beilpsched.c

/*
 * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
 *
 * This file is part of libFirm.
 *
 * This file may be distributed and/or modified under the terms of the
 * GNU General Public License version 2 as published by the Free Software
 * Foundation and appearing in the file LICENSE.GPL included in the
 * packaging of this file.
 *
 * Licensees holding valid libFirm Professional Edition licenses may use
 * this file in accordance with the libFirm Commercial License.
 * Agreement provided with the Software.
 *
 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE.
 */

/**
 * @file
 * @brief       ILP based instruction scheduling.
 * @author      Christian Wuerdig
 * @date        22.10.2006
 * @version     $Id$
 *
 * 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
 */
#include "config.h"

#ifdef WITH_ILP

#include <math.h>

#ifndef _WIN32
#include <strings.h>
#endif /* _WIN32 */

#include "irnode_t.h"
#include "irgwalk.h"
#include "irbitset.h"
#include "irphase_t.h"
#include "heights.h"
#include "iredges.h"
#include "pdeq.h"
#include "debug.h"
#include "irtools.h"
#include "irdump.h"
#include "irprintf.h"
#include "plist.h"
#include "irprintf.h"
#include "timing.h"

#include <lpp/lpp.h>
#include <lpp/lpp_net.h>

#include "lc_opts.h"
#include "lc_opts_enum.h"

#include "be.h"
#include "benode.h"
#include "besched.h"
#include "beilpsched.h"
#include "beutil.h"
#include "bestat.h"
#include "beirg.h"
#include "bemachine.h"
#include "belistsched.h"
#include "bemodule.h"

typedef struct ilpsched_options_t {
	unsigned regpress;
	unsigned time_limit;
	char     log_file[1024];
} ilpsched_options_t;

typedef struct unit_type_info_t {
	int                            n_units;
	const be_execution_unit_type_t *tp;
} unit_type_info_t;

/**
 * holding the ILP variables of the different types
 */
typedef struct ilp_var_types_t {
	int *x;   /* x_{nt}^k variables */
	int *a;   /* a_{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 */
	unsigned alap;                     /**< The ALAP scheduling control step */
	unsigned latency;                  /**< Latency of this node (needed for sorting) */
	unsigned sched_point;              /**< the step in which the node is finally scheduled */
	unsigned visit_idx;                /**< Index of the node having visited this node last */
	unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
	unsigned n_consumer;               /**< Number of consumers */
	ir_node  **block_consumer;         /**< List of consumer being in the same block */
	waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
	unsigned block_idx     : 30;       /**< A unique per block index */
	unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
	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 */
	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;

/* attributes for a block */
typedef struct ilpsched_block_attr_t {
	unsigned block_last_idx;        /**< The highest node index in block so far */
	unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
	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_ {
	ilpsched_node_attr_t  node_attr;
	ilpsched_block_attr_t block_attr;
} ilpsched_attr_t;

/* A irn for the phase and it's attributes (either node or block) */
typedef struct {
	const ir_node   *irn;
	ilpsched_attr_t attr;
} be_ilpsched_irn_t;

/* The ILP scheduling environment */
typedef struct {
	ir_phase             ph;            /**< The phase */
	ir_graph             *irg;          /**< The current irg */
	ir_heights_t         *height;       /**< The heights object of the irg */
	void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
	void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
	const arch_env_t     *arch_env;
	const be_machine_t   *cpu;          /**< the current abstract machine */
	ilpsched_options_t   *opts;         /**< the ilp options for current irg */
	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;

/* convenience macros to handle phase irn data */
#define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
#define is_ilpsched_block(node)             (is_Block((node)->irn))
#define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
#define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)

/* check if node is considered for ILP scheduling */
#define consider_for_sched(env, irn) \
	(! (is_Block(irn)            ||  \
		is_normal_Proj(env, irn) ||  \
		is_Phi(irn)              ||  \
		is_NoMem(irn)            ||  \
		is_Unknown(irn)          ||  \
		is_End(irn)                  \
		))

/* gives the valid scheduling time step interval for a node */
#define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)

