rematerialized values are like reloads
[libfirm] / ir / be / bespill.c
index e1f93bb..04e801a 100644 (file)
@@ -1,8 +1,28 @@
+/*
+ * Copyright (C) 1995-2007 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.
+ */
+
 /**
- * Author:      Daniel Grund, Sebastian Hack
- * Date:               29.09.2005
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+ * @file
+ * @brief       implementation of the spill/reload placement abstraction layer
+ * @author      Daniel Grund, Sebastian Hack, Matthias Braun
+ * @date               29.09.2005
+ * @version     $Id$
  */
 #ifdef HAVE_CONFIG_H
 #include "config.h"
@@ -14,6 +34,8 @@
 #include "irnode_t.h"
 #include "ircons_t.h"
 #include "iredges_t.h"
+#include "irbackedge_t.h"
+#include "irprintf.h"
 #include "ident_t.h"
 #include "type_t.h"
 #include "entity_t.h"
 #include "irgwalk.h"
 #include "array.h"
 #include "pdeq.h"
+#include "execfreq.h"
+#include "irnodeset.h"
+#include "error.h"
 
+#include "bearch_t.h"
 #include "belive_t.h"
 #include "besched_t.h"
 #include "bespill.h"
+#include "belive_t.h"
 #include "benode_t.h"
 #include "bechordal_t.h"
+#include "bejavacoal.h"
+#include "benodesets.h"
+#include "bespilloptions.h"
+#include "bestatevent.h"
+#include "bessaconstr.h"
+#include "beirg_t.h"
+#include "beintlive_t.h"
+#include "bemodule.h"
 
-#define REMAT
-/* This enables re-computation of values. Current state: Unfinished and buggy. */
-#undef BUGGY_REMAT
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
-typedef struct _reloader_t reloader_t;
-typedef struct _spill_info_t spill_info_t;
+#define REMAT_COST_INFINITE  1000
 
-struct _reloader_t {
+typedef struct reloader_t reloader_t;
+struct reloader_t {
        reloader_t *next;
-       ir_node *reloader;
+       ir_node    *can_spill_after;
+       ir_node    *reloader;
+       ir_node    *rematted_node;
+       int         remat_cost_delta; /** costs needed for rematerialization,
+                                          compared to placing a reload */
 };
 
-struct _spill_info_t {
-       ir_node *spilled_node;
-       reloader_t *reloaders;
+typedef struct spill_t spill_t;
+struct spill_t {
+       spill_t *next;
+       ir_node *before;   /**< spill has to be placed before this node (or earlier) */
+       ir_node *spill;
 };
 
-typedef struct _spill_ctx_t {
-       ir_node *spilled;  /**< The spilled node. */
-       ir_node *user;     /**< The node this spill is for. */
-       ir_node *spill;    /**< The spill itself. */
-} spill_ctx_t;
-
-struct _spill_env_t {
-       const arch_register_class_t *cls;
-       const be_chordal_env_t *chordal_env;
-       struct obstack obst;
-       set *spill_ctxs;
-       set *spills;                            /**< all spill_info_t's, which must be placed */
-       pset *mem_phis;                         /**< set of all special spilled phis. allocated and freed separately */
-       ir_node **copies;                       /**< set of copies placed because of phi spills */
-       DEBUG_ONLY(firm_dbg_module_t *dbg;)
+typedef struct spill_info_t spill_info_t;
+struct spill_info_t {
+       ir_node    *to_spill;  /**< the value that should get spilled */
+       reloader_t *reloaders; /**< list of places where the value should get
+                                   reloaded */
+       spill_t    *spills;    /**< list of latest places where spill must be
+                                   placed */
+       double      spill_costs; /**< costs needed for spilling the value */
+       const arch_register_class_t *reload_cls; /** the register class in which the
+                                                    reload should be placed */
 };
 
-/* associated Phi -> Spill*/
-typedef struct _phi_spill_assoc_t {
-       ir_node *phi;
-       ir_node *spill;
-} phi_spill_assoc_t;
+struct spill_env_t {
+       const arch_env_t *arch_env;
+       ir_graph         *irg;
+       struct obstack    obst;
+       be_irg_t         *birg;
+       int               spill_cost;     /**< the cost of a single spill node */
+       int               reload_cost;    /**< the cost of a reload node */
+       set              *spills;         /**< all spill_info_t's, which must be
+                                              placed */
+       ir_nodeset_t      mem_phis;       /**< set of all spilled phis. */
+       ir_exec_freq     *exec_freq;
+       unsigned          new_nodes_idx;  /**< all old nodes idx is smaller than
+                                              this */
+
+#ifdef FIRM_STATISTICS
+       unsigned          spill_count;
+       unsigned          reload_count;
+       unsigned          remat_count;
+       unsigned          spilled_phi_count;
+#endif
+};
 
 /**
- * Compare two Phi->Spill associations.
+ * Compare two spill infos.
  */
-static int cmp_phi_spill_assoc(const void *a, const void *b, size_t n) {
-       const phi_spill_assoc_t *p1 = a;
-       const phi_spill_assoc_t *p2 = b;
-       return p1->phi != p2->phi;
-}
+static int cmp_spillinfo(const void *x, const void *y, size_t size)
+{
+       const spill_info_t *xx = x;
+       const spill_info_t *yy = y;
+       (void) size;
 
-/**
- * compare two spill contexts.
- */
-static int cmp_spillctx(const void *a, const void *b, size_t n) {
-       const spill_ctx_t *p = a;
-       const spill_ctx_t *q = b;
-       return p->user != q->user || p->spilled != q->spilled;
+       return xx->to_spill != yy->to_spill;
 }
 
 /**
- * Compare two spill infos.
+ * Returns spill info for a specific value (the value that is to be spilled)
  */
-static int cmp_spillinfo(const void *x, const void *y, size_t size) {
-       const spill_info_t *xx = x;
-       const spill_info_t *yy = y;
-       return xx->spilled_node != yy->spilled_node;
-}
+static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
+{
+       spill_info_t info, *res;
+       int hash = nodeset_hash(value);
+
+       info.to_spill = value;
+       res = set_find(env->spills, &info, sizeof(info), hash);
+
+       if (res == NULL) {
+               info.reloaders   = NULL;
+               info.spills      = NULL;
+               info.spill_costs = -1;
+               info.reload_cls  = NULL;
+               res = set_insert(env->spills, &info, sizeof(info), hash);
+       }
 
-DEBUG_ONLY(
-/* Sets the debug module of a spill environment. */
-void be_set_spill_env_dbg_module(spill_env_t *env, firm_dbg_module_t *dbg) {
-       env->dbg = dbg;
+       return res;
 }
-)
 
