X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;ds=sidebyside;f=ir%2Fbe%2Fbespillbelady3.c;h=1b4d40c7dc74fff5d97635ed9ed36137cbb05232;hb=fcb579b8959da1d7563b1a7b9f008a423ffdf75a;hp=32f9ef20d70b153df9203f2f1b8a5b513c3ccddc;hpb=cd038d3027364295ee3056eebc9929b55f4514ef;p=libfirm

diff --git a/ir/be/bespillbelady3.c b/ir/be/bespillbelady3.c
index 32f9ef20d..1b4d40c7d 100644
--- a/ir/be/bespillbelady3.c
+++ b/ir/be/bespillbelady3.c
@@ -25,14 +25,16 @@
  * @version     $Id$
  *
  * TODO:
- *   - handle phis correctly, decide wether we should spill them
- *   - merge multiple start worksets of blocks
+ *   - handle phis correctly, decide whether we should spill them ~ok
+ *   - merge multiple start worksets of blocks ~ok
  *   - how to and when to do the tentative phase...
  */
 #ifdef HAVE_CONFIG_H
 #include "config.h"
 #endif
 
+#include <stdbool.h>
+
 #include "debug.h"
 #include "list.h"
 #include "pdeq.h"
@@ -40,6 +42,8 @@
 #include "irnode_t.h"
 #include "irprintf.h"
 #include "iredges_t.h"
+#include "irloop_t.h"
+#include "irgwalk.h"
 #include "execfreq.h"
 
 #include "bemodule.h"
@@ -49,36 +53,68 @@
 #include "besched_t.h"
 #include "be_t.h"
 
+#ifndef NDEBUG
+#define EXPENSIVE_CHECKS
+#endif
+
 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
+typedef struct loop_edge_t loop_edge_t;
+struct loop_edge_t {
+	ir_node     *block;
+	int          pos;
+	loop_edge_t *next;
+};
+
+typedef struct loop_info_t loop_info_t;
+struct loop_info_t {
+	ir_loop     *loop;
+	loop_edge_t *entry_edges;
+	loop_edge_t *exit_edges;
+	size_t       max_register_pressure;
+};
+
 typedef struct worklist_entry_t worklist_entry_t;
 struct worklist_entry_t {
 	struct list_head  head;
 	ir_node          *value;
-	unsigned          timestep;
 	ir_node          *reload_point;
+	ir_loop          *unused_livethrough_loop;
 };
 
 typedef struct worklist_t worklist_t;
 struct worklist_t {
 	struct list_head  live_values;
 	size_t            n_live_values;
-	unsigned          current_timestep;
+	ir_visited_t      visited;
+};
+
+typedef struct block_info_t block_info_t;
+struct block_info_t {
+	worklist_t *start_worklist;
+	worklist_t *end_worklist;
 };
 
-static const arch_env_t            *arch_env;
 static const arch_register_class_t *cls;
 static struct obstack               obst;
 static spill_env_t                 *senv;
 static size_t                       n_regs;
-static int                          tentative_mode;
+static size_t                       max_register_pressure;
+static bool                         tentative_mode;
+static bool                         should_have_reached_fixpoint;
+static bool                         do_push_unused_livethroughs;
 static ir_exec_freq                *exec_freq;
+static ir_visited_t                 worklist_visited;
 
-static void init_worklist(worklist_t *worklist, unsigned timestep)
+static worklist_t *new_worklist(void)
 {
+	worklist_t *worklist = obstack_alloc(&obst, sizeof(worklist[0]));
+	memset(worklist, 0, sizeof(worklist[0]));
+
 	INIT_LIST_HEAD(&worklist->live_values);
 	worklist->n_live_values    = 0;
-	worklist->current_timestep = timestep;
+
+	return worklist;
 }
 
 static void mark_irn_not_visited(ir_node *node)
@@ -86,6 +122,23 @@ static void mark_irn_not_visited(ir_node *node)
 	set_irn_visited(node, get_irg_visited(current_ir_graph) - 1);
 }
 
+static bool worklist_contains(const ir_node *node)
+{
+	return irn_visited(node);
+}
+
+static block_info_t *get_block_info(ir_node *block)
+{
+	block_info_t *info = get_irn_link(block);
+	if (info != NULL)
+		return info;
+
+	info = obstack_alloc(&obst, sizeof(info[0]));
+	memset(info, 0, sizeof(info[0]));
+	set_irn_link(block, info);
+	return info;
+}
+
 static void deactivate_worklist(const worklist_t *worklist)
 {
 	struct list_head *entry;
@@ -93,7 +146,7 @@ static void deactivate_worklist(const worklist_t *worklist)
 	list_for_each(entry, &worklist->live_values) {
 		worklist_entry_t *wl_entry
 			= list_entry(entry, worklist_entry_t, head);
-		assert(irn_visited(wl_entry->value));
+		assert(worklist_contains(wl_entry->value));
 		mark_irn_not_visited(wl_entry->value);
 		set_irn_link(wl_entry->value, NULL);
 	}
@@ -106,51 +159,83 @@ static void activate_worklist(const worklist_t *worklist)
 	list_for_each(entry, &worklist->live_values) {
 		worklist_entry_t *wl_entry
 			= list_entry(entry, worklist_entry_t, head);
-		ir_node          *value = wl_entry->value;
-
-		assert(irn_not_visited(value));
-		mark_irn_visited(value);
-		set_irn_link(value, wl_entry);
+		assert(!worklist_contains(wl_entry->value));
+		mark_irn_visited(wl_entry->value);
+		set_irn_link(wl_entry->value, wl_entry);
 	}
 }
 
