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 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #ifdef WITH_ILP
  32
  33 #include <math.h>
  34
  35 #include "hashptr.h"
  36 #include "debug.h"
  37 #include "obst.h"
  38 #include "set.h"
  39 #include "list.h"
  40 #include "pmap.h"
  41
  42 #include "irprintf.h"
  43 #include "irgwalk.h"
  44 #include "irdump_t.h"
  45 #include "irnode_t.h"
  46 #include "ircons_t.h"
  47 #include "irloop_t.h"
  48 #include "irnodeset.h"
  49 #include "phiclass.h"
  50 #include "iredges_t.h"
  51 #include "execfreq.h"
  52 #include "irvrfy.h"
  53 #include "irbackedge_t.h"
  54 #include "irprofile.h"
  55
  56 #include <lpp/lpp.h>
  57 #include <lpp/mps.h>
  58 #include <lpp/lpp_net.h>
  59 #include <lpp/lpp_cplex.h>
  60
  61 #include "be_t.h"
  62 #include "beirg_t.h"
  63 #include "belive_t.h"
  64 #include "besched_t.h"
  65 #include "bessaconstr.h"
  66 #include "bearch_t.h"
  67 #include "beintlive_t.h"
  68 #include "beabi.h"
  69 #include "benode_t.h"
  70 #include "beutil.h"
  71 #include "bespillremat.h"
  72 #include "bespill.h"
  73 #include "bepressurestat.h"
  74 #include "bespilloptions.h"
  75 #include "bechordal_t.h"
  76 #include "bemodule.h"
  77
  78 #include "lc_opts.h"
  79 #include "lc_opts_enum.h"
  80
  81 #define DUMP_PROBLEM       1
  82 #define DUMP_MPS           2
  83 #define DUMP_SOLUTION      4
  84 #define DUMP_STATS         8
  85 #define DUMP_PRESSURE      16
  86
  87 #define KEEPALIVE_REMATS   1
  88 #define KEEPALIVE_SPILLS   2
  89 #define KEEPALIVE_RELOADS  4
  90
  91 #define VERIFY_MEMINTERF   1
  92 #define VERIFY_DOMINANCE   2
  93
  94 #define REMATS_NONE        0
  95 #define REMATS_BRIGGS      1
  96 #define REMATS_NOINVERSE   2
  97 #define REMATS_ALL         3
  98
  99 static unsigned opt_dump_flags = 0;
 100 static int opt_log = 0;
 101 static unsigned opt_keep_alive = 0;
 102 static int opt_goodwin = 1;
 103 static int opt_memcopies = 1;
 104 static int opt_memoperands = 1;
 105 static int opt_verify = VERIFY_MEMINTERF;
 106 static unsigned opt_remats = REMATS_ALL;
 107 static int opt_repair_schedule = 0;
 108 static int opt_no_enlarge_liveness = 0;
 109 static int opt_remat_while_live = 1;
 110 static int opt_timeout = 300;
 111 static double opt_cost_reload = 8.0;
 112 static double opt_cost_memoperand =  7.0;
 113 static double opt_cost_spill =  15.0;
 114 static double opt_cost_remat =  1.0;
 115
 116
 117 static const lc_opt_enum_mask_items_t dump_items[] = {
 118         { "problem",  DUMP_PROBLEM  },
 119         { "mps",      DUMP_MPS      },
 120         { "solution", DUMP_SOLUTION },
 121         { "stats",    DUMP_STATS },
 122         { "pressure", DUMP_PRESSURE },
 123         { NULL,       0 }
 124 };
 125
 126 static lc_opt_enum_mask_var_t dump_var = {
 127         &opt_dump_flags, dump_items
 128 };
 129
 130 static const lc_opt_enum_mask_items_t keepalive_items[] = {
 131         { "remats",  KEEPALIVE_REMATS  },
 132         { "spills",  KEEPALIVE_SPILLS  },
 133         { "reloads", KEEPALIVE_RELOADS },
 134         { NULL,      0 }
 135 };
 136
 137 static lc_opt_enum_mask_var_t keep_alive_var = {
 138         &opt_keep_alive, keepalive_items
 139 };
 140
 141 static const lc_opt_enum_mask_items_t remats_items[] = {
 142         { "none",      REMATS_NONE      },
 143         { "briggs",    REMATS_BRIGGS    },
 144         { "noinverse", REMATS_NOINVERSE },
 145         { "all",       REMATS_ALL       },
 146         { NULL,        0 }
 147 };
 148
 149 static lc_opt_enum_mask_var_t remats_var = {
 150         &opt_remats, remats_items
 151 };
 152
 153 static const lc_opt_table_entry_t options[] = {
 154         LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes",                              &keep_alive_var),
 155
 156         LC_OPT_ENT_BOOL     ("goodwin",  "activate goodwin reduction",                              &opt_goodwin),
 157         LC_OPT_ENT_BOOL     ("memcopies",  "activate memcopy handling",                             &opt_memcopies),
 158         LC_OPT_ENT_BOOL     ("memoperands",  "activate memoperands",                                &opt_memoperands),
 159         LC_OPT_ENT_ENUM_INT ("remats",  "type of remats to insert",                                 &remats_var),
 160         LC_OPT_ENT_BOOL     ("repair_schedule",  "repair the schedule by rematting once used nodes",&opt_repair_schedule),
 161         LC_OPT_ENT_BOOL     ("no_enlage_liveness",  "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
 162         LC_OPT_ENT_BOOL     ("remat_while_live",  "only remat where rematted value was live",       &opt_remat_while_live),
 163
 164         LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data",                  &dump_var),
 165         LC_OPT_ENT_BOOL     ("log",  "activate the lpp log",                                        &opt_log),
 166         LC_OPT_ENT_INT      ("timeout",  "ILP solver timeout",                                      &opt_timeout),
 167
 168         LC_OPT_ENT_DBL      ("cost_reload",  "cost of a reload",                                    &opt_cost_reload),
 169         LC_OPT_ENT_DBL      ("cost_memoperand",  "cost of a memory operand",                        &opt_cost_memoperand),
 170         LC_OPT_ENT_DBL      ("cost_spill",  "cost of a spill instruction",                          &opt_cost_spill),
 171         LC_OPT_ENT_DBL      ("cost_remat",  "cost of a rematerialization",                          &opt_cost_remat),
 172         LC_OPT_LAST
 173 };
 174
 175 //#define EXECFREQ_LOOPDEPH   /* compute execution frequency from loop depth only */
 176 //#define SCHEDULE_PHIM   /* insert phim nodes into schedule */
 177
 178 #define  SOLVE
 179 //#define  SOLVE_LOCAL
 180 #define LPP_SERVER "i44pc52"
 181 #define LPP_SOLVER "cplex"
 182
 183
 184 #define MAX_PATHS      INT_MAX
 185 #define ILP_UNDEF               -1
 186
 187 typedef struct _spill_ilp_t {
 188         const arch_register_class_t  *cls;
 189         int                           n_regs;
 190         be_irg_t                     *birg;
 191         be_lv_t                      *lv;
 192         lpp_t                        *lpp;
 193         struct obstack               *obst;
 194         set                          *remat_info;
 195         pset                         *all_possible_remats;
 196         pset                         *inverse_ops;
 197         ir_node                      *keep;
 198         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 199         pset                         *spills;
 200         set                          *interferences;
 201         ir_node                      *m_unknown;
 202         set                          *memoperands;
 203         phi_classes_t                *pc;
 204 #ifndef SCHEDULE_PHIM
 205         pset                         *phims;
 206 #endif
 207         DEBUG_ONLY(firm_dbg_module_t * dbg);
 208 } spill_ilp_t;
 209
 210 typedef int ilp_var_t;
 211 typedef int ilp_cst_t;
 212
 213 typedef struct _spill_bb_t {
 214         set      *ilp;
 215         set      *reloads;
 216 } spill_bb_t;
 217
 218 typedef struct _remat_t {
 219         const ir_node        *op;      /**< for copy_irn */
 220         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 221         const ir_node        *proj;    /**< not NULL if the above op produces a tuple */
 222         int                   cost;    /**< cost of this remat */
 223         int                   inverse; /**< nonzero if this is an inverse remat */
 224 } remat_t;
 225
 226 /**
 227  * Data to be attached to each IR node. For remats this contains the ilp_var
 228  * for this remat and for normal ops this contains the ilp_vars for
 229  * reloading each operand
 230  */
 231 typedef struct _op_t {
 232         int             is_remat;
 233         union {
 234                 struct {
 235                         ilp_var_t       ilp;
 236                         const remat_t  *remat; /** the remat this op belongs to */
 237                         int             pre; /** 1, if this is a pressure-increasing remat */
 238                 } remat;
 239                 struct {
 240                         ilp_var_t       ilp;
 241                         ir_node        *op; /** the operation this live range belongs to */
 242                         union {
 243                                 ilp_var_t      *reloads;
 244                                 ilp_var_t      *copies;
 245                         } args;
 246                 } live_range;
 247         } attr;
 248 } op_t;
 249
 250 typedef struct _defs_t {
 251         const ir_node   *value;
 252         ir_node         *spills;  /**< points to the first spill for this value (linked by link field) */
 253         ir_node         *remats;  /**< points to the first definition for this value (linked by link field) */
 254 } defs_t;
 255
 256 typedef struct _remat_info_t {
 257         const ir_node       *irn; /**< the irn to which these remats belong */
 258         pset                *remats; /**< possible remats for this value */
 259         pset                *remats_by_operand; /**< remats with this value as operand */
 260 } remat_info_t;
 261
 262 typedef struct _keyval_t {
 263         const void          *key;
 264         const void          *val;
 265 } keyval_t;
 266
 267 typedef struct _spill_t {
 268         ir_node            *irn;
 269         ilp_var_t           reg_in;
 270         ilp_var_t           mem_in;
 271         ilp_var_t           reg_out;
 272         ilp_var_t           mem_out;
 273         ilp_var_t           spill;
 274 } spill_t;
 275
 276 typedef struct _memoperand_t {
 277         ir_node             *irn; /**< the irn */
 278         unsigned int         pos; /**< the position of the argument */
 279         ilp_var_t            ilp; /**< the ilp var for this memory operand */
 280 } memoperand_t;
 281
 282 static INLINE int
 283 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 284 {
 285         return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
 286                                               si->cls, irn);
 287 }
 288
 289 #if 0
 290 static int
 291 cmp_remat(const void *a, const void *b)
 292 {
 293         const keyval_t *p = a;
 294         const keyval_t *q = b;
 295         const remat_t  *r = p->val;
 296         const remat_t  *s = q->val;
 297
 298         assert(r && s);
 299
 300         return !(r == s || r->op == s->op);
 301 }
 302 #endif
 303 static int
 304 cmp_remat(const void *a, const void *b)
 305 {
 306         const remat_t  *r = a;
 307         const remat_t  *s = b;
 308
 309         return !(r == s || r->op == s->op);
 310 }
 311
 312 static int
 313 cmp_spill(const void *a, const void *b, size_t size)
 314 {
 315         const spill_t *p = a;
 316         const spill_t *q = b;
 317         (void) size;
 318
 319 //      return !(p->irn == q->irn && p->bb == q->bb);
 320         return !(p->irn == q->irn);
 321 }
 322
 323 static int
 324 cmp_memoperands(const void *a, const void *b, size_t size)
 325 {
 326         const memoperand_t *p = a;
 327         const memoperand_t *q = b;
 328         (void) size;
 329
 330         return !(p->irn == q->irn && p->pos == q->pos);
 331 }
 332
 333 static keyval_t *
 334 set_find_keyval(set * set, const void * key)
 335 {
 336         keyval_t     query;
 337
 338         query.key = key;
 339         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 340 }
 341
 342 static keyval_t *
 343 set_insert_keyval(set * set, void * key, void * val)
 344 {
 345         keyval_t     query;
 346
 347         query.key = key;
 348         query.val = val;
 349         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 350 }
 351
 352 static defs_t *
 353 set_find_def(set * set, const ir_node * value)
 354 {
 355         defs_t     query;
 356
 357         query.value = value;
 358         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 359 }
 360
 361 static defs_t *
 362 set_insert_def(set * set, const ir_node * value)
 363 {
 364         defs_t     query;
 365
 366         query.value = value;
 367         query.spills = NULL;
 368         query.remats = NULL;
 369         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 370 }
 371
 372 static memoperand_t *
 373 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
 374 {
 375         memoperand_t     query;
 376
 377         query.irn = irn;
 378         query.pos = pos;
 379         query.ilp = ilp;
 380         return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 381 }
 382
 383 static memoperand_t *
 384 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
 385 {
 386         memoperand_t     query;
 387
 388         query.irn = (ir_node*)irn;
 389         query.pos = pos;
 390         return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
 391 }
 392
 393
 394 static spill_t *
 395 set_find_spill(set * set, const ir_node * value)
 396 {
 397         spill_t     query;
 398
 399         query.irn = (ir_node*)value;
 400         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 401 }
 402
 403 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 404 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 405 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 406 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 407 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 408
 409 static int
 410 cmp_remat_info(const void *a, const void *b, size_t size)
 411 {
 412         const remat_info_t *p = a;
 413         const remat_info_t *q = b;
 414         (void) size;
 415
 416         return !(p->irn == q->irn);
 417 }
 418
 419 static int
 420 cmp_defs(const void *a, const void *b, size_t size)
 421 {
 422         const defs_t *p = a;
 423         const defs_t *q = b;
 424         (void) size;
 425
 426         return !(p->value == q->value);
 427 }
 428
 429 static int
 430 cmp_keyval(const void *a, const void *b, size_t size)
 431 {
 432         const keyval_t *p = a;
 433         const keyval_t *q = b;
 434         (void) size;
 435
 436         return !(p->key == q->key);
 437 }
 438
 439 static double
 440 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 441 {
 442 #define FUDGE 0.001
 443         if(ir_profile_has_data())
 444                 return ((double)ir_profile_get_block_execcount(get_block_const(irn))) + FUDGE;
 445
 446 #ifndef EXECFREQ_LOOPDEPH
 447         return get_block_execfreq(si->birg->exec_freq, get_block_const(irn)) + FUDGE;
 448 #else
 449         if(is_Block(irn))
 450                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 451         else
 452                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 453 #endif
 454 }
 455
 456 static double
 457 get_cost(const spill_ilp_t * si, const ir_node * irn)
 458 {
 459         if(be_is_Spill(irn)) {
 460                 return opt_cost_spill;
 461         } else if(be_is_Reload(irn)){
 462                 return opt_cost_reload;
 463         } else {
 464                 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
 465         }
 466 }
 467
 468 /**
 469  * Checks, whether node and its operands have suitable reg classes
 470  */
 471 static INLINE int
 472 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 473 {
 474         int               n;
 475         const arch_env_t *arch_env = si->birg->main_env->arch_env;
 476         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 477
 478 #if 0
 479         if(!remat)
 480                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 481 #endif
 482
 483         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 484                 ir_node        *op = get_irn_n(irn, n);
 485                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 486
 487 //              if(!remat)
 488 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 489         }
 490
 491         return remat;
 492 }
 493
 494 /**
 495  * Try to create a remat from @p op with destination value @p dest_value
 496  */
 497 static INLINE remat_t *
 498 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 499 {
 500         remat_t  *remat = NULL;
 501
 502 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 503 //              return NULL;
 504
 505         if(dest_value == op) {
 506                 const ir_node *proj = NULL;
 507
 508                 if(is_Proj(dest_value)) {
 509                         op = get_Proj_pred(op);
 510                         proj = dest_value;
 511                 }
 512
 513                 if(!is_rematerializable(si, op))
 514                         return NULL;
 515
 516                 remat          = obstack_alloc(si->obst, sizeof(*remat));
 517                 remat->op      = op;
 518                 remat->cost    = (int)get_cost(si, op);
 519                 remat->value   = dest_value;
 520                 remat->proj    = proj;
 521                 remat->inverse = 0;
 522         } else {
 523                 arch_inverse_t     inverse;
 524                 int                n;
 525
 526                 /* get the index of the operand we want to retrieve by the inverse op */
 527                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 528                         ir_node        *arg = get_irn_n(op, n);
 529
 530                         if(arg == dest_value) break;
 531                 }
 532                 if(n<0) return NULL;
 533
 534                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 535
 536                 /* else ask the backend to give an inverse op */
 537                 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 538                         int   i;
 539
 540                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 541
 542                         assert(inverse.n > 0 && "inverse op should have at least one node");
 543
 544                         for(i=inverse.n-1; i>=0; --i) {
 545                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 546                         }
 547
 548                         if(inverse.n <= 2) {
 549                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 550                                 remat->op = inverse.nodes[0];
 551                                 remat->cost = inverse.costs;
 552                                 remat->value = dest_value;
 553                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 554                                 remat->inverse = 1;
 555
 556                                 // Matze: commented this out, this doesn't seem to be true if
 557                                 // the inverse is a simple operation with only 1 result...
