* @author Michael Beck
* @version $Id$
*/
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
#include <string.h>
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
-#include "array.h"
+#include "array_t.h"
#include "irhooks.h"
#include "iredges.h"
-#include "irtools.h"
+#include "irpass.h"
#include "opt_polymorphy.h"
#include "irmemory.h"
-#include "xmalloc.h"
#include "irphase_t.h"
#include "irgopt.h"
+#include "set.h"
#include "debug.h"
/** The debug handle. */
ldst_info_t *info = get_irn_link(node);
if (! info) {
- info = obstack_alloc(obst, sizeof(*info));
- memset(info, 0, sizeof(*info));
+ info = OALLOCZ(obst, ldst_info_t);
set_irn_link(node, info);
}
return info;
block_info_t *info = get_irn_link(node);
if (! info) {
- info = obstack_alloc(obst, sizeof(*info));
- memset(info, 0, sizeof(*info));
+ info = OALLOCZ(obst, block_info_t);
set_irn_link(node, info);
}
return info;
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.
+ * at least its length.
*/
assert(get_SymConst_kind(ptr) == symconst_addr_ent);
root = get_SymConst_entity(ptr);
static void reduce_adr_usage(ir_node *ptr);
/**
- * Update a Load that may lost it's usage.
+ * Update a Load that may have lost its users.
*/
static void handle_load_update(ir_node *load) {
ldst_info_t *info = get_irn_link(load);
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 */
+ /* a Load whose 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)));
+ exchange(info->projs[pn_Load_X_regular], new_r_Jmp(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
+ * A use of an address node has vanished. Check if this was a Proj
* node and update the counters.
*/
static void reduce_adr_usage(ir_node *ptr) {
ldst_info_t *info = get_irn_link(pred);
info->projs[get_Proj_proj(ptr)] = NULL;
- /* this node lost it's result proj, handle that */
+ /* this node lost its result proj, handle that */
handle_load_update(pred);
}
}
store_mode = get_irn_mode(get_Store_value(store));
store_mode_len = get_mode_size_bytes(store_mode);
delta = load_offset - store_offset;
- if (delta < 0 || delta + load_mode_len > store_mode_len)
- return 0;
+ store_value = get_Store_value(store);
- if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
- get_mode_arithmetic(load_mode) != irma_twos_complement)
- return 0;
+ if (delta != 0 || store_mode != load_mode) {
+ if (delta < 0 || delta + load_mode_len > store_mode_len)
+ return 0;
+
+ if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
+ get_mode_arithmetic(load_mode) != irma_twos_complement)
+ return 0;
- store_value = get_Store_value(store);
- /* produce a shift to adjust offset delta */
- if (delta > 0) {
- ir_node *cnst;
+ /* produce a shift to adjust offset delta */
+ if (delta > 0) {
+ ir_node *cnst;
- /* FIXME: only true for little endian */
- cnst = new_Const_long(mode_Iu, delta * 8);
- store_value = new_r_Shr(current_ir_graph, get_nodes_block(load),
- store_value, cnst, store_mode);
- }
+ /* FIXME: only true for little endian */
+ cnst = new_Const_long(mode_Iu, delta * 8);
+ store_value = new_r_Shr(get_nodes_block(load),
+ store_value, cnst, store_mode);
+ }
- /* add an convert if needed */
- if (store_mode != load_mode) {
- store_value = new_r_Conv(current_ir_graph, get_nodes_block(load),
- store_value, load_mode);
+ /* add an convert if needed */
+ if (store_mode != load_mode) {
+ store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
+ }
}
DBG_OPT_RAW(load, store_value);
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)));
+ exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
res |= CF_CHANGED;
}
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);
+ pred_info->projs[pn_Load_res] = new_r_Proj(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);
+ value = new_r_Conv(get_nodes_block(load), value, load_mode);
}
exchange(info->projs[pn_Load_res], value);
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)));
+ exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
res |= CF_CHANGED;
}
/* copy the value from the const code irg and cast it */
res = copy_const_value(dbg, c);
- res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
+ res = new_rd_Conv(dbg, block, res, l_mode);
}
} else {
/* copy the value from the const code irg */
*/
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)));
+ exchange(info->projs[pn_Load_X_regular], new_r_Jmp(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);
+ 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 */
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)));
+ exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
res |= CF_CHANGED;
}
kill_node(load);
value = NULL;
/* check if we can determine the entity that will be loaded */
ent = find_constant_entity(ptr);
- if (ent != NULL) {
- 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 even if we cannot replace the load itself. */
-
- /* 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 (ent != NULL &&
+ 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 even if we cannot replace the load itself. */
+
+ /* 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(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
- */
- value = get_atomic_ent_value(ent);
- } else {
- if (ent->has_initializer) {
- /* new style initializer */
- value = find_compound_ent_value(ptr);
- } else {
- /* old style initializer */
- compound_graph_path *path = get_accessed_path(ptr);
-
- if (path != NULL) {
- assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
-
- value = get_compound_ent_value_by_path(ent, path);
- DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
- free_compound_graph_path(path);
- }
- }
+ 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 */
+ value = get_atomic_ent_value(ent);
+ } else if (ent->has_initializer) {
+ /* new style initializer */
+ value = find_compound_ent_value(ptr);
+ } else {
+ /* old style initializer */
+ compound_graph_path *path = get_accessed_path(ptr);
+
+ if (path != NULL) {
+ assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
+
+ value = get_compound_ent_value_by_path(ent, path);
+ DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
+ free_compound_graph_path(path);
}
- if (value != NULL)
- value = can_replace_load_by_const(load, value);
}
+ if (value != NULL)
+ value = can_replace_load_by_const(load, value);
}
}
}
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)));
+ exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
info->projs[pn_Load_X_regular] = NULL;
res |= CF_CHANGED;
}
return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
} /* is_completely_overwritten */
+/**
+ * Check whether small is a part of large (starting at same address).
+ */
+static int is_partially_same(ir_node *small, ir_node *large)
+{
+ ir_mode *sm = get_irn_mode(small);
+ ir_mode *lm = get_irn_mode(large);
+
+ /* FIXME: Check endianness */
+ return is_Conv(small) && get_Conv_op(small) == large
+ && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
+ && get_mode_arithmetic(sm) == irma_twos_complement
+ && get_mode_arithmetic(lm) == irma_twos_complement;
+} /* is_partially_same */
+
/**
* follow the memory chain as long as there are only Loads and alias free Stores.
*
* 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 ...
+ * However, if the size of the mode that is written is bigger or equal the
+ * size of 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)) {
+ get_nodes_MacroBlock(pred) == mblk) {
/*
* 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.
+ */
+
+ /*
+ * We may remove the first Store, if the old value is completely
+ * overwritten or the old value is a part of the new value,
+ * and 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;
+ if (get_Store_volatility(pred) != volatility_is_volatile
+ && !pred_info->projs[pn_Store_X_except]) {
+ ir_node *predvalue = get_Store_value(pred);
+ ir_mode *predmode = get_irn_mode(predvalue);
+
+ if(is_completely_overwritten(predmode, mode)
+ || is_partially_same(predvalue, value)) {
+ 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;
+ }
+ }
+
+ /*
+ * We may remove the Store, if the old value already contains
+ * the new value, and if it does not have an exception handler.
+ *
+ * TODO: What, if both have the same exception handler ???
+ */
+ if (get_Store_volatility(store) != volatility_is_volatile
+ && !info->projs[pn_Store_X_except]) {
+ ir_node *predvalue = get_Store_value(pred);
+
+ if(is_partially_same(value, predvalue)) {
+ DBG_OPT_WAW(pred, store);
+ exchange(info->projs[pn_Store_M], mem);
+ kill_node(store);
+ reduce_adr_usage(ptr);
+ return DF_CHANGED;
+ }
}
} else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
value == pred_info->projs[pn_Load_res]) {
if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
ldst_info_t *info = get_irn_link(store);
if (info->projs[pn_Store_X_except] == NULL) {
+ DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity));
exchange(info->projs[pn_Store_M], get_Store_mem(store));
kill_node(store);
reduce_adr_usage(ptr);
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);
+ phiM = new_rd_Phi(get_irn_dbg_info(phi), 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);
+ phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
/* rewire memory and kill the node */
for (i = n - 1; i >= 0; --i) {
}
/* fourth step: create the Store */
- store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
+ store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
#ifdef DO_CACHEOPT
co_set_irn_name(store, co_get_irn_ident(old_store));
#endif
- projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
+ projM = new_rd_Proj(NULL, block, store, mode_M, pn_Store_M);
info = get_ldst_info(store, &wenv->obst);
info->projs[pn_Store_M] = projM;
/* fifths step: repair exception flow */
if (exc) {
- ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
+ ir_node *projX = new_rd_Proj(NULL, block, store, mode_X, pn_Store_X_except);
info->projs[pn_Store_X_except] = projX;
info->exc_block = exc;
typedef struct node_entry {
unsigned DFSnum; /**< the DFS number of this node */
unsigned low; /**< the low number of this node */
- ir_node *header; /**< the header of this node */
int in_stack; /**< flag, set if the node is on the stack */
ir_node *next; /**< link to the next node the the same scc */
scc *pscc; /**< the scc of this node */
phi_entry *next;
};
+/**
+ * An entry in the avail set.
+ */
+typedef struct avail_entry_t {
+ ir_node *ptr; /**< the address pointer */
+ ir_mode *mode; /**< the load mode */
+ ir_node *load; /**< the associated Load */
+} avail_entry_t;
+
+/**
+ * Compare two avail entries.
+ */
+static int cmp_avail_entry(const void *elt, const void *key, size_t size) {
+ const avail_entry_t *a = elt;
+ const avail_entry_t *b = key;
+ (void) size;
+
+ return a->ptr != b->ptr || a->mode != b->mode;
+} /* cmp_avail_entry */
+
+/**
+ * Calculate the hash value of an avail entry.
+ */
+static unsigned hash_cache_entry(const avail_entry_t *entry) {
+ return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
+} /* hash_cache_entry */
+
/**
* Move loops out of loops if possible.
*
ir_entity *ent;
int j;
phi_entry *phi_list = NULL;
+ set *avail;
+
+ avail = new_set(cmp_avail_entry, 8);
/* collect all outer memories */
for (phi = pscc->head; phi != NULL; phi = next) {
if (! is_Phi(phi))
continue;
- assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
+ assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
ir_node *pred = get_irn_n(phi, j);
/* no Phis no fun */
assert(phi_list != NULL && "DFS found a loop without Phi");
+ /* for now, we cannot handle more than one input (only reducible cf) */
+ if (phi_list->next != NULL)
+ return;
+
for (load = pscc->head; load; load = next) {
ir_mode *load_mode;
node_entry *ne = get_irn_ne(load, env);
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(Global) */
+ /* for now, we can only move Load(Global) */
if (! is_Global(ptr))
continue;
- ent = get_Global_entity(ptr);
+ ent = get_Global_entity(ptr);
load_mode = get_Load_mode(load);
for (other = pscc->head; other != NULL; other = next_other) {
node_entry *ne = get_irn_ne(other, env);
if (rel != ir_no_alias)
break;
}
- /* only pure Calls are allowed here, so ignore them */
+ /* only Phis and pure Calls are allowed here, so ignore them */
}
if (other == NULL) {
ldst_info_t *ninfo;
phi_entry *pe;
dbg_info *db;
- /* for now, we cannot handle more than one input */
- if (phi_list->next != NULL)
- return;
-
/* yep, no aliasing Store found, Load can be moved */
DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
ir_node *blk = get_nodes_block(phi);
ir_node *pred = get_Block_cfgpred_block(blk, pos);
ir_node *irn, *mem;
+ avail_entry_t entry, *res;
- pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
+ entry.ptr = ptr;
+ entry.mode = load_mode;
+ res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
+ if (res != NULL) {
+ irn = res->load;
+ } else {
+ irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
+ entry.load = irn;
+ set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
+ DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
+ }
+ pe->load = irn;
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);
+ ninfo->projs[pn_Load_M] = mem = new_r_Proj(pred, irn, mode_M, pn_Load_M);
set_Phi_pred(phi, pos, mem);
- ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
-
- DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
+ ninfo->projs[pn_Load_res] = new_r_Proj(pred, irn, load_mode, pn_Load_res);
}
/* now kill the old Load */
}
}
}
+ del_set(avail);
} /* move_loads_out_of_loops */
/**
next = e->next;
b = get_irn_ne(block, env);
- if (header) {
+ if (header != NULL) {
if (h->POnum < b->POnum) {
header = block;
h = b;
}
- }
- else {
+ } else {
header = block;
h = b;
}
/* not a memory loop */
goto fail;
}
- if (! out_rc) {
+ if (out_rc == NULL) {
+ /* first region constant */
out_rc = pred;
++num_outside;
} else if (out_rc != pred) {
+ /* another region constant */
++num_outside;
}
}
for (irn = pscc->head; irn; irn = next) {
node_entry *e = get_irn_ne(irn, env);
next = e->next;
- e->header = NULL;
exchange(irn, out_rc);
}
env->changes |= DF_CHANGED;
return;
}
- /* set the header for every node in this scc */
+#ifdef DEBUG_libfirm
for (irn = pscc->head; irn; irn = next) {
node_entry *e = get_irn_ne(irn, env);
- e->header = header;
next = e->next;
DB((dbg, LEVEL_2, " %+F,", irn));
}
DB((dbg, LEVEL_2, "\n"));
-
+#endif
move_loads_out_of_loops(pscc, env);
fail:
ir_node *pred = get_irn_n(irn, i);
node_entry *o = get_irn_ne(pred, env);
- if (irn_not_visited(pred)) {
+ if (!irn_visited(pred)) {
dfs(pred, env);
node->low = MIN(node->low, o->low);
}
ir_node *pred = get_fragile_op_mem(irn);
node_entry *o = get_irn_ne(pred, env);
- if (irn_not_visited(pred)) {
+ if (!irn_visited(pred)) {
dfs(pred, env);
node->low = MIN(node->low, o->low);
}
ir_node *pred = get_Proj_pred(irn);
node_entry *o = get_irn_ne(pred, env);
- if (irn_not_visited(pred)) {
+ if (!irn_visited(pred)) {
dfs(pred, env);
node->low = MIN(node->low, o->low);
}
for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
ir_node *ka = get_End_keepalive(end, i);
- if (is_Phi(ka) && irn_not_visited(ka))
+ if (is_Phi(ka) && !irn_visited(ka))
dfs(ka, env);
}
current_ir_graph = rem;
/*
* do the load store optimization
*/
-void optimize_load_store(ir_graph *irg) {
+int optimize_load_store(ir_graph *irg) {
walk_env_t env;
FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
have Bad() predecessors. */
set_irg_doms_inconsistent(irg);
}
+ return env.changes != 0;
} /* optimize_load_store */
+
+ir_graph_pass_t *optimize_load_store_pass(const char *name)
+{
+ return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
+} /* optimize_load_store_pass */