/*
- * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
#undef IMAX
#define IMAX(a,b) ((a) > (b) ? (a) : (b))
-#define MAX_PROJ IMAX(pn_Load_max, pn_Store_max)
+#define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
enum changes_t {
DF_CHANGED = 1, /**< data flow changed */
unsigned changes; /**< a bitmask of graph changes */
} walk_env_t;
-/**
- * flags for Load/Store
- */
-enum ldst_flags_t {
- LDST_VISITED = 1 /**< if set, this Load/Store is already visited */
-};
-
/** 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;
*/
static void collect_nodes(ir_node *node, void *env)
{
- ir_op *op = get_irn_op(node);
+ ir_opcode opcode = get_irn_opcode(node);
ir_node *pred, *blk, *pred_blk;
ldst_info_t *ldst_info;
walk_env_t *wenv = env;
- if (op == op_Proj) {
- ir_node *adr;
- ir_op *op;
+ if (opcode == iro_Proj) {
+ pred = get_Proj_pred(node);
+ opcode = get_irn_opcode(pred);
- pred = get_Proj_pred(node);
- op = get_irn_op(pred);
-
- if (op == op_Load) {
+ 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);
- if ((ldst_info->flags & LDST_VISITED) == 0) {
- adr = get_Load_ptr(pred);
- ldst_info->flags |= LDST_VISITED;
- }
-
/*
* Place the Proj's to the same block as the
* predecessor Load. This is always ok and prevents
wenv->changes |= DF_CHANGED;
set_nodes_block(node, pred_blk);
}
- } else if (op == op_Store) {
- ldst_info = get_ldst_info(pred, &wenv->obst);
-
- wenv->changes |= update_projs(ldst_info, node);
-
- if ((ldst_info->flags & LDST_VISITED) == 0) {
- adr = get_Store_ptr(pred);
- ldst_info->flags |= LDST_VISITED;
- }
-
- /*
- * Place the Proj's to the same block as the
- * predecessor Store. 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 (op == op_Block) {
+ } else if (opcode == iro_Block) {
int i;
for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
else if (is_irn_forking(pred))
bl_info->flags |= BLOCK_HAS_COND;
- if (is_exc && (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store)) {
+ 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);
*/
static long get_Sel_array_index_long(ir_node *n, int dim) {
ir_node *index = get_Sel_index(n, dim);
- assert(get_irn_op(index) == op_Const);
+ assert(is_Const(index));
return get_tarval_long(get_Const_tarval(index));
} /* get_Sel_array_index_long */
/* a Load which value is neither used nor exception checked, remove it */
exchange(info->projs[pn_Load_M], mem);
- exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
- exchange(load, new_Bad());
+ 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 */
} /* can_use_stored_value */
/**
- * Follow the memory chain as long as there are only Loads
- * and alias free Stores and try to replace current Load or Store
- * by a previous ones.
+ * 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.
if (info->projs[pn_Load_res])
exchange(info->projs[pn_Load_res], value);
- exchange(load, new_Bad());
+ kill_node(load);
reduce_adr_usage(ptr);
return res | DF_CHANGED;
}
res |= CF_CHANGED;
}
- exchange(load, new_Bad());
+ kill_node(load);
reduce_adr_usage(ptr);
return res |= DF_CHANGED;
}
if (rel != no_alias)
break;
pred = skip_Proj(get_Store_mem(pred));
- } else if (get_irn_op(pred) == op_Load) {
+ } 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;
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)
- break;
+ return res;
}
}
exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
res |= CF_CHANGED;
}
- exchange(load, new_Bad());
+ kill_node(load);
reduce_adr_usage(ptr);
return res | DF_CHANGED;
}
if (info->projs[pn_Load_res])
exchange(info->projs[pn_Load_res], new_node);
- exchange(load, new_Bad());
+ kill_node(load);
reduce_adr_usage(ptr);
return res | DF_CHANGED;
}
res |= CF_CHANGED;
}
- if (variability_constant == get_entity_variability(ent)
- && 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;
- }
- }
- exchange(load, new_Bad());
- reduce_adr_usage(ptr);
- return res;
- } else if (variability_constant == get_entity_variability(ent)) {
- compound_graph_path *path = get_accessed_path(ptr);
-
- if (path) {
- ir_node *c;
-
- assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
- /*
- {
- int j;
- for (j = 0; j < get_compound_graph_path_length(path); ++j) {
- ir_entity *node = get_compound_graph_path_node(path, j);
- fprintf(stdout, ".%s", get_entity_name(node));
- if (is_Array_type(get_entity_owner(node)))
- fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
- }
- printf("\n");
- }
- */
-
- c = get_compound_ent_value_by_path(ent, path);
- free_compound_graph_path(path);
-
- /* printf(" cons: "); DDMN(c); */
-
+ 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]) {
- exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
- res |= DF_CHANGED;
+ 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;
+ }
}
- exchange(load, new_Bad());
+ 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)));
- printf(" load: "); DDMN(load);
- printf(" ptr: "); DDMN(ptr);
- */
+ compound_graph_path *path = get_accessed_path(ptr);
+
+ if (path) {
+ ir_node *c;
+
+ assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
+ /*
+ {
+ int j;
+ for (j = 0; j < get_compound_graph_path_length(path); ++j) {
+ ir_entity *node = get_compound_graph_path_node(path, j);
+ fprintf(stdout, ".%s", get_entity_name(node));
+ if (is_Array_type(get_entity_owner(node)))
+ fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
+ }
+ printf("\n");
+ }
+ */
+
+ c = get_compound_ent_value_by_path(ent, path);
+ free_compound_graph_path(path);
+
+ /* printf(" cons: "); DDMN(c); */
+
+ 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)));
+ printf(" load: "); DDMN(load);
+ printf(" ptr: "); DDMN(ptr);
+ */
+ }
}
}
}
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));
- exchange(pred, new_Bad());
+ 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 after a Load -- a write after read.
