make code a bit more readble
[libfirm] / ir / be / bechordal.c
index dc6ea1a..f5015e4 100644 (file)
@@ -28,6 +28,7 @@
 #include "bitset.h"
 #include "iterator.h"
 #include "bipartite.h"
+#include "hungarian.h"
 
 #include "irmode_t.h"
 #include "irgraph_t.h"
@@ -41,7 +42,6 @@
 
 #include "beutil.h"
 #include "besched.h"
-#include "benumb_t.h"
 #include "besched_t.h"
 #include "belive_t.h"
 #include "benode_t.h"
@@ -49,6 +49,8 @@
 #include "beirgmod.h"
 #include "beifg.h"
 #include "beinsn_t.h"
+#include "bestatevent.h"
+#include "beirg_t.h"
 
 #include "bechordal_t.h"
 #include "bechordal_draw.h"
@@ -132,8 +134,8 @@ static INLINE border_t *border_add(be_chordal_env_t *env, struct list_head *head
                 */
                set_irn_link(irn, def);
 
-               b->magic = BORDER_FOURCC;
-               def->magic = BORDER_FOURCC;
+               DEBUG_ONLY(b->magic = BORDER_FOURCC);
+               DEBUG_ONLY(def->magic = BORDER_FOURCC);
        }
 
        /*
@@ -221,14 +223,69 @@ static be_insn_t *chordal_scan_insn(be_chordal_env_t *env, ir_node *irn)
 
 static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
 {
-       be_insn_t *insn      = chordal_scan_insn(env, irn);
-       bitset_t *def_constr = bitset_alloca(env->cls->n_regs);
-       bitset_t *tmp        = bitset_alloca(env->cls->n_regs);
-       int i;
+       const arch_env_t *aenv = env->birg->main_env->arch_env;
+       bitset_t *def_constr   = bitset_alloca(env->cls->n_regs);
+       bitset_t *tmp          = bitset_alloca(env->cls->n_regs);
+       ir_node *bl            = get_nodes_block(irn);
+       be_lv_t *lv            = env->birg->lv;
+
+       be_insn_t *insn;
+       int i, j;
+
+       for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
+               ir_node *op = get_irn_n(irn, i);
+
+               const arch_register_t *reg;
+               arch_register_req_t req;
+
+               if (arch_get_irn_reg_class(aenv, irn, i) == env->cls) {
+                       reg = arch_get_irn_register(aenv, op);
+
+                       if (reg && arch_register_type_is(reg, ignore)) {
+                               arch_get_register_req(aenv, &req, irn, i);
+
+                               if (arch_register_req_is(&req, limited)) {
+                                       bitset_clear_all(tmp);
+                                       req.limited(req.limited_env, tmp);
+
+                                       if (! bitset_is_set(tmp, reg->index)) {
+                                               ir_node *copy = be_new_Copy(env->cls, env->irg, bl, op);
+                                               be_stat_ev("constr_copy", 1);
+
+                                               sched_add_before(irn, copy);
+                                               set_irn_n(irn, i, copy);
+                                               DBG((env->dbg, LEVEL_3, "inserting ignore arg copy %+F for %+F pos %d\n", copy, irn, i));
+                                       }
+                               }
+                       }
+               }
+       }
+
+    insn = chordal_scan_insn(env, irn);
 
        if(!insn->has_constraints)
                goto end;
 
