- Split bearch.h correctly into bearch.h and bearch_t.h
[libfirm] / ir / be / beilpsched.c
index 04f22f0..a6a3eb4 100644 (file)
@@ -3,6 +3,7 @@
  * An ILP scheduler based on
  * "ILP-based Instruction Scheduling for IA-64"
  * by Daniel Kaestner and Sebastian Winkel
+ * extended with register pressure constraints by Christian Wuerdig
  *
  * @date   22.10.2005
  * @author Christian Wuerdig
 #include "debug.h"
 #include "irtools.h"
 #include "irdump.h"
+#include "irprintf.h"
 #include "plist.h"
+#include "irprintf.h"
 
 #include <lpp/lpp.h>
 #include <lpp/lpp_net.h>
 
-#ifdef WITH_LIBCORE
 #include <libcore/lc_opts.h>
 #include <libcore/lc_opts_enum.h>
 #include <libcore/lc_timing.h>
-#endif /* WITH_LIBCORE */
 
 #include "be.h"
 #include "benode_t.h"
 #include "besched_t.h"
 #include "beilpsched.h"
+#include "beutil.h"
+#include "bestat.h"
+#include "beirg_t.h"
 
 typedef struct _ilpsched_options_t {
-       unsigned limit_dead;
+       unsigned regpress;
        unsigned time_limit;
        char     log_file[1024];
 } ilpsched_options_t;
@@ -63,10 +67,18 @@ typedef struct _unit_type_info_t {
 typedef struct _ilp_var_types_t {
        int *x;   /* x_{nt}^k variables */
        int *a;   /* a_{nt}^k variables */
-       int *d;   /* d_{nt}^k variables */
        int *y;   /* y_{nt}^k variables */
 } ilp_var_types_t;
 
+/**
+ * Holds alive variables for a node live-in to a block.
+ */
+typedef struct _ilp_livein_node_t {
+       ir_node  *irn;
+       unsigned max_alive_steps;
+       int      *a;
+} ilp_livein_node_t;
+
 /* attributes for a node */
 typedef struct _ilpsched_node_attr_t {
        unsigned asap;                     /**< The ASAP scheduling control step */
@@ -83,7 +95,7 @@ typedef struct _ilpsched_node_attr_t {
        unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
        bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
                                                                                        for ASAP, successors for ALAP */
-       unsigned n_unit_types;             /**< number of allowed execution unit types */
+       unsigned         n_unit_types;     /**< number of allowed execution unit types */
        unit_type_info_t *type_info;       /**< list of allowed execution unit types */
        ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
 } ilpsched_node_attr_t;
@@ -95,6 +107,7 @@ typedef struct _ilpsched_block_attr_t {
        unsigned max_steps;             /**< Upper bound for block execution */
        plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
        ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
+       pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
 } ilpsched_block_attr_t;
 
 typedef union _ilpsched_attr_ {
@@ -110,7 +123,7 @@ typedef struct {
 
 /* The ILP scheduling environment */
 typedef struct {
-       phase_t              ph;            /**< The phase */
+       ir_phase             ph;            /**< The phase */
        ir_graph             *irg;          /**< The current irg */
        heights_t            *height;       /**< The heights object of the irg */
        void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
@@ -120,6 +133,8 @@ typedef struct {
        const be_main_env_t  *main_env;
        const be_machine_t   *cpu;          /**< the current abstract machine */
        ilpsched_options_t   *opts;         /**< the ilp options for current irg */
+       const be_irg_t       *birg;         /**< The birg object */
+       be_options_t         *be_opts;      /**< backend options */
        const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
        DEBUG_ONLY(firm_dbg_module_t *dbg);
 } be_ilpsched_env_t;
@@ -130,9 +145,6 @@ typedef struct {
 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
 
-/* iterate over a list of ir_nodes linked by link field */
-#define foreach_linked_irns(head, iter) for ((iter) = (head); (iter); (iter) = get_irn_link((iter)))
-
 /* check if node is considered for ILP scheduling */
 #define consider_for_sched(isa, irn) \
        (! (is_Block(irn)            ||  \
@@ -160,32 +172,24 @@ typedef struct {
 /* check if a double value is within an epsilon environment of 0 */
 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
 
-#ifdef WITH_LIBCORE
-       #define ilp_timer_push(t)         lc_timer_push((t))
-       #define ilp_timer_pop()           lc_timer_pop()
-       #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
-#else /* WITH_LIBCORE */
-       #define ilp_timer_push(t)
-       #define ilp_timer_pop()
-       #define ilp_timer_elapsed_usec(t) 0.0
-#endif /* WITH_LIBCORE */
+#define ilp_timer_push(t)         lc_timer_push((t))
+#define ilp_timer_pop()           lc_timer_pop()
+#define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
 
 /* option variable */
 static ilpsched_options_t ilp_opts = {
-       70,    /* if we have more than 70 nodes: use alive nodes constraint */
+       1,     /* default is with register pressure constraints */
        300,   /* 300 sec per block time limit */
        ""     /* no log file */
 };
 
