X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=9b3a72b8aeb47500ef2fd301e7bca51095848841;hb=893a9141a398d8a8e9fa4fd6db8b3fbd4e2d55a5;hp=0742f4a14e346817ba657f2e17e5cd02bbffbdac;hpb=81d146ecea7e0f3bceb7949fa91af98be2884a8c;p=libfirm

diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c
index 0742f4a14..9b3a72b8a 100644
--- a/ir/be/bespillbelady.c
+++ b/ir/be/bespillbelady.c
@@ -1,117 +1,159 @@
-/**
- * Author:      Daniel Grund
- * Date:		20.09.2005
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+/*
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
  *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
+/**
+ * @file
+ * @brief       Beladys spillalgorithm.
+ * @author      Daniel Grund, Matthias Braun
+ * @date        20.09.2005
+ * @version     $Id$
  */
 #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 <stdbool.h>
 
 #include "obst.h"
-#include "set.h"
-#include "pset.h"
 #include "irprintf_t.h"
 #include "irgraph.h"
 #include "irnode.h"
 #include "irmode.h"
 #include "irgwalk.h"
+#include "irloop.h"
 #include "iredges_t.h"
 #include "ircons_t.h"
 #include "irprintf.h"
+#include "irnodeset.h"
+#include "xmalloc.h"
+#include "pdeq.h"
 
 #include "beutil.h"
-#include "bearch.h"
-#include "bespillbelady.h"
-#include "beuses_t.h"
+#include "bearch_t.h"
+#include "beuses.h"
 #include "besched_t.h"
 #include "beirgmod.h"
 #include "belive_t.h"
 #include "benode_t.h"
 #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
+#include "bespilloptions.h"
+#include "beloopana.h"
+#include "beirg_t.h"
+#include "bespill.h"
+#include "bemodule.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 _belady_env_t {
-	struct obstack ob;
-	const be_chordal_env_t *cenv;
-	const arch_env_t *arch;
-	const arch_register_class_t *cls;
-	int n_regs;			/** number of regs in this reg-class */
+/* factor to weight the different costs of reloading/rematerializing a node
+   (see bespill.h be_get_reload_costs_no_weight) */
+#define RELOAD_COST_FACTOR   10
 
-	workset_t *ws;		/**< the main workset used while processing a block. ob-allocated */
-	be_uses_t *uses;	/**< env for the next-use magic */
-	ir_node *instr;		/**< current instruction */
-	unsigned instr_nr;	/**< current instruction number (relative to block start) */
-	pset *used;
+#define TIME_UNDEFINED 6666
 
-	spill_env_t *senv;	/**< see bespill.h */
-} belady_env_t;
+#define PLACE_SPILLS
 
-struct _workset_t {
-	int len;			/**< current length */
-	loc_t vals[0];		/**< inlined array of the values/distances in this working set */
-};
+/**
+ * An association between a node and a point in time.
+ */
+typedef struct loc_t {
+	ir_node          *node;
+	unsigned          time;     /**< A use time (see beuses.h). */
+	bool              spilled;  /**< the value was already spilled on this path */
+} loc_t;
+
+typedef struct _workset_t {
+	int   len;          /**< current length */
+	loc_t vals[0];      /**< inlined array of the values/distances in this working set */
+} workset_t;
+
+static struct obstack               obst;
+static const arch_env_t            *arch_env;
+static const arch_register_class_t *cls;
+static const be_lv_t               *lv;
+static be_loopana_t                *loop_ana;
+static int                          n_regs;
+static workset_t                   *ws;     /**< the main workset used while
+	                                             processing a block. */
+static be_uses_t                   *uses;   /**< env for the next-use magic */
+static ir_node                     *instr;  /**< current instruction */
+static unsigned                     instr_nr; /**< current instruction number
+	                                               (relative to block start) */
+static ir_nodeset_t                 used;
+static spill_env_t                 *senv;   /**< see bespill.h */
+static pdeq                        *worklist;
+
+static int loc_compare(const void *a, const void *b)
+{
+	const loc_t *p = a;
+	const loc_t *q = b;
+	return p->time - q->time;
+}
 
 void workset_print(const workset_t *w)
 {
 	int i;
 
 	for(i = 0; i < w->len; ++i) {
-		ir_fprintf(stderr, "%+F %d\n", w->vals[i].irn, w->vals[i].time);
+		ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
 	}
 }
 
 /**
  * Alloc a new workset on obstack @p ob with maximum size @p max
  */
-static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) {
+static workset_t *new_workset(void)
+{
 	workset_t *res;
-	size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
-	res = obstack_alloc(ob, size);
+	size_t     size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
+
+	res  = obstack_alloc(&obst, size);
 	memset(res, 0, size);
 	return res;
 }
 
 /**
- * Alloc a new instance on obstack and make it equal to @param ws
+ * Alloc a new instance on obstack and make it equal to @param workset
  */
-static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) {
+static workset_t *workset_clone(workset_t *workset)
+{
 	workset_t *res;
-	size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
-	res = obstack_alloc(ob, size);
-	memcpy(res, ws, size);
+	size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
+	res = obstack_alloc(&obst, size);
+	memcpy(res, workset, size);
 	return res;
 }
 
 /**
- * Do NOT alloc anything. Make @param tgt equal to @param src.
- * returns @param tgt for convenience
+ * Copy workset @param src to @param tgt
  */
-static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src) {
-	size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]);
-	memcpy(tgt, src, size);
-	return tgt;
+static void workset_copy(workset_t *dest, const workset_t *src)
+{
+	size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
+	memcpy(dest, src, size);
 }
 
 /**
@@ -119,7 +161,8 @@ static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset
  * @param count locations given at memory @param locs.
  * Set the length of @param ws to count.
  */
-static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) {
+static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
+{
 	workset->len = count;
 	memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
 }
@@ -128,52 +171,66 @@ static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t
  * 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(belady_env_t *env, workset_t *ws, ir_node *val) {
-	int i;
+static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
+{
+	loc_t *loc;
+	int    i;
 	/* check for current regclass */
-	if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
-		DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val));
-		return;
-	}
+	assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val));
 