+       /* insert copies for nodes that occur constrained more than once. */
+       for(i = insn->use_start; i < insn->n_ops; ++i) {
+               be_operand_t *op = &insn->ops[i];
+
+               if(op->has_constraints) {
+                       for(j = i + 1; j < insn->n_ops; ++j) {
+                               be_operand_t *a_op = &insn->ops[j];
+
+                               if(a_op->carrier == op->carrier && a_op->has_constraints) {
+                                       ir_node *copy = be_new_Copy(env->cls, env->irg, bl, op->carrier);
+                                       be_stat_ev("constr_copy", 1);
+
+                                       sched_add_before(insn->irn, copy);
+                                       set_irn_n(insn->irn, a_op->pos, copy);
+                                       DBG((env->dbg, LEVEL_3, "inserting multiple constr copy %+F for %+F pos %d\n", copy, insn->irn, a_op->pos));
+                               }
+                       }
+               }
+       }
+
        /* collect all registers occuring in out constraints. */
        for(i = 0; i < insn->use_start; ++i) {
                be_operand_t *op = &insn->ops[i];
@@ -252,9 +309,7 @@ static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
                        2) lives through the node.
                        3) is constrained to a register occuring in out constraints.
                */
-               if(op->has_constraints && values_interfere(env->lv, insn->irn, op->carrier) && bitset_popcnt(tmp) > 0) {
-                       ir_node *bl   = get_nodes_block(insn->irn);
-
+               if(op->has_constraints && values_interfere(lv, insn->irn, op->carrier) && bitset_popcnt(tmp) > 0) {
                        /*
                                only create the copy if the operand is no copy.
                                this is necessary since the assure constraints phase inserts
@@ -267,7 +322,7 @@ static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
                                sched_add_before(insn->irn, copy);
                                set_irn_n(insn->irn, op->pos, copy);
                                DBG((env->dbg, LEVEL_3, "inserting constr copy %+F for %+F pos %d\n", copy, insn->irn, op->pos));
-                               be_liveness_update(env->lv, op->carrier);
+                               be_liveness_update(lv, op->carrier);
                        }
                }
        }
@@ -298,6 +353,7 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t
        int n_defs   = be_insn_n_defs(insn);
        bitset_t *bs = bitset_alloca(env->cls->n_regs);
        int *pairing = alloca(MAX(n_defs, n_uses) * sizeof(pairing[0]));
+       be_lv_t *lv  = env->birg->lv;
 
        int i, j;
 
@@ -315,7 +371,7 @@ static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t
                        int n_total;
                        const be_operand_t *op = &insn->ops[i];
 
-                       if (! values_interfere(env->lv, op->irn, op->carrier) && ! op->partner) {
+                       if (! values_interfere(lv, op->irn, op->carrier) && ! op->partner) {
                                bitset_clear_all(bs);
                                bitset_copy(bs, op->regs);
                                bitset_and(bs, out_op->regs);
@@ -347,6 +403,7 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, be_in
        ir_node *perm               = NULL;
        bitset_t *out_constr        = bitset_alloca(env->cls->n_regs);
        bitset_t *bs                = bitset_alloca(env->cls->n_regs);
+       be_lv_t *lv                 = env->birg->lv;
        DEBUG_ONLY(firm_dbg_module_t *dbg      = alloc_env->constr_dbg;)
 
        int i;
@@ -365,47 +422,22 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, be_in
                        bitset_or(out_constr, op->regs);
        }
 
-       /*
-               Now, figure out which input operand must be copied since it has input
-               constraints which are also output constraints.
-       */
        (void) bl;
        (void) copy;
        (void) bs;
-       (void) dbg;
-#if 0
-       for(i = insn->use_start; i < insn->n_ops; ++i) {
-               be_operand_t *op = &insn->ops[i];
-               if(op->has_constraints && (values_interfere(env->lv, op->carrier, insn->irn) || arch_irn_is(aenv, op->carrier, ignore))) {
-                       bitset_copy(bs, op->regs);
-                       bitset_and(bs, out_constr);
-
-                       /*
-                               The operand (interfering with the node) has input constraints
-                               which also occur as output constraints, so insert a copy.
-                       */
-                       if(bitset_popcnt(bs) > 0) {
-                               copy        = be_new_Copy(op->req.cls, env->irg, bl, op->carrier);
-                               op->carrier = copy;
-                               sched_add_before(insn->irn, copy);
-                               set_irn_n(insn->irn, op->pos, op->carrier);
-
-                               DBG((dbg, LEVEL_2, "adding copy for interfering and constrained op %+F\n", op->carrier));
-                       }
-               }
-       }
-#endif
+       DEBUG_ONLY((void) dbg;)
 
        /*
                Make the Perm, recompute liveness and re-scan the insn since the
                in operands are now the Projs of the Perm.
        */
-       perm = insert_Perm_after(aenv, env->lv, env->cls, env->dom_front, sched_prev(insn->irn));
+       perm = insert_Perm_after(aenv, lv, env->cls, env->birg->dom_front, sched_prev(insn->irn));
 
