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

   1 /*
   2  * Copyright (C) 1995-2011 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  */
  25 #include "config.h"
  26
  27 #include <limits.h>
  28 #include <stdbool.h>
  29 #include <assert.h>
  30
  31 #include "irnode_t.h"
  32 #include "irgraph_t.h"
  33 #include "irprog_t.h"
  34
  35 #include "iroptimize.h"
  36 #include "ircons_t.h"
  37 #include "iropt_t.h"
  38 #include "irgopt.h"
  39 #include "irgmod.h"
  40 #include "irgwalk.h"
  41
  42 #include "array_t.h"
  43 #include "list.h"
  44 #include "pset.h"
  45 #include "pmap.h"
  46 #include "pdeq.h"
  47 #include "xmalloc.h"
  48 #include "pqueue.h"
  49
  50 #include "irouts.h"
  51 #include "irloop_t.h"
  52 #include "irbackedge_t.h"
  53 #include "opt_init.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 #include "irpass_t.h"
  65 #include "irnodemap.h"
  66
  67 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  68
  69 /*------------------------------------------------------------------*/
  70 /* Routines for dead node elimination / copying garbage collection  */
  71 /* of the obstack.                                                  */
  72 /*------------------------------------------------------------------*/
  73
  74 /**
  75  * Remember the new node in the old node by using a field all nodes have.
  76  */
  77 static void set_new_node(ir_node *node, ir_node *new_node)
  78 {
  79         set_irn_link(node, new_node);
  80 }
  81
  82 /**
  83  * Get this new node, before the old node is forgotten.
  84  */
  85 static inline ir_node *get_new_node(ir_node *old_node)
  86 {
  87         assert(irn_visited(old_node));
  88         return (ir_node*) get_irn_link(old_node);
  89 }
  90
  91 /*--------------------------------------------------------------------*/
  92 /*  Functionality for inlining                                         */
  93 /*--------------------------------------------------------------------*/
  94
  95 /**
  96  * Copy node for inlineing.  Updates attributes that change when
  97  * inlineing but not for dead node elimination.
  98  *
  99  * Copies the node by calling copy_node() and then updates the entity if
 100  * it's a local one.  env must be a pointer of the frame type of the
 101  * inlined procedure. The new entities must be in the link field of
 102  * the entities.
 103  */
 104 static void copy_node_inline(ir_node *node, void *env)
 105 {
 106         ir_graph *new_irg  = (ir_graph*) env;
 107         ir_node  *new_node = irn_copy_into_irg(node, new_irg);
 108
 109         set_new_node(node, new_node);
 110         if (is_Sel(node)) {
 111                 ir_graph  *old_irg        = get_irn_irg(node);
 112                 ir_type   *old_frame_type = get_irg_frame_type(old_irg);
 113                 ir_entity *old_entity     = get_Sel_entity(node);
 114                 assert(is_Sel(new_node));
 115                 /* use copied entities from the new frame */
 116                 if (get_entity_owner(old_entity) == old_frame_type) {
 117                         ir_entity *new_entity = (ir_entity*)get_entity_link(old_entity);
 118                         assert(new_entity != NULL);
 119                         set_Sel_entity(new_node, new_entity);
 120                 }
 121         } else if (is_Block(new_node)) {
 122                 new_node->attr.block.irg.irg = new_irg;
 123         }
 124 }
 125
 126 static void set_preds_inline(ir_node *node, void *env)
 127 {
 128         ir_node *new_node;
 129
 130         irn_rewire_inputs(node);
 131
 132         /* move constants into start block */
 133         new_node = get_new_node(node);
 134         if (is_irn_constlike(new_node)) {
 135                 ir_graph *new_irg     = (ir_graph *) env;
 136                 ir_node  *start_block = get_irg_start_block(new_irg);
 137                 set_nodes_block(new_node, start_block);
 138         }
 139 }
 140
 141 /**
 142  * Walker: checks if P_value_arg_base is used.
 143  */
 144 static void find_addr(ir_node *node, void *env)
 145 {
 146         bool *allow_inline = (bool*)env;
 147
 148         if (is_Block(node) && get_Block_entity(node)) {
 149                 /**
 150                  * Currently we can't handle blocks whose address was taken correctly
 151                  * when inlining
 152                  */
 153                 *allow_inline = false;
 154         } else if (is_Sel(node)) {
 155                 ir_graph *irg = current_ir_graph;
 156                 if (get_Sel_ptr(node) == get_irg_frame(irg)) {
 157                         /* access to frame */
 158                         ir_entity *ent = get_Sel_entity(node);
 159                         if (get_entity_owner(ent) != get_irg_frame_type(irg)) {
 160                                 /* access to value_type */
 161                                 *allow_inline = false;
 162                         }
 163                         if (is_parameter_entity(ent)) {
 164                                 *allow_inline = false;
 165                         }
 166                 }
 167         } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
 168                 /* From GCC:
 169                  * Refuse to inline alloca call unless user explicitly forced so as this
 170                  * may change program's memory overhead drastically when the function
 171                  * using alloca is called in loop.  In GCC present in SPEC2000 inlining
 172                  * into schedule_block cause it to require 2GB of ram instead of 256MB.
 173                  *
 174                  * Sorrily this is true with our implementation also.
 175                  * Moreover, we cannot differentiate between alloca() and VLA yet, so
 176                  * this disables inlining of functions using VLA (which are completely
 177                  * save).
 178                  *
 179                  * 2 Solutions:
 180                  * - add a flag to the Alloc node for "real" alloca() calls
 181                  * - add a new Stack-Restore node at the end of a function using
 182                  *   alloca()
 183                  */
 184                 *allow_inline = false;
 185         }
 186 }
 187
 188 /**
 189  * Check if we can inline a given call.
 190  * Currently, we cannot inline two cases:
 191  * - call with compound arguments
 192  * - graphs that take the address of a parameter
 193  *
 194  * check these conditions here
 195  */
 196 static bool can_inline(ir_node *call, ir_graph *called_graph)
 197 {
 198         ir_entity          *called      = get_irg_entity(called_graph);
 199         ir_type            *called_type = get_entity_type(called);
 200         ir_type            *call_type   = get_Call_type(call);
 201         size_t              n_params    = get_method_n_params(called_type);
 202         size_t              n_arguments = get_method_n_params(call_type);
 203         size_t              n_res       = get_method_n_ress(called_type);
 204         mtp_additional_properties props = get_entity_additional_properties(called);
 205         size_t              i;
 206         bool                res;
 207
 208         if (props & mtp_property_noinline)
 209                 return false;
 210
 211         if (n_arguments != n_params) {
 212                 /* this is a bad feature of C: without a prototype, we can
 213                  * call a function with less parameters than needed. Currently
 214                  * we don't support this, although we could use Unknown than. */
 215                 return false;
 216         }
 217         if (n_res != get_method_n_ress(call_type)) {
 218                 return false;
 219         }
 220
 221         /* Argh, compiling C has some bad consequences:
 222          * It is implementation dependent what happens in that case.
 223          * We support inlining, if the bitsize of the types matches AND
 224          * the same arithmetic is used. */
 225         for (i = 0; i < n_params; ++i) {
 226                 ir_type *param_tp = get_method_param_type(called_type, i);
 227                 ir_type *arg_tp   = get_method_param_type(call_type, i);
 228
 229                 if (param_tp != arg_tp) {
 230                         ir_mode *pmode = get_type_mode(param_tp);
 231                         ir_mode *amode = get_type_mode(arg_tp);
 232
 233                         if (pmode == NULL || amode == NULL)
 234                                 return false;
 235                         if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
 236                                 return false;
 237                         if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
 238                                 return false;
 239                         /* otherwise we can simply "reinterpret" the bits */
 240                 }
 241         }
 242         for (i = 0; i < n_res; ++i) {
 243                 ir_type *decl_res_tp = get_method_res_type(called_type, i);
 244                 ir_type *used_res_tp = get_method_res_type(call_type, i);
 245
 246                 if (decl_res_tp != used_res_tp) {
 247                         ir_mode *decl_mode = get_type_mode(decl_res_tp);
 248                         ir_mode *used_mode = get_type_mode(used_res_tp);
 249                         if (decl_mode == NULL || used_mode == NULL)
 250                                 return false;
 251                         if (get_mode_size_bits(decl_mode) != get_mode_size_bits(used_mode))
 252                                 return false;
 253                         if (get_mode_arithmetic(decl_mode) != get_mode_arithmetic(used_mode))
 254                                 return false;
 255                         /* otherwise we can "reinterpret" the bits */
 256                 }
 257         }
 258
 259         /* check parameters for compound arguments */
 260         for (i = 0; i < n_params; ++i) {
 261                 ir_type *p_type = get_method_param_type(call_type, i);
 262
 263                 if (is_compound_type(p_type))
 264                         return false;
 265         }
 266
 267         /* check results for compound arguments */
 268         for (i = 0; i < n_res; ++i) {
 269                 ir_type *r_type = get_method_res_type(call_type, i);
 270
 271                 if (is_compound_type(r_type))
 272                         return false;
 273         }
 274
 275         res = true;
 276         irg_walk_graph(called_graph, find_addr, NULL, &res);
 277
 278         return res;
 279 }
 280
 281 enum exc_mode {
 282         exc_handler,    /**< There is a handler. */
 283         exc_no_handler  /**< Exception handling not represented. */
 284 };
 285
 286 /**
 287  * copy all entities on the stack frame on 1 irg to the stackframe of another.
