Adapted to previous commit.
[libfirm] / ir / be / becopyilp2.c
index e40698a..52a36f6 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
  */
 
 /**
- * Author:      Daniel Grund
- * Date:               28.02.2006
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
- *
+ * @file
+ * @brief       ILP based copy minimization.
+ * @author      Daniel Grund
+ * @date        28.02.2006
+ * @version     $Id$
  *
  * ILP formalization using G=(V, E, Q):
  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
  *
  *             x_nc, y_ij \in N,   w_ij \in R^+
  */
-
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif /* HAVE_CONFIG_H */
 
 #ifdef WITH_ILP
 
 #include "irgwalk.h"
 #include "becopyilp_t.h"
 #include "beifg_t.h"
-#include "besched_t.h"
-#include "benodesets.h"
+#include "besched.h"
+#include "bemodule.h"
 
 #define DEBUG_LVL 1
 
 typedef struct _local_env_t {
        double time_limit;
        int first_x_var, last_x_var;
+       int n_colors;
+       bitset_t *normal_colors;
        pmap *nr_2_irn;
        DEBUG_ONLY(firm_dbg_module_t *dbg;)
 } local_env_t;
 
 static void build_coloring_cstr(ilp_env_t *ienv) {
-       be_ifg_t *ifg = ienv->co->cenv->ifg;
-       void *iter = be_ifg_nodes_iter_alloca(ifg);
+       be_ifg_t *ifg     = ienv->co->cenv->ifg;
+       void *iter        = be_ifg_nodes_iter_alloca(ifg);
        bitset_t *colors;
        ir_node *irn;
        char buf[16];
@@ -82,22 +81,24 @@ static void build_coloring_cstr(ilp_env_t *ienv) {
 
        be_ifg_foreach_node(ifg, iter, irn)
                if (!sr_is_removed(ienv->sr, irn)) {
-                       int col, cst_idx;
+                       bitset_pos_t col;
+                       int cst_idx;
                        const arch_register_req_t *req;
-                       int curr_node_color = get_irn_col(ienv->co, irn);
-                       int node_nr = (int)get_irn_node_nr(irn);
+                       int curr_node_color = get_irn_col(irn);
+                       int node_nr = (int)get_irn_idx(irn);
                        local_env_t *lenv = ienv->env;
 
                        pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
 
-                       req = arch_get_register_req(ienv->co->aenv, irn, -1);
+                       req = arch_get_register_req_out(irn);
+
+                       bitset_clear_all(colors);
 
                        /* get assignable colors */
                        if (arch_register_req_is(req, limited)) {
                                rbitset_copy_to_bitset(req->limited, colors);
                        } else {
-                               arch_register_class_put(req->cls, colors);
-                               // bitset_andnot(colors, ienv->co->cenv->ignore_colors);
+                               bitset_copy(colors, lenv->normal_colors);
                        }
 
                        /* add the coloring constraint */
@@ -105,7 +106,7 @@ static void build_coloring_cstr(ilp_env_t *ienv) {
 
                        bitset_foreach(colors, col) {
                                int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
-                               lpp_set_start_value(ienv->lp, var_idx, (col == curr_node_color) ? 1.0 : 0.0);
+                               lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
                                lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 
                                lenv->last_x_var = var_idx;
@@ -125,15 +126,17 @@ static void build_coloring_cstr(ilp_env_t *ienv) {
                }
 }
 
-static void build_interference_cstr(ilp_env_t *ienv) {
-       lpp_t *lpp = ienv->lp;
-       be_ifg_t *ifg = ienv->co->cenv->ifg;
-       int n_colors = arch_register_class_n_regs(ienv->co->cls);
-       int i, col;
-
-       void *iter = be_ifg_cliques_iter_alloca(ifg);
-       ir_node **clique = alloca(sizeof(*clique) * n_colors);
-       int size;
+static void build_interference_cstr(ilp_env_t *ienv)
+{
+       lpp_t       *lpp      = ienv->lp;
+       local_env_t *lenv     = ienv->env;
+       be_ifg_t    *ifg      = ienv->co->cenv->ifg;
+       int          n_colors = lenv->n_colors;
+       void        *iter     = be_ifg_cliques_iter_alloca(ifg);
+       ir_node    **clique   = ALLOCAN(ir_node*, n_colors);
+       int          size;
+       int          col;
+       int          i;
 
        char buf[16];
 
