         query.key = key;
 279         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 280 }
 281
 282 static keyval_t *
 283 set_insert_keyval(set * set, void * key, void * val)
 284 {
 285         keyval_t     query;
 286
 287         query.key = key;
 288         query.val = val;
 289         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 290 }
 291
 292 static defs_t *
 293 set_find_def(set * set, const ir_node * value)
 294 {
 295         defs_t     query;
 296
 297         query.value = value;
 298         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 299 }
 300
 301 static defs_t *
 302 set_insert_def(set * set, const ir_node * value)
 303 {
 304         defs_t     query;
 305
 306         query.value = value;
 307         query.spills = NULL;
 308         query.remats = NULL;
 309         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 310 }
 311
 312 #ifdef WITH_MEMOPERANDS
 313 static memoperand_t *
 314 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 315 {
 316         memoperand_t     query;
 317
 318         query.irn = irn;
 319         query.pos = pos;
 320         query.ilp = ilp;
 321         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 322 }
 323
 324 static memoperand_t *
 325 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 326 {
 327         memoperand_t     query;
 328
 329         query.irn = (ir_node*)irn;
 330         query.pos = pos;
 331         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 332 }
 333 #endif
 334
 335
 336 static spill_t *
 337 set_find_spill(set * set, const ir_node * value)
 338 {
 339         spill_t     query;
 340
 341         query.irn = (ir_node*)value;
 342         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 343 }
 344
 345 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 346 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 347 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 348 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 349 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 350
 351 static int
 352 cmp_remat_info(const void *a, const void *b, size_t size)
 353 {
 354         const remat_info_t *p = a;
 355         const remat_info_t *q = b;
 356
 357         return !(p->irn == q->irn);
 358 }
 359
 360 static int
 361 cmp_defs(const void *a, const void *b, size_t size)
 362 {
 363         const defs_t *p = a;
 364         const defs_t *q = b;
 365
 366         return !(p->value == q->value);
 367 }
 368
 369 static int
 370 cmp_keyval(const void *a, const void *b, size_t size)
 371 {
 372         const keyval_t *p = a;
 373         const keyval_t *q = b;
 374
 375         return !(p->key == q->key);
 376 }
 377
 378 static double
 379 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 380 {
 381 #define FUDGE 0.001
 382 #ifndef EXECFREQ_LOOPDEPH
 383         return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
 384 #else
 385         if(is_Block(irn))
 386                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 387         else
 388                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 389 #endif
 390 }
 391
 392 static double
 393 get_cost(const spill_ilp_t * si, const ir_node * irn)
 394 {
 395         if(be_is_Spill(irn)) {
 396                 return COST_STORE;
 397         } else if(be_is_Reload(irn)){
 398                 return COST_LOAD;
 399         } else {
 400                 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
 401         }
 402 }
 403
 404 /**
 405  * Checks, whether node and its operands have suitable reg classes
 406  */
 407 static INLINE int
 408 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 409 {
 410         int               n;
 411         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 412         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 413
 414 #if 0
 415         if(!remat)
 416                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 417 #endif
 418
 419         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 420                 ir_node        *op = get_irn_n(irn, n);
 421                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 422
 423 //              if(!remat)
 424 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 425         }
 426
 427         return remat;
 428 }
 429
 430 /**
 431  * Try to create a remat from @p op with destination value @p dest_value
 432  */
 433 static INLINE remat_t *
 434 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 435 {
 436         remat_t  *remat = NULL;
 437
 438 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 439 //              return NULL;
 440
 441         if(dest_value == op) {
 442                 const ir_node *proj = NULL;
 443
 444                 if(is_Proj(dest_value)) {
 445                         op = get_irn_n(op, 0);
 446                         proj = dest_value;
 447                 }
 448
 449                 if(!is_rematerializable(si, op))
 450                         return NULL;
 451
 452                 remat = obstack_alloc(si->obst, sizeof(*remat));
 453                 remat->op = op;
 454                 remat->cost = get_cost(si, op);
 455                 remat->value = dest_value;
 456                 remat->proj = proj;
 457                 remat->inverse = 0;
 458         } else {
 459                 arch_inverse_t     inverse;
 460                 int                n;
 461
 462                 /* get the index of the operand we want to retrieve by the inverse op */
 463                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 464                         ir_node        *arg = get_irn_n(op, n);
 465
 466                         if(arg == dest_value) break;
 467                 }
 468                 if(n<0) return NULL;
 469
 470                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 471
 472                 /* else ask the backend to give an inverse op */
 473                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 474                         int   i;
 475
 476                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 477
 478                         assert(inverse.n > 0 && "inverse op should have at least one node");
 479
 480                         for(i=inverse.n-1; i>=0; --i) {
 481                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 482                         }
 483
 484                         if(inverse.n <= 2) {
 485                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 486                                 remat->op = inverse.nodes[0];
 487                                 remat->cost = inverse.costs;
 488                                 remat->value = dest_value;
 489                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 490                                 remat->inverse = 1;
 491
 492                                 assert(is_Proj(remat->proj));
 493                         } else {
 494                                 assert(0 && "I can not handle remats with more than 2 nodes");
 495                         }
 496                 }
 497         }
 498
 499         if(remat) {
 500                 if(remat->proj) {
 501                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 502                 } else {
 503                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 504                 }
 505         }
 506         return remat;
 507 }
 508
 509
 510 static INLINE void
 511 add_remat(const spill_ilp_t * si, const remat_t * remat)
 512 {
 513         remat_info_t    *remat_info,
 514                      query;
 515         int              n;
 516
 517         assert(remat->op);
 518         assert(remat->value);
 519
 520         query.irn = remat->value;
 521         query.remats = NULL;
 522         query.remats_by_operand = NULL;
 523         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 524
 525         if(remat_info->remats == NULL) {
 526                 remat_info->remats = new_pset(cmp_remat, 4096);
 527         }
 528         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 529
 530         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 531         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 532                 ir_node        *arg = get_irn_n(remat->op, n);
 533
 534                 query.irn = arg;
 535                 query.remats = NULL;
 536                 query.remats_by_operand = NULL;
 537                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 538
 539                 if(remat_info->remats_by_operand == NULL) {
 540                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 541                 }
 542                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 543         }
 544 }
 545
 546 #ifdef NO_SINGLE_USE_REMATS
 547 static int
 548 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 549 {
 550         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 551         int                i = 0;
 552
 553         while(edge) {
 554                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 555                         ++i;
 556                 }
 557                 edge = get_irn_out_edge_next(irn, edge);
 558         }
 559
 560         return i;
 561 }
 562 #endif
 563
 564 #ifdef ONLY_BRIGGS_REMATS
 565 static int
 566 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 567 {
 568         int n;
 569         unsigned int ret = 0;
 570
 571         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 572                 if(has_reg_class(si, irn)) ++ret;
 573         }
 574
 575         return ret;
 576 }
 577 #endif
 578
 579 static INLINE void
 580 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 581 {
 582         int      n;
 583         remat_t *remat;
 584
 585         if( has_reg_class(si, op)
 586 #ifdef NO_SINGLE_USE_REMATS
 587         && (get_irn_n_nonremat_edges(si, op) > 1)
 588 #endif
 589 #ifdef ONLY_BRIGGS_REMATS
 590         && (get_irn_n_nonignore_args(si, op) == 0)
 591 #endif
 592         ) {
 593                 remat = get_remat_from_op(si, op, op);
 594                 if(remat) {
 595                         add_remat(si, remat);
 596                 }
 597         }
 598
 599 #if defined(COLLECT_INVERSE_REMATS) && !defined(ONLY_BRIGGS_REMATS)
 600         /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 601            for each arg */
 602         for (n = get_irn_arity(op)-1; n>=0; --n) {
 603                 ir_node        *arg = get_irn_n(op, n);
 604
 605                 if(has_reg_class(si, arg)) {
 606                         /* try to get an inverse remat */
 607                         remat = get_remat_from_op(si, arg, op);
 608                         if(remat) {
 609                                 add_remat(si, remat);
 610                         }
 611                 }
 612         }
 613 #endif
 614
 615 }
 616
 617 static INLINE int
 618 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 619 {
 620         ir_node *block;
 621         ir_node *def_block = get_nodes_block(val);
 622         int      ret;
 623
 624         if(val == pos)
 625                 return 0;
 626
 627         /* if pos is at end of a basic block */
 628         if(is_Block(pos)) {
 629                 ret = (pos == def_block || block_dominates(def_block, pos));
 630 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 631                 return ret;
 632         }
 633
 634         /* else if this is a normal operation */
 635         block = get_nodes_block(pos);
 636         if(block == def_block) {
 637                 if(!sched_is_scheduled(val)) return 1;
 638
 639                 ret = sched_comes_after(val, pos);
 640 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 641                 return ret;
 642         }
 643
 644         ret = block_dominates(def_block, block);
 645 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 646         return ret;
 647 }
 648
 649 static INLINE ir_node *
 650 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 651 {
 652     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 653 }
 654
 655 /**
 656  * Returns first non-Phi node of block @p bb
 657  */
 658 static INLINE ir_node *
 659 sched_block_first_nonphi(const ir_node * bb)
 660 {
 661         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 662 }
 663
 664 static int
 665 sched_skip_proj_predicator(const ir_node * irn, void * data)
 666 {
 667         return (is_Proj(irn));
 668 }
 669
 670 static INLINE ir_node *
 671 sched_next_nonproj(const ir_node * irn, int forward)
 672 {
 673         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 674 }
 675
 676 /**
 677  * Returns next operation node (non-Proj) after @p irn
 678  * or the basic block of this node
 679  */
 680 static INLINE ir_node *
 681 sched_next_op(const ir_node * irn)
 682 {
 683         ir_node *next = sched_next(irn);
 684
 685         if(is_Block(next))
 686                 return next;
 687
 688         return sched_next_nonproj(next, 1);
 689 }
 690
 691 /**
 692  * Returns previous operation node (non-Proj) before @p irn
 693  * or the basic block of this node
 694  */
 695 static INLINE ir_node *
 696 sched_prev_op(const ir_node * irn)
 697 {
 698         ir_node *prev = sched_prev(irn);
 699
 700         if(is_Block(prev))
 701                 return prev;
 702
 703         return sched_next_nonproj(prev, 0);
 704 }
 705
 706 static void
 707 sched_put_after(ir_node * insert, ir_node * irn)
 708 {
 709         if(is_Block(insert)) {
 710                 insert = sched_block_first_nonphi(insert);
 711         } else {
 712                 insert = sched_next_op(insert);
 713         }
 714         sched_add_before(insert, irn);
 715 }
 716
 717 static void
 718 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 719 {
 720   if(is_Block(insert)) {
 721           insert = sched_block_last_noncf(si, insert);
 722   } else {
 723           insert = sched_next_nonproj(insert, 0);
 724           insert = sched_prev(insert);
 725   }
 726   sched_add_after(insert, irn);
 727 }
 728
 729 /**
 730  * Tells you whether a @p remat can be placed before the irn @p pos
 731  */
 732 static INLINE int
 733 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 734 {
 735         const ir_node   *op = remat->op;
 736         const ir_node   *prev;
 737         int        n,
 738                            res = 1;
 739
 740         if(is_Block(pos)) {
 741                 prev = sched_block_last_noncf(si, pos);
 742                 prev = sched_next_nonproj(prev, 0);
 743         } else {
 744                 prev = sched_prev_op(pos);
 745         }
 746         /* do not remat if the rematted value is defined immediately before this op */
 747         if(prev == remat->op) {
 748                 return 0;
 749         }
 750
 751 #if 0
 752         /* this should be just fine, the following OP will be using this value, right? */
 753
 754         /* only remat AFTER the real definition of a value (?) */
 755         if(!value_is_defined_before(si, pos, remat->value)) {
 756 //              ir_fprintf(stderr, "error(not defined)");
 757                 return 0;
 758         }
 759 #endif
 760
 761         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 762                 const ir_node   *arg = get_irn_n(op, n);
 763
 764 #ifdef NO_ENLARGE_L1V3N355
 765                 if(has_reg_class(si, arg) && live) {
 766                         res &= pset_find_ptr(live, arg)?1:0;
 767                 } else {
 768                         res &= value_is_defined_before(si, pos, arg);
 769                 }
 770 #else
 771                 res &= value_is_defined_before(si, pos, arg);
 772 #endif
 773         }
 774
 775         return res;
 776 }
 777
 778 /**
 779  * Tells you whether a @p remat can be placed after the irn @p pos
 780  */
 781 static INLINE int
 782 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 783 {
 784         if(is_Block(pos)) {
 785                 pos = sched_block_first_nonphi(pos);
 786         } else {
 787                 pos = sched_next_op(pos);
 788         }
 789
 790         /* only remat AFTER the real definition of a value (?) */
 791         if(!value_is_defined_before(si, pos, remat->value)) {
 792                 return 0;
 793         }
 794
 795         return can_remat_before(si, remat, pos, live);
 796 }
 797
 798 /**
 799  * Collect potetially rematerializable OPs
 800  */
 801 static void
 802 walker_remat_collector(ir_node * irn, void * data)
 803 {
 804         spill_ilp_t    *si = data;
 805
 806         if(!is_Block(irn) && !is_Phi(irn)) {
 807                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 808                 get_remats_from_op(si, irn);
 809         }
 810 }
 811
 812 /**
 813  * Inserts a copy of @p irn before @p pos
 814  */
 815 static ir_node *
 816 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 817 {
 818         ir_node     *bb;
 819         ir_node     *copy;
 820
 821         bb = is_Block(pos)?pos:get_nodes_block(pos);
 822         copy = exact_copy(irn);
 823
 824         _set_phi_class(copy, NULL);
 825         set_nodes_block(copy, bb);
 826         sched_put_before(si, pos, copy);
 827
 828         return copy;
 829 }
 830
 831 /**
 832  * Inserts a copy of @p irn after @p pos
 833  */
 834 static ir_node *
 835 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 836 {
 837         ir_node     *bb;
 838         ir_node     *copy;
 839
 840         bb = is_Block(pos)?pos:get_nodes_block(pos);
 841         copy = exact_copy(irn);
 842
 843         _set_phi_class(copy, NULL);
 844         set_nodes_block(copy, bb);
 845         sched_put_after(pos, copy);
 846
 847         return copy;
 848 }
 849
 850 static ir_node *
 851 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 852 {
 853         char     buf[256];
 854
 855         if(can_remat_after(si, remat, pos, live)) {
 856                 ir_node         *copy,
 857                                                 *proj_copy;
 858                 op_t            *op;
 859
 860                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 861
 862                 copy = insert_copy_after(si, remat->op, pos);
 863
 864                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 865                 op = obstack_alloc(si->obst, sizeof(*op));
 866                 op->is_remat = 1;
 867                 op->attr.remat.remat = remat;
 868                 op->attr.remat.pre = 0;
 869                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 870
 871                 set_irn_link(copy, op);
 872                 pset_insert_ptr(si->all_possible_remats, copy);
 873                 if(remat->proj) {
 874                         proj_copy = insert_copy_after(si, remat->proj, copy);
 875                         set_irn_n(proj_copy, 0, copy);
 876                         set_irn_link(proj_copy, op);
 877                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 878                 } else {
 879                         proj_copy = NULL;
 880                 }
 881
 882                 return copy;
 883         }
 884
 885         return NULL;
 886 }
 887
 888 static ir_node *
 889 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 890 {
 891         char     buf[256];
 892
 893         if(can_remat_before(si, remat, pos, live)) {
 894                 ir_node         *copy,
 895                                                 *proj_copy;
 896                 op_t            *op;
 897
 898                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 899
 900                 copy = insert_copy_before(si, remat->op, pos);
 901
 902                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 903                 op = obstack_alloc(si->obst, sizeof(*op));
 904                 op->is_remat = 1;
 905                 op->attr.remat.remat = remat;
 906                 op->attr.remat.pre = 1;
 907                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 908
 909                 set_irn_link(copy, op);
 910                 pset_insert_ptr(si->all_possible_remats, copy);
 911                 if(remat->proj) {
 912                         proj_copy = insert_copy_after(si, remat->proj, copy);
 913                         set_irn_n(proj_copy, 0, copy);
 914                         set_irn_link(proj_copy, op);
 915                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 916                 } else {
