Put opening curly brace of functions on a separate line.
[libfirm] / ir / opt / ldstopt.c
index 39ce34c..ac7137e 100644 (file)
@@ -23,9 +23,7 @@
  * @author  Michael Beck
  * @version $Id$
  */
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
 
 #include <string.h>
 
 #include "dbginfo_t.h"
 #include "iropt_dbg.h"
 #include "irflag_t.h"
-#include "array.h"
+#include "array_t.h"
 #include "irhooks.h"
 #include "iredges.h"
-#include "irtools.h"
+#include "irpass.h"
 #include "opt_polymorphy.h"
 #include "irmemory.h"
-#include "xmalloc.h"
 #include "irphase_t.h"
 #include "irgopt.h"
+#include "set.h"
 #include "debug.h"
 
 /** The debug handle. */
@@ -111,12 +109,12 @@ static unsigned master_visited = 0;
 /**
  * get the Load/Store info of a node
  */
-static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
+static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
+{
        ldst_info_t *info = get_irn_link(node);
 
        if (! info) {
-               info = obstack_alloc(obst, sizeof(*info));
-               memset(info, 0, sizeof(*info));
+               info = OALLOCZ(obst, ldst_info_t);
                set_irn_link(node, info);
        }
        return info;
@@ -125,12 +123,12 @@ static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
 /**
  * get the Block info of a node
  */
-static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
+static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
+{
        block_info_t *info = get_irn_link(node);
 
        if (! info) {
-               info = obstack_alloc(obst, sizeof(*info));
-               memset(info, 0, sizeof(*info));
+               info = OALLOCZ(obst, block_info_t);
                set_irn_link(node, info);
        }
        return info;
@@ -257,10 +255,7 @@ static ir_entity *find_constant_entity(ir_node *ptr)
 {
        for (;;) {
                if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
-                       ir_entity *ent = get_SymConst_entity(ptr);
-                       if (variability_constant == get_entity_variability(ent))
-                               return ent;
-                       return NULL;
+                       return get_SymConst_entity(ptr);
                } else if (is_Sel(ptr)) {
                        ir_entity *ent = get_Sel_entity(ptr);
                        ir_type   *tp  = get_entity_owner(ent);
@@ -302,7 +297,7 @@ static ir_entity *find_constant_entity(ir_node *ptr)
                                }
                        }
 
-                       if (variability_constant == get_entity_variability(ent))
+                       if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
                                return ent;
 
                        /* try next */
@@ -340,7 +335,8 @@ static ir_entity *find_constant_entity(ir_node *ptr)
 /**
  * Return the Selection index of a Sel node from dimension n
  */
-static long get_Sel_array_index_long(ir_node *n, int dim) {
+static long get_Sel_array_index_long(ir_node *n, int dim)
+{
        ir_node *index = get_Sel_index(n, dim);
        assert(is_Const(index));
        return get_tarval_long(get_Const_tarval(index));
@@ -354,7 +350,8 @@ static long get_Sel_array_index_long(ir_node *n, int dim) {
  * @param depth  current depth in steps upward from the root
  *               of the address
  */
-static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
+static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth)
+{
        compound_graph_path *res = NULL;
        ir_entity           *root, *field, *ent;
        int                 path_len, pos, idx;
@@ -364,7 +361,7 @@ static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
        if (is_SymConst(ptr)) {
                /* a SymConst. If the depth is 0, this is an access to a global
                 * entity and we don't need a component path, else we know
-                * at least it's length.
+                * at least its length.
                 */
                assert(get_SymConst_kind(ptr) == symconst_addr_ent);
                root = get_SymConst_entity(ptr);
@@ -386,11 +383,12 @@ static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
                        set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
                }
        } else if (is_Add(ptr)) {
-               ir_node *l = get_Add_left(ptr);
-               ir_node *r = get_Add_right(ptr);
-               ir_mode *mode;
+               ir_node *l    = get_Add_left(ptr);
+               ir_node *r    = get_Add_right(ptr);
+               ir_mode *mode = get_irn_mode(ptr);
+               tarval  *tmp;
 
-               if (is_Const(r)) {
+               if (is_Const(r) && get_irn_mode(l) == mode) {
                        ptr = l;
                        tv  = get_Const_tarval(r);
                } else {
@@ -399,6 +397,7 @@ static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
                }
 ptr_arith:
                mode = get_tarval_mode(tv);
+               tmp  = tv;
 