 288  * Sets entity links of the old entities to the copies
 289  */
 290 static void copy_frame_entities(ir_graph *from, ir_graph *to)
 291 {
 292         ir_type *from_frame = get_irg_frame_type(from);
 293         ir_type *to_frame   = get_irg_frame_type(to);
 294         size_t   n_members  = get_class_n_members(from_frame);
 295         size_t   i;
 296         assert(from_frame != to_frame);
 297
 298         for (i = 0; i < n_members; ++i) {
 299                 ir_entity *old_ent = get_class_member(from_frame, i);
 300                 ir_entity *new_ent = copy_entity_own(old_ent, to_frame);
 301                 set_entity_link(old_ent, new_ent);
 302                 assert (!is_parameter_entity(old_ent));
 303         }
 304 }
 305
 306 /* Inlines a method at the given call site. */
 307 int inline_method(ir_node *call, ir_graph *called_graph)
 308 {
 309         /* we cannot inline some types of calls */
 310         if (! can_inline(call, called_graph))
 311                 return 0;
 312
 313         /* We cannot inline a recursive call. The graph must be copied before
 314          * the call the inline_method() using create_irg_copy(). */
 315         ir_graph *irg = get_irn_irg(call);
 316         if (called_graph == irg)
 317                 return 0;
 318
 319         ir_entity *ent      = get_irg_entity(called_graph);
 320         ir_type   *mtp      = get_entity_type(ent);
 321         ir_type   *ctp      = get_Call_type(call);
 322         int        n_params = get_method_n_params(mtp);
 323
 324         ir_graph *rem = current_ir_graph;
 325         current_ir_graph = irg;
 326
 327         DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
 328
 329         /* optimizations can cause problems when allocating new nodes */
 330         int rem_opt = get_opt_optimize();
 331         set_optimize(0);
 332
 333         /* Handle graph state */
 334         assert(get_irg_pinned(irg) == op_pin_state_pinned);
 335         assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
 336         clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
 337                            | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
 338         set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
 339         clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
 340         edges_deactivate(irg);
 341
 342         /* here we know we WILL inline, so inform the statistics */
 343         hook_inline(call, called_graph);
 344
 345         /* -- Decide how to handle exception control flow: Is there a handler
 346            for the Call node, or do we branch directly to End on an exception?
 347            exc_handling:
 348            0 There is a handler.
 349            2 Exception handling not represented in Firm. -- */
 350         ir_node *Xproj = NULL;
 351         for (ir_node *proj = (ir_node*)get_irn_link(call); proj != NULL;
 352                  proj = (ir_node*)get_irn_link(proj)) {
 353                 long proj_nr = get_Proj_proj(proj);
 354                 if (proj_nr == pn_Call_X_except) Xproj = proj;
 355         }
 356         enum exc_mode exc_handling = Xproj != NULL ? exc_handler : exc_no_handler;
 357
 358         /* create the argument tuple */
 359         ir_node **args_in = ALLOCAN(ir_node*, n_params);
 360
 361         ir_node *block = get_nodes_block(call);
 362         for (int i = n_params - 1; i >= 0; --i) {
 363                 ir_node *arg      = get_Call_param(call, i);
 364                 ir_type *param_tp = get_method_param_type(mtp, i);
 365                 ir_mode *mode     = get_type_mode(param_tp);
 366
 367                 if (mode != get_irn_mode(arg)) {
 368                         arg = new_r_Conv(block, arg, mode);
 369                 }
 370                 args_in[i] = arg;
 371         }
 372
 373         /* the procedure and later replaces the Start node of the called graph.
 374          * Post_call is the old Call node and collects the results of the called
 375          * graph. Both will end up being a tuple. */
 376         ir_node *post_bl = get_nodes_block(call);
 377         /* XxMxPxPxPxT of Start + parameter of Call */
 378         ir_node *in[pn_Start_max+1];
 379         in[pn_Start_M]              = get_Call_mem(call);
 380         in[pn_Start_X_initial_exec] = new_r_Jmp(post_bl);
 381         in[pn_Start_P_frame_base]   = get_irg_frame(irg);
 382         in[pn_Start_T_args]         = new_r_Tuple(post_bl, n_params, args_in);
 383         ir_node *pre_call = new_r_Tuple(post_bl, pn_Start_max+1, in);
 384         ir_node *post_call = call;
 385
 386         /* --
 387            The new block gets the ins of the old block, pre_call and all its
 388            predecessors and all Phi nodes. -- */
 389         part_block(pre_call);
 390
 391         /* increment visited flag for later walk */
 392         inc_irg_visited(called_graph);
 393
 394         /* link some nodes to nodes in the current graph so instead of copying
 395          * the linked nodes will get used.
 396          * So the copier will use the created Tuple instead of copying the start
 397          * node, similar for singleton nodes like NoMem and Bad.
 398          * Note: this will prohibit predecessors to be copied - only do it for
 399          *       nodes without predecessors */
 400         ir_node *start_block = get_irg_start_block(called_graph);
 401         set_new_node(start_block, get_nodes_block(pre_call));
 402         mark_irn_visited(start_block);
 403
 404         ir_node *start = get_irg_start(called_graph);
 405         set_new_node(start, pre_call);
 406         mark_irn_visited(start);
 407
 408         ir_node *nomem = get_irg_no_mem(called_graph);
 409         set_new_node(nomem, get_irg_no_mem(irg));
 410         mark_irn_visited(nomem);
 411
 412         /* entitiy link is used to link entities on old stackframe to the
 413          * new stackframe */
 414         irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
 415
 416         /* copy entities and nodes */
 417         assert(!irn_visited(get_irg_end(called_graph)));
 418         copy_frame_entities(called_graph, irg);
 419         irg_walk_core(get_irg_end(called_graph), copy_node_inline, set_preds_inline,
 420                       irg);
 421
 422         irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
 423
 424         /* -- Merge the end of the inlined procedure with the call site -- */
 425         /* We will turn the old Call node into a Tuple with the following
 426            predecessors:
 427            -1:  Block of Tuple.
 428            0: Phi of all Memories of Return statements.
 429            1: Jmp from new Block that merges the control flow from all exception
 430            predecessors of the old end block.
 431            2: Tuple of all arguments.
 432            3: Phi of Exception memories.
 433            In case the old Call directly branches to End on an exception we don't
 434            need the block merging all exceptions nor the Phi of the exception
 435            memories.
 436         */
 437
 438         /* Precompute some values */
 439         ir_node *end_bl = get_new_node(get_irg_end_block(called_graph));
 440         ir_node *end    = get_new_node(get_irg_end(called_graph));
 441         int      arity  = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret  */
 442         int      n_res  = get_method_n_ress(get_Call_type(call));
 443
 444         ir_node **res_pred = XMALLOCN(ir_node*, n_res);
 445         ir_node **cf_pred  = XMALLOCN(ir_node*, arity);
 446
 447         /* archive keepalives */
 448         int irn_arity = get_irn_arity(end);
 449         for (int i = 0; i < irn_arity; i++) {
 450                 ir_node *ka = get_End_keepalive(end, i);
 451                 if (! is_Bad(ka))
 452                         add_End_keepalive(get_irg_end(irg), ka);
 453         }
 454
 455         /* replace Return nodes by Jump nodes */
 456         int n_ret = 0;
 457         for (int i = 0; i < arity; i++) {
 458                 ir_node *ret = get_Block_cfgpred(end_bl, i);
 459                 if (is_Return(ret)) {
 460                         ir_node *block = get_nodes_block(ret);
 461                         cf_pred[n_ret] = new_r_Jmp(block);
 462                         n_ret++;
 463                 }
 464         }
 465         set_irn_in(post_bl, n_ret, cf_pred);
 466
 467         /* build a Tuple for all results of the method.
 468          * add Phi node if there was more than one Return. */
 469         turn_into_tuple(post_call, pn_Call_max+1);
 470         /* First the Memory-Phi */
 471         int n_mem_phi = 0;
 472         for (int i = 0; i < arity; i++) {
 473                 ir_node *ret = get_Block_cfgpred(end_bl, i);
 474                 if (is_Return(ret)) {
 475                         cf_pred[n_mem_phi++] = get_Return_mem(ret);
 476                 }
 477                 /* memory output for some exceptions is directly connected to End */
 478                 if (is_Call(ret)) {
 479                         cf_pred[n_mem_phi++] = new_r_Proj(ret, mode_M, 3);
 480                 } else if (is_fragile_op(ret)) {
 481                         /* We rely that all cfops have the memory output at the same position. */
 482                         cf_pred[n_mem_phi++] = new_r_Proj(ret, mode_M, 0);
 483                 } else if (is_Raise(ret)) {
 484                         cf_pred[n_mem_phi++] = new_r_Proj(ret, mode_M, 1);
 485                 }
 486         }
 487         ir_node *phi = new_r_Phi(post_bl, n_mem_phi, cf_pred, mode_M);
 488         set_Tuple_pred(call, pn_Call_M, phi);
 489         /* Conserve Phi-list for further inlinings -- but might be optimized */
 490         if (get_nodes_block(phi) == post_bl) {
 491                 set_irn_link(phi, get_irn_link(post_bl));
 492                 set_irn_link(post_bl, phi);
 493         }
 494         /* Now the real results */
 495         if (n_res > 0) {
 496                 for (int j = 0; j < n_res; j++) {
 497                         ir_type *res_type = get_method_res_type(ctp, j);
 498                         ir_mode *res_mode = get_type_mode(res_type);
 499                         int n_ret = 0;
 500                         for (int i = 0; i < arity; i++) {
 501                                 ir_node *ret = get_Block_cfgpred(end_bl, i);
 502                                 if (is_Return(ret)) {
 503                                         ir_node *res = get_Return_res(ret, j);
 504                                         if (get_irn_mode(res) != res_mode) {
 505                                                 ir_node *block = get_nodes_block(res);
 506                                                 res = new_r_Conv(block, res, res_mode);
 507                                         }
 508                                         cf_pred[n_ret] = res;
 509                                         n_ret++;
 510                                 }
 511                         }
 512                         if (n_ret > 0) {
 513                                 phi = new_r_Phi(post_bl, n_ret, cf_pred, res_mode);
 514                         } else {
 515                                 phi = new_r_Bad(irg, res_mode);
 516                         }
 517                         res_pred[j] = phi;
 518                         /* Conserve Phi-list for further inlinings -- but might be optimized */
 519                         if (get_nodes_block(phi) == post_bl) {
 520                                 set_Phi_next(phi, get_Block_phis(post_bl));
 521                                 set_Block_phis(post_bl, phi);
 522                         }
 523                 }
 524                 ir_node *result_tuple = new_r_Tuple(post_bl, n_res, res_pred);
 525                 set_Tuple_pred(call, pn_Call_T_result, result_tuple);
 526         } else {
 527                 set_Tuple_pred(call, pn_Call_T_result, new_r_Bad(irg, mode_T));
 528         }
 529         /* handle the regular call */
 530         set_Tuple_pred(call, pn_Call_X_regular, new_r_Jmp(post_bl));
 531
 532         /* Finally the exception control flow.
 533            We have two possible situations:
 534            First if the Call branches to an exception handler:
 535            We need to add a Phi node to
 536            collect the memory containing the exception objects.  Further we need
 537            to add another block to get a correct representation of this Phi.  To
 538            this block we add a Jmp that resolves into the X output of the Call
 539            when the Call is turned into a tuple.
 540            Second: There is no exception edge. Just add all inlined exception
 541            branches to the End node.
 542          */
 543         if (exc_handling == exc_handler) {
 544                 int n_exc = 0;
 545                 for (int i = 0; i < arity; i++) {
 546                         ir_node *ret = get_Block_cfgpred(end_bl, i);
 547                         ir_node *irn = skip_Proj(ret);
 548                         if (is_fragile_op(irn) || is_Raise(irn)) {
 549                                 cf_pred[n_exc] = ret;
 550                                 ++n_exc;
 551                         }
 552                 }
 553                 if (n_exc > 0) {
 554                         if (n_exc == 1) {
 555                                 /* simple fix */
 556                                 set_Tuple_pred(call, pn_Call_X_except, cf_pred[0]);
 557                         } else {
 558                                 ir_node *block = new_r_Block(irg, n_exc, cf_pred);
 559                                 set_Tuple_pred(call, pn_Call_X_except, new_r_Jmp(block));
 560                         }
 561                 } else {
 562                         set_Tuple_pred(call, pn_Call_X_except, new_r_Bad(irg, mode_X));
 563                 }
 564         } else {
 565                 /* assert(exc_handling == 1 || no exceptions. ) */
 566                 int n_exc = 0;
 567                 for (int i = 0; i < arity; i++) {
 568                         ir_node *ret = get_Block_cfgpred(end_bl, i);
 569                         ir_node *irn = skip_Proj(ret);
 570
 571                         if (is_fragile_op(irn) || is_Raise(irn)) {
 572                                 cf_pred[n_exc] = ret;
 573                                 n_exc++;
 574                         }
 575                 }
 576                 ir_node  *main_end_bl       = get_irg_end_block(irg);
 577                 int       main_end_bl_arity = get_irn_arity(main_end_bl);
 578                 ir_node **end_preds         = XMALLOCN(ir_node*, n_exc+main_end_bl_arity);
 579
 580                 for (int i = 0; i < main_end_bl_arity; ++i)
 581                         end_preds[i] = get_irn_n(main_end_bl, i);
 582                 for (int i = 0; i < n_exc; ++i)
 583                         end_preds[main_end_bl_arity + i] = cf_pred[i];
 584                 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
 585                 set_Tuple_pred(call, pn_Call_X_except, new_r_Bad(irg, mode_X));
 586                 free(end_preds);
 587         }
 588         free(res_pred);
 589         free(cf_pred);
 590
 591         /* --  Turn CSE back on. -- */
 592         set_optimize(rem_opt);
 593         current_ir_graph = rem;
 594
 595         return 1;
 596 }
 597
 598 /********************************************************************/
 599 /* Apply inlining to small methods.                                 */
 600 /********************************************************************/
 601
 602 static struct obstack  temp_obst;
 603
 604 /** Represents a possible inlinable call in a graph. */
 605 typedef struct call_entry {
 606         ir_node    *call;       /**< The Call node. */
 607         ir_graph   *callee;     /**< The callee IR-graph. */
 608         list_head  list;        /**< List head for linking the next one. */
 609         int        loop_depth;  /**< The loop depth of this call. */
 610         int        benefice;    /**< The calculated benefice of this call. */
 611         unsigned   local_adr:1; /**< Set if this call gets an address of a local variable. */
 612         unsigned   all_const:1; /**< Set if this call has only constant parameters. */
 613 } call_entry;
 614
 615 /**
 616  * environment for inlining small irgs
 617  */
 618 typedef struct inline_env_t {
 619         struct obstack obst;  /**< An obstack where call_entries are allocated on. */
 620         list_head      calls; /**< The call entry list. */
 621 } inline_env_t;
 622
 623 /**
 624  * Returns the irg called from a Call node. If the irg is not
 625  * known, NULL is returned.
 626  *
 627  * @param call  the call node
 628  */
 629 static ir_graph *get_call_called_irg(ir_node *call)
 630 {
 631         ir_node *addr;
 632
 633         addr = get_Call_ptr(call);
 634         if (is_SymConst_addr_ent(addr)) {
 635                 ir_entity *ent = get_SymConst_entity(addr);
 636                 /* we don't know which function gets finally bound to a weak symbol */
 637                 if (get_entity_linkage(ent) & IR_LINKAGE_WEAK)
 638                         return NULL;
 639
 640                 return get_entity_irg(ent);
 641         }
 642
 643         return NULL;
 644 }
 645
 646 /**
 647  * Walker: Collect all calls to known graphs inside a graph.
