ldstopt.c

/*
 * Copyright (C) 1995-2011 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   Load/Store optimizations.
 * @author  Michael Beck
 * @version $Id$
 */
#include "config.h"

#include <string.h>

#include "iroptimize.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irmode_t.h"
#include "iropt_t.h"
#include "ircons_t.h"
#include "irgmod.h"
#include "irgwalk.h"
#include "tv_t.h"
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
#include "array_t.h"
#include "irhooks.h"
#include "iredges.h"
#include "irpass.h"
#include "opt_polymorphy.h"
#include "irmemory.h"
#include "irphase_t.h"
#include "irgopt.h"
#include "set.h"
#include "be.h"
#include "debug.h"
#include "opt_manage.h"

/** The debug handle. */
DEBUG_ONLY(static firm_dbg_module_t *dbg;)

#undef IMAX
#define IMAX(a,b)   ((a) > (b) ? (a) : (b))

#define MAX_PROJ    IMAX(IMAX((long)pn_Load_max, (long)pn_Store_max), (long)pn_Call_max)

enum changes_t {
	DF_CHANGED = 1,       /**< data flow changed */
	CF_CHANGED = 2,       /**< control flow changed */
};

/**
 * walker environment
 */
typedef struct walk_env_t {
	struct obstack obst;          /**< list of all stores */
	unsigned changes;             /**< a bitmask of graph changes */
} walk_env_t;

/** A Load/Store info. */
typedef struct ldst_info_t {
	ir_node  *projs[MAX_PROJ+1];  /**< list of Proj's of this node */
	ir_node  *exc_block;          /**< the exception block if available */
	int      exc_idx;             /**< predecessor index in the exception block */
	unsigned visited;             /**< visited counter for breaking loops */
} ldst_info_t;

/**
 * flags for control flow.
 */
enum block_flags_t {
	BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
	BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
};

/**
 * a Block info.
 */
typedef struct block_info_t {
	unsigned flags;               /**< flags for the block */
} block_info_t;

/** the master visited flag for loop detection. */
static unsigned master_visited = 0;

#define INC_MASTER()       ++master_visited
#define MARK_NODE(info)    (info)->visited = master_visited
#define NODE_VISITED(info) (info)->visited >= master_visited

/**
 * get the Load/Store info of a node
 */
static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
{
	ldst_info_t *info = (ldst_info_t*)get_irn_link(node);

	if (! info) {
		info = OALLOCZ(obst, ldst_info_t);
		set_irn_link(node, info);
	}
	return info;
}  /* get_ldst_info */

/**
 * get the Block info of a node
 */
static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
{
	block_info_t *info = (block_info_t*)get_irn_link(node);

	if (! info) {
		info = OALLOCZ(obst, block_info_t);
		set_irn_link(node, info);
	}
	return info;
}  /* get_block_info */

/**
 * update the projection info for a Load/Store
 */
static unsigned update_projs(ldst_info_t *info, ir_node *proj)
{
	long nr = get_Proj_proj(proj);

	assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");

	if (info->projs[nr]) {
		/* there is already one, do CSE */
		exchange(proj, info->projs[nr]);
		return DF_CHANGED;
	}
	else {
		info->projs[nr] = proj;
		return 0;
	}
}  /* update_projs */

/**
 * update the exception block info for a Load/Store node.
 *
 * @param info   the load/store info struct
 * @param block  the exception handler block for this load/store
 * @param pos    the control flow input of the block
 */
static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
{
	assert(info->exc_block == NULL && "more than one exception block found");

	info->exc_block = block;
	info->exc_idx   = pos;
	return 0;
}  /* update_exc */

/** Return the number of uses of an address node */
#define get_irn_n_uses(adr)     get_irn_n_edges(adr)

/**
 * walker, collects all Load/Store/Proj nodes
 *
 * walks from Start -> End
 */
static void collect_nodes(ir_node *node, void *env)
{
	walk_env_t  *wenv   = (walk_env_t *)env;
	unsigned     opcode = get_irn_opcode(node);
	ir_node     *pred, *blk, *pred_blk;
	ldst_info_t *ldst_info;

	if (opcode == iro_Proj) {
		pred   = get_Proj_pred(node);
		opcode = get_irn_opcode(pred);

		if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
			ldst_info = get_ldst_info(pred, &wenv->obst);

			wenv->changes |= update_projs(ldst_info, node);