-#ifdef WITH_LIBCORE
 /* ILP options */
 static const lc_opt_table_entry_t ilpsched_option_table[] = {
-       LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead),
+       LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
        LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
        LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
        { NULL }
 };
-#endif /* WITH_LIBCORE */
 
 /*
        We need this global variable as we compare nodes dependent on heights,
@@ -218,6 +222,12 @@ static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, v
        return lat;
 }
 
+static int cmp_live_in_nodes(const void *a, const void *b) {
+       const ilp_livein_node_t *n1 = a;
+       const ilp_livein_node_t *n2 = b;
+
+       return n1->irn != n2->irn;
+}
 
 /**
  * Compare scheduling time steps of two be_ilpsched_irn's.
@@ -250,7 +260,7 @@ static int cmp_ilpsched_irn(const void *a, const void *b) {
 /**
  * In case there is no phase information for irn, initialize it.
  */
-static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
+static void *init_ilpsched_irn(ir_phase *ph, ir_node *irn, void *old) {
        be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
 
        if (res == old) {
@@ -287,6 +297,7 @@ static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
                ba->block_last_idx      = 0;
                ba->root_nodes          = plist_new();
                ba->head_ilp_nodes      = NULL;
+               ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
                ba->max_steps           = 0;
        }
        else {
@@ -306,6 +317,7 @@ static void build_block_idx(ir_node *irn, void *walk_env) {
        ilpsched_node_attr_t  *na;
        ilpsched_block_attr_t *ba;
 
+       set_irn_link(irn, NULL);
        if (! consider_for_sched(env->arch_env->isa, irn))
                return;
 
@@ -420,6 +432,9 @@ static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
                                /* mark user visited by this node */
                                ua->consumer_idx = idx;
                        }
+                       else if (get_nodes_block(user) != block) {
+                               n_consumer++;
+                       }
                }
        }
 
@@ -589,7 +604,7 @@ static void calculate_block_alap(ir_node *block, void *walk_env) {
 }
 
 /**
- * We can free the list of root nodes here.
+ * Free list of root nodes and the set of live-in nodes.
  */
 static void clear_unwanted_data(ir_node *block, void *walk_env) {
        be_ilpsched_env_t     *env        = walk_env;
@@ -598,6 +613,8 @@ static void clear_unwanted_data(ir_node *block, void *walk_env) {
 
        plist_free(ba->root_nodes);
        ba->root_nodes = NULL;
+       del_pset(ba->livein_nodes);
+       ba->livein_nodes = NULL;
 }
 
 /**
@@ -690,8 +707,11 @@ static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, u
                foreach_out_edge(irn, edge) {
                        ir_node *user = get_edge_src_irn(edge);
 
-                       if (to_appear_in_schedule(user) || get_irn_mode(user) == mode_b)
+                       if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
+                               get_irn_n_edges(user) > 0)
+                       {
                                notified_sched_add_before(env, block, user, cycle);
+                       }
 
                        check_for_keeps(keeps, block, user);
                }
@@ -750,6 +770,20 @@ static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
                        cur_var = 0;
                        found   = 0;
 
+                       if (! na->is_dummy_node) {
+                               for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
+                                       for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
+                                               double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
+
+                                               if (! LPP_VALUE_IS_0(cost)) {
+                                                       DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
+                                                       found = 1;
+                                               }
+                                       }
+                               }
+                       }
+
+                       found = 0;
                        /* go over all variables of a node until the non-zero one is found */
                        for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
                                for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
@@ -759,7 +793,7 @@ static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
                                        if (! LPP_VALUE_IS_0(val)) {
                                                na->sched_point = t;
                                                ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
-                                               DBG((env->dbg, LEVEL_1, "Schedpoint of %+F is %u at unit type %s\n",
+                                               DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
                                                        irn, t, na->type_info[tp_idx].tp->name));
                                                found = 1;
                                        }
@@ -845,20 +879,74 @@ static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp
  *
  ************************************************/
 