 648  */
 649 static void collect_calls(ir_node *call, void *env)
 650 {
 651         (void) env;
 652         if (is_Call(call)) {
 653                 ir_graph *called_irg = get_call_called_irg(call);
 654
 655                 if (called_irg != NULL) {
 656                         /* The Call node calls a locally defined method.  Remember to inline. */
 657                         inline_env_t *ienv  = (inline_env_t*)env;
 658                         call_entry   *entry = OALLOC(&ienv->obst, call_entry);
 659                         entry->call       = call;
 660                         entry->callee     = called_irg;
 661                         entry->loop_depth = 0;
 662                         entry->benefice   = 0;
 663                         entry->local_adr  = 0;
 664                         entry->all_const  = 0;
 665
 666                         list_add_tail(&entry->list, &ienv->calls);
 667                 }
 668         }
 669 }
 670
 671 /**
 672  * Inlines all small methods at call sites where the called address comes
 673  * from a Const node that references the entity representing the called
 674  * method.
 675  * The size argument is a rough measure for the code size of the method:
 676  * Methods where the obstack containing the firm graph is smaller than
 677  * size are inlined.
 678  */
 679 void inline_small_irgs(ir_graph *irg, int size)
 680 {
 681         ir_graph *rem = current_ir_graph;
 682         inline_env_t env;
 683
 684         current_ir_graph = irg;
 685         free_callee_info(irg);
 686
 687         /* Find Call nodes to inline.
 688            (We can not inline during a walk of the graph, as inlining the same
 689            method several times changes the visited flag of the walked graph:
 690            after the first inlining visited of the callee equals visited of
 691            the caller.  With the next inlining both are increased.) */
 692         obstack_init(&env.obst);
 693         INIT_LIST_HEAD(&env.calls);
 694         irg_walk_graph(irg, NULL, collect_calls, &env);
 695
 696         if (! list_empty(&env.calls)) {
 697                 /* There are calls to inline */
 698                 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
 699                 collect_phiprojs(irg);
 700
 701                 list_for_each_entry(call_entry, entry, &env.calls, list) {
 702                         ir_graph  *callee = entry->callee;
 703                         ir_entity *called = get_irg_entity(callee);
 704                         mtp_additional_properties props
 705                                 = get_entity_additional_properties(called);
 706
 707                         if (props & mtp_property_noinline)
 708                                 continue;
 709
 710                         if ((props & mtp_property_always_inline) ||
 711                             _obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst) < size) {
 712                                 inline_method(entry->call, callee);
 713                         }
 714                 }
 715                 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
 716         }
 717         obstack_free(&env.obst, NULL);
 718         current_ir_graph = rem;
 719 }
 720
 721 typedef struct inline_small_irgs_pass_t {
 722         ir_graph_pass_t pass;
 723         int            size;
 724 } inline_small_irgs_pass_t;
 725
 726 /**
 727  * Wrapper to run inline_small_irgs() as a pass.
 728  */
 729 static int inline_small_irgs_wrapper(ir_graph *irg, void *context)
 730 {
 731         inline_small_irgs_pass_t *pass = (inline_small_irgs_pass_t*)context;
 732
 733         inline_small_irgs(irg, pass->size);
 734         return 0;
 735 }
 736
 737 /* create a pass for inline_small_irgs() */
 738 ir_graph_pass_t *inline_small_irgs_pass(const char *name, int size)
 739 {
 740         inline_small_irgs_pass_t *pass = XMALLOCZ(inline_small_irgs_pass_t);
 741
 742         pass->size = size;
 743         return def_graph_pass_constructor(
 744                 &pass->pass, name ? name : "inline_small_irgs", inline_small_irgs_wrapper);
 745 }
 746
 747 /**
 748  * Environment for inlining irgs.
 749  */
 750 typedef struct {
 751         list_head calls;             /**< List of of all call nodes in this graph. */
 752         unsigned  *local_weights;    /**< Once allocated, the beneficial weight for transmitting local addresses. */
 753         unsigned  n_nodes;           /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
 754         unsigned  n_blocks;          /**< Number of Blocks in graph without Start and End block. */
 755         unsigned  n_nodes_orig;      /**< for statistics */
 756         unsigned  n_call_nodes;      /**< Number of Call nodes in the graph. */
 757         unsigned  n_call_nodes_orig; /**< for statistics */
 758         unsigned  n_callers;         /**< Number of known graphs that call this graphs. */
 759         unsigned  n_callers_orig;    /**< for statistics */
 760         unsigned  got_inline:1;      /**< Set, if at least one call inside this graph was inlined. */
 761         unsigned  recursive:1;       /**< Set, if this function is self recursive. */
 762 } inline_irg_env;
 763
 764 /**
 765  * Allocate a new environment for inlining.
 766  */
 767 static inline_irg_env *alloc_inline_irg_env(void)
 768 {
 769         inline_irg_env *env    = OALLOC(&temp_obst, inline_irg_env);
 770         INIT_LIST_HEAD(&env->calls);
 771         env->local_weights     = NULL;
 772         env->n_nodes           = -2; /* do not count count Start, End */
 773         env->n_blocks          = -2; /* do not count count Start, End Block */
 774         env->n_nodes_orig      = -2; /* do not count Start, End */
 775         env->n_call_nodes      = 0;
 776         env->n_call_nodes_orig = 0;
 777         env->n_callers         = 0;
 778         env->n_callers_orig    = 0;
 779         env->got_inline        = 0;
 780         env->recursive         = 0;
 781         return env;
 782 }
 783
 784 typedef struct walker_env {
 785         inline_irg_env *x;     /**< the inline environment */
 786         char ignore_runtime;   /**< the ignore runtime flag */
 787         char ignore_callers;   /**< if set, do change callers data */
 788 } wenv_t;
 789
 790 /**
 791  * post-walker: collect all calls in the inline-environment
 792  * of a graph and sum some statistics.
 793  */
 794 static void collect_calls2(ir_node *call, void *ctx)
 795 {
 796         wenv_t         *env = (wenv_t*)ctx;
 797         inline_irg_env *x = env->x;
 798         unsigned        code = get_irn_opcode(call);
 799         ir_graph       *callee;
 800         call_entry     *entry;
 801
 802         /* count meaningful nodes in irg */
 803         if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
 804                 if (code != iro_Block) {
 805                         ++x->n_nodes;
 806                         ++x->n_nodes_orig;
 807                 } else {
 808                         ++x->n_blocks;
 809                 }
 810         }
 811
 812         if (code != iro_Call) return;
 813
 814         /* check, if it's a runtime call */
 815         if (env->ignore_runtime) {
 816                 ir_node *symc = get_Call_ptr(call);
 817
 818                 if (is_SymConst_addr_ent(symc)) {
 819                         ir_entity *ent = get_SymConst_entity(symc);
 820
 821                         if (get_entity_additional_properties(ent) & mtp_property_runtime)
 822                                 return;
 823                 }
 824         }
 825
 826         /* collect all call nodes */
 827         ++x->n_call_nodes;
 828         ++x->n_call_nodes_orig;
 829
 830         callee = get_call_called_irg(call);
 831         if (callee != NULL) {
 832                 if (! env->ignore_callers) {
 833                         inline_irg_env *callee_env = (inline_irg_env*)get_irg_link(callee);
 834                         /* count all static callers */
 835                         ++callee_env->n_callers;
 836                         ++callee_env->n_callers_orig;
 837                 }
 838                 if (callee == current_ir_graph)
 839                         x->recursive = 1;
 840
 841                 /* link it in the list of possible inlinable entries */
 842                 entry = OALLOC(&temp_obst, call_entry);
 843                 entry->call       = call;
 844                 entry->callee     = callee;
 845                 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
 846                 entry->benefice   = 0;
 847                 entry->local_adr  = 0;
 848                 entry->all_const  = 0;
 849
 850                 list_add_tail(&entry->list, &x->calls);
 851         }
 852 }
 853
 854 /**
 855  * Returns TRUE if the number of callers is 0 in the irg's environment,
 856  * hence this irg is a leaf.
 857  */
 858 inline static int is_leaf(ir_graph *irg)
 859 {
 860         inline_irg_env *env = (inline_irg_env*)get_irg_link(irg);
 861         return env->n_call_nodes == 0;
 862 }
 863
 864 /**
 865  * Returns TRUE if the number of nodes in the callee is
 866  * smaller then size in the irg's environment.
 867  */
 868 inline static int is_smaller(ir_graph *callee, unsigned size)
 869 {
 870         inline_irg_env *env = (inline_irg_env*)get_irg_link(callee);
 871         return env->n_nodes < size;
 872 }
 873
 874 /**
 875  * Duplicate a call entry.
