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 "execfreq.h"
  33 #include "phiclass.h"
  34
  35 #include <lpp/lpp.h>
  36 #include <lpp/lpp_net.h>
  37 #include <lpp/lpp_cplex.h>
  38 //#include <lc_pset.h>
  39 #include <libcore/lc_bitset.h>
  40
  41 #include "be_t.h"
  42 #include "belive_t.h"
  43 #include "besched_t.h"
  44 #include "beirgmod.h"
  45 #include "bearch.h"
  46 #include "benode_t.h"
  47 #include "beutil.h"
  48 #include "bespillremat.h"
  49 #include "bespill.h"
  50
  51 #include "bechordal_t.h"
  52
  53 #define BIGM 100000.0
  54
  55 #define DUMP_SOLUTION
  56 #define DUMP_ILP
  57 #define KEEPALIVE
  58 #define COLLECT_REMATS
  59 #define REMAT_WHILE_LIVE
  60 #define NO_ENLARGE_L1V3N355
  61 #define EXECFREQ_LOOPDEPH
  62 #define MAY_DIE_AT_PRE_REMAT
  63
  64 #define  SOLVE
  65 #undef  SOLVE_LOCAL
  66 #define LPP_SERVER "i44pc52"
  67 #define LPP_SOLVER "cplex"
  68
  69 #define COST_LOAD      10
  70 #define COST_STORE     50
  71 #define COST_REMAT     1
  72
  73 #define ILP_TIMEOUT    20
  74
  75 #define ILP_UNDEF               -1
  76
  77 typedef struct _spill_ilp_t {
  78         const arch_register_class_t  *cls;
  79         int                           n_regs;
  80         const be_chordal_env_t       *chordal_env;
  81         spill_env_t                  *senv;
  82         lpp_t                        *lpp;
  83         struct obstack               *obst;
  84         set                          *remat_info;
  85         pset                         *all_possible_remats;
  86 #ifdef KEEPALIVE
  87         ir_node                      *keep;
  88 #endif
  89         set                          *values; /**< for collecting all definitions of values before running ssa-construction */
  90 #ifndef EXECFREQ_LOOPDEPH
  91         set                          *execfreqs;
  92 #endif
  93         DEBUG_ONLY(firm_dbg_module_t * dbg);
  94 } spill_ilp_t;
  95
  96 typedef int ilp_var_t;
  97 typedef int ilp_cst_t;
  98
  99 typedef struct _spill_bb_t {
 100         set          *ilp;
 101         pset         *copys_needed;
 102         ilp_var_t    *reloads;
 103 } spill_bb_t;
 104
 105 typedef struct _remat_t {
 106         const ir_node        *op; /**< for copy_irn */
 107         const ir_node        *proj; /**< not NULL if the above op produces a tuple */
 108         const ir_node        *value; /**< the value which is being recomputed by this remat */
 109         int                   cost; /**< cost of this remat */
 110 } remat_t;
 111
 112 /**
 113  * Data to be attached to each IR node. For remats this contains the ilp_var
 114  * for this remat and for normal ops this contains the ilp_vars for
 115  * reloading each operand
 116  */
 117 typedef struct _op_t {
 118         int             is_remat;
 119         union {
 120                 struct {
 121                         ilp_var_t       ilp;
 122                         remat_t        *remat; /** the remat this op belongs to */
 123                         int             pre; /** 1, if this is a pressure-increasing remat */
 124                 } remat;
 125                 struct {
 126                         ilp_var_t       ilp;
 127                         ir_node        *op; /** the operation this live range belongs to */
 128                         ilp_var_t      *reloads;
 129                 } live_range;
 130         } attr;
 131 } op_t;
 132
 133 typedef struct _defs_t {
 134         ir_node   *value;
 135         ir_node   *spills;  /**< points to the first spill for this value (linked by link field) */
 136         ir_node   *remats;  /**< points to the first definition for this value (linked by link field) */
 137 } defs_t;
 138
 139 typedef struct _remat_info_t {
 140         const ir_node       *irn; /**< the irn to which these remats belong */
 141         pset                *remats; /**< possible remats for this value */
 142         pset                *remats_by_operand; /**< remats with this value as operand */
 143 } remat_info_t;
 144
 145 typedef struct _keyval_t {
 146         const void          *key;
 147         const void          *val;
 148 } keyval_t;
 149
 150 typedef struct _spill_t {
 151         ir_node      *irn;
 152         ilp_var_t     reg_in;
 153         ilp_var_t     mem_in;
 154         ilp_var_t     reg_out;
 155         ilp_var_t     mem_out;
 156         ilp_var_t     spill;
 157 } spill_t;
 158
 159 static INLINE int
 160 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
 161 {
 162         return chordal_has_class(si->chordal_env, irn);
 163 }
 164
 165 #if 0
 166 static int
 167 cmp_remat(const void *a, const void *b)
 168 {
 169         const keyval_t *p = a;
 170         const keyval_t *q = b;
 171         const remat_t  *r = p->val;
 172         const remat_t  *s = q->val;
 173
 174         assert(r && s);
 175
 176         return !(r == s || r->op == s->op);
 177 }
 178 #endif
 179 static int
 180 cmp_remat(const void *a, const void *b)
 181 {
 182         const remat_t  *r = a;
 183         const remat_t  *s = a;
 184
 185         return !(r == s || r->op == s->op);
 186 }
 187
 188 static int
 189 cmp_spill(const void *a, const void *b, size_t size)
 190 {
 191         const spill_t *p = a;
 192         const spill_t *q = b;
 193
 194 //      return !(p->irn == q->irn && p->bb == q->bb);
 195         return !(p->irn == q->irn);
 196 }
 197
 198 static keyval_t *
 199 set_find_keyval(set * set, void * key)
 200 {
 201         keyval_t     query;
 202
 203         query.key = key;
 204         return set_find(set, &query, sizeof(query), HASH_PTR(key));
 205 }
 206
 207 static keyval_t *
 208 set_insert_keyval(set * set, void * key, void * val)
 209 {
 210         keyval_t     query;
 211
 212         query.key = key;
 213         query.val = val;
 214         return set_insert(set, &query, sizeof(query), HASH_PTR(key));
 215 }
 216
 217 static defs_t *
 218 set_find_def(set * set, ir_node * value)
 219 {
 220         defs_t     query;
 221
 222         query.value = value;
 223         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 224 }
 225
 226 static defs_t *
 227 set_insert_def(set * set, ir_node * value)
 228 {
 229         defs_t     query;
 230
 231         query.value = value;
 232         query.spills = NULL;
 233         query.remats = NULL;
 234         return set_insert(set, &query, sizeof(query), HASH_PTR(value));
 235 }
 236
 237 static spill_t *
 238 set_find_spill(set * set, ir_node * value)
 239 {
 240         spill_t     query;
 241
 242         query.irn = value;
 243         return set_find(set, &query, sizeof(query), HASH_PTR(value));
 244 }
 245
 246 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
 247 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
 248 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
 249 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
 250 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
 251
 252 static int
 253 cmp_remat_info(const void *a, const void *b, size_t size)
 254 {
 255         const remat_info_t *p = a;
 256         const remat_info_t *q = b;
 257
 258         return !(p->irn == q->irn);
 259 }
 260
 261 static int
 262 cmp_defs(const void *a, const void *b, size_t size)
 263 {
 264         const defs_t *p = a;
 265         const defs_t *q = b;
 266
 267         return !(p->value == q->value);
 268 }
 269
 270 static int
 271 cmp_keyval(const void *a, const void *b, size_t size)
 272 {
 273         const keyval_t *p = a;
 274         const keyval_t *q = b;
 275
 276         return !(p->key == q->key);
 277 }
 278
 279 static float
 280 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
 281 {
 282 #ifdef EXECFREQ_LOOPDEPH
 283         if(is_Block(irn))
 284                 return expf((float)get_loop_depth(get_irn_loop(irn)) * logf(10));
 285         else
 286                 return expf((float)get_loop_depth(get_irn_loop(get_nodes_block(irn))) * logf(10));
 287 #else
 288         if(is_Block(irn)) {
 289                 return get_block_execfreq(si->execfreqs, irn);
 290         } else {
 291                 return get_block_execfreq(si->execfreqs, get_nodes_block(irn));
 292         }
 293 #endif
 294 }
 295
 296 /**
 297  * Checks, whether node and its operands have suitable reg classes
 298  */
 299 static INLINE int
 300 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
 301 {
 302         int             i,
 303                         n;
 304         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
 305         int               remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
 306
 307 #if 0
 308         if(!remat)
 309                 ir_fprintf(stderr, "  Node %+F is not rematerializable\n", irn);
 310 #endif
 311
 312         for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
 313                 ir_node        *op = get_irn_n(irn, i);
 314                 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
 315
 316 //              if(!remat)
 317 //                      ir_fprintf(stderr, "  Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
 318         }
 319
 320         return remat;
 321 }
 322
 323 /**
 324  * Try to create a remat from @p op with destination value @p dest_value
 325  */
 326 static INLINE remat_t *
 327 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
 328 {
 329         remat_t  *remat = NULL;
 330
 331 //      if(!mode_is_datab(get_irn_mode(dest_value)))
 332 //              return NULL;
 333
 334         if(dest_value == op) {
 335                 const ir_node *proj = NULL;
 336
 337                 if(is_Proj(dest_value)) {
 338                         op = get_irn_n(op, 0);
 339                         proj = dest_value;
 340                 }
 341
 342                 if(!is_rematerializable(si, op))
 343                         return NULL;
 344
 345                 remat = obstack_alloc(si->obst, sizeof(*remat));
 346                 remat->op = op;
 347                 remat->cost = COST_REMAT; /* TODO ask backend for real cost */
 348                 remat->value = dest_value;
 349                 remat->proj = proj;
 350         } else {
 351                 arch_inverse_t     inverse;
 352                 int                i,
 353                                                    n;
 354
 355                 /* get the index of the operand we want to retrieve by the inverse op */
 356                 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
 357                         ir_node        *arg = get_irn_n(op, i);
 358
 359                         if(arg == dest_value) break;
 360                 }
 361                 if(i == n) return NULL;
 362
 363                 DBG((si->dbg, LEVEL_5, "\t  requesting inverse op for argument %d of op %+F\n", i, op));
 364
 365                 /* else ask the backend to give an inverse op */
 366                 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
 367                         DBG((si->dbg, LEVEL_4, "\t  backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
 368
 369                         assert(inverse.n > 0 && "inverse op should have at least one node");
 370
 371                         if(inverse.n <= 2) {
 372                                 remat = obstack_alloc(si->obst, sizeof(*remat));
 373                                 remat->op = inverse.nodes[0];
 374                                 remat->cost = inverse.costs;
 375                                 remat->value = dest_value;
 376                                 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
 377
 378                                 assert(is_Proj(remat->proj));
 379                         } else {
 380                                 DBG((si->dbg, LEVEL_1, "\t  I can not handle remats with %d nodes\n", inverse.n));
 381                         }
 382                 }
 383         }
 384
 385         if(remat) {
 386                 if(remat->proj) {
 387                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
 388                 } else {
 389                         DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
 390                 }
 391         }
 392         return remat;
 393 }
 394
 395
 396 static INLINE void
 397 add_remat(const spill_ilp_t * si, const remat_t * remat)
 398 {
 399         remat_info_t    *remat_info,
 400                      query;
 401         int              i,
 402                                          n;
 403
 404         assert(remat->op);
 405         assert(remat->value);
 406
 407         query.irn = remat->value;
 408         query.remats = NULL;
 409         query.remats_by_operand = NULL;
 410         remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
 411
 412         if(remat_info->remats == NULL) {
 413                 remat_info->remats = new_pset(cmp_remat, 4096);
 414         }
 415         pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
 416
 417         /* insert the remat into the remats_be_operand set of each argument of the remat op */
 418         for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
 419                 ir_node        *arg = get_irn_n(remat->op, i);
 420
 421                 query.irn = arg;
 422                 query.remats = NULL;
 423                 query.remats_by_operand = NULL;
 424                 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 425
 426                 if(remat_info->remats_by_operand == NULL) {
 427                         remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
 428                 }
 429                 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
 430         }
 431 }
 432
 433 static INLINE void
 434 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
 435 {
 436         int       i,
 437                       n;
 438         remat_t *remat;
 439
 440         if(has_reg_class(si, op)) {
 441                 remat = get_remat_from_op(si, op, op);
 442                 if(remat) {
 443                         add_remat(si, remat);
 444                 }
 445         }
 446
 447         /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
 448            for each arg */
 449         for (i = 0, n = get_irn_arity(op); i < n; ++i) {
 450                 ir_node        *arg = get_irn_n(op, i);
 451
 452                 if(has_reg_class(si, arg)) {
 453                         /* try to get an inverse remat */
 454                         remat = get_remat_from_op(si, arg, op);
 455                         if(remat) {
 456                                 add_remat(si, remat);
 457                         }
 458                 }
 459         }
 460
 461 }
 462
 463 static INLINE int
 464 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
 465 {
 466         ir_node *block;
 467         ir_node *def_block = get_nodes_block(val);
 468         int      ret;
 469
 470         if(val == pos)
 471                 return 0;
 472
 473         /* if pos is at end of a basic block */
 474         if(is_Block(pos)) {
 475                 ret = (pos == def_block || block_dominates(def_block, pos));
 476 //              ir_fprintf(stderr, "(def(bb)=%d) ", ret);
 477                 return ret;
 478         }
 479
 480         /* else if this is a normal operation */
 481         block = get_nodes_block(pos);
 482         if(block == def_block) {
 483                 ret = sched_comes_after(val, pos);
 484 //              ir_fprintf(stderr, "(def(same block)=%d) ",ret);
 485                 return ret;
 486         }
 487
 488         ret = block_dominates(def_block, block);
