X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fldstopt.c;h=6cf4e32a5ec20a068ee37e1ef60995c5b3e9daf2;hb=4ec6eeb8cbb20b6dcb511509e3cb947f4357cec6;hp=e49010d842ec6fbba7366ba15fdd097070348204;hpb=75d2e78bc1fe1f465ea8437e6b40275ad6f022b5;p=libfirm diff --git a/ir/opt/ldstopt.c b/ir/opt/ldstopt.c index e49010d84..6cf4e32a5 100644 --- a/ir/opt/ldstopt.c +++ b/ir/opt/ldstopt.c @@ -2,7 +2,7 @@ * Project: libFIRM * File name: ir/opt/ldstopt.c * Purpose: load store optimizations - * Author: + * Author: Michael Beck * Created: * CVS-ID: $Id$ * Copyright: (c) 1998-2004 Universität Karlsruhe @@ -22,21 +22,22 @@ # include #endif -# include "irnode_t.h" -# include "irgraph_t.h" -# include "irmode_t.h" -# include "iropt_t.h" -# include "ircons_t.h" -# include "irgmod.h" -# include "irgwalk.h" -# include "irvrfy.h" -# include "tv_t.h" -# include "dbginfo_t.h" -# include "iropt_dbg.h" -# include "irflag_t.h" -# include "array.h" -# include "irhooks.h" -# include "opt_polymorphy.h" +#include "irnode_t.h" +#include "irgraph_t.h" +#include "irmode_t.h" +#include "iropt_t.h" +#include "ircons_t.h" +#include "irgmod.h" +#include "irgwalk.h" +#include "irvrfy.h" +#include "tv_t.h" +#include "dbginfo_t.h" +#include "iropt_dbg.h" +#include "irflag_t.h" +#include "array.h" +#include "irhooks.h" +#include "irtools.h" +#include "opt_polymorphy.h" #ifdef DO_CACHEOPT #include "cacheopt/cachesim.h" @@ -47,12 +48,17 @@ #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; + struct obstack obst; /**< list of all stores */ + unsigned changes; /**< a bitmask of graph changes */ } walk_env_t; /** @@ -62,44 +68,41 @@ enum ldst_flags_t { LDST_VISITED = 1 /**< if set, this Load/Store is already visited */ }; -/** - * a Load/Store info - */ +/** 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 */ unsigned flags; /**< flags */ + unsigned visited; /**< visited counter for breaking loops */ } ldst_info_t; /** - * flags for control flow + * flags for control flow. */ 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 */ }; /** - * a Block info + * a Block info. */ 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 */ -static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) -{ +static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) { ldst_info_t *info = get_irn_link(node); if (! info) { @@ -132,7 +135,7 @@ 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); @@ -141,7 +144,7 @@ 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; @@ -150,9 +153,13 @@ 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"); @@ -161,18 +168,20 @@ 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)) +/** Return the number of uses of an address node */ +#define get_irn_n_uses(adr) (unsigned)PTR_TO_INT(get_irn_link(adr)) +/** Sets the number of uses of an address node */ +#define set_irn_n_uses(adr, 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; @@ -190,10 +199,23 @@ static void collect_nodes(ir_node *node, void *env) if ((ldst_info->flags & LDST_VISITED) == 0) { adr = get_Load_ptr(pred); - set_irn_out_n(adr, get_irn_out_n(adr) + 1); + set_irn_n_uses(adr, get_irn_n_uses(adr) + 1); 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); @@ -202,16 +224,29 @@ static void collect_nodes(ir_node *node, void *env) if ((ldst_info->flags & LDST_VISITED) == 0) { adr = get_Store_ptr(pred); - set_irn_out_n(adr, get_irn_out_n(adr) + 1); + set_irn_n_uses(adr, get_irn_n_uses(adr) + 1); 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; @@ -219,15 +254,15 @@ 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); @@ -250,17 +285,17 @@ static entity *find_constant_entity(ir_node *ptr) return get_SymConst_entity(ptr); } else if (op == op_Sel) { - entity *ent = get_Sel_entity(ptr); - type *tp = get_entity_owner(ent); + entity *ent = get_Sel_entity(ptr); + ir_type *tp = get_entity_owner(ent); - /* Do not fiddle about polymorphy. */ + /* Do not fiddle with polymorphism. */ if (is_Class_type(get_entity_owner(ent)) && - ((get_entity_n_overwrites(ent) != 0) || - (get_entity_n_overwrittenby(ent) != 0) ) ) - return NULL; + ((get_entity_n_overwrites(ent) != 0) || + (get_entity_n_overwrittenby(ent) != 0) ) ) + return NULL; if (variability_constant == get_entity_variability(ent)) - return ent; + return ent; if (is_Array_type(tp)) { /* check bounds */ @@ -301,29 +336,48 @@ 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); - field = get_Sel_entity(ptr); + + /* 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; + 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)); @@ -332,51 +386,238 @@ static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) { 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. */ + * 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); } +/* forward */ +static void reduce_adr_usage(ir_node *ptr); + +/** + * Update a Load that may lost it's usage. + */ +static void handle_load_update(ir_node *load) { + ldst_info_t *info = get_irn_link(load); + + /* do NOT touch volatile loads for now */ + if (get_Load_volatility(load) == volatility_is_volatile) + return; + + if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) { + 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 */ + exchange(info->projs[pn_Load_M], mem); + reduce_adr_usage(ptr); + } +} + +/** + * A Use of an address node is vanished. Check if this was a Proj + * node and update the counters. + */ +static void reduce_adr_usage(ir_node *ptr) { + int use_count = get_irn_n_uses(ptr); + --use_count; + assert(use_count >= 0); + set_irn_n_uses(ptr, use_count); + + if (is_Proj(ptr)) { + if (use_count <= 0) { + /* this Proj is now dead, update the Load/Store info */ + ir_node *pred = get_Proj_pred(ptr); + opcode code = get_irn_opcode(pred); + + if (code == iro_Load) { + ldst_info_t *info = get_irn_link(pred); + info->projs[get_Proj_proj(ptr)] = NULL; + + /* this node lost it's result proj, handle that */ + handle_load_update(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() must be called before dive into + */ +static unsigned follow_Load_chain(ir_node *load, ir_node *curr) { + unsigned res = 0; + ldst_info_t *info = get_irn_link(load); + ir_node *pred; + ir_node *ptr = get_Load_ptr(load); + ir_node *mem = get_Load_mem(load); + ir_mode *load_mode = get_Load_mode(load); + + for (pred = curr; 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 in C were the following + * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ... + */ + + if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr && + get_irn_mode(get_Store_value(pred)) == load_mode) { + /* + * 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 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)) { + 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]) { + exchange( info->projs[pn_Load_X_except], new_Bad()); + res |= CF_CHANGED; + } + + if (info->projs[pn_Load_res]) + exchange(info->projs[pn_Load_res], value); + + reduce_adr_usage(ptr); + return res | DF_CHANGED; + } + } + 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. + * 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 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)) { + DBG_OPT_RAR(load, pred); + + if (pred_info->projs[pn_Load_res]) { + /* we need a data proj from the previous load for this optimization */ + 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); + } + else { + 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 + properly!!! */ + exchange(info->projs[pn_Load_M], mem); + } + } + + /* no exception */ + if (info->projs[pn_Load_X_except]) { + exchange(info->projs[pn_Load_X_except], new_Bad()); + res |= CF_CHANGED; + } + + reduce_adr_usage(ptr); + 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); + } + + if (get_irn_op(pred) == op_Sync) { + int i; + + /* handle all Sync predecessors */ + for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) { + res |= follow_Load_chain(load, skip_Proj(get_Sync_pred(pred, i))); + if (res) + break; + } + } + + return res; +} + /** * optimize a Load */ -static int optimize_load(ir_node *load) +static unsigned optimize_load(ir_node *load) { ldst_info_t *info = get_irn_link(load); ir_mode *load_mode = get_Load_mode(load); - ir_node *pred, *mem, *ptr, *new_node; + ir_node *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; */ if (info->projs[pn_Load_X_except]) { - if (get_irn_op(ptr) == op_Sel) { + if (is_Sel(ptr)) { 