+/*
+ * 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"
#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 "irphase_t.h"
#include "irprintf_t.h"
-
-#include "bearch.h"
+#include "bemodule.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"
#include "benodesets.h"
#include "bejavacoal.h"
#include "bestatevent.h"
+#include "beirg_t.h"
+#include "error.h"
-#ifdef WITH_LIBCORE
#include <libcore/lc_timing.h>
#include <libcore/lc_opts.h>
-#endif /* WITH_LIBCORE */
+#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 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;
-#ifdef WITH_LIBCORE
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "before", DUMP_BEFORE },
{ "after", DUMP_AFTER },
#ifdef WITH_JVM
{ "heur3", CO_ALGO_HEUR3 },
#endif /* WITH_JVM */
+ { "heur4", CO_ALGO_HEUR4 },
#ifdef WITH_ILP
{ "ilp", CO_ALGO_ILP },
#endif /* WITH_ILP */
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, 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 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;
}
-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(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
int cost = 0;
ir_loop *loop;
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->exec_freq, copy_bl);
+ res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
/* don't allow values smaller than one. */
return res < 1 ? 1 : res;
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;
/* 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 occured 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->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(skip_Proj(irn), req->other_same);
+ if (!arch_irn_is(co->aenv, other, ignore) &&
+ !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) {
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++;
}
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(skip_Proj(irn), req->other_same);
+ if (! arch_irn_is(co->aenv, other, ignore))
+ 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);
-
+ }
}
}
typedef struct _appel_clique_walker_t {
- phase_t ph;
+ ir_phase ph;
const copy_opt_t *co;
int curr_nr;
int node_count;
#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;
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;
irn = insn->next_insn;
}
- 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);
+ DBG((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) {
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);
+ 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;
{
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");
*/
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 }
};
/*
*/
+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)
{
-#ifdef WITH_LIBCORE
- lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
-#endif
+ 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 < 0 || algo >= CO_ALGO_LAST)
return;
co = new_copy_opt(cenv, cost_func);
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");
+ if (dump_flags & DUMP_APPEL) {
+ 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");
+ 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;
-#ifdef WITH_LIBCORE
+ /* perform actual copy minimization */
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("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");
+ 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