                /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
                if (is_Sel(ptr)) {
@@ -419,10 +418,10 @@ ptr_arith:
                        size = get_type_size_bytes(get_entity_type(ent));
                        sz   = new_tarval_from_long(size, mode);
 
-                       tv_index = tarval_div(tv, sz);
-                       tv       = tarval_mod(tv, sz);
+                       tv_index = tarval_div(tmp, sz);
+                       tmp      = tarval_mod(tmp, sz);
 
-                       if (tv_index == tarval_bad || tv == tarval_bad)
+                       if (tv_index == tarval_bad || tmp == tarval_bad)
                                return NULL;
 
                        assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
@@ -442,7 +441,7 @@ ptr_arith:
                        /* ok, bounds check finished */
                        ++idx;
                }
-               if (! tarval_is_null(tv)) {
+               if (! tarval_is_null(tmp)) {
                        /* access to some struct/union member */
                        return NULL;
                }
@@ -499,8 +498,10 @@ ptr_arith:
  * Returns an access path or NULL.  The access path is only
  * valid, if the graph is in phase_high and _no_ address computation is used.
  */
-static compound_graph_path *get_accessed_path(ir_node *ptr) {
-       return rec_get_accessed_path(ptr, 0);
+static compound_graph_path *get_accessed_path(ir_node *ptr)
+{
+       compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
+       return gr;
 }  /* get_accessed_path */
 
 typedef struct path_entry {
@@ -509,7 +510,8 @@ typedef struct path_entry {
        long              index;
 } path_entry;
 
-static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
+static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
+{
        path_entry       entry, *p;
        ir_entity        *ent, *field;
        ir_initializer_t *initializer;
@@ -683,7 +685,8 @@ ptr_arith:
        return NULL;
 }
 
-static ir_node *find_compound_ent_value(ir_node *ptr) {
+static ir_node *find_compound_ent_value(ir_node *ptr)
+{
        return rec_find_compound_ent_value(ptr, NULL);
 }
 
@@ -691,9 +694,10 @@ static ir_node *find_compound_ent_value(ir_node *ptr) {
 static void reduce_adr_usage(ir_node *ptr);
 
 /**
- * Update a Load that may lost it's usage.
+ * Update a Load that may have lost its users.
  */
-static void handle_load_update(ir_node *load) {
+static void handle_load_update(ir_node *load)
+{
        ldst_info_t *info = get_irn_link(load);
 
        /* do NOT touch volatile loads for now */
@@ -704,20 +708,21 @@ static void handle_load_update(ir_node *load) {
                ir_node *ptr = get_Load_ptr(load);
                ir_node *mem = get_Load_mem(load);
 
-               /* a Load which value is neither used nor exception checked, remove it */
+               /* a Load whose value is neither used nor exception checked, remove it */
                exchange(info->projs[pn_Load_M], mem);
                if (info->projs[pn_Load_X_regular])
-                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
+                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
                kill_node(load);
                reduce_adr_usage(ptr);
        }
 }  /* handle_load_update */
 
 /**
- * A Use of an address node is vanished. Check if this was a Proj
+ * A use of an address node has vanished. Check if this was a Proj
  * node and update the counters.
  */
-static void reduce_adr_usage(ir_node *ptr) {
+static void reduce_adr_usage(ir_node *ptr)
+{
        if (is_Proj(ptr)) {
                if (get_irn_n_edges(ptr) <= 0) {
                        /* this Proj is dead now */
@@ -727,7 +732,7 @@ static void reduce_adr_usage(ir_node *ptr) {
                                ldst_info_t *info = get_irn_link(pred);
                                info->projs[get_Proj_proj(ptr)] = NULL;
 
-                               /* this node lost it's result proj, handle that */
+                               /* this node lost its result proj, handle that */
                                handle_load_update(pred);
                        }
                }
@@ -738,7 +743,8 @@ static void reduce_adr_usage(ir_node *ptr) {
  * Check, if an already existing value of mode old_mode can be converted
  * into the needed one new_mode without loss.
  */
-static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
+static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
+{
        if (old_mode == new_mode)
                return 1;
 
@@ -754,7 +760,8 @@ static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 /**
  * Check whether a Call is at least pure, ie. does only read memory.
  */
-static unsigned is_Call_pure(ir_node *call) {
+static unsigned is_Call_pure(ir_node *call)
+{
        ir_type *call_tp = get_Call_type(call);
        unsigned prop = get_method_additional_properties(call_tp);
 
