fixed some minor bugs
[libfirm] / ir / be / becopyilp2.c
index 33efd78..b323464 100644 (file)
  * Copyright:   (c) Universitaet Karlsruhe
  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
  *
+ *
  * ILP formalization using G=(V, E, Q):
- *  - 1 class of variables: equal color vars
+ *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
  *  - Path constraints
- *  - Clique path constraints
+ *  - Clique-star constraints
  *
  *
  *     \min \sum_{ (i,j) \in Q }  w_ij y_ij
  *
- *             y_ij                            =  1                    (i,j) \in E
+ *             \sum_c x_nc                     =  1                    n \in N, c \in C
  *
- *             \sum_c y_nc                     =  |C| - 1              n \in N, c \in C
+ *             x_nc                            =  0                    n \in N, c \not\in C(n)
  *
- *             y_nc                            =  1                    n \in N, c \not\in C(n)
+ *             \sum x_nc                       <= 1                    x_nc \in Clique \in AllCliques,  c \in C
  *
  *             \sum_{e \in p} y_e      >= 1                    p \in P         path constraints
  *
- *             \sum_{e \in cp} y_e     >= |cp| - 1             cp \in CP       clique-path constraints
+ *             \sum_{e \in cs} y_e     >= |cs| - 1             cs \in CP       clique-star constraints
  *
- *             y_ij \in N,   w_ij \in R^+
+ *             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 "irtools.h"
+#include "irgwalk.h"
 #include "becopyilp_t.h"
+#include "beifg_t.h"
+#include "besched_t.h"
+
+#define DEBUG_LVL 1
+
+typedef struct _local_env_t {
+       firm_dbg_module_t *dbg;
+       double time_limit;
+       int first_x_var, last_x_var;
+       pmap *nr_2_irn;
+} 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);
+       bitset_t *colors;
+       ir_node *irn;
+       char buf[16];
+
+       colors = bitset_alloca(arch_register_class_n_regs(ienv->co->cls));
+
+       be_ifg_foreach_node(ifg, iter, irn)
+               if (!sr_is_removed(ienv->sr, irn)) {
+                       int col, cst_idx;
+                       arch_register_req_t req;
+                       int curr_node_color = get_irn_col(ienv->co, irn);
+                       int node_nr = (int)get_irn_node_nr(irn);
+                       local_env_t *lenv = ienv->env;
+
+                       pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
+
+                       arch_get_register_req(ienv->co->aenv, &req, irn, -1);
+
+                       /* get assignable colors */
+                       if (arch_register_req_is(&req, limited))
+                               req.limited(req.limited_env, colors);
+                       else
+                               arch_put_non_ignore_regs(ienv->co->aenv, req.cls, colors);
+
+                       /* add the coloring constraint */
+                       cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
+
+                       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_factor_fast(ienv->lp, cst_idx, var_idx, 1);
+
+                               lenv->last_x_var = var_idx;
+                               if (lenv->first_x_var == -1)
+                                       lenv->first_x_var = var_idx;
+                       }
+
+                       /* add register constraint constraints */
+                       bitset_foreach_clear(colors, col) {
+                               int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 0.0);
+                               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, 0.0);
+                               lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
+
+                               lenv->last_x_var = var_idx;
+                       }
+               }
+}
+
+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;
+
+       char buf[16];
+
+       /* for each maximal clique */
+       be_ifg_foreach_clique(ifg, iter, clique, &size) {
+
+               if (size < 2)
+                       continue;
+
+               /* for all colors */
+               for (col=0; col<n_colors; ++col) {
+                       int cst_idx = lpp_add_cst(lpp, NULL, lpp_less, 1.0);
+
+                       /* for each member of this clique */
+                       for (i=0; i<size; ++i) {
+                               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));
+                                       lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
+                               }
+                       }
+               }
+       }
+}
+
+static void build_affinity_cstr(ilp_env_t *ienv) {
+       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) {
+               ir_node *root, *arg;
+               int root_nr, arg_nr, i, col, y_idx, root_idx, arg_idx;
+               char buf[16];
+               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);
+