 489 //      ir_fprintf(stderr, "(def(other block)=%d) ", ret);
 490         return ret;
 491 }
 492
 493 static INLINE ir_node *
 494 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
 495 {
 496     return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
 497 }
 498
 499 /**
 500  * Returns first non-Phi node of block @p bb
 501  */
 502 static INLINE ir_node *
 503 sched_block_first_nonphi(const ir_node * bb)
 504 {
 505         return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
 506 }
 507
 508 static int
 509 sched_skip_proj_predicator(const ir_node * irn, void * data)
 510 {
 511         return (is_Proj(irn));
 512 }
 513
 514 static INLINE ir_node *
 515 sched_next_nonproj(const ir_node * irn, int forward)
 516 {
 517         return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
 518 }
 519
 520 /**
 521  * Returns next operation node (non-Proj) after @p irn
 522  * or the basic block of this node
 523  */
 524 static INLINE ir_node *
 525 sched_next_op(const ir_node * irn)
 526 {
 527         ir_node *next = sched_next(irn);
 528
 529         if(is_Block(next))
 530                 return next;
 531
 532         return sched_next_nonproj(next, 1);
 533 }
 534
 535 /**
 536  * Returns previous operation node (non-Proj) before @p irn
 537  * or the basic block of this node
 538  */
 539 static INLINE ir_node *
 540 sched_prev_op(const ir_node * irn)
 541 {
 542         ir_node *prev = sched_prev(irn);
 543
 544         if(is_Block(prev))
 545                 return prev;
 546
 547         return sched_next_nonproj(prev, 0);
 548 }
 549
 550 static void
 551 sched_put_after(ir_node * insert, ir_node * irn)
 552 {
 553         if(is_Block(insert)) {
 554                 insert = sched_block_first_nonphi(insert);
 555         } else {
 556                 insert = sched_next_op(insert);
 557         }
 558         sched_add_before(insert, irn);
 559 }
 560
 561 static void
 562 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
 563 {
 564   if(is_Block(insert)) {
 565           insert = sched_block_last_noncf(si, insert);
 566   } else {
 567           insert = sched_next_nonproj(insert, 0);
 568           insert = sched_prev(insert);
 569   }
 570   sched_add_after(insert, irn);
 571 }
 572
 573 /**
 574  * Tells you whether a @p remat can be placed before the irn @p pos
 575  */
 576 static INLINE int
 577 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 578 {
 579         const ir_node   *op = remat->op;
 580         const ir_node   *prev;
 581         int        i,
 582                            n,
 583                            res = 1;
 584
 585         if(is_Block(pos)) {
 586                 prev = sched_block_last_noncf(si, pos);
 587                 prev = sched_next_nonproj(prev, 0);
 588         } else {
 589                 prev = sched_prev_op(pos);
 590         }
 591         /* do not remat if the rematted value is defined immediately before this op */
 592         if(prev == remat->op) {
 593                 return 0;
 594         }
 595
 596 #if 0
 597         /* this should be just fine, the following OP will be using this value, right? */
 598
 599         /* only remat AFTER the real definition of a value (?) */
 600         if(!value_is_defined_before(si, pos, remat->value)) {
 601 //              ir_fprintf(stderr, "error(not defined)");
 602                 return 0;
 603         }
 604 #endif
 605
 606         for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
 607                 const ir_node   *arg = get_irn_n(op, i);
 608
 609 #ifdef NO_ENLARGE_L1V3N355
 610                 if(has_reg_class(si, arg) && live) {
 611                         res &= pset_find_ptr(live, arg)?1:0;
 612                 } else {
 613                         res &= value_is_defined_before(si, pos, arg);
 614                 }
 615 #else
 616                 res &= value_is_defined_before(si, pos, arg);
 617 #endif
 618         }
 619
 620         return res;
 621 }
 622
 623 /**
 624  * Tells you whether a @p remat can be placed after the irn @p pos
 625  */
 626 static INLINE int
 627 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 628 {
 629         if(is_Block(pos)) {
 630                 pos = sched_block_first_nonphi(pos);
 631         } else {
 632                 pos = sched_next_op(pos);
 633         }
 634
 635         /* only remat AFTER the real definition of a value (?) */
 636         if(!value_is_defined_before(si, pos, remat->value)) {
 637                 return 0;
 638         }
 639
 640         return can_remat_before(si, remat, pos, live);
 641 }
 642
 643 /**
 644  * Collect potetially rematerializable OPs
 645  */
 646 static void
 647 walker_remat_collector(ir_node * irn, void * data)
 648 {
 649         spill_ilp_t    *si = data;
 650
 651         if(!is_Block(irn) && !is_Phi(irn)) {
 652                 DBG((si->dbg, LEVEL_4, "\t  Processing %+F\n", irn));
 653                 get_remats_from_op(si, irn);
 654         }
 655 }
 656
 657 /**
 658  * Inserts a copy of @p irn before @p pos
 659  */
 660 static ir_node *
 661 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 662 {
 663         ir_node     *bb;
 664         ir_node     *copy;
 665
 666         bb = is_Block(pos)?pos:get_nodes_block(pos);
 667         copy = exact_copy(irn);
 668         set_nodes_block(copy, bb);
 669         sched_put_before(si, pos, copy);
 670
 671         return copy;
 672 }
 673
 674 /**
 675  * Inserts a copy of @p irn after @p pos
 676  */
 677 static ir_node *
 678 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
 679 {
 680         ir_node     *bb;
 681         ir_node     *copy;
 682
 683         bb = is_Block(pos)?pos:get_nodes_block(pos);
 684         copy = exact_copy(irn);
 685         set_nodes_block(copy, bb);
 686         sched_put_after(pos, copy);
 687
 688         return copy;
 689 }
 690
 691 static void
 692 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 693 {
 694         char     buf[256];
 695
 696         if(can_remat_after(si, remat, pos, live)) {
 697                 ir_node         *copy,
 698                                                 *proj_copy;
 699                 op_t            *op;
 700
 701                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 702
 703                 copy = insert_copy_after(si, remat->op, pos);
 704
 705 //              ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
 706                 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
 707                 op = obstack_alloc(si->obst, sizeof(*op));
 708                 op->is_remat = 1;
 709                 op->attr.remat.remat = remat;
 710                 op->attr.remat.pre = 0;
 711                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 712
 713                 set_irn_link(copy, op);
 714                 pset_insert_ptr(si->all_possible_remats, copy);
 715                 if(remat->proj) {
 716                         proj_copy = insert_copy_after(si, remat->proj, copy);
 717                         set_irn_n(proj_copy, 0, copy);
 718                         set_irn_link(proj_copy, op);
 719                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 720                 } else {
 721                         proj_copy = NULL;
 722                 }
 723         }
 724 }
 725
 726 static void
 727 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
 728 {
 729         char     buf[256];
 730
 731         if(can_remat_before(si, remat, pos, live)) {
 732                 ir_node         *copy,
 733                                                 *proj_copy;
 734                 op_t            *op;
 735
 736                 DBG((si->dbg, LEVEL_3, "\t  >inserting remat %+F\n", remat->op));
 737
 738                 copy = insert_copy_before(si, remat->op, pos);
 739
 740 //              ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos);
 741                 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
 742                 op = obstack_alloc(si->obst, sizeof(*op));
 743                 op->is_remat = 1;
 744                 op->attr.remat.remat = remat;
 745                 op->attr.remat.pre = 1;
 746                 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
 747
 748                 set_irn_link(copy, op);
 749                 pset_insert_ptr(si->all_possible_remats, copy);
 750                 if(remat->proj) {
 751                         proj_copy = insert_copy_after(si, remat->proj, copy);
 752                         set_irn_n(proj_copy, 0, copy);
 753                         set_irn_link(proj_copy, op);
 754                         pset_insert_ptr(si->all_possible_remats, proj_copy);
 755                 } else {
 756                         proj_copy = NULL;
 757                 }
 758         }
 759 }
 760
 761
 762 /**
 763  * Insert (so far unused) remats into the irg to
 764  * recompute the potential liveness of all values
 765  */
 766 static void
 767 walker_remat_insertor(ir_node * bb, void * data)
 768 {
 769         spill_ilp_t    *si = data;
 770         spill_bb_t     *spill_bb;
 771         ir_node        *irn;
 772         int             i,
 773                                         n;
 774         irn_live_t     *li;
 775         pset           *live = pset_new_ptr_default();
 776
 777         DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
 778
 779         live_foreach(bb, li) {
 780                 ir_node        *value = (ir_node *) li->irn;
 781
 782                 /* add remats at end of block */
 783                 if (live_is_end(li) && has_reg_class(si, value)) {
 784                         pset_insert_ptr(live, value);
 785                 }
 786         }
 787
 788         spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
 789         set_irn_link(bb, spill_bb);
 790
 791         irn = sched_last(bb);
 792         while(!sched_is_end(irn)) {
 793                 ir_node   *next;
 794                 op_t      *op;
 795                 pset      *args;
 796                 ir_node   *arg;
 797
 798                 next = sched_prev(irn);
 799
 800                 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
 801
 802                 if(is_Phi(irn) || is_Proj(irn)) {
 803                         op_t      *op;
 804
 805                         if(has_reg_class(si, irn)) {
 806                                 pset_remove_ptr(live, irn);
 807                         }
 808
 809                         op = obstack_alloc(si->obst, sizeof(*op));
 810                         op->is_remat = 0;
 811                         op->attr.live_range.reloads = NULL;
 812                         op->attr.live_range.ilp = ILP_UNDEF;
 813                         set_irn_link(irn, op);
 814
 815                         irn = next;
 816                         continue;
 817                 }
 818
 819                 op = obstack_alloc(si->obst, sizeof(*op));
 820                 op->is_remat = 0;
 821                 op->attr.live_range.ilp = ILP_UNDEF;
 822                 op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
 823                 memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
 824                 set_irn_link(irn, op);
 825
 826                 args = pset_new_ptr_default();
 827
 828                 /* collect arguments of op */
 829                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
 830                         ir_node        *arg = get_irn_n(irn, i);
 831
 832                         pset_insert_ptr(args, arg);
 833                 }
 834
 835                 /* set args of op live in epilog */
 836                 pset_foreach(args, arg) {
 837                         if(has_reg_class(si, arg)) {
 838                                 pset_insert_ptr(live, arg);
 839                         }
 840                 }
 841
 842                 /* insert all possible remats after irn */
 843                 pset_foreach(args, arg) {
 844                         remat_info_t   *remat_info,
 845                                                     query;
 846                         remat_t        *remat;
 847
 848                         /* continue if the operand has the wrong reg class
 849                          */
 850                         if(!has_reg_class(si, arg))
 851                                 continue;
 852
 853                         query.irn = arg;
 854                         query.remats = NULL;
 855                         query.remats_by_operand = NULL;
 856                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 857
 858                         if(!remat_info) {
 859                                 continue;
 860                         }
 861
 862                         /* do not place post remats after jumps */
 863                         if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
 864
 865                         if(remat_info->remats_by_operand) {
 866                                 pset_foreach(remat_info->remats_by_operand, remat) {
 867                                         /* do not insert remats producing the same value as one of the operands */
 868                                         if(!pset_find_ptr(args, remat->value)) {
 869                                                 DBG((si->dbg, LEVEL_4, "\t  considering remat %+F with arg %+F\n", remat->op, arg));
 870 #ifdef REMAT_WHILE_LIVE
 871                                                 if(pset_find_ptr(live, remat->value)) {
 872                                                         insert_remat_after(si, remat, irn, live);
 873                                                 }
 874 #else
 875                                                 insert_remat_after(si, remat, irn, live);
 876 #endif
 877                                         }
 878                                 }
 879                         }
 880                 }
 881
 882                 /* delete defined value from live set */
 883                 if(has_reg_class(si, irn)) {
 884                         pset_remove_ptr(live, irn);
 885                 }
 886
 887                 /* insert all possible remats before irn */
 888                 pset_foreach(args, arg) {
 889                         remat_info_t   *remat_info,
 890                                                     query;
 891                         remat_t        *remat;
 892
 893                         /* continue if the operand has the wrong reg class
 894                          */
 895                         if(!has_reg_class(si, arg))
 896                                 continue;
 897
 898                         query.irn = arg;
 899                         query.remats = NULL;
 900                         query.remats_by_operand = NULL;
 901                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
 902
 903                         if(!remat_info) {
 904                                 continue;
 905                         }
 906
 907                         if(remat_info->remats) {
 908                                 pset_foreach(remat_info->remats, remat) {
 909                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F for arg %+F\n", remat->op, arg));
 910 #ifdef REMAT_WHILE_LIVE
 911                                         if(pset_find_ptr(live, remat->value)) {
 912                                                 insert_remat_before(si, remat, irn, live);
 913                                         }
 914 #else