 917                         proj_copy = NULL;
 918                 }
 919
 920                 return copy;
 921         }
 922
 923         return NULL;
 924 }
 925
 926 static int
 927 get_block_n_succs(const ir_node *block) {
 928         const ir_edge_t *edge;
 929
 930         assert(edges_activated(current_ir_graph));
 931
 932         edge = get_block_succ_first(block);
 933         if (! edge)
 934                 return 0;
 935
 936         edge = get_block_succ_next(block, edge);
 937         return edge ? 2 : 1;
 938 }
 939
 940 static int
 941 is_merge_edge(const ir_node * bb)
 942 {
 943 #ifdef GOODWIN_REDUCTION
 944         return get_block_n_succs(bb) == 1;
 945 #else
 946         return 1;
 947 #endif
 948 }
 949
 950 static int
 951 is_diverge_edge(const ir_node * bb)
 952 {
 953 #ifdef GOODWIN_REDUCTION
 954         return get_Block_n_cfgpreds(bb) == 1;
 955 #else
 956         return 1;
 957 #endif
 958 }
 959
 960 static void
 961 walker_regclass_copy_insertor(ir_node * irn, void * data)
 962 {
 963         spill_ilp_t    *si = data;
 964
 965         if(is_Phi(irn) && has_reg_class(si, irn)) {
 966                 int n;
 967
 968                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
 969                         ir_node  *phi_arg = get_irn_n(irn, n);
 970                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
 971
 972                         if(!has_reg_class(si, phi_arg)) {
 973                                 ir_node   *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
 974                                 ir_node   *pos = sched_block_last_noncf(si, bb);
 975                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
 976
 977                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
 978                                 sched_add_after(pos, copy);
 979                                 set_irn_n(irn, n, copy);
 980
 981                                 op->is_remat = 0;
 982                                 op->attr.live_range.args.reloads = NULL;
 983                                 op->attr.live_range.ilp = ILP_UNDEF;
 984                                 set_irn_link(copy, op);
 985                         }
 986                 }
 987         }
 988 }
 989
 990
 991 /**
 992  * Insert (so far unused) remats into the irg to
 993  * recompute the potential liveness of all values
 994  */
 995 static void
 996 walker_remat_insertor(ir_node * bb, void * data)
 997 {
 998         spill_ilp_t    *si = data;
 999         spill_bb_t     *spill_bb;
1000         ir_node        *irn;
1001         int             n, i;
1002         pset           *live = pset_new_ptr_default();
1003
1004         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1005
1006         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1007                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1008
1009                 /* add remats at end of block */
1010                 if (has_reg_class(si, value)) {
1011                         pset_insert_ptr(live, value);
1012                 }
1013         }
1014
1015         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1016         set_irn_link(bb, spill_bb);
1017
1018         irn = sched_last(bb);
1019         while(!sched_is_end(irn)) {
1020                 ir_node   *next;
1021                 op_t      *op;
1022                 pset      *args;
1023                 ir_node   *arg;
1024                 pset      *remat_args;
1025
1026                 next = sched_prev(irn);
1027
1028                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1029
1030                 if(is_Phi(irn) || is_Proj(irn)) {
1031                         op_t      *op;
1032
1033                         if(has_reg_class(si, irn)) {
1034                                 pset_remove_ptr(live, irn);
1035                         }
1036
1037                         op = obstack_alloc(si->obst, sizeof(*op));
1038                         op->is_remat = 0;
1039                         op->attr.live_range.args.reloads = NULL;
1040                         op->attr.live_range.ilp = ILP_UNDEF;
1041                         set_irn_link(irn, op);
1042
1043                         irn = next;
1044                         continue;
1045                 }
1046
1047                 op = obstack_alloc(si->obst, sizeof(*op));
1048                 op->is_remat = 0;
1049                 op->attr.live_range.ilp = ILP_UNDEF;
1050                 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1051                 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1052                 set_irn_link(irn, op);
1053
1054                 args = pset_new_ptr_default();
1055
1056                 /* collect arguments of op */
1057                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1058                         ir_node        *arg = get_irn_n(irn, n);
1059
1060                         pset_insert_ptr(args, arg);
1061                 }
1062
1063                 /* set args of op already live in epilog */
1064                 pset_foreach(args, arg) {
1065                         if(has_reg_class(si, arg)) {
1066                                 pset_insert_ptr(live, arg);
1067                         }
1068                 }
1069                 /* delete defined value from live set */
1070                 if(has_reg_class(si, irn)) {
1071                         pset_remove_ptr(live, irn);
1072                 }
1073
1074
1075                 remat_args = pset_new_ptr_default();
1076
1077                 /* insert all possible remats before irn */
1078                 pset_foreach(args, arg) {
1079                         remat_info_t   *remat_info,
1080                                                     query;
1081                         remat_t        *remat;
1082
1083                         /* continue if the operand has the wrong reg class
1084                          */
1085                         if(!has_reg_class(si, arg))
1086                                 continue;
1087
1088                         query.irn = arg;
1089                         query.remats = NULL;
1090                         query.remats_by_operand = NULL;
1091                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1092
1093                         if(!remat_info) {
1094                                 continue;
1095                         }
1096
1097                         if(remat_info->remats) {
1098                                 pset_foreach(remat_info->remats, remat) {
1099                                         ir_node  *remat_irn = NULL;
1100
1101                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1102 #ifdef REMAT_WHILE_LIVE
1103                                         if(pset_find_ptr(live, remat->value)) {
1104                                                 remat_irn = insert_remat_before(si, remat, irn, live);
1105                                         }
1106 #else
1107                                         remat_irn = insert_remat_before(si, remat, irn, live);
1108 #endif
1109                                         if(remat_irn) {
1110                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1111                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1112
1113                                                         if(!has_reg_class(si, remat_arg)) continue;
1114
1115                                                         pset_insert_ptr(remat_args, remat_arg);
1116                                                 }
1117                                         }
1118                                 }
1119                         }
1120                 }
1121
1122                 /* now we add remat args to op's args because they could also die at this op */
1123                 pset_foreach(args,arg) {
1124                         if(pset_find_ptr(remat_args, arg)) {
1125                                 pset_remove_ptr(remat_args, arg);
1126                         }
1127                 }
1128                 pset_foreach(remat_args,arg) {
1129                         pset_insert_ptr(args, arg);
1130                 }
1131
1132                 /* insert all possible remats after irn */
1133                 pset_foreach(args, arg) {
1134                         remat_info_t   *remat_info,
1135                                                     query;
1136                         remat_t        *remat;
1137
1138                         /* continue if the operand has the wrong reg class */
1139                         if(!has_reg_class(si, arg))
1140                                 continue;
1141
1142                         query.irn = arg;
1143                         query.remats = NULL;
1144                         query.remats_by_operand = NULL;
1145                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1146
1147                         if(!remat_info) {
1148                                 continue;
1149                         }
1150
1151                         /* do not place post remats after jumps */
1152                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1153
1154                         if(remat_info->remats_by_operand) {
1155                                 pset_foreach(remat_info->remats_by_operand, remat) {
1156                                         /* do not insert remats producing the same value as one of the operands */
1157                                         if(!pset_find_ptr(args, remat->value)) {
1158                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1159 #ifdef REMAT_WHILE_LIVE
1160                                                 if(pset_find_ptr(live, remat->value)) {
1161                                                         insert_remat_after(si, remat, irn, live);
1162                                                 }
1163 #else
1164                                                 insert_remat_after(si, remat, irn, live);
1165 #endif
1166                                         }
1167                                 }
1168                         }
1169                 }
1170
1171                 del_pset(remat_args);
1172                 del_pset(args);
1173                 irn = next;
1174         }
1175
1176         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1177                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1178
1179                 /* add remats at end if successor has multiple predecessors */
1180                 if(is_merge_edge(bb)) {
1181                         /* add remats at end of block */
1182                         if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1183                                 remat_info_t   *remat_info,
1184                                                            query;
1185                                 remat_t        *remat;
1186
1187                                 query.irn = value;
1188                                 query.remats = NULL;
1189                                 query.remats_by_operand = NULL;
1190                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1191
1192                                 if(remat_info && remat_info->remats) {
1193                                         pset_foreach(remat_info->remats, remat) {
1194                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1195
1196                                                 insert_remat_before(si, remat, bb, NULL);
1197                                         }
1198                                 }
1199                         }
1200                 }
1201                 if(is_diverge_edge(bb)) {
1202                         /* add remat2s at beginning of block */
1203                         if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1204                                 remat_info_t   *remat_info,
1205                                                            query;
1206                                 remat_t        *remat;
1207
1208                                 query.irn = value;
1209                                 query.remats = NULL;
1210                                 query.remats_by_operand = NULL;
1211                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1212
1213                                 if(remat_info && remat_info->remats) {
1214                                         pset_foreach(remat_info->remats, remat) {
1215                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
1216
1217                                                 /* put the remat here if all its args are available */
1218                                                 insert_remat_after(si, remat, bb, NULL);
1219
1220                                         }
1221                                 }
1222                         }
1223                 }
1224         }
1225 }
1226
1227 /**
1228  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1229  */
1230 static void
1231 luke_endwalker(ir_node * bb, void * data)
1232 {
1233         spill_ilp_t    *si = (spill_ilp_t*)data;
1234         pset           *live;
1235         pset           *use_end;
1236         char            buf[256];
1237         ilp_cst_t       cst;
1238         ir_node        *irn;
1239         spill_bb_t     *spill_bb = get_irn_link(bb);
1240         int             i;
1241
1242
1243         live = pset_new_ptr_default();
1244         use_end = pset_new_ptr_default();
1245
1246         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1247                 irn = be_lv_get_irn(si->lv, bb, i);
1248                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1249                         op_t      *op;
1250
1251                         pset_insert_ptr(live, irn);
1252                         op = get_irn_link(irn);
1253                         assert(!op->is_remat);
1254                 }
1255         }
1256
1257         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1258         /* their reg_out must always be set */
1259         sched_foreach_reverse(bb, irn) {
1260                 int   n;
1261
1262                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1263
1264                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1265                         ir_node        *irn_arg = get_irn_n(irn, n);
1266
1267                         if(has_reg_class(si, irn_arg)) {
1268                                 pset_insert_ptr(use_end, irn_arg);
1269                         }
1270                 }
1271         }
1272
1273         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1274         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1275         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1276
1277         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1278
1279         /* if this is a merge edge we can reload at the end of this block */
1280         if(is_merge_edge(bb)) {
1281                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1282         } else if(pset_count(use_end)){
1283                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1284         } else {
1285                 spill_bb->reloads = NULL;
1286         }
1287
1288         pset_foreach(live,irn) {
1289                 spill_t     query,
1290                                         *spill;
1291                 double      spill_cost;
1292                 int         default_spilled;
1293
1294
1295                 /* handle values used by control flow nodes later separately */
1296                 if(pset_find_ptr(use_end, irn)) continue;
1297
1298                 query.irn = irn;
1299                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1300
1301                 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1302
1303                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1304                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1305                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1306
1307                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1308                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1309
1310                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1311                 /* by default spill value right after definition */
1312                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1313                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1314
1315                 if(is_merge_edge(bb)) {
1316                         ilp_var_t   reload;
1317                         ilp_cst_t   rel_cst;
1318
1319                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1320                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 0.0);
1321                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1322
1323                         /* reload <= mem_out */
1324                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1325                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1326                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1327                 }
1328
1329                 spill->reg_in = ILP_UNDEF;
1330                 spill->mem_in = ILP_UNDEF;
1331         }
1332
1333         pset_foreach(use_end,irn) {
1334                 spill_t     query,
1335                                         *spill;
1336                 double      spill_cost;
1337                 ilp_cst_t   end_use_req,
1338                                         rel_cst;
1339                 ilp_var_t   reload;
1340                 int         default_spilled;
1341
1342                 query.irn = irn;
1343                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1344
1345                 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1346
1347                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1348                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1349                 /* if irn is used at the end of the block, then it is live anyway */
1350                 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1351
1352                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1353                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1354
1355                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1356                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1357                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1358
1359                 /* reload for use be control flow op */
1360                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1361                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 1.0);
1362                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1363
1364                 /* reload <= mem_out */
1365                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1366                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1367                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1368
1369                 spill->reg_in = ILP_UNDEF;
1370                 spill->mem_in = ILP_UNDEF;
1371
1372                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1373                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1374                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1375         }
1376
1377         del_pset(live);
1378         del_pset(use_end);
1379 }
1380
1381 static ir_node *
1382 next_post_remat(const ir_node * irn)
1383 {
1384         op_t      *op;
1385     ir_node   *next;
1386
1387         if(is_Block(irn)) {
1388                 next = sched_block_first_nonphi(irn);
1389         } else {
1390                 next = sched_next_op(irn);
1391         }
1392
1393         if(sched_is_end(next))
1394                 return NULL;
1395
1396         op = get_irn_link(next);
1397         if(op->is_remat && !op->attr.remat.pre) {
1398                 return next;
1399         }
1400
1401         return NULL;
1402 }
1403
1404
1405 static ir_node *
1406 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1407 {
1408         op_t      *op;
1409         ir_node   *ret;
1410
1411         if(is_Block(irn)) {
1412                 ret = sched_block_last_noncf(si, irn);
1413                 ret = sched_next(ret);
1414                 ret = sched_prev_op(ret);
1415         } else {
1416                 ret = sched_prev_op(irn);
1417         }
1418
1419         if(sched_is_end(ret) || is_Phi(ret))