@@ -856,29 +863,31 @@ static int try_load_after_store(ir_node *load,
        store_mode     = get_irn_mode(get_Store_value(store));
        store_mode_len = get_mode_size_bytes(store_mode);
        delta          = load_offset - store_offset;
-       if (delta < 0 || delta + load_mode_len > store_mode_len)
-               return 0;
+       store_value    = get_Store_value(store);
 
-       if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
-           get_mode_arithmetic(load_mode)  != irma_twos_complement)
-               return 0;
+       if (delta != 0 || store_mode != load_mode) {
+               if (delta < 0 || delta + load_mode_len > store_mode_len)
+                       return 0;
 
-       store_value = get_Store_value(store);
+               if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
+                       get_mode_arithmetic(load_mode)  != irma_twos_complement)
+                       return 0;
 
-       /* produce a shift to adjust offset delta */
-       if (delta > 0) {
-               ir_node *cnst;
 
-               /* FIXME: only true for little endian */
-               cnst        = new_Const_long(mode_Iu, delta * 8);
-               store_value = new_r_Shr(current_ir_graph, get_nodes_block(load),
-                                       store_value, cnst, store_mode);
-       }
+               /* produce a shift to adjust offset delta */
+               if (delta > 0) {
+                       ir_node *cnst;
 
-       /* add an convert if needed */
-       if (store_mode != load_mode) {
-               store_value = new_r_Conv(current_ir_graph, get_nodes_block(load),
-                                        store_value, load_mode);
+                       /* FIXME: only true for little endian */
+                       cnst        = new_Const_long(mode_Iu, delta * 8);
+                       store_value = new_r_Shr(get_nodes_block(load),
+                                                                       store_value, cnst, store_mode);
+               }
+
+               /* add an convert if needed */
+               if (store_mode != load_mode) {
+                       store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
+               }
        }
 
        DBG_OPT_RAW(load, store_value);
@@ -894,7 +903,7 @@ static int try_load_after_store(ir_node *load,
                res |= CF_CHANGED;
        }
        if (info->projs[pn_Load_X_regular]) {
-               exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
+               exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
                res |= CF_CHANGED;
        }
 
@@ -917,7 +926,8 @@ static int try_load_after_store(ir_node *load,
  *
  * INC_MASTER() must be called before dive into
  */
-static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
+static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
+{
        unsigned    res = 0;
        ldst_info_t *info = get_irn_link(load);
        ir_node     *pred;
@@ -971,13 +981,13 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                                if (info->projs[pn_Load_res]) {
                                        if (pred_info->projs[pn_Load_res] == NULL) {
                                                /* create a new Proj again */
-                                               pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
+                                               pred_info->projs[pn_Load_res] = new_r_Proj(get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
                                        }
                                        value = pred_info->projs[pn_Load_res];
 
                                        /* add an convert if needed */
                                        if (get_Load_mode(pred) != load_mode) {
-                                               value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
+                                               value = new_r_Conv(get_nodes_block(load), value, load_mode);
                                        }
 
                                        exchange(info->projs[pn_Load_res], value);
@@ -992,7 +1002,7 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                                        res |= CF_CHANGED;
                                }
                                if (info->projs[pn_Load_X_regular]) {
-                                       exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
+                                       exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
                                        res |= CF_CHANGED;
                                }
 
@@ -1053,7 +1063,8 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
  * Check if we can replace the load by a given const from
  * the const code irg.
  */
-ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
+ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
+{
        ir_mode *c_mode = get_irn_mode(c);
        ir_mode *l_mode = get_Load_mode(load);
        ir_node *res    = NULL;
@@ -1067,7 +1078,7 @@ ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
 
                        /* copy the value from the const code irg and cast it */
                        res = copy_const_value(dbg, c);
-                       res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
+                       res = new_rd_Conv(dbg, block, res, l_mode);
                }
        } else {
                /* copy the value from the const code irg */
@@ -1118,14 +1129,14 @@ static unsigned optimize_load(ir_node *load)
                         */
                        exchange(info->projs[pn_Load_X_except], new_Bad());
                        info->projs[pn_Load_X_except] = NULL;
-                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
+                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
                        info->projs[pn_Load_X_regular] = NULL;
                        res |= CF_CHANGED;
                }
        }
 
        /* The mem of the Load. Must still be returned after optimization. */
-       mem  = get_Load_mem(load);
+       mem = get_Load_mem(load);
 
        if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
                /* a Load which value is neither used nor exception checked, remove it */
@@ -1133,7 +1144,7 @@ static unsigned optimize_load(ir_node *load)
 
                if (info->projs[pn_Load_X_regular]) {
                        /* should not happen, but if it does, remove it */
-                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
+                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
                        res |= CF_CHANGED;
                }
                kill_node(load);
@@ -1148,53 +1159,42 @@ static unsigned optimize_load(ir_node *load)
                value = NULL;
                /* check if we can determine the entity that will be loaded */
                ent = find_constant_entity(ptr);
-               if (ent != NULL) {
-                       if ((allocation_static == get_entity_allocation(ent)) &&
-                               (visibility_external_allocated != get_entity_visibility(ent))) {
-                               /* a static allocation that is not external: there should be NO exception
-                                * when loading even if we cannot replace the load itself. */
+               if (ent != NULL
+                               && get_entity_visibility(ent) != ir_visibility_external) {
+                       /* a static allocation that is not external: there should be NO
+                        * exception when loading even if we cannot replace the load itself.
+                        */
 
-                               /* no exception, clear the info field as it might be checked later again */
-                               if (info->projs[pn_Load_X_except]) {
-                                       exchange(info->projs[pn_Load_X_except], new_Bad());
-                                       info->projs[pn_Load_X_except] = NULL;
-                                       res |= CF_CHANGED;
-                               }
-                               if (info->projs[pn_Load_X_regular]) {
-                                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
-                                       info->projs[pn_Load_X_regular] = NULL;
-                                       res |= CF_CHANGED;
-                               }
+                       /* no exception, clear the info field as it might be checked later again */
+                       if (info->projs[pn_Load_X_except]) {
+                               exchange(info->projs[pn_Load_X_except], new_Bad());
+                               info->projs[pn_Load_X_except] = NULL;
+                               res |= CF_CHANGED;
+                       }
+                       if (info->projs[pn_Load_X_regular]) {
+                               exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
+                               info->projs[pn_Load_X_regular] = NULL;
+                               res |= CF_CHANGED;
+                       }
 
-                               if (variability_constant == get_entity_variability(ent)) {
-                                       if (is_atomic_entity(ent)) {
-                                               /* Might not be atomic after
-                                                  lowering of Sels.  In this
-                                                  case we could also load, but
-                                                  it's more complicated. */
-                                               /* more simpler case: we load the content of a constant value:
-                                                * replace it by the constant itself
-                                                */
-                                               value = get_atomic_ent_value(ent);
-                                       } else {
-                                               if (ent->has_initializer) {
-                                                       /* new style initializer */
-                                                       value = find_compound_ent_value(ptr);
-                                               } else {
-                                                       /* old style initializer */
-                                                       compound_graph_path *path = get_accessed_path(ptr);
-
-                                                       if (path != NULL) {
-                                                               assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
-
-                                                               value = get_compound_ent_value_by_path(ent, path);
-                                                               free_compound_graph_path(path);
-                                                       }
-                                               }
+                       if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
+                               if (ent->initializer != NULL) {
+                                       /* new style initializer */
+                                       value = find_compound_ent_value(ptr);
+                               } else if (entity_has_compound_ent_values(ent)) {
+                                       /* old style initializer */
+                                       compound_graph_path *path = get_accessed_path(ptr);
+
+                                       if (path != NULL) {
+                                               assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
+
+                                               value = get_compound_ent_value_by_path(ent, path);
+                                               DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
+                                               free_compound_graph_path(path);
                                        }
-                                       if (value != NULL)
-                                               value = can_replace_load_by_const(load, value);
                                }
+                               if (value != NULL)
+                                       value = can_replace_load_by_const(load, value);
                        }
                }
        }
@@ -1206,7 +1206,7 @@ static unsigned optimize_load(ir_node *load)
                        res |= CF_CHANGED;
                }
                if (info->projs[pn_Load_X_regular]) {
-                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
+                       exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
                        info->projs[pn_Load_X_regular] = NULL;
                        res |= CF_CHANGED;
                }
@@ -1249,12 +1249,28 @@ static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
        return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 }  /* is_completely_overwritten */
 
