+/*
+ * Copyright (C) 1995-2008 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"
-#endif
-#ifdef HAVE_ALLOCA_H
-#include <alloca.h>
-#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
#include "execfreq.h"
#include "xmalloc.h"
#include "debug.h"
#include "pmap.h"
+#include "raw_bitset.h"
+#include "irnode.h"
#include "irgraph.h"
#include "irgwalk.h"
#include "irprog.h"
#include "irloop_t.h"
#include "iredges_t.h"
-#include "phiclass.h"
#include "irbitset.h"
#include "irphase_t.h"
#include "irprintf_t.h"
-
+#include "bemodule.h"
#include "bearch.h"
-#include "benode_t.h"
+#include "benode.h"
#include "beutil.h"
-#include "beifg_t.h"
+#include "beifg.h"
+#include "beintlive_t.h"
#include "becopyopt_t.h"
#include "becopystat.h"
#include "belive_t.h"
#include "beinsn_t.h"
-#include "besched_t.h"
-#include "benodesets.h"
-#include "bejavacoal.h"
+#include "besched.h"
#include "bestatevent.h"
+#include "beirg.h"
+#include "error.h"
-#ifdef WITH_LIBCORE
-#include <libcore/lc_timing.h>
-#include <libcore/lc_opts.h>
-#endif /* WITH_LIBCORE */
+#include "lc_opts.h"
+#include "lc_opts_enum.h"
#define DUMP_BEFORE 1
#define DUMP_AFTER 2
#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 int improve = 1;
-#ifdef WITH_LIBCORE
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "before", DUMP_BEFORE },
{ "after", DUMP_AFTER },
{ NULL, 0 }
};
-static const lc_opt_enum_mask_items_t algo_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[] = {
&style_flags, style_items
};
-static lc_opt_enum_mask_var_t algo_var = {
- &algo, algo_items
-};
-
static lc_opt_enum_func_ptr_var_t cost_func_var = {
(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", &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 }
+ LC_OPT_ENT_BOOL ("improve", "run heur1 before if algo can exploit start solutions", &improve),
+ LC_OPT_LAST
};
-/* Insert additional options registration functions here. */
-extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
-extern void be_co2_register_options(lc_opt_entry_t *grp);
-extern void be_co3_register_options(lc_opt_entry_t *grp);
+static be_module_list_entry_t *copyopts = NULL;
+static const co_algo_info *selected_copyopt = NULL;
+
+void be_register_copyopt(const char *name, co_algo_info *copyopt)
+{
+ if (selected_copyopt == NULL)
+ selected_copyopt = copyopt;
+ be_add_module_to_list(©opts, name, copyopt);
+}
-void co_register_options(lc_opt_entry_t *grp)
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt)
+void be_init_copyopt(void)
{
- lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
+ 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");
+
lc_opt_add_table(co_grp, options);
+ be_add_module_list_opt(co_grp, "algo", "select copy optimization algo",
+ ©opts, (void**) &selected_copyopt);
+}
- be_co2_register_options(co_grp);
- be_co3_register_options(co_grp);
-#ifdef WITH_ILP
- be_co_ilp_register_options(co_grp);
-#endif
+static int void_algo(copy_opt_t *co)
+{
+ (void) co;
+ return 0;
}
-#endif
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone)
+void be_init_copynone(void)
+{
+ static co_algo_info copyheur = {
+ void_algo, 0
+ };
+
+ be_register_copyopt("none", ©heur);
+}
#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 {
+ be_lv_t *lv = be_get_irg_liveness(env->irg);
+ return be_values_interfere(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)
{
const char *s1, *s2, *s3;
- int len;
+ size_t len;
copy_opt_t *co;
FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
- co = xcalloc(1, sizeof(*co));
+ co = XMALLOCZ(copy_opt_t);
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();
+ s1 = get_irp_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);
+ co->name = XMALLOCN(char, len);
snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
return co;
}
-void free_copy_opt(copy_opt_t *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;
+/**
+ * Checks if a node is optimizable, viz. has something to do with coalescing
+ * @param irn The irn to check
+ */
+static int co_is_optimizable_root(ir_node *irn)
+{
+ const arch_register_req_t *req;
- reg = arch_get_irn_register(co->aenv, irn);
- if (arch_register_type_is(reg, ignore))
+ if (arch_irn_is_ignore(irn))
return 0;
- if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
+ if (is_Reg_Phi(irn) || is_Perm_Proj(irn))
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;
- }
- }
+ req = arch_get_register_req_out(irn);
+ if (is_2addr_code(req))
+ return 1;
return 0;
}
-int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
+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 */
return 1+cost;
}
-int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
+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;
- res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
+ ir_exec_freq *exec_freq = be_get_irg_exec_freq(co->cenv->irg);
+ (void) arg;
+ res = get_block_execfreq_ulong(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)
+{
+ (void) co;
+ (void) root;
+ (void) arg;
+ (void) pos;
return 1;
}
* Determines a maximum weighted independent set with respect to
* the interference and conflict edges of all nodes in a qnode.
