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

   1 /** vim: set sw=4 ts=4:
   2  * @file   bespillremat.c
   3  * @date   2006-04-06
   4  * @author Adam M. Szalkowski & Sebastian Hack
   5  *
   6  * ILP based spilling & rematerialization
   7  *
   8  * Copyright (C) 2006 Universitaet Karlsruhe
   9  * Released under the GPL
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 #include "config.h"
  13 #endif
  14
  15 #ifdef WITH_ILP
  16
  17 #include <math.h>
  18
  19 #include "hashptr.h"
  20 #include "debug.h"
  21 #include "obst.h"
  22 #include "set.h"
  23 #include "list.h"
  24 #include "pmap.h"
  25
  26 #include "irprintf.h"
  27 #include "irgwalk.h"
  28 #include "irdump_t.h"
  29 #include "irnode_t.h"
  30 #include "ircons_t.h"
  31 #include "irloop_t.h"
  32 #include "phiclass_t.h"
  33 #include "iredges.h"
  34 #include "execfreq.h"
  35 #include "irvrfy.h"
  36
  37 #include <lpp/lpp.h>
  38 #include <lpp/lpp_net.h>
  39 #include <lpp/lpp_cplex.h>
  40 //#include <lc_pset.h>
  41 #include <libcore/lc_bitset.h>
  42
  43 #include "be_t.h"
  44 #include "belive_t.h"
  45 #include "besched_t.h"
  46 #include "beirgmod.h"
  47 #include "bearch.h"
  48 #include "benode_t.h"
  49 #include "beutil.h"
  50 #include "bespillremat.h"
  51 #include "bespill.h"
  52 #include "bepressurestat.h"
  53
  54 #include "bechordal_t.h"
  55
  56 #define BIGM 100000.0
  57
  58 #define DUMP_SOLUTION
  59 #define DUMP_ILP
  60 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
  61 #define COLLECT_REMATS /* enable rematerialization */
  62 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
  63 #define REMAT_WHILE_LIVE /* only remat values that are live */
  64 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
  65 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
  66 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
  67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
  68 //#define KEEPALIVE_SPILLS
  69 //#define KEEPALIVE_RELOADS
  70 #define GOODWIN_REDUCTION
  71 //#define NO_MEMCOPIES
  72 //#define VERIFY_DOMINANCE
  73
  74 #define  SOLVE
  75 //#define  SOLVE_LOCAL
  76 #define LPP_SERVER "i44pc52"
  77 #define LPP_SOLVER "cplex"
  78
  79 #define COST_LOAD      8
  80 #define COST_STORE     50
  81 #define COST_REMAT     1
  82
  83 #define ILP_TIMEOUT    300
  84 #define MAX_PATHS      16
  85 #define ILP_UNDEF               -1
  86
  87 typedef struct _spill_ilp_t {
  88         const arch_register_class_t  *cls;
  89         int                           n_regs;
  90         const be_chordal_env_t       *chordal_env;
  91         lpp_t                        *lpp;
  92         struct obstack               *obst;
  93         set                          *remat_info;
  94         pset                         *all_possible_remats;
  95         pset                         *inverse_ops;
  96 #ifdef KEEPALIVE
  97         ir_node                      *keep;
  98 #endif
  99         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
 100         pset                         *spills;
 101         set                          *interferences;
 102         ir_node                      *m_unknown;
 103         DEBUG_ONLY(firm_dbg_module_t * dbg);
 104 } spill_ilp_t;
 105
 106 typedef int ilp_var_t;
 107 typedef int ilp_cst_t;
 108
 109 typedef struct _spill_bb_t {
 110         set      *ilp;
 111         set      *reloads;
 112 } spill_bb_t;
 113
 114 typedef struct _remat_t {
 115         const ir_node        *op;      /**< for copy_irn */
 116         const ir_node        *value;   /**< the value which is being recomputed by this remat */
 117         ir_node              *proj;    /**< not NULL if the above op produces a tuple */
 118         int                   cost;    /**< cost of this remat */
 119         int                   inverse; /**< nonzero if this is an inverse remat */
 120 } remat_t;
 121
 122 /**
 123  * Data to be attached to each IR node. For remats this contains the ilp_var
 124  * for this remat and for normal ops this contains the ilp_vars for
 125  * reloading each operand
 126  */
 127 typedef struct _op_t {
 128         int             is_remat;
 129         union {
 130                 struct {
 131                         ilp_var_t       ilp;
 132                         remat_t        *remat; /** the remat this op belongs to */
 133                         int             pre; /** 1, if this is a pressure-increasing remat */
 134                 } remat;
 135                 struct {
 136                         ilp_var_t       ilp;
 137                         ir_node        *op; /** the operation this live range belongs to */
 138                         union {
 139                                 ilp_var_t      *reloads;
 140                                 ilp_var_t      *copies;
 141                         } args;
 142                 } live_range;
 143         } attr;
 144 } op_t;
 145
 146 typedef struct _defs_t {
 147         ir_node   *value;
 148         ir_node   *spills;  /**< points to the first spill for this value (linked by link field) */
 149         ir_node   *remats;  /**< points to the first definition for this value (linked by link field) */
 150 } defs_t;
 151
 152 typedef struct _remat_info_t {
 153         const ir_node       *irn; /**< the irn to which these remats belong */
 154         pset                *remats; /**< possible remats for this value */
 155         pset                *remats_by_operand; /**< remats with this value as operand */
 156 } remat_info_t;
 157
 158 typedef struct _keyval_t {
 159         const void          *key;
 160         const void          *val;
 161 } keyval_t;
 162
 163 typedef struct _spill_t {
 164         ir_node      *irn;
 165         ilp_var_t     reg_in;
 166         ilp_var_t     mem_in;
 167         ilp_var_t     reg_out;
 168         ilp_var_t     mem_out;
 169         ilp_var_t     spill;
 170 } spill_t;
 171
 172 static INLINE int
 173 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 174 {
 175         return chordal_has_class(si->chordal_env, irn);
 176 }
 177
 178 #if 0
 179 static int
 180 cmp_remat(const void *a, const void *b)
 181 {
 182         const keyval_t *p = a;
 183         const keyval_t *q = b;
 184         const remat_t  *r = p->val;
 185         const remat_t  *s = q->val;
 186
 187         assert(r && s);
 188
 189         return !(r == s || r->op == s->op);
 190 }
 191 #endif
 192 static int
 193 cmp_remat(const void *a, const void *b)
 194 {
 195         const remat_t  *r = a;
 196         const remat_t  *s = a;
 197
 198         return !(r == s || r->op == s->op);
 199 }
 200
 201 static int
 202 cmp_spill(const void *a, const void *b, size_t size)
 203 {
 204         const spill_t *p = a;
 205         const spill_t *q = b;
 206
 207 //      return !(p->irn == q->irn && p->bb == q->bb);
 208         return !(p->irn == q->irn);
 209 }
 210
 211 static keyval_t *
 212 set_find_keyval(set * set, void * key)
 213 {
 214         keyval_t     query;
 215
 216         query.key = key;
 217         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 218 }
 219
 220 static keyval_t *
 221 set_insert_keyval(set * set, void * key, void * val)
 222 {
 223         keyval_t     query;
 224
 225         query.key = key;
 226         query.val = val;
 227         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 228 }
 229
 230 static defs_t *
 231 set_find_def(set * set, ir_node * value)
 232 {
 233         defs_t     query;
 234
 235         query.value = value;
 236         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 237 }
 238
 239 static defs_t *
 240 set_insert_def(set * set, ir_node * value)
 241 {
 242         defs_t     query;
 243
 244         query.value = value;
 245         query.spills = NULL;
 246         query.remats = NULL;
 247         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 248 }
 249
 250 static spill_t *
 251 set_find_spill(set * set, ir_node * value)
 252 {
 253         spill_t     query;
 254
 255         query.irn = value;
 256         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 257 }
 258
 259 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 260 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 261 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 262 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 263 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 264
 265 static int
 266 cmp_remat_info(const void *a, const void *b, size_t size)
 267 {
 268         const remat_info_t *p = a;
 269         const remat_info_t *q = b;
 270
 271         return !(p->irn == q->irn);
 272 }
 273
 274 static int
 275 cmp_defs(const void *a, const void *b, size_t size)
 276 {
 277         const defs_t *p = a;
 278         const defs_t *q = b;
 279
 280         return !(p->value == q->value);
 281 }
 282
 283 static int
 284 cmp_keyval(const void *a, const void *b, size_t size)
 285 {
 286         const keyval_t *p = a;
 287         const keyval_t *q = b;
 288
 289         return !(p->key == q->key);
 290 }
 291
 292 static double
 293 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
 294 {
 295 #define FUDGE 0.001
 296 #ifndef EXECFREQ_LOOPDEPH
 297         return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
 298 #else
 299         if(is_Block(irn))
 300                 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
 301         else
 302                 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
 303 #endif
 304 }
 305
 306 static double
 307 get_cost(const spill_ilp_t * si, const ir_node * irn)
 308 {
 309         if(be_is_Spill(irn)) {
 310                 return COST_STORE;
 311         } else if(be_is_Reload(irn)){
 312                 return COST_LOAD;
 313         } else {
 314                 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
 315         }
 316 }
 317
 318 /**
 319  * Checks, whether node and its operands have suitable reg classes
 320  */
 321 static INLINE int
 322 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 323 {
 324         int               n;
 325         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 326         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 327
 328 #if 0
 329         if(!remat)
 330                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 331 #endif
 332
 333         for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
 334                 ir_node        *op = get_irn_n(irn, n);
 335                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 336
 337 //              if(!remat)
 338 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 339         }
 340
 341         return remat;
 342 }
 343
 344 /**
 345  * Try to create a remat from @p op with destination value @p dest_value
 346  */
 347 static INLINE remat_t *
 348 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 349 {
 350         remat_t  *remat = NULL;
 351
 352 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 353 //              return NULL;
 354
 355         if(dest_value == op) {
 356                 const ir_node *proj = NULL;
 357
 358                 if(is_Proj(dest_value)) {
 359                         op = get_irn_n(op, 0);
 360                         proj = dest_value;
 361                 }
 362
 363                 if(!is_rematerializable(si, op))
 364                         return NULL;
 365
 366                 remat = obstack_alloc(si->obst, sizeof(*remat));
 367                 remat->op = op;
 368                 remat->cost = get_cost(si, op);
 369                 remat->value = dest_value;
 370                 remat->proj = proj;
 371                 remat->inverse = 0;
 372         } else {
 373                 arch_inverse_t     inverse;
 374                 int                n;
 375
 376                 /* get the index of the operand we want to retrieve by the inverse op */
 377                 for (n = get_irn_arity(op)-1; n>=0; --n) {
 378                         ir_node        *arg = get_irn_n(op, n);
 379
 380                         if(arg == dest_value) break;
 381                 }
 382                 if(n<0) return NULL;
 383
 384                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", n, op));
 385
 386                 /* else ask the backend to give an inverse op */
 387                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
 388                         int   i;
 389
 390                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 391
 392                         assert(inverse.n > 0 && "inverse op should have at least one node");
 393
 394                         for(i=inverse.n-1; i>=0; --i) {
 395                                 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
 396                         }
 397
 398                         if(inverse.n <= 2) {
 399                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 400                                 remat->op = inverse.nodes[0];
 401                                 remat->cost = inverse.costs;
 402                                 remat->value = dest_value;
 403                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 404                                 remat->inverse = 1;
 405
 406                                 assert(is_Proj(remat->proj));
 407                         } else {
 408                                 assert(0 && "I can not handle remats with more than 2 nodes");
 409                         }
 410                 }
 411         }
 412
 413         if(remat) {
 414                 if(remat->proj) {
 415                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 416                 } else {
 417                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 418                 }
 419         }
 420         return remat;
 421 }
 422
 423
 424 static INLINE void
 425 add_remat(const spill_ilp_t * si, const remat_t * remat)
 426 {
 427         remat_info_t    *remat_info,
 428                      query;
 429         int              n;
 430
 431         assert(remat->op);
 432         assert(remat->value);
 433
 434         query.irn = remat->value;
 435         query.remats = NULL;
 436         query.remats_by_operand = NULL;
 437         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 438
 439         if(remat_info->remats == NULL) {
 440                 remat_info->remats = new_pset(cmp_remat, 4096);
 441         }
 442         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 443
 444         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 445         for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
 446                 ir_node        *arg = get_irn_n(remat->op, n);
 447
 448                 query.irn = arg;
 449                 query.remats = NULL;
 450                 query.remats_by_operand = NULL;
 451                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 452
 453                 if(remat_info->remats_by_operand == NULL) {
 454                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 455                 }
 456                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 457         }
 458 }
 459
 460 static int
 461 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
 462 {
 463         const ir_edge_t   *edge = get_irn_out_edge_first(irn);
 464         int                i = 0;
 465
 466         while(edge) {
 467                 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
 468                         ++i;
 469                 }
 470                 edge = get_irn_out_edge_next(irn, edge);
 471         }
 472
 473         return i;
 474 }
 475
 476 static INLINE void
 477 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 478 {
 479         int      n;
 480         remat_t *remat;
 481
 482 #ifdef NO_SINGLE_USE_REMATS
 483         if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
 484 #else
 485         if(has_reg_class(si, op)) {
 486 #endif
 487                 remat = get_remat_from_op(si, op, op);
 488                 if(remat) {
 489                         add_remat(si, remat);
 490                 }
 491         }
 492
 493 #ifdef COLLECT_INVERSE_REMATS
 494         /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 495            for each arg */
 496         for (n = get_irn_arity(op)-1; n>=0; --n) {
 497                 ir_node        *arg = get_irn_n(op, n);
 498
 499                 if(has_reg_class(si, arg)) {
 500                         /* try to get an inverse remat */
 501                         remat = get_remat_from_op(si, arg, op);
 502                         if(remat) {
 503                                 add_remat(si, remat);
 504                         }
 505                 }
 506         }
 507 #endif
 508
 509 }
 510
 511 static INLINE int
 512 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 513 {
 514         ir_node *block;
 515         ir_node *def_block = get_nodes_block(val);
 516         int      ret;
 517
 518         if(val == pos)
 519                 return 0;
 520
 521         /* if pos is at end of a basic block */
 522         if(is_Block(pos)) {
 523                 ret = (pos == def_block || block_dominates(def_block, pos));
 524 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 525                 return ret;
 526         }
 527
 528         /* else if this is a normal operation */
 529         block = get_nodes_block(pos);
 530         if(block == def_block) {
 531                 if(!sched_is_scheduled(val)) return 1;
 532
 533                 ret = sched_comes_after(val, pos);
 534 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 535                 return ret;
 536         }
 537
 538         ret = block_dominates(def_block, block);
 539 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 540         return ret;
 541 }
 542
 543 static INLINE ir_node *
 544 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 545 {
 546     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 547 }
 548
 549 /**
 550  * Returns first non-Phi node of block @p bb
 551  */
 552 static INLINE ir_node *
 553 sched_block_first_nonphi(const ir_node * bb)
 554 {
 555         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 556 }
 557
 558 static int
 559 sched_skip_proj_predicator(const ir_node * irn, void * data)
 560 {
 561         return (is_Proj(irn));
 562 }
 563
 564 static INLINE ir_node *
 565 sched_next_nonproj(const ir_node * irn, int forward)
 566 {
 567         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 568 }
 569
 570 /**
 571  * Returns next operation node (non-Proj) after @p irn
 572  * or the basic block of this node
 573  */
 574 static INLINE ir_node *
 575 sched_next_op(const ir_node * irn)
 576 {
 577         ir_node *next = sched_next(irn);
 578
 579         if(is_Block(next))
 580                 return next;
 581
 582         return sched_next_nonproj(next, 1);
 583 }
 584
 585 /**
 586  * Returns previous operation node (non-Proj) before @p irn
 587  * or the basic block of this node
 588  */
 589 static INLINE ir_node *
 590 sched_prev_op(const ir_node * irn)
 591 {
 592         ir_node *prev = sched_prev(irn);
 593
 594         if(is_Block(prev))
 595                 return prev;
 596
 597         return sched_next_nonproj(prev, 0);
 598 }
 599
 600 static void
 601 sched_put_after(ir_node * insert, ir_node * irn)
 602 {
 603         if(is_Block(insert)) {
 604                 insert = sched_block_first_nonphi(insert);
 605         } else {
 606                 insert = sched_next_op(insert);
 607         }
 608         sched_add_before(insert, irn);
 609 }
 610
 611 static void
