Fix warnings.
[libfirm] / ir / opt / funccall.c
index 82b3828..e379ddf 100644 (file)
 /*
- * Project:     libFIRM
- * File name:   ir/opt/ldstopt.c
- * Purpose:     optimization of function calls
- * Author:
- * Created:
- * CVS-ID:      $Id$
- * Copyright:   (c) 1998-2004 Universität Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+ * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
  */
+
+/**
+ * @file
+ * @brief   Optimization of function calls.
+ * @author  Michael Beck
+ */
+#include "config.h"
+
+#include "opt_init.h"
+#include <stdbool.h>
+
 #include "irnode_t.h"
 #include "irgraph_t.h"
 #include "irgmod.h"
 #include "irgwalk.h"
-#include "irvrfy.h"
 #include "dbginfo_t.h"
 #include "irflag_t.h"
+#include "irloop_t.h"
 #include "ircons.h"
-#include "eset.h"
-#include "funccall.h"
+#include "iredges_t.h"
+#include "irpass_t.h"
+#include "iroptimize.h"
+#include "analyze_irg_args.h"
 #include "irhooks.h"
+#include "raw_bitset.h"
+#include "debug.h"
+
+DEBUG_ONLY(static firm_dbg_module_t *dbg;)
 
 /**
- * The walker environment for rem_mem_from_real_fkt_calls
+ * The walker environment for updating function calls.
  */
-typedef struct _env_t {
-  eset *pure_fkt;   /**< set containing all real functions */
-  int  changed;     /**< flag, is set if a graph was changed */
+typedef struct env_t {
+       ir_node  *float_const_call_list;    /**< The list of all floating const function calls that will be changed. */
+       ir_node  *nonfloat_const_call_list; /**< The list of all non-floating const function calls that will be changed. */
+       ir_node  *pure_call_list;           /**< The list of all pure function calls that will be changed. */
+       ir_node  *nothrow_call_list;        /**< The list of all nothrow function calls that will be changed. */
+       ir_node  *proj_list;                /**< The list of all potential Proj nodes that must be fixed. */
 } env_t;
 
+/** Ready IRG's are marked in the ready set. */
+static unsigned *ready_set;
+
+/** IRG's that are in progress are marked here. */
+static unsigned *busy_set;
+
+/**
+ * Walker: Collect all calls to const and pure functions
+ * to lists. Collect all Proj(Call) nodes into a Proj list.
+ */
+static void collect_const_and_pure_calls(ir_node *node, void *env)
+{
+       env_t *ctx = (env_t*)env;
+
+       if (is_Call(node)) {
+               ir_node *call = node;
+
+               /* set the link to NULL for all non-const/pure calls */
+               set_irn_link(call, NULL);
+               ir_node *ptr = get_Call_ptr(call);
+               if (!is_SymConst_addr_ent(ptr))
+                       return;
+
+               ir_entity *ent = get_SymConst_entity(ptr);
+
+               unsigned prop = get_entity_additional_properties(ent);
+               if ((prop & (mtp_property_const|mtp_property_pure)) == 0)
+                       return;
+
+               /* ok, if we get here we found a call to a const or a pure function */
+               if (prop & mtp_property_pure) {
+                       set_irn_link(call, ctx->pure_call_list);
+                       ctx->pure_call_list = call;
+               } else {
+                       if (prop & mtp_property_has_loop) {
+                               set_irn_link(call, ctx->nonfloat_const_call_list);
+                               ctx->nonfloat_const_call_list = call;
+                       } else {
+                               set_irn_link(call, ctx->float_const_call_list);
+                               ctx->float_const_call_list = call;
+                       }
+               }
+       } else if (is_Proj(node)) {
+               /*
+                * Collect all memory and exception Proj's from
+                * calls.
+                */
+               ir_node *call = get_Proj_pred(node);
+               if (!is_Call(call))
+                       return;
+
+               /* collect the Proj's in the Proj list */
+               switch (get_Proj_proj(node)) {
+               case pn_Call_M:
+               case pn_Call_X_except:
+               case pn_Call_X_regular:
+                       set_irn_link(node, ctx->proj_list);
+                       ctx->proj_list = node;
+                       break;
+               default:
+                       break;
+               }
+       }
+}
+
 /**
- * remove memory from real function calls by rerouting
- * it's ProjM and connection the call with a NoMem node.
+ * Fix the list of collected Calls.
  *
- * Note: By "real function" we understand a function that did neither
- * read nor write memory.
+ * @param irg  the graph that contained calls to pure functions
+ * @param ctx  context
  */
-static void rem_mem_from_real_fkt_calls(ir_node *node, void *env)
+static void fix_const_call_lists(ir_graph *irg, env_t *ctx)
 {
-  env_t *ctx = env;
-  ir_node *call, *ptr, *mem;
-  entity *ent;
+       bool exc_changed = false;
+
+       /* First step: fix all calls by removing their memory input and let
+        * them floating.
+        * The original memory input is preserved in their link fields. */
+       ir_node *next;
+       for (ir_node *call = ctx->float_const_call_list; call != NULL; call = next) {
+               next = (ir_node*)get_irn_link(call);
+               ir_node *mem = get_Call_mem(call);
+
+               set_irn_link(call, mem);
+               set_Call_mem(call, get_irg_no_mem(irg));
+
+               /*
+                * Unfortunately we cannot simply set the node to 'float'.
+                * There is a reason for that:
+                *
+                * - The call might be inside a loop/if that is NOT entered
+                *   and calls a endless function. Setting the call to float
+                *   would allow to move it out from the loop/if causing this
+                *   function be called even if the loop/if is not entered ...
+                *
+                * This could be fixed using post-dominators for calls and Pin nodes
+                * but need some more analyzes to ensure that a call that potential
+                * never returns is not executed before some code that generates
+                * observable states...
+                */
+
+               /* finally, this call can float */
+               set_irn_pinned(call, op_pin_state_floats);
+               hook_func_call(irg, call);
+       }
 