*/
-static int ou_max_ind_set_costs(unit_t *ou) {
+static int ou_max_ind_set_costs(unit_t *ou)
+{
be_chordal_env_t *chordal_env = ou->co->cenv;
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;
+ size_t pos;
+ int curr_weight, best_weight = 0;
/* assign the nodes into two groups.
* safe: node has no interference, hence it is in every max stable set.
* unsafe: node has an interference
*/
- safe = alloca((ou->node_count-1) * sizeof(*safe));
- safe_costs = 0;
- safe_count = 0;
- unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
- unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
+ safe = ALLOCAN(ir_node*, ou->node_count - 1);
+ safe_costs = 0;
+ safe_count = 0;
+ unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
+ unsafe_costs = ALLOCAN(int, ou->node_count - 1);
unsafe_count = 0;
- for(i=1; i<ou->node_count; ++i) {
+ for (i=1; i<ou->node_count; ++i) {
int is_safe = 1;
- for(o=1; o<ou->node_count; ++o) {
+ for (o=1; o<ou->node_count; ++o) {
if (i==o)
continue;
if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
/* now compute the best set out of the unsafe nodes*/
if (unsafe_count > MIS_HEUR_TRIGGER) {
bitset_t *best = bitset_alloca(unsafe_count);
- /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
+ /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
for (i=0; i<unsafe_count; ++i) {
bitset_set(best, i);
/* check if it is a stable set */
/* Exact Algorithm: Brute force */
curr = bitset_alloca(unsafe_count);
bitset_set_all(curr);
- while ((max = bitset_popcnt(curr)) != 0) {
+ while (bitset_popcount(curr) != 0) {
/* check if curr is a stable set */
for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
for (o=bitset_next_set(curr, i+1); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to qnode_max_ind_set(): NOT (curr, i) */
goto no_stable_set;
/* if we arrive here, we have a stable set */
- /* compute the weigth of the stable set*/
+ /* compute the weight of the stable set*/
curr_weight = 0;
bitset_foreach(curr, pos)
curr_weight += unsafe_costs[pos];
return safe_costs+best_weight;
}
-static void co_collect_units(ir_node *irn, void *env) {
- copy_opt_t *co = env;
+static void co_collect_units(ir_node *irn, void *env)
+{
+ const arch_register_req_t *req;
+ copy_opt_t *co = (copy_opt_t*)env;
unit_t *unit;
- arch_register_req_t req;
- if (!is_curr_reg_class(co, irn))
+ if (get_irn_mode(irn) == mode_T)
+ return;
+ req = arch_get_register_req_out(irn);
+ if (req->cls != co->cls)
return;
- if (!co_is_optimizable_root(co, irn))
+ if (!co_is_optimizable_root(irn))
return;
/* Init a new unit */
- unit = xcalloc(1, sizeof(*unit));
+ unit = XMALLOCZ(unit_t);
unit->co = co;
unit->node_count = 1;
INIT_LIST_HEAD(&unit->queue);
/* init */
arity = get_irn_arity(irn);
- unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
- unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
+ unit->nodes = XMALLOCN(ir_node*, arity + 1);
+ unit->costs = XMALLOCN(int, arity + 1);
unit->nodes[0] = irn;
/* fill */
int o, arg_pos;
ir_node *arg = get_irn_n(irn, i);
- assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
+ assert(arch_get_irn_reg_class_out(arg) == co->cls && "Argument not in same register class.");
if (arg == irn)
continue;
if (nodes_interfere(co->cenv, irn, arg)) {
/* Else insert the argument of the phi to the members of this ou */
DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
+ if (arch_irn_is_ignore(arg))
+ continue;
+
/* Check if arg has occurred at a prior position in the arg/list */
arg_pos = 0;
- for (o=0; o<unit->node_count; ++o)
+ 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 */
+ } 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));
- unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
- } else
-
- /* Proj of a perm with corresponding arg */
- if (is_Perm_Proj(co->aenv, irn)) {