 876  *
 877  * @param entry     the original entry to duplicate
 878  * @param new_call  the new call node
 879  * @param loop_depth_delta
 880  *                  delta value for the loop depth
 881  */
 882 static call_entry *duplicate_call_entry(const call_entry *entry,
 883                                         ir_node *new_call, int loop_depth_delta)
 884 {
 885         call_entry *nentry = OALLOC(&temp_obst, call_entry);
 886         nentry->call       = new_call;
 887         nentry->callee     = entry->callee;
 888         nentry->benefice   = entry->benefice;
 889         nentry->loop_depth = entry->loop_depth + loop_depth_delta;
 890         nentry->local_adr  = entry->local_adr;
 891         nentry->all_const  = entry->all_const;
 892
 893         return nentry;
 894 }
 895
 896 /**
 897  * Append all call nodes of the source environment to the nodes of in the destination
 898  * environment.
 899  *
 900  * @param dst         destination environment
 901  * @param src         source environment
 902  * @param loop_depth  the loop depth of the call that is replaced by the src list
 903  */
 904 static void append_call_list(inline_irg_env *dst, inline_irg_env *src, int loop_depth)
 905 {
 906         call_entry *nentry;
 907
 908         /* Note that the src list points to Call nodes in the inlined graph, but
 909            we need Call nodes in our graph. Luckily the inliner leaves this information
 910            in the link field. */
 911         list_for_each_entry(call_entry, entry, &src->calls, list) {
 912                 nentry = duplicate_call_entry(entry, (ir_node*)get_irn_link(entry->call), loop_depth);
 913                 list_add_tail(&nentry->list, &dst->calls);
 914         }
 915         dst->n_call_nodes += src->n_call_nodes;
 916         dst->n_nodes      += src->n_nodes;
 917 }
 918
 919 /*
 920  * Inlines small leaf methods at call sites where the called address comes
 921  * from a Const node that references the entity representing the called
 922  * method.
 923  * The size argument is a rough measure for the code size of the method:
 924  * Methods where the obstack containing the firm graph is smaller than
 925  * size are inlined.
 926  */
 927 void inline_leaf_functions(unsigned maxsize, unsigned leafsize,
 928                            unsigned size, int ignore_runtime)
 929 {
 930         inline_irg_env   *env;
 931         ir_graph         *irg;
 932         size_t           i, n_irgs;
 933         ir_graph         *rem;
 934         int              did_inline;
 935         wenv_t           wenv;
 936         pmap             *copied_graphs;
 937         pmap_entry       *pm_entry;
 938
 939         rem = current_ir_graph;
 940         obstack_init(&temp_obst);
 941
 942         /* a map for the copied graphs, used to inline recursive calls */
 943         copied_graphs = pmap_create();
 944
 945         /* extend all irgs by a temporary data structure for inlining. */
 946         n_irgs = get_irp_n_irgs();
 947         for (i = 0; i < n_irgs; ++i)
 948                 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
 949
 950         /* Pre-compute information in temporary data structure. */
 951         wenv.ignore_runtime = ignore_runtime;
 952         wenv.ignore_callers = 0;
 953         for (i = 0; i < n_irgs; ++i) {
 954                 ir_graph *irg = get_irp_irg(i);
 955
 956                 free_callee_info(irg);
 957
 958                 assure_irg_properties(irg,
 959                         IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
 960                 wenv.x = (inline_irg_env*)get_irg_link(irg);
 961                 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
 962                 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
 963         }
 964
 965         /* -- and now inline. -- */
 966
 967         /* Inline leafs recursively -- we might construct new leafs. */
 968         do {
 969                 did_inline = 0;
 970
 971                 for (i = 0; i < n_irgs; ++i) {
 972                         int phiproj_computed = 0;
 973
 974                         current_ir_graph = get_irp_irg(i);
 975                         env              = (inline_irg_env*)get_irg_link(current_ir_graph);
 976
 977                         ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
 978                         list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
 979                                 if (env->n_nodes > maxsize)
 980                                         break;
 981
 982                                 ir_node   *call   = entry->call;
 983                                 ir_graph  *callee = entry->callee;
 984                                 ir_entity *called = get_irg_entity(callee);
 985                                 mtp_additional_properties props
 986                                         = get_entity_additional_properties(called);
 987                                 if (props & mtp_property_noinline)
 988                                         continue;
 989
 990                                 if (is_leaf(callee) && (
 991                                     is_smaller(callee, leafsize) || (props & mtp_property_always_inline))) {
 992                                         if (!phiproj_computed) {
 993                                                 phiproj_computed = 1;
 994                                                 collect_phiprojs(current_ir_graph);
 995                                         }
 996                                         did_inline = inline_method(call, callee);
 997
 998                                         if (did_inline) {
 999                                                 inline_irg_env *callee_env = (inline_irg_env*)get_irg_link(callee);
1000
1001                                                 /* call was inlined, Phi/Projs for current graph must be recomputed */
1002                                                 phiproj_computed = 0;
1003
1004                                                 /* Do some statistics */
1005                                                 env->got_inline = 1;
1006                                                 --env->n_call_nodes;
1007                                                 env->n_nodes += callee_env->n_nodes;
1008                                                 --callee_env->n_callers;
1009
1010                                                 /* remove this call from the list */
1011                                                 list_del(&entry->list);
1012                                                 continue;
1013                                         }
1014                                 }
1015                         }
1016                         ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1017                 }
1018         } while (did_inline);
1019
1020         /* inline other small functions. */
1021         for (i = 0; i < n_irgs; ++i) {
1022                 int phiproj_computed = 0;
1023
1024                 current_ir_graph = get_irp_irg(i);
1025                 env              = (inline_irg_env*)get_irg_link(current_ir_graph);
1026
1027                 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1028
1029                 /* note that the list of possible calls is updated during the process */
1030                 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1031                         ir_node   *call   = entry->call;
1032                         ir_graph  *callee = entry->callee;
1033                         ir_entity *called = get_irg_entity(callee);
1034
1035                         mtp_additional_properties props = get_entity_additional_properties(called);
1036                         if (props & mtp_property_noinline)
1037                                 continue;
1038
1039                         ir_graph *calleee = pmap_get(ir_graph, copied_graphs, callee);
1040                         if (calleee != NULL) {
1041                                 /*
1042                                  * Remap callee if we have a copy.
1043                                  * FIXME: Should we do this only for recursive Calls ?
1044                                  */
1045                                 callee = calleee;
1046                         }
1047
1048                         if ((props & mtp_property_always_inline) ||
1049                             (is_smaller(callee, size) && env->n_nodes < maxsize) /* small function */) {
1050                                 if (current_ir_graph == callee) {
1051                                         /*
1052                                          * Recursive call: we cannot directly inline because we cannot walk
1053                                          * the graph and change it. So we have to make a copy of the graph
1054                                          * first.
1055                                          */
1056
1057                                         inline_irg_env *callee_env;
1058                                         ir_graph       *copy;
1059
1060                                         ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1061
1062                                         /*
1063                                          * No copy yet, create one.
1064                                          * Note that recursive methods are never leafs, so it is sufficient
1065                                          * to test this condition here.
1066                                          */
1067                                         copy = create_irg_copy(callee);
1068
1069                                         /* create_irg_copy() destroys the Proj links, recompute them */
1070                                         phiproj_computed = 0;
1071
1072                                         ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1073
1074                                         /* allocate new environment */
1075                                         callee_env = alloc_inline_irg_env();
1076                                         set_irg_link(copy, callee_env);
1077
1078                                         assure_irg_properties(copy,
1079                                                 IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
1080                                         wenv.x              = callee_env;
1081                                         wenv.ignore_callers = 1;
1082                                         irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1083
1084                                         /*
1085                                          * Enter the entity of the original graph. This is needed
1086                                          * for inline_method(). However, note that ent->irg still points
1087                                          * to callee, NOT to copy.
