#include <malloc.h>
#endif
-#include <libcore/lc_timing.h>
-
#include "xmalloc.h"
#include "debug.h"
#include "pmap.h"
#include "bearch.h"
#include "beutil.h"
+#include "beifg_t.h"
#include "becopyopt_t.h"
#include "becopystat.h"
+#undef QUICK_AND_DIRTY_HACK
+
+/******************************************************************************
+ _____ _
+ / ____| | |
+ | | __ ___ _ __ ___ _ __ __ _| |
+ | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
+ | |__| | __/ | | | __/ | | (_| | |
+ \_____|\___|_| |_|\___|_| \__,_|_|
+
+ ******************************************************************************/
static firm_dbg_module_t *dbg = NULL;
+void be_copy_opt_init(void) {
+}
+
+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;
+
+ FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
+
+ co = xcalloc(1, sizeof(*co));
+ co->cenv = chordal_env;
+ co->aenv = chordal_env->birg->main_env->arch_env;
+ co->irg = chordal_env->irg;
+ co->cls = chordal_env->cls;
+ co->get_costs = get_costs;
+
+ s1 = get_irp_prog_name();
+ s2 = get_entity_name(get_irg_entity(co->irg));
+ s3 = chordal_env->cls->name;
+ len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
+ co->name = xmalloc(len);
+ snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
+
+ return co;
+}
+
+void free_copy_opt(copy_opt_t *co) {
+ xfree(co->name);
+ free(co);
+}
+
+int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
+ arch_register_req_t req;
+ const arch_register_t *reg;
+
+ if (arch_irn_is(co->aenv, irn, ignore))
+ return 0;
+
+ reg = arch_get_irn_register(co->aenv, irn);
+ if (arch_register_type_is(reg, ignore))
+ return 0;
+
+ if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
+ return 1;
+
+ return 0;
+}
+
+int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
+ const ir_edge_t *edge;
+ const arch_register_t *reg;
+
+ assert(0 && "Is buggy and obsolete. Do not use");
+
+ if (arch_irn_is(co->aenv, irn, ignore))
+ return 0;
+
+ reg = arch_get_irn_register(co->aenv, irn);
+ if (arch_register_type_is(reg, ignore))
+ return 0;
+
+ foreach_out_edge(irn, edge) {
+ ir_node *n = edge->src;
+
+ if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
+ arch_register_req_t req;
+ arch_get_register_req(co->aenv, &req, n, -1);
+
+ if(is_Reg_Phi(n) ||
+ is_Perm(co->aenv, n) ||
+ (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
+ )
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+int co_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);
+
+ if (is_Phi(root)) {
+ /* for phis the copies are placed in the corresponding pred-block */
+ loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
+ } else {
+ /* a perm places the copy in the same block as it resides */
+ loop = get_irn_loop(root_block);
+ }
+ if (loop) {
+ int d = get_loop_depth(loop);
+ cost = d*d;
+ }
+ return cost+1;
+}
+
+int co_get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
+ return 1;
+}
+
+/******************************************************************************
+ ____ _ _ _ _ _ _____ _
+ / __ \ | | | | | | (_) | / ____| |
+ | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
+ | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
+ | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
+ \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
+ | | __/ |
+ |_| |___/
+ ******************************************************************************/
+
/**
* Determines a maximum weighted independent set with respect to
* the interference and conflict edges of all nodes in a qnode.
if (!is_curr_reg_class(co, irn))
return;
- if (!co_is_optimizable(co->aenv, irn, &req))
+ if (!co_is_optimizable_root(co, irn))
return;
/* Init a new unit */
/* Proj of a perm with corresponding arg */
if (is_Perm_Proj(co->aenv, irn)) {
- assert(!nodes_interfere(co->cenv, irn, get_Copy_src(irn)));
+ assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
unit->costs = xmalloc(2 * sizeof(*unit->costs));
unit->node_count = 2;
unit->nodes[0] = irn;
- unit->nodes[1] = get_Copy_src(irn);
+ unit->nodes[1] = get_Perm_src(irn);
unit->costs[1] = co->get_costs(irn, unit->nodes[1], -1);
} else
unit->node_count = 2;
unit->nodes[0] = irn;
unit->nodes[1] = other;
- unit->costs[1] = co->get_costs(irn, other, -120480);
+ unit->costs[1] = co->get_costs(irn, other, -1);
}
} else
assert(0 && "This is not an optimizable node!");
/* Determine the minimal costs this unit will cause: min_nodes_costs */
unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
-
/* Insert the new ou according to its sort_key */
tmp = &co->units;
while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
}
}
-void be_copy_opt_init(void) {
- dbg = firm_dbg_register("ir.be.copyoptmain");
-}
+#ifdef QUICK_AND_DIRTY_HACK
-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;
+static int compare_ous(const void *k1, const void *k2) {
+ const unit_t *u1 = *((const unit_t **) k1);
+ const unit_t *u2 = *((const unit_t **) k2);
+ int i, o, u1_has_constr, u2_has_constr;
+ arch_register_req_t req;
+ const arch_env_t *aenv = u1->co->aenv;
+
+ /* Units with constraints come first */
+ u1_has_constr = 0;
+ for (i=0; i<u1->node_count; ++i) {
+ arch_get_register_req(aenv, &req, u1->nodes[i], -1);
+ if (arch_register_req_is(&req, limited)) {
+ u1_has_constr = 1;
+ break;
+ }
+ }
- dbg = firm_dbg_register("ir.be.copyopt");
+ u2_has_constr = 0;
+ for (i=0; i<u2->node_count; ++i) {
+ arch_get_register_req(aenv, &req, u2->nodes[i], -1);
+ if (arch_register_req_is(&req, limited)) {
+ u2_has_constr = 1;
+ break;
+ }
+ }
- co = xcalloc(1, sizeof(*co));
- co->cenv = chordal_env;
- co->aenv = chordal_env->birg->main_env->arch_env;
- co->irg = chordal_env->irg;
- co->cls = chordal_env->cls;
- co->get_costs = get_costs;
+ if (u1_has_constr != u2_has_constr)
+ return u2_has_constr - u1_has_constr;
- s1 = get_irp_prog_name();
- s2 = get_entity_name(get_irg_entity(co->irg));
- s3 = chordal_env->cls->name;
- len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
- co->name = xmalloc(len);
- snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
+ /* Now check, whether the two units are connected */
+#if 0
+ for (i=0; i<u1->node_count; ++i)
+ for (o=0; o<u2->node_count; ++o)
+ if (u1->nodes[i] == u2->nodes[o])
+ return 0;
+#endif
- DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
- INIT_LIST_HEAD(&co->units);
- irg_walk_graph(co->irg, co_collect_units, NULL, co);
- return co;
-}
+ /* After all, the sort key decides. Greater keys come first. */
+ return u2->sort_key - u1->sort_key;
-void free_copy_opt(copy_opt_t *co) {
- unit_t *curr, *tmp;
- xfree(co->name);
- list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
- xfree(curr->nodes);
- xfree(curr->costs);
- xfree(curr);
- }
}
-int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
- const ir_edge_t *edge;
+/**
+ * Sort the ou's according to constraints and their sort_key
+ */
+static void co_sort_units(copy_opt_t *co) {
+ int i, count = 0, costs;
+ unit_t *ou, **ous;
- if (arch_irn_is_ignore(co->aenv, irn))
- return 0;
+ /* get the number of ous, remove them form the list and fill the array */
+ list_for_each_entry(unit_t, ou, &co->units, units)
+ count++;
+ ous = alloca(count * sizeof(*ous));
- foreach_out_edge(irn, edge) {
- ir_node *n = edge->src;
+ costs = co_get_max_copy_costs(co);
- if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
- arch_register_req_t req;
- arch_get_register_req(co->aenv, &req, n, -1);
+ i = 0;
+ list_for_each_entry(unit_t, ou, &co->units, units)
+ ous[i++] = ou;
- if(is_Reg_Phi(n) ||
- is_Perm(co->aenv, n) ||
- (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
- )
- return 1;
- }
- }
+ INIT_LIST_HEAD(&co->units);
- return 0;
+ assert(count == i && list_empty(&co->units));
+
+ for (i=0; i<count; ++i)
+ ir_printf("%+F\n", ous[i]->nodes[0]);
+
+ qsort(ous, count, sizeof(*ous), compare_ous);
+
+ ir_printf("\n\n");
+ for (i=0; i<count; ++i)
+ ir_printf("%+F\n", ous[i]->nodes[0]);
+
+ /* reinsert into list in correct order */
+ for (i=0; i<count; ++i)
+ list_add_tail(&ous[i]->units, &co->units);
+
+ assert(costs == co_get_max_copy_costs(co));
}
+#endif
-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);
+void co_build_ou_structure(copy_opt_t *co) {
+ DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
+ INIT_LIST_HEAD(&co->units);
+ irg_walk_graph(co->irg, co_collect_units, NULL, co);
+#ifdef QUICK_AND_DIRTY_HACK
+ co_sort_units(co);
+#endif
+}
- if (is_Phi(root)) {
- /* for phis the copies are placed in the corresponding pred-block */
- loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
- } else {
- /* a perm places the copy in the same block as it resides */
- loop = get_irn_loop(root_block);
- }
- if (loop) {
- int d = get_loop_depth(loop);
- cost = d*d;
+void co_free_ou_structure(copy_opt_t *co) {
+ unit_t *curr, *tmp;
+ ASSERT_OU_AVAIL(co);
+ list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
+ xfree(curr->nodes);
+ xfree(curr->costs);
+ xfree(curr);
}
- return cost+1;
+ co->units.next = NULL;
}
-int get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
- return 1;
-}
+/* co_solve_heuristic() is implemented in becopyheur.c */
int co_get_max_copy_costs(const copy_opt_t *co) {
int i, res = 0;
unit_t *curr;
+ ASSERT_OU_AVAIL(co);
+
list_for_each_entry(unit_t, curr, &co->units, units) {
res += curr->inevitable_costs;
for (i=1; i<curr->node_count; ++i)
int res = 0;
unit_t *curr;
+ ASSERT_OU_AVAIL(co);
+
list_for_each_entry(unit_t, curr, &co->units, units)
res += curr->inevitable_costs;
return res;
int i, res = 0;
unit_t *curr;
+ ASSERT_OU_AVAIL(co);
+
list_for_each_entry(unit_t, curr, &co->units, units) {
int root_col = get_irn_col(co, curr->nodes[0]);
DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
int co_get_lower_bound(const copy_opt_t *co) {
int res = 0;
unit_t *curr;
+
+ ASSERT_OU_AVAIL(co);
+
list_for_each_entry(unit_t, curr, &co->units, units)
res += curr->inevitable_costs + curr->min_nodes_costs;
return res;
}
+/******************************************************************************
+ _____ _ _____ _
+ / ____| | | / ____| |
+ | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
+ | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
+ | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
+ \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
+ | | __/ |
+ |_| |___/
+ ******************************************************************************/
+
+static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
+ const affinity_node_t *n1 = k1;
+ const affinity_node_t *n2 = k2;
+
+ return (n1->irn != n2->irn);
+}
-
-#define DO_HEUR
-#undef DO_CLASSES
-#undef DO_ILP
-
-
-/**
- * Helpers for saving and restoring colors of nodes.
- * Used to get dependable and comparable benchmark results.
- */
-#if (defined(DO_HEUR) && defined(DO_BETTER)) || (defined(DO_HEUR) && defined(DO_ILP)) || (defined(DO_BETTER) && defined(DO_ILP))
-
-typedef struct color_saver {
- arch_env_t *arch_env;
- be_chordal_env_t *chordal_env;
- pmap *saved_colors;
- int flag; /* 0 save, 1 load */
-} color_save_t;
-
-static void save_load(ir_node *irn, void *env) {
- color_save_t *saver = env;
- if (saver->chordal_env->cls == arch_get_irn_reg_class(saver->arch_env, irn, -1)) {
- if (saver->flag == 0) { /* save */
- const arch_register_t *reg = arch_get_irn_register(saver->arch_env, irn);
- pmap_insert(saver->saved_colors, irn, (void *) reg);
- } else { /*load */
- arch_register_t *reg = pmap_get(saver->saved_colors, irn);
- arch_set_irn_register(saver->arch_env, irn, reg);
+static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
+ affinity_node_t new_node, *node;
+ neighb_t new_nbr, *nbr;
+ int allocnew;
+
+ new_node.irn = n1;
+ new_node.degree = 0;
+ new_node.neighbours = NULL;
+ node = set_insert(co->nodes, &new_node, sizeof(new_node), HASH_PTR(new_node.irn));
+
+ allocnew = 1;
+ for (nbr = node->neighbours; nbr; nbr = nbr->next)
+ if (nbr->irn == n2) {
+ allocnew = 0;
+ break;
}
+
+ /* if we did not find n2 in n1's neighbourhood insert it */
+ if (allocnew) {
+ obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
+ nbr = obstack_finish(&co->obst);
+ nbr->irn = n2;
+ nbr->costs = 0;
+ nbr->next = node->neighbours;
+ node->neighbours = nbr;
+ node->degree++;
}
-}
-static void save_colors(color_save_t *color_saver) {
- color_saver->flag = 0;
- irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
+ /* now nbr points to n1's neighbour-entry of n2 */
+ nbr->costs += costs;
}
-static void load_colors(color_save_t *color_saver) {
- color_saver->flag = 1;
- irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
+static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
+ if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
+ add_edge(co, n1, n2, costs);
+ add_edge(co, n2, n1, costs);
+ }
}
-#endif /* Need save/load stuff */
-
-
-
-void co_compare_solvers(be_chordal_env_t *chordal_env) {
- copy_opt_t *co;
-
-#ifdef DO_STAT
- lc_timer_t *timer;
- color_save_t saver;
- int costs, costs_inevit, costs_init, costs_heur, costs_classes, costs_ilp, lower_bound;
-#endif
-
- co = new_copy_opt(chordal_env, get_costs_loop_depth);
- DBG((dbg, LEVEL_1, "----> CO: %s\n", co->name));
- phi_class_compute(chordal_env->irg);
-
-
-#ifdef DO_STAT
-#if (defined(DO_HEUR) && defined(DO_BETTER)) || (defined(DO_HEUR) && defined(DO_ILP)) || (defined(DO_BETTER) && defined(DO_ILP))
- saver.arch_env = chordal_env->main_env->arch_env;
- saver.chordal_env = chordal_env;
- saver.saved_colors = pmap_create();
- save_colors(&saver);
-#endif
-
- costs_inevit = co_get_inevit_copy_costs(co);
- lower_bound = co_get_lower_bound(co);
- costs_init = co_get_copy_costs(co);
-
- DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
- DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
- DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
-
- copystat_add_inevit_costs(costs_inevit);
- copystat_add_init_costs(costs_init);
- copystat_add_max_costs(co_get_max_copy_costs(co));
-#endif
-
-
-#ifdef DO_HEUR
-#ifdef DO_STAT
- timer = lc_timer_register("heur", NULL);
- lc_timer_reset_and_start(timer);
-#endif
-
- co_solve_heuristic(co);
+static void build_graph_walker(ir_node *irn, void *env) {
+ copy_opt_t *co = env;
+ int pos, max;
+ arch_register_req_t req;
+ const arch_register_t *reg;
-#ifdef DO_STAT
- lc_timer_stop(timer);
- costs_heur = co_get_copy_costs(co);
- DBG((dbg, LEVEL_1, "Heur costs: %3d\n", costs_heur));
- copystat_add_heur_time(lc_timer_elapsed_msec(timer));
- copystat_add_heur_costs(costs_heur);
- assert(lower_bound <= costs_heur);
-#endif
-#endif /* DO_HEUR */
+ if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
+ return;
+ reg = arch_get_irn_register(co->aenv, irn);
+ if (arch_register_type_is(reg, ignore))
+ return;
+ /* Phis */
+ if (is_Reg_Phi(irn))
+ for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
+ ir_node *arg = get_irn_n(irn, pos);
+ add_edges(co, irn, arg, co->get_costs(irn, arg, pos));
+ }
-#ifdef DO_CLASSES
-#ifdef DO_STAT
-#ifdef DO_HEUR
- load_colors(&saver);
-#endif
- timer = lc_timer_register("classes", NULL);
- lc_timer_reset_and_start(timer);
-#endif
+ /* Perms */
+ else if (is_Perm_Proj(co->aenv, irn)) {
+ ir_node *arg = get_Perm_src(irn);
+ add_edges(co, irn, arg, co->get_costs(irn, arg, 0));
+ }
- co_classes_opt(co);
+ /* 2-address code */
+ else if (is_2addr_code(co->aenv, irn, &req))
+ add_edges(co, irn, req.other_same, co->get_costs(irn, req.other_same, 0));
+}
-#ifdef DO_STAT
- lc_timer_stop(timer);
- costs_classes = co_get_copy_costs(co);
- DBG((dbg, LEVEL_1, "Classes costs: %3d\n", costs_classes));
- copystat_add_classes_time(lc_timer_elapsed_msec(timer));
- copystat_add_classes_costs(costs_heur);
- assert(lower_bound <= costs_classes);
-#endif
-#endif /* DO_CLASSES */
+void co_build_graph_structure(copy_opt_t *co) {
+ obstack_init(&co->obst);
+ co->nodes = new_set(compare_affinity_node_t, 32);
+ irg_walk_graph(co->irg, build_graph_walker, NULL, co);
+}
+void co_free_graph_structure(copy_opt_t *co) {
+ ASSERT_GS_AVAIL(co);
-#ifdef DO_ILP
-#ifdef DO_STAT
-#if defined(DO_HEUR) || defined(DO_CLASSES)
- load_colors(&saver);
-#endif
-#endif
+ del_set(co->nodes);
+ obstack_free(&co->obst, NULL);
+ co->nodes = NULL;
+}
- co_solve_ilp1(co, 60.0);
+/* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
-#ifdef DO_STAT
- costs_ilp = co_get_copy_costs(co);
- DBG((dbg, LEVEL_1, "Opt costs: %3d\n", costs_ilp));
- copystat_add_opt_costs(costs_ilp);
- assert(lower_bound <= costs_ilp);
-#endif
-#endif /* DO_ILP */
+int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
+ affinity_node_t new_node, *n;
+ ASSERT_GS_AVAIL(co);
-#ifdef DO_STAT
-#if (defined(DO_HEUR) && defined(DO_BETTER)) || (defined(DO_HEUR) && defined(DO_ILP)) || (defined(DO_BETTER) && defined(DO_ILP))
- pmap_destroy(saver.saved_colors);
-#endif
-#endif
- free_copy_opt(co);
+ new_node.irn = irn;
+ n = set_find(co->nodes, &new_node, sizeof(new_node), HASH_PTR(new_node.irn));
+ if (n) {
+ return (n->degree > 0);
+ } else
+ return 0;
}
projects / libfirm / blobdiff