+/**
+ * Check whether small is a part of large (starting at same address).
+ */
+static int is_partially_same(ir_node *small, ir_node *large)
+{
+       ir_mode *sm = get_irn_mode(small);
+       ir_mode *lm = get_irn_mode(large);
+
+       /* FIXME: Check endianness */
+       return is_Conv(small) && get_Conv_op(small) == large
+           && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
+           && get_mode_arithmetic(sm) == irma_twos_complement
+           && get_mode_arithmetic(lm) == irma_twos_complement;
+}  /* is_partially_same */
+
 /**
  * follow the memory chain as long as there are only Loads and alias free Stores.
  *
  * INC_MASTER() must be called before dive into
  */
-static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
+static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
+{
        unsigned res = 0;
        ldst_info_t *info = get_irn_link(store);
        ir_node *pred;
@@ -1273,24 +1289,55 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
                 * if the pointers are identical, they refer to the same object.
                 * This is only true in strong typed languages, not is C were the following
                 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
-                * However, if the mode that is written have a bigger  or equal size the the old
-                * one, the old value is completely overwritten and can be killed ...
+                * However, if the size of the mode that is written is bigger or equal the
+                * size of the old one, the old value is completely overwritten and can be
+                * killed ...
                 */
                if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
-                   get_nodes_MacroBlock(pred) == mblk &&
-                   is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
+            get_nodes_MacroBlock(pred) == mblk) {
                        /*
                         * a Store after a Store in the same MacroBlock -- a write after write.
-                        * We may remove the first Store, if it does not have an exception handler.
+                        */
+
+                       /*
+                        * We may remove the first Store, if the old value is completely
+                        * overwritten or the old value is a part of the new value,
+                        * and if it does not have an exception handler.
                         *
                         * TODO: What, if both have the same exception handler ???
                         */
-                       if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
-                               DBG_OPT_WAW(pred, store);
-                               exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
-                               kill_node(pred);
-                               reduce_adr_usage(ptr);
-                               return DF_CHANGED;
+                       if (get_Store_volatility(pred) != volatility_is_volatile
+                               && !pred_info->projs[pn_Store_X_except]) {
+                               ir_node *predvalue = get_Store_value(pred);
+                               ir_mode *predmode  = get_irn_mode(predvalue);
+
+                               if(is_completely_overwritten(predmode, mode)
+                                       || is_partially_same(predvalue, value)) {
+                                       DBG_OPT_WAW(pred, store);
+                                       exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
+                                       kill_node(pred);
+                                       reduce_adr_usage(ptr);
+                                       return DF_CHANGED;
+                               }
+                       }
+
+                       /*
+                        * We may remove the Store, if the old value already contains
+                        * the new value, and if it does not have an exception handler.
+                        *
+                        * TODO: What, if both have the same exception handler ???
+                        */
+                       if (get_Store_volatility(store) != volatility_is_volatile
+                               && !info->projs[pn_Store_X_except]) {
+                               ir_node *predvalue = get_Store_value(pred);
+
+                               if(is_partially_same(value, predvalue)) {
+                                       DBG_OPT_WAW(pred, store);
+                                       exchange(info->projs[pn_Store_M], mem);
+                                       kill_node(store);
+                                       reduce_adr_usage(ptr);
+                                       return DF_CHANGED;
+                               }
                        }
                } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
                           value == pred_info->projs[pn_Load_res]) {
@@ -1386,7 +1433,8 @@ static ir_entity *find_entity(ir_node *ptr)
  *
  * @param store  the Store node
  */
-static unsigned optimize_store(ir_node *store) {
+static unsigned optimize_store(ir_node *store)
+{
        ir_node   *ptr;
        ir_node   *mem;
        ir_entity *entity;
@@ -1401,6 +1449,7 @@ static unsigned optimize_store(ir_node *store) {
        if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
                ldst_info_t *info = get_irn_link(store);
                if (info->projs[pn_Store_X_except] == NULL) {
+                       DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
                        exchange(info->projs[pn_Store_M], get_Store_mem(store));
                        kill_node(store);
                        reduce_adr_usage(ptr);
@@ -1569,10 +1618,10 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
        block = get_nodes_block(phi);
 
        /* second step: create a new memory Phi */
-       phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
+       phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
 
        /* third step: create a new data Phi */
-       phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
+       phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
 