 558                                 //assert(is_Proj(remat->proj));
 559                         } else {
 560                                 assert(0 && "I can not handle remats with more than 2 nodes");
 561                         }
 562                 }
 563         }
 564
 565         if(remat) {
 566                 if(remat->proj) {
 567                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 568                 } else {
 569                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 570                 }
 571         }
 572         return remat;
 573 }
 574
 575
 576 static INLINE void
 577 add_remat(const spill_ilp_t * si, const remat_t * remat)
 578 {
 579         remat_info_t    *remat_info,
 580                      query;
 581         int              n;
 582
 583         assert(remat->op);
 584         assert(remat->value);
 585
 586         query.irn = remat->value;
 587         query.remats = NULL;
 588         query.remats_by_operand = NULL;
 589         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 590
 591         if(remat_info->remats == NULL) {
 592                 remat_info->remats = new_pset(cmp_remat, 4096);
 593         }
 594         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 595
 596         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 597         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 598                 ir_node        *arg = get_irn_n(remat->op, n);
 599
 600                 query.irn = arg;
 601                 query.remats = NULL;
 602                 query.remats_by_operand = NULL;
 603                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 604
 605                 if(remat_info->remats_by_operand == NULL) {
 606                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 607                 }
 608                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 609         }
 610 }
 611
 612 static int
 613 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 614 {
 615         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 616         int                i = 0;
 617
 618         while(edge) {
 619                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 620                         ++i;
 621                 }
 622                 edge = get_irn_out_edge_next(irn, edge);
 623         }
 624
 625         return i;
 626 }
 627
 628 static int
 629 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
 630 {
 631         int n;
 632         int ret = 0;
 633
 634         if(is_Proj(irn))
 635                 irn = get_Proj_pred(irn);
 636
 637         for(n=get_irn_arity(irn)-1; n>=0; --n) {
 638                 const ir_node  *arg = get_irn_n(irn, n);
 639
 640                 if(has_reg_class(si, arg)) ++ret;
 641         }
 642
 643         return ret;
 644 }
 645
 646 static INLINE void
 647 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 648 {
 649         int      n;
 650         remat_t *remat;
 651
 652         if( has_reg_class(si, op)
 653         && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
 654         && (opt_remats !=  REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
 655         ) {
 656                 remat = get_remat_from_op(si, op, op);
 657                 if(remat) {
 658                         add_remat(si, remat);
 659                 }
 660         }
 661
 662         if(opt_remats == REMATS_ALL) {
 663                 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 664                    for each arg */
 665                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 666                         ir_node        *arg = get_irn_n(op, n);
 667
 668                         if(has_reg_class(si, arg)) {
 669                                 /* try to get an inverse remat */
 670                                 remat = get_remat_from_op(si, arg, op);
 671                                 if(remat) {
 672                                         add_remat(si, remat);
 673                                 }
 674                         }
 675                 }
 676         }
 677 }
 678
 679 static INLINE int
 680 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 681 {
 682         ir_node *block;
 683         ir_node *def_block = get_nodes_block(val);
 684         int      ret;
 685         (void) si;
 686
 687         if(val == pos)
 688                 return 0;
 689
 690         /* if pos is at end of a basic block */
 691         if(is_Block(pos)) {
 692                 ret = (pos == def_block || block_dominates(def_block, pos));
 693 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 694                 return ret;
 695         }
 696
 697         /* else if this is a normal operation */
 698         block = get_nodes_block(pos);
 699         if(block == def_block) {
 700                 if(!sched_is_scheduled(val)) return 1;
 701
 702                 ret = sched_comes_after(val, pos);
 703 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 704                 return ret;
 705         }
 706
 707         ret = block_dominates(def_block, block);
 708 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 709         return ret;
 710 }
 711
 712 static INLINE ir_node *
 713 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 714 {
 715     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
 716 }
 717
 718 /**
 719  * Returns first non-Phi node of block @p bb
 720  */
 721 static INLINE ir_node *
 722 sched_block_first_nonphi(const ir_node * bb)
 723 {
 724         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 725 }
 726
 727 static int
 728 sched_skip_proj_predicator(const ir_node * irn, void * data)
 729 {
 730         (void) data;
 731         return (is_Proj(irn));
 732 }
 733
 734 static INLINE ir_node *
 735 sched_next_nonproj(const ir_node * irn, int forward)
 736 {
 737         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 738 }
 739
 740 /**
 741  * Returns next operation node (non-Proj) after @p irn
 742  * or the basic block of this node
 743  */
 744 static INLINE ir_node *
 745 sched_next_op(const ir_node * irn)
 746 {
 747         ir_node *next = sched_next(irn);
 748
 749         if(is_Block(next))
 750                 return next;
 751
 752         return sched_next_nonproj(next, 1);
 753 }
 754
 755 /**
 756  * Returns previous operation node (non-Proj) before @p irn
 757  * or the basic block of this node
 758  */
 759 static INLINE ir_node *
 760 sched_prev_op(const ir_node * irn)
 761 {
 762         ir_node *prev = sched_prev(irn);
 763
 764         if(is_Block(prev))
 765                 return prev;
 766
 767         return sched_next_nonproj(prev, 0);
 768 }
 769
 770 static void
 771 sched_put_after(ir_node * insert, ir_node * irn)
 772 {
 773         if(is_Block(insert)) {
 774                 insert = sched_block_first_nonphi(insert);
 775         } else {
 776                 insert = sched_next_op(insert);
 777         }
 778         sched_reset(irn);
 779         sched_add_before(insert, irn);
 780 }
 781
 782 static void
 783 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 784 {
 785   if(is_Block(insert)) {
 786           insert = sched_block_last_noncf(si, insert);
 787   } else {
 788           insert = sched_next_nonproj(insert, 0);
 789           insert = sched_prev(insert);
 790   }
 791   sched_reset(irn);
 792   sched_add_after(insert, irn);
 793 }
 794
 795 static ir_node *
 796 next_post_remat(const ir_node * irn)
 797 {
 798         op_t      *op;
 799     ir_node   *next;
 800
 801         if(is_Block(irn)) {
 802                 next = sched_block_first_nonphi(irn);
 803         } else {
 804                 next = sched_next_op(irn);
 805         }
 806
 807         if(sched_is_end(next))
 808                 return NULL;
 809
 810         op = get_irn_link(next);
 811         if(op->is_remat && !op->attr.remat.pre) {
 812                 return next;
 813         }
 814
 815         return NULL;
 816 }
 817
 818
 819 static ir_node *
 820 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
 821 {
 822         op_t      *op;
 823         ir_node   *ret;
 824
 825         if(is_Block(irn)) {
 826                 ret = sched_block_last_noncf(si, irn);
 827                 ret = sched_next(ret);
 828                 ret = sched_prev_op(ret);
 829         } else {
 830                 ret = sched_prev_op(irn);
 831         }
 832
 833         if(sched_is_end(ret) || is_Phi(ret))
 834                 return NULL;
 835
 836         op = (op_t*)get_irn_link(ret);
 837         if(op->is_remat && op->attr.remat.pre) {
 838                 return ret;
 839         }
 840
 841         return NULL;
 842 }
 843
 844 /**
 845  * Tells you whether a @p remat can be placed before the irn @p pos
 846  */
 847 static INLINE int
 848 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 849 {
 850         const ir_node   *op = remat->op;
 851         const ir_node   *prev;
 852         int        n,
 853                            res = 1;
 854
 855         if(is_Block(pos)) {
 856                 prev = sched_block_last_noncf(si, pos);
 857                 prev = sched_next_nonproj(prev, 0);
 858         } else {
 859                 prev = sched_prev_op(pos);
 860         }
 861         /* do not remat if the rematted value is defined immediately before this op */
 862         if(prev == remat->op) {
 863                 return 0;
 864         }
 865
 866 #if 0
 867         /* this should be just fine, the following OP will be using this value, right? */
 868
 869         /* only remat AFTER the real definition of a value (?) */
 870         if(!value_is_defined_before(si, pos, remat->value)) {
 871 //              ir_fprintf(stderr, "error(not defined)");
 872                 return 0;
 873         }
 874 #endif
 875
 876         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 877                 const ir_node   *arg = get_irn_n(op, n);
 878
 879                 if(opt_no_enlarge_liveness) {
 880                         if(has_reg_class(si, arg) && live) {
 881                                 res &= pset_find_ptr((pset*)live, arg)?1:0;
 882                         } else {
 883                                 res &= value_is_defined_before(si, pos, arg);
 884                         }
 885                 } else {
 886                         res &= value_is_defined_before(si, pos, arg);
 887                 }
 888         }
 889
 890         return res;
 891 }
 892
 893 /**
 894  * Tells you whether a @p remat can be placed after the irn @p pos
 895  */
 896 static INLINE int
 897 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 898 {
 899         if(is_Block(pos)) {
 900                 pos = sched_block_first_nonphi(pos);
 901         } else {
 902                 pos = sched_next_op(pos);
 903         }
 904
 905         /* only remat AFTER the real definition of a value (?) */
 906         if(!value_is_defined_before(si, pos, remat->value)) {
 907                 return 0;
 908         }
 909
 910         return can_remat_before(si, remat, pos, live);
 911 }
 912
 913 /**
 914  * Collect potetially rematerializable OPs
 915  */
 916 static void
 917 walker_remat_collector(ir_node * irn, void * data)
 918 {
 919         spill_ilp_t    *si = data;
 920
 921         if(!is_Block(irn) && !is_Phi(irn)) {
 922                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 923                 get_remats_from_op(si, irn);
 924         }
 925 }
 926
 927 /**
 928  * Inserts a copy of @p irn before @p pos
 929  */
 930 static ir_node *
 931 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 932 {
 933         ir_node     *bb;
 934         ir_node     *copy;
 935
 936         bb = is_Block(pos)?pos:get_nodes_block(pos);
 937         copy = exact_copy(irn);
 938
 939         set_phi_class(si->pc, copy, NULL);
 940         set_nodes_block(copy, bb);
 941         sched_put_before(si, pos, copy);
 942
 943         return copy;
 944 }
 945
 946 /**
 947  * Inserts a copy of @p irn after @p pos
 948  */
 949 static ir_node *
 950 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 951 {
 952         ir_node     *bb;
 953         ir_node     *copy;
 954
 955         bb = is_Block(pos)?pos:get_nodes_block(pos);
 956         copy = exact_copy(irn);
 957
 958         set_phi_class(si->pc, copy, NULL);
 959         set_nodes_block(copy, bb);
 960         sched_put_after(pos, copy);
 961
 962         return copy;
 963 }
 964
 965 static ir_node *
 966 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
 967 {
 968         char     buf[256];
 969
 970         if(can_remat_after(si, remat, pos, live)) {
 971                 ir_node         *copy,
 972                                                 *proj_copy;
 973                 op_t            *op;
 974
 975                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat2 %+F\n", remat->op));
 976
 977                 copy = insert_copy_after(si, remat->op, pos);
 978
 979                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 980                 op = obstack_alloc(si->obst, sizeof(*op));
 981                 op->is_remat = 1;
 982                 op->attr.remat.remat = remat;
 983                 op->attr.remat.pre = 0;
 984                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
 985
 986                 set_irn_link(copy, op);
 987                 pset_insert_ptr(si->all_possible_remats, copy);
 988                 if(remat->proj) {
 989                         proj_copy = insert_copy_after(si, remat->proj, copy);
 990                         set_irn_n(proj_copy, 0, copy);
 991                         set_irn_link(proj_copy, op);
 992                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 993                 } else {
 994                         proj_copy = NULL;
 995                 }
 996
 997                 return copy;
 998         }
 999
1000         return NULL;
1001 }
1002
1003 static ir_node *
1004 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
1005 {
1006         char     buf[256];
1007
1008         if(can_remat_before(si, remat, pos, live)) {
1009                 ir_node         *copy,
1010                                                 *proj_copy;
1011                 op_t            *op;
1012
1013                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
1014
1015                 copy = insert_copy_before(si, remat->op, pos);
1016
1017                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
1018                 op = obstack_alloc(si->obst, sizeof(*op));
1019                 op->is_remat = 1;
1020                 op->attr.remat.remat = remat;
1021                 op->attr.remat.pre = 1;
1022                 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
1023
1024                 set_irn_link(copy, op);
1025                 pset_insert_ptr(si->all_possible_remats, copy);
1026                 if(remat->proj) {
1027                         proj_copy = insert_copy_after(si, remat->proj, copy);
1028                         set_irn_n(proj_copy, 0, copy);
1029                         set_irn_link(proj_copy, op);
1030                         pset_insert_ptr(si->all_possible_remats, proj_copy);
1031                 } else {
1032                         proj_copy = NULL;
1033                 }
1034
1035                 return copy;
1036         }
1037
1038         return NULL;
1039 }
1040
1041 static int
1042 get_block_n_succs(const ir_node *block) {
1043         const ir_edge_t *edge;
1044
1045         assert(edges_activated(current_ir_graph));
1046
1047         edge = get_block_succ_first(block);
1048         if (! edge)
1049                 return 0;
1050
1051         edge = get_block_succ_next(block, edge);
1052         return edge ? 2 : 1;
1053 }
1054
1055 static int
1056 is_start_block(const ir_node * bb)
1057 {
1058         return get_irg_start_block(get_irn_irg(bb)) == bb;
1059 }
1060
1061 static int
1062 is_merge_edge(const ir_node * bb)
1063 {
1064         if(is_start_block(bb))
1065                 return 0;
1066
1067         if(opt_goodwin)
1068                 return get_block_n_succs(bb) == 1;
1069         else
1070                 return 1;
1071 }
1072
1073 static int
1074 is_diverge_edge(const ir_node * bb)
1075 {
1076         if(is_start_block(bb))
1077                 return 0;
1078
1079         if(opt_goodwin)
1080                 return get_Block_n_cfgpreds(bb) == 1;
1081         else
1082                 return 1;
1083 }
1084
1085 static void
1086 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1087 {
1088         ir_node        *irn;
1089         int i;
1090
1091         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1092                 irn = be_lv_get_irn(si->lv, bb, i);
1093
1094                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1095                         pset_insert_ptr(live, irn);
1096                 }
1097         }
1098
1099         irn = sched_last(bb);
1100
1101         /* all values eaten by control flow operations are also live until the end of the block */
1102         sched_foreach_reverse(bb, irn) {
1103                 int  i;
1104
1105                 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1106
1107                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1108                         ir_node *arg = get_irn_n(irn,i);
1109
1110                         if(has_reg_class(si, arg)) {
1111                                 pset_insert_ptr(live, arg);
1112                         }
1113                 }
1114         }
1115         /*
1116          * find values that are used by remats at end of block
1117          * and insert them into live set
1118          */
1119         foreach_pre_remat(si, bb, irn) {
1120                 int       n;
1121
1122                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1123                         ir_node        *remat_arg = get_irn_n(irn, n);
1124
1125                         if(!has_reg_class(si, remat_arg)) continue;
1126
1127                         /* if value is becoming live through use by remat */
1128                         if(!pset_find_ptr(live, remat_arg)) {
1129                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1130
1131                                 pset_insert_ptr(live, remat_arg);
1132                         }
1133                 }
1134         }
1135 }
1136
1137 static void
1138 walker_regclass_copy_insertor(ir_node * irn, void * data)
1139 {
1140         spill_ilp_t    *si = data;
1141
1142         if(is_Phi(irn) && has_reg_class(si, irn)) {
1143                 int n;
1144
1145                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1146                         ir_node  *phi_arg = get_irn_n(irn, n);
1147                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1148
1149                         if(!has_reg_class(si, phi_arg)) {
1150                                 ir_node   *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1151                                 ir_node   *pos = sched_block_last_noncf(si, bb);
1152                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
1153
1154                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1155                                 sched_add_after(pos, copy);
1156                                 set_irn_n(irn, n, copy);
1157
1158                                 op->is_remat = 0;
1159                                 op->attr.live_range.args.reloads = NULL;
1160                                 op->attr.live_range.ilp = ILP_UNDEF;
1161                                 set_irn_link(copy, op);
1162                         }
1163                 }
1164         }
1165 }
1166
1167 /**
1168  * Insert (so far unused) remats into the irg to
1169  * recompute the potential liveness of all values
1170  */
1171 static void
1172 walker_remat_insertor(ir_node * bb, void * data)
1173 {
1174         spill_ilp_t    *si = data;
1175         ir_node        *irn;
1176         int             n, i;
1177         pset           *live;
1178         pset           *post_remats;
1179         remat_t        *remat;
1180
1181         /* skip start block, no remats to do there */
1182         if(is_start_block(bb)) return;
1183
1184         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1185
1186         live = pset_new_ptr_default();
1187         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1188                 ir_node        *value = be_lv_get_irn(si->lv, bb, i);
1189
1190                 /* add remats at end of block */
1191                 if (has_reg_class(si, value)) {
1192                         pset_insert_ptr(live, value);
1193                 }
1194         }
1195
1196         irn = sched_last(bb);
1197         while(!sched_is_end(irn)) {
1198                 ir_node   *next;
1199                 pset      *args;
1200                 ir_node   *arg;
1201                 pset      *used;
1202
1203                 next = sched_prev(irn);
1204
1205                 /* delete defined value from live set */
1206                 if(has_reg_class(si, irn)) {
1207                         pset_remove_ptr(live, irn);
1208                 }
1209
1210                 if(is_Phi(irn) || is_Proj(irn)) {
1211                         irn = next;
1212                         continue;
1213                 }
1214
1215                 args = pset_new_ptr_default();
1216                 used = pset_new_ptr_default();
1217
1218                 /* collect arguments of op and set args of op already live in epilog */
1219                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1220                         ir_node        *arg = get_irn_n(irn, n);
1221
1222                         pset_insert_ptr(args, arg);
1223                         if(has_reg_class(si, arg)) {
1224                                 pset_insert_ptr(live, arg);
1225                                 pset_insert_ptr(used, arg);
1226                         }
1227                 }
1228
1229                 /* insert all possible remats before irn */
1230                 pset_foreach(args, arg) {
1231                         remat_info_t   *remat_info,
1232                                                     query;
1233
1234                         /* continue if the operand has the wrong reg class */
1235                         if(!has_reg_class(si, arg))
1236                                 continue;
1237
1238                         query.irn = arg;
1239                         query.remats = NULL;
1240                         query.remats_by_operand = NULL;
1241                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1242
1243                         if(!remat_info) {
1244                                 continue;
1245                         }
1246
1247                         if(remat_info->remats) {
1248                                 pset_foreach(remat_info->remats, remat) {
1249                                         ir_node  *remat_irn = NULL;
1250
1251                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
1252                                         remat_irn = insert_remat_before(si, remat, irn, live);
1253
1254                                         if(remat_irn) {
1255                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1256                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1257
1258                                                         /* collect args of remats which are not args of op */
1259                                                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1260                                                                 pset_insert_ptr(used, remat_arg);
1261                                                         }
1262                                                 }
1263                                         }
1264                                 }
1265                         }
1266                 }
1267
1268                 /* do not place post remats after jumps */
1269                 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1270                         del_pset(used);
