- belady spiller places its copy nodes smarter now
[libfirm] / ir / be / bespillbelady.c
index 0bf4d62..cd9d4b0 100644 (file)
@@ -3,18 +3,31 @@
  * Date:               20.09.2005
  * Copyright:   (c) Universitaet Karlsruhe
  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+ *
  */
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
 
+#ifdef HAVE_ALLOCA_H
 #include <alloca.h>
+#endif
+
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
+
 #include "obst.h"
 #include "set.h"
 #include "pset.h"
+#include "irprintf_t.h"
 #include "irgraph.h"
 #include "irnode.h"
 #include "irmode.h"
-#include "ircons.h"
 #include "irgwalk.h"
 #include "iredges_t.h"
+#include "ircons_t.h"
+#include "irprintf.h"
 
 #include "beutil.h"
 #include "bearch.h"
 #include "beirgmod.h"
 #include "belive_t.h"
 #include "benode_t.h"
-
-#define MIN(a,b) (((a)<(b))?(a):(b))
-
-#define DBG_DECIDE 1
-#define DBG_WSETS 2
-#define DBG_FIX 4
-#define DBG_SPILL 8
-#define DBG_TRACE 32
-#define DEBUG_LVL SET_LEVEL_0 //(DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
-static firm_dbg_module_t *dbg = NULL;
+#include "bechordal_t.h"
+
+#define DBG_SPILL   1
+#define DBG_WSETS   2
+#define DBG_FIX     4
+#define DBG_DECIDE  8
+#define DBG_START  16
+#define DBG_SLOTS  32
+#define DBG_TRACE  64
+#define DBG_WORKSET 128
+#define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
 typedef struct _workset_t workset_t;
-typedef struct _block_info_t block_info_t;
-typedef struct _reloader_t reloader_t;
-typedef struct _spill_info_t spill_info_t;
-typedef struct _belady_env_t belady_env_t;
 
-struct _workset_t {
-       belady_env_t *bel;
-       int i;                          /**< used for iteration */
-       int len;                        /**< current length */
-       loc_t vals[1];          /**< inlined array of the values/distances in this working set */
-};
-
-struct _block_info_t {
-       workset_t *ws_start, *ws_end;
-};
-
-struct _reloader_t {
-       reloader_t *next;
-       ir_node *reloader;
-};
-
-struct _spill_info_t {
-       ir_node *spilled_node;
-       reloader_t *reloaders;
-};
-
-struct _belady_env_t {
+typedef struct _belady_env_t {
        struct obstack ob;
-       const be_node_factory_t *factory;
        const arch_env_t *arch;
        const arch_register_class_t *cls;
        int n_regs;                     /** number of regs in this reg-class */
@@ -74,18 +63,25 @@ struct _belady_env_t {
        ir_node *instr;         /**< current instruction */
        unsigned instr_nr;      /**< current instruction number (relative to block start) */
        pset *used;                     /**< holds the values used (so far) in the current BB */
-       set *spills;            /**< all spill_info_t's, which must be placed */
+       ir_node **copies;       /**< holds all copies placed due to phi-spilling */
+
+       spill_env_t *senv;      /**< see bespill.h */
+} belady_env_t;
 
-       spill_env_t senv;       /* see bespill.h */
-       pset *reloads;          /**< all reload nodes placed */
+struct _workset_t {
+       belady_env_t *bel;
+       int len;                        /**< current length */
+       loc_t vals[1];          /**< inlined array of the values/distances in this working set */
 };
 
-static int set_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);
-}
+void workset_print(const workset_t *w)
+{
+       int i;
 
+       for(i = 0; i < w->len; ++i) {
+               ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time);
+       }
+}
 
 /**
  * Alloc a new workset on obstack @p ob with maximum size @p max
@@ -99,6 +95,9 @@ static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
        return res;
 }
 
+/**
+ * Alloc a new instance on obstack and make it equal to @param ws
+ */
 static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
        workset_t *res;
        size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
@@ -107,16 +106,33 @@ static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
        return res;
 }
 
+/**
+ * Do NOT alloc anything. Make @param tgt equal to @param src.
+ * returns @param tgt for convinience
+ */
+static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
+       size_t size = sizeof(*src) + (src->bel->n_regs-1)*sizeof(src->vals[0]);
+       memcpy(tgt, src, size);
+       return tgt;
+}
+
+/**
+ * Overwrites the current content array of @param ws with the
+ * @param count locations given at memory @param locs.
+ * Set the length of @param ws to count.
+ */
+#define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0]));
+
+
 /**
  * Inserts the value @p val into the workset, iff it is not
  * already contained. The workset must not be full.
  */
 static INLINE void workset_insert(workset_t *ws, ir_node *val) {
        int i;
-       assert(ws->len < ws->bel->n_regs && "Workset already full!");
        /* check for current regclass */
-       if (arch_get_irn_reg_class(ws->bel->arch, val, 0) != ws->bel->cls) {
-               DBG((dbg, 0, "Dropped %+F\n", val));
+       if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
+               DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val));
                return;
        }
 
@@ -126,6 +142,7 @@ static INLINE void workset_insert(workset_t *ws, ir_node *val) {
                        return;
 
        /* insert val */
+       assert(ws->len < ws->bel->n_regs && "Workset already full!");
        ws->vals[ws->len++].irn = val;
 }
 
@@ -136,13 +153,12 @@ static INLINE void workset_insert(workset_t *ws, ir_node *val) {
  */
 static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
        int i, o;
-       assert(ws->len + cnt <= ws->bel->n_regs && "Workset does not have enough room!");
 
        for(o=0; o<cnt; ++o) {
                ir_node *val = vals[o];
                DBG((dbg, DBG_TRACE, "Bulk insert %+F\n", val));
                /* check for current regclass */
-               if (arch_get_irn_reg_class(ws->bel->arch, val, 0) != ws->bel->cls) {
+               if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
                        DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
                        goto no_insert;
                }
@@ -155,6 +171,7 @@ static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
                        }
 
                /* insert val */
+               assert(ws->len < ws->bel->n_regs && "Workset does not have enough room!");
                ws->vals[ws->len++].irn = val;
                DBG((dbg, DBG_TRACE, "Inserted\n"));
 
@@ -163,12 +180,6 @@ no_insert:
        }
 }
 
-/**
- * Overwrites the current contents of @p ws with the
- * locations given in @p locs
- */
-#define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0]));
-
 /**
  * Removes all entries from this workset
  */
@@ -186,7 +197,7 @@ static INLINE void workset_remove(workset_t *ws, ir_node *val) {
                }
 }
 
-static INLINE int workset_contains(const workset_t *ws, ir_node *val) {
+static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
        int i;
        for(i=0; i<ws->len; ++i)
                if (ws->vals[i].irn == val)
@@ -194,9 +205,15 @@ static INLINE int workset_contains(const workset_t *ws, ir_node *val) {
        return 0;
 }
 
-#define workset_foreach(ws, v) for(ws->i=0; \
-                                                                       v=(ws->i < ws->len) ? ws->vals[ws->i].irn : NULL, ws->i < ws->len; \
-                                                                       ws->i++)
+/**
+ * Iterates over all values in the working set.
+ * @p ws The workset to iterate
+ * @p v  A variable to put the current value in
+ * @p i  An integer for internal use
+ */
+#define workset_foreach(ws, v, i)      for(i=0; \
+                                                                               v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
+                                                                               ++i)
 
 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 #define workset_set_length(ws, length) (ws)->len = length
@@ -204,54 +221,50 @@ static INLINE int workset_contains(const workset_t *ws, ir_node *val) {
 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
 
+typedef struct _block_info_t {
+       workset_t *ws_start, *ws_end;
+} block_info_t;
 
-/**
- * Collects all values live-in at block @p blk and all phi results in this block.
- * Then it adds the best values (at most n_regs) to the ws.
- */
-static void build_start_set(belady_env_t *bel, ir_node *blk) {
-       workset_t *ws = bel->ws;
-       ir_node *irn, *first;
-       irn_live_t *li;
-       int count;
-       loc_t loc, *starters;
-       struct obstack ob;
 
-       obstack_init(&ob);
+static INLINE void *new_block_info(struct obstack *ob) {
+       block_info_t *res = obstack_alloc(ob, sizeof(*res));
+       res->ws_start = NULL;
+       res->ws_end = NULL;
 
-       /* get all values */
-       first = sched_first(blk);
-       count = 0;
-       sched_foreach(blk, irn)
-               if (is_Phi(irn) && arch_get_irn_reg_class(bel->arch, irn, 0) == bel->cls) {
-                       loc.irn = irn;
-                       loc.time = be_get_next_use(bel->uses, first, 0, irn, 1);
-                       obstack_grow(&ob, &loc, sizeof(loc));
-                       count++;
-               } else
-                       break;
+       return res;
+}
 
-       live_foreach(blk, li)
-               if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, 0) == bel->cls) {
-                       loc.irn = (ir_node *)li->irn;
-                       loc.time = be_get_next_use(bel->uses, first, 0, li->irn, 1);
-                       obstack_grow(&ob, &loc, sizeof(loc));
-                       count++;
-               }
-       starters = obstack_finish(&ob);
+#define get_block_info(blk)                    ((block_info_t *)get_irn_link(blk))
+#define set_block_info(blk, info)      set_irn_link(blk, info)
 
-       /* sort all values */
-       qsort(starters, count, sizeof(starters[0]), loc_compare);
+static int is_mem_phi(const ir_node *irn, void *data) {
+       workset_t *sws;
+       ir_node *blk = get_nodes_block(irn);
 
-       /* copy the best ones to the ws */
-       count = MIN(count, ws->bel->n_regs);
-       workset_bulk_fill(ws, count, starters);
+       DBG((dbg, DBG_SPILL, "Is %+F a mem-phi?\n", irn));
+       sws = get_block_info(blk)->ws_start;
+       DBG((dbg, DBG_SPILL, "  %d\n", !workset_contains(sws, irn)));
+       return !workset_contains(sws, irn);
+}
 
-       obstack_free(&ob, NULL);
+/**
+ * @return The distance to the next use
+ *         Or 0 if irn is an ignore node
+ */
+
+static INLINE unsigned get_distance(belady_env_t *bel, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
+{
+       arch_irn_flags_t fl = arch_irn_get_flags(bel->arch, def);
+       unsigned dist = be_get_next_use(bel->uses, from, from_step, def, skip_from_uses);
+
+       if(!USES_IS_INIFINITE(dist) && (fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0)
+               return 0;
+
+       return dist;
 }
 
 /**
- * Performs the actions neccessary to grant the request that:
+ * Performs the actions necessary to grant the request that:
  * - new_vals can be held in registers
  * - as few as possible other values are disposed
  * - the worst values get disposed
@@ -261,7 +274,7 @@ static void build_start_set(belady_env_t *bel, ir_node *blk) {
  */
 static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
        ir_node *val;
-       int i, len, max_allowed, demand;
+       int i, len, max_allowed, demand, iter;
        workset_t *ws = bel->ws;
        ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert));
 
@@ -269,7 +282,7 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
         * 1. Identify the number of needed slots and the values to reload
         */
        demand = 0;
-       workset_foreach(new_vals, val) {
+       workset_foreach(new_vals, val, iter) {
                /* mark value as used */
                if (is_usage)
                        pset_insert_ptr(bel->used, val);
@@ -277,39 +290,27 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
                if (!workset_contains(ws, val)) {
                        DBG((dbg, DBG_DECIDE, "    insert %+F\n", val));
                        to_insert[demand++] = val;
-
-                       if (is_usage) {
-                               spill_info_t si, *found;
-                               reloader_t *rld;
-
-                               /* find the spill info or create it */
-                               si.spilled_node = val;
-                               si.reloaders = NULL;
-                               found = set_insert(bel->spills, &si, sizeof(si), HASH_PTR(si.spilled_node));
-
-                               /* insert the reloader into the linked list */
-                               rld = obstack_alloc(&bel->ob, sizeof(*rld));
-                               rld->reloader = bel->instr;
-                               rld->next = found->reloaders;
-                               found->reloaders = rld;
-                       }
-               } else
+                       if (is_usage)
+                               be_add_reload(bel->senv, val, bel->instr);
+               } else {
                        DBG((dbg, DBG_DECIDE, "    skip %+F\n", val));
+               }
        }
        DBG((dbg, DBG_DECIDE, "    demand = %d\n", demand));
 
-
        /*
         * 2. Make room for at least 'demand' slots
         */
        len = workset_get_length(ws);
        max_allowed = bel->n_regs - demand;
 
+       DBG((dbg, DBG_DECIDE, "    disposing %d values\n", ws->len - max_allowed));
+
        /* Only make more free room if we do not have enough */
        if (len > max_allowed) {
                /* get current next-use distance */
                for (i=0; i<ws->len; ++i)
-                       workset_set_time(ws, i, be_get_next_use(bel->uses, bel->instr, bel->instr_nr, workset_get_val(ws, i), is_usage));
+                       workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
 
                /* sort entries by increasing nextuse-distance*/
                workset_sort(ws);
@@ -318,14 +319,16 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
                   before its first usage, remove it from start workset */
                for (i=max_allowed; i<ws->len; ++i) {
                        ir_node *irn = ws->vals[i].irn;
+
                        if (!pset_find_ptr(bel->used, irn)) {
                                ir_node *curr_bb = get_nodes_block(bel->instr);
-                               workset_t *ws_start = ((block_info_t *) get_irn_link(curr_bb))->ws_start;
+                               workset_t *ws_start = get_block_info(curr_bb)->ws_start;
                                workset_remove(ws_start, irn);
 
                                DBG((dbg, DBG_DECIDE, "    dispose %+F dumb\n", irn));
-                       } else
+                       } else {
                                DBG((dbg, DBG_DECIDE, "    dispose %+F\n", irn));
+                       }
                }
 
                /* kill the last 'demand' entries in the array */
@@ -338,26 +341,185 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
        workset_bulk_insert(bel->ws, demand, to_insert);
 }
 
+static void belady(ir_node *blk, void *env);
+
+/**
+ * Inserts a spill of a value at the earliest possible location in a block.
+ * That is after the last use of the value or at the beginning of the block if
+ * there is no use
+ */
+static ir_node *insert_copy(belady_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);
+
+       // walk schedule backwards until we find a usage, or until we have reached the first phi
+       // TODO can we do this faster somehow? This makes insert_copy O(n) in block_size...
+       sched_foreach_reverse(block, node) {
+               int i, arity;
+
+               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;
+                       }
+               }
+       }
+       // 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:
+
+       return copy;
+}
+
+/**
+ * Collects all values live-in at block @p blk and all phi results in this block.
+ * Then it adds the best values (at most n_regs) to the blocks start_workset.
+ * The phis among the remaining values get spilled: Introduce psudo-copies of
+ *  their args to break interference and make it possible to spill them to the
+ *  same spill slot.
+ */
+static block_info_t *compute_block_start_info(ir_node *blk, void *data) {
+       belady_env_t *env = data;
+       ir_node *irn, *first;
+       irn_live_t *li;
+       int i, count, ws_count;
+       loc_t loc, *starters;
+       ir_graph *irg = get_irn_irg(blk);
+       struct obstack ob;
+       block_info_t *res = get_block_info(blk);
+
+       /* Have we seen this block before? */
+       if (res)
+               return res;
+
+       /* Create the block info for this block. */
+       res = new_block_info(&env->ob);
+       set_block_info(blk, res);
+
+
+       /* Get all values living at the block start sorted by next use*/
+       obstack_init(&ob);
+
+       DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
+       first = sched_first(blk);
+       count = 0;
+       sched_foreach(blk, irn) {
+               if (is_Phi(irn) && arch_get_irn_reg_class(env->arch, irn, -1) == env->cls) {
+                       loc.irn = irn;
+                       loc.time = get_distance(env, first, 0, irn, 0);
+                       obstack_grow(&ob, &loc, sizeof(loc));
+                       DBG((dbg, DBG_START, "    %+F:\n", irn));
+                       count++;
+               } else
+                       break;
+       }
+
+       live_foreach(blk, li) {
+               if (live_is_in(li) && arch_get_irn_reg_class(env->arch, li->irn, -1) == env->cls) {
+                       loc.irn = (ir_node *)li->irn;
+                       loc.time = get_distance(env, first, 0, li->irn, 0);
+                       obstack_grow(&ob, &loc, sizeof(loc));
+                       DBG((dbg, DBG_START, "    %+F:\n", li->irn));
+                       count++;
+               }
+       }
+
+       starters = obstack_finish(&ob);
+       qsort(starters, count, sizeof(starters[0]), loc_compare);
+
+
+       /* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */
+       if (get_Block_n_cfgpreds(blk) == 1 && blk != get_irg_start_block(get_irn_irg(blk))) {
+               ir_node *pred_blk       = get_Block_cfgpred_block(blk, 0);
+               block_info_t *pred_info = get_block_info(pred_blk);
+
+               /* if pred block has not been processed yet, do it now */
+               if (! pred_info) {
+                       belady(pred_blk, env);
+                       pred_info = get_block_info(pred_blk);
+               }
+
+               /* now we have an end_set of pred */
+               assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
+               res->ws_start = workset_clone(&env->ob, pred_info->ws_end);
+
+       } else
+
+       /* Else we want the start_set to be the values used 'the closest' */
+       {
+               /* Copy the best ones from starters to start workset */
+               ws_count = MIN(count, env->n_regs);
+               res->ws_start = new_workset(&env->ob, env);
+               workset_bulk_fill(res->ws_start, ws_count, starters);
+       }
+
+
+
+       /* The phis of this block which are not in the start set have to be spilled later.
+        * Therefore we add temporary copies in the pred_blocks so the spills can spill
+        * into the same spill slot.
+        * After spilling these copies get deleted. */
+       for (i=workset_get_length(res->ws_start); i<count; ++i) {
+               int o, max;
+
+               irn = starters[i].irn;
+               if (!is_Phi(irn) || get_nodes_block(irn) != blk)
+                       continue;
+
+               DBG((dbg, DBG_START, "For %+F:\n", irn));
+
+               for (max=get_irn_arity(irn), o=0; o<max; ++o) {
+                       ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
+                       ir_node *arg = get_irn_n(irn, o);
+                       ir_node* copy = insert_copy(env, pred_block, arg);
+
+                       set_irn_n(irn, o, copy);
+               }
+       }
+
+       obstack_free(&ob, NULL);
+       return res;
+}
+
+
 /**
  * For the given block @p blk, decide for each values
  * whether it is used from a register or is reloaded
  * before the use.
  */
-static void decide(ir_node *blk, void *env) {
+static void belady(ir_node *blk, void *env) {
        belady_env_t *bel = env;
        workset_t *new_vals;
        ir_node *irn;
-       block_info_t *blk_info = obstack_alloc(&bel->ob, sizeof(*blk_info));
-       set_irn_link(blk, blk_info);
+       int iter;
+       block_info_t *blk_info;
+
+       /* Don't do a block twice */
+       if (get_block_info(blk))
+               return;
+
+       /* get the starting workset for this block */
+       blk_info = compute_block_start_info(blk, bel);
 
        DBG((dbg, DBG_DECIDE, "\n"));
        DBG((dbg, DBG_DECIDE, "Decide for %+F\n", blk));
 
-       /* build starting-workset for this block */
-       build_start_set(bel, blk);
-       blk_info->ws_start = workset_clone(&bel->ob, bel->ws);
-       DBG((dbg, DBG_WSETS, "Initial start workset for %+F:\n", blk));
-       workset_foreach(blk_info->ws_start, irn)
+       workset_copy(bel->ws, blk_info->ws_start);
+       DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk));
+       workset_foreach(bel->ws, irn, iter)
                DBG((dbg, DBG_WSETS, "  %+F\n", irn));
 
        /* process the block from start to end */
@@ -366,19 +528,17 @@ static void decide(ir_node *blk, void *env) {
        bel->instr_nr = 0;
        new_vals = new_workset(&bel->ob, bel);
        sched_foreach(blk, irn) {
-               ir_node *iii;
-               DBG((dbg, DBG_WSETS, "Current workset for %+F:\n", blk));
-               workset_foreach(bel->ws, iii)
-                       DBG((dbg, DBG_WSETS, "  %+F\n", iii));
                assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!");
 
-               DBG((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 
                /* projs are handled with the tuple value.
-                * Phis are no real instr (see insert_starters)
+                * Phis are no real instr (see insert_starters())
                 * instr_nr does not increase */
-               if (is_Proj(irn) || is_Phi(irn))
+               if (is_Proj(irn) || is_Phi(irn)) {
+                       DBG((dbg, DBG_DECIDE, "  ...%+F skipped\n", irn));
                        continue;
+               }
+               DBG((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 
                /* set instruction in the workset */
                bel->instr = irn;
@@ -405,64 +565,52 @@ static void decide(ir_node *blk, void *env) {
 
        /* Remember end-workset for this block */
        blk_info->ws_end = workset_clone(&bel->ob, bel->ws);
-       DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk));
-       workset_foreach(blk_info->ws_start, irn)
-               DBG((dbg, DBG_WSETS, "  %+F\n", irn));
        DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk));
-       workset_foreach(blk_info->ws_end, irn)
+       workset_foreach(blk_info->ws_end, irn, iter)
                DBG((dbg, DBG_WSETS, "  %+F\n", irn));
 }
 
 /**
- * 'decide' is block-local and makes assumtions
+ * 'decide' is block-local and makes assumptions
  * about the set of live-ins. Thus we must adapt the
  * live-outs to the live-ins at each block-border.
  */
 static void fix_block_borders(ir_node *blk, void *env) {
+       workset_t *wsb;
        belady_env_t *bel = env;
-       int i, max;
+       int i, max, iter, iter2;
 
        DBG((dbg, DBG_FIX, "\n"));
        DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
 
-       workset_t *wsb = ((block_info_t *)get_irn_link(blk))->ws_start;
+       wsb = get_block_info(blk)->ws_start;
 
        /* process all pred blocks */
        for (i=0, max=get_irn_arity(blk); i<max; ++i) {
                ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
-               workset_t *wsp = ((block_info_t *)get_irn_link(pred))->ws_end;
+               workset_t *wsp = get_block_info(pred)->ws_end;
 
                DBG((dbg, DBG_FIX, "  Pred %+F\n", pred));
 
-               workset_foreach(wsb, irnb) {
-                       spill_info_t si, *found;
-                       reloader_t *rld;
-
+               workset_foreach(wsb, irnb, iter) {
                        /* if irnb is a phi of the current block we reload
                         * the corresponding argument, else irnb itself */
                        if(is_Phi(irnb) && blk == get_nodes_block(irnb))
                                irnb = get_irn_n(irnb, i);
 
+                       /* Unknowns are available everywhere */
+                       if(get_irn_opcode(irnb) == iro_Unknown)
+                               continue;
+
                        /* check if irnb is in a register at end of pred */
-                       workset_foreach(wsp, irnp)
+                       workset_foreach(wsp, irnp, iter2) {
                                if (irnb == irnp)
                                        goto next_value;
+                       }
 
                        /* irnb is in memory at the end of pred, so we have to reload it */
-
-                       /* find the spill info or create it */
-                       si.spilled_node = irnb;
-                       si.reloaders = NULL;
-                       found = set_insert(bel->spills, &si, sizeof(si), HASH_PTR(si.spilled_node));
-
-                       /* insert the reloader into the linked list.
-                        * the schedule position depends on the cf-situation of the block */
-                       rld = obstack_alloc(&bel->ob, sizeof(*rld));
-                       rld->reloader = (max==1) ? sched_skip(sched_first(blk), 1, sched_skip_phi_predicator, NULL) : pred;
-                       rld->next = found->reloaders;
-                       found->reloaders = rld;
-
-                       DBG((dbg, DBG_FIX, "    reload %+F before %+F\n", irnb, rld->reloader));
+                       DBG((dbg, DBG_FIX, "    reload %+F\n", irnb));
+                       be_add_reload_on_edge(bel->senv, irnb, blk, i);
 
 next_value:
                        /*epsilon statement :)*/;
@@ -470,106 +618,69 @@ next_value:
        }
 }
 
-static void insert_spills_reloads(ir_graph *irg, belady_env_t *bel) {
-       ir_node *irn;
-       spill_info_t *si;
-       struct obstack ob;
-       obstack_init(&ob);
-
-       /* get all special spilled phis */
-       for(si = set_first(bel->spills); si; si = set_next(bel->spills)) {
-               irn = si->spilled_node;
-               if (is_Phi(irn)) {
-                       ir_node *blk = get_nodes_block(irn);
-                       workset_t *sws = ((block_info_t *)get_irn_link(blk))->ws_start;
-                       if (!workset_contains(sws, irn))
-                               pset_insert_ptr(bel->senv.mem_phis, irn);
-               }
-       }
+/**
+ * Removes all copies introduced for phi-spills
+ */
+static void remove_copies(belady_env_t *env) {
+       int i;
 
-       /* process each spilled node */
-       for(si = set_first(bel->spills); si; si = set_next(bel->spills)) {
-               reloader_t *rld;
-               ir_node **reloads;
-               int n_reloads = 0;
-               ir_mode *mode = get_irn_mode(si->spilled_node);
-
-               /* go through all reloads for this spill */
-               for(rld = si->reloaders; rld; rld = rld->next) {
-                       /* the spill for this reloader */
-                       ir_node *spill   = be_spill_node(&bel->senv, si->spilled_node);
-
-                       /* the reload */
-                       ir_node *bl      = is_Block(rld->reloader) ? rld->reloader : get_nodes_block(rld->reloader);
-                       ir_node *reload  = new_Reload(bel->factory, bel->cls, irg, bl, mode, spill);
-                       DBG((dbg, DBG_SPILL, " RELOADER %+F   Reload %+F of %+F\n", rld->reloader, reload, si->spilled_node));
-                       pset_insert_ptr(bel->reloads, reload);
-
-                       /* remember the reaload */
-                       obstack_ptr_grow(&ob, reload);
-                       sched_add_before(rld->reloader, reload);
-                       n_reloads++;
-               }
+       for(i = 0; i < ARR_LEN(env->copies); ++i) {
+               ir_node *node = env->copies[i];
+               ir_node *src;
+               const ir_edge_t *edge, *ne;
 
-               assert(n_reloads > 0);
-               reloads = obstack_finish(&ob);
-               be_introduce_copies_ignore(bel->senv.session->dom_front, si->spilled_node, n_reloads, reloads, bel->senv.mem_phis);
-               obstack_free(&ob, reloads);
-       }
+               assert(be_is_Copy(node));
 