        /* rewire memory and kill the node */
        for (i = n - 1; i >= 0; --i) {
@@ -1586,19 +1635,19 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
        }
 
        /* fourth step: create the Store */
-       store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
+       store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
 #ifdef DO_CACHEOPT
        co_set_irn_name(store, co_get_irn_ident(old_store));
 #endif
 
-       projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
+       projM = new_rd_Proj(NULL, block, store, mode_M, pn_Store_M);
 
        info = get_ldst_info(store, &wenv->obst);
        info->projs[pn_Store_M] = projM;
 
        /* fifths step: repair exception flow */
        if (exc) {
-               ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
+               ir_node *projX = new_rd_Proj(NULL, block, store, mode_X, pn_Store_X_except);
 
                info->projs[pn_Store_X_except] = projX;
                info->exc_block                = exc;
@@ -1624,7 +1673,8 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 /**
  * walker, do the optimizations
  */
-static void do_load_store_optimize(ir_node *n, void *env) {
+static void do_load_store_optimize(ir_node *n, void *env)
+{
        walk_env_t *wenv = env;
 
        switch (get_irn_opcode(n)) {
@@ -1655,7 +1705,6 @@ typedef struct scc {
 typedef struct node_entry {
        unsigned DFSnum;    /**< the DFS number of this node */
        unsigned low;       /**< the low number of this node */
-       ir_node  *header;   /**< the header of this node */
        int      in_stack;  /**< flag, set if the node is on the stack */
        ir_node  *next;     /**< link to the next node the the same scc */
        scc      *pscc;     /**< the scc of this node */
@@ -1676,7 +1725,8 @@ typedef struct loop_env {
 /**
 * Gets the node_entry of a node
 */
-static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
+static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
+{
        ir_phase   *ph = &env->ph;
        node_entry *e  = phase_get_irn_data(&env->ph, irn);
 
@@ -1694,7 +1744,8 @@ static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
  * @param env   the loop environment
  * @param n     the node to push
  */
-static void push(loop_env *env, ir_node *n) {
+static void push(loop_env *env, ir_node *n)
+{
        node_entry *e;
 
        if (env->tos == ARR_LEN(env->stack)) {
@@ -1713,7 +1764,8 @@ static void push(loop_env *env, ir_node *n) {
  *
  * @return  The topmost node
  */
-static ir_node *pop(loop_env *env) {
+static ir_node *pop(loop_env *env)
+{
        ir_node *n = env->stack[--env->tos];
        node_entry *e = get_irn_ne(n, env);
 
@@ -1728,7 +1780,8 @@ static ir_node *pop(loop_env *env) {
  * @param irn           the node to check
  * @param header_block  the header block of the induction variable
  */
-static int is_rc(ir_node *irn, ir_node *header_block) {
+static int is_rc(ir_node *irn, ir_node *header_block)
+{
        ir_node *block = get_nodes_block(irn);
 
        return (block != header_block) && block_dominates(block, header_block);
@@ -1742,17 +1795,50 @@ struct phi_entry {
        phi_entry *next;
 };
 
+/**
+ * An entry in the avail set.
+ */
+typedef struct avail_entry_t {
+       ir_node *ptr;   /**< the address pointer */
+       ir_mode *mode;  /**< the load mode */
+       ir_node *load;  /**< the associated Load */
+} avail_entry_t;
+
+/**
+ * Compare two avail entries.
+ */
+static int cmp_avail_entry(const void *elt, const void *key, size_t size)
+{
+       const avail_entry_t *a = elt;
+       const avail_entry_t *b = key;
+       (void) size;
+
+       return a->ptr != b->ptr || a->mode != b->mode;
+}  /* cmp_avail_entry */
+
+/**
+ * Calculate the hash value of an avail entry.
+ */
+static unsigned hash_cache_entry(const avail_entry_t *entry)
+{
+       return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
+}  /* hash_cache_entry */
+
 /**
  * Move loops out of loops if possible.
  *
  * @param pscc   the loop described by an SCC
  * @param env    the loop environment
  */
-static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
+static void move_loads_out_of_loops(scc *pscc, loop_env *env)
+{
        ir_node   *phi, *load, *next, *other, *next_other;
        ir_entity *ent;
        int       j;
        phi_entry *phi_list = NULL;
+       set       *avail;
+
+       avail = new_set(cmp_avail_entry, 8);
 