1271                         del_pset(args);
1272                         break;
1273                 }
1274
1275                 /* insert all possible remats after irn */
1276                 post_remats = pset_new_ptr_default();
1277                 pset_foreach(used, arg) {
1278                         remat_info_t   *remat_info,
1279                                                     query;
1280
1281                         /* continue if the operand has the wrong reg class */
1282                         if(!has_reg_class(si, arg))
1283                                 continue;
1284
1285                         query.irn = arg;
1286                         query.remats = NULL;
1287                         query.remats_by_operand = NULL;
1288                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1289
1290                         if(!remat_info) {
1291                                 continue;
1292                         }
1293
1294                         if(remat_info->remats_by_operand) {
1295                                 pset_foreach(remat_info->remats_by_operand, remat) {
1296                                         /* do not insert remats producing the same value as one of the operands */
1297                                         if(!pset_find_ptr(args, remat->value)) {
1298                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1299
1300                                                 /* only remat values that can be used by real ops */
1301                                                 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1302                                                         pset_insert_ptr(post_remats, remat);
1303                                                 }
1304                                         }
1305                                 }
1306                         }
1307                 }
1308                 pset_foreach(post_remats, remat) {
1309                         insert_remat_after(si, remat, irn, live);
1310                 }
1311                 del_pset(post_remats);
1312
1313                 del_pset(used);
1314                 del_pset(args);
1315                 irn = next;
1316         }
1317
1318         /* add remats at end if successor has multiple predecessors */
1319         if(is_merge_edge(bb)) {
1320                 pset     *live_out = pset_new_ptr_default();
1321                 ir_node  *value;
1322
1323                 get_live_end(si, bb, live_out);
1324
1325                 /* add remats at end of block */
1326                 pset_foreach(live_out, value) {
1327                         remat_info_t   *remat_info,
1328                                                    query;
1329
1330                         query.irn = value;
1331                         query.remats = NULL;
1332                         query.remats_by_operand = NULL;
1333                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1334
1335                         if(remat_info && remat_info->remats) {
1336                                 pset_foreach(remat_info->remats, remat) {
1337                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1338
1339                                         insert_remat_before(si, remat, bb, live_out);
1340                                 }
1341                         }
1342                 }
1343                 del_pset(live_out);
1344         }
1345
1346         if(is_diverge_edge(bb)) {
1347                 pset     *live_in = pset_new_ptr_default();
1348                 ir_node  *value;
1349
1350                 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1351                         value = be_lv_get_irn(si->lv, bb, i);
1352
1353                         if(has_reg_class(si, value)) {
1354                                 pset_insert_ptr(live_in, value);
1355                         }
1356                 }
1357                 /* add phis to live_in */
1358                 sched_foreach(bb, value) {
1359                         if(!is_Phi(value)) break;
1360
1361                         if(has_reg_class(si, value)) {
1362                                 pset_insert_ptr(live_in, value);
1363                         }
1364                 }
1365
1366                 /* add remat2s at beginning of block */
1367                 post_remats = pset_new_ptr_default();
1368                 pset_foreach(live_in, value) {
1369                         remat_info_t   *remat_info,
1370                                                    query;
1371
1372                         query.irn = value;
1373                         query.remats = NULL;
1374                         query.remats_by_operand = NULL;
1375                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1376
1377                         if(remat_info && remat_info->remats_by_operand) {
1378                                 pset_foreach(remat_info->remats_by_operand, remat) {
1379                                         DBG((si->dbg, LEVEL_4, "\t  considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1380
1381                                         /* put the remat here if all its args are available and result is still live */
1382                                         if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1383                                                 pset_insert_ptr(post_remats, remat);
1384                                         }
1385                                 }
1386                         }
1387                 }
1388                 pset_foreach(post_remats, remat) {
1389                         insert_remat_after(si, remat, bb, live_in);
1390                 }
1391                 del_pset(post_remats);
1392                 del_pset(live_in);
1393         }
1394 }
1395
1396 static int
1397 can_be_copied(const ir_node * bb, const ir_node * irn)
1398 {
1399         const ir_edge_t *edge    = get_block_succ_first(bb);
1400         const ir_node   *next_bb = edge->src;
1401         int             pos      = edge->pos;
1402         const ir_node   *phi;
1403
1404         assert(is_merge_edge(bb));
1405
1406         sched_foreach(next_bb, phi) {
1407                 const ir_node  *phi_arg;
1408
1409                 if(!is_Phi(phi)) break;
1410
1411                 phi_arg = get_irn_n(phi, pos);
1412
1413                 if(phi_arg == irn) {
1414                         return 1;
1415                 }
1416         }
1417         return 0;
1418 }
1419
1420 /**
1421  * Initialize additional node info
1422  */
1423 static void
1424 luke_initializer(ir_node * bb, void * data)
1425 {
1426         spill_ilp_t    *si = (spill_ilp_t*)data;
1427         spill_bb_t     *spill_bb;
1428         ir_node        *irn;
1429
1430         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1431         set_irn_link(bb, spill_bb);
1432
1433         sched_foreach(bb, irn) {
1434                 op_t      *op;
1435
1436                 op = obstack_alloc(si->obst, sizeof(*op));
1437                 op->is_remat = 0;
1438                 op->attr.live_range.ilp = ILP_UNDEF;
1439                 if(is_Phi(irn)) {
1440                         if(opt_memcopies) {
1441                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1442                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1443                         }
1444                 } else if(!is_Proj(irn)) {
1445                         op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1446                         memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1447                 } else {
1448                         op->attr.live_range.args.reloads = NULL;
1449                 }
1450                 set_irn_link(irn, op);
1451         }
1452 }
1453
1454
1455 /**
1456  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1457  */
1458 static void
1459 luke_endwalker(ir_node * bb, void * data)
1460 {
1461         spill_ilp_t    *si = (spill_ilp_t*)data;
1462         pset           *live;
1463         pset           *use_end;
1464         char            buf[256];
1465         ilp_cst_t       cst;
1466         ir_node        *irn;
1467         spill_bb_t     *spill_bb = get_irn_link(bb);
1468         int             i;
1469
1470         live = pset_new_ptr_default();
1471         use_end = pset_new_ptr_default();
1472
1473         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1474                 irn = be_lv_get_irn(si->lv, bb, i);
1475                 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1476                         pset_insert_ptr(live, irn);
1477                 }
1478         }
1479         /*
1480          * find values that are used by remats at end of block
1481          * and insert them into live set
1482          */
1483         foreach_pre_remat(si, bb, irn) {
1484                 int       n;
1485
1486                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1487                         ir_node        *remat_arg = get_irn_n(irn, n);
1488
1489                         if(has_reg_class(si, remat_arg)) {
1490                                 pset_insert_ptr(live, remat_arg);
1491                         }
1492                 }
1493         }
1494
1495         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1496         /* their reg_out must always be set */
1497         sched_foreach_reverse(bb, irn) {
1498                 int   n;
1499
1500                 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1501
1502                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1503                         ir_node        *irn_arg = get_irn_n(irn, n);
1504
1505                         if(has_reg_class(si, irn_arg)) {
1506                                 pset_insert_ptr(use_end, irn_arg);
1507                         }
1508                 }
1509         }
1510
1511         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1512         //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1513         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1514
1515         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1516
1517         /* if this is a merge edge we can reload at the end of this block */
1518         if(is_merge_edge(bb)) {
1519                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1520         } else if(pset_count(use_end)){
1521                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1522         } else {
1523                 spill_bb->reloads = NULL;
1524         }
1525
1526         pset_foreach(live,irn) {
1527                 spill_t     query,
1528                                         *spill;
1529                 double      spill_cost;
1530                 int         default_spilled;
1531
1532
1533                 /* handle values used by control flow nodes later separately */
1534                 if(pset_find_ptr(use_end, irn)) continue;
1535
1536                 query.irn = irn;
1537                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1538
1539                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1540
1541                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1542                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1543                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1544
1545                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1546                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1547
1548                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1549                 /* by default spill value right after definition */
1550                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1551                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1552
1553                 if(is_merge_edge(bb)) {
1554                         ilp_var_t   reload;
1555                         ilp_cst_t   rel_cst;
1556
1557                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1558                         reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1559                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1560
1561                         /* reload <= mem_out */
1562                         rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1563                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1564                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1565                 }
1566
1567                 spill->reg_in = ILP_UNDEF;
1568                 spill->mem_in = ILP_UNDEF;
1569         }
1570
1571         pset_foreach(use_end,irn) {
1572                 spill_t     query,
1573                                         *spill;
1574                 double      spill_cost;
1575                 ilp_cst_t   end_use_req,
1576                                         rel_cst;
1577                 ilp_var_t   reload;
1578                 int         default_spilled;
1579
1580                 query.irn = irn;
1581                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1582
1583                 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1584
1585                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1586                 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1587
1588                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1589                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1590
1591                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1592                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1593                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1594
1595                 /* reload for use be control flow op */
1596                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1597                 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1598                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1599
1600                 /* reload <= mem_out */
1601                 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1602                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1603                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1604
1605                 spill->reg_in = ILP_UNDEF;
1606                 spill->mem_in = ILP_UNDEF;
1607
1608                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1609                 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1610                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1611         }
1612
1613         del_pset(live);
1614         del_pset(use_end);
1615 }
1616
1617 #ifndef NDEBUG
1618 /**
1619  * Find a remat of value @p value in the epilog of @p pos
1620  */
1621 static ir_node *
1622 find_post_remat(const ir_node * value, const ir_node * pos)
1623 {
1624         while((pos = next_post_remat(pos)) != NULL) {
1625                 op_t   *op;
1626
1627                 op = get_irn_link(pos);
1628                 assert(op->is_remat && !op->attr.remat.pre);
1629
1630                 if(op->attr.remat.remat->value == value)
1631                         return (ir_node*)pos;
1632
1633 #if 0
1634         const ir_edge_t *edge;
1635                 foreach_out_edge(pos, edge) {
1636                         ir_node   *proj = get_edge_src_irn(edge);
1637                         assert(is_Proj(proj));
1638                 }
1639 #endif
1640
1641         }
1642
1643         return NULL;
1644 }
1645 #endif
1646
1647 static spill_t *
1648 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1649 {
1650         spill_bb_t  *spill_bb = get_irn_link(bb);
1651         spill_t     *spill,
1652                                  query;
1653         char         buf[256];
1654         int          default_spilled;
1655
1656         query.irn = irn;
1657         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1658         if(!spill) {
1659                 double   spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1660
1661                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1662
1663                 spill->reg_out = ILP_UNDEF;
1664                 spill->reg_in  = ILP_UNDEF;
1665                 spill->mem_in  = ILP_UNDEF;
1666
1667                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1668                 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1669
1670                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1671                 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1672                 spill->spill    = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1673         }
1674
1675         return spill;
1676 }
1677
1678 /**
1679  *  Inserts ILP-constraints and variables for memory copying before the given position
1680  */
1681 static void
1682 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1683 {
1684         const ir_node    *succ;
1685         const ir_edge_t  *edge;
1686         spill_bb_t       *spill_bb = get_irn_link(block);
1687         ir_node          *phi;
1688         int               pos;
1689         ilp_cst_t         cst;
1690         ilp_var_t         copyreg;
1691         char              buf[256];
1692         ir_node          *tmp;
1693
1694
1695         assert(edges_activated(current_ir_graph));
1696
1697         edge = get_block_succ_first(block);
1698         if(!edge) return;
1699
1700         succ = edge->src;
1701         pos = edge->pos;
1702
1703         edge = get_block_succ_next(block, edge);
1704         /* next block can only contain phis, if this is a merge edge */
1705         if(edge) return;
1706
1707         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1708         copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1709
1710         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1711         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1712
1713         pset_foreach(live, tmp) {
1714                 spill_t  *spill;
1715 #if 0
1716                 op_t  *op = get_irn_link(irn);
1717                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1718 #endif
1719                 spill = set_find_spill(spill_bb->ilp, tmp);
1720                 assert(spill);
1721
1722                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1723         }
1724         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1725
1726         sched_foreach(succ, phi) {
1727                 const ir_node  *to_copy;
1728                 op_t           *to_copy_op;
1729                 spill_t        *to_copy_spill;
1730                 op_t           *phi_op = get_irn_link(phi);
1731                 ilp_var_t       reload = ILP_UNDEF;
1732
1733
1734                 if(!is_Phi(phi)) break;
1735                 if(!has_reg_class(si, phi)) continue;
1736
1737                 to_copy = get_irn_n(phi, pos);
1738                 to_copy_op = get_irn_link(to_copy);
1739
1740                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1741                 assert(to_copy_spill);
1742
1743                 if(spill_bb->reloads) {
1744                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1745
1746                         if(keyval) {
1747                                 reload = PTR_TO_INT(keyval->val);
1748                         }
1749                 }
1750
1751                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1752                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1753
1754                 /* copy - reg_out - reload - remat - live_range <= 0 */
1755                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1756                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1757                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1758                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1759                 foreach_pre_remat(si, block, tmp) {
1760                         op_t     *remat_op = get_irn_link(tmp);
1761                         if(remat_op->attr.remat.remat->value == to_copy) {
1762                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1763                         }
1764                 }
1765
1766                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1767                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1768
1769                 /* copy - reg_out - copyreg <= 0 */
1770                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1771                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1772                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1773         }
1774 }
1775
1776
1777 /**
1778  * Walk all irg blocks and emit this ILP
1779  */
1780 static void
1781 luke_blockwalker(ir_node * bb, void * data)
1782 {
1783         spill_ilp_t    *si = (spill_ilp_t*)data;
1784         ir_node        *irn;
1785         pset           *live;
1786         char            buf[256];
1787         ilp_cst_t       cst;
1788         spill_bb_t     *spill_bb = get_irn_link(bb);
1789         ir_node        *tmp;
1790         spill_t        *spill;
1791         pset           *defs = pset_new_ptr_default();
1792         const arch_env_t *arch_env = si->birg->main_env->arch_env;
1793
1794         live = pset_new_ptr_default();
1795
1796         /****************************************
1797          *      B A S I C  B L O C K  E N D
1798          ***************************************/
1799
1800
1801         /* init live values at end of block */
1802         get_live_end(si, bb, live);
1803
1804         pset_foreach(live, irn) {
1805                 op_t           *op;
1806                 ilp_var_t       reload = ILP_UNDEF;
1807
1808                 spill = set_find_spill(spill_bb->ilp, irn);
1809                 assert(spill);
1810
1811                 if(spill_bb->reloads) {
1812                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1813
1814                         if(keyval) {
1815                                 reload = PTR_TO_INT(keyval->val);
1816                         }
1817                 }
1818
1819                 op = get_irn_link(irn);
1820                 assert(!op->is_remat);
1821
1822                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1823                 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1824                 op->attr.live_range.op = bb;
1825
1826                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1827                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1828
1829                 /* reg_out - reload - remat - live_range <= 0 */
1830                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1831                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1832                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1833                 foreach_pre_remat(si, bb, tmp) {
1834                         op_t     *remat_op = get_irn_link(tmp);
1835                         if(remat_op->attr.remat.remat->value == irn) {
1836                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1837                         }
1838                 }
1839                 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1840                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1841
1842                 /* value may only die at bb end if it is used for a mem copy */
1843                 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1844                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1845                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1846                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1847                 foreach_pre_remat(si, bb, tmp) {
1848                         op_t     *remat_op = get_irn_link(tmp);
1849                         if(remat_op->attr.remat.remat->value == irn) {
1850                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1851                         }
1852                 }
1853                 if(is_merge_edge(bb)) {
1854                         const ir_edge_t *edge = get_block_succ_first(bb);
