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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       ILP based spilling & rematerialization
  23  * @author      Adam M. Szalkowski
  24  * @date        06.04.2006
  25  * @version     $Id$
  26  */
  27 #include "config.h"
  28
  29 #ifdef WITH_ILP
  30
  31 #include <math.h>
  32
  33 #include "array_t.h"
  34 #include "hashptr.h"
  35 #include "debug.h"
  36 #include "obst.h"
  37 #include "set.h"
  38 #include "list.h"
  39 #include "pmap.h"
  40
  41 #include "irprintf.h"
  42 #include "irgwalk.h"
  43 #include "irdump_t.h"
  44 #include "irnode_t.h"
  45 #include "ircons_t.h"
  46 #include "irloop_t.h"
  47 #include "irnodeset.h"
  48 #include "phiclass.h"
  49 #include "iredges_t.h"
  50 #include "execfreq.h"
  51 #include "irvrfy.h"
  52 #include "irbackedge_t.h"
  53 #include "irprofile.h"
  54
  55 #include <lpp/lpp.h>
  56 #include <lpp/mps.h>
  57 #include <lpp/lpp_net.h>
  58 #include <lpp/lpp_cplex.h>
  59
  60 #include "be_t.h"
  61 #include "beirg_t.h"
  62 #include "belive_t.h"
  63 #include "besched_t.h"
  64 #include "bessaconstr.h"
  65 #include "bearch_t.h"
  66 #include "beintlive_t.h"
  67 #include "beabi.h"
  68 #include "benode_t.h"
  69 #include "beutil.h"
  70 #include "bespillremat.h"
  71 #include "bespill.h"
  72 #include "bepressurestat.h"
  73 #include "bespilloptions.h"
  74 #include "bechordal_t.h"
  75 #include "bemodule.h"
  76
  77 #include "lc_opts.h"
  78 #include "lc_opts_enum.h"
  79
  80 #define DUMP_PROBLEM       1
  81 #define DUMP_MPS           2
  82 #define DUMP_SOLUTION      4
  83 #define DUMP_STATS         8
  84 #define DUMP_PRESSURE      16
  85
  86 #define KEEPALIVE_REMATS   1
  87 #define KEEPALIVE_SPILLS   2
  88 #define KEEPALIVE_RELOADS  4
  89
  90 #define VERIFY_MEMINTERF   1
  91 #define VERIFY_DOMINANCE   2
  92
  93 #define REMATS_NONE        0
  94 #define REMATS_BRIGGS      1
  95 #define REMATS_NOINVERSE   2
  96 #define REMATS_ALL         3
  97
  98 static unsigned opt_dump_flags = 0;
  99 static int opt_log = 0;
 100 static unsigned opt_keep_alive = 0;
 101 static int opt_goodwin = 1;
 102 static int opt_memcopies = 1;
 103 static int opt_memoperands = 1;
 104 static int opt_verify = VERIFY_MEMINTERF;
 105 static unsigned opt_remats = REMATS_ALL;
 106 static int opt_repair_schedule = 0;
 107 static int opt_no_enlarge_liveness = 0;
 108 static int opt_remat_while_live = 1;
 109 static int opt_timeout = 300;
 110 static double opt_cost_reload = 8.0;
 111 static double opt_cost_memoperand =  7.0;
 112 static double opt_cost_spill =  15.0;
 113 static double opt_cost_remat =  1.0;
 114
 115
 116 static const lc_opt_enum_mask_items_t dump_items[] = {
 117         { "problem",  DUMP_PROBLEM  },
 118         { "mps",      DUMP_MPS      },
 119         { "solution", DUMP_SOLUTION },
 120         { "stats",    DUMP_STATS },
 121         { "pressure", DUMP_PRESSURE },
 122         { NULL,       0 }
 123 };
 124
 125 static lc_opt_enum_mask_var_t dump_var = {
 126         &opt_dump_flags, dump_items
 127 };
 128
 129 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 130         { "remats",  KEEPALIVE_REMATS  },
 131         { "spills",  KEEPALIVE_SPILLS  },
 132         { "reloads", KEEPALIVE_RELOADS },
 133         { NULL,      0 }
 134 };
 135
 136 static lc_opt_enum_mask_var_t keep_alive_var = {
 137         &opt_keep_alive, keepalive_items
 138 };
 139
 140 static const lc_opt_enum_mask_items_t remats_items[] = {
 141         { "none",      REMATS_NONE      },
 142         { "briggs",    REMATS_BRIGGS    },
 143         { "noinverse", REMATS_NOINVERSE },
 144         { "all",       REMATS_ALL       },
 145         { NULL,        0 }
 146 };
 147
 148 static lc_opt_enum_mask_var_t remats_var = {
 149         &opt_remats, remats_items
 150 };
 151
 152 static const lc_opt_table_entry_t options[] = {
 153         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes",                              &keep_alive_var),
 154
 155         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 156         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 157         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 158         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert",                                 &remats_var),
 159         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 160         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 161         LC_OPT_ENT_BOOL     ("remat_while_live",  "only remat where rematted value was live",       &opt_remat_while_live),
 162
 163         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data",                  &dump_var),
 164         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 165         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 166
 167         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 168         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 169         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 170         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 171         LC_OPT_LAST
 172 };
 173
 174 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 175 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 176
 177 #define  SOLVE
 178 //#define  SOLVE_LOCAL
 179 #define LPP_SERVER "i44pc52"
 180 #define LPP_SOLVER "cplex"
 181
 182
 183 #define MAX_PATHS      INT_MAX
 184 #define ILP_UNDEF               -1
 185
 186 typedef struct _spill_ilp_t {
 187         const arch_register_class_t  *cls;
 188         int                           n_regs;
 189         be_irg_t                     *birg;
 190         be_lv_t                      *lv;
 191         lpp_t                        *lpp;
 192         struct obstack               *obst;
 193         set                          *remat_info;
 194         pset                         *all_possible_remats;
 195         pset                         *inverse_ops;
 196         ir_node                      *keep;
 197         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 198         pset                         *spills;
 199         set                          *interferences;
 200         ir_node                      *m_unknown;
 201         set                          *memoperands;
 202         phi_classes_t                *pc;
 203 #ifndef SCHEDULE_PHIM
 204         pset                         *phims;
 205 #endif
 206         DEBUG_ONLY(firm_dbg_module_t * dbg);
 207 } spill_ilp_t;
 208
 209 typedef int ilp_var_t;
 210 typedef int ilp_cst_t;
 211
 212 typedef struct _spill_bb_t {
 213         set      *ilp;
 214         set      *reloads;
 215 } spill_bb_t;
 216
 217 typedef struct _remat_t {
 218         const ir_node        *op;      /**< for copy_irn */
 219         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 220         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 221         int                   cost;    /**< cost of this remat */
 222         int                   inverse; /**< nonzero if this is an inverse remat */
 223 } remat_t;
 224
 225 /**
 226  * Data to be attached to each IR node. For remats this contains the ilp_var
 227  * for this remat and for normal ops this contains the ilp_vars for
 228  * reloading each operand
 229  */
 230 typedef struct _op_t {
 231         int             is_remat;
 232         union {
 233                 struct {
 234                         ilp_var_t       ilp;
 235                         const remat_t  *remat; /** the remat this op belongs to */
 236                         int             pre; /** 1, if this is a pressure-increasing remat */
 237                 } remat;
 238                 struct {
 239                         ilp_var_t       ilp;
 240                         ir_node        *op; /** the operation this live range belongs to */
 241                         union {
 242                                 ilp_var_t      *reloads;
 243                                 ilp_var_t      *copies;
 244                         } args;
 245                 } live_range;
 246         } attr;
 247 } op_t;
 248
 249 typedef struct _defs_t {
 250         const ir_node   *value;
 251         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 252         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 253 } defs_t;
 254
 255 typedef struct _remat_info_t {
 256         const ir_node       *irn; /**< the irn to which these remats belong */
 257         pset                *remats; /**< possible remats for this value */
 258         pset                *remats_by_operand; /**< remats with this value as operand */
 259 } remat_info_t;
 260
 261 typedef struct _keyval_t {
 262         const void          *key;
 263         const void          *val;
 264 } keyval_t;
 265
 266 typedef struct _spill_t {
 267         ir_node            *irn;
 268         ilp_var_t           reg_in;
 269         ilp_var_t           mem_in;
 270         ilp_var_t           reg_out;
 271         ilp_var_t           mem_out;
 272         ilp_var_t           spill;
 273 } spill_t;
 274
 275 typedef struct _memoperand_t {
 276         ir_node             *irn; /**< the irn */
 277         unsigned int         pos; /**< the position of the argument */
 278         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 279 } memoperand_t;
 280
 281 static INLINE int
 282 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 283 {
 284         return arch_irn_consider_in_reg_alloc(si->cls, irn);
 285 }
 286
 287 #if 0
 288 static int
 289 cmp_remat(const void *a, const void *b)
 290 {
 291         const keyval_t *p = a;
 292         const keyval_t *q = b;
 293         const remat_t  *r = p->val;
 294         const remat_t  *s = q->val;
 295
 296         assert(r && s);
 297
 298         return !(r == s || r->op == s->op);
 299 }
 300 #endif
 301 static int
 302 cmp_remat(const void *a, const void *b)
 303 {
 304         const remat_t  *r = a;
 305         const remat_t  *s = b;
 306
 307         return !(r == s || r->op == s->op);
 308 }
 309
 310 static int
 311 cmp_spill(const void *a, const void *b, size_t size)
 312 {
 313         const spill_t *p = a;
 314         const spill_t *q = b;
 315         (void) size;
 316
 317 //      return !(p->irn == q->irn && p->bb == q->bb);
 318         return !(p->irn == q->irn);
 319 }
 320
 321 static int
 322 cmp_memoperands(const void *a, const void *b, size_t size)
 323 {
 324         const memoperand_t *p = a;
 325         const memoperand_t *q = b;
 326         (void) size;
 327
 328         return !(p->irn == q->irn && p->pos == q->pos);
 329 }
 330
 331 static keyval_t *
 332 set_find_keyval(set * set, const void * key)
 333 {
 334         keyval_t     query;
 335
 336         query.key = key;
 337         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 338 }
 339
 340 static keyval_t *
 341 set_insert_keyval(set * set, void * key, void * val)
 342 {
 343         keyval_t     query;
 344
 345         query.key = key;
 346         query.val = val;
 347         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 348 }
 349
 350 static defs_t *
 351 set_find_def(set * set, const ir_node * value)
 352 {
 353         defs_t     query;
 354
 355         query.value = value;
 356         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 357 }
 358
 359 static defs_t *
 360 set_insert_def(set * set, const ir_node * value)
 361 {
 362         defs_t     query;
 363
 364         query.value = value;
 365         query.spills = NULL;
 366         query.remats = NULL;
 367         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 368 }
 369
 370 static memoperand_t *
 371 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 372 {
 373         memoperand_t     query;
 374
 375         query.irn = irn;
 376         query.pos = pos;
 377         query.ilp = ilp;
 378         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 379 }
 380
 381 static memoperand_t *
 382 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 383 {
 384         memoperand_t     query;
 385
 386         query.irn = (ir_node*)irn;
 387         query.pos = pos;
 388         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 389 }
 390
 391
 392 static spill_t *
 393 set_find_spill(set * set, const ir_node * value)
 394 {
 395         spill_t     query;
 396
 397         query.irn = (ir_node*)value;
 398         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 399 }
 400
 401 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 402 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 403 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 404 #define foreach_pre_remat(s,i) for((i)=next_pre_remat((s)); (i); (i)=next_pre_remat((i)))
 405 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 406
 407 static int
 408 cmp_remat_info(const void *a, const void *b, size_t size)
 409 {
 410         const remat_info_t *p = a;
 411         const remat_info_t *q = b;
 412         (void) size;
 413
 414         return !(p->irn == q->irn);
 415 }
 416
 417 static int
 418 cmp_defs(const void *a, const void *b, size_t size)
 419 {
 420         const defs_t *p = a;
 421         const defs_t *q = b;
 422         (void) size;
 423
 424         return !(p->value == q->value);
 425 }
 426
 427 static int
 428 cmp_keyval(const void *a, const void *b, size_t size)
 429 {
 430         const keyval_t *p = a;
 431         const keyval_t *q = b;
 432         (void) size;
 433
 434         return !(p->key == q->key);
 435 }
 436
 437 static double
 438 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 439 {
 440 #define FUDGE 0.001
 441         if(ir_profile_has_data())
 442                 return ((double)ir_profile_get_block_execcount(get_block_const(irn))) + FUDGE;
 443
 444 #ifndef EXECFREQ_LOOPDEPH
 445         return get_block_execfreq(si->birg->exec_freq, get_block_const(irn)) + FUDGE;
 446 #else
 447         if(is_Block(irn))
 448                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 449         else
 450                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 451 #endif
 452 }
 453
 454 static double get_cost(const ir_node *irn)
 455 {
 456         if(be_is_Spill(irn)) {
 457                 return opt_cost_spill;
 458         } else if(be_is_Reload(irn)){
 459                 return opt_cost_reload;
 460         } else {
 461                 return arch_get_op_estimated_cost(irn);
 462         }
 463 }
 464
 465 /**
 466  * Checks, whether node and its operands have suitable reg classes
 467  */
 468 static INLINE int
 469 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 470 {
 471         int n;
 472         int remat = (arch_irn_get_flags(irn) & arch_irn_flags_rematerializable) != 0;
 473
 474 #if 0
 475         if(!remat)
 476                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 477 #endif
 478
 479         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 480                 ir_node        *op = get_irn_n(irn, n);
 481                 remat &= has_reg_class(si, op) || arch_irn_get_flags(op) & arch_irn_flags_ignore || is_NoMem(op);
 482
 483 //              if(!remat)
 484 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 485         }
 486
 487         return remat;
 488 }
 489
 490 /**
 491  * Try to create a remat from @p op with destination value @p dest_value
 492  */
 493 static INLINE remat_t *
 494 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 495 {
 496         remat_t  *remat = NULL;
 497
 498 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 499 //              return NULL;
 500
 501         if(dest_value == op) {
 502                 const ir_node *proj = NULL;
 503
 504                 if(is_Proj(dest_value)) {
 505                         op = get_Proj_pred(op);
 506                         proj = dest_value;
 507                 }
 508
 509                 if(!is_rematerializable(si, op))
 510                         return NULL;
 511
 512                 remat          = obstack_alloc(si->obst, sizeof(*remat));
 513                 remat->op      = op;
 514                 remat->cost    = (int)get_cost(op);
 515                 remat->value   = dest_value;
 516                 remat->proj    = proj;
 517                 remat->inverse = 0;
 518         } else {
 519                 arch_inverse_t     inverse;
 520                 int                n;
 521
 522                 /* get the index of the operand we want to retrieve by the inverse op */
 523                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 524                         ir_node        *arg = get_irn_n(op, n);
 525
 526                         if(arg == dest_value) break;
 527                 }
 528                 if(n<0) return NULL;
 529
 530                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 531
 532                 /* else ask the backend to give an inverse op */
 533                 if(arch_get_inverse(op, n, &inverse, si->obst)) {
 534                         int   i;
 535
 536                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 537
 538                         assert(inverse.n > 0 && "inverse op should have at least one node");
 539
 540                         for(i=inverse.n-1; i>=0; --i) {
 541                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 542                         }
 543
 544                         if(inverse.n <= 2) {
 545                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 546                                 remat->op = inverse.nodes[0];
 547                                 remat->cost = inverse.costs;
 548                                 remat->value = dest_value;
 549                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 550                                 remat->inverse = 1;
 551
 552                                 // Matze: commented this out, this doesn't seem to be true if
 553                                 // the inverse is a simple operation with only 1 result...