        /* Registers are propagated by insert_Perm_after(). Clean them here! */
        if(perm) {
                const ir_edge_t *edge;
 
+               be_stat_ev("constr_perm", get_irn_arity(perm));
                foreach_out_edge(perm, edge) {
                        ir_node *proj = get_edge_src_irn(edge);
                        arch_set_irn_register(aenv, proj, NULL);
@@ -416,7 +448,7 @@ static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, be_in
                        the Perm. Recomputing liveness is also a good idea if a Perm is inserted, since
                        the live sets may change.
                */
-               // be_liveness_recompute(env->lv);
+               // be_liveness_recompute(lv);
                obstack_free(&env->obst, insn);
                *the_insn = insn = chordal_scan_insn(env, insn->irn);
 
@@ -447,6 +479,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
        be_insn_t *insn        = chordal_scan_insn(env, irn);
        ir_node *res           = insn->next_insn;
        int be_silent          = *silent;
+       be_lv_t *lv            = env->birg->lv;
 
        if(insn->pre_colored) {
                int i;
@@ -476,7 +509,8 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                int n_regs             = env->cls->n_regs;
                bitset_t *bs           = bitset_alloca(n_regs);
                ir_node **alloc_nodes  = alloca(n_regs * sizeof(alloc_nodes[0]));
-               bipartite_t *bp        = bipartite_new(n_regs, n_regs);
+               hungarian_problem_t *bp= hungarian_new(n_regs, n_regs, 2, HUNGARIAN_MATCH_PERFECT);
+//             bipartite_t *bp        = bipartite_new(n_regs, n_regs);
                int *assignment        = alloca(n_regs * sizeof(assignment[0]));
                pmap *partners         = pmap_create();
                DEBUG_ONLY(firm_dbg_module_t *dbg = alloc_env->constr_dbg;)
@@ -485,6 +519,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                long col;
                const ir_edge_t *edge;
                ir_node *perm = NULL;
+               int match_res, cost;
 
                /*
                        prepare the constraint handling of this node.
@@ -522,14 +557,15 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                                DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier, bs));
 
                                bitset_foreach(bs, col)
-                                       bipartite_add(bp, n_alloc, col);
+                                       hungarian_add(bp, n_alloc, col, 1);
+//                                     bipartite_add(bp, n_alloc, col);
 
                                n_alloc++;
                        }
                }
 
                /*
-                       Put all nodes which live by the constrained instruction also to the
+                       Put all nodes which live through the constrained instruction also to the
                        allocation bipartite graph. They are considered unconstrained.
                */
                if(perm) {
@@ -538,15 +574,17 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
 
                                assert(is_Proj(proj));
 
-                               if(values_interfere(env->lv, proj, irn) && !pmap_contains(partners, proj)) {
+                               if(values_interfere(lv, proj, irn) && !pmap_contains(partners, proj)) {
                                        assert(n_alloc < n_regs);
                                        alloc_nodes[n_alloc] = proj;
                                        pmap_insert(partners, proj, NULL);
 
                                        bitset_clear_all(bs);
                                        arch_put_non_ignore_regs(aenv, env->cls, bs);
+                                       bitset_andnot(bs, env->ignore_colors);
                                        bitset_foreach(bs, col)
-                                               bipartite_add(bp, n_alloc, col);
+                                               hungarian_add(bp, n_alloc, col, 1);
+//                                             bipartite_add(bp, n_alloc, col);
 
                                        n_alloc++;
                                }
@@ -554,7 +592,10 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                }
 
                /* Compute a valid register allocation. */
-               bipartite_matching(bp, assignment);
+               hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
+               match_res = hungarian_solve(bp, assignment, &cost, 1);
+               assert(match_res == 0 && "matching failed");
+               //bipartite_matching(bp, assignment);
 
                /* Assign colors obtained from the matching. */
                for(i = 0; i < n_alloc; ++i) {
@@ -573,7 +614,7 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                                        continue;
 
