nsz Git - libfirm/blob - ir/opt/opt_inline.c

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief    Dead node elimination and Procedure Inlining.
  23  * @author   Michael Beck, Goetz Lindenmaier
  24  * @version  $Id$
  25  */
  26 #ifdef HAVE_CONFIG_H
  27 # include "config.h"
  28 #endif
  29
  30 #include <limits.h>
  31 #include <assert.h>
  32
  33 #include "irnode_t.h"
  34 #include "irgraph_t.h"
  35 #include "irprog_t.h"
  36
  37 #include "iroptimize.h"
  38 #include "ircons_t.h"
  39 #include "iropt_t.h"
  40 #include "irgopt.h"
  41 #include "irgmod.h"
  42 #include "irgwalk.h"
  43
  44 #include "adt/array.h"
  45 #include "adt/pset.h"
  46 #include "adt/pmap.h"
  47 #include "adt/pdeq.h"
  48 #include "adt/xmalloc.h"
  49
  50 #include "irouts.h"
  51 #include "irloop_t.h"
  52 #include "irbackedge_t.h"
  53 #include "opt_inline_t.h"
  54 #include "cgana.h"
  55 #include "trouts.h"
  56 #include "error.h"
  57
  58 #include "analyze_irg_args.h"
  59 #include "iredges_t.h"
  60 #include "irflag_t.h"
  61 #include "irhooks.h"
  62 #include "irtools.h"
  63
  64 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  65
  66 /*------------------------------------------------------------------*/
  67 /* Routines for dead node elimination / copying garbage collection  */
  68 /* of the obstack.                                                  */
  69 /*------------------------------------------------------------------*/
  70
  71 /**
  72  * Remember the new node in the old node by using a field all nodes have.
  73  */
  74 #define set_new_node(oldn, newn)  set_irn_link(oldn, newn)
  75
  76 /**
  77  * Get this new node, before the old node is forgotten.
  78  */
  79 #define get_new_node(oldn) get_irn_link(oldn)
  80
  81 /**
  82  * Check if a new node was set.
  83  */
  84 #define has_new_node(n) (get_new_node(n) != NULL)
  85
  86 /**
  87  * We use the block_visited flag to mark that we have computed the
  88  * number of useful predecessors for this block.
  89  * Further we encode the new arity in this flag in the old blocks.
  90  * Remembering the arity is useful, as it saves a lot of pointer
  91  * accesses.  This function is called for all Phi and Block nodes
  92  * in a Block.
  93  */
  94 static INLINE int
  95 compute_new_arity(ir_node *b) {
  96         int i, res, irn_arity;
  97         int irg_v, block_v;
  98
  99         irg_v = get_irg_block_visited(current_ir_graph);
 100         block_v = get_Block_block_visited(b);
 101         if (block_v >= irg_v) {
 102                 /* we computed the number of preds for this block and saved it in the
 103                    block_v flag */
 104                 return block_v - irg_v;
 105         } else {
 106                 /* compute the number of good predecessors */
 107                 res = irn_arity = get_irn_arity(b);
 108                 for (i = 0; i < irn_arity; i++)
 109                         if (is_Bad(get_irn_n(b, i))) res--;
 110                         /* save it in the flag. */
 111                         set_Block_block_visited(b, irg_v + res);
 112                         return res;
 113         }
 114 }
 115
 116 /**
 117  * Copies the node to the new obstack. The Ins of the new node point to
 118  * the predecessors on the old obstack.  For block/phi nodes not all
 119  * predecessors might be copied.  n->link points to the new node.
 120  * For Phi and Block nodes the function allocates in-arrays with an arity
 121  * only for useful predecessors.  The arity is determined by counting
 122  * the non-bad predecessors of the block.
 123  *
 124  * @param n    The node to be copied
 125  * @param env  if non-NULL, the node number attribute will be copied to the new node
 126  *
 127  * Note: Also used for loop unrolling.
 128  */
 129 static void copy_node(ir_node *n, void *env) {
 130         ir_node *nn, *block;
 131         int new_arity;
 132         ir_op *op = get_irn_op(n);
 133         (void) env;
 134
 135         if (op == op_Bad) {
 136                 /* node copied already */
 137                 return;
 138         } else if (op == op_Block) {
 139                 block = NULL;
 140                 new_arity = compute_new_arity(n);
 141                 n->attr.block.graph_arr = NULL;
 142         } else {
 143                 block = get_nodes_block(n);
 144                 if (op == op_Phi) {
 145                         new_arity = compute_new_arity(block);
 146                 } else {
 147                         new_arity = get_irn_arity(n);
 148                 }
 149         }
 150         nn = new_ir_node(get_irn_dbg_info(n),
 151                 current_ir_graph,
 152                 block,
 153                 op,
 154                 get_irn_mode(n),
 155                 new_arity,
 156                 get_irn_in(n) + 1);
 157         /* Copy the attributes.  These might point to additional data.  If this
 158            was allocated on the old obstack the pointers now are dangling.  This
 159            frees e.g. the memory of the graph_arr allocated in new_immBlock. */
 160         if (op == op_Block) {
 161                 /* we cannot allow blocks WITHOUT macroblock input */
 162                 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
 163         }
 164         copy_node_attr(n, nn);
 165
 166 #ifdef DEBUG_libfirm
 167         {
 168                 int copy_node_nr = env != NULL;
 169                 if (copy_node_nr) {
 170                         /* for easier debugging, we want to copy the node numbers too */
 171                         nn->node_nr = n->node_nr;
 172                 }
 173         }
 174 #endif
 175
 176         set_new_node(n, nn);
 177         hook_dead_node_elim_subst(current_ir_graph, n, nn);
 178 }
 179
 180 /**
 181  * Copies new predecessors of old node to new node remembered in link.
 182  * Spare the Bad predecessors of Phi and Block nodes.
 183  */
 184 static void copy_preds(ir_node *n, void *env) {
 185         ir_node *nn, *block;
 186         int i, j, irn_arity;
 187         (void) env;
 188
 189         nn = get_new_node(n);
 190
 191         if (is_Block(n)) {
 192                 /* copy the macro block header */
 193                 ir_node *mbh = get_Block_MacroBlock(n);
 194
 195                 if (mbh == n) {
 196                         /* this block is a macroblock header */
 197                         set_Block_MacroBlock(nn, nn);
 198                 } else {
 199                         /* get the macro block header */
 200                         ir_node *nmbh = get_new_node(mbh);
 201                         assert(nmbh != NULL);
 202                         set_Block_MacroBlock(nn, nmbh);
 203                 }
 204
 205                 /* Don't copy Bad nodes. */
 206                 j = 0;
 207                 irn_arity = get_irn_arity(n);
 208                 for (i = 0; i < irn_arity; i++) {
 209                         if (! is_Bad(get_irn_n(n, i))) {
 210                                 ir_node *pred = get_irn_n(n, i);
 211                                 set_irn_n(nn, j, get_new_node(pred));
 212                                 j++;
 213                         }
 214                 }
 215                 /* repair the block visited flag from above misuse. Repair it in both
 216                    graphs so that the old one can still be used. */
 217                 set_Block_block_visited(nn, 0);
 218                 set_Block_block_visited(n, 0);
 219                 /* Local optimization could not merge two subsequent blocks if
 220                    in array contained Bads.  Now it's possible.
 221                    We don't call optimize_in_place as it requires
 222                    that the fields in ir_graph are set properly. */
 223                 if ((get_opt_control_flow_straightening()) &&
 224                         (get_Block_n_cfgpreds(nn) == 1) &&
 225                         is_Jmp(get_Block_cfgpred(nn, 0))) {
 226                         ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
 227                         if (nn == old) {
 228                                 /* Jmp jumps into the block it is in -- deal self cycle. */
 229                                 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
 230                                 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
 231                         } else {
 232                                 exchange(nn, old);
 233                         }
 234                 }
 235         } else if (is_Phi(n) && get_irn_arity(n) > 0) {
 236                 /* Don't copy node if corresponding predecessor in block is Bad.
 237                    The Block itself should not be Bad. */
 238                 block = get_nodes_block(n);
 239                 set_nodes_block(nn, get_new_node(block));
 240                 j = 0;
 241                 irn_arity = get_irn_arity(n);
 242                 for (i = 0; i < irn_arity; i++) {
 243                         if (! is_Bad(get_irn_n(block, i))) {
 244                                 ir_node *pred = get_irn_n(n, i);
 245                                 set_irn_n(nn, j, get_new_node(pred));
 246                                 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
 247                                 j++;
 248                         }
 249                 }
 250                 /* If the pre walker reached this Phi after the post walker visited the
 251                    block block_visited is > 0. */
 252                 set_Block_block_visited(get_nodes_block(n), 0);
 253                 /* Compacting the Phi's ins might generate Phis with only one
 254                    predecessor. */
 255                 if (get_irn_arity(nn) == 1)
 256                         exchange(nn, get_irn_n(nn, 0));
 257         } else {
 258                 irn_arity = get_irn_arity(n);
 259                 for (i = -1; i < irn_arity; i++)
 260                         set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
 261         }
 262         /* Now the new node is complete.  We can add it to the hash table for CSE.
 263            @@@ inlining aborts if we identify End. Why? */
 264         if (!is_End(nn))
 265                 add_identities(current_ir_graph->value_table, nn);
 266 }
 267
 268 /**
 269  * Copies the graph recursively, compacts the keep-alives of the end node.
