X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fldstopt.c;h=5aa9947dbc147cb0a3384daea240fb06df88158a;hb=2eef4468e02e351bae78e54edf578fd449b0c7c6;hp=c8b58f58e06127c61e593657257367548bcc03ce;hpb=a227f30355070599cfd5266cf66d1b35f845b01e;p=libfirm diff --git a/ir/opt/ldstopt.c b/ir/opt/ldstopt.c index c8b58f58e..5aa9947db 100644 --- a/ir/opt/ldstopt.c +++ b/ir/opt/ldstopt.c @@ -36,6 +36,7 @@ # include "irflag_t.h" # include "array.h" # include "irhooks.h" +# include "irtools.h" # include "opt_polymorphy.h" #ifdef DO_CACHEOPT @@ -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; + unsigned changes; /**< a bitmask of graph changes */ } walk_env_t; /** @@ -70,6 +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; /** @@ -77,7 +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 */ }; /** @@ -87,13 +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 @@ -132,7 +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); @@ -141,7 +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; @@ -150,9 +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"); @@ -161,18 +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; @@ -194,6 +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); @@ -206,12 +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; @@ -219,15 +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); @@ -250,10 +286,10 @@ 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) ) ) @@ -361,13 +397,13 @@ static compound_graph_path *get_accessed_path(ir_node *ptr) { /** * 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; entity *ent; - int res = 0; + unsigned res = 0; /* do NOT touch volatile loads for now */ if (get_Load_volatility(load) == volatility_is_volatile) @@ -390,17 +426,17 @@ static int optimize_load(ir_node *load) 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 */ exchange(info->projs[pn_Load_X_except], new_Bad()); info->projs[pn_Load_X_except] = NULL; - res = 1; + res |= CF_CHANGED; } } } @@ -413,7 +449,7 @@ static int optimize_load(ir_node *load) */ exchange(info->projs[pn_Load_X_except], new_Bad()); info->projs[pn_Load_X_except] = NULL; - res = 1; + res |= CF_CHANGED; } } @@ -424,11 +460,11 @@ static int optimize_load(ir_node *load) /* a Load which value is neither used nor exception checked, remove it */ exchange(info->projs[pn_Load_M], mem); - return 1; + return res | DF_CHANGED; } /* 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]) { @@ -441,7 +477,7 @@ static int optimize_load(ir_node *load) } if (info->projs[pn_Load_res]) exchange(info->projs[pn_Load_res], new_node); - return 1; + return res | DF_CHANGED; } /* check if we can determine the entity that will be loaded */ @@ -456,7 +492,7 @@ 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 = 1; + res |= CF_CHANGED; } if (variability_constant == get_entity_variability(ent) @@ -472,17 +508,17 @@ static int optimize_load(ir_node *load) /* no memory */ if (info->projs[pn_Load_M]) { exchange(info->projs[pn_Load_M], mem); - res = 1; + 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; + return DF_CHANGED | res; } } } @@ -513,11 +549,11 @@ static int optimize_load(ir_node *load) if (info->projs[pn_Load_M]) { exchange(info->projs[pn_Load_M], mem); - res = 1; + res |= DF_CHANGED; } if (info->projs[pn_Load_res]) { - exchange(info->projs[pn_Load_res], copy_const_value(c)); - return 1; + exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c)); + return res | DF_CHANGED; } } else { @@ -539,21 +575,26 @@ static int optimize_load(ir_node *load) if (get_irn_out_n(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 + /* + * 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. */ - for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) { + 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 in C were the following - * is possible a = *(type1 *)p; b = *(type2 *)p ... + * 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) { - 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 @@ -575,13 +616,15 @@ static int optimize_load(ir_node *load) 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()); + res |= CF_CHANGED; + } if (info->projs[pn_Load_res]) exchange(info->projs[pn_Load_res], value); - return 1; + return res | DF_CHANGED; } } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr && @@ -596,8 +639,6 @@ static int optimize_load(ir_node *load) * 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]) { @@ -622,29 +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; @@ -653,7 +701,7 @@ static int optimize_store(ir_node *store) * 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 ... + * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ... */ ptr = get_Store_ptr(store); @@ -669,7 +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 && @@ -683,7 +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 && @@ -695,35 +744,42 @@ 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; @@ -734,6 +790,7 @@ 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) @@ -748,11 +805,11 @@ 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; @@ -769,7 +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); @@ -778,11 +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; @@ -790,17 +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. */ @@ -852,16 +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 optimizations */ static void do_load_store_optimize(ir_node *n, void *env) { @@ -896,14 +955,15 @@ void optimize_load_store(ir_graph *irg) 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); @@ -914,9 +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 (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); } }