* Project: libFIRM
* File name: ir/opt/ldstopt.c
* Purpose: load store optimizations
- * Author:
+ * Author: Michael Beck
* Created:
* CVS-ID: $Id$
* Copyright: (c) 1998-2004 Universit\81ät Karlsruhe
# include <string.h>
#endif
-# 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 "irtools.h"
-# include "opt_polymorphy.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 "irtools.h"
+#include "opt_polymorphy.h"
#ifdef DO_CACHEOPT
#include "cacheopt/cachesim.h"
* walker environment
*/
typedef struct _walk_env_t {
- struct obstack obst; /**< list of all stores */
+ struct obstack obst; /**< list of all stores */
unsigned changes; /**< a bitmask of graph changes */
} walk_env_t;
LDST_VISITED = 1 /**< if set, this Load/Store is already visited */
};
-/**
- * a Load/Store info
- */
+/** 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 flags; /**< flags */
+ unsigned visited; /**< visited counter for breaking loops */
} ldst_info_t;
/**
- * flags for control flow
+ * flags for control flow.
*/
enum block_flags_t {
BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
- BLOCK_HAS_EXC = 2 /**< Block has exceptionl control flow */
+ BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
};
/**
- * a Block info
+ * 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, walk_env_t *env)
-{
+static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
ldst_info_t *info = get_irn_link(node);
if (! info) {
}
/**
- * update the exception block info for a Load/Store
+ * 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)
{
return 0;
}
-#define get_irn_out_n(node) (unsigned)PTR_TO_INT(get_irn_link(node))
-#define set_irn_out_n(node, n) set_irn_link(adr, INT_TO_PTR(n))
+/** Return the number of uses of an address node */
+#define get_irn_n_uses(adr) (unsigned)PTR_TO_INT(get_irn_link(adr))
+/** Sets the number of uses of an address node */
+#define set_irn_n_uses(adr, n) set_irn_link(adr, INT_TO_PTR(n))
/**
* walker, collects all Load/Store/Proj nodes
*
- * walks form Start -> End
+ * walks from Start -> End
*/
static void collect_nodes(ir_node *node, void *env)
{
if ((ldst_info->flags & LDST_VISITED) == 0) {
adr = get_Load_ptr(pred);
- set_irn_out_n(adr, get_irn_out_n(adr) + 1);
+ set_irn_n_uses(adr, get_irn_n_uses(adr) + 1);
ldst_info->flags |= LDST_VISITED;
}
if ((ldst_info->flags & LDST_VISITED) == 0) {
adr = get_Store_ptr(pred);
- set_irn_out_n(adr, get_irn_out_n(adr) + 1);
+ set_irn_n_uses(adr, get_irn_n_uses(adr) + 1);
ldst_info->flags |= LDST_VISITED;
}
}
}
else if (op == op_Block) { /* check, if it's an exception block */
- int i, n;
+ int i;
- for (i = 0, n = get_Block_n_cfgpreds(node); i < n; ++i) {
+ for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
ir_node *pred_block;
block_info_t *bl_info;
/* ignore Bad predecessors, they will be removed later */
if (is_Bad(pred))
- continue;
+ continue;
pred_block = get_nodes_block(pred);
bl_info = get_block_info(pred_block, wenv);
if (is_fragile_op(pred))
- bl_info->flags |= BLOCK_HAS_EXC;
- else if (is_forking_op(pred))
- bl_info->flags |= BLOCK_HAS_COND;
+ bl_info->flags |= BLOCK_HAS_EXC;
+ else if (is_irn_forking(pred))
+ bl_info->flags |= BLOCK_HAS_COND;
if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
ldst_info = get_ldst_info(pred, wenv);
return get_SymConst_entity(ptr);
}
else if (op == op_Sel) {
- entity *ent = get_Sel_entity(ptr);
- type *tp = get_entity_owner(ent);
+ entity *ent = get_Sel_entity(ptr);
+ ir_type *tp = get_entity_owner(ent);
- /* Do not fiddle about polymorphy. */
+ /* 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 rec_get_accessed_path(ptr, 0);
}
+/* 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);
+ reduce_adr_usage(ptr);
+ }
+}
+
+/**
+ * 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) {
+ int use_count = get_irn_n_uses(ptr);
+ --use_count;
+ assert(use_count >= 0);
+ set_irn_n_uses(ptr, use_count);
+
+ if (is_Proj(ptr)) {
+ if (use_count <= 0) {
+ /* this Proj is now dead, update the Load/Store info */
+ ir_node *pred = get_Proj_pred(ptr);
+ opcode code = get_irn_opcode(pred);
+
+ if (code == iro_Load) {
+ 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);
+ }
+ }
+ }
+}
+
+/**
+ * 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() must be called before dive into
+ */
+static unsigned follow_Load_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; pred = skip_Proj(get_Load_mem(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 (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
+ 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 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 ((!pred_info->projs[pn_Store_X_except] && !info->projs[pn_Load_X_except]) ||
+ get_nodes_block(load) == get_nodes_block(pred)) {
+ ir_node *value = get_Store_value(pred);
+
+ DBG_OPT_RAW(load, 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_res])
+ exchange(info->projs[pn_Load_res], value);
+
+ reduce_adr_usage(ptr);
+ return res | DF_CHANGED;
+ }
+ }
+ else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
+ 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...