+static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
+       const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
+       unsigned                  n_unit_types = 0;
+       be_ilpsched_irn_t         *node;
+       ilpsched_node_attr_t      *na;
+       unsigned                  unit_idx, tp_idx;
+
+       /* count number of available unit types for this node */
+       for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
+               /* just count */ ;
+
+       node = get_ilpsched_irn(env, irn);
+       na   = get_ilpsched_node_attr(node);
+
+       if (! na->type_info) {
+               na->n_unit_types = n_unit_types;
+               na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
+
+               /* fill the type info array */
+               for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
+                       for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
+                               /* beware: we also count number of available units here */
+                               if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
+                                       na->is_dummy_node = 1;
+                       }
+
+                       na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
+                       na->type_info[tp_idx].n_units = unit_idx;
+               }
+       }
+
+       return n_unit_types;
+}
+
+/**
+ * Returns the largest alap time of a user of @p irn.
+ * The user must be in block @p block.
+ */
+static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
+       const ir_edge_t *edge;
+       unsigned        max_alap = 0;
+
+       foreach_out_edge(irn, edge) {
+               ir_node *user = get_edge_src_irn(edge);
+
+               if (get_nodes_block(user) == block) {
+                       be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
+                       ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
+
+                       max_alap = MAX(max_alap, na->alap);
+               }
+       }
+
+       assert(max_alap > 0);
+       return max_alap;
+}
+
 /**
  * Create the following variables:
  * - x_{nt}^k    binary     weigthed with: t
  *      node n is scheduled at time step t to unit type k
  * ==>> These variables represent the schedule
  *
- * - d_{nt}^k    binary     weighted with: t
- *      node n dies at time step t on unit type k
  * - a_{nt}^k    binary     weighted with num_nodes
  *      node n is alive at time step t on unit type k
  *
- * - y_{nt}^k    binary     weighted with: num_nodes^2
- *      node n is scheduled at time step t to unit type k
- *      although all units of this type are occupied
+ * - y_{nt}^k    continuous  weighted with: num_nodes^2
+ *      register pressure over limit for unit type k
  * ==>> These variables represent the register pressure
  *
  */
@@ -866,42 +954,23 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn
        char                  buf[1024];
        ir_node               *irn;
        unsigned              num_block_var, num_nodes;
+       ilp_livein_node_t     *livein;
        ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
        unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
-#ifdef WITH_LIBCORE
        lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
-#endif /* WITH_LIBCORE */
 
        ilp_timer_push(t_var);
        num_block_var = num_nodes = 0;
        foreach_linked_irns(ba->head_ilp_nodes, irn) {
-               const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
-               be_ilpsched_irn_t         *node;
-               ilpsched_node_attr_t      *na;
-               unsigned                  n_unit_types, tp_idx, unit_idx, n_var, cur_unit;
-               unsigned                  cur_var_ad, cur_var_x, cur_var_y, num_ad;
-
-               /* count number of available unit types for this node */
-               for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
-                       /* just count */ ;
-
-               node = get_ilpsched_irn(env, irn);
-               na   = get_ilpsched_node_attr(node);
-
-               na->n_unit_types = n_unit_types;
-               na->type_info    = NEW_ARR_D(unit_type_info_t, var_obst, n_unit_types);
+               be_ilpsched_irn_t    *node;
+               ilpsched_node_attr_t *na;
+               unsigned             n_unit_types, tp_idx, n_var, cur_unit;
+               unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
+               int                  i;
 
-               /* fill the type info array */
-               for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
-                       for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
-                               /* beware: we also count number of available units here */
-                               if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
-                                       na->is_dummy_node = 1;
-                       }
-
-                       na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
-                       na->type_info[tp_idx].n_units = unit_idx;
-               }
+               node         = get_ilpsched_irn(env, irn);
+               na           = get_ilpsched_node_attr(node);
+               n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
 
                /* allocate space for ilp variables */
                na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
@@ -912,16 +981,9 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn
                        na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
                        memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
 
-                       num_ad = ba->max_steps - na->asap + 1;
-
-                       if (ba->n_interesting_nodes > env->opts->limit_dead) {
-                               na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
-                               memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
-                       }
-                       else {
-                               na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
-                               memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0]));
-                       }
+                       num_ad         = ba->max_steps - na->asap + 1;
+                       na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
+                       memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
                }
 
                DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
@@ -946,7 +1008,7 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn
                                        /* y_{nt}^k variables */
                                        snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
                                                get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
-                                       na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_binary, (double)(weigth_y));
+                                       na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
                                        DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 
                                        /* variable counter */
@@ -959,18 +1021,10 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn
                        if (! na->is_dummy_node) {
                                for (t = na->asap - 1; t <= ba->max_steps; ++t) {
 
-                                       if (ba->n_interesting_nodes > env->opts->limit_dead) {
-                                               /* a_{nt}^k variables */
-                                               snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
-                                                       get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
-                                               na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
-                                       }
-                                       else {
-                                               /* d_{nt}^k variables */
-                                               snprintf(buf, sizeof(buf), "d_n%u_%s_%u",
-                                                       get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
-                                               na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
-                                       }
+                                       /* a_{nt}^k variables */
+                                       snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
+                                               get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
+                                       na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
                                        DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
 
                                        /* variable counter */
@@ -978,11 +1032,58 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn
                                        num_block_var++;
                                }
                        }
+
+                       /* collect live-in nodes */
+                       for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
+                               ir_node *pred = get_irn_n(irn, i);
+
+                               if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
+                                       be_ilpsched_set_type_info(env, pred, var_obst);
+                                       if (! na->is_dummy_node) {
+                                               ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
+                                               entry->irn = pred;
+                                               entry->a   = NULL;
+                                               pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
+                                       }
+                               }
+                       }
                }
 
                DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
                num_nodes++;
        }
+
+       /* create alive variables a_{nt}^k for live-ins */
+       foreach_pset(ba->livein_nodes, livein) {
+               be_ilpsched_irn_t    *node;
+               ilpsched_node_attr_t *na;
+               unsigned             tp_idx, var_idx;
+               ir_node              *irn;
+
+               irn  = livein->irn;
+               node = get_ilpsched_irn(env, irn);
+               na   = get_ilpsched_node_attr(node);
+
+               livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
+
+               livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
+               var_idx   = 0;
+
+               /* create variables */
+               for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
+                       unsigned t;
+
+                       for (t = 0; t < livein->max_alive_steps; ++t) {
+                               /* a_{nt}^k variables */
+                               snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
+                                       get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
+                               livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
+                               DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
+                               num_block_var++;
+                       }
+               }
+       }
+
        ilp_timer_pop();
        DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
                num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
@@ -998,6 +1099,33 @@ static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn
  *
  *******************************************************/
 
+/**
+ * Collect all operands and nodes @p irn depends on.
+ * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
+ */
+static nodeset *sta_collect_in_deps(ir_node *irn, nodeset *deps) {
+       int i;
+
+       for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
+               ir_node *p = get_irn_in_or_dep(irn, i);
+
+               if (is_Proj(p)) {
+                       const ir_edge_t *edge;
+
+                       p = get_Proj_pred(p);
+                       foreach_out_edge(p, edge) {
+                               ir_node *src = get_edge_src_irn(edge);
+                               nodeset_insert(deps, src);
+                       }
+               }
+               else {
+                       nodeset_insert(deps, p);
+               }
+       }
+
+       return deps;
+}
+
 /**
  * Create following ILP constraints:
  * - the assignment constraints:
@@ -1013,18 +1141,17 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env,
        ir_node               *irn;
        ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
        bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
-#ifdef WITH_LIBCORE
-       lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign",      "create assignment constraints");
-       lc_timer_t            *t_cst_dead    = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints");
-       lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",        "create precedence constraints");
-#endif /* WITH_LIBCORE */
+       lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
+       lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",   "create precedence constraints");
 
        num_cst_assign = num_cst_prec = num_cst_dead = 0;
        foreach_linked_irns(ba->head_ilp_nodes, irn) {
-               int                  cst, tp_idx, i;
+               int                  cst, tp_idx;
                unsigned             cur_var;
                be_ilpsched_irn_t    *node;
                ilpsched_node_attr_t *na;
+               ir_node              *pred;
+               nodeset              *deps = new_nodeset(16);
 
                node    = get_ilpsched_irn(env, irn);
                na      = get_ilpsched_node_attr(node);
@@ -1040,17 +1167,6 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env,
                lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
                ilp_timer_pop();
 
-               /* the dead node assignment constraint */
-               if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) {
-                       ilp_timer_push(t_cst_dead);
-                       snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn));
-                       cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
-                       DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
-
-                       lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0);
-                       ilp_timer_pop();
-               }
-
                /* We have separate constraints for Projs and Keeps */
                // ILP becomes infeasible ?!?
 //             if (is_Proj(irn) || be_is_Keep(irn))
@@ -1059,13 +1175,15 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env,
                /* the precedence constraints */
                ilp_timer_push(t_cst_prec);
                bs_block_irns = bitset_clear_all(bs_block_irns);
-               for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
-                       ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
+
+               deps = sta_collect_in_deps(irn, deps);
+               foreach_nodeset(deps, pred) {
                        unsigned             t_low, t_high, t;
                        be_ilpsched_irn_t    *pred_node;
                        ilpsched_node_attr_t *pna;
                        unsigned             delay;
 
+                       pred = skip_normal_Proj(env->arch_env->isa, pred);
                        if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
                                continue;
 
@@ -1132,6 +1250,7 @@ static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env,
                                DEL_ARR_F(tmp_var_idx);
                        }
                }
+               del_nodeset(deps);
                ilp_timer_pop();
        }
        DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
@@ -1150,9 +1269,7 @@ static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_
        char                  buf[1024];
        unsigned              num_cst_resrc = 0;
        ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
        lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
-#endif /* WITH_LIBCORE */
 
        ilp_timer_push(t_cst_rsrc);
        for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
@@ -1208,9 +1325,7 @@ static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilp
        unsigned              num_cst_bundle = 0;
        unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
        ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
        lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
-#endif /* WITH_LIBCORE */
 
        ilp_timer_push(t_cst_bundle);
        for (t = 0; t < ba->max_steps; ++t) {
@@ -1258,101 +1373,15 @@ static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilp
 }
 