 270  *
 271  * @param irg           the graph to be copied
 272  * @param copy_node_nr  If non-zero, the node number will be copied
 273  */
 274 static void copy_graph(ir_graph *irg, int copy_node_nr) {
 275         ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
 276         ir_node *ka;      /* keep alive */
 277         int i, irn_arity;
 278         unsigned long vfl;
 279
 280         /* Some nodes must be copied by hand, sigh */
 281         vfl = get_irg_visited(irg);
 282         set_irg_visited(irg, vfl + 1);
 283
 284         oe = get_irg_end(irg);
 285         mark_irn_visited(oe);
 286         /* copy the end node by hand, allocate dynamic in array! */
 287         ne = new_ir_node(get_irn_dbg_info(oe),
 288                 irg,
 289                 NULL,
 290                 op_End,
 291                 mode_X,
 292                 -1,
 293                 NULL);
 294         /* Copy the attributes.  Well, there might be some in the future... */
 295         copy_node_attr(oe, ne);
 296         set_new_node(oe, ne);
 297
 298         /* copy the Bad node */
 299         ob = get_irg_bad(irg);
 300         mark_irn_visited(ob);
 301         nb = new_ir_node(get_irn_dbg_info(ob),
 302                 irg,
 303                 NULL,
 304                 op_Bad,
 305                 mode_T,
 306                 0,
 307                 NULL);
 308         copy_node_attr(ob, nb);
 309         set_new_node(ob, nb);
 310
 311         /* copy the NoMem node */
 312         om = get_irg_no_mem(irg);
 313         mark_irn_visited(om);
 314         nm = new_ir_node(get_irn_dbg_info(om),
 315                 irg,
 316                 NULL,
 317                 op_NoMem,
 318                 mode_M,
 319                 0,
 320                 NULL);
 321         copy_node_attr(om, nm);
 322         set_new_node(om, nm);
 323
 324         /* copy the live nodes */
 325         set_irg_visited(irg, vfl);
 326         irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 327
 328         /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
 329
 330         /* visit the anchors as well */
 331         for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
 332                 ir_node *n = get_irg_anchor(irg, i);
 333
 334                 if (n && (get_irn_visited(n) <= vfl)) {
 335                         set_irg_visited(irg, vfl);
 336                         irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 337                 }
 338         }
 339
 340         /* copy_preds for the end node ... */
 341         set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
 342
 343         /*- ... and now the keep alives. -*/
 344         /* First pick the not marked block nodes and walk them.  We must pick these
 345            first as else we will oversee blocks reachable from Phis. */
 346         irn_arity = get_End_n_keepalives(oe);
 347         for (i = 0; i < irn_arity; i++) {
 348                 ka = get_End_keepalive(oe, i);
 349                 if (is_Block(ka)) {
 350                         if (get_irn_visited(ka) <= vfl) {
 351                                 /* We must keep the block alive and copy everything reachable */
 352                                 set_irg_visited(irg, vfl);
 353                                 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 354                         }
 355                         add_End_keepalive(ne, get_new_node(ka));
 356                 }
 357         }
 358
 359         /* Now pick other nodes.  Here we will keep all! */
 360         irn_arity = get_End_n_keepalives(oe);
 361         for (i = 0; i < irn_arity; i++) {
 362                 ka = get_End_keepalive(oe, i);
 363                 if (!is_Block(ka)) {
 364                         if (get_irn_visited(ka) <= vfl) {
 365                                 /* We didn't copy the node yet.  */
 366                                 set_irg_visited(irg, vfl);
 367                                 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 368                         }
 369                         add_End_keepalive(ne, get_new_node(ka));
 370                 }
 371         }
 372
 373         /* start block sometimes only reached after keep alives */
 374         set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
 375         set_nodes_block(nm, get_new_node(get_nodes_block(om)));
 376 }
 377
 378 /**
 379  * Copies the graph reachable from current_ir_graph->end to the obstack
 380  * in current_ir_graph and fixes the environment.
 381  * Then fixes the fields in current_ir_graph containing nodes of the
 382  * graph.
 383  *
 384  * @param copy_node_nr  If non-zero, the node number will be copied
 385  */
 386 static void
 387 copy_graph_env(int copy_node_nr) {
 388         ir_graph *irg = current_ir_graph;
 389         ir_node *old_end, *new_anchor;
 390         int i;
 391
 392         /* remove end_except and end_reg nodes */
 393         old_end = get_irg_end(irg);
 394         set_irg_end_except (irg, old_end);
 395         set_irg_end_reg    (irg, old_end);
 396
 397         /* Not all nodes remembered in irg might be reachable
 398            from the end node.  Assure their link is set to NULL, so that
 399            we can test whether new nodes have been computed. */
 400         for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
 401                 ir_node *n = get_irg_anchor(irg, i);
 402                 if (n != NULL)
 403                         set_new_node(n, NULL);
 404         }
 405         /* we use the block walk flag for removing Bads from Blocks ins. */
 406         inc_irg_block_visited(irg);
 407
 408         /* copy the graph */
 409         copy_graph(irg, copy_node_nr);
 410
 411         /* fix the anchor */
 412         old_end    = get_irg_end(irg);
 413         new_anchor = new_Anchor(irg);
 414
 415         for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
 416                 ir_node *n = get_irg_anchor(irg, i);
 417                 if (n)
 418                         set_irn_n(new_anchor, i, get_new_node(n));
 419         }
 420         free_End(old_end);
 421         irg->anchor = new_anchor;
 422
 423         /* ensure the new anchor is placed in the endblock */
 424         set_nodes_block(new_anchor, get_irg_end_block(irg));
 425 }
 426
 427 /**
 428  * Copies all reachable nodes to a new obstack.  Removes bad inputs
 429  * from block nodes and the corresponding inputs from Phi nodes.
 430  * Merges single exit blocks with single entry blocks and removes
 431  * 1-input Phis.
 432  * Adds all new nodes to a new hash table for CSE.  Does not
 433  * perform CSE, so the hash table might contain common subexpressions.
 434  */
 435 void dead_node_elimination(ir_graph *irg) {
 436         ir_graph *rem;
 437 #ifdef INTERPROCEDURAL_VIEW
 438         int rem_ipview = get_interprocedural_view();
 439 #endif
 440         struct obstack *graveyard_obst = NULL;
 441         struct obstack *rebirth_obst   = NULL;
 442         assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
 443
 444         /* inform statistics that we started a dead-node elimination run */
 445         hook_dead_node_elim(irg, 1);
 446
 447         /* Remember external state of current_ir_graph. */
 448         rem = current_ir_graph;
 449         current_ir_graph = irg;
 450 #ifdef INTERPROCEDURAL_VIEW
 451         set_interprocedural_view(0);
 452 #endif
 453
 454         assert(get_irg_phase_state(irg) != phase_building);
 455
 456         /* Handle graph state */
 457         free_callee_info(irg);
 458         free_irg_outs(irg);
 459         free_trouts();
 460
 461         /* @@@ so far we loose loops when copying */
 462         free_loop_information(irg);
 463
 464         set_irg_doms_inconsistent(irg);
 465
 466         /* A quiet place, where the old obstack can rest in peace,
 467            until it will be cremated. */
 468         graveyard_obst = irg->obst;
 469
 470         /* A new obstack, where the reachable nodes will be copied to. */
 471         rebirth_obst = xmalloc(sizeof(*rebirth_obst));
 472         irg->obst = rebirth_obst;
 473         obstack_init(irg->obst);
 474         irg->last_node_idx = 0;
 475
 476         /* We also need a new value table for CSE */
 477         del_identities(irg->value_table);
 478         irg->value_table = new_identities();
 479
 480         /* Copy the graph from the old to the new obstack */
 481         copy_graph_env(/*copy_node_nr=*/1);
 482
 483         /* Free memory from old unoptimized obstack */
 484         obstack_free(graveyard_obst, 0);  /* First empty the obstack ... */
 485         xfree(graveyard_obst);            /* ... then free it.           */
 486
 487         /* inform statistics that the run is over */
 488         hook_dead_node_elim(irg, 0);
 489
 490         current_ir_graph = rem;
 491 #ifdef INTERPROCEDURAL_VIEW
 492         set_interprocedural_view(rem_ipview);
 493 #endif
 494 }
 495
 496 /**
 497  * Relink bad predecessors of a block and store the old in array to the
 498  * link field. This function is called by relink_bad_predecessors().
 499  * The array of link field starts with the block operand at position 0.
 500  * If block has bad predecessors, create a new in array without bad preds.
 501  * Otherwise let in array untouched.
 502  */
 503 static void relink_bad_block_predecessors(ir_node *n, void *env) {
 504         ir_node **new_in, *irn;
 505         int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
 506         (void) env;
 507
 508         /* if link field of block is NULL, look for bad predecessors otherwise
 509            this is already done */
 510         if (is_Block(n) && get_irn_link(n) == NULL) {
 511                 /* save old predecessors in link field (position 0 is the block operand)*/
 512                 set_irn_link(n, get_irn_in(n));
 513
 514                 /* count predecessors without bad nodes */
 515                 old_irn_arity = get_irn_arity(n);
 516                 for (i = 0; i < old_irn_arity; i++)
 517                         if (!is_Bad(get_irn_n(n, i)))
 518                                 ++new_irn_arity;
 519
 520                 /* arity changing: set new predecessors without bad nodes */
 521                 if (new_irn_arity < old_irn_arity) {
 522                         /* Get new predecessor array. We do not resize the array, as we must
 523                            keep the old one to update Phis. */
 524                         new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
 525
 526                         /* set new predecessors in array */
 527                         new_in[0] = NULL;
 528                         new_irn_n = 1;
 529                         for (i = 0; i < old_irn_arity; i++) {
 530                                 irn = get_irn_n(n, i);
 531                                 if (!is_Bad(irn)) {
 532                                         new_in[new_irn_n] = irn;
 533                                         is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
 534                                         ++new_irn_n;
 535                                 }
 536                         }
 537                         /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
 538                         ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
 539                         n->in = new_in;
 540                 } /* ir node has bad predecessors */
 541         } /* Block is not relinked */
 542 }
 543
 544 /**
 545  * Relinks Bad predecessors from Blocks and Phis called by walker
 546  * remove_bad_predecesors(). If n is a Block, call
 547  * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
 548  * function of Phi's Block. If this block has bad predecessors, relink preds
 549  * of the Phi-node.