 554                                 //assert(is_Proj(remat->proj));
 555                         } else {
 556                                 assert(0 && "I can not handle remats with more than 2 nodes");
 557                         }
 558                 }
 559         }
 560
 561         if(remat) {
 562                 if(remat->proj) {
 563                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 564                 } else {
 565                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 566                 }
 567         }
 568         return remat;
 569 }
 570
 571
 572 static INLINE void
 573 add_remat(const spill_ilp_t * si, const remat_t * remat)
 574 {
 575         remat_info_t    *remat_info,
 576                      query;
 577         int              n;
 578
 579         assert(remat->op);
 580         assert(remat->value);
 581
 582         query.irn = remat->value;
 583         query.remats = NULL;
 584         query.remats_by_operand = NULL;
 585         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 586
 587         if(remat_info->remats == NULL) {
 588                 remat_info->remats = new_pset(cmp_remat, 4096);
 589         }
 590         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 591
 592         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 593         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 594                 ir_node        *arg = get_irn_n(remat->op, n);
 595
 596                 query.irn = arg;
 597                 query.remats = NULL;
 598                 query.remats_by_operand = NULL;
 599                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 600
 601                 if(remat_info->remats_by_operand == NULL) {
 602                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 603                 }
 604                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 605         }
 606 }
 607
 608 static int
 609 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 610 {
 611         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 612         int                i = 0;
 613
 614         while(edge) {
 615                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 616                         ++i;
 617                 }
 618                 edge = get_irn_out_edge_next(irn, edge);
 619         }
 620
 621         return i;
 622 }
 623
 624 static int
 625 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 626 {
 627         int n;
 628         int ret = 0;
 629
 630         if(is_Proj(irn))
 631                 irn = get_Proj_pred(irn);
 632
 633         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 634                 const ir_node  *arg = get_irn_n(irn, n);
 635
 636                 if(has_reg_class(si, arg)) ++ret;
 637         }
 638
 639         return ret;
 640 }
 641
 642 static INLINE void
 643 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 644 {
 645         int      n;
 646         remat_t *remat;
 647
 648         if( has_reg_class(si, op)
 649         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 650         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 651         ) {
 652                 remat = get_remat_from_op(si, op, op);
 653                 if(remat) {
 654                         add_remat(si, remat);
 655                 }
 656         }
 657
 658         if(opt_remats == REMATS_ALL) {
 659                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 660                    for each arg */
 661                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 662                         ir_node        *arg = get_irn_n(op, n);
 663
 664                         if(has_reg_class(si, arg)) {
 665                                 /* try to get an inverse remat */
 666                                 remat = get_remat_from_op(si, arg, op);
 667                                 if(remat) {
 668                                         add_remat(si, remat);
 669                                 }
 670                         }
 671                 }
 672         }
 673 }
 674
 675 static INLINE int
 676 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 677 {
 678         ir_node *block;
 679         ir_node *def_block = get_nodes_block(val);
 680         int      ret;
 681         (void) si;
 682
 683         if(val == pos)
 684                 return 0;
 685
 686         /* if pos is at end of a basic block */
 687         if(is_Block(pos)) {
 688                 ret = (pos == def_block || block_dominates(def_block, pos));
 689 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 690                 return ret;
 691         }
 692
 693         /* else if this is a normal operation */
 694         block = get_nodes_block(pos);
 695         if(block == def_block) {
 696                 if(!sched_is_scheduled(val)) return 1;
 697
 698                 ret = sched_comes_after(val, pos);
 699 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 700                 return ret;
 701         }
 702
 703         ret = block_dominates(def_block, block);
 704 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 705         return ret;
 706 }
 707
 708 static INLINE ir_node *sched_block_last_noncf(const ir_node * bb)
 709 {
 710         return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, NULL);
 711 }
 712
 713 /**
 714  * Returns first non-Phi node of block @p bb
 715  */
 716 static INLINE ir_node *
 717 sched_block_first_nonphi(const ir_node * bb)
 718 {
 719         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 720 }
 721
 722 static int
 723 sched_skip_proj_predicator(const ir_node * irn, void * data)
 724 {
 725         (void) data;
 726         return (is_Proj(irn));
 727 }
 728
 729 static INLINE ir_node *
 730 sched_next_nonproj(const ir_node * irn, int forward)
 731 {
 732         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 733 }
 734
 735 /**
 736  * Returns next operation node (non-Proj) after @p irn
 737  * or the basic block of this node
 738  */
 739 static INLINE ir_node *
 740 sched_next_op(const ir_node * irn)
 741 {
 742         ir_node *next = sched_next(irn);
 743
 744         if(is_Block(next))
 745                 return next;
 746
 747         return sched_next_nonproj(next, 1);
 748 }
 749
 750 /**
 751  * Returns previous operation node (non-Proj) before @p irn
 752  * or the basic block of this node
 753  */
 754 static INLINE ir_node *
 755 sched_prev_op(const ir_node * irn)
 756 {
 757         ir_node *prev = sched_prev(irn);
 758
 759         if(is_Block(prev))
 760                 return prev;
 761
 762         return sched_next_nonproj(prev, 0);
 763 }
 764
 765 static void
 766 sched_put_after(ir_node * insert, ir_node * irn)
 767 {
 768         if(is_Block(insert)) {
 769                 insert = sched_block_first_nonphi(insert);
 770         } else {
 771                 insert = sched_next_op(insert);
 772         }
 773         sched_reset(irn);
 774         sched_add_before(insert, irn);
 775 }
 776
 777 static void sched_put_before(ir_node * insert, ir_node * irn)
 778 {
 779   if(is_Block(insert)) {
 780           insert = sched_block_last_noncf(insert);
 781   } else {
 782           insert = sched_next_nonproj(insert, 0);
 783           insert = sched_prev(insert);
 784   }
 785   sched_reset(irn);
 786   sched_add_after(insert, irn);
 787 }
 788
 789 static ir_node *
 790 next_post_remat(const ir_node * irn)
 791 {
 792         op_t      *op;
 793     ir_node   *next;
 794
 795         if(is_Block(irn)) {
 796                 next = sched_block_first_nonphi(irn);
 797         } else {
 798                 next = sched_next_op(irn);
 799         }
 800
 801         if(sched_is_end(next))
 802                 return NULL;
 803
 804         op = get_irn_link(next);
 805         if(op->is_remat && !op->attr.remat.pre) {
 806                 return next;
 807         }
 808
 809         return NULL;
 810 }
 811
 812
 813 static ir_node *next_pre_remat(const ir_node * irn)
 814 {
 815         op_t      *op;
 816         ir_node   *ret;
 817
 818         if(is_Block(irn)) {
 819                 ret = sched_block_last_noncf(irn);
 820                 ret = sched_next(ret);
 821                 ret = sched_prev_op(ret);
 822         } else {
 823                 ret = sched_prev_op(irn);
 824         }
 825
 826         if(sched_is_end(ret) || is_Phi(ret))
 827                 return NULL;
 828
 829         op = (op_t*)get_irn_link(ret);
 830         if(op->is_remat && op->attr.remat.pre) {
 831                 return ret;
 832         }
 833
 834         return NULL;
 835 }
 836
 837 /**
 838  * Tells you whether a @p remat can be placed before the irn @p pos
 839  */
 840 static INLINE int
 841 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 842 {
 843         const ir_node   *op = remat->op;
 844         const ir_node   *prev;
 845         int        n,
 846                            res = 1;
 847
 848         if(is_Block(pos)) {
 849                 prev = sched_block_last_noncf(pos);
 850                 prev = sched_next_nonproj(prev, 0);
 851         } else {
 852                 prev = sched_prev_op(pos);
 853         }
 854         /* do not remat if the rematted value is defined immediately before this op */
 855         if(prev == remat->op) {
 856                 return 0;
 857         }
 858
 859 #if 0
 860         /* this should be just fine, the following OP will be using this value, right? */
 861
 862         /* only remat AFTER the real definition of a value (?) */
 863         if(!value_is_defined_before(si, pos, remat->value)) {
 864 //              ir_fprintf(stderr, "error(not defined)");
 865                 return 0;
 866         }
 867 #endif
 868
 869         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 870                 const ir_node   *arg = get_irn_n(op, n);
 871
 872                 if(opt_no_enlarge_liveness) {
 873                         if(has_reg_class(si, arg) && live) {
 874                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 875                         } else {
 876                                 res &= value_is_defined_before(si, pos, arg);
 877                         }
 878                 } else {
 879                         res &= value_is_defined_before(si, pos, arg);
 880                 }
 881         }
 882
 883         return res;
 884 }
 885
 886 /**
 887  * Tells you whether a @p remat can be placed after the irn @p pos
 888  */
 889 static INLINE int
 890 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 891 {
 892         if(is_Block(pos)) {
 893                 pos = sched_block_first_nonphi(pos);
 894         } else {
 895                 pos = sched_next_op(pos);
 896         }
 897
 898         /* only remat AFTER the real definition of a value (?) */
 899         if(!value_is_defined_before(si, pos, remat->value)) {
 900                 return 0;
 901         }
 902
 903         return can_remat_before(si, remat, pos, live);
 904 }
 905
 906 /**
 907  * Collect potetially rematerializable OPs
 908  */
 909 static void
 910 walker_remat_collector(ir_node * irn, void * data)
 911 {
 912         spill_ilp_t    *si = data;
 913
 914         if(!is_Block(irn) && !is_Phi(irn)) {
 915                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 916                 get_remats_from_op(si, irn);
 917         }
 918 }
 919
 920 /**
 921  * Inserts a copy of @p irn before @p pos
 922  */
 923 static ir_node *
 924 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 925 {
 926         ir_node     *bb;
 927         ir_node     *copy;
 928
 929         bb = is_Block(pos)?pos:get_nodes_block(pos);
 930         copy = exact_copy(irn);
 931
 932         set_phi_class(si->pc, copy, NULL);
 933         set_nodes_block(copy, bb);
 934         sched_put_before(pos, copy);
 935
 936         return copy;
 937 }
 938
 939 /**
 940  * Inserts a copy of @p irn after @p pos
 941  */
 942 static ir_node *
 943 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 944 {
 945         ir_node     *bb;
 946         ir_node     *copy;
 947
 948         bb = is_Block(pos)?pos:get_nodes_block(pos);
 949         copy = exact_copy(irn);
 950
 951         set_phi_class(si->pc, copy, NULL);
 952         set_nodes_block(copy, bb);
 953         sched_put_after(pos, copy);
 954
 955         return copy;
 956 }
 957
 958 static ir_node *
 959 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 960 {
 961         char     buf[256];
 962
 963         if(can_remat_after(si, remat, pos, live)) {
 964                 ir_node         *copy,
 965                                                 *proj_copy;
 966                 op_t            *op;
 967
 968                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 969
 970                 copy = insert_copy_after(si, remat->op, pos);
 971
 972                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 973                 op = obstack_alloc(si->obst, sizeof(*op));
 974                 op->is_remat = 1;
 975                 op->attr.remat.remat = remat;
 976                 op->attr.remat.pre = 0;
 977                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 978
 979                 set_irn_link(copy, op);
 980                 pset_insert_ptr(si->all_possible_remats, copy);
 981                 if(remat->proj) {
 982                         proj_copy = insert_copy_after(si, remat->proj, copy);
 983                         set_irn_n(proj_copy, 0, copy);
 984                         set_irn_link(proj_copy, op);
 985                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 986                 } else {
 987                         proj_copy = NULL;
 988                 }
 989
 990                 return copy;
 991         }
 992
 993         return NULL;
 994 }
 995
 996 static ir_node *
 997 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 998 {
 999         char     buf[256];
1000
1001         if(can_remat_before(si, remat, pos, live)) {
1002                 ir_node         *copy,
1003                                                 *proj_copy;
1004                 op_t            *op;
1005
1006                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
1007
1008                 copy = insert_copy_before(si, remat->op, pos);
1009
1010                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
1011                 op = obstack_alloc(si->obst, sizeof(*op));
1012                 op->is_remat = 1;
1013                 op->attr.remat.remat = remat;
1014                 op->attr.remat.pre = 1;
1015                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
1016
1017                 set_irn_link(copy, op);
1018                 pset_insert_ptr(si->all_possible_remats, copy);
1019                 if(remat->proj) {
1020                         proj_copy = insert_copy_after(si, remat->proj, copy);
1021                         set_irn_n(proj_copy, 0, copy);
1022                         set_irn_link(proj_copy, op);
1023                         pset_insert_ptr(si->all_possible_remats, proj_copy);
1024                 } else {
1025                         proj_copy = NULL;
1026                 }
1027
1028                 return copy;
1029         }
1030
1031         return NULL;
1032 }
1033
1034 static int
1035 get_block_n_succs(const ir_node *block) {
1036         const ir_edge_t *edge;
1037
1038         assert(edges_activated(current_ir_graph));
1039
1040         edge = get_block_succ_first(block);
1041         if (! edge)
1042                 return 0;
1043
1044         edge = get_block_succ_next(block, edge);
1045         return edge ? 2 : 1;
1046 }
1047
1048 static int
1049 is_start_block(const ir_node * bb)
1050 {
1051         return get_irg_start_block(get_irn_irg(bb)) == bb;
1052 }
1053
1054 static int
1055 is_merge_edge(const ir_node * bb)
1056 {
1057         if(is_start_block(bb))
1058                 return 0;
1059
1060         if(opt_goodwin)
1061                 return get_block_n_succs(bb) == 1;
1062         else
1063                 return 1;
1064 }
1065
1066 static int
1067 is_diverge_edge(const ir_node * bb)
1068 {
1069         if(is_start_block(bb))
1070                 return 0;
1071
1072         if(opt_goodwin)
1073                 return get_Block_n_cfgpreds(bb) == 1;
1074         else
1075                 return 1;
1076 }
1077
1078 static void
1079 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1080 {
1081         ir_node        *irn;
1082         int i;
1083
1084         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1085                 irn = be_lv_get_irn(si->lv, bb, i);
1086
1087                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1088                         pset_insert_ptr(live, irn);
1089                 }
1090         }
1091
1092         irn = sched_last(bb);
1093
1094         /* all values eaten by control flow operations are also live until the end of the block */
1095         sched_foreach_reverse(bb, irn) {
1096                 int  i;
1097
1098                 if (!sched_skip_cf_predicator(irn, NULL)) break;
1099
1100                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1101                         ir_node *arg = get_irn_n(irn,i);
1102
1103                         if(has_reg_class(si, arg)) {
1104                                 pset_insert_ptr(live, arg);
1105                         }
1106                 }
1107         }
1108         /*
1109          * find values that are used by remats at end of block
1110          * and insert them into live set
1111          */
1112         foreach_pre_remat(bb, irn) {
1113                 int       n;
1114
1115                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1116                         ir_node        *remat_arg = get_irn_n(irn, n);
1117
1118                         if(!has_reg_class(si, remat_arg)) continue;
1119
1120                         /* if value is becoming live through use by remat */
1121                         if(!pset_find_ptr(live, remat_arg)) {
1122                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1123
1124                                 pset_insert_ptr(live, remat_arg);
1125                         }
1126                 }
1127         }
1128 }
1129
1130 static void
1131 walker_regclass_copy_insertor(ir_node * irn, void * data)
1132 {
1133         spill_ilp_t    *si = data;
1134
1135         if(is_Phi(irn) && has_reg_class(si, irn)) {
1136                 int n;
1137
1138                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1139                         ir_node  *phi_arg = get_irn_n(irn, n);
1140                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1141
1142                         if(!has_reg_class(si, phi_arg)) {
1143                                 ir_node   *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1144                                 ir_node   *pos = sched_block_last_noncf(bb);
1145                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1146
1147                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1148                                 sched_add_after(pos, copy);
1149                                 set_irn_n(irn, n, copy);
1150
1151                                 op->is_remat = 0;
1152                                 op->attr.live_range.args.reloads = NULL;
1153                                 op->attr.live_range.ilp = ILP_UNDEF;
1154                                 set_irn_link(copy, op);
1155                         }
1156                 }
1157         }
1158 }
1159
1160 /**
1161  * Insert (so far unused) remats into the irg to
1162  * recompute the potential liveness of all values
1163  */
1164 static void
1165 walker_remat_insertor(ir_node * bb, void * data)
1166 {
1167         spill_ilp_t    *si = data;
1168         ir_node        *irn;
1169         int             n, i;
1170         pset           *live;
1171         pset           *post_remats;
1172         remat_t        *remat;
1173
1174         /* skip start block, no remats to do there */
1175         if(is_start_block(bb)) return;
1176
1177         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1178
1179         live = pset_new_ptr_default();
1180         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1181                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1182
1183                 /* add remats at end of block */
1184                 if (has_reg_class(si, value)) {
1185                         pset_insert_ptr(live, value);
1186                 }
1187         }
1188
1189         irn = sched_last(bb);
1190         while(!sched_is_end(irn)) {
1191                 ir_node   *next;
1192                 pset      *args;
1193                 ir_node   *arg;
1194                 pset      *used;
1195
1196                 next = sched_prev(irn);
1197
1198                 /* delete defined value from live set */
1199                 if(has_reg_class(si, irn)) {
1200                         pset_remove_ptr(live, irn);
1201                 }
1202
1203                 if(is_Phi(irn) || is_Proj(irn)) {
1204                         irn = next;
1205                         continue;
1206                 }
1207
1208                 args = pset_new_ptr_default();
1209                 used = pset_new_ptr_default();
1210
1211                 /* collect arguments of op and set args of op already live in epilog */
1212                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1213                         ir_node        *arg = get_irn_n(irn, n);
1214
1215                         pset_insert_ptr(args, arg);
1216                         if(has_reg_class(si, arg)) {
1217                                 pset_insert_ptr(live, arg);
1218                                 pset_insert_ptr(used, arg);
1219                         }
1220                 }
1221
1222                 /* insert all possible remats before irn */
1223                 pset_foreach(args, arg) {
1224                         remat_info_t   *remat_info,
1225                                                     query;
1226
1227                         /* continue if the operand has the wrong reg class */
1228                         if(!has_reg_class(si, arg))
1229                                 continue;
1230
1231                         query.irn = arg;
1232                         query.remats = NULL;
1233                         query.remats_by_operand = NULL;
1234                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1235
1236                         if(!remat_info) {
1237                                 continue;
1238                         }
1239
1240                         if(remat_info->remats) {
1241                                 pset_foreach(remat_info->remats, remat) {
1242                                         ir_node  *remat_irn = NULL;
1243
1244                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1245                                         remat_irn = insert_remat_before(si, remat, irn, live);
1246
1247                                         if(remat_irn) {
1248                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1249                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1250
1251                                                         /* collect args of remats which are not args of op */
1252                                                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1253                                                                 pset_insert_ptr(used, remat_arg);
1254                                                         }
1255                                                 }
1256                                         }
1257                                 }
1258                         }
1259                 }
1260
1261                 /* do not place post remats after jumps */