+               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);
+
+                       /* 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]);
+                       lpp_set_start_value(ienv->lp, y_idx, (root_col==arg_col) ? 0.0 : 1.0);
+
+                       /* add constraints relating the affinity var to the color vars */
+                       for (col=0; col<n_colors; ++col) {
+                               int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_less, 0.0);
+                               root_idx = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', root_nr, col));
+                               arg_idx  = lpp_get_var_idx(ienv->lp, name_cdd(buf, 'x', arg_nr,  col));
+
+                               lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
+                               lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
+                               lpp_set_factor_fast(ienv->lp, cst_idx, root_idx, -1.0);
+                       }
+               }
+       }
+}
+
+/**
+ * Helping stuff for build_clique_star_cstr
+ */
+typedef struct _edge_t {
+       ir_node *n1, *n2;
+} 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;
+
+       return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
+}
+
+#define HASH_EDGE(e) (HASH_PTR((e)->n1) ^ HASH_PTR((e)->n2))
+
+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)) {
+               new_edge.n1 = n1;
+               new_edge.n2 = n2;
+       } else {
+               new_edge.n1 = n2;
+               new_edge.n2 = n1;
+       }
+       *counter++;
+       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) {
+       edge_t new_edge;
+
+       if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
+               new_edge.n1 = n1;
+               new_edge.n2 = n2;
+       } else {
+               new_edge.n1 = n2;
+               new_edge.n2 = n1;
+       }
+       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) {
+       edge_t new_edge, *e;
+
+       if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
+               new_edge.n1 = n1;
+               new_edge.n2 = n2;
+       } else {
+               new_edge.n1 = n2;
+               new_edge.n2 = n1;
+       }
+       e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
+       if (e) {
+               e->n1 = NULL;
+               e->n2 = NULL;
+               *counter--;
+       }
+}
+
+#define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
+
+/**
+ * Search for an interference clique and an external node
+ * with affinity edges to all nodes of the clique.
+ * At most 1 node of the clique can be colored equally with the external node.
+ */
+static void build_clique_star_cstr(ilp_env_t *ienv) {
+       node_t *node;
+
+       /* for each node with affinity edges */
+       co_gs_foreach_node(ienv->co, node) {
+               struct obstack ob;
+               neighb_t *nbr;
+               ir_node *center = node->irn;
+               ir_node **nodes;
+               set *edges;
+               int i, o, n_nodes, n_edges;
+
+               obstack_init(&ob);
+               edges = new_set(compare_edge_t, 8);
+
+               /* get all affinity neighbours */
+               n_nodes = 0;
+               co_gs_foreach_neighb(node, nbr) {
+                       obstack_ptr_grow(&ob, nbr->irn);
+                       ++n_nodes;
+               }
+               nodes = obstack_finish(&ob);
+
+               /* get all interference edges between these */
+               n_edges = 0;
+               for (i=0; i<n_nodes; ++i)
+                       for (o=0; o<i; ++o)
+                               if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
+                                       add_edge(edges, nodes[i], nodes[o], &n_edges);
+
+               /* cover all these interference edges with maximal cliques */
+               while (n_edges) {
+                       edge_t *e;
+                       pset *clique = pset_new_ptr(8);
+                       int growed;
+
+                       /* get 2 starting nodes to form a clique */
+                       for (e=set_first(edges); !e->n1; e=set_next(edges))
+                               /*nothing*/ ;
+
+                       remove_edge(edges, e->n1, e->n2, &n_edges);
+                       pset_insert_ptr(clique, e->n1);
+                       pset_insert_ptr(clique, e->n2);
+
+                       /* while the clique is growing */
+                       do {
+                               growed = 0;
+
+                               /* search for a candidate to extend the clique */
+                               for (i=0; i<n_nodes; ++i) {
+                                       ir_node *member, *cand = nodes[i];
+                                       int is_cand;
+
+                                       /* if its already in the clique try the next */
+                                       if (pset_find_ptr(clique, cand))