/* gives the valid interval where a node can die */
#define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)

/* gives the corresponding ILP variable for given node, unit and time step */
#define ILPVAR_IDX(na, unit, control_step) \
	((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)

/* gives the corresponding dead nodes ILP variable for given node, unit and time step */
#define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
	((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)

/* check if a double value is within an epsilon environment of 0 */
#define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)

/* option variable */
static ilpsched_options_t ilp_opts = {
	1,     /* default is with register pressure constraints */
	300,   /* 300 sec per block time limit */
	""     /* no log file */
};

/* ILP options */
static const lc_opt_table_entry_t ilpsched_option_table[] = {
	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)),
	LC_OPT_LAST
};

/*
	We need this global variable as we compare nodes dependent on heights,
	but we cannot pass any information to the qsort compare function.
*/
static ir_heights_t *glob_heights;

/**
 * Check if irn is a Proj, which has no execution units assigned.
 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
 */
static inline int is_normal_Proj(const arch_env_t *env, const ir_node *irn)
{
	return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
}

/**
 * Skips normal Projs.
 * @return predecessor if irn is a normal Proj, otherwise irn.
 */
static inline ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn)
{
	if (is_normal_Proj(env, irn))
		return get_Proj_pred(irn);
	return irn;
}

static inline int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env)
{
	unsigned lat = be_ilp_sched_latency(sel, irn, env);
	if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
		lat = 1;
	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.
 */
static int cmp_ilpsched_irn(const void *a, const void *b)
{
	be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
	be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
	ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
	ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);

	if (n1_a->sched_point == n2_a->sched_point) {
		const ir_node *irn_a = n1->irn;
		const ir_node *irn_b = n2->irn;

		if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
			return 1;
		if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
			return -1;

		/*
			Ok, timestep is equal and the nodes are parallel,
			so check latency and schedule high latency first.
		*/
		return QSORT_CMP(n2_a->latency, n1_a->latency);
	}
	else
		return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
}

static void *reinit_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old)
{
	be_ilpsched_irn_t *res = old;

	/* if we have already some data: check for reinitialization */
	if (! is_Block(irn)) {
		ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);

		if (! na->transitive_block_nodes) {
			ir_node               *block      = get_nodes_block(irn);
			be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
			ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);

			/* we are called after the block indices have been build: create bitset */
			na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
		} else {
			/* we are called from reinit block data: clear the bitset */
			bitset_clear_all(na->transitive_block_nodes);
			na->visit_idx    = 0;
			na->alap_changed = 1;
		}
	}
	return res;
}

/**
 * In case there is no phase information for irn, initialize it.
 */
static void *init_ilpsched_irn(ir_phase *phase, const ir_node *irn)
{
	be_ilpsched_irn_t *res = phase_alloc(phase, sizeof(res[0]));
	res->irn = irn;

	/* set ilpsched irn attributes (either block or irn) */
	if (is_Block(irn)) {
		ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);

		ba->n_interesting_nodes = 0;
		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 {
		ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
		memset(na, 0, sizeof(*na));
	}

	return res;
}

/**
 * Assign a per block unique number to each node.
 */
static void build_block_idx(ir_node *irn, void *walk_env)
{
	be_ilpsched_env_t     *env = walk_env;
	be_ilpsched_irn_t     *node, *block_node;
	ilpsched_node_attr_t  *na;
	ilpsched_block_attr_t *ba;

	set_irn_link(irn, NULL);
	if (! consider_for_sched(env->arch_env, irn))
		return;

	node       = get_ilpsched_irn(env, irn);
	na         = get_ilpsched_node_attr(node);
	block_node = get_ilpsched_irn(env, get_nodes_block(irn));
	ba         = get_ilpsched_block_attr(block_node);

	na->block_idx = ba->block_last_idx++;
}

/********************************************************
 *                              __        _
 *                             / /       | |
 *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
 *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
 * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
 *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
 *                 | |                           | |
 *                 |_|                           |_|
 ********************************************************/

/**
 * Add all nodes having no user in current block to last_nodes list.
 */
static void collect_alap_root_nodes(ir_node *irn, void *walk_env)
{
	ir_node               *block;
	const ir_edge_t       *edge;
	be_ilpsched_irn_t     *block_node, *node;
	ilpsched_block_attr_t *ba;
	ilpsched_node_attr_t  *na;
	int                   i, j;
	be_ilpsched_env_t     *env           = walk_env;
	int                   has_block_user = 0;
	unsigned              n_consumer     = 0;
	ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
	ir_node               **consumer;
	unsigned              idx;

	if (! consider_for_sched(env->arch_env, irn))
		return;

	block    = get_nodes_block(irn);
    idx      = get_irn_idx(irn);
	consumer = NEW_ARR_F(ir_node *, 0);

	DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));

	/* check data and dependency out edges */
	for (i = 0; i < 2 && ! has_block_user; ++i) {
		foreach_out_edge_kind(irn, edge, ekind[i]) {
			ir_node *user = get_edge_src_irn(edge);

			if (is_normal_Proj(env->arch_env, user)) {
				const ir_edge_t *user_edge;

				if (get_irn_mode(user) == mode_X)
					continue;

				/* The ABI ensures, that there will be no ProjT nodes in the graph. */
				for (j = 0; j < 2; ++j) {
					foreach_out_edge_kind(user, user_edge, ekind[j]) {
						ir_node *real_user = get_edge_src_irn(user_edge);

						if (! is_Phi(real_user) && ! is_Block(real_user)) {
							be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
							ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);

							/* skip already visited nodes */
							if (ua->consumer_idx == idx)
								continue;

							/* check if node has user in this block and collect the user if it's a data user */
							if (get_nodes_block(real_user) == block) {
								if (i == 0 && j == 0)
									ARR_APP1(ir_node *, consumer, real_user);
								has_block_user = 1;
							}

							/* only count data consumer */
							if (i == 0)
								n_consumer++;

							/* mark user as visited by this node */
							ua->consumer_idx = idx;
						}
					}
				}
			}
			else if (is_Block(user)) {
				continue;
			}
			else if (! is_Phi(user)) {
				be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
				ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);

				/* skip already visited nodes */
				if (ua->consumer_idx == idx)
					continue;

				/* check if node has user in this block and collect the user if it's a data user */
				if (get_nodes_block(user) == block) {
					if (i == 0)
						ARR_APP1(ir_node *, consumer, user);
					has_block_user = 1;
				}

				/* only count data consumer */
				if (i == 0)
					n_consumer++;

				/* mark user visited by this node */
				ua->consumer_idx = idx;
			}
			else if (get_nodes_block(user) != block) {
				n_consumer++;
			}
		}
	}

	block_node = get_ilpsched_irn(env, block);
	ba         = get_ilpsched_block_attr(block_node);

	ba->n_interesting_nodes++;

	/* current irn has no user inside this block, add to queue */
	if (! has_block_user) {
		DB((env->dbg, LEVEL_3, "root node\n"));
		plist_insert_back(ba->root_nodes, irn);
	}
	else {
		DB((env->dbg, LEVEL_3, "normal node\n"));
	}

	/* record number of all consumer and the consumer within the same block */
	node = get_ilpsched_irn(env, irn);
	na   = get_ilpsched_node_attr(node);
	na->n_consumer     = n_consumer;
	na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
	memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
	DEL_ARR_F(consumer);
}

/**
 * Calculate the ASAP scheduling step for current irn.
 */
static void calculate_irn_asap(ir_node *irn, void *walk_env)
{
	be_ilpsched_env_t     *env = walk_env;
	int                   i;
	ir_node               *block;
	be_ilpsched_irn_t     *node, *block_node;
	ilpsched_node_attr_t  *na;
	ilpsched_block_attr_t *ba;

	/* These nodes are handled separate */
	if (! consider_for_sched(env->arch_env, irn))
		return;

	DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));

	block    = get_nodes_block(irn);
	node     = get_ilpsched_irn(env, irn);
	na       = get_ilpsched_node_attr(node);
	na->asap = 1;

	for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
		ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));

		/* check for greatest distance to top */
		if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
			be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
			ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
			unsigned             lat;

			lat         = fixed_latency(env->sel, pred, env->block_env);
			na->latency = lat;
			na->asap    = MAX(na->asap, pna->asap + lat);
		}
	}

	/* add node to ILP node list and update max_steps */
	block_node = get_ilpsched_irn(env, block);
	ba         = get_ilpsched_block_attr(block_node);

	set_irn_link(irn, ba->head_ilp_nodes);
	ba->head_ilp_nodes = irn;
	ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);

	DB((env->dbg, LEVEL_2, "%u\n", na->asap));
}

/**
 * Calculate the ALAP scheduling step of all irns in current block.
 * Depends on max_steps being calculated.
 */
static void calculate_block_alap(ir_node *block, void *walk_env)
{
	be_ilpsched_env_t     *env        = walk_env;
	be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
	ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
	waitq                 *cur_queue  = new_waitq();
	plist_element_t       *el;

	assert(is_Block(block));

	DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
		block, ba->n_interesting_nodes, ba->max_steps));

	/* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
	//phase_reinit_block_irn_data(&env->ph, block);

	/* init start queue */
	foreach_plist(ba->root_nodes, el) {
		waitq_put(cur_queue, plist_element_get_value(el));
	}

	/* repeat until all nodes are processed */
	while (! waitq_empty(cur_queue)) {
		waitq *next_queue = new_waitq();

		/* process all nodes in current step */
		while (! waitq_empty(cur_queue)) {
			ir_node              *cur_irn = waitq_get(cur_queue);
			be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
			ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
			int                  i;

			/* cur_node has no alap set -> it's a root node, set to max alap */
			if (na->alap == 0) {
				na->alap = ba->max_steps;
				DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
					cur_irn, na->alap));
			}
			else {
				DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
					cur_irn, na->alap));
			}

			/* set the alap's of all predecessors */
			for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
				ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));

				/* check for greatest distance to bottom */
				if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
					be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
					ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
					unsigned             lat;

					/* mark the predecessor as visited by current irn */
					if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
						continue;
					pna->visit_idx = get_irn_idx(cur_irn);

					lat = fixed_latency(env->sel, pred, env->block_env);

					/* set ALAP of current pred */
					if (pna->alap == 0) {
						/* current ALAP is 0: node has not yet been visited */
						pna->alap_changed = 1;
						pna->alap         = na->alap - lat;
					}
					else if (pna->alap > na->alap - lat) {
						/* we found a longer path to root node: change ALAP */
						pna->alap         = na->alap - lat;
						pna->alap_changed = 1;
					}
					else {
						/* current ALAP is best found so far: keep it */
						pna->alap_changed = 0;
					}

					DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));

					/* enqueue node for next iteration */
					if (get_irn_ins_or_deps(pred) > 0)
						waitq_put(next_queue, pred);
				}
			}
		}

		/* prepare for next iteration */
		del_waitq(cur_queue);
		cur_queue = next_queue;
	}
}

/**
 * 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;
	be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
	ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);

	plist_free(ba->root_nodes);
	ba->root_nodes = NULL;
	del_pset(ba->livein_nodes);
	ba->livein_nodes = NULL;
}

/**
 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
 */
static void refine_asap_alap_times(ir_node *irn, void *walk_env)
{
	be_ilpsched_env_t    *env  = walk_env;
	ir_node              *pred = irn;
	be_ilpsched_irn_t    *node, *pred_node;
	ilpsched_node_attr_t *na, *pna;

	if (! consider_for_sched(env->arch_env, irn))
		return;

	if (! is_Proj(irn) && ! be_is_Keep(irn))
		return;

	/* go to the ancestor */
	if (be_is_Keep(irn))
		pred = get_irn_n(irn, 0);
	pred = skip_Proj(pred);

	node      = get_ilpsched_irn(env, irn);
	pred_node = get_ilpsched_irn(env, pred);
	na        = get_ilpsched_node_attr(node);
	pna       = get_ilpsched_node_attr(pred_node);

	na->asap = pna->asap;
	na->alap = pna->alap;