			/*
			 * Place the Proj's to the same block as the
			 * predecessor Load. This is always ok and prevents
			 * "non-SSA" form after optimizations if the Proj
			 * is in a wrong block.
			 */
			blk      = get_nodes_block(node);
			pred_blk = get_nodes_block(pred);
			if (blk != pred_blk) {
				wenv->changes |= DF_CHANGED;
				set_nodes_block(node, pred_blk);
			}
		}
	} else if (opcode == iro_Block) {
		int i;

		for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
			ir_node      *pred_block, *proj;
			block_info_t *bl_info;
			int          is_exc = 0;

			pred = proj = get_Block_cfgpred(node, i);

			if (is_Proj(proj)) {
				pred   = get_Proj_pred(proj);
				is_exc = is_x_except_Proj(proj);
			}

			/* ignore Bad predecessors, they will be removed later */
			if (is_Bad(pred))
				continue;

			pred_block = get_nodes_block(pred);
			bl_info    = get_block_info(pred_block, &wenv->obst);

			if (is_fragile_op(pred) && is_exc)
				bl_info->flags |= BLOCK_HAS_EXC;
			else if (is_irn_forking(pred))
				bl_info->flags |= BLOCK_HAS_COND;

			opcode = get_irn_opcode(pred);
			if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
				ldst_info = get_ldst_info(pred, &wenv->obst);

				wenv->changes |= update_exc(ldst_info, node, i);
			}
		}
	}
}  /* collect_nodes */

/**
 * Returns an entity if the address ptr points to a constant one.
 *
 * @param ptr  the address
 *
 * @return an entity or NULL
 */
static ir_entity *find_constant_entity(ir_node *ptr)
{
	for (;;) {
		if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
			return get_SymConst_entity(ptr);
		} else if (is_Sel(ptr)) {
			ir_entity *ent = get_Sel_entity(ptr);
			ir_type   *tp  = get_entity_owner(ent);

			/* Do not fiddle with polymorphism. */
			if (is_Class_type(get_entity_owner(ent)) &&
				((get_entity_n_overwrites(ent)    != 0) ||
				(get_entity_n_overwrittenby(ent) != 0)   ) )
				return NULL;

			if (is_Array_type(tp)) {
				/* check bounds */
				int i, n;

				for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
					ir_node   *bound;
					ir_tarval *tlower, *tupper;
					ir_node   *index = get_Sel_index(ptr, i);
					ir_tarval *tv    = computed_value(index);

					/* check if the index is constant */
					if (tv == tarval_bad)
						return NULL;

					bound  = get_array_lower_bound(tp, i);
					tlower = computed_value(bound);
					bound  = get_array_upper_bound(tp, i);
					tupper = computed_value(bound);

					if (tlower == tarval_bad || tupper == tarval_bad)
						return NULL;

					if (tarval_cmp(tv, tlower) == ir_relation_less)
						return NULL;
					if (tarval_cmp(tupper, tv) == ir_relation_less)
						return NULL;

					/* ok, bounds check finished */
				}
			}

			if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
				return ent;

			/* try next */
			ptr = get_Sel_ptr(ptr);
		} else if (is_Add(ptr)) {
			ir_node *l = get_Add_left(ptr);
			ir_node *r = get_Add_right(ptr);

			if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
				ptr = l;
			else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
				ptr = r;
			else
				return NULL;

			/* for now, we support only one addition, reassoc should fold all others */
			if (! is_SymConst(ptr) && !is_Sel(ptr))
				return NULL;
		} else if (is_Sub(ptr)) {
			ir_node *l = get_Sub_left(ptr);
			ir_node *r = get_Sub_right(ptr);

			if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
				ptr = l;
			else
				return NULL;
			/* for now, we support only one substraction, reassoc should fold all others */
			if (! is_SymConst(ptr) && !is_Sel(ptr))
				return NULL;
		} else
			return NULL;
	}
}  /* find_constant_entity */

/**
 * Return the Selection index of a Sel node from dimension n
 */
static long get_Sel_array_index_long(ir_node *n, int dim)
{
	ir_node *index = get_Sel_index(n, dim);
	assert(is_Const(index));
	return get_tarval_long(get_Const_tarval(index));
}  /* get_Sel_array_index_long */

