X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=c5e58b727dc4f628439b472e7fd9eed32e22ba14;hb=205396c4f4f5abe7abd6dc2350c8c398a7623afc;hp=0447d0398cebe2dd52d796558075c076044bccef;hpb=b80dc0bbc9697becc681c9119df8589d1ea39c9f;p=libfirm

diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c
index 0447d0398..c5e58b727 100644
--- a/ir/be/bespillbelady.c
+++ b/ir/be/bespillbelady.c
@@ -1,6 +1,6 @@
 /**
- * Author:      Daniel Grund
- * Date:		20.09.2005
+ * Author:      Daniel Grund, Matthias Braun
+ * Date:        20.09.2005
  * Copyright:   (c) Universitaet Karlsruhe
  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
  *
@@ -9,26 +9,22 @@
 #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 "irgwalk.h"
+#include "irloop.h"
 #include "iredges_t.h"
 #include "ircons_t.h"
+#include "irprintf.h"
+#include "xmalloc.h"
 
 #include "beutil.h"
-#include "bearch.h"
+#include "bearch_t.h"
 #include "bespillbelady.h"
 #include "beuses_t.h"
 #include "besched_t.h"
@@ -36,6 +32,8 @@
 #include "belive_t.h"
 #include "benode_t.h"
 #include "bechordal_t.h"
+#include "bespilloptions.h"
+#include "beloopana.h"
 
 #define DBG_SPILL   1
 #define DBG_WSETS   2
@@ -44,54 +42,72 @@
 #define DBG_START  16
 #define DBG_SLOTS  32
 #define DBG_TRACE  64
-#define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
-static firm_dbg_module_t *dbg = NULL;
+#define DBG_WORKSET 128
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
-#define MIN(a,b) (((a)<(b))?(a):(b))
+/**
+ * An association between a node and a point in time.
+ */
+typedef struct _loc_t {
+  ir_node *irn;        /**< A node. */
+  unsigned time;       /**< A use time (see beuses.h). */
+} loc_t;
 
-typedef struct _workset_t workset_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;
 
 typedef struct _belady_env_t {
 	struct obstack ob;
 	const arch_env_t *arch;
 	const arch_register_class_t *cls;
+	const be_lv_t *lv;
+	be_loopana_t *loop_ana;
 	int n_regs;			/** number of regs in this reg-class */
 
 	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;			/**< holds the values used (so far) in the current BB */
-	pset *copies;		/**< holds all copies placed due to phi-spilling */
+	pset *used;
 
-	spill_env_t *senv;	/* see bespill.h */
-	pset *reloads;		/**< all reload nodes placed */
+	spill_env_t *senv;	/**< see bespill.h */
 } belady_env_t;
 
-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 loc_compare(const void *a, const void *b)
+{
+	const loc_t *p = a;
+	const loc_t *q = b;
+	return p->time - q->time;
+}
+
+static INLINE 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);
+	}
+}
 
 /**
  * Alloc a new workset on obstack @p ob with maximum size @p max
  */
-static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
+static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) {
 	workset_t *res;
-	size_t size = sizeof(*res) + (bel->n_regs-1)*sizeof(res->vals[0]);
+	size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
 	res = obstack_alloc(ob, size);
 	memset(res, 0, size);
-	res->bel = 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) {
+static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) {
 	workset_t *res;
-	size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
+	size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
 	res = obstack_alloc(ob, size);
 	memcpy(res, ws, size);
 	return res;
@@ -99,10 +115,10 @@ static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
 
 /**
  * Do NOT alloc anything. Make @param tgt equal to @param src.
- * returns @param tgt for convinience
+ * returns @param tgt for convenience
  */
-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]);
+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;
 }
@@ -112,18 +128,20 @@ static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
  * @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]));
-
+static INLINE 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]));
+}
 
 /**
  * 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) {
+static INLINE void workset_insert(belady_env_t *env, workset_t *ws, ir_node *val) {
 	int i;
 	/* check for current regclass */
-	if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
-		DBG((dbg, DBG_DECIDE, "Dropped %+F\n", val));
+	if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
+		DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
 		return;
 	}
 
@@ -133,66 +151,37 @@ static INLINE void workset_insert(workset_t *ws, ir_node *val) {
 			return;
 
