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 #include "iropt_dbg.h"
  64
  65 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  66
  67 /*------------------------------------------------------------------*/
  68 /* Routines for dead node elimination / copying garbage collection  */
  69 /* of the obstack.                                                  */
  70 /*------------------------------------------------------------------*/
  71
  72 /**
  73  * Remember the new node in the old node by using a field all nodes have.
  74  */
  75 #define set_new_node(oldn, newn)  set_irn_link(oldn, newn)
  76
  77 /**
  78  * Get this new node, before the old node is forgotten.
  79  */
  80 #define get_new_node(oldn) get_irn_link(oldn)
  81
  82 /**
  83  * Check if a new node was set.
  84  */
  85 #define has_new_node(n) (get_new_node(n) != NULL)
  86
  87 /**
  88  * We use the block_visited flag to mark that we have computed the
  89  * number of useful predecessors for this block.
  90  * Further we encode the new arity in this flag in the old blocks.
  91  * Remembering the arity is useful, as it saves a lot of pointer
  92  * accesses.  This function is called for all Phi and Block nodes
  93  * in a Block.
  94  */
  95 static INLINE int
  96 compute_new_arity(ir_node *b) {
  97         int i, res, irn_arity;
  98         int irg_v, block_v;
  99
 100         irg_v = get_irg_block_visited(current_ir_graph);
 101         block_v = get_Block_block_visited(b);
 102         if (block_v >= irg_v) {
 103                 /* we computed the number of preds for this block and saved it in the
 104                    block_v flag */
 105                 return block_v - irg_v;
 106         } else {
 107                 /* compute the number of good predecessors */
 108                 res = irn_arity = get_irn_arity(b);
 109                 for (i = 0; i < irn_arity; i++)
 110                         if (is_Bad(get_irn_n(b, i))) res--;
 111                         /* save it in the flag. */
 112                         set_Block_block_visited(b, irg_v + res);
 113                         return res;
 114         }
 115 }
 116
 117 /**
 118  * Copies the node to the new obstack. The Ins of the new node point to
 119  * the predecessors on the old obstack.  For block/phi nodes not all
 120  * predecessors might be copied.  n->link points to the new node.
 121  * For Phi and Block nodes the function allocates in-arrays with an arity
 122  * only for useful predecessors.  The arity is determined by counting
 123  * the non-bad predecessors of the block.
 124  *
 125  * @param n    The node to be copied
 126  * @param env  if non-NULL, the node number attribute will be copied to the new node
 127  *
 128  * Note: Also used for loop unrolling.
 129  */
 130 static void copy_node(ir_node *n, void *env) {
 131         ir_node *nn, *block;
 132         int new_arity;
 133         ir_op *op = get_irn_op(n);
 134         (void) env;
 135
 136         if (op == op_Bad) {
 137                 /* node copied already */
 138                 return;
 139         } else if (op == op_Block) {
 140                 block = NULL;
 141                 new_arity = compute_new_arity(n);
 142                 n->attr.block.graph_arr = NULL;
 143         } else {
 144                 block = get_nodes_block(n);
 145                 if (op == op_Phi) {
 146                         new_arity = compute_new_arity(block);
 147                 } else {
 148                         new_arity = get_irn_arity(n);
 149                 }
 150         }
 151         nn = new_ir_node(get_irn_dbg_info(n),
 152                 current_ir_graph,
 153                 block,
 154                 op,
 155                 get_irn_mode(n),
 156                 new_arity,
 157                 get_irn_in(n) + 1);
 158         /* Copy the attributes.  These might point to additional data.  If this
 159            was allocated on the old obstack the pointers now are dangling.  This
 160            frees e.g. the memory of the graph_arr allocated in new_immBlock. */
 161         if (op == op_Block) {
 162                 /* we cannot allow blocks WITHOUT macroblock input */
 163                 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
 164         }
 165         copy_node_attr(n, nn);
 166
 167 #ifdef DEBUG_libfirm
 168         {
 169                 int copy_node_nr = env != NULL;
 170                 if (copy_node_nr) {
 171                         /* for easier debugging, we want to copy the node numbers too */
 172                         nn->node_nr = n->node_nr;
 173                 }
 174         }
 175 #endif
 176
 177         set_new_node(n, nn);
 178         hook_dead_node_elim_subst(current_ir_graph, n, nn);
 179 }
 180
 181 /**
 182  * Copies new predecessors of old node to new node remembered in link.
 183  * Spare the Bad predecessors of Phi and Block nodes.
 184  */
 185 static void copy_preds(ir_node *n, void *env) {
 186         ir_node *nn, *block;
 187         int i, j, irn_arity;
 188         (void) env;
 189
 190         nn = get_new_node(n);
 191
 192         if (is_Block(n)) {
 193                 /* copy the macro block header */
 194                 ir_node *mbh = get_Block_MacroBlock(n);
 195
 196                 if (mbh == n) {
 197                         /* this block is a macroblock header */
 198                         set_Block_MacroBlock(nn, nn);
 199                 } else {
 200                         /* get the macro block header */
 201                         ir_node *nmbh = get_new_node(mbh);
 202                         assert(nmbh != NULL);
 203                         set_Block_MacroBlock(nn, nmbh);
 204                 }
 205
 206                 /* Don't copy Bad nodes. */
 207                 j = 0;
 208                 irn_arity = get_irn_arity(n);
 209                 for (i = 0; i < irn_arity; i++) {
 210                         if (! is_Bad(get_irn_n(n, i))) {
 211                                 ir_node *pred = get_irn_n(n, i);
 212                                 set_irn_n(nn, j, get_new_node(pred));
 213                                 j++;
 214                         }
 215                 }
 216                 /* repair the block visited flag from above misuse. Repair it in both
 217                    graphs so that the old one can still be used. */
 218                 set_Block_block_visited(nn, 0);
 219                 set_Block_block_visited(n, 0);
 220                 /* Local optimization could not merge two subsequent blocks if
 221                    in array contained Bads.  Now it's possible.
 222                    We don't call optimize_in_place as it requires
 223                    that the fields in ir_graph are set properly. */
 224                 if ((get_opt_control_flow_straightening()) &&
 225                         (get_Block_n_cfgpreds(nn) == 1) &&
 226                         is_Jmp(get_Block_cfgpred(nn, 0))) {
 227                         ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
 228                         if (nn == old) {
 229                                 /* Jmp jumps into the block it is in -- deal self cycle. */
 230                                 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
 231                                 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
 232                         } else {
 233                                 exchange(nn, old);
 234                         }
 235                 }
 236         } else if (is_Phi(n) && get_irn_arity(n) > 0) {
 237                 /* Don't copy node if corresponding predecessor in block is Bad.
 238                    The Block itself should not be Bad. */
 239                 block = get_nodes_block(n);
 240                 set_nodes_block(nn, get_new_node(block));
 241                 j = 0;
 242                 irn_arity = get_irn_arity(n);
 243                 for (i = 0; i < irn_arity; i++) {
 244                         if (! is_Bad(get_irn_n(block, i))) {
 245                                 ir_node *pred = get_irn_n(n, i);
 246                                 set_irn_n(nn, j, get_new_node(pred));
 247                                 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
 248                                 j++;
 249                         }
 250                 }
 251                 /* If the pre walker reached this Phi after the post walker visited the
 252                    block block_visited is > 0. */
 253                 set_Block_block_visited(get_nodes_block(n), 0);
 254                 /* Compacting the Phi's ins might generate Phis with only one
 255                    predecessor. */
 256                 if (get_irn_arity(nn) == 1)
 257                         exchange(nn, get_irn_n(nn, 0));
 258         } else {
 259                 irn_arity = get_irn_arity(n);
 260                 for (i = -1; i < irn_arity; i++)
 261                         set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
 262         }
 263         /* Now the new node is complete.  We can add it to the hash table for CSE.
 264            @@@ inlining aborts if we identify End. Why? */
 265         if (!is_End(nn))
 266                 add_identities(current_ir_graph->value_table, nn);
 267 }
 268
 269 /**
 270  * Copies the graph recursively, compacts the keep-alives of the end node.