 612 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 613 {
 614   if(is_Block(insert)) {
 615           insert = sched_block_last_noncf(si, insert);
 616   } else {
 617           insert = sched_next_nonproj(insert, 0);
 618           insert = sched_prev(insert);
 619   }
 620   sched_add_after(insert, irn);
 621 }
 622
 623 /**
 624  * Tells you whether a @p remat can be placed before the irn @p pos
 625  */
 626 static INLINE int
 627 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 628 {
 629         const ir_node   *op = remat->op;
 630         const ir_node   *prev;
 631         int        n,
 632                            res = 1;
 633
 634         if(is_Block(pos)) {
 635                 prev = sched_block_last_noncf(si, pos);
 636                 prev = sched_next_nonproj(prev, 0);
 637         } else {
 638                 prev = sched_prev_op(pos);
 639         }
 640         /* do not remat if the rematted value is defined immediately before this op */
 641         if(prev == remat->op) {
 642                 return 0;
 643         }
 644
 645 #if 0
 646         /* this should be just fine, the following OP will be using this value, right? */
 647
 648         /* only remat AFTER the real definition of a value (?) */
 649         if(!value_is_defined_before(si, pos, remat->value)) {
 650 //              ir_fprintf(stderr, "error(not defined)");
 651                 return 0;
 652         }
 653 #endif
 654
 655         for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
 656                 const ir_node   *arg = get_irn_n(op, n);
 657
 658 #ifdef NO_ENLARGE_L1V3N355
 659                 if(has_reg_class(si, arg) && live) {
 660                         res &= pset_find_ptr(live, arg)?1:0;
 661                 } else {
 662                         res &= value_is_defined_before(si, pos, arg);
 663                 }
 664 #else
 665                 res &= value_is_defined_before(si, pos, arg);
 666 #endif
 667         }
 668
 669         return res;
 670 }
 671
 672 /**
 673  * Tells you whether a @p remat can be placed after the irn @p pos
 674  */
 675 static INLINE int
 676 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 677 {
 678         if(is_Block(pos)) {
 679                 pos = sched_block_first_nonphi(pos);
 680         } else {
 681                 pos = sched_next_op(pos);
 682         }
 683
 684         /* only remat AFTER the real definition of a value (?) */
 685         if(!value_is_defined_before(si, pos, remat->value)) {
 686                 return 0;
 687         }
 688
 689         return can_remat_before(si, remat, pos, live);
 690 }
 691
 692 /**
 693  * Collect potetially rematerializable OPs
 694  */
 695 static void
 696 walker_remat_collector(ir_node * irn, void * data)
 697 {
 698         spill_ilp_t    *si = data;
 699
 700         if(!is_Block(irn) && !is_Phi(irn)) {
 701                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 702                 get_remats_from_op(si, irn);
 703         }
 704 }
 705
 706 /**
 707  * Inserts a copy of @p irn before @p pos
 708  */
 709 static ir_node *
 710 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 711 {
 712         ir_node     *bb;
 713         ir_node     *copy;
 714
 715         bb = is_Block(pos)?pos:get_nodes_block(pos);
 716         copy = exact_copy(irn);
 717
 718         _set_phi_class(copy, NULL);
 719         set_nodes_block(copy, bb);
 720         sched_put_before(si, pos, copy);
 721
 722         return copy;
 723 }
 724
 725 /**
 726  * Inserts a copy of @p irn after @p pos
 727  */
 728 static ir_node *
 729 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 730 {
 731         ir_node     *bb;
 732         ir_node     *copy;
 733
 734         bb = is_Block(pos)?pos:get_nodes_block(pos);
 735         copy = exact_copy(irn);
 736
 737         _set_phi_class(copy, NULL);
 738         set_nodes_block(copy, bb);
 739         sched_put_after(pos, copy);
 740
 741         return copy;
 742 }
 743
 744 static ir_node *
 745 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 746 {
 747         char     buf[256];
 748
 749         if(can_remat_after(si, remat, pos, live)) {
 750                 ir_node         *copy,
 751                                                 *proj_copy;
 752                 op_t            *op;
 753
 754                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 755
 756                 copy = insert_copy_after(si, remat->op, pos);
 757
 758                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 759                 op = obstack_alloc(si->obst, sizeof(*op));
 760                 op->is_remat = 1;
 761                 op->attr.remat.remat = remat;
 762                 op->attr.remat.pre = 0;
 763                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 764
 765                 set_irn_link(copy, op);
 766                 pset_insert_ptr(si->all_possible_remats, copy);
 767                 if(remat->proj) {
 768                         proj_copy = insert_copy_after(si, remat->proj, copy);
 769                         set_irn_n(proj_copy, 0, copy);
 770                         set_irn_link(proj_copy, op);
 771                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 772                 } else {
 773                         proj_copy = NULL;
 774                 }
 775
 776                 return copy;
 777         }
 778
 779         return NULL;
 780 }
 781
 782 static ir_node *
 783 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 784 {
 785         char     buf[256];
 786
 787         if(can_remat_before(si, remat, pos, live)) {
 788                 ir_node         *copy,
 789                                                 *proj_copy;
 790                 op_t            *op;
 791
 792                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 793
 794                 copy = insert_copy_before(si, remat->op, pos);
 795
 796                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 797                 op = obstack_alloc(si->obst, sizeof(*op));
 798                 op->is_remat = 1;
 799                 op->attr.remat.remat = remat;
 800                 op->attr.remat.pre = 1;
 801                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 802
 803                 set_irn_link(copy, op);
 804                 pset_insert_ptr(si->all_possible_remats, copy);
 805                 if(remat->proj) {
 806                         proj_copy = insert_copy_after(si, remat->proj, copy);
 807                         set_irn_n(proj_copy, 0, copy);
 808                         set_irn_link(proj_copy, op);
 809                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 810                 } else {
 811                         proj_copy = NULL;
 812                 }
 813
 814                 return copy;
 815         }
 816
 817         return NULL;
 818 }
 819
 820 static int
 821 get_block_n_succs(const ir_node *block) {
 822         const ir_edge_t *edge;
 823
 824         assert(edges_activated(current_ir_graph));
 825
 826         edge = get_block_succ_first(block);
 827         if (! edge)
 828                 return 0;
 829
 830         edge = get_block_succ_next(block, edge);
 831         return edge ? 2 : 1;
 832 }
 833
 834 static int
 835 is_merge_edge(const ir_node * bb)
 836 {
 837 #ifdef GOODWIN_REDUCTION
 838         return get_block_n_succs(bb) == 1;
 839 #else
 840         return 1;
 841 #endif
 842 }
 843
 844 static int
 845 is_diverge_edge(const ir_node * bb)
 846 {
 847 #ifdef GOODWIN_REDUCTION
 848         return get_Block_n_cfgpreds(bb) == 1;
 849 #else
 850         return 1;
 851 #endif
 852 }
 853
 854 static void
 855 walker_regclass_copy_insertor(ir_node * irn, void * data)
 856 {
 857         spill_ilp_t    *si = data;
 858
 859         if(is_Phi(irn) && has_reg_class(si, irn)) {
 860                 int n;
 861
 862                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
 863                         ir_node  *phi_arg = get_irn_n(irn, n);
 864                         ir_node  *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
 865
 866                         if(!has_reg_class(si, phi_arg)) {
 867                                 ir_node   *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
 868                                 ir_node   *pos = sched_block_last_noncf(si, bb);
 869                                 op_t      *op = obstack_alloc(si->obst, sizeof(*op));
 870
 871                                 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
 872                                 sched_add_after(pos, copy);
 873                                 set_irn_n(irn, n, copy);
 874
 875                                 op->is_remat = 0;
 876                                 op->attr.live_range.args.reloads = NULL;
 877                                 op->attr.live_range.ilp = ILP_UNDEF;
 878                                 set_irn_link(copy, op);
 879                         }
 880                 }
 881         }
 882 }
 883
 884
 885 /**
 886  * Insert (so far unused) remats into the irg to
 887  * recompute the potential liveness of all values
 888  */
 889 static void
 890 walker_remat_insertor(ir_node * bb, void * data)
 891 {
 892         spill_ilp_t    *si = data;
 893         spill_bb_t     *spill_bb;
 894         ir_node        *irn;
 895         int             n;
 896         irn_live_t     *li;
 897         pset           *live = pset_new_ptr_default();
 898
 899         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
 900
 901         live_foreach(bb, li) {
 902                 ir_node        *value = (ir_node *) li->irn;
 903
 904                 /* add remats at end of block */
 905                 if (live_is_end(li) && has_reg_class(si, value)) {
 906                         pset_insert_ptr(live, value);
 907                 }
 908         }
 909
 910         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
 911         set_irn_link(bb, spill_bb);
 912
 913         irn = sched_last(bb);
 914         while(!sched_is_end(irn)) {
 915                 ir_node   *next;
 916                 op_t      *op;
 917                 pset      *args;
 918                 ir_node   *arg;
 919                 pset      *remat_args;
 920
 921                 next = sched_prev(irn);
 922
 923                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
 924
 925                 if(is_Phi(irn) || is_Proj(irn)) {
 926                         op_t      *op;
 927
 928                         if(has_reg_class(si, irn)) {
 929                                 pset_remove_ptr(live, irn);
 930                         }
 931
 932                         op = obstack_alloc(si->obst, sizeof(*op));
 933                         op->is_remat = 0;
 934                         op->attr.live_range.args.reloads = NULL;
 935                         op->attr.live_range.ilp = ILP_UNDEF;
 936                         set_irn_link(irn, op);
 937
 938                         irn = next;
 939                         continue;
 940                 }
 941
 942                 op = obstack_alloc(si->obst, sizeof(*op));
 943                 op->is_remat = 0;
 944                 op->attr.live_range.ilp = ILP_UNDEF;
 945                 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
 946                 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
 947                 set_irn_link(irn, op);
 948
 949                 args = pset_new_ptr_default();
 950
 951                 /* collect arguments of op */
 952                 for (n = get_irn_arity(irn)-1; n>=0; --n) {
 953                         ir_node        *arg = get_irn_n(irn, n);
 954
 955                         pset_insert_ptr(args, arg);
 956                 }
 957
 958                 /* set args of op already live in epilog */
 959                 pset_foreach(args, arg) {
 960                         if(has_reg_class(si, arg)) {
 961                                 pset_insert_ptr(live, arg);
 962                         }
 963                 }
 964                 /* delete defined value from live set */
 965                 if(has_reg_class(si, irn)) {
 966                         pset_remove_ptr(live, irn);
 967                 }
 968
 969
 970                 remat_args = pset_new_ptr_default();
 971
 972                 /* insert all possible remats before irn */
 973                 pset_foreach(args, arg) {
 974                         remat_info_t   *remat_info,
 975                                                     query;
 976                         remat_t        *remat;
 977
 978                         /* continue if the operand has the wrong reg class
 979                          */
 980                         if(!has_reg_class(si, arg))
 981                                 continue;
 982
 983                         query.irn = arg;
 984                         query.remats = NULL;
 985                         query.remats_by_operand = NULL;
 986                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 987
 988                         if(!remat_info) {
 989                                 continue;
 990                         }
 991
 992                         if(remat_info->remats) {
 993                                 pset_foreach(remat_info->remats, remat) {
 994                                         ir_node  *remat_irn = NULL;
 995
 996                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
 997 #ifdef REMAT_WHILE_LIVE
 998                                         if(pset_find_ptr(live, remat->value)) {
 999                                                 remat_irn = insert_remat_before(si, remat, irn, live);
1000                                         }
1001 #else
1002                                         remat_irn = insert_remat_before(si, remat, irn, live);
1003 #endif
1004                                         if(remat_irn) {
1005                                                 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1006                                                         ir_node  *remat_arg = get_irn_n(remat_irn, n);
1007
1008                                                         if(!has_reg_class(si, remat_arg)) continue;
1009
1010                                                         pset_insert_ptr(remat_args, remat_arg);
1011                                                 }
1012                                         }
1013                                 }
1014                         }
1015                 }
1016
1017                 /* now we add remat args to op's args because they could also die at this op */
1018                 pset_foreach(args,arg) {
1019                         if(pset_find_ptr(remat_args, arg)) {
1020                                 pset_remove_ptr(remat_args, arg);
1021                         }
1022                 }
1023                 pset_foreach(remat_args,arg) {
1024                         pset_insert_ptr(args, arg);
1025                 }
1026
1027                 /* insert all possible remats after irn */
1028                 pset_foreach(args, arg) {
1029                         remat_info_t   *remat_info,
1030                                                     query;
1031                         remat_t        *remat;
1032
1033                         /* continue if the operand has the wrong reg class */
1034                         if(!has_reg_class(si, arg))
1035                                 continue;
1036
1037                         query.irn = arg;
1038                         query.remats = NULL;
1039                         query.remats_by_operand = NULL;
1040                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1041
1042                         if(!remat_info) {
1043                                 continue;
1044                         }
1045
1046                         /* do not place post remats after jumps */
1047                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1048
1049                         if(remat_info->remats_by_operand) {
1050                                 pset_foreach(remat_info->remats_by_operand, remat) {
1051                                         /* do not insert remats producing the same value as one of the operands */
1052                                         if(!pset_find_ptr(args, remat->value)) {
1053                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
1054 #ifdef REMAT_WHILE_LIVE
1055                                                 if(pset_find_ptr(live, remat->value)) {
1056                                                         insert_remat_after(si, remat, irn, live);
1057                                                 }
1058 #else
1059                                                 insert_remat_after(si, remat, irn, live);
1060 #endif
1061                                         }
1062                                 }
1063                         }
1064                 }
1065
1066                 del_pset(remat_args);
1067                 del_pset(args);
1068                 irn = next;
1069         }
1070
1071         live_foreach(bb, li) {
1072                 ir_node        *value = (ir_node *) li->irn;
1073
1074                 /* add remats at end if successor has multiple predecessors */
1075                 if(is_merge_edge(bb)) {
1076                         /* add remats at end of block */
1077                         if (live_is_end(li) && has_reg_class(si, value)) {
1078                                 remat_info_t   *remat_info,
1079                                                            query;
1080                                 remat_t        *remat;
1081
1082                                 query.irn = value;
1083                                 query.remats = NULL;
1084                                 query.remats_by_operand = NULL;
1085                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1086
1087                                 if(remat_info && remat_info->remats) {
1088                                         pset_foreach(remat_info->remats, remat) {
1089                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
1090
1091                                                 insert_remat_before(si, remat, bb, NULL);
1092                                         }
1093                                 }
1094                         }
1095                 }
1096                 if(is_diverge_edge(bb)) {
1097                         /* add remat2s at beginning of block */
1098                         if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1099                                 remat_info_t   *remat_info,
1100                                                            query;
1101                                 remat_t        *remat;
1102
1103                                 query.irn = value;
1104                                 query.remats = NULL;
1105                                 query.remats_by_operand = NULL;
1106                                 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1107
1108                                 if(remat_info && remat_info->remats) {
1109                                         pset_foreach(remat_info->remats, remat) {
1110                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
1111
1112                                                 /* put the remat here if all its args are available */
1113                                                 insert_remat_after(si, remat, bb, NULL);
1114
1115                                         }
1116                                 }
1117                         }
