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

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

#define MAX_PROJ MAX(MAX((unsigned)pn_Load_max, (unsigned)pn_Store_max), (unsigned)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;

typedef struct base_offset_t {
	ir_node *base;
	long     offset;
} base_offset_t;

typedef struct track_load_env_t {
	ir_node      *load;
	base_offset_t base_offset;
	ir_node      *ptr; /* deprecated: alternative representation of
	                      base_offset */
} track_load_env_t;

/**
 * flags for control flow.
 */
typedef enum block_flags_t {
	BLOCK_HAS_COND = (1 << 0), /**< Block has conditional control flow */
	BLOCK_HAS_EXC  = (1 << 1), /**< Block has exceptional control flow */
} block_flags_t;

/**
 * a Block info.
 */
typedef struct block_info_t {
	block_flags_t 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)
{
	unsigned nr = get_Proj_num(proj);
	assert(nr <= MAX_PROJ);

	if (info->projs[nr] != NULL) {
		/* 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);
	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 (is_memop(node)) {
		/* Just initialize a ldst_info */
		(void)get_ldst_info(node, &wenv->obst);
	}
}

/* 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_num(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);
	}
}

static void get_base_and_offset(ir_node *ptr, base_offset_t *base_offset)
{
	/* TODO: long might not be enough, we should probably use some tarval
	 * thingy, or at least detect long overflows and abort */
	long     offset = 0;
	ir_mode *mode   = get_irn_mode(ptr);
	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_Const_long(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_Const_long(r);
			ptr     = l;
		} else if (is_Sel(ptr)) {
			ir_node *index = get_Sel_index(ptr);
			if (!is_Const(index))
				break;

			ir_type *type         = get_Sel_type(ptr);
			ir_type *element_type = get_array_element_type(type);
			if (get_type_state(element_type) != layout_fixed)
				break;

			/* TODO: may overflow here */
			int size = get_type_size_bytes(element_type);
			offset += size * get_Const_long(index);
			ptr     = get_Sel_ptr(ptr);
		} else if (is_Member(ptr)) {
			ir_entity *entity = get_Member_entity(ptr);
			ir_type   *owner  = get_entity_owner(entity);
			if (get_type_state(owner) != layout_fixed)
				break;
			offset += get_entity_offset(entity);
			ptr     = get_Member_ptr(ptr);
		} else
			break;
	}

	base_offset->offset = offset;
	base_offset->base   = ptr;
}

/**
 * Checks whether the memory of the first access
 * is contained within the second one.
 */
static bool is_contained_in(
	ir_mode *const load_mode, const base_offset_t *const load_bo,
	ir_mode *const prev_mode, const base_offset_t *const prev_bo)
{
	if (load_bo->base != prev_bo->base)
		return false;

	/* ensure the load value is completely contained in the previous one */
	long delta = load_bo->offset - prev_bo->offset;
	if (delta < 0)
		return false;
	long load_mode_len = get_mode_size_bytes(load_mode);
	long prev_mode_len = get_mode_size_bytes(prev_mode);
	if (delta+load_mode_len > prev_mode_len)
		return false;

	/* simple case: previous value has the same mode */
	if (load_mode == prev_mode)
		return true;

	ir_mode_arithmetic prev_arithmetic = get_mode_arithmetic(prev_mode);
	ir_mode_arithmetic load_arithmetic = get_mode_arithmetic(load_mode);
	return (prev_arithmetic == irma_twos_complement &&
	        load_arithmetic == irma_twos_complement)
	       || (prev_arithmetic != load_arithmetic
	           && load_mode_len == prev_mode_len);
}


/**
 * This is called for load-after-load and load-after-store.
 * If possible the value of the previous load/store is transformed in a way
 * so the 2nd load can be left out/replaced by arithmetic on the previous
 * value.
 */
static ir_node *transform_previous_value(ir_mode *const load_mode,
	const long load_offset, ir_mode *const prev_mode,
	const long prev_offset, ir_node *const prev_value,
	ir_node *const block)
{
	/* simple case: previous value has the same mode */
	if (load_mode == prev_mode)
		return prev_value;

	/* two complement values can be transformed with bitops */
	long               load_mode_len   = get_mode_size_bytes(load_mode);
	long               prev_mode_len   = get_mode_size_bytes(prev_mode);
	ir_mode_arithmetic prev_arithmetic = get_mode_arithmetic(prev_mode);
	ir_mode_arithmetic load_arithmetic = get_mode_arithmetic(load_mode);
	if (prev_arithmetic == irma_twos_complement &&
	    load_arithmetic == irma_twos_complement) {
		/* produce a shift to adjust offset delta */
		long           delta = load_offset - prev_offset;
		unsigned const shift = be_get_backend_param()->byte_order_big_endian
			? prev_mode_len - load_mode_len - delta
			: delta;
		ir_node *new_value = prev_value;
		if (shift != 0) {
			ir_graph *const irg  = get_irn_irg(block);
			ir_node  *const cnst = new_r_Const_long(irg, mode_Iu, shift * 8);
			new_value = new_r_Shr(block, new_value, cnst, prev_mode);
		}

		return new_r_Conv(block, new_value, load_mode);
	} else {
		assert(prev_arithmetic != load_arithmetic && load_mode_len == prev_mode_len);
		return new_r_Bitcast(block, prev_value, load_mode);
	}
}

static changes_t replace_load(ir_node *load, ir_node *new_value)
{
	const 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] != NULL) {
		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;

		assert(info->projs[pn_Load_X_regular] != NULL);
		ir_node *jmp = new_r_Jmp(get_nodes_block(load));
		exchange(info->projs[pn_Load_X_regular], jmp);
	}

	/* loads without user should already be optimized away */
	assert(info->projs[pn_Load_res] != NULL);
	exchange(info->projs[pn_Load_res], new_value);

	kill_and_reduce_usage(load);
	return res | DF_CHANGED;
}

/**
 * returns false if op cannot be reached through the X_regular proj of
 * @p prev_op or if there are no exceptions possible.
 */
static bool on_regular_path(ir_node *op, ir_node *prev_op)
{
	/* TODO: create a real test, for now we just make sure the previous node
	 * does not throw an exception. */
	(void)op;
	return !is_fragile_op(prev_op) || !ir_throws_exception(prev_op);
}

static changes_t try_load_after_store(track_load_env_t *env, ir_node *store)
{
	ir_node *const load = env->load;
	if (!on_regular_path(load, store))
		return NO_CHANGES;

	ir_node       *store_ptr        = get_Store_ptr(store);
	base_offset_t  prev_base_offset;
	get_base_and_offset(store_ptr, &prev_base_offset);
	ir_mode       *const load_mode   = get_Load_mode(load);
	ir_mode       *const store_mode  = get_irn_mode(get_Store_value(store));
	base_offset_t *const base_offset = &env->base_offset;

	/* load value completely contained in previous store? */
	if (is_contained_in(load_mode, base_offset, store_mode, &prev_base_offset)) {
		ir_node *const store_value = get_Store_value(store);
		ir_node *const block       = get_nodes_block(load);
		ir_node *const new_value
			= transform_previous_value(load_mode, base_offset->offset, store_mode,
			                           prev_base_offset.offset, store_value, block);
		DBG_OPT_RAW(load, new_value);
		return replace_load(load, new_value);
	}

	return NO_CHANGES;

}

static changes_t try_load_after_load(track_load_env_t *env, ir_node *prev_load)
{
	ir_node *const load = env->load;
	if (!on_regular_path(load, prev_load))
		return NO_CHANGES;

	ldst_info_t *const info       = (ldst_info_t *)get_irn_link(prev_load);
	ir_node     *const prev_value = info->projs[pn_Load_res];
	/* the other load is unused and will get removed later anyway */
	if (prev_value == NULL)
		return NO_CHANGES;

	ir_node *const prev_ptr        = get_Load_ptr(prev_load);
	base_offset_t prev_base_offset;
	get_base_and_offset(prev_ptr, &prev_base_offset);
	ir_mode       *const load_mode   = get_Load_mode(load);
	ir_mode       *const prev_mode   = get_Load_mode(prev_load);
	base_offset_t *const base_offset = &env->base_offset;

	/* load value completely contained in previous load? */
	if (is_contained_in(load_mode, base_offset, prev_mode, &prev_base_offset)) {
		ir_node *const block     = get_nodes_block(load);
		ir_node *const new_value
			= transform_previous_value(load_mode, base_offset->offset, prev_mode,
			                           prev_base_offset.offset, prev_value, block);
		DBG_OPT_RAR(prev_load, load);
		return replace_load(load, new_value);
	}

	/* previous load value completely contained in current load? */
	if (get_Load_type(load) == get_Load_type(prev_load) &&
	    is_contained_in(prev_mode, &prev_base_offset, load_mode, base_offset)) {
		ir_graph *const irg   = get_irn_irg(prev_load);
		ir_node  *const block = get_nodes_block(prev_load);

		/* Adapt pointer of previous load. */
		const long        delta         = base_offset->offset - prev_base_offset.offset;
		assert(delta <= 0);
		ir_mode    *const ptr_mode      = get_irn_mode(prev_ptr);
		ir_mode    *const ptr_mode_offs = get_reference_mode_unsigned_eq(ptr_mode);
		ir_node    *const c             = new_r_Const_long(irg, ptr_mode_offs, delta);
		ir_node    *const add           = new_r_Add(block, prev_ptr, c, ptr_mode);
		set_Load_ptr(prev_load, add);

		/* Change mode of previous load. */
		set_Load_mode(prev_load, load_mode);
		ir_node   *const original_proj = info->projs[pn_Load_res];
		ir_node   *const result_proj   = new_r_Proj(prev_load, load_mode, pn_Load_res);
		changes_t        changed       = replace_load(load, result_proj);
		ir_node   *const new_prev_value
			= transform_previous_value(prev_mode, prev_base_offset.offset, load_mode,
			                           base_offset->offset, result_proj, block);
		exchange(original_proj, new_prev_value);
		info->projs[pn_Load_res] = result_proj;
		changed |= DF_CHANGED;
		DBG_OPT_RAR(prev_load, load);
		return changed;
	}

	return NO_CHANGES;

}

static bool try_update_ptr_CopyB(track_load_env_t *env, ir_node *copyb)
{
	ir_node       *copyb_dst       = get_CopyB_dst(copyb);
	base_offset_t  dst_base_offset;
	get_base_and_offset(copyb_dst, &dst_base_offset);
	if (dst_base_offset.base != env->base_offset.base)
		return false;

	/* see if bytes loaded are fully contained in the CopyB */
	if (env->base_offset.offset < dst_base_offset.offset)
		return false;

	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);
	base_offset_t  src_base_offset;
	get_base_and_offset(copyb_src, &src_base_offset);

	long     delta     = env->base_offset.offset - dst_base_offset.offset;
	ir_node *load      = env->load;
	ir_mode *load_mode = get_Load_mode(load);
	long     load_size = get_mode_size_bytes(load_mode);
	if (delta + load_size > n_copy)
		return false;

	/* track src input */
	env->base_offset.base   = src_base_offset.base;
	env->base_offset.offset = src_base_offset.offset + delta;

	/*
	 * Everything is OK, we can replace
	 *   ptr = (load_base_ptr + load_offset)
	 * with
	 *   new_load_ptr = (src_base_ptr + delta)
	 */
	ir_graph *irg          = get_irn_irg(load);
	ir_node  *block        = get_nodes_block(load);
	ir_mode  *mode_ref     = get_irn_mode(src_base_offset.base);
	ir_mode  *mode_ref_int = get_reference_mode_unsigned_eq(mode_ref);
	ir_node  *cnst         = new_r_Const_long(irg, mode_ref_int,
	                                          src_base_offset.offset + delta);
	ir_node  *new_load_ptr = new_r_Add(block, src_base_offset.base, cnst, mode_P);
	env->ptr = new_load_ptr;
	return true;
}

/* Note on terminology:
 * In the following functions, two types are associated with each Load
 * and Store: load_type/store_type are the types of the actual
 * operations (derived from the modes of the IR nodes), whereas
 * load_objt/store_objt are the types of the objects in memory which are
 * accessed by the Load/Store nodes.
 */

/**
 * 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_load_mem_chain(track_load_env_t *env, ir_node *start)
{
	ir_node  *load      = env->load;
	ir_type  *load_type = get_Load_type(load);
	unsigned  load_size = get_mode_size_bytes(get_Load_mode(load));

	ir_node   *node = start;
	changes_t  res  = NO_CHANGES;
	for (;;) {
		ldst_info_t *node_info = (ldst_info_t *)get_irn_link(node);

		if (is_Store(node)) {
			/* first try load-after-store */
			changes_t changes = try_load_after_store(env, node);
			if (changes != NO_CHANGES)
				return changes | res;

			/* check if we can pass through this store */
			const ir_node     *ptr        = get_Store_ptr(node);
			const ir_node     *value      = get_Store_value(node);
			const ir_type     *store_type = get_Store_type(node);
			unsigned           store_size = get_mode_size_bytes(get_irn_mode(value));
			ir_alias_relation  rel        = get_alias_relation(
				ptr, store_type, store_size,
				env->ptr, load_type, load_size);
			/* if the might be an alias, we cannot pass this Store */
			if (rel != ir_no_alias)
				break;
			node = skip_Proj(get_Store_mem(node));
		} else if (is_Load(node)) {
			/* try load-after-load */
			changes_t changes = try_load_after_load(env, node);
			if (changes != NO_CHANGES)
				return changes | res;
			/* we can skip any load */
			node = skip_Proj(get_Load_mem(node));
		} else if (is_CopyB(node)) {
			/*
			 * 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.
			 */
			bool updated = try_update_ptr_CopyB(env, node);
			if (updated) {
				/* Special case: If new_ptr points to
				 * a constant, we *can* replace the
				 * Load immediately.
				 */
				res |= NODES_CREATED;
				ir_mode *load_mode = get_Load_mode(load);
				ir_node *new_value = predict_load(env->ptr, load_mode);
				if (new_value != NULL)
					return replace_load(load, new_value) | res;
			}

			/* check aliasing with the CopyB */
			ir_node           *dst  = get_CopyB_dst(node);
			ir_type           *type = get_CopyB_type(node);
			unsigned           size = get_type_size_bytes(type);
			ir_alias_relation  rel  = get_alias_relation(
				dst, type, size,
				env->ptr, load_type, load_size);
			/* possible alias => we cannot continue */
			if (rel != ir_no_alias)
				break;
			node = skip_Proj(get_CopyB_mem(node));
		} else if (is_irn_const_memory(node)) {
			node = skip_Proj(get_memop_mem(node));
		} else {
			/* be conservative about any other node and assume aliasing
			 * that changes the loaded value */
			break;
		}

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

	if (is_Sync(node)) {
		/* handle all Sync predecessors */
		foreach_irn_in(node, i, in) {
			ir_node *skipped = skip_Proj(in);
			res |= follow_load_mem_chain(env, skipped);
			if ((res & ~NODES_CREATED) != NO_CHANGES)
				break;
		}
	}
	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)
{
	const 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) {
		assert(info->projs[pn_Load_X_regular] == NULL);
		/* the value is never used and we don't care about exceptions, remove */
		exchange(info->projs[pn_Load_M], mem);
		kill_and_reduce_usage(load);
		return res | DF_CHANGED;
	}

	track_load_env_t env = { .ptr = ptr };
	get_base_and_offset(ptr, &env.base_offset);

	/* Check, if the base address of this load is used more than once.
	 * If not, we won't find another store/load/CopyB anyway.
	 * TODO: can we miss values with multiple users in between? */
	if (get_irn_n_edges(ptr) <= 1 && get_irn_n_edges(env.base_offset.base) <= 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();
	env.load = load;
	res = follow_load_mem_chain(&env, skip_Proj(mem));
	return res;
}

/**
 * Check whether small is a part of large (starting at same address).
 */
static bool is_partially_same(ir_node *small, ir_node *large)
{
	const ir_mode *sm = get_irn_mode(small);
	const 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_store_mem_chain(ir_node *store, ir_node *start,
                                        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_type     *type  = get_Store_type(store);
	unsigned     size  = get_mode_size_bytes(get_irn_mode(value));
	ir_node     *block = get_nodes_block(store);

	ir_node *node = start;
	while (node != store) {
		ldst_info_t *node_info = (ldst_info_t *)get_irn_link(node);

		/*
		 * 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(node) && !had_split && get_Store_ptr(node) == ptr &&
		    get_nodes_block(node) == 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(node) != volatility_is_volatile
			    && !node_info->projs[pn_Store_X_except]) {
				ir_node  *predvalue = get_Store_value(node);
				ir_mode  *predmode  = get_irn_mode(predvalue);
				unsigned  predsize  = get_mode_size_bytes(predmode);

				if (predsize <= size || is_partially_same(predvalue, value)) {
					DBG_OPT_WAW(node, store);
					DB((dbg, LEVEL_1, "  killing store %+F (override by %+F)\n",
					    node, store));
					exchange(node_info->projs[pn_Store_M], get_Store_mem(node));
					kill_and_reduce_usage(node);
					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(node);

				if (is_partially_same(value, predvalue)) {
					DBG_OPT_WAW(node, store);
					DB((dbg, LEVEL_1, "  killing store %+F (override by %+F)\n",
					    node, store));
					exchange(info->projs[pn_Store_M], mem);
					kill_and_reduce_usage(store);
					return DF_CHANGED;
				}
			}
		} else if (is_Load(node) && get_Load_ptr(node) == ptr &&
		           value == node_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, node);
				DB((dbg, LEVEL_1,
				    "  killing store %+F (read %+F from same address)\n",
				    store, node));
				exchange(info->projs[pn_Store_M], mem);
				kill_and_reduce_usage(store);
				return DF_CHANGED;
			}
		}

		if (is_Store(node)) {
			/* check if we can pass through this store */
			ir_node           *store_ptr   = get_Store_ptr(node);
			ir_node           *store_value = get_Store_value(node);
			unsigned           store_size  = get_mode_size_bytes(get_irn_mode(store_value));
			ir_type           *store_type  = get_Store_type(node);
			ir_alias_relation  rel         = get_alias_relation(
				store_ptr, store_type, store_size,
				ptr, type, size);
			/* if the might be an alias, we cannot pass this Store */
			if (rel != ir_no_alias)
				break;
			node = skip_Proj(get_Store_mem(node));
		} else if (is_Load(node)) {
			ir_node           *load_ptr  = get_Load_ptr(node);
			ir_type           *load_type = get_Load_type(node);
			unsigned           load_size = get_mode_size_bytes(get_Load_mode(node));
			ir_alias_relation  rel       = get_alias_relation(
				load_ptr, load_type, load_size,
				ptr, type, size);
			if (rel != ir_no_alias)
				break;

			node = skip_Proj(get_Load_mem(node));
		} else if (is_CopyB(node)) {
			ir_node           *copyb_src  = get_CopyB_src(node);
			ir_type           *copyb_type = get_CopyB_type(node);
			unsigned           copyb_size = get_type_size_bytes(copyb_type);
			ir_alias_relation  src_rel    = get_alias_relation(
				copyb_src, copyb_type, copyb_size,
				ptr, type, size);
			if (src_rel != ir_no_alias)
				break;
			ir_node           *copyb_dst = get_CopyB_dst(node);
			ir_alias_relation  dst_rel   = get_alias_relation(
				copyb_dst, copyb_type, copyb_size,
				ptr, type, size);
			if (dst_rel != ir_no_alias)
				break;
		} else {
			/* follow only Load chains */
			break;
		}

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

	if (is_Sync(node)) {
		/* handle all Sync predecessors */
		foreach_irn_in(node, i, in) {
			ir_node *skipped = skip_Proj(in);
			res |= follow_store_mem_chain(store, skipped, true);
			if (res != NO_CHANGES)
				break;
		}
	}
	return res;
}