Modified everything
[libfirm] / ir / be / becopyilp.c
index 111bef0..4ff9f25 100644 (file)
 
 #include "irprog.h"
 
-#include "lpp.h"
-#include "lpp_local.h"
-#include "lpp_remote.h"
+#include <lpp/lpp.h>
+#include <lpp/lpp_net.h>
 #include "xmalloc.h"
 #include "becopyopt.h"
 #include "becopystat.h"
+#include "besched_t.h"
+
+#define LPP_HOST "i44pc52"
+#define LPP_SOLVER "cplex"
 
 #undef DUMP_MPS
-#define DEBUG_LVL SET_LEVEL_1
 static firm_dbg_module_t *dbg = NULL;
 
+#define MAX(a,b) ((a<b)?(b):(a))
 #define EPSILON 0.00001
 #define SLOTS_LIVING 32
 
@@ -52,7 +55,7 @@ typedef struct _problem_instance_t {
 
 #define is_removed(irn) pset_find_ptr(pi->removed, irn)
 
-#define is_color_possible(irn,color) arch_reg_is_allocatable(pi->co->chordal_env->arch_env, irn, arch_pos_make_out(0), arch_register_for_index(pi->co->chordal_env->cls, color))
+#define is_color_possible(irn,color) arch_reg_is_allocatable(get_arch_env(pi->co), irn, arch_pos_make_out(0), arch_register_for_index(pi->co->chordal_env->cls, color))
 
 /*
  * Some stuff for variable name handling.
@@ -109,8 +112,8 @@ static void pi_find_simplicials(problem_instance_t *pi) {
        while (redo) {
                redo = 0;
                for (ifn = set_first(if_nodes); ifn; ifn = set_next(if_nodes)) {
-                       ir_node *irn = get_irn_for_graph_nr(pi->co->chordal_env->irg, ifn->nnr);
-                       if (!is_removed(irn) && !is_optimizable(pi->co->chordal_env->arch_env, irn) &&
+                       ir_node *irn = get_irn_for_graph_nr(get_irg(pi->co), ifn->nnr);
+                       if (!is_removed(irn) && !is_optimizable(get_arch_env(pi->co), irn) &&
           !is_optimizable_arg(pi->co, irn) && pi_is_simplicial(pi, ifn)) {
                                simpl_t *s = xmalloc(sizeof(*s));
                                s->ifn = ifn;
@@ -145,15 +148,15 @@ static void pi_add_constr_A(problem_instance_t *pi) {
 
                                nnr = get_irn_graph_nr(curr->irn);
                                mangle_cst(pi->buf, 'A', nnr);
-                               cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, equal, 1);
+                               cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_equal, 1);
 
                                // iterate over all possible colors in order
                                bitset_clear_all(pos_regs);
-                               arch_get_allocatable_regs(pi->co->chordal_env->arch_env, curr->irn, arch_pos_make_out(0), pi->co->chordal_env->cls, pos_regs);
+                               arch_get_allocatable_regs(get_arch_env(pi->co), curr->irn, arch_pos_make_out(0), pi->co->chordal_env->cls, pos_regs);
                                bitset_foreach(pos_regs, col) {
                                        int var_idx;
                                        mangle_var(pi->buf, 'x', nnr, col);
-                                       var_idx = lpp_add_var(pi->curr_lp, pi->buf, binary, 0);
+                                       var_idx = lpp_add_var(pi->curr_lp, pi->buf, lpp_binary, 0);
                                        pi->last_x_var = var_idx;
                                        lpp_set_factor_fast(pi->curr_lp, cst_idx, var_idx, 1);
                                }
@@ -213,7 +216,7 @@ static void pi_add_constr_B(problem_instance_t *pi, int color) {
                                        int cst_idx;
                                        ir_node *n;
                                        mangle_cst(pi->buf, 'B', pi->cst_counter);
-                                       cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, less, 1);
+                                       cst_idx = lpp_add_cst(pi->curr_lp, pi->buf, lpp_less, 1);
                                        for (n = pset_first(living); n; n = pset_next(living)) {
                                                int var_idx;
                                                mangle_var_irn(pi->buf, 'x', n, color);
@@ -233,16 +236,16 @@ static void pi_add_constr_B(problem_instance_t *pi, int color) {
 
