* - Register-constrained nodes
* - Two-address code instructions
*/
-#ifdef HAVE_CONFIG_H
#include "config.h"
-#endif
#include "execfreq.h"
#include "xmalloc.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_t.h"
-#include "benode_t.h"
+#include "bearch.h"
+#include "benode.h"
#include "beutil.h"
#include "beifg_t.h"
#include "beintlive_t.h"
#include "becopystat.h"
#include "belive_t.h"
#include "beinsn_t.h"
-#include "besched_t.h"
-#include "bejavacoal.h"
+#include "besched.h"
#include "bestatevent.h"
-#include "beirg_t.h"
+#include "beirg.h"
#include "error.h"
#include "lc_opts.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_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 be_init_copycoal(void)
+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);
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt);
+
+static int void_algo(copy_opt_t *co)
+{
+ (void) co;
+ return 0;
+}
+
+void be_init_copynone(void)
+{
+ static co_algo_info copyheur = {
+ void_algo, 0
+ };
+
+ be_register_copyopt("none", ©heur);
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone);
#undef QUICK_AND_DIRTY_HACK
if (env->ifg)
return be_ifg_connected(env->ifg, a, b);
else
- return values_interfere(env->birg, a, b);
+ return be_values_interfere(env->birg->lv, a, b);
}
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;
free(co);
}
-int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
+/**
+ * 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;
- const arch_register_t *reg;
+ const arch_register_t *reg;
- if (arch_irn_is(co->aenv, irn, ignore))
+ if (arch_irn_is_ignore(irn))
return 0;
- reg = arch_get_irn_register(co->aenv, irn);
+ reg = arch_get_irn_register(irn);
if (arch_register_type_is(reg, ignore))
return 0;
- req = arch_get_register_req(irn, -1);
- if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req))
+ if (is_Reg_Phi(irn) || is_Perm_Proj(irn))
+ return 1;
+
+ req = arch_get_register_req_out(irn);
+ if (is_2addr_code(req))
return 1;
return 0;
* 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) {
int is_safe = 1;
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 = arch_get_register_req_out(irn);
+ copy_opt_t *co = env;
unit_t *unit;
- if (!is_curr_reg_class(co, 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 */
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(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 (arch_irn_is_ignore(arg))
+ continue;
- 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);
+ /* 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 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, ir_node*, unit->node_count);
unit->costs = XREALLOC(unit->costs, int, unit->node_count);
- } else if (is_Perm_Proj(co->aenv, irn)) {
+ } 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 = XMALLOCN(ir_node*, 2);
unit->nodes[1] = get_Perm_src(irn);
unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
} else {
- const arch_register_req_t *req = arch_get_register_req(irn, -1);
-
/* Src == Tgt of a 2-addr-code instruction */
if (is_2addr_code(req)) {
const unsigned other = req->other_same;
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(co->aenv, o, ignore) &&
- !nodes_interfere(co->cenv, irn, o)) {
- ++count;
- }
+ if (arch_irn_is_ignore(o))
+ continue;
+ if (nodes_interfere(co->cenv, irn, o))
+ continue;
+ ++count;
}
}
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(co->aenv, o, ignore) &&
+ 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);
/* Units with constraints come first */
u1_has_constr = 0;
for (i=0; i<u1->node_count; ++i) {
- arch_get_register_req(&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(&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;
/* 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);
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];
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 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) {
- copy_opt_t *co = env;
+ const arch_register_req_t *req = arch_get_register_req_out(irn);
+ copy_opt_t *co = env;
int pos, max;
const arch_register_t *reg;
- if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
+ if (req->cls != co->cls || arch_irn_is_ignore(irn))
return;
- reg = arch_get_irn_register(co->aenv, irn);
+ reg = arch_get_irn_register(irn);
if (arch_register_type_is(reg, ignore))
return;
ir_node *arg = get_irn_n(irn, pos);
add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
}
- }
- else if (is_Perm_Proj(co->aenv, irn)) { /* Perms */
+ } 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 */
- const arch_register_req_t *req = arch_get_register_req(irn, -1);
+ } 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(co->aenv, other, ignore))
+ if (!arch_irn_is_ignore(other))
add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
}
}
{
ir_node *nodes[] = { a, b };
bitset_t *constr[] = { NULL, NULL };
- const arch_register_req_t *req;
int j;
constr[0] = bitset_alloca(co->cls->n_regs);
constr[1] = bitset_alloca(co->cls->n_regs);
for (j = 0; j < 2; ++j) {
- req = arch_get_register_req(nodes[j], BE_OUT_POS(0));
+ 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
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]));
- int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
+ 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_node *irn;
void *it, *nit;
n = n_regs;
be_ifg_foreach_node(ifg, it, irn) {
- if(!arch_irn_is(co->aenv, irn, ignore))
- node_map[get_irn_idx(irn)] = n++;
+ 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 = node_map[get_irn_idx(irn)];
- affinity_node_t *a = get_affinity_info(co, 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;
- const arch_register_req_t *req;
- ir_node *adj;
-
- req = arch_get_register_req(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)
}
be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
- if(!arch_irn_is(co->aenv, adj, ignore) && !co_dump_appel_disjoint_constraints(co, 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);
neighb_t *n;
co_gs_foreach_neighb(a, n) {
- if(!arch_irn_is(co->aenv, n->irn, ignore)) {
+ 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);
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);
- const arch_register_req_t *req;
- int limited;
-
- req = arch_get_register_req(irn, BE_OUT_POS(0));
- limited = arch_register_req_is(req, limited);
+ co_ifg_dump_t *env = 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) {
ir_fprintf(f, "label=\"%+F", irn);
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) {
(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_timer_t *timer = ir_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;
- if (algo >= CO_ALGO_LAST)
- return;
+ assert(selected_copyopt);
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));