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;
/**
*/
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 */
};
/**
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
*/
}
/**
- * 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)
{
/**
* walker, collects all Load/Store/Proj nodes
*
- * walks form Start -> End
+ * walks from Start -> End
*/
static void collect_nodes(ir_node *node, void *env)
{
}
}
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) ) )
* 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 */
}
/* 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]) {
/* 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);
res |= DF_CHANGED;
}
if (info->projs[pn_Load_res]) {
- exchange(info->projs[pn_Load_res], copy_const_value(c));
+ exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
return res | DF_CHANGED;
}
}
if (get_irn_out_n(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))) {
+ INC_MASTER();
+ for (pred = skip_Proj(mem); 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 = *(type1 *)p; b = *(type2 *)p ...
+ * 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) {
- 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
* 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]) {
/* 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);
+ }
return res;
}
* 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 ...
+ * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
*/
ptr = get_Store_ptr(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))) {
+ INC_MASTER();
+ for (pred = skip_Proj(mem); pred != store; pred = skip_Proj(get_Load_mem(pred))) {
ldst_info_t *pred_info = get_irn_link(pred);
if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
/* 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);
}
return res;
}
/**
- * walker, optimizes Phi after Stores:
+ * 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)
{
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.
*/
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)
{
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 */
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);
+ set_irg_doms_inconsistent(current_ir_graph);
}
}