+ */
+ if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
+ DBG_OPT_RAR(load, pred);
+
+ if (pred_info->projs[pn_Load_res]) {
+ /* we need a data proj from the previous load for this optimization */
+ if (info->projs[pn_Load_res])
+ exchange(info->projs[pn_Load_res], pred_info->projs[pn_Load_res]);
+
+ if (info->projs[pn_Load_M])
+ exchange(info->projs[pn_Load_M], mem);
+ }
+ else {
+ if (info->projs[pn_Load_res]) {
+ set_Proj_pred(info->projs[pn_Load_res], pred);
+ set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred));
+ pred_info->projs[pn_Load_res] = info->projs[pn_Load_res];
+ }
+ if (info->projs[pn_Load_M]) {
+ /* Actually, this if should not be necessary. Construct the Loads
+ properly!!! */
+ 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;
+ }
+
+ reduce_adr_usage(ptr);
+ return res |= DF_CHANGED;
+ }
+ }
+
+ /* follow only Load chains */
+ if (get_irn_op(pred) != op_Load)
+ break;
+
+ /* check for cycles */
+ if (NODE_VISITED(pred_info))
+ break;
+ MARK_NODE(pred_info);
+ }
+
+ if (get_irn_op(pred) == op_Sync) {
+ int i;
+
+ /* handle all Sync predecessors */
+ for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
+ res |= follow_Load_chain(load, skip_Proj(get_Sync_pred(pred, i)));
+ if (res)
+ break;
+ }
+ }
+
+ return res;
+}
+
/**
* optimize a Load
*/
{
ldst_info_t *info = get_irn_link(load);
ir_mode *load_mode = get_Load_mode(load);
- ir_node *pred, *mem, *ptr, *new_node;
+ ir_node *mem, *ptr, *new_node;
entity *ent;
unsigned res = 0;
* like x = new O; x->t;
*/
if (info->projs[pn_Load_X_except]) {
- if (get_irn_op(ptr) == op_Sel) {
+ if (is_Sel(ptr)) {
ir_node *mem = get_Sel_mem(ptr);
if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
/* ok, check the types */
- entity *ent = get_Sel_entity(ptr);
- type *s_type = get_entity_type(ent);
- type *a_type = get_Alloc_type(mem);
+ 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)) {
+ 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 */
/* a Load which value is neither used nor exception checked, remove it */
exchange(info->projs[pn_Load_M], mem);
+ reduce_adr_usage(ptr);
return res | DF_CHANGED;
}
/* Load from a constant polymorphic field, where we can resolve
- polymorphy. */
+ polymorphism. */
new_node = transform_node_Load(load);
if (new_node != load) {
if (info->projs[pn_Load_M]) {
}
if (info->projs[pn_Load_res])
exchange(info->projs[pn_Load_res], new_node);
+
+ reduce_adr_usage(ptr);
return res | DF_CHANGED;
}
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_atomic_ent_value(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);
- return DF_CHANGED | res;
+ res |= DF_CHANGED;
}
}
+ reduce_adr_usage(ptr);
+ return res;
}
else if (variability_constant == get_entity_variability(ent)) {
compound_graph_path *path = get_accessed_path(ptr);
res |= DF_CHANGED;
}
if (info->projs[pn_Load_res]) {
- exchange(info->projs[pn_Load_res], copy_const_value(c));
- return res | DF_CHANGED;
+ exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
+ res |= DF_CHANGED;
}
+ reduce_adr_usage(ptr);
+ return res;
}
else {
/* We can not determine a correct access path. E.g., in jack, we load
/* 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_out_n(ptr) <= 1)
+ 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
+ /*
+ * 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.