1262                 if (sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1263                         del_pset(used);
1264                         del_pset(args);
1265                         break;
1266                 }
1267
1268                 /* insert all possible remats after irn */
1269                 post_remats = pset_new_ptr_default();
1270                 pset_foreach(used, arg) {
1271                         remat_info_t   *remat_info,
1272                                                     query;
1273
1274                         /* continue if the operand has the wrong reg class */
1275                         if(!has_reg_class(si, arg))
1276                                 continue;
1277
1278                         query.irn = arg;
1279                         query.remats = NULL;
1280                         query.remats_by_operand = NULL;
1281                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1282
1283                         if(!remat_info) {
1284                                 continue;
1285                         }
1286
1287                         if(remat_info->remats_by_operand) {
1288                                 pset_foreach(remat_info->remats_by_operand, remat) {
1289                                         /* do not insert remats producing the same value as one of the operands */
1290                                         if(!pset_find_ptr(args, remat->value)) {
1291                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1292
1293                                                 /* only remat values that can be used by real ops */
1294                                                 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1295                                                         pset_insert_ptr(post_remats, remat);
1296                                                 }
1297                                         }
1298                                 }
1299                         }
1300                 }
1301                 pset_foreach(post_remats, remat) {
1302                         insert_remat_after(si, remat, irn, live);
1303                 }
1304                 del_pset(post_remats);
1305
1306                 del_pset(used);
1307                 del_pset(args);
1308                 irn = next;
1309         }
1310
1311         /* add remats at end if successor has multiple predecessors */
1312         if(is_merge_edge(bb)) {
1313                 pset     *live_out = pset_new_ptr_default();
1314                 ir_node  *value;
1315
1316                 get_live_end(si, bb, live_out);
1317
1318                 /* add remats at end of block */
1319                 pset_foreach(live_out, value) {
1320                         remat_info_t   *remat_info,
1321                                                    query;
1322
1323                         query.irn = value;
1324                         query.remats = NULL;
1325                         query.remats_by_operand = NULL;
1326                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1327
1328                         if(remat_info && remat_info->remats) {
1329                                 pset_foreach(remat_info->remats, remat) {
1330                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1331
1332                                         insert_remat_before(si, remat, bb, live_out);
1333                                 }
1334                         }
1335                 }
1336                 del_pset(live_out);
1337         }
1338
1339         if(is_diverge_edge(bb)) {
1340                 pset     *live_in = pset_new_ptr_default();
1341                 ir_node  *value;
1342
1343                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1344                         value = be_lv_get_irn(si->lv, bb, i);
1345
1346                         if(has_reg_class(si, value)) {
1347                                 pset_insert_ptr(live_in, value);
1348                         }
1349                 }
1350                 /* add phis to live_in */
1351                 sched_foreach(bb, value) {
1352                         if(!is_Phi(value)) break;
1353
1354                         if(has_reg_class(si, value)) {
1355                                 pset_insert_ptr(live_in, value);
1356                         }
1357                 }
1358
1359                 /* add remat2s at beginning of block */
1360                 post_remats = pset_new_ptr_default();
1361                 pset_foreach(live_in, value) {
1362                         remat_info_t   *remat_info,
1363                                                    query;
1364
1365                         query.irn = value;
1366                         query.remats = NULL;
1367                         query.remats_by_operand = NULL;
1368                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1369
1370                         if(remat_info && remat_info->remats_by_operand) {
1371                                 pset_foreach(remat_info->remats_by_operand, remat) {
1372                                         DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1373
1374                                         /* put the remat here if all its args are available and result is still live */
1375                                         if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1376                                                 pset_insert_ptr(post_remats, remat);
1377                                         }
1378                                 }
1379                         }
1380                 }
1381                 pset_foreach(post_remats, remat) {
1382                         insert_remat_after(si, remat, bb, live_in);
1383                 }
1384                 del_pset(post_remats);
1385                 del_pset(live_in);
1386         }
1387 }
1388
1389 static int
1390 can_be_copied(const ir_node * bb, const ir_node * irn)
1391 {
1392         const ir_edge_t *edge    = get_block_succ_first(bb);
1393         const ir_node   *next_bb = edge->src;
1394         int             pos      = edge->pos;
1395         const ir_node   *phi;
1396
1397         assert(is_merge_edge(bb));
1398
1399         sched_foreach(next_bb, phi) {
1400                 const ir_node  *phi_arg;
1401
1402                 if(!is_Phi(phi)) break;
1403
1404                 phi_arg = get_irn_n(phi, pos);
1405
1406                 if(phi_arg == irn) {
1407                         return 1;
1408                 }
1409         }
1410         return 0;
1411 }
1412
1413 /**
1414  * Initialize additional node info
1415  */
1416 static void
1417 luke_initializer(ir_node * bb, void * data)
1418 {
1419         spill_ilp_t    *si = (spill_ilp_t*)data;
1420         spill_bb_t     *spill_bb;
1421         ir_node        *irn;
1422
1423         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1424         set_irn_link(bb, spill_bb);
1425
1426         sched_foreach(bb, irn) {
1427                 op_t      *op;
1428
1429                 op = obstack_alloc(si->obst, sizeof(*op));
1430                 op->is_remat = 0;
1431                 op->attr.live_range.ilp = ILP_UNDEF;
1432                 if(is_Phi(irn)) {
1433                         if(opt_memcopies) {
1434                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1435                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1436                         }
1437                 } else if(!is_Proj(irn)) {
1438                         op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1439                         memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1440                 } else {
1441                         op->attr.live_range.args.reloads = NULL;
1442                 }
1443                 set_irn_link(irn, op);
1444         }
1445 }
1446
1447
1448 /**
1449  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1450  */
1451 static void
1452 luke_endwalker(ir_node * bb, void * data)
1453 {
1454         spill_ilp_t    *si = (spill_ilp_t*)data;
1455         pset           *live;
1456         pset           *use_end;
1457         char            buf[256];
1458         ilp_cst_t       cst;
1459         ir_node        *irn;
1460         spill_bb_t     *spill_bb = get_irn_link(bb);
1461         int             i;
1462
1463         live = pset_new_ptr_default();
1464         use_end = pset_new_ptr_default();
1465
1466         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1467                 irn = be_lv_get_irn(si->lv, bb, i);
1468                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1469                         pset_insert_ptr(live, irn);
1470                 }
1471         }
1472         /*
1473          * find values that are used by remats at end of block
1474          * and insert them into live set
1475          */
1476         foreach_pre_remat(bb, irn) {
1477                 int       n;
1478
1479                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1480                         ir_node        *remat_arg = get_irn_n(irn, n);
1481
1482                         if(has_reg_class(si, remat_arg)) {
1483                                 pset_insert_ptr(live, remat_arg);
1484                         }
1485                 }
1486         }
1487
1488         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1489         /* their reg_out must always be set */
1490         sched_foreach_reverse(bb, irn) {
1491                 int   n;
1492
1493                 if (!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1494
1495                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1496                         ir_node        *irn_arg = get_irn_n(irn, n);
1497
1498                         if(has_reg_class(si, irn_arg)) {
1499                                 pset_insert_ptr(use_end, irn_arg);
1500                         }
1501                 }
1502         }
1503
1504         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1505         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1506         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1507
1508         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1509
1510         /* if this is a merge edge we can reload at the end of this block */
1511         if(is_merge_edge(bb)) {
1512                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1513         } else if(pset_count(use_end)){
1514                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1515         } else {
1516                 spill_bb->reloads = NULL;
1517         }
1518
1519         pset_foreach(live,irn) {
1520                 spill_t     query,
1521                                         *spill;
1522                 double      spill_cost;
1523                 int         default_spilled;
1524
1525
1526                 /* handle values used by control flow nodes later separately */
1527                 if(pset_find_ptr(use_end, irn)) continue;
1528
1529                 query.irn = irn;
1530                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1531
1532                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1533
1534                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1535                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1536                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1537
1538                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1539                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1540
1541                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1542                 /* by default spill value right after definition */
1543                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1544                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1545
1546                 if(is_merge_edge(bb)) {
1547                         ilp_var_t   reload;
1548                         ilp_cst_t   rel_cst;
1549
1550                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1551                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1552                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1553
1554                         /* reload <= mem_out */
1555                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1556                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1557                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1558                 }
1559
1560                 spill->reg_in = ILP_UNDEF;
1561                 spill->mem_in = ILP_UNDEF;
1562         }
1563
1564         pset_foreach(use_end,irn) {
1565                 spill_t     query,
1566                                         *spill;
1567                 double      spill_cost;
1568                 ilp_cst_t   end_use_req,
1569                                         rel_cst;
1570                 ilp_var_t   reload;
1571                 int         default_spilled;
1572
1573                 query.irn = irn;
1574                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1575
1576                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1577
1578                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1579                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1580
1581                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1582                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1583
1584                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1585                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1586                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1587
1588                 /* reload for use be control flow op */
1589                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1590                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1591                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1592
1593                 /* reload <= mem_out */
1594                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1595                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1596                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1597
1598                 spill->reg_in = ILP_UNDEF;
1599                 spill->mem_in = ILP_UNDEF;
1600
1601                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1602                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1603                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1604         }
1605
1606         del_pset(live);
1607         del_pset(use_end);
1608 }
1609
1610 #ifndef NDEBUG
1611 /**
1612  * Find a remat of value @p value in the epilog of @p pos
1613  */
1614 static ir_node *
1615 find_post_remat(const ir_node * value, const ir_node * pos)
1616 {
1617         while((pos = next_post_remat(pos)) != NULL) {
1618                 op_t   *op;
1619
1620                 op = get_irn_link(pos);
1621                 assert(op->is_remat && !op->attr.remat.pre);
1622
1623                 if(op->attr.remat.remat->value == value)
1624                         return (ir_node*)pos;
1625
1626 #if 0
1627         const ir_edge_t *edge;
1628                 foreach_out_edge(pos, edge) {
1629                         ir_node   *proj = get_edge_src_irn(edge);
1630                         assert(is_Proj(proj));
1631                 }
1632 #endif
1633
1634         }
1635
1636         return NULL;
1637 }
1638 #endif
1639
1640 static spill_t *
1641 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1642 {
1643         spill_bb_t  *spill_bb = get_irn_link(bb);
1644         spill_t     *spill,
1645                                  query;
1646         char         buf[256];
1647         int          default_spilled;
1648
1649         query.irn = irn;
1650         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1651         if(!spill) {
1652                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1653
1654                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1655
1656                 spill->reg_out = ILP_UNDEF;
1657                 spill->reg_in  = ILP_UNDEF;
1658                 spill->mem_in  = ILP_UNDEF;
1659
1660                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1661                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1662
1663                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1664                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1665                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1666         }
1667
1668         return spill;
1669 }
1670
1671 /**
1672  *  Inserts ILP-constraints and variables for memory copying before the given position
1673  */
1674 static void
1675 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1676 {
1677         const ir_node    *succ;
1678         const ir_edge_t  *edge;
1679         spill_bb_t       *spill_bb = get_irn_link(block);
1680         ir_node          *phi;
1681         int               pos;
1682         ilp_cst_t         cst;
1683         ilp_var_t         copyreg;
1684         char              buf[256];
1685         ir_node          *tmp;
1686
1687
1688         assert(edges_activated(current_ir_graph));
1689
1690         edge = get_block_succ_first(block);
1691         if(!edge) return;
1692
1693         succ = edge->src;
1694         pos = edge->pos;
1695
1696         edge = get_block_succ_next(block, edge);
1697         /* next block can only contain phis, if this is a merge edge */
1698         if(edge) return;
1699
1700         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1701         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1702
1703         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1704         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1705
1706         pset_foreach(live, tmp) {
1707                 spill_t  *spill;
1708 #if 0
1709                 op_t  *op = get_irn_link(irn);
1710                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1711 #endif
1712                 spill = set_find_spill(spill_bb->ilp, tmp);
1713                 assert(spill);
1714
1715                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1716         }
1717         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1718
1719         sched_foreach(succ, phi) {
1720                 const ir_node  *to_copy;
1721                 op_t           *to_copy_op;
1722                 spill_t        *to_copy_spill;
1723                 op_t           *phi_op = get_irn_link(phi);
1724                 ilp_var_t       reload = ILP_UNDEF;
1725
1726
1727                 if(!is_Phi(phi)) break;
1728                 if(!has_reg_class(si, phi)) continue;
1729
1730                 to_copy = get_irn_n(phi, pos);
1731                 to_copy_op = get_irn_link(to_copy);
1732
1733                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1734                 assert(to_copy_spill);
1735
1736                 if(spill_bb->reloads) {
1737                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1738
1739                         if(keyval) {
1740                                 reload = PTR_TO_INT(keyval->val);
1741                         }
1742                 }
1743
1744                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1745                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1746
1747                 /* copy - reg_out - reload - remat - live_range <= 0 */
1748                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1749                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1750                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1751                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1752                 foreach_pre_remat(block, tmp) {
1753                         op_t     *remat_op = get_irn_link(tmp);
1754                         if(remat_op->attr.remat.remat->value == to_copy) {
1755                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1756                         }
1757                 }
1758
1759                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1760                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1761
1762                 /* copy - reg_out - copyreg <= 0 */
1763                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1764                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1765                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1766         }
1767 }
1768
1769
1770 /**
1771  * Walk all irg blocks and emit this ILP
1772  */
1773 static void
1774 luke_blockwalker(ir_node * bb, void * data)
1775 {
1776         spill_ilp_t *si = (spill_ilp_t*)data;
1777         ir_node     *irn;
1778         pset        *live;
1779         char         buf[256];
1780         ilp_cst_t    cst;
1781         spill_bb_t  *spill_bb = get_irn_link(bb);
1782         ir_node     *tmp;
1783         spill_t     *spill;
1784         pset        *defs = pset_new_ptr_default();
1785
1786         live = pset_new_ptr_default();
1787
1788         /****************************************
1789          *      B A S I C  B L O C K  E N D
1790          ***************************************/
1791
1792
1793         /* init live values at end of block */
1794         get_live_end(si, bb, live);
1795
1796         pset_foreach(live, irn) {
1797                 op_t           *op;
1798                 ilp_var_t       reload = ILP_UNDEF;
1799
1800                 spill = set_find_spill(spill_bb->ilp, irn);
1801                 assert(spill);
1802
1803                 if(spill_bb->reloads) {
1804                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1805
1806                         if(keyval) {
1807                                 reload = PTR_TO_INT(keyval->val);
1808                         }
1809                 }
1810
1811                 op = get_irn_link(irn);
1812                 assert(!op->is_remat);
1813
1814                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1815                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1816                 op->attr.live_range.op = bb;
1817
1818                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1819                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1820
1821                 /* reg_out - reload - remat - live_range <= 0 */
1822                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1823                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1824                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1825                 foreach_pre_remat(bb, tmp) {
1826                         op_t     *remat_op = get_irn_link(tmp);
1827                         if(remat_op->attr.remat.remat->value == irn) {
1828                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1829                         }
1830                 }
1831                 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1832                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1833
1834                 /* value may only die at bb end if it is used for a mem copy */
1835                 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1836                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1837                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1838                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1839                 foreach_pre_remat(bb, tmp) {
1840                         op_t     *remat_op = get_irn_link(tmp);
1841                         if(remat_op->attr.remat.remat->value == irn) {
1842                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1843                         }
1844                 }