 	/* check if val is already contained */
-	for(i=0; i<ws->len; ++i)
-		if (ws->vals[i].irn == val)
+	for (i = 0; i < workset->len; ++i) {
+		loc = &workset->vals[i];
+		if (loc->node == val) {
+			if (spilled) {
+				loc->spilled = true;
+			}
 			return;
+		}
+	}
 
 	/* insert val */
-	assert(ws->len < env->n_regs && "Workset already full!");
-	ws->vals[ws->len++].irn = val;
+	assert(workset->len < n_regs && "Workset already full!");
+	loc           = &workset->vals[workset->len];
+	loc->node     = val;
+	loc->spilled  = spilled;
+	loc->time     = TIME_UNDEFINED;
+	workset->len++;
 }
 
 /**
  * Removes all entries from this workset
  */
-static INLINE void workset_clear(workset_t *ws) {
-	ws->len = 0;
+static void workset_clear(workset_t *workset)
+{
+	workset->len = 0;
 }
 
 /**
  * Removes the value @p val from the workset if present.
  */
-static INLINE void workset_remove(workset_t *ws, ir_node *val) {
+static INLINE void workset_remove(workset_t *workset, ir_node *val)
+{
 	int i;
-	for(i=0; i<ws->len; ++i) {
-		if (ws->vals[i].irn == val) {
-			ws->vals[i] = ws->vals[--ws->len];
+	for(i = 0; i < workset->len; ++i) {
+		if (workset->vals[i].node == val) {
+			workset->vals[i] = workset->vals[--workset->len];
 			return;
 		}
 	}
 }
 
-static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
+static INLINE const loc_t *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)
-			return 1;
+
+	for (i = 0; i < ws->len; ++i) {
+		if (ws->vals[i].node == val)
+			return &ws->vals[i];
 	}
 
-	return 0;
+	return NULL;
 }
 
 /**
@@ -183,89 +240,60 @@ static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
  * @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; \
+										v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
 										++i)
 
 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
 #define workset_get_time(ws, i) (ws)->vals[i].time
 #define workset_set_length(ws, length) (ws)->len = length
 #define workset_get_length(ws) ((ws)->len)
-#define workset_get_val(ws, i) ((ws)->vals[i].irn)
+#define workset_get_val(ws, i) ((ws)->vals[i].node)
 #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;
-	int processed;
+typedef struct _block_info_t
+{
+	workset_t *start_workset;
+	workset_t *end_workset;
 } block_info_t;
 
 
-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;
-	res->processed = 0;
+static void *new_block_info(void)
+{
+	block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
+	memset(res, 0, sizeof(res[0]));
 
 	return res;
 }
 
-#define get_block_info(blk)			((block_info_t *)get_irn_link(blk))
-#define set_block_info(blk, info)	set_irn_link(blk, info)
+#define get_block_info(block)        ((block_info_t *)get_irn_link(block))
+#define set_block_info(block, info)  set_irn_link(block, info)
 
 /**
  * @return The distance to the next use or 0 if irn has dont_spill flag set
  */
-static INLINE unsigned get_distance(belady_env_t *env, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
+static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
+                                    const ir_node *def, int skip_from_uses)
 {
-	int flags = arch_irn_get_flags(env->arch, def);
-	unsigned dist = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
+	be_next_use_t use;
+	int           flags = arch_irn_get_flags(arch_env, def);
+	unsigned      costs;
+	unsigned      time;
 
 	assert(! (flags & arch_irn_flags_ignore));
 
-	/* we have to keep nonspillable nodes in the workingset */
-	if(flags & arch_irn_flags_dont_spill)
-		return 0;
-
-	return dist;
-}
-
-/**
- * Fix to remove dead nodes (especially don't spill nodes) from workset.
- */
-static void fix_dead_values(workset_t *ws, ir_node *irn) {
-	int idx;
-	ir_node *node;
-	ir_node *block = get_nodes_block(irn);
-
-	DBG((dbg, DBG_DECIDE, "fixing dead values at %+F:\n", irn));
-
-	workset_foreach(ws, node, idx) {
-		const ir_edge_t *edge;
-		int             fixme = 1;
-
-		/* skip already fixed nodes */
-		if (workset_get_time(ws, idx) == INT_MAX)
-			continue;
+	use = be_get_next_use(uses, from, from_step, def, skip_from_uses);
+	if (USES_IS_INFINITE(use.time))
+		return USES_INFINITY;
 
-		/* check all users */
-		foreach_out_edge(node, edge) {
-			ir_node *user = get_edge_src_irn(edge);
-
-			if ((get_nodes_block(user) != block)                           ||  /* user is in a different block */
-				(sched_is_scheduled(user) && sched_comes_after(irn, user)) ||  /* user is scheduled after irn */
-				user == irn)                                                   /* irn is the user */
-			{                                                                  /* => don't fix distance */
-				fixme = 0;
-				break;
-			}
-		}
+	/* We have to keep nonspillable nodes in the workingset */
+	if (flags & arch_irn_flags_dont_spill)
+		return 0;
 
-		/* all users scheduled prior to current irn in in same block as irn -> fix */
-		if (fixme) {
-			workset_set_time(ws, idx, INT_MAX);
-			DBG((dbg, DBG_DECIDE, "\tfixing time for %+F to INT_MAX\n", node));
-		}
-	}
+	costs = be_get_reload_costs_no_weight(senv, def, use.before);
+	assert(costs * RELOAD_COST_FACTOR < 1000);
+	time  = use.time + 1000 - (costs * RELOAD_COST_FACTOR);
 