        /* collect all outer memories */
        for (phi = pscc->head; phi != NULL; phi = next) {
@@ -1763,7 +1849,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
                if (! is_Phi(phi))
                        continue;
 
-               assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
+               assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
 
                for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
                        ir_node    *pred = get_irn_n(phi, j);
@@ -1783,6 +1869,10 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
        /* no Phis no fun */
        assert(phi_list != NULL && "DFS found a loop without Phi");
 
+       /* for now, we cannot handle more than one input (only reducible cf) */
+       if (phi_list->next != NULL)
+               return;
+
        for (load = pscc->head; load; load = next) {
                ir_mode *load_mode;
                node_entry *ne = get_irn_ne(load, env);
@@ -1796,10 +1886,10 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
                        if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
                                continue;
 
-                       /* for now, we can only handle Load(Global) */
+                       /* for now, we can only move Load(Global) */
                        if (! is_Global(ptr))
                                continue;
-                       ent = get_Global_entity(ptr);
+                       ent       = get_Global_entity(ptr);
                        load_mode = get_Load_mode(load);
                        for (other = pscc->head; other != NULL; other = next_other) {
                                node_entry *ne = get_irn_ne(other, env);
@@ -1815,17 +1905,13 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
                                        if (rel != ir_no_alias)
                                                break;
                                }
-                               /* only pure Calls are allowed here, so ignore them */
+                               /* only Phis and pure Calls are allowed here, so ignore them */
                        }
                        if (other == NULL) {
                                ldst_info_t *ninfo;
                                phi_entry   *pe;
                                dbg_info    *db;
 
-                               /* for now, we cannot handle more than one input */
-                               if (phi_list->next != NULL)
-                                       return;
-
                                /* yep, no aliasing Store found, Load can be moved */
                                DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
 
@@ -1836,16 +1922,26 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
                                        ir_node *blk  = get_nodes_block(phi);
                                        ir_node *pred = get_Block_cfgpred_block(blk, pos);
                                        ir_node *irn, *mem;
+                                       avail_entry_t entry, *res;
 
-                                       pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
+                                       entry.ptr  = ptr;
+                                       entry.mode = load_mode;
+                                       res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
+                                       if (res != NULL) {
+                                               irn = res->load;
+                                       } else {
+                                               irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
+                                               entry.load = irn;
+                                               set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
+                                               DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
+                                       }
+                                       pe->load = irn;
                                        ninfo = get_ldst_info(irn, phase_obst(&env->ph));
 
-                                       ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
+                                       ninfo->projs[pn_Load_M] = mem = new_r_Proj(pred, irn, mode_M, pn_Load_M);
                                        set_Phi_pred(phi, pos, mem);
 
-                                       ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
-
-                                       DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
+                                       ninfo->projs[pn_Load_res] = new_r_Proj(pred, irn, load_mode, pn_Load_res);
                                }
 
                                /* now kill the old Load */
@@ -1856,6 +1952,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
                        }
                }
        }
+       del_set(avail);
 }  /* move_loads_out_of_loops */
 
 /**
@@ -1864,7 +1961,8 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
  * @param pscc  the SCC
  * @param env   the loop environment
  */
-static void process_loop(scc *pscc, loop_env *env) {
+static void process_loop(scc *pscc, loop_env *env)
+{
        ir_node *irn, *next, *header = NULL;
        node_entry *b, *h = NULL;
        int j, only_phi, num_outside, process = 0;
@@ -1878,13 +1976,12 @@ static void process_loop(scc *pscc, loop_env *env) {
                next = e->next;
                b = get_irn_ne(block, env);
 
-               if (header) {
+               if (header != NULL) {
                        if (h->POnum < b->POnum) {
                                header = block;
                                h      = b;
                        }
-               }
-               else {
+               } else {
                        header = block;
                        h      = b;
                }
@@ -1939,10 +2036,12 @@ static void process_loop(scc *pscc, loop_env *env) {
                                                /* not a memory loop */
                                                goto fail;
                                        }
-                                       if (! out_rc) {
+                                       if (out_rc == NULL) {
+                                               /* first region constant */
                                                out_rc = pred;
                                                ++num_outside;
                                        } else if (out_rc != pred) {
+                                               /* another region constant */
                                                ++num_outside;
                                        }
                                }
@@ -1962,22 +2061,20 @@ static void process_loop(scc *pscc, loop_env *env) {
                for (irn = pscc->head; irn; irn = next) {
                        node_entry *e = get_irn_ne(irn, env);
                        next = e->next;
-                       e->header = NULL;
                        exchange(irn, out_rc);
                }
                env->changes |= DF_CHANGED;
                return;
        }
 
