* Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
*/
#ifdef HAVE_CONFIG_H
-#include "config.h"
+#include <config.h>
#endif
+
#ifdef HAVE_ALLOCA_H
#include <alloca.h>
#endif
#include "xmalloc.h"
#include "debug.h"
#include "pmap.h"
+#include "raw_bitset.h"
#include "irgraph.h"
#include "irgwalk.h"
#include "irprog.h"
#include "irphase_t.h"
#include "irprintf_t.h"
-
+#include "bemodule.h"
#include "bearch.h"
#include "benode_t.h"
#include "beutil.h"
#include "beinsn_t.h"
#include "besched_t.h"
#include "benodesets.h"
+#include "bejavacoal.h"
+#include "bestatevent.h"
+#include "beirg_t.h"
+#include "error.h"
+
+#include <libcore/lc_timing.h>
+#include <libcore/lc_opts.h>
+#include <libcore/lc_opts_enum.h>
#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 unsigned algo = CO_ALGO_HEUR2;
+static int improve = 1;
-#ifdef WITH_LIBCORE
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "before", DUMP_BEFORE },
{ "after", DUMP_AFTER },
{ "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 }
};
extern void be_co2_register_options(lc_opt_entry_t *grp);
extern void be_co3_register_options(lc_opt_entry_t *grp);
-void co_register_options(lc_opt_entry_t *grp)
+void be_init_copycoal(void)
{
- lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
- lc_opt_add_table(co_grp, options);
+ 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 *co_grp = lc_opt_get_grp(chordal_grp, "co");
- be_co2_register_options(co_grp);
- be_co3_register_options(co_grp);
-#ifdef WITH_ILP
- be_co_ilp_register_options(co_grp);
-#endif
+ lc_opt_add_table(co_grp, options);
}
-#endif
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
#undef QUICK_AND_DIRTY_HACK
+static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
+{
+ if(env->ifg)
+ return be_ifg_connected(env->ifg, a, b);
+ else
+ return values_interfere(env->birg->lv, a, b);
+}
+
+
/******************************************************************************
_____ _
/ ____| | |
DEBUG_ONLY(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, cost_fct_t get_costs)
{
}
int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
- arch_register_req_t req;
+ const arch_register_req_t *req;
const arch_register_t *reg;
if (arch_irn_is(co->aenv, irn, ignore))
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))
+ req = arch_get_register_req(co->aenv, irn, -1);
+ if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req))
return 1;
return 0;
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);
+ const arch_register_req_t *req;
+ req = arch_get_register_req(co->aenv, 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;
+ (arch_register_req_is(req, should_be_same))) {
+ ir_node *other = get_irn_n(irn, req->other_same);
+ if(other == irn)
+ return 1;
+ }
}
}
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->birg->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;
}
static void co_collect_units(ir_node *irn, void *env) {
copy_opt_t *co = env;
unit_t *unit;
- arch_register_req_t req;
if (!is_curr_reg_class(co, irn))
return;
unit->nodes[0] = irn;
unit->nodes[1] = get_Perm_src(irn);
unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
- } else
-
- /* Src == Tgt of a 2-addr-code instruction */
- if (is_2addr_code(co->aenv, irn, &req)) {
- ir_node *other = req.other_same;
- if (!nodes_interfere(co->cenv, irn, other)) {
- 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] = other;
- unit->costs[1] = co->get_costs(co, irn, other, -1);
+ } else {
+ const arch_register_req_t *req =
+ arch_get_register_req(co->aenv, irn, -1);
+
+ /* Src == Tgt of a 2-addr-code instruction */
+ if (is_2addr_code(req)) {
+ ir_node *other = get_irn_n(irn, req->other_same);
+ if (!nodes_interfere(co->cenv, irn, other)) {
+ 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] = other;
+ unit->costs[1] = co->get_costs(co, irn, other, -1);
+ }
+ } else {
+ assert(0 && "This is not an optimizable node!");
}
- } else
- assert(0 && "This is not an optimizable node!");
+ }
/* Insert the new unit at a position according to its costs */
if (unit->node_count > 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);
+}
+
/******************************************************************************
_____ _ _____ _
/ ____| | | / ____| |
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;
if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
}
/* 2-address code */