                                arch_set_irn_register(aenv, nodes[j], reg);
-                               pset_hinsert_ptr(alloc_env->pre_colored, nodes[j]);
+                               (void) pset_hinsert_ptr(alloc_env->pre_colored, nodes[j]);
                                DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", nodes[j], reg->name));
                        }
                }
@@ -611,7 +652,8 @@ static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *i
                        }
                }
 
-               bipartite_free(bp);
+               //bipartite_free(bp);
+               hungarian_free(bp);
                pmap_destroy(partners);
        }
 
@@ -669,14 +711,15 @@ static void pressure(ir_node *block, void *env_ptr)
        be_chordal_env_t *env             = alloc_env->chordal_env;
        bitset_t *live                    = alloc_env->live;
        ir_node *irn;
+       be_lv_t *lv                       = env->birg->lv;
        DEBUG_ONLY(firm_dbg_module_t *dbg            = env->dbg;)
 
        int i, n;
        unsigned step = 0;
        unsigned pressure = 0;
        struct list_head *head;
-       pset *live_in  = be_lv_pset_put_in(env->lv, block, pset_new_ptr_default());
-       pset *live_end = be_lv_pset_put_end(env->lv, block, pset_new_ptr_default());
+       pset *live_in  = be_lv_pset_put_in(lv, block, pset_new_ptr_default());
+       pset *live_end = be_lv_pset_put_end(lv, block, pset_new_ptr_default());
 
        DBG((dbg, LEVEL_1, "Computing pressure in block %+F\n", block));
        bitset_clear_all(live);
@@ -770,7 +813,8 @@ static void assign(ir_node *block, void *env_ptr)
        bitset_t *in_colors         = alloc_env->in_colors;
        const arch_env_t *arch_env  = env->birg->main_env->arch_env;
        struct list_head *head      = get_block_border_head(env, block);
-       pset *live_in               = be_lv_pset_put_in(env->lv, block, pset_new_ptr_default());
+       be_lv_t *lv                 = env->birg->lv;
+       pset *live_in               = be_lv_pset_put_in(lv, block, pset_new_ptr_default());
 
        const ir_node *irn;
        border_t *b;
@@ -812,8 +856,7 @@ static void assign(ir_node *block, void *env_ptr)
        }
 
        /*
-        * Mind that the sequence
-        * of defs from back to front defines a perfect
+        * Mind that the sequence of defs from back to front defines a perfect
         * elimination order. So, coloring the definitions from first to last
         * will work.
         */
@@ -826,7 +869,7 @@ static void assign(ir_node *block, void *env_ptr)
                 * Assign a color, if it is a local def. Global defs already have a
                 * color.
                 */
-               if(b->is_def && !be_is_live_in(env->lv, block, irn)) {
+               if(b->is_def && !be_is_live_in(lv, block, irn)) {
                        const arch_register_t *reg;
                        int col = NO_COLOR;
 
@@ -874,11 +917,14 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
 {
        be_chordal_alloc_env_t env;
        char buf[256];
+       be_irg_t *birg = chordal_env->birg;
 
        int colors_n          = arch_register_class_n_regs(chordal_env->cls);
        ir_graph *irg         = chordal_env->irg;
 
 
+       be_assure_dom_front(birg);
+       be_assure_liveness(birg);
        assure_doms(irg);
 
        env.chordal_env   = chordal_env;
@@ -898,7 +944,6 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
                be_dump(chordal_env->irg, buf, dump_ir_block_graph_sched);
        }
 
-       be_numbering(irg);
        env.live = bitset_malloc(get_irg_last_idx(chordal_env->irg));
 
        /* First, determine the pressure */
@@ -907,8 +952,6 @@ void be_ra_chordal_color(be_chordal_env_t *chordal_env)
        /* Assign the colors */
        dom_tree_walk_irg(irg, assign, NULL, &env);
 
-       be_numbering_done(irg);
-
        if(chordal_env->opts->dump_flags & BE_CH_DUMP_TREE_INTV) {
                plotter_t *plotter;
                ir_snprintf(buf, sizeof(buf), "ifg_%s_%F.eps", chordal_env->cls->name, irg);