From bc6aa15e71b0fe97accfbe566c027049828ea513 Mon Sep 17 00:00:00 2001 From: Daniel Grund Date: Fri, 22 Jul 2005 11:30:03 +0000 Subject: [PATCH] Copy costs instead of numbers --- ir/be/becopyheur.c | 117 ++++++++++------------ ir/be/becopyilp.c | 17 ++-- ir/be/becopyopt.c | 242 ++++++++++++++++++++++----------------------- ir/be/becopyopt.h | 54 ++++++---- 4 files changed, 208 insertions(+), 222 deletions(-) diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c index 55092fea2..0745f2c33 100644 --- a/ir/be/becopyheur.c +++ b/ir/be/becopyheur.c @@ -26,6 +26,7 @@ #include "xmalloc.h" #include "becopyopt.h" #include "becopystat.h" +#include "bitset.h" #define DEBUG_LVL 0 //SET_LEVEL_1 static firm_dbg_module_t *dbg = NULL; @@ -63,7 +64,8 @@ typedef struct _qnode_t { const unit_t *ou; /**< the opt unit this qnode belongs to */ int color; /**< target color */ set *conflicts; /**< contains conflict_t's. All internal conflicts */ - int mis_size; /**< number of nodes in the mis. */ + int mis_costs; /**< costs of nodes/copies in the mis. */ + int mis_size; /**< size of the array below */ ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */ set *changed_nodes; /**< contains node_stat_t's. */ } qnode_t; @@ -365,71 +367,54 @@ static int qnode_try_color(const qnode_t *qn) { } /** - * Determines a maximum independent set with respect to the interference and - * conflict edges of all nodes in a qnode. + * Determines a maximum weighted independent set with respect to + * the interference and conflict edges of all nodes in a qnode. */ static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) { - int all_size, curr_size, i, o; - int *which; - ir_node **curr, **all = alloca(ou->node_count * sizeof(*all)); + ir_node **irns; + int max, next, pos, curr_weight, best_weight = 0; + bitset_t *best, *curr; + + irns = alloca((ou->node_count-1) * sizeof(*irns)); + best = bitset_alloca(ou->node_count-1); + curr = bitset_alloca(ou->node_count-1); + + /* brute force the best set */ + bitset_set_all(curr); + while ((max = bitset_popcnt(curr)) != 0) { + /* check if curr is a stable set */ + int i, o, is_stable_set = 1; + bitset_foreach(curr, pos) + irns[pos] = ou->nodes[1+pos]; + for(i=0; inode_count; ++i) - if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i])) - all[all_size++] = ou->nodes[i]; - - /* which[i] says which element to take out of all[] and put into curr[i] */ - which = alloca(all_size*sizeof(*which)); - for (curr_size=0; curr_sizemis_size = curr_size; - for (i=0; imis[i] = curr[i]; - return; + if (is_stable_set) { + /* calc current weigth */ + curr_weight = 0; + bitset_foreach(curr, pos) + curr_weight += ou->costs[1+pos]; -conflict_found: - /* We had a conflict. Generate next set */ - if (which[all_size-curr_size+1] == all_size-curr_size+1) { - curr_size--; - for (i=0; i=which[i+1]) { - redo = 1; - break; - } + /* any better ? */ + if (curr_weight > best_weight) { + best_weight = curr_weight; + bitset_copy(best, curr); } } + + bitset_minus1(curr); } + + /* transfer the best set into the qn */ + qn->mis_size = bitset_popcnt(best); + qn->mis_costs = best_weight; + next = 0; + bitset_foreach(best, pos) + qn->mis[next++] = ou->nodes[1+pos]; } /** @@ -470,11 +455,11 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { qnode_max_ind_set(qn, ou); /* do the insertion */ - DBG((dbg, LEVEL_4, "\t Insert qnode color %d with size %d\n", qn->color, qn->mis_size)); + DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs)); lh = &ou->queue; while (lh->next != &ou->queue) { qnode_t *curr = list_entry_queue(lh->next); - if (curr->mis_size <= qn->mis_size) + if (curr->mis_costs <= qn->mis_costs) break; lh = lh->next; } @@ -482,10 +467,10 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { } /** - * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower - * number of copy instructions placed during SSA-destruction and lowering. + * Tries to re-allocate colors of nodes in this opt unit, to achieve lower + * costs of copy instructions placed during SSA-destruction and lowering. * Works only for opt units with exactly 1 root node, which is the - * case for approximately 80% of all phi classes and all register constrained + * case for approximately 80% of all phi classes and 100% of register constrained * nodes. (All other phi classes are reduced to this case.) */ static void ou_optimize(unit_t *ou) { @@ -508,7 +493,7 @@ static void ou_optimize(unit_t *ou) { /* get head of queue */ curr = list_entry_queue(ou->queue.next); list_del(&curr->queue); - DBG((dbg, LEVEL_2, "\t Examine qnode color %d with size %d\n", curr->color, curr->mis_size)); + DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs)); /* try */ if (qnode_try_color(curr)) @@ -523,7 +508,7 @@ static void ou_optimize(unit_t *ou) { if (curr->mis_size >= 2) { node_stat_t *ns; - DBG((dbg, LEVEL_1, "\t Best color: %d Copies: %d/%d\n", curr->color, ou->interf+ou->node_count-curr->mis_size, ou->interf+ou->node_count-1)); + DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d/%d\n", curr->color, ou->complete_costs - curr->mis_costs, ou->complete_costs)); /* globally pin root and eventually others */ pset_insert_ptr(pinned_global, ou->nodes[0]); for (i=1; inode_count; ++i) { diff --git a/ir/be/becopyilp.c b/ir/be/becopyilp.c index 462e802ec..111bef007 100644 --- a/ir/be/becopyilp.c +++ b/ir/be/becopyilp.c @@ -31,12 +31,6 @@ static firm_dbg_module_t *dbg = NULL; #define EPSILON 0.00001 #define SLOTS_LIVING 32 -/** - * Represents the _costs_ if node n and m have different colors. - * Must be >=0. - **/ -#define get_weight(n,m) 1 - typedef struct _simpl_t { struct list_head chain; if_node_t *ifn; @@ -253,7 +247,7 @@ static void pi_add_constr_E(problem_instance_t *pi) { DBG((dbg, LEVEL_2, "Add E constraints...\n")); /* for all roots of optimization units */ list_for_each_entry(unit_t, curr, &pi->co->units, units) { - const ir_node *root, *arg; + ir_node *root, *arg; int rootnr, argnr, color; int y_idx, i; char buf[32]; @@ -272,7 +266,8 @@ static void pi_add_constr_E(problem_instance_t *pi) { /* Introduce new variable and set factor in objective function */ mangle_var(buf, 'y', rootnr, argnr); - y_idx = lpp_add_var(pi->curr_lp, buf, continous, get_weight(root, arg)); + y_idx = lpp_add_var(pi->curr_lp, buf, continous, curr->costs[i]); + /* set starting value */ //lpp_set_start_value(pi->curr_lp, y_idx, (get_irn_col(pi->co, root) != get_irn_col(pi->co, arg))); @@ -320,13 +315,13 @@ static void pi_add_constr_M(problem_instance_t *pi) { int cst_idx, y_idx, i; char buf[32]; - if (curr->ifg_mis_size == curr->node_count) + if (curr->minimal_costs == 0) continue; root = curr->nodes[0]; rootnr = get_irn_graph_nr(root); mangle_cst(buf, 'M', cst_counter++); - cst_idx = lpp_add_cst(pi->curr_lp, buf, greater, curr->node_count - curr->ifg_mis_size); + cst_idx = lpp_add_cst(pi->curr_lp, buf, greater, curr->minimal_costs); /* for all arguments */ for (i = 1; i < curr->node_count; ++i) { @@ -356,7 +351,7 @@ static problem_instance_t *new_pi(const copy_opt_t *co) { /* problem size reduction */ pi_find_simplicials(pi); - //TODO dump_ifg_w/o_removed + //TODO If you wish to see it: dump_ifg_w/o_removed if (pi->all_simplicial) return pi; diff --git a/ir/be/becopyopt.c b/ir/be/becopyopt.c index 667bf90bf..c544e9cc9 100644 --- a/ir/be/becopyopt.c +++ b/ir/be/becopyopt.c @@ -15,6 +15,7 @@ #endif #include "irprog.h" +#include "irloop_t.h" #include "xmalloc.h" #include "bechordal_t.h" @@ -29,71 +30,49 @@ static firm_dbg_module_t *dbg = NULL; #define MIN(a,b) ((anode_count * sizeof(*all)); - - /* all contains all nodes */ - all_size = 0; - for (i=0; inode_count; ++i) - all[all_size++] = ou->nodes[i]; - - /* which[i] says which element to take out of all[] and put into curr[i] */ - which = alloca(all_size*sizeof(*which)); - for (curr_size=0; curr_sizeco->chordal_env, curr[i], curr[o])) - goto conflict_found; - - /* We had no conflict. This is the (one) max indep. set */ - return curr_size; - -conflict_found: - /* We had a conflict. Generate next set */ - if (which[all_size-curr_size+1] == all_size-curr_size+1) { - curr_size--; - for (i=0; i=which[i+1]) { - redo = 1; - break; - } - } +static int ou_max_ind_set_costs(unit_t *ou) { + ir_node **irns; + int max, pos, curr_weight, best_weight = 0; + bitset_t *curr; + + irns = alloca((ou->node_count-1) * sizeof(*irns)); + curr = bitset_alloca(ou->node_count-1); + + /* brute force the best set */ + bitset_set_all(curr); + while ((max = bitset_popcnt(curr)) != 0) { + /* check if curr is a stable set */ + int i, o, is_stable_set = 1; + bitset_foreach(curr, pos) + irns[pos] = ou->nodes[1+pos]; + for(i=0; ico->chordal_env, irns[i], irns[o])) { + is_stable_set = 0; + break; + } + + if (is_stable_set) { + /* calc current weigth */ + curr_weight = 0; + bitset_foreach(curr, pos) + curr_weight += ou->costs[1+pos]; + + /* any better ? */ + if (curr_weight > best_weight) + best_weight = curr_weight; } + + bitset_minus1(curr); } - assert(0 && "How did you get here?"); + return best_weight; } /** @@ -106,6 +85,8 @@ conflict_found: static void co_append_unit(copy_opt_t *co, ir_node *root) { int i, arity; unit_t *unit; + struct list_head *tmp; + DBG((dbg, LEVEL_1, "\t Root: %n %N\n", root, root)); /* check if we encountered this root earlier */ if (pset_find_ptr(co->roots, root)) @@ -118,10 +99,12 @@ static void co_append_unit(copy_opt_t *co, ir_node *root) { arity = get_irn_arity(root); unit = xcalloc(1, sizeof(*unit)); unit->co = co; - unit->interf = 0; unit->node_count = 1; unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes)); + unit->costs = xmalloc((arity+1) * sizeof(*unit->costs)); unit->nodes[0] = root; + unit->complete_costs = 0; + unit->avg_costs = 0; INIT_LIST_HEAD(&unit->queue); /* check all args */ @@ -130,26 +113,61 @@ static void co_append_unit(copy_opt_t *co, ir_node *root) { ir_node *arg = get_irn_n(root, i); assert(is_curr_reg_class(arg) && "Argument not in same register class."); if (arg != root) { - if (!nodes_interfere(co->chordal_env, root, arg)) { - DBG((dbg, LEVEL_1, "\t Member: %n %N\n", arg, arg)); + int o, arg_pos = 0; + if (nodes_interfere(co->chordal_env, root, arg)) + assert(0 && "root and arg interfere"); + //TODO do not insert duplicate args + DBG((dbg, LEVEL_1, "\t Member: %n %N\n", arg, arg)); + + /* check if arg has occurred at a prior position in the arg/list */ + for (o=0; onode_count; ++o) + if (unit->nodes[o] == arg) { + arg_pos = o; + break; + } + + if (!arg_pos) { /* a new argument */ + /* TODO Think about the next 2 lines. (inserting in arg-order) */ if (is_optimizable(co->chordal_env->arch_env, arg)) co_append_unit(co, arg); - unit->nodes[unit->node_count++] = arg; - } else - unit->interf++; + /* insert node, set costs */ + unit->nodes[unit->node_count] = arg; + unit->costs[unit->node_count] = co->get_costs(root, arg, i); + unit->node_count++; + } else { /* arg has occured before in same phi */ + /* increase costs for existing arg */ + unit->costs[arg_pos] = co->get_costs(root, arg, i); + } } } unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes)); + unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs)); } else if (is_Copy(co->chordal_env->arch_env, root)) { assert(!nodes_interfere(co->chordal_env, root, get_Copy_src(root))); - unit->nodes[unit->node_count++] = get_Copy_src(root); + unit->nodes[1] = get_Copy_src(root); + unit->costs[1] = co->get_costs(root, unit->nodes[1], -1); + unit->node_count = 2; unit->nodes = xrealloc(unit->nodes, 2 * sizeof(*unit->nodes)); + unit->costs = xrealloc(unit->costs, 2 * sizeof(*unit->costs)); } else assert(0 && "This is not an optimizable node!"); + /* TODO add ou's for 2-addr-code instructions */ + + + for(i=1; inode_count; ++i) + unit->complete_costs += unit->costs[i]; + + assert(unit->node_count > 1); + unit->avg_costs = (100 * unit->complete_costs) / (unit->node_count-1); + + /* insert according to average costs */ + tmp = &co->units; + while (tmp->next != &co->units && list_entry_units(tmp->next)->avg_costs > unit->avg_costs) + tmp = tmp->next; + list_add(&unit->units, tmp); - list_add_tail(&unit->units, &co->units); /* Init ifg_mis_size to