1845                 if(is_merge_edge(bb)) {
1846                         const ir_edge_t *edge = get_block_succ_first(bb);
1847                         const ir_node   *next_bb = edge->src;
1848                         int              pos = edge->pos;
1849                         const ir_node   *phi;
1850
1851                         sched_foreach(next_bb, phi) {
1852                                 const ir_node  *phi_arg;
1853
1854                                 if(!is_Phi(phi)) break;
1855
1856                                 phi_arg = get_irn_n(phi, pos);
1857
1858                                 if(phi_arg == irn) {
1859                                         op_t      *phi_op = get_irn_link(phi);
1860                                         ilp_var_t  copy = phi_op->attr.live_range.args.copies[pos];
1861
1862                                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1863                                 }
1864                         }
1865                 }
1866         }
1867
1868         if(opt_memcopies)
1869                 insert_mem_copy_position(si, live, bb);
1870
1871         /*
1872          * assure the remat args are available
1873          */
1874         foreach_pre_remat(bb, tmp) {
1875                 op_t     *remat_op = get_irn_link(tmp);
1876                 int       n;
1877
1878                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1879                         ir_node        *remat_arg = get_irn_n(tmp, n);
1880                         op_t           *arg_op = get_irn_link(remat_arg);
1881
1882                         if(!has_reg_class(si, remat_arg)) continue;
1883
1884                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1885                         assert(spill);
1886
1887                         /* arguments of remats have to be live until the very end of the block
1888                          * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1889                          * no remats, they could be in wrong order
1890                          */
1891
1892                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1893                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1894
1895                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1896                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1897                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1898
1899                         /* use reload placed for this argument */
1900                         if(spill_bb->reloads) {
1901                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1902
1903                                 if(keyval) {
1904                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1905
1906                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1907                                 }
1908                         }
1909                 }
1910         }
1911         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1912
1913
1914
1915
1916         /**************************************
1917          *    B A S I C  B L O C K  B O D Y
1918          **************************************/
1919
1920         sched_foreach_reverse_from(sched_block_last_noncf(bb), irn) {
1921                 op_t       *op;
1922                 op_t       *tmp_op;
1923                 int         n,
1924                                         u = 0,
1925                                         d = 0;
1926                 ilp_cst_t       check_pre,
1927                                         check_post;
1928                 set        *args;
1929                 pset       *used;
1930                 pset       *remat_defs;
1931                 keyval_t   *keyval;
1932                 ilp_cst_t   one_memoperand = -1;
1933
1934                 /* iterate only until first phi */
1935                 if(is_Phi(irn))
1936                         break;
1937
1938                 op = get_irn_link(irn);
1939                 /* skip remats */
1940                 if(op->is_remat) continue;
1941
1942                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1943
1944                 /* collect defined values */
1945                 if(has_reg_class(si, irn)) {
1946                         pset_insert_ptr(defs, irn);
1947                 }
1948
1949                 /* skip projs */
1950                 if(is_Proj(irn)) continue;
1951
1952                 /*
1953                  * init set of irn's arguments
1954                  * and all possibly used values around this op
1955                  * and values defined by post remats
1956                  */
1957                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1958                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1959                 remat_defs = pset_new_ptr(pset_count(live));
1960
1961                 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1962                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
1963                                 ir_node        *irn_arg = get_irn_n(irn, n);
1964                                 if(has_reg_class(si, irn_arg)) {
1965                                         set_insert_keyval(args, irn_arg, (void*)n);
1966                                         pset_insert_ptr(used, irn_arg);
1967                                 }
1968                         }
1969                         foreach_post_remat(irn, tmp) {
1970                                 op_t    *remat_op = get_irn_link(tmp);
1971
1972                                 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1973
1974                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1975                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1976                                         if(has_reg_class(si, remat_arg)) {
1977                                                 pset_insert_ptr(used, remat_arg);
1978                                         }
1979                                 }
1980                         }
1981                         foreach_pre_remat(irn, tmp) {
1982                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1983                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1984                                         if(has_reg_class(si, remat_arg)) {
1985                                                 pset_insert_ptr(used, remat_arg);
1986                                         }
1987                                 }
1988                         }
1989                 }
1990
1991                 /**********************************
1992                  *   I N  E P I L O G  O F  irn
1993                  **********************************/
1994
1995                 /* ensure each dying value is used by only one post remat */
1996                 pset_foreach(used, tmp) {
1997                         ir_node     *value = tmp;
1998                         op_t        *value_op = get_irn_link(value);
1999                         ir_node     *remat;
2000                         int          n_remats = 0;
2001
2002                         cst = ILP_UNDEF;
2003                         foreach_post_remat(irn, remat) {
2004                                 op_t  *remat_op = get_irn_link(remat);
2005
2006                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
2007                                         ir_node   *remat_arg = get_irn_n(remat, n);
2008
2009                                         /* if value is used by this remat add it to constraint */
2010                                         if(remat_arg == value) {
2011                                                 if(n_remats == 0) {
2012                                                         /* sum remat2s <= 1 + n_remats*live_range */
2013                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
2014                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2015                                                 }
2016
2017                                                 n_remats++;
2018                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2019                                                 break;
2020                                         }
2021                                 }
2022                         }
2023
2024                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2025                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2026                         }
2027                 }
2028
2029         /* ensure at least one value dies at post remat */
2030         foreach_post_remat(irn, tmp) {
2031             op_t     *remat_op = get_irn_link(tmp);
2032             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
2033             ir_node  *remat_arg;
2034
2035             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2036                 remat_arg = get_irn_n(tmp, n);
2037
2038                 if(has_reg_class(si, remat_arg)) {
2039
2040                     /* does arg always die at this op? */
2041                     if(!pset_find_ptr(live, remat_arg))
2042                         goto skip_one_must_die;
2043
2044                     pset_insert_ptr(remat_args, remat_arg);
2045                 }
2046             }
2047
2048             /* remat + \sum live_range(remat_arg) <= |args| */
2049             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2050             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2051             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2052
2053             pset_foreach(remat_args, remat_arg) {
2054                 op_t  *arg_op = get_irn_link(remat_arg);
2055
2056                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2057             }
2058
2059 skip_one_must_die:
2060             del_pset(remat_args);
2061         }
2062
2063                 /* new live ranges for values from L\U defined by post remats */
2064                 pset_foreach(live, tmp) {
2065                         ir_node     *value = tmp;
2066                         op_t        *value_op = get_irn_link(value);
2067
2068                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2069                                 ilp_var_t    prev_lr = ILP_UNDEF;
2070                                 ir_node     *remat;
2071
2072                                 if(pset_find_ptr(remat_defs, value)) {
2073
2074                                         /* next_live_range <= prev_live_range + sum remat2s */
2075                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2076                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2077
2078                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2079                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2080
2081                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2082                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2083
2084                                         foreach_post_remat(irn, remat) {
2085                                                 op_t        *remat_op = get_irn_link(remat);
2086
2087                                                 /* if value is being rematerialized by this remat */
2088                                                 if(value == remat_op->attr.remat.remat->value) {
2089                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2090                                                 }
2091                                         }
2092
2093                                         value_op->attr.live_range.ilp = prev_lr;
2094                                         value_op->attr.live_range.op = irn;
2095                                 }
2096                         }
2097                 }
2098
2099                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2100                 foreach_post_remat(irn, tmp) {
2101                         op_t        *remat_op = get_irn_link(tmp);
2102                         int          n;
2103
2104                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2105                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2106                                 op_t           *arg_op = get_irn_link(remat_arg);
2107
2108                                 if(!has_reg_class(si, remat_arg)) continue;
2109
2110                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2111                                 if(!pset_find_ptr(used, remat_arg)) {
2112                                         /* remat <= live_range(remat_arg) */
2113                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2114                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2115
2116                                         /* if value is becoming live through use by remat2 */
2117                                         if(!pset_find_ptr(live, remat_arg)) {
2118                                                 ilp_var_t     lr;
2119
2120                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2121                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2122
2123                                                 arg_op->attr.live_range.ilp = lr;
2124                                                 arg_op->attr.live_range.op = irn;
2125
2126                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2127
2128                                                 pset_insert_ptr(live, remat_arg);
2129                                                 add_to_spill_bb(si, bb, remat_arg);
2130                                         }
2131
2132                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2133                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2134                                 }
2135                         }
2136                 }
2137
2138                 d = pset_count(defs);
2139                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2140
2141                 /* count how many regs irn needs for arguments */
2142                 u = set_count(args);
2143
2144
2145                 /* check the register pressure in the epilog */
2146                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2147                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2148                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2149
2150                 /* add L\U' to check_post */
2151                 pset_foreach(live, tmp) {
2152                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2153                                 /* if a live value is not used by irn */
2154                                 tmp_op = get_irn_link(tmp);
2155                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2156                         }
2157                 }
2158
2159                 /***********************************************************
2160                  *  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
2161                  **********************************************************/
2162
2163
2164                 pset_foreach(used, tmp) {
2165                         ilp_var_t       prev_lr;
2166                         ilp_var_t       post_use;
2167                         int             p = 0;
2168                         spill_t        *spill;
2169                         ir_node        *arg = tmp;
2170                         op_t           *arg_op = get_irn_link(arg);
2171                         ir_node        *remat;
2172
2173                         spill = add_to_spill_bb(si, bb, arg);
2174
2175                         /* new live range for each used value */
2176                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2177                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2178
2179                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2180                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2181                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2182
2183                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2184
2185                         /* arg is live throughout epilog if the next live_range is in a register */
2186                         if(pset_find_ptr(live, arg)) {
2187                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2188
2189                                 /* post_use >= next_lr + remat */
2190                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2191                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2192                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2193                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2194                         }
2195
2196                         /* forall post remat which use arg add a similar cst */
2197                         foreach_post_remat(irn, remat) {
2198                                 int      n;
2199
2200                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2201                                         ir_node    *remat_arg = get_irn_n(remat, n);
2202                                         op_t       *remat_op = get_irn_link(remat);
2203
2204                                         if(remat_arg == arg) {
2205                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2206
2207                                                 /* post_use >= remat */
2208                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2209                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2210                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2211                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2212                                         }
2213                                 }
2214                         }
2215
2216                         /* if value is not an arg of op and not possibly defined by post remat
2217                          * then it may only die and not become live
2218                          */
2219                         if(!set_find_keyval(args, arg)) {
2220                                 /* post_use <= prev_lr */
2221                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2222                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2223                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2224                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2225
2226                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2227                                         /* next_lr <= prev_lr */
2228                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2229                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2230                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2231                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2232                                 }
2233                         }
2234
2235                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2236                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2237                                         if (get_irn_n(irn, n) == arg &&
2238                                                         arch_possible_memory_operand(irn, n)) {
2239                                                 ilp_var_t       memoperand;
2240
2241                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2242                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2243                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2244
2245                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2246                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2247
2248                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2249                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2250                                         }
2251                                 }
2252                         }
2253
2254                         /* new live range begins for each used value */
2255                         arg_op->attr.live_range.ilp = prev_lr;
2256                         arg_op->attr.live_range.op = irn;
2257
2258                         pset_insert_ptr(live, arg);
2259                 }
2260
2261                 /* just to be sure */
2262                 check_post = ILP_UNDEF;
2263
2264                 /* allow original defintions to be removed */
2265                 if(opt_repair_schedule) {
2266                         pset_foreach(defs, tmp) {
2267                                 op_t      *tmp_op = get_irn_link(tmp);
2268                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2269 #if 1
2270                                 ilp_var_t  delete;
2271                                 assert(spill);
2272
2273                                 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2274                                 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0 * get_cost(irn) * execution_frequency(si, bb), 0.0);
2275
2276                                 /* op may not be killed if its first live_range is 1 */
2277                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2278                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2279                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2280                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2281
2282                                 /* op may not be killed if it is spilled after the definition */
2283                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2284                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2285                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2286                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2287 #else
2288                                 ilp_var_t  keep;
2289                                 assert(spill);
2290
2291                                 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2292                                 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(irn) * execution_frequency(si, bb), 1.0);
2293
2294                                 /* op may not be killed if its first live_range is 1 */
2295                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2296                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2297                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2298                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2299
2300                                 /* op may not be killed if it is spilled after the definition */
2301                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2302                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2303                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2304                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2305 #endif
2306                         }
2307                 } else {
2308 #if 0
2309                         pset_foreach(defs, tmp) {
2310                                 op_t      *tmp_op = get_irn_link(tmp);
2311                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2312                                 assert(spill);
2313
2314                                 /* live_range or spill should be 1
2315                                    TODO: lr should be live until first use */
2316                                 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2317                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2318                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2319                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2320                         }
2321 #endif
2322                 }
2323
2324
2325                 /******************