1420                 return NULL;
1421
1422         op = (op_t*)get_irn_link(ret);
1423         if(op->is_remat && op->attr.remat.pre) {
1424                 return ret;
1425         }
1426
1427         return NULL;
1428 }
1429
1430 /**
1431  * Find a remat of value @p value in the epilog of @p pos
1432  */
1433 static ir_node *
1434 find_post_remat(const ir_node * value, const ir_node * pos)
1435 {
1436         while((pos = next_post_remat(pos)) != NULL) {
1437                 op_t   *op;
1438
1439                 op = get_irn_link(pos);
1440                 assert(op->is_remat && !op->attr.remat.pre);
1441
1442                 if(op->attr.remat.remat->value == value)
1443                         return (ir_node*)pos;
1444
1445 #if 0
1446         const ir_edge_t *edge;
1447                 foreach_out_edge(pos, edge) {
1448                         ir_node   *proj = get_edge_src_irn(edge);
1449                         assert(is_Proj(proj));
1450                 }
1451 #endif
1452
1453         }
1454
1455         return NULL;
1456 }
1457
1458 static spill_t *
1459 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1460 {
1461         spill_bb_t  *spill_bb = get_irn_link(bb);
1462         spill_t     *spill,
1463                                  query;
1464         char         buf[256];
1465         int          default_spilled;
1466
1467         query.irn = irn;
1468         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1469         if(!spill) {
1470                 double   spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1471
1472                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1473
1474                 spill->reg_out = ILP_UNDEF;
1475                 spill->reg_in  = ILP_UNDEF;
1476                 spill->mem_in  = ILP_UNDEF;
1477
1478                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1479                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1480
1481                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1482                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1483                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1484         }
1485
1486         return spill;
1487 }
1488
1489 static void
1490 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1491 {
1492         ir_node        *irn;
1493         int i;
1494
1495         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1496                 irn = be_lv_get_irn(si->lv, bb, i);
1497
1498                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1499                         pset_insert_ptr(live, irn);
1500                 }
1501         }
1502
1503         irn = sched_last(bb);
1504
1505         /* all values eaten by control flow operations are also live until the end of the block */
1506         sched_foreach_reverse(bb, irn) {
1507                 int  i;
1508
1509                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1510
1511                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1512                         ir_node *arg = get_irn_n(irn,i);
1513
1514                         if(has_reg_class(si, arg)) {
1515                                 pset_insert_ptr(live, arg);
1516                         }
1517                 }
1518         }
1519 }
1520
1521 /**
1522  *  Inserts ILP-constraints and variables for memory copying before the given position
1523  */
1524 static void
1525 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1526 {
1527         const ir_node    *succ;
1528         const ir_edge_t  *edge;
1529         spill_bb_t       *spill_bb = get_irn_link(block);
1530         ir_node          *phi;
1531         int               pos;
1532         ilp_cst_t         cst;
1533         ilp_var_t         copyreg;
1534         char              buf[256];
1535         ir_node          *tmp;
1536
1537
1538         assert(edges_activated(current_ir_graph));
1539
1540         edge = get_block_succ_first(block);
1541         if(!edge) return;
1542
1543         succ = edge->src;
1544         pos = edge->pos;
1545
1546         edge = get_block_succ_next(block, edge);
1547         /* next block can only contain phis, if this is a merge edge */
1548         if(edge) return;
1549
1550         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1551         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1552
1553         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1554         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1555
1556         pset_foreach(live, tmp) {
1557                 spill_t  *spill;
1558 #if 0
1559                 op_t  *op = get_irn_link(irn);
1560                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1561 #endif
1562                 spill = set_find_spill(spill_bb->ilp, tmp);
1563                 assert(spill);
1564
1565                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1566         }
1567         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1568
1569         sched_foreach(succ, phi) {
1570                 const ir_node  *to_copy;
1571                 op_t           *to_copy_op;
1572                 spill_t        *to_copy_spill;
1573                 op_t           *phi_op = get_irn_link(phi);
1574                 ilp_var_t       reload = ILP_UNDEF;
1575
1576
1577                 if(!is_Phi(phi)) break;
1578                 if(!has_reg_class(si, phi)) continue;
1579
1580                 to_copy = get_irn_n(phi, pos);
1581
1582                 to_copy_op = get_irn_link(to_copy);
1583
1584                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1585                 assert(to_copy_spill);
1586
1587                 if(spill_bb->reloads) {
1588                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1589
1590                         if(keyval) {
1591                                 reload = PTR_TO_INT(keyval->val);
1592                         }
1593                 }
1594
1595                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1596                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1597
1598                 /* copy - reg_out - reload - remat - live_range <= 0 */
1599                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1600                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1601                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1602                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1603                 foreach_pre_remat(si, block, tmp) {
1604                         op_t     *remat_op = get_irn_link(tmp);
1605                         if(remat_op->attr.remat.remat->value == to_copy) {
1606                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1607                         }
1608                 }
1609
1610                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1611                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1612
1613                 /* copy - reg_out - copyreg <= 0 */
1614                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1615                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1616                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1617         }
1618 }
1619
1620
1621 /**
1622  * Walk all irg blocks and emit this ILP
1623  */
1624 static void
1625 luke_blockwalker(ir_node * bb, void * data)
1626 {
1627         spill_ilp_t    *si = (spill_ilp_t*)data;
1628         ir_node        *irn;
1629         pset           *live;
1630         char            buf[256];
1631         ilp_cst_t       cst;
1632         spill_bb_t     *spill_bb = get_irn_link(bb);
1633         ir_node        *tmp;
1634         spill_t        *spill;
1635         pset           *defs = pset_new_ptr_default();
1636 #ifdef WITH_MEMOPERANDS
1637         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1638 #endif
1639
1640
1641         live = pset_new_ptr_default();
1642
1643         /****************************************
1644          *      B A S I C  B L O C K  E N D
1645          ***************************************/
1646
1647
1648         /* init live values at end of block */
1649         get_live_end(si, bb, live);
1650
1651         pset_foreach(live, irn) {
1652                 op_t           *op;
1653                 ilp_var_t       reload = ILP_UNDEF;
1654
1655                 spill = set_find_spill(spill_bb->ilp, irn);
1656                 assert(spill);
1657
1658                 if(spill_bb->reloads) {
1659                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1660
1661                         if(keyval) {
1662                                 reload = PTR_TO_INT(keyval->val);
1663                         }
1664                 }
1665
1666                 op = get_irn_link(irn);
1667                 assert(!op->is_remat);
1668
1669                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1670                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1671                 op->attr.live_range.op = bb;
1672
1673                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1674                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1675
1676                 /* reg_out - reload - remat - live_range <= 0 */
1677                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1678                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1679                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1680                 foreach_pre_remat(si, bb, tmp) {
1681                         op_t     *remat_op = get_irn_link(tmp);
1682                         if(remat_op->attr.remat.remat->value == irn) {
1683                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1684                         }
1685                 }
1686                 /* maybe we should also assure that reg_out >= live_range etc. */
1687         }
1688
1689 #ifndef NO_MEMCOPIES
1690         insert_mem_copy_position(si, live, bb);
1691 #endif
1692
1693         /*
1694          * start new live ranges for values used by remats at end of block
1695          * and assure the remat args are available
1696          */
1697         foreach_pre_remat(si, bb, tmp) {
1698                 op_t     *remat_op = get_irn_link(tmp);
1699                 int       n;
1700
1701                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1702                         ir_node        *remat_arg = get_irn_n(tmp, n);
1703                         op_t           *arg_op = get_irn_link(remat_arg);
1704                         ilp_var_t       prev_lr;
1705
1706                         if(!has_reg_class(si, remat_arg)) continue;
1707
1708                         /* if value is becoming live through use by remat */
1709                         if(!pset_find_ptr(live, remat_arg)) {
1710                                 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1711                                 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1712
1713                                 arg_op->attr.live_range.ilp = prev_lr;
1714                                 arg_op->attr.live_range.op = bb;
1715
1716                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1717
1718                                 pset_insert_ptr(live, remat_arg);
1719                                 add_to_spill_bb(si, bb, remat_arg);
1720                         }
1721
1722                         /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1723                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1724                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1725
1726                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1727                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1728
1729                         /* use reload placed for this argument */
1730                         if(spill_bb->reloads) {
1731                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1732
1733                                 if(keyval) {
1734                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1735
1736                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1737                                 }
1738                         }
1739                 }
1740         }
1741         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1742
1743
1744
1745
1746         /**************************************
1747          *    B A S I C  B L O C K  B O D Y
1748          **************************************/
1749
1750         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1751                 op_t       *op;
1752                 op_t       *tmp_op;
1753                 int         n,
1754                                         u = 0,
1755                                         d = 0;
1756                 ilp_cst_t       check_pre,
1757                                         check_post;
1758                 set        *args;
1759                 pset       *used;
1760                 pset       *remat_defs;
1761                 keyval_t   *keyval;
1762 #ifdef WITH_MEMOPERANDS
1763                 ilp_cst_t   one_memoperand;
1764 #endif
1765
1766                 /* iterate only until first phi */
1767                 if(is_Phi(irn))
1768                         break;
1769
1770                 op = get_irn_link(irn);
1771                 /* skip remats */
1772                 if(op->is_remat) continue;
1773                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1774
1775                 /* collect defined values */
1776                 if(has_reg_class(si, irn)) {
1777                         pset_insert_ptr(defs, irn);
1778                 }
1779
1780                 /* skip projs */
1781                 if(is_Proj(irn)) continue;
1782
1783                 /*
1784                  * init set of irn's arguments
1785                  * and all possibly used values around this op
1786                  * and values defined by post remats
1787                  */
1788                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1789                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1790                 remat_defs = pset_new_ptr(pset_count(live));
1791
1792                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1793                         ir_node        *irn_arg = get_irn_n(irn, n);
1794                         if(has_reg_class(si, irn_arg)) {
1795                                 set_insert_keyval(args, irn_arg, (void*)n);
1796                                 pset_insert_ptr(used, irn_arg);
1797                         }
1798                 }
1799                 foreach_post_remat(irn, tmp) {
1800                         op_t    *remat_op = get_irn_link(tmp);
1801
1802                         pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1803
1804                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1805                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1806                                 if(has_reg_class(si, remat_arg)) {
1807                                         pset_insert_ptr(used, remat_arg);
1808                                 }
1809                         }
1810                 }
1811                 foreach_pre_remat(si, irn, tmp) {
1812                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1813                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1814                                 if(has_reg_class(si, remat_arg)) {
1815                                         pset_insert_ptr(used, remat_arg);
1816                                 }
1817                         }
1818                 }
1819
1820                 /**********************************
1821                  *   I N  E P I L O G  O F  irn
1822                  **********************************/
1823
1824                 /* ensure each dying value is used by only one post remat */
1825                 pset_foreach(used, tmp) {
1826                         ir_node     *value = tmp;
1827                         op_t        *value_op = get_irn_link(value);
1828                         ir_node     *remat;
1829                         int          n_remats = 0;
1830
1831                         cst = ILP_UNDEF;
1832                         foreach_post_remat(irn, remat) {
1833                                 op_t  *remat_op = get_irn_link(remat);
1834
1835                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1836                                         ir_node   *remat_arg = get_irn_n(remat, n);
1837
1838                                         /* if value is used by this remat add it to constraint */
1839                                         if(remat_arg == value) {
1840                                                 if(n_remats == 0) {
1841                                                         /* sum remat2s <= 1 + n_remats*live_range */
1842                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1843                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1844                                                 }
1845
1846                                                 n_remats++;
1847                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1848                                                 break;
1849                                         }
1850                                 }
1851                         }
1852
1853             // value_op->attr.live_range.ilp != ILP_UNDEF
1854                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1855                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1856                         }
1857                 }
1858
1859         /* ensure at least one value dies at post remat */
1860         foreach_post_remat(irn, tmp) {
1861             op_t     *remat_op = get_irn_link(tmp);
1862             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
1863             ir_node  *remat_arg;
1864
1865             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1866                 remat_arg = get_irn_n(tmp, n);
1867
1868                 if(has_reg_class(si, remat_arg)) {
1869
1870                     /* does arg always die at this op? */
1871                     if(!pset_find_ptr(live, remat_arg))
1872                         goto skip_one_must_die;
1873
1874                     pset_insert_ptr(remat_args, remat_arg);
1875                 }
1876             }
1877
1878             /* remat + \sum live_range(remat_arg) <= |args| */
1879             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1880             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1881             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1882
1883             pset_foreach(remat_args, remat_arg) {
1884                 op_t  *arg_op = get_irn_link(remat_arg);
1885
1886                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1887             }
1888
1889 skip_one_must_die:
1890             del_pset(remat_args);
1891         }
1892
1893                 /* new live ranges for values from L\U defined by post remats */
1894                 pset_foreach(live, tmp) {
1895                         ir_node     *value = tmp;
1896                         op_t        *value_op = get_irn_link(value);
1897
1898                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1899                                 ilp_var_t    prev_lr = ILP_UNDEF;
1900                                 ir_node     *remat;
1901
1902                                 if(pset_find_ptr(remat_defs, value)) {
1903
1904                                         /* next_live_range <= prev_live_range + sum remat2s */
1905                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1906                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1907
1908                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1909                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1910
1911                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1912                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1913
1914                                         foreach_post_remat(irn, remat) {
1915                                                 op_t        *remat_op = get_irn_link(remat);
1916
1917                                                 /* if value is being rematerialized by this remat */