1118                 }
1119         }
1120 }
1121
1122 /**
1123  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1124  */
1125 static void
1126 luke_endwalker(ir_node * bb, void * data)
1127 {
1128         spill_ilp_t    *si = (spill_ilp_t*)data;
1129         irn_live_t     *li;
1130         pset           *live;
1131         pset           *use_end;
1132         char            buf[256];
1133         ilp_cst_t       cst;
1134         ir_node        *irn;
1135         spill_bb_t     *spill_bb = get_irn_link(bb);
1136
1137
1138         live = pset_new_ptr_default();
1139         use_end = pset_new_ptr_default();
1140
1141         live_foreach(bb, li) {
1142                 irn = (ir_node *) li->irn;
1143                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1144                         op_t      *op;
1145
1146                         pset_insert_ptr(live, irn);
1147                         op = get_irn_link(irn);
1148                         assert(!op->is_remat);
1149                 }
1150         }
1151
1152         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1153         /* their reg_out must always be set */
1154         sched_foreach_reverse(bb, irn) {
1155                 int   n;
1156
1157                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1158
1159                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1160                         ir_node        *irn_arg = get_irn_n(irn, n);
1161
1162                         if(has_reg_class(si, irn_arg)) {
1163                                 pset_insert_ptr(use_end, irn_arg);
1164                         }
1165                 }
1166         }
1167
1168         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1169         //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1170         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1171
1172         spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1173
1174         /* if this is a merge edge we can reload at the end of this block */
1175         if(is_merge_edge(bb)) {
1176                 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1177         } else if(pset_count(use_end)){
1178                 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1179         } else {
1180                 spill_bb->reloads = NULL;
1181         }
1182
1183         pset_foreach(live,irn) {
1184                 spill_t     query,
1185                                         *spill;
1186                 double      spill_cost;
1187
1188
1189                 /* handle values used by control flow nodes later separately */
1190                 if(pset_find_ptr(use_end, irn)) continue;
1191
1192                 query.irn = irn;
1193                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1194
1195                 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1196
1197                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1198                 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1199                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1200
1201                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1202                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1203
1204                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1205                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1206
1207                 if(is_merge_edge(bb)) {
1208                         ilp_var_t   reload;
1209                         ilp_cst_t   rel_cst;
1210
1211                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1212                         reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1213                         set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1214
1215                         /* reload <= mem_out */
1216                         rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1217                         lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1218                         lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1219         }
1220
1221                 spill->reg_in = ILP_UNDEF;
1222                 spill->mem_in = ILP_UNDEF;
1223         }
1224
1225         pset_foreach(use_end,irn) {
1226                 spill_t     query,
1227                                         *spill;
1228                 double      spill_cost;
1229                 ilp_cst_t   end_use_req,
1230                                         rel_cst;
1231                 ilp_var_t   reload;
1232
1233                 query.irn = irn;
1234                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1235
1236                 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1237
1238                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1239                 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1240                 /* if irn is used at the end of the block, then it is live anyway */
1241                 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1242
1243                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1244                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1245
1246                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1247                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1248
1249                 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1250                 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1251                 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1252
1253                 /* reload <= mem_out */
1254                 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1255                 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1256                 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1257
1258                 spill->reg_in = ILP_UNDEF;
1259                 spill->mem_in = ILP_UNDEF;
1260
1261                 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1262                 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1263                 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1264         }
1265
1266         del_pset(live);
1267         del_pset(use_end);
1268 }
1269
1270 static ir_node *
1271 next_post_remat(const ir_node * irn)
1272 {
1273         op_t      *op;
1274
1275         if(is_Block(irn)) {
1276                 irn = sched_block_first_nonphi(irn);
1277         } else {
1278                 irn = sched_next_op(irn);
1279         }
1280
1281         if(sched_is_end(irn))
1282                 return NULL;
1283
1284         op = (op_t*)get_irn_link(irn);
1285         if(op->is_remat && !op->attr.remat.pre) {
1286                 return irn;
1287         }
1288
1289         return NULL;
1290 }
1291
1292
1293 static ir_node *
1294 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1295 {
1296         op_t      *op;
1297         ir_node   *ret;
1298
1299         if(is_Block(irn)) {
1300                 ret = sched_block_last_noncf(si, irn);
1301                 ret = sched_next(ret);
1302                 ret = sched_prev_op(ret);
1303         } else {
1304                 ret = sched_prev_op(irn);
1305         }
1306
1307         if(sched_is_end(ret) || is_Phi(ret))
1308                 return NULL;
1309
1310         op = (op_t*)get_irn_link(ret);
1311         if(op->is_remat && op->attr.remat.pre) {
1312                 return ret;
1313         }
1314
1315         return NULL;
1316 }
1317
1318 /**
1319  * Find a remat of value @p value in the epilog of @p pos
1320  */
1321 static ir_node *
1322 find_post_remat(const ir_node * value, const ir_node * pos)
1323 {
1324         while((pos = next_post_remat(pos)) != NULL) {
1325                 op_t   *op;
1326
1327                 op = get_irn_link(pos);
1328                 assert(op->is_remat && !op->attr.remat.pre);
1329
1330                 if(op->attr.remat.remat->value == value)
1331                         return (ir_node*)pos;
1332
1333 #if 0
1334         const ir_edge_t *edge;
1335                 foreach_out_edge(pos, edge) {
1336                         ir_node   *proj = get_edge_src_irn(edge);
1337                         assert(is_Proj(proj));
1338                 }
1339 #endif
1340
1341         }
1342
1343         return NULL;
1344 }
1345
1346 static spill_t *
1347 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1348 {
1349         spill_bb_t  *spill_bb = get_irn_link(bb);
1350         spill_t     *spill,
1351                                  query;
1352         char         buf[256];
1353
1354         query.irn = irn;
1355         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1356         if(!spill) {
1357                 double   spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1358
1359                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1360
1361                 spill->reg_out = ILP_UNDEF;
1362                 spill->reg_in  = ILP_UNDEF;
1363                 spill->mem_in  = ILP_UNDEF;
1364
1365                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1366                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1367
1368                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1369                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1370         }
1371
1372         return spill;
1373 }
1374
1375 static void
1376 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1377 {
1378         irn_live_t     *li;
1379         ir_node        *irn;
1380
1381         live_foreach(bb, li) {
1382                 irn = (ir_node *) li->irn;
1383
1384                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1385                         pset_insert_ptr(live, irn);
1386                 }
1387         }
1388
1389         irn = sched_last(bb);
1390
1391         /* all values eaten by control flow operations are also live until the end of the block */
1392         sched_foreach_reverse(bb, irn) {
1393                 int  i;
1394
1395                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1396
1397                 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1398                         ir_node *arg = get_irn_n(irn,i);
1399
1400                         if(has_reg_class(si, arg)) {
1401                                 pset_insert_ptr(live, arg);
1402                         }
1403                 }
1404         }
1405 }
1406
1407 /**
1408  *  Inserts ILP-constraints and variables for memory copying before the given position
1409  */
1410 static void
1411 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1412 {
1413         const ir_node    *succ;
1414         const ir_edge_t  *edge;
1415         spill_bb_t       *spill_bb = get_irn_link(block);
1416         ir_node          *phi;
1417         int               pos;
1418         ilp_cst_t         cst;
1419         ilp_var_t         copyreg;
1420         char              buf[256];
1421         ir_node          *tmp;
1422
1423
1424         assert(edges_activated(current_ir_graph));
1425
1426         edge = get_block_succ_first(block);
1427         if(!edge) return;
1428
1429         succ = edge->src;
1430         pos = edge->pos;
1431
1432         edge = get_block_succ_next(block, edge);
1433         /* next block can only contain phis, if this is a merge edge */
1434         if(edge) return;
1435
1436         ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1437         copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1438
1439         ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1440         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1441
1442         pset_foreach(live, tmp) {
1443                 spill_t  *spill;
1444 #if 0
1445                 op_t  *op = get_irn_link(irn);
1446                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1447 #endif
1448                 spill = set_find_spill(spill_bb->ilp, tmp);
1449                 assert(spill);
1450
1451                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1452         }
1453         lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1454
1455         sched_foreach(succ, phi) {
1456                 const ir_node  *to_copy;
1457                 op_t           *to_copy_op;
1458                 spill_t        *to_copy_spill;
1459                 op_t           *phi_op = get_irn_link(phi);
1460                 ilp_var_t       reload = ILP_UNDEF;
1461
1462
1463                 if(!is_Phi(phi)) break;
1464                 if(!has_reg_class(si, phi)) continue;
1465
1466                 to_copy = get_irn_n(phi, pos);
1467
1468                 to_copy_op = get_irn_link(to_copy);
1469
1470                 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1471                 assert(to_copy_spill);
1472
1473                 if(spill_bb->reloads) {
1474                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1475
1476                         if(keyval) {
1477                                 reload = PTR_TO_INT(keyval->val);
1478                         }
1479                 }
1480
1481                 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1482                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1483
1484                 /* copy - reg_out - reload - remat - live_range <= 0 */
1485                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1486                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1487                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1488                 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1489                 foreach_pre_remat(si, block, tmp) {
1490                         op_t     *remat_op = get_irn_link(tmp);
1491                         if(remat_op->attr.remat.remat->value == to_copy) {
1492                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1493                         }
1494                 }
1495
1496                 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1497                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1498
1499                 /* copy - reg_out - copyreg <= 0 */
1500                 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1501                 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1502                 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1503         }
1504 }
1505
1506
1507 /**
1508  * Walk all irg blocks and emit this ILP
1509  */
1510 static void
1511 luke_blockwalker(ir_node * bb, void * data)
1512 {
1513         spill_ilp_t    *si = (spill_ilp_t*)data;
1514         ir_node        *irn;
1515         pset           *live;
1516         char            buf[256];
1517         ilp_cst_t       cst;
1518         spill_bb_t     *spill_bb = get_irn_link(bb);
1519         ir_node        *tmp;
1520         spill_t        *spill;
1521         pset           *defs = pset_new_ptr_default();
1522
1523
1524         live = pset_new_ptr_default();
1525
1526         /****************************************
1527          *      B A S I C  B L O C K  E N D
1528          ***************************************/
1529
1530
1531         /* init live values at end of block */
1532         get_live_end(si, bb, live);
1533
1534         pset_foreach(live, irn) {
1535                 op_t           *op;
1536                 ilp_var_t       reload = ILP_UNDEF;
1537
1538                 spill = set_find_spill(spill_bb->ilp, irn);
1539                 assert(spill);
1540
1541                 if(spill_bb->reloads) {
1542                         keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1543
1544                         if(keyval) {
1545                                 reload = PTR_TO_INT(keyval->val);
1546                         }
1547                 }
1548
1549                 op = get_irn_link(irn);
1550                 assert(!op->is_remat);
1551
1552                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1553                 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1554                 op->attr.live_range.op = bb;
1555
1556                 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1557                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1558
1559                 /* reg_out - reload - remat - live_range <= 0 */
1560                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1561                 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1562                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1563                 foreach_pre_remat(si, bb, tmp) {
1564                         op_t     *remat_op = get_irn_link(tmp);
1565                         if(remat_op->attr.remat.remat->value == irn) {
1566                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1567                         }
1568                 }
1569                 /* maybe we should also assure that reg_out >= live_range etc. */
1570         }
1571
1572 #ifndef NO_MEMCOPIES
1573         insert_mem_copy_position(si, live, bb);
1574 #endif
1575
1576         /*
1577          * start new live ranges for values used by remats at end of block
1578          * and assure the remat args are available
1579          */
1580         foreach_pre_remat(si, bb, tmp) {
1581                 op_t     *remat_op = get_irn_link(tmp);
1582                 int       n;
1583
1584                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1585                         ir_node        *remat_arg = get_irn_n(tmp, n);
1586                         op_t           *arg_op = get_irn_link(remat_arg);
1587                         ilp_var_t       prev_lr;
1588
1589                         if(!has_reg_class(si, remat_arg)) continue;
1590
1591                         /* if value is becoming live through use by remat */
1592                         if(!pset_find_ptr(live, remat_arg)) {
1593                                 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1594                                 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1595
1596                                 arg_op->attr.live_range.ilp = prev_lr;
1597                                 arg_op->attr.live_range.op = bb;
1598
1599                                 DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1600
1601                                 pset_insert_ptr(live, remat_arg);
1602                                 add_to_spill_bb(si, bb, remat_arg);
1603                         }
1604
1605                         /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1606                         ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1607                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1608
1609                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1610                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1611
1612                         /* use reload placed for this argument */
1613                         if(spill_bb->reloads) {
1614                                 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1615
1616                                 if(keyval) {
1617                                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
1618
1619                                         lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1620                                 }
1621                         }
1622                 }
1623         }
1624         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1625
1626
1627
1628
1629         /**************************************
1630          *    B A S I C  B L O C K  B O D Y