/**
 * Returns the accessed component graph path for an
 * node computing an address.
 *
 * @param ptr    the node computing the address
 * @param depth  current depth in steps upward from the root
 *               of the address
 */
static compound_graph_path *rec_get_accessed_path(ir_node *ptr, size_t depth)
{
	compound_graph_path *res = NULL;
	ir_entity           *root, *field, *ent;
	size_t              path_len, pos, idx;
	ir_tarval           *tv;
	ir_type             *tp;

	if (is_SymConst(ptr)) {
		/* a SymConst. If the depth is 0, this is an access to a global
		 * entity and we don't need a component path, else we know
		 * at least its length.
		 */
		assert(get_SymConst_kind(ptr) == symconst_addr_ent);
		root = get_SymConst_entity(ptr);
		res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
	} else if (is_Sel(ptr)) {
		/* it's a Sel, go up until we find the root */
		res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
		if (res == NULL)
			return NULL;

		/* fill up the step in the path at the current position */
		field    = get_Sel_entity(ptr);
		path_len = get_compound_graph_path_length(res);
		pos      = path_len - depth - 1;
		set_compound_graph_path_node(res, pos, field);

		if (is_Array_type(get_entity_owner(field))) {
			assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
			set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
		}
	} else if (is_Add(ptr)) {
		ir_mode   *mode;
		ir_tarval *tmp;

		{
			ir_node   *l    = get_Add_left(ptr);
			ir_node   *r    = get_Add_right(ptr);
			if (is_Const(r) && get_irn_mode(l) == get_irn_mode(ptr)) {
				ptr = l;
				tv  = get_Const_tarval(r);
			} else {
				ptr = r;
				tv  = get_Const_tarval(l);
			}
		}
ptr_arith:
		mode = get_tarval_mode(tv);
		tmp  = tv;

		/* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
		if (is_Sel(ptr)) {
			field = get_Sel_entity(ptr);
		} else {
			field = get_SymConst_entity(ptr);
		}
		idx = 0;
		for (ent = field;;) {
			unsigned   size;
			ir_tarval *sz, *tv_index, *tlower, *tupper;
			ir_node   *bound;

			tp = get_entity_type(ent);
			if (! is_Array_type(tp))
				break;
			ent = get_array_element_entity(tp);
			size = get_type_size_bytes(get_entity_type(ent));
			sz   = new_tarval_from_long(size, mode);

			tv_index = tarval_div(tmp, sz);
			tmp      = tarval_mod(tmp, sz);

			if (tv_index == tarval_bad || tmp == tarval_bad)
				return NULL;

			assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
			bound  = get_array_lower_bound(tp, 0);
			tlower = computed_value(bound);
			bound  = get_array_upper_bound(tp, 0);
			tupper = computed_value(bound);

			if (tlower == tarval_bad || tupper == tarval_bad)
				return NULL;

			if (tarval_cmp(tv_index, tlower) == ir_relation_less)
				return NULL;
			if (tarval_cmp(tupper, tv_index) == ir_relation_less)
				return NULL;

			/* ok, bounds check finished */
			++idx;
		}
		if (! tarval_is_null(tmp)) {
			/* access to some struct/union member */
			return NULL;
		}

		/* should be at least ONE array */
		if (idx == 0)
			return NULL;

		res = rec_get_accessed_path(ptr, depth + idx);
		if (res == NULL)
			return NULL;

		path_len = get_compound_graph_path_length(res);
		pos      = path_len - depth - idx;

		for (ent = field;;) {
			unsigned   size;
			ir_tarval *sz, *tv_index;
			long       index;

			tp = get_entity_type(ent);
			if (! is_Array_type(tp))
				break;
			ent = get_array_element_entity(tp);
			set_compound_graph_path_node(res, pos, ent);

			size = get_type_size_bytes(get_entity_type(ent));
			sz   = new_tarval_from_long(size, mode);

			tv_index = tarval_div(tv, sz);
			tv       = tarval_mod(tv, sz);

			/* worked above, should work again */
			assert(tv_index != tarval_bad && tv != tarval_bad);