-  if (get_irn_op(node) != op_Proj || get_irn_mode(node) != mode_M)
-    return;
+       /* Last step: fix all Proj's */
+       for (ir_node *proj = ctx->proj_list; proj != NULL; proj = next) {
+               next = (ir_node*)get_irn_link(proj);
+               ir_node *call = get_Proj_pred(proj);
+               ir_node *mem  = (ir_node*)get_irn_link(call);
 
-  /* look at Memory Proj's */
-  call = get_Proj_pred(node);
-  if (get_irn_op(call) != op_Call)
-    return;
+               /* beware of calls in the pure call list */
+               if (!mem || is_Call(mem))
+                       continue;
+               assert(get_irn_mode(mem) == mode_M);
 
-  ptr = get_Call_ptr(call);
-  if (get_irn_op(ptr) != op_SymConst || get_SymConst_kind(ptr) != symconst_addr_ent)
-    return;
+               switch (get_Proj_proj(proj)) {
+               case pn_Call_M: {
+                       /* in dead code there might be cycles where proj == mem */
+                       if (proj != mem)
+                               exchange(proj, mem);
+                        break;
+               }
+               case pn_Call_X_except:
+                       exc_changed = true;
+                       exchange(proj, new_r_Bad(irg, mode_X));
+                       break;
+               case pn_Call_X_regular: {
+                       ir_node *block = get_nodes_block(call);
+                       exc_changed = true;
+                       exchange(proj, new_r_Jmp(block));
+                       break;
+               }
+               default:
+                       break;
+               }
+       }
 
-  ent = get_SymConst_entity(ptr);
+       if (exc_changed) {
+               /* ... including exception edges */
+               clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
+                                  | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
+       }
+}
+
+/**
+ * Walker: Collect all calls to nothrow functions
+ * to lists. Collect all Proj(Call) nodes into a Proj list.
+ */
+static void collect_nothrow_calls(ir_node *node, void *env)
+{
+       env_t *ctx = (env_t*)env;
 
-  if (! eset_contains(ctx->pure_fkt, ent))
-    return;
+       if (is_Call(node)) {
+               ir_node *call = node;
 
-  /* ok, if we get here we found a call to a pure function,
-   * route the NoMem node to the call */
-  mem   = get_Call_mem(call);
+               /* set the link to NULL for all non-const/pure calls */
+               set_irn_link(call, NULL);
+               ir_node *ptr = get_Call_ptr(call);
+               if (!is_SymConst_addr_ent(ptr))
+                       return;
 
-  exchange(node, mem);
-  set_Call_mem(call, new_r_NoMem(current_ir_graph));
+               ir_entity *ent = get_SymConst_entity(ptr);
 