+ unit->nodes = XREALLOC(unit->nodes, ir_node*, unit->node_count);
+ unit->costs = XREALLOC(unit->costs, int, unit->node_count);
+ } else if (is_Perm_Proj(irn)) {
+ /* Proj of a perm with corresponding arg */
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->nodes = XMALLOCN(ir_node*, 2);
+ unit->costs = XMALLOCN(int, 2);
unit->node_count = 2;
unit->nodes[0] = irn;
unit->nodes[1] = get_Perm_src(irn);
unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
- } else
+ } else {
+ /* Src == Tgt of a 2-addr-code instruction */
+ if (is_2addr_code(req)) {
+ const unsigned other = req->other_same;
+ int count = 0;
+ int i;
+
+ for (i = 0; (1U << i) <= other; ++i) {
+ if (other & (1U << i)) {
+ ir_node *o = get_irn_n(skip_Proj(irn), i);
+ if (arch_irn_is_ignore(o))
+ continue;
+ if (nodes_interfere(co->cenv, irn, o))
+ continue;
+ ++count;
+ }
+ }
- /* 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);
+ if (count != 0) {
+ int k = 0;
+ ++count;
+ unit->nodes = XMALLOCN(ir_node*, count);
+ unit->costs = XMALLOCN(int, count);
+ unit->node_count = count;
+ unit->nodes[k++] = irn;
+
+ for (i = 0; 1U << i <= other; ++i) {
+ if (other & (1U << i)) {
+ ir_node *o = get_irn_n(skip_Proj(irn), i);
+ if (!arch_irn_is_ignore(o) &&
+ !nodes_interfere(co->cenv, irn, o)) {
+ unit->nodes[k] = o;
+ unit->costs[k] = co->get_costs(co, irn, o, -1);
+ ++k;
+ }
+ }
+ }
+ }
+ } 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) {
struct list_head *tmp;
/* Determine the maximum costs this unit can cause: all_nodes_cost */
- for(i=1; i<unit->node_count; ++i) {
+ for (i=1; i<unit->node_count; ++i) {
unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
unit->all_nodes_costs += unit->costs[i];
}
#ifdef QUICK_AND_DIRTY_HACK
-static int compare_ous(const void *k1, const void *k2) {
+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);
+ arch_get_register_req_out(&req, u1->nodes[i]);
if (arch_register_req_is(&req, limited)) {
u1_has_constr = 1;
break;
u2_has_constr = 0;
for (i=0; i<u2->node_count; ++i) {
- arch_get_register_req(aenv, &req, u2->nodes[i], -1);
+ arch_get_register_req_out(&req, u2->nodes[i]);
if (arch_register_req_is(&req, limited)) {
u2_has_constr = 1;
break;
/**
* Sort the ou's according to constraints and their sort_key
*/
-static void co_sort_units(copy_opt_t *co) {
+static void co_sort_units(copy_opt_t *co)
+{
int i, count = 0, costs;
unit_t *ou, **ous;
/* 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));
+ ous = ALLOCAN(unit_t, count);
costs = co_get_max_copy_costs(co);
}
#endif
-void co_build_ou_structure(copy_opt_t *co) {
+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);
#endif
}
-void co_free_ou_structure(copy_opt_t *co) {
+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) {
/* co_solve_heuristic() is implemented in becopyheur.c */
-int co_get_max_copy_costs(const copy_opt_t *co) {
+int co_get_max_copy_costs(const copy_opt_t *co)
+{
int i, res = 0;
unit_t *curr;
return res;
}
-int co_get_inevit_copy_costs(const copy_opt_t *co) {
+int co_get_inevit_copy_costs(const copy_opt_t *co)
+{
int res = 0;
unit_t *curr;
return res;
}
-int co_get_copy_costs(const copy_opt_t *co) {
+int co_get_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) {
- int root_col = get_irn_col(co, curr->nodes[0]);
+ int root_col = get_irn_col(curr->nodes[0]);
DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
res += curr->inevitable_costs;
for (i=1; i<curr->node_count; ++i) {
- int arg_col = get_irn_col(co, curr->nodes[i]);
+ int arg_col = get_irn_col(curr->nodes[i]);
if (root_col != arg_col) {
DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
res += curr->costs[i];
return res;
}
-int co_get_lower_bound(const copy_opt_t *co) {
+int co_get_lower_bound(const copy_opt_t *co)
+{
int res = 0;
unit_t *curr;
stat->aff_nodes += 1;