1918                                                 if(value == remat_op->attr.remat.remat->value) {
1919                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1920                                                 }
1921                                         }
1922
1923                                         value_op->attr.live_range.ilp = prev_lr;
1924                                         value_op->attr.live_range.op = irn;
1925                                 }
1926                         }
1927                 }
1928
1929                 /* requirements for post remats and start live ranges from L/U' for values dying here */
1930                 foreach_post_remat(irn, tmp) {
1931                         op_t        *remat_op = get_irn_link(tmp);
1932                         int          n;
1933
1934                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1935                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1936                                 op_t           *arg_op = get_irn_link(remat_arg);
1937
1938                                 if(!has_reg_class(si, remat_arg)) continue;
1939
1940                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1941                                 if(!pset_find_ptr(used, remat_arg)) {
1942                                         /* remat <= live_range(remat_arg) */
1943                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1944                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1945
1946                                         /* if value is becoming live through use by remat2 */
1947                                         if(!pset_find_ptr(live, remat_arg)) {
1948                                                 ilp_var_t     lr;
1949
1950                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1951                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1952
1953                                                 arg_op->attr.live_range.ilp = lr;
1954                                                 arg_op->attr.live_range.op = irn;
1955
1956                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1957
1958                                                 pset_insert_ptr(live, remat_arg);
1959                                                 add_to_spill_bb(si, bb, remat_arg);
1960                                         }
1961
1962                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1963                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1964                                 }
1965                         }
1966                 }
1967
1968                 d = pset_count(defs);
1969                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
1970
1971                 /* count how many regs irn needs for arguments */
1972                 u = set_count(args);
1973
1974
1975                 /* check the register pressure in the epilog */
1976                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1977                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1978                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
1979
1980                 /* add L\U' to check_post */
1981                 pset_foreach(live, tmp) {
1982                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1983                                 /* if a live value is not used by irn */
1984                                 tmp_op = get_irn_link(tmp);
1985                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1986                         }
1987                 }
1988
1989                 /***********************************************************
1990                  *  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
1991                  **********************************************************/
1992
1993
1994                 pset_foreach(used, tmp) {
1995                         ilp_var_t       prev_lr;
1996                         ilp_var_t       post_use;
1997                         int             p = 0;
1998                         spill_t        *spill;
1999                         ir_node        *arg = tmp;
2000                         op_t           *arg_op = get_irn_link(arg);
2001                         ir_node        *remat;
2002
2003                         spill = add_to_spill_bb(si, bb, arg);
2004
2005                         /* new live range for each used value */
2006                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2007                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2008
2009                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2010                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2011                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2012
2013                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2014
2015                         /* arg is live throughout epilog if the next live_range is in a register */
2016                         if(pset_find_ptr(live, arg)) {
2017                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2018
2019                                 /* post_use >= next_lr + remat */
2020                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2021                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2022                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2023                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2024
2025                         }
2026
2027                         /* if value is not an arg of op and not possibly defined by post remat
2028                          * then it may only die and not become live
2029                          */
2030                         if(!set_find_keyval(args, arg)) {
2031                                 /* post_use <= prev_lr */
2032                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2033                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2034                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2035                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2036
2037                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2038                                         /* next_lr <= prev_lr */
2039                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2040                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2041                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2042                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2043                                 }
2044                         }
2045
2046
2047                         /* forall post remat which use arg add a similar cst */
2048                         foreach_post_remat(irn, remat) {
2049                                 int      n;
2050
2051                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2052                                         ir_node    *remat_arg = get_irn_n(remat, n);
2053                                         op_t       *remat_op = get_irn_link(remat);
2054
2055                                         if(remat_arg == arg) {
2056                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2057
2058                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2059                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2060                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2061                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2062                                         }
2063                                 }
2064                         }
2065
2066 #ifdef WITH_MEMOPERANDS
2067                         for(n = get_irn_arity(irn)-1; n>=0; --n) {
2068                                 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2069                                         ilp_var_t       memoperand;
2070
2071                                         ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2072                                         memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_MEMOPERAND*execution_frequency(si, bb), 0.0);
2073                                         set_insert_memoperand(si->memoperands, irn, n, memoperand);
2074
2075                                         ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2076                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2077
2078                                         lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2079                                         lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2080 //                                      if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2081 //                                              lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2082                                 }
2083                         }
2084 #endif
2085
2086                         /* new live range begins for each used value */
2087                         arg_op->attr.live_range.ilp = prev_lr;
2088                         arg_op->attr.live_range.op = irn;
2089
2090                         /*if(!pset_find_ptr(live, arg)) {
2091                                 pset_insert_ptr(live, arg);
2092                                 add_to_spill_bb(si, bb, arg);
2093                         }*/
2094                         pset_insert_ptr(live, arg);
2095
2096                 }
2097
2098                 /* just to be sure */
2099                 check_post = ILP_UNDEF;
2100
2101
2102
2103
2104                 /******************
2105                  *   P R O L O G
2106                  ******************/
2107
2108                 /* check the register pressure in the prolog */
2109                 /* sum_{L\U} lr <= k - |U| */
2110                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2111                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2112
2113                 /* for the prolog remove defined values from the live set */
2114                 pset_foreach(defs, tmp) {
2115                         pset_remove_ptr(live, tmp);
2116                 }
2117
2118 #ifdef WITH_MEMOPERANDS
2119                 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2120                 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2121 #endif
2122
2123                 /***********************************************************
2124                  *  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
2125                  **********************************************************/
2126
2127
2128                 set_foreach(args, keyval) {
2129                         spill_t          *spill;
2130                         const ir_node    *arg = keyval->key;
2131                         int               i = PTR_TO_INT(keyval->val);
2132                         op_t             *arg_op = get_irn_link(arg);
2133                         ilp_cst_t         requirements;
2134 #ifdef WITH_MEMOPERANDS
2135                         int               n_memoperands;
2136 #endif
2137
2138                         spill = set_find_spill(spill_bb->ilp, arg);
2139                         assert(spill);
2140
2141                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2142                         op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 1.0);
2143
2144                         /* reload <= mem_out */
2145                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2146                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2147                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2148                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2149
2150                         /* requirement: arg must be in register for use */
2151                         /* reload + remat + live_range == 1 */
2152                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2153                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2154
2155                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2156                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2157                         foreach_pre_remat(si, irn, tmp) {
2158                                 op_t     *remat_op = get_irn_link(tmp);
2159                                 if(remat_op->attr.remat.remat->value == arg) {
2160                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2161                                 }
2162                         }
2163
2164 #ifdef WITH_MEMOPERANDS
2165                         n_memoperands = 0;
2166                         for(n = get_irn_arity(irn)-1; n>=0; --n) {
2167                                 if(get_irn_n(irn, n) == arg) {
2168                                         n_memoperands++;
2169                                 }
2170                         }
2171                         for(n = get_irn_arity(irn)-1; n>=0; --n) {
2172                                 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2173                                         memoperand_t  *memoperand;
2174                                         memoperand = set_find_memoperand(si->memoperands, irn, n);
2175
2176                                         /* memoperand <= mem_out */
2177                                         ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2178                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2179                                         lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2180                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2181
2182                                         /* the memoperand is only sufficient if it is used once by the op */
2183                                         if(n_memoperands == 1)
2184                                                 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2185
2186                                         lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2187
2188                                         /* we have one more free register if we use a memory operand */
2189                                         lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2190                                 }
2191                         }
2192 #endif
2193                 }
2194
2195                 /* iterate over L\U */
2196                 pset_foreach(live, tmp) {
2197                         if(!set_find_keyval(args, tmp)) {
2198                                 /* if a live value is not used by irn */
2199                                 tmp_op = get_irn_link(tmp);
2200                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2201                         }
2202                 }
2203
2204
2205                 /* requirements for remats */
2206                 /* start new live ranges for values used by remats */
2207                 foreach_pre_remat(si, irn, tmp) {
2208                         op_t        *remat_op = get_irn_link(tmp);
2209                         int          n;
2210
2211                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2212                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2213                                 op_t           *arg_op = get_irn_link(remat_arg);
2214
2215                                 if(!has_reg_class(si, remat_arg)) continue;
2216
2217                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2218                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2219                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2220
2221                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2222                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2223
2224                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2225                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2226                                         int    index = (int)keyval->val;
2227
2228                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2229                                 }
2230                         }
2231                 }
2232
2233
2234
2235
2236                 /*************************
2237                  *  D O N E  W I T H  O P
2238                  *************************/
2239
2240                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2241
2242                 pset_foreach(live, tmp) {
2243                         assert(has_reg_class(si, tmp));
2244                 }
2245
2246                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2247                         ir_node        *arg = get_irn_n(irn, n);
2248
2249                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2250                 }
2251
2252                 del_pset(remat_defs);
2253                 del_pset(used);
2254                 del_set(args);
2255                 del_pset(defs);
2256                 defs = pset_new_ptr_default();
2257         }
2258
2259
2260
2261         /***************************************
2262          *   B E G I N N I N G  O F  B L O C K
2263          ***************************************/
2264
2265
2266         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2267         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2268
2269         pset_foreach(live, irn) {
2270                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2271         }
2272
2273         /* construct mem_outs for all values */
2274
2275         set_foreach(spill_bb->ilp, spill) {
2276                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2277                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2278
2279                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2280                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2281
2282                 if(pset_find_ptr(live, spill->irn)) {
2283                         int default_spilled;
2284                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2285
2286                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2287                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2288                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2289                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2290
2291                         if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2292                                 int   n;
2293                                 op_t *op = get_irn_link(spill->irn);
2294
2295                                 /* do we have to copy a phi argument? */
2296                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2297                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2298
2299                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2300                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2301                                         double          freq=0.0;
2302                                         int             m;
2303                                         ilp_var_t       var;
2304
2305
2306                                         /* argument already done? */
2307                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2308
2309                                         /* get sum of execution frequencies of blocks with the same phi argument */
2310                                         for(m=n; m>=0; --m) {
2311                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2312
2313                                                 if(arg==arg2) {
2314                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2315                                                 }
2316                                         }
2317
2318                                         /* copies are not for free */
2319                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2320                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_STORE * freq, 1.0);
2321
2322                                         for(m=n; m>=0; --m) {
2323                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2324
2325                                                 if(arg==arg2) {
2326                                                         op->attr.live_range.args.copies[m] = var;
