+/*
+ * Copyright (C) 1995-2007 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.
+ */
+
/**
- * Author: Daniel Grund
- * Date: 12.04.2005
- * Copyright: (c) Universitaet Karlsruhe
- * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
+ * @file
+ * @brief Copy minimization driver.
+ * @author Daniel Grund
+ * @date 12.04.2005
+ * @version $Id$
+ *
+ * Main file for the optimization reducing the copies needed for:
+ * - Phi coalescing
+ * - Register-constrained nodes
+ * - Two-address code instructions
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#include "debug.h"
#include "pmap.h"
#include "raw_bitset.h"
+#include "irnode.h"
#include "irgraph.h"
#include "irgwalk.h"
#include "irprog.h"
#include "irprintf_t.h"
#include "bemodule.h"
-#include "bearch.h"
+#include "bearch_t.h"
#include "benode_t.h"
#include "beutil.h"
#include "beifg_t.h"
+#include "beintlive_t.h"
#include "becopyopt_t.h"
#include "becopystat.h"
#include "belive_t.h"
#define COST_FUNC_LOOP 2
#define COST_FUNC_ALL_ONE 3
-static unsigned dump_flags = 0;
-static unsigned style_flags = 0;
-static unsigned 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 unsigned algo = CO_ALGO_HEUR2;
-static int improve = 1;
+static unsigned algo = CO_ALGO_HEUR4;
+static int improve = 1;
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "before", DUMP_BEFORE },
#ifdef WITH_JVM
{ "heur3", CO_ALGO_HEUR3 },
#endif /* WITH_JVM */
+ { "heur4", CO_ALGO_HEUR4 },
#ifdef WITH_ILP
{ "ilp", CO_ALGO_ILP },
#endif /* WITH_ILP */
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 }
+ LC_OPT_LAST
};
/* Insert additional options registration functions here. */
static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
{
- if(env->ifg)
+ if (env->ifg)
return be_ifg_connected(env->ifg, a, b);
else
- return values_interfere(env->birg->lv, a, b);
+ return values_interfere(env->birg, a, b);
}
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) {
- 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))) {
- ir_node *other = get_irn_n(irn, req->other_same);
- if(other == irn)
- return 1;
- }
- }
- }
-
- return 0;
-}
-
int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
int cost = 0;
ir_loop *loop;
ir_node *root_block = get_nodes_block(root);
+ (void) co;
+ (void) arg;
if (is_Phi(root)) {
/* for phis the copies are placed in the corresponding pred-block */
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;
+ (void) arg;
res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
/* don't allow values smaller than one. */
int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
+ (void) co;
+ (void) root;
+ (void) arg;
+ (void) pos;
return 1;
}
ir_node **safe, **unsafe;
int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
bitset_t *curr;
- int max, pos, curr_weight, best_weight = 0;
+ bitset_pos_t pos;
+ int max, curr_weight, best_weight = 0;
/* assign the nodes into two groups.
* safe: node has no interference, hence it is in every max stable set.
/* Else insert the argument of the phi to the members of this ou */
DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
- /* Check if arg has occurred at a prior position in the arg/list */
- arg_pos = 0;
- for (o=0; o<unit->node_count; ++o)
- if (unit->nodes[o] == arg) {
- arg_pos = o;
- break;
+ if (! arch_irn_is(co->aenv, arg, ignore)) {
+ /* Check if arg has occurred at a prior position in the arg/list */
+ arg_pos = 0;
+ for (o=1; o<unit->node_count; ++o) {
+ if (unit->nodes[o] == arg) {
+ arg_pos = o;
+ break;
+ }
}
- if (!arg_pos) { /* a new argument */
- /* insert node, set costs */
- unit->nodes[unit->node_count] = arg;
- unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
- unit->node_count++;
- } else { /* arg has occured before in same phi */
- /* increase costs for existing arg */
- unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
+ if (!arg_pos) { /* a new argument */
+ /* insert node, set costs */
+ unit->nodes[unit->node_count] = arg;
+ unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
+ unit->node_count++;
+ } else { /* arg has occurred before in same phi */
+ /* increase costs for existing arg */
+ unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
+ }
}
}
unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
/* Src == Tgt of a 2-addr-code instruction */
if (is_2addr_code(req)) {
- ir_node *other = get_irn_n(irn, req->other_same);
+ ir_node *other = get_irn_n(skip_Proj(irn), req->other_same[0]); /* TODO handle second should-be-same constraint */
if (!arch_irn_is(co->aenv, other, ignore) &&
!nodes_interfere(co->cenv, irn, other)) {
unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
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;
+ (void) size;
return (n1->irn != n2->irn);
}
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;
+ neighb_t *nbr;
+ int allocnew = 1;
new_node.irn = n1;
new_node.degree = 0;
new_node.neighbours = NULL;
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)
if (nbr->irn == n2) {
allocnew = 0;
/* 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 = obstack_alloc(&co->obst, sizeof(*nbr));
nbr->irn = n2;
nbr->costs = 0;
nbr->next = node->neighbours;
+
node->neighbours = nbr;
node->degree++;
}
if (arch_register_type_is(reg, ignore))
return;
- /* Phis */
- if (is_Reg_Phi(irn))
+ if (is_Reg_Phi(irn)) { /* Phis */
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(co, irn, arg, pos));
}
-
- /* Perms */
- else if (is_Perm_Proj(co->aenv, irn)) {
+ }
+ else if (is_Perm_Proj(co->aenv, irn)) { /* Perms */
ir_node *arg = get_Perm_src(irn);
add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
}
-
- /* 2-address code */
- else {
- const arch_register_req_t *req =
- arch_get_register_req(co->aenv, irn, -1);
+ else { /* 2-address code */
+ 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);
- if(!arch_irn_is(co->aenv, other, ignore))
- add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
+ const int *i;
+ for (i = req->other_same; i != ENDOF(req->other_same); ++i) {
+ ir_node *other;
+
+ if (*i == -1) break;
+
+ other = get_irn_n(skip_Proj(irn), *i);
+ if (! arch_irn_is(co->aenv, other, ignore))
+ add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
+ }
}
}
}
}
typedef struct _appel_clique_walker_t {
- phase_t ph;
+ ir_phase ph;
const copy_opt_t *co;
int curr_nr;
int node_count;
return res == 0 ? 1 : res;
#else
ir_loop *loop = get_irn_loop(bl);
+ (void) env;
if(loop) {
int d = get_loop_depth(loop);
return 1 + d * d;
#endif
}
-static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
+static void *appel_clique_walker_irn_init(ir_phase *phase, ir_node *irn, void *old)
{
appel_block_info_t *res = NULL;
+ (void) old;
if(is_Block(irn)) {
appel_clique_walker_t *d = (void *) phase;
return -1;
}
-static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
+static int appel_dump_clique(appel_clique_walker_t *env, const ir_nodeset_t *live, ir_node *bl, int curr_nr, int start_nr)
{
ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
ir_node *irn;
int n_live;
int j;
+ ir_nodeset_iterator_t iter;
n_live = 0;
- foreach_pset(live, irn)
+ foreach_ir_nodeset(live, irn, iter)
live_arr[n_live++] = irn;
/* dump the live after clique */
appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
struct obstack *obst = &env->obst;
void *base = obstack_base(obst);
- pset *live = pset_new_ptr_default();
+ ir_nodeset_t live;
+ ir_nodeset_iterator_t iter;
be_lv_t *lv = env->co->cenv->birg->lv;
int n_insns = 0;
n_nodes++;
bli->n_phi = 0;
- insns = malloc(n_nodes * sizeof(insns[0]));
+ insns = xmalloc(n_nodes * sizeof(insns[0]));
/* Put all insns in an array. */
irn = sched_first(bl);
irn = insn->next_insn;
}
- DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
- be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
+ DBG((dbg, LEVEL_2, "%+F\n", bl));
+ ir_nodeset_init(&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) {
/* The first live set has to be saved in the block border set. */
if(i == n_insns - 1) {
j = 0;
- foreach_pset(live, irn) {
+ foreach_ir_nodeset(&live, irn, iter) {
bli->live_end[j] = irn;
bli->live_end_nr[j] = curr_nr + j;
++j;
ir_node *op = insn->ops[j].carrier;
bitset_t *adm = insn->ops[j].regs;
int k;
- int nr;
+ size_t nr;
if(!insn->ops[j].has_constraints)
continue;
nr = 0;
- foreach_pset(live, irn) {
+ foreach_ir_nodeset(&live, irn, iter) {
if(irn == op) {
- pset_break(live);
break;
}
++nr;
}
- assert(nr < pset_count(live));
+ assert(nr < ir_nodeset_size(&live));