 550  */
 551 static void relink_bad_predecessors(ir_node *n, void *env) {
 552         ir_node *block, **old_in;
 553         int i, old_irn_arity, new_irn_arity;
 554
 555         /* relink bad predecessors of a block */
 556         if (is_Block(n))
 557                 relink_bad_block_predecessors(n, env);
 558
 559         /* If Phi node relink its block and its predecessors */
 560         if (is_Phi(n)) {
 561                 /* Relink predecessors of phi's block */
 562                 block = get_nodes_block(n);
 563                 if (get_irn_link(block) == NULL)
 564                         relink_bad_block_predecessors(block, env);
 565
 566                 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
 567                 old_irn_arity = ARR_LEN(old_in);
 568
 569                 /* Relink Phi predecessors if count of predecessors changed */
 570                 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
 571                         /* set new predecessors in array
 572                            n->in[0] remains the same block */
 573                         new_irn_arity = 1;
 574                         for(i = 1; i < old_irn_arity; i++)
 575                                 if (!is_Bad(old_in[i])) {
 576                                         n->in[new_irn_arity] = n->in[i];
 577                                         is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
 578                                         ++new_irn_arity;
 579                                 }
 580
 581                                 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
 582                                 ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
 583                 }
 584         } /* n is a Phi node */
 585 }
 586
 587 /*
 588  * Removes Bad Bad predecessors from Blocks and the corresponding
 589  * inputs to Phi nodes as in dead_node_elimination but without
 590  * copying the graph.
 591  * On walking up set the link field to NULL, on walking down call
 592  * relink_bad_predecessors() (This function stores the old in array
 593  * to the link field and sets a new in array if arity of predecessors
 594  * changes).
 595  */
 596 void remove_bad_predecessors(ir_graph *irg) {
 597         panic("Fix backedge handling first");
 598         irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
 599 }
 600
 601
 602 /*
 603    __                      _  __ __
 604   (_     __    o     _    | \/  |_
 605   __)|_| | \_/ | \_/(/_   |_/\__|__
 606
 607   The following stuff implements a facility that automatically patches
 608   registered ir_node pointers to the new node when a dead node elimination occurs.
 609 */
 610
 611 struct _survive_dce_t {
 612         struct obstack obst;
 613         pmap *places;
 614         pmap *new_places;
 615         hook_entry_t dead_node_elim;
 616         hook_entry_t dead_node_elim_subst;
 617 };
 618
 619 typedef struct _survive_dce_list_t {
 620         struct _survive_dce_list_t *next;
 621         ir_node **place;
 622 } survive_dce_list_t;
 623
 624 static void dead_node_hook(void *context, ir_graph *irg, int start) {
 625         survive_dce_t *sd = context;
 626         (void) irg;
 627
 628         /* Create a new map before the dead node elimination is performed. */
 629         if (start) {
 630                 sd->new_places = pmap_create_ex(pmap_count(sd->places));
 631         } else {
 632                 /* Patch back all nodes if dead node elimination is over and something is to be done. */
 633                 pmap_destroy(sd->places);
 634                 sd->places     = sd->new_places;
 635                 sd->new_places = NULL;
 636         }
 637 }
 638
 639 /**
 640  * Hook called when dead node elimination replaces old by nw.
 641  */
 642 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
 643         survive_dce_t *sd = context;
 644         survive_dce_list_t *list = pmap_get(sd->places, old);
 645         (void) irg;
 646
 647         /* If the node is to be patched back, write the new address to all registered locations. */
 648         if (list) {
 649                 survive_dce_list_t *p;
 650
 651                 for (p = list; p; p = p->next)
 652                         *(p->place) = nw;
 653
 654                 pmap_insert(sd->new_places, nw, list);
 655         }
 656 }
 657
 658 /**
 659  * Make a new Survive DCE environment.
 660  */
 661 survive_dce_t *new_survive_dce(void) {
 662         survive_dce_t *res = xmalloc(sizeof(res[0]));
 663         obstack_init(&res->obst);
 664         res->places     = pmap_create();
 665         res->new_places = NULL;
 666
 667         res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
 668         res->dead_node_elim.context                   = res;
 669         res->dead_node_elim.next                      = NULL;
 670
 671         res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
 672         res->dead_node_elim_subst.context = res;
 673         res->dead_node_elim_subst.next    = NULL;
 674
 675 #ifndef FIRM_ENABLE_HOOKS
 676         assert(0 && "need hooks enabled");
 677 #endif
 678
 679         register_hook(hook_dead_node_elim, &res->dead_node_elim);
 680         register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
 681         return res;
 682 }
 683
 684 /**
 685  * Free a Survive DCE environment.
 686  */
 687 void free_survive_dce(survive_dce_t *sd) {
 688         obstack_free(&sd->obst, NULL);
 689         pmap_destroy(sd->places);
 690         unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
 691         unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
 692         xfree(sd);
 693 }
 694
 695 /**
 696  * Register a node pointer to be patched upon DCE.
 697  * When DCE occurs, the node pointer specified by @p place will be
 698  * patched to the new address of the node it is pointing to.
 699  *
 700  * @param sd    The Survive DCE environment.
 701  * @param place The address of the node pointer.
 702  */
 703 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
 704         if (*place != NULL) {
 705                 ir_node *irn      = *place;
 706                 survive_dce_list_t *curr = pmap_get(sd->places, irn);
 707                 survive_dce_list_t *nw   = obstack_alloc(&sd->obst, sizeof(nw[0]));
 708
 709                 nw->next  = curr;
 710                 nw->place = place;
 711
 712                 pmap_insert(sd->places, irn, nw);
 713         }
 714 }
 715
 716 /*--------------------------------------------------------------------*/
 717 /*  Functionality for inlining                                         */
 718 /*--------------------------------------------------------------------*/
 719
 720 /**
 721  * Copy node for inlineing.  Updates attributes that change when
 722  * inlineing but not for dead node elimination.
 723  *
 724  * Copies the node by calling copy_node() and then updates the entity if
 725  * it's a local one.  env must be a pointer of the frame type of the
 726  * inlined procedure. The new entities must be in the link field of
 727  * the entities.
 728  */
 729 static INLINE void
 730 copy_node_inline(ir_node *n, void *env) {
 731         ir_node *nn;
 732         ir_type *frame_tp = (ir_type *)env;
 733
 734         copy_node(n, NULL);
 735         if (is_Sel(n)) {
 736                 nn = get_new_node (n);
 737                 assert(is_Sel(nn));
 738                 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
 739                         set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
 740                 }
 741         } else if (is_Block(n)) {
 742                 nn = get_new_node (n);
 743                 nn->attr.block.irg = current_ir_graph;
 744         }
 745 }
 746
 747 /**
 748  * Walker: checks if P_value_arg_base is used.
 749  */
 750 static void find_addr(ir_node *node, void *env) {
 751         int *allow_inline = env;
 752         if (is_Proj(node) &&
 753                         is_Start(get_Proj_pred(node)) &&
 754                         get_Proj_proj(node) == pn_Start_P_value_arg_base) {
 755                 *allow_inline = 0;
 756         } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
 757                 /* From GCC:
 758                  * Refuse to inline alloca call unless user explicitly forced so as this
 759                  * may change program's memory overhead drastically when the function
 760                  * using alloca is called in loop.  In GCC present in SPEC2000 inlining
 761                  * into schedule_block cause it to require 2GB of ram instead of 256MB.
 762                  *
 763                  * Sorryly this is true with our implementation also.
 764                  * Moreover, we cannot differentiate between alloca() and VLA yet, so this
 765                  * disables inlining of functions using VLA (with are completely save).
 766                  *
 767                  * 2 Solutions:
 768                  * - add a flag to the Alloc node for "real" alloca() calls
 769                  * - add a new Stack-Restore node at the end of a function using alloca()
 770                  */
 771                 *allow_inline = 0;
 772         }
 773 }
 774
 775 /**
 776  * Check if we can inline a given call.