 915                                         insert_remat_before(si, remat, irn, live);
 916 #endif
 917                                 }
 918                         }
 919                 }
 920
 921                 del_pset(args);
 922                 irn = next;
 923         }
 924
 925         live_foreach(bb, li) {
 926                 ir_node        *value = (ir_node *) li->irn;
 927
 928                 /* add remats at end of block */
 929                 if (live_is_end(li) && has_reg_class(si, value)) {
 930                         remat_info_t   *remat_info,
 931                                                     query;
 932                         remat_t        *remat;
 933
 934                         query.irn = value;
 935                         query.remats = NULL;
 936                         query.remats_by_operand = NULL;
 937                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
 938
 939                         if(remat_info && remat_info->remats) {
 940                                 pset_foreach(remat_info->remats, remat) {
 941                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at end of block %+F\n", remat->op, bb));
 942 #ifdef REMAT_WHILE_LIVE
 943                                         if(is_live_end(bb, remat->value)) {
 944                                                 insert_remat_before(si, remat, bb, NULL); //TODO
 945                                         }
 946 #else
 947                                         insert_remat_before(si, remat, bb, NULL);
 948 #endif
 949                                 }
 950                         }
 951                 }
 952
 953
 954                 /* add remat2s at beginning of block */
 955                 if ((live_is_in(li) || is_Phi(value)) && has_reg_class(si, value)) {
 956                         remat_info_t   *remat_info,
 957                                                     query;
 958                         remat_t        *remat;
 959
 960                         query.irn = value;
 961                         query.remats = NULL;
 962                         query.remats_by_operand = NULL;
 963                         remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
 964
 965                         if(remat_info && remat_info->remats_by_operand) {
 966                                 pset_foreach(remat_info->remats_by_operand, remat) {
 967                                         DBG((si->dbg, LEVEL_4, "\t  considering remat %+F at beginning of block %+F\n", remat->op, bb));
 968 #ifdef REMAT_WHILE_LIVE
 969                                         if(is_live_in(bb, remat->value)) {
 970                                                 insert_remat_after(si, remat, bb, NULL);
 971                                         }
 972 #else
 973                                         insert_remat_after(si, remat, bb, NULL);
 974 #endif
 975                                 }
 976                         }
 977                 }
 978         }
 979 }
 980
 981 /**
 982  * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
 983  */
 984 static void
 985 luke_endwalker(ir_node * bb, void * data)
 986 {
 987         spill_ilp_t    *si = (spill_ilp_t*)data;
 988         irn_live_t     *li;
 989         pset           *live;
 990         pset           *use_end;
 991         char            buf[256];
 992         ilp_cst_t       cst;
 993         ir_node        *irn;
 994         spill_bb_t     *spill_bb = get_irn_link(bb);
 995
 996
 997         live = pset_new_ptr_default();
 998         use_end = pset_new_ptr_default();
 999
1000         live_foreach(bb, li) {
1001                 irn = (ir_node *) li->irn;
1002                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1003                         op_t      *op;
1004
1005                         pset_insert_ptr(live, irn);
1006                         op = get_irn_link(irn);
1007                         assert(!op->is_remat);
1008                 }
1009         }
1010
1011         /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1012         /* their reg_out is unimportant because it can always be set */
1013         sched_foreach_reverse(bb, irn) {
1014                 int   i,
1015                           n;
1016
1017                 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1018
1019                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1020                         ir_node        *irn_arg = get_irn_n(irn, i);
1021                         if(has_reg_class(si, irn_arg)) {
1022                                 pset_insert_ptr(use_end, irn);
1023                         }
1024                 }
1025         }
1026
1027         ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1028         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1029
1030         spill_bb->ilp = new_set(cmp_spill, 16);
1031
1032         live_foreach(bb, li) {
1033                 irn = (ir_node *) li->irn;
1034                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1035                         spill_t     query,
1036                                    *spill;
1037
1038                         query.irn = irn;
1039                         spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1040
1041                         ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1042                         spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1043                         /* if irn is used at the end of the block, then it is live anyway */
1044                         if(!pset_find_ptr(use_end, irn))
1045                                 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1046
1047                         ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", bb, irn);
1048                         spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1049
1050                         ir_snprintf(buf, sizeof(buf), "spill_%N_%N", bb, irn);
1051                         spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, irn));
1052
1053                         spill->reg_in = ILP_UNDEF;
1054                         spill->mem_in = ILP_UNDEF;
1055                 }
1056         }
1057
1058         del_pset(live);
1059         del_pset(use_end);
1060 }
1061
1062 static ir_node *
1063 next_post_remat(const ir_node * irn)
1064 {
1065         op_t      *op;
1066
1067         if(is_Block(irn)) {
1068                 irn = sched_block_first_nonphi(irn);
1069         } else {
1070                 irn = sched_next_op(irn);
1071         }
1072
1073         if(sched_is_end(irn))
1074                 return NULL;
1075
1076         op = (op_t*)get_irn_link(irn);
1077         if(op->is_remat && !op->attr.remat.pre) {
1078                 return irn;
1079         }
1080
1081         return NULL;
1082 }
1083
1084
1085 static ir_node *
1086 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1087 {
1088         op_t      *op;
1089         ir_node   *ret;
1090
1091         if(is_Block(irn)) {
1092                 ret = sched_block_last_noncf(si, irn);
1093                 ret = sched_next(ret);
1094                 ret = sched_prev_op(ret);
1095         } else {
1096                 ret = sched_prev_op(irn);
1097         }
1098
1099         if(sched_is_end(ret) || is_Phi(ret))
1100                 return NULL;
1101
1102         op = (op_t*)get_irn_link(ret);
1103         if(op->is_remat && op->attr.remat.pre) {
1104                 return ret;
1105         }
1106
1107         return NULL;
1108 }
1109
1110 /**
1111  * Find a remat of value @p value in the epilog of @p pos
1112  */
1113 static ir_node *
1114 find_post_remat(const ir_node * value, const ir_node * pos)
1115 {
1116         while((pos = next_post_remat(pos)) != NULL) {
1117                 op_t   *op;
1118
1119                 op = get_irn_link(pos);
1120                 assert(op->is_remat && !op->attr.remat.pre);
1121
1122                 if(op->attr.remat.remat->value == value)
1123                         return (ir_node*)pos;
1124
1125 #if 0
1126         const ir_edge_t *edge;
1127                 foreach_out_edge(pos, edge) {
1128                         ir_node   *proj = get_edge_src_irn(edge);
1129                         assert(is_Proj(proj));
1130                 }
1131 #endif
1132
1133         }
1134
1135         return NULL;
1136 }
1137
1138 /**
1139  * Find a remat of value @p value in the prolog of @p pos
1140  */
1141 static ir_node *
1142 find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos)
1143 {
1144         while((pos = next_pre_remat(si,pos)) != NULL) {
1145                 op_t   *op;
1146
1147                 op = get_irn_link(pos);
1148                 assert(op->is_remat && op->attr.remat.pre);
1149
1150                 if(op->attr.remat.remat->value == value)
1151                         return (ir_node*)pos;
1152         }
1153
1154         return NULL;
1155 }
1156
1157 static spill_t *
1158 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1159 {
1160         spill_bb_t  *spill_bb = get_irn_link(bb);
1161         spill_t     *spill,
1162                                  query;
1163         char         buf[256];
1164
1165         query.irn = irn;
1166         spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1167         if(!spill) {
1168                 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1169
1170                 spill->reg_out = ILP_UNDEF;
1171                 spill->reg_in  = ILP_UNDEF;
1172                 spill->mem_in  = ILP_UNDEF;
1173
1174                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1175                 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1176
1177                 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1178                 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1179         }
1180
1181         return spill;
1182 }
1183
1184 /**
1185  * Walk all irg blocks and emit this ILP
1186  */
1187 static void
1188 luke_blockwalker(ir_node * bb, void * data)
1189 {
1190         spill_ilp_t    *si = (spill_ilp_t*)data;
1191         ir_node        *irn;
1192         irn_live_t     *li;
1193         pset           *live;
1194         char            buf[256];
1195         ilp_cst_t       cst;
1196         spill_bb_t     *spill_bb = get_irn_link(bb);
1197         int             i;
1198         ir_node        *tmp;
1199         spill_t        *spill;
1200
1201
1202         live = pset_new_ptr_default();
1203
1204         /* do something at the end of the block */
1205
1206         /* init live values at end of block */
1207         live_foreach(bb, li) {
1208                 ir_node        *irn = (ir_node *) li->irn;
1209
1210                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1211                         pset_insert_ptr(live, irn);
1212                 }
1213         }
1214
1215         spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1216         memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1217
1218         i=0;
1219         live_foreach(bb, li) {
1220                 ir_node        *irn = (ir_node *) li->irn;
1221                 op_t           *op;
1222
1223                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1224                         spill = set_find_spill(spill_bb->ilp, irn);
1225                         assert(spill);
1226
1227                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1228                         spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1229
1230                         /* reload <= mem_out */
1231                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1232                         lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0);
1233                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1234
1235                         op = get_irn_link(irn);
1236                         assert(!op->is_remat);
1237
1238                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1239                         op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1240                         op->attr.live_range.op = bb;
1241
1242                         ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1243                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1244
1245                         /* reg_out - reload - remat - live_range <= 0 */
1246                         lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1247                         lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0);
1248                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1249                         foreach_pre_remat(si, bb, tmp) {
1250                                 op_t     *remat_op = get_irn_link(tmp);
1251                                 if(remat_op->attr.remat.remat->value == irn) {
1252                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1253                                 }
1254                         }
1255
1256                         ++i;
1257                 }
1258         }
1259         DBG((si->dbg, LEVEL_4, "\t   %d values live at end of block %+F\n", pset_count(live), bb));
1260
1261         sched_foreach_reverse(bb, irn) {
1262                 op_t       *op;
1263                 op_t       *tmp_op;
1264                 int         n,
1265                                         k,
1266                                         d = 0;
1267                 ilp_cst_t       check_pre,
1268                                         check_post,
1269                                         check_post_remat;
1270                 set        *args = new_set(cmp_keyval, get_irn_arity(irn));
1271                 keyval_t   *keyval;
1272
1273                 if(is_Phi(irn))
1274                         break;
1275
1276                 op = get_irn_link(irn);
1277                 /* skip remats */
1278                 if(op->is_remat) continue;
1279                 DBG((si->dbg, LEVEL_4, "\t  at node %+F\n", irn));
1280
1281                 if(has_reg_class(si, irn)) {
1282                         assert(pset_find_ptr(live, irn));
1283                         pset_remove_ptr(live, irn);
1284                 }
1285
1286                 /* init set of irn's arguments */
1287                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1288                         ir_node        *irn_arg = get_irn_n(irn, i);
1289                         if(has_reg_class(si, irn_arg)) {
1290                                 set_insert_keyval(args, irn_arg, (void*)i);
1291                         }
1292                 }
1293
1294                 /* check the register pressure after the epilog */
1295                 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1296                 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1297
1298                 /* iterate over L\U */
1299                 pset_foreach(live, tmp) {
1300                         if(!set_find_keyval(args, tmp)) {
1301                                 /* if a live value is not used by irn */
1302                                 tmp_op = get_irn_link(tmp);
1303 //                              assert(tmp_op->attr.live_range.op != irn);
1304                                 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1305                         }
1306                 }
1307                 /* iterate over following remats and remove possibly defined values again from check_post_remat */
1308                 foreach_post_remat(irn, tmp) {
1309                         op_t           *remat_op = get_irn_link(tmp);