 271  *
 272  * @param irg           the graph to be copied
 273  * @param copy_node_nr  If non-zero, the node number will be copied
 274  */
 275 static void copy_graph(ir_graph *irg, int copy_node_nr) {
 276         ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
 277         ir_node *ka;      /* keep alive */
 278         int i, irn_arity;
 279         unsigned long vfl;
 280
 281         /* Some nodes must be copied by hand, sigh */
 282         vfl = get_irg_visited(irg);
 283         set_irg_visited(irg, vfl + 1);
 284
 285         oe = get_irg_end(irg);
 286         mark_irn_visited(oe);
 287         /* copy the end node by hand, allocate dynamic in array! */
 288         ne = new_ir_node(get_irn_dbg_info(oe),
 289                 irg,
 290                 NULL,
 291                 op_End,
 292                 mode_X,
 293                 -1,
 294                 NULL);
 295         /* Copy the attributes.  Well, there might be some in the future... */
 296         copy_node_attr(oe, ne);
 297         set_new_node(oe, ne);
 298
 299         /* copy the Bad node */
 300         ob = get_irg_bad(irg);
 301         mark_irn_visited(ob);
 302         nb = new_ir_node(get_irn_dbg_info(ob),
 303                 irg,
 304                 NULL,
 305                 op_Bad,
 306                 mode_T,
 307                 0,
 308                 NULL);
 309         copy_node_attr(ob, nb);
 310         set_new_node(ob, nb);
 311
 312         /* copy the NoMem node */
 313         om = get_irg_no_mem(irg);
 314         mark_irn_visited(om);
 315         nm = new_ir_node(get_irn_dbg_info(om),
 316                 irg,
 317                 NULL,
 318                 op_NoMem,
 319                 mode_M,
 320                 0,
 321                 NULL);
 322         copy_node_attr(om, nm);
 323         set_new_node(om, nm);
 324
 325         /* copy the live nodes */
 326         set_irg_visited(irg, vfl);
 327         irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 328
 329         /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
 330
 331         /* visit the anchors as well */
 332         for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
 333                 ir_node *n = get_irg_anchor(irg, i);
 334
 335                 if (n && (get_irn_visited(n) <= vfl)) {
 336                         set_irg_visited(irg, vfl);
 337                         irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 338                 }
 339         }
 340
 341         /* copy_preds for the end node ... */
 342         set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
 343
 344         /*- ... and now the keep alives. -*/
 345         /* First pick the not marked block nodes and walk them.  We must pick these
 346            first as else we will oversee blocks reachable from Phis. */
 347         irn_arity = get_End_n_keepalives(oe);
 348         for (i = 0; i < irn_arity; i++) {
 349                 ka = get_End_keepalive(oe, i);
 350                 if (is_Block(ka)) {
 351                         if (get_irn_visited(ka) <= vfl) {
 352                                 /* We must keep the block alive and copy everything reachable */
 353                                 set_irg_visited(irg, vfl);
 354                                 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 355                         }
 356                         add_End_keepalive(ne, get_new_node(ka));
 357                 }
 358         }
 359
 360         /* Now pick other nodes.  Here we will keep all! */
 361         irn_arity = get_End_n_keepalives(oe);
 362         for (i = 0; i < irn_arity; i++) {
 363                 ka = get_End_keepalive(oe, i);
 364                 if (!is_Block(ka)) {
 365                         if (get_irn_visited(ka) <= vfl) {
 366                                 /* We didn't copy the node yet.  */
 367                                 set_irg_visited(irg, vfl);
 368                                 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
 369                         }
 370                         add_End_keepalive(ne, get_new_node(ka));
 371                 }
 372         }
 373
 374         /* start block sometimes only reached after keep alives */
 375         set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
 376         set_nodes_block(nm, get_new_node(get_nodes_block(om)));
 377 }
 378
 379 /**
 380  * Copies the graph reachable from current_ir_graph->end to the obstack
 381  * in current_ir_graph and fixes the environment.
 382  * Then fixes the fields in current_ir_graph containing nodes of the
 383  * graph.
 384  *
 385  * @param copy_node_nr  If non-zero, the node number will be copied
 386  */
 387 static void
 388 copy_graph_env(int copy_node_nr) {
 389         ir_graph *irg = current_ir_graph;
 390         ir_node *old_end, *new_anchor;
 391         int i;
 392
 393         /* remove end_except and end_reg nodes */
 394         old_end = get_irg_end(irg);
 395         set_irg_end_except (irg, old_end);
 396         set_irg_end_reg    (irg, old_end);
 397
 398         /* Not all nodes remembered in irg might be reachable
 399            from the end node.  Assure their link is set to NULL, so that
 400            we can test whether new nodes have been computed. */
 401         for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
 402                 ir_node *n = get_irg_anchor(irg, i);
 403                 if (n != NULL)
 404                         set_new_node(n, NULL);
 405         }
 406         /* we use the block walk flag for removing Bads from Blocks ins. */
 407         inc_irg_block_visited(irg);
 408
 409         /* copy the graph */
 410         copy_graph(irg, copy_node_nr);
 411
 412         /* fix the anchor */
 413         old_end    = get_irg_end(irg);
 414         new_anchor = new_Anchor(irg);
 415
 416         for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
 417                 ir_node *n = get_irg_anchor(irg, i);
 418                 if (n)
 419                         set_irn_n(new_anchor, i, get_new_node(n));
 420         }
 421         free_End(old_end);
 422         irg->anchor = new_anchor;
 423
 424         /* ensure the new anchor is placed in the endblock */
 425         set_nodes_block(new_anchor, get_irg_end_block(irg));
 426 }
 427
 428 /**
 429  * Copies all reachable nodes to a new obstack.  Removes bad inputs
 430  * from block nodes and the corresponding inputs from Phi nodes.
 431  * Merges single exit blocks with single entry blocks and removes
 432  * 1-input Phis.
 433  * Adds all new nodes to a new hash table for CSE.  Does not
 434  * perform CSE, so the hash table might contain common subexpressions.
 435  */
 436 void dead_node_elimination(ir_graph *irg) {
 437         ir_graph *rem;
 438 #ifdef INTERPROCEDURAL_VIEW
 439         int rem_ipview = get_interprocedural_view();
 440 #endif
 441         struct obstack *graveyard_obst = NULL;
 442         struct obstack *rebirth_obst   = NULL;
 443
 444         edges_deactivate(irg);
 445
 446         /* inform statistics that we started a dead-node elimination run */
 447         hook_dead_node_elim(irg, 1);
 448
 449         /* Remember external state of current_ir_graph. */
 450         rem = current_ir_graph;
 451         current_ir_graph = irg;
 452 #ifdef INTERPROCEDURAL_VIEW
 453         set_interprocedural_view(0);
 454 #endif
 455
 456         assert(get_irg_phase_state(irg) != phase_building);
 457
 458         /* Handle graph state */
 459         free_callee_info(irg);
 460         free_irg_outs(irg);
 461         free_trouts();
 462
 463         /* @@@ so far we loose loops when copying */
 464         free_loop_information(irg);
 465
 466         set_irg_doms_inconsistent(irg);
 467
 468         /* A quiet place, where the old obstack can rest in peace,
 469            until it will be cremated. */
 470         graveyard_obst = irg->obst;
 471
 472         /* A new obstack, where the reachable nodes will be copied to. */
 473         rebirth_obst = xmalloc(sizeof(*rebirth_obst));
 474         irg->obst = rebirth_obst;
 475         obstack_init(irg->obst);
 476         irg->last_node_idx = 0;
 477
 478         /* We also need a new value table for CSE */
 479         del_identities(irg->value_table);
 480         irg->value_table = new_identities();
 481
 482         /* Copy the graph from the old to the new obstack */
 483         copy_graph_env(/*copy_node_nr=*/1);
 484
 485         /* Free memory from old unoptimized obstack */
 486         obstack_free(graveyard_obst, 0);  /* First empty the obstack ... */
 487         xfree(graveyard_obst);            /* ... then free it.           */
 488
 489         /* inform statistics that the run is over */
 490         hook_dead_node_elim(irg, 0);
 491
 492         current_ir_graph = rem;
 493 #ifdef INTERPROCEDURAL_VIEW
 494         set_interprocedural_view(rem_ipview);
 495 #endif
 496 }
 497
 498 /**
 499  * Relink bad predecessors of a block and store the old in array to the
 500  * link field. This function is called by relink_bad_predecessors().
 501  * The array of link field starts with the block operand at position 0.
 502  * If block has bad predecessors, create a new in array without bad preds.
 503  * Otherwise let in array untouched.
 504  */
 505 static void relink_bad_block_predecessors(ir_node *n, void *env) {
 506         ir_node **new_in, *irn;
 507         int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
 508         (void) env;
 509
 510         /* if link field of block is NULL, look for bad predecessors otherwise
 511            this is already done */
 512         if (is_Block(n) && get_irn_link(n) == NULL) {
 513                 /* save old predecessors in link field (position 0 is the block operand)*/
 514                 set_irn_link(n, get_irn_in(n));
 515
 516                 /* count predecessors without bad nodes */
 517                 old_irn_arity = get_irn_arity(n);
 518                 for (i = 0; i < old_irn_arity; i++)
 519                         if (!is_Bad(get_irn_n(n, i)))
 520                                 ++new_irn_arity;
 521
 522                 /* arity changing: set new predecessors without bad nodes */
 523                 if (new_irn_arity < old_irn_arity) {
 524                         /* Get new predecessor array. We do not resize the array, as we must
 525                            keep the old one to update Phis. */
 526                         new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
 527
 528                         /* set new predecessors in array */
 529                         new_in[0] = NULL;
 530                         new_irn_n = 1;
 531                         for (i = 0; i < old_irn_arity; i++) {
 532                                 irn = get_irn_n(n, i);
 533                                 if (!is_Bad(irn)) {
 534                                         new_in[new_irn_n] = irn;
 535                                         is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
 536                                         ++new_irn_n;
 537                                 }
 538                         }
 539                         /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
 540                         ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
 541                         n->in = new_in;
 542                 } /* ir node has bad predecessors */
 543         } /* Block is not relinked */
 544 }
 545
 546 /**
 547  * Relinks Bad predecessors from Blocks and Phis called by walker
 548  * remove_bad_predecesors(). If n is a Block, call
 549  * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
 550  * function of Phi's Block. If this block has bad predecessors, relink preds
 551  * of the Phi-node.
 552  */
 553 static void relink_bad_predecessors(ir_node *n, void *env) {
 554         ir_node *block, **old_in;
 555         int i, old_irn_arity, new_irn_arity;
 556
 557         /* relink bad predecessors of a block */
 558         if (is_Block(n))
 559                 relink_bad_block_predecessors(n, env);
 560
 561         /* If Phi node relink its block and its predecessors */
 562         if (is_Phi(n)) {
 563                 /* Relink predecessors of phi's block */
 564                 block = get_nodes_block(n);
 565                 if (get_irn_link(block) == NULL)
 566                         relink_bad_block_predecessors(block, env);
 567
 568                 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
 569                 old_irn_arity = ARR_LEN(old_in);
 570
 571                 /* Relink Phi predecessors if count of predecessors changed */
 572                 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
 573                         /* set new predecessors in array
 574                            n->in[0] remains the same block */
 575                         new_irn_arity = 1;
 576                         for(i = 1; i < old_irn_arity; i++)
 577                                 if (!is_Bad(old_in[i])) {
 578                                         n->in[new_irn_arity] = n->in[i];
 579                                         is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
 580                                         ++new_irn_arity;
 581                                 }
 582
 583                                 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
 584                                 ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
 585                 }
 586         } /* n is a Phi node */
 587 }
 588
 589 /*
 590  * Removes Bad Bad predecessors from Blocks and the corresponding
 591  * inputs to Phi nodes as in dead_node_elimination but without
 592  * copying the graph.