2326                  *   P R O L O G
2327                  ******************/
2328
2329                 /* check the register pressure in the prolog */
2330                 /* sum_{L\U} lr <= k - |U| */
2331                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2332                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2333
2334                 /* for the prolog remove defined values from the live set */
2335                 pset_foreach(defs, tmp) {
2336                         pset_remove_ptr(live, tmp);
2337                 }
2338
2339                 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2340                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2341                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2342                 }
2343
2344                 /***********************************************************
2345                  *  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
2346                  **********************************************************/
2347
2348
2349                 set_foreach(args, keyval) {
2350                         spill_t          *spill;
2351                         const ir_node    *arg = keyval->key;
2352                         int               i = PTR_TO_INT(keyval->val);
2353                         op_t             *arg_op = get_irn_link(arg);
2354                         ilp_cst_t         requirements;
2355                         int               n_memoperands;
2356
2357                         spill = set_find_spill(spill_bb->ilp, arg);
2358                         assert(spill);
2359
2360                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2361                         op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2362
2363                         /* reload <= mem_out */
2364                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2365                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2366                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2367                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2368
2369                         /* requirement: arg must be in register for use */
2370                         /* reload + remat + live_range == 1 */
2371                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2372                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2373
2374                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2375                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2376                         foreach_pre_remat(irn, tmp) {
2377                                 op_t     *remat_op = get_irn_link(tmp);
2378                                 if(remat_op->attr.remat.remat->value == arg) {
2379                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2380                                 }
2381                         }
2382
2383                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2384                                 n_memoperands = 0;
2385                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2386                                         if(get_irn_n(irn, n) == arg) {
2387                                                 n_memoperands++;
2388                                         }
2389                                 }
2390                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2391                                         if (get_irn_n(irn, n) == arg &&
2392                                                         arch_possible_memory_operand(irn, n)) {
2393                                                 memoperand_t  *memoperand;
2394                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2395
2396                                                 /* memoperand <= mem_out */
2397                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2398                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2399                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2400                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2401
2402                                                 /* the memoperand is only sufficient if it is used once by the op */
2403                                                 if(n_memoperands == 1)
2404                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2405
2406                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2407
2408                                                 /* we have one more free register if we use a memory operand */
2409                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2410                                         }
2411                                 }
2412                         }
2413                 }
2414
2415                 /* iterate over L\U */
2416                 pset_foreach(live, tmp) {
2417                         if(!set_find_keyval(args, tmp)) {
2418                                 /* if a live value is not used by irn */
2419                                 tmp_op = get_irn_link(tmp);
2420                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2421                         }
2422                 }
2423
2424                 /* requirements for remats */
2425                 foreach_pre_remat(irn, tmp) {
2426                         op_t        *remat_op = get_irn_link(tmp);
2427                         int          n;
2428
2429                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2430                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2431                                 op_t           *arg_op = get_irn_link(remat_arg);
2432
2433                                 if(!has_reg_class(si, remat_arg)) continue;
2434
2435                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2436                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2437                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2438
2439                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2440                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2441
2442                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2443                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2444                                         int    index = (int)keyval->val;
2445
2446                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2447                                 }
2448                         }
2449                 }
2450
2451
2452
2453
2454                 /*************************
2455                  *  D O N E  W I T H  O P
2456                  *************************/
2457
2458                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2459
2460                 pset_foreach(live, tmp) {
2461                         assert(has_reg_class(si, tmp));
2462                 }
2463
2464 #ifndef NDEBUG
2465                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2466                         ir_node        *arg = get_irn_n(irn, n);
2467
2468                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2469                 }
2470 #endif
2471
2472                 del_pset(remat_defs);
2473                 del_pset(used);
2474                 del_set(args);
2475                 del_pset(defs);
2476                 defs = pset_new_ptr_default();
2477
2478                 /* skip everything above barrier in start block */
2479                 if(is_start_block(bb) && be_is_Barrier(irn)) {
2480                         assert(pset_count(live) == 0);
2481                         break;
2482                 }
2483
2484         }
2485         del_pset(defs);
2486
2487
2488
2489         /***************************************
2490          *   B E G I N N I N G  O F  B L O C K
2491          ***************************************/
2492
2493
2494         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2495         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2496
2497         pset_foreach(live, irn) {
2498                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2499         }
2500
2501         /* construct mem_outs for all values */
2502         set_foreach(spill_bb->ilp, spill) {
2503                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2504                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2505
2506                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2507                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2508
2509                 if(pset_find_ptr(live, spill->irn)) {
2510                         int default_spilled;
2511                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2512
2513                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2514                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2515                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2516                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2517
2518                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2519                                 int   n;
2520                                 op_t *op = get_irn_link(spill->irn);
2521
2522                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2523                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2524                                         double          freq=0.0;
2525                                         int             m;
2526                                         ilp_var_t       var;
2527
2528
2529                                         /* argument already done? */
2530                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2531
2532                                         /* get sum of execution frequencies of blocks with the same phi argument */
2533                                         for(m=n; m>=0; --m) {
2534                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2535
2536                                                 if(arg==arg2) {
2537                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2538                                                 }
2539                                         }
2540
2541                                         /* copies are not for free */
2542                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2543                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2544
2545                                         for(m=n; m>=0; --m) {
2546                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2547
2548                                                 if(arg==arg2) {
2549                                                         op->attr.live_range.args.copies[m] = var;
2550                                                 }
2551                                         }
2552
2553 #if 0
2554                                         /* copy <= mem_in */
2555                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2556                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2557                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2558                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2559 #endif
2560                                 }
2561                         }
2562                 }
2563         }
2564
2565         foreach_post_remat(bb, tmp) {
2566                 int         n;
2567                 op_t       *remat_op = get_irn_link(tmp);
2568                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2569                 ir_node    *remat_arg;
2570
2571                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2572                         remat_arg = get_irn_n(tmp, n);
2573
2574                         if(has_reg_class(si, remat_arg)) {
2575                                 pset_insert_ptr(remat_args, remat_arg);
2576                         }
2577                 }
2578
2579                 /* remat + \sum live_range(remat_arg) <= |args| */
2580                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2581                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2582                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2583
2584                 pset_foreach(remat_args, remat_arg) {
2585                         if(pset_find_ptr(live, remat_arg)) {
2586                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2587                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2588                         }
2589                 }
2590                 del_pset(remat_args);
2591         }
2592
2593         foreach_post_remat(bb, tmp) {
2594                 int  n;
2595
2596                 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2597                         ir_node  *remat_arg = get_irn_n(tmp, n);
2598
2599                         /* if value is becoming live through use by remat2 */
2600                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2601                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2602                                 ilp_cst_t   nomem;
2603
2604                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2605
2606                                 pset_insert_ptr(live, remat_arg);
2607                                 spill = add_to_spill_bb(si, bb, remat_arg);
2608                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2609
2610                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2611                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2612                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2613                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2614                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2615
2616
2617                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2618                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2619                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2620                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2621                         }
2622                 }
2623         }
2624
2625         /* L\U is empty at bb start */
2626         /* arg is live throughout epilog if it is reg_in into this block */
2627
2628         /* check the register pressure at the beginning of the block
2629          * including remats
2630          */
2631         /* reg_in entspricht post_use */
2632
2633         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2634         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2635
2636         pset_foreach(live, irn) {
2637         ilp_cst_t  nospill;
2638
2639                 spill = set_find_spill(spill_bb->ilp, irn);
2640                 assert(spill);
2641
2642                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2643                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2644
2645                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2646
2647                 /* spill + mem_in <= 1 */
2648                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2649                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2650
2651                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2652                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2653
2654         } /* post_remats are NOT included in register pressure check because
2655            they do not increase regpressure */
2656
2657         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2658         pset_foreach(live, irn) {
2659                 int          p = 0,
2660                                          n;
2661
2662                 spill = set_find_spill(spill_bb->ilp, irn);
2663                 assert(spill && spill->irn == irn);
2664
2665                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2666                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2667                                 ilp_cst_t       mem_in,
2668                                                                 reg_in;
2669                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2670                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2671                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2672                                 spill_t        *spill_p;
2673                                 op_t           *op = get_irn_link(irn);
2674
2675                                 /* although the phi is in the right regclass one or more of
2676                                  * its arguments can be in a different one or at least to
2677                                  * ignore
2678                                  */
2679                                 if(has_reg_class(si, phi_arg)) {
2680                                         /* mem_in < mem_out_arg + copy */
2681                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2682                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2683
2684                                         /* reg_in < reg_out_arg */
2685                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2686                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2687
2688                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2689                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2690
2691                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2692                                         assert(spill_p);
2693
2694                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2695                                         if(opt_memcopies)
2696                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2697
2698                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2699                                 }
2700                         }
2701                 } else {
2702                         /* else assure the value arrives on all paths in the same resource */
2703
2704                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2705                                 ilp_cst_t       mem_in,
2706                                                                 reg_in;
2707                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2708                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2709                                 spill_t        *spill_p;
2710
2711                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2712                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2713                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2714                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2715
2716                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2717                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2718
2719                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2720                                 assert(spill_p);
2721
2722                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2723                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2724                         }
2725                 }
2726         }
2727
2728         foreach_post_remat(bb, tmp) {
2729                 int         n;
2730
2731                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2732                         ir_node    *remat_arg = get_irn_n(tmp, n);
2733                         op_t       *remat_op = get_irn_link(tmp);
2734
2735                         if(!has_reg_class(si, remat_arg)) continue;
2736
2737                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2738                         assert(spill);
2739
2740                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2741                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2742                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2743                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2744                 }
2745         }
2746
2747         pset_foreach(live, irn) {
2748                 const op_t      *op = get_irn_link(irn);
2749                 const ir_node   *remat;
2750                 int              n_remats = 0;
2751
2752                 cst = ILP_UNDEF;
2753
2754                 foreach_post_remat(bb, remat) {
2755                         int   n;
2756
2757                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2758                                 const ir_node  *arg = get_irn_n(remat, n);
2759
2760                                 if(arg == irn) {
2761                                         const op_t   *remat_op = get_irn_link(remat);
2762
2763                                         if(cst == ILP_UNDEF) {
2764                                                 /* sum remat2s <= 1 + n_remats*live_range */
2765                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2766                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2767                                         }
2768                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2769                                         ++n_remats;
2770                                         break;
2771                                 }
2772                         }
2773                 }
2774                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2775                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2776                 }
2777         }
2778
2779         /* first live ranges from reg_ins */
2780         pset_foreach(live, irn) {
2781                 op_t      *op = get_irn_link(irn);
2782
2783                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2784
2785                         spill = set_find_spill(spill_bb->ilp, irn);
2786                         assert(spill && spill->irn == irn);
2787
2788                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2789                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2790                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2791                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2792
2793                         foreach_post_remat(bb, tmp) {
2794                                 op_t     *remat_op = get_irn_link(tmp);
2795
2796                                 if(remat_op->attr.remat.remat->value == irn) {
2797                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2798                                 }
2799                         }
2800                 }
2801         }
2802
2803         /* walk forward now and compute constraints for placing spills */
2804         /* this must only be done for values that are not defined in this block */
2805         pset_foreach(live, irn) {
2806                 /*
2807                  * if value is defined in this block we can anways place the spill directly after the def
2808                  *    -> no constraint necessary
2809                  */
2810                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2811                         assert(0);
2812                 }
2813
2814
2815                 spill = set_find_spill(spill_bb->ilp, irn);
2816                 assert(spill);
2817
2818                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2819                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2820
2821                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2822                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2823
2824                 if(!is_Phi(irn)) {
2825                         sched_foreach_op(bb, tmp) {
2826                                 op_t   *op = get_irn_link(tmp);
2827