node_count. So get_lower_bound returns correct results. */ - unit->ifg_mis_size = get_ifg_mis_size(unit); + unit->minimal_costs = unit->complete_costs - ou_max_ind_set_costs(unit); } static void co_collect_in_block(ir_node *block, void *env) { @@ -169,16 +187,17 @@ static void co_collect_units(copy_opt_t *co) { del_pset(co->roots); } -copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env) { +copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)) { const char *s1, *s2, *s3; int len; - copy_opt_t *co; + copy_opt_t *co; dbg = firm_dbg_register("ir.be.copyopt"); firm_dbg_set_mask(dbg, DEBUG_LVL); co = xcalloc(1, sizeof(*co)); co->chordal_env = chordal_env; + co->get_costs = get_costs; s1 = get_irp_prog_name(); s2 = get_entity_name(get_irg_entity(co->chordal_env->irg)); @@ -215,17 +234,37 @@ int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) { return 0; } -int co_get_copy_count(const copy_opt_t *co) { +int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos) { + int cost = 0; + ir_loop *loop; + ir_node *root_block = get_nodes_block(root); + + assert(pos==-1 || is_Phi(root)); + if (pos == -1) { + /* a perm places the copy in the same block as it resides */ + loop = get_irn_loop(root_block); + } else { + /* for phis the copies are placed in the corresponding pred-block */ + loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos)); + } + if (loop) + cost = 2*get_loop_depth(loop); + return cost+1; +} + +int get_costs_all_one(ir_node *root, ir_node* arg, int pos) { + return 1; +} + +int co_get_copy_costs(const copy_opt_t *co) { int i, res = 0; unit_t *curr; list_for_each_entry(unit_t, curr, &co->units, units) { int root_col = get_irn_col(co, curr->nodes[0]); - res += curr->interf; - DBG((dbg, LEVEL_1, "%n %N has %d intf\n", curr->nodes[0], curr->nodes[0], curr->interf)); for (i=1; inode_count; ++i) if (root_col != get_irn_col(co, curr->nodes[i])) { DBG((dbg, LEVEL_1, " %n %N\n", curr->nodes[i], curr->nodes[i])); - res++; + res += curr->costs[i]; } } return res; @@ -235,55 +274,6 @@ int co_get_lower_bound(const copy_opt_t *co) { int res = 0; unit_t *curr; list_for_each_entry(unit_t, curr, &co->units, units) - res += curr->interf + curr->node_count - curr->ifg_mis_size; - return res; -} - -int co_get_interferer_count(const copy_opt_t *co) { - int res = 0; - unit_t *curr; - list_for_each_entry(unit_t, curr, &co->units, units) - res += curr->interf; + res += curr->minimal_costs; return res; } - -/** - * Needed for result checking - */ -static void co_collect_for_checker(ir_node *block, void *env) { - copy_opt_t *co = env; - struct list_head *head = get_block_border_head(co->chordal_env, block); - border_t *curr; - - list_for_each_entry_reverse(border_t, curr, head, list) - if (curr->is_def && curr->is_real && is_curr_reg_class(curr->irn)) - obstack_ptr_grow(&co->ob, curr->irn); -} - -/** - * This O(n^2) checker checks if - * two ifg-connected nodes have the same color - * register constraint are satisfied - */ -void co_check_allocation(copy_opt_t *co) { - ir_node **nodes, *n1, *n2; - int i, o; - - obstack_init(&co->ob); - dom_tree_walk_irg(co->chordal_env->irg, co_collect_for_checker, NULL, co); - obstack_ptr_grow(&co->ob, NULL); - - nodes = (ir_node **) obstack_finish(&co->ob); - for (i = 0, n1 = nodes[i]; n1; n1 = nodes[++i]) { - assert(arch_reg_is_allocatable(co->chordal_env->arch_env, n1, arch_pos_make_out(0), - arch_get_irn_register(co->chordal_env->arch_env, n1, 0)) && "Constraint does not hold"); - for (o = i+1, n2 = nodes[o]; n2; n2 = nodes[++o]) - if (nodes_interfere(co->chordal_env, n1, n2) - && get_irn_col(co, n1) == get_irn_col(co, n2)) { - DBG((dbg, 0, "Error in graph %s: %n %d and %n %d have the same color %d.\n", co->name, n1, get_irn_graph_nr(n1), n2, get_irn_graph_nr(n2), get_irn_col(co, n1))); - assert(0 && "Interfering values have the same color!"); - } - } - obstack_free(&co->ob, NULL); - DBG((dbg, 2, "The checker seems to be happy!