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); + entity *ent = get_Sel_entity(ptr); + ir_type *s_type = get_entity_type(ent); + ir_type *a_type = get_Alloc_type(mem); - if (is_subclass_of(s_type, a_type)) { + 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; } } } @@ -389,26 +630,23 @@ 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; + reduce_adr_usage(ptr); + 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. */ + polymorphism. */ new_node = transform_node_Load(load); if (new_node != load) { if (info->projs[pn_Load_M]) { @@ -421,7 +659,9 @@ static int optimize_load(ir_node *load) } if (info->projs[pn_Load_res]) exchange(info->projs[pn_Load_res], new_node); - return 1; + + reduce_adr_usage(ptr); + return res | DF_CHANGED; } /* check if we can determine the entity that will be loaded */ @@ -436,6 +676,7 @@ 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) @@ -449,42 +690,30 @@ static int optimize_load(ir_node *load) */ /* 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]) { if (is_atomic_entity(ent)) { - ir_node *c = copy_const_value(get_atomic_ent_value(ent)); + ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent)); DBG_OPT_RC(load, c); exchange(info->projs[pn_Load_res], c); - - return 1; + res |= DF_CHANGED; } } + reduce_adr_usage(ptr); + return res; } 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); + 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)); - 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; @@ -496,158 +725,79 @@ static int optimize_load(ir_node *load) } 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 */ - res = 1; - } - } - - /* 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))) { - - /* - * 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 - * 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 :-( - * 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)) { - ir_node *value = get_Store_value(pred); + */ - DBG_OPT_RAW(load, value); - exchange(info->projs[pn_Load_res], value); - - if (info->projs[pn_Load_M]) - exchange(info->projs[pn_Load_M], mem); - - /* no exception */ - if (info->projs[pn_Load_X_except]) - exchange( info->projs[pn_Load_X_except], new_Bad()); - return 1; - } - } - 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. - * 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 - * 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); + c = get_compound_ent_value_by_path(ent, path); + free_compound_graph_path(path); - DBG_OPT_RAR(load, pred); + /* printf(" cons: "); DDMN(c); */ - 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_M]) - exchange(info->projs[pn_Load_M], mem); - } - else { - 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)); - } if (info->projs[pn_Load_M]) { - /* Actually, this if should not be necessary. Construct the Loads - properly!!! */ 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(get_irn_dbg_info(load), c)); + res |= DF_CHANGED; + } + reduce_adr_usage(ptr); + return res; + } + 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); + */ } - - /* no exception */ - if (info->projs[pn_Load_X_except]) - exchange(info->projs[pn_Load_X_except], new_Bad()); - - return 1; } } + } - /* follow only Load chains */ - if (get_irn_op(pred) != op_Load) - break; - } + /* 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_n_uses(ptr) <= 1) + return res; + + /* + * follow the memory chain as long as there are only Loads + * and try to replace current Load or Store by a previous one. + * 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(); + res = follow_Load_chain(load, skip_Proj(mem)); return res; } /** - * optimize a Store + * follow the memory chain as long as there are only Loads. + * + * INC_MASTER() must be called before dive into */ -static int optimize_store(ir_node *store) -{ +static unsigned follow_Load_chain_for_Store(ir_node *store, ir_node *curr) { + unsigned res = 0; ldst_info_t *info = get_irn_link(store); - ir_node *pred, *mem, *ptr, *value, *block; - ir_mode *mode; - int res = 0; - - if (get_Store_volatility(store) == volatility_is_volatile) - return 0; - - /* - * 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 - * is possible *(type1 *)p = a; *(type2 *)p = b ... - */ - - 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 */ - for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) { + ir_node *pred; + ir_node *ptr = get_Store_ptr(store); + ir_node *mem = get_Store_mem(store); + ir_node *value = get_Store_value(store); + ir_mode *mode = get_irn_mode(value); + ir_node *block = get_nodes_block(store); + + for (pred = curr; pred != store; 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 + * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ... + */ 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) { /* @@ -659,7 +809,8 @@ 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; + reduce_adr_usage(ptr); + return DF_CHANGED; } } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr && @@ -671,37 +822,80 @@ 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; + reduce_adr_usage(ptr); + 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); + } + + if (get_irn_op(pred) == op_Sync) { + int i; + + /* handle all Sync predecessors */ + for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) { + res |= follow_Load_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i))); + if (res) + break; + } } return res; } /** - * walker, optimizes Phi after Stores: + * optimize a Store + */ +static unsigned optimize_store(ir_node *store) +{ + ldst_info_t *info = get_irn_link(store); + ir_node *ptr, *mem; + + if (get_Store_volatility(store) == volatility_is_volatile) + return 0; + + 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_n_uses(ptr) <= 1) + return 0; + + mem = get_Store_mem(store); + + /* follow the memory chain as long as there are only Loads */ + INC_MASTER(); + return follow_Load_chain_for_Store(store, skip_Proj(mem)); +} + +/** + * 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, walk_env_t *wenv) { - walk_env_t *wenv = env; int i, n; ir_node *store, *old_store, *ptr, *block, *phiM, *phiD, *exc, *projM; ir_mode *mode; @@ -710,6 +904,7 @@ static int optimize_phi(ir_node *phi, void *env) 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) @@ -724,11 +919,11 @@ static int optimize_phi(ir_node *phi, void *env) 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; @@ -745,7 +940,7 @@ 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); @@ -754,11 +949,11 @@ 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; @@ -766,17 +961,17 @@ 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. */ @@ -793,6 +988,12 @@ static int optimize_phi(ir_node *phi, void *env) inM[i] = get_Store_mem(pred); inD[i] = get_Store_value(pred); idx[i] = info->exc_idx; + + /* Should we here replace the Proj after the Store by + * the Store's memory? Would be save but should not be needed, + * because we checked that all pred blocks have only one + * control flow successor. + */ } block = get_nodes_block(phi); @@ -807,6 +1008,8 @@ static int optimize_phi(ir_node *phi, void *env) #ifdef DO_CACHEOPT co_set_irn_name(store, co_get_irn_ident(old_store)); #endif + /* we replaced n uses by 1 */ + set_irn_n_uses(ptr, get_irn_n_uses(ptr) - n + 1); projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M); @@ -828,16 +1031,18 @@ 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 optimizations */ static void do_load_store_optimize(ir_node *n, void *env) { @@ -854,7 +1059,7 @@ static void do_load_store_optimize(ir_node *n, void *env) break; case iro_Phi: - wenv->changes |= optimize_phi(n, env); + wenv->changes |= optimize_phi(n, wenv); default: ; @@ -868,18 +1073,23 @@ void optimize_load_store(ir_graph *irg) { walk_env_t env; + /* must set current it graph here, because new nodes are + constructed in the fixpoint iteration */ + current_ir_graph = irg; + 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()) + if (! get_opt_redundant_loadstore()) return; obstack_init(&env.obst); 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); @@ -890,9 +1100,10 @@ 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 (get_irg_dom_state(current_ir_graph) == dom_consistent) - set_irg_dom_inconsistent(current_ir_graph); + if (env.changes & CF_CHANGED) { + /* is this really needed: Yes, control flow changed, block might get Bad. */ + set_irg_doms_inconsistent(current_ir_graph); } }