			/* bounds already checked above */
			index = get_tarval_long(tv_index);
			set_compound_graph_path_array_index(res, pos, index);
			++pos;
		}
	} else if (is_Sub(ptr)) {
		ir_node *l = get_Sub_left(ptr);
		ir_node *r = get_Sub_right(ptr);

		ptr = l;
		tv  = get_Const_tarval(r);
		tv  = tarval_neg(tv);
		goto ptr_arith;
	}
	return res;
}  /* rec_get_accessed_path */

/**
 * Returns an access path or NULL.  The access path is only
 * valid, if the graph is in phase_high and _no_ address computation is used.
 */
static compound_graph_path *get_accessed_path(ir_node *ptr)
{
	compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
	return gr;
}  /* get_accessed_path */

typedef struct path_entry {
	ir_entity         *ent;
	struct path_entry *next;
	size_t            index;
} path_entry;

static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
{
	path_entry       entry, *p;
	ir_entity        *ent, *field;
	ir_initializer_t *initializer;
	ir_tarval        *tv;
	ir_type          *tp;
	size_t           n;

	entry.next = next;
	if (is_SymConst(ptr)) {
		/* found the root */
		ent         = get_SymConst_entity(ptr);
		initializer = get_entity_initializer(ent);
		for (p = next; p != NULL;) {
			if (initializer->kind != IR_INITIALIZER_COMPOUND)
				return NULL;
			n  = get_initializer_compound_n_entries(initializer);
			tp = get_entity_type(ent);

			if (is_Array_type(tp)) {
				ent = get_array_element_entity(tp);
				if (ent != p->ent) {
					/* a missing [0] */
					if (0 >= n)
						return NULL;
					initializer = get_initializer_compound_value(initializer, 0);
					continue;
				}
			}
			if (p->index >= n)
				return NULL;
			initializer = get_initializer_compound_value(initializer, p->index);

			ent = p->ent;
			p   = p->next;
		}
		tp = get_entity_type(ent);
		while (is_Array_type(tp)) {
			ent = get_array_element_entity(tp);
			tp = get_entity_type(ent);
			/* a missing [0] */
			n  = get_initializer_compound_n_entries(initializer);
			if (0 >= n)
				return NULL;
			initializer = get_initializer_compound_value(initializer, 0);
		}

		switch (initializer->kind) {
		case IR_INITIALIZER_CONST:
			return get_initializer_const_value(initializer);
		case IR_INITIALIZER_TARVAL:
		case IR_INITIALIZER_NULL:
		default:
			return NULL;
		}
	} else if (is_Sel(ptr)) {
		entry.ent = field = get_Sel_entity(ptr);
		tp = get_entity_owner(field);
		if (is_Array_type(tp)) {
			assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
			entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
		} else {
			size_t i, n_members = get_compound_n_members(tp);
			for (i = 0; i < n_members; ++i) {
				if (get_compound_member(tp, i) == field)
					break;
			}
			if (i >= n_members) {
				/* not found: should NOT happen */
				return NULL;
			}
			entry.index = i;
		}
		return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
	}  else if (is_Add(ptr)) {
		ir_mode  *mode;
		unsigned pos;

		{
			ir_node *l = get_Add_left(ptr);
			ir_node *r = get_Add_right(ptr);
			if (is_Const(r)) {
				ptr = l;
				tv  = get_Const_tarval(r);
			} else {
				ptr = r;
				tv  = get_Const_tarval(l);
			}
		}
ptr_arith:
		mode = get_tarval_mode(tv);

		/* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
		if (is_Sel(ptr)) {
			field = get_Sel_entity(ptr);
		} else {
			field = get_SymConst_entity(ptr);
		}

		/* count needed entries */
		pos = 0;
		for (ent = field;;) {
			tp = get_entity_type(ent);
			if (! is_Array_type(tp))
				break;
			ent = get_array_element_entity(tp);
			++pos;
		}
		/* should be at least ONE entry */
		if (pos == 0)
			return NULL;

		/* allocate the right number of entries */
		NEW_ARR_A(path_entry, p, pos);