*/
- for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(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 = *(type1 *)p; b = *(type2 *)p ...
- */
-
- if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
- get_irn_mode(get_Store_value(pred)) == load_mode) {
- ldst_info_t *pred_info = 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 ((!pred_info->projs[pn_Store_X_except] && !info->projs[pn_Load_X_except]) ||
- get_nodes_block(load) == get_nodes_block(pred)) {
- ir_node *value = get_Store_value(pred);
-
- DBG_OPT_RAW(load, 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_res])
- exchange(info->projs[pn_Load_res], value);
-
- return res | DF_CHANGED;
- }
- }
- else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
- 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...
- */
- if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
- ldst_info_t *pred_info = get_irn_link(pred);
-
- DBG_OPT_RAR(load, pred);
-
- if (pred_info->projs[pn_Load_res]) {
- /* we need a data proj from the previous load for this optimization */
- if (info->projs[pn_Load_res])
- exchange(info->projs[pn_Load_res], pred_info->projs[pn_Load_res]);
-
- if (info->projs[pn_Load_M])
- exchange(info->projs[pn_Load_M], mem);
- }
- else {
- if (info->projs[pn_Load_res]) {
- set_Proj_pred(info->projs[pn_Load_res], pred);
- set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred));
- pred_info->projs[pn_Load_res] = info->projs[pn_Load_res];
- }
- if (info->projs[pn_Load_M]) {
- /* Actually, this if should not be necessary. Construct the Loads
- properly!!! */
- 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;
- }
-
- return res |= DF_CHANGED;
- }
- }
-
- /* follow only Load chains */
- if (get_irn_op(pred) != op_Load)
- break;
- }
+ INC_MASTER();
+ res = follow_Load_chain(load, skip_Proj(mem));
return res;
}
/**
- * optimize a Store
+ * follow the memory chain as long as there are only Loads.
+ *
+ * INC_MASTER() must be called before dive into
*/
-static unsigned optimize_store(ir_node *store)
-{
- ldst_info_t *info = get_irn_link(store);
- ir_node *pred, *mem, *ptr, *value, *block;
- ir_mode *mode;
+static unsigned follow_Load_chain_for_Store(ir_node *store, ir_node *curr) {
unsigned res = 0;
-
- if (get_Store_volatility(store) == volatility_is_volatile)
- return 0;
-
- /*
- * 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 *(type1 *)p = a; *(type2 *)p = b ...
- */
-
- ptr = get_Store_ptr(store);
-
- /* 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_out_n(ptr) <= 1)
- return 0;
-
- block = get_nodes_block(store);
- mem = get_Store_mem(store);
- value = get_Store_value(store);
- mode = get_irn_mode(value);
-
- /* follow the memory chain as long as there are only Loads */
- for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) {
+ 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);
+
+ for (pred = curr; pred != store; pred = skip_Proj(get_Load_mem(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 is C were the following
+ * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
+ */
if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
get_nodes_block(pred) == block && get_irn_mode(get_Store_value(pred)) == mode) {
/*
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) );
+ reduce_adr_usage(ptr);
return DF_CHANGED;
}
}
if (! info->projs[pn_Store_X_except]) {
DBG_OPT_WAR(store, pred);
exchange( info->projs[pn_Store_M], mem );
+ reduce_adr_usage(ptr);
return DF_CHANGED;
}
}
/* follow only Load chains */
if (get_irn_op(pred) != op_Load)
break;
+
+ /* check for cycles */
+ if (NODE_VISITED(pred_info))
+ break;
+ MARK_NODE(pred_info);
+ }
+
+ if (get_irn_op(pred) == op_Sync) {
+ int i;
+
+ /* handle all Sync predecessors */
+ for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
+ res |= follow_Load_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
+ if (res)
+ break;
+ }
}
return res;
}
/**
- * walker, optimizes Phi after Stores:
+ * optimize a Store
+ */
+static unsigned optimize_store(ir_node *store)
+{
+ ldst_info_t *info = get_irn_link(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 load is used more than once.
