#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 "irgraph.h"
#include "irgwalk.h"
#include "irprog.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
static unsigned algo = CO_ALGO_HEUR2;
static int improve = 1;
-#ifdef WITH_LIBCORE
static const lc_opt_enum_mask_items_t dump_items[] = {
{ "before", DUMP_BEFORE },
{ "after", DUMP_AFTER },
}
BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
-#endif
-
#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->lv, a, b);
+}
+
+
/******************************************************************************
_____ _
/ ____| | |
}
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;
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);
+ const arch_register_req_t *req;
+ req = arch_get_register_req(co->aenv, n, -1);
if(is_Reg_Phi(n) ||
is_Perm(co->aenv, n) ||
- (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
- )
- return 1;
+ (arch_register_req_is(req, should_be_same))) {
+ ir_node *other = get_irn_n(irn, req->other_same);
+ if(other == irn)
+ return 1;
+ }
}
}
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;
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(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 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(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;
{
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");
*/
+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
co_complete_stats_t before, after;
copy_opt_t *co;
co_algo_t *algo_func;
/* Dump the interference graph in Appel's format. */
if(dump_flags & DUMP_APPEL) {
- FILE *f = be_chordal_open(cenv, "", ".apl");
+ 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");
+ FILE *f = my_open(cenv, "", "-before.dot");
co_dump_ifg_dot(co, f, style_flags);
fclose(f);
}
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
be_stat_ev_ull("co_optimal", was_optimal);
if(dump_flags & DUMP_AFTER) {
- FILE *f = be_chordal_open(cenv, "", "-after.dot");
+ FILE *f = my_open(cenv, "", "-after.dot");
co_dump_ifg_dot(co, f, style_flags);
fclose(f);
}