fix handling of default pns in the backend (they don't need table entries)
[libfirm] / ir / be / becopyopt.c
index 8464541..fe51312 100644 (file)
@@ -38,6 +38,7 @@
 #include "debug.h"
 #include "pmap.h"
 #include "raw_bitset.h"
+#include "irnode.h"
 #include "irgraph.h"
 #include "irgwalk.h"
 #include "irprog.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 "beinsn_t.h"
 #include "besched_t.h"
-#include "benodesets.h"
 #include "bejavacoal.h"
 #include "bestatevent.h"
 #include "beirg_t.h"
 #include "error.h"
-#include "bera.h"
 
 #include <libcore/lc_timing.h>
 #include <libcore/lc_opts.h>
@@ -106,13 +106,9 @@ static const lc_opt_enum_mask_items_t algo_items[] = {
        { "none",   CO_ALGO_NONE  },
        { "heur",   CO_ALGO_HEUR  },
        { "heur2",  CO_ALGO_HEUR2 },
-#ifdef WITH_JVM
        { "heur3",  CO_ALGO_HEUR3 },
-#endif /* WITH_JVM */
        { "heur4",  CO_ALGO_HEUR4 },
-#ifdef WITH_ILP
        { "ilp",    CO_ALGO_ILP   },
-#endif /* WITH_ILP */
        { NULL,     0 }
 };
 
@@ -148,7 +144,7 @@ static const lc_opt_table_entry_t options[] = {
        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),
-       { NULL }
+       LC_OPT_LAST
 };
 
 /* Insert additional options registration functions here. */
@@ -175,7 +171,7 @@ static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const
        if (env->ifg)
                return be_ifg_connected(env->ifg, a, b);
        else
-               return values_interfere(env->birg->lv, a, b);
+               return values_interfere(env->birg, a, b);
 }
 
 
@@ -240,43 +236,12 @@ int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
        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) {
-                       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))) {
-                               ir_node *other = get_irn_n(irn, req->other_same);
-                               if(other == 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;
        ir_node *root_block = get_nodes_block(root);
+       (void) co;
+       (void) arg;
 
        if (is_Phi(root)) {
                /* for phis the copies are placed in the corresponding pred-block */
@@ -296,6 +261,7 @@ 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;
+       (void) arg;
        res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
 
        /* don't allow values smaller than one. */
@@ -304,6 +270,10 @@ int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, in
 
 
 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;
+       (void) pos;
        return 1;
 }
 
@@ -327,7 +297,8 @@ static int ou_max_ind_set_costs(unit_t *ou) {
        ir_node **safe, **unsafe;
        int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
        bitset_t *curr;
-       int max, pos, curr_weight, best_weight = 0;
+       bitset_pos_t pos;
+       int max, curr_weight, best_weight = 0;
 
        /* assign the nodes into two groups.
         * safe: node has no interference, hence it is in every max stable set.
@@ -461,7 +432,7 @@ static void co_collect_units(ir_node *irn, void *env) {
                                        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 */
+                               } else { /* arg has occurred before in same phi */
                                        /* increase costs for existing arg */
                                        unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
                                }
@@ -469,10 +440,8 @@ static void co_collect_units(ir_node *irn, void *env) {
                }
                unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
                unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
-       } else
-
-       /* Proj of a perm with corresponding arg */
-       if (is_Perm_Proj(co->aenv, irn)) {
+       } else if (is_Perm_Proj(co->aenv, irn)) {
+               /* Proj of a perm with corresponding arg */
                assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
                unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
                unit->costs = xmalloc(2 * sizeof(*unit->costs));
@@ -486,15 +455,39 @@ static void co_collect_units(ir_node *irn, void *env) {
 
                /* 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);
+                       const unsigned other = req->other_same;
+                       int            count = 0;
+                       int            i;
+
+                       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 (count != 0) {
+                               int k = 0;
+                               ++count;
+                               unit->nodes = xmalloc(count * sizeof(*unit->nodes));
+                               unit->costs = xmalloc(count * sizeof(*unit->costs));
+                               unit->node_count = count;
+                               unit->nodes[k++] = irn;
+
+                               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)) {
+                                                       unit->nodes[k] = o;
+                                                       unit->costs[k] = co->get_costs(co, irn, o, -1);
+                                                       ++k;
+                                               }
+                                       }
+                               }
                        }
                } else {
                        assert(0 && "This is not an optimizable node!");
@@ -734,21 +727,21 @@ void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
 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) {
        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));
+       node = set_insert(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
 