-       /* set the header for every node in this scc */
+#ifdef DEBUG_libfirm
        for (irn = pscc->head; irn; irn = next) {
                node_entry *e = get_irn_ne(irn, env);
-               e->header = header;
                next = e->next;
                DB((dbg, LEVEL_2, " %+F,", irn));
        }
        DB((dbg, LEVEL_2, "\n"));
-
+#endif
        move_loads_out_of_loops(pscc, env);
 
 fail:
@@ -1990,7 +2087,8 @@ fail:
  * @param pscc  the SCC
  * @param env   the loop environment
  */
-static void process_scc(scc *pscc, loop_env *env) {
+static void process_scc(scc *pscc, loop_env *env)
+{
        ir_node *head = pscc->head;
        node_entry *e = get_irn_ne(head, env);
 
@@ -2040,7 +2138,7 @@ static void dfs(ir_node *irn, loop_env *env)
                        ir_node *pred = get_irn_n(irn, i);
                        node_entry *o = get_irn_ne(pred, env);
 
-                       if (irn_not_visited(pred)) {
+                       if (!irn_visited(pred)) {
                                dfs(pred, env);
                                node->low = MIN(node->low, o->low);
                        }
@@ -2051,7 +2149,7 @@ static void dfs(ir_node *irn, loop_env *env)
                ir_node *pred = get_fragile_op_mem(irn);
                node_entry *o = get_irn_ne(pred, env);
 
-               if (irn_not_visited(pred)) {
+               if (!irn_visited(pred)) {
                        dfs(pred, env);
                        node->low = MIN(node->low, o->low);
                }
@@ -2061,7 +2159,7 @@ static void dfs(ir_node *irn, loop_env *env)
                ir_node *pred = get_Proj_pred(irn);
                node_entry *o = get_irn_ne(pred, env);
 
-               if (irn_not_visited(pred)) {
+               if (!irn_visited(pred)) {
                        dfs(pred, env);
                        node->low = MIN(node->low, o->low);
                }
@@ -2097,7 +2195,8 @@ static void dfs(ir_node *irn, loop_env *env)
  * @param irg  the graph to process
  * @param env  the loop environment
  */
-static void do_dfs(ir_graph *irg, loop_env *env) {
+static void do_dfs(ir_graph *irg, loop_env *env)
+{
        ir_graph *rem = current_ir_graph;
        ir_node  *endblk, *end;
        int      i;
@@ -2127,7 +2226,7 @@ static void do_dfs(ir_graph *irg, loop_env *env) {
        for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
                ir_node *ka = get_End_keepalive(end, i);
 
-               if (is_Phi(ka) && irn_not_visited(ka))
+               if (is_Phi(ka) && !irn_visited(ka))
                        dfs(ka, env);
        }
        current_ir_graph = rem;
@@ -2136,7 +2235,8 @@ static void do_dfs(ir_graph *irg, loop_env *env) {
 /**
  * Initialize new phase data. We do this always explicit, so return NULL here
  */
-static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
+static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data)
+{
        (void)ph;
        (void)irn;
        (void)data;
@@ -2148,7 +2248,8 @@ static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
  *
  * @param irg  the graph
  */
-static int optimize_loops(ir_graph *irg) {
+static int optimize_loops(ir_graph *irg)
+{
        loop_env env;
 
        env.stack         = NEW_ARR_F(ir_node *, 128);
@@ -2170,7 +2271,8 @@ static int optimize_loops(ir_graph *irg) {
 /*
  * do the load store optimization
  */
-void optimize_load_store(ir_graph *irg) {
+int optimize_load_store(ir_graph *irg)
+{
        walk_env_t env;
 
        FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
@@ -2217,4 +2319,10 @@ void optimize_load_store(ir_graph *irg) {
                have Bad() predecessors. */
                set_irg_doms_inconsistent(irg);
        }
+       return env.changes != 0;
 }  /* optimize_load_store */
+
+ir_graph_pass_t *optimize_load_store_pass(const char *name)
+{
+       return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
+}  /* optimize_load_store_pass */