 /**
- * Create ILP dying nodes constraints:
- * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k
+ * Create ILP alive nodes constraints:
+ * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
  */
-static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
-       char                  buf[1024];
-       unsigned              t;
-       unsigned              num_cst = 0;
-       ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
-       lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints");
-#endif /* WITH_LIBCORE */
-
-       ilp_timer_push(t_cst);
-       /* check all time_steps */
-       for (t = 0; t < ba->max_steps; ++t) {
-               ir_node *irn;
-
-               /* for all nodes */
-               foreach_linked_irns(ba->head_ilp_nodes, irn) {
-                       be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
-                       ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
-
-                       /* if node has no consumer within current block, it cannot die here */
-                       /* we also ignore nodes assigned to dummy unit */
-                       if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node)
-                               continue;
-
-                       /* node can only die here if t at least asap(n) */
-                       if (t >= na->asap - 1) {
-                               int node_tp_idx;
-
-                               /* for all unit types */
-                               for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
-                                       int tp_idx, i, cst;
-                                       int *tmp_var_idx = NEW_ARR_F(int, 0);
-
-                                       snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn));
-                                       cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1));
-                                       DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
-                                       num_cst++;
-
-                                       /* number of consumer scheduled till t */
-                                       for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
-                                               be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
-                                               ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
-
-                                               for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
-                                                       unsigned tm;
-
-                                                       for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) {
-                                                               int idx = ILPVAR_IDX(ca, tp_idx, tm);
-                                                               ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]);
-                                                       }
-                                               }
-                                       }
-
-                                       /* could be that no consumer can be scheduled at this point */
-                                       if (ARR_LEN(tmp_var_idx)) {
-                                               int      idx;
-                                               unsigned tn;
-
-                                               /* subtract possible prior kill points */
-                                               for (tn = na->asap - 1; tn < t; ++tn) {
-                                                       idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn);
-                                                       lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0);
-                                               }
-
-                                               idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
-                                               lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer));
-                                               lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
-                                       }
-
-                                       DEL_ARR_F(tmp_var_idx);
-                               }
-                       }
-
-               }
-       }
-       ilp_timer_pop();
-       DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n",
-               num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
-}
-
-/**
-* Create ILP alive nodes constraints:
-* - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
-*/
 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
        char                  buf[1024];
        ir_node               *irn;
        unsigned              num_cst = 0;
        ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
        lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
-#endif /* WITH_LIBCORE */
 
        ilp_timer_push(t_cst);
        /* for each node */
@@ -1426,103 +1455,80 @@ static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be
 }
 
 /**
- * Create ILP pressure constraints, based on dead nodes:
- * - add additional costs to objective function if a node is scheduled
- *   on a unit although all units of this type are currently occupied
+ * Create ILP alive nodes constraints for live-in nodes:
+ * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
  */
-static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
+static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
        char                  buf[1024];
-       ir_node               *cur_irn;
+       ilp_livein_node_t     *livein;
        unsigned              num_cst = 0;
        ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
-       lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
-#endif /* WITH_LIBCORE */
+       lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
 
        ilp_timer_push(t_cst);
-       /* y_{nt}^k is set for each node and timestep and unit type */
-       foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
-               unsigned             cur_idx   = get_irn_idx(cur_irn);
-               be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
-               ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
-               int                  glob_type_idx;
-
-               /* we ignore nodes assigned to DUMMY unit here */
-               if (cur_na->is_dummy_node)
-                       continue;
-
-               /* for all types */
-               for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
-                       be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
-                       int                      cur_tp_idx;
-                       unsigned                 t;
-
-                       /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
-
-                       /* check if node can be executed on this unit type */
-                       cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
-                       if (cur_tp_idx < 0)
-                               continue;
+       /* for each node */
+       foreach_pset(ba->livein_nodes, livein) {
+               ir_node              *irn  = livein->irn;
+               be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
+               ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
+               unsigned             t;
 
-                       /* check all time_steps */
-                       for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
-                               int     cst, y_idx;
-                               ir_node *irn;
-                               int     *tmp_idx_1  = NEW_ARR_F(int, 0);
-                               int     *tmp_idx_m1 = NEW_ARR_F(int, 0);
+               /* check check all time steps: 0 <= t < max_alive_steps */
+               for (t = 0; t < livein->max_alive_steps; ++t) {
+                       int node_tp_idx;
 
-                               snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
-                               cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
-                               DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
-                               num_cst++;
+                       /* for all unit types available for this node */
+                       for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
+                               const ir_edge_t *edge;
+                               unsigned idx;
+                               int      cst, num_block_user;
+                               int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
 
-                               /*
-                                       - accumulate all nodes scheduled on unit type k till t
-                                       - subtract all nodes died on unit type k till t
-                               */
-                               foreach_linked_irns(ba->head_ilp_nodes, irn) {
-                                       be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
-                                       ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
-                                       unsigned             tn, tmax;
+                               /* check the number of consumer scheduled so far */
+                               num_block_user = 0;
+                               foreach_out_edge(irn, edge) {
+                                       ir_node              *user = get_edge_src_irn(edge);
+                                       be_ilpsched_irn_t    *cons;
+                                       ilpsched_node_attr_t *ca;
                                        int                  tp_idx;
+                                       unsigned             tm, tm_max;
 
-                                       tmax   = MIN(t, na->alap - 1);
-                                       tp_idx = is_valid_unit_type_for_node(cur_tp, node);
-
-                                       /* current unit type is not suitable for current node */
-                                       if (tp_idx < 0)
+                                       /* check only users within current block */
+                                       if (get_nodes_block(user) != block_node->irn)
                                                continue;
 