-  /* finally, this call can floats */
-  set_irn_pinned(call, op_pin_state_floats);
+               unsigned prop = get_entity_additional_properties(ent);
+               if ((prop & mtp_property_nothrow) == 0)
+                       return;
 
-  hook_func_call(current_ir_graph, call);
+               /* ok, if we get here we found a call to a nothrow function */
+               set_irn_link(call, ctx->nothrow_call_list);
+               ctx->nothrow_call_list = call;
+       } else if (is_Proj(node)) {
+               /*
+                * Collect all memory and exception Proj's from
+                * calls.
+                */
+               ir_node *call = get_Proj_pred(node);
+               if (! is_Call(call))
+                       return;
 
-  ctx->changed = 1;
+               /* collect the Proj's in the Proj list */
+               switch (get_Proj_proj(node)) {
+               case pn_Call_M:
+               case pn_Call_X_except:
+               case pn_Call_X_regular:
+                       set_irn_link(node, ctx->proj_list);
+                       ctx->proj_list = node;
+                       break;
+               default:
+                       break;
+               }
+       }
 }
 
-/*
- * optimize function calls by handling real functions
+/**
+ * Fix the list of collected nothrow Calls.
+ *
+ * @param irg        the graph that contained calls to pure functions
+ * @param call_list  the list of all call sites of const functions
+ * @param proj_list  the list of all memory/exception Proj's of this call sites
  */
-void optimize_funccalls(void)
+static void fix_nothrow_call_list(ir_graph *irg, ir_node *call_list,
+                                  ir_node *proj_list)
 {
-  int i, n, j, k;
-  int change;
-  unsigned num_pure = 0;
-  eset *pure_fkt = eset_create();
+       bool exc_changed = false;
 
-  if (! get_opt_real_func_call())
-    return;
+       /* First step: go through the list of calls and mark them. */
+       ir_node *next;
+       for (ir_node *call = call_list; call; call = next) {
+               next = (ir_node*)get_irn_link(call);
 
-  /* first step: detect, which functions are pure, ie do NOT change any memory */
-  for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
-    ir_graph *irg  = get_irp_irg(i);
-    ir_node *end   = get_irg_end(irg);
-    ir_node *endbl = get_nodes_block(end);
+               /* current_ir_graph is in memory anyway, so it's a good marker */
+               set_irn_link(call, &current_ir_graph);
+               hook_func_call(irg, call);
+       }
 
-    change = 0;
+       /* Second step: Remove all exception Proj's */
+       for (ir_node *proj = proj_list; proj; proj = next) {
+               next = (ir_node*)get_irn_link(proj);
+               ir_node *call = get_Proj_pred(proj);
 
-    /* visit every Return */
-    for (j = 0, k = get_Block_n_cfgpreds(endbl); j < k; ++j) {
-      ir_node *ret = get_Block_cfgpred(endbl, j);
-      ir_node *mem;
+               /* handle only marked calls */
+               if (get_irn_link(call) != &current_ir_graph)
+                       continue;
 
-      /* Bad nodes usually do NOT produce anything, so it's ok */
-      if (is_Bad(ret))
-        continue;
+               /* kill any exception flow */
+               switch (get_Proj_proj(proj)) {
+               case pn_Call_X_except:
+                       exc_changed = true;
+                       exchange(proj, new_r_Bad(irg, mode_X));
+                       break;
+               case pn_Call_X_regular: {
+                       ir_node *block = get_nodes_block(call);
+                       exc_changed = true;
+                       exchange(proj, new_r_Jmp(block));
+                       break;
+               }
+               default:
+                       break;
+               }
+       }
 
-      mem = get_Return_mem(ret);
+       /* changes were done ... */
+       if (exc_changed) {
+               /* ... including exception edges */
+               clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
+                                  | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
+       }
+}
 
-      /* Bad nodes usually do NOT produce anything, so it's ok */
-      if (is_Bad(mem))
-        continue;
+/* marking */
+#define SET_IRG_READY(irg)  rbitset_set(ready_set, get_irg_idx(irg))
+#define IS_IRG_READY(irg)   rbitset_is_set(ready_set, get_irg_idx(irg))
+#define SET_IRG_BUSY(irg)   rbitset_set(busy_set, get_irg_idx(irg))
+#define CLEAR_IRG_BUSY(irg) rbitset_clear(busy_set, get_irg_idx(irg))
+#define IS_IRG_BUSY(irg)    rbitset_is_set(busy_set, get_irg_idx(irg))
 