+ * If not, this load 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_Load_chain_for_Store(store, skip_Proj(mem));
+}
+
+/**
+ * walker, optimizes Phi after Stores to identical places:
* Does the following optimization:
+ * @verbatim
*
* val1 val2 val3 val1 val2 val3
* | | | \ | /
* Str Str Str \ | /
- * \ | / Phi
+ * \ | / PhiData
* \ | / |
* \ | / Str
- * Phi
+ * PhiM
*
- * This removes the number of stores and allows for predicated execution.
- * Moves Stores back to the end of a function which may be bad
+ * @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.
*
- * Is only allowed if the predecessor blocks have only one successor.
+ * This is only possible if the predecessor blocks have only one successor.
*/
-static unsigned optimize_phi(ir_node *phi, void *env)
+static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
{
- walk_env_t *wenv = env;
int i, n;
ir_node *store, *old_store, *ptr, *block, *phiM, *phiD, *exc, *projM;
ir_mode *mode;
if (is_Block_dead(get_nodes_block(store)))
return 0;
- /* check if the block has only one output */
+ /* check if the block has only one successor */
bl_info = get_irn_link(get_nodes_block(store));
if (bl_info->flags)
return 0;
if (get_irn_op(pred) != op_Store)
return 0;
- if (mode != get_irn_mode(get_Store_value(pred)) || ptr != get_Store_ptr(pred))
+ if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
return 0;
info = get_irn_link(pred);
if (is_Block_dead(get_nodes_block(store)))
return 0;
- /* check if the block has only one output */
+ /* check if the block has only one successor */
bl_info = get_irn_link(get_nodes_block(store));
if (bl_info->flags)
return 0;
/*
* ok, when we are here, we found all predecessors of a Phi that
- * are Stores to the same address. 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:
+ * 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 | \ | /
- * \ | / Phi
+ * \ | / PhiData
* \ | / |
* \ | / Str
- * Phi
+ * PhiM
*
* Is only allowed if the predecessor blocks have only one successor.
*/
inM[i] = get_Store_mem(pred);
inD[i] = get_Store_value(pred);
idx[i] = info->exc_idx;
+
+ /* Should we here replace the Proj after the Store by
+ * the Store's memory? Would be save but should not be needed,
+ * because we checked that all pred blocks have only one
+ * control flow successor.
+ */
}
block = get_nodes_block(phi);
#ifdef DO_CACHEOPT
co_set_irn_name(store, co_get_irn_ident(old_store));
#endif
+ /* we replaced n uses by 1 */
+ set_irn_n_uses(ptr, get_irn_n_uses(ptr) - n + 1);
projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
res |= CF_CHANGED;
}
- /* sixt step: replace old Phi */
+ /* sixth step: replace old Phi */
exchange(phi, projM);
return res | DF_CHANGED;
}
/**
- * walker, collects all Load/Store/Proj nodes
+ * walker, do the optimizations
*/
static void do_load_store_optimize(ir_node *n, void *env)
{
break;
case iro_Phi:
- wenv->changes |= optimize_phi(n, env);
+ wenv->changes |= optimize_phi(n, wenv);
default:
;
assert(get_irg_pinned(irg) != op_pin_state_floats &&
"LoadStore optimization needs pinned graph");
- if (!get_opt_redundant_LoadStore())
+ if (! get_opt_redundant_loadstore())
return;
obstack_init(&env.obst);
env.changes = 0;
/* init the links, then collect Loads/Stores/Proj's in lists */
+ master_visited = 0;
irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
/* now we have collected enough information, optimize */
/* Handle graph state */
if (env.changes) {
- if (get_irg_outs_state(current_ir_graph) == outs_consistent)
- set_irg_outs_inconsistent(current_ir_graph);
+ if (get_irg_outs_state(irg) == outs_consistent)
+ set_irg_outs_inconsistent(irg);
}
if (env.changes & CF_CHANGED) {
- /* is this really needed: Yes, control flow changed, block might get Bad. */
- if (get_irg_dom_state(current_ir_graph) == dom_consistent)
- set_irg_dom_inconsistent(current_ir_graph);
+ /* is this really needed: Yes, control flow changed, block might
+ have Bad() predecessors. */
+ set_irg_doms_inconsistent(irg);
}
}