-                                       for (tn = na->asap - 1; tn <= tmax; ++tn) {
-                                               int idx;
-
-                                               /* node scheduled */
-                                               idx = ILPVAR_IDX(na, tp_idx, tn);
-                                               ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]);
+                                       num_block_user++;
+                                       cons = get_ilpsched_irn(env, user);
+                                       ca   = get_ilpsched_node_attr(cons);
 
-                                               /* node dead */
-                                               idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn);
-                                               ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]);
+                                       tm_max = MIN(ca->alap - 1, t);
+                                       for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
+                                               for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
+                                                       int idx = ILPVAR_IDX(ca, tp_idx, tm);
+                                                       ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
+                                               }
                                        }
                                }
 
-                               if (ARR_LEN(tmp_idx_1) > 0)
-                                       lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0);
-
-                               if (ARR_LEN(tmp_idx_m1) > 0)
-                                       lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0);
+                               snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
+                                       t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
+                               cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
+                               DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
+                               num_cst++;
 
-                               /* BEWARE: t is unsigned, so (double)(-t) won't work */
-                               y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
-                               lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(t));
+                               /* sum(scheduled users) */
+                               if (ARR_LEN(tmp_var_idx_m) > 0)
+                                       lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
+                               DEL_ARR_F(tmp_var_idx_m);
 
-                               DEL_ARR_F(tmp_idx_1);
-                               DEL_ARR_F(tmp_idx_m1);
+                               /* + c * a_{nt}^k */
+                               idx = node_tp_idx * livein->max_alive_steps + t;
+                               lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
                        }
                }
        }
        ilp_timer_pop();
-       DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
+       DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
                num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
 }
 
@@ -1536,9 +1542,7 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp,
        ir_node               *cur_irn;
        unsigned              num_cst = 0;
        ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
        lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
-#endif /* WITH_LIBCORE */
 
        ilp_timer_push(t_cst);
        /* y_{nt}^k is set for each node and timestep and unit type */
@@ -1570,6 +1574,7 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp,
                                int     cst, y_idx;
                                ir_node *irn;
                                int     *tmp_var_idx = NEW_ARR_F(int, 0);
+                               ilp_livein_node_t *livein;
 
                                snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
                                cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
@@ -1595,6 +1600,25 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp,
                                        a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
                                        ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
                                }
+                               /* do the same for livein nodes */
+                               foreach_pset(ba->livein_nodes, livein) {
+                                       ir_node              *irn  = livein->irn;
+                                       be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
+                                       int                  a_idx, tp_idx;
+
+                                       /* check if node can be alive here */
+                                       if (t >= livein->max_alive_steps)
+                                               continue;
+
+                                       tp_idx = is_valid_unit_type_for_node(cur_tp, node);
+
+                                       /* current type is not suitable */
+                                       if (tp_idx < 0)
+                                               continue;
+
+                                       a_idx = tp_idx * livein->max_alive_steps + t;
+                                       ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
+                               }
 
                                if (ARR_LEN(tmp_var_idx) > 0)
                                        lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
@@ -1602,7 +1626,7 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp,
 
                                /* - num_nodes * y_{nt}^k */
                                y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
-                               lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], 0.0 - (double)(ba->n_interesting_nodes));
+                               lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
                        }
                }
        }
@@ -1611,15 +1635,109 @@ static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp,
                num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
 }
 
+/**
+ * Create ILP branch constraints:
+ * Assure, alle nodes are scheduled prior to cfg op.
+ */
+static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
+       char                  buf[1024];
+       ir_node               *cur_irn, *cfop;
+       unsigned              num_cst          = 0;
+       unsigned              num_non_branches = 0;
+       ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
+       lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
+
+       ilp_timer_push(t_cst);
+       cfop = NULL;
+       /* determine number of non-branch nodes and the one and only branch node */
+       foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
+               switch (get_irn_opcode(cur_irn)) {
+                       case iro_Phi:
+                       case iro_Start:
+                       case iro_End:
+                       case iro_Proj:
+                       case iro_Bad:
+                       case iro_Unknown:
+                               num_non_branches++;
+                               break;
+                       default:
+                               if (is_cfop(cur_irn)) {
+                                       assert(cfop == NULL && "Highlander - there can be only one to be constrained");
+                                       cfop = cur_irn;
+                               }
+                               else {
+                                       num_non_branches++;
+                               }
+                               break;
+               }
+       }
+
+       if (cfop) {
+               be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
+               ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
+               unsigned t;
+
+               /* for each time step */
+               for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
+                       int *non_branch_vars, *branch_vars;
+                       int cst;
+
+                       snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
+                       cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
+                       DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
+                       num_cst++;
+
+                       /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
+                       non_branch_vars = NEW_ARR_F(int, 0);
+                       branch_vars     = NEW_ARR_F(int, 0);
+                       foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
+                               be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
+                               ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
+                               int                  tp_idx;
+
+                               if (cur_irn == cfop) {
+                                       /* for all unit types available for this node */
+                                       for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
+                                               unsigned idx = ILPVAR_IDX(na, tp_idx, t);
+                                               ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
+                                       }
+                               }
+                               else {
+                                       /* sum up all possible schedule points for this node upto current timestep */
+                                       for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
+                                               unsigned tn;
+                                               unsigned tmax = MIN(t, na->alap - 1);
+
+                                               for (tn = na->asap - 1; tn <= tmax; ++tn) {
+                                                       unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
+                                                       ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
+                                               }
+                                       }
+                               }
+
+                       }
+
+                       if (ARR_LEN(non_branch_vars) > 0)
+                               lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
+                       if (ARR_LEN(branch_vars) > 0)
+                               lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
+
+                       DEL_ARR_F(branch_vars);
+                       DEL_ARR_F(non_branch_vars);
+               }
+       }
+       ilp_timer_pop();
+       DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
+               num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
+}
+
 #if 0
 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
        char                  buf[1024];
        ir_node               *irn;
        unsigned              num_cst = 0;
        ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
-#ifdef WITH_LIBCORE
        lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
-#endif /* WITH_LIBCORE */
 