	/* record all Projs and Keeps for this node */
	if (! pna->projkeeps)
		pna->projkeeps = new_waitq();
	waitq_put(pna->projkeeps, irn);

	DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
}

/*******************************************
 *           _              _       _
 *          | |            | |     | |
 *  ___  ___| |__   ___  __| |_   _| | ___
 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
 * \__ \ (__| | | |  __/ (_| | |_| | |  __/
 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
 *
 *******************************************/

static inline void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn)
{
	const ir_edge_t *edge;
        (void) block;

	foreach_out_edge(irn, edge) {
		ir_node *user = get_edge_src_irn(edge);

		if (be_is_Keep(user)) {
			assert(get_nodes_block(user) == block && "Keep must not be in different block.");
			waitq_put(keeps, user);
		}
	}
}

/**
 * Inserts @p irn before @p before into schedule and notifies backend.
 */
static inline void notified_sched_add_before(be_ilpsched_env_t *env,
	const ir_node *before, const ir_node *irn, unsigned cycle)
{
	be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
	sched_add_before((ir_node*) before, (ir_node*) irn);
}

/**
 * Adds a node, it's Projs (in case of mode_T nodes) and
 * it's Keeps to schedule.
 */
static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle)
{
	const ir_edge_t *edge;
	waitq           *keeps = new_waitq();

	/* mode_M nodes are not scheduled */
	if (get_irn_mode(irn) == mode_M)
		return;

	if (! sched_is_scheduled(irn))
		notified_sched_add_before(env, block, irn, cycle);

	/* add Projs */
	if (get_irn_mode(irn) == mode_T) {
		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) &&
				get_irn_n_edges(user) > 0)
			{
				notified_sched_add_before(env, block, user, cycle);
			}

			check_for_keeps(keeps, block, user);
		}
	}
	else {
		check_for_keeps(keeps, block, irn);
	}

	/* add Keeps */
	while (! waitq_empty(keeps)) {
		ir_node *keep = waitq_get(keeps);
		if (! sched_is_scheduled(keep))
			notified_sched_add_before(env, block, keep, cycle);
	}

	del_waitq(keeps);
}

/**
 * Schedule all nodes in the given block, according to the ILP solution.
 */
static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block)
{
	be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
	ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
	be_ilpsched_irn_t     **sched_nodes;
	unsigned              i, l;
	ir_node               *cfop, *irn;
	const ir_edge_t       *edge;

	/* init block schedule list */
	sched_init_block(block);

	/* collect nodes and their scheduling time step */
	sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
	if (ba->n_interesting_nodes == 0) {
		/* ignore */
	}
	else if (ba->n_interesting_nodes == 1) {
		be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);

		/* add the single node */
		ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
	}
	else {
		/* check all nodes for their positive solution */
		foreach_linked_irns(ba->head_ilp_nodes, irn) {
			be_ilpsched_irn_t    *node;
			ilpsched_node_attr_t *na;
			int                  tp_idx, found;
			unsigned             cur_var, t;

			node    = get_ilpsched_irn(env, irn);
			na      = get_ilpsched_node_attr(node);
			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) {
					double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);

					/* check, if variable is set to one (it's not zero then :) */
					if (! LPP_VALUE_IS_0(val)) {
						na->sched_point = t;
						ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
						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;
					}
				}
			}
		}

		glob_heights = env->height;
		/* sort nodes ascending by scheduling time step */
		qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
	}

	/* make all Phis ready and remember the single cf op */
	cfop = NULL;
	foreach_out_edge(block, edge) {
		irn = get_edge_src_irn(edge);

		switch (get_irn_opcode(irn)) {
			case iro_Phi:
				add_to_sched(env, block, irn, 0);
				break;
			case beo_Start:
			case iro_End:
			case iro_Proj:
			case iro_Bad:
			case iro_Unknown:
				break;
			default:
				if (is_cfop(irn)) {
					assert(cfop == NULL && "Highlander - there can be only one");
					cfop = irn;
				}
			break;
		}
	}

	/* add all nodes from list */
	for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
		ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
		if (sched_nodes[i]->irn != cfop)
			add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
	}