2327                                                 }
2328                                         }
2329
2330 #if 0
2331                                         /* copy <= mem_in */
2332                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2333                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2334                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2335                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2336 #endif
2337                                 }
2338                         }
2339                 }
2340         }
2341
2342
2343         /* L\U is empty at bb start */
2344         /* arg is live throughout epilog if it is reg_in into this block */
2345
2346         /* check the register pressure at the beginning of the block
2347          * including remats
2348          */
2349         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2350         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2351
2352         pset_foreach(live, irn) {
2353         ilp_cst_t  nospill;
2354
2355                 spill = set_find_spill(spill_bb->ilp, irn);
2356                 assert(spill);
2357
2358                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2359                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2360
2361                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2362
2363                 /* spill + mem_in <= 1 */
2364                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2365                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2366
2367                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2368                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2369
2370         }
2371         foreach_post_remat(bb, irn) {
2372                 op_t     *remat_op = get_irn_link(irn);
2373
2374                 DBG((si->dbg, LEVEL_4, "\t  next post remat: %+F\n", irn));
2375                 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2376
2377                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2378         }
2379
2380         /* forall post remats add requirements */
2381         foreach_post_remat(bb, tmp) {
2382                 int         n;
2383
2384                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2385                         ir_node    *remat_arg = get_irn_n(tmp, n);
2386                         op_t       *remat_op = get_irn_link(tmp);
2387
2388                         if(!has_reg_class(si, remat_arg)) continue;
2389
2390                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2391                         assert(spill);
2392
2393                         /* remat <= reg_in_argument */
2394                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2395                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2396                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2397                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2398                 }
2399         }
2400
2401         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2402         pset_foreach(live, irn) {
2403                 int          p = 0,
2404                                          n;
2405
2406                 spill = set_find_spill(spill_bb->ilp, irn);
2407                 assert(spill && spill->irn == irn);
2408
2409                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2410                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2411                                 ilp_cst_t       mem_in,
2412                                                                 reg_in;
2413                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2414                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2415                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2416                                 spill_t        *spill_p;
2417                                 op_t           *op = get_irn_link(irn);
2418
2419                                 /* although the phi is in the right regclass one or more of
2420                                  * its arguments can be in a different one or at least to
2421                                  * ignore
2422                                  */
2423                                 if(has_reg_class(si, phi_arg)) {
2424                                         /* mem_in < mem_out_arg + copy */
2425                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2426                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2427
2428                                         /* reg_in < reg_out_arg */
2429                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2430                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2431
2432                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2433                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2434
2435                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2436                                         assert(spill_p);
2437
2438                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2439                                         lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2440                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2441                                 }
2442                         }
2443                 } else {
2444                         /* else assure the value arrives on all paths in the same resource */
2445
2446                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2447                                 ilp_cst_t       mem_in,
2448                                                                 reg_in;
2449                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2450                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2451                                 spill_t        *spill_p;
2452
2453                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2454                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2455                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2456                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2457
2458                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2459                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2460
2461                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2462                                 assert(spill_p);
2463
2464                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2465                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2466                         }
2467                 }
2468         }
2469
2470         /* first live ranges from reg_ins */
2471         pset_foreach(live, irn) {
2472                 op_t      *op = get_irn_link(irn);
2473
2474                 spill = set_find_spill(spill_bb->ilp, irn);
2475                 assert(spill && spill->irn == irn);
2476
2477                 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2478                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2479                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2480                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2481
2482                 foreach_post_remat(bb, tmp) {
2483                         op_t     *remat_op = get_irn_link(tmp);
2484
2485                         if(remat_op->attr.remat.remat->value == irn) {
2486                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2487                         }
2488                 }
2489         }
2490
2491         /* walk forward now and compute constraints for placing spills */
2492         /* this must only be done for values that are not defined in this block */
2493         /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2494         pset_foreach(live, irn) {
2495                 /*
2496                  * if value is defined in this block we can anways place the spill directly after the def
2497                  *    -> no constraint necessary
2498                  */
2499                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2500
2501
2502                 spill = set_find_spill(spill_bb->ilp, irn);
2503                 assert(spill);
2504
2505                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2506                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2507
2508                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2509                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2510
2511                 if(!is_Phi(irn)) {
2512                         sched_foreach_op(bb, tmp) {
2513                                 op_t   *op = get_irn_link(tmp);
2514
2515                                 if(is_Phi(tmp)) continue;
2516                                 assert(!is_Proj(tmp));
2517
2518                                 if(op->is_remat) {
2519                                         const ir_node   *value = op->attr.remat.remat->value;
2520
2521                                         if(value == irn) {
2522                                                 /* only collect remats up to the first real use of a value */
2523                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2524                                         }
2525                                 } else {
2526                                         int   n;
2527
2528                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2529                                                 ir_node    *arg = get_irn_n(tmp, n);
2530
2531                                                 if(arg == irn) {
2532                                                         /* if a value is used stop collecting remats */
2533                             goto next_live;
2534                                                 }
2535                                         }
2536                                 }
2537                         }
2538                 }
2539 next_live: ;
2540         }
2541
2542         del_pset(live);
2543 }
2544
2545 typedef struct _irnlist_t {
2546         struct list_head   list;
2547         ir_node           *irn;
2548 } irnlist_t;
2549
2550 typedef struct _interference_t {
2551         struct list_head    blocklist;
2552         ir_node            *a;
2553         ir_node            *b;
2554 } interference_t;
2555
2556 static int
2557 cmp_interference(const void *a, const void *b, size_t size)
2558 {
2559         const interference_t *p = a;
2560         const interference_t *q = b;
2561
2562         return !(p->a == q->a && p->b == q->b);
2563 }
2564
2565 static interference_t *
2566 set_find_interference(set * set, ir_node * a, ir_node * b)
2567 {
2568         interference_t     query;
2569
2570         query.a = (a>b)?a:b;
2571         query.b = (a>b)?b:a;
2572
2573         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2574 }
2575
2576 static interference_t *
2577 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2578 {
2579         interference_t     query,
2580                                           *result;
2581         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2582
2583         list->irn = bb;
2584
2585         result = set_find_interference(set, a, b);
2586         if(result) {
2587
2588                 list_add(&list->list, &result->blocklist);
2589                 return result;
2590         }
2591
2592         query.a = (a>b)?a:b;
2593         query.b = (a>b)?b:a;
2594
2595         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2596
2597         INIT_LIST_HEAD(&result->blocklist);
2598         list_add(&list->list, &result->blocklist);
2599
2600         return result;
2601 }
2602
2603 static int
2604 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2605 {
2606         const ir_edge_t *edge;
2607
2608         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2609                 /* both values are live in, so they interfere */
2610                 return 1;
2611         }
2612
2613         /* ensure a dominates b */
2614         if(value_dominates(b,a)) {
2615                 const ir_node * t;
2616                 t = b;
2617                 b = a;
2618                 a = t;
2619         }
2620         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2621
2622
2623         /* the following code is stolen from bera.c */
2624         if(be_is_live_end(si->lv, bb, a))
2625                 return 1;
2626
2627         foreach_out_edge(a, edge) {
2628                 const ir_node *user = edge->src;
2629                 if(get_nodes_block(user) == bb
2630                                 && !is_Phi(user)
2631                                 && b != user
2632                                 && value_dominates(b, user))
2633                         return 1;
2634         }
2635
2636         return 0;
2637 }
2638
2639 /**
2640  * Walk all irg blocks and collect interfering values inside of phi classes
2641  */
2642 static void
2643 luke_interferencewalker(ir_node * bb, void * data)
2644 {
2645         spill_ilp_t    *si = (spill_ilp_t*)data;
2646         int             l1, l2;
2647
2648         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2649                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2650                 op_t           *a_op = get_irn_link(a);
2651
2652
2653                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2654                 if (has_reg_class(si, a) && get_phi_class(a)) {
2655                         if(a_op->is_remat)
2656                                 continue;
2657
2658                         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)) {
2659                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
2660                                 op_t           *b_op = get_irn_link(b);
2661
2662
2663                                 /* a and b are only interesting if they are in the same phi class */
2664                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2665                                         if(b_op->is_remat)
2666                                                 continue;
2667
2668                                         if(values_interfere_in_block(si, bb, a, b)) {
2669                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2670                                                 set_insert_interference(si, si->interferences, a, b, bb);
2671                                         }
2672                                 }
2673                         }
2674                 }
2675         }
2676 }
2677
2678 static unsigned int copy_path_id = 0;
2679
2680 static void
2681 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2682 {
2683         ilp_cst_t  cst;
2684         ilp_var_t  copy;
2685         char       buf[256];
2686         void      *ptr;
2687
2688         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2689         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2690
2691         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2692
2693         pset_foreach(copies, ptr) {
2694                 copy = PTR_TO_INT(ptr);
2695                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2696         }
2697 }
2698
2699 /**
2700  * @parameter copies   contains a path of copies which lead us to irn
2701  * @parameter visited  contains a set of nodes already visited on this path
2702  */
2703 static int
2704 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2705 {
2706         const ir_edge_t *edge;
2707         op_t            *op = get_irn_link(irn);
2708     pset            *visited_users = pset_new_ptr_default();
2709         int              paths = 0;
2710
2711         if(op->is_remat) return 0;
2712
2713         pset_insert_ptr(visited, irn);
2714
2715         if(is_Phi(irn)) {
2716                 int    n;
2717         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2718
2719                 /* visit all operands */
2720                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2721                         ir_node  *arg = get_irn_n(irn, n);
2722                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2723
2724                         if(!has_reg_class(si, arg)) continue;
2725             if(pset_find_ptr(visited_operands, arg)) continue;
2726             pset_insert_ptr(visited_operands, arg);
2727
2728                         if(arg == target) {
2729                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2730                                         del_pset(visited_operands);
2731                                         del_pset(visited_users);
2732                                         pset_remove_ptr(visited, irn);
2733                                         return paths;
2734                                 }
2735                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2736                                 write_copy_path_cst(si, copies, any_interfere);
2737                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2738                         } else if(!pset_find_ptr(visited, arg)) {
2739                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2740                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2741                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2742
2743                 /*if(paths > MAX_PATHS) {
2744                     if(pset_count(copies) == 0) {
2745                         ilp_cst_t  cst;
2746                         char       buf[256];
2747
2748                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2749                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2750                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2751                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2752                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2753
2754                         paths = 0;
2755                     } else {
2756                         del_pset(visited_operands);
2757                         del_pset(visited_users);
2758                         pset_remove_ptr(visited, irn);
2759                         return paths;
2760                     }
2761                 } else if(pset_count(copies) == 0) {
2762                                         paths = 0;
2763                                 }*/
2764                         }
2765                 }
2766
2767         del_pset(visited_operands);
2768         }
2769
2770         /* visit all uses which are phis */
2771         foreach_out_edge(irn, edge) {
2772                 ir_node  *user = edge->src;
2773                 int       pos  = edge->pos;
2774                 op_t     *op = get_irn_link(user);
2775                 ilp_var_t copy;
2776
2777                 if(!is_Phi(user)) continue;
2778                 if(!has_reg_class(si, user)) continue;
2779         if(pset_find_ptr(visited_users, user)) continue;
2780         pset_insert_ptr(visited_users, user);
2781
2782                 copy = op->attr.live_range.args.copies[pos];
2783
2784                 if(user == target) {
2785                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2786                                 del_pset(visited_users);
2787                                 pset_remove_ptr(visited, irn);
2788                                 return paths;
2789                         }
2790                         pset_insert(copies, INT_TO_PTR(copy), copy);
2791                         write_copy_path_cst(si, copies, any_interfere);
2792                         pset_remove(copies, INT_TO_PTR(copy), copy);
2793                 } else if(!pset_find_ptr(visited, user)) {
2794                         pset_insert(copies, INT_TO_PTR(copy), copy);
2795                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2796                         pset_remove(copies, INT_TO_PTR(copy), copy);
2797
2798             /*if(paths > MAX_PATHS) {
2799                 if(pset_count(copies) == 0) {
2800                     ilp_cst_t  cst;
2801                     char       buf[256];
2802
2803                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2804                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2805                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2806                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2807                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2808