-      change = mem != get_irg_initial_mem(irg);
-      if (change)
-        break;
-    }
+/* forward */
+static mtp_additional_properties check_const_or_pure_function(ir_graph *irg, bool top);
 
-    if (! change) {
-      /* check, if a keep-alive exists */
-      for (j = 0, k = get_End_n_keepalives(end); j < k; ++j) {
-        ir_node *mem = get_End_keepalive(end, j);
+/**
+ * Calculate the bigger property of two. Handle the temporary flag right.
+ */
+static mtp_additional_properties max_property(mtp_additional_properties a,
+                                              mtp_additional_properties b)
+{
+       mtp_additional_properties t = (a | b) & mtp_temporary;
+       a &= ~mtp_temporary;
+       b &= ~mtp_temporary;
 
-        if (mode_M != get_irn_mode(mem))
-          continue;
+       if (a == mtp_no_property || b == mtp_no_property)
+               return mtp_no_property;
+       mtp_additional_properties r = a > b ? a : b;
+       return r | t;
+}
 
-        change = mem != get_irg_initial_mem(irg);
-        if (change)
-          break;
-      }
-    }
+/**
+ * Follow the memory chain starting at node and determine
+ * the mtp_property.
+ *
+ * @return mtp_property_const if only calls of const functions are detected
+ *         mtp_property_pure  if only Loads and const/pure calls detected
+ *         mtp_no_property    else
+ */
+static mtp_additional_properties follow_mem_(ir_node *node)
+{
+       mtp_additional_properties mode = mtp_property_const;
+
+       for (;;) {
+               if (mode == mtp_no_property)
+                       return mtp_no_property;
+
+               if (irn_visited_else_mark(node))
+                       return mode;
+
+               switch (get_irn_opcode(node)) {
+               case iro_Proj:
+                       node = get_Proj_pred(node);
+                       break;
 
-    if (! change) {
-      eset_insert(pure_fkt, get_irg_entity(irg));
-      ++num_pure;
-    }
-  }
+               case iro_NoMem:
+                       /* finish here */
+                       return mode;
 
-  if (num_pure > 0) {
-    env_t ctx;
+               case iro_Phi:
+               case iro_Sync:
+                       /* do a dfs search */
+                       for (int i = get_irn_arity(node) - 1; i >= 0; --i) {
+                               mtp_additional_properties m = follow_mem_(get_irn_n(node, i));
+                               mode = max_property(mode, m);
+                               if (mode == mtp_no_property)
+                                       return mtp_no_property;
+                       }
+                       return mode;
 
-    ctx.pure_fkt = pure_fkt;
+               case iro_Load:
+                       /* Beware volatile Loads are NOT allowed in pure functions. */
+                       if (get_Load_volatility(node) == volatility_is_volatile)
+                               return mtp_no_property;
+                       mode = max_property(mode, mtp_property_pure);
+                       node = get_Load_mem(node);
+                       break;
+
+               case iro_Call: {
+                       /* A call is only tolerable if its either constant or pure. */
+                       ir_node *ptr = get_Call_ptr(node);
+                       if (!is_SymConst_addr_ent(ptr))
+                               return mtp_no_property;
+
+                       ir_entity *ent = get_SymConst_entity(ptr);
+                       ir_graph  *irg = get_entity_irg(ent);
+
+                       mtp_additional_properties m;
+                       if (irg == NULL) {
+                               m = get_entity_additional_properties(ent) & (mtp_property_const|mtp_property_pure);
+                               mode = max_property(mode, m);
+                       } else {
+                               /* we have a graph, analyze it. */
+                               m = check_const_or_pure_function(irg, false);
+                               mode = max_property(mode, m);
+                       }
+                       node = get_Call_mem(node);
+                       break;
+               }
+
+               default:
+                       return mtp_no_property;
+               }
+       }
+}
 
