lower_mode_b: refactoring, make it more robust
[libfirm] / ir / be / beprefalloc.c
index 4f0bf00..af2ece4 100644 (file)
@@ -271,7 +271,7 @@ static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
                        ir_node *node)
 {
-       const arch_register_req_t *req = arch_get_register_req_out(node);
+       const arch_register_req_t *req = arch_get_irn_register_req(node);
        if (req->type & arch_register_req_type_limited) {
                const unsigned *limited = req->limited;
                float           penalty = weight * DEF_FACTOR;
@@ -352,7 +352,7 @@ static void analyze_block(ir_node *block, void *data)
                info = get_allocation_info(node);
                for (i = 0; i < arity; ++i) {
                        ir_node                   *op  = get_irn_n(node, i);
-                       const arch_register_req_t *req = arch_get_register_req_out(op);
+                       const arch_register_req_t *req = arch_get_irn_register_req(op);
                        if (req->cls != cls)
                                continue;
 
@@ -374,7 +374,7 @@ static void analyze_block(ir_node *block, void *data)
                                if (!arch_irn_consider_in_reg_alloc(cls, op))
                                        continue;
 
-                               req = arch_get_register_req(node, i);
+                               req = arch_get_irn_register_req_in(node, i);
                                if (!(req->type & arch_register_req_type_limited))
                                        continue;
 
@@ -390,7 +390,7 @@ static void analyze_block(ir_node *block, void *data)
 
 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
 {
-       const arch_register_req_t *req = arch_get_register_req_out(node);
+       const arch_register_req_t *req = arch_get_irn_register_req(node);
 
        /* should be same constraint? */
        if (req->type & arch_register_req_type_should_be_same) {
@@ -650,7 +650,7 @@ static bool try_optimistic_split(ir_node *to_split, ir_node *before,
         * (so we don't split away the values produced because of
         *  must_be_different constraints) */
        original_insn = skip_Proj(info->original_value);
-       if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
+       if (arch_get_irn_flags(original_insn) & arch_irn_flags_dont_spill)
                return false;
 
        from_reg        = arch_get_irn_register(to_split);
@@ -721,7 +721,7 @@ static bool try_optimistic_split(ir_node *to_split, ir_node *before,
                return false;
 
        reg  = arch_register_for_index(cls, r);
-       copy = be_new_Copy(cls, block, to_split);
+       copy = be_new_Copy(block, to_split);
        mark_as_copy_of(copy, to_split);
        /* hacky, but correct here */
        if (assignments[arch_register_get_index(from_reg)] == to_split)
@@ -759,7 +759,7 @@ static void assign_reg(const ir_node *block, ir_node *node,
                return;
        }
 
-       req = arch_get_register_req_out(node);
+       req = arch_get_irn_register_req(node);
        /* ignore reqs must be preassigned */
        assert (! (req->type & arch_register_req_type_ignore));
 
@@ -882,7 +882,7 @@ static void assign_reg(const ir_node *block, ir_node *node,
  *                     registers.
  */
 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
-                             unsigned *permutation)
+                           unsigned *permutation)
 {
        unsigned  *n_used = ALLOCANZ(unsigned, n_regs);
        ir_node   *block;
@@ -922,7 +922,7 @@ static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
 
                /* create a copy */
                src  = assignments[old_r];
-               copy = be_new_Copy(cls, block, src);
+               copy = be_new_Copy(block, src);
                sched_add_before(before, copy);
                reg = arch_register_for_index(cls, r);
                DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
@@ -1131,7 +1131,7 @@ static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
                if (!arch_irn_consider_in_reg_alloc(cls, op))
                        continue;
 
-               req = arch_get_register_req(node, i);
+               req = arch_get_irn_register_req_in(node, i);
                if (!(req->type & arch_register_req_type_limited))
                        continue;
 
@@ -1146,8 +1146,15 @@ static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
                }
        }
 
-       /* add all combinations, then remove not allowed ones */
+       /* add all combinations, except for not allowed ones */
        for (l = 0; l < n_regs; ++l) {
+               if (!rbitset_is_set(normal_regs, l)) {
+                       char name[15];
+                       snprintf(name, sizeof(name), "%u_to_%u", l, l);
+                       lpp_vars[l*n_regs+l] = lpp_add_var(lpp, name, lpp_binary, 1);
+                       continue;
+               }
+
                for (r = 0; r < n_regs; ++r) {
                        if (!rbitset_is_set(normal_regs, r))
                                continue;
@@ -1158,26 +1165,48 @@ static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
                            && rbitset_is_set(forbidden_regs, r))
                                continue;
 
+                       char name[15];
+                       snprintf(name, sizeof(name), "%u_to_%u", l, r);
+
                        double costs = l==r ? 9 : 8;
                        lpp_vars[l*n_regs+r]
-                               = lpp_add_var(lpp, NULL, lpp_binary, costs);
+                               = lpp_add_var(lpp, name, lpp_binary, costs);
                        assert(lpp_vars[l*n_regs+r] > 0);
                }
        }
        /* add constraints */
        for (l = 0; l < n_regs; ++l) {
+               int constraint;
                /* only 1 destination per register */
-               int constraint = lpp_add_cst(lpp, NULL, lpp_equal, 1);
+               constraint = -1;
                for (r = 0; r < n_regs; ++r) {
                        int var = lpp_vars[l*n_regs+r];
                        if (var == 0)
                                continue;
+                       if (constraint < 0) {
+                               char name[64];
+                               snprintf(name, sizeof(name), "%u_to_dest", l);
+                               constraint = lpp_add_cst(lpp, name, lpp_equal, 1);
+                       }
                        lpp_set_factor_fast(lpp, constraint, var, 1);
                }
                /* each destination used by at most 1 value */
-               constraint = lpp_add_cst(lpp, NULL, lpp_less_equal, 1);
+               constraint = -1;
+               for (r = 0; r < n_regs; ++r) {
+                       int var = lpp_vars[r*n_regs+l];
+                       if (var == 0)
+                               continue;
+                       if (constraint < 0) {
+                               char name[64];
+                               snprintf(name, sizeof(name), "one_to_%u", l);
+                               constraint = lpp_add_cst(lpp, name, lpp_less_equal, 1);
+                       }
+                       lpp_set_factor_fast(lpp, constraint, var, 1);
+               }
        }
 
+       lpp_dump_plain(lpp, fopen("lppdump.txt", "w"));
+
        /* solve lpp */
        {
                ir_graph     *irg     = get_irn_irg(node);
@@ -1201,8 +1230,15 @@ static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
                                }
                        }
                        assert(dest_reg != (unsigned)-1);
-                       assignment[l] = dest_reg;
+                       assignment[dest_reg] = l;
+               }
+
+               fprintf(stderr, "Assignment: ");
+               for (l = 0; l < n_regs; ++l) {
+                       fprintf(stderr, "%u ", assignment[l]);
                }
+               fprintf(stderr, "\n");
+               fflush(stdout);
                permute_values(live_nodes, node, assignment);
        }
        lpp_free(lpp);
@@ -1241,7 +1277,7 @@ static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
                        continue;
 
                /* are there any limitations for the i'th operand? */
-               req = arch_get_register_req(node, i);
+               req = arch_get_irn_register_req_in(node, i);
                if (req->width > 1)
                        double_width = true;
                if (!(req->type & arch_register_req_type_limited))
@@ -1281,6 +1317,7 @@ static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
        }
 
        if (double_width) {
+               /* only the ILP variant can solve this yet */
                solve_lpp(live_nodes, node, forbidden_regs, live_through_regs);
                return;
        }
@@ -1322,7 +1359,7 @@ static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
                if (!arch_irn_consider_in_reg_alloc(cls, op))
                        continue;
 
-               req = arch_get_register_req(node, i);
+               req = arch_get_irn_register_req_in(node, i);
                if (!(req->type & arch_register_req_type_limited))
                        continue;
 
@@ -1678,7 +1715,7 @@ static void allocate_coalesce_block(ir_node *block, void *data)
                int                        p;
 
                node = be_lv_get_irn(lv, block, i);
-               req  = arch_get_register_req_out(node);
+               req  = arch_get_irn_register_req(node);
                if (req->cls != cls)
                        continue;
 
@@ -1995,7 +2032,7 @@ static void spill(void)
        be_pre_spill_prepare_constr(irg, cls);
        be_timer_pop(T_RA_CONSTR);
 
-       dump(DUMP_RA, irg, "-spillprepare");
+       dump(DUMP_RA, irg, "spillprepare");
 
        /* spill */
        be_timer_push(T_RA_SPILL);
@@ -2006,7 +2043,7 @@ static void spill(void)
        check_for_memory_operands(irg);
        be_timer_pop(T_RA_SPILL_APPLY);
 
-       dump(DUMP_RA, irg, "-spill");
+       dump(DUMP_RA, irg, "spill");
 }
 
 /**