 	/* insert val */
-	assert(ws->len < ws->bel->n_regs && "Workset already full!");
+	assert(ws->len < env->n_regs && "Workset already full!");
 	ws->vals[ws->len++].irn = val;
 }
 
-/**
- * Inserts all values in array @p vals of length @p cnt
- * into the workset. There must be enough space for the
- * entries.
- */
-static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
-	int i, o;
-
-	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, -1) != ws->bel->cls) {
-			DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
-			goto no_insert;
-		}
-
-		/* check if val is already contained */
-		for(i=0; i<ws->len; ++i)
-			if (ws->vals[i].irn == val) {
-				DBG((dbg, DBG_TRACE, "Already contained\n"));
-				goto no_insert;
-			}
-
-		/* 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"));
-
-no_insert:
-		/*epsilon statement :)*/;
-	}
-}
-
 /**
  * Removes all entries from this workset
  */
-#define workset_clear(ws) (ws)->len = 0;
+static INLINE void workset_clear(workset_t *ws) {
+	ws->len = 0;
+}
 
 /**
  * Removes the value @p val from the workset if present.
  */
 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
 	int i;
-	for(i=0; i<ws->len; ++i)
+	for(i=0; i<ws->len; ++i) {
 		if (ws->vals[i].irn == val) {
 			ws->vals[i] = ws->vals[--ws->len];
 			return;
 		}
+	}
 }
 
 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
 	int i;
-	for(i=0; i<ws->len; ++i)
+	for(i=0; i<ws->len; ++i) {
 		if (ws->vals[i].irn == val)
 			return 1;
+	}
+
 	return 0;
 }
 
@@ -207,6 +196,7 @@ static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
 										++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)
@@ -214,6 +204,7 @@ static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
 
 typedef struct _block_info_t {
 	workset_t *ws_start, *ws_end;
+	int processed;
 } block_info_t;
 
 
@@ -221,33 +212,37 @@ 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;
 
 	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)
-
-static int is_mem_phi(const ir_node *irn, void *data) {
-	workset_t *sws;
-	ir_node *blk = get_nodes_block(irn);
-
-	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);
-}
+#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 is an ignore node
+ * @return The distance to the next use or 0 if irn has dont_spill flag set
  */
-#define get_distance(bel, from, from_step, def, skip_from_uses) \
-		((arch_irn_is_ignore(bel->arch, def) ) ? 0 : be_get_next_use(bel->uses, from, from_step, def, skip_from_uses))
+static INLINE unsigned get_distance(belady_env_t *env, ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
+{
+	be_next_use_t use;
+	int flags = arch_irn_get_flags(env->arch, def);
+
+	assert(! (flags & arch_irn_flags_ignore));
 
+	use = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
+	if(USES_IS_INFINITE(use.time))
+		return USES_INFINITY;
+
+	/* We have to keep nonspillable nodes in the workingset */
+	if(flags & arch_irn_flags_dont_spill)
+		return 0;
+
+	return use.time;
+}
 
 /**
- * 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
@@ -255,59 +250,76 @@ static int is_mem_phi(const ir_node *irn, void *data) {
  * @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 *bel, workset_t *new_vals, int is_usage) {
+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 = bel->ws;
-	ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert));
+	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
-	 */
+		1. Identify the number of needed slots and the values to reload
+	*/
 	demand = 0;
 	workset_foreach(new_vals, val, iter) {
 		/* mark value as used */
 		if (is_usage)
-			pset_insert_ptr(bel->used, val);
+			pset_insert_ptr(env->used, val);
 
-		if (!workset_contains(ws, val)) {
+		if (! workset_contains(ws, val)) {
 			DBG((dbg, DBG_DECIDE, "    insert %+F\n", val));
 			to_insert[demand++] = val;
-			if (is_usage)
-				be_add_reload(bel->senv, val, bel->instr);
-		} else
+			if (is_usage) {
+				DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr));
+				be_add_reload(env->senv, val, env->instr, env->cls, 1);
+			}
+		}
+		else {
+			assert(is_usage || "Defined value already in workset?!?");
 			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;
+		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));
 