bitset_add_irn(seen, an->irn);
co_gs_foreach_neighb(an, neigh) {
- if(!bitset_contains_irn(seen, neigh->irn)) {
+ 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)) {
+ if (get_irn_col(an->irn) != get_irn_col(neigh->irn)) {
stat->costs += neigh->costs;
stat->unsatisfied_edges += 1;
}
- if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
+ if (nodes_interfere(co->cenv, an->irn, neigh->irn)) {
stat->aff_int += 1;
stat->inevit_costs += neigh->costs;
}
|_| |___/
******************************************************************************/
-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;
+static int compare_affinity_node_t(const void *k1, const void *k2, size_t size)
+{
+ const affinity_node_t *n1 = (const affinity_node_t*)k1;
+ const affinity_node_t *n2 = (const affinity_node_t*)k2;
+ (void) size;
return (n1->irn != n2->irn);
}
-static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
+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));
+ node = (affinity_node_t*)set_insert(co->nodes, &new_node, sizeof(new_node), hash_irn(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 = OALLOC(&co->obst, neighb_t);
nbr->irn = n2;
nbr->costs = 0;
nbr->next = node->neighbours;
+
node->neighbours = nbr;
node->degree++;
}
nbr->costs += costs;
}
-static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
+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);
}
}
-static void build_graph_walker(ir_node *irn, void *env) {
- copy_opt_t *co = env;
+static void build_graph_walker(ir_node *irn, void *env)
+{
+ const arch_register_req_t *req;
+ copy_opt_t *co = (copy_opt_t*)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))
+ if (get_irn_mode(irn) == mode_T)
return;
-
- reg = arch_get_irn_register(co->aenv, irn);
- if (arch_register_type_is(reg, ignore))
+ req = arch_get_register_req_out(irn);
+ if (req->cls != co->cls || arch_irn_is_ignore(irn))
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(irn)) { /* Perms */
ir_node *arg = get_Perm_src(irn);
add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
+ } else { /* 2-address code */
+ if (is_2addr_code(req)) {
+ const unsigned other = req->other_same;
+ int i;
+
+ for (i = 0; 1U << i <= other; ++i) {
+ if (other & (1U << i)) {
+ ir_node *other = get_irn_n(skip_Proj(irn), i);
+ if (!arch_irn_is_ignore(other))
+ add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
+ }
+ }
+ }
}
-
- /* 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));
}
-void co_build_graph_structure(copy_opt_t *co) {
+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) {
+void co_free_graph_structure(copy_opt_t *co)
+{
ASSERT_GS_AVAIL(co);
del_set(co->nodes);
co->nodes = NULL;
}
-/* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
-
-int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
+int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn)
+{
affinity_node_t new_node, *n;
ASSERT_GS_AVAIL(co);
new_node.irn = irn;
- n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
+ n = (affinity_node_t*)set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
if (n) {
return (n->degree > 0);
} else
return 0;
}
+static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b)
+{
+ ir_node *nodes[] = { a, b };
+ bitset_t *constr[] = { NULL, NULL };
+ int j;
+
+ constr[0] = bitset_alloca(co->cls->n_regs);
+ constr[1] = bitset_alloca(co->cls->n_regs);
+
+ for (j = 0; j < 2; ++j) {
+ const arch_register_req_t *req = arch_get_register_req_out(nodes[j]);
+ if (arch_register_req_is(req, limited))
+ rbitset_copy_to_bitset(req->limited, constr[j]);
+ else
+ bitset_set_all(constr[j]);
+
+ }
+
+ return !bitset_intersect(constr[0], constr[1]);
+}
+
void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
{
- 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);
+ be_ifg_t *ifg = co->cenv->ifg;
+ int *color_map = ALLOCAN(int, co->cls->n_regs);