-    /* all calls of pure functions can be transformed into FuncCalls */
-    for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
-      ir_graph *irg  = get_irp_irg(i);
+/**
+ * Follow the memory chain starting at node and determine
+ * the mtp_property.
+ *
+ * @return mtp_property_const if only calls of const functions are detected
+ *         mtp_property_pure  if only Loads and const/pure calls detected
+ *         mtp_no_property else
+ */
+static mtp_additional_properties follow_mem(ir_node *node, mtp_additional_properties mode)
+{
+       mtp_additional_properties m = follow_mem_(node);
+       return max_property(mode, m);
+}
+
+/**
+ * Check if a graph represents a const or a pure function.
+ *
+ * @param irg  the graph to check
+ * @param top  if set, this is the top call
+ */
+static mtp_additional_properties check_const_or_pure_function(ir_graph *irg, bool top)
+{
+       ir_entity *entity   = get_irg_entity(irg);
+       ir_type   *type     = get_entity_type(entity);
+       size_t     n_params = get_method_n_params(type);
+       mtp_additional_properties may_be_const = mtp_property_const;
+       mtp_additional_properties prop = get_entity_additional_properties(entity);
 
-      ctx.changed = 0;
-      irg_walk_graph(irg, NULL, rem_mem_from_real_fkt_calls, &ctx);
+       /* libfirm handles aggregate parameters by passing around pointers to
+        * stuff in memory, so if we have compound parameters we are never const */
+       for (size_t i = 0; i < n_params; ++i) {
+               ir_type *param = get_method_param_type(type, i);
+               if (is_compound_type(param)) {
+                       prop        &= ~mtp_property_const;
+                       may_be_const = mtp_no_property;
+               }
+       }
+
+       if (prop & mtp_property_const) {
+               /* already marked as a const function */
+               return mtp_property_const;
+       }
+       if (prop & mtp_property_pure) {
+               /* already marked as a pure function */
+               return mtp_property_pure;
+       }
+
+       if (IS_IRG_READY(irg)) {
+               /* already checked */
+               return mtp_no_property;
+       }
+       if (IS_IRG_BUSY(irg)) {
+               /* We are still evaluate this method.
+                * The function (indirectly) calls itself and thus may not terminate. */
+               return mtp_no_property;
+       }
+       SET_IRG_BUSY(irg);
+
+       ir_node *end   = get_irg_end(irg);
+       ir_node *endbl = get_nodes_block(end);
+       prop = may_be_const;
+
+       ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
+       inc_irg_visited(irg);
+       /* mark the initial mem: recursion of follow_mem() stops here */
+       mark_irn_visited(get_irg_initial_mem(irg));
+
+       /* visit every Return */
+       for (int j = get_Block_n_cfgpreds(endbl) - 1; j >= 0; --j) {
+               ir_node  *node = get_Block_cfgpred(endbl, j);
+               unsigned  code = get_irn_opcode(node);
+
+               /* Bad nodes usually do NOT produce anything, so it's ok */
+               if (code == iro_Bad)
+                       continue;
+
+               if (code == iro_Return) {
+                       ir_node *mem = get_Return_mem(node);
+
+                       /* Bad nodes usually do NOT produce anything, so it's ok */
+                       if (is_Bad(mem))
+                               continue;
+
+                       if (mem != get_irg_initial_mem(irg))
+                               prop = max_property(prop, follow_mem(mem, prop));
+               } else {
+                       /* Exception found. Cannot be const or pure. */
+                       prop = mtp_no_property;
+                       break;
+               }
+               if (prop == mtp_no_property)
+                       break;
+       }
+
+       if (prop != mtp_no_property) {
+               /* check, if a keep-alive exists */
+               for (int j = get_End_n_keepalives(end) - 1; j >= 0; --j) {
+                       ir_node *kept = get_End_keepalive(end, j);
+
+                       if (is_Block(kept)) {
+                               prop = mtp_no_property;
+                               break;
+                       }
+
+                       if (mode_M != get_irn_mode(kept))
+                               continue;
+
+                       prop = max_property(prop, follow_mem(kept, prop));
+                       if (prop == mtp_no_property)
+                               break;
+               }
+       }
+
+       if (top) {
+               /* Set the property only if we are at top-level. */
+               if (prop != mtp_no_property) {
+                       add_entity_additional_properties(entity, prop);
+               }
+               SET_IRG_READY(irg);
+       }
+       CLEAR_IRG_BUSY(irg);
+       ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
+       return prop;
+}
+
+/**
+ * Handle calls to const functions.
+ *
+ * @param ctx  context
+ */
+static void handle_const_Calls(env_t *ctx)
+{
+       /* all calls of const functions can be transformed */
+       size_t n = get_irp_n_irgs();
+       for (size_t i = 0; i < n; ++i) {
+               ir_graph *irg = get_irp_irg(i);
 