 	/* 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(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
+		for (i = 0; i < ws->len; ++i) {
+			unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !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 */
-		for (i=max_allowed; i<ws->len; ++i) {
+		/*
+			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;
-			if (!pset_find_ptr(bel->used, irn)) {
-				ir_node *curr_bb = get_nodes_block(bel->instr);
+
+			DBG((dbg, DBG_DECIDE, "    disposing %+F (%u)\n", irn, workset_get_time(ws, i)));
+
+            if (is_Phi(irn))
+                continue;
+
+			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));
-			} else
-				DBG((dbg, DBG_DECIDE, "    dispose %+F\n", irn));
+				DBG((dbg, DBG_DECIDE, "    (and removing %+F from start workset)\n", irn));
+			}
 		}
 
 		/* kill the last 'demand' entries in the array */
@@ -315,161 +327,246 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
 	}
 
 	/*
-	 * 3. Insert the new values into the workset
-	 */
-	workset_bulk_insert(bel->ws, demand, to_insert);
+		3. Insert the new values into the workset
+	*/
+	for (i = 0; i < demand; ++i)
+		workset_insert(env, env->ws, to_insert[i]);
 }
 
-static void belady(ir_node *blk, void *env);
+static void belady(ir_node *block, void *env);
 
-/**
- * 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.
+/** 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 block    the block of the node
+ * @param loop     the loop of the node
  */
-static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
-	belady_env_t *bel = env;
-	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;
+static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first,
+		                            ir_node *node, ir_node *block,
+									ir_loop *loop)
+{
+	be_next_use_t next_use;
+	loc_t loc;
+	loc.time = USES_INFINITY;
+	loc.irn = node;
+
+	if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) {
+		loc.time = USES_INFINITY;
+		return loc;
+	}
 
-	/* Create the block info for this block. */
-	res = new_block_info(&bel->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(bel->arch, irn, -1) == bel->cls) {
-			loc.irn = irn;
-			loc.time = get_distance(bel, first, 0, irn, 0);
-			obstack_grow(&ob, &loc, sizeof(loc));
-			DBG((dbg, DBG_START, "    %+F:\n", irn));
-			count++;
-		} else
-			break;
+	/* We have to keep nonspillable nodes in the workingset */
+	if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) {
+		loc.time = 0;
+		DBG((dbg, DBG_START, "    %+F taken (dontspill node)\n", node, loc.time));
+		return loc;
+	}
 
-	live_foreach(blk, li)
-		if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) {
-			loc.irn = (ir_node *)li->irn;
-			loc.time = get_distance(bel, first, 0, li->irn, 0);
-			obstack_grow(&ob, &loc, sizeof(loc));
-			DBG((dbg, DBG_START, "    %+F:\n", irn));
-			count++;
+	next_use = be_get_next_use(env->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;
+		DBG((dbg, DBG_START, "    %+F not taken (dead)\n", node));
+		if(is_Phi(node)) {
+			be_spill_phi(env->senv, node);
 		}
+		return loc;
+	}
 
-	starters = obstack_finish(&ob);
-	qsort(starters, count, sizeof(starters[0]), loc_compare);
+	loc.time = next_use.time;
 
+	if(next_use.outermost_loop >= get_loop_depth(loop)) {
+		DBG((dbg, DBG_START, "    %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
+	} else {
+		loc.time = USES_PENDING;
+		DBG((dbg, DBG_START, "    %+F delayed (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
+	}
+	return loc;
+}
+
+/*
+ * Computes set of live-ins for each block with multiple predecessors
+ * and notifies spill algorithm which phis need to be spilled
+ */
+static void compute_live_ins(ir_node *block, void *data) {
+	belady_env_t  *env  = data;
+	ir_loop       *loop = get_irn_loop(block);
+	const be_lv_t *lv   = env->lv;
+	block_info_t  *block_info;
+	ir_node       *first, *irn;
+	loc_t         loc, *starters, *delayed;
+	int           i, len, ws_count;
+	int	          free_slots, free_pressure_slots;
+	unsigned      pressure;
+
+	if (get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
+		return;
 
-	/* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */
-	if (get_irn_arity(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);
+	block_info = new_block_info(&env->ob);
+	set_block_info(block, block_info);
+
+	/* Collect all values living at start of block */
+	starters = NEW_ARR_F(loc_t, 0);
+	delayed  = NEW_ARR_F(loc_t, 0);
+
+	DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
+	first = sched_first(block);
+
+	/* check all Phis first */
+	sched_foreach(block, irn) {
+		if (! is_Phi(irn))
+			break;
 