-static worklist_t *duplicate_worklist(const worklist_t *worklist,
-                                      ir_node *block,
-                                      ir_node *succ_block, int succ_pos)
+/**
+ * duplicate a worklist and directly activates it
+ */
+static void fill_and_activate_worklist(worklist_t *new_worklist,
+		const worklist_t *worklist, ir_node *block, ir_node *succ_block,
+		int succ_pos)
 {
-	ir_node          *reload_point = NULL;
+	ir_node          *reload_point  = NULL;
+	size_t            n_live_values = 0;
 	struct list_head *entry;
-	worklist_t       *new_worklist = obstack_alloc(&obst, sizeof(new_worklist[0]));
 
-	INIT_LIST_HEAD(&new_worklist->live_values);
-
-	if(succ_block != NULL && get_Block_n_cfgpreds(succ_block) > 1) {
+	if (succ_block != NULL &&
+			(get_Block_n_cfgpreds(succ_block) > 1
+			 || get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) > 1)) {
 		reload_point = be_get_end_of_block_insertion_point(block);
 	}
 
-	new_worklist->current_timestep = worklist->current_timestep;
-	new_worklist->n_live_values    = worklist->n_live_values;
-
 	list_for_each(entry, &worklist->live_values) {
-		worklist_entry_t *wl_entry
-			= list_entry(entry, worklist_entry_t, head);
-		worklist_entry_t *new_entry
-			= obstack_alloc(&obst, sizeof(new_entry[0]));
-		ir_node          *value = wl_entry->value;
+		worklist_entry_t *wl_entry  = list_entry(entry, worklist_entry_t, head);
+		ir_node          *value     = wl_entry->value;
+		worklist_entry_t *new_entry;
 
-		if(is_Phi(value) && get_nodes_block(value) == succ_block) {
+		if (new_worklist->n_live_values >= n_regs)
+			break;
+
+		if (is_Phi(value) && get_nodes_block(value) == succ_block) {
 			value = get_Phi_pred(value, succ_pos);
+
+			/* can happen for unknown phi preds */
+			if (!arch_irn_consider_in_reg_alloc(cls, value))
+				continue;
 		}
 
-		new_entry->value        = value;
-		new_entry->timestep     = wl_entry->timestep;
-		if(reload_point != NULL) {
+		if (irn_visited_else_mark(value))
+			continue;
+
+		new_entry = obstack_alloc(&obst, sizeof(new_entry[0]));
+		memset(new_entry, 0, sizeof(new_entry[0]));
+
+		new_entry->value = value;
+		if (reload_point != NULL) {
 			new_entry->reload_point = reload_point;
 		} else {
 			new_entry->reload_point = wl_entry->reload_point;
 		}
 
 		list_add_tail(&new_entry->head, &new_worklist->live_values);
+		++n_live_values;
+
+		set_irn_link(value, new_entry);
+		new_worklist->n_live_values++;
+	}
+}
+
+static worklist_t *duplicate_worklist(const worklist_t *worklist)
+{
+	worklist_t       *new_worklist;
+	struct list_head *entry;
+
+	new_worklist = obstack_alloc(&obst, sizeof(new_worklist[0]));
+	memset(new_worklist, 0, sizeof(new_worklist[0]));
+	INIT_LIST_HEAD(&new_worklist->live_values);
+	new_worklist->n_live_values = worklist->n_live_values;
+
+	list_for_each(entry, &worklist->live_values) {
+		worklist_entry_t *wl_entry  = list_entry(entry, worklist_entry_t, head);
+		worklist_entry_t *new_entry
+			= obstack_alloc(&obst, sizeof(new_entry[0]));
+
+		memcpy(new_entry, wl_entry, sizeof(new_entry[0]));
+		list_add_tail(&new_entry->head, &new_worklist->live_values);
 	}
 
 	return new_worklist;
@@ -161,66 +246,147 @@ static void print_worklist(const worklist_t *worklist, int level)
 {
 	struct list_head *entry;
 
-	DB((dbg, level, "%d values (TS %u): ", worklist->n_live_values,
-	    worklist->current_timestep));
+	DB((dbg, level, "%d values: ", worklist->n_live_values));
 	list_for_each(entry, &worklist->live_values) {
 		worklist_entry_t *wl_entry
 			= list_entry(entry, worklist_entry_t, head);
 
-		//DB((dbg, level, "%+F(%+F) ", wl_entry->value,
-		//    wl_entry->reload_point));
 		DB((dbg, level, "%+F ", wl_entry->value));
 	}
 }
 #endif
 