	/* schedule control flow node if not already done */
	if (cfop && ! sched_is_scheduled(cfop))
		add_to_sched(env, block, cfop, 0);

	DEL_ARR_F(sched_nodes);
}

/***************************************************************
 *   _____ _      _____     _____           _   _
 *  |_   _| |    |  __ \   / ____|         | | (_)
 *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
 *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
 *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
 *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
 *
 ***************************************************************/

/**
 * Check if node can be executed on given unit type.
 */
static inline int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node)
{
	int                  i;
	ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);

	for (i = na->n_unit_types - 1; i >= 0; --i) {
		if (na->type_info[i].tp == tp)
			return i;
	}

	return -1;
}

/************************************************
 *                   _       _     _
 *                  (_)     | |   | |
 *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
 *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
 *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
 *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
 *
 ************************************************/

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_env_get_allowed_execution_units(env->arch_env, 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, const ir_node *irn, const 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
 *
 * - a_{nt}^k    binary     weighted with num_nodes
 *      node n is alive at time step t on unit type k
 *
 * - y_{nt}^k    continuous  weighted with: num_nodes^2
 *      register pressure over limit for unit type k
 * ==>> These variables represent the register pressure
 *
 */
static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst)
{
	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;

	stat_ev_tim_push();

	num_block_var = num_nodes = 0;
	foreach_linked_irns(ba->head_ilp_nodes, irn) {
		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;

		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));
		memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));

		/* we need these variables only for "real" nodes */
		if (! na->is_dummy_node) {
			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;
			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",
			irn, na->asap, na->alap, na->n_unit_types));

		cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
		/* create variables */
		for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
			unsigned t;

			for (t = na->asap - 1; t <= na->alap - 1; ++t) {
				/* x_{nt}^k variables */
				snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
					get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
				na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
				DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
				/* variable counter */
				n_var++;
				num_block_var++;

				if (! na->is_dummy_node) {
					/* 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_continous, (double)(weigth_y));
					DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));

					/* variable counter */
					n_var++;
					num_block_var++;
				}
			}

			/* a node can die at any step t: asap(n) <= t <= U */
			if (! na->is_dummy_node) {
				for (t = na->asap - 1; t <= ba->max_steps; ++t) {

					/* 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 */
					n_var++;
					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, pred)) {
					be_ilpsched_set_type_info(env, pred, var_obst);
					if (! na->is_dummy_node) {
						ilp_livein_node_t *entry = OALLOC(var_obst, ilp_livein_node_t);
						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++;
			}
		}
	}

	stat_ev_tim_pop("beilpsched_var");
}

/*******************************************************
 *                      _             _       _
 *                     | |           (_)     | |
 *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
 *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
 *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
 *
 *******************************************************/

/**
 * 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 void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *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);
				ir_nodeset_insert(deps, src);
			}
		}
		else {
			ir_nodeset_insert(deps, p);
		}
	}
}

/**
 * Create following ILP constraints:
 * - the assignment constraints:
 *     assure each node is executed once by exactly one (allowed) execution unit
 * - the dead node assignment constraints:
 *     assure a node can only die at most once
 * - the precedence constraints:
 *     assure that no data dependencies are violated
 */
static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node)
{
	unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
	char                  buf[1024];
	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);

	num_cst_assign = num_cst_prec = num_cst_dead = 0;
	foreach_linked_irns(ba->head_ilp_nodes, irn) {
		int                  cst, tp_idx;
		unsigned             cur_var;
		be_ilpsched_irn_t    *node;
		ilpsched_node_attr_t *na;
		ir_node              *pred;
		ir_nodeset_t          deps;
		ir_nodeset_iterator_t iter;

		ir_nodeset_init(&deps);

		node    = get_ilpsched_irn(env, irn);
		na      = get_ilpsched_node_attr(node);
		cur_var = 0;

		/* the assignment constraint */
		stat_ev_tim_push();
		snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
		cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
		DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
		num_cst_assign++;

		lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
		stat_ev_tim_pop("beilpsched_cst_assign");