#include "irprintf.h"
#include "irtools.h"
+#include "bemodule.h"
#include "beabi.h"
#include "benode_t.h"
#include "becopyopt.h"
#define DUMP_BEFORE 1
#define DUMP_AFTER 2
#define DUMP_CLOUD 4
+#define DUMP_ALL 2 * DUMP_CLOUD - 1
-static int dump_flags = 0;
+static unsigned dump_flags = 0;
+static int subtree_iter = 4;
+static int max_depth = 20;
+static double constr_factor = 0.9;
/* Options using libcore */
#ifdef WITH_LIBCORE
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "before", DUMP_BEFORE },
- { "after", DUMP_AFTER },
- { "cloud", DUMP_CLOUD },
- { "all", 2 * DUMP_CLOUD - 1 },
+ { "after", DUMP_AFTER },
+ { "cloud", DUMP_CLOUD },
+ { "all", DUMP_ALL },
{ NULL, 0 }
};
};
static const lc_opt_table_entry_t options[] = {
- LC_OPT_ENT_ENUM_MASK("dump", "dump ifg before, after or after each cloud", &dump_var),
+ 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),
{ NULL }
};
-void be_co2_register_options(lc_opt_entry_t *grp)
+void be_init_copyheur2(void)
{
- lc_opt_entry_t *co2_grp = lc_opt_get_grp(grp, "co2");
+ 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");
+
lc_opt_add_table(co2_grp, options);
}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur2);
#endif
/*
col_t tmp_col;
col_t orig_col;
int last_color_change;
- unsigned fixed : 1;
- unsigned tmp_fixed : 1;
+ bitset_t *adm_cache;
+ unsigned fixed : 1;
+ unsigned tmp_fixed : 1;
+ unsigned is_constrained : 1;
struct list_head changed_list;
};
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;
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 list;
};
+typedef struct {
+ co2_cloud_irn_t *src, *tgt;
+ int costs;
+} edge_t;
+
#define FRONT_BASE(ci,col) ((ci)->fronts + col * (ci)->mst_n_childs)
#define get_co2_irn(co2, irn) ((co2_irn_t *) phase_get_or_set_irn_data(&co2->ph, irn))
return ci;
}
+#define CLOUD_WEIGHT(c) ((1 - constr_factor) * (c)->costs + constr_factor * (c)->freedom)
static int cmp_clouds_gt(const void *a, const void *b)
{
- const co2_cloud_t **p = a;
- const co2_cloud_t **q = b;
- int c = (*p)->costs;
- int d = (*q)->costs;
- return CMP(d, c);
+ const co2_cloud_t * const *p = a;
+ const co2_cloud_t * const *q = b;
+ double c = CLOUD_WEIGHT(*p);
+ double d = CLOUD_WEIGHT(*q);
+ return QSORT_CMP(d, c);
}
/**
const col_cost_pair_t *q = b;
int c = p->costs;
int d = q->costs;
- return CMP(c, d);
+ return QSORT_CMP(c, d);
+}
+
+int cmp_edges(const void *a, const void *b)
+{
+ const edge_t *p = a;
+ const edge_t *q = b;
+ return QSORT_CMP(q->costs, p->costs);
}
static col_t get_col(co2_t *env, ir_node *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) {
+ arch_register_req_t req;
+ ci->adm_cache = bitset_obstack_alloc(phase_obst(&env->ph), env->n_regs);
+ 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, ci->adm_cache);
+ ci->is_constrained = 1;
+ }
+ else {
+ bitset_copy(ci->adm_cache, env->ignore_regs);
+ bitset_flip_all(ci->adm_cache);
+ }
}
+ 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 INLINE int is_constrained(co2_t *env, co2_irn_t *ci)
+{
+ if(!ci->adm_cache)
+ get_adm(env, ci);
+ return ci->is_constrained;
+}
+
static void incur_constraint_costs(co2_t *env, ir_node *irn, col_cost_pair_t *col_costs, int costs)
{
bitset_t *aux = bitset_alloca(env->co->cls->n_regs);
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;
col_costs[col].costs = add_saturated(col_costs[col].costs, 8 * be_ifg_degree(ifg, pos));
}
}
+ be_ifg_neighbours_break(ifg, it);
/* Set the costs to infinity for each color which is not allowed at this node. */
bitset_foreach(forb, elm) {
}
-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;
seq[i].costs = INT_MAX;
}
+ assert(is_color_admissible(env, ci, col));
seq[col].col = 0;
seq[0].col = col;
seq[0].costs = 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 recolor(co2_t *env, ir_node *irn, col_cost_pair_t *col_list, struct list_head *parent_changed, int depth)
{
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;
+ 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;
break;
}
}
+ be_ifg_neighbours_break(ifg, it);
/*
We managed to assign the target color to all neighbors, so from the perspective
list_add(&ci->changed_list, parent_changed);
}
- DB((env->dbg, LEVEL_3, "\t\t%2{firm:indent}ok\n", depth));
- return 1;
+ res = 1;
+ goto end;
}
- if(!color_is_fix(env, irn)) {
+ 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]));
/* 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\t%2{firm:indent}color %d %s for %+F\n", depth, 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;
}
-static void front_inval_color(co2_cloud_irn_t *ci, col_t col)
-{
- int *base = FRONT_BASE(ci, col);
- memset(base, -1, ci->mst_n_childs * sizeof(base[0]));
-}
-
-typedef struct {
- co2_cloud_irn_t *src, *tgt;
- int costs;
-} edge_t;
-
-int cmp_edges(const void *a, const void *b)
-{
- const edge_t *p = a;
- const edge_t *q = b;
- return CMP(p->costs, q->costs);
-}
-
-static co2_cloud_irn_t *find_mst_root(co2_cloud_irn_t *ci)
-{
- while(ci->mst_parent != ci->mst_parent)
- ci = ci->mst_parent;
- return ci;
-}
-
-
-static int cmp_parent(const void *a, const void *b)
-{
- const co2_cloud_irn_t *p = a;
- const co2_cloud_irn_t *q = b;
- return CMP(q->mst_costs, p->mst_costs);
-}
-
-static void fill_tmp_coloring(co2_cloud_irn_t *ci, col_t col)
+/**
+ * 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)
{
- int n_regs = ci->cloud->env->n_regs;
- int i, j;
+ int *front = FRONT_BASE(ci, col);
+ int cost = 0;
+ int i;
for(i = 0; i < ci->mst_n_childs; ++i) {
- co2_cloud_irn_t *c = ci->mst_childs[i];
- for(j = 0; j < n_regs; ++j) {
- int costs = c->col_costs[j];
- if(INFEASIBLE(costs))
- c->tmp_coloring[j].costs = INT_MAX;
- else {
- int add = j != (int) col ? c->mst_costs : 0;
- c->tmp_coloring[j].costs = add + costs;
- }
- c->tmp_coloring[j].col = j;
- }
- qsort(c->tmp_coloring, n_regs, sizeof(c->tmp_coloring[0]), col_cost_pair_lt);
- }
-}
-
-static void determine_start_colors(co2_cloud_irn_t *ci, col_cost_pair_t *seq)
-{
- int n_regs = ci->cloud->env->n_regs;
- bitset_t *adm = bitset_alloca(n_regs);
- int i, j;
-
- // TODO: Prefer some colors depending on the neighbors, etc.
-
- admissible_colors(ci->cloud->env, &ci->inh, adm);
- for(i = 0; i < n_regs; ++i) {
- seq[i].col = i;
+ co2_cloud_irn_t *chld = ci->mst_childs[i];
+ col_t chld_col = front[i];
- if (!bitset_is_set(adm, i))
- seq[i].costs = INT_MAX;
- else {
- seq[i].costs = 0;
- for(j = 0; j < ci->mst_n_childs; ++j) {
- co2_cloud_irn_t *child = ci->mst_childs[j];
- if (!INFEASIBLE(child->col_costs[i]))
- seq[i].costs -= ci->mst_childs[j]->col_costs[i];
- }
- }
+ cost += examine_subtree_coloring(chld, chld_col);
+ cost += col != chld_col ? chld->mst_costs : 0;
}
- qsort(seq, n_regs, sizeof(seq[0]), col_cost_pair_lt);
+ return cost;
}
-static int push_front(co2_cloud_irn_t *ci, int *front)
+/**
+ * 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 <code>badness</code> is not cleared.
+ */
+static void node_color_badness(co2_cloud_irn_t *ci, int *badness)
{
- co2_t *env = ci->cloud->env;
- int n_regs = env->n_regs;
- int min_diff = INT_MAX;
- int min_chld = -1;
- int i;
-
- for(i = 0; i < ci->mst_n_childs; ++i) {
- co2_cloud_irn_t *child = ci->mst_childs[i];
- int idx = front[i];
-
+ 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);
- if(idx + 1 < n_regs) {
- int diff = child->tmp_coloring[idx].costs - child->tmp_coloring[idx + 1].costs;
- if(diff < min_diff) {
- min_diff = diff;
- min_chld = i;
- }
- }
- }
-
- if(min_chld >= 0) {
- co2_cloud_irn_t *child = ci->mst_childs[min_chld];
- DBG((env->dbg, LEVEL_3, "\tsmallest diff with child %+F on index %d is %d\n", child->inh.irn, front[min_chld], min_diff));
- front[min_chld] += 1;
- }
+ bitset_pos_t elm;
+ ir_node *irn;
+ void *it;
- return min_chld;
-}
+ admissible_colors(env, &ci->inh, bs);
+ bitset_flip_all(bs);
+ bitset_foreach(bs, elm)
+ badness[elm] = ci->costs;
-static int color_subtree(co2_cloud_irn_t *ci, col_t col, struct list_head *changed, int depth)
-{
- int n_childs = ci->mst_n_childs;
- /*
- select the front for the given color.
- The front will determine the colors of the children.
- */
- int *front = FRONT_BASE(ci, col);
- int i, ok = 1;
+ /* Use constrained/fixed interfering neighbors to influence the color badness */
+ it = be_ifg_neighbours_iter_alloca(ifg);
+ be_ifg_foreach_neighbour(ifg, it, ir->irn, irn) {
+ co2_irn_t *ni = get_co2_irn(env, irn);
- ok = change_color_single(ci->cloud->env, ci->inh.irn, col, changed, 0);
- for(i = 0; i < n_childs && ok; ++i) {
- co2_cloud_irn_t *child = ci->mst_childs[i];
- col_t col = front[i];
+ admissible_colors(env, ni, bs);
+ if(bitset_popcnt(bs) == 1) {
+ bitset_pos_t c = bitset_next_set(bs, 0);
+ badness[c] += ci->costs;
+ }
- ok = color_subtree(child, col, changed, depth + 1);
+ else if(ni->fixed) {
+ col_t c = get_col(env, ni->irn);
+ badness[c] += ci->costs;
+ }
}
-
- return ok;
+ be_ifg_neighbours_break(ifg, it);
}
-static int try_coloring(co2_cloud_irn_t *ci, col_t col, int *front, int *initial_ok, int depth)
+/**
+ * Determine the badness of a MST subtree.
+ * The badness is written into the <code>color_badness</code> 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)
{
- co2_t *env = ci->cloud->env;
- struct list_head changed;
- int i, ok = 1;
+ co2_t *env = ci->cloud->env;
+ int i, j;
- INIT_LIST_HEAD(&changed);
- *initial_ok = ok = change_color_single(env, ci->inh.irn, col, &changed, depth + 1);
+ node_color_badness(ci, ci->color_badness);
- for (i = 0; i < ci->mst_n_childs && ok; ++i) {
+ /* 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];
- col_t tgt_col = child->tmp_coloring[front[i]].col;
+ determine_color_badness(child, depth + 1);
- ok = color_subtree(child, tgt_col, &changed, depth + 1);
+ for(j = 0; j < env->n_regs; ++j)
+ ci->color_badness[j] += child->color_badness[j];
}
- reject_coloring(&changed);
-
- return ok;
+ 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 examine_subtree_coloring(co2_cloud_irn_t *ci, col_t col)
+/**
+ * Unfix all nodes in a MST subtree.