        ilp_timer_push(t_cst);
        /* check all nodes */
@@ -1685,7 +1803,7 @@ static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_
        DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
                num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
 }
-#endif
+#endif /* if 0 */
 
 /***************************************************
  *  _____ _      _____                    _
@@ -1706,6 +1824,7 @@ static void create_ilp(ir_node *block, void *walk_env) {
        ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
        FILE                  *logfile       = NULL;
        lpp_t                 *lpp           = NULL;
+       int                   need_heur      = 0;
        struct obstack        var_obst;
        char                  name[1024];
 
@@ -1721,13 +1840,13 @@ static void create_ilp(ir_node *block, void *walk_env) {
 
        /* if we have less than two interesting nodes, there is no need to create the ILP */
        if (ba->n_interesting_nodes > 1) {
-               double fact_var        = ba->n_interesting_nodes > 25 ? 1.1 : 1.2;
-               double fact_cst        = ba->n_interesting_nodes > 25 ? 0.7 : 1.5;
+               double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
+               double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
                int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
                int    estimated_n_var = (int)((double)base_num * fact_var);
                int    estimated_n_cst = (int)((double)base_num * fact_cst);
 
-               DBG((env->dbg, LEVEL_1, "Creating LPP with estimed numbers: %d vars, %d cst\n",
+               DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
                        estimated_n_var, estimated_n_cst));
 
                /* set up the LPP object */
@@ -1736,9 +1855,9 @@ static void create_ilp(ir_node *block, void *walk_env) {
                lpp = new_lpp_userdef(
                        (const char *)name,
                        lpp_minimize,
-                       estimated_n_cst + 1,  /* num vars */
-                       estimated_n_cst + 20, /* num cst */
-                       1.2);                 /* grow factor */
+                       estimated_n_cst,     /* num vars */
+                       estimated_n_cst + 1, /* num cst */
+                       1.3);                /* grow factor */
                obstack_init(&var_obst);
 
                /* create ILP variables */
@@ -1749,22 +1868,20 @@ static void create_ilp(ir_node *block, void *walk_env) {
                create_assignment_and_precedence_constraints(env, lpp, block_node);
                create_ressource_constraints(env, lpp, block_node);
                create_bundle_constraints(env, lpp, block_node);
+               create_branch_constraint(env, lpp, block_node);
                //create_proj_keep_constraints(env, lpp, block_node);
-#if 0
-               if (ba->n_interesting_nodes > env->opts->limit_dead) {
+
+               if (env->opts->regpress) {
                        create_alive_nodes_constraint(env, lpp, block_node);
+                       create_alive_livein_nodes_constraint(env, lpp, block_node);
                        create_pressure_alive_constraint(env, lpp, block_node);
-               } else {
-                       create_dying_nodes_constraint(env, lpp, block_node);
-                       create_pressure_dead_constraint(env, lpp, block_node);
                }
-#endif
 
                DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
 
                /* debug stuff, dump lpp when debugging is on  */
                DEBUG_ONLY(
-                       if (firm_dbg_get_mask(env->dbg) > 0) {
+                       if (firm_dbg_get_mask(env->dbg) > 1) {
                                char buf[1024];
                                FILE *f;
 
@@ -1806,20 +1923,26 @@ static void create_ilp(ir_node *block, void *walk_env) {
                        char buf[1024];
                        FILE *f;
 
-                       snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block));
-                       f = fopen(buf, "w");
-                       lpp_dump_plain(lpp, f);
-                       fclose(f);
-                       snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block));
-                       lpp_dump(lpp, buf);
-                       dump_ir_block_graph(env->irg, "-assert");
+                       DEBUG_ONLY(
+                               if (firm_dbg_get_mask(env->dbg) >= 2) {
+                                       snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
+                                       f = fopen(buf, "w");
+                                       lpp_dump_plain(lpp, f);
+                                       fclose(f);
+                                       snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
+                                       lpp_dump(lpp, buf);
+                                       dump_ir_block_graph(env->irg, "-infeasible");
+                               }
+                       )
 