2828                                 if(is_Phi(tmp)) continue;
2829                                 assert(!is_Proj(tmp));
2830
2831                                 if(op->is_remat) {
2832                                         const ir_node   *value = op->attr.remat.remat->value;
2833
2834                                         if(value == irn) {
2835                                                 /* only collect remats up to the first real use of a value */
2836                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2837                                         }
2838                                 } else {
2839                                         int   n;
2840
2841                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2842                                                 ir_node    *arg = get_irn_n(tmp, n);
2843
2844                                                 if(arg == irn) {
2845                                                         /* if a value is used stop collecting remats */
2846                             goto next_live;
2847                                                 }
2848                                         }
2849                                 }
2850                         }
2851                 }
2852 next_live: ;
2853         }
2854
2855         del_pset(live);
2856 }
2857
2858 typedef struct _irnlist_t {
2859         struct list_head   list;
2860         ir_node           *irn;
2861 } irnlist_t;
2862
2863 typedef struct _interference_t {
2864         struct list_head    blocklist;
2865         ir_node            *a;
2866         ir_node            *b;
2867 } interference_t;
2868
2869 static int
2870 cmp_interference(const void *a, const void *b, size_t size)
2871 {
2872         const interference_t *p = a;
2873         const interference_t *q = b;
2874         (void) size;
2875
2876         return !(p->a == q->a && p->b == q->b);
2877 }
2878
2879 static interference_t *
2880 set_find_interference(set * set, ir_node * a, ir_node * b)
2881 {
2882         interference_t     query;
2883
2884         query.a = (a>b)?a:b;
2885         query.b = (a>b)?b:a;
2886
2887         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2888 }
2889
2890 static interference_t *
2891 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2892 {
2893         interference_t     query,
2894                                           *result;
2895         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2896
2897         list->irn = bb;
2898
2899         result = set_find_interference(set, a, b);
2900         if(result) {
2901
2902                 list_add(&list->list, &result->blocklist);
2903                 return result;
2904         }
2905
2906         query.a = (a>b)?a:b;
2907         query.b = (a>b)?b:a;
2908
2909         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2910
2911         INIT_LIST_HEAD(&result->blocklist);
2912         list_add(&list->list, &result->blocklist);
2913
2914         return result;
2915 }
2916
2917 static
2918 int values_interfere_in_block(const spill_ilp_t *si, const ir_node *bb, const ir_node *a, const ir_node *b)
2919 {
2920         const ir_edge_t *edge;
2921
2922         if (get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2923                 /* both values are live in, so they interfere */
2924                 return 1;
2925         }
2926
2927         /* ensure a dominates b */
2928         if (value_dominates(b, a)) {
2929                 const ir_node *t;
2930                 t = b;
2931                 b = a;
2932                 a = t;
2933         }
2934         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2935
2936
2937         /* the following code is stolen from bera.c */
2938         if (be_is_live_end(si->lv, bb, a))
2939                 return 1;
2940
2941         foreach_out_edge(a, edge) {
2942                 const ir_node *user = edge->src;
2943                 if (get_nodes_block(user) == bb
2944                                 && ! is_Phi(user)
2945                                 && b != user
2946                                 && ! pset_find_ptr(si->inverse_ops, user)
2947                                 && value_dominates(b, user))
2948                         return 1;
2949         }
2950
2951         return 0;
2952 }
2953
2954 /**
2955  * Walk all irg blocks and collect interfering values inside of phi classes
2956  */
2957 static void
2958 luke_interferencewalker(ir_node * bb, void * data)
2959 {
2960         spill_ilp_t    *si = (spill_ilp_t*)data;
2961         int             l1, l2;
2962
2963         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2964                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2965                 op_t           *a_op = get_irn_link(a);
2966
2967
2968                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2969                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2970                         if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2971                                 continue;
2972
2973                         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)) {
2974                                 ir_node *b    = be_lv_get_irn(si->lv, bb, l2);
2975                                 op_t    *b_op = get_irn_link(b);
2976
2977                                 /* a and b are only interesting if they are in the same phi class */
2978                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2979                                         if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2980                                                 continue;
2981
2982                                         if (values_interfere_in_block(si, bb, a, b)) {
2983                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2984                                                 set_insert_interference(si, si->interferences, a, b, bb);
2985                                         }
2986                                 }
2987                         }
2988                 }
2989         }
2990 }
2991
2992 static unsigned int copy_path_id = 0;
2993
2994 static void
2995 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2996 {
2997         ilp_cst_t  cst;
2998         ilp_var_t  copy;
2999         char       buf[256];
3000         void      *ptr;
3001
3002         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
3003         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3004
3005         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
3006
3007         pset_foreach(copies, ptr) {
3008                 copy = PTR_TO_INT(ptr);
3009                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
3010         }
3011 }
3012
3013 /**
3014  * @parameter copies   contains a path of copies which lead us to irn
3015  * @parameter visited  contains a set of nodes already visited on this path
3016  */
3017 static int
3018 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3019 {
3020         const ir_edge_t *edge;
3021         op_t            *op = get_irn_link(irn);
3022     pset            *visited_users = pset_new_ptr_default();
3023         int              paths = 0;
3024
3025         if(op->is_remat) return 0;
3026
3027         pset_insert_ptr(visited, irn);
3028
3029         if(is_Phi(irn)) {
3030                 int    n;
3031         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
3032
3033                 /* visit all operands */
3034                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3035                         ir_node  *arg = get_irn_n(irn, n);
3036                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
3037
3038                         if(!has_reg_class(si, arg)) continue;
3039             if(pset_find_ptr(visited_operands, arg)) continue;
3040             pset_insert_ptr(visited_operands, arg);
3041
3042                         if(arg == target) {
3043                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3044                                         del_pset(visited_operands);
3045                                         del_pset(visited_users);
3046                                         pset_remove_ptr(visited, irn);
3047                                         return paths;
3048                                 }
3049                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3050                                 write_copy_path_cst(si, copies, any_interfere);
3051                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3052                         } else if(!pset_find_ptr(visited, arg)) {
3053                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3054                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3055                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3056
3057                 if(paths > MAX_PATHS) {
3058                     if(pset_count(copies) == 0) {
3059                         ilp_cst_t  cst;
3060                         char       buf[256];
3061
3062                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3063                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3064                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3065                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3066                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3067
3068                         paths = 0;
3069                     } else {
3070                         del_pset(visited_operands);
3071                         del_pset(visited_users);
3072                         pset_remove_ptr(visited, irn);
3073                         return paths;
3074                     }
3075                 } else if(pset_count(copies) == 0) {
3076                                         paths = 0;
3077                                 }
3078                         }
3079                 }
3080
3081         del_pset(visited_operands);
3082         }
3083
3084         /* visit all uses which are phis */
3085         foreach_out_edge(irn, edge) {
3086                 ir_node  *user = edge->src;
3087                 int       pos  = edge->pos;
3088                 op_t     *op = get_irn_link(user);
3089                 ilp_var_t copy;
3090
3091                 if(!is_Phi(user)) continue;
3092                 if(!has_reg_class(si, user)) continue;
3093         if(pset_find_ptr(visited_users, user)) continue;
3094         pset_insert_ptr(visited_users, user);
3095
3096                 copy = op->attr.live_range.args.copies[pos];
3097
3098                 if(user == target) {
3099                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3100                                 del_pset(visited_users);
3101                                 pset_remove_ptr(visited, irn);
3102                                 return paths;
3103                         }
3104                         pset_insert(copies, INT_TO_PTR(copy), copy);
3105                         write_copy_path_cst(si, copies, any_interfere);
3106                         pset_remove(copies, INT_TO_PTR(copy), copy);
3107                 } else if(!pset_find_ptr(visited, user)) {
3108                         pset_insert(copies, INT_TO_PTR(copy), copy);
3109                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3110                         pset_remove(copies, INT_TO_PTR(copy), copy);
3111
3112             if(paths > MAX_PATHS) {
3113                 if(pset_count(copies) == 0) {
3114                     ilp_cst_t  cst;
3115                     char       buf[256];
3116
3117                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3118                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3119                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3120                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3121                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3122
3123                     paths = 0;
3124                 } else {
3125                     del_pset(visited_users);
3126                     pset_remove_ptr(visited, irn);
3127                     return paths;
3128                 }
3129             } else if(pset_count(copies) == 0) {
3130                                 paths = 0;
3131                         }
3132                 }
3133         }
3134
3135     del_pset(visited_users);
3136         pset_remove_ptr(visited, irn);
3137         return paths;
3138 }
3139
3140 static void
3141 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3142 {
3143         pset * copies = pset_new_ptr_default();
3144         pset * visited = pset_new_ptr_default();
3145
3146         find_copy_path(si, a, b, any_interfere, copies, visited);
3147
3148         del_pset(visited);
3149         del_pset(copies);
3150 }
3151
3152
3153 static void
3154 memcopyhandler(spill_ilp_t * si)
3155 {
3156         interference_t   *interference;
3157         char              buf[256];
3158         /* teste Speicherwerte auf Interferenz */
3159
3160         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3161         irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3162
3163         /* now lets emit the ILP unequations for the crap */
3164         set_foreach(si->interferences, interference) {
3165                 irnlist_t      *irnlist;
3166                 ilp_var_t      interfere, any_interfere;
3167                 ilp_cst_t      any_interfere_cst, cst;
3168                 const ir_node  *a  = interference->a;
3169                 const ir_node  *b  = interference->b;
3170
3171                 /* any_interf <= \sum interf */
3172                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3173                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3174                 any_interfere     = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3175
3176                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3177
3178                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3179                         const ir_node  *bb = irnlist->irn;
3180                         spill_bb_t     *spill_bb = get_irn_link(bb);
3181                         spill_t        *spilla,
3182                                                    *spillb;
3183                         char           buf[256];
3184
3185                         spilla = set_find_spill(spill_bb->ilp, a);
3186                         assert(spilla);
3187
3188                         spillb = set_find_spill(spill_bb->ilp, b);
3189                         assert(spillb);
3190
3191                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3192                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3193                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3194                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3195                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3196                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3197
3198                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3199                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3200
3201                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3202                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3203                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3204                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3205                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3206
3207                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3208                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3209
3210                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3211                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3212                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3213
3214                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3215                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3216
3217                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3218                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3219                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3220
3221
3222                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3223
3224                         /* any_interfere >= interf */
3225                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3226                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3227
3228                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3229                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3230                 }
3231
3232                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3233                 gen_copy_constraints(si,a,b,any_interfere);
3234         }
3235 }
3236
3237
3238 static INLINE int
3239 is_zero(double x)
3240 {
3241         return fabs(x) < 0.00001;
3242 }
3243
3244 /**
3245  * node attribute hook for changing colors
3246  */
3247 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3248 {
3249         spill_ilp_t *si = get_irg_link(current_ir_graph);
3250         (void) l;
3251
3252         if(pset_find_ptr(si->all_possible_remats, n)) {
3253                 op_t   *op = (op_t*)get_irn_link(n);
3254                 assert(op && op->is_remat);
3255
3256                 if(!op->attr.remat.remat->inverse) {
3257                         if(op->attr.remat.pre) {
3258                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3259                         } else {
3260                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3261                         }
3262
3263                         return 1;
3264                 } else {
3265                         op_t   *op = (op_t*)get_irn_link(n);
3266                         assert(op && op->is_remat);
3267
3268                         if(op->attr.remat.pre) {
3269                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3270                         } else {
3271                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3272                         }
3273
3274                         return 1;
3275                 }
3276         }
3277
3278         return 0;
3279 }
3280
3281 static void
3282 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3283 {
3284         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3285         be_dump(irg, suffix, dump_ir_block_graph_sched);
3286         set_dump_node_vcgattr_hook(NULL);
3287 }
3288
3289 /**
3290  * Edge hook to dump the schedule edges with annotated register pressure.
3291  */
3292 static int
3293 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3294 {
3295         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3296                 ir_node *prev = sched_prev(irn);
3297                 fprintf(F, "edge:{sourcename:\"");
3298                 PRINT_NODEID(irn);
3299                 fprintf(F, "\" targetname:\"");
3300                 PRINT_NODEID(prev);
3301                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3302                 fprintf(F, "\" color:magenta}\n");
3303         }
3304         return 1;
3305 }
3306
3307 static void
3308 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3309 {
3310         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3311
3312         dump_consts_local(0);
3313         set_dump_node_edge_hook(sched_pressure_edge_hook);
3314         dump_ir_block_graph(irg, suffix);
3315         set_dump_node_edge_hook(old_edge_hook);
3316 }
3317
3318 static void
3319 walker_pressure_annotator(ir_node * bb, void * data)
3320 {
3321         spill_ilp_t  *si = data;
3322         ir_node      *irn;
3323         int           n, i;
3324         pset         *live = pset_new_ptr_default();
3325         int           projs = 0;
3326
3327         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3328                 irn = be_lv_get_irn(si->lv, bb, i);
3329
3330                 if (has_reg_class(si, irn)) {
3331                         pset_insert_ptr(live, irn);
3332                 }
3333         }
3334
3335         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3336
3337         sched_foreach_reverse(bb, irn) {
3338                 if(is_Phi(irn)) {
3339                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3340                         continue;
3341                 }
3342
3343                 if(has_reg_class(si, irn)) {
3344                         pset_remove_ptr(live, irn);
3345                         if(is_Proj(irn)) ++projs;
3346                 }
3347
3348                 if(!is_Proj(irn)) projs = 0;
3349
3350                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3351                         ir_node    *arg = get_irn_n(irn, n);
3352
3353                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3354                 }
3355                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3356         }
3357
3358         del_pset(live);
3359 }
3360
3361 static void
3362 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3363 {
3364         be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3365 }
3366
3367 static void
3368 connect_all_remats_with_keep(spill_ilp_t * si)
3369 {
3370         ir_node   *irn;
3371         ir_node  **ins,
3372                          **pos;
3373         int        n_remats;
3374
3375
3376         n_remats = pset_count(si->all_possible_remats);
3377         if(n_remats) {
3378                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3379
3380                 pos = ins;
3381                 pset_foreach(si->all_possible_remats, irn) {
3382                         *pos = irn;
3383                         ++pos;
3384                 }
3385
3386                 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3387
3388                 obstack_free(si->obst, ins);
3389         }
3390 }
3391
3392 static void
3393 connect_all_spills_with_keep(spill_ilp_t * si)
3394 {
3395         ir_node   *irn;
3396         ir_node  **ins,
3397                          **pos;
3398         int        n_spills;
3399         ir_node   *keep;
3400
3401
3402         n_spills = pset_count(si->spills);
3403         if(n_spills) {
3404                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3405
3406                 pos = ins;
3407                 pset_foreach(si->spills, irn) {
3408                         *pos = irn;
3409                         ++pos;
3410                 }
3411
3412                 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3413
3414                 obstack_free(si->obst, ins);
3415         }
3416 }
3417
3418 /** insert a spill at an arbitrary position */
3419 static ir_node *be_spill2(ir_node *irn, ir_node *insert)
3420 {
3421         ir_node  *bl    = is_Block(insert) ? insert : get_nodes_block(insert);
3422         ir_graph *irg   = get_irn_irg(bl);
3423         ir_node  *frame = get_irg_frame(irg);
3424         ir_node  *spill;
3425         ir_node  *next;
3426         const arch_register_class_t *cls       = arch_get_irn_reg_class(irn, -1);
3427         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame, -1);
3428
3429         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3430
3431         /*
3432          * search the right insertion point. a spill of a phi cannot be put
3433          * directly after the phi, if there are some phis behind the one which
3434          * is spilled. Also, a spill of a Proj must be after all Projs of the
3435          * same tuple node.
3436          *
3437          * Here's one special case:
3438          * If the spill is in the start block, the spill must be after the frame
3439          * pointer is set up. This is done by setting insert to the end of the block
3440          * which is its default initialization (see above).