+ */
+static void unfix_subtree(co2_cloud_irn_t *ci)
{
- int *front = FRONT_BASE(ci, col);
- int cost = 0;
int i;
- for(i = 0; i < ci->mst_n_childs; ++i) {
- co2_cloud_irn_t *chld = ci->mst_childs[i];
- col_t chld_col = front[i];
-
- cost += examine_subtree_coloring(chld, chld_col);
- cost += col != chld_col ? chld->mst_costs : 0;
- }
-
- return cost;
+ ci->inh.fixed = 0;
+ for(i = 0; i < ci->mst_n_childs; ++i)
+ unfix_subtree(ci->mst_childs[i]);
}
-static int cloud_mst_build_colorings(co2_cloud_irn_t *ci, int depth)
+static int coalesce_top_down(co2_cloud_irn_t *ci, int child_nr, int depth)
{
- co2_t *env = ci->cloud->env;
- int n_regs = env->n_regs;
- col_cost_pair_t *seq = alloca(n_regs * sizeof(seq[0]));
- int *front = alloca(ci->mst_n_childs * sizeof(front[0]));
- int best_col = -1;
- int best_cost = INT_MAX;
+ co2_t *env = ci->cloud->env;
+ col_cost_pair_t *seq = alloca(env->n_regs * sizeof(seq[0]));
+ int is_root = ci->mst_parent == ci;
+ col_t parent_col = is_root ? -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;
+ struct list_head changed;
+ int ok, i, j;
- int i;
-
- DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}build colorings: %+F\n", depth, ci->inh.irn));
-
- for (i = 0; i < ci->mst_n_childs; ++i)
- cloud_mst_build_colorings(ci->mst_childs[i], depth + 1);
+ for(i = 0; i < n_regs; ++i) {
+ int badness = ci->color_badness[i];
- for (i = 0; i < n_regs; ++i)
- ci->col_costs[i] = INT_MAX;
+ seq[i].col = i;
+ seq[i].costs = is_color_admissible(env, &ci->inh, i) ? badness : INT_MAX;
- /* Sort the children according to the cost of the affinity edge they have to the current node. */
- // qsort(child, ci->mst_n_childs, sizeof(childs[0]), cmp_parent);
+ min_badness = MIN(min_badness, badness);
+ }
- determine_start_colors(ci, seq);
- // qsort(seq, n_regs, sizeof(seq[0]), col_cost_pair_lt);
+ /* 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;
- for(i = 0; i < n_regs; ++i) {
- col_t col = seq[i].col;
- int costs = seq[i].costs;
- int done = 0;
+ /* Sort the colors. The will be processed in that ordering. */
+ qsort(seq, env->n_regs, sizeof(seq[0]), col_cost_pair_lt);
- if(INFEASIBLE(costs))
- break;
+ 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;
- /*
- Judge, if it is worthwhile trying this color.
- If another color was so good that we cannot get any better, bail out here.
- Perhaps???
- */
+ int subtree_costs, sum_costs;
DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}%+F trying color %d\n", depth, ci->inh.irn, col));
- /* This sorts the tmp_coloring array in the children according to the costs of the current color. */
- fill_tmp_coloring(ci, col);
+ 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;
+ }
- /* Initialize the front. It gives the indexes into the color tmp_coloring array. */
- memset(front, 0, ci->mst_n_childs * sizeof(front));
+ 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;
+ }
- /*
- As long as we have color configurations to try.
- We try the best ones first and get worse over and over.
- */
- while (!done) {
- int j, try_push;
-
- if (try_coloring(ci, col, front, &try_push, depth + 1)) {
- int *res_front = FRONT_BASE(ci, col);
- int costs;
-
- for(j = 0; j < ci->mst_n_childs; ++j) {
- co2_cloud_irn_t *child = ci->mst_childs[j];
- col_t col = child->tmp_coloring[front[j]].col;
- res_front[j] = col;
- }
-
- costs = examine_subtree_coloring(ci, col);
- ci->col_costs[col] = costs;
- done = 1;
-
- /* Set the current best color. */
- if(costs < best_cost) {
- best_cost = costs;
- best_col = col;
- }
- }
+ /* If the subtree could not be colored, we have to try another color. */
+ if(!ok)
+ continue;
- DBG((env->dbg, LEVEL_2, "\t%2{firm:indent}-> %s\n", depth, done ? "ok" : "failed"));
+ 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));
- /* Worsen the configuration, if that one didn't succeed. */
- if (!done)
- done = try_push ? push_front(ci, front) < 0 : 1;
+ if(sum_costs < best_col_costs) {
+ best_col = col;
+ best_col_costs = sum_costs;
+ ci->col_costs[col] = subtree_costs;
}
+
+ if(sum_costs == 0)
+ break;
}
- DBG((env->dbg, LEVEL_2, "%2{firm:indent} %+F\n", depth, ci->inh.irn));
- for(i = 0; i < n_regs; ++i)
- DBG((env->dbg, LEVEL_2, "%2{firm:indent} color %d costs %d\n", depth, i, ci->col_costs[i]));
+ if(!is_root) {
+ int *front = FRONT_BASE(ci->mst_parent, parent_col);
+ front[child_nr] = best_col;
+ }
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;
int costs = 0;
neighb_t *n;
- 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, "\t%+F\n", ci->inh.irn));
+ DB((env->dbg, LEVEL_2, "\t%+F\n", ci->inh.irn));
/* determine the nodes costs */
co_gs_foreach_neighb(a, n) {
/* add the node's cost to the total costs of the cloud. */
ci->costs = costs;
cloud->costs += costs;
+ cloud->n_constr += is_constrained(env, &ci->inh);
+ cloud->freedom += bitset_popcnt(get_adm(env, &ci->inh));
cloud->max_degree = MAX(cloud->max_degree, ci->inh.aff->degree);
cloud->n_memb++;
cloud->env = env;
env->visited++;
populate_cloud(env, cloud, a, 0);
-
- /* Allocate space for the best colors array, where the best coloring is saved. */
- // cloud->best_cols = phase_alloc(&env->ph, cloud->n_memb * sizeof(cloud->best_cols[0]));
+ cloud->freedom = (cloud->n_memb * env->n_regs) / cloud->freedom;
/* Also allocate space for the node sequence and compute that sequence. */
cloud->seq = phase_alloc(&env->ph, cloud->n_memb * sizeof(cloud->seq[0]));
ci->index = i;
cloud->seq[i++] = ci;
}
- DBG((env->dbg, LEVEL_2, "cloud cost %d\n", cloud->costs));
+ DBG((env->dbg, LEVEL_2, "cloud cost %d, freedom %f\n", cloud->costs, cloud->freedom));
return cloud;
}
-static void apply_coloring(co2_cloud_irn_t *ci, col_t col, struct list_head *changed, int depth)
+static void apply_coloring(co2_cloud_irn_t *ci, col_t col, int depth)
{
ir_node *irn = ci->inh.irn;
int *front = FRONT_BASE(ci, col);
int i, ok;
+ struct list_head changed;
+
+ INIT_LIST_HEAD(&changed);
DBG((ci->cloud->env->dbg, LEVEL_2, "%2{firm:indent}setting %+F to %d\n", depth, irn, col));
- ok = change_color_single(ci->cloud->env, irn, col, changed, depth);
- assert(ok && "Color changing may not fail while committing the coloring");
+ ok = change_color_single(ci->cloud->env, irn, col, &changed, depth);
+ // assert(ok && "Color changing may not fail while committing the coloring");
+ materialize_coloring(&changed);
- for(i = 0; i < ci->mst_n_childs; ++i) {
- apply_coloring(ci->mst_childs[i], front[i], changed, depth + 1);
+ for(i = 0; i < ci->mst_n_childs; ++i) {
+ apply_coloring(ci->mst_childs[i], front[i], depth + 1);
}
}
+static co2_cloud_irn_t *find_mst_root(co2_cloud_irn_t *ci)
+{
+ while(ci != ci->mst_parent)
+ ci = ci->mst_parent;
+ return ci;
+}
+
+
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 = xmalloc(cloud->n_memb * cloud->n_memb * sizeof(mst_edges[0]));
+ pdeq *q;
- struct list_head changed;
edge_t *edges;
int i;
int best_col;
+ memset(mst_edges, 0, cloud->n_memb * cloud->n_memb * sizeof(mst_edges[0]));
+
/* 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) {
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 *p_src = find_mst_root(e->src);
- co2_cloud_irn_t *p_tgt = find_mst_root(e->tgt);
+ 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;
- if(p_src != p_tgt) {
- p_tgt->mst_n_childs++;
- p_src->mst_parent = p_tgt;
- p_src->mst_costs = e->costs;
+ /* 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));
- DBG((env->dbg, LEVEL_2, "\tadding edge %+F -- %+F cost %d\n", p_src->inh.irn, p_tgt->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 = pdeq_getl(q);
+ int ofs = ci->index * cloud->n_memb;
+ int end = ofs + cloud->n_memb;
+ int i;
+
+ 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];
+
+ /* put the child to the worklist */
+ pdeq_putr(q, child);
+
+ /* 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);
+
+ mst_edges[other * cloud->n_memb + ci->index] = 0;
+ mst_edges[i] = 0;
+ }
+ }
+
+ obstack_ptr_grow(&cloud->obst, NULL);
+ ci->mst_childs = obstack_finish(&cloud->obst);
+ }
+ del_pdeq(q);
+ free(mst_edges);
+
+
+ DBG((env->dbg, LEVEL_3, "mst:\n"));
for(i = 0; i < cloud->n_memb; ++i) {
- co2_cloud_irn_t *ci = cloud->seq[i];
- ci->mst_childs = obstack_alloc(&cloud->obst, ci->mst_n_childs * sizeof(ci->mst_childs));
- ci->col_costs = obstack_alloc(&cloud->obst, env->n_regs * sizeof(ci->col_costs[0]));
- ci->tmp_coloring = obstack_alloc(&cloud->obst, env->n_regs * sizeof(ci->tmp_coloring[0]));
- ci->fronts = obstack_alloc(&cloud->obst, env->n_regs * ci->mst_n_childs * sizeof(ci->fronts[0]));
- memset(ci->col_costs, 0, env->n_regs * sizeof(ci->col_costs[0]));
- memset(ci->tmp_coloring, 0, env->n_regs * sizeof(ci->tmp_coloring[0]));
- memset(ci->fronts, 0, env->n_regs * ci->mst_n_childs * sizeof(ci->fronts[0]));
-
- ci->mst_n_childs = 0;
+ 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));
}
- /* build the child arrays in the nodes */
for(i = 0; i < cloud->n_memb; ++i) {
co2_cloud_irn_t *ci = cloud->seq[i];
- if(ci->mst_parent != ci)
- ci->mst_parent->mst_childs[ci->mst_parent->mst_n_childs++] = ci;
- else {
- cloud->mst_root = ci;
- cloud->mst_costs = 0;
- }
- }
+ int n_childs = ci->mst_n_childs;
+ int j;
+
+ ci->col_costs = obstack_alloc(&cloud->obst, n_regs * sizeof(ci->col_costs[0]));
+ ci->tmp_coloring = obstack_alloc(&cloud->obst, n_regs * sizeof(ci->tmp_coloring[0]));
+ ci->fronts = obstack_alloc(&cloud->obst, n_regs * n_childs * sizeof(ci->fronts[0]));
+ ci->color_badness = obstack_alloc(&cloud->obst, n_regs * sizeof(ci->fronts[0]));
+ memset(ci->color_badness, 0, n_regs * sizeof(ci->color_badness[0]));
+ memset(ci->col_costs, 0, n_regs * sizeof(ci->col_costs[0]));
+ memset(ci->tmp_coloring, 0, n_regs * sizeof(ci->tmp_coloring[0]));
+ memset(ci->fronts, 0, n_regs * n_childs * sizeof(ci->fronts[0]));
- /* Compute the "best" colorings. */
- best_col = cloud_mst_build_colorings(cloud->mst_root, 0);
+ for(j = 0; j < env->n_regs; j++)
+ ci->col_costs[j] = INT_MAX;
- for(i = 0; i < env->n_regs; ++i) {
- int c;
- c = examine_subtree_coloring(cloud->mst_root, i);
- DBG((env->dbg, LEVEL_2, "color %d costs %d\n", i, c));
}
- /* Apply the coloring for the best color in the root node and fix all nodes in this cloud */
- INIT_LIST_HEAD(&changed);
- apply_coloring(cloud->mst_root, best_col, &changed, 0);
- materialize_coloring(&changed);
+ 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);
+
+ /* 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)));
+
+ /* Fix all nodes in the cloud. */
for(i = 0; i < cloud->n_memb; ++i)
cloud->seq[i]->inh.fixed = 1;
+ /* Free all space used while optimizing this cloud. */
obstack_free(&cloud->obst, NULL);
}
co2_cloud_irn_t *ci = get_co2_cloud_irn(env, a->irn);
if(!ci->cloud) {
- co2_cloud_t *cloud = new_cloud(env, a);
+ new_cloud(env, a);
n_clouds++;
}
}
for(i = 0; i < n_clouds; ++i) {
init_costs += cloud_costs(clouds[i]);
+
+ /* Process the cloud. */
process_cloud(clouds[i]);
+
all_costs += clouds[i]->costs;
final_costs += cloud_costs(clouds[i]);
FILE *f;
ir_snprintf(buf, sizeof(buf), "ifg_%F_%s_cloud_%d.dot", env->co->irg, env->co->cls->name, i);
- if(f = fopen(buf, "wt")) {
+ f = fopen(buf, "wt");
+ if(f != NULL) {
be_ifg_dump_dot(env->co->cenv->ifg, env->co->irg, f, &ifg_dot_cb, env);
fclose(f);
}
co2_irn_t *ci = get_co2_irn(env, irn);
int peri = 1;
+ char buf[128] = "";
+
if(ci->aff) {
co2_cloud_irn_t *cci = (void *) ci;
if (cci->cloud && cci->cloud->mst_root == cci)
peri = 2;
+
+ if(cci->cloud && cci->cloud->mst_root)
+ ir_snprintf(buf, sizeof(buf), "%+F", cci->cloud->mst_root->inh.irn);
}
- ir_fprintf(f, "label=\"%+F\" style=filled peripheries=%d color=%s shape=%s", irn, peri,
+ ir_fprintf(f, "label=\"%+F%s\" style=filled peripheries=%d color=%s shape=%s", irn, buf, peri,
get_dot_color_name(get_col(env, irn)), get_dot_shape_name(env, ci));
}
int idx = get_irn_idx(a->irn);
neighb_t *n;
+ if(ai->mst_parent != ai)
+ fprintf(file, "\tn%d -- n%d [style=dotted color=blue arrowhead=normal];\n", idx, get_irn_idx(ai->mst_parent->inh.irn));
+
co_gs_foreach_neighb(a, n) {
int nidx = get_irn_idx(n->irn);
co2_cloud_irn_t *ci = get_co2_cloud_irn(env, n->irn);
};
-void co_solve_heuristic_new(copy_opt_t *co)
+int co_solve_heuristic_new(copy_opt_t *co)
{
+ char buf[256];
co2_t env;
FILE *f;
INIT_LIST_HEAD(&env.cloud_head);
if(dump_flags & DUMP_BEFORE) {
- if(f = be_chordal_open(co->cenv, "ifg_before_", "dot")) {
+ ir_snprintf(buf, sizeof(buf), "ifg_%F_%s_before.dot", co->irg, co->cls->name);
+ f = fopen(buf, "wt");
+ if (f != NULL) {
be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &env);
fclose(f);
}
process(&env);
if(dump_flags & DUMP_AFTER) {
- if(f = be_chordal_open(co->cenv, "ifg_after_", "dot")) {
+ ir_snprintf(buf, sizeof(buf), "ifg_%F_%s_after.dot", co->irg, co->cls->name);
+ f = fopen(buf, "wt");
+ if (f != NULL) {
be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &env);
fclose(f);
}
writeback_colors(&env);
phase_free(&env.ph);
+ return 0;
}