no strange stuff for including alloca anymore, use xmalloc.h
[libfirm] / ir / be / becopyopt.c
index a6317d2..2de1592 100644 (file)
@@ -7,17 +7,12 @@
 #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 "irgraph.h"
 #include "irgwalk.h"
 #include "irprog.h"
@@ -28,7 +23,7 @@
 #include "irphase_t.h"
 #include "irprintf_t.h"
 
-
+#include "bemodule.h"
 #include "bearch.h"
 #include "benode_t.h"
 #include "beutil.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
@@ -63,7 +59,6 @@ static cost_fct_t cost_func   = co_get_costs_exec_freq;
 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  },
@@ -134,22 +129,29 @@ 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);
 
-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->lv, a, b);
+}
+
+
 /******************************************************************************
     _____                           _
    / ____|                         | |
@@ -163,9 +165,6 @@ void co_register_options(lc_opt_entry_t *grp)
 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;
@@ -197,7 +196,7 @@ void free_copy_opt(copy_opt_t *co) {
 }
 
 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))
@@ -207,7 +206,8 @@ int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
        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;
@@ -230,14 +230,16 @@ int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
                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;
+                       }
                }
        }
 
@@ -267,7 +269,7 @@ int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, in
        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;
@@ -379,7 +381,6 @@ static int ou_max_ind_set_costs(unit_t *ou) {
 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;
@@ -449,21 +450,26 @@ static void co_collect_units(ir_node *irn, void *env) {
                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) {
@@ -744,7 +750,6 @@ static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs
 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))
@@ -768,8 +773,15 @@ static void build_graph_walker(ir_node *irn, void *env) {
        }
 
        /* 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) {
@@ -806,7 +818,6 @@ 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);
 
        ir_node *irn;
        void *it, *nit;
@@ -839,17 +850,15 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
                        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);
-
+                               }
                        }
 
 
@@ -1001,6 +1010,7 @@ static void appel_walker(ir_node *bl, void *data)
        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;
@@ -1034,7 +1044,7 @@ static void appel_walker(ir_node *bl, void *data)
        }
 
        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);
+       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) {
@@ -1156,8 +1166,9 @@ void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
        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);
        env.curr_nr = co->cls->n_regs;
@@ -1265,25 +1276,27 @@ 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;
+       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");
@@ -1388,11 +1401,23 @@ static co_algo_info_t algos[] = {
 
 */
 
+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;
@@ -1418,13 +1443,13 @@ void co_driver(be_chordal_env_t *cenv)
 
        /* 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);
        }
@@ -1449,21 +1474,17 @@ void co_driver(be_chordal_env_t *cenv)
 
        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);
        }