1631          **************************************/
1632
1633         sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1634                 op_t       *op;
1635                 op_t       *tmp_op;
1636                 int         n,
1637                                         u = 0,
1638                                         d = 0;
1639                 ilp_cst_t       check_pre,
1640                                         check_post;
1641                 set        *args;
1642                 pset       *used;
1643                 pset       *remat_defs;
1644                 keyval_t   *keyval;
1645
1646                 /* iterate only until first phi */
1647                 if(is_Phi(irn))
1648                         break;
1649
1650                 op = get_irn_link(irn);
1651                 /* skip remats */
1652                 if(op->is_remat) continue;
1653                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1654
1655                 /* collect defined values */
1656                 if(has_reg_class(si, irn)) {
1657                         pset_insert_ptr(defs, irn);
1658                 }
1659
1660                 /* skip projs */
1661                 if(is_Proj(irn)) continue;
1662
1663                 /*
1664                  * init set of irn's arguments
1665                  * and all possibly used values around this op
1666                  * and values defined by post remats
1667                  */
1668                 args =       new_set(cmp_keyval, get_irn_arity(irn));
1669                 used =       pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1670                 remat_defs = pset_new_ptr(pset_count(live));
1671
1672                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1673                         ir_node        *irn_arg = get_irn_n(irn, n);
1674                         if(has_reg_class(si, irn_arg)) {
1675                                 set_insert_keyval(args, irn_arg, (void*)n);
1676                                 pset_insert_ptr(used, irn_arg);
1677                         }
1678                 }
1679                 foreach_post_remat(irn, tmp) {
1680                         op_t    *remat_op = get_irn_link(tmp);
1681
1682                         pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1683
1684                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1685                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1686                                 if(has_reg_class(si, remat_arg)) {
1687                                         pset_insert_ptr(used, remat_arg);
1688                                 }
1689                         }
1690                 }
1691                 foreach_pre_remat(si, irn, tmp) {
1692                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1693                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1694                                 if(has_reg_class(si, remat_arg)) {
1695                                         pset_insert_ptr(used, remat_arg);
1696                                 }
1697                         }
1698                 }
1699
1700                 /**********************************
1701                  *   I N  E P I L O G  O F  irn
1702                  **********************************/
1703
1704                 /* ensure each dying value is used by only one post remat */
1705                 pset_foreach(used, tmp) {
1706                         ir_node     *value = tmp;
1707                         op_t        *value_op = get_irn_link(value);
1708                         ir_node     *remat;
1709                         int          n_remats = 0;
1710
1711                         cst = ILP_UNDEF;
1712                         foreach_post_remat(irn, remat) {
1713                                 op_t  *remat_op = get_irn_link(remat);
1714
1715                                 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1716                                         ir_node   *remat_arg = get_irn_n(remat, n);
1717
1718                                         /* if value is used by this remat add it to constraint */
1719                                         if(remat_arg == value) {
1720                                                 if(n_remats == 0) {
1721                                                         /* sum remat2s <= 1 + n_remats*live_range */
1722                                                         ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1723                                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
1724                                                 }
1725
1726                                                 n_remats++;
1727                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1728                                                 break;
1729                                         }
1730                                 }
1731                         }
1732
1733             // value_op->attr.live_range.ilp != ILP_UNDEF
1734                         if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1735                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1736                         }
1737                 }
1738
1739         /* ensure at least one value dies at post remat */
1740         foreach_post_remat(irn, tmp) {
1741             op_t     *remat_op = get_irn_link(tmp);
1742             pset     *remat_args = pset_new_ptr(get_irn_arity(tmp));
1743             ir_node  *remat_arg;
1744
1745             for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1746                 remat_arg = get_irn_n(tmp, n);
1747
1748                 if(has_reg_class(si, remat_arg)) {
1749
1750                     /* does arg always die at this op? */
1751                     if(!pset_find_ptr(live, remat_arg))
1752                         goto skip_one_must_die;
1753
1754                     pset_insert_ptr(remat_args, remat_arg);
1755                 }
1756             }
1757
1758             /* remat + \sum live_range(remat_arg) <= |args| */
1759             ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1760             cst = lpp_add_cst(si->lpp, buf, lpp_less, pset_count(remat_args));
1761             lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1762
1763             pset_foreach(remat_args, remat_arg) {
1764                 op_t  *arg_op = get_irn_link(remat_arg);
1765
1766                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1767             }
1768
1769 skip_one_must_die:
1770             del_pset(remat_args);
1771         }
1772
1773                 /* new live ranges for values from L\U defined by post remats */
1774                 pset_foreach(live, tmp) {
1775                         ir_node     *value = tmp;
1776                         op_t        *value_op = get_irn_link(value);
1777
1778                         if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1779                                 ilp_var_t    prev_lr = ILP_UNDEF;
1780                                 ir_node     *remat;
1781
1782                                 if(pset_find_ptr(remat_defs, value)) {
1783
1784                                         /* next_live_range <= prev_live_range + sum remat2s */
1785                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1786                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1787
1788                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1789                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1790
1791                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1792                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1793
1794                                         foreach_post_remat(irn, remat) {
1795                                                 op_t        *remat_op = get_irn_link(remat);
1796
1797                                                 /* if value is being rematerialized by this remat */
1798                                                 if(value == remat_op->attr.remat.remat->value) {
1799                                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1800                                                 }
1801                                         }
1802
1803                                         value_op->attr.live_range.ilp = prev_lr;
1804                                         value_op->attr.live_range.op = irn;
1805                                 }
1806                         }
1807                 }
1808
1809                 /* requirements for post remats and start live ranges from L/U' for values dying here */
1810                 foreach_post_remat(irn, tmp) {
1811                         op_t        *remat_op = get_irn_link(tmp);
1812                         int          n;
1813
1814                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1815                                 ir_node        *remat_arg = get_irn_n(tmp, n);
1816                                 op_t           *arg_op = get_irn_link(remat_arg);
1817
1818                                 if(!has_reg_class(si, remat_arg)) continue;
1819
1820                                 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1821                                 if(!pset_find_ptr(used, remat_arg)) {
1822                                         /* remat <= live_range(remat_arg) */
1823                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1824                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1825
1826                                         /* if value is becoming live through use by remat2 */
1827                                         if(!pset_find_ptr(live, remat_arg)) {
1828                                                 ilp_var_t     lr;
1829
1830                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1831                                                 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1832
1833                                                 arg_op->attr.live_range.ilp = lr;
1834                                                 arg_op->attr.live_range.op = irn;
1835
1836                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1837
1838                                                 pset_insert_ptr(live, remat_arg);
1839                                                 add_to_spill_bb(si, bb, remat_arg);
1840                                         }
1841
1842                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1843                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1844                                 }
1845                         }
1846                 }
1847
1848                 d = pset_count(defs);
1849                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
1850
1851                 /* count how many regs irn needs for arguments */
1852                 u = set_count(args);
1853
1854
1855                 /* check the register pressure in the epilog */
1856                 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1857                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1858                 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1859
1860                 /* add L\U' to check_post */
1861                 pset_foreach(live, tmp) {
1862                         if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1863                                 /* if a live value is not used by irn */
1864                                 tmp_op = get_irn_link(tmp);
1865                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1866                         }
1867                 }
1868
1869                 /***********************************************************
1870                  *  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
1871                  **********************************************************/
1872
1873
1874                 pset_foreach(used, tmp) {
1875                         ilp_var_t       prev_lr;
1876                         ilp_var_t       post_use;
1877                         int             p = 0;
1878                         spill_t        *spill;
1879                         ir_node        *arg = tmp;
1880                         op_t           *arg_op = get_irn_link(arg);
1881                         ir_node        *remat;
1882
1883                         spill = add_to_spill_bb(si, bb, arg);
1884
1885                         /* new live range for each used value */
1886                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1887                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1888
1889                         /* the epilog stuff - including post_use, check_post, check_post_remat */
1890                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1891                         post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1892
1893                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1894
1895                         /* arg is live throughout epilog if the next live_range is in a register */
1896                         if(pset_find_ptr(live, arg)) {
1897                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
1898
1899                                 /* post_use >= next_lr + remat */
1900                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1901                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1902                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1903                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1904
1905                         }
1906
1907                         /* if value is not an arg of op and not possibly defined by post remat
1908                          * then it may only die and not become live
1909                          */
1910                         if(!set_find_keyval(args, arg)) {
1911                                 /* post_use <= prev_lr */
1912                                 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1913                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1914                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1915                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1916
1917                                 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1918                                         /* next_lr <= prev_lr */
1919                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1920                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1921                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1922                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1923                                 }
1924                         }
1925
1926
1927
1928                         /* forall post remat which use arg add a similar cst */
1929                         foreach_post_remat(irn, remat) {
1930                                 int      n;
1931
1932                                 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1933                                         ir_node    *remat_arg = get_irn_n(remat, n);
1934                                         op_t       *remat_op = get_irn_link(remat);
1935
1936                                         if(remat_arg == arg) {
1937                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
1938
1939                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1940                                                 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1941                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1942                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1943                                         }
1944                                 }
1945                         }
1946
1947                         /* new live range begins for each used value */
1948                         arg_op->attr.live_range.ilp = prev_lr;
1949                         arg_op->attr.live_range.op = irn;
1950
1951                         /*if(!pset_find_ptr(live, arg)) {
1952                                 pset_insert_ptr(live, arg);
1953                                 add_to_spill_bb(si, bb, arg);
1954                         }*/
1955                         pset_insert_ptr(live, arg);
1956
1957                 }
1958
1959                 /* just to be sure */
1960                 check_post = ILP_UNDEF;
1961
1962
1963
1964
1965                 /******************
1966                  *   P R O L O G
1967                  ******************/
1968
1969                 /* check the register pressure in the prolog */
1970                 /* sum_{L\U} lr <= k - |U| */
1971                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1972                 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1973
1974                 /* for the prolog remove defined values from the live set */
1975                 pset_foreach(defs, tmp) {
1976                         pset_remove_ptr(live, tmp);
1977                 }
1978
1979                 /***********************************************************
1980                  *  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
1981                  **********************************************************/
1982
1983
1984                 set_foreach(args, keyval) {
1985                         spill_t        *spill;
1986                         ir_node        *arg = keyval->key;
1987                         int             i = PTR_TO_INT(keyval->val);
1988                         op_t           *arg_op = get_irn_link(arg);
1989
1990                         spill = set_find_spill(spill_bb->ilp, arg);
1991                         assert(spill);
1992
1993                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1994                         op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1995
1996                         /* reload <= mem_out */
1997                         ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1998                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1999                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2000                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2001
2002                         /* requirement: arg must be in register for use */
2003                         /* reload + remat + live_range == 1 */
2004                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2005                         cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
2006
2007                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2008                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2009                         foreach_pre_remat(si, irn, tmp) {
2010                                 op_t     *remat_op = get_irn_link(tmp);
2011                                 if(remat_op->attr.remat.remat->value == arg) {
2012                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2013                                 }
2014                         }
2015                 }
2016
2017                 /* iterate over L\U */
2018                 pset_foreach(live, tmp) {
2019                         if(!set_find_keyval(args, tmp)) {
2020                                 /* if a live value is not used by irn */
2021                                 tmp_op = get_irn_link(tmp);
2022                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2023                         }