1855                         const ir_node   *next_bb = edge->src;
1856                         int              pos = edge->pos;
1857                         const ir_node   *phi;
1858
1859                         sched_foreach(next_bb, phi) {
1860                                 const ir_node  *phi_arg;
1861
1862                                 if(!is_Phi(phi)) break;
1863
1864                                 phi_arg = get_irn_n(phi, pos);
1865
1866                                 if(phi_arg == irn) {
1867                                         op_t      *phi_op = get_irn_link(phi);
1868                                         ilp_var_t  copy = phi_op->attr.live_range.args.copies[pos];
1869
1870                                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1871                                 }
1872                         }
1873                 }
1874         }
1875
1876         if(opt_memcopies)
1877                 insert_mem_copy_position(si, live, bb);
1878
1879         /*
1880          * assure the remat args are available
1881          */
1882         foreach_pre_remat(si, bb, tmp) {
1883                 op_t     *remat_op = get_irn_link(tmp);
1884                 int       n;
1885
1886                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1887                         ir_node        *remat_arg = get_irn_n(tmp, n);
1888                         op_t           *arg_op = get_irn_link(remat_arg);
1889
1890                         if(!has_reg_class(si, remat_arg)) continue;
1891
1892                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1893                         assert(spill);
1894
1895                         /* arguments of remats have to be live until the very end of the block
1896                          * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1897                          * no remats, they could be in wrong order
1898                          */
1899
1900                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1901                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1902
1903                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1904                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1905                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1906
1907                         /* use reload placed for this argument */
1908                         if(spill_bb->reloads) {
1909                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1910
1911                                 if(keyval) {
1912                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1913
1914                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1915                                 }
1916                         }
1917                 }
1918         }
1919         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1920
1921
1922
1923
1924         /**************************************
1925          *    B A S I C  B L O C K  B O D Y
1926          **************************************/
1927
1928         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1929                 op_t       *op;
1930                 op_t       *tmp_op;
1931                 int         n,
1932                                         u = 0,
1933                                         d = 0;
1934                 ilp_cst_t       check_pre,
1935                                         check_post;
1936                 set        *args;
1937                 pset       *used;
1938                 pset       *remat_defs;
1939                 keyval_t   *keyval;
1940                 ilp_cst_t   one_memoperand = -1;
1941
1942                 /* iterate only until first phi */
1943                 if(is_Phi(irn))
1944                         break;
1945
1946                 op = get_irn_link(irn);
1947                 /* skip remats */
1948                 if(op->is_remat) continue;
1949
1950                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1951
1952                 /* collect defined values */
1953                 if(has_reg_class(si, irn)) {
1954                         pset_insert_ptr(defs, irn);
1955                 }
1956
1957                 /* skip projs */
1958                 if(is_Proj(irn)) continue;
1959
1960                 /*
1961                  * init set of irn's arguments
1962                  * and all possibly used values around this op
1963                  * and values defined by post remats
1964                  */
1965                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1966                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1967                 remat_defs = pset_new_ptr(pset_count(live));
1968
1969                 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1970                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
1971                                 ir_node        *irn_arg = get_irn_n(irn, n);
1972                                 if(has_reg_class(si, irn_arg)) {
1973                                         set_insert_keyval(args, irn_arg, (void*)n);
1974                                         pset_insert_ptr(used, irn_arg);
1975                                 }
1976                         }
1977                         foreach_post_remat(irn, tmp) {
1978                                 op_t    *remat_op = get_irn_link(tmp);
1979
1980                                 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1981
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                         foreach_pre_remat(si, irn, tmp) {
1990                                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1991                                         ir_node        *remat_arg = get_irn_n(tmp, n);
1992                                         if(has_reg_class(si, remat_arg)) {
1993                                                 pset_insert_ptr(used, remat_arg);
1994                                         }
1995                                 }
1996                         }
1997                 }
1998
1999                 /**********************************
2000                  *   I N  E P I L O G  O F  irn
2001                  **********************************/
2002
2003                 /* ensure each dying value is used by only one post remat */
2004                 pset_foreach(used, tmp) {
2005                         ir_node     *value = tmp;
2006                         op_t        *value_op = get_irn_link(value);
2007                         ir_node     *remat;
2008                         int          n_remats = 0;
2009
2010                         cst = ILP_UNDEF;
2011                         foreach_post_remat(irn, remat) {
2012                                 op_t  *remat_op = get_irn_link(remat);
2013
2014                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
2015                                         ir_node   *remat_arg = get_irn_n(remat, n);
2016
2017                                         /* if value is used by this remat add it to constraint */
2018                                         if(remat_arg == value) {
2019                                                 if(n_remats == 0) {
2020                                                         /* sum remat2s <= 1 + n_remats*live_range */
2021                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
2022                                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2023                                                 }
2024
2025                                                 n_remats++;
2026                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2027                                                 break;
2028                                         }
2029                                 }
2030                         }
2031
2032                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2033                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2034                         }
2035                 }
2036
2037         /* ensure at least one value dies at post remat */
2038         foreach_post_remat(irn, tmp) {
2039             op_t     *remat_op = get_irn_link(tmp);
2040             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
2041             ir_node  *remat_arg;
2042
2043             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2044                 remat_arg = get_irn_n(tmp, n);
2045
2046                 if(has_reg_class(si, remat_arg)) {
2047
2048                     /* does arg always die at this op? */
2049                     if(!pset_find_ptr(live, remat_arg))
2050                         goto skip_one_must_die;
2051
2052                     pset_insert_ptr(remat_args, remat_arg);
2053                 }
2054             }
2055
2056             /* remat + \sum live_range(remat_arg) <= |args| */
2057             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2058             cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2059             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2060
2061             pset_foreach(remat_args, remat_arg) {
2062                 op_t  *arg_op = get_irn_link(remat_arg);
2063
2064                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2065             }
2066
2067 skip_one_must_die:
2068             del_pset(remat_args);
2069         }
2070
2071                 /* new live ranges for values from L\U defined by post remats */
2072                 pset_foreach(live, tmp) {
2073                         ir_node     *value = tmp;
2074                         op_t        *value_op = get_irn_link(value);
2075
2076                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2077                                 ilp_var_t    prev_lr = ILP_UNDEF;
2078                                 ir_node     *remat;
2079
2080                                 if(pset_find_ptr(remat_defs, value)) {
2081
2082                                         /* next_live_range <= prev_live_range + sum remat2s */
2083                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2084                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2085
2086                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2087                                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2088
2089                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2090                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2091
2092                                         foreach_post_remat(irn, remat) {
2093                                                 op_t        *remat_op = get_irn_link(remat);
2094
2095                                                 /* if value is being rematerialized by this remat */
2096                                                 if(value == remat_op->attr.remat.remat->value) {
2097                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2098                                                 }
2099                                         }
2100
2101                                         value_op->attr.live_range.ilp = prev_lr;
2102                                         value_op->attr.live_range.op = irn;
2103                                 }
2104                         }
2105                 }
2106
2107                 /* requirements for post remats and start live ranges from L/U' for values dying here */
2108                 foreach_post_remat(irn, tmp) {
2109                         op_t        *remat_op = get_irn_link(tmp);
2110                         int          n;
2111
2112                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2113                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2114                                 op_t           *arg_op = get_irn_link(remat_arg);
2115
2116                                 if(!has_reg_class(si, remat_arg)) continue;
2117
2118                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2119                                 if(!pset_find_ptr(used, remat_arg)) {
2120                                         /* remat <= live_range(remat_arg) */
2121                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2122                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2123
2124                                         /* if value is becoming live through use by remat2 */
2125                                         if(!pset_find_ptr(live, remat_arg)) {
2126                                                 ilp_var_t     lr;
2127
2128                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2129                                                 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2130
2131                                                 arg_op->attr.live_range.ilp = lr;
2132                                                 arg_op->attr.live_range.op = irn;
2133
2134                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2135
2136                                                 pset_insert_ptr(live, remat_arg);
2137                                                 add_to_spill_bb(si, bb, remat_arg);
2138                                         }
2139
2140                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2141                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2142                                 }
2143                         }
2144                 }
2145
2146                 d = pset_count(defs);
2147                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
2148
2149                 /* count how many regs irn needs for arguments */
2150                 u = set_count(args);
2151
2152
2153                 /* check the register pressure in the epilog */
2154                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2155                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2156                 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2157
2158                 /* add L\U' to check_post */
2159                 pset_foreach(live, tmp) {
2160                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2161                                 /* if a live value is not used by irn */
2162                                 tmp_op = get_irn_link(tmp);
2163                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2164                         }
2165                 }
2166
2167                 /***********************************************************
2168                  *  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
2169                  **********************************************************/
2170
2171
2172                 pset_foreach(used, tmp) {
2173                         ilp_var_t       prev_lr;
2174                         ilp_var_t       post_use;
2175                         int             p = 0;
2176                         spill_t        *spill;
2177                         ir_node        *arg = tmp;
2178                         op_t           *arg_op = get_irn_link(arg);
2179                         ir_node        *remat;
2180
2181                         spill = add_to_spill_bb(si, bb, arg);
2182
2183                         /* new live range for each used value */
2184                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2185                         prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2186
2187                         /* the epilog stuff - including post_use, check_post, check_post_remat */
2188                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2189                         post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2190
2191                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2192
2193                         /* arg is live throughout epilog if the next live_range is in a register */
2194                         if(pset_find_ptr(live, arg)) {
2195                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
2196
2197                                 /* post_use >= next_lr + remat */
2198                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2199                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2200                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2201                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2202                         }
2203
2204                         /* forall post remat which use arg add a similar cst */
2205                         foreach_post_remat(irn, remat) {
2206                                 int      n;
2207
2208                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2209                                         ir_node    *remat_arg = get_irn_n(remat, n);
2210                                         op_t       *remat_op = get_irn_link(remat);
2211
2212                                         if(remat_arg == arg) {
2213                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
2214
2215                                                 /* post_use >= remat */
2216                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2217                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2218                                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2219                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2220                                         }
2221                                 }
2222                         }
2223
2224                         /* if value is not an arg of op and not possibly defined by post remat
2225                          * then it may only die and not become live
2226                          */
2227                         if(!set_find_keyval(args, arg)) {
2228                                 /* post_use <= prev_lr */
2229                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2230                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2231                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2232                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2233
2234                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2235                                         /* next_lr <= prev_lr */
2236                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2237                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2238                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2239                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2240                                 }
2241                         }
2242
2243                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2244                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2245                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2246                                                 ilp_var_t       memoperand;
2247
2248                                                 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2249                                                 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2250                                                 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2251
2252                                                 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2253                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2254
2255                                                 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2256                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2257                                         }
2258                                 }
2259                         }
2260
2261                         /* new live range begins for each used value */
2262                         arg_op->attr.live_range.ilp = prev_lr;
2263                         arg_op->attr.live_range.op = irn;
2264
2265                         pset_insert_ptr(live, arg);
2266                 }
2267
2268                 /* just to be sure */
2269                 check_post = ILP_UNDEF;
2270
2271                 /* allow original defintions to be removed */
2272                 if(opt_repair_schedule) {
2273                         pset_foreach(defs, tmp) {
2274                                 op_t      *tmp_op = get_irn_link(tmp);
2275                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2276 #if 1
2277                                 ilp_var_t  delete;
2278                                 assert(spill);
2279
2280                                 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2281                                 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2282
2283                                 /* op may not be killed if its first live_range is 1 */
2284                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2285                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2286                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2287                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2288
2289                                 /* op may not be killed if it is spilled after the definition */
2290                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2291                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2292                                 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2293                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2294 #else
2295                                 ilp_var_t  keep;
2296                                 assert(spill);
2297
2298                                 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2299                                 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2300
2301                                 /* op may not be killed if its first live_range is 1 */
2302                                 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2303                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2304                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2305                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2306
2307                                 /* op may not be killed if it is spilled after the definition */
2308                                 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2309                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2310                                 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2311                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2312 #endif
2313                         }
2314                 } else {
2315 #if 0
2316                         pset_foreach(defs, tmp) {
2317                                 op_t      *tmp_op = get_irn_link(tmp);
2318                                 spill_t   *spill = set_find_spill(spill_bb->ilp, tmp);
2319                                 assert(spill);
2320
2321                                 /* live_range or spill should be 1
2322                                    TODO: lr should be live until first use */
2323                                 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2324                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2325                                 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2326                                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2327                         }
2328 #endif
2329                 }
2330
2331
2332                 /******************
2333                  *   P R O L O G
2334                  ******************/
2335
2336                 /* check the register pressure in the prolog */
2337                 /* sum_{L\U} lr <= k - |U| */
2338                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2339                 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2340