+	return time;
 }
 
 /**
@@ -277,285 +305,644 @@ static void fix_dead_values(workset_t *ws, ir_node *irn) {
  * @p is_usage indicates that the values in new_vals are used (not defined)
  * In this case reloads must be performed
  */
-static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) {
-	ir_node *val;
-	int     i, len, max_allowed, demand, iter;
-
-	workset_t *ws         = env->ws;
-	ir_node   **to_insert = alloca(env->n_regs * sizeof(*to_insert));
-
-	/*
-		1. Identify the number of needed slots and the values to reload
-	*/
+static void displace(workset_t *new_vals, int is_usage)
+{
+	ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
+	bool     *spilled   = alloca(n_regs * sizeof(spilled[0]));
+	ir_node  *val;
+	int       i;
+	int       len;
+	int       spills_needed;
+	int       demand;
+	int       iter;
+
+	/* 1. Identify the number of needed slots and the values to reload */
 	demand = 0;
 	workset_foreach(new_vals, val, iter) {
+		bool reloaded = false;
+
 		/* mark value as used */
 		if (is_usage)
-			pset_insert_ptr(env->used, val);
+			ir_nodeset_insert(&used, val);
 
 		if (! workset_contains(ws, val)) {
-			DBG((dbg, DBG_DECIDE, "    insert %+F\n", val));
-			to_insert[demand++] = val;
-			if (is_usage)
-				be_add_reload(env->senv, val, env->instr);
-		}
-		else {
-			assert(is_usage || "Defined value already in workset?!?");
-			DBG((dbg, DBG_DECIDE, "    skip %+F\n", val));
+			DB((dbg, DBG_DECIDE, "    insert %+F\n", val));
+			if (is_usage) {
+				DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
+				be_add_reload(senv, val, instr, cls, 1);
+				reloaded = true;
+			}
+		} else {
+			DB((dbg, DBG_DECIDE, "    %+F already in workset\n", val));
+			assert(is_usage);
+			/* remove the value from the current workset so it is not accidently
+			 * spilled */
+			workset_remove(ws, val);
 		}
+		spilled[demand]   = reloaded;
+		to_insert[demand] = val;
+		++demand;
 	}
-	DBG((dbg, DBG_DECIDE, "    demand = %d\n", demand));
 
-	/*
-		2. Make room for at least 'demand' slots
-	*/
-	len         = workset_get_length(ws);
-	max_allowed = env->n_regs - demand;
-
-	DBG((dbg, DBG_DECIDE, "    disposing %d values\n", ws->len - max_allowed));
+	/* 2. Make room for at least 'demand' slots */
+	len           = workset_get_length(ws);
+	spills_needed = len + demand - n_regs;
+	assert(spills_needed <= len);
 
 	/* 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, get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage));
-
-		/*
-			FIX for don't spill nodes:
-			Problem is that get_distance always returns 0 for those nodes even if they are not
-			needed anymore (all their usages have already been visited).
-			Even if we change this behavior, get_distance doesn't distinguish between not
-			used anymore (dead) and live out of block.
-			Solution: Set distances of all nodes having all their usages in schedule prior to
-			current instruction to MAX_INT.
-		*/
-		fix_dead_values(ws, env->instr);
+	if (spills_needed > 0) {
+#ifndef PLACE_SPILLS
+		ir_node   *curr_bb  = get_nodes_block(instr);
+		workset_t *ws_start = get_block_info(curr_bb)->start_workset;
+#endif
+
+		DB((dbg, DBG_DECIDE, "    disposing %d values\n", spills_needed));
+
+		/* calculate current next-use distance for live values */
+		for (i = 0; i < len; ++i) {
+			ir_node  *val  = workset_get_val(ws, i);
+			unsigned  dist = get_distance(instr, instr_nr, val, !is_usage);
+			workset_set_time(ws, i, dist);
+		}
 
 		/* sort entries by increasing nextuse-distance*/
 		workset_sort(ws);
 
-		/*
-			Logic for not needed live-ins: If a value is disposed
-			before its first usage, remove it from start workset
-			We don't do this for phis though
-		*/
-		for (i = max_allowed; i < ws->len; ++i) {
-			ir_node *irn = ws->vals[i].irn;
+		for (i = len - spills_needed; i < len; ++i) {
+			ir_node *val = ws->vals[i].node;
 
-            if (is_Phi(irn))
-                continue;
+			DB((dbg, DBG_DECIDE, "    disposing node %+F (%u)\n", val,
+			     workset_get_time(ws, i)));
 
-			if (! pset_find_ptr(env->used, irn)) {
-				ir_node   *curr_bb  = get_nodes_block(env->instr);
-				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));
+#ifdef PLACE_SPILLS
+			if (!USES_IS_INFINITE(ws->vals[i].time) && !ws->vals[i].spilled) {
+				ir_node *after_pos = sched_prev(instr);
+				DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val,
+				    after_pos));
+				be_add_spill(senv, val, after_pos);
 			}
-			else {
-				DBG((dbg, DBG_DECIDE, "    dispose %+F\n", irn));
+#endif
+
+#ifndef PLACE_SPILLS
+			/* Logic for not needed live-ins: If a value is disposed
+			 * before its first use, remove it from start workset
+			 * We don't do this for phis though	*/
+			if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) {
+				workset_remove(ws_start, val);
+				DB((dbg, DBG_DECIDE, "    (and removing %+F from start workset)\n", val));
 			}
+#endif
 		}
 
 		/* kill the last 'demand' entries in the array */
-		workset_set_length(ws, max_allowed);
+		workset_set_length(ws, len - spills_needed);
 	}
 
-	/*
-		3. Insert the new values into the workset
-	*/
-	for (i = 0; i < demand; ++i)
-		workset_insert(env, env->ws, to_insert[i]);
+	/* 3. Insert the new values into the workset */
+	for (i = 0; i < demand; ++i) {
+		ir_node *val = to_insert[i];
+
+		workset_insert(ws, val, spilled[i]);
+	}
 }
 