2809                     paths = 0;
2810                 } else {
2811                     del_pset(visited_users);
2812                     pset_remove_ptr(visited, irn);
2813                     return paths;
2814                 }
2815             } else if(pset_count(copies) == 0) {
2816                                 paths = 0;
2817                         }*/
2818                 }
2819         }
2820
2821     del_pset(visited_users);
2822         pset_remove_ptr(visited, irn);
2823         return paths;
2824 }
2825
2826 static void
2827 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2828 {
2829         pset * copies = pset_new_ptr_default();
2830         pset * visited = pset_new_ptr_default();
2831
2832         find_copy_path(si, a, b, any_interfere, copies, visited);
2833
2834         del_pset(visited);
2835         del_pset(copies);
2836 }
2837
2838
2839 static void
2840 memcopyhandler(spill_ilp_t * si)
2841 {
2842         interference_t   *interference;
2843         char              buf[256];
2844         /* teste Speicherwerte auf Interferenz */
2845
2846         /* analyze phi classes */
2847         phi_class_compute(si->chordal_env->irg);
2848
2849         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2850         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2851
2852         /* now lets emit the ILP unequations for the crap */
2853         set_foreach(si->interferences, interference) {
2854                 irnlist_t      *irnlist;
2855                 ilp_var_t       interfere,
2856                                                 any_interfere;
2857                 ilp_cst_t       any_interfere_cst,
2858                                                 cst;
2859                 const ir_node  *a  = interference->a;
2860                 const ir_node  *b  = interference->b;
2861
2862                 /* any_interf <= \sum interf */
2863                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2864                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2865                 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2866
2867                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2868
2869                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2870                         const ir_node  *bb = irnlist->irn;
2871                         spill_bb_t     *spill_bb = get_irn_link(bb);
2872                         spill_t        *spilla,
2873                                                    *spillb;
2874                         char           buf[256];
2875
2876                         spilla = set_find_spill(spill_bb->ilp, a);
2877                         assert(spilla);
2878
2879                         spillb = set_find_spill(spill_bb->ilp, b);
2880                         assert(spillb);
2881
2882                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2883                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2884                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
2885                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
2886                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2887                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2888
2889                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2890                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2891
2892                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2893                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2894                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2895                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2896                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2897
2898                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2899                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2900
2901                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2902                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2903                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2904
2905                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2906                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2907
2908                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2909                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2910                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2911
2912
2913                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2914
2915                         /* any_interfere >= interf */
2916                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2917                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2918
2919                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2920                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2921                 }
2922
2923                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2924                 gen_copy_constraints(si,a,b,any_interfere);
2925         }
2926 }
2927
2928
2929 static INLINE int
2930 is_zero(double x)
2931 {
2932         return fabs(x) < 0.00001;
2933 }
2934
2935 #ifdef KEEPALIVE
2936 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2937 {
2938         spill_ilp_t *si = get_irg_link(current_ir_graph);
2939
2940         if(pset_find_ptr(si->all_possible_remats, n)) {
2941                 op_t   *op = (op_t*)get_irn_link(n);
2942                 assert(op && op->is_remat);
2943
2944                 if(!op->attr.remat.remat->inverse) {
2945                         if(op->attr.remat.pre) {
2946                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2947                         } else {
2948                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2949                         }
2950
2951                         return 1;
2952                 } else {
2953                         op_t   *op = (op_t*)get_irn_link(n);
2954                         assert(op && op->is_remat);
2955
2956                         if(op->attr.remat.pre) {
2957                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2958                         } else {
2959                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2960                         }
2961
2962                         return 1;
2963                 }
2964         }
2965
2966         return 0;
2967 }
2968
2969 static void
2970 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2971 {
2972         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2973         be_dump(irg, suffix, dump_ir_block_graph_sched);
2974         set_dump_node_vcgattr_hook(NULL);
2975 }
2976 #endif
2977
2978 /**
2979  * Edge hook to dump the schedule edges with annotated register pressure.
2980  */
2981 static int
2982 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2983 {
2984         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2985                 ir_node *prev = sched_prev(irn);
2986                 fprintf(F, "edge:{sourcename:\"");
2987                 PRINT_NODEID(irn);
2988                 fprintf(F, "\" targetname:\"");
2989                 PRINT_NODEID(prev);
2990                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2991                 fprintf(F, "\" color:magenta}\n");
2992         }
2993         return 1;
2994 }
2995
2996 static void
2997 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2998 {
2999         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3000
3001         dump_consts_local(0);
3002         set_dump_node_edge_hook(sched_pressure_edge_hook);
3003         dump_ir_block_graph(irg, suffix);
3004         set_dump_node_edge_hook(old_edge_hook);
3005 }
3006
3007 static void
3008 walker_pressure_annotator(ir_node * bb, void * data)
3009 {
3010         spill_ilp_t  *si = data;
3011         ir_node      *irn;
3012         int           n, i;
3013         pset         *live = pset_new_ptr_default();
3014         int           projs = 0;
3015
3016         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3017                 irn = be_lv_get_irn(si->lv, bb, i);
3018
3019                 if (has_reg_class(si, irn)) {
3020                         pset_insert_ptr(live, irn);
3021                 }
3022         }
3023
3024         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3025
3026         sched_foreach_reverse(bb, irn) {
3027                 if(is_Phi(irn)) {
3028                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3029                         continue;
3030                 }
3031
3032                 if(has_reg_class(si, irn)) {
3033                         pset_remove_ptr(live, irn);
3034                         if(is_Proj(irn)) ++projs;
3035                 }
3036
3037                 if(!is_Proj(irn)) projs = 0;
3038
3039                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3040                         ir_node    *arg = get_irn_n(irn, n);
3041
3042                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3043                 }
3044                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3045         }
3046
3047         del_pset(live);
3048 }
3049
3050 static void
3051 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3052 {
3053         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3054 }
3055
3056 #ifdef KEEPALIVE
3057 static void
3058 connect_all_remats_with_keep(spill_ilp_t * si)
3059 {
3060         ir_node   *irn;
3061         ir_node  **ins,
3062                          **pos;
3063         int        n_remats;
3064
3065
3066         n_remats = pset_count(si->all_possible_remats);
3067         if(n_remats) {
3068                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3069
3070                 pos = ins;
3071                 pset_foreach(si->all_possible_remats, irn) {
3072                         *pos = irn;
3073                         ++pos;
3074                 }
3075
3076                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3077
3078                 obstack_free(si->obst, ins);
3079         }
3080 }
3081 #endif
3082
3083 static void
3084 connect_all_spills_with_keep(spill_ilp_t * si)
3085 {
3086         ir_node   *irn;
3087         ir_node  **ins,
3088                          **pos;
3089         int        n_spills;
3090         ir_node   *keep;
3091
3092
3093         n_spills = pset_count(si->spills);
3094         if(n_spills) {
3095                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3096
3097                 pos = ins;
3098                 pset_foreach(si->spills, irn) {
3099                         *pos = irn;
3100                         ++pos;
3101                 }
3102
3103                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3104
3105                 obstack_free(si->obst, ins);
3106         }
3107 }
3108
3109 /** insert a spill at an arbitrary position */
3110 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3111 {
3112         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
3113         ir_graph *irg   = get_irn_irg(bl);
3114         ir_node *frame  = get_irg_frame(irg);
3115         ir_node *spill;
3116         ir_node *next;
3117
3118         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3119         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3120
3121         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3122
3123         /*
3124          * search the right insertion point. a spill of a phi cannot be put
3125          * directly after the phi, if there are some phis behind the one which
3126          * is spilled. Also, a spill of a Proj must be after all Projs of the
3127          * same tuple node.
3128          *
3129          * Here's one special case:
3130          * If the spill is in the start block, the spill must be after the frame
3131          * pointer is set up. This is done by setting insert to the end of the block
3132          * which is its default initialization (see above).
3133          */
3134
3135         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3136                 insert = frame;
3137
3138         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3139                 insert = next;
3140
3141         sched_add_after(insert, spill);
3142         return spill;
3143 }
3144
3145 static void
3146 delete_remat(spill_ilp_t * si, ir_node * remat) {
3147         int       n;
3148         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3149
3150         sched_remove(remat);
3151
3152         /* kill links to operands */
3153         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3154                 set_irn_n(remat, n, bad);
3155         }
3156 }
3157
3158 static void
3159 clean_remat_info(spill_ilp_t * si)
3160 {
3161         int            n;
3162         remat_t       *remat;
3163         remat_info_t  *remat_info;
3164         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
3165
3166         set_foreach(si->remat_info, remat_info) {
3167                 if(!remat_info->remats) continue;
3168
3169                 pset_foreach(remat_info->remats, remat)
3170                 {
3171                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3172                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3173                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3174                         }
3175
3176                         if(get_irn_n_edges(remat->op) == 0) {
3177                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3178                                         set_irn_n((ir_node*)remat->op, n, bad);
3179                                 }
3180                         }
3181                 }
3182
3183                 if(remat_info->remats) del_pset(remat_info->remats);
3184                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3185         }
3186 }
3187
3188 static void
3189 delete_unnecessary_remats(spill_ilp_t * si)
3190 {
3191 #ifdef KEEPALIVE
3192         int       n;
3193         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3194
3195         if(si->keep) {
3196                 ir_node   *end = get_irg_end(si->chordal_env->irg);
3197                 ir_node  **keeps;
3198
3199                 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3200                         ir_node        *keep_arg = get_irn_n(si->keep, n);
3201                         op_t           *arg_op = get_irn_link(keep_arg);
3202                         lpp_name_t     *name;
3203
3204                         assert(arg_op->is_remat);
3205
3206                         name = si->lpp->vars[arg_op->attr.remat.ilp];
3207
3208                         if(is_zero(name->value)) {
3209                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3210                                 /* TODO check whether reload is preferred over remat (could be bug) */
3211                                 delete_remat(si, keep_arg);
3212                         } else {
3213                                 if(!arg_op->attr.remat.remat->inverse) {
3214                                         if(arg_op->attr.remat.pre) {
3215                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3216                                         } else {
3217                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3218                                         }
3219                                 } else {
3220                                         if(arg_op->attr.remat.pre) {
3221                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3222                                         } else {
3223                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3224                                         }
3225                                 }
3226                         }
3227
3228                         set_irn_n(si->keep, n, bad);
3229                 }
3230 #if 0
3231                 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3232                         ir_node        *end_arg = get_End_keepalive(end, i);
3233
3234                         if(end_arg != si->keep) {
3235                                 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3236                         }
3237                 }
3238                 keeps = obstack_finish(si->obst);
3239                 set_End_keepalives(end, n-1, keeps);
3240                 obstack_free(si->obst, keeps);
3241 #endif
3242         } else {
3243                 DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3244         }
3245 #else
3246         ir_node  *remat;
3247
3248         pset_foreach(si->all_possible_remats, remat) {
3249                 op_t           *remat_op = get_irn_link(remat);
3250                 lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3251
3252                 if(is_zero(name->value)) {
3253                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3254                         /* TODO check whether reload is preferred over remat (could be bug) */
3255                         delete_remat(si, remat);
3256                 } else {
3257                         if(!remat_op->attr.remat.remat->inverse) {
3258                                 if(remat_op->attr.remat.pre) {
3259                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3260                                 } else {
3261                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3262                                 }
3263                         } else {
3264                                 if(remat_op->attr.remat.pre) {
3265                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3266                                 } else {
3267                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3268                                 }
3269                         }
3270                 }
3271         }
3272 #endif
3273 }
3274
3275 static pset *
3276 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3277 {
3278         pset     *spills = pset_new_ptr_default();
3279
3280         const ir_node  *next;
3281         defs_t         *defs;
3282
3283         defs = set_find_def(si->values, value);
3284
3285         if(defs && defs->spills) {
3286                 for(next = defs->spills; next; next = get_irn_link(next)) {
3287                         pset_insert_ptr(spills, next);
3288                 }
3289         }
3290
3291         return spills;
3292 }
3293
3294 /**
3295  * @param before   The node after which the spill will be placed in the schedule
3296  */
3297 /* TODO set context properly */
3298 static ir_node *
3299 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3300 {
3301         defs_t   *defs;
3302         ir_node  *spill;
3303         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3304
3305         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3306
3307         spill = be_spill2(arch_env, irn, before);
3308
3309         defs = set_insert_def(si->values, value);
3310         assert(defs);
3311
3312         /* enter into the linked list */
3313         set_irn_link(spill, defs->spills);
3314         defs->spills = spill;
3315
3316 #ifdef KEEPALIVE_SPILLS
3317         pset_insert_ptr(si->spills, spill);
3318 #endif
3319
3320         return spill;
3321 }
3322
3323 /**
3324  * @param before   The Phi node which has to be spilled
3325  */
3326 static ir_node *
3327 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3328 {
3329         ir_node   *mem_phi;
3330         ir_node  **ins;
3331         defs_t    *defs;
3332         int        n;
3333
3334         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3335
3336         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3337                 ins[n] = si->m_unknown;
3338         }
3339
3340         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3341
3342         defs = set_insert_def(si->values, phi);
3343         assert(defs);
3344
3345         /* enter into the linked list */
3346         set_irn_link(mem_phi, defs->spills);
3347         defs->spills = mem_phi;
3348
3349         sched_add_after(phi, mem_phi);
3350
3351 #ifdef KEEPALIVE_SPILLS
3352         pset_insert_ptr(si->spills, mem_phi);
3353 #endif
3354
3355
3356         return mem_phi;
3357 }
3358
3359 /**
3360  * Add remat to list of defs, destroys link field!