-/* Creates a new spill environment. */
-spill_env_t *be_new_spill_env(const be_chordal_env_t *chordal_env) {
+spill_env_t *be_new_spill_env(be_irg_t *birg)
+{
+       const arch_env_t *arch_env = birg->main_env->arch_env;
+
        spill_env_t *env        = xmalloc(sizeof(env[0]));
-       env->spill_ctxs         = new_set(cmp_spillctx, 1024);
        env->spills                     = new_set(cmp_spillinfo, 1024);
-       env->cls                        = chordal_env->cls;
-       env->chordal_env        = chordal_env;
-       env->mem_phis           = pset_new_ptr_default();
-       env->copies                     = NEW_ARR_F(ir_node*, 0);
+       env->irg            = be_get_birg_irg(birg);
+       env->birg           = birg;
+       env->arch_env       = arch_env;
+       ir_nodeset_init(&env->mem_phis);
+       env->spill_cost     = arch_env->isa->spill_cost;
+       env->reload_cost    = arch_env->isa->reload_cost;
+       env->exec_freq      = be_get_birg_exec_freq(birg);
        obstack_init(&env->obst);
+
+#ifdef FIRM_STATISTICS
+       env->spill_count       = 0;
+       env->reload_count      = 0;
+       env->remat_count       = 0;
+       env->spilled_phi_count = 0;
+#endif
+
        return env;
 }
 
-/* Deletes a spill environment. */
-void be_delete_spill_env(spill_env_t *env) {
-       del_set(env->spill_ctxs);
+void be_delete_spill_env(spill_env_t *env)
+{
        del_set(env->spills);
-       del_pset(env->mem_phis);
-       DEL_ARR_F(env->copies);
+       ir_nodeset_destroy(&env->mem_phis);
        obstack_free(&env->obst, NULL);
        free(env);
 }
 
-/**
- * Returns a spill context. If the context did not exists, create one.
+/*
+ *  ____  _                  ____      _                 _
+ * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
+ * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
+ * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
+ * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
  *
- * @param sc        the set containing all spill contexts
- * @param to_spill  the node that should be spilled
- * @param ctx_irn   an user of the spilled node
- *
- * @return a spill context.
  */
-static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
-       spill_ctx_t templ;
 
-       templ.spilled = to_spill;
-       templ.user    = ctx_irn;
-       templ.spill   = NULL;
+void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *before)
+{
+       spill_info_t *spill_info = get_spillinfo(env, to_spill);
+       spill_t      *spill;
+       spill_t      *s;
+       spill_t      *last;
+
+       DB((dbg, LEVEL_1, "Add spill of %+F before %+F\n", to_spill, before));
+
+       /* spills that are dominated by others are not needed */
+       last = NULL;
+       s    = spill_info->spills;
+       for( ; s != NULL; s = s->next) {
+               /* no need to add this spill if it is dominated by another */
+               if(value_dominates(s->before, before)) {
+                       DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->before));
+                       return;
+               }
+               /* remove spills that we dominate */
+               if(value_dominates(before, s->before)) {
+                       DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->before));
+                       if(last != NULL) {
+                               last->next         = s->next;
+                       } else {
+                               spill_info->spills = s->next;
+                       }
+               } else {
+                       last = s;
+               }
+       }
+
+       spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
+       spill->before = before;
+       spill->next   = spill_info->spills;
+       spill->spill  = NULL;
 
-       return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
+       spill_info->spills = spill;
 }
 
-/**
- * Creates a spill.
- *
- * @param senv      the spill environment
- * @param irn       the node that should be spilled
- * @param ctx_irn   an user of the spilled node
- *
- * @return a be_Spill node
- */
-static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
-       spill_ctx_t *ctx;
-       const be_main_env_t *env = senv->chordal_env->birg->main_env;
-       DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
-
-       // Has the value already been spilled?
-       ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
-       if(ctx->spill)
-               return ctx->spill;
-
-       /* Trying to spill an already spilled value, no need for a new spill
-        * node then, we can simply connect to the same one for this reload
-        */
-       if(be_is_Reload(irn)) {
-               return get_irn_n(irn, be_pos_Reload_mem);
+void be_add_remat(spill_env_t *env, ir_node *to_spill, ir_node *before,
+                  ir_node *rematted_node)
+{
+       spill_info_t *spill_info;
+       reloader_t *reloader;
+
+       spill_info = get_spillinfo(env, to_spill);
+
+       /* add the remat information */
+       reloader                   = obstack_alloc(&env->obst, sizeof(reloader[0]));
+       reloader->next             = spill_info->reloaders;
+       reloader->reloader         = before;
+       reloader->rematted_node    = rematted_node;
+       reloader->remat_cost_delta = 0; /* We will never have a cost win over a
+                                          reload since we're not even allowed to
+                                          create a reload */
+
+       spill_info->reloaders  = reloader;
+
+       DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be rematerialized before %+F\n",
+               to_spill, before));
+}
+
+void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
+               ir_node *can_spill_after, const arch_register_class_t *reload_cls,
+               int allow_remat)
+{
+       spill_info_t *info;
+       reloader_t *rel;
+
+       info = get_spillinfo(env, to_spill);
+
+       if (is_Phi(to_spill)) {
+               int i, arity;
+
+               /* create spillinfos for the phi arguments */
+               for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
+                       ir_node *arg = get_irn_n(to_spill, i);
+                       get_spillinfo(env, arg);
+               }
        }
 
-       ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
+       assert(!is_Proj(before) && !be_is_Keep(before));
+
+       /* put reload into list */
+       rel                   = obstack_alloc(&env->obst, sizeof(rel[0]));
+       rel->next             = info->reloaders;
+       rel->reloader         = before;
+       rel->rematted_node    = NULL;
+       rel->can_spill_after  = can_spill_after;
+       rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
+
+       info->reloaders  = rel;
+       assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
+       info->reload_cls = reload_cls;
 
-       return ctx->spill;
+       DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
+               to_spill, before, allow_remat ? "" : " not"));
+}
+
+void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
+                   const arch_register_class_t *reload_cls, int allow_remat)
+{
+       be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
+
+}
+
+ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
+{
+       ir_node *last = sched_last(block);
+
+       /* we might have keeps behind the jump... */
+       while(be_is_Keep(last)) {
+               last = sched_prev(last);
+               assert(!sched_is_end(last));
+       }
+
+       assert(is_cfop(last));
+
+       /* add the reload before the (cond-)jump */
+       return last;
 }
 
 /**
- * Removes all copies introduced for phi-spills
+ * Returns the point at which you can insert a node that should be executed
+ * before block @p block when coming from pred @p pos.
  */
-static void remove_copies(spill_env_t *env) {
-       int i;
+static ir_node *get_block_insertion_point(ir_node *block, int pos)
+{
+       ir_node *predblock;
+
+       /* simply add the reload to the beginning of the block if we only have 1
+        * predecessor. We don't need to check for phis as there can't be any in a
+        * block with only 1 pred. */
+       if(get_Block_n_cfgpreds(block) == 1) {
+               assert(!is_Phi(sched_first(block)));
+               return sched_first(block);
+       }
+
+       /* We have to reload the value in pred-block */
+       predblock = get_Block_cfgpred_block(block, pos);
+       return be_get_end_of_block_insertion_point(predblock);
+}
 
-       for(i = 0; i < ARR_LEN(env->copies); ++i) {
-               ir_node *node = env->copies[i];
-               ir_node *src;
-               const ir_edge_t *edge, *ne;
+void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
+                          const ir_node *block,
+                          const arch_register_class_t *reload_cls,
+                          int allow_remat)
+{
+       ir_node *before = be_get_end_of_block_insertion_point(block);
+       be_add_reload(env, to_spill, before, reload_cls, allow_remat);
+}
 
-               assert(be_is_Copy(node));
+void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
+                           int pos,    const arch_register_class_t *reload_cls,
+                           int allow_remat)
+{
+       ir_node *before = get_block_insertion_point(block, pos);
+       be_add_reload(env, to_spill, before, reload_cls, allow_remat);
+}
 
-               src = be_get_Copy_op(node);
-               foreach_out_edge_safe(node, edge, ne) {
-                       ir_node *user = get_edge_src_irn(edge);
-                       int user_pos  = get_edge_src_pos(edge);
+void be_spill_phi(spill_env_t *env, ir_node *node)
+{
+       ir_node *block;
+       spill_info_t* spill;
+       int i, arity;
 
-                       set_irn_n(user, user_pos, src);
-               }
-       }
+       assert(is_Phi(node));
 
-       ARR_SETLEN(ir_node*, env->copies, 0);
+       ir_nodeset_insert(&env->mem_phis, node);
+
+       /* create spills for the phi arguments */
+       block = get_nodes_block(node);
+       spill = get_spillinfo(env, node);
+       for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
+               ir_node *arg        = get_irn_n(node, i);
+               ir_node *pred_block = get_Block_cfgpred_block(block, i);
+               ir_node *insert     = be_get_end_of_block_insertion_point(pred_block);
+               //get_spillinfo(env, arg);
+               be_add_spill(env, arg, insert);
+       }
 }
 
-/**
- * Inserts a copy (needed for spilled phi handling) of a value at the earliest
- * possible location in a block. That is after the last use/def of the value or at
- * the beginning of the block if there is no use/def.
+/*
+ *   ____                _         ____        _ _ _
+ *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
+ * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
+ * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
+ *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
+ *                                      |_|
  */
-static ir_node *insert_copy(spill_env_t *env, ir_node *block, ir_node *value) {
-       ir_node* node;
-       ir_graph *irg = get_irn_irg(block);
-       ir_node *copy = be_new_Copy(env->cls, irg, block, value);
 
-       ARR_APP1(ir_node*, env->copies, copy);
+static ir_node *skip_keeps_phis(ir_node *node)
+{
+       node = sched_next(node);
+       while(is_Phi(node) || be_is_Keep(node)) {
+               node = sched_next(node);
+       }
+       return node;
+}
 
-       // walk schedule backwards until we find a use/def, or until we have reached the first phi
-       // TODO we could also do this by iterating over all uses and checking the
-       // sched_get_time_step value. Need benchmarks to decide this...
-       sched_foreach_reverse(block, node) {
-               int i, arity;
+static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 
-               if(is_Phi(node)) {
-                       sched_add_after(node, copy);
-                       goto placed;
-               }
-               if(value == node) {
-                       sched_add_after(node, copy);
-                       goto placed;
-               }
-               for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
-                       ir_node *arg = get_irn_n(node, i);
-                       if(arg == value) {
-                               sched_add_after(node, copy);
-                               goto placed;
-                       }
-               }
+/**
+ * Creates a spill.
+ *
+ * @param senv      the spill environment
+ * @param irn       the node that should be spilled
+ * @param ctx_irn   an user of the spilled node
+ *
+ * @return a be_Spill node
+ */
+static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
+{
+       ir_node *to_spill = spillinfo->to_spill;
+       spill_t *spill;
+
+       /* determine_spill_costs must have been run before */
+       assert(spillinfo->spill_costs >= 0);
+
+       /* some backends have virtual noreg/unknown nodes that are not scheduled
+        * and simply always available. */
+       if(!sched_is_scheduled(to_spill)) {
+               /* override spillinfos or create a new one */
+               spillinfo->spills->spill = new_NoMem();
+               DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
+               return;
        }
-       // we didn't find a use or a phi yet, so place the copy at the beginning of the block
-       sched_add_before(sched_first(block), copy);
 
-placed:
+       DBG((dbg, LEVEL_1, "spilling %+F ... ", to_spill));
+       spill = spillinfo->spills;
+       for( ; spill != NULL; spill = spill->next) {
+               ir_node *block  = get_block(spill->before);
+               ir_node *before = spill->before;
 
-       return copy;
+               /* place all spills before the reloads (as we can't guarantee the
+                * same order as the be_add_spill and be_add_reload calls */
+               while(get_irn_idx(sched_prev(before)) > env->new_nodes_idx) {
+                       before = sched_prev(before);
+               }
+
+               spill->spill    = be_spill(env->arch_env, block, to_spill);
+               sched_add_before(before, spill->spill);
+               DB((dbg, LEVEL_1, "\t%+F before %+F,", spill->spill, before));
+#ifdef FIRM_STATISTICS
+               env->spill_count++;
+#endif
+       }
+       DBG((dbg, LEVEL_1, "\n"));
 }
 
+static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
+
 /**
  * If the first usage of a Phi result would be out of memory
  * there is no sense in allocating a register for it.
@@ -251,65 +455,52 @@ placed:
  * @param senv      the spill environment
  * @param phi       the Phi node that should be spilled
  * @param ctx_irn   an user of the spilled node
- *
- * @return a be_Spill node
  */
-static ir_node *spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn, set *already_visited_phis, bitset_t *bs) {
-       int         i;
-       int arity = get_irn_arity(phi);
-       ir_graph    *irg      = senv->chordal_env->irg;
-       ir_node     *bl       = get_nodes_block(phi);
-       ir_node     **ins, *phi_spill;
-       phi_spill_assoc_t key;
-       spill_ctx_t *ctx;
+static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
+{
+       ir_graph *irg   = env->irg;
+       ir_node  *phi   = spillinfo->to_spill;
+       ir_node  *block = get_nodes_block(phi);
+       ir_node  *unknown;
+       ir_node **ins;
+       spill_t  *spill;
+       int       i;
+       int       arity;
 
        assert(is_Phi(phi));
-       DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
+       assert(!get_opt_cse());
+       DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 
        /* build a new PhiM */
-       NEW_ARR_A(ir_node *, ins, arity);
-       for (i = 0; i < arity; ++i) {
-               ins[i] = new_r_Bad(irg);
+       arity   = get_irn_arity(phi);
+       ins     = alloca(sizeof(ir_node*) * arity);
+       unknown = new_r_Unknown(irg, mode_M);
+       for(i = 0; i < arity; ++i) {
+               ins[i] = unknown;
        }
-       phi_spill = new_r_Phi(senv->chordal_env->irg, bl, arity, ins, mode_M);
-       key.phi   = phi;
-       key.spill = phi_spill;
-       set_insert(already_visited_phis, &key, sizeof(key), HASH_PTR(phi));
-       bitset_set(bs, get_irn_idx(phi));
-
-       /* search an existing spill for this context */
-       ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
-
-       /* if not found spill the phi */
-       if (! ctx->spill) {
-               /* collect all arguments of the phi */
-               for (i = 0; i < arity; ++i) {
-                       ir_node *arg = get_irn_n(phi, i);
-                       ir_node *sub_res;
-                       phi_spill_assoc_t *entry;
-
-                       if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg)) {
-                               if (! bitset_is_set(bs, get_irn_idx(arg)))
-                                       sub_res = spill_phi(senv, arg, ctx_irn, already_visited_phis, bs);
-                               else {
-                                       /* we already visited the argument phi: get it's spill */
-                                       key.phi   = arg;
-                                       key.spill = NULL;
-                                       entry     = set_find(already_visited_phis, &key, sizeof(key), HASH_PTR(arg));
-                                       assert(entry && "argument phi already visited, but no spill found?!?");
-                                       sub_res   = entry->spill;
-                                       assert(sub_res && "spill missing?!?");
-                               }
-                       }
-                       else
-                               sub_res = be_spill_irn(senv, arg, ctx_irn);
 
-                       set_irn_n(phi_spill, i, sub_res);
-               }
+       /* override or replace spills list... */
+       spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
+       spill->before = skip_keeps_phis(phi);
+       spill->spill  = new_r_Phi(irg, block, arity, ins, mode_M);
+       spill->next   = NULL;
+
+       spillinfo->spills = spill;
+#ifdef FIRM_STATISTICS
+       env->spilled_phi_count++;
+#endif
+
+       for(i = 0; i < arity; ++i) {
+               ir_node      *arg      = get_irn_n(phi, i);
+               spill_info_t *arg_info = get_spillinfo(env, arg);
 
-               ctx->spill = phi_spill;
+               determine_spill_costs(env, arg_info);
+               spill_node(env, arg_info);
+
+               set_irn_n(spill->spill, i, arg_info->spills->spill);
        }
-       return ctx->spill;
+       DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
+            spill->spill));
 }
 
 /**
@@ -317,110 +508,158 @@ static ir_node *spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn, set
  *
  * @param senv      the spill environment
  * @param to_spill  the node that should be spilled
- *
- * @return a be_Spill node
  */
-static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
-       ir_graph *irg = get_irn_irg(to_spill);
-       ir_node  *res;
-
-       if (pset_find_ptr(senv->mem_phis, to_spill)) {
-               set *already_visited_phis = new_set(cmp_phi_spill_assoc, 10);
-               bitset_t *bs = bitset_alloca(get_irg_last_idx(irg));
-               res = spill_phi(senv, to_spill, to_spill, already_visited_phis, bs);
-               del_set(already_visited_phis);
+static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
+{
+       ir_node *to_spill;
+
+       /* node is already spilled */
+       if(spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
+               return;
+
+       to_spill = spillinfo->to_spill;
+
+       if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
+               spill_phi(env, spillinfo);
        } else {
-               res = be_spill_irn(senv, to_spill, to_spill);
+               spill_irn(env, spillinfo);
        }
-
-       return res;
 }
 
-#ifdef REMAT
-
-#ifdef BUGGY_REMAT
+/*
+ *
+ *  ____                      _            _       _ _
+ * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
+ * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
+ * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
+ * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
+ *
+ */
 
 /**
- * Check if a spilled node could be rematerialized.
- *
- * @param senv      the spill environment
- * @param spill     the Spill node
- * @param spilled   the node that was spilled
- * @param reloader  a irn that requires a reload
+ * Tests whether value @p arg is available before node @p reloader
+ * @returns 1 if value is available, 0 otherwise
  */
-static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
-       int pos, max;
+static int is_value_available(spill_env_t *env, const ir_node *arg,
+                              const ir_node *reloader)
+{
+       if(is_Unknown(arg) || arg == new_NoMem())
+               return 1;
 
-       /* check for 'normal' spill and general remat condition */
-       if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
-               return 0;
+       if(be_is_Spill(arg))
+               return 1;
 
-       /* check availability of original arguments */
-       if (is_Block(reloader)) {
+       if(arg == get_irg_frame(env->irg))
+               return 1;
 
-               /* we want to remat at the end of a block.
-                * thus all arguments must be alive at the end of the block
-                */
-               for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
-                       ir_node *arg = get_irn_n(spilled, pos);
-                       if (!is_live_end(reloader, arg))
-                               return 0;
-               }
+       /* hack for now (happens when command should be inserted at end of block) */
+       if(is_Block(reloader)) {
+               return 0;
+       }
 
-       } else {
+       /*
+        * Ignore registers are always available
+        */
+       if(arch_irn_is(env->arch_env, arg, ignore)) {
+               return 1;
+       }
 
-               /* we want to remat before the insn reloader
-                * thus an arguments is alive if
-                *   - it interferes with the reloaders result
-                * or
-                *   - or it is (last-) used by reloader itself
-                */
-               for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
-                       ir_node *arg = get_irn_n(spilled, pos);
-                       int i, m;
-
-                       if (values_interfere(reloader, arg))
-                               goto is_alive;
-
-                       for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
-                               ir_node *rel_arg = get_irn_n(reloader, i);
-                               if (rel_arg == arg)
-                                       goto is_alive;
-                       }
+       /* the following test does not work while spilling,
+        * because the liveness info is not adapted yet to the effects of the
+        * additional spills/reloads.
+        */
+#if 0
+       /* we want to remat before the insn reloader
+        * thus an arguments is alive if
+        *   - it interferes with the reloaders result
+        *   - or it is (last-) used by reloader itself
+        */
+       if (values_interfere(env->birg->lv, reloader, arg)) {
+               return 1;
+       }
 
-                       /* arg is not alive before reloader */
-                       return 0;
+       arity = get_irn_arity(reloader);
+       for (i = 0; i < arity; ++i) {
+               ir_node *rel_arg = get_irn_n(reloader, i);
+               if (rel_arg == arg)
+                       return 1;
+       }
+#endif
 
-is_alive:      ;
+       return 0;
+}
 
-               }
+/**
+ * Checks whether the node can principally be rematerialized
+ */
+static int is_remat_node(spill_env_t *env, const ir_node *node)
+{
+       const arch_env_t *arch_env = env->arch_env;
 
-       }
+       assert(!be_is_Spill(node));
 
-       return 1;
-}
+       if(arch_irn_is(arch_env, node, rematerializable))
+               return 1;
 
-#else /* BUGGY_REMAT */
+       return 0;
+}
 
 /**
- * A very simple rematerialization checker.
+ * Check if a node is rematerializable. This tests for the following conditions:
  *
- * @param senv      the spill environment
- * @param spill     the Spill node
- * @param spilled   the node that was spilled
- * @param reloader  a irn that requires a reload
+ * - The node itself is rematerializable
+ * - All arguments of the node are available or also rematerialisable
+ * - The costs for the rematerialisation operation is less or equal a limit
+ *
+ * Returns the costs needed for rematerialisation or something
+ * >= REMAT_COST_INFINITE if remat is not possible.
  */
-static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
-       const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
+static int check_remat_conditions_costs(spill_env_t *env,
+               const ir_node *spilled, const ir_node *reloader, int parentcosts)
+{
+       int i, arity;
+       int argremats;
+       int costs = 0;
+
+       if(!is_remat_node(env, spilled))
+               return REMAT_COST_INFINITE;
+
+       if(be_is_Reload(spilled)) {
+               costs += 2;
+       } else {
+               costs += arch_get_op_estimated_cost(env->arch_env, spilled);
+       }
+       if(parentcosts + costs >= env->reload_cost + env->spill_cost) {
+               return REMAT_COST_INFINITE;
+       }
+       if(arch_irn_is(env->arch_env, spilled, modify_flags)) {
+               return REMAT_COST_INFINITE;
+       }
 
-       return get_irn_arity(spilled) == 0 &&
-                  be_is_Spill(spill) &&
-                  arch_irn_is(aenv, spilled, rematerializable);
-}
+       argremats = 0;
+       for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
+               ir_node *arg = get_irn_n(spilled, i);
+
+               if(is_value_available(env, arg, reloader))
+                       continue;
+
+               /* we have to rematerialize the argument as well */
+               if(argremats >= 1) {
+                       /* we only support rematerializing 1 argument at the moment,
+                        * so that we don't have to care about register pressure
+                        */
+                       return REMAT_COST_INFINITE;
+               }
+               argremats++;
 
-#endif /* BUGGY_REMAT */
+               costs += check_remat_conditions_costs(env, arg, reloader,
+                                                     parentcosts + costs);
+               if(parentcosts + costs >= env->reload_cost + env->spill_cost)
+                       return REMAT_COST_INFINITE;
+       }
 
-#endif /* REMAT */
+       return costs;
+}
 
 /**
  * Re-materialize a node.
@@ -429,406 +668,373 @@ static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *sp
  * @param spilled   the node that was spilled
  * @param reloader  a irn that requires a reload
  */
-static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
+static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
+{
+       int i, arity;
        ir_node *res;
-       ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
-
-       /* recompute the value */
-       res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
-               get_irn_op(spilled),
-               get_irn_mode(spilled),
-               get_irn_arity(spilled),
-               get_irn_in(spilled) + 1);
+       ir_node *bl;
+       ir_node **ins;
+
+       if(is_Block(reloader)) {
+               bl = reloader;
+       } else {
+               bl = get_nodes_block(reloader);
+       }
+
+       ins = alloca(get_irn_arity(spilled) * sizeof(ins[0]));
+       for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
+               ir_node *arg = get_irn_n(spilled, i);
+
+               if(is_value_available(env, arg, reloader)) {
+                       ins[i] = arg;
+               } else {
+                       ins[i] = do_remat(env, arg, reloader);
+#ifdef FIRM_STATISTICS
+                       /* don't count the recursive call as remat */
+                       env->remat_count--;
+#endif
+               }
+       }
+
+       /* create a copy of the node */
+       res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
+                         get_irn_op(spilled), get_irn_mode(spilled),
+                         get_irn_arity(spilled), ins);
        copy_node_attr(spilled, res);
+       new_backedge_info(res);
 
-       DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
+       DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 
-       /* insert in schedule */
-       if (is_Block(reloader)) {
-               ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
-               sched_add_after(insert, res);
-       } else {
+       if (! is_Proj(res)) {
+               /* insert in schedule */
+               sched_reset(res);
                sched_add_before(reloader, res);
+#ifdef FIRM_STATISTICS
+               env->remat_count++;
+#endif
        }
 
        return res;
 }
 
-void be_spill_phi(spill_env_t *env, ir_node *node) {
-       assert(is_Phi(node));
+double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
+{
+       ir_node *block = get_nodes_block(before);
+       double   freq  = get_block_execfreq(env->exec_freq, block);
+       (void) to_spill;
 
-       pset_insert_ptr(env->mem_phis, node);
+       return env->spill_cost * freq;
 }
 
-void be_insert_spills_reloads(spill_env_t *env) {
-       const arch_env_t *arch_env = env->chordal_env->birg->main_env->arch_env;
-       ir_node *node;
-       spill_info_t *si;
-
-       DBG((env->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
-       foreach_pset(env->mem_phis, node) {
-               const ir_edge_t *e;
-               int i, arity;
-
-               /* We have to place copy nodes in the predecessor blocks to temporarily
-               * produce new values that get separate spill slots
-               */
-               for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
-                       ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(node), i);
-                       ir_node *arg = get_irn_n(node, i);
-                       ir_node* copy = insert_copy(env, pred_block, arg);
-
-                       set_irn_n(node, i, copy);
-               }
+double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
+{
+       ir_node      *block = get_nodes_block(before);
+       double        freq  = get_block_execfreq(env->exec_freq, block);
 
-               /* Add reloads for mem_phis */
-               /* BETTER: These reloads (1) should only be inserted, if they are really needed */
-               DBG((env->dbg, LEVEL_1, " Mem-phi %+F\n", node));
-               foreach_out_edge(node, e) {
-                       ir_node *user = e->src;
-                       if (is_Phi(user) && !pset_find_ptr(env->mem_phis, user)) {
-                               ir_node *use_bl = get_nodes_block(user);
-                               DBG((env->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
-                               be_add_reload_on_edge(env, node, use_bl, e->pos); /* (1) */
-                       }
-               }
+       if(be_do_remats) {
+               /* is the node rematerializable? */
+               int costs = check_remat_conditions_costs(env, to_spill, before, 0);
+               if(costs < env->reload_cost)
+                       return costs * freq;
        }
 
-       /* process each spilled node */
-       DBG((env->dbg, LEVEL_1, "Insert spills and reloads:\n"));
-       for(si = set_first(env->spills); si; si = set_next(env->spills)) {
-               reloader_t *rld;
-               ir_mode *mode = get_irn_mode(si->spilled_node);
-               //ir_node *value;
-               pset *values = pset_new_ptr(16);
+       return env->reload_cost * freq;
+}
 
-               /* go through all reloads for this spill */
-               for(rld = si->reloaders; rld; rld = rld->next) {
-                       ir_node *new_val;
+int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
+                           const ir_node *before)
+{
+       return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
+}
 
-                       /* the spill for this reloader */
-                       ir_node *spill   = be_spill_node(env, si->spilled_node);
+double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
+                                   ir_node *block, int pos)
+{
+       ir_node *before = get_block_insertion_point(block, pos);
+       return be_get_reload_costs(env, to_spill, before);
+}
 
-#ifdef REMAT
-                       if (check_remat_conditions(env, spill, si->spilled_node, rld->reloader)) {
-                               new_val = do_remat(env, si->spilled_node, rld->reloader);
-                               //pdeq_putl(possibly_dead, spill);
-                       }
-                       else
-#endif
-                               /* do a reload */
-                               new_val = be_reload(arch_env, env->cls, rld->reloader, mode, spill);
+/*
+ *  ___                     _     ____      _                 _
+ * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
+ *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
+ *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
+ * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
+ *
+ */
 
-                       DBG((env->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
-                       pset_insert_ptr(values, new_val);
-               }
+/**
+ * analyzes how to best spill a node and determine costs for that
+ */
+static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
+{
+       ir_node *to_spill = spillinfo->to_spill;
+       ir_node *spill_block;
+       spill_t *spill;
+       double   spill_execfreq;
+
+       /* already calculated? */
+       if(spillinfo->spill_costs >= 0)
+               return;
 
-               /* introduce copies, rewire the uses */
-               assert(pset_count(values) > 0 && "???");
-               pset_insert_ptr(values, si->spilled_node);
-               be_ssa_constr_set_ignore(env->chordal_env->dom_front, values, env->mem_phis);
+       assert(! arch_irn_is(env->arch_env, to_spill, dont_spill));
+       assert(!be_is_Reload(to_spill));
 
-               del_pset(values);
-       }
+       /* some backends have virtual noreg/unknown nodes that are not scheduled
+        * and simply always available.
+        * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
+        * predecessor (of a PhiM) but this test might match other things too...
+        */
+       if(!sched_is_scheduled(to_spill)) {
+               /* override spillinfos or create a new one */
+               spill_t *spill = obstack_alloc(&env->obst, sizeof(spill[0]));
+               spill->before  = NULL;
+               spill->next    = NULL;
+               spill->spill   = new_NoMem();
 
-       remove_copies(env);
+               spillinfo->spills      = spill;
+               spillinfo->spill_costs = 0;
 
-       // reloads are placed now, but we might reuse the spill environment for further spilling decisions
-       del_set(env->spills);
-       env->spills = new_set(cmp_spillinfo, 1024);
-}
+               DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
+               return;
+       }
 
-void be_add_reload(spill_env_t *env, ir_node *to_spill, ir_node *before) {
-       spill_info_t templ, *res;
-       reloader_t *rel;
+       spill_block    = get_nodes_block(to_spill);
+       spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 
-       assert(arch_irn_consider_in_reg_alloc(env->chordal_env->birg->main_env->arch_env, env->cls, to_spill));
+       if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
+               /* TODO calculate correct costs...
+                * (though we can't remat this node anyway so no big problem) */
+               spillinfo->spill_costs = env->spill_cost * spill_execfreq;
+               return;
+       }
 
-       templ.spilled_node = to_spill;
-       templ.reloaders    = NULL;
-       res = set_insert(env->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
+       if(spillinfo->spills != NULL) {
+               spill_t *s;
+               double   spills_execfreq;
 
-       rel           = obstack_alloc(&env->obst, sizeof(rel[0]));
-       rel->reloader = before;
-       rel->next     = res->reloaders;
-       res->reloaders = rel;
-}
+               /* calculate sum of executaion frequencies of individual spills */
+               spills_execfreq = 0;
+               s               = spillinfo->spills;
+               for( ; s != NULL; s = s->next) {
+                       ir_node *spill_block = s->before;
+                       double   freq;
 
-void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *bl, int pos) {
-       ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
-       be_add_reload(env, to_spill, insert_bl);
-}
+                       if(!is_Block(spill_block)) {
+                               spill_block = get_nodes_block(spill_block);
+                       }
+                       freq = get_block_execfreq(env->exec_freq, spill_block);
 
+                       spills_execfreq += freq;
+               }
 
+               DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
+                   spills_execfreq * env->spill_cost,
+                   spill_execfreq * env->spill_cost));
 
-/****************************************
+               /* multi-/latespill is advantageous -> return*/
+               if(spills_execfreq < spill_execfreq) {
+                       DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
+                       spillinfo->spill_costs = spills_execfreq * env->spill_cost;
+                       return;
+               }
+       }
 
-       SPILL SLOT MANAGEMENT AND OPTS
+       /* override spillinfos or create a new one */
+       spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
+       spill->before = skip_keeps_phis(to_spill);
+       spill->next   = NULL;
+       spill->spill  = NULL;
 
-****************************************/
+       spillinfo->spills      = spill;
+       spillinfo->spill_costs = spill_execfreq * env->spill_cost;
+       DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
+}
 
-typedef struct _spill_slot_t {
-       unsigned size;
-       unsigned align;
-       pset *members;
-       ir_mode *largest_mode;  /* the mode of all members with largest size */
-} spill_slot_t;
+void be_insert_spills_reloads(spill_env_t *env)
+{
+       ir_graph              *irg       = env->irg;
+       const arch_env_t      *arch_env  = env->arch_env;
+       const ir_exec_freq    *exec_freq = env->exec_freq;
+       spill_info_t          *si;
+       ir_nodeset_iterator_t  iter;
+       ir_node               *node;
+
+       env->new_nodes_idx = get_irg_last_idx(irg);
+
+       /* create all phi-ms first, this is needed so, that phis, hanging on
+          spilled phis work correctly */
+       foreach_ir_nodeset(&env->mem_phis, node, iter) {
+               spill_info_t *info = get_spillinfo(env, node);
+               spill_node(env, info);
+       }
 
-typedef struct _ss_env_t {
-       struct obstack ob;
-       be_chordal_env_t *cenv;
-       pmap *slots;            /* maps spill_contexts to spill_slots */
-       pmap *types;    /* maps modes to types */
-       DEBUG_ONLY(firm_dbg_module_t *dbg;)
-} ss_env_t;
+       /* process each spilled node */
+       for (si = set_first(env->spills); si; si = set_next(env->spills)) {
+               reloader_t *rld;
+               ir_node  *to_spill        = si->to_spill;
+               ir_mode  *mode            = get_irn_mode(to_spill);
+               ir_node **copies          = NEW_ARR_F(ir_node*, 0);
+               double    all_remat_costs = 0; /** costs when we would remat all nodes */
+               int       force_remat     = 0;
+
+               DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
+
+               determine_spill_costs(env, si);
+
+               /* determine possibility of rematerialisations */
+               if(be_do_remats) {
+                       /* calculate cost savings for each indivial value when it would
+                          be rematted instead of reloaded */
+                       for (rld = si->reloaders; rld != NULL; rld = rld->next) {
+                               double   freq;
+                               int      remat_cost;
+                               int      remat_cost_delta;
+                               ir_node *block;
+                               ir_node *reloader = rld->reloader;
+
+                               if(rld->rematted_node != NULL) {
+                                       DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
+                                            rld->rematted_node, reloader));
+                                       continue;
+                               }
+                               if(rld->remat_cost_delta >= REMAT_COST_INFINITE) {
+                                       DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
+                                            reloader));
+                                       all_remat_costs = REMAT_COST_INFINITE;
+                                       continue;
+                               }
 
+                               remat_cost  = check_remat_conditions_costs(env, to_spill,
+                                                                          reloader, 0);
+                               if(remat_cost >= REMAT_COST_INFINITE) {
+                                       DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
+                                            reloader));
+                                       rld->remat_cost_delta = REMAT_COST_INFINITE;
+                                       all_remat_costs       = REMAT_COST_INFINITE;
+                                       continue;
+                               }
 
-/**
- * Walker: compute the spill slots
- */
-static void compute_spill_slots_walker(ir_node *spill, void *env) {
-       ss_env_t *ssenv = env;
-       ir_node *ctx;
-       pmap_entry *entry;
-       spill_slot_t *ss;
+                               remat_cost_delta      = remat_cost - env->reload_cost;
+                               rld->remat_cost_delta = remat_cost_delta;
+                               block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
+                               freq                  = get_block_execfreq(exec_freq, block);
+                               all_remat_costs      += remat_cost_delta * freq;
+                               DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
+                                    "%d (rel %f)\n", reloader, remat_cost_delta,
+                                    remat_cost_delta * freq));
+                       }
+                       if(all_remat_costs < REMAT_COST_INFINITE) {
+                               /* we don't need the costs for the spill if we can remat
+                                  all reloaders */
+                               all_remat_costs -= si->spill_costs;
 
-       if (!be_is_Spill(spill))
-               return;
+                               DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
+                                    env->spill_cost, si->spill_costs));
+                       }
 
-       /* check, if this spill is for a context already known */
-       ctx = be_get_Spill_context(spill);
-       entry = pmap_find(ssenv->slots, ctx);
-
-       if (!entry) {
-               struct _arch_env_t *arch_env     = ssenv->cenv->birg->main_env->arch_env;
-               const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
-               ir_mode *largest_mode            = arch_register_class_mode(cls);
-
-               /* this is a new spill context */
-               ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
-               ss->members      = pset_new_ptr(8);
-               ss->largest_mode = largest_mode;
-               ss->size         = get_mode_size_bytes(ss->largest_mode);
-               ss->align        = arch_isa_get_reg_class_alignment(arch_env->isa, cls);
-               pmap_insert(ssenv->slots, ctx, ss);
-       } else {
-               /* values with the same spill_ctx must go into the same spill slot */
-               ss = entry->value;
-
-#ifndef NDEBUG
-               /* ugly mega assert :-) */
-               {
-                       ir_node *irn;
-                       struct _arch_env_t *arch_env     = ssenv->cenv->birg->main_env->arch_env;
-                       const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
-                       int size = get_mode_size_bytes(arch_register_class_mode(cls));
-                       assert((int) ss->size == size && "Different sizes for the same spill slot are not allowed.");
-                       for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
-                               /* use values_interfere here, because it uses the dominance check,
-                                        which does work for values in memory */
-                               assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
+                       if(all_remat_costs < 0) {
+                               DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
+                                    all_remat_costs));
+                               force_remat = 1;
                        }
                }
-#endif /* NDEBUG */
-       }
-
-       pset_insert_ptr(ss->members, spill);
-}
-
-/**
- * qsort compare function, sort spill slots by size.
- */
-static int ss_sorter(const void *v1, const void *v2) {
-       const spill_slot_t **ss1 = (const spill_slot_t **)v1;
-       const spill_slot_t **ss2 = (const spill_slot_t **)v2;
-       return ((int) (*ss2)->size) - ((int) (*ss1)->size);
-}
 
+               /* go through all reloads for this spill */
+               for (rld = si->reloaders; rld != NULL; rld = rld->next) {
+                       ir_node *copy; /* a reload is a "copy" of the original value */
+
+                       if (rld->rematted_node != NULL) {
+                               copy = rld->rematted_node;
+                               sched_add_before(rld->reloader, copy);
+                       } else if (be_do_remats &&
+                                       (force_remat || rld->remat_cost_delta < 0)) {
+                               copy = do_remat(env, to_spill, rld->reloader);
+                       } else {
+                               /* make sure we have a spill */
+                               spill_node(env, si);
+
+                               /* create a reload, use the first spill for now SSA
+                                * reconstruction for memory comes below */
+                               assert(si->spills != NULL);
+                               copy = be_reload(arch_env, si->reload_cls, rld->reloader, mode,
+                                                si->spills->spill);
+#ifdef FIRM_STATISTICS
+                               env->reload_count++;
+#endif
+                       }
 
-/**
- * This function should optimize the spill slots.
- *  - Coalescing of multiple slots
- *  - Ordering the slots
- *
- * Input slots are in @p ssenv->slots
- * @p size The count of initial spill slots in @p ssenv->slots
- *         This also is the size of the preallocated array @p ass
- *
- * @return An array of spill slots @p ass in specific order
- **/
-static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t *ass[]) {
-       int i, o, used_slots;
-       pmap_entry *entr;
-
-       i=0;
-       pmap_foreach(ssenv->slots, entr)
-               ass[i++] = entr->value;
-
-       /* Sort the array to minimize fragmentation and cache footprint.
-          Large slots come first */
-       qsort(ass, size, sizeof(ass[0]), ss_sorter);
-
-       /* For each spill slot:
-               - assign a new offset to this slot
-           - xor find another slot to coalesce with */
-       used_slots = 0;
-       for (i=0; i<size; ++i) { /* for each spill slot */
-               ir_node *n1;
-               int tgt_slot = -1;
-
-               DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
-               for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
-                       DBG((ssenv->dbg, LEVEL_1, "  %+F\n", n1));
-
-
-               for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
-                       /* check inter-slot-pairs for interference */
-                       ir_node *n2;
-                       for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
-                               for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
-                                       if(values_interfere(n1, n2)) {
-                                               pset_break(ass[i]->members);
-                                               pset_break(ass[o]->members);
-                                               DBG((ssenv->dbg, LEVEL_1, "    Interf %+F -- %+F\n", n1, n2));
-                                               goto interf_detected;
-                                       }
-
-                       /* if we are here, there is no interference between ass[i] and ass[o] */
-                       tgt_slot = o;
-
-interf_detected: /*nothing*/ ;
+                       DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
+                            copy, to_spill, rld->reloader));
+                       ARR_APP1(ir_node*, copies, copy);
                }
 
-               /* now the members of ass[i] join the members of ass[tgt_slot] */
-
-               /* do we need a new slot? */
-               if (tgt_slot == -1) {
-                       tgt_slot = used_slots;
-                       used_slots++;
-
-                       /* init slot */
-                       if (tgt_slot != i) {
-                               ass[tgt_slot]->size = ass[i]->size;
-                               del_pset(ass[tgt_slot]->members);
-                               ass[tgt_slot]->members = pset_new_ptr(8);
+               /* if we had any reloads or remats, then we need to reconstruct the
+                * SSA form for the spilled value */
+               if (ARR_LEN(copies) > 0) {
+                       be_ssa_construction_env_t senv;
+                       /* be_lv_t *lv = be_get_birg_liveness(env->birg); */
+
+                       be_ssa_construction_init(&senv, env->birg);
+                       be_ssa_construction_add_copy(&senv, to_spill);
+                       be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
+                       be_ssa_construction_fix_users(&senv, to_spill);
+
+#if 0
+                       /* no need to enable this as long as we invalidate liveness
+                          after this function... */
+                       be_ssa_construction_update_liveness_phis(&senv);
+                       be_liveness_update(to_spill);
+                       len = ARR_LEN(copies);
+                       for(i = 0; i < len; ++i) {
+                               be_liveness_update(lv, copies[i]);
                        }
+#endif
+                       be_ssa_construction_destroy(&senv);
                }
+               /* need to reconstruct SSA form if we had multiple spills */
+               if (si->spills != NULL && si->spills->next != NULL) {
+                       spill_t *spill;
+                       int      spill_count = 0;
+
+                       be_ssa_construction_env_t senv;
+
+                       be_ssa_construction_init(&senv, env->birg);
+                       spill = si->spills;
+                       for( ; spill != NULL; spill = spill->next) {
+                               /* maybe we rematerialized the value and need no spill */
+                               if(spill->spill == NULL)
+                                       continue;
+                               be_ssa_construction_add_copy(&senv, spill->spill);
+                               spill_count++;
+                       }
+                       if(spill_count > 1) {
+                               /* all reloads are attached to the first spill, fix them now */
+                               be_ssa_construction_fix_users(&senv, si->spills->spill);
+                       }
 
-               /* copy the members to the target pset */
-               /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
-               if (tgt_slot != i)
-                       for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
-                                       pset_insert_ptr(ass[tgt_slot]->members, n1);
-       }
-}
-
-#define ALIGN_SPILL_AREA 16
-#define pset_foreach(pset, elm)  for(elm=pset_first(pset); elm; elm=pset_next(pset))
+                       be_ssa_construction_destroy(&senv);
+               }
 
-/**
- * Returns a spill type for a mode. Keep them in a map to reduce
- * the number of types.
- *
- * @param types  a map containing all created types
- * @param ss     the spill slot
- *
- * Note that type types should are identical for every mode.
- * This rule might break if two different register classes return the same
- * mode but different alignments.
- */
-static ir_type *get_spill_type(pmap *types, spill_slot_t *ss) {
-  pmap_entry *e = pmap_find(types, ss->largest_mode);
-  ir_type *res;
-
-  if (! e) {
-               char buf[64];
-    snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(ss->largest_mode));
-    res = new_type_primitive(new_id_from_str(buf), ss->largest_mode);
-               set_type_alignment_bytes(res, ss->align);
-    pmap_insert(types, ss->largest_mode, res);
-  }
-  else {
-    res = e->value;
-               assert(get_type_alignment_bytes(res) == (int)ss->align);
+               DEL_ARR_F(copies);
+               si->reloaders = NULL;
        }
-  return res;
-}
 
-/**
- * Create spill slot entities on the frame type.
- *
- * @param ssenv   the spill environment
- * @param n       number of spill slots
- * @param ss      array of spill slots
- */
-static void assign_entities(ss_env_t *ssenv, int n_slots, spill_slot_t *ss[]) {
-       int i, offset, frame_align;
-       ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
-
-       /* aligning by increasing frame size */
-       offset = get_type_size_bits(frame) / 8;
-       offset = round_up2(offset, ALIGN_SPILL_AREA);
-       set_type_size_bytes(frame, -1);
-
-       /* create entities and assign offsets according to size and alignment*/
-       for (i = 0; i < n_slots; ++i) {
-               char buf[64];
-               ident *name;
-               entity *spill_ent;
-               ir_node *irn;
-
-               /* build entity */
-               snprintf(buf, sizeof(buf), "spill_slot_%d", i);
-               name = new_id_from_str(buf);
-
-               spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]));
-
-               /* align */
-               offset = round_up2(offset, ss[i]->align);
-               /* set */
-               set_entity_offset_bytes(spill_ent, offset);
-               /* next possible offset */
-               offset += round_up2(ss[i]->size, ss[i]->align);
-
-               pset_foreach(ss[i]->members, irn)
-                       be_set_Spill_entity(irn, spill_ent);
-       }
+       stat_ev_dbl("spill_spills", env->spill_count);
+       stat_ev_dbl("spill_reloads", env->reload_count);
+       stat_ev_dbl("spill_remats", env->remat_count);
+       stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
 
+       /* Matze: In theory be_ssa_construction should take care of the liveness...
+        * try to disable this again in the future */
+       be_liveness_invalidate(env->birg->lv);
 
-       /* set final size of stack frame */
-       frame_align = get_type_alignment_bytes(frame);
-       set_type_size_bytes(frame, round_up2(offset, frame_align));
+       be_remove_dead_nodes_from_schedule(env->birg);
 }
 
-void be_compute_spill_offsets(be_chordal_env_t *cenv) {
-       ss_env_t ssenv;
-       spill_slot_t **ss;
-       int ss_size;
-       pmap_entry *pme;
-
-       obstack_init(&ssenv.ob);
-       ssenv.cenv  = cenv;
-       ssenv.slots = pmap_create();
-       ssenv.types = pmap_create();
-       FIRM_DBG_REGISTER(ssenv.dbg, "ir.be.spillslots");
-
-       /* Get initial spill slots */
-       irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
-
-       /* Build an empty array for optimized spill slots */
-       ss_size = pmap_count(ssenv.slots);
-       ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
-       optimize_slots(&ssenv, ss_size, ss);
-
-       /* Integrate slots into the stack frame entity */
-       assign_entities(&ssenv, ss_size, ss);
-
-       /* Clean up */
-       pmap_foreach(ssenv.slots, pme)
-       del_pset(((spill_slot_t *)pme->value)->members);
-       pmap_destroy(ssenv.slots);
-       pmap_destroy(ssenv.types);
-       obstack_free(&ssenv.ob, NULL);
-
-       be_copy_entities_to_reloads(cenv->irg);
+void be_init_spill(void)
+{
+       FIRM_DBG_REGISTER(dbg, "firm.be.spill");
 }
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill);