X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur2.c;h=1d727c010ea0e2be9df52152dd2705a13cd595c5;hb=df2faee01a5832057bb3ca0ba5f67e979c916e19;hp=72fc6d27c9576206e0c7fab58082c8f84ffddb96;hpb=e6bd37b3fb4fc9f852a072d9d102e463bed66828;p=libfirm diff --git a/ir/be/becopyheur2.c b/ir/be/becopyheur2.c index 72fc6d27c..1d727c010 100644 --- a/ir/be/becopyheur2.c +++ b/ir/be/becopyheur2.c @@ -1,9 +1,32 @@ +/* + * 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. + */ /** - * More experiments on coalescing. - * @author Sebastian Hack - * @date 14.04.2006 + * @file + * @brief More experiments on coalescing. + * @author Sebastian Hack + * @date 14.04.2006 */ +#include "config.h" + +#include "lc_opts.h" +#include "lc_opts_enum.h" #include #include @@ -11,120 +34,195 @@ #include "list.h" #include "pdeq.h" #include "bitset.h" +#include "raw_bitset.h" #include "debug.h" #include "bitfiddle.h" -#include "irphase_t.h" #include "irgraph_t.h" #include "irnode_t.h" #include "irprintf.h" - +#include "util.h" +#include "irtools.h" +#include "irnodemap.h" +#include "be_t.h" +#include "bemodule.h" #include "beabi.h" -#include "benode_t.h" +#include "benode.h" #include "becopyopt.h" #include "becopyopt_t.h" #include "bechordal_t.h" -#define INFEASIBLE(cost) ((cost) == INT_MAX) +#define DUMP_BEFORE 1 +#define DUMP_AFTER 2 +#define DUMP_CLOUD 4 +#define DUMP_ALL 2 * DUMP_CLOUD - 1 + +static unsigned dump_flags = 0; +static int subtree_iter = 4; +static int max_depth = 20; +static double constr_factor = 0.9; + +static const lc_opt_enum_mask_items_t dump_items[] = { + { "before", DUMP_BEFORE }, + { "after", DUMP_AFTER }, + { "cloud", DUMP_CLOUD }, + { "all", DUMP_ALL }, + { NULL, 0 } +}; -static be_ifg_dump_dot_cb_t ifg_dot_cb; +static lc_opt_enum_mask_var_t dump_var = { + &dump_flags, dump_items +}; + +static const lc_opt_table_entry_t options[] = { + LC_OPT_ENT_ENUM_MASK("dump", "dump ifg cloud", &dump_var), + LC_OPT_ENT_INT ("iter", "iterations for subtree nodes", &subtree_iter), + LC_OPT_ENT_DBL ("cf", "factor of constraint importance (between 0.0 and 1.0)", &constr_factor), + LC_OPT_ENT_INT ("max", "maximum recursion depth", &max_depth), + LC_OPT_LAST +}; + +/* + ____ _ _ + / ___|| |_ __ _ _ __| |_ + \___ \| __/ _` | '__| __| + ___) | || (_| | | | |_ + |____/ \__\__,_|_| \__| + +*/ + +#define INFEASIBLE(cost) ((cost) == INT_MAX) typedef unsigned col_t; -typedef struct _co2_irn_t co2_irn_t; -typedef struct _co2_cloud_t co2_cloud_t; +typedef struct co2_irn_t co2_irn_t; +typedef struct co2_cloud_t co2_cloud_t; +typedef struct co2_cloud_irn_t co2_cloud_irn_t; typedef struct { - phase_t ph; + col_t col; + int costs; +} col_cost_pair_t; + +typedef struct { + ir_nodemap map; + struct obstack obst; copy_opt_t *co; - bitset_t *ignore_regs; - co2_irn_t *touched; - int visited; + bitset_t *allocatable_regs; + co2_irn_t *touched; + int visited; + int n_regs; struct list_head cloud_head; DEBUG_ONLY(firm_dbg_module_t *dbg;) } co2_t; -struct _co2_irn_t { - ir_node *irn; - co2_cloud_t *cloud; - co2_irn_t *touched_next; +struct co2_irn_t { + const ir_node *irn; affinity_node_t *aff; - int costs; + co2_irn_t *touched_next; col_t tmp_col; col_t orig_col; - int visited; - int rank; - unsigned fixed : 1; - unsigned tmp_fixed : 1; + int last_color_change; + bitset_t *adm_cache; + unsigned fixed : 1; + unsigned tmp_fixed : 1; + unsigned is_constrained : 1; struct list_head changed_list; - struct list_head cloud_list; }; -struct _co2_cloud_t { - int costs; - int inevit; - int best_costs; - int n_memb; - int max_degree; - co2_irn_t *master; - co2_irn_t **seq; - col_t *best_cols; - struct list_head members_head; - struct list_head list; +struct co2_cloud_irn_t { + struct co2_irn_t inh; + co2_cloud_t *cloud; + int visited; + int index; + co2_cloud_irn_t *mst_parent; + int mst_costs; + int mst_n_childs; + co2_cloud_irn_t **mst_childs; + int *col_costs; + int costs; + int *fronts; + int *color_badness; + col_cost_pair_t *tmp_coloring; + struct list_head cloud_list; + struct list_head mst_list; }; -#define NEIGHBOR_FIXED 1 -#define NEIGHBOR_CONSTR 2 -#define SELF_CONSTR 4 -#define DONT_WANT 8 +struct co2_cloud_t { + co2_t *env; + struct obstack obst; + int costs; + int mst_costs; + int inevit; + int best_costs; + int n_memb; + int n_constr; + int max_degree; + int ticks; + double freedom; + co2_cloud_irn_t *master; + co2_cloud_irn_t *mst_root; + co2_cloud_irn_t **seq; + struct list_head members_head; + struct list_head list; +}; typedef struct { - col_t col; + co2_cloud_irn_t *src, *tgt; int costs; - unsigned flags; -} col_cost_pair_t; +} edge_t; -#define get_co2_irn(co2, irn) ((co2_irn_t *) phase_get_or_set_irn_data(&co2->ph, irn)) +#define FRONT_BASE(ci,col) ((ci)->fronts + col * (ci)->mst_n_childs) -static void co2_irn_init(phase_t *ph, const ir_node *irn, void *data) +static co2_irn_t *get_co2_irn(co2_t *env, const ir_node *node) { - co2_t *env = (co2_t *) ph; - co2_irn_t *ci = data; - - memset(ci, 0, sizeof(ci[0])); - INIT_LIST_HEAD(&ci->changed_list); - INIT_LIST_HEAD(&ci->cloud_list); - ci->irn = irn; - ci->touched_next = env->touched; - ci->orig_col = get_irn_col(env->co, irn); - ci->aff = get_affinity_info(env->co, (ir_node *)irn); - env->touched = ci; + co2_irn_t *ci = ir_nodemap_get(co2_irn_t, &env->map, node); + if (ci == NULL) { + ci = OALLOCZ(&env->obst, co2_irn_t); + + INIT_LIST_HEAD(&ci->changed_list); + ci->touched_next = env->touched; + ci->orig_col = get_irn_col(node); + env->touched = ci; + ci->irn = node; + ci->aff = NULL; + + ir_nodemap_insert(&env->map, node, ci); + } + return ci; } - -static int co2_irn_cmp(const void *a, const void *b) +static co2_cloud_irn_t *get_co2_cloud_irn(co2_t *env, const ir_node *node) { - const co2_irn_t **p = a; - const co2_irn_t **q = b; - return (*q)->costs - (*p)->costs; -} + co2_cloud_irn_t *ci = ir_nodemap_get(co2_cloud_irn_t, &env->map, node); + if (ci == NULL) { + ci = OALLOCZ(&env->obst, co2_cloud_irn_t); -static int cmp_clouds(const void *a, const void *b) -{ - const co2_cloud_t **p = a; - const co2_cloud_t **q = b; - return (*q)->costs - (*p)->costs; + INIT_LIST_HEAD(&ci->inh.changed_list); + ci->inh.touched_next = env->touched; + ci->inh.orig_col = get_irn_col(node); + env->touched = &ci->inh; + ci->inh.irn = node; + ci->inh.aff = get_affinity_info(env->co, node); + + INIT_LIST_HEAD(&ci->cloud_list); + ci->mst_parent = ci; + + ir_nodemap_insert(&env->map, node, ci); + } + return