 /**
  * Generates constraints which interrelate x with y variables.
- * y=1 ==> x1 and x2 must have the same color.
- *     <== is achieved automatically by minimization.
+ * x1 and x2 have the different colors ==> y_12 = 1
  */
 static void pi_add_constr_E(problem_instance_t *pi) {
        unit_t *curr;
-       bitset_t *root_regs, *arg_regs;
+       bitset_t *root_regs, *arg_regs, *work_regs;
        int cst_counter = 0;
        unsigned nregs = pi->co->chordal_env->cls->n_regs;
        root_regs = bitset_alloca(nregs);
        arg_regs = bitset_alloca(nregs);
+       work_regs = bitset_alloca(nregs);
 
        DBG((dbg, LEVEL_2, "Add E constraints...\n"));
        /* for all roots of optimization units */
@@ -255,25 +258,27 @@ static void pi_add_constr_E(problem_instance_t *pi) {
                root = curr->nodes[0];
                rootnr = get_irn_graph_nr(root);
                bitset_clear_all(root_regs);
-               arch_get_allocatable_regs(pi->co->chordal_env->arch_env, root, arch_pos_make_out(0), pi->co->chordal_env->cls, root_regs);
+               arch_get_allocatable_regs(get_arch_env(pi->co), root, arch_pos_make_out(0), pi->co->chordal_env->cls, root_regs);
 
                /* for all arguments of root */
                for (i = 1; i < curr->node_count; ++i) {
                        arg = curr->nodes[i];
                        argnr = get_irn_graph_nr(arg);
                        bitset_clear_all(arg_regs);
-                       arch_get_allocatable_regs(pi->co->chordal_env->arch_env, arg, arch_pos_make_out(0), pi->co->chordal_env->cls, arg_regs);
+                       arch_get_allocatable_regs(get_arch_env(pi->co), arg, arch_pos_make_out(0), pi->co->chordal_env->cls, arg_regs);
 
                        /* Introduce new variable and set factor in objective function */
                        mangle_var(buf, 'y', rootnr, argnr);
-                       y_idx = lpp_add_var(pi->curr_lp, buf, continous, curr->costs[i]);
+                       y_idx = lpp_add_var(pi->curr_lp, buf, lpp_continous, curr->costs[i]);
 
+                       //BETTER: y vars as binary or continous vars ??
                        /* set starting value */
                        //lpp_set_start_value(pi->curr_lp, y_idx, (get_irn_col(pi->co, root) != get_irn_col(pi->co, arg)));
 
                        /* For all colors root and arg have in common, add 2 constraints to E */
-                       bitset_and(arg_regs, root_regs);
-                       bitset_foreach(arg_regs, color) {
+                       bitset_copy(work_regs, root_regs);
+                       bitset_and(work_regs, arg_regs);
+                       bitset_foreach(work_regs, color) {
                                int root_idx, arg_idx, cst_idx;
                                mangle_var(buf, 'x', rootnr, color);
                                root_idx = lpp_get_var_idx(pi->curr_lp, buf);
@@ -282,31 +287,58 @@ static void pi_add_constr_E(problem_instance_t *pi) {
 
                                /* add root-arg-y <= 0 */
                                mangle_cst(buf, 'E', cst_counter++);
-                               cst_idx = lpp_add_cst(pi->curr_lp, buf, less, 0);
+                               cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
                                lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
                                lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, -1);
                                lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
 
                                /* add arg-root-y <= 0 */
                                mangle_cst(buf, 'E', cst_counter++);
-                               cst_idx = lpp_add_cst(pi->curr_lp, buf, less, 0);
+                               cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
                                lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, -1);
                                lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
                                lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
                        }
+                       /* For all colors root and arg have "disjunct", add 1 constraints to E.
+                        * If root gets a color the arg is not possible to get then they will
+                        * definetly get different colors. So y has to be 1.
+                        * Vice versa for arg.
+                        */
+                       bitset_copy(work_regs, root_regs);
+                       bitset_xor(work_regs, arg_regs);
+                       bitset_foreach(work_regs, color) {
+                               int root_idx, arg_idx, cst_idx;
+                               mangle_var(buf, 'x', rootnr, color);
+                               root_idx = lpp_get_var_idx(pi->curr_lp, buf);
+                               mangle_var(buf, 'x', argnr, color);
+                               arg_idx = lpp_get_var_idx(pi->curr_lp, buf);
+
+                               mangle_cst(buf, 'E', cst_counter++);
+                               cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_less, 0);
+                               if (bitset_is_set(root_regs, color)) {
+                                       /* add root-y <= 0 */
+                                       lpp_set_factor_fast(pi->curr_lp, cst_idx, root_idx, 1);
+                                       lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
+                               } else {
+                                       assert(bitset_is_set(arg_regs, color) && "bitset_xor is buggy");
+                                       /* add arg-y <= 0 */
+                                       lpp_set_factor_fast(pi->curr_lp, cst_idx, arg_idx, 1);
+                                       lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, -1);
+                               }
+                       }
                }
        }
 }
 