1310                         const ir_node  *value = remat_op->attr.remat.remat->value;
1311                         op_t           *val_op = get_irn_link(value);
1312
1313                         assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1314
1315                         /* values that are defined by remats are not counted */
1316                         /* TODO assert(val_op->attr.live_range.ilp)) */
1317                         if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1318                                 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1319                         }
1320                 }
1321
1322
1323                 /* new live ranges for values from L\U defined by remat2s or used by remats */
1324                 pset_foreach(live, tmp) {
1325                         ir_node     *value = tmp;//remat_op->attr.remat.remat->value;
1326                         op_t        *value_op = get_irn_link(value);
1327
1328                         if(!set_find_keyval(args, value)) {
1329                                 ilp_var_t    prev_lr = ILP_UNDEF;
1330                                 ir_node     *remat;
1331                                 cst = ILP_UNDEF;
1332
1333                                 foreach_post_remat(irn, remat) {
1334                                         op_t        *remat_op = get_irn_link(remat);
1335
1336                                         /* if value is being rematerialized by this remat */
1337                                         if(value == remat_op->attr.remat.remat->value) {
1338                                                 if(cst == ILP_UNDEF) {
1339                                                         /* next_live_range <= prev_live_range + sum remat2s */
1340                                                         ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1341                                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1342                                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1343                                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1344                                                         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1345                                                         lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1346                                                 }
1347
1348                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1349                                         }
1350                                 }
1351
1352 #ifdef MAY_DIE_AT_PRE_REMAT
1353                                 if(cst == ILP_UNDEF) {
1354                                         foreach_pre_remat(si, irn, remat) {
1355                                                 int          i,
1356                                                                          n;
1357
1358                                                 for (i = 0, n = get_irn_arity(remat); i < n; ++i) {
1359                                                         ir_node        *remat_arg = get_irn_n(remat, i);
1360
1361                                                         /* if value is being used by this remat */
1362                                                         if(value == remat_arg) {
1363                                                                 /* next_live_range <= prev_live_range */
1364                                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1365                                                                 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1366
1367                                                                 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1368                                                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1369                                                                 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1370                                                                 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1371                                                                 goto fertig;
1372                                                         }
1373                                                         /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1374                                                 }
1375                                         }
1376                                 }
1377 #endif
1378
1379 fertig:
1380                                 if(prev_lr != ILP_UNDEF) {
1381                                         value_op->attr.live_range.ilp = prev_lr;
1382                                         value_op->attr.live_range.op = irn;
1383                                 }
1384                         }
1385                 }
1386
1387                 /* get count of values in my register class defined by irn */
1388                 /* also add defined values to check_post_remat; do this before iterating over args */
1389                 if(get_irn_mode(irn) == mode_T) {
1390                         ir_node  *proj = sched_next(irn);
1391                         op_t     *proj_op = get_irn_link(proj);
1392
1393                         while(is_Proj(proj)) {
1394                                 if(has_reg_class(si, proj)) {
1395                                         ++d;
1396                                         lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0);
1397                                 }
1398                                 proj = sched_next(proj);
1399                                 proj_op = get_irn_link(proj);
1400                         }
1401                 } else {
1402                         if(has_reg_class(si, irn)) {
1403                                  d = 1;
1404                                  lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0);
1405                         }
1406                 }
1407                 DBG((si->dbg, LEVEL_4, "\t   %+F produces %d values in my register class\n", irn, d));
1408
1409                 /* count how many regs irn needs for arguments */
1410                 k = set_count(args);
1411
1412                 /* check the register pressure in the prolog */
1413                 /* sum_{L\U} lr <= n - |U| */
1414                 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1415                 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k);
1416
1417                 /* check the register pressure in the epilog */
1418                 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1419                 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1420
1421                 set_foreach(args, keyval) {
1422                         ilp_var_t       next_lr;
1423                         op_t           *arg_op;
1424                         ilp_var_t       post_use;
1425                         int             p = 0;
1426                         spill_t        *spill;
1427                         ir_node        *arg = keyval->key;
1428
1429                         spill = add_to_spill_bb(si, bb, arg);
1430
1431                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1432                         next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1433
1434                         i = (int)keyval->val;
1435                         assert(i<n);
1436
1437                         ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1438                         op->attr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, irn));
1439
1440                         /* reload <= mem_out */
1441                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1442                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1443                         lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1444
1445                         arg_op = get_irn_link(arg);
1446
1447                         /* requirement: arg must be in register for use */
1448                         /* reload + remat + live_range == 1 */
1449                         ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1450                         cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1451
1452                         lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0);
1453                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1454                         foreach_pre_remat(si, irn, tmp) {
1455                                 op_t     *remat_op = get_irn_link(tmp);
1456                                 if(remat_op->attr.remat.remat->value == arg) {
1457                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1458                                 }
1459                         }
1460
1461                         /* the epilog stuff - including post_use, post, post_remat */
1462                         ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1463                         post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1464
1465                         lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1466
1467                         /* arg is live throughout epilog if the next live_range is in a register */
1468                         if(pset_find_ptr(live, arg)) {
1469                                 DBG((si->dbg, LEVEL_3, "\t  arg %+F is possibly live in epilog of %+F\n", arg, irn));
1470
1471                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1472                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1473                                 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1474                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1475
1476                                 lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1477                         }
1478
1479                         /*forall remat2 which use arg add a similar cst*/
1480                         foreach_post_remat(irn, tmp) {
1481                                 int         i,
1482                                                         n;
1483
1484                                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1485                                         ir_node    *remat_arg = get_irn_n(tmp, i);
1486                                         op_t       *remat_op = get_irn_link(tmp);
1487
1488                                         if(remat_arg == arg) {
1489                                                 DBG((si->dbg, LEVEL_3, "\t  found remat with arg %+F in epilog of %+F\n", arg, irn));
1490
1491                                                 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1492                                                 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1493                                                 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1494                                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1495                                         }
1496                                 }
1497                         }
1498
1499                         /* new live range begins for each argument */
1500                         arg_op->attr.live_range.ilp = next_lr;
1501                         arg_op->attr.live_range.op = irn;
1502
1503                         pset_insert_ptr(live, arg);
1504                 }
1505
1506                 /* start new live ranges for values used by remats */
1507                 foreach_pre_remat(si, irn, tmp) {
1508                         int          i,
1509                                                  n;
1510
1511                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1512                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1513                                 op_t           *arg_op = get_irn_link(remat_arg);
1514                                 ilp_var_t       prev_lr;
1515
1516                                 if(!has_reg_class(si, remat_arg)) continue;
1517
1518                                 /* if value is becoming live through use by remat */
1519                                 if(!pset_find_ptr(live, remat_arg)) {
1520                                         ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1521                                         prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1522
1523                                         arg_op->attr.live_range.ilp = prev_lr;
1524                                         arg_op->attr.live_range.op = irn;
1525
1526                                         DBG((si->dbg, LEVEL_4, "  value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1527
1528                                         /* TODO ist das hier die richtige Stelle???? */
1529                                         pset_insert_ptr(live, remat_arg);
1530                                         add_to_spill_bb(si, bb, remat_arg);
1531                                 }
1532                                 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1533                         }
1534                 }
1535
1536                 /* iterate over L\U */
1537                 pset_foreach(live, tmp) {
1538                         if(!set_find_keyval(args, tmp)) {
1539                                 /* if a live value is not used by irn */
1540                                 tmp_op = get_irn_link(tmp);
1541 //                              assert(tmp_op->attr.live_range.op != irn);
1542                                 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1543                                 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1544                         }
1545                 }
1546
1547                 /* requirements for remats */
1548                 foreach_pre_remat(si, irn, tmp) {
1549                         op_t        *remat_op = get_irn_link(tmp);
1550                         int          i,
1551                                                  n;
1552
1553                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1554                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1555                                 op_t           *arg_op = get_irn_link(remat_arg);
1556
1557                                 if(!has_reg_class(si, remat_arg)) continue;
1558
1559                                 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1560                                 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1561                                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1562
1563                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1564                                 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1565
1566                                 /* if remat arg is also used by current op then we can use reload placed for this argument */
1567                                 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1568                                         int    index = (int)keyval->val;
1569
1570                                         lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0);
1571                                 }
1572                         }
1573                 }
1574
1575                 /* requirements for remats2
1576                  *
1577                  *  TODO unsure if this does the right thing.
