X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fldstopt.c;h=5733efe7a42cdb36a44e721f3857c217b1716ad3;hb=d303af48c97698c818c67366ed02d79798a02518;hp=cd2ba7b347e30354810dbcf96bf81f8f50841b7c;hpb=5d2f7c6eca2173a427ccf35a4b8ac793706de2e9;p=libfirm diff --git a/ir/opt/ldstopt.c b/ir/opt/ldstopt.c index cd2ba7b34..5733efe7a 100644 --- a/ir/opt/ldstopt.c +++ b/ir/opt/ldstopt.c @@ -35,19 +35,30 @@ # include "iropt_dbg.h" # include "irflag_t.h" # include "array.h" -# include "firmstat.h" +# include "irhooks.h" +# include "irtools.h" +# include "opt_polymorphy.h" + +#ifdef DO_CACHEOPT +#include "cacheopt/cachesim.h" +#endif #undef IMAX #define IMAX(a,b) ((a) > (b) ? (a) : (b)) #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; /** @@ -65,6 +76,7 @@ typedef struct _ldst_info_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; /** @@ -72,7 +84,7 @@ typedef struct _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 */ }; /** @@ -82,13 +94,12 @@ typedef struct _block_info_t { 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 @@ -127,7 +138,7 @@ static block_info_t *get_block_info(ir_node *node, walk_env_t *env) /** * 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); @@ -136,7 +147,7 @@ static int update_projs(ldst_info_t *info, ir_node *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; @@ -145,9 +156,13 @@ static int update_projs(ldst_info_t *info, ir_node *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"); @@ -156,18 +171,18 @@ 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)) +#define get_irn_out_n(node) (unsigned)PTR_TO_INT(get_irn_link(node)) +#define set_irn_out_n(node, n) set_irn_link(adr, INT_TO_PTR(n)) /** * 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; @@ -189,6 +204,19 @@ static void collect_nodes(ir_node *node, void *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); @@ -201,12 +229,25 @@ static void collect_nodes(ir_node *node, void *env) 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; @@ -214,15 +255,15 @@ static void collect_nodes(ir_node *node, void *env) /* 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); @@ -234,7 +275,7 @@ static void collect_nodes(ir_node *node, void *env) } /** - * returns a entity if the address ptr points to a constant one. + * Returns an entity if the address ptr points to a constant one. */ static entity *find_constant_entity(ir_node *ptr) { @@ -248,7 +289,16 @@ static entity *find_constant_entity(ir_node *ptr) entity *ent = get_Sel_entity(ptr); type *tp = get_entity_owner(ent); - if (is_array_type(tp)) { + /* 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; @@ -270,9 +320,9 @@ static entity *find_constant_entity(ir_node *ptr) if (tlower == tarval_bad || tupper == tarval_bad) return NULL; - if (tarval_cmp(tv, tlower) & Lt) + if (tarval_cmp(tv, tlower) & pn_Cmp_Lt) return NULL; - if (tarval_cmp(tupper, tv) & Lt) + if (tarval_cmp(tupper, tv) & pn_Cmp_Lt) return NULL; /* ok, bounds check finished */ @@ -287,24 +337,85 @@ static entity *find_constant_entity(ir_node *ptr) } } +/** + * Return the Selection index of a Sel node from dimension n + */ +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)); +} + +/** + * Returns the accessed component graph path for an + * node computing an address. + * + * @param ptr the node computing the address + * @param depth current depth in steps upward from the root + * of the address + */ +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) { + /* 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. + */ + 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); + /* it's a Sel, go up until we find the root */ + res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1); + + /* fill up the step in the path at the current position */ + 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 */ -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; + 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) + return 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. + * Check if we can remove the exception from a Load: + * This can be done, if the address is from an Sel(Alloc) and + * the Sel type is a subtype of the allocated type. * * This optimizes some often used OO constructs, * like x = new O; x->t; @@ -313,7 +424,7 @@ static int optimize_load(ir_node *load) if (get_irn_op(ptr) == op_Sel) { ir_node *mem = get_Sel_mem(ptr); - if (get_irn_op(mem) == op_Alloc) { + 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); @@ -321,14 +432,16 @@ static int optimize_load(ir_node *load) 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 */ + * Alloc guarantees that enough memory was allocated */ exchange(info->projs[pn_Load_X_except], new_Bad()); info->projs[pn_Load_X_except] = NULL; + res |= CF_CHANGED; } } } - else if (get_irn_op(ptr) == op_Alloc) { + 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. @@ -336,23 +449,36 @@ static int optimize_load(ir_node *load) */ exchange(info->projs[pn_Load_X_except], new_Bad()); info->projs[pn_Load_X_except] = NULL; + res |= CF_CHANGED; } } - /* do NOT touch volatile loads for now */ - if (get_Load_volatility(load) == volatility_is_volatile) - return 0; + /* the mem of the Load. Must still be returned after optimization */ + 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 */ - mem = get_Load_mem(load); exchange(info->projs[pn_Load_M], mem); - return 1; + return res | DF_CHANGED; } - /* the mem of the Load. Must still be returned after optimization */ - mem = get_Load_mem(load); + /* 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 res | DF_CHANGED; + } /* check if we can determine the entity that will be loaded */ ent = find_constant_entity(ptr); @@ -366,20 +492,24 @@ static int optimize_load(ir_node *load) 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 (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. */ + && 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]) + if (info->projs[pn_Load_M]) { exchange(info->projs[pn_Load_M], mem); + res |= DF_CHANGED; + } /* no result :-) */ if (info->projs[pn_Load_res]) { @@ -388,63 +518,113 @@ static int optimize_load(ir_node *load) DBG_OPT_RC(load, c); exchange(info->projs[pn_Load_res], c); - - return 1; + return DF_CHANGED | res; } } } else if (variability_constant == get_entity_variability(ent)) { - printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent), - get_entity_name(get_irg_entity(current_ir_graph))); - } + 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) { + 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"); + } + */ - /* we changed the irg, but try further */ - res = 1; + 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(c)); + return res | DF_CHANGED; + } + } + 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); + */ + } + } } } /* 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; + return res; - /* follow the memory chain as long as there are only Loads */ - for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) { + /* + * 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. + */ + 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 is C were the following + * This is only true in strong typed languages, not in C were the following * is possible a = *(type1 *)p; b = *(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 * 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 :-( + * Why we need to check for Store Exception? 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 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(load, pred); - exchange( info->projs[pn_Load_res], get_Store_value(pred) ); + ir_node *value = get_Store_value(pred); + DBG_OPT_RAW(load, value); if (info->projs[pn_Load_M]) 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()); - return 1; + res |= CF_CHANGED; + } + + if (info->projs[pn_Load_res]) + exchange(info->projs[pn_Load_res], value); + + return res | DF_CHANGED; } } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr && @@ -455,17 +635,17 @@ static int optimize_load(ir_node *load) * 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 - * they would have exact the same exception... + * Here, there is no need to check if the previous 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(load, pred); if (pred_info->projs[pn_Load_res]) { /* we need a data proj from the previous load for this optimization */ - exchange( info->projs[pn_Load_res], pred_info->projs[pn_Load_res] ); + if (info->projs[pn_Load_res]) + exchange(info->projs[pn_Load_res], pred_info->projs[pn_Load_res]); + if (info->projs[pn_Load_M]) exchange(info->projs[pn_Load_M], mem); } @@ -473,6 +653,7 @@ static int optimize_load(ir_node *load) if (info->projs[pn_Load_res]) { set_Proj_pred(info->projs[pn_Load_res], pred); set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred)); + pred_info->projs[pn_Load_res] = info->projs[pn_Load_res]; } if (info->projs[pn_Load_M]) { /* Actually, this if should not be necessary. Construct the Loads @@ -482,29 +663,36 @@ static int optimize_load(ir_node *load) } /* 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; @@ -529,7 +717,8 @@ static int optimize_store(ir_node *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 && @@ -543,7 +732,7 @@ static int optimize_store(ir_node *store) 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 && @@ -555,45 +744,53 @@ static int optimize_store(ir_node *store) 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; - 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; 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) @@ -604,14 +801,15 @@ 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; - /* 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; @@ -628,7 +826,7 @@ static int optimize_phi(ir_node *phi, void *env) 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); @@ -637,11 +835,11 @@ static int optimize_phi(ir_node *phi, void *env) 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; @@ -649,17 +847,17 @@ static int optimize_phi(ir_node *phi, void *env) /* * 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. */ @@ -687,6 +885,10 @@ 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); @@ -707,16 +909,18 @@ static int optimize_phi(ir_node *phi, void *env) 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) { @@ -748,6 +952,8 @@ void optimize_load_store(ir_graph *irg) walk_env_t env; 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; @@ -756,7 +962,8 @@ void optimize_load_store(ir_graph *irg) 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); @@ -767,8 +974,10 @@ void optimize_load_store(ir_graph *irg) 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); }