 /**
- * Matrix M: maximum independent set constraints
+ * Matrix S: maximum independent set constraints
  * Generates lower bound-cuts for optimization units with inner interferences.
  * Sum(y_{root, arg}, arg \in Args) <= max_indep_set_size - 1
  */
-static void pi_add_constr_M(problem_instance_t *pi) {
+static void pi_add_constr_S(problem_instance_t *pi) {
        unit_t *curr;
        int cst_counter = 0;
-       DBG((dbg, LEVEL_2, "Add M constraints...\n"));
+       DBG((dbg, LEVEL_2, "Add S constraints...\n"));
 
        /* for all optimization units */
        list_for_each_entry(unit_t, curr, &pi->co->units, units) {
@@ -315,13 +347,13 @@ static void pi_add_constr_M(problem_instance_t *pi) {
                int cst_idx, y_idx, i;
                char buf[32];
 
-               if (curr->minimal_costs == 0)
+               if (curr->min_nodes_costs == 0)
                        continue;
 
                root = curr->nodes[0];
                rootnr = get_irn_graph_nr(root);
-               mangle_cst(buf, 'M', cst_counter++);
-               cst_idx = lpp_add_cst(pi->curr_lp, buf, greater, curr->minimal_costs);
+               mangle_cst(buf, 'S', cst_counter++);
+               cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, curr->min_nodes_costs);
 
                /* for all arguments */
                for (i = 1; i < curr->node_count; ++i) {
@@ -329,11 +361,132 @@ static void pi_add_constr_M(problem_instance_t *pi) {
                        argnr = get_irn_graph_nr(arg);
                        mangle_var(buf, 'y', rootnr, argnr);
                        y_idx = lpp_get_var_idx(pi->curr_lp, buf);
-                       lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, 1);
+                       lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, curr->costs[i]);
                }
        }
 }
 