		/* fill them up */
		pos = 0;
		for (ent = field;;) {
			unsigned   size;
			ir_tarval *sz, *tv_index, *tlower, *tupper;
			long       index;
			ir_node   *bound;

			tp = get_entity_type(ent);
			if (! is_Array_type(tp))
				break;
			ent = get_array_element_entity(tp);
			p[pos].ent  = ent;
			p[pos].next = &p[pos + 1];

			size = get_type_size_bytes(get_entity_type(ent));
			sz   = new_tarval_from_long(size, mode);

			tv_index = tarval_div(tv, sz);
			tv       = tarval_mod(tv, sz);

			if (tv_index == tarval_bad || tv == tarval_bad)
				return NULL;

			assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
			bound  = get_array_lower_bound(tp, 0);
			tlower = computed_value(bound);
			bound  = get_array_upper_bound(tp, 0);
			tupper = computed_value(bound);

			if (tlower == tarval_bad || tupper == tarval_bad)
				return NULL;

			if (tarval_cmp(tv_index, tlower) == ir_relation_less)
				return NULL;
			if (tarval_cmp(tupper, tv_index) == ir_relation_less)
				return NULL;

			/* ok, bounds check finished */
			index = get_tarval_long(tv_index);
			p[pos].index = index;
			++pos;
		}
		if (! tarval_is_null(tv)) {
			/* hmm, wrong access */
			return NULL;
		}
		p[pos - 1].next = next;
		return rec_find_compound_ent_value(ptr, p);
	} else if (is_Sub(ptr)) {
		ir_node *l = get_Sub_left(ptr);
		ir_node *r = get_Sub_right(ptr);

		ptr = l;
		tv  = get_Const_tarval(r);
		tv  = tarval_neg(tv);
		goto ptr_arith;
	}
	return NULL;
}

static ir_node *find_compound_ent_value(ir_node *ptr)
{
	return rec_find_compound_ent_value(ptr, NULL);
}

/* forward */
static void reduce_adr_usage(ir_node *ptr);

/**
 * Update a Load that may have lost its users.
 */
static void handle_load_update(ir_node *load)
{
	ldst_info_t *info = (ldst_info_t*)get_irn_link(load);

	/* do NOT touch volatile loads for now */
	if (get_Load_volatility(load) == volatility_is_volatile)
		return;

	if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
		ir_node *ptr = get_Load_ptr(load);
		ir_node *mem = get_Load_mem(load);

		/* a Load whose value is neither used nor exception checked, remove it */
		exchange(info->projs[pn_Load_M], mem);
		if (info->projs[pn_Load_X_regular])
			exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
		kill_node(load);
		reduce_adr_usage(ptr);
	}
}  /* handle_load_update */

/**
 * A use of an address node has vanished. Check if this was a Proj
 * node and update the counters.
 */
static void reduce_adr_usage(ir_node *ptr)
{
	ir_node *pred;
	if (!is_Proj(ptr))
		return;
	if (get_irn_n_edges(ptr) > 0)
		return;

	/* this Proj is dead now */
	pred = get_Proj_pred(ptr);
	if (is_Load(pred)) {
		ldst_info_t *info = (ldst_info_t*)get_irn_link(pred);
		info->projs[get_Proj_proj(ptr)] = NULL;

		/* this node lost its result proj, handle that */
		handle_load_update(pred);
	}
}  /* reduce_adr_usage */

/**
 * Check, if an already existing value of mode old_mode can be converted
 * into the needed one new_mode without loss.
 */
static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
{
	unsigned old_size;
	unsigned new_size;
	if (old_mode == new_mode)
		return true;

	old_size = get_mode_size_bits(old_mode);
	new_size = get_mode_size_bits(new_mode);

	/* if both modes are two-complement ones, we can always convert the
	   Stored value into the needed one. (on big endian machines we currently
	   only support this for modes of same size) */
	if (old_size >= new_size &&
		  get_mode_arithmetic(old_mode) == irma_twos_complement &&
		  get_mode_arithmetic(new_mode) == irma_twos_complement &&
		  (!be_get_backend_param()->byte_order_big_endian
	        || old_size == new_size)) {
		return true;
	}
	return false;
}

/**
 * Check whether a Call is at least pure, i.e. does only read memory.
 */
static unsigned is_Call_pure(ir_node *call)
{
	ir_type *call_tp = get_Call_type(call);
	unsigned prop = get_method_additional_properties(call_tp);

