X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fldstopt.c;h=e2977883622c2c6eb544640ef626434a0f4019fb;hb=4e0288d4c454f3cded36e0b81e97be03dca528db;hp=15964a359d7f71fadb5461172c654f51fe78c706;hpb=e88385016800d3c56c3fa09770e9f7995c42e106;p=libfirm diff --git a/ir/opt/ldstopt.c b/ir/opt/ldstopt.c index 15964a359..e29778836 100644 --- a/ir/opt/ldstopt.c +++ b/ir/opt/ldstopt.c @@ -1,27 +1,12 @@ /* - * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. - * * This file is part of libFirm. - * - * This file may be distributed and/or modified under the terms of the - * GNU General Public License version 2 as published by the Free Software - * Foundation and appearing in the file LICENSE.GPL included in the - * packaging of this file. - * - * Licensees holding valid libFirm Professional Edition licenses may use - * this file in accordance with the libFirm Commercial License. - * Agreement provided with the Software. - * - * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE - * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR - * PURPOSE. + * Copyright (C) 2012 University of Karlsruhe. */ /** * @file * @brief Load/Store optimizations. * @author Michael Beck - * @version $Id$ */ #include "config.h" @@ -35,6 +20,7 @@ #include "ircons_t.h" #include "irgmod.h" #include "irgwalk.h" +#include "irtools.h" #include "tv_t.h" #include "dbginfo_t.h" #include "iropt_dbg.h" @@ -43,20 +29,16 @@ #include "irhooks.h" #include "iredges.h" #include "irpass.h" -#include "opt_polymorphy.h" #include "irmemory.h" -#include "irphase_t.h" +#include "irnodehashmap.h" #include "irgopt.h" #include "set.h" +#include "be.h" #include "debug.h" /** The debug handle. */ DEBUG_ONLY(static firm_dbg_module_t *dbg;) -#ifdef DO_CACHEOPT -#include "cacheopt/cachesim.h" -#endif - #undef IMAX #define IMAX(a,b) ((a) > (b) ? (a) : (b)) @@ -77,7 +59,7 @@ typedef struct walk_env_t { /** A Load/Store info. */ typedef struct ldst_info_t { - ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */ + ir_node *projs[MAX_PROJ+1]; /**< 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 visited; /**< visited counter for breaking loops */ @@ -117,7 +99,7 @@ static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) set_irn_link(node, info); } return info; -} /* get_ldst_info */ +} /** * get the Block info of a node @@ -131,7 +113,7 @@ static block_info_t *get_block_info(ir_node *node, struct obstack *obst) set_irn_link(node, info); } return info; -} /* get_block_info */ +} /** * update the projection info for a Load/Store @@ -151,7 +133,7 @@ static unsigned update_projs(ldst_info_t *info, ir_node *proj) info->projs[nr] = proj; return 0; } -} /* update_projs */ +} /** * update the exception block info for a Load/Store node. @@ -167,10 +149,7 @@ static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos) info->exc_block = block; info->exc_idx = pos; return 0; -} /* update_exc */ - -/** Return the number of uses of an address node */ -#define get_irn_n_uses(adr) get_irn_n_edges(adr) +} /** * walker, collects all Load/Store/Proj nodes @@ -218,7 +197,7 @@ static void collect_nodes(ir_node *node, void *env) if (is_Proj(proj)) { pred = get_Proj_pred(proj); - is_exc = get_Proj_proj(proj) == pn_Generic_X_except; + is_exc = is_x_except_Proj(proj); } /* ignore Bad predecessors, they will be removed later */ @@ -241,7 +220,7 @@ static void collect_nodes(ir_node *node, void *env) } } } -} /* collect_nodes */ +} /** * Returns an entity if the address ptr points to a constant one. @@ -260,7 +239,7 @@ static ir_entity *find_constant_entity(ir_node *ptr) ir_type *tp = get_entity_owner(ent); /* Do not fiddle with polymorphism. */ - if (is_Class_type(get_entity_owner(ent)) && + if (is_Class_type(tp) && ((get_entity_n_overwrites(ent) != 0) || (get_entity_n_overwrittenby(ent) != 0) ) ) return NULL; @@ -287,9 +266,9 @@ static ir_entity *find_constant_entity(ir_node *ptr) if (tlower == tarval_bad || tupper == tarval_bad) return NULL; - if (tarval_cmp(tv, tlower) & pn_Cmp_Lt) + if (tarval_cmp(tv, tlower) == ir_relation_less) return NULL; - if (tarval_cmp(tupper, tv) & pn_Cmp_Lt) + if (tarval_cmp(tupper, tv) == ir_relation_less) return NULL; /* ok, bounds check finished */ @@ -329,7 +308,7 @@ static ir_entity *find_constant_entity(ir_node *ptr) } else return NULL; } -} /* find_constant_entity */ +} /** * Return the Selection index of a Sel node from dimension n @@ -337,178 +316,13 @@ static ir_entity *find_constant_entity(ir_node *ptr) static long get_Sel_array_index_long(ir_node *n, int dim) { ir_node *index = get_Sel_index(n, dim); - assert(is_Const(index)); return get_tarval_long(get_Const_tarval(index)); -} /* get_Sel_array_index_long */ - -/** - * 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; - ir_entity *root, *field, *ent; - int path_len, pos, idx; - ir_tarval *tv; - ir_type *tp; - - 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 its 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 if (is_Sel(ptr)) { - /* it's a Sel, go up until we find the root */ - res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1); - if (res == NULL) - return NULL; - - /* 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)); - } - } else if (is_Add(ptr)) { - ir_mode *mode; - ir_tarval *tmp; - - { - ir_node *l = get_Add_left(ptr); - ir_node *r = get_Add_right(ptr); - if (is_Const(r) && get_irn_mode(l) == get_irn_mode(ptr)) { - ptr = l; - tv = get_Const_tarval(r); - } else { - ptr = r; - tv = get_Const_tarval(l); - } - } -ptr_arith: - mode = get_tarval_mode(tv); - tmp = tv; - - /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */ - if (is_Sel(ptr)) { - field = get_Sel_entity(ptr); - } else { - field = get_SymConst_entity(ptr); - } - idx = 0; - for (ent = field;;) { - unsigned size; - ir_tarval *sz, *tv_index, *tlower, *tupper; - ir_node *bound; - - tp = get_entity_type(ent); - if (! is_Array_type(tp)) - break; - ent = get_array_element_entity(tp); - size = get_type_size_bytes(get_entity_type(ent)); - sz = new_tarval_from_long(size, mode); - - tv_index = tarval_div(tmp, sz); - tmp = tarval_mod(tmp, sz); - - if (tv_index == tarval_bad || tmp == tarval_bad) - return NULL; - - assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented"); - bound = get_array_lower_bound(tp, 0); - tlower = computed_value(bound); - bound = get_array_upper_bound(tp, 0); - tupper = computed_value(bound); - - if (tlower == tarval_bad || tupper == tarval_bad) - return NULL; - - if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt) - return NULL; - if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt) - return NULL; - - /* ok, bounds check finished */ - ++idx; - } - if (! tarval_is_null(tmp)) { - /* access to some struct/union member */ - return NULL; - } - - /* should be at least ONE array */ - if (idx == 0) - return NULL; - - res = rec_get_accessed_path(ptr, depth + idx); - if (res == NULL) - return NULL; - - path_len = get_compound_graph_path_length(res); - pos = path_len - depth - idx; - - for (ent = field;;) { - unsigned size; - ir_tarval *sz, *tv_index; - long index; - - tp = get_entity_type(ent); - if (! is_Array_type(tp)) - break; - ent = get_array_element_entity(tp); - set_compound_graph_path_node(res, pos, ent); - - size = get_type_size_bytes(get_entity_type(ent)); - sz = new_tarval_from_long(size, mode); - - tv_index = tarval_div(tv, sz); - tv = tarval_mod(tv, sz); - - /* worked above, should work again */ - assert(tv_index != tarval_bad && tv != tarval_bad); - - /* bounds already checked above */ - index = get_tarval_long(tv_index); - set_compound_graph_path_array_index(res, pos, index); - ++pos; - } - } else if (is_Sub(ptr)) { - ir_node *l = get_Sub_left(ptr); - ir_node *r = get_Sub_right(ptr); - - ptr = l; - tv = get_Const_tarval(r); - tv = tarval_neg(tv); - goto ptr_arith; - } - return res; -} /* rec_get_accessed_path */ - -/** - * 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) -{ - compound_graph_path *gr = rec_get_accessed_path(ptr, 0); - return gr; -} /* get_accessed_path */ +} typedef struct path_entry { ir_entity *ent; struct path_entry *next; - long index; + size_t index; } path_entry; static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) @@ -518,7 +332,7 @@ static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) ir_initializer_t *initializer; ir_tarval *tv; ir_type *tp; - unsigned n; + size_t n; entry.next = next; if (is_SymConst(ptr)) { @@ -541,7 +355,7 @@ static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) continue; } } - if (p->index >= (int) n) + if (p->index >= n) return NULL; initializer = get_initializer_compound_value(initializer, p->index); @@ -574,7 +388,7 @@ static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented"); entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0); } else { - int i, n_members = get_compound_n_members(tp); + size_t i, n_members = get_compound_n_members(tp); for (i = 0; i < n_members; ++i) { if (get_compound_member(tp, i) == field) break; @@ -660,9 +474,9 @@ ptr_arith: if (tlower == tarval_bad || tupper == tarval_bad) return NULL; - if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt) + if (tarval_cmp(tv_index, tlower) == ir_relation_less) return NULL; - if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt) + if (tarval_cmp(tupper, tv_index) == ir_relation_less) return NULL; /* ok, bounds check finished */ @@ -718,7 +532,7 @@ static void handle_load_update(ir_node *load) kill_node(load); reduce_adr_usage(ptr); } -} /* handle_load_update */ +} /** * A use of an address node has vanished. Check if this was a Proj @@ -741,7 +555,7 @@ static void reduce_adr_usage(ir_node *ptr) /* this node lost its result proj, handle that */ handle_load_update(pred); } -} /* reduce_adr_usage */ +} /** * Check, if an already existing value of mode old_mode can be converted @@ -749,20 +563,29 @@ static void reduce_adr_usage(ir_node *ptr) */ static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) { + unsigned old_size; + unsigned new_size; if (old_mode == new_mode) - return 1; + return true; + + old_size = get_mode_size_bits(old_mode); + new_size = get_mode_size_bits(new_mode); /* if both modes are two-complement ones, we can always convert the - Stored value into the needed one. */ - if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) && + Stored value into the needed one. (on big endian machines we currently + only support this for modes of same size) */ + if (old_size >= new_size && get_mode_arithmetic(old_mode) == irma_twos_complement && - get_mode_arithmetic(new_mode) == irma_twos_complement) - return 1; - return 0; -} /* can_use_stored_value */ + get_mode_arithmetic(new_mode) == irma_twos_complement && + (!be_get_backend_param()->byte_order_big_endian + || old_size == new_size)) { + return true; + } + return false; +} /** - * Check whether a Call is at least pure, ie. does only read memory. + * Check whether a Call is at least pure, i.e. does only read memory. */ static unsigned is_Call_pure(ir_node *call) { @@ -774,14 +597,14 @@ static unsigned is_Call_pure(ir_node *call) /* try the called entity */ ir_node *ptr = get_Call_ptr(call); - if (is_Global(ptr)) { - ir_entity *ent = get_Global_entity(ptr); + if (is_SymConst_addr_ent(ptr)) { + ir_entity *ent = get_SymConst_entity(ptr); prop = get_entity_additional_properties(ent); } } return (prop & (mtp_property_const|mtp_property_pure)) != 0; -} /* is_Call_pure */ +} static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset) { @@ -869,30 +692,28 @@ static int try_load_after_store(ir_node *load, delta = load_offset - store_offset; store_value = get_Store_value(store); - 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; + if (delta < 0 || delta+load_mode_len > store_mode_len) + return 0; + if (store_mode != load_mode && + get_mode_arithmetic(store_mode) == irma_twos_complement && + get_mode_arithmetic(load_mode) == irma_twos_complement) { /* produce a shift to adjust offset delta */ - if (delta > 0) { - ir_node *cnst; - ir_graph *irg = get_irn_irg(load); - - /* FIXME: only true for little endian */ - cnst = new_r_Const_long(irg, mode_Iu, delta * 8); + unsigned const shift = be_get_backend_param()->byte_order_big_endian + ? store_mode_len - load_mode_len - delta + : delta; + if (shift != 0) { + ir_graph *const irg = get_irn_irg(load); + ir_node *const cnst = new_r_Const_long(irg, mode_Iu, shift * 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(get_nodes_block(load), store_value, load_mode); - } + store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode); + } else { + /* we would need some kind of bitcast node here */ + return 0; } DBG_OPT_RAW(load, store_value); @@ -905,7 +726,7 @@ static int try_load_after_store(ir_node *load, /* no exception */ if (info->projs[pn_Load_X_except]) { ir_graph *irg = get_irn_irg(load); - exchange( info->projs[pn_Load_X_except], new_r_Bad(irg)); + exchange( info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); res |= CF_CHANGED; } if (info->projs[pn_Load_X_regular]) { @@ -979,6 +800,9 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) * Here, there is no need to check if the previous Load has an * exception hander because they would have exact the same * exception... + * + * TODO: implement load-after-load with different mode for big + * endian */ if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) { @@ -1008,7 +832,7 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) /* no exception */ if (info->projs[pn_Load_X_except]) { ir_graph *irg = get_irn_irg(load); - exchange(info->projs[pn_Load_X_except], new_r_Bad(irg)); + exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); res |= CF_CHANGED; } if (info->projs[pn_Load_X_regular]) { @@ -1066,12 +890,8 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) } return res; -} /* follow_Mem_chain */ +} -/* - * Check if we can replace the load by a given const from - * the const code irg. - */ ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) { ir_mode *c_mode = get_irn_mode(c); @@ -1085,8 +905,9 @@ ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) if (is_reinterpret_cast(c_mode, l_mode)) { /* copy the value from the const code irg and cast it */ res = new_rd_Conv(dbgi, block, res, l_mode); + } else { + return NULL; } - return NULL; } return res; } @@ -1129,50 +950,51 @@ static unsigned optimize_load(ir_node *load) return res | DF_CHANGED; } - /* Load from a constant polymorphic field, where we can resolve - polymorphism. */ - value = transform_polymorph_Load(load); - if (value == load) { - value = NULL; - /* check if we can determine the entity that will be loaded */ - ent = find_constant_entity(ptr); - if (ent != NULL - && get_entity_visibility(ent) != ir_visibility_external) { - /* 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]) { - ir_graph *irg = get_irn_irg(load); - exchange(info->projs[pn_Load_X_except], new_r_Bad(irg)); - 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 (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) { - if (ent->initializer != NULL) { - /* new style initializer */ - value = find_compound_ent_value(ptr); - } else if (entity_has_compound_ent_values(ent)) { - /* old style initializer */ - compound_graph_path *path = get_accessed_path(ptr); + value = NULL; + /* check if we can determine the entity that will be loaded */ + ent = find_constant_entity(ptr); + if (ent != NULL + && get_entity_visibility(ent) != ir_visibility_external) { + /* a static