+static INLINE int get_costs(problem_instance_t *pi, ir_node *phi, ir_node *irn) {
+       int i;
+       unit_t *curr;
+       /* search optimization unit for phi */
+       list_for_each_entry(unit_t, curr, &pi->co->units, units)
+               if (curr->nodes[0] == phi) {
+                       for (i=1; i<curr->node_count; ++i)
+                               if (curr->nodes[i] == irn)
+                                       return curr->costs[i];
+                       assert(0 && "irn must occur in this ou");
+               }
+       assert(0 && "phi must be found in a ou");
+       return 0;
+}
+
+/*
+ * TODO: Because this here uses a phi-walker and not the ou's,
+ * it is possible, that the interfering args of a phi will cause a bug ??!!
+ */
+static void M_constr_walker(ir_node *block, void *env) {
+       problem_instance_t *pi = env;
+       int count, arity, row, col, other_row, *costs;
+       ir_node **phis, *phi, *irn, **phi_matrix;
+       pset *done;
+       bitset_t *candidates;
+
+       /* Count all phi nodes of this block */
+       for (count=0, irn = sched_first(block); is_Phi(irn); irn = sched_next(irn))
+               count++;
+
+       /* We at least 2 phi nodes for this class of inequalities */
+       if (count < 2)
+               return;
+
+       /* Build the \Phi-Matrix */
+       arity = get_irn_arity(sched_first(block));
+       phis = alloca(count * sizeof(*phis));
+       costs = alloca(count * sizeof(costs));
+       phi_matrix = alloca(count*arity * sizeof(*phi_matrix));
+       candidates = bitset_alloca(count);
+
+       phi = sched_first(block);
+       for (row=0; row<count; ++row) {
+               phis[row] = phi;
+               for (col=0; col<arity; ++col)
+                       phi_matrix[row*arity + col] = get_irn_n(phi, col);
+               phi = sched_next(phi);
+       }
+
+       /* Now find the interesting patterns in the matrix:
+        * All nodes which are used at least twice in a column. */
+       /* columnwise ... */
+       for (col=0; col<arity; ++col) {
+               done = pset_new_ptr_default();
+               for (row=0; row<count; ++row) {
+                       irn = phi_matrix[row*arity + col];
+                       /* has the irn already been processed in this col? */
+                       if (pset_find_ptr(done, irn))
+                               continue;
+                       else
+                               pset_insert_ptr(done, irn);
+
+                       /* insert irn in candidates */
+                       bitset_clear_all(candidates);
+                       bitset_set(candidates, row);
+                       /* search the irn in the rows below */
+                       for (other_row = row+1; other_row<count; ++other_row)
+                               if (irn == phi_matrix[other_row*arity + col]) {
+                                       /* found the irn in the same col in another row */
+                                       bitset_set(candidates, other_row);
+                               }
+
+                       /* now we know all occurences of irn in this col */
+                       if (bitset_popcnt(candidates) < 2)
+                               continue;
+
+                       /* compute the minimal costs (rhs) */
+                       int phi_nr, sum=0, max=-1, minimal_costs;
+                       bitset_foreach(candidates, phi_nr) {
+                               costs[phi_nr] = get_costs(pi, phis[phi_nr], irn);
+                               sum += costs[phi_nr];
+                               max = MAX(max, costs[phi_nr]);
+                       }
+                       minimal_costs = sum - max;
+
+                       /* generate an unequation finally.
+                        * phis are indexed in the bitset,
+                        * shared argument is irn
+                        * rhs is minimal_costs */
+                       {
+                               char buf[32];
+                               ir_node *root;
+                               int pos, irnnr, rootnr, cst_idx, y_idx, cst_counter = 0;
+
+                               irnnr = get_irn_graph_nr(irn);
+                               mangle_cst(buf, 'M', cst_counter++);
+                               cst_idx = lpp_add_cst(pi->curr_lp, buf, lpp_greater, minimal_costs);
+
+                               /* for all phis */
+                               bitset_foreach(candidates, pos) {
+                                       root = phis[pos];
+                                       rootnr = get_irn_graph_nr(root);
+                                       mangle_var(buf, 'y', rootnr, irnnr);
+                                       y_idx = lpp_get_var_idx(pi->curr_lp, buf);
+                                       lpp_set_factor_fast(pi->curr_lp, cst_idx, y_idx, costs[pos]);
+                               }
+                       }
+               }
+               del_pset(done); /* clear set for next row */
+       } /*next col*/
+}
+
+/**
+ * Matrix M: Multi-Arg-Use. Interrelates different \phi-functions
+ * in the same block, iff they use the same arg at the same pos.
+ * Only one of the phis can get the arg.
+ */
+static void pi_add_constr_M(problem_instance_t *pi) {
+       dom_tree_walk_irg(get_irg(pi->co), M_constr_walker, NULL, pi);
+}
+
 /**
  * Generate the initial problem matrices and vectors.
  */
@@ -346,12 +499,12 @@ static problem_instance_t *new_pi(const copy_opt_t *co) {
        pi->co = co;
        pi->removed = pset_new_ptr_default();
        INIT_LIST_HEAD(&pi->simplicials);
-       pi->dilp = new_lpp(co->name, minimize);
+       pi->dilp = new_lpp(co->name, lpp_minimize);
        pi->last_x_var = -1;
 
        /* problem size reduction */
        pi_find_simplicials(pi);
-       //TODO If you wish to see it: dump_ifg_w/o_removed
+       //BETTER If you wish to see it: dump_ifg_w/o_removed
        if (pi->all_simplicial)
                return pi;
 
@@ -361,6 +514,7 @@ static problem_instance_t *new_pi(const copy_opt_t *co) {
        for (col = 0; col < pi->co->chordal_env->cls->n_regs; ++col)
                pi_add_constr_B(pi, col);
        pi_add_constr_E(pi);
+       pi_add_constr_S(pi);
        pi_add_constr_M(pi);
 
        return pi;
@@ -395,8 +549,8 @@ static void pi_set_start_sol(problem_instance_t *pi) {
                lpp_get_var_name(pi->curr_lp, i, var_name, sizeof(var_name));
                /* split into components */
                if (split_var(var_name, &nnr, &col) == 2) {
-                       assert(get_irn_col(pi->co, get_irn_for_graph_nr(pi->co->chordal_env->irg, nnr)) != -1);
-                       val = (get_irn_col(pi->co, get_irn_for_graph_nr(pi->co->chordal_env->irg, nnr)) == col) ? 1 : 0;
+                       assert(get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) != -1);
+                       val = (get_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr)) == col) ? 1 : 0;
                        lpp_set_start_value(pi->curr_lp, i, val);
                } else {
                        fprintf(stderr, "Variable name is: %s\n", var_name);
@@ -408,9 +562,9 @@ static void pi_set_start_sol(problem_instance_t *pi) {
 /**
  * Invoke a solver
  */
-static void pi_solve_ilp(problem_instance_t *pi, void (*lpp_solve)(lpp_t *)) {
+static void pi_solve_ilp(problem_instance_t *pi) {
        pi_set_start_sol(pi);
-       lpp_solve(pi->curr_lp);
+       lpp_solve_net(pi->curr_lp, LPP_HOST, LPP_SOLVER);
 }
 