+ int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
+ ir_graph *irg = co->irg;
+ be_irg_t *birg = be_birg_from_irg(irg);
ir_node *irn;
- void *it, *nit;
- int i, n, n_regs;
+ nodes_iter_t it;
+ neighbours_iter_t nit;
+ int n, n_regs;
+ unsigned i;
n_regs = 0;
- for(i = 0; i < co->cls->n_regs; ++i) {
+ for (i = 0; i < co->cls->n_regs; ++i) {
const arch_register_t *reg = &co->cls->regs[i];
- color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
+ if (rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
+ color_map[i] = n_regs++;
+ } else {
+ color_map[i] = -1;
+ }
}
/*
* the values below n are the pre-colored register nodes
*/
- it = be_ifg_nodes_iter_alloca(ifg);
- nit = be_ifg_neighbours_iter_alloca(ifg);
-
n = n_regs;
- be_ifg_foreach_node(ifg, it, irn) {
- if(!arch_irn_is(co->aenv, irn, ignore))
- set_irn_link(irn, INT_TO_PTR(n++));
+ be_ifg_foreach_node(ifg, &it, irn) {
+ if (arch_irn_is_ignore(irn))
+ continue;
+ node_map[get_irn_idx(irn)] = n++;
}
fprintf(f, "%d %d\n", n, n_regs);
- be_ifg_foreach_node(ifg, it, irn) {
- if(!arch_irn_is(co->aenv, irn, ignore)) {
- int idx = PTR_TO_INT(get_irn_link(irn));
- affinity_node_t *a = get_affinity_info(co, irn);
-
- arch_register_req_t req;
- ir_node *adj;
+ be_ifg_foreach_node(ifg, &it, irn) {
+ if (!arch_irn_is_ignore(irn)) {
+ int idx = node_map[get_irn_idx(irn)];
+ affinity_node_t *a = get_affinity_info(co, irn);
+ const arch_register_req_t *req = arch_get_register_req_out(irn);
+ 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)
+ 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);
-
+ }
}
-
- be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
- if(!arch_irn_is(co->aenv, adj, ignore)) {
- int adj_idx = PTR_TO_INT(get_irn_link(adj));
- if(idx < adj_idx)
+ be_ifg_foreach_neighbour(ifg, &nit, irn, adj) {
+ if (!arch_irn_is_ignore(adj) &&
+ !co_dump_appel_disjoint_constraints(co, irn, adj)) {
+ int adj_idx = node_map[get_irn_idx(adj)];
+ if (idx < adj_idx)
fprintf(f, "%d %d -1\n", idx, adj_idx);
}
}
- if(a) {
+ if (a) {
neighb_t *n;
co_gs_foreach_neighb(a, n) {
- if(!arch_irn_is(co->aenv, n->irn, ignore)) {
- int n_idx = PTR_TO_INT(get_irn_link(n->irn));
- if(idx < n_idx)
+ if (!arch_irn_is_ignore(n->irn)) {
+ int n_idx = node_map[get_irn_idx(n->irn)];
+ if (idx < n_idx)
fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
}
}
}
}
}
-}
-typedef struct _appel_clique_walker_t {
- phase_t ph;
- const copy_opt_t *co;
- int curr_nr;
- int node_count;
- FILE *f;
- int dumb;
- int *color_map;
- struct obstack obst;
-} appel_clique_walker_t;
-
-typedef struct _appel_block_info_t {
- int *live_end_nr;
- int *live_in_nr;
- int *phi_nr;
- ir_node **live_end;
- ir_node **live_in;
- ir_node **phi;
- int n_live_end;
- int n_live_in;
- int n_phi;
-} appel_block_info_t;
-
-static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
-{
-#if 0
- double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
- int res = (int) freq;
- return res == 0 ? 1 : res;
-#else
- ir_loop *loop = get_irn_loop(bl);
- if(loop) {
- int d = get_loop_depth(loop);
- return 1 + d * d;
- }
- return 1;
-#endif
-}
-
-static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
-{
- appel_block_info_t *res = NULL;
-
- if(is_Block(irn)) {
- appel_clique_walker_t *d = (void *) phase;
- res = phase_alloc(phase, sizeof(res[0]));
- res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
- res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
- res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
- res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
- res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
- res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
- }
-
- return res;
-}
-
-typedef struct _insn_list_t {
- be_insn_t *insn;
- struct list_head list;
-} insn_list_t;