-static void place_reload(worklist_entry_t *entry)
+
+
+static void clear_loop_info(ir_loop *loop)
 {
-	if(tentative_mode)
-		return;
-	DB((dbg, LEVEL_1, "reload of %+F before %+F", entry->value,
-	    entry->reload_point));
-	be_add_reload(senv, entry->value, entry->reload_point, cls, 1);
+	int n_elements = get_loop_n_elements(loop);
+	int i;
+
+	loop->link = NULL;
+	for (i = 0; i < n_elements; ++i) {
+		loop_element element = get_loop_element(loop, i);
+		if (*element.kind != k_ir_loop)
+			continue;
+
+		clear_loop_info(element.son);
+	}
 }
 
-static void spill_non_live_nodes(const worklist_t *worklist)
+static loop_info_t *get_loop_info(ir_loop *loop)
 {
-	struct list_head *entry;
+	loop_info_t *info = loop->link;
+	if (info != NULL)
+		return info;
+
+	info = obstack_alloc(&obst, sizeof(info[0]));
+	memset(info, 0, sizeof(info[0]));
+	info->loop = loop;
+	loop->link = info;
+	return info;
+}
 
-	list_for_each(entry, &worklist->live_values) {
-		worklist_entry_t *wl_entry
-			= list_entry(entry, worklist_entry_t, head);
-		ir_node          *value = wl_entry->value;
+/**
+ * constructs loop in+out edges when called in a block walker
+ */
+static void construct_loop_edges(ir_node *block, void *data)
+{
+	int      n_cfgpreds = get_Block_n_cfgpreds(block);
+	ir_loop *loop       = get_irn_loop(block);
+	int      i;
+
+	(void) data;
+
+	for (i = 0; i < n_cfgpreds; ++i) {
+		ir_node     *cfgpred_block = get_Block_cfgpred_block(block, i);
+		ir_loop     *cfgpred_loop  = get_irn_loop(cfgpred_block);
+		loop_edge_t *edge;
+		bool        is_exit_edge;
+		ir_loop     *l, *goal;
 
-		if(irn_visited(value))
+		if (cfgpred_loop == loop)
 			continue;
 
-		place_reload(wl_entry);
+		/* critical edges are splitted, so we can't jump from 1 loop
+		 * directly into another without going out of it */
+		assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
+
+		/* edge out of loop */
+		if (get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
+			is_exit_edge = true;
+			l            = cfgpred_loop;
+			goal         = loop;
+		} else {
+			is_exit_edge = false;
+			l            = loop;
+			goal         = cfgpred_loop;
+		}
+
+		/* this might be a jump out/in multiple loops */
+		do {
+			loop_info_t *l_info = get_loop_info(l);
+
+			edge = obstack_alloc(&obst, sizeof(edge[0]));
+			memset(edge, 0, sizeof(edge[0]));
+			edge->block = block;
+			edge->pos   = i;
+
+			if (is_exit_edge) {
+				edge->next         = l_info->exit_edges;
+				l_info->exit_edges = edge;
+				assert(get_loop_depth(loop) < get_loop_depth(l));
+			} else {
+				edge->next          = l_info->entry_edges;
+				l_info->entry_edges = edge;
+				assert(get_loop_depth(cfgpred_loop) < get_loop_depth(l));
+			}
+			l = get_loop_outer_loop(l);
+		} while(l != goal);
 	}
 }
 
+
+
+
+static void place_reload(worklist_entry_t *entry)
+{
+	if (tentative_mode)
+		return;
+
+	DB((dbg, LEVEL_1, "reload of %+F before %+F\n", entry->value,
+		entry->reload_point));
+	assert(entry->reload_point != NULL);
+	be_add_reload(senv, entry->value, entry->reload_point, cls, 1);
+	entry->reload_point = NULL;
+}
+
 /**
  * makes sure the worklist contains not more than n_regs - room_needed entries
  */
 static void make_room(worklist_t *worklist, size_t room_needed)
 {
-	int spills_needed = worklist->n_live_values + room_needed - n_regs;
-	if(spills_needed > 0) {
-		int               i     = spills_needed;
-		struct list_head *entry = worklist->live_values.next;
-		for(i = spills_needed; i > 0; --i) {
-			struct list_head *next = entry->next;
-			worklist_entry_t *wl_entry
-				= list_entry(entry, worklist_entry_t, head);
-			assert(irn_visited(wl_entry->value));
-			mark_irn_not_visited(wl_entry->value);
-			place_reload(wl_entry);
-			list_del(entry);
+	int               i;
+	int               spills_needed;
+	struct list_head *entry;
 
-			entry = next;
-		}
-		worklist->n_live_values -= spills_needed;
+	spills_needed = worklist->n_live_values + room_needed - n_regs;
+	if (spills_needed <= 0)
+		return;
+
+	entry = worklist->live_values.next;
+	for(i = spills_needed; i > 0; --i) {
+		struct list_head *next = entry->next;
+		worklist_entry_t *wl_entry
+			= list_entry(entry, worklist_entry_t, head);
+
+		assert(worklist_contains(wl_entry->value));
+		mark_irn_not_visited(wl_entry->value);
+		place_reload(wl_entry);
+		list_del(entry);
+
+		entry = next;
 	}
+	worklist->n_live_values -= spills_needed;
 }
 