2024                 }
2025
2026
2027                 /* requirements for remats */
2028                 /* start new live ranges for values used by remats */
2029                 foreach_pre_remat(si, irn, tmp) {
2030                         op_t        *remat_op = get_irn_link(tmp);
2031                         int          n;
2032
2033                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2034                                 ir_node        *remat_arg = get_irn_n(tmp, n);
2035                                 op_t           *arg_op = get_irn_link(remat_arg);
2036                                 ilp_var_t       prev_lr;
2037
2038                                 if(!has_reg_class(si, remat_arg)) continue;
2039
2040                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2041                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2042                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2043
2044                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2045                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2046
2047                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
2048                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2049                                         int    index = (int)keyval->val;
2050
2051                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2052                                 }
2053                         }
2054                 }
2055
2056
2057
2058
2059                 /*************************
2060                  *  D O N E  W I T H  O P
2061                  *************************/
2062
2063                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
2064
2065                 pset_foreach(live, tmp) {
2066                         assert(has_reg_class(si, tmp));
2067                 }
2068
2069                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2070                         ir_node        *arg = get_irn_n(irn, n);
2071
2072                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2073                 }
2074
2075                 del_pset(remat_defs);
2076                 del_pset(used);
2077                 del_set(args);
2078                 del_pset(defs);
2079                 defs = pset_new_ptr_default();
2080         }
2081
2082
2083
2084         /***************************************
2085          *   B E G I N N I N G  O F  B L O C K
2086          ***************************************/
2087
2088
2089         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2090         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
2091
2092         pset_foreach(live, irn) {
2093                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2094         }
2095
2096         /* construct mem_outs for all values */
2097
2098         set_foreach(spill_bb->ilp, spill) {
2099                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2100                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2101
2102                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2103                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2104
2105                 if(pset_find_ptr(live, spill->irn)) {
2106                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
2107
2108                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2109                         spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2110                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2111
2112                         if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2113                                 int   n;
2114                                 op_t *op = get_irn_link(spill->irn);
2115
2116                                 /* do we have to copy a phi argument? */
2117                                 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2118                                 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2119
2120                                 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2121                                         const ir_node  *arg = get_irn_n(spill->irn, n);
2122                                         double          freq=0.0;
2123                                         int             m;
2124                                         ilp_var_t       var;
2125
2126
2127                                         /* argument already done? */
2128                                         if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2129
2130                                         /* get sum of execution frequencies of blocks with the same phi argument */
2131                                         for(m=n; m>=0; --m) {
2132                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2133
2134                                                 if(arg==arg2) {
2135                                                         freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2136                                                 }
2137                                         }
2138
2139                                         /* copies are not for free */
2140                                         ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2141                                         var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2142
2143                                         for(m=n; m>=0; --m) {
2144                                                 const ir_node  *arg2 = get_irn_n(spill->irn, m);
2145
2146                                                 if(arg==arg2) {
2147                                                         op->attr.live_range.args.copies[m] = var;
2148                                                 }
2149                                         }
2150
2151                                         /* copy <= mem_in */
2152                                         ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2153                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2154                                         lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2155                                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2156                                 }
2157                         }
2158                 }
2159         }
2160
2161
2162         /* L\U is empty at bb start */
2163         /* arg is live throughout epilog if it is reg_in into this block */
2164
2165         /* check the register pressure at the beginning of the block
2166          * including remats
2167          */
2168         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2169         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2170
2171         pset_foreach(live, irn) {
2172         ilp_cst_t  nospill;
2173
2174                 spill = set_find_spill(spill_bb->ilp, irn);
2175                 assert(spill);
2176
2177                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2178                 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2179
2180                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2181
2182                 /* spill + mem_in <= 1 */
2183                 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2184                 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2185
2186                 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2187                 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2188
2189         }
2190         foreach_post_remat(bb, irn) {
2191                 op_t     *remat_op = get_irn_link(irn);
2192
2193                 DBG((si->dbg, LEVEL_4, "\t  next post remat: %+F\n", irn));
2194                 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2195
2196                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2197         }
2198
2199         /* forall post remats add requirements */
2200         foreach_post_remat(bb, tmp) {
2201                 int         n;
2202
2203                 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2204                         ir_node    *remat_arg = get_irn_n(tmp, n);
2205                         op_t       *remat_op = get_irn_link(tmp);
2206
2207                         if(!has_reg_class(si, remat_arg)) continue;
2208
2209                         spill = set_find_spill(spill_bb->ilp, remat_arg);
2210                         assert(spill);
2211
2212                         /* remat <= reg_in_argument */
2213                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2214                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2215                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2216                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2217                 }
2218         }
2219
2220         /* mem_in/reg_in for live_in values, especially phis and their arguments */
2221         pset_foreach(live, irn) {
2222                 int          p = 0,
2223                                          n;
2224
2225                 spill = set_find_spill(spill_bb->ilp, irn);
2226                 assert(spill && spill->irn == irn);
2227
2228                 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2229                         for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2230                                 ilp_cst_t       mem_in,
2231                                                                 reg_in;
2232                                 ir_node        *phi_arg = get_Phi_pred(irn, n);
2233                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2234                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2235                                 spill_t        *spill_p;
2236                                 op_t           *op = get_irn_link(irn);
2237
2238                                 /* although the phi is in the right regclass one or more of
2239                                  * its arguments can be in a different one or at least to
2240                                  * ignore
2241                                  */
2242                                 if(has_reg_class(si, phi_arg)) {
2243                                         /* mem_in < mem_out_arg + copy */
2244                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2245                                         mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2246
2247                                         /* reg_in < reg_out_arg */
2248                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2249                                         reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2250
2251                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2252                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2253
2254                                         spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2255                                         assert(spill_p);
2256
2257                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2258                                         lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2259                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2260                                 }
2261                         }
2262                 } else {
2263                         /* else assure the value arrives on all paths in the same resource */
2264
2265                         for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2266                                 ilp_cst_t       mem_in,
2267                                                                 reg_in;
2268                                 ir_node        *bb_p = get_Block_cfgpred_block(bb, n);
2269                                 spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
2270                                 spill_t        *spill_p;
2271
2272                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2273                                 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2274                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2275                                 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2276
2277                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2278                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2279
2280                                 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2281                                 assert(spill_p);
2282
2283                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2284                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2285                         }
2286                 }
2287         }
2288
2289         /* first live ranges from reg_ins */
2290         pset_foreach(live, irn) {
2291                 op_t      *op = get_irn_link(irn);
2292
2293                 spill = set_find_spill(spill_bb->ilp, irn);
2294                 assert(spill && spill->irn == irn);
2295
2296                 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2297                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2298                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2299                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2300
2301                 foreach_post_remat(bb, tmp) {
2302                         op_t     *remat_op = get_irn_link(tmp);
2303
2304                         if(remat_op->attr.remat.remat->value == irn) {
2305                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2306                         }
2307                 }
2308         }
2309
2310         /* walk forward now and compute constraints for placing spills */
2311         /* this must only be done for values that are not defined in this block */
2312         /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2313         pset_foreach(live, irn) {
2314                 /*
2315                  * if value is defined in this block we can anways place the spill directly after the def
2316                  *    -> no constraint necessary
2317                  */
2318                 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2319
2320
2321                 spill = set_find_spill(spill_bb->ilp, irn);
2322                 assert(spill);
2323
2324                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2325                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2326
2327                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2328                 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2329
2330                 if(!is_Phi(irn)) {
2331                         sched_foreach_op(bb, tmp) {
2332                                 op_t   *op = get_irn_link(tmp);
2333
2334                                 if(is_Phi(tmp)) continue;
2335                                 assert(!is_Proj(tmp));
2336
2337                                 if(op->is_remat) {
2338                                         ir_node   *value = op->attr.remat.remat->value;
2339
2340                                         if(value == irn) {
2341                                                 /* only collect remats up to the first real use of a value */
2342                                                 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2343                                         }
2344                                 } else {
2345                                         int   n;
2346
2347                                         for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2348                                                 ir_node    *arg = get_irn_n(tmp, n);
2349
2350                                                 if(arg == irn) {
2351                                                         /* if a value is used stop collecting remats */
2352                             goto next_live;
2353                                                 }
2354                                         }
2355                                 }
2356                         }
2357                 }
2358 next_live:
2359         }
2360
2361         del_pset(live);
2362 }
2363
2364 typedef struct _irnlist_t {
2365         struct list_head   list;
2366         ir_node           *irn;
2367 } irnlist_t;
2368
2369 typedef struct _interference_t {
2370         struct list_head    blocklist;
2371         ir_node            *a;
2372         ir_node            *b;
2373 } interference_t;
2374
2375 static int
2376 cmp_interference(const void *a, const void *b, size_t size)
2377 {
2378         const interference_t *p = a;
2379         const interference_t *q = b;
2380
2381         return !(p->a == q->a && p->b == q->b);
2382 }
2383
2384 static interference_t *
2385 set_find_interference(set * set, ir_node * a, ir_node * b)
2386 {
2387         interference_t     query;
2388
2389         query.a = (a>b)?a:b;
2390         query.b = (a>b)?b:a;
2391
2392         return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2393 }
2394
2395 static interference_t *
2396 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2397 {
2398         interference_t     query,
2399                                           *result;
2400         irnlist_t         *list = obstack_alloc(si->obst, sizeof(*list));
2401
2402         list->irn = bb;
2403
2404         result = set_find_interference(set, a, b);
2405         if(result) {
2406
2407                 list_add(&list->list, &result->blocklist);
2408                 return result;
2409         }
2410
2411         query.a = (a>b)?a:b;
2412         query.b = (a>b)?b:a;
2413
2414         result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2415
2416         INIT_LIST_HEAD(&result->blocklist);
2417         list_add(&list->list, &result->blocklist);
2418
2419         return result;
2420 }
2421
2422 static int
2423 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2424 {
2425         const ir_edge_t *edge;
2426
2427         if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2428                 /* both values are live in, so they interfere */
2429                 return 1;
2430         }
2431
2432         /* ensure a dominates b */
2433         if(value_dominates(b,a)) {
2434                 const ir_node * t;
2435                 t = b;
2436                 b = a;
2437                 a = t;
2438         }
2439         assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2440
2441
2442         /* the following code is stolen from bera.c */
2443         if(is_live_end(bb, a))
2444                 return 1;
2445
2446         foreach_out_edge(a, edge) {
2447                 const ir_node *user = edge->src;
2448                 if(get_nodes_block(user) == bb
2449                                 && !is_Phi(user)
2450                                 && b != user
2451                                 && value_dominates(b, user))
2452                         return 1;
2453         }
2454
2455         return 0;
2456 }
2457
2458 /**
2459  * Walk all irg blocks and collect interfering values inside of phi classes
2460  */
2461 static void
2462 luke_interferencewalker(ir_node * bb, void * data)
2463 {
2464         spill_ilp_t    *si = (spill_ilp_t*)data;
2465         irn_live_t     *li1,
2466                        *li2;
2467
2468         live_foreach(bb, li1) {
2469                 ir_node        *a = (ir_node *) li1->irn;
2470                 op_t           *a_op = get_irn_link(a);
2471