 777  * Currently, we cannot inline two cases:
 778  * - call with compound arguments
 779  * - graphs that take the address of a parameter
 780  *
 781  * check these conditions here
 782  */
 783 static int can_inline(ir_node *call, ir_graph *called_graph) {
 784         ir_type *call_type = get_Call_type(call);
 785         int params, ress, i, res;
 786         assert(is_Method_type(call_type));
 787
 788         params = get_method_n_params(call_type);
 789         ress   = get_method_n_ress(call_type);
 790
 791         /* check parameters for compound arguments */
 792         for (i = 0; i < params; ++i) {
 793                 ir_type *p_type = get_method_param_type(call_type, i);
 794
 795                 if (is_compound_type(p_type))
 796                         return 0;
 797         }
 798
 799         /* check results for compound arguments */
 800         for (i = 0; i < ress; ++i) {
 801                 ir_type *r_type = get_method_res_type(call_type, i);
 802
 803                 if (is_compound_type(r_type))
 804                         return 0;
 805         }
 806
 807         res = 1;
 808         irg_walk_graph(called_graph, find_addr, NULL, &res);
 809
 810         return res;
 811 }
 812
 813 enum exc_mode {
 814         exc_handler    = 0, /**< There is a handler. */
 815         exc_to_end     = 1, /**< Branches to End. */
 816         exc_no_handler = 2  /**< Exception handling not represented. */
 817 };
 818
 819 /* Inlines a method at the given call site. */
 820 int inline_method(ir_node *call, ir_graph *called_graph) {
 821         ir_node             *pre_call;
 822         ir_node             *post_call, *post_bl;
 823         ir_node             *in[pn_Start_max];
 824         ir_node             *end, *end_bl;
 825         ir_node             **res_pred;
 826         ir_node             **cf_pred;
 827         ir_node             *ret, *phi;
 828         int                 arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity;
 829         enum exc_mode       exc_handling;
 830         ir_type             *called_frame, *curr_frame;
 831         ir_entity           *ent;
 832         ir_graph            *rem, *irg;
 833         irg_inline_property prop = get_irg_inline_property(called_graph);
 834
 835         if (prop == irg_inline_forbidden)
 836                 return 0;
 837
 838         ent = get_irg_entity(called_graph);
 839
 840         /* Do not inline variadic functions. */
 841         if (get_method_variadicity(get_entity_type(ent)) == variadicity_variadic) {
 842                 /* Arg, KR functions are marked as variadic one's, so check further */
 843                 ir_type *mtp     = get_entity_type(ent);
 844                 ir_type *ctp     = get_Call_type(call);
 845                 int     n_params = get_method_n_params(mtp);
 846                 int     i;
 847
 848                 /* This is too strong, but probably ok. Function calls with a wrong number of
 849                    parameters should not be inlined. */
 850                 if (n_params != get_method_n_params(ctp))
 851                         return 0;
 852
 853                 /* check types: for K&R calls, this was not done by the compiler. Again, this is
 854                    too strong, but ok for now. */
 855                 for (i = n_params - 1; i >= 0; --i) {
 856                         ir_type *param_tp = get_method_param_type(mtp, i);
 857                         ir_type *arg_tp   = get_method_param_type(ctp, i);
 858
 859                         if (param_tp != arg_tp)
 860                                 return 0;
 861                 }
 862                 DB((dbg, LEVEL_1, "Inlining allowed for variadic function %+F\n", called_graph));
 863                 /* types match, fine: when the frame is access, the inliner stops at can_inline() */
 864         }
 865
 866         assert(get_method_n_params(get_entity_type(ent)) ==
 867                get_method_n_params(get_Call_type(call)));
 868
 869         irg = get_irn_irg(call);
 870
 871         /*
 872          * We cannot inline a recursive call. The graph must be copied before
 873          * the call the inline_method() using create_irg_copy().
 874          */
 875         if (called_graph == irg)
 876                 return 0;
 877
 878         /*
 879          * currently, we cannot inline two cases:
 880          * - call with compound arguments
 881          * - graphs that take the address of a parameter
 882          */
 883         if (! can_inline(call, called_graph))
 884                 return 0;
 885
 886         rem = current_ir_graph;
 887         current_ir_graph = irg;
 888
 889         DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
 890
 891         /* --  Turn off optimizations, this can cause problems when allocating new nodes. -- */
 892         rem_opt = get_opt_optimize();
 893         set_optimize(0);
 894
 895         /* Handle graph state */
 896         assert(get_irg_phase_state(irg) != phase_building);
 897         assert(get_irg_pinned(irg) == op_pin_state_pinned);
 898         assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
 899         set_irg_outs_inconsistent(irg);
 900         set_irg_extblk_inconsistent(irg);
 901         set_irg_doms_inconsistent(irg);
 902         set_irg_loopinfo_inconsistent(irg);
 903         set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
 904
 905         /* -- Check preconditions -- */
 906         assert(is_Call(call));
 907
 908         /* here we know we WILL inline, so inform the statistics */
 909         hook_inline(call, called_graph);
 910
 911         /* -- Decide how to handle exception control flow: Is there a handler
 912            for the Call node, or do we branch directly to End on an exception?
 913            exc_handling:
 914            0 There is a handler.
 915            1 Branches to End.
 916            2 Exception handling not represented in Firm. -- */
 917         {
 918                 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
 919                 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
 920                         long proj_nr = get_Proj_proj(proj);
 921                         if (proj_nr == pn_Call_X_except) Xproj = proj;
 922                         if (proj_nr == pn_Call_M_except) Mproj = proj;
 923                 }
 924                 if      (Mproj) { assert(Xproj); exc_handling = exc_handler; } /*  Mproj           */
 925                 else if (Xproj) {                exc_handling = exc_to_end; } /* !Mproj &&  Xproj   */
 926                 else            {                exc_handling = exc_no_handler; } /* !Mproj && !Xproj   */
 927         }
 928
 929         /* --
 930            the procedure and later replaces the Start node of the called graph.
 931            Post_call is the old Call node and collects the results of the called
 932            graph. Both will end up being a tuple.  -- */
 933         post_bl = get_nodes_block(call);
 934         set_irg_current_block(irg, post_bl);
 935         /* XxMxPxPxPxT of Start + parameter of Call */
 936         in[pn_Start_X_initial_exec]   = new_Jmp();
 937         in[pn_Start_M]                = get_Call_mem(call);
 938         in[pn_Start_P_frame_base]     = get_irg_frame(irg);
 939         in[pn_Start_P_tls]            = get_irg_tls(irg);
 940         in[pn_Start_T_args]           = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
 941         /* in[pn_Start_P_value_arg_base] = ??? */
 942         assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
 943         pre_call = new_Tuple(pn_Start_max - 1, in);
 944         post_call = call;
 945
 946         /* --
 947            The new block gets the ins of the old block, pre_call and all its
 948            predecessors and all Phi nodes. -- */
 949         part_block(pre_call);
 950
 951         /* -- Prepare state for dead node elimination -- */
 952         /* Visited flags in calling irg must be >= flag in called irg.
 953            Else walker and arity computation will not work. */
 954         if (get_irg_visited(irg) <= get_irg_visited(called_graph))
 955                 set_irg_visited(irg, get_irg_visited(called_graph)+1);
 956         if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
 957                 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
 958         /* Set pre_call as new Start node in link field of the start node of
 959            calling graph and pre_calls block as new block for the start block
 960            of calling graph.
 961            Further mark these nodes so that they are not visited by the
 962            copying. */
 963         set_irn_link(get_irg_start(called_graph), pre_call);
 964         set_irn_visited(get_irg_start(called_graph), get_irg_visited(irg));
 965         set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
 966         set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(irg));
 967         set_irn_link(get_irg_bad(called_graph), get_irg_bad(irg));
 968         set_irn_visited(get_irg_bad(called_graph), get_irg_visited(irg));
 969
 970         /* Initialize for compaction of in arrays */
 971         inc_irg_block_visited(irg);
 972
 973         /* -- Replicate local entities of the called_graph -- */
 974         /* copy the entities. */
 975         called_frame = get_irg_frame_type(called_graph);
 976         curr_frame   = get_irg_frame_type(irg);
 977         for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
 978                 ir_entity *new_ent, *old_ent;
 979                 old_ent = get_class_member(called_frame, i);
 980                 new_ent = copy_entity_own(old_ent, curr_frame);
 981                 set_entity_link(old_ent, new_ent);
 982         }
 983
 984         /* visited is > than that of called graph.  With this trick visited will
 985            remain unchanged so that an outer walker, e.g., searching the call nodes
 986             to inline, calling this inline will not visit the inlined nodes. */
 987         set_irg_visited(irg, get_irg_visited(irg)-1);
 988
 989         /* -- Performing dead node elimination inlines the graph -- */
 990         /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
 991            entities. */
 992         irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
 993                  get_irg_frame_type(called_graph));
 994
 995         /* Repair called_graph */
 996         set_irg_visited(called_graph, get_irg_visited(irg));
 997         set_irg_block_visited(called_graph, get_irg_block_visited(irg));
 998         set_Block_block_visited(get_irg_start_block(called_graph), 0);
 999
1000         /* -- Merge the end of the inlined procedure with the call site -- */
1001         /* We will turn the old Call node into a Tuple with the following
1002            predecessors:
1003            -1:  Block of Tuple.
1004            0: Phi of all Memories of Return statements.
1005            1: Jmp from new Block that merges the control flow from all exception
1006            predecessors of the old end block.
1007            2: Tuple of all arguments.
1008            3: Phi of Exception memories.
1009            In case the old Call directly branches to End on an exception we don't
1010            need the block merging all exceptions nor the Phi of the exception
1011            memories.
1012         */
1013
1014         /* -- Precompute some values -- */
1015         end_bl = get_new_node(get_irg_end_block(called_graph));
1016         end = get_new_node(get_irg_end(called_graph));
1017         arity = get_irn_arity(end_bl);    /* arity = n_exc + n_ret  */
1018         n_res = get_method_n_ress(get_Call_type(call));
1019
1020         res_pred = xmalloc(n_res * sizeof(*res_pred));
1021         cf_pred  = xmalloc(arity * sizeof(*res_pred));
1022
1023         set_irg_current_block(irg, post_bl); /* just to make sure */
1024
1025         /* -- archive keepalives -- */
1026         irn_arity = get_irn_arity(end);
1027         for (i = 0; i < irn_arity; i++) {
1028                 ir_node *ka = get_End_keepalive(end, i);
1029                 if (! is_Bad(ka))
1030                         add_End_keepalive(get_irg_end(irg), ka);
1031         }
1032
1033         /* The new end node will die.  We need not free as the in array is on the obstack:
1034            copy_node() only generated 'D' arrays. */
1035
1036         /* -- Replace Return nodes by Jump nodes. -- */
1037         n_ret = 0;
1038         for (i = 0; i < arity; i++) {
1039                 ir_node *ret;
1040                 ret = get_irn_n(end_bl, i);
1041                 if (is_Return(ret)) {
1042                         cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1043                         n_ret++;
1044                 }
1045         }
1046         set_irn_in(post_bl, n_ret, cf_pred);
1047
1048         /* -- Build a Tuple for all results of the method.
1049            Add Phi node if there was more than one Return.  -- */
1050         turn_into_tuple(post_call, pn_Call_max);
1051         /* First the Memory-Phi */
1052         n_ret = 0;
1053         for (i = 0; i < arity; i++) {
1054                 ret = get_irn_n(end_bl, i);
1055                 if (is_Return(ret)) {
1056                         cf_pred[n_ret] = get_Return_mem(ret);
1057                         n_ret++;
1058                 }
1059         }
1060         phi = new_Phi(n_ret, cf_pred, mode_M);
1061         set_Tuple_pred(call, pn_Call_M_regular, phi);
1062         /* Conserve Phi-list for further inlinings -- but might be optimized */
1063         if (get_nodes_block(phi) == post_bl) {
1064                 set_irn_link(phi, get_irn_link(post_bl));
1065                 set_irn_link(post_bl, phi);
1066         }
1067         /* Now the real results */
1068         if (n_res > 0) {
1069                 for (j = 0; j < n_res; j++) {
1070                         n_ret = 0;
1071                         for (i = 0; i < arity; i++) {
1072                                 ret = get_irn_n(end_bl, i);
1073                                 if (is_Return(ret)) {
1074                                         cf_pred[n_ret] = get_Return_res(ret, j);
1075                                         n_ret++;
1076                                 }
1077                         }
1078                         if (n_ret > 0)
1079                                 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1080                         else
1081                                 phi = new_Bad();
1082                         res_pred[j] = phi;
1083                         /* Conserve Phi-list for further inlinings -- but might be optimized */
1084                         if (get_nodes_block(phi) == post_bl) {
1085                                 set_Phi_next(phi, get_Block_phis(post_bl));
1086                                 set_Block_phis(post_bl, phi);
1087                         }
1088                 }
1089                 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1090         } else {
1091                 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1092         }
1093         /* handle the regular call */
1094         set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1095
1096         /* For now, we cannot inline calls with value_base */
1097         set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1098
1099         /* Finally the exception control flow.
1100            We have two (three) possible situations:
1101            First if the Call branches to an exception handler: We need to add a Phi node to
1102            collect the memory containing the exception objects.  Further we need
1103            to add another block to get a correct representation of this Phi.  To
1104            this block we add a Jmp that resolves into the X output of the Call
1105            when the Call is turned into a tuple.
1106            Second the Call branches to End, the exception is not handled.  Just
1107            add all inlined exception branches to the End node.
1108            Third: there is no Exception edge at all. Handle as case two. */
1109         if (exc_handling == exc_handler) {
1110                 n_exc = 0;
1111                 for (i = 0; i < arity; i++) {
1112                         ir_node *ret, *irn;
1113                         ret = get_irn_n(end_bl, i);
1114                         irn = skip_Proj(ret);
1115                         if (is_fragile_op(irn) || is_Raise(irn)) {
1116                                 cf_pred[n_exc] = ret;
1117                                 ++n_exc;
1118                         }
1119                 }
1120                 if (n_exc > 0) {
1121                         new_Block(n_exc, cf_pred);      /* watch it: current_block is changed! */
1122                         set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1123                         /* The Phi for the memories with the exception objects */
1124                         n_exc = 0;
1125                         for (i = 0; i < arity; i++) {
1126                                 ir_node *ret;
1127                                 ret = skip_Proj(get_irn_n(end_bl, i));
1128                                 if (is_Call(ret)) {
1129                                         cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1130                                         n_exc++;
1131                                 } else if (is_fragile_op(ret)) {
1132                                         /* We rely that all cfops have the memory output at the same position. */
1133                                         cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1134                                         n_exc++;
1135                                 } else if (is_Raise(ret)) {
1136                                         cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1137                                         n_exc++;
1138                                 }
1139                         }
1140                         set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1141                 } else {
1142                         set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1143                         set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1144                 }
1145         } else {
1146                 ir_node *main_end_bl;
1147                 int main_end_bl_arity;
1148                 ir_node **end_preds;
1149
1150                 /* assert(exc_handling == 1 || no exceptions. ) */
1151                 n_exc = 0;
1152                 for (i = 0; i < arity; i++) {
1153                         ir_node *ret = get_irn_n(end_bl, i);
1154                         ir_node *irn = skip_Proj(ret);
1155
1156                         if (is_fragile_op(irn) || is_Raise(irn)) {
1157                                 cf_pred[n_exc] = ret;
1158                                 n_exc++;
1159                         }
1160                 }
1161                 main_end_bl = get_irg_end_block(irg);
1162                 main_end_bl_arity = get_irn_arity(main_end_bl);
1163                 end_preds =  xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1164
1165                 for (i = 0; i < main_end_bl_arity; ++i)
1166                         end_preds[i] = get_irn_n(main_end_bl, i);
1167                 for (i = 0; i < n_exc; ++i)
1168                         end_preds[main_end_bl_arity + i] = cf_pred[i];
1169                 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1170                 set_Tuple_pred(call, pn_Call_X_except,  new_Bad());
1171                 set_Tuple_pred(call, pn_Call_M_except,  new_Bad());
1172                 free(end_preds);
1173         }
1174         free(res_pred);
1175         free(cf_pred);
1176
1177         /* --  Turn CSE back on. -- */
1178         set_optimize(rem_opt);
1179         current_ir_graph = rem;
1180
1181         return 1;
1182 }
1183
1184 /********************************************************************/
1185 /* Apply inlineing to small methods.                                */
1186 /********************************************************************/
1187
1188 static struct obstack  temp_obst;
1189
1190 /** Represents a possible inlinable call in a graph. */
1191 typedef struct _call_entry call_entry;
1192 struct _call_entry {
1193         ir_node    *call;      /**< the Call node */
1194         ir_graph   *callee;    /**< the callee IR-graph called here */
1195         call_entry *next;      /**< for linking the next one */
1196         int        loop_depth; /**< the loop depth of this call */
1197 };
1198
1199 /**
1200  * environment for inlining small irgs
1201  */
1202 typedef struct _inline_env_t {
1203         struct obstack obst;  /**< an obstack where call_entries are allocated on. */
1204         call_entry *head;     /**< the head of the call entry list */
1205         call_entry *tail;     /**< the tail of the call entry list */
1206 } inline_env_t;
1207
1208 /**
1209  * Returns the irg called from a Call node. If the irg is not
1210  * known, NULL is returned.
1211  *
1212  * @param call  the call node
1213  */
1214 static ir_graph *get_call_called_irg(ir_node *call) {
1215         ir_node *addr;
1216
1217         addr = get_Call_ptr(call);
1218         if (is_Global(addr)) {
1219                 ir_entity *ent = get_Global_entity(addr);
1220                 return get_entity_irg(ent);
1221         }
1222
1223         return NULL;
1224 }
1225
1226 /**
1227  * Walker: Collect all calls to known graphs inside a graph.
1228  */
1229 static void collect_calls(ir_node *call, void *env) {
1230         if (is_Call(call)) {
1231                 ir_graph *called_irg = get_call_called_irg(call);
1232
1233                 if (called_irg != NULL) {
1234                         /* The Call node calls a locally defined method.  Remember to inline. */
1235                         inline_env_t *ienv  = env;
1236                         call_entry   *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1237                         entry->call       = call;
1238                         entry->callee     = called_irg;
1239                         entry->next       = NULL;
1240                         entry->loop_depth = 0;
1241
1242                         if (ienv->tail == NULL)
1243                                 ienv->head = entry;
1244                         else
1245                                 ienv->tail->next = entry;
1246                         ienv->tail = entry;
1247                 }
1248         }
1249 }
1250
1251 /**
1252  * Inlines all small methods at call sites where the called address comes
1253  * from a Const node that references the entity representing the called
1254  * method.
1255  * The size argument is a rough measure for the code size of the method:
1256  * Methods where the obstack containing the firm graph is smaller than
1257  * size are inlined.
1258  */
1259 void inline_small_irgs(ir_graph *irg, int size) {
1260         ir_graph *rem = current_ir_graph;
1261         inline_env_t env;
1262         call_entry *entry;
1263
1264         current_ir_graph = irg;
1265         /* Handle graph state */
1266         assert(get_irg_phase_state(irg) != phase_building);
1267         free_callee_info(irg);
1268
1269         /* Find Call nodes to inline.
1270            (We can not inline during a walk of the graph, as inlineing the same
1271            method several times changes the visited flag of the walked graph:
1272            after the first inlineing visited of the callee equals visited of
1273            the caller.  With the next inlineing both are increased.) */
1274         obstack_init(&env.obst);
1275         env.head = env.tail = NULL;
1276         irg_walk_graph(irg, NULL, collect_calls, &env);
1277
1278         if (env.head != NULL) {
1279                 /* There are calls to inline */
1280                 collect_phiprojs(irg);
1281                 for (entry = env.head; entry != NULL; entry = entry->next) {
1282                         ir_graph *callee = entry->callee;
1283                         if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1284                             (get_irg_inline_property(callee) >= irg_inline_forced)) {
1285                                 inline_method(entry->call, callee);
1286                         }
1287                 }
1288         }
1289         obstack_free(&env.obst, NULL);
1290         current_ir_graph = rem;
1291 }
1292
1293 /**
1294  * Environment for inlining irgs.
1295  */
1296 typedef struct {
1297         int n_nodes;             /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1298         int n_blocks;            /**< Number of Blocks in graph without Start and End block. */
1299         int n_nodes_orig;        /**< for statistics */
1300         int n_call_nodes;        /**< Number of Call nodes in the graph. */
1301         int n_call_nodes_orig;   /**< for statistics */
1302         int n_callers;           /**< Number of known graphs that call this graphs. */
1303         int n_callers_orig;      /**< for statistics */
1304         unsigned got_inline:1;   /**< Set, if at least one call inside this graph was inlined. */
1305         unsigned local_vars:1;   /**< Set, if a inlined function gets the address of an inlined variable. */
1306         call_entry *call_head;   /**< The head of the list of all call nodes in this graph. */
1307         call_entry *call_tail;   /**< The tail of the list of all call nodes in this graph .*/
1308         unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1309 } inline_irg_env;
1310
1311 /**
1312  * Allocate a new environment for inlining.
1313  */
1314 static inline_irg_env *alloc_inline_irg_env(void) {
1315         inline_irg_env *env    = obstack_alloc(&temp_obst, sizeof(*env));
1316         env->n_nodes           = -2; /* do not count count Start, End */
1317         env->n_blocks          = -2; /* do not count count Start, End Block */
1318         env->n_nodes_orig      = -2; /* do not count Start, End */
1319         env->call_head         = NULL;
1320         env->call_tail         = NULL;
1321         env->n_call_nodes      = 0;
1322         env->n_call_nodes_orig = 0;
1323         env->n_callers         = 0;
1324         env->n_callers_orig    = 0;
1325         env->got_inline        = 0;
1326         env->local_vars        = 0;
1327         env->local_weights     = NULL;
1328         return env;
1329 }
1330
1331 typedef struct walker_env {
1332         inline_irg_env *x;     /**< the inline environment */
1333         call_entry *last_call; /**< points to the last inserted call */
1334         char ignore_runtime;   /**< the ignore runtime flag */
1335         char ignore_callers;   /**< if set, do change callers data */
1336 } wenv_t;
1337
1338 /**
1339  * post-walker: collect all calls in the inline-environment
1340  * of a graph and sum some statistics.
1341  */
1342 static void collect_calls2(ir_node *call, void *ctx) {
1343         wenv_t         *env = ctx;
1344         inline_irg_env *x = env->x;
1345         ir_opcode      code = get_irn_opcode(call);
1346         ir_graph       *callee;
1347         call_entry     *entry;
1348
1349         /* count meaningful nodes in irg */
1350         if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1351                 if (code != iro_Block) {
1352                         ++x->n_nodes;
1353                         ++x->n_nodes_orig;
1354                 } else {
1355                         ++x->n_blocks;
1356                 }
1357         }
1358
1359         if (code != iro_Call) return;
1360
1361         /* check, if it's a runtime call */
1362         if (env->ignore_runtime) {
1363                 ir_node *symc = get_Call_ptr(call);
1364
1365                 if (is_Global(symc)) {
1366                         ir_entity *ent = get_Global_entity(symc);
1367
1368                         if (get_entity_additional_properties(ent) & mtp_property_runtime)
1369                                 return;
1370                 }
1371         }
1372
1373         /* collect all call nodes */
1374         ++x->n_call_nodes;
1375         ++x->n_call_nodes_orig;
1376
1377         callee = get_call_called_irg(call);
1378         if (callee != NULL) {
1379                 if (! env->ignore_callers) {
1380                         inline_irg_env *callee_env = get_irg_link(callee);
1381                         /* count all static callers */
1382                         ++callee_env->n_callers;
1383                         ++callee_env->n_callers_orig;
1384                 }
1385
1386                 /* link it in the list of possible inlinable entries */
1387                 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1388                 entry->call       = call;
1389                 entry->callee     = callee;
1390                 entry->next       = NULL;
1391                 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1392
1393                 /* note: we use call_tail here as a pointer to the last inserted */
1394                 if (x->call_head == NULL) {
1395                         x->call_head = entry;
1396                 } else {
1397                         if (entry->loop_depth == env->last_call->loop_depth) {
1398                                 /* same depth as the last one, enqueue after it */
1399                                 entry->next          = env->last_call->next;
1400                                 env->last_call->next = entry;
1401                         } else if (entry->loop_depth > x->call_head->loop_depth) {
1402                                 /* put first */
1403                                 entry->next  = x->call_head;
1404                                 x->call_head = entry;
1405                         } else {
1406                                 /* search the insertion point */
1407                                 call_entry *p;
1408
1409                                 for (p = x->call_head; p->next != NULL; p = p->next)
1410                                         if (entry->loop_depth > p->next->loop_depth)
1411                                                 break;
1412                                 entry->next = p->next;
1413                                 p->next     = entry;
1414                         }
1415                 }
1416                 env->last_call = entry;
1417                 if (entry->next == NULL) {
1418                         /* keep tail up to date */
1419                         x->call_tail = entry;
1420                 }
1421         }
1422 }
1423
1424 /**
1425  * Returns TRUE if the number of callers is 0 in the irg's environment,
1426  * hence this irg is a leave.
1427  */
1428 INLINE static int is_leave(ir_graph *irg) {
1429         inline_irg_env *env = get_irg_link(irg);
1430         return env->n_call_nodes == 0;
1431 }
1432
1433 /**
1434  * Returns TRUE if the number of nodes in the callee is
1435  * smaller then size in the irg's environment.
1436  */
1437 INLINE static int is_smaller(ir_graph *callee, int size) {
1438         inline_irg_env *env = get_irg_link(callee);
1439         return env->n_nodes < size;
1440 }
1441
1442 /**
1443  * Append the nodes of the list src to the nodes of the list in environment dst.
1444  */
1445 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1446         call_entry *entry, *nentry;
1447
1448         /* Note that the src list points to Call nodes in the inlined graph, but
1449            we need Call nodes in our graph. Luckily the inliner leaves this information
1450            in the link field. */
1451         for (entry = src; entry != NULL; entry = entry->next) {
1452                 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1453                 nentry->call         = get_irn_link(entry->call);
1454                 nentry->callee       = entry->callee;
1455                 nentry->next         = NULL;
1456                 nentry->loop_depth   = entry->loop_depth;
1457                 dst->call_tail->next = nentry;
1458                 dst->call_tail       = nentry;
1459         }
1460 }
1461
1462 /**
1463  * Add the nodes of the list src in front to the nodes of the list dst.
1464  */
1465 static call_entry *replace_entry_by_call_list(call_entry *dst, call_entry *src) {
1466         call_entry *entry, *nentry, *head, *tail;
1467
1468         /* Note that the src list points to Call nodes in the inlined graph, but
1469            we need Call nodes in our graph. Luckily the inliner leaves this information
1470            in the link field. */
1471         head = tail = NULL;
1472         for (entry = src; entry != NULL; entry = entry->next) {
1473                 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1474                 nentry->call         = get_irn_link(entry->call);
1475                 nentry->callee       = entry->callee;
1476                 nentry->next         = NULL;
1477                 nentry->loop_depth   = entry->loop_depth + dst->loop_depth;
1478                 if (head == NULL)
1479                         head = nentry;
1480                 else
1481                         tail->next = nentry;
1482                 tail = nentry;
1483         }
1484         /* skip the head of dst */
1485         if (head != NULL) {
1486                 tail->next = dst->next;
1487         } else {
1488                 head = dst->next;
1489         }
1490         return head;
1491 }
1492
1493 /*
1494  * Inlines small leave methods at call sites where the called address comes
1495  * from a Const node that references the entity representing the called
1496  * method.
1497  * The size argument is a rough measure for the code size of the method:
1498  * Methods where the obstack containing the firm graph is smaller than
1499  * size are inlined.
1500  */
1501 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1502         inline_irg_env   *env;
1503         ir_graph         *irg;
1504         int              i, n_irgs;
1505         ir_graph         *rem;
1506         int              did_inline;
1507         wenv_t           wenv;
1508         call_entry       *entry, *tail;
1509         const call_entry *centry;
1510         pmap             *copied_graphs;
1511         pmap_entry       *pm_entry;
1512
1513         rem = current_ir_graph;
1514         obstack_init(&temp_obst);
1515
1516         /* a map for the copied graphs, used to inline recursive calls */
1517         copied_graphs = pmap_create();
1518
1519         /* extend all irgs by a temporary data structure for inlining. */
1520         n_irgs = get_irp_n_irgs();
1521         for (i = 0; i < n_irgs; ++i)
1522                 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1523
1524         /* Precompute information in temporary data structure. */
1525         wenv.ignore_runtime = ignore_runtime;
1526         wenv.ignore_callers = 0;
1527         for (i = 0; i < n_irgs; ++i) {
1528                 ir_graph *irg = get_irp_irg(i);
1529
1530                 assert(get_irg_phase_state(irg) != phase_building);
1531                 free_callee_info(irg);
1532
1533                 wenv.x = get_irg_link(irg);
1534                 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1535         }
1536
1537         /* -- and now inline. -- */
1538
1539         /* Inline leaves recursively -- we might construct new leaves. */
1540         do {
1541                 did_inline = 0;
1542
1543                 for (i = 0; i < n_irgs; ++i) {
1544                         ir_node *call;
1545                         int phiproj_computed = 0;
1546
1547                         current_ir_graph = get_irp_irg(i);
1548                         env = (inline_irg_env *)get_irg_link(current_ir_graph);
1549
1550                         tail = NULL;
1551                         for (entry = env->call_head; entry != NULL; entry = entry->next) {
1552                                 ir_graph *callee;
1553
1554                                 if (env->n_nodes > maxsize) break;
1555
1556                                 call   = entry->call;
1557                                 callee = entry->callee;
1558
1559                                 if (is_leave(callee) && (
1560                                     is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1561                                         if (!phiproj_computed) {
1562                                                 phiproj_computed = 1;
1563                                                 collect_phiprojs(current_ir_graph);
1564                                         }
1565                                         did_inline = inline_method(call, callee);
1566
1567                                         if (did_inline) {
1568                                                 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1569
1570                                                 /* was inlined, must be recomputed */
1571                                                 phiproj_computed = 0;
1572
1573                                                 /* Do some statistics */
1574                                                 env->got_inline = 1;
1575                                                 --env->n_call_nodes;
1576                                                 env->n_nodes += callee_env->n_nodes;
1577                                                 --callee_env->n_callers;
1578
1579                                                 /* remove this call from the list */
1580                                                 if (tail != NULL)
1581                                                         tail->next = entry->next;
1582                                                 else
1583                                                         env->call_head = entry->next;
1584                                                 continue;
1585                                         }
1586                                 }
1587                                 tail = entry;
1588                         }
1589                         env->call_tail = tail;
1590                 }
1591         } while (did_inline);
1592
1593         /* inline other small functions. */
1594         for (i = 0; i < n_irgs; ++i) {
1595                 ir_node *call;
1596                 int phiproj_computed = 0;
1597
1598                 current_ir_graph = get_irp_irg(i);
1599                 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1600
1601                 /* note that the list of possible calls is updated during the process */
1602                 tail = NULL;
1603                 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1604                         ir_graph   *callee;
1605                         pmap_entry *e;
1606
1607                         call   = entry->call;
1608                         callee = entry->callee;
1609
1610                         e = pmap_find(copied_graphs, callee);
1611                         if (e != NULL) {
1612                                 /*
1613                                  * Remap callee if we have a copy.
1614                                  * FIXME: Should we do this only for recursive Calls ?
1615                                  */
1616                                 callee = e->value;
1617                         }
1618
1619                         if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) ||    /* small function */
1620                                 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1621                                 if (current_ir_graph == callee) {
1622                                         /*
1623                                          * Recursive call: we cannot directly inline because we cannot walk
1624                                          * the graph and change it. So we have to make a copy of the graph
1625                                          * first.
1626                                          */
1627
1628                                         inline_irg_env *callee_env;
1629                                         ir_graph       *copy;
1630
1631                                         /*
1632                                          * No copy yet, create one.
1633                                          * Note that recursive methods are never leaves, so it is sufficient
1634                                          * to test this condition here.
1635                                          */
1636                                         copy = create_irg_copy(callee);
1637
1638                                         /* create_irg_copy() destroys the Proj links, recompute them */
1639                                         phiproj_computed = 0;
1640
1641                                         /* allocate new environment */
1642                                         callee_env = alloc_inline_irg_env();
1643                                         set_irg_link(copy, callee_env);
1644
1645                                         wenv.x              = callee_env;
1646                                         wenv.ignore_callers = 1;
1647                                         irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1648
1649                                         /*
1650                                          * Enter the entity of the original graph. This is needed
1651                                          * for inline_method(). However, note that ent->irg still points
1652                                          * to callee, NOT to copy.
1653                                          */
1654                                         set_irg_entity(copy, get_irg_entity(callee));
1655
1656                                         pmap_insert(copied_graphs, callee, copy);
1657                                         callee = copy;
1658
1659                                         /* we have only one caller: the original graph */
1660                                         callee_env->n_callers      = 1;
1661                                         callee_env->n_callers_orig = 1;
1662                                 }
1663                                 if (! phiproj_computed) {
1664                                         phiproj_computed = 1;
1665                                         collect_phiprojs(current_ir_graph);
1666                                 }
1667                                 did_inline = inline_method(call, callee);
1668                                 if (did_inline) {
1669                                         inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1670
1671                                         /* was inlined, must be recomputed */
1672                                         phiproj_computed = 0;
1673
1674                                         /* callee was inline. Append it's call list. */
1675                                         env->got_inline = 1;
1676                                         --env->n_call_nodes;
1677                                         append_call_list(env, callee_env->call_head);
1678                                         env->n_call_nodes += callee_env->n_call_nodes;
1679                                         env->n_nodes += callee_env->n_nodes;
1680                                         --callee_env->n_callers;
1681
1682                                         /* after we have inlined callee, all called methods inside callee
1683                                            are now called once more */
1684                                         for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1685                                                 inline_irg_env *penv = get_irg_link(centry->callee);
1686                                                 ++penv->n_callers;
1687                                         }
1688
1689                                         /* remove this call from the list */
1690                                         if (tail != NULL)
1691                                                 tail->next = entry->next;
1692                                         else
1693                                                 env->call_head = entry->next;
1694                                         continue;
1695                                 }
1696                         }
1697                         tail = entry;
1698                 }
1699                 env->call_tail = tail;
1700         }
1701
1702         for (i = 0; i < n_irgs; ++i) {
1703                 irg = get_irp_irg(i);
1704                 env = (inline_irg_env *)get_irg_link(irg);
1705
1706                 if (env->got_inline) {
1707                         optimize_graph_df(irg);
1708                         optimize_cf(irg);
1709                 }
1710                 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1711                         DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1712                         env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1713                         env->n_callers_orig, env->n_callers,
1714                         get_entity_name(get_irg_entity(irg))));
1715                 }
1716         }
1717
1718         /* kill the copied graphs: we don't need them anymore */
1719         foreach_pmap(copied_graphs, pm_entry) {
1720                 ir_graph *copy = pm_entry->value;
1721
1722                 /* reset the entity, otherwise it will be deleted in the next step ... */
1723                 set_irg_entity(copy, NULL);
1724                 free_ir_graph(copy);
1725         }
1726         pmap_destroy(copied_graphs);
1727
1728         obstack_free(&temp_obst, NULL);
1729         current_ir_graph = rem;
1730 }
1731
1732 /**
1733  * Calculate the parameter weights for transmitting the address of a local variable.
1734  */
1735 static unsigned calc_method_local_weight(ir_node *arg) {
1736         int      i, j, k;
1737         unsigned v, weight = 0;
1738
1739         for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1740                 ir_node *succ = get_irn_out(arg, i);
1741
1742                 switch (get_irn_opcode(succ)) {
1743                 case iro_Load:
1744                 case iro_Store:
1745                         /* Loads and Store can be removed */
1746                         weight += 3;
1747                         break;
1748                 case iro_Sel:
1749                         /* check if all args are constant */
1750                         for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1751                                 ir_node *idx = get_Sel_index(succ, j);
1752                                 if (! is_Const(idx))
1753                                         return 0;
1754                         }
1755                         /* Check users on this Sel. Note: if a 0 is returned here, there was
1756                            some unsupported node. */
1757                         v = calc_method_local_weight(succ);
1758                         if (v == 0)
1759                                 return 0;
1760                         /* we can kill one Sel with constant indexes, this is cheap */
1761                         weight += v + 1;
1762                         break;
1763                 case iro_Id:
1764                         /* when looking backward we might find Id nodes */
1765                         weight += calc_method_local_weight(succ);
1766                         break;
1767                 case iro_Tuple:
1768                         /* unoptimized tuple */
1769                         for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1770                                 ir_node *pred = get_Tuple_pred(succ, j);
1771                                 if (pred == arg) {
1772                                         /* look for Proj(j) */
1773                                         for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1774                                                 ir_node *succ_succ = get_irn_out(succ, k);
1775                                                 if (is_Proj(succ_succ)) {
1776                                                         if (get_Proj_proj(succ_succ) == j) {
1777                                                                 /* found */
1778                                                                 weight += calc_method_local_weight(succ_succ);
1779                                                         }
1780                                                 } else {
1781                                                         /* this should NOT happen */
1782                                                         return 0;
1783                                                 }
1784                                         }
1785                                 }
1786                         }
1787                         break;
1788                 default:
1789                         /* any other node: unsupported yet or bad. */
1790                         return 0;
1791                 }
1792         }
1793         return weight;
1794 }
1795
1796 /**
1797  * Calculate the parameter weights for transmitting the address of a local variable.
1798  */
1799 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1800         ir_entity *ent = get_irg_entity(irg);
1801         ir_type  *mtp;
1802         int      nparams, i, proj_nr;
1803         ir_node  *irg_args, *arg;
1804
1805         mtp      = get_entity_type(ent);
1806         nparams  = get_method_n_params(mtp);
1807
1808         /* allocate a new array. currently used as 'analysed' flag */
1809         env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1810
1811         /* If the method haven't parameters we have nothing to do. */
1812         if (nparams <= 0)
1813                 return;
1814
1815         assure_irg_outs(irg);
1816         irg_args = get_irg_args(irg);
1817         for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1818                 arg     = get_irn_out(irg_args, i);
1819                 proj_nr = get_Proj_proj(arg);
1820                 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1821         }
1822 }
1823
1824 /**
1825  * Calculate the benefice for transmitting an local variable address.
1826  * After inlining, the local variable might be transformed into a
1827  * SSA variable by scalar_replacement().
1828  */
1829 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1830         inline_irg_env *env = get_irg_link(callee);
1831
1832         if (env->local_weights != NULL) {
1833                 if (pos < ARR_LEN(env->local_weights))
1834                         return env->local_weights[pos];
1835                 return 0;
1836         }
1837
1838         analyze_irg_local_weights(env, callee);
1839
1840         if (pos < ARR_LEN(env->local_weights))
1841                 return env->local_weights[pos];
1842         return 0;
1843 }
1844
1845 /**
1846  * calculate a benefice value for inlining the given call.