for(k = 0; k < env->co->cls->n_regs; ++k) {
int mapped_col = env->color_map[k];
}
/* dump the clique and update the stuff. */
- curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
+ curr_nr = appel_dump_clique(env, &live, bl, curr_nr, start_nr);
/* remove all defs. */
for(j = 0; j < insn->use_start; ++j)
- pset_remove_ptr(live, insn->ops[j].carrier);
+ ir_nodeset_remove(&live, insn->ops[j].carrier);
if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
bli->phi[bli->n_phi] = insn->irn;
/* add all uses */
else
for(j = insn->use_start; j < insn->n_ops; ++j)
- pset_insert_ptr(live, insn->ops[j].carrier);
+ ir_nodeset_insert(&live, insn->ops[j].carrier);
}
/* print the start clique. */
- curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
+ curr_nr = appel_dump_clique(env, &live, bl, curr_nr, start_nr);
i = 0;
- foreach_pset(live, irn) {
+ foreach_ir_nodeset(&live, irn, iter) {
bli->live_in[i] = irn;
bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
++i;
}
bli->n_live_in = i;
- del_pset(live);
+ ir_nodeset_destroy(&live);
free(insns);
obstack_free(obst, base);
env->curr_nr = curr_nr;
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);
+ phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init, NULL);
env.curr_nr = co->cls->n_regs;
env.co = co;
env.f = f;
|_| |___/
*/
-static const char *get_dot_color_name(int col)
+static const char *get_dot_color_name(size_t col)
{
static const char *names[] = {
"blue",
static void ifg_dump_graph_attr(FILE *f, void *self)
{
+ (void) self;
fprintf(f, "overlap=scale");
}
void co_solve_park_moon(copy_opt_t *opt)
{
-
+ (void) opt;
}
static int void_algo(copy_opt_t *co)
{
+ (void) co;
return 0;
}
*/
typedef struct {
- co_algo_t *algo;
+ co_algo_t *algo;
const char *name;
- int can_improve_existing;
+ 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 },
+ { void_algo, "none", 0 },
+ { co_solve_heuristic, "heur1", 0 },
+ { co_solve_heuristic_new, "heur2", 0 },
+ { co_solve_heuristic_java, "heur3", 0 },
+ { co_solve_heuristic_mst, "heur4", 0 },
#ifdef WITH_ILP
- { co_solve_ilp2, "ilp", 1 },
+ { co_solve_ilp2, "ilp", 1 },
#endif
- { NULL, "", 0 }
+ { NULL, "", 0 }
};
/*
void co_driver(be_chordal_env_t *cenv)
{
- lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
+ 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 was_optimal = 0;
+ copy_opt_t *co;
+ co_algo_t *algo_func;
+ int was_optimal = 0;
- if(algo < 0 || algo >= CO_ALGO_LAST)
+ if (algo >= CO_ALGO_LAST)
return;
+ be_liveness_assure_chk(be_get_birg_liveness(cenv->birg));
+
co = new_copy_opt(cenv, cost_func);
co_build_ou_structure(co);
co_build_graph_structure(co);
be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
/* Dump the interference graph in Appel's format. */
- if(dump_flags & DUMP_APPEL) {
+ if (dump_flags & DUMP_APPEL) {
FILE *f = my_open(cenv, "", ".apl");
co_dump_appel_graph(co, f);
fclose(f);
}
- if(dump_flags & DUMP_BEFORE) {
+ if (dump_flags & DUMP_BEFORE) {
FILE *f = my_open(cenv, "", "-before.dot");
co_dump_ifg_dot(co, f, style_flags);
fclose(f);
}
/* if the algo can improve results, provide an initial solution with heur3 */
- if(improve && algos[algo].can_improve_existing) {
+ if (improve && algos[algo].can_improve_existing) {
co_complete_stats_t stats;
- /* produce a heuristical solution */
+ /* produce a heuristic solution */
+#ifdef WITH_JVM
co_solve_heuristic_java(co);
+#else
+ co_solve_heuristic(co);
+#endif /* WITH_JVM */
/* do the stats and provide the current costs */
co_complete_stats(co, &stats);
#ifdef WITH_JVM
/* start the JVM here so that it does not tamper the timing. */
- if(algo == CO_ALGO_HEUR3)
+ if (algo == CO_ALGO_HEUR3)
be_java_coal_start_jvm();
-#endif
+#endif /* WITH_JVM */
algo_func = algos[algo].algo;
+ /* perform actual copy minimization */
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("co_time", lc_timer_elapsed_msec(timer));
be_stat_ev_ull("co_optimal", was_optimal);
- if(dump_flags & DUMP_AFTER) {
+ if (dump_flags & DUMP_AFTER) {
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) {
+ if (do_stats) {
ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
ulong64 evitable = after.costs - after.inevit_costs;
projects / libfirm / blobdiff