-		/* if pred block has not been processed yet, do it now */
-		if (! pred_info) {
-			belady(pred_blk, bel);
-			pred_info = get_block_info(pred_blk);
+		loc = to_take_or_not_to_take(env, first, irn, block, loop);
+
+		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);
 		}
+	}
+
+	/* check all Live-Ins */
+	be_lv_foreach(lv, block, be_lv_state_in, i) {
+		ir_node *node = be_lv_get_irn(lv, block, i);
+
+		loc = to_take_or_not_to_take(env, first, node, block, loop);
 
-		/* 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(&bel->ob, pred_info->ws_end);
+		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
+	pressure            = be_get_loop_pressure(env->loop_ana, env->cls, loop);
+	assert(ARR_LEN(delayed) <= pressure);
+	free_slots          = env->n_regs - ARR_LEN(starters);
+	free_pressure_slots = env->n_regs - (pressure - ARR_LEN(delayed));
+	free_slots          = MIN(free_slots, free_pressure_slots);
+	/* append nodes delayed due to loop structure until start set is full */
+	for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
+		DBG((dbg, DBG_START, "    delayed %+F taken\n", delayed[i].irn));
+		ARR_APP1(loc_t, starters, delayed[i]);
+		delayed[i].irn = NULL;
+	}
 
-	/* 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, bel->n_regs);
-		res->ws_start = new_workset(&bel->ob, bel);
-		workset_bulk_fill(res->ws_start, ws_count, starters);
+	/* spill all delayed phis which didn't make it into start workset */
+	for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
+		ir_node *irn = delayed[i].irn;
+		if (irn && is_Phi(irn)) {
+			DBG((dbg, DBG_START, "    spilling delayed phi %+F\n", irn));
+			be_spill_phi(env->senv, irn);
+		}
 	}
+	DEL_ARR_F(delayed);
 
+	/* Sort start values by first use */
+	qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
 
-	/* 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;
+	/* 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. */
+	len = ARR_LEN(starters);
+	for (i = ws_count; i < len; ++i) {
 		irn = starters[i].irn;
-		if (!is_Phi(irn) || get_nodes_block(irn) != blk)
+		if (! is_Phi(irn) || get_nodes_block(irn) != block)
 			continue;
 
-		DBG((dbg, DBG_START, "For %+F:\n", irn));
+		be_spill_phi(env->senv, irn);
+	}
 
-		for (max=get_irn_arity(irn), o=0; o<max; ++o) {
-			ir_node *arg = get_irn_n(irn, o);
-			ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
-			ir_node *cpy = be_new_Copy(bel->cls, irg, pred_block, arg);
-			pset_insert_ptr(bel->copies, cpy);
-			DBG((dbg, DBG_START, "    place a %+F of %+F in %+F\n", cpy, arg, pred_block));
-			sched_add_before(pred_block, cpy);
-			set_irn_n(irn, o, cpy);
-		}
+	DEL_ARR_F(starters);
+}
+
+/**
+ * Collects all values live-in at block @p block 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;
+
+	/* Have we seen this block before? */
+	res = get_block_info(block);
+	if (res)
+		return res;
+
+	/* Create the block info for this block. */
+	res = new_block_info(&env->ob);
+	set_block_info(block, res);
+
+	/* 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);
+
+	/* 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);
 	}
 
-	obstack_free(&ob, NULL);
+	/* 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);
+
 	return res;
 }
 
 
 /**
- * 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 *blk, void *env) {
-	belady_env_t *bel = env;
+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 *blk_info;
+	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 (get_block_info(blk))
+	if(block_info->processed)
 		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));
+	DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
 
-	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));
+	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 (%u)\n", irn, workset_get_time(env->ws, iter)));
 
 	/* process the block from start to end */
 	DBG((dbg, DBG_WSETS, "Processing...\n"));
-	bel->used = pset_new_ptr(32);
-	bel->instr_nr = 0;
-	new_vals = new_workset(&bel->ob, bel);
-	sched_foreach(blk, irn) {
-		assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!");
-
+	env->used = pset_new_ptr_default();
+	env->instr_nr = 0;
+	new_vals = new_workset(env, &env->ob);
+	sched_foreach(block, irn) {
+		int i, arity;
+		assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
 
 		/* 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)) {
 			DBG((dbg, DBG_DECIDE, "  ...%+F skipped\n", irn));
@@ -478,33 +575,36 @@ static void belady(ir_node *blk, void *env) {
 		DBG((dbg, DBG_DECIDE, "  ...%+F\n", irn));
 
 		/* set instruction in the workset */
-		bel->instr = irn;
+		env->instr = irn;
 
 		/* allocate all values _used_ by this instruction */
 		workset_clear(new_vals);
-		workset_bulk_insert(new_vals, get_irn_arity(irn)+1, get_irn_in(irn));
-		displace(bel, new_vals, 1);
+		for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
+			workset_insert(env, new_vals, get_irn_n(irn, i));
+		}
+		displace(env, 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(new_vals, proj);
+				workset_insert(env, new_vals, proj);
 		} else {
-			workset_insert(new_vals, irn);
+			workset_insert(env, new_vals, irn);
 		}
-		displace(bel, new_vals, 0);
+		displace(env, new_vals, 0);
 
-		bel->instr_nr++;
+		env->instr_nr++;
 	}
-	del_pset(bel->used);
+	del_pset(env->used);
 
 	/* Remember end-workset for this block */
-	blk_info->ws_end = workset_clone(&bel->ob, bel->ws);
-	DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk));
-	workset_foreach(blk_info->ws_end, irn, iter)
-		DBG((dbg, DBG_WSETS, "  %+F\n", irn));
+	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 (%u)\n", irn, workset_get_time(block_info->ws_end, iter)));
 }
 
 /**
@@ -512,19 +612,24 @@ static void belady(ir_node *blk, void *env) {
  * 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) {
+static void fix_block_borders(ir_node *block, void *data) {
+	belady_env_t *env = data;
 	workset_t *wsb;
-	belady_env_t *bel = env;
+        ir_graph *irg = get_irn_irg(block);
+        ir_node *startblock = get_irg_start_block(irg);
 	int i, max, iter, iter2;
 
+        if(block == startblock)
+            return;
+
 	DBG((dbg, DBG_FIX, "\n"));
-	DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
+	DBG((dbg, DBG_FIX, "Fixing %+F\n", block));
 
-	wsb = get_block_info(blk)->ws_start;
+	wsb = get_block_info(block)->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);
+	for (i=0, max=get_irn_arity(block); i<max; ++i) {
+		ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(block, i);
 		workset_t *wsp = get_block_info(pred)->ws_end;
 
 		DBG((dbg, DBG_FIX, "  Pred %+F\n", pred));
@@ -532,21 +637,28 @@ static void fix_block_borders(ir_node *blk, void *env) {
 		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))
+			if(is_Phi(irnb) && block == 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;
+			}
+
 			/* 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, iter2)
+			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 */
+			/* 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(bel->senv, irnb, blk, i);
+			DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", irnb, block, i));
+			be_add_reload_on_edge(env->senv, irnb, block, i, env->cls, 1);
 
 next_value:
 			/*epsilon statement :)*/;
@@ -555,91 +667,76 @@ next_value:
 }
 
 /**
- * Removes all used reloads from bel->reloads.
- * The remaining nodes in bel->reloads will be removed from the graph.
+ * Do spilling for a register class on a graph using the belady heuristic.
+ * In the transformed graph, the register pressure never exceeds the number
+ * of available registers.
+ *
+ * @param birg  The backend graph
+ * @param cls   The register class to spill
  */
-static void rescue_used_reloads(ir_node *irn, void *env) {
-	pset *rlds = (pset *)env;
-	if (pset_find_ptr(rlds, irn))
-		pset_remove_ptr(rlds, irn);
+static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) {
+	be_spill_belady_spill_env(birg, cls, NULL);
 }
 
-/**
- * Removes all copies introduced for phi-spills
- */
-static void remove_copies(belady_env_t *bel) {
-	ir_node *irn;
-
-	for (irn = pset_first(bel->copies); irn; irn = pset_next(bel->copies)) {
-		ir_node *src, *spill;
-
-		assert(be_is_Copy(irn));
-		assert(get_irn_n_edges(irn) == 1 && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
+void be_spill_belady_spill_env(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) {
+	belady_env_t env;
+	ir_graph *irg = be_get_birg_irg(birg);
+	int n_regs;
 
-		spill = get_irn_edge(get_irn_irg(irn), irn, 0)->src;
-		assert(be_is_Spill(spill) && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
+	/* some special classes contain only ignore regs, nothing to do then */
+	n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
+	if(n_regs == 0)
+		return;
 
-		src = get_irn_n(irn, 0);
-		set_irn_n(spill, 0, src);
+	be_invalidate_liveness(birg);
+	be_assure_liveness(birg);
+	/* construct control flow loop tree */
+	if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
+		construct_cf_backedges(irg);
 	}
-}
 
-/**
- * Finds all unused reloads and remove them from the schedule
- * Also removes spills if they are not used anymore after removing reloads
- */
-static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) {
-	ir_node *irn;
-
-	irg_walk_graph(irg, rescue_used_reloads, NULL, bel->reloads);
-	for(irn = pset_first(bel->reloads); irn; irn = pset_next(bel->reloads)) {
-		ir_node *spill;
-		DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn)));
-
-		spill = get_irn_n(irn, 0);
-
-		/* remove reload */
-		set_irn_n(irn, 0, new_Bad());
-		sched_remove(irn);
-
-		/* if spill not used anymore, remove it too
-		 * test of regclass is necessary since spill may be a phi-M */
-		if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) {
-			set_irn_n(spill, 0, new_Bad());
-			sched_remove(spill);
-		}
+	/* init belady env */
+	obstack_init(&env.ob);
+	env.arch      = birg->main_env->arch_env;
+	env.cls       = cls;
+	env.lv        = be_get_birg_liveness(birg);
+	env.n_regs    = n_regs;
+	env.ws        = new_workset(&env, &env.ob);
+	env.uses      = be_begin_uses(irg, env.lv);
+	env.loop_ana  = be_new_loop_pressure(birg);
+	if(spill_env == NULL) {
+		env.senv = be_new_spill_env(birg);
+	} else {
+		env.senv = spill_env;
 	}
-}
+	DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
+
+	be_clear_links(irg);
+	/* Decide which phi nodes will be spilled and place copies for them into the graph */
+	irg_block_walk_graph(irg, compute_live_ins, NULL, &env);
+	/* Fix high register pressure with belady algorithm */
+	irg_block_walk_graph(irg, NULL, belady, &env);
+	/* belady was block-local, fix the global flow by adding reloads on the edges */
+	irg_block_walk_graph(irg, fix_block_borders, NULL, &env);
+	/* Insert spill/reload nodes into the graph and fix usages */
+	be_insert_spills_reloads(env.senv);
 
-void be_spill_belady(const be_chordal_env_t *chordal_env) {
-	belady_env_t bel;
-
-	dbg = firm_dbg_register("ir.be.spillbelady");
+	/* clean up */
+	if(spill_env == NULL)
+		be_delete_spill_env(env.senv);
+	be_end_uses(env.uses);
+	be_free_loop_pressure(env.loop_ana);
+	obstack_free(&env.ob, NULL);
+}
 
-	/* init belady env */
-	obstack_init(&bel.ob);
-	bel.arch    = chordal_env->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->main_env->arch_env, bel.cls);
-	bel.senv    = be_new_spill_env(dbg, chordal_env, is_mem_phi, NULL);
-	bel.reloads = pset_new_ptr_default();
-	bel.copies  = pset_new_ptr_default();
-
-	DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
-
-	/* do the work */
-	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, bel.reloads);
-	remove_unused_reloads(chordal_env->irg, &bel);
-	remove_copies(&bel);
+void be_init_spillbelady(void)
+{
+	static be_spiller_t belady_spiller = {
+		be_spill_belady
+	};
 
-	/* clean up */
-	del_pset(bel.reloads);
-	be_delete_spill_env(bel.senv);
-	be_end_uses(bel.uses);
-	obstack_free(&bel.ob, NULL);
+	be_register_spiller("belady", &belady_spiller);
+	FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
 }
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);