-static void belady(ir_node *blk, void *env);
+enum {
+	AVAILABLE_EVERYWHERE,
+	AVAILABLE_NOWHERE,
+	AVAILABLE_PARTLY,
+	AVAILABLE_UNKNOWN
+};
 
-/*
- * Computes set of live-ins for each block with multiple predecessors
- * and notifies spill algorithm which phis need to be spilled
+static unsigned available_in_all_preds(workset_t* const* pred_worksets,
+                                       size_t n_pred_worksets,
+                                       const ir_node *value, bool is_local_phi)
+{
+	size_t i;
+	bool   avail_everywhere = true;
+	bool   avail_nowhere    = true;
+
+	assert(n_pred_worksets > 0);
+
+	/* value available in all preds? */
+	for (i = 0; i < n_pred_worksets; ++i) {
+		bool             found     = false;
+		const workset_t *p_workset = pred_worksets[i];
+		int              p_len     = workset_get_length(p_workset);
+		int              p_i;
+		const ir_node   *l_value;
+
+		if (is_local_phi) {
+			assert(is_Phi(value));
+			l_value = get_irn_n(value, i);
+		} else {
+			l_value = value;
+		}
+
+		for (p_i = 0; p_i < p_len; ++p_i) {
+			const loc_t *p_l = &p_workset->vals[p_i];
+			if (p_l->node != l_value)
+				continue;
+
+			found = true;
+			break;
+		}
+
+		if (found) {
+			avail_nowhere = false;
+		} else {
+			avail_everywhere = false;
+		}
+	}
+
+	if (avail_everywhere) {
+		assert(!avail_nowhere);
+		return AVAILABLE_EVERYWHERE;
+	} else if (avail_nowhere) {
+		return AVAILABLE_NOWHERE;
+	} else {
+		return AVAILABLE_PARTLY;
+	}
+}
+
+/** Decides whether a specific node should be in the start workset or not
+ *
+ * @param env      belady environment
+ * @param first
+ * @param node     the node to test
+ * @param loop     the loop of the node
  */
-static void spill_phi_walker(ir_node *block, void *data) {
-	belady_env_t *env = data;
-	block_info_t *block_info;
-	ir_node *first, *irn;
-	loc_t loc, *starters;
-	int i, len, ws_count;
-
-	if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
-		return;
+static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
+                                    ir_loop *loop, unsigned available)
+{
+	be_next_use_t next_use;
+	loc_t         loc;
 
-	block_info = new_block_info(&env->ob);
-	set_block_info(block, block_info);
+	loc.time    = USES_INFINITY;
+	loc.node    = node;
+	loc.spilled = false;
+
+	if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
+		loc.time = USES_INFINITY;
+		return loc;
+	}
+
+	/* We have to keep nonspillable nodes in the workingset */
+	if (arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
+		loc.time = 0;
+		DB((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
+		return loc;
+	}
+
+	next_use = be_get_next_use(uses, first, 0, node, 0);
+	if (USES_IS_INFINITE(next_use.time)) {
+		// the nodes marked as live in shouldn't be dead, so it must be a phi
+		assert(is_Phi(node));
+		loc.time = USES_INFINITY;
+		DB((dbg, DBG_START, "    %+F not taken (dead)\n", node));
+		return loc;
+	}
+
+	loc.time = next_use.time;
+
+	if (available == AVAILABLE_EVERYWHERE) {
+		DB((dbg, DBG_START, "    %+F taken (%u, live in all preds)\n", node,
+		    loc.time));
+		return loc;
+	} else if(available == AVAILABLE_NOWHERE) {
+		DB((dbg, DBG_START, "    %+F not taken (%u, live in no pred)\n", node,
+		    loc.time));
+		loc.time = USES_INFINITY;
+		return loc;
+	}
+
+	if (next_use.outermost_loop >= get_loop_depth(loop)) {
+		DB((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time,
+		    next_use.outermost_loop));
+	} else {
+		loc.time = USES_PENDING;
+		DB((dbg, DBG_START, "    %+F delayed (outerdepth %d < loopdepth %d)\n",
+		    node, next_use.outermost_loop, get_loop_depth(loop)));
+	}
+	return loc;
+}
+
+/**
+ * Computes the start-workset for a block with multiple predecessors. We assume
+ * that at least 1 of the predeccesors is a back-edge which means we're at the
+ * beginning of a loop. We try to reload as much values as possible now so they
+ * don't get reloaded inside the loop.
+ */
+static void decide_start_workset(const ir_node *block)
+{
+	ir_loop    *loop = get_irn_loop(block);
+	ir_node    *first;
+	ir_node    *node;
+	loc_t       loc;
+	loc_t      *starters;
+	loc_t      *delayed;
+	int         i, len, ws_count;
+	int	        free_slots, free_pressure_slots;
+	unsigned    pressure;
+	int         arity;
+	workset_t **pred_worksets;
+	bool        all_preds_known;
+
+	/* check predecessors */
+	arity           = get_irn_arity(block);
+	pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
+	all_preds_known = true;
+	for(i = 0; i < arity; ++i) {
+		ir_node      *pred_block = get_Block_cfgpred_block(block, i);
+		block_info_t *pred_info  = get_block_info(pred_block);
+
+		if (pred_info == NULL) {
+			pred_worksets[i] = NULL;
+			all_preds_known  = false;
+		} else {
+			pred_worksets[i] = pred_info->end_workset;
+		}
+	}
 
 	/* Collect all values living at start of block */
 	starters = NEW_ARR_F(loc_t, 0);
+	delayed  = NEW_ARR_F(loc_t, 0);
 
