#include "belive_t.h"
#include "beinsn_t.h"
#include "besched_t.h"
+#include "benodesets.h"
+#include "bejavacoal.h"
+#include "bestatevent.h"
+
+#ifdef WITH_LIBCORE
+#include <libcore/lc_timing.h>
+#include <libcore/lc_opts.h>
+#endif /* WITH_LIBCORE */
#define DUMP_BEFORE 1
#define DUMP_AFTER 2
#define COST_FUNC_LOOP 2
#define COST_FUNC_ALL_ONE 3
-static int dump_flags = 0;
-static int style_flags = 0;
-static int do_stats = 0;
+static unsigned dump_flags = 0;
+static unsigned style_flags = 0;
+static unsigned do_stats = 0;
static cost_fct_t cost_func = co_get_costs_exec_freq;
static int algo = CO_ALGO_HEUR2;
+static int improve = 1;
#ifdef WITH_LIBCORE
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "none", CO_ALGO_NONE },
{ "heur", CO_ALGO_HEUR },
{ "heur2", CO_ALGO_HEUR2 },
+#ifdef WITH_JVM
{ "heur3", CO_ALGO_HEUR3 },
+#endif /* WITH_JVM */
+#ifdef WITH_ILP
{ "ilp", CO_ALGO_ILP },
+#endif /* WITH_ILP */
{ NULL, 0 }
};
+typedef int (*opt_funcptr)(void);
+
static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
- { "freq", co_get_costs_exec_freq },
- { "loop", co_get_costs_loop_depth },
- { "one", co_get_costs_all_one },
- { NULL, 0 }
+ { "freq", (opt_funcptr) co_get_costs_exec_freq },
+ { "loop", (opt_funcptr) co_get_costs_loop_depth },
+ { "one", (opt_funcptr) co_get_costs_all_one },
+ { NULL, NULL }
};
static lc_opt_enum_mask_var_t dump_var = {
};
static lc_opt_enum_func_ptr_var_t cost_func_var = {
- &cost_func, cost_func_items
+ (opt_funcptr*) &cost_func, cost_func_items
};
static const lc_opt_table_entry_t options[] = {
- LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo (heur, heur2, heur3, ilp)", &algo_var),
- LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function (freq, loop, one)", &cost_func_var),
- LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
- LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
- LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
+ LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var),
+ LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
+ LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
+ LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
+ LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
+ LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
{ NULL }
};
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
+
void be_copy_opt_init(void) {
}
int d = get_loop_depth(loop);
cost = d*d;
}
- return cost+1;
+ return 1+cost;
}
int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
+ int res;
ir_node *root_bl = get_nodes_block(root);
ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
- unsigned long freq = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
- return freq > 0 ? (int) freq : 1;
+ res = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
+
+ /* don't allow values smaller than one. */
+ return res < 1 ? 1 : res;
}
-int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
+int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
return 1;
}
return res;
}
+void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
+{
+ bitset_t *seen = bitset_irg_malloc(co->irg);
+ affinity_node_t *an;
+
+ memset(stat, 0, sizeof(stat[0]));
+
+ /* count affinity edges. */
+ co_gs_foreach_aff_node(co, an) {
+ neighb_t *neigh;
+ stat->aff_nodes += 1;
+ bitset_add_irn(seen, an->irn);
+ co_gs_foreach_neighb(an, neigh) {
+ if(!bitset_contains_irn(seen, neigh->irn)) {
+ stat->aff_edges += 1;
+ stat->max_costs += neigh->costs;
+
+ if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
+ stat->costs += neigh->costs;
+ stat->unsatisfied_edges += 1;
+ }
+
+ if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
+ stat->aff_int += 1;
+ stat->inevit_costs += neigh->costs;
+ }
+
+ }
+ }
+ }
+
+ bitset_free(seen);
+}
+
/******************************************************************************
_____ _ _____ _
/ ____| | | / ____| |
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));
+ node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
allocnew = 1;
for (nbr = node->neighbours; nbr; nbr = nbr->next)
ASSERT_GS_AVAIL(co);
new_node.irn = irn;
- n = set_find(co->nodes, &new_node, sizeof(new_node), HASH_PTR(new_node.irn));
+ n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
if (n) {
return (n->degree > 0);
} else
if(!arch_irn_is(co->aenv, n->irn, ignore)) {
int n_idx = PTR_TO_INT(get_irn_link(n->irn));
if(idx < n_idx)
- fprintf(f, "%d %d %d\n", idx, n_idx, n->costs);
+ fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
}
}
}
for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
ir_node *pred = get_Block_cfgpred_block(bl, j);
- appel_block_info_t *pred_bli = phase_get_irn_data(&env->ph, pred);
int nr = appel_get_live_end_nr(env, pred, irn);
assert(nr >= 0);
for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
ir_node *pred = get_Block_cfgpred_block(bl, j);
- appel_block_info_t *pred_bli = phase_get_irn_data(&env->ph, pred);
ir_node *op = get_irn_n(irn, j);
int nr = appel_get_live_end_nr(env, pred, op);
}
else
- fprintf(f, "label=\"\"" );
+ fprintf(f, "label=\"\" shape=point " );
if(env->flags & CO_IFG_DUMP_SHAPE)
fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
if(env->flags & CO_IFG_DUMP_COLORS)
fprintf(file, "color=%s ", color);
else
- fprintf(file, "style=dashed");
+ fprintf(file, "style=dotted");
fprintf(file, "];\n");
}
}
|___/
*/
-static co_algo_t *algos[] = {
- void_algo,
- co_solve_heuristic,
- co_solve_heuristic_new,
- co_solve_heuristic_java,
+typedef struct {
+ co_algo_t *algo;
+ const char *name;
+ int can_improve_existing;
+} co_algo_info_t;
+
+static co_algo_info_t algos[] = {
+ { void_algo, "none", 0 },
+ { co_solve_heuristic, "heur1", 0 },
+ { co_solve_heuristic_new, "heur2", 0 },
+ { co_solve_heuristic_java, "heur3", 0 },
#ifdef WITH_ILP
- co_solve_ilp2
+ { co_solve_ilp2, "ilp", 1 },
#endif
+ { NULL, "", 0 }
};
/*
void co_driver(be_chordal_env_t *cenv)
{
+#ifdef WITH_LIBCORE
+ lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
+#endif
+ co_complete_stats_t before, after;
copy_opt_t *co;
co_algo_t *algo_func;
+ int was_optimal = 0;
if(algo < 0 || algo >= CO_ALGO_LAST)
return;
co = new_copy_opt(cenv, cost_func);
co_build_ou_structure(co);
co_build_graph_structure(co);
- if(do_stats) {
- ir_printf("%30F %10s %10d%10d%10d%10d", cenv->irg, cenv->cls->name,
- co_get_max_copy_costs(co), co_get_copy_costs(co),
- co_get_inevit_copy_costs(co), co_get_lower_bound(co));
- }
+
+ co_complete_stats(co, &before);
+
+ be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
+ be_stat_ev_ull("co_aff_edges", before.aff_edges);
+ be_stat_ev_ull("co_max_costs", before.max_costs);
+ be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
+ be_stat_ev_ull("co_aff_int", before.aff_int);
+
+ be_stat_ev_ull("co_init_costs", before.costs);
+ be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
/* Dump the interference graph in Appel's format. */
if(dump_flags & DUMP_APPEL) {
fclose(f);
}
- algo_func = algos[algo];
- algo_func(co);
+ /* if the algo can improve results, provide an initial solution with heur3 */
+ if(improve && algos[algo].can_improve_existing) {
+ co_complete_stats_t stats;
+
+ /* produce a heuristical solution */
+ co_solve_heuristic_java(co);
+
+ /* do the stats and provide the current costs */
+ co_complete_stats(co, &stats);
+ be_stat_ev_ull("co_prepare_costs", stats.costs);
+ }
+
+#ifdef WITH_JVM
+ /* start the JVM here so that it does not tamper the timing. */
+ if(algo == CO_ALGO_HEUR3)
+ be_java_coal_start_jvm();
+#endif
+
+ algo_func = algos[algo].algo;
+
+#ifdef WITH_LIBCORE
+ lc_timer_reset_and_start(timer);
+#endif
+
+ was_optimal = algo_func(co);
+
+#ifdef WITH_LIBCORE
+ lc_timer_stop(timer);
+ be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
+#endif
+
+ be_stat_ev_ull("co_optimal", was_optimal);
if(dump_flags & DUMP_AFTER) {
FILE *f = be_chordal_open(cenv, "", "-after.dot");
fclose(f);
}
+ co_complete_stats(co, &after);
+
if(do_stats) {
- int optimizable_costs = co_get_max_copy_costs(co) - co_get_lower_bound(co);
- int remaining = co_get_copy_costs(co);
- int evitable = remaining - co_get_lower_bound(co);
+ ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
+ ulong64 evitable = after.costs - after.inevit_costs;
+
+ ir_printf("%30F ", cenv->irg);
+ printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
if(optimizable_costs > 0)
- printf("%10d %5.2f\n", remaining, (evitable * 100.0) / optimizable_costs);
+ printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
else
- printf("%10d %5s\n", remaining, "-");
+ printf("%10" ULL_FMT " %5s\n", after.costs, "-");
}
+ be_stat_ev_ull("co_after_costs", after.costs);
+ be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
+
co_free_graph_structure(co);
co_free_ou_structure(co);
free_copy_opt(co);
projects / libfirm / blobdiff