-      if (ctx.changed) {
-        /* changes were done */
-        set_irg_outs_inconsistent(irg);
-        set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
-      }
-    }
-  }
+               ctx->float_const_call_list    = NULL;
+               ctx->nonfloat_const_call_list = NULL;
+               ctx->pure_call_list           = NULL;
+               ctx->proj_list                = NULL;
 
-  eset_destroy(pure_fkt);
+               ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
+               irg_walk_graph(irg, NULL, collect_const_and_pure_calls, ctx);
+               fix_const_call_lists(irg, ctx);
+               ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+
+               confirm_irg_properties(irg,
+                       IR_GRAPH_PROPERTIES_CONTROL_FLOW
+                       | IR_GRAPH_PROPERTY_ONE_RETURN
+                       | IR_GRAPH_PROPERTY_MANY_RETURNS);
+       }
+}
+
+/**
+ * Handle calls to nothrow functions.
+ *
+ * @param ctx  context
+ */
+static void handle_nothrow_Calls(env_t *ctx)
+{
+       /* all calls of const functions can be transformed */
+       size_t n = get_irp_n_irgs();
+       for (size_t i = 0; i < n; ++i) {
+               ir_graph *irg  = get_irp_irg(i);
+
+               ctx->nothrow_call_list = NULL;
+               ctx->proj_list         = NULL;
+
+               ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
+               irg_walk_graph(irg, NULL, collect_nothrow_calls, ctx);
+
+               if (ctx->nothrow_call_list)
+                       fix_nothrow_call_list(irg, ctx->nothrow_call_list, ctx->proj_list);
+               ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+       }
+}
+
+/**
+ * Check, whether a given node represents a return value of
+ * a malloc like function (ie, new heap allocated memory).
+ *
+ * @param node  the node to check
+ */
+static bool is_malloc_call_result(const ir_node *node)
+{
+       if (is_Alloc(node) && get_Alloc_where(node) == heap_alloc) {
+               /* Firm style high-level allocation */
+               return true;
+       }
+       /* TODO: check mtp_malloc */
+       return false;
+}
+
+/**
+ * Update a property depending on a call property.
+ */
+static mtp_additional_properties update_property(mtp_additional_properties orig_prop, mtp_additional_properties call_prop)
+{
+       mtp_additional_properties t = (orig_prop | call_prop) & mtp_temporary;
+       mtp_additional_properties r = orig_prop & call_prop;
+       return r | t;
+}
+
+/**
+ * Check if a node is stored.
+ */
+static bool is_stored(const ir_node *n)
+{
+       const ir_node *ptr;
+
+       foreach_out_edge(n, edge) {
+               const ir_node *succ = get_edge_src_irn(edge);
+
+               switch (get_irn_opcode(succ)) {
+               case iro_Return:
+               case iro_Load:
+               case iro_Cmp:
+                       /* ok */
+                       break;
+               case iro_Store:
+                       if (get_Store_value(succ) == n)
+                               return true;
+                       /* ok if its only the address input */
+                       break;
+               case iro_Sel:
+               case iro_Cast:
+               case iro_Confirm:
+                       if (is_stored(succ))
+                               return true;
+                       break;
+               case iro_Call:
+                       ptr = get_Call_ptr(succ);
+                       if (is_SymConst_addr_ent(ptr)) {
+                               ir_entity *ent = get_SymConst_entity(ptr);
+                               size_t    i;
+
+                               /* we know the called entity */
+                               for (i = get_Call_n_params(succ); i > 0;) {
+                                       if (get_Call_param(succ, --i) == n) {
+                                               /* n is the i'th param of the call */
+                                               if (get_method_param_access(ent, i) & ptr_access_store) {
+                                                       /* n is store in ent */
+                                                       return true;
+                                               }
+                                       }
+                               }
+                       } else {
+                               /* unknown call address */
+                               return true;
+                       }
+                       break;
+               default:
+                       /* bad, potential alias */
+                       return true;
+               }
+       }
+       return false;
+}
+
+/**
+ * Check that the return value of an irg is not stored anywhere.
+ *
+ * return ~mtp_property_malloc if return values are stored, ~0 else
+ */
+static mtp_additional_properties check_stored_result(ir_graph *irg)
+{
+       ir_node *end_blk = get_irg_end_block(irg);
+       mtp_additional_properties res = ~mtp_no_property;
+
+       assure_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES);
+
+       for (int i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