+                                               continue;
+
+                                       /* are there all necessary interferences? */
+                                       is_cand = 1;
+                                       pset_foreach(clique, member) {
+                                               if (!find_edge(edges, cand, member)) {
+                                                       is_cand = 0;
+                                                       pset_break(clique);
+                                                       break;
+                                               }
+                                       }
+
+                                       /* now we know if we have a clique extender */
+                                       if (is_cand) {
+                                               /* first remove all covered edges */
+                                               pset_foreach(clique, member)
+                                                       remove_edge(edges, cand, member, &n_edges);
+
+                                               /* insert into clique */
+                                               pset_insert_ptr(clique, cand);
+                                               growed = 1;
+                                               break;
+                                       }
+                               }
+                       } while (growed);
+
+                       /* now the clique is maximal. Finally add the constraint */
+                       {
+                               ir_node *member;
+                               int var_idx, cst_idx, center_nr, member_nr;
+                               char buf[16];
+
+                               cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
+                               center_nr = get_irn_node_nr(center);
+
+                               pset_foreach(clique, member) {
+                                       member_nr = get_irn_node_nr(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);
+                               }
+                       }
+
+                       del_pset(clique);
+               }
+
+               del_set(edges);
+               obstack_free(&ob, NULL);
+       }
+}
+
+/**
+ *
+ */
+static void build_path_cstr(ilp_env_t *ienv) {
+
+}
+
+static void ilp2_build(ilp_env_t *ienv) {
+       local_env_t *lenv = ienv->env;
+       int lower_bound;
+
+       ienv->lp = new_lpp(ienv->co->name, lpp_minimize);
+       build_coloring_cstr(ienv);
+       build_interference_cstr(ienv);
+       build_affinity_cstr(ienv);
+       build_clique_star_cstr(ienv);
+       build_path_cstr(ienv);
+
+       lower_bound = co_get_lower_bound(ienv->co) - co_get_inevit_copy_costs(ienv->co);
+       lpp_set_bound(ienv->lp, lower_bound);
+       lpp_set_time_limit(ienv->lp, lenv->time_limit);
+}
+
+static void ilp2_apply(ilp_env_t *ienv) {
+       local_env_t *lenv = ienv->env;
+       double *sol;
+       lpp_sol_state_t state;
+       int i, count;
+
+       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!");
+       }
+
+       for (i=0; i<count; ++i) {
+               int nodenr, color;
+               char var_name[16];
+
+               if (sol[i] > 1-EPSILON) { /* split variable name into components */
+                       lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
+
+                       if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
+                               ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
+                               assert(irn && "This node number must be present in the map");
+
+                               set_irn_col(ienv->co, irn, color);
+                       } else
+                               assert(0 && "This should be a x-var");
+               }
+       }
+
+#ifdef COPYOPT_STAT
+       /* TODO adapt to multiple possible ILPs */
+       copystat_add_ilp_time((int)(1000.0*lpp_get_sol_time(pi->curr_lp)));  //now we have ms
+       copystat_add_ilp_vars(lpp_get_var_count(pi->curr_lp));
+       copystat_add_ilp_csts(lpp_get_cst_count(pi->curr_lp));
+       copystat_add_ilp_iter(lpp_get_iter_cnt(pi->curr_lp));
+#endif
+}
 
 int co_solve_ilp2(copy_opt_t *co, double time_limit) {
-       int res = 1;
+       lpp_sol_state_t sol_state;
+       ilp_env_t *ienv;
+       local_env_t my;
+
+       my.time_limit  = time_limit;
+       my.first_x_var = -1;
+       my.last_x_var  = -1;
+       my.nr_2_irn    = pmap_create();
+       my.dbg         = firm_dbg_register("ir.be.coilp2");
+       firm_dbg_set_mask(my.dbg, DEBUG_LVL);
+
+       ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 
-       return res;
+       sol_state = ilp_go(ienv);
+
+       pmap_destroy(my.nr_2_irn);
+       free_ilp_env(ienv);
+
+       return sol_state == lpp_optimal;
+}
+
+#else /* WITH_ILP */
+
+static void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 }
+
+#endif /* WITH_ILP */