1088                                          */
1089                                         set_irg_entity(copy, get_irg_entity(callee));
1090
1091                                         pmap_insert(copied_graphs, callee, copy);
1092                                         callee = copy;
1093
1094                                         /* we have only one caller: the original graph */
1095                                         callee_env->n_callers      = 1;
1096                                         callee_env->n_callers_orig = 1;
1097                                 }
1098                                 if (! phiproj_computed) {
1099                                         phiproj_computed = 1;
1100                                         collect_phiprojs(current_ir_graph);
1101                                 }
1102                                 did_inline = inline_method(call, callee);
1103                                 if (did_inline) {
1104                                         inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1105
1106                                         /* call was inlined, Phi/Projs for current graph must be recomputed */
1107                                         phiproj_computed = 0;
1108
1109                                         /* callee was inline. Append its call list. */
1110                                         env->got_inline = 1;
1111                                         --env->n_call_nodes;
1112                                         append_call_list(env, callee_env, entry->loop_depth);
1113                                         --callee_env->n_callers;
1114
1115                                         /* after we have inlined callee, all called methods inside callee
1116                                            are now called once more */
1117                                         list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
1118                                                 inline_irg_env *penv = (inline_irg_env*)get_irg_link(centry->callee);
1119                                                 ++penv->n_callers;
1120                                         }
1121
1122                                         /* remove this call from the list */
1123                                         list_del(&entry->list);
1124                                         continue;
1125                                 }
1126                         }
1127                 }
1128                 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1129         }
1130
1131         for (i = 0; i < n_irgs; ++i) {
1132                 irg = get_irp_irg(i);
1133                 env = (inline_irg_env*)get_irg_link(irg);
1134
1135                 if (env->got_inline) {
1136                         optimize_graph_df(irg);
1137                         optimize_cf(irg);
1138                 }
1139                 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1140                         DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1141                         env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1142                         env->n_callers_orig, env->n_callers,
1143                         get_entity_name(get_irg_entity(irg))));
1144                 }
1145         }
1146
1147         /* kill the copied graphs: we don't need them anymore */
1148         foreach_pmap(copied_graphs, pm_entry) {
1149                 ir_graph *copy = (ir_graph*)pm_entry->value;
1150
1151                 /* reset the entity, otherwise it will be deleted in the next step ... */
1152                 set_irg_entity(copy, NULL);
1153                 free_ir_graph(copy);
1154         }
1155         pmap_destroy(copied_graphs);
1156
1157         obstack_free(&temp_obst, NULL);
1158         current_ir_graph = rem;
1159 }
1160
1161 typedef struct inline_leaf_functions_pass_t {
1162         ir_prog_pass_t pass;
1163         unsigned       maxsize;
1164         unsigned       leafsize;
1165         unsigned       size;
1166         int            ignore_runtime;
1167 } inline_leaf_functions_pass_t;
1168
1169 /**
1170  * Wrapper to run inline_leaf_functions() as a ir_prog pass.
1171  */
1172 static int inline_leaf_functions_wrapper(ir_prog *irp, void *context)
1173 {
1174         inline_leaf_functions_pass_t *pass = (inline_leaf_functions_pass_t*)context;
1175
1176         (void)irp;
1177         inline_leaf_functions(
1178                 pass->maxsize, pass->leafsize,
1179                 pass->size, pass->ignore_runtime);
1180         return 0;
1181 }
1182
1183 /* create a pass for inline_leaf_functions() */
1184 ir_prog_pass_t *inline_leaf_functions_pass(
1185         const char *name, unsigned maxsize, unsigned leafsize,
1186         unsigned size, int ignore_runtime)
1187 {
1188         inline_leaf_functions_pass_t *pass = XMALLOCZ(inline_leaf_functions_pass_t);
1189
1190         pass->maxsize        = maxsize;
1191         pass->leafsize       = leafsize;
1192         pass->size           = size;
1193         pass->ignore_runtime = ignore_runtime;
1194
1195         return def_prog_pass_constructor(
1196                 &pass->pass,
1197                 name ? name : "inline_leaf_functions",
1198                 inline_leaf_functions_wrapper);
1199 }
1200
1201 /**
1202  * Calculate the parameter weights for transmitting the address of a local variable.
1203  */
1204 static unsigned calc_method_local_weight(ir_node *arg)
1205 {
1206         int      j;
1207         unsigned v, weight = 0;
1208
1209         for (unsigned i = get_irn_n_outs(arg); i-- > 0; ) {
1210                 ir_node *succ = get_irn_out(arg, i);
1211
1212                 switch (get_irn_opcode(succ)) {
1213                 case iro_Load:
1214                 case iro_Store:
1215                         /* Loads and Store can be removed */
1216                         weight += 3;
1217                         break;
1218                 case iro_Sel:
1219                         /* check if all args are constant */
1220                         for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1221                                 ir_node *idx = get_Sel_index(succ, j);
1222                                 if (! is_Const(idx))
1223                                         return 0;
1224                         }
1225                         /* Check users on this Sel. Note: if a 0 is returned here, there was
1226                            some unsupported node. */
1227                         v = calc_method_local_weight(succ);
1228                         if (v == 0)
1229                                 return 0;
1230                         /* we can kill one Sel with constant indexes, this is cheap */
1231                         weight += v + 1;
1232                         break;
1233                 case iro_Id:
1234                         /* when looking backward we might find Id nodes */
1235                         weight += calc_method_local_weight(succ);
1236                         break;
1237                 case iro_Tuple:
1238                         /* unoptimized tuple */
1239                         for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1240                                 ir_node *pred = get_Tuple_pred(succ, j);
1241                                 if (pred == arg) {
1242                                         /* look for Proj(j) */
1243                                         for (unsigned k = get_irn_n_outs(succ); k-- > 0; ) {
1244                                                 ir_node *succ_succ = get_irn_out(succ, k);
1245                                                 if (is_Proj(succ_succ)) {
1246                                                         if (get_Proj_proj(succ_succ) == j) {
1247                                                                 /* found */
1248                                                                 weight += calc_method_local_weight(succ_succ);
1249                                                         }
1250                                                 } else {
1251                                                         /* this should NOT happen */
1252                                                         return 0;
1253                                                 }
1254                                         }
1255                                 }
1256                         }
1257                         break;
1258                 default:
1259                         /* any other node: unsupported yet or bad. */
1260                         return 0;
1261                 }
1262         }
1263         return weight;
1264 }
1265
1266 /**
1267  * Calculate the parameter weights for transmitting the address of a local variable.
1268  */
1269 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg)
1270 {
1271         ir_entity *ent = get_irg_entity(irg);
1272         ir_type  *mtp;
1273         size_t   nparams;
1274         long     proj_nr;
1275         ir_node  *irg_args, *arg;
1276
1277         mtp      = get_entity_type(ent);
1278         nparams  = get_method_n_params(mtp);
1279
1280         /* allocate a new array. currently used as 'analysed' flag */
1281         env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1282
1283         /* If the method haven't parameters we have nothing to do. */
1284         if (nparams <= 0)
1285                 return;
1286
1287         assure_irg_outs(irg);
1288         irg_args = get_irg_args(irg);
1289         for (unsigned i = get_irn_n_outs(irg_args); i-- > 0; ) {
1290                 arg     = get_irn_out(irg_args, i);
1291                 proj_nr = get_Proj_proj(arg);
1292                 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1293         }
1294 }
1295
1296 /**
1297  * Calculate the benefice for transmitting an local variable address.
1298  * After inlining, the local variable might be transformed into a
1299  * SSA variable by scalar_replacement().
1300  */
1301 static unsigned get_method_local_adress_weight(ir_graph *callee, size_t pos)
1302 {
1303         inline_irg_env *env = (inline_irg_env*)get_irg_link(callee);
1304
1305         if (env->local_weights == NULL)
1306                 analyze_irg_local_weights(env, callee);
1307
1308         if (pos < ARR_LEN(env->local_weights))
1309                 return env->local_weights[pos];
1310         return 0;
1311 }
1312
1313 /**
1314  * Calculate a benefice value for inlining the given call.