-
-static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
-{
- appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
- int i;
-
- for(i = 0; i < bli->n_live_end; ++i)
- if(bli->live_end[i] == irn)
- return bli->live_end_nr[i];
-
- return -1;
-}
-
-static int appel_dump_clique(appel_clique_walker_t *env, pset *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;
-
- n_live = 0;
- foreach_pset(live, irn)
- live_arr[n_live++] = irn;
-
- /* dump the live after clique */
- if(!env->dumb) {
- for(j = 0; j < n_live; ++j) {
- int k;
-
- for(k = j + 1; k < n_live; ++k) {
- fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
- }
- fprintf(env->f, "\n");
- }
- }
-
- /* dump the affinities */
- for(j = 0; !env->dumb && j < n_live; ++j) {
- ir_node *irn = live_arr[j];
- int old_nr = PTR_TO_INT(get_irn_link(irn));
-
- /* if the node was already live in the last insn dump the affinity */
- if(old_nr > start_nr) {
- int weight = appel_aff_weight(env, bl);
- fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
- }
- }
-
- /* set the current numbers into the link field. */
- for(j = 0; j < n_live; ++j) {
- ir_node *irn = live_arr[j];
- set_irn_link(irn, INT_TO_PTR(curr_nr + j));
- }
-
- return curr_nr + n_live;
-}
-
-static void appel_walker(ir_node *bl, void *data)
-{
- appel_clique_walker_t *env = data;
- 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();
- be_lv_t *lv = env->co->cenv->birg->lv;
-
- int n_insns = 0;
- int n_nodes = 0;
- int start_nr = env->curr_nr;
- int curr_nr = start_nr;
-
- be_insn_env_t insn_env;
- int i, j;
- ir_node *irn;
- be_insn_t **insns;
-
- insn_env.aenv = env->co->aenv;
- insn_env.cls = env->co->cls;
- insn_env.obst = obst;
- insn_env.ignore_colors = env->co->cenv->ignore_colors;
-
- /* Guess how many insns will be in this block. */
- sched_foreach(bl, irn)
- n_nodes++;
-
- bli->n_phi = 0;
- insns = malloc(n_nodes * sizeof(insns[0]));
-
- /* Put all insns in an array. */
- irn = sched_first(bl);
- while(!sched_is_end(irn)) {
- be_insn_t *insn;
- insn = be_scan_insn(&insn_env, irn);
- insns[n_insns++] = insn;
- 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);
-
- /* Generate the bad and ugly. */
- for(i = n_insns - 1; i >= 0; --i) {
- be_insn_t *insn = insns[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) {
- bli->live_end[j] = irn;
- bli->live_end_nr[j] = curr_nr + j;
- ++j;
- }
- bli->n_live_end = j;
- }
-
- if(!env->dumb) {
- for(j = 0; j < insn->use_start; ++j) {
- ir_node *op = insn->ops[j].carrier;
- bitset_t *adm = insn->ops[j].regs;
- int k;
- int nr;
-
- if(!insn->ops[j].has_constraints)
- continue;
-
- nr = 0;
- foreach_pset(live, irn) {
- if(irn == op) {
- pset_break(live);
- break;
- }
- ++nr;
- }
-
- assert(nr < pset_count(live));
-
- for(k = 0; k < env->co->cls->n_regs; ++k) {
- int mapped_col = env->color_map[k];
- if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
- fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
- }
- }
- }
-
- /* dump the clique and update the stuff. */
- 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);
-
- 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;
- bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
- bli->n_phi++;
- }
-
- /* add all uses */
- else
- for(j = insn->use_start; j < insn->n_ops; ++j)
- pset_insert_ptr(live, insn->ops[j].carrier);
- }
-
- /* print the start clique. */
- curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
-
- i = 0;
- foreach_pset(live, irn) {
- 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);
- free(insns);
- obstack_free(obst, base);
- env->curr_nr = curr_nr;
-}
-
-static void appel_inter_block_aff(ir_node *bl, void *data)
-{
- appel_clique_walker_t *env = data;
- appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
-
- int i, j, n;
-
- for(i = 0; i < bli->n_live_in; ++i) {