3361  */
3362 static void
3363 insert_remat(spill_ilp_t * si, ir_node * remat)
3364 {
3365         defs_t   *defs;
3366         op_t     *remat_op = get_irn_link(remat);
3367
3368         assert(remat_op->is_remat);
3369
3370         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3371         assert(defs);
3372
3373         /* enter into the linked list */
3374         set_irn_link(remat, defs->remats);
3375         defs->remats = remat;
3376 }
3377
3378
3379 /**
3380  * Add reload before operation and add to list of defs
3381  */
3382 static ir_node *
3383 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3384 {
3385         defs_t   *defs;
3386         ir_node  *reload,
3387                          *spill;
3388         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3389
3390         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3391
3392         defs = set_find_def(si->values, value);
3393
3394         spill = defs->spills;
3395         assert(spill && "no spill placed before reload");
3396
3397         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3398
3399         /* enter into the linked list */
3400         set_irn_link(reload, defs->remats);
3401         defs->remats = reload;
3402
3403         return reload;
3404 }
3405
3406 #ifdef WITH_MEMOPERANDS
3407 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3408 {
3409         defs_t           *defs;
3410         ir_node          *reload;
3411         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3412         ir_node          *spill;
3413         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3414
3415         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3416
3417         defs = set_find_def(si->values, value);
3418
3419         spill = defs->spills;
3420         assert(spill && "no spill placed before reload");
3421
3422         reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3423
3424         arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3425         sched_remove(reload);
3426 }
3427
3428 void insert_memoperands(spill_ilp_t * si)
3429 {
3430         memoperand_t   *memoperand;
3431         lpp_name_t     *name;
3432
3433         set_foreach(si->memoperands, memoperand) {
3434                 name = si->lpp->vars[memoperand->ilp];
3435                 if(!is_zero(name->value)) {
3436                         perform_memory_operand(si, memoperand);
3437                 }
3438         }
3439 }
3440 #endif
3441
3442 static void
3443 walker_spill_placer(ir_node * bb, void * data) {
3444         spill_ilp_t   *si = (spill_ilp_t*)data;
3445         ir_node       *irn;
3446         spill_bb_t    *spill_bb = get_irn_link(bb);
3447         pset          *spills_to_do = pset_new_ptr_default();
3448         spill_t       *spill;
3449
3450         set_foreach(spill_bb->ilp, spill) {
3451                 lpp_name_t    *name;
3452
3453                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3454                         name = si->lpp->vars[spill->mem_in];
3455                         if(!is_zero(name->value)) {
3456                                 ir_node   *mem_phi;
3457
3458                                 mem_phi = insert_mem_phi(si, spill->irn);
3459
3460                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3461                         }
3462                 }
3463
3464                 name = si->lpp->vars[spill->spill];
3465                 if(!is_zero(name->value)) {
3466                         /* place spill directly after definition */
3467                         if(get_nodes_block(spill->irn) == bb) {
3468                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3469                                 continue;
3470                         }
3471
3472                         /* place spill at bb start */
3473                         if(spill->reg_in > 0) {
3474                                 name = si->lpp->vars[spill->reg_in];
3475                                 if(!is_zero(name->value)) {
3476                                         insert_spill(si, spill->irn, spill->irn, bb);
3477                                         continue;
3478                                 }
3479                         }
3480                         /* place spill after a remat */
3481                         pset_insert_ptr(spills_to_do, spill->irn);
3482                 }
3483         }
3484         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3485
3486
3487         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3488                 op_t     *op = get_irn_link(irn);
3489
3490                 if(be_is_Spill(irn)) continue;
3491
3492                 if(op->is_remat) {
3493                         /* TODO fix this if we want to support remats with more than two nodes */
3494                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3495                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3496
3497                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3498                         }
3499                 } else {
3500                         if(pset_find_ptr(spills_to_do, irn)) {
3501                                 pset_remove_ptr(spills_to_do, irn);
3502
3503                                 insert_spill(si, irn, irn, irn);
3504                         }
3505                 }
3506
3507         }
3508
3509         assert(pset_count(spills_to_do) == 0);
3510
3511         /* afterwards free data in block */
3512         del_pset(spills_to_do);
3513 }
3514
3515 static ir_node *
3516 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3517 {
3518         ir_node          *insert_pos = bb;
3519         ir_node          *spill;
3520         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3521
3522         /* find last definition of arg value in block */
3523         ir_node  *next;
3524         defs_t   *defs;
3525         int       last = 0;
3526
3527         defs = set_find_def(si->values, value);
3528
3529         if(defs && defs->remats) {
3530                 for(next = defs->remats; next; next = get_irn_link(next)) {
3531                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3532                                 last = sched_get_time_step(next);
3533                                 insert_pos = next;
3534                         }
3535                 }
3536         }
3537
3538         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3539                 last = sched_get_time_step(value);
3540                 insert_pos = value;
3541         }
3542
3543         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3544
3545         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3546
3547         return spill;
3548 }
3549
3550 static void
3551 phim_fixer(spill_ilp_t *si) {
3552         defs_t  *defs;
3553
3554         set_foreach(si->values, defs) {
3555                 const ir_node  *phi = defs->value;
3556                 op_t           *op = get_irn_link(phi);
3557                 ir_node        *phi_m = NULL;
3558                 ir_node        *next = defs->spills;
3559                 int             n;
3560
3561                 if(!is_Phi(phi)) continue;
3562
3563                 while(next) {
3564                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3565                                 phi_m = next;
3566                                 break;
3567                         } else {
3568                                 next = get_irn_link(next);
3569                         }
3570                 }
3571                 if(!phi_m) continue;
3572
3573                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3574                         ir_node        *value = get_irn_n(phi, n);
3575                         defs_t         *val_defs = set_find_def(si->values, value);
3576
3577                         /* a spill of this value */
3578                         ir_node      *spill;
3579
3580
3581 #ifndef NO_MEMCOPIES
3582                         ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3583                         lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3584
3585                         if(!is_zero(name->value)) {
3586                                 spill = insert_mem_copy(si, pred, value);
3587                         } else {
3588                                 spill = val_defs->spills;
3589                         }
3590 #else
3591                         spill = val_defs->spills;
3592 #endif
3593                         assert(spill && "no spill placed before PhiM");
3594                         set_irn_n(phi_m, n, spill);
3595                 }
3596         }
3597 }
3598
3599 static void
3600 walker_reload_placer(ir_node * bb, void * data) {
3601         spill_ilp_t   *si = (spill_ilp_t*)data;
3602         ir_node       *irn;
3603         spill_bb_t    *spill_bb = get_irn_link(bb);
3604
3605         /* reloads at end of block */
3606         if(spill_bb->reloads) {
3607                 keyval_t    *keyval;
3608
3609                 set_foreach(spill_bb->reloads, keyval) {
3610                         ir_node        *irn = (ir_node*)keyval->key;
3611                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3612                         lpp_name_t     *name;
3613
3614                         name = si->lpp->vars[reload];
3615                         if(!is_zero(name->value)) {
3616                                 ir_node    *reload;
3617                                 ir_node    *insert_pos = bb;
3618                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3619                                 op_t       *prev_op = get_irn_link(prev);
3620
3621                                 while(be_is_Spill(prev)) {
3622                                         prev = sched_prev(prev);
3623                                 }
3624
3625                                 prev_op = get_irn_link(prev);
3626
3627                                 /* insert reload before pre-remats */
3628                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3629                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3630                                         insert_pos = prev;
3631
3632                                         do {
3633                                                 prev = sched_prev(prev);
3634                                         } while(be_is_Spill(prev));
3635
3636                                         prev_op = get_irn_link(prev);
3637
3638                                 }
3639
3640                                 reload = insert_reload(si, irn, insert_pos);
3641
3642 #ifdef KEEPALIVE_RELOADS
3643                                 pset_insert_ptr(si->spills, reload);
3644 #endif
3645                         }
3646                 }
3647         }
3648
3649         /* walk and insert more reloads and collect remats */
3650         sched_foreach_reverse(bb, irn) {
3651                 op_t     *op = get_irn_link(irn);
3652
3653                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3654                 if(is_Phi(irn)) break;
3655
3656                 if(op->is_remat) {
3657                         if(get_irn_mode(irn) != mode_T) {
3658                                 insert_remat(si, irn);
3659                         }
3660                 } else {
3661                         int    n;
3662
3663                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3664                                 ir_node    *arg = get_irn_n(irn, n);
3665
3666                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3667                                         lpp_name_t    *name;
3668
3669                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3670                                         if(!is_zero(name->value)) {
3671                                                 ir_node    *reload;
3672                                                 ir_node    *insert_pos = irn;
3673                                                 ir_node    *prev = sched_prev(insert_pos);
3674                                                 op_t       *prev_op;
3675
3676                                                 while(be_is_Spill(prev)) {
3677                                                         prev = sched_prev(prev);
3678                                                 }
3679
3680                                                 prev_op = get_irn_link(prev);
3681
3682                                                 /* insert reload before pre-remats */
3683                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3684                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3685                                                         insert_pos = prev;
3686
3687                                                         do {
3688                                                                 prev = sched_prev(prev);
3689                                                         } while(be_is_Spill(prev));
3690
3691                                                         prev_op = get_irn_link(prev);
3692
3693                                                 }
3694
3695                                                 reload = insert_reload(si, arg, insert_pos);
3696
3697                                                 set_irn_n(irn, n, reload);
3698
3699 #ifdef KEEPALIVE_RELOADS
3700                                                 pset_insert_ptr(si->spills, reload);
3701 #endif
3702                                         }
3703                                 }
3704                         }
3705                 }
3706         }
3707
3708         del_set(spill_bb->ilp);
3709         if(spill_bb->reloads) del_set(spill_bb->reloads);
3710 }
3711
3712 static void
3713 walker_collect_used(ir_node * irn, void * data)
3714 {
3715         lc_bitset_t   *used = data;
3716
3717         lc_bitset_set(used, get_irn_idx(irn));
3718 }
3719
3720 struct kill_helper {
3721         lc_bitset_t  *used;
3722         spill_ilp_t  *si;
3723 };
3724
3725 static void
3726 walker_kill_unused(ir_node * bb, void * data)
3727 {
3728         struct kill_helper *kh = data;
3729         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
3730         ir_node            *irn;
3731
3732
3733         for(irn=sched_first(bb); !sched_is_end(irn);) {
3734                 ir_node     *next = sched_next(irn);
3735                 int          n;
3736
3737                 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3738                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3739                                 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)));
3740 #if 0
3741                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3742 #endif
3743                         }
3744
3745                         sched_remove(irn);
3746
3747                         set_nodes_block(irn, bad);
3748                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3749                                 set_irn_n(irn, n, bad);
3750                         }
3751                 }
3752                 irn = next;
3753         }
3754 }
3755
3756 static void
3757 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3758 {
3759         struct kill_helper kh;
3760
3761         kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3762         kh.si = si;
3763
3764         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3765         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3766
3767         lc_bitset_free(kh.used);
3768 }
3769
3770 void
3771 print_irn_pset(pset * p)
3772 {
3773         ir_node   *irn;
3774
3775         pset_foreach(p, irn) {
3776                 ir_printf("%+F\n", irn);
3777         }
3778 }
3779
3780 void
3781 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3782 {
3783     FILE           *f = fopen(file, "w");
3784     ir_node        *irn;
3785     interference_t *interference;
3786
3787     pset_break(phiclass);
3788     set_break(si->interferences);
3789
3790     ir_fprintf(f, "digraph phiclass {\n");
3791
3792     pset_foreach(phiclass, irn) {
3793         if(is_Phi(irn))
3794             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
3795     }
3796
3797     pset_foreach(phiclass, irn) {
3798         int n;
3799
