X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=4b1ebb72fea6d7183b08ae6f0b482d42d7257b35;hb=8974b94349462d835b33ad4f896d1821986458bd;hp=eaf93d036edc159510a929782e370ff48b2e366c;hpb=44d51ff21a4a48494fd8e42dd10f5b489b4fca78;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index eaf93d036..4b1ebb72f 100644 --- a/ir/be/bespillbelady.c +++ b/ir/be/bespillbelady.c @@ -68,13 +68,9 @@ #define DBG_WORKSET 128 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) -/* factor to weight the different costs of reloading/rematerializing a node - (see bespill.h be_get_reload_costs_no_weight) */ -#define RELOAD_COST_FACTOR 10 - #define TIME_UNDEFINED 6666 -#define PLACE_SPILLS +//#define LOOK_AT_LOOPDEPTH /** * An association between a node and a point in time. @@ -106,6 +102,21 @@ static ir_nodeset_t used; static spill_env_t *senv; /**< see bespill.h */ static pdeq *worklist; +static bool move_spills = true; +static bool respectloopdepth = true; +static bool improve_known_preds = true; +/* factor to weight the different costs of reloading/rematerializing a node + (see bespill.h be_get_reload_costs_no_weight) */ +static int remat_bonus = 10; + +static const lc_opt_table_entry_t options[] = { + LC_OPT_ENT_BOOL ("movespills", "try to move spills out of loops", &move_spills), + LC_OPT_ENT_BOOL ("respectloopdepth", "exprimental (outermost loop cutting)", &respectloopdepth), + LC_OPT_ENT_BOOL ("improveknownpreds", "experimental (known preds cutting)", &improve_known_preds), + LC_OPT_ENT_INT ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus), + LC_OPT_LAST +}; + static int loc_compare(const void *a, const void *b) { const loc_t *p = a; @@ -282,16 +293,19 @@ static INLINE unsigned get_distance(ir_node *from, unsigned from_step, assert(! (flags & arch_irn_flags_ignore)); use = be_get_next_use(uses, from, from_step, def, skip_from_uses); - if(USES_IS_INFINITE(use.time)) + if (USES_IS_INFINITE(use.time)) return USES_INFINITY; /* We have to keep nonspillable nodes in the workingset */ - if(flags & arch_irn_flags_dont_spill) + if (flags & arch_irn_flags_dont_spill) return 0; - costs = be_get_reload_costs_no_weight(senv, def, use.before); - assert(costs * RELOAD_COST_FACTOR < 1000); - time = use.time + 1000 - (costs * RELOAD_COST_FACTOR); + /* give some bonus to rematerialisable nodes */ + if (remat_bonus > 0) { + costs = be_get_reload_costs_no_weight(senv, def, use.before); + assert(costs * remat_bonus < 1000); + time = use.time + 1000 - (costs * remat_bonus); + } return time; } @@ -351,10 +365,13 @@ static void displace(workset_t *new_vals, int is_usage) /* Only make more free room if we do not have enough */ if (spills_needed > 0) { -#ifndef PLACE_SPILLS - ir_node *curr_bb = get_nodes_block(instr); - workset_t *ws_start = get_block_info(curr_bb)->start_workset; -#endif + ir_node *curr_bb = NULL; + workset_t *ws_start = NULL; + + if (move_spills) { + curr_bb = get_nodes_block(instr); + ws_start = get_block_info(curr_bb)->start_workset; + } DB((dbg, DBG_DECIDE, " disposing %d values\n", spills_needed)); @@ -374,22 +391,23 @@ static void displace(workset_t *new_vals, int is_usage) DB((dbg, DBG_DECIDE, " disposing node %+F (%u)\n", val, workset_get_time(ws, i))); -#ifdef PLACE_SPILLS - if(!USES_IS_INFINITE(ws->vals[i].time) && !ws->vals[i].spilled) { - ir_node *after_pos = sched_prev(instr); - be_add_spill(senv, val, after_pos); - } -#endif - -#ifndef PLACE_SPILLS - /* Logic for not needed live-ins: If a value is disposed - * before its first use, remove it from start workset - * We don't do this for phis though */ - if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) { - workset_remove(ws_start, val); - DB((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", val)); + if (move_spills) { + if (!USES_IS_INFINITE(ws->vals[i].time) + && !ws->vals[i].spilled) { + ir_node *after_pos = sched_prev(instr); + DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val, + after_pos)); + be_add_spill(senv, val, after_pos); + } + } else { + /* Logic for not needed live-ins: If a value is disposed + * before its first use, remove it from start workset + * We don't do this for phis though */ + if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) { + workset_remove(ws_start, val); + DB((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", val)); + } } -#endif } /* kill the last 'demand' entries in the array */ @@ -404,6 +422,64 @@ static void displace(workset_t *new_vals, int is_usage) } } +enum { + AVAILABLE_EVERYWHERE, + AVAILABLE_NOWHERE, + AVAILABLE_PARTLY, + AVAILABLE_UNKNOWN +}; + +static unsigned available_in_all_preds(workset_t* const* pred_worksets, + size_t n_pred_worksets, + const ir_node *value, bool is_local_phi) +{ + size_t i; + bool avail_everywhere = true; + bool avail_nowhere = true; + + assert(n_pred_worksets > 0); + + /* value available in all preds? */ + for (i = 0; i < n_pred_worksets; ++i) { + bool found = false; + const workset_t *p_workset = pred_worksets[i]; + int p_len = workset_get_length(p_workset); + int p_i; + const ir_node *l_value; + + if (is_local_phi) { + assert(is_Phi(value)); + l_value = get_irn_n(value, i); + } else { + l_value = value; + } + + for (p_i = 0; p_i < p_len; ++p_i) { + const loc_t *p_l = &p_workset->vals[p_i]; + if (p_l->node != l_value) + continue; + + found = true; + break; + } + + if (found) { + avail_nowhere = false; + } else { + avail_everywhere = false; + } + } + + if (avail_everywhere) { + assert(!avail_nowhere); + return AVAILABLE_EVERYWHERE; + } else if (avail_nowhere) { + return AVAILABLE_NOWHERE; + } else { + return AVAILABLE_PARTLY; + } +} + /** Decides whether a specific node should be in the start workset or not * * @param env belady environment @@ -412,7 +488,7 @@ static void displace(workset_t *new_vals, int is_usage) * @param loop the loop of the node */ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, - ir_loop *loop) + ir_loop *loop, unsigned available) { be_next_use_t next_use; loc_t loc; @@ -427,27 +503,37 @@ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, } /* We have to keep nonspillable nodes in the workingset */ - if(arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) { + if (arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) { loc.time = 0; DB((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time)); return loc; } next_use = be_get_next_use(uses, first, 0, node, 0); - if(USES_IS_INFINITE(next_use.time)) { + if (USES_IS_INFINITE(next_use.time)) { // the nodes marked as live in shouldn't be dead, so it must be a phi assert(is_Phi(node)); loc.time = USES_INFINITY; DB((dbg, DBG_START, " %+F not taken (dead)\n", node)); - if(is_Phi(node)) { - be_spill_phi(senv, node); - } return loc; } loc.time = next_use.time; - if(next_use.outermost_loop >= get_loop_depth(loop)) { + if (improve_known_preds) { + if (available == AVAILABLE_EVERYWHERE) { + DB((dbg, DBG_START, " %+F taken (%u, live in all preds)\n", + node, loc.time)); + return loc; + } else if(available == AVAILABLE_NOWHERE) { + DB((dbg, DBG_START, " %+F not taken (%u, live in no pred)\n", + node, loc.time)); + loc.time = USES_INFINITY; + return loc; + } + } + + if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) { DB((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop)); } else { @@ -455,6 +541,7 @@ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, DB((dbg, DBG_START, " %+F delayed (outerdepth %d < loopdepth %d)\n", node, next_use.outermost_loop, get_loop_depth(loop))); } + return loc; } @@ -477,6 +564,23 @@ static void decide_start_workset(const ir_node *block) unsigned pressure; int arity; workset_t **pred_worksets; + bool all_preds_known; + + /* check predecessors */ + arity = get_irn_arity(block); + pred_worksets = alloca(sizeof(pred_worksets[0]) * arity); + all_preds_known = true; + for(i = 0; i < arity; ++i) { + ir_node *pred_block = get_Block_cfgpred_block(block, i); + block_info_t *pred_info = get_block_info(pred_block); + + if (pred_info == NULL) { + pred_worksets[i] = NULL; + all_preds_known = false; + } else { + pred_worksets[i] = pred_info->end_workset; + } + } /* Collect all values living at start of block */ starters = NEW_ARR_F(loc_t, 0); @@ -487,24 +591,43 @@ static void decide_start_workset(const ir_node *block) /* check all Phis first */ sched_foreach(block, node) { + unsigned available; + if (! is_Phi(node)) break; + if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) + continue; - loc = to_take_or_not_to_take(first, node, loop); + if (all_preds_known) { + available = available_in_all_preds(pred_worksets, arity, node, true); + } else { + available = AVAILABLE_UNKNOWN; + } + + loc = to_take_or_not_to_take(first, node, loop, available); if (! USES_IS_INFINITE(loc.time)) { if (USES_IS_PENDING(loc.time)) ARR_APP1(loc_t, delayed, loc); else ARR_APP1(loc_t, starters, loc); + } else { + be_spill_phi(senv, node); } } /* check all Live-Ins */ be_lv_foreach(lv, block, be_lv_state_in, i) { ir_node *node = be_lv_get_irn(lv, block, i); + unsigned available; + + if (all_preds_known) { + available = available_in_all_preds(pred_worksets, arity, node, false); + } else { + available = AVAILABLE_UNKNOWN; + } - loc = to_take_or_not_to_take(first, node, loop); + loc = to_take_or_not_to_take(first, node, loop, available); if (! USES_IS_INFINITE(loc.time)) { if (USES_IS_PENDING(loc.time)) @@ -523,10 +646,12 @@ static void decide_start_workset(const ir_node *block) /* so far we only put nodes into the starters list that are used inside * the loop. If register pressure in the loop is low then we can take some * values and let them live through the loop */ + DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n", + pressure, free_slots)); if (free_slots > 0) { qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare); - for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) { + for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) { int p, arity; loc_t *loc = & delayed[i]; @@ -551,6 +676,7 @@ static void decide_start_workset(const ir_node *block) DB((dbg, DBG_START, " delayed %+F taken\n", loc->node)); ARR_APP1(loc_t, starters, *loc); loc->node = NULL; + --free_slots; skip_delayed: ; } @@ -560,7 +686,7 @@ static void decide_start_workset(const ir_node *block) * but not in the start workset */ for (i = ARR_LEN(delayed) - 1; i >= 0; --i) { ir_node *node = delayed[i].node; - if(node == NULL || !is_Phi(node) || get_nodes_block(node) != block) + if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block) continue; DB((dbg, DBG_START, " spilling delayed phi %+F\n", node)); @@ -593,18 +719,6 @@ static void decide_start_workset(const ir_node *block) /* determine spill status of the values: If there's 1 pred block (which * is no backedge) where the value is spilled then we must set it to * spilled here. */ - arity = get_irn_arity(block); - pred_worksets = alloca(sizeof(pred_worksets[0]) * arity); - for(i = 0; i < arity; ++i) { - ir_node *pred_block = get_Block_cfgpred_block(block, i); - block_info_t *pred_info = get_block_info(pred_block); - - if(pred_info == NULL) - pred_worksets[i] = NULL; - else - pred_worksets[i] = pred_info->end_workset; - } - for(i = 0; i < ws_count; ++i) { loc_t *loc = &ws->vals[i]; ir_node *value = loc->node; @@ -612,7 +726,7 @@ static void decide_start_workset(const ir_node *block) int n; /* phis from this block aren't spilled */ - if(get_nodes_block(value) == block) { + if (get_nodes_block(value) == block) { assert(is_Phi(value)); loc->spilled = false; continue; @@ -646,78 +760,6 @@ static void decide_start_workset(const ir_node *block) } } -#if 0 -static void decide_start_workset2(const ir_node *block) -{ - int arity; - workset_t **pred_worksets; - int p; - int len; - - /* check if all predecessors are known */ - arity = get_irn_arity(block); - pred_worksets = alloca(sizeof(pred_worksets[0]) * arity); - for (i = 0; i < arity; ++i) { - ir_node *pred_block = get_Block_cfgpred_block(block, i); - block_info_t *pred_info = get_block_info(pred_block); - - if (pred_info == NULL) { - /* not all predecessors known, use decide_start_workset */ - decide_start_workset(block); - return; - } - - pred_worksets[i] = pred_info->end_workset; - } - - /* take values live in all pred blocks */ - len = workset_get_length(pred_workset[0]); - for (p = 0; p < p_len; ++p) { - const loc_t *l = &pred_workset[0]->vals[p]; - ir_node *value; - bool spilled = false; - - if (USES_IS_INFINITE(l->time)) - continue; - - /* value available in all preds? */ - value = l->node; - for (i = 0; i < arity; ++i) { - bool found = false; - workset_t p_workset = &pred_worksets[i]; - int p_len = workset_get_length(p_workset); - int p_i; - - for (p_i = 0; p_i < p_len; ++p_i) { - const loc_t *p_l = &p_workset->vals[p_i]; - if (p_l->node != value) - continue; - - found = true; - if (p_l->spilled) - spilled = true; - break; - } - - if (!found) - break; - } - - /* it was available in all preds, TODO: insert spills... */ - if (i >= arity) { - workset_insert(ws, value, spilled); - } - } - - - /* Copy the best ones from starters to start workset */ - ws_count = MIN(ARR_LEN(starters), n_regs); - workset_clear(ws); - workset_bulk_fill(ws, ws_count, starters); -} - -#endif - /** * For the given block @p block, decide for each values * whether it is used from a register or is reloaded @@ -735,7 +777,7 @@ static void belady(ir_node *block) const ir_edge_t *edge; /* no need to process a block twice */ - if(get_block_info(block) != NULL) { + if (get_block_info(block) != NULL) { return; } @@ -746,18 +788,18 @@ static void belady(ir_node *block) ir_node *pred_block = get_Block_cfgpred_block(block, i); block_info_t *pred_info = get_block_info(pred_block); - if(pred_info == NULL) { + if (pred_info == NULL) { /* process predecessor first (it will be in the queue already) */ - if(!is_backedge(block, i)) { + if (!is_backedge(block, i)) { return; } has_backedges = 1; } } (void) has_backedges; - if(arity == 0) { + if (arity == 0) { workset_clear(ws); - } else if(arity == 1) { + } else if (arity == 1) { ir_node *pred_block = get_Block_cfgpred_block(block, 0); block_info_t *pred_info = get_block_info(pred_block); @@ -885,7 +927,7 @@ static void fix_block_borders(ir_node *block, void *data) int iter2; bool found = false; workset_foreach(start_workset, n2, iter2) { - if(n2 == node) { + if (n2 == node) { found = true; break; } @@ -898,8 +940,7 @@ static void fix_block_borders(ir_node *block, void *data) if (found) continue; -#ifdef PLACE_SPILLS - if(be_is_live_in(lv, block, node) + if (move_spills && be_is_live_in(lv, block, node) && !pred_end_workset->vals[iter].spilled) { ir_node *insert_point; if (arity > 1) { @@ -912,7 +953,6 @@ static void fix_block_borders(ir_node *block, void *data) insert_point)); be_add_spill(senv, node, insert_point); } -#endif } /* reload missing values in predecessors, add missing spills */ @@ -922,21 +962,20 @@ static void fix_block_borders(ir_node *block, void *data) /* if node is a phi of the current block we reload * the corresponding argument, else node itself */ - if(is_Phi(node) && get_nodes_block(node) == block) { + if (is_Phi(node) && get_nodes_block(node) == block) { node = get_irn_n(node, i); assert(!l->spilled); /* we might have unknowns as argument for the phi */ - if(!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) + if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) continue; } /* check if node is in a register at end of pred */ pred_loc = workset_contains(pred_end_workset, node); if (pred_loc != NULL) { -#ifdef PLACE_SPILLS /* we might have to spill value on this path */ - if (!pred_loc->spilled && l->spilled) { + if (move_spills && !pred_loc->spilled && l->spilled) { ir_node *insert_point = be_get_end_of_block_insertion_point(pred); insert_point = sched_prev(insert_point); @@ -944,7 +983,6 @@ static void fix_block_borders(ir_node *block, void *data) insert_point)); be_add_spill(senv, node, insert_point); } -#endif } else { /* node is not in register at the end of pred -> reload it */ DB((dbg, DBG_FIX, " reload %+F\n", node)); @@ -962,7 +1000,7 @@ static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) be_liveness_assure_sets(be_assure_liveness(birg)); /* construct control flow loop tree */ - if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) { + if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) { construct_cf_backedges(irg); } @@ -1010,6 +1048,9 @@ void be_init_spillbelady(void) static be_spiller_t belady_spiller = { be_spill_belady }; + lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady"); + lc_opt_add_table(belady_group, options); be_register_spiller("belady", &belady_spiller); FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");