- ir_node *irn = bli->live_in[i];
-
- for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
- ir_node *pred = get_Block_cfgpred_block(bl, j);
-
- int nr = appel_get_live_end_nr(env, pred, irn);
- assert(nr >= 0);
- fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
- }
- }
-
- for(i = 0; i < bli->n_phi; ++i) {
- ir_node *irn = bli->phi[i];
-
- for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
- ir_node *pred = get_Block_cfgpred_block(bl, j);
- ir_node *op = get_irn_n(irn, j);
-
- int nr = appel_get_live_end_nr(env, pred, op);
- assert(nr >= 0);
- fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
- }
- }
-
-}
-
-void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
-{
- int 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(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;
- env.co = co;
- env.f = f;
-
- bitset_copy(adm, co->cenv->ignore_colors);
- bitset_flip_all(adm);
-
- /* Make color map. */
- env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
- for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
- const arch_register_t *reg = &co->cls->regs[i];
- env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
- }
-
- env.dumb = 1;
- env.curr_nr = n_colors;
- irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
- irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
-
- fprintf(f, "%d %d\n", env.curr_nr, n_colors);
-
- /* make the first k nodes interfere */
- for(i = 0; i < n_colors; ++i) {
- int j;
- for(j = i + 1; j < n_colors; ++j)
- fprintf(f, "%d %d -1 ", i, j);
- fprintf(f, "\n");
- }
-
- env.dumb = 0;
- env.curr_nr = n_colors;
- irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
- irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
- irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
- obstack_free(&env.obst, NULL);
+ xfree(node_map);
}
/*
|_| |___/
*/
-static const char *get_dot_color_name(int col)
+static const char *get_dot_color_name(size_t col)
{
static const char *names[] = {
"blue",
return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
}
-typedef struct _co_ifg_dump_t {
+typedef struct co_ifg_dump_t {
const copy_opt_t *co;
unsigned flags;
} co_ifg_dump_t;
static void ifg_dump_graph_attr(FILE *f, void *self)
{
+ (void) self;
fprintf(f, "overlap=scale");
}
static int ifg_is_dump_node(void *self, ir_node *irn)
{
- co_ifg_dump_t *cod = self;
- return !arch_irn_is(cod->co->aenv, irn, ignore);
+ (void)self;
+ return !arch_irn_is_ignore(irn);
}
static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
{
- co_ifg_dump_t *env = self;
- const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
- 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);
+ co_ifg_dump_t *env = (co_ifg_dump_t*)self;
+ const arch_register_t *reg = arch_get_irn_register(irn);
+ const arch_register_req_t *req = arch_get_register_req_out(irn);
+ int limited = arch_register_req_is(req, limited);
- if(env->flags & CO_IFG_DUMP_LABELS) {
+ if (env->flags & CO_IFG_DUMP_LABELS) {
ir_fprintf(f, "label=\"%+F", irn);
- if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
+ if ((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
- req.limited(req.limited_env, bs);
+ rbitset_copy_to_bitset(req->limited, bs);
ir_fprintf(f, "\\n%B", bs);
}
ir_fprintf(f, "\" ");
- }
-
- else
+ } else {
fprintf(f, "label=\"\" shape=point " );
+ }
- if(env->flags & CO_IFG_DUMP_SHAPE)
+ if (env->flags & CO_IFG_DUMP_SHAPE)
fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
- if(env->flags & CO_IFG_DUMP_COLORS)
+ if (env->flags & CO_IFG_DUMP_COLORS)
fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
}
static void ifg_dump_at_end(FILE *file, void *self)
{
- co_ifg_dump_t *env = self;
+ co_ifg_dump_t *env = (co_ifg_dump_t*)self;
affinity_node_t *a;
co_gs_foreach_aff_node(env->co, a) {
- const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