ci; } -static co2_cloud_t *new_cloud(co2_t *env) +#define CLOUD_WEIGHT(c) ((1 - constr_factor) * (c)->costs + constr_factor * (c)->freedom) + +static int cmp_clouds_gt(const void *a, const void *b) { - co2_cloud_t *cloud = phase_alloc(&env->ph, sizeof(cloud[0])); - memset(cloud, 0, sizeof(cloud[0])); - INIT_LIST_HEAD(&cloud->members_head); - INIT_LIST_HEAD(&cloud->list); - cloud->best_costs = INT_MAX; - return cloud; + const co2_cloud_t * const *p = (const co2_cloud_t*const*)a; + const co2_cloud_t * const *q = (const co2_cloud_t*const*)b; + double c = CLOUD_WEIGHT(*p); + double d = CLOUD_WEIGHT(*q); + return QSORT_CMP(d, c); } /** @@ -133,76 +231,89 @@ static co2_cloud_t *new_cloud(co2_t *env) */ static int col_cost_pair_lt(const void *a, const void *b) { - const col_cost_pair_t *p = a; - const col_cost_pair_t *q = b; + const col_cost_pair_t *p = (const col_cost_pair_t*)a; + const col_cost_pair_t *q = (const col_cost_pair_t*)b; int c = p->costs; int d = q->costs; - - return (c > d) - (c < d); + return QSORT_CMP(c, d); } -const char *flag_str(unsigned int fl) +static int cmp_edges(const void *a, const void *b) { - static char buf[10]; - - buf[0] = fl & NEIGHBOR_CONSTR ? 'c' : '-'; - buf[1] = fl & NEIGHBOR_FIXED ? 'n' : '-'; - buf[2] = fl & SELF_CONSTR ? 'C' : '-'; - buf[3] = fl & DONT_WANT ? 'd' : '-'; - buf[4] = '\0'; - return buf; + const edge_t *p = (const edge_t*)a; + const edge_t *q = (const edge_t*)b; + return QSORT_CMP(q->costs, p->costs); } -static col_t get_col(co2_t *env, ir_node *irn) +static col_t get_col(co2_t *env, const ir_node *irn) { co2_irn_t *ci = get_co2_irn(env, irn); return ci->tmp_fixed ? ci->tmp_col : ci->orig_col; } -static INLINE int color_is_fix(co2_t *env, ir_node *irn) +static inline int color_is_fix(co2_t *env, const ir_node *irn) { co2_irn_t *ci = get_co2_irn(env, irn); return ci->fixed || ci->tmp_fixed; } -static bitset_t *admissible_colors(co2_t *env, co2_irn_t *ci, bitset_t *bs) +static inline bitset_t *get_adm(co2_t *env, co2_irn_t *ci) { - arch_register_req_t req; - - arch_get_register_req(env->co->aenv, &req, ci->irn, BE_OUT_POS(0)); - if(arch_register_req_is(&req, limited)) - req.limited(req.limited_env, bs); - else { - bitset_copy(bs, env->ignore_regs); - bitset_flip_all(bs); + if (ci->adm_cache == NULL) { + const arch_register_req_t *req; + ci->adm_cache = bitset_obstack_alloc(&env->obst, env->n_regs); + req = arch_get_irn_register_req(ci->irn); + + if (arch_register_req_is(req, limited)) { + int i, n; + + n = env->n_regs; + for (i = 0; i < n; ++i) { + if (rbitset_is_set(req->limited, i)) + bitset_set(ci->adm_cache, i); + } + ci->is_constrained = 1; + } else { + bitset_copy(ci->adm_cache, env->allocatable_regs); + } } + return ci->adm_cache; +} + +static inline bitset_t *admissible_colors(co2_t *env, co2_irn_t *ci, bitset_t *bs) +{ + bitset_copy(bs, get_adm(env, ci)); return bs; } -static int is_color_admissible(co2_t *env, co2_irn_t *ci, col_t col) +static inline int is_color_admissible(co2_t *env, co2_irn_t *ci, col_t col) { - bitset_t *bs = bitset_alloca(env->co->cls->n_regs); - admissible_colors(env, ci, bs); + bitset_t *bs = get_adm(env, ci); return bitset_is_set(bs, col); } -static void incur_constraint_costs(co2_t *env, ir_node *irn, col_cost_pair_t *col_costs, int costs) +static inline int is_constrained(co2_t *env, co2_irn_t *ci) { - bitset_t *aux = bitset_alloca(env->co->cls->n_regs); - arch_register_req_t req; + if (!ci->adm_cache) + get_adm(env, ci); + return ci->is_constrained; +} - arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0)); +static void incur_constraint_costs(co2_t *env, const ir_node *irn, col_cost_pair_t *col_costs, int costs) +{ + const arch_register_req_t *req = arch_get_irn_register_req(irn); - if(arch_register_req_is(&req, limited)) { - bitset_pos_t elm; - int n_constr; + if (arch_register_req_is(req, limited)) { + unsigned n_regs = env->co->cls->n_regs; + unsigned n_constr = 0; + unsigned i; - req.limited(req.limited_env, aux); - n_constr = bitset_popcnt(aux); - bitset_foreach(aux, elm) { - col_costs[elm].costs = add_saturated(col_costs[elm].costs, costs / n_constr); - col_costs[elm].flags |= NEIGHBOR_CONSTR; + n_constr = rbitset_popcount(req->limited, n_regs); + for (i = 0; i < n_regs; ++i) { + if (rbitset_is_set(req->limited, i)) { + col_costs[i].costs = add_saturated(col_costs[i].costs, costs / n_constr); + } } } } @@ -219,36 +330,27 @@ static void incur_constraint_costs(co2_t *env, ir_node *irn, col_cost_pair_t *co */ static void determine_color_costs(co2_t *env, co2_irn_t *ci, col_cost_pair_t *col_costs) { - ir_node *irn = ci->irn; + const ir_node *irn = ci->irn; be_ifg_t *ifg = env->co->cenv->ifg; int n_regs = env->co->cls->n_regs; - bitset_t *forb = bitset_alloca(n_regs); - affinity_node_t *a = get_affinity_info(env->co, irn); + affinity_node_t *a = ci->aff; - bitset_pos_t elm; - ir_node *pos; - void *it; + const ir_node *pos; + neighbours_iter_t it; int i; - if(get_irn_node_nr(irn) == 2040) { - printf("Hallo"); - } - /* Put all forbidden colors into the aux bitset. */ - admissible_colors(env, ci, forb); - bitset_flip_all(forb); + bitset_t *const admissible = bitset_alloca(n_regs); + admissible_colors(env, ci, admissible); - for(i = 0; i < n_regs; ++i) { + for (i = 0; i < n_regs; ++i) { col_costs[i].col = i; col_costs[i].costs = 0; - col_costs[i].flags = 0; } - if(a) { - neighb_t *n; - + if (a) { co_gs_foreach_neighb(a, n) { - if(color_is_fix(env, n->irn)) { + if (color_is_fix(env, n->irn)) { col_t col = get_col(env, n->irn); col_costs[col].costs = add_saturated(col_costs[col].costs, -n->costs * 128); } @@ -257,145 +359,91 @@ static void determine_color_costs(co2_t *env, co2_irn_t *ci, col_cost_pair_t *co } } - it = be_ifg_neighbours_iter_alloca(ifg); - be_ifg_foreach_neighbour(ifg, it, irn, pos) { + be_ifg_foreach_neighbour(ifg, &it, irn, pos) { col_t col = get_col(env, pos); - if(color_is_fix(env, pos)) { + if (color_is_fix(env, pos)) { col_costs[col].costs = INT_MAX; - col_costs[col].flags |= NEIGHBOR_FIXED; } else { incur_constraint_costs(env, pos, col_costs, INT_MAX); col_costs[col].costs = add_saturated(col_costs[col].costs, 8 * be_ifg_degree(ifg, pos)); } } + be_ifg_neighbours_break(&it); /* Set the costs to infinity for each color which is not allowed at this node. */ - bitset_foreach(forb, elm) { + bitset_foreach_clear(admissible, elm) { col_costs[elm].costs = INT_MAX; - col_costs[elm].flags |= SELF_CONSTR; } } -static void single_color_cost(co2_t *env, col_t col, col_cost_pair_t *seq) +static void single_color_cost(co2_t *env, co2_irn_t *ci, col_t col, col_cost_pair_t *seq) { int n_regs = env->co->cls->n_regs; int i; - for(i = 0; i < n_regs; ++i) { + for (i = 0; i < n_regs; ++i) { seq[i].col = i; seq[i].costs = INT_MAX; - seq[i].flags = 0; - seq[i].flags = DONT_WANT; } + (void) ci; + assert(is_color_admissible(env, ci, col)); seq[col].col = 0; seq[0].col = col; seq[0].costs = 0; - seq[0].flags = 0; -} - -static int curr_costs(co2_t *env, affinity_node_t *a) -{ - col_t a_col = get_col(env, a->irn); - int costs = 0; - neighb_t *n; - - co_gs_foreach_neighb(a, n) { - col_t n_col = get_col(env, n->irn); - costs += n_col != a_col ? n->costs : 0; - } - - return costs; -} - -static int cloud_costs(co2_t *env, co2_cloud_t *cloud) -{ - int costs = 0; - co2_irn_t *ci; - - list_for_each_entry(co2_irn_t, ci, &cloud->members_head, cloud_list) { - affinity_node_t *a = get_affinity_info(env->co, ci->irn); - costs += curr_costs(env, a); - } - - return costs; } static void reject_coloring(struct list_head *h) { - co2_irn_t *pos; - list_for_each_entry(co2_irn_t, pos, h, changed_list) pos->tmp_fixed = 0; } static void materialize_coloring(struct list_head *h) { - co2_irn_t *pos; - list_for_each_entry(co2_irn_t, pos, h, changed_list) { - pos->orig_col = pos->tmp_col; + pos->orig_col = pos->tmp_col; pos->tmp_fixed = 0; } } -typedef struct { - co2_irn_t *ci; - col_t col; -} col_entry_t; - -static col_entry_t *save_coloring(struct obstack *obst, struct list_head *changed) -{ - co2_irn_t *pos; - col_entry_t ent; - - list_for_each_entry(co2_irn_t, pos, changed, changed_list) { - ent.ci = pos; - ent.col = pos->tmp_col; - pos->tmp_col = 0; - obstack_grow(obst, &ent, sizeof(ent)); - } - memset(&ent, 0, sizeof(ent)); - obstack_grow(obst, &ent, sizeof(ent)); - return obstack_finish(obst); -} - -static int change_color_not(co2_t *env, ir_node *irn, col_t not_col, struct list_head *parent_changed, int depth); -static int change_color_single(co2_t *env, ir_node *irn, col_t tgt_col, struct list_head *parent_changed, int depth); +static int change_color_not(co2_t *env, const ir_node *irn, col_t not_col, struct list_head *parent_changed, int depth); -static int recolor(co2_t *env, ir_node *irn, col_cost_pair_t *col_list, struct list_head *parent_changed, int depth) +static int recolor(co2_t *env, const ir_node *irn, col_cost_pair_t *col_list, struct list_head *parent_changed, int depth) { int n_regs = env->co->cls->n_regs; be_ifg_t *ifg = env->co->cenv->ifg; co2_irn_t *ci = get_co2_irn(env, irn); int res = 0; - int n_aff = 0; int i; - for(i = 0; i < n_regs; ++i) { + if (depth >= max_depth) + return 0; + + for (i = 0; i < n_regs; ++i) { col_t tgt_col = col_list[i].col; unsigned costs = col_list[i].costs; int neigh_ok = 1; struct list_head changed; - ir_node *n; - void *it; + const ir_node *n; + neighbours_iter_t it; DBG((env->dbg, LEVEL_3, "\t\t%2{firm:indent}trying color %d(%d) on %+F\n", depth, tgt_col, costs, irn)); /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */ - if(INFEASIBLE(costs)) { - DB((env->dbg, LEVEL_4, "\t\t%2{firm:indent}color %d infeasible due to %s\n", depth, tgt_col, flag_str(col_list[i].flags))); + if (INFEASIBLE(costs)) { + DB((env->dbg, LEVEL_4, "\t\t%2{firm:indent}color %d infeasible\n", depth, tgt_col)); ci->tmp_fixed = 0; return 0; } /* Set the new color of the node and mark the node as temporarily fixed. */ - ci->tmp_col = tgt_col; - ci->tmp_fixed = 1; + ci->tmp_col = tgt_col; + ci->tmp_fixed = 1; /* If that color has costs > 0, there's at least one neighbor having that color, @@ -404,11 +452,10 @@ static int recolor(co2_t *env, ir_node *irn, col_cost_pair_t *col_list, struct l INIT_LIST_HEAD(&changed); list_add(&ci->changed_list, &changed); - it = be_ifg_neighbours_iter_alloca(ifg); - be_ifg_foreach_neighbour(ifg, it, irn, n) { + be_ifg_foreach_neighbour(ifg, &it, irn, n) { /* try to re-color the neighbor if it has the target color. */ - if(get_col(env, n) == tgt_col) { + if (get_col(env, n) == tgt_col) { struct list_head tmp; /* @@ -420,16 +467,17 @@ static int recolor(co2_t *env, ir_node *irn, col_cost_pair_t *col_list, struct l INIT_LIST_HEAD(&tmp); neigh_ok = change_color_not(env, n, tgt_col, &tmp, depth + 1); list_splice(&tmp, &changed); - if(!neigh_ok) + if (!neigh_ok) break; } } + be_ifg_neighbours_break(&it); /* We managed to assign the target color to all neighbors, so from the perspective of the current node, every thing was ok and we can return safely. */ - if(neigh_ok) { + if (neigh_ok) { DBG((env->dbg, LEVEL_3, "\t\t%2{firm:indent}color %d(%d) was ok\n", depth, tgt_col, costs)); list_splice(&changed, parent_changed); res = 1; @@ -447,7 +495,7 @@ static int recolor(co2_t *env, ir_node *irn, col_cost_pair_t *col_list, struct l return res; } -static int change_color_not(co2_t *env, ir_node *irn, col_t not_col, struct list_head *parent_changed, int depth) +static int change_color_not(co2_t *env, const ir_node *irn, col_t not_col, struct list_head *parent_changed, int depth) { co2_irn_t *ci = get_co2_irn(env, irn); int res = 0; @@ -456,10 +504,10 @@ static int change_color_not(co2_t *env, ir_node *irn, col_t not_col, struct list DBG((env->dbg, LEVEL_3, "\t\t%2{firm:indent}clearing %+F(%d) of color %d\n", depth, irn, col, not_col)); /* the node does not have to forbidden color. That's fine, mark it as visited and return. */ - if(col != not_col) { - if(!ci->tmp_fixed) { - ci->tmp_col = col; - ci->tmp_fixed = 1; + if (col != not_col) { + if (!ci->tmp_fixed) { + ci->tmp_col = col; + ci->tmp_fixed = 1; } list_add(&ci->changed_list, parent_changed); @@ -467,9 +515,9 @@ static int change_color_not(co2_t *env, ir_node *irn, col_t not_col, struct list } /* The node has the color it should not have _and_ has not been visited yet. */ - if(!color_is_fix(env, irn)) { + if (!color_is_fix(env, irn)) { int n_regs = env->co->cls->n_regs; - col_cost_pair_t *csts = alloca(n_regs * sizeof(csts[0])); + col_cost_pair_t *csts = ALLOCAN(col_cost_pair_t, n_regs); /* Get the costs for giving the node a specific color. */ determine_color_costs(env, ci, csts); @@ -488,7 +536,7 @@ static int change_color_not(co2_t *env, ir_node *irn, col_t not_col, struct list return res; } -static int change_color_single(co2_t *env, ir_node *irn, col_t tgt_col, struct list_head *parent_changed, int depth) +static int change_color_single(co2_t *env, const ir_node *irn, col_t tgt_col, struct list_head *parent_changed, int depth) { co2_irn_t *ci = get_co2_irn(env, irn); col_t col = get_col(env, irn); @@ -496,454 +544,261 @@ static int change_color_single(co2_t *env, ir_node *irn, col_t tgt_col, struct l DBG((env->dbg, LEVEL_3, "\t\t%2{firm:indent}trying to set %+F(%d) to color %d\n", depth, irn, col, tgt_col)); - /* If the color is already fix, bail out. */ - if(color_is_fix(env, irn)) - return 0; - /* the node has the wanted color. That's fine, mark it as visited and return. */ - if(col == tgt_col) { - if(!ci->tmp_fixed) { - ci->tmp_col = col; - ci->tmp_fixed = 1; + if (col == tgt_col) { + if (!ci->tmp_fixed) { + ci->tmp_col = col; + ci->tmp_fixed = 1; + list_add(&ci->changed_list, parent_changed); } - list_add(&ci->changed_list, parent_changed); - DB((env->dbg, LEVEL_3, "\t\tok\n")); - return 1; + res = 1; + goto end; } - else { + if (!color_is_fix(env, irn) && is_color_admissible(env, ci, tgt_col)) { int n_regs = env->co->cls->n_regs; - col_cost_pair_t *seq = alloca(n_regs * sizeof(seq[0])); + col_cost_pair_t *seq = ALLOCAN(col_cost_pair_t, n_regs); /* Get the costs for giving the node a specific color. */ - single_color_cost(env, tgt_col, seq); + single_color_cost(env, ci, tgt_col, seq); /* Try recoloring the node using the color list. */ res = recolor(env, irn, seq, parent_changed, depth); - DB((env->dbg, LEVEL_3, "\t\tcolor %d %s for %+F\n", tgt_col, res ? "was ok" : "failed", irn)); } +end: + DB((env->dbg, LEVEL_3, "\t\t%2{firm:indent}color %d %s for %+F\n", depth, tgt_col, res ? "was ok" : "failed", irn)); return res; } - -#if 0 -static void try_color(co2_t *env, co2_irn_t *ci, col_t col, struct list_head *parent_changed) +/** + * Examine the costs of the current coloring concerning a MST subtree. + * @param ci The subtree root. + * @param col The color of @p ci. + * @return The best coloring for that subtree under the assumption that @p ci has color @p col. + */ +static int examine_subtree_coloring(co2_cloud_irn_t *ci, col_t col) { - be_ifg_t *ifg = env->co->cenv->ifg; - int n_regs = env->co->cls->n_regs; - col_cost_pair_t *col_seq = alloca(n_regs * sizeof(col_seq[0])); - affinity_node_t *a = get_affinity_info(env->co, ci->irn); - co2_irn_t **nbs = alloca(a->degree * sizeof(nbs[0])); - int ok = 0; - - col_t new_col; - neighb_t *n; + int *front = FRONT_BASE(ci, col); + int cost = 0; + int i; - assert(a != NULL && "This node must be an affinity node"); + for (i = 0; i < ci->mst_n_childs; ++i) { + co2_cloud_irn_t *chld = ci->mst_childs[i]; + col_t chld_col = front[i]; - /* If that node has already been fixed, leave it alone. */ - if(color_is_fix(env, ci->irn) || !is_color_admissible(env, ci, col)) { - // DB((env->dbg, LEVEL_2, "\t-> color is already fix: %d\n", get_col(env, ci->irn))); - return; + cost += examine_subtree_coloring(chld, chld_col); + cost += col != chld_col ? chld->mst_costs : 0; } - DB((env->dbg, LEVEL_1, "\taffinity node %+F cost %d trying color %d\n", ci->irn, ci->costs, col)); - - single_color_cost(env, col, col_seq); - recolor(env, ci->irn, col_seq, parent_changed, 0); - new_col = get_col(env, ci->irn); - - ci->tmp_fixed = 1; - ci->tmp_col = new_col; - - DB((env->dbg, LEVEL_2, "\t-> has color %d now. %d wanted\n", new_col, col)); - - i = 0; - co_gs_foreach_neighb(a, n) - nbs[i++] = get_co2_irn(env, n->irn); - - co_gs_foreach_neighb(a, n) { - co2_irn_t *ni = get_co2_irn(env, n->irn); - col_t tgt_col = be_ifg_connected(ifg, ci->irn, ni->irn) ? get_col(env, ni->irn) : new_col; - try_color(env, ni, tgt_col, parent_changed); - } + return cost; } -static void process_cloud(co2_t *env, co2_cloud_t *cloud) +/** + * Determine color badnesses of a node. + * Badness means that it is unlikely that the node in question can + * obtain a color. The higher the badness, the more unlikely it is that + * the node can be assigned that color. + * @param ci The node. + * @param badness An integer array as long as there are registers. + * @note The array badness is not cleared. + */ +static void node_color_badness(co2_cloud_irn_t *ci, int *badness) { - int n_regs = env->co->cls->n_regs; - col_cost_pair_t *cols = alloca(n_regs * sizeof(cols[0])); - int best_costs = cloud_costs(env, cloud); - int best_col = 0; - - struct list_head changed; - co2_irn_t *ci; - int i; - - - i = 0; - DB((env->dbg, LEVEL_2, "processing cloud with costs %d and master %+F containing:\n", cloud->costs, cloud->master->irn)); - list_for_each_entry(co2_irn_t, ci, &cloud->members_head, cloud_list) { - DB((env->dbg, LEVEL_2, "\t%+F %d\n", ci->irn, ci->costs)); - } - - determine_color_costs(env, cloud->master, cols); - qsort(cols, n_regs, sizeof(cols[0]), col_cost_pair_lt); - - best_col = cols[0].col; - for(i = 0; i < n_regs; ++i) { - col_t col = cols[i].col; - int reject = 1; - int costs; - - INIT_LIST_HEAD(&changed); - DBG((env->dbg, LEVEL_2, "\n\ttrying color %d. current costs: %d\n", col, best_costs)); - - /* try to recolor all the cloud members. */ - try_color(env, cloud->master, col, &changed); - - /* recoloring of all nodes did succeed. measure the costs and decide if the coloring shall be kept. */ - costs = cloud_costs(env, cloud); - - /* materialize the new coloring. */ - if(costs < best_costs) { - materialize_coloring(&changed); - best_costs = costs; - best_col = col; - reject = 0; + co2_t *env = ci->cloud->env; + co2_irn_t *ir = &ci->inh; + int n_regs = env->n_regs; + be_ifg_t *ifg = env->co->cenv->ifg; + bitset_t *bs = bitset_alloca(n_regs); + + const ir_node *irn; + neighbours_iter_t it; + + admissible_colors(env, &ci->inh, bs); + bitset_foreach_clear(bs, elm) + badness[elm] = ci->costs; + + /* Use constrained/fixed interfering neighbors to influence the color badness */ + be_ifg_foreach_neighbour(ifg, &it, ir->irn, irn) { + co2_irn_t *ni = get_co2_irn(env, irn); + + admissible_colors(env, ni, bs); + if (bitset_popcount(bs) == 1) { + size_t c = bitset_next_set(bs, 0); + badness[c] += ci->costs; } - /* We won't get the cloud any better so stop it. */ - if(costs == 0) - break; - - if(reject) - reject_coloring(&changed); - } - - DB((env->dbg, LEVEL_2, "\tfinished cloud with costs %d\n", best_costs)); - - /* fix all cloud members */ - list_for_each_entry(co2_irn_t, ci, &cloud->members_head, cloud_list) { - ci->fixed = 1; - } - - { - char buf[256]; - FILE *f; - - ir_snprintf(buf, sizeof(buf), "ifg_%F_%s_cloud_%F.dot", env->co->irg, env->co->cls->name, cloud->master); - if(f = fopen(buf, "wt")) { - be_ifg_dump_dot(env->co->cenv->ifg, env->co->irg, f, &ifg_dot_cb, env); - fclose(f); + else if (ni->fixed) { + col_t c = get_col(env, ni->irn); + badness[c] += ci->costs; } } + be_ifg_neighbours_break(&it); } -static void try_affinity_node(co2_t *env, co2_irn_t *ci, col_t preferred, struct list_head *parent_changed) +/** + * Determine the badness of a MST subtree. + * The badness is written into the color_badness array of each node and accumulated in the parents. + * @see node_color_badness() for a definition of badness. + * @param ci The root of the subtree. + * @param depth Depth for debugging purposes. + */ +static void determine_color_badness(co2_cloud_irn_t *ci, int depth) { - ir_node *irn = ci->irn; - - if(!color_is_fix(env, irn)) { - int n_regs = env->co->cls->n_regs; - bitset_t *tried = bitset_alloca(n_regs); - bitset_t *adm = bitset_alloca(n_regs); - col_cost_pair_t *seq = alloca(n_regs * sizeof(seq[0])); - - affinity_node_t *a = get_affinity_info(env->co, irn); - int best_costs = cloud_costs(env, ci->cloud); - int best_col = get_col(env, ci->irn); - - int i; - - determine_color_costs(env, ci, seq); - if(!INFEASIBLE(seq[preferred].costs)) - seq[preferred].costs = INT_MIN; - - qsort(seq, n_regs, sizeof(seq[0]), col_cost_pair_lt); - - for(i = 0; i < n_regs; ++i) { - col_t col = seq[i].col; - - struct list_head changed; - int ok, costs; - - INIT_LIST_HEAD(&changed); - ok = change_color_single(env, irn, col, &changed, 0); - col = get_col(env, irn); - - if(!bitset_is_set(tried, col)) { - neighb_t *n; - - if(!ci->tmp_col) { - ci->tmp_col = col; - ci->tmp_fixed = 1; - list_add(&ci->changed_list, &changed); - } + co2_t *env = ci->cloud->env; + int i, j; - co_gs_foreach_neighb(a, n) { - co2_irn_t *ni = get_co2_irn(env, n->irn); - try_affinity_node(env, ni, col, &changed); - } + node_color_badness(ci, ci->color_badness); - examine_coloring(env, ci->cloud); - reject_coloring(&changed); + /* Collect the color badness for the whole subtree */ + for (i = 0; i < ci->mst_n_childs; ++i) { + co2_cloud_irn_t *child = ci->mst_childs[i]; + determine_color_badness(child, depth + 1); - bitset_set(tried, col); - } - } + for (j = 0; j < env->n_regs; ++j) + ci->color_badness[j] += child->color_badness[j]; } -} -#endif - -static void examine_cloud_coloring(co2_t *env, co2_cloud_t *cloud) -{ - int costs = cloud_costs(env, cloud); - if(costs < cloud->best_costs) { - int i; - - for(i = 0; i < cloud->n_memb; ++i) - cloud->best_cols[i] = get_col(env, cloud->seq[i]->irn); - - cloud->best_costs = costs; - } + for (j = 0; j < env->n_regs; ++j) + DBG((env->dbg, LEVEL_2, "%2{firm:indent}%+F col %d badness %d\n", depth, ci->inh.irn, j, ci->color_badness[j])); } -static int color_change_balance(co2_t *env, co2_irn_t *ci, bitset_t *tried_colors) +/** + * Unfix all nodes in a MST subtree. + */ +static void unfix_subtree(co2_cloud_irn_t *ci) { - col_t col = get_col(env, ci->irn); - neighb_t *n; - int balance = 0; - - co_gs_foreach_neighb(ci->aff, n) { - col_t nc = get_col(env, n->irn); - int fixed = color_is_fix(env, n->irn); - - if(nc == col) - balance -= n->costs; - else if(!fixed || !bitset_is_set(tried_colors, nc)) - balance += n->costs; - } + int i; - DBG((env->dbg, LEVEL_4, "\t\tbalance for changing %+F color %d\n", ci->irn, balance)); - return balance; + ci->inh.fixed = 0; + for (i = 0; i < ci->mst_n_childs; ++i) + unfix_subtree(ci->mst_childs[i]); } -static void adjust_start_colors(co2_t *env, co2_cloud_t *cloud, col_cost_pair_t *seq) +static int coalesce_top_down(co2_cloud_irn_t *ci, int child_nr, int depth) { - int n_regs = env->co->cls->n_regs; - bitset_t *adm = bitset_alloca(n_regs); - bitset_pos_t col; - int i; + co2_t *env = ci->cloud->env; + col_cost_pair_t *seq = ALLOCAN(col_cost_pair_t, env->n_regs); + int is_root = ci->mst_parent == ci; + col_t parent_col = is_root ? (col_t) -1 : get_col(env, ci->mst_parent->inh.irn); + int min_badness = INT_MAX; + int best_col_costs = INT_MAX; + int best_col = -1; + int n_regs = env->n_regs; + int n_iter = is_root ? MIN(n_regs, subtree_iter) : 1; - for(i = 0; i < cloud->n_memb; ++i) { - co2_irn_t *ci = cloud->seq[i]; - int n_constr; + struct list_head changed; + int ok, i, j; - /* Prefer precolored neighbors. */ - bitset_clear_all(adm); - admissible_colors(env, ci, adm); - n_constr = bitset_popcnt(adm); + for (i = 0; i < n_regs; ++i) { + int badness = ci->color_badness[i]; - bitset_foreach(adm, col) { - seq[col].costs = add_saturated(seq[col].costs, -128 * (n_regs - n_constr)); - } + seq[i].col = i; + seq[i].costs = is_color_admissible(env, &ci->inh, i) ? badness : INT_MAX; - bitset_foreach_clear(adm, col) { - seq[col].costs = add_saturated(seq[col].costs, 128); - } + min_badness = MIN(min_badness, badness); } - admissible_colors(env, cloud->master, adm); - bitset_flip_all(adm); + /* If we are not the root and the parent's color is allowed for this node give it top prio. */ + if (!is_root && is_color_admissible(env, &ci->inh, parent_col)) + seq[parent_col].costs = min_badness - 1; - bitset_foreach(adm, col) - seq[col].costs = INT_MAX; -} + /* Sort the colors. The will be processed in that ordering. */ + qsort(seq, env->n_regs, sizeof(seq[0]), col_cost_pair_lt); -static int process_node(co2_t *env, co2_cloud_t *cloud, int index) -{ - struct list_head changed; - int res = 0; - - if(index < cloud->n_memb) { - co2_irn_t *ci = cloud->seq[index]; - int n_regs = env->co->cls->n_regs; - col_cost_pair_t *seq = alloca(n_regs * sizeof(seq[0])); - bitset_t *cols_tried = bitset_alloca(n_regs); - int done = 0; - //col_cost_pair_t *single = alloca(n_regs * sizeof(seq[0])); - - int i; - - determine_color_costs(env, ci, seq); - if(index == 0) - adjust_start_colors(env, cloud, seq); + DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}starting top-down coalesce for %+F\n", depth, ci->inh.irn)); + INIT_LIST_HEAD(&changed); + for (i = 0; i < (best_col < 0 ? n_regs : n_iter); ++i) { + col_t col = seq[i].col; + int add_cost = !is_root && col != parent_col ? ci->mst_costs : 0; -#if 0 - if(index == 0) { - col_t col = get_col(env, ci->irn); - int min_costs = INT_MAX; - int i; + int subtree_costs, sum_costs; - for(i = 0; i < n_regs; ++i) - min_costs = MIN(min_costs, seq[i].costs); + DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}%+F trying color %d\n", depth, ci->inh.irn, col)); - seq[col].costs = min_costs - 1; + unfix_subtree(ci); + INIT_LIST_HEAD(&changed); + ok = change_color_single(env, ci->inh.irn, col, &changed, depth); + if (ok) { + materialize_coloring(&changed); + ci->inh.fixed = 1; } -#endif - - qsort(seq, n_regs, sizeof(seq[0]), col_cost_pair_lt); -#if 0 - if(index == cloud->n_memb - 1) { - for(i = 0; i < n_regs; ++i) - if(seq[i].costs >= 0) - seq[i].costs = INT_MAX; + else + continue; + + for (j = 0; j < ci->mst_n_childs; ++j) { + co2_cloud_irn_t *child = ci->mst_childs[j]; + ok = coalesce_top_down(child, j, depth + 1) >= 0; + if (ok) + child->inh.fixed = 1; + else + break; } -#endif - - - for(i = 0; i < n_regs && !done; ++i) { - col_t col = seq[i].col; - int costs = seq[i].costs; - int ok; - - /* - if all affinity neighbors fixed, - try only color changes to affinity colors. - all other colors do no good. - */ - DB((env->dbg, LEVEL_2, "\t%2{firm:indent}trying %+F index %d for color %d\n", index, ci->irn, index, col)); - if(INFEASIBLE(costs)) { - DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}-> color is infeasible due to %s\n", index, flag_str(seq[i].flags))); - break; - } + /* If the subtree could not be colored, we have to try another color. */ + if (!ok) + continue; - bitset_set(cols_tried, col); - INIT_LIST_HEAD(&changed); - ok = change_color_single(env, ci->irn, col, &changed, 0); - DB((env->dbg, LEVEL_2, "\t%2{firm:indent}-> %s\n", index, ok ? "ok" : "failed")); - - /* if we succeeded changing the color, we will figure out the next node. */ - if(ok) { - int finish; - - /* materialize the coloring and fix the node's color. */ - ci->fixed = 1; - - /* process the next nodes. if the function returns one, we found an optimal coloring already, so get out. */ - finish = process_node(env, cloud, index + 1); - - /* if this is the last node in the coloring sequence, examine the coloring */ - if(index == cloud->n_memb - 1) { - examine_cloud_coloring(env, cloud); - DB((env->dbg, LEVEL_2, "\t%2{firm:indent}-> current best coloring %d\n", index, cloud->best_costs)); - if(cloud->best_costs == cloud->inevit) { - done = 1; - res = 1; - } - } - - /* unfix the node. */ - reject_coloring(&changed); - ci->fixed = 0; - - if(finish || color_change_balance(env, ci, cols_tried) <= 0) { - res = finish; - done = 1; - } - } + subtree_costs = examine_subtree_coloring(ci, col); + sum_costs = subtree_costs + add_cost; + DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}-> %+F costing %d + %d is ok.\n", depth, ci->inh.irn, subtree_costs, add_cost)); + if (sum_costs < best_col_costs) { + best_col = col; + best_col_costs = sum_costs; + ci->col_costs[col] = subtree_costs; } - } - return res; -} - -static co2_irn_t **get_neighb_arr(co2_t *env, co2_irn_t *ci, co2_irn_t **nbs) -{ - int i; - neighb_t *n; - - i = 0; - co_gs_foreach_neighb(ci->aff, n) { - nbs[i++] = get_co2_irn(env, n->irn); + if (sum_costs == 0) + break; } - qsort(nbs, ci->aff->degree, sizeof(nbs[0]), co2_irn_cmp); - return nbs; -} - -static void determine_coloring_sequence(co2_t *env, co2_cloud_t *cloud) -{ - pdeq *q = new_pdeq1(cloud->master); - bitset_t *seen = bitset_malloc(get_irg_last_idx(env->co->irg)); - co2_irn_t **nbs = alloca(cloud->max_degree * sizeof(nbs[0])); - int i, j; - - j = 0; - bitset_set(seen, get_irn_idx(cloud->master->irn)); - while(!pdeq_empty(q)) { - co2_irn_t *curr = pdeq_getl(q); - - cloud->seq[j++] = curr; - get_neighb_arr(env, curr, nbs); - - for(i = 0; i < curr->aff->degree; ++i) { - co2_irn_t *ni = nbs[i]; - int idx = get_irn_idx(ni->irn); - if(!bitset_is_set(seen, idx)) { - pdeq_putr(q, ni); - bitset_set(seen, idx); - } - } + if (!is_root) { + int *front = FRONT_BASE(ci->mst_parent, parent_col); + front[child_nr] = best_col; } - del_pdeq(q); - bitset_free(seen); + return best_col; } static void populate_cloud(co2_t *env, co2_cloud_t *cloud, affinity_node_t *a, int curr_costs) { - be_ifg_t *ifg = env->co->cenv->ifg; - co2_irn_t *ci = get_co2_irn(env, a->irn); - int costs = 0; - neighb_t *n; + be_ifg_t *ifg = env->co->cenv->ifg; + co2_cloud_irn_t *ci = get_co2_cloud_irn(env, a->irn); + int costs = 0; - if(ci->visited >= env->visited) + if (ci->cloud) return; /* mark the node as visited and add it to the cloud. */ - ci->visited = env->visited; ci->cloud = cloud; list_add(&ci->cloud_list, &cloud->members_head); - DB((env->dbg, LEVEL_3, "%+F\n", ci->irn)); + DB((env->dbg, LEVEL_2, "\t%+F\n", ci->inh.irn)); /* determine the nodes costs */ co_gs_foreach_neighb(a, n) { costs += n->costs; - DB((env->dbg, LEVEL_3, "\t%+F\n", n->irn)); - if(be_ifg_connected(ifg, a->irn, n->irn)) + DB((env->dbg, LEVEL_3, "\t\tneigh %+F cost %d\n", n->irn, n->costs)); + if (be_ifg_connected(ifg, a->irn, n->irn)) cloud->inevit += n->costs; } /* add the node's cost to the total costs of the cloud. */ ci->costs = costs; cloud->costs += costs; - cloud->max_degree = MAX(cloud->max_degree, ci->aff->degree); + cloud->n_constr += is_constrained(env, &ci->inh); + cloud->freedom += bitset_popcount(get_adm(env, &ci->inh)); + cloud->max_degree = MAX(cloud->max_degree, ci->inh.aff->degree); cloud->n_memb++; /* If this is the heaviest node in the cloud, set it as the cloud's master. */ - if(costs >= curr_costs) { - cloud->master = ci; + if (costs >= curr_costs) { curr_costs = costs; + cloud->master = ci; } /* add all the neighbors of the node to the cloud. */ @@ -954,272 +809,291 @@ static void populate_cloud(co2_t *env, co2_cloud_t *cloud, affinity_node_t *a, i } } -static void init_cloud(co2_t *env, co2_cloud_t *cloud, affinity_node_t *a) +static co2_cloud_t *new_cloud(co2_t *env, affinity_node_t *a) { + co2_cloud_t *cloud = OALLOC(&env->obst, co2_cloud_t); + int i; + + DBG((env->dbg, LEVEL_2, "new cloud with %+F\n", a->irn)); + memset(cloud, 0, sizeof(cloud[0])); + INIT_LIST_HEAD(&cloud->members_head); + INIT_LIST_HEAD(&cloud->list); + list_add(&cloud->list, &env->cloud_head); + cloud->best_costs = INT_MAX; + cloud->env = env; env->visited++; populate_cloud(env, cloud, a, 0); + cloud->freedom = (cloud->n_memb * env->n_regs) / cloud->freedom; - cloud->best_cols = phase_alloc(&env->ph, cloud->n_memb * sizeof(cloud->best_cols[0])); - cloud->seq = phase_alloc(&env->ph, cloud->n_memb * sizeof(cloud->seq[0])); - env->visited++; - cloud->seq[0] = cloud->master; - determine_coloring_sequence(env, cloud); + /* Also allocate space for the node sequence and compute that sequence. */ + cloud->seq = OALLOCN(&env->obst, co2_cloud_irn_t*, cloud->n_memb); + + i = 0; + list_for_each_entry(co2_cloud_irn_t, ci, &cloud->members_head, cloud_list) { + ci->index = i; + cloud->seq[i++] = ci; + } + DBG((env->dbg, LEVEL_2, "cloud cost %d, freedom %f\n", cloud->costs, cloud->freedom)); + + return cloud; } -static void process_cloud(co2_t *env, co2_cloud_t *cloud, int nr) +static void apply_coloring(co2_cloud_irn_t *ci, col_t col, int depth) { - struct list_head changed; + const ir_node *irn = ci->inh.irn; + int *front = FRONT_BASE(ci, col); int i; + struct list_head changed; - /* initialize the best coloring. */ - examine_cloud_coloring(env, cloud); + INIT_LIST_HEAD(&changed); - DB((env->dbg, LEVEL_1, "\nnew cloud\nall costs %d, initial costs %d, inevit %d\n", cloud->costs, cloud->best_costs, cloud->inevit)); - for(i = 0; i < cloud->n_memb; ++i) { - co2_irn_t *ci = cloud->seq[i]; - DB((env->dbg, LEVEL_1, "\tmember %+F cost %d col %d\n", ci->irn, ci->costs, get_col(env, ci->irn))); - } + DBG((ci->cloud->env->dbg, LEVEL_2, "%2{firm:indent}setting %+F to %d\n", depth, irn, col)); + change_color_single(ci->cloud->env, irn, col, &changed, depth); + materialize_coloring(&changed); - process_node(env, cloud, 0); - DB((env->dbg, LEVEL_1, "final coloring costs %d\n", cloud->best_costs)); + for (i = 0; i < ci->mst_n_childs; ++i) { + apply_coloring(ci->mst_childs[i], front[i], depth + 1); + } +} - /* re-try the best coloring. */ - INIT_LIST_HEAD(&changed); - for(i = 0; i < cloud->n_memb; ++i) { - co2_irn_t *ci = cloud->seq[i]; - col_t col = cloud->best_cols[i]; +static co2_cloud_irn_t *find_mst_root(co2_cloud_irn_t *ci) +{ + while (ci != ci->mst_parent) + ci = ci->mst_parent; + return ci; +} - int ok; - DB((env->dbg, LEVEL_2, "\tsetting %+F to %d\n", ci->irn, col)); - ok = change_color_single(env, ci->irn, col, &changed, 0); - assert(ok); - ci->fixed = 1; - } - materialize_coloring(&changed); +static void process_cloud(co2_cloud_t *cloud) +{ + co2_t *env = cloud->env; + int n_regs = env->n_regs; + int n_edges = 0; + int *mst_edges = XMALLOCNZ(int, cloud->n_memb * cloud->n_memb); + pdeq *q; - { - co2_irn_t *ci; - int some_fixed = 0; - for(ci = env->touched; ci; ci = ci->touched_next) { - if(ci->tmp_fixed) { - some_fixed = 1; - ir_printf("%+F is still temp fixed\n", ci->irn); + edge_t *edges; + int i; + int best_col; + + /* Collect all edges in the cloud on an obstack and sort the increasingly */ + obstack_init(&cloud->obst); + for (i = 0; i < cloud->n_memb; ++i) { + co2_cloud_irn_t *ci = cloud->seq[i]; + + co_gs_foreach_neighb(ci->inh.aff, n) { + co2_cloud_irn_t *ni = get_co2_cloud_irn(cloud->env, n->irn); + if (ci->index < ni->index) { + edge_t e; + e.src = ci; + e.tgt = ni; + e.costs = n->costs; + obstack_grow(&cloud->obst, &e, sizeof(e)); + n_edges++; } } - assert(!some_fixed); } - - { - char buf[256]; - FILE *f; - - ir_snprintf(buf, sizeof(buf), "ifg_%F_%s_cloud_%d.dot", env->co->irg, env->co->cls->name, nr); - if(f = fopen(buf, "wt")) { - be_ifg_dump_dot(env->co->cenv->ifg, env->co->irg, f, &ifg_dot_cb, env); - fclose(f); + edges = (edge_t*)obstack_finish(&cloud->obst); + qsort(edges, n_edges, sizeof(edges[0]), cmp_edges); + + /* Compute the maximum spanning tree using Kruskal/Union-Find */ + DBG((env->dbg, LEVEL_2, "computing spanning tree of cloud with master %+F\n", cloud->master->inh.irn)); + for (i = 0; i < n_edges; ++i) { + edge_t *e = &edges[i]; + co2_cloud_irn_t *rs = find_mst_root(e->src); + co2_cloud_irn_t *rt = find_mst_root(e->tgt); + + /* if the union/find roots are different */ + if (rs != rt) { + int si = e->src->index; + int ti = e->tgt->index; + + /* unify the sets */ + rs->mst_parent = rt; + DBG((env->dbg, LEVEL_2, "\tadding edge %+F -- %+F cost %d\n", rs->inh.irn, rt->inh.irn, e->costs)); + + /* this edge is in the MST, so set it in the bitset. */ + mst_edges[si * cloud->n_memb + ti] = e->costs; + mst_edges[ti * cloud->n_memb + si] = e->costs; } } + obstack_free(&cloud->obst, edges); + + cloud->master->mst_parent = cloud->master; + cloud->mst_root = cloud->master; + q = new_pdeq1(cloud->master); + while (!pdeq_empty(q)) { + co2_cloud_irn_t *ci = (co2_cloud_irn_t*)pdeq_getl(q); + int ofs = ci->index * cloud->n_memb; + int end = ofs + cloud->n_memb; + int i; -} - -static void process(co2_t *env) -{ - affinity_node_t *a; - co2_cloud_t *pos; - co2_cloud_t **clouds; - int n_clouds; - int i; - int init_costs = 0; - int all_costs = 0; - int final_costs = 0; - + ci->mst_n_childs = 0; + for (i = ofs; i < end; ++i) { + if (mst_edges[i] != 0) { + int other = i - ofs; + co2_cloud_irn_t *child = cloud->seq[i - ofs]; - n_clouds = 0; - co_gs_foreach_aff_node(env->co, a) { - co2_irn_t *ci = get_co2_irn(env, a->irn); + /* put the child to the worklist */ + pdeq_putr(q, child); - if(!ci->cloud) { - co2_cloud_t *cloud = new_cloud(env); + /* make ci the parent of the child and add the child to the children array of the parent */ + child->mst_parent = ci; + child->mst_costs = mst_edges[i]; + ci->mst_n_childs++; + obstack_ptr_grow(&cloud->obst, child); - init_cloud(env, cloud, a); - list_add(&cloud->list, &env->cloud_head); - n_clouds++; + mst_edges[other * cloud->n_memb + ci->index] = 0; + mst_edges[i] = 0; + } } - } - i = 0; - clouds = xmalloc(n_clouds * sizeof(clouds[0])); - list_for_each_entry(co2_cloud_t, pos, &env->cloud_head, list) - clouds[i++] = pos; - qsort(clouds, n_clouds, sizeof(clouds[0]), cmp_clouds); - - for(i = 0; i < n_clouds; ++i) { - init_costs += cloud_costs(env, clouds[i]); - process_cloud(env, clouds[i], i); - all_costs += clouds[i]->costs; - final_costs += clouds[i]->best_costs; + obstack_ptr_grow(&cloud->obst, NULL); + ci->mst_childs = (co2_cloud_irn_t**)obstack_finish(&cloud->obst); } + del_pdeq(q); + free(mst_edges); - DB((env->dbg, LEVEL_1, "all costs: %d, init costs: %d, final costs: %d\n", all_costs, init_costs, final_costs)); - - xfree(clouds); -} - -static void writeback_colors(co2_t *env) -{ - const arch_env_t *aenv = env->co->aenv; - co2_irn_t *irn; - for(irn = env->touched; irn; irn = irn->touched_next) { - const arch_register_t *reg = arch_register_for_index(env->co->cls, irn->orig_col); - arch_set_irn_register(aenv, irn->irn, reg); + DBG((env->dbg, LEVEL_3, "mst:\n")); + for (i = 0; i < cloud->n_memb; ++i) { + DEBUG_ONLY(co2_cloud_irn_t *ci = cloud->seq[i];) + DBG((env->dbg, LEVEL_3, "\t%+F -> %+F\n", ci->inh.irn, ci->mst_parent->inh.irn)); } -} - -/* - ___ _____ ____ ____ ___ _____ ____ _ - |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _ - | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` | - | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| | - |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, | - |_| |___/ -*/ -static const char *get_dot_color_name(int col) -{ - static const char *names[] = { - "blue", - "red", - "green", - "yellow", - "cyan", - "magenta", - "orange", - "chocolate", - "beige", - "navy", - "darkgreen", - "darkred", - "lightPink", - "chartreuse", - "lightskyblue", - "linen", - "pink", - "lightslateblue", - "mintcream", - "red", - "darkolivegreen", - "mediumblue", - "mistyrose", - "salmon", - "darkseagreen", - "mediumslateblue" - "moccasin", - "tomato", - "forestgreen", - "darkturquoise", - "palevioletred" - }; + for (i = 0; i < cloud->n_memb; ++i) { + co2_cloud_irn_t *ci = cloud->seq[i]; + int n_childs = ci->mst_n_childs; + int j; - return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white"; -} + ci->col_costs = OALLOCNZ(&cloud->obst, int, n_regs); + ci->tmp_coloring = OALLOCNZ(&cloud->obst, col_cost_pair_t, n_regs); + ci->fronts = OALLOCNZ(&cloud->obst, int, n_regs * n_childs); + ci->color_badness = OALLOCNZ(&cloud->obst, int, n_regs); -static const char *get_dot_shape_name(co2_t *env, co2_irn_t *ci) -{ - arch_register_req_t req; + for (j = 0; j < env->n_regs; j++) + ci->col_costs[j] = INT_MAX; + } - arch_get_register_req(env->co->aenv, &req, ci->irn, BE_OUT_POS(0)); - if(arch_register_req_is(&req, limited)) - return "diamond"; + determine_color_badness(cloud->mst_root, 0); + best_col = coalesce_top_down(cloud->mst_root, -1, 0); + unfix_subtree(cloud->mst_root); + apply_coloring(cloud->mst_root, best_col, 0); - if(ci->fixed) - return "rectangle"; + /* The coloring should represent the one with the best costs. */ + //materialize_coloring(&changed); + DBG((env->dbg, LEVEL_2, "\tbest coloring for root %+F was %d costing %d\n", + cloud->mst_root->inh.irn, best_col, examine_subtree_coloring(cloud->mst_root, best_col))); - if(ci->tmp_fixed) - return "hexagon"; + /* Fix all nodes in the cloud. */ + for (i = 0; i < cloud->n_memb; ++i) + cloud->seq[i]->inh.fixed = 1; - return "ellipse"; + /* Free all space used while optimizing this cloud. */ + obstack_free(&cloud->obst, NULL); } -static void ifg_dump_graph_attr(FILE *f, void *self) +static int cloud_costs(co2_cloud_t *cloud) { - fprintf(f, "overlay=false"); -} + int i, costs = 0; + + for (i = 0; i < cloud->n_memb; ++i) { + co2_irn_t *ci = (co2_irn_t *) cloud->seq[i]; + col_t col = get_col(cloud->env, ci->irn); + co_gs_foreach_neighb(ci->aff, n) { + col_t n_col = get_col(cloud->env, n->irn); + costs += col != n_col ? n->costs : 0; + } + } -static int ifg_is_dump_node(void *self, ir_node *irn) -{ - co2_t *env = self; - return !arch_irn_is(env->co->aenv, irn, ignore); + return costs / 2; } -static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn) +static void writeback_colors(co2_t *env) { - co2_t *env = self; - co2_irn_t *ci = get_co2_irn(env, irn); + co2_irn_t *irn; - ir_fprintf(f, "label=\"%+F,%d\" style=filled color=%s shape=%s", irn, ci->costs, - get_dot_color_name(get_col(env, irn)), get_dot_shape_name(env, ci)); + for (irn = env->touched; irn; irn = irn->touched_next) { + const arch_register_t *reg = arch_register_for_index(env->co->cls, irn->orig_col); + arch_set_irn_register((ir_node*)irn->irn, reg); + } } -static void ifg_dump_at_end(FILE *file, void *self) +static void process(co2_t *env) { - co2_t *env = self; - affinity_node_t *a; + co2_cloud_t **clouds; + int n_clouds; + int i; + int init_costs = 0; + int all_costs = 0; + int final_costs = 0; + n_clouds = 0; co_gs_foreach_aff_node(env->co, a) { - int idx = get_irn_idx(a->irn); - neighb_t *n; + co2_cloud_irn_t *ci = get_co2_cloud_irn(env, a->irn); - co_gs_foreach_neighb(a, n) { - int nidx = get_irn_idx(n->irn); - - if(idx < nidx) { - const char *style = get_col(env, a->irn) == get_col(env, n->irn) ? "dashed" : "dotted"; - fprintf(file, "\tn%d -- n%d [label=\"%d\" style=%s weight=0.01];\n", idx, nidx, n->costs, style); - } + if (!ci->cloud) { + new_cloud(env, a); + n_clouds++; } } -} + i = 0; + clouds = XMALLOCN(co2_cloud_t*, n_clouds); + list_for_each_entry(co2_cloud_t, pos, &env->cloud_head, list) + clouds[i++] = pos; + qsort(clouds, n_clouds, sizeof(clouds[0]), cmp_clouds_gt); -static be_ifg_dump_dot_cb_t ifg_dot_cb = { - ifg_is_dump_node, - ifg_dump_graph_attr, - ifg_dump_node_attr, - NULL, - NULL, - ifg_dump_at_end -}; + for (i = 0; i < n_clouds; ++i) { + init_costs += cloud_costs(clouds[i]); -void co_solve_heuristic_new(copy_opt_t *co) + /* Process the cloud. */ + process_cloud(clouds[i]); + + all_costs += clouds[i]->costs; + final_costs += cloud_costs(clouds[i]); + } + + DB((env->dbg, LEVEL_1, "all costs: %d, init costs: %d, final costs: %d\n", all_costs, init_costs, final_costs)); + + xfree(clouds); +} + +static int co_solve_heuristic_new(copy_opt_t *co) { co2_t env; - FILE *f; - phase_init(&env.ph, "co2", co->cenv->birg->irg, sizeof(co2_irn_t), PHASE_DEFAULT_GROWTH, co2_irn_init); + ir_nodemap_init(&env.map, co->irg); + obstack_init(&env.obst); env.touched = NULL; env.visited = 0; env.co = co; - env.ignore_regs = bitset_alloca(co->cls->n_regs); - arch_put_non_ignore_regs(co->aenv, co->cls, env.ignore_regs); - bitset_flip_all(env.ignore_regs); - be_abi_put_ignore_regs(co->cenv->birg->abi, co->cls, env.ignore_regs); + env.n_regs = co->cls->n_regs; + env.allocatable_regs = bitset_alloca(co->cls->n_regs); + be_put_allocatable_regs(co->cenv->irg, co->cls, env.allocatable_regs); FIRM_DBG_REGISTER(env.dbg, "firm.be.co2"); INIT_LIST_HEAD(&env.cloud_head); - if(f = be_chordal_open(co->cenv, "ifg_before_", "dot")) { - be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &env); - fclose(f); - } - process(&env); - if(f = be_chordal_open(co->cenv, "ifg_after_", "dot")) { - be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &env); - fclose(f); - } - writeback_colors(&env); - phase_free(&env.ph); + obstack_free(&env.obst, NULL); + ir_nodemap_destroy(&env.map); + return 0; +} + +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur2) +void be_init_copyheur2(void) +{ + lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra"); + lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal"); + lc_opt_entry_t *co2_grp = lc_opt_get_grp(chordal_grp, "co2"); + + static co_algo_info copyheur = { + co_solve_heuristic_new, 0 + }; + + lc_opt_add_table(co2_grp, options); + be_register_copyopt("heur2", ©heur); }