3800         if(!is_Phi(irn)) continue;
3801
3802         for(n=get_irn_arity(irn)-1; n>=0; --n) {
3803             ir_node  *arg = get_irn_n(irn, n);
3804
3805             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
3806         }
3807     }
3808
3809     set_foreach(si->interferences, interference) {
3810         const ir_node  *a  = interference->a;
3811         const ir_node  *b  = interference->b;
3812         if(get_phi_class(a) == phiclass) {
3813             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3814         }
3815     }
3816
3817     ir_fprintf(f, "}");
3818     fclose(f);
3819 }
3820
3821 static void
3822 rewire_uses(spill_ilp_t * si)
3823 {
3824         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3825         defs_t               *defs;
3826         pset                 *ignore = pset_new_ptr(1);
3827
3828         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3829
3830         /* then fix uses of spills */
3831         set_foreach(si->values, defs) {
3832                 pset           *reloads;
3833                 pset           *spills;
3834                 const ir_node  *next = defs->remats;
3835                 int remats = 0;
3836
3837                 reloads = pset_new_ptr_default();
3838
3839                 while(next) {
3840                         if(be_is_Reload(next)) {
3841                                 pset_insert_ptr(reloads, next);
3842                         } else {
3843                                 ++remats;
3844                         }
3845                         next = get_irn_link(next);
3846                 }
3847
3848                 spills = get_spills_for_value(si, defs->value);
3849                 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));
3850                 if(pset_count(spills) > 1) {
3851                         //assert(pset_count(reloads) > 0);
3852                         //                              print_irn_pset(spills);
3853                         //                              print_irn_pset(reloads);
3854
3855                         be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3856                 }
3857
3858                 del_pset(reloads);
3859                 del_pset(spills);
3860         }
3861
3862         /* first fix uses of remats and reloads */
3863         set_foreach(si->values, defs) {
3864                 pset           *nodes;
3865                 const ir_node  *next = defs->remats;
3866
3867                 if(next) {
3868                         nodes = pset_new_ptr_default();
3869                         pset_insert_ptr(nodes, defs->value);
3870
3871                         while(next) {
3872                                 pset_insert_ptr(nodes, next);
3873                                 next = get_irn_link(next);
3874                         }
3875
3876                         if(pset_count(nodes) > 1) {
3877                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3878                                 be_ssa_constr_set(dfi, si->lv, nodes);
3879                         }
3880
3881                         del_pset(nodes);
3882                 }
3883         }
3884
3885 //      remove_unused_defs(si);
3886
3887         be_free_dominance_frontiers(dfi);
3888 }
3889
3890
3891 static void
3892 writeback_results(spill_ilp_t * si)
3893 {
3894         /* walk through the graph and collect all spills, reloads and remats for a value */
3895
3896         si->values = new_set(cmp_defs, 4096);
3897
3898         DBG((si->dbg, LEVEL_1, "Applying results\n"));
3899         delete_unnecessary_remats(si);
3900         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3901         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3902         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3903 #ifdef WITH_MEMOPERANDS
3904         insert_memoperands(si);
3905 #endif
3906         phim_fixer(si);
3907
3908         /* clean the remat info! there are still back-edges leading there! */
3909         clean_remat_info(si);
3910
3911         rewire_uses(si);
3912
3913         connect_all_spills_with_keep(si);
3914
3915         del_set(si->values);
3916 }
3917
3918 static int
3919 get_n_regs(spill_ilp_t * si)
3920 {
3921         int     arch_n_regs = arch_register_class_n_regs(si->cls);
3922         int     free = 0;
3923         int     i;
3924
3925         for(i=0; i<arch_n_regs; i++) {
3926                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3927                         free++;
3928                 }
3929         }
3930
3931         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3932         return free;
3933 }
3934
3935 static void
3936 walker_reload_mover(ir_node * bb, void * data)
3937 {
3938         spill_ilp_t   *si = data;
3939         ir_node           *tmp;
3940
3941         sched_foreach(bb, tmp) {
3942                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3943                         ir_node       *reload = tmp;
3944                         ir_node       *irn = tmp;
3945
3946                         /* move reload upwards */
3947
3948                         int pressure = (int)get_irn_link(reload);
3949                         if(pressure < si->n_regs) {
3950                                 irn = sched_prev(reload);
3951                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3952                                 sched_remove(reload);
3953                                 pressure = (int)get_irn_link(irn);
3954
3955                                 while(pressure < si->n_regs) {
3956                                         if( sched_is_end(irn) ||
3957                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3958                                            /* do not move reload before its spill */
3959                                            (irn == be_get_Reload_mem(reload)) ||
3960                                            /* do not move before phi */
3961                                            is_Phi(irn)) break;
3962
3963                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
3964                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3965                                         irn = sched_prev(irn);
3966
3967                                         pressure = (int)get_irn_link(irn);
3968                                 }
3969
3970                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3971                                 sched_put_after(irn, reload);
3972                         }
3973                 }
3974         }
3975 }
3976
3977 static void
3978 move_reloads_upward(spill_ilp_t * si)
3979 {
3980         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3981 }
3982
3983
3984 /**
3985  * Walk all irg blocks and check for interfering spills inside of phi classes
3986  */
3987 static void
3988 luke_meminterferencechecker(ir_node * bb, void * data)
3989 {
3990         spill_ilp_t    *si = (spill_ilp_t*)data;
3991         int             l1, l2;
3992
3993         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
3994                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
3995
3996                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3997
3998                 /* a is only interesting if it is in my register class and if it is inside a phi class */
3999                 if (has_reg_class(si, a) && get_phi_class(a)) {
4000                         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)) {
4001                                 ir_node        *b = be_lv_get_irn(si->lv, bb, l2);
4002
4003                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4004
4005                                 /* a and b are only interesting if they are in the same phi class */
4006                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4007                                         if(values_interfere_in_block(si, bb, a, b)) {
4008                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4009                                         }
4010                                 }
4011                         }
4012                 }
4013         }
4014 }
4015
4016 static void
4017 verify_phiclasses(spill_ilp_t * si)
4018 {
4019         /* analyze phi classes */
4020         phi_class_compute(si->chordal_env->irg);
4021
4022         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4023         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4024 }
4025
4026 static void
4027 walker_spillslotassigner(ir_node * irn, void * data)
4028 {
4029         void                   *cls;
4030
4031         if(!be_is_Spill(irn)) return;
4032
4033         /* set spill context to phi class if it has one ;) */
4034         (void) cls;
4035 #if 0
4036         // Matze: not needed anymore
4037         cls = get_phi_class(irn);
4038         if(cls)
4039                 be_set_Spill_context(irn, cls);
4040         else
4041                 be_set_Spill_context(irn, irn);
4042 #endif
4043 }
4044
4045
4046 static void
4047 assign_spillslots(spill_ilp_t * si)
4048 {
4049         DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4050         irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4051 }
4052
4053 void
4054 be_spill_remat(const be_chordal_env_t * chordal_env)
4055 {
4056         char            buf[256];
4057         char            problem_name[256];
4058         char            dump_suffix[256];
4059         char            dump_suffix2[256];
4060         struct obstack  obst;
4061         spill_ilp_t     si;
4062
4063         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4064         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4065         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4066
4067         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4068         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4069
4070 #ifdef VERIFY_DOMINANCE
4071     be_check_dominance(chordal_env->irg);
4072 #endif
4073
4074         obstack_init(&obst);
4075         si.chordal_env = chordal_env;
4076         si.obst = &obst;
4077         si.cls = chordal_env->cls;
4078         si.lpp = new_lpp(problem_name, lpp_minimize);
4079         si.remat_info = new_set(cmp_remat_info, 4096);
4080         si.interferences = new_set(cmp_interference, 32);
4081 #ifdef WITH_MEMOPERANDS
4082         si.memoperands = new_set(cmp_memoperands, 128);
4083 #endif
4084         si.all_possible_remats = pset_new_ptr_default();
4085         si.spills = pset_new_ptr_default();
4086         si.inverse_ops = pset_new_ptr_default();
4087         si.lv = chordal_env->lv;
4088 #ifdef KEEPALIVE
4089         si.keep = NULL;
4090 #endif
4091         si.n_regs = get_n_regs(&si);
4092
4093         set_irg_link(chordal_env->irg, &si);
4094         compute_doms(chordal_env->irg);
4095
4096         /* compute phi classes */
4097 //      phi_class_compute(chordal_env->irg);
4098
4099         be_analyze_regpressure(chordal_env, "-pre");
4100
4101 #ifdef COLLECT_REMATS
4102         /* collect remats */
4103         DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4104         irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4105 #endif
4106
4107         /* insert possible remats */
4108         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4109         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4110         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4111
4112 #ifdef KEEPALIVE
4113         DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4114         connect_all_remats_with_keep(&si);
4115         /* dump graph with inserted remats */
4116         dump_graph_with_remats(chordal_env->irg, dump_suffix);
4117 #endif
4118
4119         /* insert copies for phi arguments not in my regclass */
4120         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4121
4122         /* recompute liveness */
4123         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4124         be_liveness_recompute(si.lv);
4125
4126         /* build the ILP */
4127
4128         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4129         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4130         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4131
4132         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4133         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4134
4135 #ifndef NO_MEMCOPIES
4136         DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4137         memcopyhandler(&si);
4138 #endif
4139
4140         if(dump_flags & DUMP_PROBLEM) {
4141                 FILE           *f;
4142                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4143                 if ((f = fopen(buf, "wt")) != NULL) {
4144                         lpp_dump_plain(si.lpp, f);
4145                         fclose(f);
4146                 }
4147         }
4148
4149         if(dump_flags & DUMP_MPS) {
4150                 FILE *f;
4151
4152                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4153                 if((f = fopen(buf, "wt")) != NULL) {
4154                         mps_write_mps(si.lpp, s_mps_fixed, f);
4155                         fclose(f);
4156                 }
4157
4158                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4159                 if((f = fopen(buf, "wt")) != NULL) {
4160                         mps_write_mst(si.lpp, s_mps_fixed, f);
4161                         fclose(f);
4162                 }
4163         }
4164
4165         lpp_check_startvals(si.lpp);
4166
4167 #ifdef SOLVE
4168         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4169 #ifdef ILP_TIMEOUT
4170         lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
4171 #endif
4172
4173         if(activate_log)
4174                 lpp_set_log(si.lpp, stdout);
4175
4176 #ifdef SOLVE_LOCAL
4177         lpp_solve_cplex(si.lpp);
4178 #else
4179         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4180 #endif
4181         assert(lpp_is_sol_valid(si.lpp)
4182                && "solution of ILP must be valid");
4183
4184         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
4185
4186         if(dump_flags & DUMP_SOLUTION) {
4187                 FILE           *f;
4188                 char            buf[256];
4189
4190                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4191                 if ((f = fopen(buf, "wt")) != NULL) {
4192                         int             i;
4193                         for (i = 0; i < si.lpp->var_next; ++i) {
4194                                 lpp_name_t     *name = si.lpp->vars[i];
4195                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4196                         }
4197                         fclose(f);
4198                 }
4199         }
4200
4201         writeback_results(&si);
4202
4203 #endif                          /* SOLVE */
4204
4205         kill_all_unused_values_in_schedule(&si);
4206
4207 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
4208         be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4209 #endif
4210
4211         // move reloads upwards
4212         be_liveness_recompute(si.lv);
4213         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4214         move_reloads_upward(&si);
4215
4216 #ifndef NO_MEMCOPIES
4217         verify_phiclasses(&si);
4218         assign_spillslots(&si);
4219 #endif
4220
4221         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4222
4223         dump_pressure_graph(&si, dump_suffix2);
4224
4225         be_analyze_regpressure(chordal_env, "-post");
4226
4227 #ifdef VERIFY_DOMINANCE
4228         be_check_dominance(chordal_env->irg);
4229 #endif
4230
4231         free_dom(chordal_env->irg);
4232         del_set(si.interferences);
4233         del_pset(si.inverse_ops);
4234         del_pset(si.all_possible_remats);
4235 #ifdef WITH_MEMOPERANDS
4236         del_set(si.memoperands);
4237 #endif
4238         del_pset(si.spills);
4239         free_lpp(si.lpp);
4240         obstack_free(&obst, NULL);
4241         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4242 }
4243
4244 #else                           /* WITH_ILP */
4245
4246 static void
4247 only_that_you_can_compile_without_WITH_ILP_defined(void)
4248 {
4249 }
4250
4251 #endif                          /* WITH_ILP */