-       allocnew = 1;
        for (nbr = node->neighbours; nbr; nbr = nbr->next)
                if (nbr->irn == n2) {
                        allocnew = 0;
@@ -757,11 +750,11 @@ static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
 
        /* 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++;
        }
@@ -789,27 +782,29 @@ static void build_graph_walker(ir_node *irn, void *env) {
        if (arch_register_type_is(reg, ignore))
                return;
 
-       /* Phis */
-       if (is_Reg_Phi(irn))
+       if (is_Reg_Phi(irn)) { /* Phis */
                for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
                        ir_node *arg = get_irn_n(irn, pos);
                        add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
                }
-
-       /* Perms */
-       else if (is_Perm_Proj(co->aenv, irn)) {
+       }
+       else if (is_Perm_Proj(co->aenv, irn)) { /* Perms */
                ir_node *arg = get_Perm_src(irn);
                add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
        }
-
-       /* 2-address code */
-       else {
-               const arch_register_req_t *req =
-                       arch_get_register_req(co->aenv, irn, -1);
+       else { /* 2-address code */
+               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));
+                       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))
+                                               add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
+                               }
+                       }
                }
        }
 }
@@ -837,7 +832,7 @@ int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
        ASSERT_GS_AVAIL(co);
 
        new_node.irn = irn;
-       n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
+       n = set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
        if (n) {
                return (n->degree > 0);
        } else
@@ -851,7 +846,8 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
 
        ir_node *irn;
        void *it, *nit;
-       int i, n, n_regs;
+       int n, n_regs;
+       unsigned i;
 
        n_regs = 0;
        for(i = 0; i < co->cls->n_regs; ++i) {
@@ -946,6 +942,7 @@ static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
        return res == 0 ? 1 : res;
 #else
        ir_loop *loop = get_irn_loop(bl);
+       (void) env;
        if(loop) {
                int d = get_loop_depth(loop);
                return 1 + d * d;
@@ -957,6 +954,7 @@ static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
 static void *appel_clique_walker_irn_init(ir_phase *phase, ir_node *irn, void *old)
 {
        appel_block_info_t *res = NULL;
+       (void) old;
 
        if(is_Block(irn)) {
                appel_clique_walker_t *d = (void *) phase;
@@ -989,15 +987,16 @@ static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_nod
        return -1;
 }
 
-static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
+static int appel_dump_clique(appel_clique_walker_t *env, const ir_nodeset_t *live, ir_node *bl, int curr_nr, int start_nr)
 {
        ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
        ir_node *irn;
        int n_live;
        int j;
+       ir_nodeset_iterator_t iter;
 
        n_live = 0;
-       foreach_pset(live, irn)
+       foreach_ir_nodeset(live, irn, iter)
                live_arr[n_live++] = irn;
 
        /* dump the live after clique */
@@ -1039,7 +1038,8 @@ static void appel_walker(ir_node *bl, void *data)
        appel_block_info_t *bli    = phase_get_or_set_irn_data(&env->ph, bl);
        struct obstack *obst       = &env->obst;
        void *base                 = obstack_base(obst);
-       pset *live                 = pset_new_ptr_default();
+       ir_nodeset_t live;
+       ir_nodeset_iterator_t iter;
        be_lv_t *lv                = env->co->cenv->birg->lv;
 
        int n_insns  = 0;
@@ -1062,7 +1062,7 @@ static void appel_walker(ir_node *bl, void *data)
                n_nodes++;
 
        bli->n_phi = 0;
-       insns = malloc(n_nodes * sizeof(insns[0]));
+       insns = xmalloc(n_nodes * sizeof(insns[0]));
 
        /* Put all insns in an array. */
        irn = sched_first(bl);
@@ -1074,7 +1074,8 @@ static void appel_walker(ir_node *bl, void *data)
        }
 
        DBG((dbg, LEVEL_2, "%+F\n", bl));
-       be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
+       ir_nodeset_init(&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) {
@@ -1083,7 +1084,7 @@ static void appel_walker(ir_node *bl, void *data)
                /* The first live set has to be saved in the block border set. */
                if(i == n_insns - 1) {
                        j = 0;
-                       foreach_pset(live, irn) {
+                       foreach_ir_nodeset(&live, irn, iter) {
                                bli->live_end[j]    = irn;
                                bli->live_end_nr[j] = curr_nr + j;
                                ++j;
@@ -1095,22 +1096,21 @@ static void appel_walker(ir_node *bl, void *data)
                        for(j = 0; j < insn->use_start; ++j) {
                                ir_node *op   = insn->ops[j].carrier;
                                bitset_t *adm = insn->ops[j].regs;
-                               int k;
-                               int nr;
+                               unsigned k;
+                               size_t nr;
 
                                if(!insn->ops[j].has_constraints)
                                        continue;
 
                                nr = 0;
-                               foreach_pset(live, irn) {
+                               foreach_ir_nodeset(&live, irn, iter) {
                                        if(irn == op) {
-                                               pset_break(live);
                                                break;
                                        }
                                        ++nr;
                                }
 
-                               assert(nr < pset_count(live));
+                               assert(nr < ir_nodeset_size(&live));
 
                                for(k = 0; k < env->co->cls->n_regs; ++k) {
                                        int mapped_col = env->color_map[k];
@@ -1121,11 +1121,11 @@ static void appel_walker(ir_node *bl, void *data)
                }
 
                /* dump the clique and update the stuff. */
-               curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
+               curr_nr = appel_dump_clique(env, &live, bl, curr_nr, start_nr);
 
                /* remove all defs. */
                for(j = 0; j < insn->use_start; ++j)
-                       pset_remove_ptr(live, insn->ops[j].carrier);
+                       ir_nodeset_remove(&live, insn->ops[j].carrier);
 
                if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
                        bli->phi[bli->n_phi]    = insn->irn;
@@ -1136,21 +1136,21 @@ static void appel_walker(ir_node *bl, void *data)
                /* add all uses */
                else
                        for(j = insn->use_start; j < insn->n_ops; ++j)
-                               pset_insert_ptr(live, insn->ops[j].carrier);
+                               ir_nodeset_insert(&live, insn->ops[j].carrier);
        }
 
        /* print the start clique. */
-       curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
+       curr_nr = appel_dump_clique(env, &live, bl, curr_nr, start_nr);
 
        i = 0;
-       foreach_pset(live, irn) {
+       foreach_ir_nodeset(&live, irn, iter) {
                bli->live_in[i]    = irn;
                bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
                ++i;
        }
        bli->n_live_in = i;
 
-       del_pset(live);
+       ir_nodeset_destroy(&live);
        free(insns);
        obstack_free(obst, base);
        env->curr_nr = curr_nr;
@@ -1192,8 +1192,8 @@ static void appel_inter_block_aff(ir_node *bl, void *data)
 
 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
 {
-       int i;
-       int n_colors;
+       unsigned i;
+       unsigned n_colors;
        appel_clique_walker_t env;
        bitset_t *adm = bitset_alloca(co->cls->n_regs);
        be_lv_t *lv = co->cenv->birg->lv;
@@ -1212,7 +1212,7 @@ void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
        env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
        for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
                const arch_register_t *reg = &co->cls->regs[i];
-               env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
+               env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : (int) n_colors++;
        }
 
        env.dumb = 1;
@@ -1224,7 +1224,7 @@ void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
 
        /* make the first k nodes interfere */
        for(i = 0; i < n_colors; ++i) {
-               int j;
+               unsigned j;
                for(j = i + 1; j < n_colors; ++j)
                        fprintf(f, "%d %d -1 ", i, j);
                fprintf(f, "\n");
@@ -1247,7 +1247,7 @@ void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
                                                                  |_|            |___/
 */
 
-static const char *get_dot_color_name(int col)
+static const char *get_dot_color_name(size_t col)
 {
        static const char *names[] = {
                "blue",
@@ -1293,6 +1293,7 @@ typedef struct _co_ifg_dump_t {
 
 static void ifg_dump_graph_attr(FILE *f, void *self)
 {
+       (void) self;
        fprintf(f, "overlap=scale");
 }
 
@@ -1388,11 +1389,12 @@ void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
 
 void co_solve_park_moon(copy_opt_t *opt)
 {
-
+       (void) opt;
 }
 
 static int void_algo(copy_opt_t *co)
 {
+       (void) co;
        return 0;
 }
 
@@ -1415,10 +1417,16 @@ 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 }
 };
@@ -1454,9 +1462,11 @@ void co_driver(be_chordal_env_t *cenv)
        co_algo_t           *algo_func;
        int                 was_optimal = 0;
 
-       if (algo < 0 || algo >= CO_ALGO_LAST)
+       if (algo >= CO_ALGO_LAST)
                return;
 
+       be_liveness_assure_chk(be_get_birg_liveness(cenv->birg));
+
        co = new_copy_opt(cenv, cost_func);
        co_build_ou_structure(co);
        co_build_graph_structure(co);