 /**
@@ -229,16 +395,20 @@ static void make_room(worklist_t *worklist, size_t room_needed)
  */
 static void val_used(worklist_t *worklist, ir_node *value, ir_node *sched_point)
 {
-	/* is the node in the worklist already? */
+	/* already in the worklist? move around, otherwise add at back */
 	worklist_entry_t *entry = get_irn_link(value);
-	if(irn_visited(value)) {
+
+	assert(arch_irn_consider_in_reg_alloc(cls, value));
+
+	if (worklist_contains(value)) {
 		assert(entry != NULL);
 
-		assert(irn_visited(value));
 		list_del(&entry->head);
 	} else {
-		if(entry == NULL) {
+		if (entry == NULL) {
 			entry        = obstack_alloc(&obst, sizeof(entry[0]));
+			memset(entry, 0, sizeof(entry[0]));
+
 			entry->value = value;
 			set_irn_link(value, entry);
 		}
@@ -247,7 +417,6 @@ static void val_used(worklist_t *worklist, ir_node *value, ir_node *sched_point)
 		mark_irn_visited(value);
 	}
 
-	entry->timestep     = worklist->current_timestep;
 	entry->reload_point = sched_point;
 	list_add_tail(&entry->head, &worklist->live_values);
 }
@@ -259,29 +428,30 @@ static void worklist_remove(worklist_t *worklist, ir_node *value)
 	list_del(&entry->head);
 	--worklist->n_live_values;
 
-	assert(irn_visited(value));
+	assert(worklist_contains(value));
 	mark_irn_not_visited(value);
 }
 
+static void update_max_pressure(worklist_t *worklist)
+{
+	if (worklist->n_live_values > max_register_pressure)
+		max_register_pressure = worklist->n_live_values;
+}
+
 static void do_spilling(ir_node *block, worklist_t *worklist)
 {
 	ir_node *node;
 
 	assert(worklist != NULL);
 
-#ifdef DEBUG_libfirm
-	DB((dbg, LEVEL_1, "worklist at end of %+F:", block));
-	print_worklist(worklist, LEVEL_1);
-	DB((dbg, LEVEL_1, "\n"));
-#endif
-
 	sched_foreach_reverse(block, node) {
 		int    i, arity;
 		size_t n_defs = 0;
 
 		DB((dbg, LEVEL_2, "\t%+F... ", node));
+		update_max_pressure(worklist);
 
-		if(is_Phi(node)) {
+		if (is_Phi(node)) {
 			ir_node *node2;
 			/* TODO: if we have some free registers, then we could decide to
 			 * not spill some phis (but not for phis where at least 1 input is
@@ -291,33 +461,35 @@ static void do_spilling(ir_node *block, worklist_t *worklist)
 			sched_foreach_reverse_from(node, node2) {
 				assert(is_Phi(node2));
 
-				if(irn_visited(node2))
+				if (worklist_contains(node2))
 					continue;
-				if(!arch_irn_consider_in_reg_alloc(arch_env, cls, node2))
+				if (!arch_irn_consider_in_reg_alloc(cls, node2))
 					continue;
 
-				be_spill_phi(senv, node2);
+				if (!tentative_mode)
+					be_spill_phi(senv, node2);
 			}
+			DB((dbg, LEVEL_2, "\n"));
 			break;
 		}
 
 		/* remove values defined by this instruction from the workset. Values
 		 * defined but not in the workset need free registers */
-		if(get_irn_mode(node) == mode_T) {
+		if (get_irn_mode(node) == mode_T) {
 			const ir_edge_t *edge;
 
 			foreach_out_edge(node, edge) {
 				ir_node *proj = get_edge_src_irn(edge);
-				if(!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
+				if (!arch_irn_consider_in_reg_alloc(cls, proj))
 					continue;
-				if(irn_visited(proj)) {
+				if (worklist_contains(proj)) {
 					worklist_remove(worklist, proj);
 				} else {
 					++n_defs;
 				}
 			}
-		} else if(arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
-			if(irn_visited(node)) {
+		} else if (arch_irn_consider_in_reg_alloc(cls, node)) {
+			if (worklist_contains(node)) {
 				worklist_remove(worklist, node);
 			} else {
 				n_defs = 1;
@@ -327,12 +499,12 @@ static void do_spilling(ir_node *block, worklist_t *worklist)
 		/* make sure we have enough free registers for the spills */
 		make_room(worklist, n_defs);
 
-		/* put all values by the instruction into the workset */
+		/* put all values used by the instruction into the workset */
 		arity = get_irn_arity(node);
 		for(i = 0; i < arity; ++i) {
 			ir_node *use = get_irn_n(node, i);
 
-			if(!arch_irn_consider_in_reg_alloc(arch_env, cls, use))
+			if (!arch_irn_consider_in_reg_alloc(cls, use))
 				continue;
 
 			val_used(worklist, use, node);
@@ -342,14 +514,14 @@ static void do_spilling(ir_node *block, worklist_t *worklist)
 		 * some */
 		make_room(worklist, 0);
 