\n")); -} diff --git a/ir/be/becopyopt.h b/ir/be/becopyopt.h index e60d4b4bd..e68717102 100644 --- a/ir/be/becopyopt.h +++ b/ir/be/becopyopt.h @@ -7,6 +7,10 @@ * Header for copy optimization problem. Analysis and set up of the problem. */ +/* + * TODO: get_nodes_block(get_irn_n(get_nodes_block(phi), i)); --> get_ifgblock_nodeblock + */ + #ifndef _BECOPYOPT_H #define _BECOPYOPT_H @@ -35,14 +39,17 @@ #define DEBUG_LVL_HEUR LEVEL_1 #define DEBUG_LVL_ILP LEVEL_1 +typedef int(*cost_fct_t)(ir_node*, ir_node*, int); + /** * Data representing the problem of copy minimization. */ typedef struct _copy_opt_t { be_chordal_env_t *chordal_env; - char *name; /**< ProgName__IrgName__RegClass */ - struct list_head units; /**< all units to optimize in right order */ - pset *roots; /**< used only temporary for detecting multiple appends */ + char *name; /**< ProgName__IrgName__RegClass */ + struct list_head units; /**< all units to optimize in right order */ + pset *roots; /**< used only temporary for detecting multiple appends */ + cost_fct_t get_costs; /**< function ptr used to get costs for copies */ struct obstack ob; } copy_opt_t; @@ -52,10 +59,14 @@ typedef struct _copy_opt_t { typedef struct _unit_t { struct list_head units; /**< chain for all units */ copy_opt_t *co; /**< the copy_opt this unit belongs to */ - int interf; /**< number of nodes dropped due to interference */ int node_count; /**< size of the nodes array */ ir_node **nodes; /**< [0] is the root-node, others are non interfering args of it. */ - int ifg_mis_size; /**< size of a mis considering only ifg (not coloring conflicts) */ + int *costs; /**< costs[i] are arising, if nodes[i] has a different color */ + int complete_costs; /**< sum of all costs[i] */ + int minimal_costs; /**< a lower bound for this ou, considering only ifg (not coloring conflicts) */ + + //TODO Think of the ordering. + int avg_costs; /**< average costs. controls the order of ou's. */ /* for heuristic */ struct list_head queue; /**< list of (mis/color) sorted by size of mis */ @@ -68,11 +79,13 @@ typedef struct _unit_t { #define get_irn_col(co, irn) \ arch_register_get_index(arch_get_irn_register(co->chordal_env->arch_env, irn, 0)) +#define list_entry_units(lh) list_entry(lh, unit_t, units) + /** * Generate the problem. Collect all infos and optimizable nodes. */ -copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env); +copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)); /** * Free the space... @@ -102,23 +115,31 @@ void free_copy_opt(copy_opt_t *co); */ int is_optimizable_arg(const copy_opt_t *co, ir_node *irn); +/** + * Computes the costs of a copy according to loop depth + * @param root, arg: clear. + * @param pos: -1 for perm-copies. + * Else the argument position of arg in the phi node root. + * @return Must be >= 0 in all cases. + */ +int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos); /** - * Returns the current number of copies needed + * All costs equal 1. Using this will reduce the number of copies. + * @return Must be >= 0 in all cases. */ -int co_get_copy_count(const copy_opt_t *co); +int get_costs_all_one(ir_node *root, ir_node* arg, int pos); /** - * Returns a lower bound for the number of copies needed based on interfering - * arguments and the size of a max indep. set (only ifg-edges) of the other args. + * Returns the current costs the copies are causing */ -int co_get_lower_bound(const copy_opt_t *co); +int co_get_copy_costs(const copy_opt_t *co); /** - * Returns the number of arguments interfering with their root node. This also - * is a (worse) lower bound for the number of copies needed. + * Returns a lower bound for the costs of copies based on interfering + * arguments and the size of a max indep. set (only ifg-edges) of the other args. */ -int co_get_interferer_count(const copy_opt_t *co); +int co_get_lower_bound(const copy_opt_t *co); /** * Solves the problem using a heuristic approach @@ -130,9 +151,4 @@ void co_heur_opt(copy_opt_t *co); */ void co_ilp_opt(copy_opt_t *co); -/** - * Checks the register allocation for correctness - */ -void co_check_allocation(copy_opt_t *co); - #endif -- 2.20.1