- * We may remove the second Store, if it does not have an exception handler.
+ * 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);
- exchange(store, new_Bad());
+ kill_node(store);
reduce_adr_usage(ptr);
return DF_CHANGED;
}
break;
pred = skip_Proj(get_Store_mem(pred));
} else if (get_irn_op(pred) == op_Load) {
+ ir_alias_relation rel = get_alias_relation(
+ current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
+ ptr, mode);
+ if (rel != no_alias)
+ break;
+
pred = skip_Proj(get_Load_mem(pred));
} else {
/* follow only Load chains */
case iro_Phi:
wenv->changes |= optimize_phi(n, wenv);
+ break;
default:
;
};
/**
- * Move loops out of loops if possible
+ * Move loops out of loops if possible.
+ *
+ * @param pscc the loop described by an SCC
+ * @param env the loop environment
*/
-static void move_loads_in_loops(scc *pscc, loop_env *env) {
+static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
ir_node *phi, *load, *next, *other, *next_other;
ir_entity *ent;
int j;
next = ne->next;
/* check all memory Phi's */
- if (! is_Phi(phi) || get_irn_mode(phi) != mode_M)
+ if (! is_Phi(phi))
continue;
+ assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
+
for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
ir_node *pred = get_irn_n(phi, j);
node_entry *pe = get_irn_ne(pred, env);
}
}
/* no Phis no fun */
- if (phi_list == NULL)
- return;
+ assert(phi_list != NULL && "DFS found a loop without Phi");
for (load = pscc->head; load; load = next) {
ir_mode *load_mode;
if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
continue;
- /* for now, we can only handle Load(SymConst) */
- if (! is_SymConst(ptr) || get_SymConst_kind(ptr) != symconst_addr_ent)
+ /* for now, we can only handle Load(Global) */
+ if (! is_Global(ptr))
continue;
- ent = get_SymConst_entity(ptr);
-
+ ent = get_Global_entity(ptr);
load_mode = get_Load_mode(load);
- if (get_entity_address_taken(ent) == ir_address_not_taken) {
- /* Shortcut: If the addres is never taken, this address if complete alias free*/
- goto can_move;
- }
-
for (other = pscc->head; other != NULL; other = next_other) {
node_entry *ne = get_irn_ne(other, env);
next_other = ne->next;
if (rel != no_alias)
break;
}
+ /* only pure Calls are allowed here, so ignore them */
}
if (other == NULL) {
ldst_info_t *ninfo;
phi_entry *pe;
dbg_info *db;
-can_move:
+
/* for now, we cannot handle more than one input */
if (phi_list->next != NULL)
return;
ir_node *irn, *mem;
pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
- ninfo = get_ldst_info(pred, phase_obst(&env->ph));
+ ninfo = get_ldst_info(irn, phase_obst(&env->ph));
ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
set_Phi_pred(phi, pos, mem);
}
}
}
-} /* move_loads_in_loops */
+} /* move_loads_out_of_loops */
/**
* Process a loop SCC.
next = e->next;
switch (get_irn_opcode(irn)) {
case iro_Call:
+ if (is_Call_pure(irn)) {
+ /* pure calls can be treated like loads */
+ only_phi = 0;
+ break;
+ }
+ /* non-pure calls must be handle like may-alias Stores */
+ goto fail;
case iro_CopyB:
- /* cannot handle Calls or CopyB yet */
+ /* cannot handle CopyB yet */
goto fail;
case iro_Load:
process = 1;
}
DB((dbg, LEVEL_2, "\n"));
- move_loads_in_loops(pscc, env);
+ move_loads_out_of_loops(pscc, env);
fail:
;
process_loop(pscc, env);
}
} /* process_scc */
+
/**
* Do Tarjan's SCC algorithm and drive load/store optimization.
*
mark_irn_visited(irn);
- /* do not put blocks into the scc */
- if (is_Block(irn)) {
- n = get_irn_arity(irn);
- for (i = 0; i < n; ++i) {
- ir_node *pred = get_irn_n(irn, i);
-
- if (irn_not_visited(pred))
- dfs(pred, env);
- }
- }
- else {
- ir_node *block = get_nodes_block(irn);
-
- node->DFSnum = env->nextDFSnum++;
- node->low = node->DFSnum;
- push(env, irn);
-
- /* handle the block */
- if (irn_not_visited(block))
- dfs(block, env);
+ node->DFSnum = env->nextDFSnum++;
+ node->low = node->DFSnum;
+ push(env, irn);
+ /* handle preds */
+ if (is_Phi(irn) || is_Sync(irn)) {
n = get_irn_arity(irn);
for (i = 0; i < n; ++i) {
ir_node *pred = get_irn_n(irn, i);
if (o->DFSnum < node->DFSnum && o->in_stack)
node->low = MIN(o->DFSnum, node->low);
}
- if (node->low == node->DFSnum) {
- scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
- ir_node *x;
-
- pscc->head = NULL;
- do {
- node_entry *e;
-
- x = pop(env);
- e = get_irn_ne(x, env);
- e->pscc = pscc;
- e->next = pscc->head;
- pscc->head = x;
- } while (x != irn);
-
- process_scc(pscc, env);
+ } else if (is_fragile_op(irn)) {
+ ir_node *pred = get_fragile_op_mem(irn);
+ node_entry *o = get_irn_ne(pred, env);
+
+ if (irn_not_visited(pred)) {
+ dfs(pred, env);
+ node->low = MIN(node->low, o->low);
+ }
+ if (o->DFSnum < node->DFSnum && o->in_stack)
+ node->low = MIN(o->DFSnum, node->low);
+ } else if (is_Proj(irn)) {
+ ir_node *pred = get_Proj_pred(irn);
+ node_entry *o = get_irn_ne(pred, env);
+
+ if (irn_not_visited(pred)) {
+ dfs(pred, env);
+ node->low = MIN(node->low, o->low);
}
+ if (o->DFSnum < node->DFSnum && o->in_stack)
+ node->low = MIN(o->DFSnum, node->low);
+ }
+ else {
+ /* IGNORE predecessors */
+ }
+
+ if (node->low == node->DFSnum) {
+ scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
+ ir_node *x;
+
+ pscc->head = NULL;
+ do {
+ node_entry *e;
+
+ x = pop(env);
+ e = get_irn_ne(x, env);
+ e->pscc = pscc;
+ e->next = pscc->head;
+ pscc->head = x;
+ } while (x != irn);
+
+ process_scc(pscc, env);
}
} /* dfs */
/**
- * Do the DFS by starting at the End node of a graph.
+ * Do the DFS on the memory edges a graph.
*
* @param irg the graph to process
* @param env the loop environment
*/
static void do_dfs(ir_graph *irg, loop_env *env) {
ir_graph *rem = current_ir_graph;
- ir_node *end = get_irg_end(irg);
- int i, n;
+ ir_node *endblk, *end;
+ int i;
current_ir_graph = irg;
inc_irg_visited(irg);
- /* visit all visible nodes */
- dfs(end, env);
+ /* visit all memory nodes */
+ endblk = get_irg_end_block(irg);
+ for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
+ ir_node *pred = get_Block_cfgpred(endblk, i);
+
+ pred = skip_Proj(pred);
+ if (is_Return(pred))
+ dfs(get_Return_mem(pred), env);
+ else if (is_Raise(pred))
+ dfs(get_Raise_mem(pred), env);
+ else if (is_fragile_op(pred))
+ dfs(get_fragile_op_mem(pred), env);
+ else {
+ assert(0 && "Unknown EndBlock predecessor");
+ }
+ }
/* visit the keep-alives */
- n = get_End_n_keepalives(end);
- for (i = 0; i < n; ++i) {
+ end = get_irg_end(irg);
+ for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
ir_node *ka = get_End_keepalive(end, i);
- if (irn_not_visited(ka))
+ if (is_Phi(ka) && irn_not_visited(ka))
dfs(ka, env);
}
current_ir_graph = rem;
/**
* Initialize new phase data. We do this always explicit, so return NULL here
*/
-static void *init_loop_data(ir_phase *ph, ir_node *irn, void *data) {
+static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
(void)ph;
(void)irn;
(void)data;
walk_env_t env;
FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
- firm_dbg_set_mask(dbg, SET_LEVEL_1);
assert(get_irg_phase_state(irg) != phase_building);
assert(get_irg_pinned(irg) != op_pin_state_floats &&
"LoadStore optimization needs pinned graph");
- if (! get_opt_redundant_loadstore())
- return;
-
/* we need landing pads */
remove_critical_cf_edges(irg);
edges_assure(irg);
- /* loop optimizations need dominators ... */
- assure_doms(irg);
-
/* for Phi optimization post-dominators are needed ... */
assure_postdoms(irg);
projects / libfirm / blobdiff