1578                  *  should insert values into set if they do not become live through remat and
1579                  *  op
1580                  */
1581                 foreach_post_remat(irn, tmp) {
1582                         op_t        *remat_op = get_irn_link(tmp);
1583                         int          i,
1584                                                  n;
1585
1586                         for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1587                                 ir_node        *remat_arg = get_irn_n(tmp, i);
1588                                 op_t           *arg_op = get_irn_link(remat_arg);
1589
1590                                 if(!has_reg_class(si, remat_arg)) continue;
1591
1592                                 /* only for values in L\U, the others are handled with post_use */
1593                                 if(!set_find_keyval(args, remat_arg)) {
1594                                         /* remat <= live_rang(remat_arg) */
1595                                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1596                                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1597
1598                                         /* if value is becoming live through use by remat2 */
1599                                         if(!pset_find_ptr(live, remat_arg)) {
1600                                                 ilp_var_t     lr;
1601
1602                                                 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1603                                                 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1604
1605                                                 arg_op->attr.live_range.ilp = lr;
1606                                                 arg_op->attr.live_range.op = irn;
1607
1608                                                 DBG((si->dbg, LEVEL_3, "  value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1609
1610                                                 pset_insert_ptr(live, remat_arg);
1611                                                 add_to_spill_bb(si, bb, remat_arg);
1612                                         }
1613
1614                                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1615                                         lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1616                                 }
1617                         }
1618                 }
1619
1620                 /* iterate over following remats and add them to check_post_remat */
1621                 foreach_post_remat(irn, tmp) {
1622                         op_t           *remat_op = get_irn_link(tmp);
1623
1624                         assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1625
1626                         lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1627                 }
1628
1629
1630                 DBG((si->dbg, LEVEL_4, "\t   %d values live at %+F\n", pset_count(live), irn));
1631
1632                 pset_foreach(live, tmp) {
1633                         assert(has_reg_class(si, tmp));
1634                 }
1635
1636                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1637                         ir_node        *arg = get_irn_n(irn, i);
1638
1639                         assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1640                 }
1641
1642                 del_set(args);
1643         }
1644
1645
1646
1647         /* do something at the beginning of the block */
1648
1649         /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1650         DBG((si->dbg, LEVEL_3, "\t   %d values live at beginning of block %+F\n", pset_count(live), bb));
1651
1652         pset_foreach(live, irn) {
1653                 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1654         }
1655
1656         /* construct mem_outs for all values */
1657
1658         set_foreach(spill_bb->ilp, spill) {
1659                 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1660                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1661
1662                 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1663                 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1664
1665                 if(pset_find_ptr(live, spill->irn)) {
1666                         DBG((si->dbg, LEVEL_5, "\t     %+F live at beginning of block %+F\n", spill->irn, bb));
1667
1668                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1669                         spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1670
1671                         lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1672                 }
1673         }
1674
1675
1676         /* L\U is empty at bb start */
1677         /* arg is live throughout epilog if it is reg_in into this block */
1678
1679         /* check the register pressure at the beginning of the block
1680          * including remats
1681          */
1682         ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1683         cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1684
1685         pset_foreach(live, irn) {
1686                         spill = set_find_spill(spill_bb->ilp, irn);
1687                         assert(spill);
1688
1689                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
1690                         spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1691
1692                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
1693         }
1694         foreach_post_remat(bb, irn) {
1695                 op_t     *remat_op = get_irn_link(irn);
1696
1697                 DBG((si->dbg, LEVEL_4, "\t  next post remat: %+F\n", irn));
1698                 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1699
1700                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1701         }
1702
1703         /* forall remat2 add requirements */
1704         foreach_post_remat(bb, tmp) {
1705                 int         i,
1706                                         n;
1707
1708                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1709                         ir_node    *remat_arg = get_irn_n(tmp, i);
1710                         op_t       *remat_op = get_irn_link(tmp);
1711
1712                         if(!has_reg_class(si, remat_arg)) continue;
1713
1714                         spill = set_find_spill(spill_bb->ilp, remat_arg);
1715                         assert(spill);
1716
1717                         /* TODO verify this is placed correctly */
1718                         ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
1719                         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1720                         lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1721                         lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1722                 }
1723         }
1724
1725         /* mem_in/reg_in for live_in values, especially phis and their arguments */
1726 //      if(get_Block_n_cfgpreds(bb) > 1) {
1727                 pset_foreach(live, irn) {
1728                         int          p = 0,
1729                                                  i,
1730                                                  n;
1731
1732                         spill = set_find_spill(spill_bb->ilp, irn);
1733                         assert(spill && spill->irn == irn);
1734
1735                         if(is_Phi(irn) && get_nodes_block(irn) == bb) {
1736                                 for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) {
1737                                         ilp_cst_t       mem_in,
1738                                                                         reg_in;
1739                                         ir_node        *phi_arg = get_Phi_pred(irn, i);
1740                                         ir_node        *bb_p = get_Block_cfgpred_block(bb, i);
1741                                         spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
1742                                         spill_t        *spill_p;
1743
1744                                         /* although the phi is in the right regclass one or more of
1745                                          * its arguments can be in a different one or at least to
1746                                          * ignore
1747                                          */
1748                                         if(has_reg_class(si, phi_arg)) {
1749                                                 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1750                                                 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1751                                                 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1752                                                 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1753
1754                                                 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1755                                                 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1756
1757                                                 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
1758                                                 assert(spill_p);
1759
1760                                                 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1761                                                 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1762                                         }
1763                                 }
1764                         } else {
1765                                 /* else assure the value arrives on all paths in the same resource */
1766
1767                                 for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) {
1768                                         ilp_cst_t       mem_in,
1769                                                                         reg_in;
1770                                         ir_node        *bb_p = get_Block_cfgpred_block(bb, i);
1771                                         spill_bb_t     *spill_bb_p = get_irn_link(bb_p);
1772                                         spill_t        *spill_p;
1773
1774                                         ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1775                                         mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1776                                         ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1777                                         reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1778
1779                                         lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1780                                         lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1781
1782                                         spill_p = set_find_spill(spill_bb_p->ilp, irn);
1783                                         assert(spill_p);
1784
1785                                         lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1786                                         lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1787                                 }
1788                         }
1789                 }
1790 //      }
1791
1792         /* first live ranges from reg_ins */
1793         pset_foreach(live, irn) {
1794                 op_t      *op = get_irn_link(irn);
1795
1796                 spill = set_find_spill(spill_bb->ilp, irn);
1797                 assert(spill && spill->irn == irn);
1798
1799                 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
1800                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1801                 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1802                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1803
1804                 foreach_post_remat(bb, tmp) {
1805                         op_t     *remat_op = get_irn_link(tmp);
1806
1807                         if(remat_op->attr.remat.remat->value == irn) {
1808                                 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1809                         }
1810                 }
1811         }
1812
1813         /* walk forward now and compute constraints for placing spills */
1814         /* this must only be done for values that are not defined in this block */
1815         pset_foreach(live, irn) {
1816                 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", spill->irn, bb);
1817                 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1818
1819                 spill = set_find_spill(spill_bb->ilp, irn);
1820                 assert(spill);
1821
1822                 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
1823                 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1824
1825                 sched_foreach_op(bb, tmp) {
1826                         op_t   *op = get_irn_link(tmp);
1827
1828                         if(is_Phi(tmp)) continue;
1829                         assert(!is_Proj(tmp));
1830
1831                         if(op->is_remat) {
1832                                 ir_node   *value = op->attr.remat.remat->value;
1833
1834                                 if(value == irn) {
1835                                         /* only collect remats up to the first use of a value */
1836                                         lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
1837                                 }
1838                         } else {
1839                                 int i,
1840                                         n;
1841
1842                                 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1843                                         ir_node    *arg = get_irn_n(tmp, i);
1844
1845                                         if(arg == irn) {
1846                                                 /* if a value is used stop collecting remats */
1847                                                 cst = ILP_UNDEF;
1848                                         }
1849                                         break;
1850                                 }
1851                         }
1852                         if(cst == ILP_UNDEF) break;
1853                 }
1854         }
1855
1856
1857         /* if a value is used by a mem-phi, then mem_in of this value is 0 (has to be spilled again into a different slot)
1858            mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor?
1859          */
1860
1861         /* mem_in of mem-phi has associated costs (but first one is free) */
1862         /* define n_mem_copies as positive integer in each predecessor block,
1863            #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor)
1864            TODO
1865          */
1866
1867
1868 /*      ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1869         cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1870         lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1871 */
1872
1873 #if 0
1874         ir_snprintf(buf, sizeof(buf), "ce2_%N_%N_%d", op, irn, i);
1875         cst = lpp_add_cst(si->lpp, buf, lpp_less, n_operands);
1876         lpp_set_factor_fast(si->lpp, cst, op_lr->in_mem_var, -1.0);
1877         op_lr->is_remat_var = lpp_add_var(si->lpp, buf, lpp_binary, COST_REMAT);
1878 #endif
1879
1880         del_pset(live);
1881 }
1882
1883
1884 #if 0
1885 static void
1886 process_block(ir_node * block, void *data)
1887 {
1888         spill_ilp_t    *si = data;
1889         int             step = 0;
1890         int             n_regs = arch_register_class_n_regs(si->cls);
1891         int             n_preds = get_irn_arity(block);
1892         irn_live_t     *li;
1893         ir_node        *irn;
1894
1895         DBG((si->dbg, LEVEL_3, "\nProcessing %+F\n", block));
1896
1897         /*
1898          * compute all possible remats, sorted by operands *and* resulting
1899          * values
1900          */
1901         ir_node        *new_r_Tuple(ir_graph * irg, ir_node * block, int arity, ir_node * in[]);
1902         ir_node        *new_r_Proj(ir_graph * irg, ir_node * block, ir_node * arg, ir_mode * mode, long proj);
1903         void           *get_irn_link(const ir_node * node);
1904         void            set_irn_link(ir_node * node, void *link);
1905
1906         int             block_dominates(const ir_node * a, const ir_node * b);
1907
1908
1909         void            compute_doms(ir_graph * irg);
1910         void            free_dom(ir_graph * irg);
1911
1912
1913         /*
1914          * nochmal: - laufe in dominatorreihenfolge die bloecke ab -
1915          * sammle dabei moeglichkeiten zur rematerialisierung von werten
1916          * (auch inverse etc.) - mehrere listen: fuer zielwert und fuer
1917          * jeden operanden (get_remats_with_operand) - ignoriere projs
1918          * (und phis?) und andere registerklassen - wenn ich nun auf eine
1919          * op treffe füge ich für alle operanden mögliche remats ein
1920          * (keine reloads und spills) in den schedule direkt vor der op
1921          * ein (in der Hoffnung, dass diese nicht wegoptimiert werden...).
1922          * - dann überspringe ich die folgenden projs - nach den projs
1923          * wird der epilog eingefügt (nur remats, keine spills) -
1924          * lebendigkeit neu berechnen
1925          *
1926          * -durchlaufe den ganzen mist nochmal und emittiere dabei das ILP:
1927          *
1928          *
1929          * datenhaltung: - remat merkt sich die zugehörige ilp_var und op
1930          * - op merkt sich (zugehörige remats) - block merkt sich ilp_vars
1931          * für spills
1932          *
1933          * nachbearbeitung: - durchlaufe alle instruktionen, dabei entferne
1934          * nicht benötigte remats und füge benötigte reloads ein - wenn
1935          * spill in Grundblock eingesetzt werden soll durchlaufen wir den
1936          * Block von oben runter und setzen das spill an der letzten
1937          * möglichen Stelle vor der ersten Benutzung/Reload ein. -
1938          * Speicherkopienminimierung: teste Speicherwerte auf Interferenz
1939          * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
1940          * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
1941          * Kontext)
1942          *
1943          * nachoptimierung: - laufe wieder von oben nach unten durch jeden
1944          * block - wenn ich ein reload finde, dann schiebe es hoch solange
1945          * es der Registerdruck erlaubt (wenn ich es über sein
1946          * zugehöriges spill schiebe ist das ein bug, oder?)
1947          *
1948          */
1949
1950 }
1951 #endif
1952
1953 static INLINE int
1954 is_zero(double x)
1955 {
1956         return fabs(x) < 0.00001;
1957 }
1958
1959 #if 0
1960 static int
1961 is_spilled(const spill_ilp_t * si, const live_range_t * lr)
1962 {
1963         return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
1964 }
1965 #endif
1966
1967 static int
1968 is_mem_phi(const ir_node * phi, void *data)
1969 {
1970         spill_ilp_t    *si = data;
1971 //      return is_spilled(si, get_use_head(si, phi)->closest_use);
1972         return 0;
1973 }
1974
1975 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
1976 {
1977         spill_ilp_t *si = get_irg_link(current_ir_graph);
1978
1979         if(pset_find_ptr(si->all_possible_remats, n)) {
1980                 op_t   *op = (op_t*)get_irn_link(n);
1981                 assert(op && op->is_remat);
1982
1983                 if(op->attr.remat.pre) {
1984                         ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
1985                 } else {
1986                         ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
1987                 }
1988
1989                 return 1;
1990         }
1991
1992         return 0;
1993 }
1994
1995 static void
1996 dump_graph_with_remats(ir_graph * irg, const char * suffix)
1997 {
1998         set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
1999         be_dump(irg, suffix, dump_ir_block_graph_sched);
2000         set_dump_node_vcgattr_hook(NULL);
2001 }
2002
2003 /**
2004  * Edge hook to dump the schedule edges with annotated register pressure.
2005  */
2006 static int
2007 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2008 {
2009         if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2010                 ir_node *prev = sched_prev(irn);
2011                 fprintf(F, "edge:{sourcename:\"");
2012                 PRINT_NODEID(irn);
2013                 fprintf(F, "\" targetname:\"");
2014                 PRINT_NODEID(prev);
2015                 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2016                 fprintf(F, "\" color:magenta}\n");
2017         }
2018         return 1;
2019 }
2020
2021 static void
2022 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2023 {
2024         DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2025
2026         dump_consts_local(0);
2027         set_dump_node_edge_hook(sched_pressure_edge_hook);
2028         dump_ir_block_graph(irg, suffix);
2029         set_dump_node_edge_hook(old);
2030 }
2031
2032 static void
2033 walker_pressure_annotator(ir_node * bb, void * data)
2034 {
2035         spill_ilp_t  *si = data;
2036         ir_node      *irn;
2037         irn_live_t   *li;
2038         int           i,
2039                                   n;
2040         pset         *live = pset_new_ptr_default();
2041         int           projs = 0;
2042
2043         live_foreach(bb, li) {
2044                 irn = (ir_node *) li->irn;
2045
2046                 if (live_is_end(li) && has_reg_class(si, irn)) {
2047                         pset_insert_ptr(live, irn);
2048                 }
2049         }
2050
2051         set_irn_link(bb, (void*)pset_count(live));
2052
2053         sched_foreach_reverse(bb, irn) {
2054                 if(is_Phi(irn)) {
2055                         set_irn_link(irn, (void*)pset_count(live));
2056                         continue;
2057                 }
2058
2059                 if(has_reg_class(si, irn)) {
2060                         pset_remove_ptr(live, irn);
2061                         if(is_Proj(irn)) ++projs;
2062                 }
2063
2064                 if(!is_Proj(irn)) projs = 0;
2065
2066                 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2067                         ir_node    *arg = get_irn_n(irn, i);
2068
2069                         if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2070                 }
2071                 set_irn_link(irn, (void*)pset_count(live)+projs);
2072         }
2073
2074         del_pset(live);
2075 }
2076
2077 static void
2078 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2079 {
2080         be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2081 }
2082
2083 #ifdef KEEPALIVE
2084 static void
2085 connect_all_remats_with_keep(spill_ilp_t * si)
2086 {
2087         ir_node   *irn;
2088         ir_node  **ins,
2089                          **pos;
2090         int        n_remats;
2091
2092
2093         n_remats = pset_count(si->all_possible_remats);
2094         if(n_remats) {
2095                 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2096
2097                 pos = ins;
2098                 pset_foreach(si->all_possible_remats, irn) {
2099                         *pos = irn;
2100                         ++pos;
2101                 }
2102
2103                 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2104
2105                 obstack_free(si->obst, ins);
2106         }
2107 }
2108 #endif
2109
2110 /** insert a spill at an arbitrary position */
2111 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2112 {
2113         ir_node *bl     = is_Block(insert)?insert:get_nodes_block(insert);
2114         ir_graph *irg   = get_irn_irg(bl);
2115         ir_node *frame  = get_irg_frame(irg);
2116         ir_node *spill;
2117         ir_node *next;
2118
2119         const arch_register_class_t *cls       = arch_get_irn_reg_class(arch_env, irn, -1);
2120         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2121
2122         spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2123
2124         /*
2125          * search the right insertion point. a spill of a phi cannot be put
2126          * directly after the phi, if there are some phis behind the one which
2127          * is spilled. Also, a spill of a Proj must be after all Projs of the
2128          * same tuple node.
2129          *
2130          * Here's one special case:
2131          * If the spill is in the start block, the spill must be after the frame
2132          * pointer is set up. This is done by setting insert to the end of the block
2133          * which is its default initialization (see above).