2472                 if(a_op->is_remat) continue;
2473
2474                 /* a is only interesting if it is in my register class and if it is inside a phi class */
2475                 if (has_reg_class(si, a) && get_phi_class(a)) {
2476                         for(li2=li1->next; li2; li2 = li2->next) {
2477                                 ir_node        *b = (ir_node *) li2->irn;
2478                                 op_t           *b_op = get_irn_link(b);
2479
2480                                 if(b_op->is_remat) continue;
2481
2482                                 /* a and b are only interesting if they are in the same phi class */
2483                                 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2484                                         if(values_interfere_in_block(bb, a, b)) {
2485                                                 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2486                                                 set_insert_interference(si, si->interferences, a, b, bb);
2487                                         }
2488                                 }
2489                         }
2490                 }
2491         }
2492 }
2493
2494 static unsigned int copy_path_id = 0;
2495
2496 static void
2497 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2498 {
2499         ilp_cst_t  cst;
2500         ilp_var_t  copy;
2501         char       buf[256];
2502         void      *ptr;
2503
2504         ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2505         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2506
2507         lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2508
2509         pset_foreach(copies, ptr) {
2510                 copy = PTR_TO_INT(ptr);
2511                 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2512         }
2513 }
2514
2515 /**
2516  * @parameter copies   contains a path of copies which lead us to irn
2517  * @parameter visited  contains a set of nodes already visited on this path
2518  */
2519 static int
2520 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2521 {
2522         ir_edge_t *edge;
2523         op_t      *op = get_irn_link(irn);
2524     pset      *visited_users = pset_new_ptr_default();
2525         int        paths = 0;
2526
2527         if(op->is_remat) return;
2528
2529         pset_insert_ptr(visited, irn);
2530
2531         if(is_Phi(irn)) {
2532                 int    n;
2533         pset  *visited_operands = pset_new_ptr(get_irn_arity(irn));
2534
2535                 /* visit all operands */
2536                 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2537                         ir_node  *arg = get_irn_n(irn, n);
2538                         ilp_var_t  copy = op->attr.live_range.args.copies[n];
2539
2540                         if(!has_reg_class(si, arg)) continue;
2541             if(pset_find_ptr(visited_operands, arg)) continue;
2542             pset_insert_ptr(visited_operands, arg);
2543
2544                         if(arg == target) {
2545                                 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2546                                         del_pset(visited_operands);
2547                                         del_pset(visited_users);
2548                                         pset_remove_ptr(visited, irn);
2549                                         return paths;
2550                                 }
2551                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2552                                 write_copy_path_cst(si, copies, any_interfere);
2553                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2554                         } else if(!pset_find_ptr(visited, arg)) {
2555                                 pset_insert(copies, INT_TO_PTR(copy), copy);
2556                                 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2557                                 pset_remove(copies, INT_TO_PTR(copy), copy);
2558
2559                 /*if(paths > MAX_PATHS) {
2560                     if(pset_count(copies) == 0) {
2561                         ilp_cst_t  cst;
2562                         char       buf[256];
2563
2564                         ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2565                         cst = lpp_add_cst(si->lpp, buf, lpp_equal, 0);
2566                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2567                         lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2568                         DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2569
2570                         paths = 0;
2571                     } else {
2572                         del_pset(visited_operands);
2573                         del_pset(visited_users);
2574                         pset_remove_ptr(visited, irn);
2575                         return paths;
2576                     }
2577                 } else if(pset_count(copies) == 0) {
2578                                         paths = 0;
2579                                 }*/
2580                         }
2581                 }
2582
2583         del_pset(visited_operands);
2584         }
2585
2586         /* visit all uses which are phis */
2587         foreach_out_edge(irn, edge) {
2588                 ir_node  *user = edge->src;
2589                 int       pos  = edge->pos;
2590                 op_t     *op = get_irn_link(user);
2591                 ilp_var_t copy;
2592
2593                 if(!is_Phi(user)) continue;
2594                 if(!has_reg_class(si, user)) continue;
2595         if(pset_find_ptr(visited_users, user)) continue;
2596         pset_insert_ptr(visited_users, user);
2597
2598                 copy = op->attr.live_range.args.copies[pos];
2599
2600                 if(user == target) {
2601                         if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2602                                 del_pset(visited_users);
2603                                 pset_remove_ptr(visited, irn);
2604                                 return paths;
2605                         }
2606                         pset_insert(copies, INT_TO_PTR(copy), copy);
2607                         write_copy_path_cst(si, copies, any_interfere);
2608                         pset_remove(copies, INT_TO_PTR(copy), copy);
2609                 } else if(!pset_find_ptr(visited, user)) {
2610                         pset_insert(copies, INT_TO_PTR(copy), copy);
2611                         paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2612                         pset_remove(copies, INT_TO_PTR(copy), copy);
2613
2614             /*if(paths > MAX_PATHS) {
2615                 if(pset_count(copies) == 0) {
2616                     ilp_cst_t  cst;
2617                     char       buf[256];
2618
2619                     ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2620                     cst = lpp_add_cst(si->lpp, buf, lpp_equal, 0);
2621                     lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2622                     lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2623                     DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2624
2625                     paths = 0;
2626                 } else {
2627                     del_pset(visited_users);
2628                     pset_remove_ptr(visited, irn);
2629                     return paths;
2630                 }
2631             } else if(pset_count(copies) == 0) {
2632                                 paths = 0;
2633                         }*/
2634                 }
2635         }
2636
2637     del_pset(visited_users);
2638         pset_remove_ptr(visited, irn);
2639         return paths;
2640 }
2641
2642 static void
2643 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2644 {
2645         pset * copies = pset_new_ptr_default();
2646         pset * visited = pset_new_ptr_default();
2647
2648         find_copy_path(si, a, b, any_interfere, copies, visited);
2649
2650         del_pset(visited);
2651         del_pset(copies);
2652 }
2653
2654
2655 static void
2656 memcopyhandler(spill_ilp_t * si)
2657 {
2658         interference_t   *interference;
2659         char              buf[256];
2660         /* teste Speicherwerte auf Interferenz */
2661
2662         /* analyze phi classes */
2663         phi_class_compute(si->chordal_env->irg);
2664
2665         DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2666         irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2667
2668         /* now lets emit the ILP unequations for the crap */
2669         set_foreach(si->interferences, interference) {
2670                 irnlist_t      *irnlist;
2671                 ilp_var_t       interfere,
2672                                                 any_interfere;
2673                 ilp_cst_t       any_interfere_cst,
2674                                                 cst;
2675                 const ir_node  *a  = interference->a;
2676                 const ir_node  *b  = interference->b;
2677
2678                 /* any_interf <= \sum interf */
2679                 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2680                 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2681                 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2682
2683                 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2684
2685                 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2686                         const ir_node  *bb = irnlist->irn;
2687                         spill_bb_t     *spill_bb = get_irn_link(bb);
2688                         spill_t        *spilla,
2689                                                    *spillb,
2690                                                    query;
2691                         char           buf[256];
2692
2693                         query.irn = a;
2694                         spilla = set_find_spill(spill_bb->ilp, a);
2695                         assert(spilla);
2696
2697                         query.irn = b;
2698                         spillb = set_find_spill(spill_bb->ilp, b);
2699                         assert(spillb);
2700
2701                         /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2702                         /* 1:   mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2703                         /* 2: - mem_in_a - spill_a + interfere <= 0 */
2704                         /* 3: - mem_in_b - spill_b + interfere <= 0 */
2705                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2706                         interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2707
2708                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2709                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2710
2711                         lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2712                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2713                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2714                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2715                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2716
2717                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2718                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2719
2720                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2721                         if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2722                         lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2723
2724                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2725                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2726
2727                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2728                         if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2729                         lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2730
2731
2732                         lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2733
2734                         /* any_interfere >= interf */
2735                         ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2736                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2737
2738                         lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2739                         lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2740                 }
2741
2742                 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2743                 gen_copy_constraints(si,a,b,any_interfere);
2744         }
2745 }
2746
2747
2748 static INLINE int
2749 is_zero(double x)
2750 {
2751         return fabs(x) < 0.00001;
2752 }
2753
2754 #ifdef KEEPALIVE
2755 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2756 {
2757         spill_ilp_t *si = get_irg_link(current_ir_graph);
2758
2759         if(pset_find_ptr(si->all_possible_remats, n)) {
2760                 op_t   *op = (op_t*)get_irn_link(n);
2761                 assert(op && op->is_remat);
2762
2763                 if(!op->attr.remat.remat->inverse) {
2764                         if(op->attr.remat.pre) {
2765                                 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2766                         } else {
2767                                 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2768                         }
2769
2770                         return 1;
2771                 } else {
2772                         op_t   *op = (op_t*)get_irn_link(n);
2773                         assert(op && op->is_remat);
2774
2775                         if(op->attr.remat.pre) {
2776                                 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2777                         } else {
2778                                 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2779                         }
2780
2781                         return 1;
2782                 }
2783         }
2784
2785         return 0;
2786 }
2787
2788 static void
2789 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2790 {
2791         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2792         be_dump(irg, suffix, dump_ir_block_graph_sched);
2793         set_dump_node_vcgattr_hook(NULL);
2794 }
2795 #endif
2796
2797 /**
2798  * Edge hook to dump the schedule edges with annotated register pressure.
2799  */
2800 static int
2801 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2802 {
2803         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2804                 ir_node *prev = sched_prev(irn);
2805                 fprintf(F, "edge:{sourcename:\"");
2806                 PRINT_NODEID(irn);
2807                 fprintf(F, "\" targetname:\"");
2808                 PRINT_NODEID(prev);
2809                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2810                 fprintf(F, "\" color:magenta}\n");
2811         }
2812         return 1;
2813 }
2814
2815 static void
2816 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2817 {
2818         DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2819
2820         dump_consts_local(0);
2821         set_dump_node_edge_hook(sched_pressure_edge_hook);
2822         dump_ir_block_graph(irg, suffix);
2823         set_dump_node_edge_hook(old_edge_hook);
2824 }
2825
2826 static void
2827 walker_pressure_annotator(ir_node * bb, void * data)
2828 {
2829         spill_ilp_t  *si = data;
2830         ir_node      *irn;
2831         irn_live_t   *li;
2832         int           n;
2833         pset         *live = pset_new_ptr_default();
2834         int           projs = 0;
2835
2836         live_foreach(bb, li) {
2837                 irn = (ir_node *) li->irn;
2838
2839                 if (live_is_end(li) && has_reg_class(si, irn)) {
2840                         pset_insert_ptr(live, irn);
2841                 }
2842         }
2843
2844         set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2845
2846         sched_foreach_reverse(bb, irn) {
2847                 if(is_Phi(irn)) {
2848                         set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2849                         continue;
2850                 }
2851
2852                 if(has_reg_class(si, irn)) {
2853                         pset_remove_ptr(live, irn);
2854                         if(is_Proj(irn)) ++projs;
2855                 }
2856
2857                 if(!is_Proj(irn)) projs = 0;
2858
2859                 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2860                         ir_node    *arg = get_irn_n(irn, n);
2861
2862                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2863                 }
2864                 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2865         }
2866
2867         del_pset(live);
2868 }
2869
2870 static void
2871 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2872 {
2873         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2874 }
2875
2876 #ifdef KEEPALIVE
2877 static void
2878 connect_all_remats_with_keep(spill_ilp_t * si)
2879 {
2880         ir_node   *irn;
2881         ir_node  **ins,
2882                          **pos;
2883         int        n_remats;
2884
2885
2886         n_remats = pset_count(si->all_possible_remats);
2887         if(n_remats) {
2888                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2889
2890                 pos = ins;
2891                 pset_foreach(si->all_possible_remats, irn) {
2892                         *pos = irn;
2893                         ++pos;
2894                 }
2895
2896                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2897
2898                 obstack_free(si->obst, ins);
2899         }
2900 }
2901 #endif
2902
2903 static void
2904 connect_all_spills_with_keep(spill_ilp_t * si)
2905 {
2906         ir_node   *irn;
2907         ir_node  **ins,
2908                          **pos;
2909         int        n_spills;
2910         ir_node   *keep;
2911
2912
2913         n_spills = pset_count(si->spills);
2914         if(n_spills) {
2915                 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2916
2917                 pos = ins;
2918                 pset_foreach(si->spills, irn) {
2919                         *pos = irn;
2920                         ++pos;
2921                 }
2922
2923                 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2924
2925                 obstack_free(si->obst, ins);
2926         }
2927 }
2928
2929 /** insert a spill at an arbitrary position */
2930 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2931 {
2932         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
2933         ir_graph *irg   = get_irn_irg(bl);
2934         ir_node *frame  = get_irg_frame(irg);
2935         ir_node *spill;
2936         ir_node *next;
2937
2938         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
2939         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2940
2941         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2942
2943         /*
2944          * search the right insertion point. a spill of a phi cannot be put
2945          * directly after the phi, if there are some phis behind the one which
2946          * is spilled. Also, a spill of a Proj must be after all Projs of the
2947          * same tuple node.
2948          *
2949          * Here's one special case:
2950          * If the spill is in the start block, the spill must be after the frame
2951          * pointer is set up. This is done by setting insert to the end of the block
2952          * which is its default initialization (see above).
