ldstopt.c

/*
 * This file is part of libFirm.
 * Copyright (C) 2012 University of Karlsruhe.
 */

/**
 * @file
 * @brief   Load/Store optimizations.
 * @author  Michael Beck
 */
#include <string.h>

#include "entity_t.h"
#include "error.h"
#include "iroptimize.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irmode_t.h"
#include "type_t.h"
#include "iropt_t.h"
#include "ircons_t.h"
#include "irgmod.h"
#include "irgwalk.h"
#include "irtools.h"
#include "tv_t.h"
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
#include "array.h"
#include "irhooks.h"
#include "iredges.h"
#include "irmemory.h"
#include "irnodehashmap.h"
#include "irgopt.h"
#include "set.h"
#include "be.h"
#include "debug.h"

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

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

typedef enum changes_t {
	NO_CHANGES = 0,
	DF_CHANGED = (1 << 0), /**< data flow changed */
	CF_CHANGED = (1 << 1), /**< control flow changed */
	/** nodes have been created but are not reachable. This is a bad hack
	 * try hard to avoid it! */
	NODES_CREATED = (1 << 2),
} changes_t;

/**
 * walker environment
 */
typedef struct walk_env_t {
	struct obstack obst;    /**< list of all stores */
	changes_t      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 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;

#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 == NULL) {
		info = OALLOCZ(obst, ldst_info_t);
		set_irn_link(node, info);
	}
	return 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 == NULL) {
		info = OALLOCZ(obst, block_info_t);
		set_irn_link(node, info);
	}
	return info;
}

/**
 * update the projection info for a Load/Store
 */
static changes_t 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 NO_CHANGES;
	}
}

/**
 * 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 void 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;
}

/**
 * walker, collects all Proj/Load/Store/Call/CopyB 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);

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

		if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
			ldst_info_t *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.
			 */
			ir_node *blk      = get_nodes_block(node);
			ir_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) {
		for (int i = get_Block_n_cfgpreds(node); i-- > 0; ) {
			bool     is_exc = false;
			ir_node *proj   = get_Block_cfgpred(node, i);
			ir_node *pred   = proj;

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

			if (is_Proj(proj)) {
				pred   = get_Proj_pred(proj);
				is_exc = is_x_except_Proj(proj);
			}
			ir_node      *pred_block = get_nodes_block(pred);
			block_info_t *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_t *ldst_info = get_ldst_info(pred, &wenv->obst);

				update_exc(ldst_info, node, i);
			}
		}
	} else if (opcode == iro_CopyB) {
		/* Just initialize a ldst_info for the CopyB node. */
		(void) get_ldst_info(node, &wenv->obst);
	}
}

/**
 * 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_Address(ptr)) {
			return get_Address_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(tp) &&
				(get_entity_n_overwrites(ent) > 0
			     || get_entity_n_overwrittenby(ent) > 0))
				return NULL;

			if (is_Array_type(tp)) {
				/* check bounds */
				ir_node   *index = get_Sel_index(ptr, 0);
				ir_tarval *tv    = computed_value(index);

				/* check if the index is constant */
				if (!tarval_is_constant(tv))
					return NULL;

				ir_node   *size  = get_array_size(tp);
				ir_tarval *tsize = computed_value(size);

				if (!tarval_is_constant(tsize))
					return NULL;

				if (tarval_cmp(tv, tsize) != 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;

			/* we support only one addition, reassoc should fold all others */
			if (!is_Address(ptr) && !is_Align(ptr) && !is_Offset(ptr)
			    && !is_Sel(ptr) && !is_Size(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;
			/* we support only one subtraction, reassoc should fold all others */
			if (!is_Address(ptr) && !is_Align(ptr) && !is_Offset(ptr)
			    && !is_Sel(ptr) && !is_Size(ptr))
				return NULL;
		} else {
			return NULL;
		}
	}
}

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

/**
 * Update a Load that may have lost its users.
 */
static void handle_load_update(ir_node *load)
{
	/* do NOT touch volatile loads for now */
	if (get_Load_volatility(load) == volatility_is_volatile)
		return;

	ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
	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]) {
			ir_node *jmp = new_r_Jmp(get_nodes_block(load));
			exchange(info->projs[pn_Load_X_regular], jmp);
		}
		kill_node(load);
		reduce_node_usage(ptr);
	}
}

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

	/* this Proj is dead now */
	ir_node *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);
	}
}

/**
 * Kill a Load or Store and all other nodes which are not needed after
 * it has been killed.
 */
static void kill_and_reduce_usage(ir_node *node)
{
	ir_node *ptr;
	ir_node *value;
	switch(get_irn_opcode(node)) {
	case iro_Load:
		ptr   = get_Load_ptr(node);
		value = NULL;
		break;
	case iro_Store:
		ptr   = get_Store_ptr(node);
		value = get_Store_value(node);
		break;
	default:
		panic("Cannot handle node %+F", node);
	}

	kill_node(node);
	reduce_node_usage(ptr);
	if (value != NULL) {
		reduce_node_usage(value);
	}
}

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

	unsigned old_size = get_mode_size_bits(old_mode);
	unsigned 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 bool 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_entity *callee = get_Call_callee(call);
		if (callee != NULL) {
			prop = get_entity_additional_properties(callee);
		}
	}
	return prop & (mtp_property_const|mtp_property_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)) {
				assert(get_Sel_n_indexs(ptr) == 1);
				ir_node *index = get_Sel_index(ptr, 0);
				if (!is_Const(index))
					break;

				ir_type *element_type = get_entity_type(ent);
				if (get_type_state(element_type) != layout_fixed)
					break;

				int size = get_type_size_bytes(element_type);
				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 changes_t try_load_after_store(ir_node *load, ir_node *load_base_ptr,
                                      long load_offset, ir_node *store)
{
	ir_node *store_ptr      = get_Store_ptr(store);
	long     store_offset;
	ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);

	if (load_base_ptr != store_base_ptr)
		return NO_CHANGES;

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

	/* Ensure that Load is completely contained in Store. */
	if (delta < 0 || delta+load_mode_len > store_mode_len)
		return NO_CHANGES;

	if (store_mode != load_mode) {
		if (get_mode_arithmetic(store_mode) == irma_twos_complement &&
		    get_mode_arithmetic(load_mode)  == irma_twos_complement) {

			/* produce a shift to adjust offset delta */
			unsigned const shift = be_get_backend_param()->byte_order_big_endian
				? store_mode_len - load_mode_len - delta
				: delta;
			if (shift != 0) {
				ir_graph *const irg  = get_irn_irg(load);
				ir_node  *const cnst = new_r_Const_long(irg, mode_Iu, shift * 8);
				store_value = new_r_Shr(get_nodes_block(load),
				                        store_value, cnst, store_mode);
			}

			store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
		} else {
			/* we would need some kind of bitcast node here */
			return NO_CHANGES;
		}
	}

	DBG_OPT_RAW(load, store_value);

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

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

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

	kill_and_reduce_usage(load);
	return res | DF_CHANGED;
}

static ir_node *try_update_ptr_CopyB(ir_node *load, ir_node *load_base_ptr,
                                     long load_offset, ir_node *copyB)
{
	ir_node *copyB_dst = get_CopyB_dst(copyB);
	long     dst_offset;
	ir_node *dst_base_ptr = get_base_and_offset(copyB_dst, &dst_offset);

	if (load_base_ptr != dst_base_ptr) {
		return NULL;
	}

	/*
	 * There are quite a few offsets here, please do not get
	 * confused: The CopyB node copies n_copy bytes from
	 * (src_base_ptr + src_offset) to (dst_base_ptr + dst_offset)
	 * (calculated in bytes; no pointer arithmetic). The Load
	 * loads from (dst_base_ptr + load_offset). This is translated
	 * into load_src_offset, which points to the same data in the
	 * CopyB's source array, if it is within its bounds.
	 */
	ir_type *copyB_type      = get_CopyB_type(copyB);
	long     n_copy          = get_type_size_bytes(copyB_type);

	ir_node *copyB_src       = get_CopyB_src(copyB);
	long     src_offset;
	ir_node *src_base_ptr    = get_base_and_offset(copyB_src, &src_offset);

	long     load_src_offset = src_offset - dst_offset + load_offset;

	ir_mode *load_mode       = get_Load_mode(load);
	long     load_size       = get_mode_size_bytes(load_mode);

	if (load_src_offset + load_size > n_copy || load_src_offset < src_offset) {
		return NULL;
	}

	/*
	 * Everything is OK, we can replace
	 *   ptr = (load_base_ptr + load_offset)
	 * with
	 *   new_load_ptr = (src_base_ptr + load_src_offset)
	 */
	ir_graph *irg          = get_irn_irg(load);
	ir_node  *block        = get_nodes_block(load);
	ir_mode  *mode_ref     = get_irn_mode(src_base_ptr);
	ir_mode  *mode_ref_int = get_reference_mode_unsigned_eq(mode_ref);
	ir_node  *cnst         = new_r_Const_long(irg, mode_ref_int, load_src_offset);
	ir_node  *new_load_ptr = new_r_Add(block, src_base_ptr, cnst, mode_ref);
	return new_load_ptr;
}

/**
 * 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 changes_t follow_Mem_chain(ir_node *load, ir_node *curr)
{
	changes_t    res  = NO_CHANGES;
	ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
	ir_node     *ptr       = get_Load_ptr(load);
	ir_node     *mem       = get_Load_mem(load);
	ir_mode     *load_mode = get_Load_mode(load);

	ir_node *pred = curr;
	while (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);
			changes_t changes  = try_load_after_store(load, base_ptr, load_offset, pred);

			if (changes != NO_CHANGES)
				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)) {
				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);
					}
					ir_node *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_and_reduce_usage(load);
				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_type_for_mode(get_irn_mode(get_Store_value(pred))),
				ptr,
				get_type_for_mode(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 if (is_CopyB(pred)) {
			/*
			 * We cannot replace the Load with another
			 * Load from the CopyB's source directly,
			 * because there may be Stores in between,
			 * destroying the source data. However, we can
			 * use the source address from this point
			 * onwards for further optimizations.
			 */
			long     load_offset;
			ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
			ir_node *new_ptr  = try_update_ptr_CopyB(load, base_ptr, load_offset, pred);

			if (new_ptr) {
				res |= NODES_CREATED;
				ptr = new_ptr;
				/*
				 * Special case: If new_ptr points to
				 * a constant, we *can* replace the
				 * Load immediately.
				 */
				if (find_constant_entity(new_ptr)) {
					set_Load_ptr(load, new_ptr);
					return res | DF_CHANGED;
				}
			} else {
				ir_alias_relation rel = get_alias_relation(
					get_CopyB_dst(pred),
					get_CopyB_type(pred),
					ptr,
					get_type_for_mode(load_mode));
				if (rel != ir_no_alias)
					break;
			}

			pred = skip_Proj(get_CopyB_mem(pred));
		} else {
			/* follow only Load chains */
			break;
		}

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

	if (is_Sync(pred)) {
		/* handle all Sync predecessors */
		for (int i = get_Sync_n_preds(pred); i-- > 0; ) {
			res |=  follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
			if (res & ~NODES_CREATED)
				return res;
		}
	}

	return res;
}

ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
{
	ir_mode  *c_mode = get_irn_mode(c);
	ir_mode  *l_mode = get_Load_mode(load);
	ir_node  *block  = get_nodes_block(load);
	dbg_info *dbgi   = get_irn_dbg_info(load);
	ir_node  *res    = duplicate_subgraph(dbgi, c, block);

	if (c_mode != l_mode) {
		/* check, if the mode matches OR can be easily converted info */
		if (is_reinterpret_cast(c_mode, l_mode)) {
			/* copy the value from the const code irg and cast it */
			res = new_rd_Conv(dbgi, block, res, l_mode);
		} else {
			return NULL;
		}
	}
	return res;
}

/**
 * optimize a Load
 *
 * @param load  the Load node
 */
static changes_t optimize_load(ir_node *load)
{
	ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
	changes_t    res  = NO_CHANGES;

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

	/* the address of the load to be optimized */
	ir_node *ptr = get_Load_ptr(load);

	/* The mem of the Load. Must still be returned after optimization. */
	ir_node *mem = get_Load_mem(load);

	if (info->projs[pn_Load_res] == NULL
			&& info->projs[pn_Load_X_except] == NULL) {
		/* the value is never used and we don't care about exceptions, remove */
		exchange(info->projs[pn_Load_M], mem);

		if (info->projs[pn_Load_X_regular]) {
			/* should not happen, but if it does, remove it */
			exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
			res |= CF_CHANGED;
		}
		kill_and_reduce_usage(load);
		return res | DF_CHANGED;
	}

	/* Check, if the address of this load is used more than once.
	 * If not, more load cannot be removed in any case. */
	long dummy;
	if (get_irn_n_edges(ptr) <= 1 && get_irn_n_edges(get_base_and_offset(ptr, &dummy)) <= 1)
		return res;

	/*
	 * follow the memory chain as long as there are only Loads
	 * and try to replace current Load or Store by a previous one.
	 * 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();
	res = follow_Mem_chain(load, skip_Proj(mem));
	return res;
}

/**
 * Check whether a value of mode new_mode would completely overwrite a value
 * of mode old_mode in memory.
 */
static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
{
	return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
}

/**
 * Check whether small is a part of large (starting at same address).
 */
static int is_partially_same(ir_node *small, ir_node *large)
{
	ir_mode *sm = get_irn_mode(small);
	ir_mode *lm = get_irn_mode(large);

	/* FIXME: Check endianness */
	return is_Conv(small) && get_Conv_op(small) == large
	    && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
	    && get_mode_arithmetic(sm) == irma_twos_complement
	    && get_mode_arithmetic(lm) == irma_twos_complement;
}

/**
 * follow the memory chain as long as there are only Loads and alias free
 * Stores.
 * INC_MASTER() must be called before dive into
 */
static changes_t follow_Mem_chain_for_Store(ir_node *store, ir_node *curr,
                                            bool had_split)
{
	changes_t    res   = NO_CHANGES;
	ldst_info_t *info  = (ldst_info_t*)get_irn_link(store);
	ir_node     *ptr   = get_Store_ptr(store);
	ir_node     *mem   = get_Store_mem(store);
	ir_node     *value = get_Store_value(store);
	ir_mode     *mode  = get_irn_mode(value);
	ir_node     *block = get_nodes_block(store);

	ir_node *pred = curr;
	while (pred != store) {
		ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);

		/*
		 * BEWARE: one might think that checking the modes is useless, because
		 * if the pointers are identical, they refer to the same object.
		 * This is only true in strong typed languages, not is C were the following
		 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
		 * However, if the size of the mode that is written is bigger or equal the
		 * size of the old one, the old value is completely overwritten and can be
		 * killed ...
		 */
		if (is_Store(pred) && !had_split && get_Store_ptr(pred) == ptr &&
		    get_nodes_block(pred) == block) {
			/*
			 * a Store after a Store in the same Block -- a write after write.
			 */

			/*
			 * We may remove the first Store, if the old value is completely
			 * overwritten or the old value is a part of the new value,
			 * and if it does not have an exception handler.
			 *
			 * TODO: What, if both have the same exception handler ???
			 */
			if (get_Store_volatility(pred) != volatility_is_volatile
			        && !pred_info->projs[pn_Store_X_except]) {
				ir_node *predvalue = get_Store_value(pred);
				ir_mode *predmode  = get_irn_mode(predvalue);

				if (is_completely_overwritten(predmode, mode)
				        || is_partially_same(predvalue, value)) {
					DBG_OPT_WAW(pred, store);
					DB((dbg, LEVEL_1, "  killing store %+F (override by %+F)\n", pred, store));
					exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
					kill_and_reduce_usage(pred);
					return DF_CHANGED;
				}
			}

			/*
			 * We may remove the Store, if the old value already contains
			 * the new value, and if it does not have an exception handler.
			 *
			 * TODO: What, if both have the same exception handler ???
			 */
			if (get_Store_volatility(store) != volatility_is_volatile
			        && !info->projs[pn_Store_X_except]) {
				ir_node *predvalue = get_Store_value(pred);

				if (is_partially_same(value, predvalue)) {
					DBG_OPT_WAW(pred, store);
					DB((dbg, LEVEL_1, "  killing store %+F (override by %+F)\n", pred, store));
					exchange(info->projs[pn_Store_M], mem);
					kill_and_reduce_usage(store);
					return DF_CHANGED;
				}
			}
		} else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
		           value == pred_info->projs[pn_Load_res]) {
			/*
			 * a Store of a value just loaded from the same address
			 * -- a write after read.
			 * We may remove the Store, if it does not have an exception
			 * handler.
			 */
			if (!info->projs[pn_Store_X_except]) {
				DBG_OPT_WAR(store, pred);
				DB((dbg, LEVEL_1, "  killing store %+F (read %+F from same address)\n", store, pred));
				exchange(info->projs[pn_Store_M], mem);
				kill_and_reduce_usage(store);
				return 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_type_for_mode(get_irn_mode(get_Store_value(