2134          */
2135
2136         if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2137                 insert = frame;
2138
2139         for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2140                 insert = next;
2141
2142         sched_add_after(insert, spill);
2143         return spill;
2144 }
2145
2146 static void
2147 delete_remat(spill_ilp_t * si, ir_node * remat) {
2148         int       i,
2149                       n;
2150         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2151
2152         sched_remove(remat);
2153
2154         /* kill links to operands */
2155         for (i = -1, n = get_irn_arity(remat); i < n; ++i) {
2156                 set_irn_n(remat, i, bad);
2157         }
2158 }
2159
2160 static void
2161 clean_remat_info(spill_ilp_t * si)
2162 {
2163         int            i,
2164                                n;
2165         remat_t       *remat;
2166         remat_info_t  *remat_info;
2167         ir_node       *bad = get_irg_bad(si->chordal_env->irg);
2168
2169         set_foreach(si->remat_info, remat_info) {
2170                 if(!remat_info->remats) continue;
2171
2172                 pset_foreach(remat_info->remats, remat)
2173                 {
2174                         if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2175                                 set_irn_n(remat->proj, -1, bad);
2176                                 set_irn_n(remat->proj, 0, bad);
2177                         }
2178
2179                         if(get_irn_n_edges(remat->op) == 0) {
2180                                 for (i = -1, n = get_irn_arity(remat->op); i < n; ++i) {
2181                                         set_irn_n(remat->op, i, bad);
2182                                 }
2183                         }
2184                 }
2185
2186                 if(remat_info->remats) del_pset(remat_info->remats);
2187                 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2188         }
2189 }
2190
2191 static void
2192 delete_unnecessary_remats(spill_ilp_t * si)
2193 {
2194         int       i,
2195                       n;
2196         ir_node  *bad = get_irg_bad(si->chordal_env->irg);
2197
2198 #ifdef KEEPALIVE
2199         if(si->keep) {
2200                 ir_node   *end = get_irg_end(si->chordal_env->irg);
2201                 ir_node  **keeps;
2202
2203                 for (i = 0, n = get_irn_arity(si->keep); i < n; ++i) {
2204                         ir_node        *keep_arg = get_irn_n(si->keep, i);
2205                         op_t           *arg_op = get_irn_link(keep_arg);
2206                         lpp_name_t     *name;
2207
2208                         assert(arg_op->is_remat);
2209
2210                         name = si->lpp->vars[arg_op->attr.remat.ilp];
2211
2212                         if(is_zero(name->value)) {
2213                                 DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", keep_arg));
2214                                 /* TODO check whether reload is preferred over remat (could be bug) */
2215                                 delete_remat(si, keep_arg);
2216                         } else {
2217                                 if(arg_op->attr.remat.pre) {
2218                                         DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2219                                 } else {
2220                                         DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2221                                 }
2222                         }
2223
2224                         set_irn_n(si->keep, i, bad);
2225                 }
2226 #if 0
2227                 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2228                         ir_node        *end_arg = get_End_keepalive(end, i);
2229
2230                         if(end_arg != si->keep) {
2231                                 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2232                         }
2233                 }
2234                 keeps = obstack_finish(si->obst);
2235                 set_End_keepalives(end, n-1, keeps);
2236                 obstack_free(si->obst, keeps);
2237 #endif
2238         } else {
2239                 DBG((si->dbg, LEVEL_2, "\t  no remats to delete (none have been inserted)\n"));
2240         }
2241 #else
2242         ir_node  *remat;
2243
2244         pset_foreach(si->all_possible_remats, remat) {
2245                 op_t           *remat_op = get_irn_link(remat);
2246                 lpp_name_t     *name = si->lpp->vars[remat_op->attr.remat.ilp];
2247
2248                 if(is_zero(name->value)) {
2249                         DBG((si->dbg, LEVEL_3, "\t  deleting remat %+F\n", remat));
2250                         /* TODO check whether reload is preferred over remat (could be bug) */
2251                         delete_remat(si, remat);
2252                 } else {
2253                         if(remat_op->attr.remat.pre) {
2254                                 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2255                         } else {
2256                                 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2257                         }
2258                 }
2259         }
2260 #endif
2261 }
2262
2263 /**
2264  * @param before   The node after which the spill will be placed in the schedule
2265  */
2266 /* TODO set context properly */
2267 static ir_node *
2268 insert_spill(spill_ilp_t * si, const ir_node * irn, const ir_node * value, const ir_node * before)
2269 {
2270         defs_t   *defs;
2271         ir_node  *spill;
2272         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2273
2274         DBG((si->dbg, LEVEL_3, "\t  inserting spill for value %+F after %+F\n", irn, before));
2275
2276         spill = be_spill2(arch_env, irn, before, irn);
2277
2278         defs = set_insert_def(si->values, value);
2279         assert(defs);
2280
2281         /* enter into the linked list */
2282         set_irn_link(spill, defs->spills);
2283         defs->spills = spill;
2284
2285 #ifdef KEEPALIVE_SPILLS
2286         keep_alive(spill);
2287 #endif
2288
2289         return spill;
2290 }
2291
2292
2293 /**
2294  * Add remat to list of defs, destroys link field!
2295  */
2296 static void
2297 insert_remat(spill_ilp_t * si, ir_node * remat)
2298 {
2299         defs_t   *defs;
2300         op_t     *remat_op = get_irn_link(remat);
2301
2302         assert(remat_op->is_remat);
2303
2304         defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2305         assert(defs);
2306
2307         /* enter into the linked list */
2308         set_irn_link(remat, defs->remats);
2309         defs->remats = remat;
2310 }
2311
2312 static void
2313 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2314 {
2315         ir_node  *next;
2316         defs_t   *defs;
2317
2318         defs = set_find_def(si->values, value);
2319
2320         if(defs && defs->spills) {
2321                 for(next = defs->spills; next; next = get_irn_link(next)) {
2322                         pset_insert_ptr(spills, next);
2323                 }
2324         } else if (is_Phi(value)) {
2325                 /* recursion */
2326                 if(!pset_find_ptr(visited, value)) {
2327                         int    i,
2328                                    n;
2329
2330                         pset_insert_ptr(visited, value);
2331                         for(i=0, n=get_irn_arity(value); i<n; ++i) {
2332                                 ir_node    *arg = get_irn_n(value, i);
2333
2334                                 collect_spills(si, arg, spills, visited);
2335                         }
2336                 }
2337         } else {
2338 //              assert(0 && "Phi operand not spilled");
2339         }
2340 }
2341
2342 static pset *
2343 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2344 {
2345         pset     *spills = pset_new_ptr_default();
2346         pset     *visited = pset_new_ptr_default();
2347
2348         collect_spills(si, value, spills, visited);
2349         del_pset(visited);
2350
2351         return spills;
2352 }
2353
2354 /**
2355  * Add reload before operation and add to list of defs
2356  */
2357 static ir_node *
2358 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
2359 {
2360         defs_t   *defs;
2361         ir_node  *reload,
2362                          *spill;
2363         const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2364
2365         DBG((si->dbg, LEVEL_3, "\t  inserting reload for value %+F before %+F\n", value, after));
2366
2367         defs = set_find_def(si->values, value);
2368         /* get a spill of this value */
2369         if((!defs || !defs->spills) && is_Phi(value)) {
2370                 pset  *spills;
2371
2372                 spills = get_spills_for_value(si, value);
2373
2374                 spill = pset_first(spills);
2375                 del_pset(spills);
2376
2377                 if(!defs) {
2378                         defs = set_insert_def(si->values, value);
2379                 }
2380         } else {
2381                 spill = defs->spills;
2382         }
2383         assert(spill && "no spill placed before reload");
2384
2385         reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
2386
2387         /* enter into the linked list */
2388         set_irn_link(reload, defs->remats);
2389         defs->remats = reload;
2390
2391         return reload;
2392
2393 }
2394
2395 static void
2396 walker_spill_placer(ir_node * bb, void * data) {
2397         spill_ilp_t   *si = (spill_ilp_t*)data;
2398         ir_node       *irn;
2399         spill_bb_t    *spill_bb = get_irn_link(bb);
2400         pset          *spills_to_do = pset_new_ptr_default();
2401         spill_t       *spill;
2402
2403         set_foreach(spill_bb->ilp, spill) {
2404                 lpp_name_t    *name;
2405
2406                 assert(spill->spill > 0);
2407
2408                 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2409                         name = si->lpp->vars[spill->mem_in];
2410                         if(!is_zero(name->value)) {
2411                                 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F\n", spill->irn));
2412                         }
2413                 }
2414
2415                 name = si->lpp->vars[spill->spill];
2416                 if(!is_zero(name->value)) {
2417                         if(spill->reg_in > 0) {
2418                                 name = si->lpp->vars[spill->reg_in];
2419                                 if(!is_zero(name->value)) {
2420                                         insert_spill(si, spill->irn, spill->irn, bb);
2421                                         continue;
2422                                 }
2423                         }
2424                         pset_insert_ptr(spills_to_do, spill->irn);
2425                 }
2426         }
2427         DBG((si->dbg, LEVEL_3, "\t  %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
2428
2429
2430         for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
2431                 op_t     *op = get_irn_link(irn);
2432
2433                 if(be_is_Spill(irn)) continue;
2434
2435                 if(op->is_remat) {
2436                         /* TODO fix this if we want to support remats with more than two nodes */
2437                         if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
2438                                 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
2439
2440                                 insert_spill(si, irn, op->attr.remat.remat->value, irn);
2441                         }
2442                 } else {
2443                         if(pset_find_ptr(spills_to_do, irn)) {
2444                                 pset_remove_ptr(spills_to_do, irn);
2445
2446                                 insert_spill(si, irn, irn, irn);
2447                         }
2448                 }
2449
2450         }
2451
2452         assert(pset_count(spills_to_do) == 0);
2453
2454         /* afterwards free data in block */
2455         del_pset(spills_to_do);
2456 }
2457
2458 static void
2459 walker_reload_placer(ir_node * bb, void * data) {
2460         spill_ilp_t   *si = (spill_ilp_t*)data;
2461         ir_node       *irn;
2462         spill_bb_t    *spill_bb = get_irn_link(bb);
2463         int            i;
2464         irn_live_t    *li;
2465
2466         sched_foreach_reverse(bb, irn) {
2467                 op_t     *op = get_irn_link(irn);
2468
2469                 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
2470                 if(is_Phi(irn)) break;
2471
2472                 if(op->is_remat) {
2473                         if(get_irn_mode(irn) != mode_T) {
2474                                 insert_remat(si, irn);
2475                         }
2476                 } else {
2477                         int    n;
2478
2479                         for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2480                                 ir_node    *arg = get_irn_n(irn, i);
2481
2482                                 if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) {
2483                                         lpp_name_t    *name;
2484
2485                                         name = si->lpp->vars[op->attr.live_range.reloads[i]];
2486                                         if(!is_zero(name->value)) {
2487                                                 ir_node    *reload;
2488                                                 ir_node    *insert_pos = irn;
2489                                                 ir_node    *prev = sched_prev(insert_pos);
2490                                                 op_t       *prev_op = get_irn_link(prev);
2491
2492                                                 /* insert reload before pre-remats */
2493                                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2494                                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
2495                                                         insert_pos = prev;
2496
2497                                                         prev = sched_prev(insert_pos);
2498                                                         prev_op = get_irn_link(prev);
2499                                                 }
2500
2501                                                 reload = insert_reload(si, arg, insert_pos);
2502
2503                                                 set_irn_n(irn, i, reload);
2504
2505
2506 #ifdef KEEPALIVE_SPILLS
2507                                                 keep_alive(reload);
2508 #endif
2509                                         }
2510                                 }
2511                         }
2512                 }
2513         }
2514
2515         i=0;
2516         live_foreach(bb, li) {
2517                 ir_node        *irn = (ir_node *) li->irn;
2518
2519                 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
2520                         lpp_name_t    *name;
2521