 593  * On walking up set the link field to NULL, on walking down call
 594  * relink_bad_predecessors() (This function stores the old in array
 595  * to the link field and sets a new in array if arity of predecessors
 596  * changes).
 597  */
 598 void remove_bad_predecessors(ir_graph *irg) {
 599         panic("Fix backedge handling first");
 600         irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
 601 }
 602
 603
 604 /*
 605    __                      _  __ __
 606   (_     __    o     _    | \/  |_
 607   __)|_| | \_/ | \_/(/_   |_/\__|__
 608
 609   The following stuff implements a facility that automatically patches
 610   registered ir_node pointers to the new node when a dead node elimination occurs.
 611 */
 612
 613 struct _survive_dce_t {
 614         struct obstack obst;
 615         pmap *places;
 616         pmap *new_places;
 617         hook_entry_t dead_node_elim;
 618         hook_entry_t dead_node_elim_subst;
 619 };
 620
 621 typedef struct _survive_dce_list_t {
 622         struct _survive_dce_list_t *next;
 623         ir_node **place;
 624 } survive_dce_list_t;
 625
 626 static void dead_node_hook(void *context, ir_graph *irg, int start) {
 627         survive_dce_t *sd = context;
 628         (void) irg;
 629
 630         /* Create a new map before the dead node elimination is performed. */
 631         if (start) {
 632                 sd->new_places = pmap_create_ex(pmap_count(sd->places));
 633         } else {
 634                 /* Patch back all nodes if dead node elimination is over and something is to be done. */
 635                 pmap_destroy(sd->places);
 636                 sd->places     = sd->new_places;
 637                 sd->new_places = NULL;
 638         }
 639 }
 640
 641 /**
 642  * Hook called when dead node elimination replaces old by nw.
 643  */
 644 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
 645         survive_dce_t *sd = context;
 646         survive_dce_list_t *list = pmap_get(sd->places, old);
 647         (void) irg;
 648
 649         /* If the node is to be patched back, write the new address to all registered locations. */
 650         if (list) {
 651                 survive_dce_list_t *p;
 652
 653                 for (p = list; p; p = p->next)
 654                         *(p->place) = nw;
 655
 656                 pmap_insert(sd->new_places, nw, list);
 657         }
 658 }
 659
 660 /**
 661  * Make a new Survive DCE environment.
 662  */
 663 survive_dce_t *new_survive_dce(void) {
 664         survive_dce_t *res = xmalloc(sizeof(res[0]));
 665         obstack_init(&res->obst);
 666         res->places     = pmap_create();
 667         res->new_places = NULL;
 668
 669         res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
 670         res->dead_node_elim.context                   = res;
 671         res->dead_node_elim.next                      = NULL;
 672
 673         res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
 674         res->dead_node_elim_subst.context = res;
 675         res->dead_node_elim_subst.next    = NULL;
 676
 677 #ifndef FIRM_ENABLE_HOOKS
 678         assert(0 && "need hooks enabled");
 679 #endif
 680
 681         register_hook(hook_dead_node_elim, &res->dead_node_elim);
 682         register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
 683         return res;
 684 }
 685
 686 /**
 687  * Free a Survive DCE environment.
 688  */
 689 void free_survive_dce(survive_dce_t *sd) {
 690         obstack_free(&sd->obst, NULL);
 691         pmap_destroy(sd->places);
 692         unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
 693         unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
 694         xfree(sd);
 695 }
 696
 697 /**
 698  * Register a node pointer to be patched upon DCE.
 699  * When DCE occurs, the node pointer specified by @p place will be
 700  * patched to the new address of the node it is pointing to.
 701  *
 702  * @param sd    The Survive DCE environment.
 703  * @param place The address of the node pointer.
 704  */
 705 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
 706         if (*place != NULL) {
 707                 ir_node *irn      = *place;
 708                 survive_dce_list_t *curr = pmap_get(sd->places, irn);
 709                 survive_dce_list_t *nw   = obstack_alloc(&sd->obst, sizeof(nw[0]));
 710
 711                 nw->next  = curr;
 712                 nw->place = place;
 713
 714                 pmap_insert(sd->places, irn, nw);
 715         }
 716 }
 717
 718 /*--------------------------------------------------------------------*/
 719 /*  Functionality for inlining                                         */
 720 /*--------------------------------------------------------------------*/
 721
 722 /**
 723  * Copy node for inlineing.  Updates attributes that change when
 724  * inlineing but not for dead node elimination.
 725  *
 726  * Copies the node by calling copy_node() and then updates the entity if
 727  * it's a local one.  env must be a pointer of the frame type of the
 728  * inlined procedure. The new entities must be in the link field of
 729  * the entities.
 730  */
 731 static void copy_node_inline(ir_node *n, void *env) {
 732         ir_node *nn;
 733         ir_type *frame_tp = (ir_type *)env;
 734
 735         copy_node(n, NULL);
 736         if (is_Sel(n)) {
 737                 nn = get_new_node (n);
 738                 assert(is_Sel(nn));
 739                 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
 740                         set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
 741                 }
 742         } else if (is_Block(n)) {
 743                 nn = get_new_node (n);
 744                 nn->attr.block.irg = current_ir_graph;
 745         }
 746 }
 747
 748 /**
 749  * Copies new predecessors of old node and move constants to
 750  * the Start Block.
 751  */
 752 static void copy_preds_inline(ir_node *n, void *env) {
 753         ir_node *nn;
 754
 755         copy_preds(n, env);
 756         nn = skip_Id(get_new_node(n));
 757         if (is_irn_constlike(nn)) {
 758                 /* move Constants into the start block */
 759                 set_nodes_block(nn, get_irg_start_block(current_ir_graph));
 760
 761                 n = identify_remember(current_ir_graph->value_table, nn);
 762                 if (nn != n) {
 763                         DBG_OPT_CSE(nn, n);
 764                         exchange(nn, n);
 765                 }
 766         }
 767 }
 768
 769 /**
 770  * Walker: checks if P_value_arg_base is used.
 771  */
 772 static void find_addr(ir_node *node, void *env) {
 773         int *allow_inline = env;
 774         if (is_Proj(node) &&
 775                         is_Start(get_Proj_pred(node)) &&
 776                         get_Proj_proj(node) == pn_Start_P_value_arg_base) {
 777                 *allow_inline = 0;
 778         } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
 779                 /* From GCC:
 780                  * Refuse to inline alloca call unless user explicitly forced so as this
 781                  * may change program's memory overhead drastically when the function
 782                  * using alloca is called in loop.  In GCC present in SPEC2000 inlining
 783                  * into schedule_block cause it to require 2GB of ram instead of 256MB.
 784                  *
 785                  * Sorryly this is true with our implementation also.
 786                  * Moreover, we cannot differentiate between alloca() and VLA yet, so this
 787                  * disables inlining of functions using VLA (with are completely save).
 788                  *
 789                  * 2 Solutions:
 790                  * - add a flag to the Alloc node for "real" alloca() calls
 791                  * - add a new Stack-Restore node at the end of a function using alloca()
 792                  */
 793                 *allow_inline = 0;
 794         }
 795 }
 796
 797 /**
 798  * Check if we can inline a given call.
 799  * Currently, we cannot inline two cases:
 800  * - call with compound arguments
 801  * - graphs that take the address of a parameter
 802  *
 803  * check these conditions here
 804  */
 805 static int can_inline(ir_node *call, ir_graph *called_graph) {
 806         ir_type *call_type = get_Call_type(call);
 807         int params, ress, i, res;
 808         assert(is_Method_type(call_type));
 809
 810         params = get_method_n_params(call_type);
 811         ress   = get_method_n_ress(call_type);
 812
 813         /* check parameters for compound arguments */
 814         for (i = 0; i < params; ++i) {
 815                 ir_type *p_type = get_method_param_type(call_type, i);
 816
 817                 if (is_compound_type(p_type))
 818                         return 0;
 819         }
 820
 821         /* check results for compound arguments */
 822         for (i = 0; i < ress; ++i) {
 823                 ir_type *r_type = get_method_res_type(call_type, i);
 824
 825                 if (is_compound_type(r_type))
 826                         return 0;
 827         }
 828
 829         res = 1;
 830         irg_walk_graph(called_graph, find_addr, NULL, &res);
 831
 832         return res;
 833 }
 834
 835 enum exc_mode {
 836         exc_handler    = 0, /**< There is a handler. */
 837         exc_to_end     = 1, /**< Branches to End. */
 838         exc_no_handler = 2  /**< Exception handling not represented. */
 839 };
 840
 841 /* Inlines a method at the given call site. */
 842 int inline_method(ir_node *call, ir_graph *called_graph) {
 843         ir_node             *pre_call;
 844         ir_node             *post_call, *post_bl;
 845         ir_node             *in[pn_Start_max];
 846         ir_node             *end, *end_bl, *block;
 847         ir_node             **res_pred;
 848         ir_node             **cf_pred;
 849         ir_node             **args_in;
 850         ir_node             *ret, *phi;
 851         int                 arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
 852         enum exc_mode       exc_handling;
 853         ir_type             *called_frame, *curr_frame, *mtp, *ctp;
 854         ir_entity           *ent;
 855         ir_graph            *rem, *irg;
 856         irg_inline_property prop = get_irg_inline_property(called_graph);
 857         unsigned long       visited;
 858
 859         if (prop == irg_inline_forbidden)
 860                 return 0;
 861
 862         ent = get_irg_entity(called_graph);
 863
 864         mtp = get_entity_type(ent);
 865         ctp = get_Call_type(call);
 866         if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
 867                 /* this is a bad feature of C: without a prototype, we can can call a function with less
 868                 parameters than needed. Currently we don't support this, although it would be
 869                 to use Unknown than. */
 870                 return 0;
 871         }
 872
 873         /* Argh, compiling C has some bad consequences:
 874            the call type AND the method type might be different.
