Let dfs() discover only memory nodes

[libfirm] / ir / be / becopyopt.c

diff --git a/ir/be/becopyopt.c b/ir/be/becopyopt.c
index 26d0686..b5d3ffc 100644 (file)
--- a/ir/be/becopyopt.c
+++ b/ir/be/becopyopt.c
@@ -445,10 +445,8 @@ 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));
@@ -462,15 +460,44 @@ 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(skip_Proj(irn), req->other_same[0]); /* TODO handle second should-be-same constraint */
-                       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);
+                       ir_node *other  = get_irn_n(skip_Proj(irn), req->other_same[0]);
+                       ir_node *other2 = NULL;
+                       int      count;
+
+                       if (arch_irn_is(co->aenv, other, ignore) ||
+                                       nodes_interfere(co->cenv, irn, other)) {
+                               other = NULL;
+                       }
+                       if (req->other_same[1] != -1) {
+                               other2 = get_irn_n(skip_Proj(irn), req->other_same[1]);
+                               if (arch_irn_is(co->aenv, other2, ignore) ||
+                                               nodes_interfere(co->cenv, irn, other2)) {
+                                       other2 = NULL;
+                               }
+                       }
+                       count = 1 + (other != NULL) + (other2 != NULL && other != other2);
+
+                       if (count > 1) {
+                               int i = 0;
+
+                               unit->nodes = xmalloc(count * sizeof(*unit->nodes));
+                               unit->costs = xmalloc(count * sizeof(*unit->costs));
+                               unit->node_count = count;
+                               unit->nodes[i] = irn;
+                               if (other != NULL) {
+                                       ++i;
+                                       unit->nodes[i] = other;
+                                       unit->costs[i] = co->get_costs(co, irn, other, -1);
+                               }
+                               if (other2 != NULL) {
+                                       if (other == other2) {
+                                               unit->costs[i] += co->get_costs(co, irn, other2, -1);
+                                       } else {
+                                               ++i;
+                                               unit->nodes[i] = other2;
+                                               unit->costs[i] = co->get_costs(co, irn, other2, -1);
+                                       }
+                               }
                        }
                } else {
                        assert(0 && "This is not an optimizable node!");
@@ -829,7 +856,8 @@ 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) {
@@ -1078,7 +1106,7 @@ 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;
+                               unsigned k;
                                size_t nr;
 
                                if(!insn->ops[j].has_constraints)
@@ -1174,8 +1202,8 @@ 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;
@@ -1194,7 +1222,7 @@ 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;
@@ -1206,7 +1234,7 @@ 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");