-       obstack_free(&ob, NULL);
+               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);
 
-       be_remove_spilled_phis(&bel->senv);
+                       set_irn_n(user, user_pos, src);
+               }
+       }
 }
 
-/**
- * Removes all used reloads from bel->reloads.
- * The remaining nodes in bel->reloads will be removed from the graph.
- */
-static void rescue_used_reloads(ir_node *irn, void *env) {
-       pset *rlds = ((belady_env_t *)env)->reloads;
-       if (pset_find_ptr(rlds, irn)) {
-               DBG((dbg, DBG_SPILL, "Removing %+F in %+F\n", irn, get_nodes_block(irn)));
-               pset_remove_ptr(rlds, irn);
-       }
+void be_spill_belady(const be_chordal_env_t *chordal_env) {
+       be_spill_belady_spill_env(chordal_env, NULL);
 }
 
-void be_spill_belady(const be_main_session_env_t *session, const arch_register_class_t *cls) {
-       ir_node *irn;
+void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
+       belady_env_t bel;
 
-       dbg = firm_dbg_register("ir.be.spillbelady");
-       firm_dbg_set_mask(dbg, DEBUG_LVL);
+       FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
 
        /* init belady env */
-       belady_env_t *bel = alloca(sizeof(*bel));
-       obstack_init(&bel->ob);
-       bel->senv.session = session;
-       bel->factory = session->main_env->node_factory;
-       bel->arch = session->main_env->arch_env;
-       bel->cls = cls;
-       bel->n_regs = arch_register_class_n_regs(cls);
-       bel->ws = new_workset(&bel->ob, bel);
-       bel->uses = be_begin_uses(session->irg, session->main_env->arch_env, cls);
-       bel->spills = new_set(set_cmp_spillinfo, 32);
-       bel->senv.spill_ctxs = new_set(be_set_cmp_spillctx, 32);
-       bel->senv.mem_phis = pset_new_ptr_default();
-       bel->reloads = pset_new_ptr_default();
+       obstack_init(&bel.ob);
+       bel.arch      = chordal_env->birg->main_env->arch_env;
+       bel.cls       = chordal_env->cls;
+       bel.n_regs    = arch_register_class_n_regs(bel.cls);
+       bel.ws        = new_workset(&bel.ob, &bel);
+       bel.uses      = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls);
+       if(spill_env == NULL) {
+               bel.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL);
+       } else {
+               bel.senv = spill_env;
+               be_set_is_spilled_phi(bel.senv, is_mem_phi, NULL);
+       }
+       DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);)
+       bel.copies    = NEW_ARR_F(ir_node*, 0);
+
+       DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
 
        /* do the work */
-       irg_block_walk_graph(session->irg, decide, NULL, bel);
-       irg_block_walk_graph(session->irg, fix_block_borders, NULL, bel);
-       insert_spills_reloads(session->irg, bel);
+       be_clear_links(chordal_env->irg);
+       irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel);
+       irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
+       be_insert_spills_reloads(bel.senv);
+       remove_copies(&bel);
+       DEL_ARR_F(bel.copies);
 
-       /* find all unused reloads and remove them from the schedule */
-       irg_walk_graph(session->irg, rescue_used_reloads, NULL, bel);
-       for(irn = pset_first(bel->reloads); irn; irn = pset_next(bel->reloads))
-               sched_remove(irn);
+       be_remove_dead_nodes_from_schedule(chordal_env->irg);
 
        /* clean up */
-       del_pset(bel->reloads);
-       del_pset(bel->senv.mem_phis);
-       del_set(bel->senv.spill_ctxs);
-       del_set(bel->spills);
-       be_end_uses(bel->uses);
-       obstack_free(&bel->ob, NULL);
+       if(spill_env == NULL)
+               be_delete_spill_env(bel.senv);
+       be_end_uses(bel.uses);
+       obstack_free(&bel.ob, NULL);
 }