-		++worklist->current_timestep;
-
 #ifdef DEBUG_libfirm
 		print_worklist(worklist, LEVEL_2);
 		DB((dbg, LEVEL_2, "\n"));
 #endif
 	}
 
+	update_max_pressure(worklist);
+
 #ifdef DEBUG_libfirm
 	DB((dbg, LEVEL_1, "worklist at begin of %+F:", block));
 	print_worklist(worklist, LEVEL_1);
@@ -357,95 +529,526 @@ static void do_spilling(ir_node *block, worklist_t *worklist)
 #endif
 }
 
-static void process_block(ir_node *block, void *env)
+static bool worklists_equal(const worklist_t *wl1, const worklist_t *wl2)
+{
+	const struct list_head *i1 = &wl1->live_values;
+	const struct list_head *i2 = &wl2->live_values;
+
+	for ( ; i1 != &wl1->live_values && i2 != &wl2->live_values;
+			i1 = i1->next, i2 = i2->next) {
+		worklist_entry_t *entry1 = list_entry(i1, worklist_entry_t, head);
+		worklist_entry_t *entry2 = list_entry(i2, worklist_entry_t, head);
+
+		if (entry1->value != entry2->value)
+			return false;
+	}
+	/* both lists at the end */
+	if (i1 != &wl1->live_values || i2 != &wl2->live_values)
+		return false;
+
+	return true;
+}
+
+static bool fill_start_worklist(worklist_t *new_worklist, ir_node *block)
 {
-	int              n_preds;
-	const ir_edge_t *edge;
-	worklist_t      *worklist        = NULL;
 	double           best_execfreq   = -1;
+	worklist_t      *best_worklist   = NULL;
 	ir_node         *best_succ_block = NULL;
-	int              best_pos        = -1;
+	int              best_pos;
+	const ir_edge_t *edge;
+
+	/* construct worklist */
+	foreach_block_succ(block, edge) {
+		ir_node      *succ_block = get_edge_src_irn(edge);
+		double       execfreq    = get_block_execfreq(exec_freq, succ_block);
+		block_info_t *block_info;
+		worklist_t   *succ_worklist;
+
+		if (execfreq < best_execfreq)
+			continue;
+
+		block_info    = get_block_info(succ_block);
+		succ_worklist = block_info->start_worklist;
+
+		if (succ_worklist == NULL || succ_worklist->visited >= worklist_visited)
+			continue;
+
+		best_execfreq   = execfreq;
+		best_worklist   = succ_worklist;
+		best_succ_block = succ_block;
+		best_pos        = get_edge_src_pos(edge);
+	}
+
+	if (best_worklist == NULL)
+		return false;
+	best_worklist->visited = worklist_visited;
+
+	fill_and_activate_worklist(new_worklist, best_worklist, block,
+			best_succ_block, best_pos);
+	return true;
+}
+
+static worklist_t *construct_start_worklist(ir_node *block)
+{
+	worklist_t *worklist = new_worklist();
+
+	++worklist_visited;
+
+	while(fill_start_worklist(worklist, block)) {
+		if (worklist->n_live_values >= n_regs)
+			break;
+	}
+
+	return worklist;
+}
+
+static void process_block(ir_node *block, void *env)
+{
+	block_info_t    *block_info;
+	worklist_t      *worklist;
+	int              n_preds;
 
 	(void) env;
 	DB((dbg, LEVEL_1, "Processing %+F\n", block));
 