2522                         name = si->lpp->vars[spill_bb->reloads[i]];
2523                         if(!is_zero(name->value)) {
2524                                 ir_node    *reload;
2525                                 ir_node    *insert_pos = bb;
2526                                 ir_node    *prev = sched_prev(insert_pos);
2527                                 op_t       *prev_op = get_irn_link(prev);
2528
2529                                 /* insert reload before pre-remats */
2530                                 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2531                                                 && prev_op->is_remat && prev_op->attr.remat.pre) {
2532                                         insert_pos = prev;
2533
2534                                         prev = sched_prev(insert_pos);
2535                                         prev_op = get_irn_link(prev);
2536                                 }
2537
2538                                 reload = insert_reload(si, irn, insert_pos);
2539
2540 #ifdef KEEPALIVE_SPILLS
2541                                 keep_alive(reload);
2542 #endif
2543                         }
2544                         ++i;
2545                 }
2546         }
2547
2548         del_set(spill_bb->ilp);
2549 }
2550
2551 static void
2552 walker_collect_used(ir_node * irn, void * data)
2553 {
2554         lc_bitset_t   *used = data;
2555
2556         lc_bitset_set(used, get_irn_idx(irn));
2557 }
2558
2559 static void
2560 walker_kill_unused(ir_node * bb, void * data)
2561 {
2562         lc_bitset_t     *used = data;
2563         const ir_node   *bad = get_irg_bad(get_irn_irg(bb));
2564         ir_node         *irn;
2565
2566
2567         for(irn=sched_first(bb); !sched_is_end(irn);) {
2568                 ir_node     *next = sched_next(irn);
2569                 int          i,
2570                                          n;
2571
2572                 if(!lc_bitset_is_set(used, get_irn_idx(irn))) {
2573                         assert(!be_is_Spill(irn) && !be_is_Reload(irn) && "something is fishy, spill or remat is unused");
2574
2575                         sched_remove(irn);
2576
2577                         set_nodes_block(irn, bad);
2578                         for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2579                                 ir_node             *arg = get_irn_n(irn, i);
2580
2581                                 set_irn_n(irn, i, bad);
2582                         }
2583                 }
2584                 irn = next;
2585         }
2586 }
2587
2588 static void
2589 kill_all_unused_values_in_schedule(spill_ilp_t * si)
2590 {
2591         lc_bitset_t   *used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
2592
2593         irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, used);
2594         irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, used);
2595
2596         lc_bitset_free(used);
2597 }
2598
2599 static void
2600 print_irn_pset(pset * p)
2601 {
2602         ir_node   *irn;
2603
2604         pset_foreach(p, irn) {
2605                 ir_printf("%+F\n", irn);
2606         }
2607 }
2608
2609 static void
2610 rewire_uses(spill_ilp_t * si)
2611 {
2612         dom_front_info_t     *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
2613         defs_t               *defs;
2614
2615         /* then fix uses of spills */
2616         set_foreach(si->values, defs) {
2617                 pset     *reloads;
2618                 pset     *spills;
2619                 ir_node  *next = defs->remats;
2620                 int remats = 0;
2621
2622                 if(next) {
2623                         reloads = pset_new_ptr_default();
2624
2625                         while(next) {
2626                                 if(be_is_Reload(next)) {
2627                                         pset_insert_ptr(reloads, next);
2628                                 } else {
2629                                         ++remats;
2630                                 }
2631                                 next = get_irn_link(next);
2632                         }
2633
2634                         spills = get_spills_for_value(si, defs->value);
2635                         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));
2636                         if(pset_count(spills) > 1) {
2637                                 assert(pset_count(reloads) > 0);
2638 //                              print_irn_pset(spills);
2639 //                              print_irn_pset(reloads);
2640                                 be_ssa_constr_set_uses(dfi, spills, reloads);
2641                         }
2642
2643                         del_pset(reloads);
2644                         del_pset(spills);
2645                 }
2646         }
2647
2648         /* first fix uses of remats and reloads */
2649         set_foreach(si->values, defs) {
2650                 pset     *nodes;
2651                 ir_node  *next = defs->remats;
2652
2653                 if(next) {
2654                         nodes = pset_new_ptr_default();
2655                         pset_insert_ptr(nodes, defs->value);
2656
2657                         while(next) {
2658                                 pset_insert_ptr(nodes, next);
2659                                 next = get_irn_link(next);
2660                         }
2661
2662                         if(pset_count(nodes) > 1) {
2663                                 DBG((si->dbg, LEVEL_4, "\t    %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
2664                                 be_ssa_constr_set(dfi, nodes);
2665                         }
2666
2667                         del_pset(nodes);
2668                 }
2669         }
2670
2671 //      remove_unused_defs(si);
2672
2673         be_free_dominance_frontiers(dfi);
2674 }
2675
2676 static void
2677 writeback_results(spill_ilp_t * si)
2678 {
2679         /* walk through the graph and collect all spills, reloads and remats for a value */
2680
2681         si->values = new_set(cmp_defs, 4096);
2682
2683         DBG((si->dbg, LEVEL_1, "Applying results\n"));
2684         delete_unnecessary_remats(si);
2685         irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
2686         irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
2687
2688         /* clean the remat info! there are still back-edges leading there! */
2689         clean_remat_info(si);
2690
2691         rewire_uses(si);
2692
2693         del_set(si->values);
2694 }
2695
2696 static int
2697 get_n_regs(spill_ilp_t * si)
2698 {
2699         int     arch_n_regs = arch_register_class_n_regs(si->cls);
2700         int     free = 0;
2701         int     i;
2702
2703         for(i=0; i<arch_n_regs; i++) {
2704                 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
2705                         free++;
2706                 }
2707         }
2708
2709         DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
2710         return free;
2711 }
2712
2713 #if 0
2714 #define  enqueue(r) if(reloads[ins_pos]) put(); \
2715                         assert(reloads[ins_pos]==NULL); \
2716                     reload[ins_pos] = (r); \
2717                     ins_pos = (ins_pos+1)%si->n_regs; \
2718                     ++count
2719 #define  put() if(reloads[del_pos]) sched_put_before(irn, reloads[del_pos]); \
2720                reloads[del_pos] = NULL; \
2721                del_pos = (del_pos+1)%si->n_regs; \
2722                --count;
2723 #endif
2724 static void
2725 walker_reload_mover(ir_node * bb, void * data)
2726 {
2727         spill_ilp_t   *si = data;
2728         ir_node           *irn;
2729
2730         sched_foreach(bb, irn) {
2731                 if(be_is_Reload(irn) && has_reg_class(si, irn)) {
2732                         ir_node       *reload = irn;
2733                         ir_node       *irn = irn;
2734
2735                         /* move reload upwards */
2736
2737                         int pressure = (int)get_irn_link(reload);
2738                         if(pressure < si->n_regs) {
2739                                 irn = sched_prev(reload);
2740                                 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
2741                                 sched_remove(reload);
2742                                 pressure = (int)get_irn_link(irn);
2743
2744                                 while(pressure < si->n_regs) {
2745                                         if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
2746
2747                                         set_irn_link(irn, (void*)(pressure+1));
2748                                         DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
2749                                         irn = sched_prev(irn);
2750
2751                                         pressure = (int)get_irn_link(irn);
2752                                 }
2753
2754                                 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
2755                                 sched_put_after(irn, reload);
2756 //                              sched_add_after(irn, reload);
2757 //                              pressure = (int)get_irn_link(sched_next(reload));
2758 //                              set_irn_link(reload, (void*)(pressure-1));
2759                         }
2760                 }
2761         }
2762 }
2763
2764 static void
2765 move_reloads_upward(spill_ilp_t * si)
2766 {
2767         irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
2768 }
2769
2770 void
2771 be_spill_remat(const be_chordal_env_t * chordal_env)
2772 {
2773         char            problem_name[256];
2774         char            dump_suffix[256];
2775         char            dump_suffix2[256];
2776         char            dump_suffix3[256];
2777         struct obstack  obst;
2778         spill_ilp_t     si;
2779
2780         ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
2781         ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
2782         ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
2783         ir_snprintf(dump_suffix2, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name);
2784
2785         FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
2786         DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
2787
2788         obstack_init(&obst);
2789         si.chordal_env = chordal_env;
2790         si.obst = &obst;
2791         si.senv = be_new_spill_env(chordal_env, is_mem_phi, &si);
2792         si.cls = chordal_env->cls;
2793         si.lpp = new_lpp(problem_name, lpp_minimize);
2794         si.remat_info = new_set(cmp_remat_info, 4096);
2795         si.all_possible_remats = pset_new_ptr_default();
2796 #ifndef EXECFREQ_LOOPDEPH
2797         si.execfreqs = compute_execfreq(chordal_env->irg);
2798 #endif
2799 #ifdef KEEPALIVE
2800         si.keep = NULL;
2801 #endif
2802         si.n_regs = get_n_regs(&si);
2803
2804         set_irg_link(chordal_env->irg, &si);
2805         compute_doms(chordal_env->irg);
2806
2807 #ifdef COLLECT_REMATS
2808 //      irg_block_walk_graph(chordal_env->irg, process_block, NULL, &si);
2809         /* collect remats */
2810         DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
2811         irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
2812 #endif
2813
2814         /* insert possible remats */
2815         DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
2816         irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
2817         DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
2818
2819 #ifdef KEEPALIVE
2820         DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
2821         connect_all_remats_with_keep(&si);
2822         /* dump graph with inserted remats */
2823         dump_graph_with_remats(chordal_env->irg, dump_suffix);
2824 #endif
2825
2826
2827         /* recompute liveness */
2828         DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
2829         be_liveness(chordal_env->irg);
2830
2831         /* build the ILP */
2832
2833         DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
2834         DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
2835         irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
2836
2837         DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
2838         irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
2839
2840 #ifdef DUMP_ILP
2841         {
2842                 FILE           *f;
2843                 char            buf[256];
2844
2845                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
2846                 if ((f = fopen(buf, "wt")) != NULL) {
2847                         lpp_dump_plain(si.lpp, f);
2848                         fclose(f);
2849                 }
2850         }
2851 #endif
2852
2853 #ifdef SOLVE
2854         DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
2855         lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
2856
2857 #ifdef SOLVE_LOCAL
2858         lpp_solve_cplex(si.lpp);
2859 #else
2860         lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
2861 #endif
2862         assert(lpp_is_sol_valid(si.lpp)
2863                && "solution of ILP must be valid");
2864
2865         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, si.lpp->objval));
2866
2867 #ifdef DUMP_SOLUTION
2868         {
2869                 FILE           *f;
2870                 char            buf[256];
2871
2872                 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
2873                 if ((f = fopen(buf, "wt")) != NULL) {
2874                         int             i;
2875                         for (i = 0; i < si.lpp->var_next; ++i) {
2876                                 lpp_name_t     *name = si.lpp->vars[i];
2877                                 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
2878                         }
2879                         fclose(f);
2880                 }
2881         }
2882 #endif
2883
2884         writeback_results(&si);
2885
2886 #endif                          /* SOLVE */
2887
2888         kill_all_unused_values_in_schedule(&si);
2889
2890         be_liveness(chordal_env->irg);
2891         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
2892
2893         dump_pressure_graph(&si, dump_suffix2);
2894
2895         // TODO fix temporarily exceeded regpressure due to remat2s
2896
2897         // TODO insert copys to fix interferences in memory
2898
2899         // move reloads upwards
2900         move_reloads_upward(&si);
2901         irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
2902         dump_pressure_graph(&si, dump_suffix3);
2903
2904         free_dom(chordal_env->irg);
2905         del_pset(si.all_possible_remats);
2906 #ifndef EXECFREQ_LOOPDEPH
2907         del_set(si.execfreqs);
2908 #endif
2909         free_lpp(si.lpp);
2910         obstack_free(&obst, NULL);
2911 //      exit(0);
2912 }
2913
2914 #else                           /* WITH_ILP */
2915
2916 static void
2917 only_that_you_can_compile_without_WITH_ILP_defined(void)
2918 {
2919 }
2920
2921 #endif                          /* WITH_ILP */