-                       assert(0 && "ILP solution is not feasible!");
+                       ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
+                       need_heur = 1;
                }
 
                DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
                DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
                DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
+               DBG((env->dbg, LEVEL_1, "\tmatrix: %u elements, density %.2f%%, size %.2fMB\n", lpp->n_elems, lpp->density, (double)lpp->matrix_mem / 1024.0 / 1024.0));
                DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
                DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
                DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
@@ -1829,7 +1952,38 @@ static void create_ilp(ir_node *block, void *walk_env) {
        }
 
        /* apply solution */
-       apply_solution(env, lpp, block);
+#ifdef FIRM_STATISTICS
+       if (be_stat_ev_is_active()) {
+               be_stat_ev("nodes", ba->block_last_idx);
+               be_stat_ev("vars", lpp ? lpp->var_next : 0);
+               be_stat_ev("csts", lpp ? lpp->cst_next : 0);
+       }
+#endif /* FIRM_STATISTICS */
+       if (need_heur) {
+#ifdef FIRM_STATISTICS
+               if (be_stat_ev_is_active()) {
+                       be_stat_ev("time", -1);
+                       be_stat_ev_dbl("opt", 0.0);
+               }
+#endif /* FIRM_STATISTICS */
+               list_sched_single_block(env->birg, block, env->be_opts);
+       }
+       else {
+#ifdef FIRM_STATISTICS
+               if (be_stat_ev_is_active()) {
+                       if (lpp) {
+                               double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
+                               be_stat_ev_dbl("time", lpp->sol_time);
+                               be_stat_ev_dbl("opt", opt);
+                       }
+                       else {
+                               be_stat_ev_dbl("time", 0.0);
+                               be_stat_ev_dbl("opt", 100.0);
+                       }
+               }
+#endif /* FIRM_STATISTICS */
+               apply_solution(env, lpp, block);
+       }
 
        if (lpp)
                free_lpp(lpp);
@@ -1841,24 +1995,35 @@ static void create_ilp(ir_node *block, void *walk_env) {
 /**
  * Perform ILP scheduling on the given irg.
  */
-void be_ilp_sched(const be_irg_t *birg) {
+void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
        be_ilpsched_env_t          env;
-       const char                 *name = "be ilp scheduling";
-       arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
-       const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
+       const char                 *name     = "be ilp scheduling";
+       ir_graph                   *irg      = be_get_birg_irg(birg);
+       const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
+       const arch_isa_t           *isa      = arch_env->isa;
+       const ilp_sched_selector_t *sel      = isa->impl->get_ilp_sched_selector(isa);
 
        FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
 
-       //firm_dbg_set_mask(env.dbg, 31);
+#ifdef FIRM_STATISTICS
+       if (be_stat_ev_is_active()) {
+               be_stat_tags[STAT_TAG_CLS] = "ilpsched";
+               be_stat_ev_push(be_stat_tags, STAT_TAG_LAST, be_stat_file);
+       }
+#endif /* FIRM_STATISTICS */
 
-       env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
+//     firm_dbg_set_mask(env.dbg, 1);
+
+       env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
        env.sel        = sel;
-       env.irg        = birg->irg;
-       env.height     = heights_new(birg->irg);
+       env.irg        = irg;
+       env.height     = heights_new(irg);
        env.main_env   = birg->main_env;
-       env.arch_env   = birg->main_env->arch_env;
-       env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
+       env.arch_env   = arch_env;
+       env.cpu        = arch_isa_get_machine(arch_env->isa);
        env.opts       = &ilp_opts;
+       env.birg       = birg;
+       env.be_opts    = be_opts;
        phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
 
        /* assign a unique per block number to all interesting nodes */
@@ -1881,7 +2046,7 @@ void be_ilp_sched(const be_irg_t *birg) {
        irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
 
        /* perform ILP scheduling */
-       irg_block_walk_graph(env.irg, clear_unwanted_data, create_ilp, &env);
+       irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
 
        DEBUG_ONLY(
                if (firm_dbg_get_mask(env.dbg)) {
@@ -1892,15 +2057,23 @@ void be_ilp_sched(const be_irg_t *birg) {
                }
        );
 
+       /* free data allocated dynamically */
+       irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
+
        /* free all allocated object */
        phase_free(&env.ph);
        heights_free(env.height);
 
        /* notify backend */
        be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
+
+#ifdef FIRM_STATISTICS
+       if (be_stat_ev_is_active()) {
+               be_stat_ev_pop();
+       }
+#endif /* FIRM_STATISTICS */
 }
 
-#ifdef WITH_LIBCORE
 /**
  * Register ILP scheduler options.
  */
@@ -1911,7 +2084,6 @@ void be_init_ilpsched(void)
 
        lc_opt_add_table(sched_grp, ilpsched_option_table);
 }
-#endif /* WITH_LIBCORE */
 
 #else /* WITH_ILP */