-	/* construct worklist */
+	worklist   = construct_start_worklist(block);
+	block_info = get_block_info(block);
+
+#ifdef DEBUG_libfirm
+	DB((dbg, LEVEL_1, "worklist at end of %+F:", block));
+	print_worklist(worklist, LEVEL_1);
+	DB((dbg, LEVEL_1, "\n"));
+
+	if (should_have_reached_fixpoint) {
+		assert(worklists_equal(block_info->end_worklist, worklist));
+	}
+#endif
+	block_info->end_worklist = duplicate_worklist(worklist);
+
+	do_spilling(block, worklist);
+	deactivate_worklist(worklist);
+
+#ifdef DEBUG_libfirm
+	if (should_have_reached_fixpoint) {
+		assert(worklists_equal(block_info->start_worklist, worklist));
+	}
+#endif
+	block_info->start_worklist = worklist;
+
+	/* we shouldn't have any live values left at the start block */
+	n_preds = get_Block_n_cfgpreds(block);
+	//assert(n_preds != 0 || worklist->n_live_values == 0);
+}
+
+typedef struct block_or_loop_t block_or_loop_t;
+struct block_or_loop_t {
+	union {
+		ir_node *block;
+		ir_loop *loop;
+	} v;
+	bool is_loop;
+};
+
+static block_or_loop_t *loop_blocks;
+static ir_loop         *current_loop;
+
+static void find_blocks(ir_node *block);
+
+static void find_in_loop(ir_loop *loop, ir_node *entry)
+{
+	loop_info_t     *loop_info = get_loop_info(loop);
+	block_or_loop_t block_or_loop;
+	loop_edge_t     *edge;
+
+	/* simply mark 1 block in the loop to indicate that the loop was already
+	 * processed */
+	ir_node *some_block = loop_info->entry_edges->block;
+	if (Block_block_visited(some_block))
+		return;
+
+	block_or_loop.v.loop  = loop;
+	block_or_loop.is_loop = true;
+	ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
+
+#ifndef NDEBUG
+	{
+		/* we should be 1 of the entry blocks */
+		loop_edge_t *edge  = loop_info->entry_edges;
+		bool         found = false;
+		for ( ; edge != NULL; edge = edge->next) {
+			if (edge->block == entry)
+				found = true;
+		}
+		assert(found);
+	}
+#else
+	(void) entry;
+#endif
+
+	/* check all loop successors */
+	for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
+		ir_node *succ      = edge->block;
+		ir_loop *succ_loop = get_irn_loop(succ);
+
+		if (succ_loop == current_loop) {
+			find_blocks(succ);
+		} else {
+			assert(get_loop_depth(succ_loop) < get_loop_depth(current_loop));
+		}
+	}
+}
+
+static void find_blocks(ir_node *block)
+{
+	const ir_edge_t *edge;
+	block_or_loop_t block_or_loop;
+
+	if (Block_block_visited(block))
+		return;
+
+	block_or_loop.v.block = block;
+	block_or_loop.is_loop = false;
+
+	ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
+	mark_Block_block_visited(block);
+
 	foreach_block_succ(block, edge) {
-		ir_node *succ_block = get_edge_src_irn(edge);
-		double   execfreq   = get_block_execfreq(exec_freq, succ_block);
-
-		if(execfreq > best_execfreq) {
-			worklist_t *succ_worklist = get_irn_link(succ_block);
-			if(succ_worklist != NULL) {
-				best_execfreq   = execfreq;
-				worklist        = succ_worklist;
-				best_succ_block = succ_block;
-				best_pos        = get_edge_src_pos(edge);
+		ir_node *succ = get_edge_src_irn(edge);
+
+		/* is it part of the current loop? */
+		ir_loop *loop = get_irn_loop(succ);
+		if (loop != current_loop) {
+			/* a sub-loop? */
+			if (get_loop_depth(loop) > get_loop_depth(current_loop)) {
+				find_in_loop(loop, succ);
+			} else {
+				/* parent loop: we're not interested in the block */
 			}
+		} else {
+			find_blocks(succ);
 		}
 	}
-	if(worklist == NULL) {
-		/* only the end-block has 0 successors */
-		assert(block == get_irg_end_block(get_irn_irg(block)));
+}
 
-		worklist = obstack_alloc(&obst, sizeof(worklist[0]));
-		init_worklist(worklist, 0);
-	} else {
-		worklist = duplicate_worklist(worklist, block, best_succ_block,
-		                              best_pos);
-		activate_worklist(worklist);
-
-		/* now we could have live values in the succ worklists that are not
-		 * live anymore in the worklist we picked. We need reloads for them.
-		 */
-		if(!tentative_mode) {
-			foreach_block_succ(block, edge) {
-				ir_node      *succ_block    = get_edge_src_irn(edge);
-				worklist_t   *succ_worklist = get_irn_link(succ_block);
-
-				spill_non_live_nodes(succ_worklist);
+/**
+ * append an entry to a worklist. WARNING: The entry must not already be in the
+ * worklist.
+ */
+static void worklist_append(worklist_t *worklist, ir_node *value,
+                            ir_node *reload_point,
+							ir_loop *unused_livethrough_loop)
+{
+	worklist_entry_t *entry     = obstack_alloc(&obst, sizeof(entry[0]));
+	memset(entry, 0, sizeof(entry[0]));
+
+#ifdef EXPENSIVE_CHECKS
+	{
+		struct list_head *entry;
+		list_for_each(entry, &worklist->live_values) {
+			worklist_entry_t *wl_entry
+				= list_entry(entry, worklist_entry_t, head);
+			assert(wl_entry->value != value);
+		}
+	}
+#endif
+
+	entry->value                   = value;
+	entry->reload_point            = reload_point;
+	entry->unused_livethrough_loop = unused_livethrough_loop;
+	list_add_tail(&entry->head, &worklist->live_values);
+	++worklist->n_live_values;
+	assert(worklist->n_live_values <= n_regs);
+}
+
+static void push_unused_livethrough(loop_info_t *loop_info, ir_node *value)
+{
+	loop_edge_t *edge;
+	++worklist_visited;
+
+	/* add the value to all loop exit and entry blocks */
+	for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
+		ir_node            *block
+			= get_Block_cfgpred_block(edge->block, edge->pos);
+		const block_info_t *info     = get_block_info(block);
+		worklist_t         *worklist = info->end_worklist;
+		ir_node            *reload_point = NULL;
+
+		if (worklist->visited >= worklist_visited)
+			continue;
+		worklist->visited = worklist_visited;
+
+		/* TODO: we need a smarter mechanism here, that makes the reloader place
+		 * reload nodes on all loop exits... */
+
+		worklist_append(worklist, value, reload_point, loop_info->loop);
+	}
+	edge = loop_info->entry_edges;
+	for ( ; edge != NULL; edge = edge->next) {
+		ir_node            *entry_block = edge->block;
+		const block_info_t *info        = get_block_info(entry_block);
+		worklist_t         *worklist    = info->start_worklist;
+		ir_node            *pred_block;
+		ir_node            *reload_point;
+
+		if (worklist->visited >= worklist_visited)
+			continue;
+		worklist->visited = worklist_visited;
+
+		pred_block   = get_Block_cfgpred_block(entry_block, edge->pos);
+		reload_point = be_get_end_of_block_insertion_point(pred_block);
+
+		worklist_append(worklist, value, reload_point, loop_info->loop);
+	}
+
+	set_irn_link(value, NULL);
+	++loop_info->max_register_pressure;
+}
+
+static void push_unused_livethroughs(loop_info_t *loop_info)
+{
+	loop_edge_t *edge;
+
+	/* we can only push unused livethroughs if register pressure inside the loop
+	 * was low enough */
+	if (loop_info->max_register_pressure >= n_regs)
+		return;
+
+	/* find unused livethroughs: register pressure in the loop was low enough
+	 * which means that we had no spills which implies that at every point in
+	 * the loop all*/
+	for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
+		ir_node            *block          = edge->block;
+		const block_info_t *info           = get_block_info(block);
+		worklist_t         *start_worklist = info->start_worklist;
+		ir_node            *exit_block;
+		const block_info_t *exit_info;
+		worklist_t         *end_worklist;
+		struct list_head   *entry;
+
+		if (start_worklist == NULL)
+			continue;
+
+		exit_block   = get_Block_cfgpred_block(edge->block, edge->pos);
+		exit_info    = get_block_info(exit_block);
+		end_worklist = exit_info->end_worklist;
+
+		activate_worklist(end_worklist);
+		/* all values contained in the start_worklist, which are not available
+		 * in the end_worklist, must be unused livethroughs */
+
+		list_for_each(entry, &start_worklist->live_values) {
+			worklist_entry_t *wl_entry
+				= list_entry(entry, worklist_entry_t, head);
+			ir_node *value = wl_entry->value;
+
+			if (loop_info->max_register_pressure >= n_regs)
+				break;
+
+
+			if (!worklist_contains(value)) {
+				/* add it to all loop exits */
+				DB((dbg, LEVEL_2, "Found unused livethrough: %+F (regpressure in loop %d)\n", value, loop_info->max_register_pressure));
+				push_unused_livethrough(loop_info, value);
+				/* the value should now be part of end_worklist, so mark it */
+				mark_irn_visited(value);
 			}
 		}
+
+		deactivate_worklist(end_worklist);
 	}
+}
 
-	do_spilling(block, worklist);
-	deactivate_worklist(worklist);
+static void process_block_or_loop(const block_or_loop_t *block_or_loop)
+{
+	if (block_or_loop->is_loop) {
+		loop_info_t *loop_info = get_loop_info(block_or_loop->v.loop);
 
-	set_irn_link(block, worklist);
+		if (do_push_unused_livethroughs)
+			push_unused_livethroughs(loop_info);
 
-	/* we shouldn't have any live values left at the start block */
-	n_preds = get_Block_n_cfgpreds(block);
-	assert(n_preds != 0 || worklist->n_live_values == 0);
+		if (loop_info->max_register_pressure > max_register_pressure)
+			max_register_pressure = loop_info->max_register_pressure;
+
+		return;
+	}
+	process_block(block_or_loop->v.block, NULL);
+}
+
+static void process_loop(ir_loop *loop)
+{
+	int         n_elements = get_loop_n_elements(loop);
+	int         i, len;
+	loop_info_t *loop_info;
+	loop_edge_t *edge;
+	ir_node     *some_block;
+
+	/* first handle all sub-loops */
+	for (i = 0; i < n_elements; ++i) {
+		loop_element element = get_loop_element(loop, i);
+		if (*element.kind != k_ir_loop)
+			continue;
+
+		process_loop(element.son);
+	}
+
+	/* create a postorder of the blocks */
+	loop_info = get_loop_info(loop);
+	edge      = loop_info->entry_edges;
+	if (edge != NULL) {
+		some_block = edge->block;
+	} else {
+		assert(loop == get_irg_loop(current_ir_graph));
+		some_block = get_irg_start_block(current_ir_graph);
+	}
+
+	loop_blocks  = NEW_ARR_F(block_or_loop_t,0);
+	current_loop = loop;
+
+	ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
+	inc_irg_block_visited(current_ir_graph);
+	find_blocks(some_block);
+	/* for endless loops the end-block might be unreachable */
+	if (loop == get_irg_loop(current_ir_graph)) {
+		find_blocks(get_irg_end_block(current_ir_graph));
+	}
+	ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
+
+	DB((dbg, LEVEL_3, "Block List for loop %p\n", loop));
+	len = ARR_LEN(loop_blocks);
+	for (i = 0; i < len; ++i) {
+		block_or_loop_t *block_or_loop = &loop_blocks[i];
+		if (block_or_loop->is_loop) {
+			DB((dbg, LEVEL_3, " L-%p", block_or_loop->v.loop));
+		} else {
+			DB((dbg, LEVEL_3, " B-%+F", block_or_loop->v.block));
+		}
+	}
+	DB((dbg, LEVEL_3, "\n"));
+
+	max_register_pressure = 0;
+
+	/* run1: tentative phase */
+	tentative_mode = true;
+	for (i = len-1; i >= 0; --i) {
+		process_block_or_loop(&loop_blocks[i]);
+	}
+
+	/* run2: tentative phase - should reach fixpoint */
+	tentative_mode = true;
+	for (i = len-1; i >= 0; --i) {
+		process_block_or_loop(&loop_blocks[i]);
+	}
+
+#ifndef NDEBUG
+	/* run3: tentative phase - check fixpoint */
+	tentative_mode               = true;
+	should_have_reached_fixpoint = true;
+	for (i = len-1; i >= 0; --i) {
+		process_block_or_loop(&loop_blocks[i]);
+	}
+	should_have_reached_fixpoint = false;
+#endif
+
+	/* run4: add spills/reloads */
+	tentative_mode              = false;
+	do_push_unused_livethroughs = true;
+	for (i = len-1; i >= 0; --i) {
+		process_block_or_loop(&loop_blocks[i]);
+	}
+	do_push_unused_livethroughs = false;
+
+	loop_info->max_register_pressure = max_register_pressure;
+	DB((dbg, LEVEL_2, "Regpressure in loop %p: %u\n", loop,
+				(unsigned) max_register_pressure));
+
+	DEL_ARR_F(loop_blocks);
+}
+
+static void fix_block_borders(ir_node *block, void *data)
+{
+	block_info_t *block_info     = get_block_info(block);
+	worklist_t   *start_worklist = block_info->start_worklist;
+	int           n_cfgpreds     = get_Block_n_cfgpreds(block);
+	ir_loop      *loop           = get_irn_loop(block);
+	int           i;
+
+	(void) data;
+	assert(start_worklist != NULL);
+
+	for (i = 0; i < n_cfgpreds; ++i) {
+		ir_node      *pred_block      = get_Block_cfgpred_block(block, i);
+		block_info_t *pred_block_info = get_block_info(pred_block);
+		worklist_t   *end_worklist    = pred_block_info->end_worklist;
+		ir_loop      *pred_loop       = get_irn_loop(pred_block);
+		bool          is_loop_entry   = false;
+		struct list_head *entry;
+
+		assert(end_worklist != NULL);
+
+		if (get_loop_depth(pred_loop) < get_loop_depth(loop)) {
+			is_loop_entry = true;
+		}
+
+		/* reload missing values */
+		activate_worklist(end_worklist);
+
+		list_for_each(entry, &start_worklist->live_values) {
+			worklist_entry_t *wl_entry
+				= list_entry(entry, worklist_entry_t, head);
+			ir_node          *value = wl_entry->value;
+
+			if (is_Phi(value) && get_nodes_block(value) == block) {
+				value = get_irn_n(value, i);
+
+				/* we might have unknowns as argument for the phi */
+				if (!arch_irn_consider_in_reg_alloc(cls, value))
+					continue;
+			}
+
+			if (worklist_contains(value))
+				continue;
+			if (wl_entry->unused_livethrough_loop != NULL && !is_loop_entry)
+				continue;
+
+			be_add_reload_on_edge(senv, value, block, i, cls, 1);
+		}
+
+		deactivate_worklist(end_worklist);
+	}
 }
 
 static void be_spill_belady3(be_irg_t *birg, const arch_register_class_t *ncls)
 {
 	ir_graph *irg = be_get_birg_irg(birg);
 
-	cls       = ncls;
-	n_regs    = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
+	cls    = ncls;
+	n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
 
-	if(n_regs == 0)
+	/* shortcut for register classes with ignore regs only */
+	if (n_regs == 0)
 		return;
 
-	arch_env       = be_get_birg_arch_env(birg);
-	exec_freq      = be_get_birg_exec_freq(birg);
-	tentative_mode = 0;
+	worklist_visited = 0;
+	exec_freq        = be_get_birg_exec_freq(birg);
 
 	be_clear_links(irg);
-	set_using_irn_link(irg);
-	set_using_irn_visited(irg);
+	ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
+			| IR_RESOURCE_LOOP_LINK);
 	inc_irg_visited(irg);
 
 	obstack_init(&obst);
 	senv = be_new_spill_env(birg);
 
-	/* do a post-order walk over the CFG to make sure we have a maximum number
-	 * of preds processed before entering a block */
-	irg_block_edges_walk(get_irg_start_block(irg), NULL, process_block, NULL);
+	assure_cf_loop(irg);
+	clear_loop_info(get_irg_loop(irg));
+	irg_block_walk_graph(irg, construct_loop_edges, NULL, NULL);
+
+	process_loop(get_irg_loop(current_ir_graph));
+
+	irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
 
-	clear_using_irn_link(irg);
-	clear_using_irn_visited(irg);
+	ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
+			| IR_RESOURCE_LOOP_LINK);
 
 	be_insert_spills_reloads(senv);