-	/* rebuild schedule time information, because it seems to be broken */
-	sched_renumber(block);
-
-	DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
+	DB((dbg, DBG_START, "Living at start of %+F:\n", block));
 	first = sched_first(block);
-	sched_foreach(block, irn) {
-		if(!is_Phi(irn))
+
+	/* check all Phis first */
+	sched_foreach(block, node) {
+		unsigned available;
+
+		if (! is_Phi(node))
 			break;
-		if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn))
+		if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
 			continue;
 
-		loc.irn = irn;
-		loc.time = get_distance(env, first, 0, irn, 0);
-		ARR_APP1(loc_t, starters, loc);
-		DBG((dbg, DBG_START, "    %+F:\n", irn));
+		if (all_preds_known) {
+			available = available_in_all_preds(pred_worksets, arity, node, true);
+		} else {
+			available = AVAILABLE_UNKNOWN;
+		}
+
+		loc = to_take_or_not_to_take(first, node, loop, available);
+
+		if (! USES_IS_INFINITE(loc.time)) {
+			if (USES_IS_PENDING(loc.time))
+				ARR_APP1(loc_t, delayed, loc);
+			else
+				ARR_APP1(loc_t, starters, loc);
+		} else {
+			be_spill_phi(senv, node);
+		}
+	}
+
+	/* check all Live-Ins */
+	be_lv_foreach(lv, block, be_lv_state_in, i) {
+		ir_node *node = be_lv_get_irn(lv, block, i);
+		unsigned available;
+
+		if (all_preds_known) {
+			available = available_in_all_preds(pred_worksets, arity, node, false);
+		} else {
+			available = AVAILABLE_UNKNOWN;
+		}
+
+		loc = to_take_or_not_to_take(first, node, loop, available);
+
+		if (! USES_IS_INFINITE(loc.time)) {
+			if (USES_IS_PENDING(loc.time))
+				ARR_APP1(loc_t, delayed, loc);
+			else
+				ARR_APP1(loc_t, starters, loc);
+		}
+	}
+
+	pressure            = be_get_loop_pressure(loop_ana, cls, loop);
+	assert(ARR_LEN(delayed) <= (signed)pressure);
+	free_slots          = n_regs - ARR_LEN(starters);
+	free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
+	free_slots          = MIN(free_slots, free_pressure_slots);
+
+	/* so far we only put nodes into the starters list that are used inside
+	 * the loop. If register pressure in the loop is low then we can take some
+	 * values and let them live through the loop */
+	if (free_slots > 0) {
+		qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
+
+		for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
+			int    p, arity;
+			loc_t *loc = & delayed[i];
+
+			/* don't use values which are dead in a known predecessors
+			 * to not induce unnecessary reloads */
+			arity = get_irn_arity(block);
+			for (p = 0; p < arity; ++p) {
+				ir_node      *pred_block = get_Block_cfgpred_block(block, p);
+				block_info_t *pred_info  = get_block_info(pred_block);
+
+				if (pred_info == NULL)
+					continue;
+
+				if (!workset_contains(pred_info->end_workset, loc->node)) {
+					DB((dbg, DBG_START,
+					    "    delayed %+F not live at pred %+F\n", loc->node,
+					    pred_block));
+					goto skip_delayed;
+				}
+			}
+
+			DB((dbg, DBG_START, "    delayed %+F taken\n", loc->node));
+			ARR_APP1(loc_t, starters, *loc);
+			loc->node = NULL;
+		skip_delayed:
+			;
+		}
 	}
 
-	be_lv_foreach(env->cenv->lv, block, be_lv_state_in, i) {
-		ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
-		if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn))
+	/* spill phis (the actual phis not just their values) that are in this block
+	 * but not in the start workset */
+	for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
+		ir_node *node = delayed[i].node;
+		if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
 			continue;
 
-		loc.irn = irn;
-		loc.time = get_distance(env, first, 0, irn, 0);
-		ARR_APP1(loc_t, starters, loc);
-		DBG((dbg, DBG_START, "    %+F:\n", irn));
+		DB((dbg, DBG_START, "    spilling delayed phi %+F\n", node));
+		be_spill_phi(senv, node);
 	}
+	DEL_ARR_F(delayed);
 
-	// Sort start values by first use
+	/* Sort start values by first use */
 	qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 
 	/* Copy the best ones from starters to start workset */
-	ws_count = MIN(ARR_LEN(starters), env->n_regs);
-	block_info->ws_start = new_workset(env, &env->ob);
-	workset_bulk_fill(block_info->ws_start, ws_count, starters);
-
-	/* The phis of this block which are not in the start set have to be spilled later. */
-	for (i = ws_count, len = ARR_LEN(starters); i < len; ++i) {
-		irn = starters[i].irn;
-		if (!is_Phi(irn) || get_nodes_block(irn) != block)
+	ws_count = MIN(ARR_LEN(starters), n_regs);
+	workset_clear(ws);
+	workset_bulk_fill(ws, ws_count, starters);
+
+	/* spill phis (the actual phis not just their values) that are in this block
+	 * but not in the start workset */
+	len = ARR_LEN(starters);
+	for (i = ws_count; i < len; ++i) {
+		ir_node *node = starters[i].node;
+		if (! is_Phi(node) || get_nodes_block(node) != block)
 			continue;
 
-		be_spill_phi(env->senv, irn);
+		DB((dbg, DBG_START, "    spilling phi %+F\n", node));
+		be_spill_phi(senv, node);
 	}
 
 	DEL_ARR_F(starters);
-}
 
-/**
- * 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(belady_env_t *env, ir_node *block) {
-	ir_node *pred_block;
-	block_info_t *res, *pred_info;
+	/* determine spill status of the values: If there's 1 pred block (which
+	 * is no backedge) where the value is spilled then we must set it to
+	 * spilled here. */
+	for(i = 0; i < ws_count; ++i) {
+		loc_t   *loc     = &ws->vals[i];
+		ir_node *value   = loc->node;
+		bool     spilled;
+		int      n;
+
+		/* phis from this block aren't spilled */
+		if (get_nodes_block(value) == block) {
+			assert(is_Phi(value));
+			loc->spilled = false;
+			continue;
+		}
 