allocation that is not external: there should be NO + * exception when loading even if we cannot replace the load itself. + */ - if (path != NULL) { - assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1)); + /* no exception, clear the info field as it might be checked later again */ + if (info->projs[pn_Load_X_except]) { + ir_graph *irg = get_irn_irg(load); + exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); + 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; + } - 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 (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) { + if (has_entity_initializer(ent)) { + /* new style initializer */ + value = find_compound_ent_value(ptr); + } + if (value != NULL) { + ir_graph *irg = get_irn_irg(load); + value = can_replace_load_by_const(load, value); + if (value != NULL && is_Sel(ptr)) { + /* frontend has inserted masking operations after bitfield accesses, + * so we might have to shift the const. */ + unsigned char bit_offset = get_entity_offset_bits_remainder(get_Sel_entity(ptr)); + if (bit_offset != 0) { + if (is_Const(value)) { + ir_tarval *tv_old = get_Const_tarval(value); + ir_tarval *tv_offset = new_tarval_from_long(bit_offset, mode_Bu); + ir_tarval *tv_new = tarval_shl(tv_old, tv_offset); + value = new_r_Const(irg, tv_new); + } else { + value = NULL; + } } } - if (value != NULL) - value = can_replace_load_by_const(load, value); } } } @@ -1180,7 +1002,7 @@ static unsigned optimize_load(ir_node *load) /* we completely replace the load by this value */ if (info->projs[pn_Load_X_except]) { ir_graph *irg = get_irn_irg(load); - exchange(info->projs[pn_Load_X_except], new_r_Bad(irg)); + exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X)); info->projs[pn_Load_X_except] = NULL; res |= CF_CHANGED; } @@ -1204,7 +1026,7 @@ static unsigned optimize_load(ir_node *load) /* Check, if the address of this load is used more than once. * If not, more load cannot be removed in any case. */ - if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1) + if (get_irn_n_edges(ptr) <= 1 && get_irn_n_edges(get_base_and_offset(ptr, &dummy)) <= 1) return res; /* @@ -1217,7 +1039,7 @@ static unsigned optimize_load(ir_node *load) INC_MASTER(); res = follow_Mem_chain(load, skip_Proj(mem)); return res; -} /* optimize_load */ +} /** * Check whether a value of mode new_mode would completely overwrite a value @@ -1226,7 +1048,7 @@ static unsigned optimize_load(ir_node *load) static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode) { 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). @@ -1241,7 +1063,7 @@ static int is_partially_same(ir_node *small, ir_node *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. @@ -1374,7 +1196,7 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) } } return res; -} /* follow_Mem_chain_for_Store */ +} /** find entity used as base for an address calculation */ static ir_entity *find_entity(ir_node *ptr) @@ -1436,7 +1258,7 @@ static unsigned optimize_store(ir_node *store) /* Check, if the address of this Store is used more than once. * If not, this Store cannot be removed in any case. */ - if (get_irn_n_uses(ptr) <= 1) + if (get_irn_n_edges(ptr) <= 1) return 0; mem = get_Store_mem(store); @@ -1445,7 +1267,23 @@ static unsigned optimize_store(ir_node *store) INC_MASTER(); return follow_Mem_chain_for_Store(store, skip_Proj(mem)); -} /* optimize_store */ +} + +/* check if a node has more than one real user. Keepalive edges do not count as + * real users */ +static bool has_multiple_users(const ir_node *node) +{ + unsigned real_users = 0; + foreach_out_edge(node, edge) { + ir_node *user = get_edge_src_irn(edge); + if (is_End(user)) + continue; + ++real_users; + if (real_users > 1) + return true; + } + return false; +} /** * walker, optimizes Phi after Stores to identical places: @@ -1469,7 +1307,10 @@ static unsigned optimize_store(ir_node *store) static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) { int i, n; - ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM; + ir_node *store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM; +#ifdef DO_CACHEOPT + ir_node *old_store; +#endif ir_mode *mode; ir_node **inM, **inD, **projMs; int *idx; @@ -1488,20 +1329,18 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) /* must be only one user */ projM = get_Phi_pred(phi, 0); - if (get_irn_n_edges(projM) != 1) + if (has_multiple_users(projM)) return 0; store = skip_Proj(projM); +#ifdef DO_CACHEOPT old_store = store; +#endif if (!is_Store(store)) return 0; block = get_nodes_block(store); - /* abort on dead blocks */ - if (is_Block_dead(block)) - return 0; - /* check if the block is post dominated by Phi-block and has no exception exit */ bl_info = (block_info_t*)get_irn_link(block); @@ -1521,7 +1360,7 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) for (i = 1; i < n; ++i) { ir_node *pred = get_Phi_pred(phi, i); - if (get_irn_n_edges(pred) != 1) + if (has_multiple_users(pred)) return 0; pred = skip_Proj(pred); @@ -1537,10 +1376,7 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) if (exc != info->exc_block) return 0; - /* abort on dead blocks */ block = get_nodes_block(pred); - if (is_Block_dead(block)) - return 0; /* check if the block is post dominated by Phi-block and has no exception exit. Note that block must be different from @@ -1580,12 +1416,9 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) memory Proj. */ for (i = n - 1; i >= 0; --i) { - ir_node *store; - projMs[i] = get_Phi_pred(phi, i); - assert(is_Proj(projMs[i])); - store = get_Proj_pred(projMs[i]); + ir_node *const store = get_Proj_pred(projMs[i]); info = (ldst_info_t*)get_irn_link(store); inM[i] = get_Store_mem(store); @@ -1645,7 +1478,45 @@ static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv) exchange(phi, projM); return res | DF_CHANGED; -} /* optimize_phi */ +} + +static int optimize_conv_load(ir_node *conv) +{ + ir_node *op = get_Conv_op(conv); + if (!is_Proj(op)) + return 0; + if (has_multiple_users(op)) + return 0; + /* shrink mode of load if possible. */ + ir_node *load = get_Proj_pred(op); + if (!is_Load(load)) + return 0; + + /* only do it if we are the only user (otherwise the risk is too + * great that we end up with 2 loads instead of one). */ + ir_mode *mode = get_irn_mode(conv); + ir_mode *load_mode = get_Load_mode(load); + int bits_diff + = get_mode_size_bits(load_mode) - get_mode_size_bits(mode); + if (mode_is_float(load_mode) || mode_is_float(mode) || bits_diff < 0) + return 0; + + if (be_get_backend_param()->byte_order_big_endian) { + if (bits_diff % 8 != 0) + return 0; + ir_graph *irg = get_irn_irg(conv); + ir_node *ptr = get_Load_ptr(load); + ir_mode *mode = get_irn_mode(ptr); + ir_node *delta = new_r_Const_long(irg, mode, bits_diff/8); + ir_node *block = get_nodes_block(load); + ir_node *add = new_r_Add(block, ptr, delta, mode); + set_Load_ptr(load, add); + } + set_Load_mode(load, mode); + set_irn_mode(op, mode); + exchange(conv, op); + return DF_CHANGED; +} /** * walker, do the optimizations @@ -1668,10 +1539,14 @@ static void do_load_store_optimize(ir_node *n, void *env) wenv->changes |= optimize_phi(n, wenv); break; + case iro_Conv: + wenv->changes |= optimize_conv_load(n); + break; + default: - ; + break; } -} /* do_load_store_optimize */ +} /** A scc. */ typedef struct scc { @@ -1690,13 +1565,14 @@ typedef struct node_entry { /** A loop entry. */ typedef struct loop_env { - ir_phase ph; /**< the phase object */ - ir_node **stack; /**< the node stack */ - size_t tos; /**< tos index */ - unsigned nextDFSnum; /**< the current DFS number */ - unsigned POnum; /**< current post order number */ - - unsigned changes; /**< a bitmask of graph