2341                 /* for the prolog remove defined values from the live set */
2342                 pset_foreach(defs, tmp) {
2343                         pset_remove_ptr(live, tmp);
2344                 }
2345
2346                 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2347                         ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2348                         one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2349                 }
2350
2351                 /***********************************************************
2352                  *  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
2353                  **********************************************************/
2354
2355
2356                 set_foreach(args, keyval) {
2357                         spill_t          *spill;
2358                         const ir_node    *arg = keyval->key;
2359                         int               i = PTR_TO_INT(keyval->val);
2360                         op_t             *arg_op = get_irn_link(arg);
2361                         ilp_cst_t         requirements;
2362                         int               n_memoperands;
2363
2364                         spill = set_find_spill(spill_bb->ilp, arg);
2365                         assert(spill);
2366
2367                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2368                         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);
2369
2370                         /* reload <= mem_out */
2371                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2372                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2373                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2374                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2375
2376                         /* requirement: arg must be in register for use */
2377                         /* reload + remat + live_range == 1 */
2378                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2379                         requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2380
2381                         lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2382                         lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2383                         foreach_pre_remat(si, irn, tmp) {
2384                                 op_t     *remat_op = get_irn_link(tmp);
2385                                 if(remat_op->attr.remat.remat->value == arg) {
2386                                         lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2387                                 }
2388                         }
2389
2390                         if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2391                                 n_memoperands = 0;
2392                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2393                                         if(get_irn_n(irn, n) == arg) {
2394                                                 n_memoperands++;
2395                                         }
2396                                 }
2397                                 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2398                                         if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2399                                                 memoperand_t  *memoperand;
2400                                                 memoperand = set_find_memoperand(si->memoperands, irn, n);
2401
2402                                                 /* memoperand <= mem_out */
2403                                                 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2404                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2405                                                 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2406                                                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2407
2408                                                 /* the memoperand is only sufficient if it is used once by the op */
2409                                                 if(n_memoperands == 1)
2410                                                         lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2411
2412                                                 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2413
2414                                                 /* we have one more free register if we use a memory operand */
2415                                                 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2416                                         }
2417                                 }
2418                         }
2419                 }
2420
2421                 /* iterate over L\U */
2422                 pset_foreach(live, tmp) {
2423                         if(!set_find_keyval(args, tmp)) {
2424                                 /* if a live value is not used by irn */
2425                                 tmp_op = get_irn_link(tmp);
2426                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2427                         }
2428                 }
2429
2430                 /* requirements for remats */
2431                 foreach_pre_remat(si, irn, tmp) {
2432                         op_t        *remat_op = get_irn_link(tmp);
2433                         int          n;
2434
2435                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2436                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2437                                 op_t           *arg_op = get_irn_link(remat_arg);
2438
2439                                 if(!has_reg_class(si, remat_arg)) continue;
2440
2441                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2442                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2443                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2444
2445                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2446                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2447
2448                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2449                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2450                                         int    index = (int)keyval->val;
2451
2452                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2453                                 }
2454                         }
2455                 }
2456
2457
2458
2459
2460                 /*************************
2461                  *  D O N E  W I T H  O P
2462                  *************************/
2463
2464                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2465
2466                 pset_foreach(live, tmp) {
2467                         assert(has_reg_class(si, tmp));
2468                 }
2469
2470 #ifndef NDEBUG
2471                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2472                         ir_node        *arg = get_irn_n(irn, n);
2473
2474                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2475                 }
2476 #endif
2477
2478                 del_pset(remat_defs);
2479                 del_pset(used);
2480                 del_set(args);
2481                 del_pset(defs);
2482                 defs = pset_new_ptr_default();
2483
2484                 /* skip everything above barrier in start block */
2485                 if(is_start_block(bb) && be_is_Barrier(irn)) {
2486                         assert(pset_count(live) == 0);
2487                         break;
2488                 }
2489
2490         }
2491         del_pset(defs);
2492
2493
2494
2495         /***************************************
2496          *   B E G I N N I N G  O F  B L O C K
2497          ***************************************/
2498
2499
2500         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2501         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2502
2503         pset_foreach(live, irn) {
2504                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2505         }
2506
2507         /* construct mem_outs for all values */
2508         set_foreach(spill_bb->ilp, spill) {
2509                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2510                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2511
2512                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2513                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2514
2515                 if(pset_find_ptr(live, spill->irn)) {
2516                         int default_spilled;
2517                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2518
2519                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2520                         default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2521                         spill->mem_in   = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2522                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2523
2524                         if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2525                                 int   n;
2526                                 op_t *op = get_irn_link(spill->irn);
2527
2528                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2529                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2530                                         double          freq=0.0;
2531                                         int             m;
2532                                         ilp_var_t       var;
2533
2534
2535                                         /* argument already done? */
2536                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2537
2538                                         /* get sum of execution frequencies of blocks with the same phi argument */
2539                                         for(m=n; m>=0; --m) {
2540                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2541
2542                                                 if(arg==arg2) {
2543                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2544                                                 }
2545                                         }
2546
2547                                         /* copies are not for free */
2548                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2549                                         var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2550
2551                                         for(m=n; m>=0; --m) {
2552                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2553
2554                                                 if(arg==arg2) {
2555                                                         op->attr.live_range.args.copies[m] = var;
2556                                                 }
2557                                         }
2558
2559 #if 0
2560                                         /* copy <= mem_in */
2561                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2562                                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2563                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2564                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2565 #endif
2566                                 }
2567                         }
2568                 }
2569         }
2570
2571         foreach_post_remat(bb, tmp) {
2572                 int         n;
2573                 op_t       *remat_op = get_irn_link(tmp);
2574                 pset       *remat_args = pset_new_ptr(get_irn_arity(tmp));
2575                 ir_node    *remat_arg;
2576
2577                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2578                         remat_arg = get_irn_n(tmp, n);
2579
2580                         if(has_reg_class(si, remat_arg)) {
2581                                 pset_insert_ptr(remat_args, remat_arg);
2582                         }
2583                 }
2584
2585                 /* remat + \sum live_range(remat_arg) <= |args| */
2586                 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2587                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2588                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2589
2590                 pset_foreach(remat_args, remat_arg) {
2591                         if(pset_find_ptr(live, remat_arg)) {
2592                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2593                                 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2594                         }
2595                 }
2596                 del_pset(remat_args);
2597         }
2598
2599         foreach_post_remat(bb, tmp) {
2600                 int  n;
2601
2602                 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2603                         ir_node  *remat_arg = get_irn_n(tmp, n);
2604
2605                         /* if value is becoming live through use by remat2 */
2606                         if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2607                                 op_t       *remat_arg_op = get_irn_link(remat_arg);
2608                                 ilp_cst_t   nomem;
2609
2610                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2611
2612                                 pset_insert_ptr(live, remat_arg);
2613                                 spill = add_to_spill_bb(si, bb, remat_arg);
2614                                 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2615
2616                                 /* we need reg_in and mem_in for this value; they will be referenced later */
2617                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2618                                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2619                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2620                                 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2621
2622
2623                                 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2624                                 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2625                                 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2626                                 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2627                         }
2628                 }
2629         }
2630
2631         /* L\U is empty at bb start */
2632         /* arg is live throughout epilog if it is reg_in into this block */
2633
2634         /* check the register pressure at the beginning of the block
2635          * including remats
2636          */
2637         /* reg_in entspricht post_use */
2638
2639         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2640         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2641
2642         pset_foreach(live, irn) {
2643         ilp_cst_t  nospill;
2644
2645                 spill = set_find_spill(spill_bb->ilp, irn);
2646                 assert(spill);
2647
2648                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2649                 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2650
2651                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2652
2653                 /* spill + mem_in <= 1 */
2654                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2655                 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2656
2657                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2658                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2659
2660         } /* post_remats are NOT included in register pressure check because
2661            they do not increase regpressure */
2662
2663         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2664         pset_foreach(live, irn) {
2665                 int          p = 0,
2666                                          n;
2667
2668                 spill = set_find_spill(spill_bb->ilp, irn);
2669                 assert(spill && spill->irn == irn);
2670
2671                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2672                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2673                                 ilp_cst_t       mem_in,
2674                                                                 reg_in;
2675                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2676                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2677                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2678                                 spill_t        *spill_p;
2679                                 op_t           *op = get_irn_link(irn);
2680
2681                                 /* although the phi is in the right regclass one or more of
2682                                  * its arguments can be in a different one or at least to
2683                                  * ignore
2684                                  */
2685                                 if(has_reg_class(si, phi_arg)) {
2686                                         /* mem_in < mem_out_arg + copy */
2687                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2688                                         mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2689
2690                                         /* reg_in < reg_out_arg */
2691                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2692                                         reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2693
2694                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2695                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2696
2697                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2698                                         assert(spill_p);
2699
2700                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2701                                         if(opt_memcopies)
2702                                                 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2703
2704                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2705                                 }
2706                         }
2707                 } else {
2708                         /* else assure the value arrives on all paths in the same resource */
2709
2710                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2711                                 ilp_cst_t       mem_in,
2712                                                                 reg_in;
2713                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2714                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2715                                 spill_t        *spill_p;
2716
2717                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2718                                 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2719                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2720                                 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2721
2722                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2723                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2724
2725                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2726                                 assert(spill_p);
2727
2728                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2729                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2730                         }
2731                 }
2732         }
2733
2734         foreach_post_remat(bb, tmp) {
2735                 int         n;
2736
2737                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2738                         ir_node    *remat_arg = get_irn_n(tmp, n);
2739                         op_t       *remat_op = get_irn_link(tmp);
2740
2741                         if(!has_reg_class(si, remat_arg)) continue;
2742
2743                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2744                         assert(spill);
2745
2746                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2747                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2748                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2749                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2750                 }
2751         }
2752
2753         pset_foreach(live, irn) {
2754                 const op_t      *op = get_irn_link(irn);
2755                 const ir_node   *remat;
2756                 int              n_remats = 0;
2757
2758                 cst = ILP_UNDEF;
2759
2760                 foreach_post_remat(bb, remat) {
2761                         int   n;
2762
2763                         for (n=get_irn_arity(remat)-1; n>=0; --n) {
2764                                 const ir_node  *arg = get_irn_n(remat, n);
2765
2766                                 if(arg == irn) {
2767                                         const op_t   *remat_op = get_irn_link(remat);
2768
2769                                         if(cst == ILP_UNDEF) {
2770                                                 /* sum remat2s <= 1 + n_remats*live_range */
2771                                                 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2772                                                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2773                                         }
2774                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2775                                         ++n_remats;
2776                                         break;
2777                                 }
2778                         }
2779                 }
2780                 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2781                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2782                 }
2783         }
2784
2785         /* first live ranges from reg_ins */
2786         pset_foreach(live, irn) {
2787                 op_t      *op = get_irn_link(irn);
2788
2789                 if(op->attr.live_range.ilp != ILP_UNDEF) {
2790
2791                         spill = set_find_spill(spill_bb->ilp, irn);
2792                         assert(spill && spill->irn == irn);
2793
2794                         ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2795                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2796                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2797                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2798
2799                         foreach_post_remat(bb, tmp) {
2800                                 op_t     *remat_op = get_irn_link(tmp);
2801
2802                                 if(remat_op->attr.remat.remat->value == irn) {
2803                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2804                                 }
2805                         }
2806                 }
2807         }
2808
2809         /* walk forward now and compute constraints for placing spills */
2810         /* this must only be done for values that are not defined in this block */
2811         pset_foreach(live, irn) {
2812                 /*
2813                  * if value is defined in this block we can anways place the spill directly after the def
2814                  *    -> no constraint necessary
2815                  */
2816                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2817                         assert(0);
2818                 }
2819
2820
2821                 spill = set_find_spill(spill_bb->ilp, irn);
2822                 assert(spill);
2823
2824                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2825                 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2826
2827                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2828                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2829