+               ir_node *pred = get_Block_cfgpred(end_blk, i);
+
+               if (! is_Return(pred))
+                       continue;
+               for (size_t j = get_Return_n_ress(pred); j > 0;) {
+                       const ir_node *irn = get_Return_res(pred, --j);
+
+                       if (is_stored(irn)) {
+                               /* bad, might create an alias */
+                               res = ~mtp_property_malloc;
+                               goto finish;
+                       }
+               }
+       }
+finish:
+       confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
+       return res;
+}
+
+/**
+ * Check if a graph represents a nothrow or a malloc function.
+ *
+ * @param irg  the graph to check
+ * @param top  if set, this is the top call
+ */
+static mtp_additional_properties check_nothrow_or_malloc(ir_graph *irg, bool top)
+{
+       mtp_additional_properties curr_prop
+               = mtp_property_malloc | mtp_property_nothrow;
+
+       ir_entity *ent = get_irg_entity(irg);
+       if (IS_IRG_READY(irg)) {
+               /* already checked */
+               return get_entity_additional_properties(ent);
+       }
+       if (IS_IRG_BUSY(irg)) {
+               /* we are still evaluate this method. Be optimistic,
+               return the best possible so far but mark the result as temporary. */
+               return mtp_temporary | mtp_property_malloc | mtp_property_nothrow;
+       }
+       SET_IRG_BUSY(irg);
+
+       ir_type *mtp = get_entity_type(ent);
+       if (get_method_n_ress(mtp) <= 0)
+               curr_prop &= ~mtp_property_malloc;
+
+       ir_node *end_blk = get_irg_end_block(irg);
+       for (int i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
+               ir_node *pred = get_Block_cfgpred(end_blk, i);
+
+               if (is_Return(pred)) {
+                       if (curr_prop & mtp_property_malloc) {
+                               /* check, if malloc is called here */
+                               for (size_t j = get_Return_n_ress(pred); j > 0;) {
+                                       ir_node *res = get_Return_res(pred, --j);
+
+                                       /* skip Confirms and Casts */
+                                       res = skip_HighLevel_ops(res);
+                                       /* skip Proj's */
+                                       while (is_Proj(res))
+                                               res = get_Proj_pred(res);
+                                       if (is_malloc_call_result(res)) {
+                                               /* ok, this is a malloc */
+                                       } else if (is_Call(res)) {
+                                               ir_node *ptr = get_Call_ptr(res);
+
+                                               if (is_SymConst_addr_ent(ptr)) {
+                                                       /* a direct call */
+                                                       ir_entity *ent    = get_SymConst_entity(ptr);
+                                                       ir_graph  *callee = get_entity_irg(ent);
+
+                                                       if (callee == irg) {
+                                                               /* A self-recursive call. The property did not depend on this call. */
+                                                       } else if (callee != NULL) {
+                                                               mtp_additional_properties prop = check_nothrow_or_malloc(callee, false);
+                                                               curr_prop = update_property(curr_prop, prop);
+                                                       } else {
+                                                               curr_prop = update_property(curr_prop, get_entity_additional_properties(ent));
+                                                       }
+                                               } else {
+                                                       /* unknown call */
+                                                       curr_prop &= ~mtp_property_malloc;
+                                               }
+                                       } else {
+                                               /* unknown return value */
+                                               curr_prop &= ~mtp_property_malloc;
+                                       }
+                               }
+                       }
+               } else if (curr_prop & mtp_property_nothrow) {
+                       /* exception flow detected */
+                       pred = skip_Proj(pred);
+
+                       if (is_Call(pred)) {
+                               ir_node *ptr = get_Call_ptr(pred);
+
+                               if (is_SymConst_addr_ent(ptr)) {
+                                       /* a direct call */
+                                       ir_entity *ent    = get_SymConst_entity(ptr);
+                                       ir_graph  *callee = get_entity_irg(ent);
+
+                                       if (callee == irg) {