@@ -157,7 +160,7 @@ static void build_interference_cstr(ilp_env_t *ienv) {
                                ir_node *irn = clique[i];
 
                                if (!sr_is_removed(ienv->sr, irn)) {
-                                       int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_node_nr(irn), col));
+                                       int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
                                        lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
                                }
                        }
@@ -172,8 +175,9 @@ static void build_interference_cstr(ilp_env_t *ienv) {
  *       does not provide this walker, yet.
  */
 static void build_affinity_cstr(ilp_env_t *ienv) {
+       local_env_t *lenv = ienv->env;
+       int n_colors      = lenv->n_colors;
        unit_t *curr;
-       int n_colors = arch_register_class_n_regs(ienv->co->cls);
 
        /* for all optimization units */
        list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
@@ -183,13 +187,13 @@ static void build_affinity_cstr(ilp_env_t *ienv) {
                int root_col, arg_col;
 
                root = curr->nodes[0];
-               root_nr = (int) get_irn_node_nr(root);
-               root_col = get_irn_col(ienv->co, root);
+               root_nr = (int) get_irn_idx(root);
+               root_col = get_irn_col(root);
 
                for (i = 1; i < curr->node_count; ++i) {
                        arg = curr->nodes[i];
-                       arg_nr = (int) get_irn_node_nr(arg);
-                       arg_col = get_irn_col(ienv->co, arg);
+                       arg_nr = (int) get_irn_idx(arg);
+                       arg_col = get_irn_col(arg);
 
                        /* add a new affinity variable */
                        y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
@@ -219,13 +223,14 @@ typedef struct _edge_t {
 static int compare_edge_t(const void *k1, const void *k2, size_t size) {
        const edge_t *e1 = k1;
        const edge_t *e2 = k2;
+       (void) size;
 
        return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
 }
 
-#define HASH_EDGE(e) (nodeset_hash((e)->n1) ^ nodeset_hash((e)->n2))
+#define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
 
-static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
+static inline edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
        edge_t new_edge;
 
        if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
@@ -239,7 +244,7 @@ static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counte
        return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 }
 
-static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
+static inline edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
        edge_t new_edge;
 
        if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
@@ -252,7 +257,7 @@ static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
        return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
 }
 
-static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
+static inline void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
        edge_t new_edge, *e;
 
        if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
@@ -284,19 +289,24 @@ static void build_clique_star_cstr(ilp_env_t *ienv) {
        co_gs_foreach_aff_node(ienv->co, aff) {
                struct obstack ob;
                neighb_t *nbr;
-               ir_node *center = aff->irn;
+               const ir_node *center = aff->irn;
                ir_node **nodes;
                set *edges;
                int i, o, n_nodes, n_edges;
 
+               if (arch_irn_is_ignore(aff->irn))
+                       continue;
+
                obstack_init(&ob);
                edges = new_set(compare_edge_t, 8);
 
                /* get all affinity neighbours */
                n_nodes = 0;
                co_gs_foreach_neighb(aff, nbr) {
-                       obstack_ptr_grow(&ob, nbr->irn);
-                       ++n_nodes;
+                       if (!arch_irn_is_ignore(nbr->irn)) {
+                               obstack_ptr_grow(&ob, nbr->irn);
+                               ++n_nodes;
+                       }
                }
                nodes = obstack_finish(&ob);
 
@@ -368,10 +378,10 @@ static void build_clique_star_cstr(ilp_env_t *ienv) {
                                char buf[16];
 
                                cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
-                               center_nr = get_irn_node_nr(center);
+                               center_nr = get_irn_idx(center);
 
                                pset_foreach(clique, member) {
-                                       member_nr = get_irn_node_nr(member);
+                                       member_nr = get_irn_idx(member);
                                        var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
                                        lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
                                }
@@ -386,7 +396,7 @@ static void build_clique_star_cstr(ilp_env_t *ienv) {
 }
 
 
-static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
+static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn) {
        be_ifg_t *ifg = ienv->co->cenv->ifg;
        int i, len;
        ir_node **curr_path;
@@ -397,14 +407,17 @@ static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
        if (pdeq_contains(path, irn))
                return;
 
+       if (arch_irn_is_ignore(irn))
+               return;
+
        /* insert the new irn */
        pdeq_putr(path, irn);
 
 
 
        /* check for forbidden interferences */
-       len = pdeq_len(path);
-       curr_path = alloca(len * sizeof(*curr_path));
+       len       = pdeq_len(path);
+       curr_path = ALLOCAN(ir_node*, len);
        pdeq_copyl(path, (const void **)curr_path);
 
        for (i=1; i<len; ++i)
@@ -423,8 +436,8 @@ static void extend_path(ilp_env_t *ienv, pdeq *path, ir_node *irn) {
                        int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
                        for (i=1; i<len; ++i) {
                                char buf[16];
-                               int nr_1    = get_irn_node_nr(curr_path[i-1]);
-                               int nr_2    = get_irn_node_nr(curr_path[i]);
+                               int nr_1    = get_irn_idx(curr_path[i-1]);
+                               int nr_2    = get_irn_idx(curr_path[i]);
                                int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
                                lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
                        }
@@ -485,16 +498,14 @@ static void ilp2_build(ilp_env_t *ienv) {
 
 static void ilp2_apply(ilp_env_t *ienv) {
        local_env_t *lenv = ienv->env;
-       double *sol;
-       lpp_sol_state_t state;
-       int i, count;
+       int i;
 
        /* first check if there was sth. to optimize */
        if (lenv->first_x_var >= 0) {
+               int              count = lenv->last_x_var - lenv->first_x_var + 1;
+               double          *sol   = XMALLOCN(double, count);
+               lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 
-               count = lenv->last_x_var - lenv->first_x_var + 1;
-               sol = xmalloc(count * sizeof(sol[0]));
-               state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
                if (state != lpp_optimal) {
                        printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
                        assert(state >= lpp_feasible && "The solution should at least be feasible!");
@@ -516,6 +527,8 @@ static void ilp2_apply(ilp_env_t *ienv) {
                                        assert(0 && "This should be a x-var");
                        }
                }
+
+               xfree(sol);
        }
 
 #ifdef COPYOPT_STAT
@@ -527,6 +540,17 @@ static void ilp2_apply(ilp_env_t *ienv) {
 #endif
 }
 
+void be_init_copyilp2(void)
+{
+       static co_algo_info copyheur = {
+               co_solve_ilp2, 1
+       };
+
+       be_register_copyopt("ilp", &copyheur);
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyilp2);
+
 int co_solve_ilp2(copy_opt_t *co) {
        lpp_sol_state_t sol_state;
        ilp_env_t *ienv;
@@ -541,6 +565,11 @@ int co_solve_ilp2(copy_opt_t *co) {
        my.nr_2_irn    = pmap_create();
        FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
 
+       my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
+       bitset_clear_all(my.normal_colors);
+       arch_put_non_ignore_regs(co->cls, my.normal_colors);
+       my.n_colors = bitset_popcnt(my.normal_colors);
+
        ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 
        sol_state = ilp_go(ienv);
@@ -553,7 +582,7 @@ int co_solve_ilp2(copy_opt_t *co) {
 
 #else /* WITH_ILP */
 
-static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
+static inline void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 }
 
 #endif /* WITH_ILP */