#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"
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_HEUR4;
static int improve = 1;
static const lc_opt_enum_mask_items_t dump_items[] = {
{ NULL, 0 }
};
-static const lc_opt_enum_mask_items_t algo_items[] = {
- { "none", CO_ALGO_NONE },
- { "heur", CO_ALGO_HEUR },
- { "heur2", CO_ALGO_HEUR2 },
- { "heur3", CO_ALGO_HEUR3 },
- { "heur4", CO_ALGO_HEUR4 },
- { "ilp", CO_ALGO_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),
+ 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_init_copycoal(void)
+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);
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt);
+void be_init_copyopt(void)
{
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 *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_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
+static int void_algo(copy_opt_t *co)
+{
+ (void) co;
+ return 0;
+}
+
+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
return co;
}
-void free_copy_opt(copy_opt_t *co) {
+void free_copy_opt(copy_opt_t *co)
+{
xfree(co->name);
free(co);
}
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);
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;
}
-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;
* 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;
- bitset_pos_t pos;
- int max, curr_weight, best_weight = 0;
+ unsigned 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 = 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])) {
/* 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) */
static void co_collect_units(ir_node *irn, void *env)
{
- const arch_register_req_t *req = arch_get_register_req_out(irn);
+ const arch_register_req_t *req;
copy_opt_t *co = env;
unit_t *unit;
+ 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(irn))
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;
/**
* 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;
}
#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;
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;
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) {
+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) {
+static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
+{
affinity_node_t new_node, *node;
neighb_t *nbr;
int allocnew = 1;
/* if we did not find n2 in n1's neighbourhood insert it */
if (allocnew) {
- nbr = obstack_alloc(&co->obst, sizeof(*nbr));
+ nbr = OALLOC(&co->obst, neighb_t);
nbr->irn = n2;
nbr->costs = 0;
nbr->next = node->neighbours;
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) {
- const arch_register_req_t *req = arch_get_register_req_out(irn);
+static void build_graph_walker(ir_node *irn, void *env)
+{
+ const arch_register_req_t *req;
copy_opt_t *co = env;
int pos, max;
const arch_register_t *reg;
+ if (get_irn_mode(irn) == mode_T)
+ return;
+ req = arch_get_register_req_out(irn);
if (req->cls != co->cls || arch_irn_is_ignore(irn))
return;
}
}
-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_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);
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))
+ if (arch_register_req_is(req, limited))
rbitset_copy_to_bitset(req->limited, constr[j]);
else
bitset_set_all(constr[j]);
int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
ir_node *irn;
- void *it, *nit;
+ 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++;
}
* 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) {
+ 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) {
+ 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;
- 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)
+ 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) {
+ 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)
+ 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_ignore(n->irn)) {
int n_idx = node_map[get_irn_idx(n->irn)];
- if(idx < n_idx)
+ if (idx < n_idx)
fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
}
}
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);
rbitset_copy_to_bitset(req->limited, bs);
ir_fprintf(f, "\\n%B", bs);
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));
}
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) opt;
}
-static int void_algo(copy_opt_t *co)
-{
- (void) 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 },
-#ifdef WITH_JVM
- { co_solve_heuristic_java, "heur3", 0 },
-#else
- { NULL, "heur3", 0 },
-#endif
- { co_solve_heuristic_mst, "heur4", 0 },
-#ifdef WITH_ILP
- { co_solve_ilp2, "ilp", 1 },
-#else
- { NULL, "ilp", 1 },
-#endif
- { NULL, "", 0 }
-};
-
/*
__ __ _ ____ _
| \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
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) {
+ if (result == NULL) {
panic("Couldn't open '%s' for writing.", buf);
}
void co_driver(be_chordal_env_t *cenv)
{
- ir_timer_t *timer = ir_timer_register("firm.be.copyopt", "runtime");
+ 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;
- if (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_birg_liveness(cenv->birg));
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 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);
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 /* WITH_JVM */
-
- algo_func = algos[algo].algo;
-
/* perform actual copy minimization */
ir_timer_reset_and_start(timer);
- was_optimal = algo_func(co);
+ 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 = my_open(cenv, "", "-after.dot");
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, "# %lld %lld\n", after.costs, after.unsatisfied_edges);
+ fprintf(f, "# %llu %llu\n", after.costs, after.unsatisfied_edges);
co_dump_appel_graph(co, f);
fclose(f);
}
projects / libfirm / blobdiff