1315  *
1316  * @param call       the call node we have to inspect
1317  * @param callee     the called graph
1318  */
1319 static int calc_inline_benefice(call_entry *entry, ir_graph *callee)
1320 {
1321         ir_node   *call = entry->call;
1322         ir_entity *ent  = get_irg_entity(callee);
1323         ir_type   *callee_frame;
1324         size_t    i, n_members, n_params;
1325         ir_node   *frame_ptr;
1326         ir_type   *mtp;
1327         int       weight = 0;
1328         int       all_const;
1329         unsigned  cc, v;
1330
1331         inline_irg_env *callee_env;
1332
1333         mtp_additional_properties props = get_entity_additional_properties(ent);
1334         if (props & mtp_property_noinline) {
1335                 DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden\n",
1336                     call, callee));
1337                 return entry->benefice = INT_MIN;
1338         }
1339
1340         callee_frame = get_irg_frame_type(callee);
1341         n_members = get_class_n_members(callee_frame);
1342         for (i = 0; i < n_members; ++i) {
1343                 ir_entity *frame_ent = get_class_member(callee_frame, i);
1344                 if (is_parameter_entity(frame_ent)) {
1345                         // TODO inliner should handle parameter entities by inserting Store operations
1346                         DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden due to parameter entity\n", call, callee));
1347                         add_entity_additional_properties(ent, mtp_property_noinline);
1348                         return entry->benefice = INT_MIN;
1349                 }
1350         }
1351
1352         if (props & mtp_property_noreturn) {
1353                 DB((dbg, LEVEL_2, "In %+F Call to %+F: not inlining noreturn or weak\n",
1354                     call, callee));
1355                 return entry->benefice = INT_MIN;
1356         }
1357
1358         /* costs for every passed parameter */
1359         n_params = get_Call_n_params(call);
1360         mtp      = get_entity_type(ent);
1361         cc       = get_method_calling_convention(mtp);
1362         if (cc & cc_reg_param) {
1363                 /* register parameter, smaller costs for register parameters */
1364                 size_t max_regs = cc & ~cc_bits;
1365
1366                 if (max_regs < n_params)
1367                         weight += max_regs * 2 + (n_params - max_regs) * 5;
1368                 else
1369                         weight += n_params * 2;
1370         } else {
1371                 /* parameters are passed an stack */
1372                 weight += 5 * n_params;
1373         }
1374
1375         /* constant parameters improve the benefice */
1376         frame_ptr = get_irg_frame(current_ir_graph);
1377         all_const = 1;
1378         for (i = 0; i < n_params; ++i) {
1379                 ir_node *param = get_Call_param(call, i);
1380
1381                 if (is_Const(param)) {
1382                         weight += get_method_param_weight(ent, i);
1383                 } else {
1384                         all_const = 0;
1385                         if (is_SymConst(param))
1386                                 weight += get_method_param_weight(ent, i);
1387                         else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1388                                 /*
1389                                  * An address of a local variable is transmitted. After
1390                                  * inlining, scalar_replacement might be able to remove the
1391                                  * local variable, so honor this.
1392                                  */
1393                                 v = get_method_local_adress_weight(callee, i);
1394                                 weight += v;
1395                                 if (v > 0)
1396                                         entry->local_adr = 1;
1397                         }
1398                 }
1399         }
1400         entry->all_const = all_const;
1401
1402         callee_env = (inline_irg_env*)get_irg_link(callee);
1403         if (callee_env->n_callers == 1 &&
1404             callee != current_ir_graph &&
1405             !entity_is_externally_visible(ent)) {
1406                 weight += 700;
1407         }
1408
1409         /* give a bonus for functions with one block */
1410         if (callee_env->n_blocks == 1)
1411                 weight = weight * 3 / 2;
1412
1413         /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1414         if (callee_env->n_nodes < 30 && !callee_env->recursive)
1415                 weight += 2000;
1416
1417         /* and finally for leafs: they do not increase the register pressure
1418            because of callee safe registers */
1419         if (callee_env->n_call_nodes == 0)
1420                 weight += 400;
1421
1422         /** it's important to inline inner loops first */
1423         if (entry->loop_depth > 30)
1424                 weight += 30 * 1024;
1425         else
1426                 weight += entry->loop_depth * 1024;
1427
1428         /*
1429          * All arguments constant is probably a good sign, give an extra bonus
1430          */
1431         if (all_const)
1432                 weight += 1024;
1433
1434         return entry->benefice = weight;
1435 }
1436
1437 typedef struct walk_env_t {
1438         ir_graph **irgs;
1439         size_t   last_irg;
1440 } walk_env_t;
1441
1442 /**
1443  * Callgraph walker, collect all visited graphs.
1444  */
1445 static void callgraph_walker(ir_graph *irg, void *data)
1446 {
1447         walk_env_t *env = (walk_env_t *)data;
1448         env->irgs[env->last_irg++] = irg;
1449 }
1450
1451 /**
1452  * Creates an inline order for all graphs.
1453  *
1454  * @return the list of graphs.
1455  */
1456 static ir_graph **create_irg_list(void)
1457 {
1458         ir_entity  **free_methods;
1459         size_t     n_irgs = get_irp_n_irgs();
1460         walk_env_t env;
1461
1462         cgana(&free_methods);
1463         xfree(free_methods);
1464
1465         compute_callgraph();
1466
1467         env.irgs     = XMALLOCNZ(ir_graph*, n_irgs);
1468         env.last_irg = 0;
1469
1470         callgraph_walk(NULL, callgraph_walker, &env);
1471         assert(n_irgs == env.last_irg);
1472
1473         free_callgraph();
1474
1475         return env.irgs;
1476 }
1477
1478 /**
1479  * Push a call onto the priority list if its benefice is big enough.
1480  *
1481  * @param pqueue   the priority queue of calls
1482  * @param call     the call entry
1483  * @param inlien_threshold
1484  *                 the threshold value
1485  */
1486 static void maybe_push_call(pqueue_t *pqueue, call_entry *call,
1487                             int inline_threshold)
1488 {
1489         ir_graph *callee   = call->callee;
1490         int       benefice = calc_inline_benefice(call, callee);
1491
1492         DB((dbg, LEVEL_2, "In %+F Call %+F to %+F has benefice %d\n",
1493             get_irn_irg(call->call), call->call, callee, benefice));
1494
1495         ir_entity                *ent   = get_irg_entity(callee);
1496         mtp_additional_properties props = get_entity_additional_properties(ent);
1497         if (!(props & mtp_property_always_inline) && benefice < inline_threshold) {
1498                 return;
1499         }
1500
1501         pqueue_put(pqueue, call, benefice);
1502 }
1503
1504 /**
1505  * Try to inline calls into a graph.
1506  *
1507  * @param irg      the graph into which we inline
1508  * @param maxsize  do NOT inline if the size of irg gets
1509  *                 bigger than this amount
1510  * @param inline_threshold
1511  *                 threshold value for inline decision
1512  * @param copied_graphs
1513  *                 map containing copied of recursive graphs
1514  */
1515 static void inline_into(ir_graph *irg, unsigned maxsize,
1516                         int inline_threshold, pmap *copied_graphs)
1517 {
1518         int            phiproj_computed = 0;
1519         inline_irg_env *env = (inline_irg_env*)get_irg_link(irg);
1520         wenv_t         wenv;
1521         pqueue_t       *pqueue;
1522
1523         if (env->n_call_nodes == 0)
1524                 return;
1525
1526         if (env->n_nodes > maxsize) {
1527                 DB((dbg, LEVEL_2, "%+F: too big (%d)\n", irg, env->n_nodes));
1528                 return;
1529         }
1530
1531         current_ir_graph = irg;
1532         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1533
1534         /* put irgs into the pqueue */
1535         pqueue = new_pqueue();
1536
1537         list_for_each_entry(call_entry, curr_call, &env->calls, list) {
1538                 assert(is_Call(curr_call->call));
1539                 maybe_push_call(pqueue, curr_call, inline_threshold);
1540         }
1541
1542         /* note that the list of possible calls is updated during the process */
1543         while (!pqueue_empty(pqueue)) {
1544                 int                 did_inline;
1545                 call_entry          *curr_call  = (call_entry*)pqueue_pop_front(pqueue);
1546                 ir_graph            *callee     = curr_call->callee;
1547                 ir_node             *call_node  = curr_call->call;
1548                 inline_irg_env      *callee_env = (inline_irg_env*)get_irg_link(callee);
1549                 ir_entity           *ent        = get_irg_entity(callee);
1550                 mtp_additional_properties props
1551                         = get_entity_additional_properties(ent);
1552                 ir_graph            *calleee;
1553                 int                 loop_depth;
1554
1555                 if (!(props & mtp_property_always_inline)
1556                     && env->n_nodes + callee_env->n_nodes > maxsize) {
1557                         DB((dbg, LEVEL_2, "%+F: too big (%d) + %+F (%d)\n", irg,
1558                                                 env->n_nodes, callee, callee_env->n_nodes));
1559                         continue;
1560                 }
1561
1562                 calleee = pmap_get(ir_graph, copied_graphs, callee);
1563                 if (calleee != NULL) {
1564                         int benefice = curr_call->benefice;
1565                         /*
1566                          * Reduce the weight for recursive function IFF not all arguments are const.
1567                          * inlining recursive functions is rarely good.
1568                          */
1569                         if (!curr_call->all_const)
1570                                 benefice -= 2000;
1571                         if (benefice < inline_threshold)
1572                                 continue;
1573
1574                         /*
1575                          * Remap callee if we have a copy.