2953          */
2954
2955         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2956                 insert = frame;
2957
2958         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2959                 insert = next;
2960
2961         sched_add_after(insert, spill);
2962         return spill;
2963 }
2964
2965 static void
2966 delete_remat(spill_ilp_t * si, ir_node * remat) {
2967         int       n;
2968         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2969
2970         sched_remove(remat);
2971
2972         /* kill links to operands */
2973         for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2974                 set_irn_n(remat, n, bad);
2975         }
2976 }
2977
2978 static void
2979 clean_remat_info(spill_ilp_t * si)
2980 {
2981         int            n;
2982         remat_t       *remat;
2983         remat_info_t  *remat_info;
2984         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
2985
2986         set_foreach(si->remat_info, remat_info) {
2987                 if(!remat_info->remats) continue;
2988
2989                 pset_foreach(remat_info->remats, remat)
2990                 {
2991                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2992                                 set_irn_n(remat->proj, -1, bad);
2993                                 set_irn_n(remat->proj, 0, bad);
2994                         }
2995
2996                         if(get_irn_n_edges(remat->op) == 0) {
2997                                 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2998                                         set_irn_n(remat->op, n, bad);
2999                                 }
3000                         }
3001                 }
3002
3003                 if(remat_info->remats) del_pset(remat_info->remats);
3004                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3005         }
3006 }
3007
3008 static void
3009 delete_unnecessary_remats(spill_ilp_t * si)
3010 {
3011 #ifdef KEEPALIVE
3012         int       n;
3013         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
3014
3015         if(si->keep) {
3016                 ir_node   *end = get_irg_end(si->chordal_env->irg);
3017                 ir_node  **keeps;
3018
3019                 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3020                         ir_node        *keep_arg = get_irn_n(si->keep, n);
3021                         op_t           *arg_op = get_irn_link(keep_arg);
3022                         lpp_name_t     *name;
3023
3024                         assert(arg_op->is_remat);
3025
3026                         name = si->lpp->vars[arg_op->attr.remat.ilp];
3027
3028                         if(is_zero(name->value)) {
3029                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
3030                                 /* TODO check whether reload is preferred over remat (could be bug) */
3031                                 delete_remat(si, keep_arg);
3032                         } else {
3033                                 if(!arg_op->attr.remat.remat->inverse) {
3034                                         if(arg_op->attr.remat.pre) {
3035                                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3036                                         } else {
3037                                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3038                                         }
3039                                 } else {
3040                                         if(arg_op->attr.remat.pre) {
3041                                                 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3042                                         } else {
3043                                                 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3044                                         }
3045                                 }
3046                         }
3047
3048                         set_irn_n(si->keep, n, bad);
3049                 }
3050 #if 0
3051                 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3052                         ir_node        *end_arg = get_End_keepalive(end, i);
3053
3054                         if(end_arg != si->keep) {
3055                                 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3056                         }
3057                 }
3058                 keeps = obstack_finish(si->obst);
3059                 set_End_keepalives(end, n-1, keeps);
3060                 obstack_free(si->obst, keeps);
3061 #endif
3062         } else {
3063                 DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
3064         }
3065 #else
3066         ir_node  *remat;
3067
3068         pset_foreach(si->all_possible_remats, remat) {
3069                 op_t           *remat_op = get_irn_link(remat);
3070                 lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
3071
3072                 if(is_zero(name->value)) {
3073                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
3074                         /* TODO check whether reload is preferred over remat (could be bug) */
3075                         delete_remat(si, remat);
3076                 } else {
3077                         if(!remat_op->attr.remat.remat->inverse) {
3078                                 if(remat_op->attr.remat.pre) {
3079                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3080                                 } else {
3081                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3082                                 }
3083                         } else {
3084                                 if(remat_op->attr.remat.pre) {
3085                                         DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3086                                 } else {
3087                                         DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3088                                 }
3089                         }
3090                 }
3091         }
3092 #endif
3093 }
3094
3095 static pset *
3096 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3097 {
3098         pset     *spills = pset_new_ptr_default();
3099
3100         ir_node  *next;
3101         defs_t   *defs;
3102
3103         defs = set_find_def(si->values, value);
3104
3105         if(defs && defs->spills) {
3106                 for(next = defs->spills; next; next = get_irn_link(next)) {
3107                         pset_insert_ptr(spills, next);
3108                 }
3109         }
3110
3111         return spills;
3112 }
3113
3114 static pset *
3115 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3116 {
3117         pset     *remats = pset_new_ptr_default();
3118
3119         ir_node  *next;
3120         defs_t   *defs;
3121
3122         pset_insert_ptr(remats, value);
3123         defs = set_find_def(si->values, value);
3124
3125         if(defs && defs->remats) {
3126                 for(next = defs->remats; next; next = get_irn_link(next)) {
3127                         pset_insert_ptr(remats, next);
3128                 }
3129         }
3130
3131         return remats;
3132 }
3133
3134
3135 /**
3136  * @param before   The node after which the spill will be placed in the schedule
3137  */
3138 /* TODO set context properly */
3139 static ir_node *
3140 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3141 {
3142         defs_t   *defs;
3143         ir_node  *spill;
3144         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3145
3146         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
3147
3148         spill = be_spill2(arch_env, irn, before, irn);
3149
3150         defs = set_insert_def(si->values, value);
3151         assert(defs);
3152
3153         /* enter into the linked list */
3154         set_irn_link(spill, defs->spills);
3155         defs->spills = spill;
3156
3157 #ifdef KEEPALIVE_SPILLS
3158         pset_insert_ptr(si->spills, spill);
3159 #endif
3160
3161         return spill;
3162 }
3163
3164 /**
3165  * @param before   The Phi node which has to be spilled
3166  */
3167 static ir_node *
3168 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3169 {
3170         ir_node   *mem_phi;
3171         ir_node  **ins;
3172         defs_t    *defs;
3173         int        n;
3174         op_t      *op = get_irn_link(phi);
3175
3176         NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3177
3178         for(n=get_irn_arity(phi)-1; n>=0; --n) {
3179                 ins[n] = si->m_unknown;
3180         }
3181
3182         mem_phi =  new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3183
3184         defs = set_insert_def(si->values, phi);
3185         assert(defs);
3186
3187         /* enter into the linked list */
3188         set_irn_link(mem_phi, defs->spills);
3189         defs->spills = mem_phi;
3190
3191         sched_add_after(phi, mem_phi);
3192
3193 #ifdef KEEPALIVE_SPILLS
3194         pset_insert_ptr(si->spills, mem_phi);
3195 #endif
3196
3197
3198         return mem_phi;
3199 }
3200
3201 /**
3202  * Add remat to list of defs, destroys link field!
3203  */
3204 static void
3205 insert_remat(spill_ilp_t * si, ir_node * remat)
3206 {
3207         defs_t   *defs;
3208         op_t     *remat_op = get_irn_link(remat);
3209
3210         assert(remat_op->is_remat);
3211
3212         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3213         assert(defs);
3214
3215         /* enter into the linked list */
3216         set_irn_link(remat, defs->remats);
3217         defs->remats = remat;
3218 }
3219
3220
3221 /**
3222  * Add reload before operation and add to list of defs
3223  */
3224 static ir_node *
3225 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3226 {
3227         defs_t   *defs;
3228         ir_node  *reload,
3229                          *spill;
3230         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3231
3232         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
3233
3234         defs = set_find_def(si->values, value);
3235
3236         spill = defs->spills;
3237         assert(spill && "no spill placed before reload");
3238
3239         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3240
3241         /* enter into the linked list */
3242         set_irn_link(reload, defs->remats);
3243         defs->remats = reload;
3244
3245         return reload;
3246 }
3247
3248 static void
3249 walker_spill_placer(ir_node * bb, void * data) {
3250         spill_ilp_t   *si = (spill_ilp_t*)data;
3251         ir_node       *irn;
3252         spill_bb_t    *spill_bb = get_irn_link(bb);
3253         pset          *spills_to_do = pset_new_ptr_default();
3254         spill_t       *spill;
3255
3256         set_foreach(spill_bb->ilp, spill) {
3257                 lpp_name_t    *name;
3258
3259                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3260                         name = si->lpp->vars[spill->mem_in];
3261                         if(!is_zero(name->value)) {
3262                                 ir_node   *mem_phi;
3263
3264                                 mem_phi = insert_mem_phi(si, spill->irn);
3265
3266                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3267                         }
3268                 }
3269
3270                 name = si->lpp->vars[spill->spill];
3271                 if(!is_zero(name->value)) {
3272                         /* place spill directly after definition */
3273                         if(get_nodes_block(spill->irn) == bb) {
3274                                 insert_spill(si, spill->irn, spill->irn, spill->irn);
3275                                 continue;
3276                         }
3277
3278                         /* place spill at bb start */
3279                         if(spill->reg_in > 0) {
3280                                 name = si->lpp->vars[spill->reg_in];
3281                                 if(!is_zero(name->value)) {
3282                                         insert_spill(si, spill->irn, spill->irn, bb);
3283                                         continue;
3284                                 }
3285                         }
3286                         /* place spill after a remat */
3287                         pset_insert_ptr(spills_to_do, spill->irn);
3288                 }
3289         }
3290         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3291
3292
3293         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3294                 op_t     *op = get_irn_link(irn);
3295
3296                 if(be_is_Spill(irn)) continue;
3297
3298                 if(op->is_remat) {
3299                         /* TODO fix this if we want to support remats with more than two nodes */
3300                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3301                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3302
3303                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3304                         }
3305                 } else {
3306                         if(pset_find_ptr(spills_to_do, irn)) {
3307                                 pset_remove_ptr(spills_to_do, irn);
3308
3309                                 insert_spill(si, irn, irn, irn);
3310                         }
3311                 }
3312
3313         }
3314
3315         assert(pset_count(spills_to_do) == 0);
3316
3317         /* afterwards free data in block */
3318         del_pset(spills_to_do);
3319 }
3320
3321 static ir_node *
3322 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3323 {
3324         ir_node          *insert_pos = bb;
3325         ir_node          *spill;
3326         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3327
3328         /* find last definition of arg value in block */
3329         ir_node  *next;
3330         defs_t   *defs;
3331         int       last = 0;
3332
3333         defs = set_find_def(si->values, value);
3334
3335         if(defs && defs->remats) {
3336                 for(next = defs->remats; next; next = get_irn_link(next)) {
3337                         if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3338                                 last = sched_get_time_step(next);
3339                                 insert_pos = next;
3340                         }
3341                 }
3342         }
3343
3344         if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3345                 last = sched_get_time_step(value);
3346                 insert_pos = value;
3347         }
3348
3349         DBG((si->dbg, LEVEL_2, "\t  inserting mem copy for value %+F after %+F\n", value, insert_pos));
3350
3351         spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3352
3353         return spill;
3354 }
3355
3356 static void
3357 phim_fixer(spill_ilp_t *si) {
3358         defs_t  *defs;
3359
3360         set_foreach(si->values, defs) {
3361                 const ir_node  *phi = defs->value;
3362                 op_t           *op = get_irn_link(phi);
3363                 ir_node        *phi_m = NULL;
3364                 ir_node        *next = defs->spills;
3365                 int             n;
3366
3367                 if(!is_Phi(phi)) continue;
3368
3369                 while(next) {
3370                         if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3371                                 phi_m = next;
3372                                 break;
3373                         } else {
3374                                 next = get_irn_link(next);
3375                         }
3376                 }
3377                 if(!phi_m) continue;
3378
3379                 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3380                         const ir_node  *value = get_irn_n(phi, n);
3381                         defs_t         *val_defs = set_find_def(si->values, value);
3382                         ir_node        *arg = get_irn_n(phi_m, n);
3383
3384                         /* a spill of this value */
3385                         ir_node      *spill;
3386
3387
3388 #ifndef NO_MEMCOPIES
3389                         ir_node    *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3390                         lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3391
3392                         if(!is_zero(name->value)) {
3393                                 spill = insert_mem_copy(si, pred, value);
3394                         } else {
3395                                 spill = val_defs->spills;
3396                         }
3397 #else
3398                         spill = val_defs->spills;
3399 #endif
3400                         assert(spill && "no spill placed before PhiM");
3401                         set_irn_n(phi_m, n, spill);
3402                 }
3403         }
3404 }
3405
3406 static void
3407 walker_reload_placer(ir_node * bb, void * data) {
3408         spill_ilp_t   *si = (spill_ilp_t*)data;
3409         ir_node       *irn;
3410         spill_bb_t    *spill_bb = get_irn_link(bb);
3411         int            i;
3412         irn_live_t    *li;
3413
3414         /* reloads at end of block */
3415         if(spill_bb->reloads) {
3416                 keyval_t    *keyval;
3417
3418                 set_foreach(spill_bb->reloads, keyval) {
3419                         ir_node        *irn = (ir_node*)keyval->key;
3420                         ilp_var_t       reload = PTR_TO_INT(keyval->val);
3421                         lpp_name_t     *name;
3422
3423                         name = si->lpp->vars[reload];
3424                         if(!is_zero(name->value)) {
3425                                 ir_node    *reload;
3426                                 ir_node    *insert_pos = bb;
3427                                 ir_node    *prev = sched_block_last_noncf(si, bb);
3428                                 op_t       *prev_op = get_irn_link(prev);
3429
3430                                 while(be_is_Spill(prev)) {
3431                                         prev = sched_prev(prev);
3432                                 }
3433
3434                                 prev_op = get_irn_link(prev);
3435
3436                                 /* insert reload before pre-remats */
3437                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3438                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3439                                         insert_pos = prev;
3440
3441                                         do {
3442                                                 prev = sched_prev(prev);
3443                                         } while(be_is_Spill(prev));
3444
3445                                         prev_op = get_irn_link(prev);
3446
3447                                 }
3448
3449                                 reload = insert_reload(si, irn, insert_pos);
3450
3451 #ifdef KEEPALIVE_RELOADS
3452                                 pset_insert_ptr(si->spills, reload);
3453 #endif
3454                         }
3455                 }
3456         }
3457
3458         /* walk and insert more reloads and collect remats */
3459         sched_foreach_reverse(bb, irn) {
3460                 op_t     *op = get_irn_link(irn);
3461
3462                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3463                 if(is_Phi(irn)) break;
3464
3465                 if(op->is_remat) {
3466                         if(get_irn_mode(irn) != mode_T) {
3467                                 insert_remat(si, irn);
3468                         }
3469                 } else {
3470                         int    n;
3471
3472                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3473                                 ir_node    *arg = get_irn_n(irn, n);
3474
3475                                 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3476                                         lpp_name_t    *name;