	/* check first the call type */
	if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
		/* try the called entity */
		ir_node *ptr = get_Call_ptr(call);

		if (is_Global(ptr)) {
			ir_entity *ent = get_Global_entity(ptr);

			prop = get_entity_additional_properties(ent);
		}
	}
	return (prop & (mtp_property_const|mtp_property_pure)) != 0;
}  /* is_Call_pure */

static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
{
	ir_mode *mode  = get_irn_mode(ptr);
	long    offset = 0;

	/* TODO: long might not be enough, we should probably use some tarval thingy... */
	for (;;) {
		if (is_Add(ptr)) {
			ir_node *l = get_Add_left(ptr);
			ir_node *r = get_Add_right(ptr);

			if (get_irn_mode(l) != mode || !is_Const(r))
				break;

			offset += get_tarval_long(get_Const_tarval(r));
			ptr     = l;
		} else if (is_Sub(ptr)) {
			ir_node *l = get_Sub_left(ptr);
			ir_node *r = get_Sub_right(ptr);

			if (get_irn_mode(l) != mode || !is_Const(r))
				break;

			offset -= get_tarval_long(get_Const_tarval(r));
			ptr     = l;
		} else if (is_Sel(ptr)) {
			ir_entity *ent = get_Sel_entity(ptr);
			ir_type   *tp  = get_entity_owner(ent);

			if (is_Array_type(tp)) {
				int     size;
				ir_node *index;

				/* only one dimensional arrays yet */
				if (get_Sel_n_indexs(ptr) != 1)
					break;
				index = get_Sel_index(ptr, 0);
				if (! is_Const(index))
					break;

				tp = get_entity_type(ent);
				if (get_type_state(tp) != layout_fixed)
					break;

				size    = get_type_size_bytes(tp);
				offset += size * get_tarval_long(get_Const_tarval(index));
			} else {
				if (get_type_state(tp) != layout_fixed)
					break;
				offset += get_entity_offset(ent);
			}
			ptr = get_Sel_ptr(ptr);
		} else
			break;
	}

	*pOffset = offset;
	return ptr;
}

static int try_load_after_store(ir_node *load,
		ir_node *load_base_ptr, long load_offset, ir_node *store)
{
	ldst_info_t *info;
	ir_node *store_ptr      = get_Store_ptr(store);
	long     store_offset;
	ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
	ir_node *store_value;
	ir_mode *store_mode;
	ir_node *load_ptr;
	ir_mode *load_mode;
	long     load_mode_len;
	long     store_mode_len;
	long     delta;
	int      res;

	if (load_base_ptr != store_base_ptr)
		return 0;

	load_mode      = get_Load_mode(load);
	load_mode_len  = get_mode_size_bytes(load_mode);
	store_mode     = get_irn_mode(get_Store_value(store));
	store_mode_len = get_mode_size_bytes(store_mode);
	delta          = load_offset - store_offset;
	store_value    = get_Store_value(store);

	if (delta != 0 || store_mode != load_mode) {
		/* TODO: implement for big-endian */
		if (delta < 0 || delta + load_mode_len > store_mode_len
				|| (be_get_backend_param()->byte_order_big_endian
				    && load_mode_len != store_mode_len))
			return 0;

		if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
			get_mode_arithmetic(load_mode)  != irma_twos_complement)
			return 0;


		/* produce a shift to adjust offset delta */
		if (delta > 0) {
			ir_node *cnst;
			ir_graph *irg = get_irn_irg(load);

			cnst        = new_r_Const_long(irg, mode_Iu, delta * 8);
			store_value = new_r_Shr(get_nodes_block(load),
									store_value, cnst, store_mode);
		}

		/* add an convert if needed */
		if (store_mode != load_mode) {
			store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
		}
	}

	DBG_OPT_RAW(load, store_value);

	info = (ldst_info_t*)get_irn_link(load);
	if (info->projs[pn_Load_M])
		exchange(info->projs[pn_Load_M], get_Load_mem(load));

	res = 0;
	/* no exception */
	if (info->projs[pn_Load_X_except]) {
		ir_graph *irg = get_irn_irg(load);
		exchange( info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
		res |= CF_CHANGED;
	}
	if (info->projs[pn_Load_X_regular]) {
		exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
		res |= CF_CHANGED;
	}

	if (info->projs[pn_Load_res])
		exchange(info->projs[pn_Load_res], store_value);

	load_ptr = get_Load_ptr(load);
	kill_node(load);
	reduce_adr_usage(load_ptr);
	return res | DF_CHANGED;
}