 /**
@@ -430,10 +584,10 @@ static void pi_set_simplicials(problem_instance_t *pi) {
 
                /* get free color by inspecting all neighbors */
                ifn = simpl->ifn;
-               irn = get_irn_for_graph_nr(pi->co->chordal_env->irg, ifn->nnr);
+               irn = get_irn_for_graph_nr(get_irg(pi->co), ifn->nnr);
                bitset_clear_all(used_cols);
                foreach_neighb(ifn, other) {
-                       other_irn = get_irn_for_graph_nr(pi->co->chordal_env->irg, other->nnr);
+                       other_irn = get_irn_for_graph_nr(get_irg(pi->co), other->nnr);
                        if (!is_removed(other_irn)) /* only inspect nodes which are in graph right now */
                                bitset_set(used_cols, get_irn_col(pi->co, other_irn));
                }
@@ -453,7 +607,7 @@ static void pi_set_simplicials(problem_instance_t *pi) {
 static void pi_apply_solution(problem_instance_t *pi) {
        int i;
        double *sol;
-       sol_state_t state;
+       lpp_sol_state_t state;
        DBG((dbg, LEVEL_2, "Applying solution...\n"));
 
 #ifdef DO_STAT
@@ -463,9 +617,9 @@ static void pi_apply_solution(problem_instance_t *pi) {
 
        sol = xmalloc((pi->last_x_var+1) * sizeof(*sol));
        state = lpp_get_solution(pi->curr_lp, sol, 1, pi->last_x_var);
-       if (state != optimal) {
+       if (state != lpp_optimal) {
                printf("Solution state is not 'optimal': %d\n", state);
-               assert(state >= feasible && "The solution should at least be feasible!");
+               assert(state >= lpp_feasible && "The solution should at least be feasible!");
        }
        for (i=0; i<pi->last_x_var; ++i) {
                int nnr, col;
@@ -474,9 +628,9 @@ static void pi_apply_solution(problem_instance_t *pi) {
                if (sol[i] > 1-EPSILON) { /* split varibale name into components */
                        lpp_get_var_name(pi->curr_lp, 1+i, var_name, sizeof(var_name));
                        if (split_var(var_name, &nnr, &col) == 2) {
-                               DBG((dbg, LEVEL_2, "Irn %n  Idx %d  Var %s  Val %f\n", get_irn_for_graph_nr(pi->co->chordal_env->irg, nnr), i, var_name, sol[i]));
+                               DBG((dbg, LEVEL_2, "Irn %n  Idx %d  Var %s  Val %f\n", get_irn_for_graph_nr(get_irg(pi->co), nnr), i, var_name, sol[i]));
                                DBG((dbg, LEVEL_2, "x%d = %d\n", nnr, col));
-                               set_irn_col(pi->co, get_irn_for_graph_nr(pi->co->chordal_env->irg, nnr), col);
+                               set_irn_col(pi->co, get_irn_for_graph_nr(get_irg(pi->co), nnr), col);
                        } else
                                assert(0 && "This should be a x-var");
                }
@@ -488,9 +642,9 @@ void co_ilp_opt(copy_opt_t *co) {
 
        dbg = firm_dbg_register("ir.be.copyoptilp");
        if (!strcmp(co->name, DEBUG_IRG))
-               firm_dbg_set_mask(dbg, DEBUG_LVL_ILP);
+               firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_ILP);
        else
-               firm_dbg_set_mask(dbg, DEBUG_LVL);
+               firm_dbg_set_mask(dbg, DEBUG_LVL_ILP);
 
        pi = new_pi(co);
        if (!pi->all_simplicial) {
@@ -499,7 +653,7 @@ void co_ilp_opt(copy_opt_t *co) {
                snprintf(buf, sizeof(buf), "%s.mps", co->name);
                lpp_dump(pi->curr_lp, buf);
 #endif
-               pi_solve_ilp(pi, lpp_solve_local);
+               pi_solve_ilp(pi);
                pi_apply_solution(pi);
                pi_set_simplicials(pi);
        }