- else if (is_2addr_code(co->aenv, irn, &req))
- add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
+ else {
+ const arch_register_req_t *req =
+ arch_get_register_req(co->aenv, irn, -1);
+ if (is_2addr_code(req)) {
+ ir_node *other = get_irn_n(irn, req->other_same);
+ add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
+ }
+ }
}
void co_build_graph_structure(copy_opt_t *co) {
{
be_ifg_t *ifg = co->cenv->ifg;
int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
- bitset_t *adm = bitset_alloca(co->cls->n_regs);
ir_node *irn;
void *it, *nit;
int idx = PTR_TO_INT(get_irn_link(irn));
affinity_node_t *a = get_affinity_info(co, irn);
- arch_register_req_t req;
+ const arch_register_req_t *req;
ir_node *adj;
- arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
- if(arch_register_req_is(&req, limited)) {
- bitset_clear_all(adm);
- req.limited(req.limited_env, adm);
- for(i = 0; i < co->cls->n_regs; ++i)
- if(!bitset_is_set(adm, i) && color_map[i] >= 0)
+ req = arch_get_register_req(co->aenv, irn, BE_OUT_POS(0));
+ if(arch_register_req_is(req, limited)) {
+ for(i = 0; i < co->cls->n_regs; ++i) {
+ if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
fprintf(f, "%d %d -1\n", color_map[i], idx);
-
+ }
}
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);
}
}
}
struct obstack *obst = &env->obst;
void *base = obstack_base(obst);
pset *live = pset_new_ptr_default();
+ be_lv_t *lv = env->co->cenv->birg->lv;
int n_insns = 0;
int n_nodes = 0;
}
DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
- be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
+ be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
/* Generate the bad and ugly. */
for(i = n_insns - 1; i >= 0; --i) {
int n_colors;
appel_clique_walker_t env;
bitset_t *adm = bitset_alloca(co->cls->n_regs);
+ be_lv_t *lv = co->cenv->birg->lv;
- be_liveness_recompute(co->cenv->lv);
+ be_liveness_recompute(lv);
obstack_init(&env.obst);
phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
env.curr_nr = co->cls->n_regs;
{
co_ifg_dump_t *env = self;
const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
- arch_register_req_t req;
+ const arch_register_req_t *req;
int limited;
- arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
- limited = arch_register_req_is(&req, limited);
+ req = arch_get_register_req(env->co->aenv, irn, BE_OUT_POS(0));
+ limited = arch_register_req_is(req, limited);
if(env->flags & CO_IFG_DUMP_LABELS) {
ir_fprintf(f, "label=\"%+F", irn);
+#if 0
+ // TODO fix this...
if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
req.limited(req.limited_env, bs);
ir_fprintf(f, "\\n%B", bs);
}
+#endif
ir_fprintf(f, "\" ");
- }
-
- else
+ } else {
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 }
};
/*
*/
+static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
+{
+ FILE *result;
+ char buf[1024];
+
+ ir_snprintf(buf, sizeof(buf), "%s%F_%s%s", prefix, env->irg, env->cls->name, suffix);
+ result = fopen(buf, "wt");
+ if(result == NULL) {
+ panic("Couldn't open '%s' for writing.", buf);
+ }
+
+ return result;
+}
+
void co_driver(be_chordal_env_t *cenv)
{
+ lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
+ co_complete_stats_t before, after;
copy_opt_t *co;
co_algo_t *algo_func;
- int init_costs;
+ 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);
- init_costs = co_get_copy_costs(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) {
- FILE *f = be_chordal_open(cenv, "", ".apl");
+ FILE *f = my_open(cenv, "", ".apl");
co_dump_appel_graph(co, f);
fclose(f);
}
if(dump_flags & DUMP_BEFORE) {
- FILE *f = be_chordal_open(cenv, "", "-before.dot");
+ FILE *f = my_open(cenv, "", "-before.dot");
co_dump_ifg_dot(co, f, style_flags);
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;
+
+ lc_timer_reset_and_start(timer);
+
+ was_optimal = algo_func(co);
+
+ lc_timer_stop(timer);
+ be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
+
+ be_stat_ev_ull("co_optimal", was_optimal);
if(dump_flags & DUMP_AFTER) {
- FILE *f = be_chordal_open(cenv, "", "-after.dot");
+ FILE *f = my_open(cenv, "", "-after.dot");
co_dump_ifg_dot(co, f, style_flags);
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 %10s %10d%10d%10d%10d", cenv->irg, cenv->cls->name,
- co_get_max_copy_costs(co), init_costs,
- co_get_inevit_copy_costs(co), co_get_lower_bound(co));
+ 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