3441          */
3442
3443         if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3444                 insert = frame;
3445
3446         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3447                 insert = next;
3448
3449         sched_add_after(insert, spill);
3450         return spill;
3451 }
3452
3453 static void
3454 delete_remat(spill_ilp_t * si, ir_node * remat) {
3455         int       n;
3456         ir_node  *bad = get_irg_bad(si->birg->irg);
3457
3458         sched_remove(remat);
3459
3460         /* kill links to operands */
3461         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3462                 set_irn_n(remat, n, bad);
3463         }
3464 }
3465
3466 static void
3467 clean_remat_info(spill_ilp_t * si)
3468 {
3469         int            n;
3470         remat_t       *remat;
3471         remat_info_t  *remat_info;
3472         ir_node       *bad = get_irg_bad(si->birg->irg);
3473
3474         set_foreach(si->remat_info, remat_info) {
3475                 if(!remat_info->remats) continue;
3476
3477                 pset_foreach(remat_info->remats, remat)
3478                 {
3479                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3480                                 if(sched_is_scheduled(remat->proj)) {
3481                                         sched_remove((ir_node*)remat->proj);
3482                                 }
3483                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3484                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3485                         }
3486
3487                         if(get_irn_n_edges(remat->op) == 0) {
3488                                 if(sched_is_scheduled(remat->op)) {
3489                                         sched_remove((ir_node*)remat->op);
3490                                 }
3491                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3492                                         set_irn_n((ir_node*)remat->op, n, bad);
3493                                 }
3494                         }
3495                 }
3496
3497                 if(remat_info->remats) del_pset(remat_info->remats);
3498                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3499         }
3500 }
3501
3502 static void
3503 delete_unnecessary_remats(spill_ilp_t * si)
3504 {
3505         if(opt_keep_alive & KEEPALIVE_REMATS) {
3506                 int       n;
3507                 ir_node  *bad = get_irg_bad(si->birg->irg);
3508
3509                 if(si->keep) {
3510                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3511                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3512                                 op_t           *arg_op = get_irn_link(keep_arg);
3513                                 lpp_name_t     *name;
3514
3515                                 assert(arg_op->is_remat);
3516
3517                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3518
3519                                 if(is_zero(name->value)) {
3520                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3521                                         /* TODO check whether reload is preferred over remat (could be bug) */
3522                                         delete_remat(si, keep_arg);
3523                                 } else {
3524                                         if(!arg_op->attr.remat.remat->inverse) {
3525                                                 if(arg_op->attr.remat.pre) {
3526                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3527                                                 } else {
3528                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3529                                                 }
3530                                         } else {
3531                                                 if(arg_op->attr.remat.pre) {
3532                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3533                                                 } else {
3534                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3535                                                 }
3536                                         }
3537                                 }
3538
3539                                 set_irn_n(si->keep, n, bad);
3540                         }
3541                 } else {
3542                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3543                 }
3544         } else {
3545                 ir_node  *remat;
3546
3547                 pset_foreach(si->all_possible_remats, remat) {
3548                         op_t           *remat_op = get_irn_link(remat);
3549                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3550
3551                         if(is_zero(name->value)) {
3552                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3553                                 /* TODO check whether reload is preferred over remat (could be bug) */
3554                                 delete_remat(si, remat);
3555                         } else {
3556                                 if(!remat_op->attr.remat.remat->inverse) {
3557                                         if(remat_op->attr.remat.pre) {
3558                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3559                                         } else {
3560                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3561                                         }
3562                                 } else {
3563                                         if(remat_op->attr.remat.pre) {
3564                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3565                                         } else {
3566                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3567                                         }
3568                                 }
3569                         }
3570                 }
3571         }
3572 }
3573
3574 static pset *
3575 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3576 {
3577         pset     *spills = pset_new_ptr_default();
3578
3579         const ir_node  *next;
3580         defs_t         *defs;
3581
3582         defs = set_find_def(si->values, value);
3583
3584         if(defs && defs->spills) {
3585                 for(next = defs->spills; next; next = get_irn_link(next)) {
3586                         pset_insert_ptr(spills, next);
3587                 }
3588         }
3589
3590         return spills;
3591 }
3592
3593 static ir_node *
3594 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3595 {
3596         ir_node  *res;
3597
3598         assert( get_irn_arity(block) == arity );
3599
3600         res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3601         res->attr.phi.u.backedge = new_backedge_arr(irg->obst, arity);
3602
3603         return res;
3604 }
3605
3606 /**
3607  * @param before   The node after which the spill will be placed in the schedule
3608  */
3609 static ir_node *
3610 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3611 {
3612         defs_t   *defs;
3613         ir_node  *spill;
3614
3615         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3616
3617         spill = be_spill2(irn, before);
3618
3619         defs = set_insert_def(si->values, value);
3620         assert(defs);
3621
3622         /* enter into the linked list */
3623         set_irn_link(spill, defs->spills);
3624         defs->spills = spill;
3625
3626         if(opt_keep_alive & KEEPALIVE_SPILLS)
3627                 pset_insert_ptr(si->spills, spill);
3628
3629         return spill;
3630 }
3631
3632 /**
3633  * @param before   The Phi node which has to be spilled
3634  */
3635 static ir_node *
3636 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3637 {
3638         ir_node   *mem_phi;
3639         ir_node  **ins;
3640         defs_t    *defs;
3641         int        n;
3642
3643         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3644
3645         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3646                 ins[n] = si->m_unknown;
3647         }
3648
3649         mem_phi =  new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3650
3651         defs = set_insert_def(si->values, phi);
3652         assert(defs);
3653
3654         /* enter into the linked list */
3655         set_irn_link(mem_phi, defs->spills);
3656         defs->spills = mem_phi;
3657
3658 #ifdef SCHEDULE_PHIM
3659         sched_add_after(phi, mem_phi);
3660 #else
3661         pset_insert_ptr(si->phims, mem_phi);
3662 #endif
3663
3664         if(opt_keep_alive & KEEPALIVE_SPILLS)
3665                 pset_insert_ptr(si->spills, mem_phi);
3666
3667
3668         return mem_phi;
3669 }
3670
3671 /**
3672  * Add remat to list of defs, destroys link field!
3673  */
3674 static void
3675 insert_remat(spill_ilp_t * si, ir_node * remat)
3676 {
3677         defs_t   *defs;
3678         op_t     *remat_op = get_irn_link(remat);
3679
3680         assert(remat_op->is_remat);
3681
3682         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3683         assert(defs);
3684
3685         /* enter into the linked list */
3686         set_irn_link(remat, defs->remats);
3687         defs->remats = remat;
3688 }
3689
3690
3691 /**
3692  * Add reload before operation and add to list of defs
3693  */
3694 static ir_node *
3695 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3696 {
3697         defs_t   *defs;
3698         ir_node  *reload,
3699                          *spill;
3700
3701         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3702
3703         defs = set_find_def(si->values, value);
3704
3705         spill = defs->spills;
3706         assert(spill && "no spill placed before reload");
3707
3708         reload = be_reload(si->cls, after, get_irn_mode(value), spill);
3709
3710         /* enter into the linked list */
3711         set_irn_link(reload, defs->remats);
3712         defs->remats = reload;
3713
3714         return reload;
3715 }
3716
3717 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3718 {
3719         defs_t  *defs;
3720         ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3721         ir_node *spill;
3722
3723         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3724
3725         defs = set_find_def(si->values, value);
3726
3727         spill = defs->spills;
3728         assert(spill && "no spill placed before reload");
3729
3730         arch_perform_memory_operand(memoperand->irn, spill, memoperand->pos);
3731 }
3732
3733 void insert_memoperands(spill_ilp_t * si)
3734 {
3735         memoperand_t   *memoperand;
3736         lpp_name_t     *name;
3737
3738         set_foreach(si->memoperands, memoperand) {
3739                 name = si->lpp->vars[memoperand->ilp];
3740                 if(!is_zero(name->value)) {
3741                         perform_memory_operand(si, memoperand);
3742                 }
3743         }
3744 }
3745
3746 static void
3747 walker_spill_placer(ir_node * bb, void * data) {
3748         spill_ilp_t   *si = (spill_ilp_t*)data;
3749         ir_node       *irn;
3750         spill_bb_t    *spill_bb = get_irn_link(bb);
3751         pset          *spills_to_do = pset_new_ptr_default();
3752         spill_t       *spill;
3753
3754         set_foreach(spill_bb->ilp, spill) {
3755                 lpp_name_t    *name;
3756
3757                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3758                         name = si->lpp->vars[spill->mem_in];
3759                         if(!is_zero(name->value)) {
3760                                 ir_node   *mem_phi;
3761
3762                                 mem_phi = insert_mem_phi(si, spill->irn);
3763
3764                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3765                         }
3766                 }
3767
3768                 name = si->lpp->vars[spill->spill];
3769                 if(!is_zero(name->value)) {
3770                         /* place spill directly after definition */
3771                         if(get_nodes_block(spill->irn) == bb) {
3772                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3773                                 continue;
3774                         }
3775
3776                         /* place spill at bb start */
3777                         if(spill->reg_in > 0) {
3778                                 name = si->lpp->vars[spill->reg_in];
3779                                 if(!is_zero(name->value)) {
3780                                         insert_spill(si, spill->irn, spill->irn, bb);
3781                                         continue;
3782                                 }
3783                         }
3784                         /* place spill after a remat */
3785                         pset_insert_ptr(spills_to_do, spill->irn);
3786                 }
3787         }
3788         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3789
3790
3791         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3792                 op_t     *op = get_irn_link(irn);
3793
3794                 if(be_is_Spill(irn)) continue;
3795
3796                 if(op->is_remat) {
3797                         /* TODO fix this if we want to support remats with more than two nodes */
3798                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3799                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3800
3801                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3802                         }
3803                 } else {
3804                         if(pset_find_ptr(spills_to_do, irn)) {
3805                                 pset_remove_ptr(spills_to_do, irn);
3806
3807                                 insert_spill(si, irn, irn, irn);
3808                         }
3809                 }
3810
3811         }
3812
3813         assert(pset_count(spills_to_do) == 0);
3814
3815         /* afterwards free data in block */
3816         del_pset(spills_to_do);
3817 }
3818
3819 static ir_node *
3820 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3821 {
3822         ir_node          *insert_pos = bb;
3823         ir_node          *spill;
3824
3825         /* find last definition of arg value in block */
3826         ir_node  *next;
3827         defs_t   *defs;
3828         int       last = 0;
3829
3830         defs = set_find_def(si->values, value);
3831
3832         if(defs && defs->remats) {
3833                 for(next = defs->remats; next; next = get_irn_link(next)) {
3834                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3835                                 last = sched_get_time_step(next);
3836                                 insert_pos = next;
3837                         }
3838                 }
3839         }
3840
3841         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3842                 last = sched_get_time_step(value);
3843                 insert_pos = value;
3844         }
3845
3846         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3847
3848         spill = be_spill2(is_Block(insert_pos)?value:insert_pos, insert_pos);
3849
3850         return spill;
3851 }
3852
3853 static void
3854 phim_fixer(spill_ilp_t *si) {
3855         defs_t  *defs;
3856
3857         set_foreach(si->values, defs) {
3858                 const ir_node  *phi = defs->value;
3859                 op_t           *op = get_irn_link(phi);
3860                 ir_node        *phi_m = NULL;
3861                 ir_node        *next = defs->spills;
3862                 int             n;
3863
3864                 if(!is_Phi(phi)) continue;
3865
3866                 while(next) {
3867                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3868                                 phi_m = next;
3869                                 break;
3870                         } else {
3871                                 next = get_irn_link(next);
3872                         }
3873                 }
3874                 if(!phi_m) continue;
3875
3876                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3877                         ir_node        *value = get_irn_n(phi, n);
3878                         defs_t         *val_defs = set_find_def(si->values, value);
3879
3880                         /* a spill of this value */
3881                         ir_node      *spill;
3882
3883
3884                         if(opt_memcopies) {
3885                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3886                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3887
3888                                 if(!is_zero(name->value)) {
3889                                         spill = insert_mem_copy(si, pred, value);
3890                                 } else {
3891                                         spill = val_defs->spills;
3892                                 }
3893                         } else {
3894                                 spill = val_defs->spills;
3895                         }
3896
3897                         assert(spill && "no spill placed before PhiM");
3898                         set_irn_n(phi_m, n, spill);
3899                 }
3900         }
3901 }
3902
3903 static void
3904 walker_reload_placer(ir_node * bb, void * data) {
3905         spill_ilp_t   *si = (spill_ilp_t*)data;
3906         ir_node       *irn;
3907         spill_bb_t    *spill_bb = get_irn_link(bb);
3908
3909         /* reloads at end of block */
3910         if(spill_bb->reloads) {
3911                 keyval_t    *keyval;
3912
3913                 set_foreach(spill_bb->reloads, keyval) {
3914                         ir_node        *irn = (ir_node*)keyval->key;
3915                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3916                         lpp_name_t     *name;
3917
3918                         name = si->lpp->vars[reload];
3919                         if(!is_zero(name->value)) {
3920                                 ir_node    *reload;
3921                                 ir_node    *insert_pos = bb;
3922                                 ir_node    *prev = sched_block_last_noncf(bb);
3923                                 op_t       *prev_op = get_irn_link(prev);
3924
3925                                 while(be_is_Spill(prev)) {
3926                                         prev = sched_prev(prev);
3927                                 }
3928
3929                                 prev_op = get_irn_link(prev);
3930
3931                                 /* insert reload before pre-remats */
3932                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3933                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3934                                         insert_pos = prev;
3935
3936                                         do {
3937                                                 prev = sched_prev(prev);
3938                                         } while(be_is_Spill(prev));
3939
3940                                         prev_op = get_irn_link(prev);
3941
3942                                 }
3943
3944                                 reload = insert_reload(si, irn, insert_pos);
3945
3946                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3947                                         pset_insert_ptr(si->spills, reload);
3948                         }
3949                 }
3950         }
3951
3952         /* walk and insert more reloads and collect remats */
3953         sched_foreach_reverse(bb, irn) {
3954                 op_t     *op = get_irn_link(irn);
3955
3956                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3957                 if(is_Phi(irn)) break;
3958
3959                 if(op->is_remat) {
3960                         if(get_irn_mode(irn) != mode_T) {
3961                                 insert_remat(si, irn);
3962                         }
3963                 } else {
3964                         int    n;
3965
3966                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3967                                 ir_node    *arg = get_irn_n(irn, n);
3968
3969                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3970                                         lpp_name_t    *name;
3971
3972                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3973                                         if(!is_zero(name->value)) {
3974                                                 ir_node    *reload;
3975                                                 ir_node    *insert_pos = irn;
3976                                                 ir_node    *prev = sched_prev(insert_pos);
3977                                                 op_t       *prev_op;
3978
3979                                                 while(be_is_Spill(prev)) {
3980                                                         prev = sched_prev(prev);
3981                                                 }
3982
3983                                                 prev_op = get_irn_link(prev);
3984
3985                                                 /* insert reload before pre-remats */
3986                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3987                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3988                                                         insert_pos = prev;
3989
3990                                                         do {
3991                                                                 prev = sched_prev(prev);
3992                                                         } while(be_is_Spill(prev));
3993
3994                                                         prev_op = get_irn_link(prev);
3995
3996                                                 }
3997
3998                                                 reload = insert_reload(si, arg, insert_pos);
3999
4000                                                 assert(reload && "no reload returned");
4001                                                 set_irn_n(irn, n, reload);
4002
4003                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
4004                                                         pset_insert_ptr(si->spills, reload);
4005                                         }
4006                                 }
4007                         }
4008                 }
4009         }
4010
4011         del_set(spill_bb->ilp);
4012         if(spill_bb->reloads) del_set(spill_bb->reloads);
4013 }
4014
4015 static void
4016 walker_collect_used(ir_node * irn, void * data)