changes */ + ir_nodehashmap_t map; + struct obstack obst; + ir_node **stack; /**< the node stack */ + size_t tos; /**< tos index */ + unsigned nextDFSnum; /**< the current DFS number */ + unsigned POnum; /**< current post order number */ + + unsigned changes; /**< a bitmask of graph changes */ } loop_env; /** @@ -1704,16 +1580,15 @@ typedef struct loop_env { */ static node_entry *get_irn_ne(ir_node *irn, loop_env *env) { - ir_phase *ph = &env->ph; - node_entry *e = (node_entry*)phase_get_irn_data(&env->ph, irn); + node_entry *e = ir_nodehashmap_get(node_entry, &env->map, irn); - if (! e) { - e = (node_entry*)phase_alloc(ph, sizeof(*e)); + if (e == NULL) { + e = OALLOC(&env->obst, node_entry); memset(e, 0, sizeof(*e)); - phase_set_irn_data(ph, irn, e); + ir_nodehashmap_insert(&env->map, irn, e); } return e; -} /* get_irn_ne */ +} /** * Push a node onto the stack. @@ -1732,7 +1607,7 @@ static void push(loop_env *env, ir_node *n) env->stack[env->tos++] = n; e = get_irn_ne(n, env); e->in_stack = 1; -} /* push */ +} /** * pop a node from the stack @@ -1748,7 +1623,7 @@ static ir_node *pop(loop_env *env) e->in_stack = 0; return n; -} /* pop */ +} /** * Check if irn is a region constant. @@ -1762,7 +1637,7 @@ static int is_rc(ir_node *irn, ir_node *header_block) ir_node *block = get_nodes_block(irn); return (block != header_block) && block_dominates(block, header_block); -} /* is_rc */ +} typedef struct phi_entry phi_entry; struct phi_entry { @@ -1791,15 +1666,15 @@ static int cmp_avail_entry(const void *elt, const void *key, size_t size) (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 */ + return get_irn_idx(entry->ptr) * 9 + hash_ptr(entry->mode); +} /** * Move loops out of loops if possible. @@ -1810,13 +1685,10 @@ static unsigned hash_cache_entry(const avail_entry_t *entry) 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; 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) { node_entry *ne = get_irn_ne(phi, env); @@ -1834,7 +1706,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) if (pe->pscc != ne->pscc) { /* not in the same SCC, is region const */ - phi_entry *pe = (phi_entry*)phase_alloc(&env->ph, sizeof(*pe)); + phi_entry *pe = OALLOC(&env->obst, phi_entry); pe->phi = phi; pe->pos = j; @@ -1850,6 +1722,8 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) if (phi_list->next != NULL) return; + avail = new_set(cmp_avail_entry, 8); + for (load = pscc->head; load; load = next) { ir_mode *load_mode; node_entry *ne = get_irn_ne(load, env); @@ -1864,9 +1738,8 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) continue; /* for now, we can only move Load(Global) */ - if (! is_Global(ptr)) + if (! is_SymConst_addr_ent(ptr)) continue; - 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); @@ -1902,20 +1775,25 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) entry.ptr = ptr; entry.mode = load_mode; - res = (avail_entry_t*)set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry)); + res = set_find(avail_entry_t, 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, cons_none); entry.load = irn; - set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry)); + (void)set_insert(avail_entry_t, 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 = get_ldst_info(irn, &env->obst); ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M); - set_Phi_pred(phi, pos, mem); + if (res == NULL) { + /* irn is from cache, so do not set phi pred again. + * There might be other Loads between phi and irn already. + */ + set_Phi_pred(phi, pos, mem); + } ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res); } @@ -1929,7 +1807,7 @@ static void move_loads_out_of_loops(scc *pscc, loop_env *env) } } del_set(avail); -} /* move_loads_out_of_loops */ +} /** * Process a loop SCC. @@ -2055,7 +1933,7 @@ static void process_loop(scc *pscc, loop_env *env) fail: ; -} /* process_loop */ +} /** * Process a SCC. @@ -2088,7 +1966,7 @@ static void process_scc(scc *pscc, loop_env *env) /* this SCC has more than one member */ process_loop(pscc, env); } -} /* process_scc */ +} /** * Do Tarjan's SCC algorithm and drive load/store optimization. @@ -2122,7 +2000,7 @@ static void dfs(ir_node *irn, loop_env *env) node->low = MIN(o->DFSnum, node->low); } } else if (is_fragile_op(irn)) { - ir_node *pred = get_fragile_op_mem(irn); + ir_node *pred = get_memop_mem(irn); node_entry *o = get_irn_ne(pred, env); if (!irn_visited(pred)) { @@ -2147,7 +2025,7 @@ static void dfs(ir_node *irn, loop_env *env) } if (node->low == node->DFSnum) { - scc *pscc = (scc*)phase_alloc(&env->ph, sizeof(*pscc)); + scc *pscc = OALLOC(&env->obst, scc); ir_node *x; pscc->head = NULL; @@ -2163,7 +2041,7 @@ static void dfs(ir_node *irn, loop_env *env) process_scc(pscc, env); } -} /* dfs */ +} /** * Do the DFS on the memory edges a graph. @@ -2184,13 +2062,15 @@ static void do_dfs(ir_graph *irg, loop_env *env) ir_node *pred = get_Block_cfgpred(endblk, i); pred = skip_Proj(pred); - if (is_Return(pred)) + if (is_Return(pred)) { dfs(get_Return_mem(pred), env); - else if (is_Raise(pred)) + } 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 { + } else if (is_fragile_op(pred)) { + dfs(get_memop_mem(pred), env); + } else if (is_Bad(pred)) { + /* ignore non-optimized block predecessor */ + } else { assert(0 && "Unknown EndBlock predecessor"); } } @@ -2203,7 +2083,7 @@ static void do_dfs(ir_graph *irg, loop_env *env) if (is_Phi(ka) && !irn_visited(ka)) dfs(ka, env); } -} /* do_dfs */ +} /** * Optimize Loads/Stores in loops. @@ -2219,40 +2099,36 @@ static int optimize_loops(ir_graph *irg) env.nextDFSnum = 0; env.POnum = 0; env.changes = 0; - phase_init(&env.ph, irg, phase_irn_init_default); + ir_nodehashmap_init(&env.map); + obstack_init(&env.obst); /* calculate the SCC's and drive loop optimization. */ do_dfs(irg, &env); DEL_ARR_F(env.stack); - phase_deinit(&env.ph); + obstack_free(&env.obst, NULL); + ir_nodehashmap_destroy(&env.map); return env.changes; -} /* optimize_loops */ +} -/* - * do the load store optimization - */ -int optimize_load_store(ir_graph *irg) +void optimize_load_store(ir_graph *irg) { walk_env_t env; + assure_irg_properties(irg, + IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE + | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES + | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES + | IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE + | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE); + FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt"); - assert(get_irg_phase_state(irg) != phase_building); assert(get_irg_pinned(irg) != op_pin_state_floats && "LoadStore optimization needs pinned graph"); - /* we need landing pads */ - remove_critical_cf_edges(irg); - - edges_assure(irg); - - /* for Phi optimization post-dominators are needed ... */ - assure_postdoms(irg); - if (get_opt_alias_analysis()) { - assure_irg_entity_usage_computed(irg); assure_irp_globals_entity_usage_computed(); } @@ -2270,21 +2146,15 @@ int optimize_load_store(ir_graph *irg) obstack_free(&env.obst, NULL); - /* Handle graph state */ - if (env.changes) { - set_irg_outs_inconsistent(irg); - set_irg_entity_usage_state(irg, ir_entity_usage_not_computed); - } - - if (env.changes & CF_CHANGED) { - /* is this really needed: Yes, control flow changed, block might - have Bad() predecessors. */ - set_irg_doms_inconsistent(irg); - } - return env.changes != 0; -} /* optimize_load_store */ + confirm_irg_properties(irg, + env.changes + ? env.changes & CF_CHANGED + ? IR_GRAPH_PROPERTIES_NONE + : IR_GRAPH_PROPERTIES_CONTROL_FLOW + : IR_GRAPH_PROPERTIES_ALL); +} 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 */ + return def_graph_pass(name ? name : "ldst", optimize_load_store); +}