X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fldstopt.c;h=6dc6ba1bf4f1d72a988cbc1cc588ecf0d4b6c08a;hb=6a6c1557afcd038930e649e29d69cb85a5dc372b;hp=c83c293d491608d507362bb0ee46128894f55de5;hpb=6f26ed069768912aa4ac49ec8bf1e2b6b64c2cee;p=libfirm diff --git a/ir/opt/ldstopt.c b/ir/opt/ldstopt.c index c83c293d4..6dc6ba1bf 100644 --- a/ir/opt/ldstopt.c +++ b/ir/opt/ldstopt.c @@ -8,6 +8,20 @@ * Copyright: (c) 1998-2004 Universität Karlsruhe * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. */ +#ifdef HAVE_CONFIG_H +# include "config.h" +#endif + +#ifdef HAVE_ALLOCA_H +#include +#endif +#ifdef HAVE_MALLOC_H +#include +#endif +#ifdef HAVE_STRING_H +# include +#endif + # include "irnode_t.h" # include "irgraph_t.h" # include "irmode_t.h" @@ -21,7 +35,12 @@ # include "iropt_dbg.h" # include "irflag_t.h" # include "array.h" -# include "firmstat.h" +# include "irhooks.h" +# include "opt_polymorphy.h" + +#ifdef DO_CACHEOPT +#include "cacheopt/cachesim.h" +#endif #undef IMAX #define IMAX(a,b) ((a) > (b) ? (a) : (b)) @@ -36,13 +55,21 @@ typedef struct _walk_env_t { int 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 */ + 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 */ } ldst_info_t; /** @@ -134,6 +161,9 @@ static int update_exc(ldst_info_t *info, ir_node *block, int pos) return 0; } +#define get_irn_out_n(node) (unsigned)get_irn_link(node) +#define set_irn_out_n(node, n) set_irn_link(adr, (void *)(n)) + /** * walker, collects all Load/Store/Proj nodes * @@ -141,20 +171,44 @@ static int update_exc(ldst_info_t *info, ir_node *block, int pos) */ static void collect_nodes(ir_node *node, void *env) { + ir_op *op = get_irn_op(node); ir_node *pred; ldst_info_t *ldst_info; walk_env_t *wenv = env; - if (get_irn_op(node) == op_Proj) { + if (op == op_Proj) { + ir_node *adr; + ir_op *op; + pred = get_Proj_pred(node); + op = get_irn_op(pred); + + if (op == op_Load) { + ldst_info = get_ldst_info(pred, wenv); + + wenv->changes |= update_projs(ldst_info, node); + + if ((ldst_info->flags & LDST_VISITED) == 0) { + adr = get_Load_ptr(pred); + set_irn_out_n(adr, get_irn_out_n(adr) + 1); - if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) { + ldst_info->flags |= LDST_VISITED; + } + } + else if (op == op_Store) { ldst_info = get_ldst_info(pred, wenv); wenv->changes |= update_projs(ldst_info, node); + + if ((ldst_info->flags & LDST_VISITED) == 0) { + adr = get_Store_ptr(pred); + set_irn_out_n(adr, get_irn_out_n(adr) + 1); + + ldst_info->flags |= LDST_VISITED; + } } } - else if (get_irn_op(node) == op_Block) { /* check, if it's an exception block */ + else if (op == op_Block) { /* check, if it's an exception block */ int i, n; for (i = 0, n = get_Block_n_cfgpreds(node); i < n; ++i) { @@ -184,6 +238,107 @@ static void collect_nodes(ir_node *node, void *env) } } +/** + * Returns an entity if the address ptr points to a constant one. + */ +static entity *find_constant_entity(ir_node *ptr) +{ + for (;;) { + ir_op *op = get_irn_op(ptr); + + if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) { + return get_SymConst_entity(ptr); + } + else if (op == op_Sel) { + entity *ent = get_Sel_entity(ptr); + type *tp = get_entity_owner(ent); + + /* Do not fiddle about polymorphy. */ + if (is_Class_type(get_entity_owner(ent)) && + ((get_entity_n_overwrites(ent) != 0) || + (get_entity_n_overwrittenby(ent) != 0) ) ) + return NULL; + + if (variability_constant == get_entity_variability(ent)) + return ent; + + if (is_Array_type(tp)) { + /* check bounds */ + int i, n; + + for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) { + ir_node *bound; + tarval *tlower, *tupper; + ir_node *index = get_Sel_index(ptr, i); + tarval *tv = computed_value(index); + + /* check if the index is constant */ + if (tv == tarval_bad) + return NULL; + + bound = get_array_lower_bound(tp, i); + tlower = computed_value(bound); + bound = get_array_upper_bound(tp, i); + tupper = computed_value(bound); + + if (tlower == tarval_bad || tupper == tarval_bad) + return NULL; + + if (tarval_cmp(tv, tlower) & pn_Cmp_Lt) + return NULL; + if (tarval_cmp(tupper, tv) & pn_Cmp_Lt) + return NULL; + + /* ok, bounds check finished */ + } + } + + /* try next */ + ptr = get_Sel_ptr(ptr); + } + else + return NULL; + } +} + +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); + return