4017 {
4018         bitset_t   *used = data;
4019
4020         bitset_set(used, get_irn_idx(irn));
4021 }
4022
4023 struct kill_helper {
4024         bitset_t  *used;
4025         spill_ilp_t  *si;
4026 };
4027
4028 static void
4029 walker_kill_unused(ir_node * bb, void * data)
4030 {
4031         struct kill_helper *kh = data;
4032         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
4033         ir_node            *irn;
4034
4035
4036         for(irn=sched_first(bb); !sched_is_end(irn);) {
4037                 ir_node     *next = sched_next(irn);
4038                 int          n;
4039
4040                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4041                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
4042                                 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(irn) * execution_frequency(kh->si, bb)));
4043 #if 0
4044                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4045 #endif
4046                         }
4047
4048                         sched_remove(irn);
4049
4050                         set_nodes_block(irn, bad);
4051                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
4052                                 set_irn_n(irn, n, bad);
4053                         }
4054                 }
4055                 irn = next;
4056         }
4057 }
4058
4059 #ifndef SCHEDULE_PHIM
4060 static void
4061 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4062 {
4063         ir_node  *phi;
4064         ir_node  *bad = get_irg_bad(si->birg->irg);
4065         int       n;
4066
4067         pset_foreach(si->phims, phi) {
4068                 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4069
4070                         set_nodes_block(phi, bad);
4071                         for (n=get_irn_arity(phi)-1; n>=0; --n) {
4072                                 set_irn_n(phi, n, bad);
4073                         }
4074                 }
4075         }
4076 }
4077 #endif
4078
4079 static void
4080 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4081 {
4082         struct kill_helper  kh;
4083
4084         kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4085         kh.si = si;
4086
4087         irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4088 #ifndef SCHEDULE_PHIM
4089         kill_unused_phims(si, &kh);
4090 #endif
4091         irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4092
4093         bitset_free(kh.used);
4094 }
4095
4096 void
4097 print_irn_pset(pset * p)
4098 {
4099         ir_node   *irn;
4100
4101         pset_foreach(p, irn) {
4102                 ir_printf("%+F\n", irn);
4103         }
4104 }
4105
4106 void
4107 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4108 {
4109     FILE           *f = fopen(file, "w");
4110     ir_node        *irn;
4111     interference_t *interference;
4112     int            i;
4113
4114     set_break(si->interferences);
4115
4116     ir_fprintf(f, "digraph phiclass {\n");
4117
4118     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4119         irn = phiclass[i];
4120         if (is_Phi(irn))
4121             ir_fprintf(f, "  %F%N [shape=box]\n", irn, irn);
4122     }
4123
4124     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4125         int n;
4126
4127         irn = phiclass[i];
4128         if (! is_Phi(irn))
4129             continue;
4130
4131         for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4132             ir_node  *arg = get_irn_n(irn, n);
4133
4134             ir_fprintf(f, "  %F%N -> %F%N\n", irn, irn, arg, arg);
4135         }
4136     }
4137
4138     set_foreach(si->interferences, interference) {
4139         const ir_node *a = interference->a;
4140         const ir_node *b = interference->b;
4141         if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4142             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4143         }
4144     }
4145
4146     ir_fprintf(f, "}");
4147     fclose(f);
4148 }
4149
4150 static void
4151 rewire_uses(spill_ilp_t * si)
4152 {
4153         defs_t               *defs;
4154         ir_nodeset_t         ignore;
4155
4156         ir_nodeset_init(&ignore);
4157         ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg));
4158
4159         /* then fix uses of spills */
4160         set_foreach(si->values, defs) {
4161                 pset           *reloads;
4162                 pset           *spills;
4163                 const ir_node  *next = defs->remats;
4164                 int remats = 0;
4165
4166                 reloads = pset_new_ptr_default();
4167
4168                 while(next) {
4169                         if(be_is_Reload(next)) {
4170                                 pset_insert_ptr(reloads, next);
4171                         } else {
4172                                 ++remats;
4173                         }
4174                         next = get_irn_link(next);
4175                 }
4176
4177                 spills = get_spills_for_value(si, defs->value);
4178                 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));
4179                 if(pset_count(spills) > 1) {
4180                         be_ssa_construction_env_t senv;
4181                         ir_node *node;
4182                         //assert(pset_count(reloads) > 0);
4183                         //                              print_irn_pset(spills);
4184                         //                              print_irn_pset(reloads);
4185
4186                         be_ssa_construction_init(&senv, si->birg);
4187                         be_ssa_construction_set_ignore_uses(&senv, &ignore);
4188                         pset_foreach(spills, node) {
4189                                 be_ssa_construction_add_copy(&senv, node);
4190                         }
4191                         pset_foreach(spills, node) {
4192                                 be_ssa_construction_fix_users(&senv, node);
4193                         }
4194                         be_ssa_construction_update_liveness_phis(&senv, si->lv);
4195                         pset_foreach(spills, node) {
4196                                 be_liveness_update(si->lv, node);
4197                         }
4198                         be_ssa_construction_destroy(&senv);
4199                 }
4200
4201                 del_pset(reloads);
4202                 del_pset(spills);
4203         }
4204
4205         /* first fix uses of remats and reloads */
4206         set_foreach(si->values, defs) {
4207                 const ir_node  *next = defs->remats;
4208                 int             orig_kept = 0;
4209
4210                 if(next) {
4211                         be_ssa_construction_env_t senv;
4212
4213                         be_ssa_construction_init(&senv, si->birg);
4214
4215                         if(sched_is_scheduled(defs->value)) {
4216                                 be_ssa_construction_add_copy(&senv, (ir_node*) defs->value);
4217                                 orig_kept = 1;
4218                         }
4219
4220                         next = defs->remats;
4221                         while(next) {
4222                                 be_ssa_construction_add_copy(&senv, (ir_node*) next);
4223                                 next = get_irn_link(next);
4224                         }
4225
4226                         if(sched_is_scheduled(defs->value)) {
4227                                 be_ssa_construction_fix_users(&senv, (ir_node*) defs->value);
4228                         }
4229
4230                         next = defs->remats;
4231                         while(next) {
4232                                 be_ssa_construction_fix_users(&senv, (ir_node*) next);
4233                                 next = get_irn_link(next);
4234                         }
4235
4236                         be_ssa_construction_update_liveness_phis(&senv, si->lv);
4237                         if(sched_is_scheduled(defs->value)) {
4238                                 be_liveness_update(si->lv, (ir_node*) defs->value);
4239                         }
4240
4241                         next = defs->remats;
4242                         while(next) {
4243                                 be_liveness_update(si->lv, (ir_node*) next);
4244                                 next = get_irn_link(next);
4245                         }
4246
4247                         be_ssa_construction_destroy(&senv);
4248                 }
4249         }
4250
4251         ir_nodeset_destroy(&ignore);
4252 //      remove_unused_defs(si);
4253 }
4254
4255
4256 static void
4257 writeback_results(spill_ilp_t * si)
4258 {
4259         /* walk through the graph and collect all spills, reloads and remats for a value */
4260
4261         si->values = new_set(cmp_defs, 4096);
4262
4263         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4264         delete_unnecessary_remats(si);
4265         si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4266         irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4267         irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4268         if(opt_memoperands)
4269                 insert_memoperands(si);
4270         phim_fixer(si);
4271
4272         /* clean the remat info! there are still back-edges leading there! */
4273         clean_remat_info(si);
4274
4275         rewire_uses(si);
4276
4277         connect_all_spills_with_keep(si);
4278
4279         del_set(si->values);
4280 }
4281
4282 static int
4283 get_n_regs(spill_ilp_t * si)
4284 {
4285         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4286
4287         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4288         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4289
4290         arch_put_non_ignore_regs(si->cls, arch_regs);
4291         be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4292
4293         bitset_andnot(arch_regs, abi_regs);
4294         arch_n_regs = bitset_popcnt(arch_regs);
4295
4296         bitset_free(arch_regs);
4297         bitset_free(abi_regs);
4298
4299         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4300         return arch_n_regs;
4301 }
4302
4303 static void
4304 walker_reload_mover(ir_node * bb, void * data)
4305 {
4306         spill_ilp_t   *si = data;
4307         ir_node           *tmp;
4308
4309         sched_foreach(bb, tmp) {
4310                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4311                         ir_node       *reload = tmp;
4312                         ir_node       *irn = tmp;
4313
4314                         /* move reload upwards */
4315
4316                         int pressure = (int)get_irn_link(reload);
4317                         if(pressure < si->n_regs) {
4318                                 irn = sched_prev(reload);
4319                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4320                                 sched_remove(reload);
4321                                 pressure = (int)get_irn_link(irn);
4322
4323                                 while(pressure < si->n_regs) {
4324                                         if( sched_is_end(irn) ||
4325                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4326                                            /* do not move reload before its spill */
4327                                            (irn == be_get_Reload_mem(reload)) ||
4328                                            /* do not move before phi */
4329                                            is_Phi(irn)) break;
4330
4331                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4332                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4333                                         irn = sched_prev(irn);
4334
4335                                         pressure = (int)get_irn_link(irn);
4336                                 }
4337
4338                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4339                                 sched_put_after(irn, reload);
4340                         }
4341                 }
4342         }
4343 }
4344
4345 static void
4346 move_reloads_upward(spill_ilp_t * si)
4347 {
4348         irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4349 }
4350
4351
4352 /**
4353  * Walk all irg blocks and check for interfering spills inside of phi classes
4354  */
4355 static void
4356 luke_meminterferencechecker(ir_node * bb, void * data)
4357 {
4358         spill_ilp_t *si = (spill_ilp_t*)data;
4359         int         l1, l2;
4360
4361         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4362                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4363
4364                 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4365                         continue;
4366
4367                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4368                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4369                         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)) {
4370                                 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4371
4372                                 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4373                                         continue;
4374
4375                                 /* a and b are only interesting if they are in the same phi class */
4376                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4377                                         if (values_interfere_in_block(si, bb, a, b)) {
4378                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4379                                         }
4380                                 }
4381                         }
4382                 }
4383         }
4384 }
4385
4386 static void
4387 verify_phiclasses(spill_ilp_t * si)
4388 {
4389         /* analyze phi classes */
4390         phi_class_free(si->pc);
4391         si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4392
4393         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4394         irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4395 }
4396
4397 void
4398 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4399 {
4400         char            buf[256];
4401         char            problem_name[256];
4402         char            dump_suffix[256];
4403         char            dump_suffix2[256];
4404         struct obstack  obst;
4405         spill_ilp_t     si;
4406         ir_graph       *irg = be_get_birg_irg(birg);
4407
4408         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4409         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4410         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4411
4412         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4413         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4414
4415         if(opt_verify & VERIFY_DOMINANCE)
4416                 be_check_dominance(irg);
4417
4418         be_assure_liveness(birg);
4419
4420         obstack_init(&obst);
4421         si.obst                = &obst;
4422         si.birg                = birg;
4423         si.cls                 = cls;
4424         si.lpp                 = new_lpp(problem_name, lpp_minimize);
4425         si.remat_info          = new_set(cmp_remat_info, 4096);
4426         si.interferences       = new_set(cmp_interference, 32);
4427         si.memoperands         = new_set(cmp_memoperands, 128);
4428         si.all_possible_remats = pset_new_ptr_default();
4429         si.spills              = pset_new_ptr_default();
4430         si.inverse_ops         = pset_new_ptr_default();
4431         si.lv                  = birg->lv;
4432         si.keep                = NULL;
4433         si.n_regs              = get_n_regs(&si);
4434
4435         set_irg_link(irg, &si);
4436         compute_doms(irg);
4437
4438         /* compute phi classes */
4439         // phi_class_compute(irg);
4440
4441         if(opt_dump_flags & DUMP_STATS)
4442                 be_analyze_regpressure(birg, cls, "-pre");
4443
4444         DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4445         irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4446
4447         if(opt_remats) {
4448                 /* collect remats */
4449                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4450                 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4451         }
4452
4453         /* insert possible remats */
4454         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4455         irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4456         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4457
4458         if(opt_keep_alive & KEEPALIVE_REMATS) {
4459                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4460                 connect_all_remats_with_keep(&si);
4461                 /* dump graph with inserted remats */
4462                 dump_graph_with_remats(irg, dump_suffix);
4463         }
4464
4465         /* insert copies for phi arguments not in my regclass */
4466         irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4467
4468         /* recompute liveness */
4469         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4470         be_liveness_recompute(si.lv);
4471
4472         /* build the ILP */
4473         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4474         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4475         irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4476
4477         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4478         irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4479
4480         si.pc = phi_class_new_from_irg(birg->irg, 0);
4481         if (opt_memcopies) {
4482                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4483                 memcopyhandler(&si);
4484         }
4485
4486         if (opt_dump_flags & DUMP_PROBLEM) {
4487                 FILE           *f;
4488                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4489                 if ((f = fopen(buf, "wt")) != NULL) {
4490                         lpp_dump_plain(si.lpp, f);
4491                         fclose(f);
4492                 }
4493         }
4494
4495         if (opt_dump_flags & DUMP_MPS) {
4496                 FILE *f;
4497
4498                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4499                 if ((f = fopen(buf, "wt")) != NULL) {
4500                         mps_write_mps(si.lpp, s_mps_fixed, f);
4501                         fclose(f);
4502                 }
4503
4504                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4505                 if ((f = fopen(buf, "wt")) != NULL) {
4506                         mps_write_mst(si.lpp, s_mps_fixed, f);
4507                         fclose(f);
4508                 }
4509         }
4510
4511         lpp_check_startvals(si.lpp);
4512
4513 #ifdef SOLVE
4514         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4515         lpp_set_time_limit(si.lpp, opt_timeout);
4516
4517         if(opt_log)
4518                 lpp_set_log(si.lpp, stdout);
4519
4520 #ifdef SOLVE_LOCAL
4521         lpp_solve_cplex(si.lpp);
4522 #else
4523         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4524 #endif
4525         assert(lpp_is_sol_valid(si.lpp)
4526                && "solution of ILP must be valid");
4527
4528         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4529
4530         if(opt_dump_flags & DUMP_SOLUTION) {
4531                 FILE           *f;
4532                 char            buf[256];
4533
4534                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4535                 if ((f = fopen(buf, "wt")) != NULL) {
4536                         int             i;
4537                         for (i = 0; i < si.lpp->var_next; ++i) {
4538                                 lpp_name_t     *name = si.lpp->vars[i];
4539                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4540                         }
4541                         fclose(f);
4542                 }
4543         }
4544
4545 #ifndef SCHEDULE_PHIM
4546         si.phims = pset_new_ptr_default();
4547 #endif
4548         writeback_results(&si);
4549
4550
4551 #endif                          /* SOLVE */
4552
4553         kill_all_unused_values_in_schedule(&si);
4554
4555 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4556         del_pset(si.phims);
4557 #endif
4558
4559         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4560                 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4561
4562         // move reloads upwards
4563         be_liveness_recompute(si.lv);
4564         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4565         move_reloads_upward(&si);
4566
4567         if(opt_memcopies) {
4568                 verify_phiclasses(&si);
4569         }
4570
4571         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4572
4573         if(opt_dump_flags & DUMP_PRESSURE)
4574                 dump_pressure_graph(&si, dump_suffix2);
4575
4576         if(opt_dump_flags & DUMP_STATS)
4577                 be_analyze_regpressure(birg, cls, "-post");
4578
4579         if(opt_verify & VERIFY_DOMINANCE)
4580                 be_check_dominance(irg);
4581
4582         free_dom(irg);
4583         del_set(si.interferences);
4584         del_pset(si.inverse_ops);
4585         del_pset(si.all_possible_remats);
4586         del_set(si.memoperands);
4587         del_pset(si.spills);
4588         free_lpp(si.lpp);
4589         phi_class_free(si.pc);
4590         obstack_free(&obst, NULL);
4591         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4592 }
4593
4594 void be_init_spillremat(void)
4595 {
4596         static be_spiller_t remat_spiller = {
4597                 be_spill_remat
4598         };
4599         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4600         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4601         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4602         lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4603
4604         be_register_spiller("remat", &remat_spiller);
4605         lc_opt_add_table(remat_grp, options);
4606 }
4607
4608 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4609
4610 #else                           /* WITH_ILP */
4611
4612 static __attribute__((unused))
4613 void only_that_you_can_compile_without_WITH_ILP_defined(void)
4614 {
4615 }
4616
4617 #endif                          /* WITH_ILP */