+ const arch_register_t *ar = arch_get_irn_register(a->irn);
unsigned aidx = get_irn_idx(a->irn);
neighb_t *n;
co_gs_foreach_neighb(a, n) {
- const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
+ const arch_register_t *nr = arch_get_irn_register(n->irn);
unsigned nidx = get_irn_idx(n->irn);
- if(aidx < nidx) {
+ if (aidx < nidx) {
const char *color = nr == ar ? "blue" : "red";
- fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
- if(env->flags & CO_IFG_DUMP_LABELS)
+ fprintf(file, "\tn%u -- n%u [weight=0.01 ", aidx, nidx);
+ if (env->flags & CO_IFG_DUMP_LABELS)
fprintf(file, "label=\"%d\" ", n->costs);
- if(env->flags & CO_IFG_DUMP_COLORS)
+ if (env->flags & CO_IFG_DUMP_COLORS)
fprintf(file, "color=%s ", color);
else
fprintf(file, "style=dotted");
void co_solve_park_moon(copy_opt_t *opt)
{
-
+ (void) opt;
}
-static int void_algo(copy_opt_t *co)
-{
- return 0;
-}
-
-/*
- _ _ _ _ _
- / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
- / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
- / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
- /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
- |___/
-*/
-
-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, "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];
+ size_t i, n;
+ char *tu_name;
+ const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
+
+ n = strlen(cup_name);
+ tu_name = XMALLOCN(char, n + 1);
+ strcpy(tu_name, cup_name);
+ for (i = 0; i < n; ++i)
+ if (tu_name[i] == '.')
+ tu_name[i] = '_';
+
+
+ ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
+ xfree(tu_name);
+ 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)
{
-#ifdef WITH_LIBCORE
- lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
-#endif
+ ir_timer_t *timer = ir_timer_new();
co_complete_stats_t before, after;
- copy_opt_t *co;
- co_algo_t *algo_func;
- int was_optimal = 0;
+ copy_opt_t *co;
+ int was_optimal = 0;
- if(algo < 0 || algo >= CO_ALGO_LAST)
+ assert(selected_copyopt);
+
+ /* skip copymin if algo is 'none' */
+ if (selected_copyopt->copyopt == void_algo)
return;
+ be_liveness_assure_chk(be_get_irg_liveness(cenv->irg));
+
co = new_copy_opt(cenv, cost_func);
co_build_ou_structure(co);
co_build_graph_structure(co);
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");
- co_dump_appel_graph(co, f);
- fclose(f);
- }
-
- if(dump_flags & DUMP_BEFORE) {
- FILE *f = be_chordal_open(cenv, "", "-before.dot");
+ 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 the algo can improve results, provide an initial solution with heur1 */
+ if (improve && selected_copyopt->can_improve_existing) {
co_complete_stats_t stats;
- /* produce a heuristical solution */
- co_solve_heuristic_java(co);
+ /* produce a heuristic solution */
+ co_solve_heuristic(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
+ /* perform actual copy minimization */
+ ir_timer_reset_and_start(timer);
+ was_optimal = selected_copyopt->copyopt(co);
+ ir_timer_stop(timer);
+ be_stat_ev("co_time", ir_timer_elapsed_msec(timer));
be_stat_ev_ull("co_optimal", was_optimal);
+ ir_timer_free(timer);
- if(dump_flags & DUMP_AFTER) {
- FILE *f = be_chordal_open(cenv, "", "-after.dot");
+ 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;
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)
+ if (optimizable_costs > 0)
printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
else
printf("%10" ULL_FMT " %5s\n", after.costs, "-");
}
+ /* Dump the interference graph in Appel's format. */
+ if (dump_flags & DUMP_APPEL) {
+ FILE *f = my_open(cenv, "", ".apl");
+ fprintf(f, "# %llu %llu\n", after.costs, after.unsatisfied_edges);
+ co_dump_appel_graph(co, f);
+ fclose(f);
+ }
+
be_stat_ev_ull("co_after_costs", after.costs);
be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);