/**
 * Follow the memory chain as long as there are only Loads,
 * alias free Stores, and constant Calls and try to replace the
 * current Load by a previous ones.
 * Note that in unreachable loops it might happen that we reach
 * load again, as well as we can fall into a cycle.
 * We break such cycles using a special visited flag.
 *
 * INC_MASTER() must be called before dive into
 */
static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
{
	unsigned    res = 0;
	ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
	ir_node     *pred;
	ir_node     *ptr       = get_Load_ptr(load);
	ir_node     *mem       = get_Load_mem(load);
	ir_mode     *load_mode = get_Load_mode(load);

	for (pred = curr; load != pred; ) {
		ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);

		/*
		 * a Load immediately after a Store -- a read after write.
		 * We may remove the Load, if both Load & Store does not have an
		 * exception handler OR they are in the same Block. In the latter
		 * case the Load cannot throw an exception when the previous Store was
		 * quiet.
		 *
		 * Why we need to check for Store Exception? If the Store cannot
		 * be executed (ROM) the exception handler might simply jump into
		 * the load Block :-(
		 * We could make it a little bit better if we would know that the
		 * exception handler of the Store jumps directly to the end...
		 */
		if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
				&& info->projs[pn_Load_X_except] == NULL)
				|| get_nodes_block(load) == get_nodes_block(pred)))
		{
			long    load_offset;
			ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
			int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);

			if (changes != 0)
				return res | changes;
		} else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
		           can_use_stored_value(get_Load_mode(pred), load_mode)) {
			/*
			 * a Load after a Load -- a read after read.
			 * We may remove the second Load, if it does not have an exception
			 * handler OR they are in the same Block. In the later case
			 * the Load cannot throw an exception when the previous Load was
			 * quiet.
			 *
			 * Here, there is no need to check if the previous Load has an
			 * exception hander because they would have exact the same
			 * exception...
			 *
			 * TODO: implement load-after-load with different mode for big
			 *       endian
			 */
			if (info->projs[pn_Load_X_except] == NULL
					|| get_nodes_block(load) == get_nodes_block(pred)) {
				ir_node *value;

				DBG_OPT_RAR(load, pred);

				/* the result is used */
				if (info->projs[pn_Load_res]) {
					if (pred_info->projs[pn_Load_res] == NULL) {
						/* create a new Proj again */
						pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
					}
					value = pred_info->projs[pn_Load_res];

					/* add an convert if needed */
					if (get_Load_mode(pred) != load_mode) {
						value = new_r_Conv(get_nodes_block(load), value, load_mode);
					}

					exchange(info->projs[pn_Load_res], value);
				}

				if (info->projs[pn_Load_M])
					exchange(info->projs[pn_Load_M], mem);

				/* no exception */
				if (info->projs[pn_Load_X_except]) {
					ir_graph *irg = get_irn_irg(load);
					exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
					res |= CF_CHANGED;
				}
				if (info->projs[pn_Load_X_regular]) {
					exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
					res |= CF_CHANGED;
				}

				kill_node(load);
				reduce_adr_usage(ptr);
				return res |= DF_CHANGED;
			}
		}

		if (is_Store(pred)) {
			/* check if we can pass through this store */
			ir_alias_relation rel = get_alias_relation(
				get_Store_ptr(pred),
				get_irn_mode(get_Store_value(pred)),
				ptr, load_mode);
			/* if the might be an alias, we cannot pass this Store */
			if (rel != ir_no_alias)
				break;
			pred = skip_Proj(get_Store_mem(pred));
		} else if (is_Load(pred)) {
			pred = skip_Proj(get_Load_mem(pred));
		} else if (is_Call(pred)) {
			if (is_Call_pure(pred)) {
				/* The called graph is at least pure, so there are no Store's
				   in it. We can handle it like a Load and skip it. */
				pred = skip_Proj(get_Call_mem(pred));
			} else {
				/* there might be Store's in the graph, stop here */
				break;
			}
		} else {
			/* follow only Load chains */
			break;
		}

		/* check for cycles */
		if (NODE_VISITED(pred_info))
			break;
		MARK_NODE(pred_info);
	}

	if (is_Sync(pred)) {
		int i;

		/* handle all Sync predecessors */
		for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
			res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred,