ldstopt.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   Load/Store optimizations.
 * @author  Michael Beck
 * @version $Id$
 */
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif

#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 "irvrfy.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 "irtools.h"
#include "opt_polymorphy.h"
#include "irmemory.h"
#include "xmalloc.h"
#include "irphase_t.h"
#include "irgopt.h"
#include "debug.h"

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

#ifdef DO_CACHEOPT
#include "cacheopt/cachesim.h"
#endif

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

#define MAX_PROJ	IMAX(IMAX(pn_Load_max, pn_Store_max), 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];    /**< 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 = get_irn_link(node);

	if (! info) {
		info = obstack_alloc(obst, sizeof(*info));
		memset(info, 0, sizeof(*info));
		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 = get_irn_link(node);

	if (! info) {
		info = obstack_alloc(obst, sizeof(*info));
		memset(info, 0, sizeof(*info));
		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)
{
	ir_opcode   opcode = get_irn_opcode(node);
	ir_node     *pred, *blk, *pred_blk;
	ldst_info_t *ldst_info;
	walk_env_t  *wenv = env;

	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 = get_Proj_proj(proj) == pn_Generic_X_except;
			}

			/* 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) {
			ir_entity *ent = get_SymConst_entity(ptr);
			if (variability_constant == get_entity_variability(ent))
				return ent;
			return NULL;
		} 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;
					tarval *tlower, *tupper;
					ir_node *index = get_Sel_index(ptr, i);
					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) & pn_Cmp_Lt)
						return NULL;
					if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
						return NULL;

					/* ok, bounds check finished */
				}
			}

			if (variability_constant == get_entity_variability(ent))
				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, int depth) {
	compound_graph_path *res = NULL;
	ir_entity           *root, *field, *ent;
	int                 path_len, pos, idx;
	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 it's 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_node *l = get_Add_left(ptr);
		ir_node *r = get_Add_right(ptr);
		ir_mode *mode;

		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);
		}
		idx = 0;
		for (ent = field;;) {
			tp = get_entity_type(ent);
			if (! is_Array_type(tp))
				break;
			ent = get_array_element_entity(tp);
			++idx;
		}
		/* 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;
			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);
			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);

			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) & pn_Cmp_Lt)
				return NULL;
			if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
				return NULL;

			/* ok, bounds check finished */
			index = get_tarval_long(tv_index);
			set_compound_graph_path_array_index(res, pos, index);
			++pos;
		}
		if (! tarval_is_null(tv)) {
			/* hmm, wrong access */
			return NULL;
		}
	} 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) {
	return rec_get_accessed_path(ptr, 0);
}  /* get_accessed_path */

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

static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
	path_entry       entry, *p;
	ir_entity        *ent;
	ir_initializer_t *initializer;

	entry.next      = next;

	if (is_Sel(ptr)) {
		ir_entity *field;
		ir_type   *tp;

		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 {
			int 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);
	}

	/* found the end */
	assert(is_SymConst(ptr));

	ent = get_SymConst_entity(ptr);
	initializer = get_entity_initializer(ent);
	for (p = next; p != NULL; p = p->next) {
		unsigned n;

		if (initializer->kind != IR_INITIALIZER_COMPOUND)
			return NULL;

		n = get_initializer_compound_n_entries(initializer);
		if (p->index >= n)
			return NULL;
		initializer = get_initializer_compound_value(initializer, p->index);
	}

	switch (initializer->kind) {
	case IR_INITIALIZER_CONST:
		return get_initializer_const_value(initializer);
	case IR_INITIALIZER_TARVAL:
	case IR_INITIALIZER_NULL:
	default:
		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 lost it's usage.
 */
static void handle_load_update(ir_node *load) {
	ldst_info_t *info = 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 which 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(current_ir_graph, get_nodes_block(load)));
		kill_node(load);
		reduce_adr_usage(ptr);
	}
}  /* handle_load_update */

/**
 * A Use of an address node is vanished. Check if this was a Proj
 * node and update the counters.
 */
static void reduce_adr_usage(ir_node *ptr) {
	if (is_Proj(ptr)) {
		if (get_irn_n_edges(ptr) <= 0) {
			/* this Proj is dead now */
			ir_node *pred = get_Proj_pred(ptr);

			if (is_Load(pred)) {
				ldst_info_t *info = get_irn_link(pred);
				info->projs[get_Proj_proj(ptr)] = NULL;

				/* this node lost it's 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) {
	if (old_mode == new_mode)
		return 1;

	/* if both modes are two-complement ones, we can always convert the
	   Stored value into the needed one. */
	if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
		  get_mode_arithmetic(old_mode) == irma_twos_complement &&
		  get_mode_arithmetic(new_mode) == irma_twos_complement)
		return 1;
	return 0;
}  /* can_use_stored_value */

/**
 * Check whether a Call is at least pure, ie. 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 */

/**
 * 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 = 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 = 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 in C were the following
		 * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
		 */
		if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
		    can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
			/*
			 * 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 MacroBlock. 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 MacroBlock :-(
			 * 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 ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
			    get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
				ir_node *value = get_Store_value(pred);

				DBG_OPT_RAW(load, value);

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

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

				/* no exception */
				if (info->projs[pn_Load_X_except]) {
					exchange( info->projs[pn_Load_X_except], new_Bad());
					res |= CF_CHANGED;
				}
				if (info->projs[pn_Load_X_regular]) {
					exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
					res |= CF_CHANGED;
				}

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

				kill_node(load);
				reduce_adr_usage(ptr);
				return res | DF_CHANGED;
			}
		} 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 MacroBlock. 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...
			 */
			if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(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(current_ir_graph, get_nodes_block(pred), 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(current_ir_graph, 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]) {
					exchange(info->projs[pn_Load_X_except], new_Bad());
					res |= CF_CHANGED;
				}
				if (info->projs[pn_Load_X_regular]) {
					exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, 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(
				current_ir_graph,
				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, i)));
			if (res)
				return res;
		}
	}

	return res;
}  /* follow_Mem_chain */

/**
 * optimize a Load
 *
 * @param load  the Load node
 */
static unsigned optimize_load(ir_node *load)
{
	ldst_info_t *info = get_irn_link(load);
	ir_node *mem, *ptr, *new_node;
	ir_entity *ent;
	unsigned res = 0;

	/* do NOT touch volatile loads for now */
	if (get_Load_volatility(load) == volatility_is_volatile)
		return 0;

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

	/*
	 * Check if we can remove the exception from a Load:
	 * This can be done, if the address is from an Sel(Alloc) and
	 * the Sel type is a subtype of the allocated type.
	 *
	 * This optimizes some often used OO constructs,
	 * like x = new O; x->t;
	 */
	if (info->projs[pn_Load_X_except]) {
		if (is_Sel(ptr)) {
			ir_node *mem = get_Sel_mem(ptr);

			/* FIXME: works with the current FE, but better use the base */
			if (is_Alloc(skip_Proj(mem))) {
				/* ok, check the types */
				ir_entity *ent    = get_Sel_entity(ptr);
				ir_type   *s_type = get_entity_type(ent);
				ir_type   *a_type = get_Alloc_type(mem);

				if (is_SubClass_of(s_type, a_type)) {
					/* ok, condition met: there can't be an exception because
					* Alloc guarantees that enough memory was allocated */

					exchange(info->projs[pn_Load_X_except], new_Bad());
					info->projs[pn_Load_X_except] = NULL;
					exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
					info->projs[pn_Load_X_regular] = NULL;
					res |= CF_CHANGED;
				}
			}
		} else if (is_Alloc(skip_Proj(skip_Cast(ptr)))) {
				/* simple case: a direct load after an Alloc. Firm Alloc throw
				 * an exception in case of out-of-memory. So, there is no way for an
				 * exception in this load.
				 * This code is constructed by the "exception lowering" in the Jack compiler.
				 */
				exchange(info->projs[pn_Load_X_except], new_Bad());
				info->projs[pn_Load_X_except] = NULL;
				exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
				info->projs[pn_Load_X_regular] = NULL;
				res |= CF_CHANGED;
		}
	}

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

	if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
		/* a Load which value is neither used nor exception checked, remove it */
		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(current_ir_graph, get_nodes_block(load)));
			res |= CF_CHANGED;
		}
		kill_node(load);
		reduce_adr_usage(ptr);
		return res | DF_CHANGED;
	}

	/* Load from a constant polymorphic field, where we can resolve
	   polymorphism. */
	new_node = transform_node_Load(load);
	if (new_node != load) {
		if (info->projs[pn_Load_M]) {
			exchange(info->projs[pn_Load_M], mem);
			info->projs[pn_Load_M] = NULL;
		}
		if (info->projs[pn_Load_X_except]) {
			exchange(info->projs[pn_Load_X_except], new_Bad());
			info->projs[pn_Load_X_except] = NULL;
			res |= CF_CHANGED;
		}
		if (info->projs[pn_Load_X_regular]) {
			exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
			info->projs[pn_Load_X_regular] = NULL;
			res |= CF_CHANGED;
		}
		if (info->projs[pn_Load_res])
			exchange(info->projs[pn_Load_res], new_node);

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

	/* check if we can determine the entity that will be loaded */
	ent = find_constant_entity(ptr);
	if (ent) {
		if ((allocation_static == get_entity_allocation(ent)) &&
			(visibility_external_allocated != get_entity_visibility(ent))) {
			/* a static allocation that is not external: there should be NO exception
			 * when loading. */

			/* no exception, clear the info field as it might be checked later again */
			if (info->projs[pn_Load_X_except]) {
				exchange(info->projs[pn_Load_X_except], new_Bad());
				info->projs[pn_Load_X_except] = NULL;
				res |= CF_CHANGED;
			}
			if (info->projs[pn_Load_X_regular]) {
				exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
				info->projs[pn_Load_X_regular] = NULL;
				res |= CF_CHANGED;
			}

			if (variability_constant == get_entity_variability(ent)) {
				if (is_atomic_entity(ent)) {
					/* Might not be atomic after
					   lowering of Sels.  In this
					   case we could also load, but
					   it's more complicated. */
					/* more simpler case: we load the content of a constant value:
					 * replace it by the constant itself
					 */

					/* no memory */
					if (info->projs[pn_Load_M]) {
						exchange(info->projs[pn_Load_M], mem);
						res |= DF_CHANGED;
					}
					/* no result :-) */
					if (info->projs[pn_Load_res]) {
						if (is_atomic_entity(ent)) {
							ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));

							DBG_OPT_RC(load, c);
							exchange(info->projs[pn_Load_res], c);
							res |= DF_CHANGED;
						}
					}
					kill_node(load);
					reduce_adr_usage(ptr);
					return res;
				} else {
					ir_node *c = NULL;
					if (ent->has_initializer) {
						/* new style initializer */
						c = find_compound_ent_value(ptr);
					} else {
						/* old style initializer */
						compound_graph_path *path = get_accessed_path(ptr);

						if (path) {
							assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));

							c = get_compound_ent_value_by_path(ent, path);
							free_compound_graph_path(path);
						}
					}
					if (c != NULL) {
						if (info->projs[pn_Load_M]) {
							exchange(info->projs[pn_Load_M], mem);
							res |= DF_CHANGED;
						}
						if (info->projs[pn_Load_res]) {
							exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
							res |= DF_CHANGED;
						}
						kill_node(load);
						reduce_adr_usage(ptr);
						return res;
					} else {
						/*  We can not determine a correct access path.  E.g., in jack, we load
						a byte from an object to generate an exception.   Happens in test program
						Reflectiontest.
						printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
						get_entity_name(get_irg_entity(current_ir_graph)));
						ir_printf("  load: %+F\n", load);
						ir_printf("  ptr:  %+F\n", ptr);
						*/
					}
				}
			}
		}
	}

	/* Check, if the address of this load is used more than once.
	 * If not, this load cannot be removed in any case. */
	if (get_irn_n_uses(ptr) <= 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;
}  /* optimize_load */

/**
 * 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);
}  /* is_completely_overwritten */

/**
 * follow the memory chain as long as there are only Loads and alias free Stores.
 *
 * INC_MASTER() must be called before dive into
 */
static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
	unsigned res = 0;
	ldst_info_t *info = get_irn_link(store);
	ir_node *pred;
	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 *mblk  = get_Block_MacroBlock(block);

	for (pred = curr; pred != store;) {
		ldst_info_t *pred_info = 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 mode that is written have a bigger  or equal size the the old
		 * one, the old value is completely overwritten and can be killed ...
		 */
		if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
		    get_nodes_MacroBlock(pred) == mblk &&
		    is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
			/*
			 * a Store after a Store in the same MacroBlock -- a write after write.
			 * We may remove the first Store, 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]) {
				DBG_OPT_WAW(pred, store);
				exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
				kill_node(pred);
				reduce_adr_usage(ptr);
				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);
				exchange(info->projs[pn_Store_M], mem);
				kill_node(store);
				reduce_adr_usage(ptr);
				return DF_CHANGED;
			}
		}

		if (is_Store(pred)) {
			/* check if we can pass thru this store */
			ir_alias_relation rel = get_alias_relation(
				current_ir_graph,
				get_Store_ptr(pred),
				get_irn_mode(get_Store_value(pred)),
				ptr, 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)) {
			ir_alias_relation rel = get_alias_relation(
				current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
				ptr, mode);
			if (rel != ir_no_alias)
				break;

			pred = skip_Proj(get_Load_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)) {
		int i;

		/* handle all Sync predecessors */
		for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
			res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
			if (res)
				break;
		}
	}
	return res;
}  /* follow_Mem_chain_for_Store */

/**
 * optimize a Store
 *
 * @param store  the Store node
 */
static unsigned optimize_store(ir_node *store) {
	ir_node *ptr, *mem;

	if (get_Store_volatility(store) == volatility_is_volatile)
		return 0;

	ptr = get_Store_ptr(store);

	/* Check, if the address of this Store is used more than once.
	 * If not, this Store cannot be removed in any case. */
	if (get_irn_n_uses(ptr) <= 1)
		return 0;

	mem = get_Store_mem(store);

	/* follow the memory chain as long as there are only Loads */
	INC_MASTER();

	return follow_Mem_chain_for_Store(store, skip_Proj(mem));
}  /* optimize_store */

/**
 * walker, optimizes Phi after Stores to identical places:
 * Does the following optimization:
 * @verbatim
 *
 *   val1   val2   val3          val1  val2  val3
 *    |      |      |               \    |    /
 *  Store  Store  Store              \   |   /
 *      \    |    /                   PhiData
 *       \   |   /                       |
 *        \  |  /                      Store
 *          PhiM
 *
 * @endverbatim
 * This reduces the number of stores and allows for predicated execution.
 * Moves Stores back to the end of a function which may be bad.
 *
 * This is only possible if the predecessor blocks have only one successor.
 */
static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
{
	int i, n;
	ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
	ir_mode *mode;
	ir_node **inM, **inD, **projMs;
	int *idx;
	dbg_info *db = NULL;
	ldst_info_t *info;
	block_info_t *bl_info;
	unsigned res = 0;

	/* Must be a memory Phi */
	if (get_irn_mode(phi) != mode_M)
		return 0;

	n = get_Phi_n_preds(phi);
	if (n <= 0)
		return 0;

	/* must be only one user */
	projM = get_Phi_pred(phi, 0);
	if (get_irn_n_edges(projM) != 1)
		return 0;

	store = skip_Proj(projM);
	old_store = store;
	if (!is_Store(store))
		return 0;

	block = get_nodes_block(store);

	/* abort on dead blocks */
	if (is_Block_dead(block))
		return 0;

	/* check if the block is post dominated by Phi-block
	   and has no exception exit */
	bl_info = get_irn_link(block);
	if (bl_info->flags & BLOCK_HAS_EXC)
		return 0;

	phi_block = get_nodes_block(phi);
	if (! block_strictly_postdominates(phi_block, block))
		return 0;

	/* this is the address of the store */
	ptr  = get_Store_ptr(store);
	mode = get_irn_mode(get_Store_value(store));
	info = get_irn_link(store);
	exc  = info->exc_block;

	for (i = 1; i < n; ++i) {
		ir_node *pred = get_Phi_pred(phi, i);

		if (get_irn_n_edges(pred) != 1)
			return 0;

		pred = skip_Proj(pred);
		if (!is_Store(pred))
			return 0;

		if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
			return 0;

		info = get_irn_link(pred);

		/* check, if all stores have the same exception flow */
		if (exc != info->exc_block)
			return 0;

		/* abort on dead blocks */
		block = get_nodes_block(pred);
		if (is_Block_dead(block))
			return 0;

		/* check if the block is post dominated by Phi-block
		   and has no exception exit. Note that block must be different from
		   Phi-block, else we would move a Store from end End of a block to its
		   Start... */
		bl_info = get_irn_link(block);
		if (bl_info->flags & BLOCK_HAS_EXC)
			return 0;
		if (block == phi_block || ! block_postdominates(phi_block, block))
			return 0;
	}

	/*
	 * ok, when we are here, we found all predecessors of a Phi that
	 * are Stores to the same address and size. That means whatever
	 * we do before we enter the block of the Phi, we do a Store.
	 * So, we can move the Store to the current block:
	 *
	 *   val1    val2    val3          val1  val2  val3
	 *    |       |       |               \    |    /
	 * | Str | | Str | | Str |             \   |   /
	 *      \     |     /                   PhiData
	 *       \    |    /                       |
	 *        \   |   /                       Str
	 *           PhiM
	 *
	 * Is only allowed if the predecessor blocks have only one successor.
	 */

	NEW_ARR_A(ir_node *, projMs, n);
	NEW_ARR_A(ir_node *, inM, n);
	NEW_ARR_A(ir_node *, inD, n);
	NEW_ARR_A(int, idx, n);

	/* Prepare: Collect all Store nodes.  We must do this
	   first because we otherwise may loose a store when exchanging its
	   memory Proj.
	 */
	for (i = n - 1; i >= 0; --i) {
		ir_node *store;

		projMs[i] = get_Phi_pred(phi, i);
		assert(is_Proj(projMs[i]));

		store = get_Proj_pred(projMs[i]);
		info  = get_irn_link(store);

		inM[i] = get_Store_mem(store);
		inD[i] = get_Store_value(store);
		idx[i] = info->exc_idx;
	}
	block = get_nodes_block(phi);

	/* second step: create a new memory Phi */
	phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);

	/* third step: create a new data Phi */
	phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);

	/* rewire memory and kill the node */
	for (i = n - 1; i >= 0; --i) {
		ir_node *proj  = projMs[i];