 875            It is implementation defendant what happens in that case.
 876            We support inlining, if the bitsize of the types matches AND
 877            the same arithmetic is used. */
 878         n_params = get_method_n_params(mtp);
 879         for (i = n_params - 1; i >= 0; --i) {
 880                 ir_type *param_tp = get_method_param_type(mtp, i);
 881                 ir_type *arg_tp   = get_method_param_type(ctp, i);
 882
 883                 if (param_tp != arg_tp) {
 884                         ir_mode *pmode = get_type_mode(param_tp);
 885                         ir_mode *amode = get_type_mode(arg_tp);
 886
 887                         if (pmode == NULL || amode == NULL)
 888                                 return 0;
 889                         if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
 890                                 return 0;
 891                         if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
 892                                 return 0;
 893                         /* otherwise we can simply "reinterpret" the bits */
 894                 }
 895         }
 896
 897         irg = get_irn_irg(call);
 898
 899         /*
 900          * We cannot inline a recursive call. The graph must be copied before
 901          * the call the inline_method() using create_irg_copy().
 902          */
 903         if (called_graph == irg)
 904                 return 0;
 905
 906         /*
 907          * currently, we cannot inline two cases:
 908          * - call with compound arguments
 909          * - graphs that take the address of a parameter
 910          */
 911         if (! can_inline(call, called_graph))
 912                 return 0;
 913
 914         rem = current_ir_graph;
 915         current_ir_graph = irg;
 916
 917         DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
 918
 919         /* --  Turn off optimizations, this can cause problems when allocating new nodes. -- */
 920         rem_opt = get_opt_optimize();
 921         set_optimize(0);
 922
 923         /* Handle graph state */
 924         assert(get_irg_phase_state(irg) != phase_building);
 925         assert(get_irg_pinned(irg) == op_pin_state_pinned);
 926         assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
 927         set_irg_outs_inconsistent(irg);
 928         set_irg_extblk_inconsistent(irg);
 929         set_irg_doms_inconsistent(irg);
 930         set_irg_loopinfo_inconsistent(irg);
 931         set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
 932
 933         /* -- Check preconditions -- */
 934         assert(is_Call(call));
 935
 936         /* here we know we WILL inline, so inform the statistics */
 937         hook_inline(call, called_graph);
 938
 939         /* -- Decide how to handle exception control flow: Is there a handler
 940            for the Call node, or do we branch directly to End on an exception?
 941            exc_handling:
 942            0 There is a handler.
 943            1 Branches to End.
 944            2 Exception handling not represented in Firm. -- */
 945         {
 946                 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
 947                 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
 948                         long proj_nr = get_Proj_proj(proj);
 949                         if (proj_nr == pn_Call_X_except) Xproj = proj;
 950                         if (proj_nr == pn_Call_M_except) Mproj = proj;
 951                 }
 952                 if      (Mproj) { assert(Xproj); exc_handling = exc_handler; } /*  Mproj           */
 953                 else if (Xproj) {                exc_handling = exc_to_end; } /* !Mproj &&  Xproj   */
 954                 else            {                exc_handling = exc_no_handler; } /* !Mproj && !Xproj   */
 955         }
 956
 957         /* create the argument tuple */
 958         NEW_ARR_A(ir_type *, args_in, n_params);
 959
 960         block = get_nodes_block(call);
 961         for (i = n_params - 1; i >= 0; --i) {
 962                 ir_node *arg      = get_Call_param(call, i);
 963                 ir_type *param_tp = get_method_param_type(mtp, i);
 964                 ir_mode *mode     = get_type_mode(param_tp);
 965
 966                 if (mode != get_irn_mode(arg)) {
 967                         arg = new_r_Conv(irg, block, arg, mode);
 968                 }
 969                 args_in[i] = arg;
 970         }
 971
 972         /* --
 973            the procedure and later replaces the Start node of the called graph.
 974            Post_call is the old Call node and collects the results of the called
 975            graph. Both will end up being a tuple.  -- */
 976         post_bl = get_nodes_block(call);
 977         set_irg_current_block(irg, post_bl);
 978         /* XxMxPxPxPxT of Start + parameter of Call */
 979         in[pn_Start_X_initial_exec]   = new_Jmp();
 980         in[pn_Start_M]                = get_Call_mem(call);
 981         in[pn_Start_P_frame_base]     = get_irg_frame(irg);
 982         in[pn_Start_P_tls]            = get_irg_tls(irg);
 983         in[pn_Start_T_args]           = new_Tuple(n_params, args_in);
 984         /* in[pn_Start_P_value_arg_base] = ??? */
 985         assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
 986         pre_call = new_Tuple(pn_Start_max - 1, in);
 987         post_call = call;
 988
 989         /* --
 990            The new block gets the ins of the old block, pre_call and all its
 991            predecessors and all Phi nodes. -- */
 992         part_block(pre_call);
 993
 994         /* -- Prepare state for dead node elimination -- */
 995         /* Visited flags in calling irg must be >= flag in called irg.
 996            Else walker and arity computation will not work. */
 997         if (get_irg_visited(irg) <= get_irg_visited(called_graph))
 998                 set_irg_visited(irg, get_irg_visited(called_graph) + 1);
 999         if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
1000                 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
1001         visited = get_irg_visited(irg);
1002
1003         /* Set pre_call as new Start node in link field of the start node of
1004            calling graph and pre_calls block as new block for the start block
1005            of calling graph.
1006            Further mark these nodes so that they are not visited by the
1007            copying. */
1008         set_irn_link(get_irg_start(called_graph), pre_call);
1009         set_irn_visited(get_irg_start(called_graph), visited);
1010         set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1011         set_irn_visited(get_irg_start_block(called_graph), visited);
1012
1013         set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1014         set_irn_visited(get_irg_bad(called_graph), visited);
1015
1016         set_irn_link(get_irg_no_mem(called_graph), get_irg_no_mem(current_ir_graph));
1017         set_irn_visited(get_irg_no_mem(called_graph), visited);
1018
1019         /* Initialize for compaction of in arrays */
1020         inc_irg_block_visited(irg);
1021
1022         /* -- Replicate local entities of the called_graph -- */
1023         /* copy the entities. */
1024         called_frame = get_irg_frame_type(called_graph);
1025         curr_frame   = get_irg_frame_type(irg);
1026         for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
1027                 ir_entity *new_ent, *old_ent;
1028                 old_ent = get_class_member(called_frame, i);
1029                 new_ent = copy_entity_own(old_ent, curr_frame);
1030                 set_entity_link(old_ent, new_ent);
1031         }
1032
1033         /* visited is > than that of called graph.  With this trick visited will
1034            remain unchanged so that an outer walker, e.g., searching the call nodes
1035             to inline, calling this inline will not visit the inlined nodes. */
1036         set_irg_visited(irg, get_irg_visited(irg)-1);
1037
1038         /* -- Performing dead node elimination inlines the graph -- */
1039         /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1040            entities. */
1041         irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
1042                  get_irg_frame_type(called_graph));
1043
1044         /* Repair called_graph */
1045         set_irg_visited(called_graph, get_irg_visited(irg));
1046         set_irg_block_visited(called_graph, get_irg_block_visited(irg));
1047         set_Block_block_visited(get_irg_start_block(called_graph), 0);
1048
1049         /* -- Merge the end of the inlined procedure with the call site -- */
1050         /* We will turn the old Call node into a Tuple with the following
1051            predecessors:
1052            -1:  Block of Tuple.
1053            0: Phi of all Memories of Return statements.
1054            1: Jmp from new Block that merges the control flow from all exception
1055            predecessors of the old end block.
1056            2: Tuple of all arguments.
1057            3: Phi of Exception memories.
1058            In case the old Call directly branches to End on an exception we don't
1059            need the block merging all exceptions nor the Phi of the exception
1060            memories.
1061         */
1062
1063         /* -- Precompute some values -- */
1064         end_bl = get_new_node(get_irg_end_block(called_graph));
1065         end = get_new_node(get_irg_end(called_graph));
1066         arity = get_irn_arity(end_bl);    /* arity = n_exc + n_ret  */
1067         n_res = get_method_n_ress(get_Call_type(call));
1068
1069         res_pred = xmalloc(n_res * sizeof(*res_pred));
1070         cf_pred  = xmalloc(arity * sizeof(*res_pred));
1071
1072         set_irg_current_block(irg, post_bl); /* just to make sure */
1073
1074         /* -- archive keepalives -- */
1075         irn_arity = get_irn_arity(end);
1076         for (i = 0; i < irn_arity; i++) {
1077                 ir_node *ka = get_End_keepalive(end, i);
1078                 if (! is_Bad(ka))
1079                         add_End_keepalive(get_irg_end(irg), ka);
1080         }
1081
1082         /* The new end node will die.  We need not free as the in array is on the obstack:
1083            copy_node() only generated 'D' arrays. */
1084
1085         /* -- Replace Return nodes by Jump nodes. -- */
1086         n_ret = 0;
1087         for (i = 0; i < arity; i++) {
1088                 ir_node *ret;
1089                 ret = get_irn_n(end_bl, i);
1090                 if (is_Return(ret)) {
1091                         cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1092                         n_ret++;
1093                 }
1094         }
1095         set_irn_in(post_bl, n_ret, cf_pred);
1096
1097         /* -- Build a Tuple for all results of the method.