+                                               /* A self-recursive call. The property did not depend on this call. */
+                                       } else if (callee != NULL) {
+                                               /* Note: we check here for nothrow only, so do NOT reset the malloc property */
+                                               mtp_additional_properties prop = check_nothrow_or_malloc(callee, false) | mtp_property_malloc;
+                                               curr_prop = update_property(curr_prop, prop);
+                                       } else {
+                                               if ((get_entity_additional_properties(ent) & mtp_property_nothrow) == 0)
+                                                       curr_prop &= ~mtp_property_nothrow;
+                                       }
+                               } else {
+                                       /* unknown call */
+                                       curr_prop &= ~mtp_property_nothrow;
+                               }
+                       } else {
+                               /* real exception flow possible. */
+                               curr_prop &= ~mtp_property_nothrow;
+                       }
+               }
+               if ((curr_prop & ~mtp_temporary) == mtp_no_property) {
+                       /* no need to search further */
+                       break;
+               }
+       }
+
+       if (curr_prop & mtp_property_malloc) {
+               /* Note that the malloc property means not only return newly allocated
+                * memory, but also that this memory is ALIAS FREE.
+                * To ensure that, we do NOT allow that the returned memory is somewhere
+                * stored. */
+               curr_prop &= check_stored_result(irg);
+       }
+
+       if (curr_prop != mtp_no_property
+           && (top || (curr_prop & mtp_temporary) == 0)) {
+               /* We use the temporary flag here to mark an optimistic result.
+                * Set the property only if we are sure that it does NOT base on
+                * temporary results OR if we are at top-level. */
+               add_entity_additional_properties(ent, curr_prop & ~mtp_temporary);
+               SET_IRG_READY(irg);
+       }
+       if (top)
+               SET_IRG_READY(irg);
+       CLEAR_IRG_BUSY(irg);
+       return curr_prop;
+}
+
+/**
+ * When a function was detected as "const", it might be moved out of loops.
+ * This might be dangerous if the graph can contain endless loops.
+ */
+static void check_for_possible_endless_loops(ir_graph *irg)
+{
+       assure_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
+
+       ir_loop *root_loop = get_irg_loop(irg);
+       if (root_loop->flags & loop_outer_loop) {
+               ir_entity *ent = get_irg_entity(irg);
+               add_entity_additional_properties(ent, mtp_property_has_loop);
+       }
+
+       confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
+}
+
+void optimize_funccalls(void)
+{
+       /* prepare: mark all graphs as not analyzed */
+       size_t last_idx = get_irp_last_idx();
+       ready_set = rbitset_malloc(last_idx);
+       busy_set  = rbitset_malloc(last_idx);
+
+       /* first step: detect, which functions are nothrow or malloc */
+       DB((dbg, LEVEL_2, "Detecting nothrow and malloc properties ...\n"));
+       for (size_t i = 0, n = get_irp_n_irgs(); i < n; ++i) {
+               ir_graph *irg  = get_irp_irg(i);
+               unsigned  prop = check_nothrow_or_malloc(irg, true);
+
+               if (prop & mtp_property_nothrow) {
+                       DB((dbg, LEVEL_2, "%+F has the nothrow property\n", irg));
+               } else if (prop & mtp_property_malloc) {
+                       DB((dbg, LEVEL_2, "%+F has the malloc property\n", irg));
+               }
+       }
+
+       /* second step: remove exception edges: this must be done before the
+          detection of const and pure functions take place. */
+       env_t ctx;
+       handle_nothrow_Calls(&ctx);
+
+       rbitset_clear_all(ready_set, last_idx);
+       rbitset_clear_all(busy_set, last_idx);
+
+       /* third step: detect, which functions are const or pure */
+       DB((dbg, LEVEL_2, "Detecting const and pure properties ...\n"));
+       for (size_t i = 0, n = get_irp_n_irgs(); i < n; ++i) {
+               ir_graph *irg  = get_irp_irg(i);
+               unsigned  prop = check_const_or_pure_function(irg, true);
+
+               if (prop & mtp_property_const) {
+                       DB((dbg, LEVEL_2, "%+F has the const property\n", irg));
+                       check_for_possible_endless_loops(irg);
+               } else if (prop & mtp_property_pure) {
+                       DB((dbg, LEVEL_2, "%+F has the pure property\n", irg));
+               }
+       }
+
+       handle_const_Calls(&ctx);
+
+       xfree(busy_set);
+       xfree(ready_set);
+}
+
+void firm_init_funccalls(void)
+{
+       FIRM_DBG_REGISTER(dbg, "firm.opt.funccalls");
+}
+
+ir_prog_pass_t *optimize_funccalls_pass(const char *name)
+{
+       return def_prog_pass(name ? name : "funccall", optimize_funccalls);
 }