1576                          */
1577                         callee     = calleee;
1578                         callee_env = (inline_irg_env*)get_irg_link(callee);
1579                 }
1580
1581                 if (current_ir_graph == callee) {
1582                         /*
1583                          * Recursive call: we cannot directly inline because we cannot
1584                          * walk the graph and change it. So we have to make a copy of
1585                          * the graph first.
1586                          */
1587                         int benefice = curr_call->benefice;
1588                         ir_graph *copy;
1589
1590                         /*
1591                          * Reduce the weight for recursive function IFF not all arguments are const.
1592                          * inlining recursive functions is rarely good.
1593                          */
1594                         if (!curr_call->all_const)
1595                                 benefice -= 2000;
1596                         if (benefice < inline_threshold)
1597                                 continue;
1598
1599                         ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1600
1601                         /*
1602                          * No copy yet, create one.
1603                          * Note that recursive methods are never leafs, so it is
1604                          * sufficient to test this condition here.
1605                          */
1606                         copy = create_irg_copy(callee);
1607
1608                         /* create_irg_copy() destroys the Proj links, recompute them */
1609                         phiproj_computed = 0;
1610
1611                         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1612
1613                         /* allocate a new environment */
1614                         callee_env = alloc_inline_irg_env();
1615                         set_irg_link(copy, callee_env);
1616
1617                         assure_irg_properties(copy, IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
1618                         memset(&wenv, 0, sizeof(wenv));
1619                         wenv.x              = callee_env;
1620                         wenv.ignore_callers = 1;
1621                         irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1622
1623                         /*
1624                          * Enter the entity of the original graph. This is needed
1625                          * for inline_method(). However, note that ent->irg still points
1626                          * to callee, NOT to copy.
1627                          */
1628                         set_irg_entity(copy, get_irg_entity(callee));
1629
1630                         pmap_insert(copied_graphs, callee, copy);
1631                         callee = copy;
1632
1633                         /* we have only one caller: the original graph */
1634                         callee_env->n_callers      = 1;
1635                         callee_env->n_callers_orig = 1;
1636                 }
1637                 if (! phiproj_computed) {
1638                         phiproj_computed = 1;
1639                         collect_phiprojs(current_ir_graph);
1640                 }
1641                 did_inline = inline_method(call_node, callee);
1642                 if (!did_inline)
1643                         continue;
1644
1645                 /* call was inlined, Phi/Projs for current graph must be recomputed */
1646                 phiproj_computed = 0;
1647
1648                 /* remove it from the caller list */
1649                 list_del(&curr_call->list);
1650
1651                 /* callee was inline. Append its call list. */
1652                 env->got_inline = 1;
1653                 --env->n_call_nodes;
1654
1655                 /* we just generate a bunch of new calls */
1656                 loop_depth = curr_call->loop_depth;
1657                 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
1658                         inline_irg_env *penv = (inline_irg_env*)get_irg_link(centry->callee);
1659                         ir_node        *new_call;
1660                         call_entry     *new_entry;
1661
1662                         /* after we have inlined callee, all called methods inside
1663                          * callee are now called once more */
1664                         ++penv->n_callers;
1665
1666                         /* Note that the src list points to Call nodes in the inlined graph,
1667                          * but we need Call nodes in our graph. Luckily the inliner leaves
1668                          * this information in the link field. */
1669                         new_call = (ir_node*)get_irn_link(centry->call);
1670                         if (get_irn_irg(new_call) != irg) {
1671                                 /* centry->call has not been copied, which means it is dead.
1672                                  * This might happen during inlining, if a const function,
1673                                  * which cannot be inlined is only used as an unused argument
1674                                  * of another function, which is inlined. */
1675                                 continue;
1676                         }
1677                         assert(is_Call(new_call));
1678
1679                         new_entry = duplicate_call_entry(centry, new_call, loop_depth);
1680                         list_add_tail(&new_entry->list, &env->calls);
1681                         maybe_push_call(pqueue, new_entry, inline_threshold);
1682                 }
1683
1684                 env->n_call_nodes += callee_env->n_call_nodes;
1685                 env->n_nodes += callee_env->n_nodes;
1686                 --callee_env->n_callers;
1687         }
1688         ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1689         del_pqueue(pqueue);
1690 }
1691
1692 /*
1693  * Heuristic inliner. Calculates a benefice value for every call and inlines
1694  * those calls with a value higher than the threshold.
1695  */
1696 void inline_functions(unsigned maxsize, int inline_threshold,
1697                       opt_ptr after_inline_opt)
1698 {
1699         inline_irg_env   *env;
1700         size_t           i, n_irgs;
1701         ir_graph         *rem;
1702         wenv_t           wenv;
1703         pmap             *copied_graphs;
1704         pmap_entry       *pm_entry;
1705         ir_graph         **irgs;
1706
1707         rem = current_ir_graph;
1708         obstack_init(&temp_obst);
1709
1710         irgs = create_irg_list();
1711
1712         /* a map for the copied graphs, used to inline recursive calls */
1713         copied_graphs = pmap_create();
1714
1715         /* extend all irgs by a temporary data structure for inlining. */
1716         n_irgs = get_irp_n_irgs();
1717         for (i = 0; i < n_irgs; ++i)
1718                 set_irg_link(irgs[i], alloc_inline_irg_env());
1719
1720         /* Pre-compute information in temporary data structure. */
1721         wenv.ignore_runtime = 0;
1722         wenv.ignore_callers = 0;
1723         for (i = 0; i < n_irgs; ++i) {
1724                 ir_graph *irg = irgs[i];
1725
1726                 free_callee_info(irg);
1727
1728                 wenv.x = (inline_irg_env*)get_irg_link(irg);
1729                 assure_loopinfo(irg);
1730                 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1731         }
1732
1733         /* -- and now inline. -- */
1734         for (i = 0; i < n_irgs; ++i) {
1735                 ir_graph *irg = irgs[i];
1736
1737                 inline_into(irg, maxsize, inline_threshold, copied_graphs);
1738         }
1739
1740         for (i = 0; i < n_irgs; ++i) {
1741                 ir_graph *irg = irgs[i];
1742
1743                 env = (inline_irg_env*)get_irg_link(irg);
1744                 if (env->got_inline && after_inline_opt != NULL) {
1745                         /* this irg got calls inlined: optimize it */
1746                         after_inline_opt(irg);
1747                 }
1748                 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1749                         DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1750                         env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1751                         env->n_callers_orig, env->n_callers,
1752                         get_entity_name(get_irg_entity(irg))));
1753                 }
1754         }
1755
1756         /* kill the copied graphs: we don't need them anymore */
1757         foreach_pmap(copied_graphs, pm_entry) {
1758                 ir_graph *copy = (ir_graph*)pm_entry->value;
1759
1760                 /* reset the entity, otherwise it will be deleted in the next step ... */
1761                 set_irg_entity(copy, NULL);
1762                 free_ir_graph(copy);
1763         }
1764         pmap_destroy(copied_graphs);
1765
1766         xfree(irgs);
1767
1768         obstack_free(&temp_obst, NULL);
1769         current_ir_graph = rem;
1770 }
1771
1772 typedef struct inline_functions_pass_t {
1773         ir_prog_pass_t pass;
1774         unsigned       maxsize;
1775         int            inline_threshold;
1776         opt_ptr        after_inline_opt;
1777 } inline_functions_pass_t;
1778
1779 /**
1780  * Wrapper to run inline_functions() as a ir_prog pass.
1781  */
1782 static int inline_functions_wrapper(ir_prog *irp, void *context)
1783 {
1784         inline_functions_pass_t *pass = (inline_functions_pass_t*)context;
1785
1786         (void)irp;
1787         inline_functions(pass->maxsize, pass->inline_threshold,
1788                          pass->after_inline_opt);
1789         return 0;
1790 }
1791
1792 /* create a ir_prog pass for inline_functions */
1793 ir_prog_pass_t *inline_functions_pass(
1794           const char *name, unsigned maxsize, int inline_threshold,
1795           opt_ptr after_inline_opt)
1796 {
1797         inline_functions_pass_t *pass = XMALLOCZ(inline_functions_pass_t);
1798
1799         pass->maxsize          = maxsize;
1800         pass->inline_threshold = inline_threshold;
1801         pass->after_inline_opt = after_inline_opt;
1802
1803         return def_prog_pass_constructor(
1804                 &pass->pass, name ? name : "inline_functions",
1805                 inline_functions_wrapper);
1806 }
1807
1808 void firm_init_inline(void)
1809 {
1810         FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1811 }