1098            Add Phi node if there was more than one Return.  -- */
1099         turn_into_tuple(post_call, pn_Call_max);
1100         /* First the Memory-Phi */
1101         n_ret = 0;
1102         for (i = 0; i < arity; i++) {
1103                 ret = get_irn_n(end_bl, i);
1104                 if (is_Return(ret)) {
1105                         cf_pred[n_ret] = get_Return_mem(ret);
1106                         n_ret++;
1107                 }
1108         }
1109         phi = new_Phi(n_ret, cf_pred, mode_M);
1110         set_Tuple_pred(call, pn_Call_M_regular, phi);
1111         /* Conserve Phi-list for further inlinings -- but might be optimized */
1112         if (get_nodes_block(phi) == post_bl) {
1113                 set_irn_link(phi, get_irn_link(post_bl));
1114                 set_irn_link(post_bl, phi);
1115         }
1116         /* Now the real results */
1117         if (n_res > 0) {
1118                 for (j = 0; j < n_res; j++) {
1119                         n_ret = 0;
1120                         for (i = 0; i < arity; i++) {
1121                                 ret = get_irn_n(end_bl, i);
1122                                 if (is_Return(ret)) {
1123                                         cf_pred[n_ret] = get_Return_res(ret, j);
1124                                         n_ret++;
1125                                 }
1126                         }
1127                         if (n_ret > 0)
1128                                 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1129                         else
1130                                 phi = new_Bad();
1131                         res_pred[j] = phi;
1132                         /* Conserve Phi-list for further inlinings -- but might be optimized */
1133                         if (get_nodes_block(phi) == post_bl) {
1134                                 set_Phi_next(phi, get_Block_phis(post_bl));
1135                                 set_Block_phis(post_bl, phi);
1136                         }
1137                 }
1138                 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1139         } else {
1140                 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1141         }
1142         /* handle the regular call */
1143         set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1144
1145         /* For now, we cannot inline calls with value_base */
1146         set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1147
1148         /* Finally the exception control flow.
1149            We have two (three) possible situations:
1150            First if the Call branches to an exception handler: We need to add a Phi node to
1151            collect the memory containing the exception objects.  Further we need
1152            to add another block to get a correct representation of this Phi.  To
1153            this block we add a Jmp that resolves into the X output of the Call
1154            when the Call is turned into a tuple.
1155            Second the Call branches to End, the exception is not handled.  Just
1156            add all inlined exception branches to the End node.
1157            Third: there is no Exception edge at all. Handle as case two. */
1158         if (exc_handling == exc_handler) {
1159                 n_exc = 0;
1160                 for (i = 0; i < arity; i++) {
1161                         ir_node *ret, *irn;
1162                         ret = get_irn_n(end_bl, i);
1163                         irn = skip_Proj(ret);
1164                         if (is_fragile_op(irn) || is_Raise(irn)) {
1165                                 cf_pred[n_exc] = ret;
1166                                 ++n_exc;
1167                         }
1168                 }
1169                 if (n_exc > 0) {
1170                         new_Block(n_exc, cf_pred);      /* watch it: current_block is changed! */
1171                         set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1172                         /* The Phi for the memories with the exception objects */
1173                         n_exc = 0;
1174                         for (i = 0; i < arity; i++) {
1175                                 ir_node *ret;
1176                                 ret = skip_Proj(get_irn_n(end_bl, i));
1177                                 if (is_Call(ret)) {
1178                                         cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1179                                         n_exc++;
1180                                 } else if (is_fragile_op(ret)) {
1181                                         /* We rely that all cfops have the memory output at the same position. */
1182                                         cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1183                                         n_exc++;
1184                                 } else if (is_Raise(ret)) {
1185                                         cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1186                                         n_exc++;
1187                                 }
1188                         }
1189                         set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1190                 } else {
1191                         set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1192                         set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1193                 }
1194         } else {
1195                 ir_node *main_end_bl;
1196                 int main_end_bl_arity;
1197                 ir_node **end_preds;
1198
1199                 /* assert(exc_handling == 1 || no exceptions. ) */
1200                 n_exc = 0;
1201                 for (i = 0; i < arity; i++) {
1202                         ir_node *ret = get_irn_n(end_bl, i);
1203                         ir_node *irn = skip_Proj(ret);
1204
1205                         if (is_fragile_op(irn) || is_Raise(irn)) {
1206                                 cf_pred[n_exc] = ret;
1207                                 n_exc++;
1208                         }
1209                 }
1210                 main_end_bl = get_irg_end_block(irg);
1211                 main_end_bl_arity = get_irn_arity(main_end_bl);
1212                 end_preds =  xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1213
1214                 for (i = 0; i < main_end_bl_arity; ++i)
1215                         end_preds[i] = get_irn_n(main_end_bl, i);
1216                 for (i = 0; i < n_exc; ++i)
1217                         end_preds[main_end_bl_arity + i] = cf_pred[i];
1218                 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1219                 set_Tuple_pred(call, pn_Call_X_except,  new_Bad());
1220                 set_Tuple_pred(call, pn_Call_M_except,  new_Bad());
1221                 free(end_preds);
1222         }
1223         free(res_pred);
1224         free(cf_pred);
1225
1226         /* --  Turn CSE back on. -- */
1227         set_optimize(rem_opt);
1228         current_ir_graph = rem;
1229
1230         return 1;
1231 }
1232
1233 /********************************************************************/
1234 /* Apply inlineing to small methods.                                */
1235 /********************************************************************/
1236
1237 static struct obstack  temp_obst;
1238
1239 /** Represents a possible inlinable call in a graph. */
1240 typedef struct _call_entry call_entry;
1241 struct _call_entry {
1242         ir_node    *call;      /**< the Call node */
1243         ir_graph   *callee;    /**< the callee IR-graph called here */
1244         call_entry *next;      /**< for linking the next one */
1245         int        loop_depth; /**< the loop depth of this call */
1246 };
1247
1248 /**
1249  * environment for inlining small irgs
1250  */
1251 typedef struct _inline_env_t {
1252         struct obstack obst;  /**< an obstack where call_entries are allocated on. */
1253         call_entry *head;     /**< the head of the call entry list */
1254         call_entry *tail;     /**< the tail of the call entry list */
1255 } inline_env_t;
1256
1257 /**
1258  * Returns the irg called from a Call node. If the irg is not
1259  * known, NULL is returned.
1260  *
1261  * @param call  the call node
1262  */
1263 static ir_graph *get_call_called_irg(ir_node *call) {
1264         ir_node *addr;
1265
1266         addr = get_Call_ptr(call);
1267         if (is_Global(addr)) {
1268                 ir_entity *ent = get_Global_entity(addr);
1269                 return get_entity_irg(ent);
1270         }
1271
1272         return NULL;
1273 }
1274
1275 /**
1276  * Walker: Collect all calls to known graphs inside a graph.
1277  */
1278 static void collect_calls(ir_node *call, void *env) {
1279         if (is_Call(call)) {
1280                 ir_graph *called_irg = get_call_called_irg(call);
1281
1282                 if (called_irg != NULL) {
1283                         /* The Call node calls a locally defined method.  Remember to inline. */
1284                         inline_env_t *ienv  = env;
1285                         call_entry   *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1286                         entry->call       = call;
1287                         entry->callee     = called_irg;
1288                         entry->next       = NULL;
1289                         entry->loop_depth = 0;
1290
1291                         if (ienv->tail == NULL)
1292                                 ienv->head = entry;
1293                         else
1294                                 ienv->tail->next = entry;
1295                         ienv->tail = entry;
1296                 }
1297         }
1298 }
1299
1300 /**
1301  * Inlines all small methods at call sites where the called address comes
1302  * from a Const node that references the entity representing the called
1303  * method.
1304  * The size argument is a rough measure for the code size of the method:
1305  * Methods where the obstack containing the firm graph is smaller than
1306  * size are inlined.
1307  */
1308 void inline_small_irgs(ir_graph *irg, int size) {
1309         ir_graph *rem = current_ir_graph;
1310         inline_env_t env;
1311         call_entry *entry;
1312
1313         current_ir_graph = irg;
1314         /* Handle graph state */
1315         assert(get_irg_phase_state(irg) != phase_building);
1316         free_callee_info(irg);
1317
1318         /* Find Call nodes to inline.
1319            (We can not inline during a walk of the graph, as inlineing the same
1320            method several times changes the visited flag of the walked graph:
1321            after the first inlineing visited of the callee equals visited of
1322            the caller.  With the next inlineing both are increased.) */
1323         obstack_init(&env.obst);
1324         env.head = env.tail = NULL;
1325         irg_walk_graph(irg, NULL, collect_calls, &env);
1326
1327         if (env.head != NULL) {
1328                 /* There are calls to inline */
1329                 collect_phiprojs(irg);
1330                 for (entry = env.head; entry != NULL; entry = entry->next) {
1331                         ir_graph *callee = entry->callee;
1332                         if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1333                             (get_irg_inline_property(callee) >= irg_inline_forced)) {
1334                                 inline_method(entry->call, callee);
1335                         }
1336                 }
1337         }
1338         obstack_free(&env.obst, NULL);
1339         current_ir_graph = rem;
1340 }
1341
1342 /**
1343  * Environment for inlining irgs.
1344  */
1345 typedef struct {
1346         int n_nodes;             /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1347         int n_blocks;            /**< Number of Blocks in graph without Start and End block. */
1348         int n_nodes_orig;        /**< for statistics */
1349         int n_call_nodes;        /**< Number of Call nodes in the graph. */
1350         int n_call_nodes_orig;   /**< for statistics */
1351         int n_callers;           /**< Number of known graphs that call this graphs. */
1352         int n_callers_orig;      /**< for statistics */
1353         unsigned got_inline:1;   /**< Set, if at least one call inside this graph was inlined. */
1354         unsigned local_vars:1;   /**< Set, if a inlined function gets the address of an inlined variable. */
1355         unsigned recursive:1;    /**< Set, if this function is self recursive. */
1356         call_entry *call_head;   /**< The head of the list of all call nodes in this graph. */
1357         call_entry *call_tail;   /**< The tail of the list of all call nodes in this graph .*/
1358         unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1359 } inline_irg_env;
1360
1361 /**
1362  * Allocate a new environment for inlining.
1363  */
1364 static inline_irg_env *alloc_inline_irg_env(void) {
1365         inline_irg_env *env    = obstack_alloc(&temp_obst, sizeof(*env));
1366         env->n_nodes           = -2; /* do not count count Start, End */
1367         env->n_blocks          = -2; /* do not count count Start, End Block */
1368         env->n_nodes_orig      = -2; /* do not count Start, End */
1369         env->call_head         = NULL;
1370         env->call_tail         = NULL;
1371         env->n_call_nodes      = 0;
1372         env->n_call_nodes_orig = 0;
1373         env->n_callers         = 0;
1374         env->n_callers_orig    = 0;
1375         env->got_inline        = 0;
1376         env->local_vars        = 0;
1377         env->recursive         = 0;
1378         env->local_weights     = NULL;
1379         return env;
1380 }
1381
1382 typedef struct walker_env {
1383         inline_irg_env *x;     /**< the inline environment */
1384         call_entry *last_call; /**< points to the last inserted call */
1385         char ignore_runtime;   /**< the ignore runtime flag */
1386         char ignore_callers;   /**< if set, do change callers data */
1387 } wenv_t;
1388
1389 /**
1390  * post-walker: collect all calls in the inline-environment
1391  * of a graph and sum some statistics.
1392  */
1393 static void collect_calls2(ir_node *call, void *ctx) {
1394         wenv_t         *env = ctx;
1395         inline_irg_env *x = env->x;
1396         ir_opcode      code = get_irn_opcode(call);
1397         ir_graph       *callee;
1398         call_entry     *entry;
1399
1400         /* count meaningful nodes in irg */
1401         if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1402                 if (code != iro_Block) {
1403                         ++x->n_nodes;
1404                         ++x->n_nodes_orig;
1405                 } else {
1406                         ++x->n_blocks;
1407                 }
1408         }
1409
1410         if (code != iro_Call) return;
1411
1412         /* check, if it's a runtime call */
1413         if (env->ignore_runtime) {
1414                 ir_node *symc = get_Call_ptr(call);
1415
1416                 if (is_Global(symc)) {
1417                         ir_entity *ent = get_Global_entity(symc);
1418
1419                         if (get_entity_additional_properties(ent) & mtp_property_runtime)
1420                                 return;
1421                 }
1422         }
1423
1424         /* collect all call nodes */
1425         ++x->n_call_nodes;
1426         ++x->n_call_nodes_orig;
1427
1428         callee = get_call_called_irg(call);
1429         if (callee != NULL) {
1430                 if (! env->ignore_callers) {
1431                         inline_irg_env *callee_env = get_irg_link(callee);
1432                         /* count all static callers */
1433                         ++callee_env->n_callers;
1434                         ++callee_env->n_callers_orig;
1435                 }
1436                 if (callee == current_ir_graph)
1437                         x->recursive = 1;
1438
1439                 /* link it in the list of possible inlinable entries */
1440                 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1441                 entry->call       = call;
1442                 entry->callee     = callee;
1443                 entry->next       = NULL;
1444                 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1445
1446                 /* note: we use call_tail here as a pointer to the last inserted */
1447                 if (x->call_head == NULL) {
1448                         x->call_head = entry;
1449                 } else {
1450                         if (entry->loop_depth == env->last_call->loop_depth) {
1451                                 /* same depth as the last one, enqueue after it */
1452                                 entry->next          = env->last_call->next;
1453                                 env->last_call->next = entry;
1454                         } else if (entry->loop_depth > x->call_head->loop_depth) {
1455                                 /* put first */
1456                                 entry->next  = x->call_head;
1457                                 x->call_head = entry;
1458                         } else {
1459                                 /* search the insertion point */
1460                                 call_entry *p;
1461
1462                                 for (p = x->call_head; p->next != NULL; p = p->next)
1463                                         if (entry->loop_depth > p->next->loop_depth)
1464                                                 break;
1465                                 entry->next = p->next;
1466                                 p->next     = entry;
1467                         }
1468                 }
1469                 env->last_call = entry;
1470                 if (entry->next == NULL) {
1471                         /* keep tail up to date */
1472                         x->call_tail = entry;
1473                 }
1474         }
1475 }
1476
1477 /**
1478  * Returns TRUE if the number of callers is 0 in the irg's environment,
1479  * hence this irg is a leave.
1480  */
1481 INLINE static int is_leave(ir_graph *irg) {
1482         inline_irg_env *env = get_irg_link(irg);
1483         return env->n_call_nodes == 0;
1484 }
1485
1486 /**
1487  * Returns TRUE if the number of nodes in the callee is
1488  * smaller then size in the irg's environment.
1489  */
1490 INLINE static int is_smaller(ir_graph *callee, int size) {
1491         inline_irg_env *env = get_irg_link(callee);
1492         return env->n_nodes < size;
1493 }
1494
1495 /**
1496  * Append the nodes of the list src to the nodes of the list in environment dst.
1497  */
1498 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1499         call_entry *entry, *nentry;
1500
1501         /* Note that the src list points to Call nodes in the inlined graph, but
1502            we need Call nodes in our graph. Luckily the inliner leaves this information
1503            in the link field. */
1504         for (entry = src; entry != NULL; entry = entry->next) {
1505                 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1506                 nentry->call         = get_irn_link(entry->call);
1507                 nentry->callee       = entry->callee;
1508                 nentry->next         = NULL;
1509                 nentry->loop_depth   = entry->loop_depth;
1510                 dst->call_tail->next = nentry;
1511                 dst->call_tail       = nentry;
1512         }
1513 }
1514
1515 /**
1516  * Add the nodes of the list src in front to the nodes of the list dst.
1517  */
1518 static call_entry *replace_entry_by_call_list(call_entry *dst, call_entry *src) {
1519         call_entry *entry, *nentry, *head, *tail;
1520
1521         /* Note that the src list points to Call nodes in the inlined graph, but
1522            we need Call nodes in our graph. Luckily the inliner leaves this information
1523            in the link field. */
1524         head = tail = NULL;
1525         for (entry = src; entry != NULL; entry = entry->next) {
1526                 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1527                 nentry->call         = get_irn_link(entry->call);
1528                 nentry->callee       = entry->callee;
1529                 nentry->next         = NULL;
1530                 nentry->loop_depth   = entry->loop_depth + dst->loop_depth;
1531                 if (head == NULL)
1532                         head = nentry;
1533                 else
1534                         tail->next = nentry;
1535                 tail = nentry;
1536         }
1537         /* skip the head of dst */
1538         if (head != NULL) {
1539                 tail->next = dst->next;
1540         } else {
1541                 head = dst->next;
1542         }
1543         return head;
1544 }
1545
1546 /*
1547  * Inlines small leave methods at call sites where the called address comes
1548  * from a Const node that references the entity representing the called
1549  * method.
1550  * The size argument is a rough measure for the code size of the method:
1551  * Methods where the obstack containing the firm graph is smaller than
1552  * size are inlined.
1553  */
1554 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1555         inline_irg_env   *env;
1556         ir_graph         *irg;
1557         int              i, n_irgs;
1558         ir_graph         *rem;
1559         int              did_inline;
1560         wenv_t           wenv;
1561         call_entry       *entry, *tail;
1562         const call_entry *centry;
1563         pmap             *copied_graphs;
1564         pmap_entry       *pm_entry;
1565
1566         rem = current_ir_graph;
1567         obstack_init(&temp_obst);
1568
1569         /* a map for the copied graphs, used to inline recursive calls */
1570         copied_graphs = pmap_create();
1571
1572         /* extend all irgs by a temporary data structure for inlining. */
1573         n_irgs = get_irp_n_irgs();
1574         for (i = 0; i < n_irgs; ++i)
1575                 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1576
1577         /* Precompute information in temporary data structure. */
1578         wenv.ignore_runtime = ignore_runtime;
1579         wenv.ignore_callers = 0;
1580         for (i = 0; i < n_irgs; ++i) {
1581                 ir_graph *irg = get_irp_irg(i);
1582
1583                 assert(get_irg_phase_state(irg) != phase_building);
1584                 free_callee_info(irg);
1585
1586                 assure_cf_loop(irg);
1587                 wenv.x = get_irg_link(irg);
1588                 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1589         }
1590
1591         /* -- and now inline. -- */
1592
1593         /* Inline leaves recursively -- we might construct new leaves. */
1594         do {
1595                 did_inline = 0;
1596
1597                 for (i = 0; i < n_irgs; ++i) {
1598                         ir_node *call;
1599                         int phiproj_computed = 0;
1600
1601                         current_ir_graph = get_irp_irg(i);
1602                         env = (inline_irg_env *)get_irg_link(current_ir_graph);
1603
1604                         tail = NULL;
1605                         for (entry = env->call_head; entry != NULL; entry = entry->next) {
1606                                 ir_graph *callee;
1607
1608                                 if (env->n_nodes > maxsize) break;
1609
1610                                 call   = entry->call;
1611                                 callee = entry->callee;
1612
1613                                 if (is_leave(callee) && (
1614                                     is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1615                                         if (!phiproj_computed) {
1616                                                 phiproj_computed = 1;
1617                                                 collect_phiprojs(current_ir_graph);
1618                                         }
1619                                         did_inline = inline_method(call, callee);
1620
1621                                         if (did_inline) {
1622                                                 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1623
1624                                                 /* was inlined, must be recomputed */
1625                                                 phiproj_computed = 0;
1626
1627                                                 /* Do some statistics */
1628                                                 env->got_inline = 1;
1629                                                 --env->n_call_nodes;
1630                                                 env->n_nodes += callee_env->n_nodes;
1631                                                 --callee_env->n_callers;
1632
1633                                                 /* remove this call from the list */
1634                                                 if (tail != NULL)
1635                                                         tail->next = entry->next;
1636                                                 else
1637                                                         env->call_head = entry->next;
1638                                                 continue;
1639                                         }
1640                                 }
1641                                 tail = entry;
1642                         }
1643                         env->call_tail = tail;
1644                 }
1645         } while (did_inline);
1646
1647         /* inline other small functions. */
1648         for (i = 0; i < n_irgs; ++i) {
1649                 ir_node *call;
1650                 int phiproj_computed = 0;
1651
1652                 current_ir_graph = get_irp_irg(i);
1653                 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1654
1655                 /* note that the list of possible calls is updated during the process */
1656                 tail = NULL;
1657                 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1658                         ir_graph   *callee;
1659                         pmap_entry *e;
1660
1661                         call   = entry->call;
1662                         callee = entry->callee;
1663
1664                         e = pmap_find(copied_graphs, callee);
1665                         if (e != NULL) {
1666                                 /*
1667                                  * Remap callee if we have a copy.
1668                                  * FIXME: Should we do this only for recursive Calls ?
1669                                  */
1670                                 callee = e->value;
1671                         }
1672
1673                         if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) ||    /* small function */
1674                                 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1675                                 if (current_ir_graph == callee) {
1676                                         /*
1677                                          * Recursive call: we cannot directly inline because we cannot walk
1678                                          * the graph and change it. So we have to make a copy of the graph
1679                                          * first.
1680                                          */
1681
1682                                         inline_irg_env *callee_env;
1683                                         ir_graph       *copy;
1684
1685                                         /*
1686                                          * No copy yet, create one.
1687                                          * Note that recursive methods are never leaves, so it is sufficient
1688                                          * to test this condition here.
1689                                          */
1690                                         copy = create_irg_copy(callee);
1691
1692                                         /* create_irg_copy() destroys the Proj links, recompute them */
1693                                         phiproj_computed = 0;
1694
1695                                         /* allocate new environment */
1696                                         callee_env = alloc_inline_irg_env();
1697                                         set_irg_link(copy, callee_env);
1698
1699                                         assure_cf_loop(copy);
1700                                         wenv.x              = callee_env;
1701                                         wenv.ignore_callers = 1;
1702                                         irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1703
1704                                         /*
1705                                          * Enter the entity of the original graph. This is needed
1706                                          * for inline_method(). However, note that ent->irg still points
1707                                          * to callee, NOT to copy.
1708                                          */
1709                                         set_irg_entity(copy, get_irg_entity(callee));
1710
1711                                         pmap_insert(copied_graphs, callee, copy);
1712                                         callee = copy;
1713
1714                                         /* we have only one caller: the original graph */
1715                                         callee_env->n_callers      = 1;
1716                                         callee_env->n_callers_orig = 1;
1717                                 }
1718                                 if (! phiproj_computed) {
1719                                         phiproj_computed = 1;
1720                                         collect_phiprojs(current_ir_graph);
1721                                 }
1722                                 did_inline = inline_method(call, callee);
1723                                 if (did_inline) {
1724                                         inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1725
1726                                         /* was inlined, must be recomputed */
1727                                         phiproj_computed = 0;
1728
1729                                         /* callee was inline. Append it's call list. */
1730                                         env->got_inline = 1;
1731                                         --env->n_call_nodes;
1732                                         append_call_list(env, callee_env->call_head);
1733                                         env->n_call_nodes += callee_env->n_call_nodes;
1734                                         env->n_nodes += callee_env->n_nodes;
1735                                         --callee_env->n_callers;
1736
1737                                         /* after we have inlined callee, all called methods inside callee
1738                                            are now called once more */
1739                                         for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1740                                                 inline_irg_env *penv = get_irg_link(centry->callee);
1741                                                 ++penv->n_callers;
1742                                         }
1743
1744                                         /* remove this call from the list */
1745                                         if (tail != NULL)
1746                                                 tail->next = entry->next;
1747                                         else
1748                                                 env->call_head = entry->next;
1749                                         continue;
1750                                 }
1751                         }
1752                         tail = entry;
1753                 }
1754                 env->call_tail = tail;
1755         }
1756
1757         for (i = 0; i < n_irgs; ++i) {
1758                 irg = get_irp_irg(i);
1759                 env = (inline_irg_env *)get_irg_link(irg);
1760
1761                 if (env->got_inline) {
1762                         optimize_graph_df(irg);
1763                         optimize_cf(irg);
1764                 }
1765                 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1766                         DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1767                         env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1768                         env->n_callers_orig, env->n_callers,
1769                         get_entity_name(get_irg_entity(irg))));
1770                 }
1771         }
1772
1773         /* kill the copied graphs: we don't need them anymore */
1774         foreach_pmap(copied_graphs, pm_entry) {
1775                 ir_graph *copy = pm_entry->value;
1776
1777                 /* reset the entity, otherwise it will be deleted in the next step ... */
1778                 set_irg_entity(copy, NULL);
1779                 free_ir_graph(copy);
1780         }
1781         pmap_destroy(copied_graphs);
1782
1783         obstack_free(&temp_obst, NULL);
1784         current_ir_graph = rem;
1785 }
1786
1787 /**
1788  * Calculate the parameter weights for transmitting the address of a local variable.
1789  */
1790 static unsigned calc_method_local_weight(ir_node *arg) {
1791         int      i, j, k;
1792         unsigned v, weight = 0;
1793
1794         for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1795                 ir_node *succ = get_irn_out(arg, i);
1796
1797                 switch (get_irn_opcode(succ)) {
1798                 case iro_Load:
1799                 case iro_Store:
1800                         /* Loads and Store can be removed */
1801                         weight += 3;
1802                         break;
1803                 case iro_Sel:
1804                         /* check if all args are constant */
1805                         for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1806                                 ir_node *idx = get_Sel_index(succ, j);
1807                                 if (! is_Const(idx))
1808                                         return 0;
1809                         }
1810                         /* Check users on this Sel. Note: if a 0 is returned here, there was
1811                            some unsupported node. */
1812                         v = calc_method_local_weight(succ);
1813                         if (v == 0)
1814                                 return 0;
1815                         /* we can kill one Sel with constant indexes, this is cheap */
1816                         weight += v + 1;
1817                         break;
1818                 case iro_Id:
1819                         /* when looking backward we might find Id nodes */
1820                         weight += calc_method_local_weight(succ);
1821                         break;
1822                 case iro_Tuple:
1823                         /* unoptimized tuple */
1824                         for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1825                                 ir_node *pred = get_Tuple_pred(succ, j);
1826                                 if (pred == arg) {
1827                                         /* look for Proj(j) */
1828                                         for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1829                                                 ir_node *succ_succ = get_irn_out(succ, k);
1830                                                 if (is_Proj(succ_succ)) {
1831                                                         if (get_Proj_proj(succ_succ) == j) {
1832                                                                 /* found */
1833                                                                 weight += calc_method_local_weight(succ_succ);
1834                                                         }
1835                                                 } else {
1836                                                         /* this should NOT happen */
1837                                                         return 0;
1838                                                 }
1839                                         }
1840                                 }
1841                         }
1842                         break;
1843                 default:
1844                         /* any other node: unsupported yet or bad. */
1845                         return 0;
1846                 }
1847         }
1848         return weight;
1849 }
1850
1851 /**
1852  * Calculate the parameter weights for transmitting the address of a local variable.
1853  */
1854 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1855         ir_entity *ent = get_irg_entity(irg);
1856         ir_type  *mtp;
1857         int      nparams, i, proj_nr;
1858         ir_node  *irg_args, *arg;
1859
1860         mtp      = get_entity_type(ent);
1861         nparams  = get_method_n_params(mtp);
1862
1863         /* allocate a new array. currently used as 'analysed' flag */
1864         env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1865
1866         /* If the method haven't parameters we have nothing to do. */
1867         if (nparams <= 0)
1868                 return;
1869
1870         assure_irg_outs(irg);
1871         irg_args = get_irg_args(irg);
1872         for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1873                 arg     = get_irn_out(irg_args, i);
1874                 proj_nr = get_Proj_proj(arg);
1875                 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1876         }
1877 }
1878
1879 /**
1880  * Calculate the benefice for transmitting an local variable address.
1881  * After inlining, the local variable might be transformed into a
1882  * SSA variable by scalar_replacement().
1883  */
1884 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1885         inline_irg_env *env = get_irg_link(callee);
1886
1887         if (env->local_weights != NULL) {
1888                 if (pos < ARR_LEN(env->local_weights))
1889                         return env->local_weights[pos];
1890                 return 0;
1891         }
1892
1893         analyze_irg_local_weights(env, callee);
1894
1895         if (pos < ARR_LEN(env->local_weights))
1896                 return env->local_weights[pos];
1897         return 0;
1898 }
1899
1900 /**
1901  * Calculate a benefice value for inlining the given call.
1902  *
1903  * @param call       the call node we have to inspect
1904  * @param callee     the called graph
1905  * @param local_adr  set after return if an address of a local variable is
1906  *                   transmitted as a parameter
1907  */
1908 static int calc_inline_benefice(ir_node *call, ir_graph *callee, unsigned *local_adr) {
1909         ir_entity *ent = get_irg_entity(callee);
1910         ir_node   *frame_ptr;
1911         ir_type   *mtp;
1912         int       weight = 0;
1913         int       i, n_params, all_const;
1914         unsigned  cc, v;
1915
1916         inline_irg_env *curr_env, *callee_env;
1917
1918         if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1919                 /* do NOT inline noreturn calls */
1920                 return INT_MIN;
1921         }
1922
1923         /* costs for every passed parameter */
1924         n_params = get_Call_n_params(call);
1925         mtp      = get_entity_type(ent);
1926         cc       = get_method_calling_convention(mtp);
1927         if (cc & cc_reg_param) {
1928                 /* register parameter, smaller costs for register parameters */
1929                 int max_regs = cc & ~cc_bits;
1930
1931                 if (max_regs < n_params)
1932                         weight += max_regs * 2 + (n_params - max_regs) * 5;
1933                 else
1934                         weight += n_params * 2;
1935         } else {
1936                 /* parameters are passed an stack */
1937                 weight += 5 * n_params;
1938         }
1939
1940         /* constant parameters improve the benefice */
1941         frame_ptr = get_irg_frame(current_ir_graph);
1942         all_const = 1;
1943         for (i = 0; i < n_params; ++i) {
1944                 ir_node *param = get_Call_param(call, i);
1945
1946                 if (is_Const(param)) {
1947                         weight += get_method_param_weight(ent, i);
1948                 } else {
1949                         all_const = 0;
1950                         if (is_SymConst(param))
1951                                 weight += get_method_param_weight(ent, i);
1952                         else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1953                                 /*
1954                                  * An address of a local variable is transmitted. After inlining,
1955                                  * scalar_replacement might be able to remove the local variable,
1956                                  * so honor this.
1957                                  */
1958                                 v = get_method_local_adress_weight(callee, i);
1959                                 weight += v;
1960                                 if (v > 0)
1961                                         *local_adr = 1;
1962                         }
1963                 }
1964         }
1965
1966         callee_env = get_irg_link(callee);
1967         if (get_entity_visibility(ent) == visibility_local &&
1968             callee_env->n_callers_orig == 1 &&
1969             callee != current_ir_graph) {
1970                 /* we are the only caller, give big bonus */
1971                 weight += 5000;
1972         }
1973
1974         /* do not inline big functions */
1975         weight -= callee_env->n_nodes;
1976
1977         /* reduce the benefice if the current function is already big */
1978         curr_env = get_irg_link(current_ir_graph);
1979         weight -= curr_env->n_nodes / 50;
1980
1981         /* give a bonus for functions with one block */
1982         if (callee_env->n_blocks == 1)
1983                 weight = weight * 3 / 2;
1984
1985         /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1986         else if (callee_env->n_nodes < 20 && !callee_env->recursive)
1987                 weight += 5000;
1988
1989         /* and finally for leaves: they do not increase the register pressure
1990            because of callee safe registers */
1991         else if (callee_env->n_call_nodes == 0)
1992                 weight += 25;
1993
1994         /*
1995          * Reduce the weight for recursive function IFF not all arguments are const.
1996          * inlining recursive functions is rarely good.
1997          */
1998         if (callee_env->recursive && !all_const)
1999                 weight -= 500;
2000
2001         /*
2002          * All arguments constant is probably a good sign, give an extra bonus
2003          */
2004         if (all_const)
2005                 weight += 100;
2006
2007         return weight;
2008 }
2009
2010 /**
2011  * Heuristic inliner. Calculates a benefice value for every call and inlines
2012  * those calls with a value higher than the threshold.
2013  */
2014 void inline_functions(int maxsize, int inline_threshold) {
2015         inline_irg_env   *env;
2016         int              i, n_irgs;
2017         ir_graph         *rem;
2018         int              did_inline;
2019         wenv_t           wenv;
2020         call_entry       *curr_call, **last_call;
2021         const call_entry *centry;
2022         pmap             *copied_graphs;
2023         pmap_entry       *pm_entry;
2024
2025         rem = current_ir_graph;
2026         obstack_init(&temp_obst);
2027
2028         /* a map for the copied graphs, used to inline recursive calls */
2029         copied_graphs = pmap_create();
2030
2031         /* extend all irgs by a temporary data structure for inlining. */
2032         n_irgs = get_irp_n_irgs();
2033         for (i = 0; i < n_irgs; ++i)
2034                 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
2035
2036         /* Precompute information in temporary data structure. */
2037         wenv.ignore_runtime = 0;
2038         wenv.ignore_callers = 0;
2039         for (i = 0; i < n_irgs; ++i) {
2040                 ir_graph *irg = get_irp_irg(i);
2041
2042                 assert(get_irg_phase_state(irg) != phase_building);
2043                 free_callee_info(irg);
2044
2045                 wenv.x         = get_irg_link(irg);
2046                 wenv.last_call = NULL;
2047                 assure_cf_loop(irg);
2048                 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
2049         }
2050
2051         /* -- and now inline. -- */
2052         for (i = 0; i < n_irgs; ++i) {
2053                 int      phiproj_computed = 0;
2054                 ir_node  *call;
2055                 ir_graph *irg = get_irp_irg(i);
2056
2057                 current_ir_graph = irg;
2058                 env = get_irg_link(irg);
2059
2060                 /* note that the list of possible calls is updated during the process */
2061                 last_call = &env->call_head;
2062                 for (curr_call = env->call_head; curr_call != NULL;) {
2063                         ir_graph   *callee;
2064                         pmap_entry *e;
2065                         int        benefice;
2066                         unsigned   local_adr;
2067
2068                         if (env->n_nodes > maxsize) break;
2069
2070                         call   = curr_call->call;
2071                         callee = curr_call->callee;
2072
2073                         e = pmap_find(copied_graphs, callee);
2074                         if (e != NULL) {
2075                                 /*
2076                                 * Remap callee if we have a copy.
2077                                 * FIXME: Should we do this only for recursive Calls ?
2078                                 */
2079                                 callee = e->value;
2080                         }
2081
2082                         /* calculate the benefice on the original call to prevent excessive inlining */
2083                         local_adr = 0;
2084                         benefice = calc_inline_benefice(call, callee, &local_adr);
2085                         DB((dbg, LEVEL_2, "In %+F Call %+F has benefice %d\n", irg, callee, benefice));
2086
2087                         if (benefice > -inline_threshold ||
2088                                 (get_irg_inline_property(callee) >= irg_inline_forced)) {
2089                                 if (current_ir_graph == callee) {
2090                                         /*
2091                                          * Recursive call: we cannot directly inline because we cannot walk
2092                                          * the graph and change it. So we have to make a copy of the graph
2093                                          * first.
2094                                          */
2095
2096                                         inline_irg_env *callee_env;
2097                                         ir_graph       *copy;
2098
2099                                         /*
2100                                          * No copy yet, create one.
2101                                          * Note that recursive methods are never leaves, so it is sufficient
2102                                          * to test this condition here.
2103                                          */
2104                                         copy = create_irg_copy(callee);
2105
2106                                         /* create_irg_copy() destroys the Proj links, recompute them */
2107                                         phiproj_computed = 0;
2108
2109                                         /* allocate new environment */
2110                                         callee_env = alloc_inline_irg_env();
2111                                         set_irg_link(copy, callee_env);
2112
2113                                         assure_cf_loop(copy);
2114                                         wenv.x              = callee_env;
2115                                         wenv.ignore_callers = 1;
2116                                         irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2117
2118                                         /*
2119                                          * Enter the entity of the original graph. This is needed
2120                                          * for inline_method(). However, note that ent->irg still points
2121                                          * to callee, NOT to copy.
2122                                          */
2123                                         set_irg_entity(copy, get_irg_entity(callee));
2124
2125                                         pmap_insert(copied_graphs, callee, copy);
2126                                         callee = copy;
2127
2128                                         /* we have only one caller: the original graph */
2129                                         callee_env->n_callers      = 1;
2130                                         callee_env->n_callers_orig = 1;
2131                                 }
2132                                 if (! phiproj_computed) {
2133                                         phiproj_computed = 1;
2134                                         collect_phiprojs(current_ir_graph);
2135                                 }
2136                                 did_inline = inline_method(call, callee);
2137                                 if (did_inline) {
2138                                         inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
2139
2140                                         /* was inlined, must be recomputed */
2141                                         phiproj_computed = 0;
2142
2143                                         /* after we have inlined callee, all called methods inside callee
2144                                         are now called once more */
2145                                         for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
2146                                                 inline_irg_env *penv = get_irg_link(centry->callee);
2147                                                 ++penv->n_callers;
2148                                         }
2149
2150                                         /* callee was inline. Append it's call list. */
2151                                         env->got_inline = 1;
2152                                         if (local_adr)
2153                                                 env->local_vars = 1;
2154                                         --env->n_call_nodes;
2155                                         curr_call = replace_entry_by_call_list(curr_call, callee_env->call_head);
2156                                         env->n_call_nodes += callee_env->n_call_nodes;
2157                                         env->n_nodes += callee_env->n_nodes;
2158                                         --callee_env->n_callers;
2159
2160                                         /* remove the current call entry from the list */
2161                                         *last_call = curr_call;
2162                                         continue;
2163                                 }
2164                         }
2165                         last_call = &curr_call->next;
2166                         curr_call = curr_call->next;
2167                 }
2168
2169         }
2170
2171         for (i = 0; i < n_irgs; ++i) {
2172                 ir_graph *irg = get_irp_irg(i);
2173
2174                 env = get_irg_link(irg);
2175                 if (env->got_inline) {
2176                         /* this irg got calls inlined: optimize it */
2177
2178                         /* scalar replacement does not work well with Tuple nodes, so optimize them away */
2179                         optimize_graph_df(irg);
2180
2181                         if (env->local_vars) {
2182                                 if (scalar_replacement_opt(irg)) {
2183                                         optimize_graph_df(irg);
2184                                 }
2185                         }
2186                         optimize_cf(irg);
2187                 }
2188                 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2189                         DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2190                         env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2191                         env->n_callers_orig, env->n_callers,
2192                         get_entity_name(get_irg_entity(irg))));
2193                 }
2194         }
2195
2196         /* kill the copied graphs: we don't need them anymore */
2197         foreach_pmap(copied_graphs, pm_entry) {
2198                 ir_graph *copy = pm_entry->value;
2199
2200                 /* reset the entity, otherwise it will be deleted in the next step ... */
2201                 set_irg_entity(copy, NULL);
2202                 free_ir_graph(copy);
2203         }
2204         pmap_destroy(copied_graphs);
2205
2206         obstack_free(&temp_obst, NULL);
2207         current_ir_graph = rem;
2208 }
2209
2210 void firm_init_inline(void) {
2211         FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
2212 }