3477
3478                                         name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3479                                         if(!is_zero(name->value)) {
3480                                                 ir_node    *reload;
3481                                                 ir_node    *insert_pos = irn;
3482                                                 ir_node    *prev = sched_prev(insert_pos);
3483                                                 op_t       *prev_op;
3484
3485                                                 while(be_is_Spill(prev)) {
3486                                                         prev = sched_prev(prev);
3487                                                 }
3488
3489                                                 prev_op = get_irn_link(prev);
3490
3491                                                 /* insert reload before pre-remats */
3492                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3493                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
3494                                                         insert_pos = prev;
3495
3496                                                         do {
3497                                                                 prev = sched_prev(prev);
3498                                                         } while(be_is_Spill(prev));
3499
3500                                                         prev_op = get_irn_link(prev);
3501
3502                                                 }
3503
3504                                                 reload = insert_reload(si, arg, insert_pos);
3505
3506                                                 set_irn_n(irn, n, reload);
3507
3508 #ifdef KEEPALIVE_RELOADS
3509                                                 pset_insert_ptr(si->spills, reload);
3510 #endif
3511                                         }
3512                                 }
3513                         }
3514                 }
3515         }
3516
3517         del_set(spill_bb->ilp);
3518         if(spill_bb->reloads) del_set(spill_bb->reloads);
3519 }
3520
3521 static void
3522 walker_collect_used(ir_node * irn, void * data)
3523 {
3524         lc_bitset_t   *used = data;
3525
3526         lc_bitset_set(used, get_irn_idx(irn));
3527 }
3528
3529 struct kill_helper {
3530         lc_bitset_t  *used;
3531         spill_ilp_t  *si;
3532 };
3533
3534 static void
3535 walker_kill_unused(ir_node * bb, void * data)
3536 {
3537         struct kill_helper *kh = data;
3538         const ir_node      *bad = get_irg_bad(get_irn_irg(bb));
3539         ir_node            *irn;
3540
3541
3542         for(irn=sched_first(bb); !sched_is_end(irn);) {
3543                 ir_node     *next = sched_next(irn);
3544                 int          n;
3545
3546                 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3547                         if(be_is_Spill(irn) || be_is_Reload(irn)) {
3548                                 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)));
3549 #if 0
3550                                 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3551 #endif
3552                         }
3553
3554                         sched_remove(irn);
3555
3556                         set_nodes_block(irn, bad);
3557                         for (n=get_irn_arity(irn)-1; n>=0; --n) {
3558                                 set_irn_n(irn, n, bad);
3559                         }
3560                 }
3561                 irn = next;
3562         }
3563 }
3564
3565 static void
3566 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3567 {
3568         struct kill_helper kh;
3569
3570         kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3571         kh.si = si;
3572
3573         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3574         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3575
3576         lc_bitset_free(kh.used);
3577 }
3578
3579 static void
3580 print_irn_pset(pset * p)
3581 {
3582         ir_node   *irn;
3583
3584         pset_foreach(p, irn) {
3585                 ir_printf("%+F\n", irn);
3586         }
3587 }
3588
3589 static void
3590 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3591 {
3592     FILE           *f = fopen(file, "w");
3593     ir_node        *irn;
3594     interference_t *interference;
3595
3596     pset_break(phiclass);
3597     set_break(si->interferences);
3598
3599     ir_fprintf(f, "digraph phiclass {\n");
3600
3601     pset_foreach(phiclass, irn) {
3602         if(is_Phi(irn))
3603             ir_fprintf(f, "  %F%N [shape=box]\n",irn,irn);
3604     }
3605
3606     pset_foreach(phiclass, irn) {
3607         int n;
3608
3609         if(!is_Phi(irn)) continue;
3610
3611         for(n=get_irn_arity(irn)-1; n>=0; --n) {
3612             ir_node  *arg = get_irn_n(irn, n);
3613
3614             ir_fprintf(f, "  %F%N -> %F%N\n",irn,irn,arg,arg);
3615         }
3616     }
3617
3618     set_foreach(si->interferences, interference) {
3619         const ir_node  *a  = interference->a;
3620         const ir_node  *b  = interference->b;
3621         if(get_phi_class(a) == phiclass) {
3622             ir_fprintf(f, "  %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3623         }
3624     }
3625
3626     ir_fprintf(f, "}");
3627     fclose(f);
3628 }
3629
3630 static void
3631 rewire_uses(spill_ilp_t * si)
3632 {
3633         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3634         defs_t               *defs;
3635         pset                 *ignore = pset_new_ptr(1);
3636
3637         pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3638
3639         /* then fix uses of spills */
3640         set_foreach(si->values, defs) {
3641                 pset     *reloads;
3642                 pset     *spills;
3643                 ir_node  *next = defs->remats;
3644                 int remats = 0;
3645
3646                 reloads = pset_new_ptr_default();
3647
3648                 while(next) {
3649                         if(be_is_Reload(next)) {
3650                                 pset_insert_ptr(reloads, next);
3651                         } else {
3652                                 ++remats;
3653                         }
3654                         next = get_irn_link(next);
3655                 }
3656
3657                 spills = get_spills_for_value(si, defs->value);
3658                 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));
3659                 if(pset_count(spills) > 1) {
3660                         //assert(pset_count(reloads) > 0);
3661                         //                              print_irn_pset(spills);
3662                         //                              print_irn_pset(reloads);
3663
3664                         be_ssa_constr_set_ignore(dfi, spills, ignore);
3665                 }
3666
3667                 del_pset(reloads);
3668                 del_pset(spills);
3669         }
3670
3671         /* first fix uses of remats and reloads */
3672         set_foreach(si->values, defs) {
3673                 pset     *nodes;
3674                 ir_node  *next = defs->remats;
3675
3676                 if(next) {
3677                         nodes = pset_new_ptr_default();
3678                         pset_insert_ptr(nodes, defs->value);
3679
3680                         while(next) {
3681                                 pset_insert_ptr(nodes, next);
3682                                 next = get_irn_link(next);
3683                         }
3684
3685                         if(pset_count(nodes) > 1) {
3686                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3687                                 be_ssa_constr_set(dfi, nodes);
3688                         }
3689
3690                         del_pset(nodes);
3691                 }
3692         }
3693
3694 //      remove_unused_defs(si);
3695
3696         be_free_dominance_frontiers(dfi);
3697 }
3698
3699
3700 static void
3701 writeback_results(spill_ilp_t * si)
3702 {
3703         /* walk through the graph and collect all spills, reloads and remats for a value */
3704
3705         si->values = new_set(cmp_defs, 4096);
3706
3707         DBG((si->dbg, LEVEL_1, "Applying results\n"));
3708         delete_unnecessary_remats(si);
3709         si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3710         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3711         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3712         phim_fixer(si);
3713
3714         /* clean the remat info! there are still back-edges leading there! */
3715         clean_remat_info(si);
3716
3717         rewire_uses(si);
3718
3719         connect_all_spills_with_keep(si);
3720
3721         del_set(si->values);
3722 }
3723
3724 static int
3725 get_n_regs(spill_ilp_t * si)
3726 {
3727         int     arch_n_regs = arch_register_class_n_regs(si->cls);
3728         int     free = 0;
3729         int     i;
3730
3731         for(i=0; i<arch_n_regs; i++) {
3732                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3733                         free++;
3734                 }
3735         }
3736
3737         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3738         return free;
3739 }
3740
3741 static void
3742 walker_reload_mover(ir_node * bb, void * data)
3743 {
3744         spill_ilp_t   *si = data;
3745         ir_node           *tmp;
3746
3747         sched_foreach(bb, tmp) {
3748                 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3749                         ir_node       *reload = tmp;
3750                         ir_node       *irn = tmp;
3751
3752                         /* move reload upwards */
3753
3754                         int pressure = (int)get_irn_link(reload);
3755                         if(pressure < si->n_regs) {
3756                                 irn = sched_prev(reload);
3757                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3758                                 sched_remove(reload);
3759                                 pressure = (int)get_irn_link(irn);
3760
3761                                 while(pressure < si->n_regs) {
3762                                         if( sched_is_end(irn) ||
3763                                            (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3764                                            /* do not move reload before its spill */
3765                                            (irn == be_get_Reload_mem(reload)) ) break;
3766
3767                                         set_irn_link(irn, INT_TO_PTR(pressure+1));
3768                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3769                                         irn = sched_prev(irn);
3770
3771                                         pressure = (int)get_irn_link(irn);
3772                                 }
3773
3774                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3775                                 sched_put_after(irn, reload);
3776                         }
3777                 }
3778         }
3779 }
3780
3781 static void
3782 move_reloads_upward(spill_ilp_t * si)
3783 {
3784         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3785 }
3786
3787
3788 /**
3789  * Walk all irg blocks and check for interfering spills inside of phi classes
3790  */
3791 static void
3792 luke_meminterferencechecker(ir_node * bb, void * data)
3793 {
3794         spill_ilp_t    *si = (spill_ilp_t*)data;
3795         irn_live_t     *li1,
3796                        *li2;
3797
3798         live_foreach(bb, li1) {
3799                 ir_node        *a = (ir_node *) li1->irn;
3800
3801                 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3802
3803                 /* a is only interesting if it is inside a phi class */
3804                 if (get_phi_class(a)) {
3805                         for(li2=li1->next; li2; li2 = li2->next) {
3806                                 ir_node        *b = (ir_node *) li2->irn;
3807
3808                                 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3809
3810                                 /* a and b are only interesting if they are in the same phi class */
3811                                 if(get_phi_class(a) == get_phi_class(b)) {
3812                                         if(values_interfere_in_block(bb, a, b)) {
3813                                                 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3814                                         }
3815                                 }
3816                         }
3817                 }
3818         }
3819 }
3820
3821 static void
3822 verify_phiclasses(spill_ilp_t * si)
3823 {
3824         /* analyze phi classes */
3825         phi_class_compute(si->chordal_env->irg);
3826
3827         DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3828         irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3829 }
3830
3831 static void
3832 walker_spillslotassigner(ir_node * irn, void * data)
3833 {
3834         spill_ilp_t            *si = (spill_ilp_t*)data;
3835         void                   *cls;
3836
3837         if(!be_is_Spill(irn)) return;
3838
3839         /* set spill context to phi class if it has one ;) */
3840
3841         cls = get_phi_class(irn);
3842         if(cls)
3843                 be_set_Spill_context(irn, cls);
3844         else
3845                 be_set_Spill_context(irn, irn);
3846 }
3847
3848
3849 static void
3850 assign_spillslots(spill_ilp_t * si)
3851 {
3852         DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3853         irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3854 }
3855
3856 void
3857 be_spill_remat(const be_chordal_env_t * chordal_env)
3858 {
3859         char            problem_name[256];
3860         char            dump_suffix[256];
3861         char            dump_suffix2[256];
3862         struct obstack  obst;
3863         spill_ilp_t     si;
3864
3865         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3866         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3867         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3868
3869         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3870         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3871
3872 #ifdef VERIFY_DOMINANCE
3873     be_check_dominance(chordal_env->irg);
3874 #endif
3875
3876         obstack_init(&obst);
3877         si.chordal_env = chordal_env;
3878         si.obst = &obst;
3879         si.cls = chordal_env->cls;
3880         si.lpp = new_lpp(problem_name, lpp_minimize);
3881         si.remat_info = new_set(cmp_remat_info, 4096);
3882         si.interferences = new_set(cmp_interference, 32);
3883         si.all_possible_remats = pset_new_ptr_default();
3884         si.spills = pset_new_ptr_default();
3885         si.inverse_ops = pset_new_ptr_default();
3886 #ifdef KEEPALIVE
3887         si.keep = NULL;
3888 #endif
3889         si.n_regs = get_n_regs(&si);
3890
3891         set_irg_link(chordal_env->irg, &si);
3892         compute_doms(chordal_env->irg);
3893
3894         /* compute phi classes */
3895 //      phi_class_compute(chordal_env->irg);
3896
3897         be_analyze_regpressure(chordal_env, "-pre");
3898
3899 #ifdef COLLECT_REMATS
3900         /* collect remats */
3901         DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3902         irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3903 #endif
3904
3905         /* insert possible remats */
3906         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3907         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3908         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3909
3910 #ifdef KEEPALIVE
3911         DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3912         connect_all_remats_with_keep(&si);
3913         /* dump graph with inserted remats */
3914         dump_graph_with_remats(chordal_env->irg, dump_suffix);
3915 #endif
3916
3917         /* insert copies for phi arguments not in my regclass */
3918         irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3919
3920         /* recompute liveness */
3921         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3922         be_liveness(chordal_env->irg);
3923
3924         /* build the ILP */
3925
3926         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3927         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3928         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3929
3930         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3931         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3932
3933 #ifndef NO_MEMCOPIES
3934         DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3935         memcopyhandler(&si);
3936 #endif
3937
3938 #ifdef DUMP_ILP
3939         {
3940                 FILE           *f;
3941                 char            buf[256];
3942
3943                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3944                 if ((f = fopen(buf, "wt")) != NULL) {
3945                         lpp_dump_plain(si.lpp, f);
3946                         fclose(f);
3947                 }
3948         }
3949 #endif
3950
3951 #ifdef SOLVE
3952         DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3953 #ifdef ILP_TIMEOUT
3954         lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3955 #endif
3956
3957 #ifdef SOLVE_LOCAL
3958         lpp_solve_cplex(si.lpp);
3959 #else
3960         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3961 #endif
3962         assert(lpp_is_sol_valid(si.lpp)
3963                && "solution of ILP must be valid");
3964
3965         DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
3966
3967 #ifdef DUMP_SOLUTION
3968         {
3969                 FILE           *f;
3970                 char            buf[256];
3971
3972                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3973                 if ((f = fopen(buf, "wt")) != NULL) {
3974                         int             i;
3975                         for (i = 0; i < si.lpp->var_next; ++i) {
3976                                 lpp_name_t     *name = si.lpp->vars[i];
3977                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3978                         }
3979                         fclose(f);
3980                 }
3981         }
3982 #endif
3983
3984         writeback_results(&si);
3985
3986 #endif                          /* SOLVE */
3987
3988         kill_all_unused_values_in_schedule(&si);
3989
3990 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3991         be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3992 #endif
3993
3994         // move reloads upwards
3995         be_liveness(chordal_env->irg);
3996         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3997         move_reloads_upward(&si);
3998
3999 #ifndef NO_MEMCOPIES
4000         verify_phiclasses(&si);
4001         assign_spillslots(&si);
4002 #endif
4003
4004         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4005
4006         dump_pressure_graph(&si, dump_suffix2);
4007
4008         be_analyze_regpressure(chordal_env, "-post");
4009
4010 #ifdef VERIFY_DOMINANCE
4011         be_check_dominance(chordal_env->irg);
4012 #endif
4013
4014         free_dom(chordal_env->irg);
4015         del_set(si.interferences);
4016         del_pset(si.inverse_ops);
4017         del_pset(si.all_possible_remats);
4018         del_pset(si.spills);
4019         free_lpp(si.lpp);
4020         obstack_free(&obst, NULL);
4021         DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4022 }
4023
4024 #else                           /* WITH_ILP */
4025
4026 static void
4027 only_that_you_can_compile_without_WITH_ILP_defined(void)
4028 {
4029 }
4030
4031 #endif                          /* WITH_ILP */