get_tarval_long(get_Const_tarval(index)); +} + +static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) { + compound_graph_path *res = NULL; + entity *root, *field; + int path_len, pos; + + if (get_irn_op(ptr) == op_SymConst) { + assert(get_SymConst_kind(ptr) == symconst_addr_ent); + root = get_SymConst_entity(ptr); + res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth); + } + else { + assert(get_irn_op(ptr) == op_Sel); + res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1); + field = get_Sel_entity(ptr); + path_len = get_compound_graph_path_length(res); + pos = path_len - depth - 1; + set_compound_graph_path_node(res, pos, field); + if (is_Array_type(get_entity_owner(field))) { + assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented"); + set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0)); + } + } + return res; +} + + +/** Returns an access path or NULL. The access path is only + * valid, if the graph is in phase_high and _no_ address computation is used. */ +static compound_graph_path *get_accessed_path(ir_node *ptr) { + return rec_get_accessed_path(ptr, 0); +} + /** * optimize a Load */ @@ -191,9 +346,52 @@ static int 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; + ir_node *pred, *mem, *ptr, *new_node; + entity *ent; int res = 0; + /* the address of the load to be optimized */ + ptr = get_Load_ptr(load); + + /* + * Check if we can remove the exception form a Load: + * this can be done, if the address is from an Sel(Alloc) and + * the Sel type is a subtype of the alloc type. + * + * This optimizes some often used OO constructs, + * like x = new O; x->t; + */ + if (info->projs[pn_Load_X_except]) { + if (get_irn_op(ptr) == op_Sel) { + 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); + + 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 */ + + exchange(info->projs[pn_Load_X_except], new_Bad()); + info->projs[pn_Load_X_except] = NULL; + } + } + } + else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) || + ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) { + /* simple case: a direct load after an Alloc. Firm Alloc throw + * an exception in case of out-of-memory. So, there is no way for an + * exception in this load. + * This code is constructed by the "exception lowering" in the Jack compiler. + */ + exchange(info->projs[pn_Load_X_except], new_Bad()); + info->projs[pn_Load_X_except] = NULL; + } + } + /* do NOT touch volatile loads for now */ if (get_Load_volatility(load) == volatility_is_volatile) return 0; @@ -206,17 +404,29 @@ static int optimize_load(ir_node *load) return 1; } - /* the address of the load to be optimized */ - ptr = get_Load_ptr(load); - - /* the mem of the load. Must still be returned after optimization */ + /* the mem of the Load. Must still be returned after optimization */ mem = get_Load_mem(load); - if ((get_irn_opcode(ptr) == iro_SymConst) && (get_SymConst_kind(ptr) == symconst_addr_ent)) { - entity *ent = get_SymConst_entity(ptr); - - assert(mode_is_reference(get_irn_mode(ptr))); + /* Load from a constant polymorphic field, where we can resolve + polymorphy. */ + new_node = transform_node_Load(load); + if (new_node != load) { + if (info->projs[pn_Load_M]) { + exchange(info->projs[pn_Load_M], mem); + info->projs[pn_Load_M] = NULL; + } + if (info->projs[pn_Load_X_except]) { + exchange(info->projs[pn_Load_X_except], new_Bad()); + info->projs[pn_Load_X_except] = NULL; + } + if (info->projs[pn_Load_res]) + exchange(info->projs[pn_Load_res], new_node); + return 1; + } + /* check if we can determine the entity that will be loaded */ + ent = find_constant_entity(ptr); + if (ent) { 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 @@ -228,7 +438,12 @@ static int optimize_load(ir_node *load) info->projs[pn_Load_X_except] = NULL; } - if (variability_constant == get_entity_variability(ent)) { + 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 */ @@ -238,10 +453,63 @@ static int optimize_load(ir_node *load) exchange(info->projs[pn_Load_M], mem); /* no result :-) */ - if (info->projs[pn_Load_res]) - exchange(info->projs[pn_Load_res], copy_const_value(get_atomic_ent_value(ent))); + if (info->projs[pn_Load_res]) { + if (is_atomic_entity(ent)) { + ir_node *c = copy_const_value(get_atomic_ent_value(ent)); - return 1; + DBG_OPT_RC(load, c); + exchange(info->projs[pn_Load_res], c); + + return 1; + } + } + } + else if (variability_constant == get_entity_variability(ent)) { + compound_graph_path *path; + /* + 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); + */ + path = get_accessed_path(ptr); + if (path) { + ir_node *c; + + assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1)); + c = get_compound_ent_value_by_path(ent, path); + + /* printf(" cons: "); DDMN(c); */ + + if (info->projs[pn_Load_M]) + exchange(info->projs[pn_Load_M], mem); + if (info->projs[pn_Load_res]) + exchange(info->projs[pn_Load_res], copy_const_value(c)); + /* + { + int j; + for (j = 0; j < get_compound_graph_path_length(path); ++j) { + 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"); + } + */ + } 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_irn_op(ptr) == op_SymConst && + get_SymConst_kind(ptr) == symconst_addr_ent) { printf(" "); DDMEO(get_SymConst_entity(ptr)); } + printf("cannot optimize.\n"); + */ + } } /* we changed the irg, but try further */ @@ -249,6 +517,11 @@ static int optimize_load(ir_node *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) + return 0; + /* follow the memory chain as long as there are only Loads */ for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) { @@ -264,19 +537,19 @@ static int optimize_load(ir_node *load) 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 + * 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 Exc? 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 teh Store jumps directly to the end... + * 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)) { - DBG_OPT_RAW(pred, load); + DBG_OPT_RAW(load, pred); exchange( info->projs[pn_Load_res], get_Store_value(pred) ); if (info->projs[pn_Load_M]) @@ -291,18 +564,18 @@ static int optimize_load(ir_node *load) 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. + * 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 previos load has an exception hander because + * Here, there is no need to check if the previos 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(pred, load); + DBG_OPT_RAR(load, pred); if (pred_info->projs[pn_Load_res]) { /* we need a data proj from the previous load for this optimization */ @@ -330,7 +603,7 @@ static int optimize_load(ir_node *load) } } - /* follow only load chains */ + /* follow only Load chains */ if (get_irn_op(pred) != op_Load) break; } @@ -357,11 +630,16 @@ static int optimize_store(ir_node *store) * is possible *(type1 *)p = a; *(type2 *)p = b ... */ - block = get_nodes_block(store); 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 */ @@ -371,7 +649,7 @@ static int optimize_store(ir_node *store) 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) { /* - * a store after a store in the same block -- a write after write. + * a Store after a Store in the same block -- a write after write. * We may remove the first Store, if it does not have an exception handler. * * TODO: What, if both have the same exception handler ??? @@ -385,17 +663,17 @@ static int optimize_store(ir_node *store) else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr && value == pred_info->projs[pn_Load_res]) { /* - * a store of a value after a load -- a write after read. + * 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. */ if (! info->projs[pn_Store_X_except]) { - DBG_OPT_WAR(pred, store); + DBG_OPT_WAR(store, pred); exchange( info->projs[pn_Store_M], mem ); return 1; } } - /* follow only load chains */ + /* follow only Load chains */ if (get_irn_op(pred) != op_Load) break; } @@ -423,7 +701,7 @@ static int optimize_phi(ir_node *phi, void *env) { walk_env_t *wenv = env; int i, n; - ir_node *store, *ptr, *block, *phiM, *phiD, *exc, *projM; + ir_node *store, *old_store, *ptr, *block, *phiM, *phiD, *exc, *projM; ir_mode *mode; ir_node **inM, **inD; int *idx; @@ -440,6 +718,7 @@ static int optimize_phi(ir_node *phi, void *env) return 0; store = skip_Proj(get_Phi_pred(phi, 0)); + old_store = store; if (get_irn_op(store) != op_Store) return 0; @@ -523,6 +802,10 @@ static int optimize_phi(ir_node *phi, void *env) /* fourth step: create the Store */ store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD); +#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); info = get_ldst_info(store, wenv);