2830                 if(!is_Phi(irn)) {
2831                         sched_foreach_op(bb, tmp) {
2832                                 op_t   *op = get_irn_link(tmp);
2833
2834                                 if(is_Phi(tmp)) continue;
2835                                 assert(!is_Proj(tmp));
2836
2837                                 if(op->is_remat) {
2838                                         const ir_node   *value = op->attr.remat.remat->value;
2839
2840                                         if(value == irn) {
2841                                                 /* only collect remats up to the first real use of a value */
2842                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2843                                         }
2844                                 } else {
2845                                         int   n;
2846
2847                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2848                                                 ir_node    *arg = get_irn_n(tmp, n);
2849
2850                                                 if(arg == irn) {
2851                                                         /* if a value is used stop collecting remats */
2852                             goto next_live;
2853                                                 }
2854                                         }
2855                                 }
2856                         }
2857                 }
2858 next_live: ;
2859         }
2860
2861         del_pset(live);
2862 }
2863
2864 typedef struct _irnlist_t {
2865         struct list_head   list;
2866         ir_node           *irn;
2867 } irnlist_t;
2868
2869 typedef struct _interference_t {
2870         struct list_head    blocklist;
2871         ir_node            *a;
2872         ir_node            *b;
2873 } interference_t;
2874
2875 static int
2876 cmp_interference(const void *a, const void *b, size_t size)
2877 {
2878         const interference_t *p = a;
2879         const interference_t *q = b;
2880         (void) size;
2881
2882         return !(p->a == q->a && p->b == q->b);
2883 }
2884
2885 static interference_t *
2886 set_find_interference(set * set, ir_node * a, ir_node * b)
2887 {
2888         interference_t     query;
2889
2890         query.a = (a>b)?a:b;
2891         query.b = (a>b)?b:a;
2892
2893         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2894 }
2895
2896 static interference_t *
2897 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2898 {
2899         interference_t     query,
2900                                           *result;
2901         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2902
2903         list->irn = bb;
2904
2905         result = set_find_interference(set, a, b);
2906         if(result) {
2907
2908                 list_add(&list->list, &result->blocklist);
2909                 return result;
2910         }
2911
2912         query.a = (a>b)?a:b;
2913         query.b = (a>b)?b:a;
2914
2915         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2916
2917         INIT_LIST_HEAD(&result->blocklist);
2918         list_add(&list->list, &result->blocklist);
2919
2920         return result;
2921 }
2922
2923 static
2924 int values_interfere_in_block(const spill_ilp_t *si, const ir_node *bb, const ir_node *a, const ir_node *b)
2925 {
2926         const ir_edge_t *edge;
2927
2928         if (get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2929                 /* both values are live in, so they interfere */
2930                 return 1;
2931         }
2932
2933         /* ensure a dominates b */
2934         if (value_dominates(b, a)) {
2935                 const ir_node *t;
2936                 t = b;
2937                 b = a;
2938                 a = t;
2939         }
2940         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2941
2942
2943         /* the following code is stolen from bera.c */
2944         if (be_is_live_end(si->lv, bb, a))
2945                 return 1;
2946
2947         foreach_out_edge(a, edge) {
2948                 const ir_node *user = edge->src;
2949                 if (get_nodes_block(user) == bb
2950                                 && ! is_Phi(user)
2951                                 && b != user
2952                                 && ! pset_find_ptr(si->inverse_ops, user)
2953                                 && value_dominates(b, user))
2954                         return 1;
2955         }
2956
2957         return 0;
2958 }
2959
2960 /**
2961  * Walk all irg blocks and collect interfering values inside of phi classes
2962  */
2963 static void
2964 luke_interferencewalker(ir_node * bb, void * data)
2965 {
2966         spill_ilp_t    *si = (spill_ilp_t*)data;
2967         int             l1, l2;
2968
2969         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2970                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
2971                 op_t           *a_op = get_irn_link(a);
2972
2973
2974                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2975                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2976                         if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2977                                 continue;
2978
2979                         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)) {
2980                                 ir_node *b    = be_lv_get_irn(si->lv, bb, l2);
2981                                 op_t    *b_op = get_irn_link(b);
2982
2983                                 /* a and b are only interesting if they are in the same phi class */
2984                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2985                                         if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2986                                                 continue;
2987
2988                                         if (values_interfere_in_block(si, bb, a, b)) {
2989                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2990                                                 set_insert_interference(si, si->interferences, a, b, bb);
2991                                         }
2992                                 }
2993                         }
2994                 }
2995         }
2996 }
2997
2998 static unsigned int copy_path_id = 0;
2999
3000 static void
3001 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
3002 {
3003         ilp_cst_t  cst;
3004         ilp_var_t  copy;
3005         char       buf[256];
3006         void      *ptr;
3007
3008         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
3009         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3010
3011         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
3012
3013         pset_foreach(copies, ptr) {
3014                 copy = PTR_TO_INT(ptr);
3015                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
3016         }
3017 }
3018
3019 /**
3020  * @parameter copies   contains a path of copies which lead us to irn
3021  * @parameter visited  contains a set of nodes already visited on this path
3022  */
3023 static int
3024 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3025 {
3026         const ir_edge_t *edge;
3027         op_t            *op = get_irn_link(irn);
3028     pset            *visited_users = pset_new_ptr_default();
3029         int              paths = 0;
3030
3031         if(op->is_remat) return 0;
3032
3033         pset_insert_ptr(visited, irn);
3034
3035         if(is_Phi(irn)) {
3036                 int    n;
3037         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
3038
3039                 /* visit all operands */
3040                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3041                         ir_node  *arg = get_irn_n(irn, n);
3042                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
3043
3044                         if(!has_reg_class(si, arg)) continue;
3045             if(pset_find_ptr(visited_operands, arg)) continue;
3046             pset_insert_ptr(visited_operands, arg);
3047
3048                         if(arg == target) {
3049                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3050                                         del_pset(visited_operands);
3051                                         del_pset(visited_users);
3052                                         pset_remove_ptr(visited, irn);
3053                                         return paths;
3054                                 }
3055                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3056                                 write_copy_path_cst(si, copies, any_interfere);
3057                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3058                         } else if(!pset_find_ptr(visited, arg)) {
3059                                 pset_insert(copies, INT_TO_PTR(copy), copy);
3060                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3061                                 pset_remove(copies, INT_TO_PTR(copy), copy);
3062
3063                 if(paths > MAX_PATHS) {
3064                     if(pset_count(copies) == 0) {
3065                         ilp_cst_t  cst;
3066                         char       buf[256];
3067
3068                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3069                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3070                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3071                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3072                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3073
3074                         paths = 0;
3075                     } else {
3076                         del_pset(visited_operands);
3077                         del_pset(visited_users);
3078                         pset_remove_ptr(visited, irn);
3079                         return paths;
3080                     }
3081                 } else if(pset_count(copies) == 0) {
3082                                         paths = 0;
3083                                 }
3084                         }
3085                 }
3086
3087         del_pset(visited_operands);
3088         }
3089
3090         /* visit all uses which are phis */
3091         foreach_out_edge(irn, edge) {
3092                 ir_node  *user = edge->src;
3093                 int       pos  = edge->pos;
3094                 op_t     *op = get_irn_link(user);
3095                 ilp_var_t copy;
3096
3097                 if(!is_Phi(user)) continue;
3098                 if(!has_reg_class(si, user)) continue;
3099         if(pset_find_ptr(visited_users, user)) continue;
3100         pset_insert_ptr(visited_users, user);
3101
3102                 copy = op->attr.live_range.args.copies[pos];
3103
3104                 if(user == target) {
3105                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3106                                 del_pset(visited_users);
3107                                 pset_remove_ptr(visited, irn);
3108                                 return paths;
3109                         }
3110                         pset_insert(copies, INT_TO_PTR(copy), copy);
3111                         write_copy_path_cst(si, copies, any_interfere);
3112                         pset_remove(copies, INT_TO_PTR(copy), copy);
3113                 } else if(!pset_find_ptr(visited, user)) {
3114                         pset_insert(copies, INT_TO_PTR(copy), copy);
3115                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3116                         pset_remove(copies, INT_TO_PTR(copy), copy);
3117
3118             if(paths > MAX_PATHS) {
3119                 if(pset_count(copies) == 0) {
3120                     ilp_cst_t  cst;
3121                     char       buf[256];
3122
3123                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3124                     cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3125                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3126                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3127                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3128
3129                     paths = 0;
3130                 } else {
3131                     del_pset(visited_users);
3132                     pset_remove_ptr(visited, irn);
3133                     return paths;
3134                 }
3135             } else if(pset_count(copies) == 0) {
3136                                 paths = 0;
3137                         }
3138                 }
3139         }
3140
3141     del_pset(visited_users);
3142         pset_remove_ptr(visited, irn);
3143         return paths;
3144 }
3145
3146 static void
3147 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3148 {
3149         pset * copies = pset_new_ptr_default();
3150         pset * visited = pset_new_ptr_default();
3151
3152         find_copy_path(si, a, b, any_interfere, copies, visited);
3153
3154         del_pset(visited);
3155         del_pset(copies);
3156 }
3157
3158
3159 static void
3160 memcopyhandler(spill_ilp_t * si)
3161 {
3162         interference_t   *interference;
3163         char              buf[256];
3164         /* teste Speicherwerte auf Interferenz */
3165
3166         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3167         irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3168
3169         /* now lets emit the ILP unequations for the crap */
3170         set_foreach(si->interferences, interference) {
3171                 irnlist_t      *irnlist;
3172                 ilp_var_t      interfere, any_interfere;
3173                 ilp_cst_t      any_interfere_cst, cst;
3174                 const ir_node  *a  = interference->a;
3175                 const ir_node  *b  = interference->b;
3176
3177                 /* any_interf <= \sum interf */
3178                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3179                 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3180                 any_interfere     = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3181
3182                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3183
3184                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3185                         const ir_node  *bb = irnlist->irn;
3186                         spill_bb_t     *spill_bb = get_irn_link(bb);
3187                         spill_t        *spilla,
3188                                                    *spillb;
3189                         char           buf[256];
3190
3191                         spilla = set_find_spill(spill_bb->ilp, a);
3192                         assert(spilla);
3193
3194                         spillb = set_find_spill(spill_bb->ilp, b);
3195                         assert(spillb);
3196
3197                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3198                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3199                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
3200                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
3201                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3202                         interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3203
3204                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3205                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3206
3207                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3208                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3209                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3210                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3211                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3212
3213                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3214                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3215
3216                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3217                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3218                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3219
3220                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3221                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3222
3223                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3224                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3225                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3226
3227
3228                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3229
3230                         /* any_interfere >= interf */
3231                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3232                         cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3233
3234                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3235                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3236                 }
3237
3238                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3239                 gen_copy_constraints(si,a,b,any_interfere);
3240         }
3241 }
3242
3243
3244 static INLINE int
3245 is_zero(double x)
3246 {
3247         return fabs(x) < 0.00001;
3248 }
3249
3250 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3251 {
3252         spill_ilp_t *si = get_irg_link(current_ir_graph);
3253         (void) l;
3254
3255         if(pset_find_ptr(si->all_possible_remats, n)) {
3256                 op_t   *op = (op_t*)get_irn_link(n);
3257                 assert(op && op->is_remat);
3258
3259                 if(!op->attr.remat.remat->inverse) {
3260                         if(op->attr.remat.pre) {
3261                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3262                         } else {
3263                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3264                         }
3265
3266                         return 1;
3267                 } else {
3268                         op_t   *op = (op_t*)get_irn_link(n);
3269                         assert(op && op->is_remat);
3270
3271                         if(op->attr.remat.pre) {
3272                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3273                         } else {
3274                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3275                         }
3276
3277                         return 1;
3278                 }
3279         }
3280
3281         return 0;
3282 }
3283
3284 static void
3285 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3286 {
3287         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3288         be_dump(irg, suffix, dump_ir_block_graph_sched);
3289         set_dump_node_vcgattr_hook(NULL);
3290 }
3291
3292 /**
3293  * Edge hook to dump the schedule edges with annotated register pressure.
3294  */
3295 static int
3296 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3297 {
3298         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3299                 ir_node *prev = sched_prev(irn);
3300                 fprintf(F, "edge:{sourcename:\"");
3301                 PRINT_NODEID(irn);
3302                 fprintf(F, "\" targetname:\"");
3303                 PRINT_NODEID(prev);
3304                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3305                 fprintf(F, "\" color:magenta}\n");
3306         }
3307         return 1;
3308 }
3309
3310 static void
3311 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3312 {
3313         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3314
3315         dump_consts_local(0);
3316         set_dump_node_edge_hook(sched_pressure_edge_hook);
3317         dump_ir_block_graph(irg, suffix);
3318         set_dump_node_edge_hook(old_edge_hook);
3319 }
3320
3321 static void
3322 walker_pressure_annotator(ir_node * bb, void * data)
3323 {
3324         spill_ilp_t  *si = data;
3325         ir_node      *irn;
3326         int           n, i;
3327         pset         *live = pset_new_ptr_default();
3328         int           projs = 0;
3329
3330         be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3331                 irn = be_lv_get_irn(si->lv, bb, i);
3332
3333                 if (has_reg_class(si, irn)) {
3334                         pset_insert_ptr(live, irn);
3335                 }
3336         }
3337
3338         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3339
3340         sched_foreach_reverse(bb, irn) {
3341                 if(is_Phi(irn)) {
3342                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3343                         continue;
3344                 }
3345
3346                 if(has_reg_class(si, irn)) {
3347                         pset_remove_ptr(live, irn);
3348                         if(is_Proj(irn)) ++projs;
3349                 }
3350
3351                 if(!is_Proj(irn)) projs = 0;
3352
3353                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3354                         ir_node    *arg = get_irn_n(irn, n);
3355
3356                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3357                 }
3358                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3359         }
3360
3361         del_pset(live);
3362 }
3363
3364 static void
3365 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3366 {
3367         be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3368 }
3369
3370 static void
3371 connect_all_remats_with_keep(spill_ilp_t * si)
3372 {
3373         ir_node   *irn;
3374         ir_node  **ins,
3375                          **pos;
3376         int        n_remats;
3377
3378
3379         n_remats = pset_count(si->all_possible_remats);
3380         if(n_remats) {
3381                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3382
3383                 pos = ins;
3384                 pset_foreach(si->all_possible_remats, irn) {
3385                         *pos = irn;
3386                         ++pos;
3387                 }
3388
3389                 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3390
3391                 obstack_free(si->obst, ins);
3392         }
3393 }
3394
3395 static void
3396 connect_all_spills_with_keep(spill_ilp_t * si)
3397 {
3398         ir_node   *irn;
3399         ir_node  **ins,
3400                          **pos;
3401         int        n_spills;
3402         ir_node   *keep;
3403
3404
3405         n_spills = pset_count(si->spills);
3406         if(n_spills) {
3407                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3408
3409                 pos = ins;
3410                 pset_foreach(si->spills, irn) {
3411                         *pos = irn;
3412                         ++pos;
3413                 }
3414
3415                 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3416
3417                 obstack_free(si->obst, ins);
3418         }
3419 }
3420
3421 /** insert a spill at an arbitrary position */
3422 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3423 {
3424         ir_node  *bl    = is_Block(insert) ? insert : get_nodes_block(insert);
3425         ir_graph *irg   = get_irn_irg(bl);
3426         ir_node  *frame = get_irg_frame(irg);
3427         ir_node  *spill;
3428         ir_node  *next;
3429         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
3430         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3431
3432         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3433
3434         /*
3435          * search the right insertion point. a spill of a phi cannot be put
3436          * directly after the phi, if there are some phis behind the one which
3437          * is spilled. Also, a spill of a Proj must be after all Projs of the
3438          * same tuple node.
3439          *
3440          * Here's one special case:
3441          * If the spill is in the start block, the spill must be after the frame
3442          * pointer is set up. This is done by setting insert to the end of the block
3443          * which is its default initialization (see above).
3444          */
3445
3446         if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3447                 insert = frame;
3448
3449         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3450                 insert = next;
3451
3452         sched_add_after(insert, spill);
3453         return spill;
3454 }
3455
3456 static void
3457 delete_remat(spill_ilp_t * si, ir_node * remat) {
3458         int       n;
3459         ir_node  *bad = get_irg_bad(si->birg->irg);
3460
3461         sched_remove(remat);
3462
3463         /* kill links to operands */
3464         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3465                 set_irn_n(remat, n, bad);
3466         }
3467 }
3468
3469 static void
3470 clean_remat_info(spill_ilp_t * si)
3471 {
3472         int            n;
3473         remat_t       *remat;
3474         remat_info_t  *remat_info;
3475         ir_node       *bad = get_irg_bad(si->birg->irg);
3476
3477         set_foreach(si->remat_info, remat_info) {
3478                 if(!remat_info->remats) continue;
3479
3480                 pset_foreach(remat_info->remats, remat)
3481                 {
3482                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3483                                 if(sched_is_scheduled(remat->proj)) {
3484                                         sched_remove((ir_node*)remat->proj);
3485                                 }
3486                                 set_irn_n((ir_node*)remat->proj, -1, bad);
3487                                 set_irn_n((ir_node*)remat->proj, 0, bad);
3488                         }
3489
3490                         if(get_irn_n_edges(remat->op) == 0) {
3491                                 if(sched_is_scheduled(remat->op)) {
3492                                         sched_remove((ir_node*)remat->op);
3493                                 }
3494                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3495                                         set_irn_n((ir_node*)remat->op, n, bad);
3496                                 }
3497                         }
3498                 }
3499
3500                 if(remat_info->remats) del_pset(remat_info->remats);
3501                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3502         }
3503 }
3504
3505 static void
3506 delete_unnecessary_remats(spill_ilp_t * si)
3507 {
3508         if(opt_keep_alive & KEEPALIVE_REMATS) {
3509                 int       n;
3510                 ir_node  *bad = get_irg_bad(si->birg->irg);
3511
3512                 if(si->keep) {
3513                         for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3514                                 ir_node        *keep_arg = get_irn_n(si->keep, n);
3515                                 op_t           *arg_op = get_irn_link(keep_arg);
3516                                 lpp_name_t     *name;
3517
3518                                 assert(arg_op->is_remat);
3519
3520                                 name = si->lpp->vars[arg_op->attr.remat.ilp];
3521
3522                                 if(is_zero(name->value)) {
3523                                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3524                                         /* TODO check whether reload is preferred over remat (could be bug) */
3525                                         delete_remat(si, keep_arg);
3526                                 } else {
3527                                         if(!arg_op->attr.remat.remat->inverse) {
3528                                                 if(arg_op->attr.remat.pre) {
3529                                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3530                                                 } else {
3531                                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3532                                                 }
3533                                         } else {
3534                                                 if(arg_op->attr.remat.pre) {
3535                                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3536                                                 } else {
3537                                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3538                                                 }
3539                                         }
3540                                 }
3541
3542                                 set_irn_n(si->keep, n, bad);
3543                         }
3544                 } else {
3545                         DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3546                 }
3547         } else {
3548                 ir_node  *remat;
3549
3550                 pset_foreach(si->all_possible_remats, remat) {
3551                         op_t           *remat_op = get_irn_link(remat);
3552                         lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3553
3554                         if(is_zero(name->value)) {
3555                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3556                                 /* TODO check whether reload is preferred over remat (could be bug) */
3557                                 delete_remat(si, remat);
3558                         } else {
3559                                 if(!remat_op->attr.remat.remat->inverse) {
3560                                         if(remat_op->attr.remat.pre) {
3561                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3562                                         } else {
3563                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3564                                         }
3565                                 } else {
3566                                         if(remat_op->attr.remat.pre) {
3567                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3568                                         } else {
3569                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3570                                         }
3571                                 }
3572                         }
3573                 }
3574         }
3575 }
3576
3577 static pset *
3578 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3579 {
3580         pset     *spills = pset_new_ptr_default();
3581
3582         const ir_node  *next;
3583         defs_t         *defs;
3584
3585         defs = set_find_def(si->values, value);
3586
3587         if(defs && defs->spills) {
3588                 for(next = defs->spills; next; next = get_irn_link(next)) {
3589                         pset_insert_ptr(spills, next);
3590                 }
3591         }
3592
3593         return spills;
3594 }
3595
3596 static ir_node *
3597 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3598 {
3599         ir_node  *res;
3600
3601         assert( get_irn_arity(block) == arity );
3602
3603         res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3604         res->attr.phi.u.backedge = new_backedge_arr(irg->obst, arity);
3605
3606         return res;
3607 }
3608
3609 /**
3610  * @param before   The node after which the spill will be placed in the schedule
3611  */
3612 static ir_node *
3613 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3614 {
3615         defs_t   *defs;
3616         ir_node  *spill;
3617         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3618
3619         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3620
3621         spill = be_spill2(arch_env, irn, before);
3622
3623         defs = set_insert_def(si->values, value);
3624         assert(defs);
3625
3626         /* enter into the linked list */
3627         set_irn_link(spill, defs->spills);
3628         defs->spills = spill;
3629
3630         if(opt_keep_alive & KEEPALIVE_SPILLS)
3631                 pset_insert_ptr(si->spills, spill);
3632
3633         return spill;
3634 }
3635
3636 /**
3637  * @param before   The Phi node which has to be spilled
3638  */
3639 static ir_node *
3640 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3641 {
3642         ir_node   *mem_phi;
3643         ir_node  **ins;
3644         defs_t    *defs;
3645         int        n;
3646
3647         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3648
3649         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3650                 ins[n] = si->m_unknown;
3651         }
3652
3653         mem_phi =  new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3654
3655         defs = set_insert_def(si->values, phi);
3656         assert(defs);
3657
3658         /* enter into the linked list */
3659         set_irn_link(mem_phi, defs->spills);
3660         defs->spills = mem_phi;
3661
3662 #ifdef SCHEDULE_PHIM
3663         sched_add_after(phi, mem_phi);
3664 #else
3665         pset_insert_ptr(si->phims, mem_phi);
3666 #endif
3667
3668         if(opt_keep_alive & KEEPALIVE_SPILLS)
3669                 pset_insert_ptr(si->spills, mem_phi);
3670
3671
3672         return mem_phi;
3673 }
3674
3675 /**
3676  * Add remat to list of defs, destroys link field!
3677  */
3678 static void
3679 insert_remat(spill_ilp_t * si, ir_node * remat)
3680 {
3681         defs_t   *defs;
3682         op_t     *remat_op = get_irn_link(remat);
3683
3684         assert(remat_op->is_remat);
3685
3686         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3687         assert(defs);
3688
3689         /* enter into the linked list */
3690         set_irn_link(remat, defs->remats);
3691         defs->remats = remat;
3692 }
3693
3694
3695 /**
3696  * Add reload before operation and add to list of defs
3697  */
3698 static ir_node *
3699 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3700 {
3701         defs_t   *defs;
3702         ir_node  *reload,
3703                          *spill;
3704         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3705
3706         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3707
3708         defs = set_find_def(si->values, value);
3709
3710         spill = defs->spills;
3711         assert(spill && "no spill placed before reload");
3712
3713         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3714
3715         /* enter into the linked list */
3716         set_irn_link(reload, defs->remats);
3717         defs->remats = reload;
3718
3719         return reload;
3720 }
3721
3722 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3723 {
3724         defs_t           *defs;
3725         ir_node          *value = get_irn_n(memoperand->irn, memoperand->pos);
3726         ir_node          *spill;
3727         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3728
3729         DBG((si->dbg, LEVEL_2, "\t  inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3730
3731         defs = set_find_def(si->values, value);
3732
3733         spill = defs->spills;
3734         assert(spill && "no spill placed before reload");
3735
3736         arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3737 }
3738
3739 void insert_memoperands(spill_ilp_t * si)
3740 {
3741         memoperand_t   *memoperand;
3742         lpp_name_t     *name;
3743
3744         set_foreach(si->memoperands, memoperand) {
3745                 name = si->lpp->vars[memoperand->ilp];
3746                 if(!is_zero(name->value)) {
3747                         perform_memory_operand(si, memoperand);
3748                 }
3749         }
3750 }
3751
3752 static void
3753 walker_spill_placer(ir_node * bb, void * data) {
3754         spill_ilp_t   *si = (spill_ilp_t*)data;
3755         ir_node       *irn;
3756         spill_bb_t    *spill_bb = get_irn_link(bb);
3757         pset          *spills_to_do = pset_new_ptr_default();
3758         spill_t       *spill;
3759
3760         set_foreach(spill_bb->ilp, spill) {
3761                 lpp_name_t    *name;
3762
3763                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3764                         name = si->lpp->vars[spill->mem_in];
3765                         if(!is_zero(name->value)) {
3766                                 ir_node   *mem_phi;
3767
3768                                 mem_phi = insert_mem_phi(si, spill->irn);
3769
3770                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3771                         }
3772                 }
3773
3774                 name = si->lpp->vars[spill->spill];
3775                 if(!is_zero(name->value)) {
3776                         /* place spill directly after definition */
3777                         if(get_nodes_block(spill->irn) == bb) {
3778                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3779                                 continue;
3780                         }
3781
3782                         /* place spill at bb start */
3783                         if(spill->reg_in > 0) {
3784                                 name = si->lpp->vars[spill->reg_in];
3785                                 if(!is_zero(name->value)) {
3786                                         insert_spill(si, spill->irn, spill->irn, bb);
3787                                         continue;
3788                                 }
3789                         }
3790                         /* place spill after a remat */
3791                         pset_insert_ptr(spills_to_do, spill->irn);
3792                 }
3793         }
3794         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3795
3796
3797         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3798                 op_t     *op = get_irn_link(irn);
3799
3800                 if(be_is_Spill(irn)) continue;
3801
3802                 if(op->is_remat) {
3803                         /* TODO fix this if we want to support remats with more than two nodes */
3804                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3805                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3806
3807                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3808                         }
3809                 } else {
3810                         if(pset_find_ptr(spills_to_do, irn)) {
3811                                 pset_remove_ptr(spills_to_do, irn);
3812
3813                                 insert_spill(si, irn, irn, irn);
3814                         }
3815                 }
3816
3817         }
3818
3819         assert(pset_count(spills_to_do) == 0);
3820
3821         /* afterwards free data in block */
3822         del_pset(spills_to_do);
3823 }
3824
3825 static ir_node *
3826 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3827 {
3828         ir_node          *insert_pos = bb;
3829         ir_node          *spill;
3830         const arch_env_t *arch_env = si->birg->main_env->arch_env;
3831
3832         /* find last definition of arg value in block */
3833         ir_node  *next;
3834         defs_t   *defs;
3835         int       last = 0;
3836
3837         defs = set_find_def(si->values, value);
3838
3839         if(defs && defs->remats) {
3840                 for(next = defs->remats; next; next = get_irn_link(next)) {
3841                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3842                                 last = sched_get_time_step(next);
3843                                 insert_pos = next;
3844                         }
3845                 }
3846         }
3847
3848         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3849                 last = sched_get_time_step(value);
3850                 insert_pos = value;
3851         }
3852
3853         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3854
3855         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3856
3857         return spill;
3858 }
3859
3860 static void
3861 phim_fixer(spill_ilp_t *si) {
3862         defs_t  *defs;
3863
3864         set_foreach(si->values, defs) {
3865                 const ir_node  *phi = defs->value;
3866                 op_t           *op = get_irn_link(phi);
3867                 ir_node        *phi_m = NULL;
3868                 ir_node        *next = defs->spills;
3869                 int             n;
3870
3871                 if(!is_Phi(phi)) continue;
3872
3873                 while(next) {
3874                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3875                                 phi_m = next;
3876                                 break;
3877                         } else {
3878                                 next = get_irn_link(next);
3879                         }
3880                 }
3881                 if(!phi_m) continue;
3882
3883                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3884                         ir_node        *value = get_irn_n(phi, n);
3885                         defs_t         *val_defs = set_find_def(si->values, value);
3886
3887                         /* a spill of this value */
3888                         ir_node      *spill;
3889
3890
3891                         if(opt_memcopies) {
3892                                 ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3893                                 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3894
3895                                 if(!is_zero(name->value)) {
3896                                         spill = insert_mem_copy(si, pred, value);
3897                                 } else {
3898                                         spill = val_defs->spills;
3899                                 }
3900                         } else {
3901                                 spill = val_defs->spills;
3902                         }
3903
3904                         assert(spill && "no spill placed before PhiM");
3905                         set_irn_n(phi_m, n, spill);
3906                 }
3907         }
3908 }
3909
3910 static void
3911 walker_reload_placer(ir_node * bb, void * data) {
3912         spill_ilp_t   *si = (spill_ilp_t*)data;
3913         ir_node       *irn;
3914         spill_bb_t    *spill_bb = get_irn_link(bb);
3915
3916         /* reloads at end of block */
3917         if(spill_bb->reloads) {
3918                 keyval_t    *keyval;
3919
3920                 set_foreach(spill_bb->reloads, keyval) {
3921                         ir_node        *irn = (ir_node*)keyval->key;
3922                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3923                         lpp_name_t     *name;
3924
3925                         name = si->lpp->vars[reload];
3926                         if(!is_zero(name->value)) {
3927                                 ir_node    *reload;
3928                                 ir_node    *insert_pos = bb;
3929                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3930                                 op_t       *prev_op = get_irn_link(prev);
3931
3932                                 while(be_is_Spill(prev)) {
3933                                         prev = sched_prev(prev);
3934                                 }
3935
3936                                 prev_op = get_irn_link(prev);
3937
3938                                 /* insert reload before pre-remats */
3939                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3940                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3941                                         insert_pos = prev;
3942
3943                                         do {
3944                                                 prev = sched_prev(prev);
3945                                         } while(be_is_Spill(prev));
3946
3947                                         prev_op = get_irn_link(prev);
3948
3949                                 }
3950
3951                                 reload = insert_reload(si, irn, insert_pos);
3952
3953                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
3954                                         pset_insert_ptr(si->spills, reload);
3955                         }
3956                 }
3957         }
3958
3959         /* walk and insert more reloads and collect remats */
3960         sched_foreach_reverse(bb, irn) {
3961                 op_t     *op = get_irn_link(irn);
3962
3963                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3964                 if(is_Phi(irn)) break;
3965
3966                 if(op->is_remat) {
3967                         if(get_irn_mode(irn) != mode_T) {
3968                                 insert_remat(si, irn);
3969                         }
3970                 } else {
3971                         int    n;
3972
3973                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3974                                 ir_node    *arg = get_irn_n(irn, n);
3975
3976                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3977                                         lpp_name_t    *name;
3978
3979                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3980                                         if(!is_zero(name->value)) {
3981                                                 ir_node    *reload;
3982                                                 ir_node    *insert_pos = irn;
3983                                                 ir_node    *prev = sched_prev(insert_pos);
3984                                                 op_t       *prev_op;
3985
3986                                                 while(be_is_Spill(prev)) {
3987                                                         prev = sched_prev(prev);
3988                                                 }
3989
3990                                                 prev_op = get_irn_link(prev);
3991
3992                                                 /* insert reload before pre-remats */
3993                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3994                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3995                                                         insert_pos = prev;
3996
3997                                                         do {
3998                                                                 prev = sched_prev(prev);
3999                                                         } while(be_is_Spill(prev));
4000
4001                                                         prev_op = get_irn_link(prev);
4002
4003                                                 }
4004
4005                                                 reload = insert_reload(si, arg, insert_pos);
4006
4007                                                 assert(reload && "no reload returned");
4008                                                 set_irn_n(irn, n, reload);
4009
4010                                                 if(opt_keep_alive & KEEPALIVE_RELOADS)
4011                                                         pset_insert_ptr(si->spills, reload);
4012                                         }
4013                                 }
4014                         }
4015                 }
4016         }
4017
4018         del_set(spill_bb->ilp);
4019         if(spill_bb->reloads) del_set(spill_bb->reloads);
4020 }
4021
4022 static void
4023 walker_collect_used(ir_node * irn, void * data)
4024 {
4025         bitset_t   *used = data;
4026
4027         bitset_set(used, get_irn_idx(irn));
4028 }
4029
4030 struct kill_helper {
4031         bitset_t  *used;
4032         spill_ilp_t  *si;
4033 };
4034
4035 static void
4036 walker_kill_unused(ir_node * bb, void * data)
4037 {
4038         struct kill_helper *kh = data;
4039         ir_node            *bad = get_irg_bad(get_irn_irg(bb));
4040         ir_node            *irn;
4041
4042
4043         for(irn=sched_first(bb); !sched_is_end(irn);) {
4044                 ir_node     *next = sched_next(irn);
4045                 int          n;
4046
4047                 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4048                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
4049                                 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(kh->si, irn)*execution_frequency(kh->si, bb)));
4050 #if 0
4051                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4052 #endif
4053                         }
4054
4055                         sched_remove(irn);
4056
4057                         set_nodes_block(irn, bad);
4058                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
4059                                 set_irn_n(irn, n, bad);
4060                         }
4061                 }
4062                 irn = next;
4063         }
4064 }
4065
4066 #ifndef SCHEDULE_PHIM
4067 static void
4068 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4069 {
4070         ir_node  *phi;
4071         ir_node  *bad = get_irg_bad(si->birg->irg);
4072         int       n;
4073
4074         pset_foreach(si->phims, phi) {
4075                 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4076
4077                         set_nodes_block(phi, bad);
4078                         for (n=get_irn_arity(phi)-1; n>=0; --n) {
4079                                 set_irn_n(phi, n, bad);
4080                         }
4081                 }
4082         }
4083 }
4084 #endif
4085
4086 static void
4087 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4088 {
4089         struct kill_helper  kh;
4090
4091         kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4092         kh.si = si;
4093
4094         irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4095 #ifndef SCHEDULE_PHIM
4096         kill_unused_phims(si, &kh);
4097 #endif
4098         irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4099
4100         bitset_free(kh.used);
4101 }
4102
4103 void
4104 print_irn_pset(pset * p)
4105 {
4106         ir_node   *irn;
4107
4108         pset_foreach(p, irn) {
4109                 ir_printf("%+F\n", irn);
4110         }
4111 }
4112
4113 void
4114 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4115 {
4116     FILE           *f = fopen(file, "w");
4117     ir_node        *irn;
4118     interference_t *interference;
4119     int            i;
4120
4121     set_break(si->interferences);
4122
4123     ir_fprintf(f, "digraph phiclass {\n");
4124
4125     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4126         irn = phiclass[i];
4127         if (is_Phi(irn))
4128             ir_fprintf(f, "  %F%N [shape=box]\n", irn, irn);
4129     }
4130
4131     for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4132         int n;
4133
4134         irn = phiclass[i];
4135         if (! is_Phi(irn))
4136             continue;
4137
4138         for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4139             ir_node  *arg = get_irn_n(irn, n);
4140
4141             ir_fprintf(f, "  %F%N -> %F%N\n", irn, irn, arg, arg);
4142         }
4143     }
4144
4145     set_foreach(si->interferences, interference) {
4146         const ir_node *a = interference->a;
4147         const ir_node *b = interference->b;
4148         if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4149             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4150         }
4151     }
4152
4153     ir_fprintf(f, "}");
4154     fclose(f);
4155 }
4156
4157 static void
4158 rewire_uses(spill_ilp_t * si)
4159 {
4160         defs_t               *defs;
4161         ir_nodeset_t         ignore;
4162
4163         ir_nodeset_init(&ignore);
4164         ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg));
4165
4166         /* then fix uses of spills */
4167         set_foreach(si->values, defs) {
4168                 pset           *reloads;
4169                 pset           *spills;
4170                 const ir_node  *next = defs->remats;
4171                 int remats = 0;
4172
4173                 reloads = pset_new_ptr_default();
4174
4175                 while(next) {
4176                         if(be_is_Reload(next)) {
4177                                 pset_insert_ptr(reloads, next);
4178                         } else {
4179                                 ++remats;
4180                         }
4181                         next = get_irn_link(next);
4182                 }
4183
4184                 spills = get_spills_for_value(si, defs->value);
4185                 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));
4186                 if(pset_count(spills) > 1) {
4187                         be_ssa_construction_env_t senv;
4188                         ir_node *node;
4189                         //assert(pset_count(reloads) > 0);
4190                         //                              print_irn_pset(spills);
4191                         //                              print_irn_pset(reloads);
4192
4193                         be_ssa_construction_init(&senv, si->birg);
4194                         be_ssa_construction_set_ignore_uses(&senv, &ignore);
4195                         pset_foreach(spills, node) {
4196                                 be_ssa_construction_add_copy(&senv, node);
4197                         }
4198                         pset_foreach(spills, node) {
4199                                 be_ssa_construction_fix_users(&senv, node);
4200                         }
4201                         be_ssa_construction_update_liveness_phis(&senv, si->lv);
4202                         pset_foreach(spills, node) {
4203                                 be_liveness_update(si->lv, node);
4204                         }
4205                         be_ssa_construction_destroy(&senv);
4206                 }
4207
4208                 del_pset(reloads);
4209                 del_pset(spills);
4210         }
4211
4212         /* first fix uses of remats and reloads */
4213         set_foreach(si->values, defs) {
4214                 const ir_node  *next = defs->remats;
4215                 int             orig_kept = 0;
4216
4217                 if(next) {
4218                         be_ssa_construction_env_t senv;
4219
4220                         be_ssa_construction_init(&senv, si->birg);
4221
4222                         if(sched_is_scheduled(defs->value)) {
4223                                 be_ssa_construction_add_copy(&senv, (ir_node*) defs->value);
4224                                 orig_kept = 1;
4225                         }
4226
4227                         next = defs->remats;
4228                         while(next) {
4229                                 be_ssa_construction_add_copy(&senv, (ir_node*) next);
4230                                 next = get_irn_link(next);
4231                         }
4232
4233                         if(sched_is_scheduled(defs->value)) {
4234                                 be_ssa_construction_fix_users(&senv, (ir_node*) defs->value);
4235                         }
4236
4237                         next = defs->remats;
4238                         while(next) {
4239                                 be_ssa_construction_fix_users(&senv, (ir_node*) next);
4240                                 next = get_irn_link(next);
4241                         }
4242
4243                         be_ssa_construction_update_liveness_phis(&senv, si->lv);
4244                         if(sched_is_scheduled(defs->value)) {
4245                                 be_liveness_update(si->lv, (ir_node*) defs->value);
4246                         }
4247
4248                         next = defs->remats;
4249                         while(next) {
4250                                 be_liveness_update(si->lv, (ir_node*) next);
4251                                 next = get_irn_link(next);
4252                         }
4253
4254                         be_ssa_construction_destroy(&senv);
4255                 }
4256         }
4257
4258         ir_nodeset_destroy(&ignore);
4259 //      remove_unused_defs(si);
4260 }
4261
4262
4263 static void
4264 writeback_results(spill_ilp_t * si)
4265 {
4266         /* walk through the graph and collect all spills, reloads and remats for a value */
4267
4268         si->values = new_set(cmp_defs, 4096);
4269
4270         DBG((si->dbg, LEVEL_1, "Applying results\n"));
4271         delete_unnecessary_remats(si);
4272         si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4273         irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4274         irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4275         if(opt_memoperands)
4276                 insert_memoperands(si);
4277         phim_fixer(si);
4278
4279         /* clean the remat info! there are still back-edges leading there! */
4280         clean_remat_info(si);
4281
4282         rewire_uses(si);
4283
4284         connect_all_spills_with_keep(si);
4285
4286         del_set(si->values);
4287 }
4288
4289 static int
4290 get_n_regs(spill_ilp_t * si)
4291 {
4292         int       arch_n_regs = arch_register_class_n_regs(si->cls);
4293
4294         bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4295         bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4296
4297         arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4298     be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4299
4300         bitset_andnot(arch_regs, abi_regs);
4301         arch_n_regs = bitset_popcnt(arch_regs);
4302
4303         bitset_free(arch_regs);
4304         bitset_free(abi_regs);
4305
4306         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4307         return arch_n_regs;
4308 }
4309
4310 static void
4311 walker_reload_mover(ir_node * bb, void * data)
4312 {
4313         spill_ilp_t   *si = data;
4314         ir_node           *tmp;
4315
4316         sched_foreach(bb, tmp) {
4317                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4318                         ir_node       *reload = tmp;
4319                         ir_node       *irn = tmp;
4320
4321                         /* move reload upwards */
4322
4323                         int pressure = (int)get_irn_link(reload);
4324                         if(pressure < si->n_regs) {
4325                                 irn = sched_prev(reload);
4326                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4327                                 sched_remove(reload);
4328                                 pressure = (int)get_irn_link(irn);
4329
4330                                 while(pressure < si->n_regs) {
4331                                         if( sched_is_end(irn) ||
4332                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4333                                            /* do not move reload before its spill */
4334                                            (irn == be_get_Reload_mem(reload)) ||
4335                                            /* do not move before phi */
4336                                            is_Phi(irn)) break;
4337
4338                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
4339                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4340                                         irn = sched_prev(irn);
4341
4342                                         pressure = (int)get_irn_link(irn);
4343                                 }
4344
4345                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4346                                 sched_put_after(irn, reload);
4347                         }
4348                 }
4349         }
4350 }
4351
4352 static void
4353 move_reloads_upward(spill_ilp_t * si)
4354 {
4355         irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4356 }
4357
4358
4359 /**
4360  * Walk all irg blocks and check for interfering spills inside of phi classes
4361  */
4362 static void
4363 luke_meminterferencechecker(ir_node * bb, void * data)
4364 {
4365         spill_ilp_t *si = (spill_ilp_t*)data;
4366         int         l1, l2;
4367
4368         be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4369                 ir_node        *a = be_lv_get_irn(si->lv, bb, l1);
4370
4371                 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4372                         continue;
4373
4374                 /* a is only interesting if it is in my register class and if it is inside a phi class */
4375                 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4376                         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)) {
4377                                 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4378
4379                                 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4380                                         continue;
4381
4382                                 /* a and b are only interesting if they are in the same phi class */
4383                                 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4384                                         if (values_interfere_in_block(si, bb, a, b)) {
4385                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4386                                         }
4387                                 }
4388                         }
4389                 }
4390         }
4391 }
4392
4393 static void
4394 verify_phiclasses(spill_ilp_t * si)
4395 {
4396         /* analyze phi classes */
4397         phi_class_free(si->pc);
4398         si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4399
4400         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4401         irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4402 }
4403
4404 void
4405 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4406 {
4407         char            buf[256];
4408         char            problem_name[256];
4409         char            dump_suffix[256];
4410         char            dump_suffix2[256];
4411         struct obstack  obst;
4412         spill_ilp_t     si;
4413         ir_graph       *irg = be_get_birg_irg(birg);
4414
4415         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4416         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4417         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4418
4419         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4420         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4421
4422         if(opt_verify & VERIFY_DOMINANCE)
4423                 be_check_dominance(irg);
4424
4425         be_assure_dom_front(birg);
4426         be_assure_liveness(birg);
4427
4428         obstack_init(&obst);
4429         si.obst                = &obst;
4430         si.birg                = birg;
4431         si.cls                 = cls;
4432         si.lpp                 = new_lpp(problem_name, lpp_minimize);
4433         si.remat_info          = new_set(cmp_remat_info, 4096);
4434         si.interferences       = new_set(cmp_interference, 32);
4435         si.memoperands         = new_set(cmp_memoperands, 128);
4436         si.all_possible_remats = pset_new_ptr_default();
4437         si.spills              = pset_new_ptr_default();
4438         si.inverse_ops         = pset_new_ptr_default();
4439         si.lv                  = birg->lv;
4440         si.keep                = NULL;
4441         si.n_regs              = get_n_regs(&si);
4442
4443         set_irg_link(irg, &si);
4444         compute_doms(irg);
4445
4446         /* compute phi classes */
4447         // phi_class_compute(irg);
4448
4449         if(opt_dump_flags & DUMP_STATS)
4450                 be_analyze_regpressure(birg, cls, "-pre");
4451
4452         DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4453         irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4454
4455         if(opt_remats) {
4456                 /* collect remats */
4457                 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4458                 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4459         }
4460
4461         /* insert possible remats */
4462         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4463         irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4464         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4465
4466         if(opt_keep_alive & KEEPALIVE_REMATS) {
4467                 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4468                 connect_all_remats_with_keep(&si);
4469                 /* dump graph with inserted remats */
4470                 dump_graph_with_remats(irg, dump_suffix);
4471         }
4472
4473         /* insert copies for phi arguments not in my regclass */
4474         irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4475
4476         /* recompute liveness */
4477         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4478         be_liveness_recompute(si.lv);
4479
4480         /* build the ILP */
4481         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4482         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4483         irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4484
4485         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4486         irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4487
4488         si.pc = phi_class_new_from_irg(birg->irg, 0);
4489         if (opt_memcopies) {
4490                 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4491                 memcopyhandler(&si);
4492         }
4493
4494         if (opt_dump_flags & DUMP_PROBLEM) {
4495                 FILE           *f;
4496                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4497                 if ((f = fopen(buf, "wt")) != NULL) {
4498                         lpp_dump_plain(si.lpp, f);
4499                         fclose(f);
4500                 }
4501         }
4502
4503         if (opt_dump_flags & DUMP_MPS) {
4504                 FILE *f;
4505
4506                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4507                 if ((f = fopen(buf, "wt")) != NULL) {
4508                         mps_write_mps(si.lpp, s_mps_fixed, f);
4509                         fclose(f);
4510                 }
4511
4512                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4513                 if ((f = fopen(buf, "wt")) != NULL) {
4514                         mps_write_mst(si.lpp, s_mps_fixed, f);
4515                         fclose(f);
4516                 }
4517         }
4518
4519         lpp_check_startvals(si.lpp);
4520
4521 #ifdef SOLVE
4522         DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4523         lpp_set_time_limit(si.lpp, opt_timeout);
4524
4525         if(opt_log)
4526                 lpp_set_log(si.lpp, stdout);
4527
4528 #ifdef SOLVE_LOCAL
4529         lpp_solve_cplex(si.lpp);
4530 #else
4531         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4532 #endif
4533         assert(lpp_is_sol_valid(si.lpp)
4534                && "solution of ILP must be valid");
4535
4536         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));
4537
4538         if(opt_dump_flags & DUMP_SOLUTION) {
4539                 FILE           *f;
4540                 char            buf[256];
4541
4542                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4543                 if ((f = fopen(buf, "wt")) != NULL) {
4544                         int             i;
4545                         for (i = 0; i < si.lpp->var_next; ++i) {
4546                                 lpp_name_t     *name = si.lpp->vars[i];
4547                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4548                         }
4549                         fclose(f);
4550                 }
4551         }
4552
4553 #ifndef SCHEDULE_PHIM
4554         si.phims = pset_new_ptr_default();
4555 #endif
4556         writeback_results(&si);
4557
4558
4559 #endif                          /* SOLVE */
4560
4561         kill_all_unused_values_in_schedule(&si);
4562
4563 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4564         del_pset(si.phims);
4565 #endif
4566
4567         if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4568                 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4569
4570         // move reloads upwards
4571         be_liveness_recompute(si.lv);
4572         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4573         move_reloads_upward(&si);
4574
4575         if(opt_memcopies) {
4576                 verify_phiclasses(&si);
4577         }
4578
4579         irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4580
4581         if(opt_dump_flags & DUMP_PRESSURE)
4582                 dump_pressure_graph(&si, dump_suffix2);
4583
4584         if(opt_dump_flags & DUMP_STATS)
4585                 be_analyze_regpressure(birg, cls, "-post");
4586
4587         if(opt_verify & VERIFY_DOMINANCE)
4588                 be_check_dominance(irg);
4589
4590         free_dom(irg);
4591         del_set(si.interferences);
4592         del_pset(si.inverse_ops);
4593         del_pset(si.all_possible_remats);
4594         del_set(si.memoperands);
4595         del_pset(si.spills);
4596         free_lpp(si.lpp);
4597         phi_class_free(si.pc);
4598         obstack_free(&obst, NULL);
4599         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4600 }
4601
4602 void be_init_spillremat(void)
4603 {
4604         static be_spiller_t remat_spiller = {
4605                 be_spill_remat
4606         };
4607         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4608         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4609         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4610         lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4611
4612         be_register_spiller("remat", &remat_spiller);
4613         lc_opt_add_table(remat_grp, options);
4614 }
4615
4616 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4617
4618 #else                           /* WITH_ILP */
4619
4620 static __attribute__((unused))
4621 void only_that_you_can_compile_without_WITH_ILP_defined(void)
4622 {
4623 }
4624
4625 #endif                          /* WITH_ILP */