1847  */
1848 static int calc_inline_benefice(ir_node *call, ir_graph *callee, unsigned *local_adr) {
1849         ir_entity *ent = get_irg_entity(callee);
1850         ir_node   *frame_ptr;
1851         ir_type   *mtp;
1852         int       weight = 0;
1853         int       i, n_params;
1854         unsigned  cc, v;
1855
1856         inline_irg_env *curr_env, *callee_env;
1857
1858         if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1859                 /* do NOT inline noreturn calls */
1860                 return INT_MIN;
1861         }
1862
1863         /* costs for every passed parameter */
1864         n_params = get_Call_n_params(call);
1865         mtp      = get_entity_type(ent);
1866         cc       = get_method_calling_convention(mtp);
1867         if (cc & cc_reg_param) {
1868                 /* register parameter, smaller costs for register parameters */
1869                 int max_regs = cc & ~cc_bits;
1870
1871                 if (max_regs < n_params)
1872                         weight += max_regs * 2 + (n_params - max_regs) * 5;
1873                 else
1874                         weight += n_params * 2;
1875         } else {
1876                 /* parameters are passed an stack */
1877                 weight += 5 * n_params;
1878         }
1879
1880         /* constant parameters improve the benefice */
1881         frame_ptr = get_irg_frame(current_ir_graph);
1882         for (i = 0; i < n_params; ++i) {
1883                 ir_node *param = get_Call_param(call, i);
1884
1885                 if (is_Const(param) || is_SymConst(param))
1886                         weight += get_method_param_weight(ent, i);
1887                 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1888                         /*
1889                          * An address of a local variable is transmitted. After inlining,
1890                          * scalar_replacement might be able to remove the local variable,
1891                          * so honor this.
1892                          */
1893                         v = get_method_local_adress_weight(callee, i);
1894                         weight += v;
1895                         if (v > 0)
1896                                 *local_adr = 1;
1897                 }
1898         }
1899
1900         callee_env = get_irg_link(callee);
1901         if (get_entity_visibility(ent) == visibility_local &&
1902             callee_env->n_callers_orig == 1 &&
1903             callee != current_ir_graph) {
1904                 /* we are the only caller, give big bonus */
1905                 weight += 5000;
1906         }
1907
1908         /* do not inline big functions */
1909         weight -= callee_env->n_nodes;
1910
1911         /* reduce the benefice if the current function is already big */
1912         curr_env = get_irg_link(current_ir_graph);
1913         weight -= curr_env->n_nodes / 100;
1914
1915         /* give a bonus for functions with one block */
1916         if (callee_env->n_blocks == 1)
1917                 weight = weight * 3 / 2;
1918
1919         /* and one for small functions: we want them to be inlined in mostly every case */
1920         else if (callee_env->n_nodes < 20)
1921                 weight += 5000;
1922
1923         /* and finally for leaves: they do not increase the register pressure
1924            because of callee safe registers */
1925         else if (callee_env->n_call_nodes == 0)
1926                 weight += 25;
1927
1928         return weight;
1929 }
1930
1931 /**
1932  * Heuristic inliner. Calculates a benefice value for every call and inlines
1933  * those calls with a value higher than the threshold.
1934  */
1935 void inline_functions(int maxsize, int inline_threshold) {
1936         inline_irg_env   *env;
1937         int              i, n_irgs;
1938         ir_graph         *rem;
1939         int              did_inline;
1940         wenv_t           wenv;
1941         call_entry       *curr_call, **last_call;
1942         const call_entry *centry;
1943         pmap             *copied_graphs;
1944         pmap_entry       *pm_entry;
1945
1946         rem = current_ir_graph;
1947         obstack_init(&temp_obst);
1948
1949         /* a map for the copied graphs, used to inline recursive calls */
1950         copied_graphs = pmap_create();
1951
1952         /* extend all irgs by a temporary data structure for inlining. */
1953         n_irgs = get_irp_n_irgs();
1954         for (i = 0; i < n_irgs; ++i)
1955                 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1956
1957         /* Precompute information in temporary data structure. */
1958         wenv.ignore_runtime = 0;
1959         wenv.ignore_callers = 0;
1960         for (i = 0; i < n_irgs; ++i) {
1961                 ir_graph *irg = get_irp_irg(i);
1962
1963                 assert(get_irg_phase_state(irg) != phase_building);
1964                 free_callee_info(irg);
1965
1966                 wenv.x         = get_irg_link(irg);
1967                 wenv.last_call = NULL;
1968                 assure_cf_loop(irg);
1969                 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1970         }
1971
1972         /* -- and now inline. -- */
1973         for (i = 0; i < n_irgs; ++i) {
1974                 int      phiproj_computed = 0;
1975                 ir_node  *call;
1976                 ir_graph *irg = get_irp_irg(i);
1977
1978                 current_ir_graph = irg;
1979                 env = get_irg_link(irg);
1980
1981                 /* note that the list of possible calls is updated during the process */
1982                 last_call = &env->call_head;
1983                 for (curr_call = env->call_head; curr_call != NULL;) {
1984                         ir_graph   *callee;
1985                         pmap_entry *e;
1986                         int        benefice;
1987                         unsigned   local_adr;
1988
1989                         if (env->n_nodes > maxsize) break;
1990
1991                         call   = curr_call->call;
1992                         callee = curr_call->callee;
1993
1994                         e = pmap_find(copied_graphs, callee);
1995                         if (e != NULL) {
1996                                 /*
1997                                 * Remap callee if we have a copy.
1998                                 * FIXME: Should we do this only for recursive Calls ?
1999                                 */
2000                                 callee = e->value;
2001                         }
2002
2003                         /* calculate the benefice on the original call to prevent excessive inlining */
2004                         local_adr = 0;
2005                         benefice = calc_inline_benefice(call, callee, &local_adr);
2006                         DB((dbg, LEVEL_2, "In %+F Call %+F has benefice %d\n", irg, callee, benefice));
2007
2008                         if (benefice > -inline_threshold ||
2009                                 (get_irg_inline_property(callee) >= irg_inline_forced)) {
2010                                 if (current_ir_graph == callee) {
2011                                         /*
2012                                          * Recursive call: we cannot directly inline because we cannot walk
2013                                          * the graph and change it. So we have to make a copy of the graph
2014                                          * first.
2015                                          */
2016
2017                                         inline_irg_env *callee_env;
2018                                         ir_graph       *copy;
2019
2020                                         /*
2021                                          * No copy yet, create one.
2022                                          * Note that recursive methods are never leaves, so it is sufficient
2023                                          * to test this condition here.
2024                                          */
2025                                         copy = create_irg_copy(callee);
2026
2027                                         /* create_irg_copy() destroys the Proj links, recompute them */
2028                                         phiproj_computed = 0;
2029
2030                                         /* allocate new environment */
2031                                         callee_env = alloc_inline_irg_env();
2032                                         set_irg_link(copy, callee_env);
2033
2034                                         assure_cf_loop(copy);
2035                                         wenv.x              = callee_env;
2036                                         wenv.ignore_callers = 1;
2037                                         irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2038
2039                                         /*
2040                                          * Enter the entity of the original graph. This is needed
2041                                          * for inline_method(). However, note that ent->irg still points
2042                                          * to callee, NOT to copy.
2043                                          */
2044                                         set_irg_entity(copy, get_irg_entity(callee));
2045
2046                                         pmap_insert(copied_graphs, callee, copy);
2047                                         callee = copy;
2048
2049                                         /* we have only one caller: the original graph */
2050                                         callee_env->n_callers      = 1;
2051                                         callee_env->n_callers_orig = 1;
2052                                 }
2053                                 if (! phiproj_computed) {
2054                                         phiproj_computed = 1;
2055                                         collect_phiprojs(current_ir_graph);
2056                                 }
2057                                 did_inline = inline_method(call, callee);
2058                                 if (did_inline) {
2059                                         inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
2060
2061                                         /* was inlined, must be recomputed */
2062                                         phiproj_computed = 0;
2063
2064                                         /* after we have inlined callee, all called methods inside callee
2065                                         are now called once more */
2066                                         for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
2067                                                 inline_irg_env *penv = get_irg_link(centry->callee);
2068                                                 ++penv->n_callers;
2069                                         }
2070
2071                                         /* callee was inline. Append it's call list. */
2072                                         env->got_inline = 1;
2073                                         if (local_adr)
2074                                                 env->local_vars = 1;
2075                                         --env->n_call_nodes;
2076                                         curr_call = replace_entry_by_call_list(curr_call, callee_env->call_head);
2077                                         env->n_call_nodes += callee_env->n_call_nodes;
2078                                         env->n_nodes += callee_env->n_nodes;
2079                                         --callee_env->n_callers;
2080
2081                                         /* remove the current call entry from the list */
2082                                         *last_call = curr_call;
2083                                         continue;
2084                                 }
2085                         }
2086                         last_call = &curr_call->next;
2087                         curr_call = curr_call->next;
2088                 }
2089
2090                 if (env->got_inline) {
2091                         /* this irg got calls inlined: optimize it */
2092
2093                         /* scalar replacement does not work well with Tuple nodes, so optimize them away */
2094                         optimize_graph_df(irg);
2095
2096                         if (env->local_vars) {
2097                                 if (scalar_replacement_opt(irg)) {
2098                                         optimize_graph_df(irg);
2099                                 }
2100                         }
2101                         optimize_cf(irg);
2102                 }
2103                 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2104                         DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2105                         env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2106                         env->n_callers_orig, env->n_callers,
2107                         get_entity_name(get_irg_entity(irg))));
2108                 }
2109         }
2110
2111         /* kill the copied graphs: we don't need them anymore */
2112         foreach_pmap(copied_graphs, pm_entry) {
2113                 ir_graph *copy = pm_entry->value;
2114
2115                 /* reset the entity, otherwise it will be deleted in the next step ... */
2116                 set_irg_entity(copy, NULL);
2117                 free_ir_graph(copy);
2118         }
2119         pmap_destroy(copied_graphs);
2120
2121         obstack_free(&temp_obst, NULL);
2122         current_ir_graph = rem;
2123 }
2124
2125 void firm_init_inline(void) {
2126         FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
2127 }