-	/* Have we seen this block before? */
-	res = get_block_info(block);
-	if (res)
-		return res;
+		/* determine if value was spilled on any predecessor */
+		spilled = false;
+		for(n = 0; n < arity; ++n) {
+			workset_t *pred_workset = pred_worksets[n];
+			int        p_len;
+			int        p;
 
-	/* Create the block info for this block. */
-	res = new_block_info(&env->ob);
-	set_block_info(block, res);
+			if (pred_workset == NULL)
+				continue;
 
-	/* Use endset of predecessor block as startset */
-	assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block)));
-	pred_block = get_Block_cfgpred_block(block, 0);
-	pred_info = get_block_info(pred_block);
+			p_len = workset_get_length(pred_workset);
+			for(p = 0; p < p_len; ++p) {
+				loc_t *l = &pred_workset->vals[p];
 
-	/* if pred block has not been processed yet, do it now */
-	if (pred_info == NULL || pred_info->processed == 0) {
-		belady(pred_block, env);
-		pred_info = get_block_info(pred_block);
-	}
+				if (l->node != value)
+					continue;
 
-	/* 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, &env->ob, pred_info->ws_end);
+				if (l->spilled) {
+					spilled = true;
+				}
+				break;
+			}
+		}
 
-	return res;
+		loc->spilled = spilled;
+	}
 }
 
+#if 0
+static void decide_start_workset2(const ir_node *block)
+{
+	int         arity;
+	workset_t **pred_worksets;
+	int         p;
+	int         i;
+	int         len;
+
+	/* check if all predecessors are known */
+	arity           = get_irn_arity(block);
+	pred_worksets   = alloca(sizeof(pred_worksets[0]) * arity);
+	for (i = 0; i < arity; ++i) {
+		ir_node      *pred_block = get_Block_cfgpred_block(block, i);
+		block_info_t *pred_info  = get_block_info(pred_block);
+
+		if (pred_info == NULL) {
+			/* not all predecessors known, use decide_start_workset */
+			decide_start_workset(block);
+			return;
+		}
+
+		pred_worksets[i] = pred_info->end_workset;
+	}
+
+	/* we construct a new workset */
+	workset_clear(ws);
+
+	/* take values live in all pred blocks */
+	len = workset_get_length(pred_worksets[0]);
+	for (p = 0; p < len; ++p) {
+		const loc_t *l = &pred_worksets[0]->vals[p];
+		ir_node     *value;
+		bool         spilled = false;
+
+		if (USES_IS_INFINITE(l->time))
+			continue;
+
+		/* value available in all preds? */
+		value = l->node;
+		for (i = 1; i < arity; ++i) {
+			bool       found     = false;
+			workset_t *p_workset = pred_worksets[i];
+			int        p_len     = workset_get_length(p_workset);
+			int        p_i;
+
+			for (p_i = 0; p_i < p_len; ++p_i) {
+				const loc_t *p_l = &p_workset->vals[p_i];
+				if (p_l->node != value)
+					continue;
+
+				found = true;
+				if (p_l->spilled)
+					spilled = true;
+				break;
+			}
+
+			if (!found)
+				break;
+		}
+
+		/* it was available in all preds */
+		if (i >= arity) {
+			workset_insert(ws, value, spilled);
+		}
+	}
+
+	/* Copy the best ones from starters to start workset */
+	ws_count = MIN(ARR_LEN(starters), n_regs);
+	workset_bulk_fill(ws, ws_count, starters);
+}
+#endif
 
 /**
- * For the given block @p blk, decide for each values
+ * For the given block @p block, decide for each values
  * whether it is used from a register or is reloaded
  * before the use.
  */
-static void belady(ir_node *block, void *data) {
-	belady_env_t *env = data;
-	workset_t *new_vals;
-	ir_node *irn;
-	int iter;
-	block_info_t *block_info;
-
-	/* make sure we have blockinfo (with startset) */
-	block_info = get_block_info(block);
-	if (block_info == NULL)
-		block_info = compute_block_start_info(env, block);
-
-	/* Don't do a block twice */
-	if(block_info->processed)
+static void belady(ir_node *block)
+{
+	workset_t       *new_vals;
+	ir_node         *irn;
+	int              iter;
+	block_info_t    *block_info;
+	int              i, arity;
+	int              has_backedges = 0;
+	//int              first         = 0;
+	const ir_edge_t *edge;
+
+	/* no need to process a block twice */
+	if (get_block_info(block) != NULL) {
 		return;
+	}
+
+	/* check if all predecessor blocks are processed yet (though for backedges
+	 * we have to make an exception as we can't process them first) */
+	arity = get_Block_n_cfgpreds(block);
+	for(i = 0; i < arity; ++i) {
+		ir_node      *pred_block = get_Block_cfgpred_block(block, i);
+		block_info_t *pred_info  = get_block_info(pred_block);
+
+		if (pred_info == NULL) {
+			/* process predecessor first (it will be in the queue already) */
+			if (!is_backedge(block, i)) {
+				return;
+			}
+			has_backedges = 1;
+		}
+	}
+	(void) has_backedges;
+	if (arity == 0) {
+		workset_clear(ws);
+	} else if (arity == 1) {
+		ir_node      *pred_block = get_Block_cfgpred_block(block, 0);
+		block_info_t *pred_info  = get_block_info(pred_block);
+
+		assert(pred_info != NULL);
+		workset_copy(ws, pred_info->end_workset);
+	} else {
+		/* we need 2 heuristics here, for the case when all predecessor blocks
+		 * are known and when some are backedges (and therefore can't be known
+		 * yet) */
+		decide_start_workset(block);
+	}
 
