#define MAX_PROJ IMAX(pn_Load_max, pn_Store_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 */
- int changes;
+ unsigned changes; /**< a bitmask of graph changes */
} walk_env_t;
/**
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;
-/**
- * walker, clears all links first
- */
-static void init_links(ir_node *n, void *env)
-{
- set_irn_link(n, NULL);
-}
+/** 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 projection info for a Load/Store
*/
-static int update_projs(ldst_info_t *info, ir_node *proj)
+static unsigned update_projs(ldst_info_t *info, ir_node *proj)
{
long nr = get_Proj_proj(proj);
if (info->projs[nr]) {
/* there is already one, do CSE */
exchange(proj, info->projs[nr]);
- return 1;
+ return DF_CHANGED;
}
else {
info->projs[nr] = proj;
}
/**
- * 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 int update_exc(ldst_info_t *info, ir_node *block, int pos)
+static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
{
assert(info->exc_block == NULL && "more than one exception block found");
/**
* walker, collects all Load/Store/Proj nodes
*
- * walks form Start -> End
+ * walks from Start -> End
*/
static void collect_nodes(ir_node *node, void *env)
{
ir_op *op = get_irn_op(node);
- ir_node *pred;
+ ir_node *pred, *blk, *pred_blk;
ldst_info_t *ldst_info;
walk_env_t *wenv = env;
ldst_info->flags |= LDST_VISITED;
}
+
+ /*
+ * 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 (op == op_Store) {
ldst_info = get_ldst_info(pred, wenv);
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) { /* 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);
/**
* optimize a Load
*/
-static int optimize_load(ir_node *load)
+static unsigned optimize_load(ir_node *load)
{
ldst_info_t *info = get_irn_link(load);
ir_mode *load_mode = get_Load_mode(load);
ir_node *pred, *mem, *ptr, *new_node;
entity *ent;
- int res = 0;
+ unsigned res = 0;
/* do NOT touch volatile loads for now */
if (get_Load_volatility(load) == volatility_is_volatile)
exchange(info->projs[pn_Load_X_except], new_Bad());
info->projs[pn_Load_X_except] = NULL;
- res = 1;
+ res |= CF_CHANGED;
}
}
}
*/
exchange(info->projs[pn_Load_X_except], new_Bad());
info->projs[pn_Load_X_except] = NULL;
- res = 1;
+ res |= CF_CHANGED;
}
}
/* a Load which value is neither used nor exception checked, remove it */
exchange(info->projs[pn_Load_M], mem);
- return 1;
+ return res | DF_CHANGED;
}
/* Load from a constant polymorphic field, where we can resolve
}
if (info->projs[pn_Load_res])
exchange(info->projs[pn_Load_res], new_node);
- return 1;
+ return res | DF_CHANGED;
}
/* check if we can determine the entity that will be loaded */
if (info->projs[pn_Load_X_except]) {
exchange(info->projs[pn_Load_X_except], new_Bad());
info->projs[pn_Load_X_except] = NULL;
- res = 1;
+ res |= CF_CHANGED;
}
if (variability_constant == get_entity_variability(ent)
/* no memory */
if (info->projs[pn_Load_M]) {
exchange(info->projs[pn_Load_M], mem);
- res = 1;
+ res |= DF_CHANGED;
}
/* no result :-) */
DBG_OPT_RC(load, c);
exchange(info->projs[pn_Load_res], c);
- return 1;
+ return DF_CHANGED | res;
}
}
}
if (info->projs[pn_Load_M]) {
exchange(info->projs[pn_Load_M], mem);
- res = 1;
+ res |= DF_CHANGED;
}
if (info->projs[pn_Load_res]) {
exchange(info->projs[pn_Load_res], copy_const_value(c));
- return 1;
+ return res | DF_CHANGED;
}
}
else {
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.
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
exchange(info->projs[pn_Load_M], mem);
/* no exception */
- if (info->projs[pn_Load_X_except])
+ 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 1;
+ return res | DF_CHANGED;
}
}
else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
* 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]) {
}
/* no exception */
- if (info->projs[pn_Load_X_except])
+ if (info->projs[pn_Load_X_except]) {
exchange(info->projs[pn_Load_X_except], new_Bad());
+ res |= CF_CHANGED;
+ }
- return 1;
+ 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);
+ }
return res;
}
/**
* optimize a Store
*/
-static int optimize_store(ir_node *store)
+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;
- int res = 0;
+ unsigned res = 0;
if (get_Store_volatility(store) == volatility_is_volatile)
return 0;
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 &&
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) );
- return 1;
+ return DF_CHANGED;
}
}
else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
if (! info->projs[pn_Store_X_except]) {
DBG_OPT_WAR(store, pred);
exchange( info->projs[pn_Store_M], mem );
- return 1;
+ 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);
}
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 int optimize_phi(ir_node *phi, void *env)
+static unsigned optimize_phi(ir_node *phi, void *env)
{
walk_env_t *wenv = env;
int i, n;
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)
if (get_irn_op(store) != op_Store)
return 0;
- /* abort on bad blocks */
- if (is_Bad(get_nodes_block(store)))
+ /* abort on dead blocks */
+ 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 (exc != info->exc_block)
return 0;
- /* abort on bad blocks */
- if (is_Bad(get_nodes_block(store)))
+ /* abort on dead blocks */
+ 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.
*/
if (n > 1) {
/* the exception block should be optimized as some inputs are identical now */
}
+
+ res |= CF_CHANGED;
}
- /* sixt step: replace old Phi */
+ /* sixth step: replace old Phi */
exchange(phi, projM);
- return 1;
+ return res | DF_CHANGED;
}
/**
- * walker, collects all Load/Store/Proj nodes
+ * walker, do the optiimizations
*/
static void do_load_store_optimize(ir_node *n, void *env)
{
env.changes = 0;
/* init the links, then collect Loads/Stores/Proj's in lists */
- irg_walk_graph(irg, init_links, collect_nodes, &env);
+ master_visited = 0;
+ irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
/* now we have collected enough information, optimize */
irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
if (env.changes) {
if (get_irg_outs_state(current_ir_graph) == outs_consistent)
set_irg_outs_inconsistent(current_ir_graph);
+ }
- /* is this really needed: Yes, as exception block may get bad but this might be tested */
+ 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);
}
projects / libfirm / blobdiff