-	/* get the starting workset for this block */
-	DBG((dbg, DBG_DECIDE, "\n"));
-	DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
+	DB((dbg, DBG_DECIDE, "\n"));
+	DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 
-	workset_copy(env, env->ws, block_info->ws_start);
-	DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
-	workset_foreach(env->ws, irn, iter)
-		DBG((dbg, DBG_WSETS, "  %+F\n", irn));
+	block_info = new_block_info();
+	set_block_info(block, block_info);
+
+	DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
+	workset_foreach(ws, irn, iter) {
+		DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
+		     workset_get_time(ws, iter)));
+	}
+
+	block_info->start_workset = workset_clone(ws);
 
 	/* process the block from start to end */
-	DBG((dbg, DBG_WSETS, "Processing...\n"));
-	env->used = pset_new_ptr_default();
-	env->instr_nr = 0;
-	new_vals = new_workset(env, &env->ob);
+	DB((dbg, DBG_WSETS, "Processing...\n"));
+	ir_nodeset_init(&used);
+	instr_nr = 0;
+	/* TODO: this leaks (into the obstack)... */
+	new_vals = new_workset();
+
 	sched_foreach(block, irn) {
 		int i, arity;
-		assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
+		assert(workset_get_length(ws) <= n_regs);
 
-		/* projs are handled with the tuple value.
-		 * Phis are no real instr (see insert_starters())
-		 * instr_nr does not increase */
-		if (is_Proj(irn) || is_Phi(irn)) {
-			DBG((dbg, DBG_DECIDE, "  ...%+F skipped\n", irn));
+		/* Phis are no real instr (see insert_starters()) */
+		if (is_Phi(irn)) {
 			continue;
 		}
-		DBG((dbg, DBG_DECIDE, "  ...%+F\n", irn));
+		DB((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 
 		/* set instruction in the workset */
-		env->instr = irn;
+		instr = irn;
 
 		/* allocate all values _used_ by this instruction */
 		workset_clear(new_vals);
 		for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
-			workset_insert(env, new_vals, get_irn_n(irn, i));
+			ir_node *in = get_irn_n(irn, i);
+			if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in))
+				continue;
+
+			/* (note that "spilled" is irrelevant here) */
+			workset_insert(new_vals, in, false);
 		}
-		displace(env, new_vals, 1);
+		displace(new_vals, 1);
 
 		/* allocate all values _defined_ by this instruction */
 		workset_clear(new_vals);
-		if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
-			ir_node *proj;
-			for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
-				workset_insert(env, new_vals, proj);
+		if (get_irn_mode(irn) == mode_T) {
+			const ir_edge_t *edge;
+
+			foreach_out_edge(irn, edge) {
+				ir_node *proj = get_edge_src_irn(edge);
+				if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
+					continue;
+				workset_insert(new_vals, proj, false);
+			}
 		} else {
-			workset_insert(env, new_vals, irn);
+			if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
+				continue;
+			workset_insert(new_vals, irn, false);
 		}
-		displace(env, new_vals, 0);
+		displace(new_vals, 0);
 
-		env->instr_nr++;
+		instr_nr++;
 	}
-	del_pset(env->used);
+	ir_nodeset_destroy(&used);
 
 	/* Remember end-workset for this block */
-	block_info->ws_end = workset_clone(env, &env->ob, env->ws);
-	block_info->processed = 1;
-	DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
-	workset_foreach(block_info->ws_end, irn, iter)
-		DBG((dbg, DBG_WSETS, "  %+F\n", irn));
+	block_info->end_workset = workset_clone(ws);
+	DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
+	workset_foreach(ws, irn, iter)
+		DB((dbg, DBG_WSETS, "  %+F (%u)\n", irn,
+		     workset_get_time(ws, iter)));
+
+	/* add successor blocks into worklist */
+	foreach_block_succ(block, edge) {
+		ir_node *succ = get_edge_src_irn(edge);
+		pdeq_putr(worklist, succ);
+	}
 }
 
 /**
@@ -563,97 +950,162 @@ static void belady(ir_node *block, void *data) {
  * 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 *data) {
-	belady_env_t *env = data;
-	workset_t *wsb;
-	int i, max, iter, iter2;
+static void fix_block_borders(ir_node *block, void *data)
+{
+	workset_t    *start_workset;
+	int           arity;
+	int           i;
+	int           iter;
+	(void) data;
 
-	DBG((dbg, DBG_FIX, "\n"));
-	DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
+	DB((dbg, DBG_FIX, "\n"));
+	DB((dbg, DBG_FIX, "Fixing %+F\n", block));
 
-	wsb = get_block_info(blk)->ws_start;
+	start_workset = get_block_info(block)->start_workset;
 
 	/* 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 = get_block_info(pred)->ws_end;
-
-		DBG((dbg, DBG_FIX, "  Pred %+F\n", pred));
-
-		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);
-
-				// we might have unknowns as argument for the phi
-				if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
-					continue;
+	arity = get_irn_arity(block);
+	for (i = 0; i < arity; ++i) {
+		ir_node   *pred = get_Block_cfgpred_block(block, i);
+		workset_t *pred_end_workset = get_block_info(pred)->end_workset;
+		ir_node   *node;
+
+		DB((dbg, DBG_FIX, "  Pred %+F\n", pred));
+
+		/* spill all values not used anymore */
+		workset_foreach(pred_end_workset, node, iter) {
+			ir_node *n2;
+			int      iter2;
+			bool     found = false;
+			workset_foreach(start_workset, n2, iter2) {
+				if (n2 == node) {
+					found = true;
+					break;
+				}
+				/* note that we do not look at phi inputs, becuase the values
+				 * will be either live-end and need no spill or
+				 * they have other users in which must be somewhere else in the
+				 * workset */
 			}
 
-			/* Unknowns are available everywhere */
-			if(get_irn_opcode(irnb) == iro_Unknown)
+			if (found)
 				continue;
 
-			/* check if irnb is in a register at end of pred */
-			workset_foreach(wsp, irnp, iter2) {
-				if (irnb == irnp)
-					goto next_value;
+#ifdef PLACE_SPILLS
+			if (be_is_live_in(lv, block, node)
+					&& !pred_end_workset->vals[iter].spilled) {
+				ir_node *insert_point;
+				if (arity > 1) {
+					insert_point = be_get_end_of_block_insertion_point(pred);
+					insert_point = sched_prev(insert_point);
+				} else {
+					insert_point = block;
+				}
+				DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
+				     insert_point));
+				be_add_spill(senv, node, insert_point);
 			}
+#endif
+		}
 
-			/* irnb is not in memory at the end of pred, so we have to reload it */
-			DBG((dbg, DBG_FIX, "    reload %+F\n", irnb));
-			be_add_reload_on_edge(env->senv, irnb, blk, i);
+		/* reload missing values in predecessors, add missing spills */
+		workset_foreach(start_workset, node, iter) {
+			const loc_t *l    = &start_workset->vals[iter];
+			const loc_t *pred_loc;
 
-next_value:
-			/*epsilon statement :)*/;
+			/* if node is a phi of the current block we reload
+			 * the corresponding argument, else node itself */
+			if (is_Phi(node) && get_nodes_block(node) == block) {
+				node = get_irn_n(node, i);
+				assert(!l->spilled);
+
+				/* we might have unknowns as argument for the phi */
+				if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
+					continue;
+			}
+
+			/* check if node is in a register at end of pred */
+			pred_loc = workset_contains(pred_end_workset, node);
+			if (pred_loc != NULL) {
+#ifdef PLACE_SPILLS
+				/* we might have to spill value on this path */
+				if (!pred_loc->spilled && l->spilled) {
+					ir_node *insert_point
+						= be_get_end_of_block_insertion_point(pred);
+					insert_point = sched_prev(insert_point);
+					DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
+					    insert_point));
+					be_add_spill(senv, node, insert_point);
+				}
+#endif
+			} else {
+				/* node is not in register at the end of pred -> reload it */
+				DB((dbg, DBG_FIX, "    reload %+F\n", node));
+				DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
+				be_add_reload_on_edge(senv, node, block, i, cls, 1);
+			}
 		}
 	}
 }
 
-void be_spill_belady(const be_chordal_env_t *chordal_env) {
-	be_spill_belady_spill_env(chordal_env, NULL);
-}
+static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
+{
+	ir_graph *irg = be_get_birg_irg(birg);
 
-void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
-	belady_env_t env;
+	be_liveness_assure_sets(be_assure_liveness(birg));
 
-	FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
-	//firm_dbg_set_mask(dbg, DBG_WSETS);
+	/* construct control flow loop tree */
+	if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
+		construct_cf_backedges(irg);
+	}
+
+	be_clear_links(irg);
 
 	/* init belady env */
-	obstack_init(&env.ob);
-	env.cenv      = chordal_env;
-	env.arch      = chordal_env->birg->main_env->arch_env;
-	env.cls       = chordal_env->cls;
-	env.n_regs    = arch_count_non_ignore_regs(env.arch, env.cls);
-	env.ws        = new_workset(&env, &env.ob);
-	env.uses      = be_begin_uses(chordal_env->irg, chordal_env->lv, chordal_env->birg->main_env->arch_env, env.cls);
-	if(spill_env == NULL) {
-		env.senv = be_new_spill_env(chordal_env);
-	} else {
-		env.senv = spill_env;
+	obstack_init(&obst);
+	arch_env = birg->main_env->arch_env;
+	cls      = rcls;
+	lv       = be_get_birg_liveness(birg);
+	n_regs   = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
+	ws       = new_workset();
+	uses     = be_begin_uses(irg, lv);
+	loop_ana = be_new_loop_pressure(birg);
+	senv     = be_new_spill_env(birg);
+	worklist = new_pdeq();
+
+	pdeq_putr(worklist, get_irg_start_block(irg));
+
+	while(!pdeq_empty(worklist)) {
+		ir_node *block = pdeq_getl(worklist);
+		belady(block);
 	}
-	DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
+	/* end block might not be reachable in endless loops */
+	belady(get_irg_end_block(irg));
 
-	DBG((dbg, LEVEL_1, "running on register class: %s\n", env.cls->name));
+	del_pdeq(worklist);
 
-	be_clear_links(chordal_env->irg);
-	/* Decide which phi nodes will be spilled and place copies for them into the graph */
-	irg_block_walk_graph(chordal_env->irg, spill_phi_walker, NULL, &env);
-	/* Fix high register pressure with belady algorithm */
-	irg_block_walk_graph(chordal_env->irg, NULL, belady, &env);
-	/* belady was block-local, fix the global flow by adding reloads on the edges */
-	irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &env);
-	/* Insert spill/reload nodes into the graph and fix usages */
-	be_insert_spills_reloads(env.senv);
+	/* belady was block-local, fix the global flow by adding reloads on the
+	 * edges */
+	irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 
-	be_remove_dead_nodes_from_schedule(chordal_env->irg);
-	be_liveness_recompute(chordal_env->lv);
+	/* Insert spill/reload nodes into the graph and fix usages */
+	be_insert_spills_reloads(senv);
 
 	/* clean up */
-	if(spill_env == NULL)
-		be_delete_spill_env(env.senv);
-	be_end_uses(env.uses);
-	obstack_free(&env.ob, NULL);
+	be_delete_spill_env(senv);
+	be_end_uses(uses);
+	be_free_loop_pressure(loop_ana);
+	obstack_free(&obst, NULL);
 }
+
+void be_init_spillbelady(void)
+{
+	static be_spiller_t belady_spiller = {
+		be_spill_belady
+	};
+
+	be_register_spiller("belady", &belady_spiller);
+	FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);