Adapted to stat events
[libfirm] / ir / be / belower.c
index 36c7aff..b6112eb 100644 (file)
@@ -1,11 +1,28 @@
-/**
- * Author:      Christian Wuerdig
- * Date:        2005/12/14
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
- * CVS-Id:      $Id$
+/*
+ * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
  *
- * Performs lowering of perm nodes and spill/reload optimization.
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
+/**
+ * @file
+ * @brief       Performs lowering of perm nodes. Inserts copies to assure register constraints.
+ * @author      Christian Wuerdig
+ * @date        14.12.2005
+ * @version     $Id$
  */
 #ifdef HAVE_CONFIG_H
 #include "config.h"
 
 #include "ircons.h"
 #include "debug.h"
-
-#include "bearch.h"
-#include "belower.h"
-#include "benode_t.h"
-#include "bechordal_t.h"
-#include "besched_t.h"
-
+#include "irhooks.h"
+#include "xmalloc.h"
+#include "irnodeset.h"
 #include "irgmod.h"
 #include "iredges_t.h"
 #include "irgwalk.h"
 
-#ifdef HAVE_MALLOC_H
- #include <malloc.h>
-#endif
-#ifdef HAVE_ALLOCA_H
- #include <alloca.h>
-#endif
-
-#undef is_Perm
-#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
-
-#undef is_Call
-#define is_Call(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_call)
+#include "bearch_t.h"
+#include "belower.h"
+#include "benode_t.h"
+#include "besched_t.h"
+#include "bestat.h"
+#include "bessaconstr.h"
+#include "benodesets.h"
+#include "beintlive_t.h"
+
+#undef KEEP_ALIVE_COPYKEEP_HACK
+
+/* associates op with it's copy and CopyKeep */
+typedef struct {
+       ir_node *op;         /* an irn which must be different */
+       ir_nodeset_t copies; /* all non-spillable copies of this irn */
+       const arch_register_class_t *cls;
+} op_copy_assoc_t;
+
+/* environment for constraints */
+typedef struct {
+       be_irg_t       *birg;
+       pset           *op_set;
+       struct obstack obst;
+       DEBUG_ONLY(firm_dbg_module_t *dbg;)
+} constraint_env_t;
 
 /* lowering walker environment */
 typedef struct _lower_env_t {
-       be_chordal_env_t  *chord_env;
-       int                do_copy;
-       firm_dbg_module_t *dbg_module;
+       be_irg_t         *birg;
+       const arch_env_t *arch_env;
+       unsigned          do_copy : 1;
+       DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
 } lower_env_t;
 
 /* holds a perm register pair */
@@ -71,6 +98,15 @@ typedef struct _perm_cycle_t {
        perm_type_t             type;        /**< type (CHAIN or CYCLE) */
 } perm_cycle_t;
 
+//
+/* Compare the two operands */
+static int cmp_op_copy_assoc(const void *a, const void *b) {
+       const op_copy_assoc_t *op1 = a;
+       const op_copy_assoc_t *op2 = b;
+
+       return op1->op != op2->op;
+}
+
 /* Compare the in registers of two register pairs */
 static int compare_reg_pair(const void *a, const void *b) {
        const reg_pair_t *pair_a = a;
@@ -246,21 +282,24 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int
  * @param walk_env The environment
  */
 static void lower_perm_node(ir_node *irn, void *walk_env) {
+       ir_graph        *irg = get_irn_irg(irn);
        const arch_register_class_t *reg_class;
        const arch_env_t            *arch_env;
-       firm_dbg_module_t           *mod;
-       lower_env_t     *env = walk_env;
+       lower_env_t     *env         = walk_env;
+       int             real_size    = 0;
+       int             keep_perm    = 0;
+       int             n, i, pn, do_copy, j, n_ops;
        reg_pair_t      *pairs;
        const ir_edge_t *edge;
        perm_cycle_t    *cycle;
-       int              n, i, pn, do_copy, j;
        ir_node         *sched_point, *block, *in[2];
        ir_node         *arg1, *arg2, *res1, *res2;
        ir_node         *cpyxchg = NULL;
+       DEBUG_ONLY(firm_dbg_module_t *mod;)
 
-       arch_env = env->chord_env->birg->main_env->arch_env;
+       arch_env = env->arch_env;
        do_copy  = env->do_copy;
-       mod      = env->dbg_module;
+       DEBUG_ONLY(mod = env->dbg_module;)
        block    = get_nodes_block(irn);
 
        /*
@@ -270,6 +309,7 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                        should be ok.
        */
        sched_point = sched_prev(irn);
+       DBG((mod, LEVEL_1, "perm: %+F\n", irn));
        DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
        assert(sched_point && "Perm is not scheduled or has no predecessor");
 
@@ -306,15 +346,19 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                        /* We have to check for a special case:
                                The in-node could be a Proj from a Perm. In this case,
                                we need to correct the projnum */
-                       if (is_Perm(arch_env, pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
+                       if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
                                set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
                        }
 
+#ifdef SCHEDULE_PROJS
                        /* remove the proj from the schedule */
                        sched_remove(pairs[i].out_node);
+#endif
 
                        /* reroute the edges from the proj to the argument */
-                       edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg);
+                       exchange(pairs[i].out_node, pairs[i].in_node);
+                       //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
+                       //set_irn_n(pairs[i].out_node, 0, new_Bad());
 
                        pairs[i].checked = 1;
                }
@@ -325,9 +369,13 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                do_copy = 0;
        }
 
+       real_size = n - get_n_checked_pairs(pairs, n);
+
+       be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
+
        /* check for cycles and chains */
        while (get_n_checked_pairs(pairs, n) < n) {
-               i = 0;
+               i = n_ops = 0;
 
                /* go to the first not-checked pair */
                while (pairs[i].checked) i++;
@@ -340,11 +388,14 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                }
                DB((mod, LEVEL_1, "\n"));
 
-               /* We don't need to do anything if we have a Perm with two
+               /*
+                       We don't need to do anything if we have a Perm with two
                        elements which represents a cycle, because those nodes
-                       already represent exchange nodes */
+                       already represent exchange nodes
+               */
                if (n == 2 && cycle->type == PERM_CYCLE) {
                        free(cycle);
+                       keep_perm = 1;
                        continue;
                }
 
@@ -358,7 +409,6 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
 
                        res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
                        res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
-
                        /*
                                If we have a cycle and don't copy: we need to create exchange nodes
                                NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
@@ -403,32 +453,38 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                DBG((mod, LEVEL_1, "%+F                        (%+F, %s) and (%+F, %s)\n",
                                        irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
 
-                               cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
+                               cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
+                               n_ops++;
 
                                if (i > 0) {
                                        /* cycle is not done yet */
                                        int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
 
                                        /* create intermediate proj */
-                                       res1 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0);
+                                       res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
 
                                        /* set as in for next Perm */
                                        pairs[pidx].in_node = res1;
                                }
                                else {
+#ifdef SCHEDULE_PROJS
                                        sched_remove(res1);
+#endif
                                }
 
+#ifdef SCHEDULE_PROJS
                                sched_remove(res2);
+#endif
 
                                set_Proj_pred(res2, cpyxchg);
                                set_Proj_proj(res2, 0);
                                set_Proj_pred(res1, cpyxchg);
                                set_Proj_proj(res1, 1);
 
+#ifdef SCHEDULE_PROJS
                                sched_add_after(sched_point, res1);
                                sched_add_after(sched_point, res2);
-
+#endif
                                arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
                                arch_set_irn_register(arch_env, res1, cycle->elems[i]);
 
@@ -444,12 +500,14 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
                                        irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
 
-                               cpyxchg = be_new_Copy(reg_class, env->chord_env->irg, block, arg1);
+                               cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
                                arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
+                               n_ops++;
 
+#ifdef SCHEDULE_PROJS
                                /* remove the proj from the schedule */
                                sched_remove(res2);
-
+#endif
                                /* exchange copy node and proj */
                                exchange(res2, cpyxchg);
 
@@ -461,12 +519,17 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                        }
                }
 
+               be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
+
                free((void *) cycle->elems);
                free(cycle);
        }
 
        /* remove the perm from schedule */
-       sched_remove(irn);
+       if (! keep_perm) {
+               sched_remove(irn);
+               be_kill_node(irn);
+       }
 }
 
 
@@ -482,32 +545,38 @@ static int get_n_out_edges(const ir_node *irn) {
        return cnt;
 }
 
-static ir_node *belower_skip_proj(ir_node *irn) {
+static INLINE ir_node *belower_skip_proj(ir_node *irn) {
        while(is_Proj(irn))
                irn = get_Proj_pred(irn);
        return irn;
 }
 
-static void fix_in(ir_node *irn, ir_node *old, ir_node *nw) {
-       int i, n;
-
-       irn = belower_skip_proj(irn);
-       n   = get_irn_arity(irn);
-
-       for (i = 0; i < n; i++) {
-               if (get_irn_n(irn, i) == old) {
-                       set_irn_n(irn, i, nw);
-                       break;
-               }
+static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) {
+       const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
+       ir_node          *block    = get_nodes_block(irn);
+       ir_node          *cur_node;
+
+       for (cur_node = sched_prev(irn);
+               ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
+               cur_node = sched_prev(cur_node))
+       {
+               if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill))
+                       return cur_node;
        }
+
+       return NULL;
 }
 
-static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node *other_different) {
-       const arch_env_t          *arch_env = birg->main_env->arch_env;
-       ir_node                   *in[2], *keep, *cpy, *temp;
-       ir_node                   *block = get_nodes_block(irn);
-       firm_dbg_module_t         *mod   = firm_dbg_register("firm.be.lower");
-       const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
+static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
+       be_irg_t                    *birg     = env->birg;
+       ir_graph                    *irg      = be_get_birg_irg(birg);
+       pset                        *op_set   = env->op_set;
+       const arch_env_t            *arch_env = be_get_birg_arch_env(birg);
+       ir_node                     *block    = get_nodes_block(irn);
+       const arch_register_class_t *cls      = arch_get_irn_reg_class(arch_env, other_different, -1);
+       ir_node                     *in[2], *keep, *cpy;
+       op_copy_assoc_t             key, *entry;
+       DEBUG_ONLY(firm_dbg_module_t *mod     = env->dbg;)
 
        if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
                DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
@@ -519,53 +588,77 @@ static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node *
        /* in block far far away                             */
        /* The copy is optimized later if not needed         */
 
-       temp = new_rd_Unknown(birg->irg, get_irn_mode(other_different));
-       cpy = be_new_Copy(cls, birg->irg, block, temp);
-       be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
+       /* check if already exists such a copy in the schedule immediatly before */
+       cpy = find_copy(env, belower_skip_proj(irn), other_different);
+       if (! cpy) {
+               cpy = be_new_Copy(cls, irg, block, other_different);
+               be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
+               DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
+       }
+       else {
+               DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
+       }
 
        in[0] = irn;
        in[1] = cpy;
 
-       /* Let the irn use the copy instead of the old other_different */
-       fix_in(irn, other_different, cpy);
-
        /* Add the Keep resp. CopyKeep and reroute the users */
        /* of the other_different irn in case of CopyKeep.   */
        if (get_n_out_edges(other_different) == 0) {
-               keep = be_new_Keep(cls, birg->irg, block, 2, in);
+               keep = be_new_Keep(cls, irg, block, 2, in);
        }
        else {
-               keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different));
-               edges_reroute(other_different, keep, birg->irg);
+               keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
+               be_node_set_reg_class(keep, 1, cls);
        }
 
-       /* after rerouting: let the copy point to the other_different irn */
-       set_irn_n(cpy, 0, other_different);
+       DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
+
+       /* insert copy and keep into schedule */
+       assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
+       if (! sched_is_scheduled(cpy))
+               sched_add_before(belower_skip_proj(irn), cpy);
+       sched_add_after(irn, keep);
+
+       /* insert the other different and it's copies into the set */
+       key.op         = other_different;
+       entry          = pset_find(op_set, &key, nodeset_hash(other_different));
+
+       if (! entry) {
+               entry         = obstack_alloc(&env->obst, sizeof(*entry));
+               ir_nodeset_init(&entry->copies);
+               entry->op     = other_different;
+               entry->cls    = cls;
+       }
 
-       DBG((mod, LEVEL_1, "created %+F for %+F to assure should_be_different\n", keep, irn));
+       /* insert copy */
+       ir_nodeset_insert(&entry->copies, cpy);
+
+       /* insert keep in case of CopyKeep */
+       if (be_is_CopyKeep(keep)) {
+               ir_nodeset_insert(&entry->copies, keep);
+       }
+
+       pset_insert(op_set, entry, nodeset_hash(other_different));
 }
 
 /**
  * Checks if node has a should_be_different constraint in output
  * and adds a Keep then to assure the constraint.
  */
-static void assure_different_constraints(ir_node *irn, be_irg_t *birg) {
-       const arch_env_t          *arch_env = birg->main_env->arch_env;
+static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
        const arch_register_req_t *req;
-       arch_register_req_t        req_temp;
-       int i, n;
+       const arch_env_t          *arch_env = be_get_birg_arch_env(env->birg);
 
-       req = arch_get_register_req(arch_env, &req_temp, irn, -1);
+       req = arch_get_register_req(arch_env, irn, -1);
 
-       if (req) {
-               if (arch_register_req_is(req, should_be_different)) {
-                       gen_assure_different_pattern(irn, birg, req->other_different);
-               }
-               else if (arch_register_req_is(req, should_be_different_from_all)) {
-                       n = get_irn_arity(belower_skip_proj(irn));
-                       for (i = 0; i < n; i++) {
-                               gen_assure_different_pattern(irn, birg, get_irn_n(belower_skip_proj(irn), i));
-                       }
+       if (arch_register_req_is(req, should_be_different)) {
+               ir_node *different_from = get_irn_n(belower_skip_proj(irn), req->other_different);
+               gen_assure_different_pattern(irn, different_from, env);
+       } else if (arch_register_req_is(req, should_be_different_from_all)) {
+               int i, n = get_irn_arity(belower_skip_proj(irn));
+               for (i = 0; i < n; i++) {
+                       gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env);
                }
        }
 }
@@ -582,13 +675,134 @@ static void assure_constraints_walker(ir_node *irn, void *walk_env) {
        if (is_Block(irn))
                return;
 
-       if (mode_is_datab(get_irn_mode(irn)))
+       if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
                assure_different_constraints(irn, walk_env);
 
        return;
 }
 
+/**
+ * Melt all copykeeps pointing to the same node
+ * (or Projs of the same node), copying the same operand.
+ */
+static void melt_copykeeps(constraint_env_t *cenv) {
+       be_irg_t *birg = cenv->birg;
+       ir_graph *irg  = be_get_birg_irg(birg);
+       op_copy_assoc_t *entry;
+
+       /* for all */
+       foreach_pset(cenv->op_set, entry) {
+               int     idx, num_ck;
+               ir_node *cp;
+               struct obstack obst;
+               ir_nodeset_iterator_t iter;
+               ir_node **ck_arr, **melt_arr;
+
+               obstack_init(&obst);
+
+               /* collect all copykeeps */
+               num_ck = idx = 0;
+               foreach_ir_nodeset(&entry->copies, cp, iter) {
+                       if (be_is_CopyKeep(cp)) {
+                               obstack_grow(&obst, &cp, sizeof(cp));
+                               ++num_ck;
+                       }
+#ifdef KEEP_ALIVE_COPYKEEP_HACK
+                       else {
+                               set_irn_mode(cp, mode_ANY);
+                               keep_alive(cp);
+                       }
+#endif /* KEEP_ALIVE_COPYKEEP_HACK */
+               }
+
+               /* compare each copykeep with all other copykeeps */
+               ck_arr = (ir_node **)obstack_finish(&obst);
+               for (idx = 0; idx < num_ck; ++idx) {
+                       ir_node *ref, *ref_mode_T;
+
+                       if (ck_arr[idx]) {
+                               int j, n_melt;
+                               ir_node **new_ck_in;
+                               ir_node *new_ck;
+                               ir_node *sched_pt = NULL;
+
+                               n_melt     = 1;
+                               ref        = ck_arr[idx];
+                               ref_mode_T = skip_Proj(get_irn_n(ref, 1));
+                               obstack_grow(&obst, &ref, sizeof(ref));
+
+                               DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
+
+                               /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
+                               for (j = 0; j < num_ck; ++j) {
+                                       ir_node *cur_ck = ck_arr[j];
+
+                                       if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
+                                               obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
+                                               ir_nodeset_remove(&entry->copies, cur_ck);
+                                               DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
+                                               ck_arr[j] = NULL;
+                                               ++n_melt;
+                                               sched_remove(cur_ck);
+                                       }
+                               }
+                               ck_arr[idx] = NULL;
+
+                               /* check, if we found some candidates for melting */
+                               if (n_melt == 1) {
+                                       DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
+                                       continue;
+                               }
+
+                               ir_nodeset_remove(&entry->copies, ref);
+                               sched_remove(ref);
+
+                               melt_arr = (ir_node **)obstack_finish(&obst);
+                               /* melt all found copykeeps */
+                               NEW_ARR_A(ir_node *, new_ck_in, n_melt);
+                               for (j = 0; j < n_melt; ++j) {
+                                       new_ck_in[j] = get_irn_n(melt_arr[j], 1);
+
+                                       /* now, we can kill the melted keep, except the */
+                                       /* ref one, we still need some information      */
+                                       if (melt_arr[j] != ref)
+                                               be_kill_node(melt_arr[j]);
+                               }
+
+#ifdef KEEP_ALIVE_COPYKEEP_HACK
+                               new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
+                               keep_alive(new_ck);
+#else
+                               new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, get_irn_mode(ref));
+#endif /* KEEP_ALIVE_COPYKEEP_HACK */
 
+                               /* set register class for all keeped inputs */
+                               for (j = 1; j <= n_melt; ++j)
+                                       be_node_set_reg_class(new_ck, j, entry->cls);
+
+                               ir_nodeset_insert(&entry->copies, new_ck);
+
+                               /* find scheduling point */
+                               if (get_irn_mode(ref_mode_T) == mode_T) {
+                                       /* walk along the Projs */
+                                       for (sched_pt = sched_next(ref_mode_T); is_Proj(sched_pt) || be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt); sched_pt = sched_next(sched_pt))
+                                               /* just walk along the schedule until a non-Proj/Keep/CopyKeep node is found*/ ;
+                               }
+                               else {
+                                       sched_pt = ref_mode_T;
+                               }
+
+                               sched_add_before(sched_pt, new_ck);
+                               DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
+
+                               /* finally: kill the reference copykeep */
+                               be_kill_node(ref);
+                       }
+               }
+
+               obstack_free(&obst, NULL);
+       }
+}
 
 /**
  * Walks over all nodes to assure register constraints.
@@ -596,10 +810,220 @@ static void assure_constraints_walker(ir_node *irn, void *walk_env) {
  * @param birg  The birg structure containing the irg
  */
 void assure_constraints(be_irg_t *birg) {
-       irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, birg);
+       ir_graph         *irg      = be_get_birg_irg(birg);
+       const arch_env_t *arch_env = be_get_birg_arch_env(birg);
+       constraint_env_t cenv;
+       op_copy_assoc_t  *entry;
+       ir_node          **nodes;
+       FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
+
+       be_assure_dom_front(birg);
+
+       DEBUG_ONLY(cenv.dbg = mod;)
+       cenv.birg   = birg;
+       cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
+       obstack_init(&cenv.obst);
+
+       irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
+
+       /* melt copykeeps, pointing to projs of */
+       /* the same mode_T node and keeping the */
+       /* same operand                         */
+       melt_copykeeps(&cenv);
+
+       /* for all */
+       foreach_pset(cenv.op_set, entry) {
+               int     n;
+               ir_node *cp;
+               ir_nodeset_iterator_t iter;
+               be_ssa_construction_env_t senv;
+
+               n     = ir_nodeset_size(&entry->copies);
+               nodes = alloca(n * sizeof(nodes[0]));
+
+               /* put the node in an array */
+               DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
+
+               /* collect all copies */
+               n = 0;
+               foreach_ir_nodeset(&entry->copies, cp, iter) {
+                       nodes[n++] = cp;
+                       DB((mod, LEVEL_1, ", %+F ", cp));
+               }
+
+               DB((mod, LEVEL_1, "\n"));
+
+               /* introduce the copies for the operand and it's copies */
+               be_ssa_construction_init(&senv, birg);
+               be_ssa_construction_add_copy(&senv, entry->op);
+               be_ssa_construction_add_copies(&senv, nodes, n);
+               be_ssa_construction_fix_users(&senv, entry->op);
+               be_ssa_construction_destroy(&senv);
+
+               /* Could be that not all CopyKeeps are really needed, */
+               /* so we transform unnecessary ones into Keeps.       */
+               foreach_ir_nodeset(&entry->copies, cp, iter) {
+                       if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
+                               ir_node *keep;
+                               int     n = get_irn_arity(cp);
+
+                               keep = be_new_Keep(arch_get_irn_reg_class(arch_env, cp, -1),
+                                       irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
+                               sched_add_before(cp, keep);
+
+                               /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
+                               sched_remove(cp);
+                               be_kill_node(cp);
+                       }
+               }
+
+               ir_nodeset_destroy(&entry->copies);
+       }
+
+       del_pset(cenv.op_set);
+       obstack_free(&cenv.obst, NULL);
+       be_liveness_invalidate(be_get_birg_liveness(birg));
 }
 
 
+/**
+ * Push nodes that do not need to be permed through the Perm.
+ * This is commonly a reload cascade at block ends.
+ * @note This routine needs interference.
+ * @note Probably, we can implement it a little more efficient.
+ *       Especially searching the frontier lazily might be better.
+ * @param perm The perm.
+ * @param data The walker data (lower_env_t).
+ * @return     1, if there is something left to perm over.
+ *             0, if removed the complete perm.
+ */
+static int push_through_perm(ir_node *perm, void *data)
+{
+       lower_env_t *env       = data;
+       const arch_env_t *aenv = env->arch_env;
+
+       ir_graph *irg     = get_irn_irg(perm);
+       ir_node *bl       = get_nodes_block(perm);
+       int n             = get_irn_arity(perm);
+       int *map          = alloca(n * sizeof(map[0]));
+       ir_node **projs   = alloca(n * sizeof(projs[0]));
+       bitset_t *keep    = bitset_alloca(n);
+       ir_node *frontier = sched_first(bl);
+       FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
+
+       int i, new_size, n_keep;
+       const ir_edge_t *edge;
+       ir_node *last_proj, *irn;
+       const arch_register_class_t *cls;
+
+       DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
+
+       /* get some proj and find out the register class of the proj. */
+       foreach_out_edge (perm, edge) {
+               last_proj  = get_edge_src_irn(edge);
+               cls = arch_get_irn_reg_class(aenv, last_proj, -1);
+               assert(is_Proj(last_proj));
+               break;
+       }
+
+       /* find the point in the schedule after which the
+        * potentially movable nodes must be defined.
+        * A perm will only be pushed up to first instruction
+        * which lets an operand of itself die.  */
+
+       sched_foreach_reverse_from (sched_prev(perm), irn) {
+               for(i = get_irn_arity(irn) - 1; i >= 0; --i) {
+                       ir_node *op = get_irn_n(irn, i);
+                       if(arch_irn_consider_in_reg_alloc(aenv, cls, op)
+                               && !values_interfere(env->birg, op, last_proj)) {
+                               frontier = sched_next(irn);
+                               goto found_front;
+                       }
+               }
+       }
+found_front:
+
+       DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
+
+       foreach_out_edge (perm, edge) {
+               ir_node *proj  = get_edge_src_irn(edge);
+               int nr         = get_Proj_proj(proj);
+               ir_node *op    = get_irn_n(perm, nr);
+
+               assert(nr < n);
+
+               /* we will need the last Proj as an insertion point
+                * for the instruction(s) pushed through the Perm */
+               if (sched_comes_after(last_proj, proj))
+                       last_proj = proj;
+
+               projs[nr] = proj;
+
+               bitset_set(keep, nr);
+               if (!is_Proj(op) && get_nodes_block(op) == bl
+                               && (op == frontier || sched_comes_after(frontier, op))) {
+                       for (i = get_irn_arity(op) - 1; i >= 0; --i) {
+                               ir_node *opop = get_irn_n(op, i);
+                               if (!arch_irn_consider_in_reg_alloc(aenv, cls, opop)) {
+                                       bitset_clear(keep, nr);
+                                       break;
+                               }
+                       }
+               }
+       }
+
+       n_keep = bitset_popcnt(keep);
+
+       /* well, we could not push enything through the perm */
+       if (n_keep == n)
+               return 1;
+
+       assert(is_Proj(last_proj));
+
+       DBG((mod, LEVEL_2, "\tkeep: %d, total: %d, mask: %b\n", n_keep, n, keep));
+       last_proj = sched_next(last_proj);
+       for (new_size = 0, i = 0; i < n; ++i) {
+               ir_node *proj  = projs[i];
+
+               if (bitset_is_set(keep, i)) {
+                       map[i] = new_size++;
+                       set_Proj_proj(proj, map[i]);
+                       DBG((mod, LEVEL_1, "\targ %d remap to %d\n", i, map[i]));
+               }
+
+               else {
+                       ir_node *move = get_irn_n(perm, i);
+
+                       DBG((mod, LEVEL_2, "\tmoving %+F before %+F, killing %+F\n", move, last_proj, proj));
+
+                       /* move the movable node in front of the Perm */
+                       sched_remove(move);
+                       sched_add_before(last_proj, move);
+
+                       /* give it the proj's register */
+                       arch_set_irn_register(aenv, move, arch_get_irn_register(aenv, proj));
+
+                       /* reroute all users of the proj to the moved node. */
+                       edges_reroute(proj, move, irg);
+
+#ifdef SCHEDULE_PROJS
+                       /* remove the proj from the schedule. */
+                       sched_remove(proj);
+#endif
+
+                       /* and like it to bad so it is no more in the use array of the perm */
+                       set_Proj_pred(proj, get_irg_bad(irg));
+
+                       map[i] = -1;
+               }
+
+       }
+
+       if (n_keep > 0)
+               be_Perm_reduce(perm, new_size, map);
+
+       return n_keep > 0;
+}
 
 /**
  * Calls the corresponding lowering function for the node.
@@ -608,36 +1032,35 @@ void assure_constraints(be_irg_t *birg) {
  * @param walk_env The walker environment
  */
 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
-       lower_env_t      *env      = walk_env;
-       const arch_env_t *arch_env = env->chord_env->birg->main_env->arch_env;
-
-       if (!is_Block(irn) && !is_Proj(irn)) {
-               if (is_Perm(arch_env, irn)) {
-                       lower_perm_node(irn, walk_env);
+       if (! is_Block(irn) && ! is_Proj(irn)) {
+               if (be_is_Perm(irn)) {
+                       int perm_stayed = 1; //push_through_perm(irn, walk_env);
+                       if (perm_stayed)
+                               lower_perm_node(irn, walk_env);
                }
        }
 
        return;
 }
 
-
-
 /**
  * Walks over all blocks in an irg and performs lowering need to be
  * done after register allocation (e.g. perm lowering).
  *
- * @param chord_env The chordal environment containing the irg
+ * @param birg      The birg object
  * @param do_copy   1 == resolve cycles with a free reg if available
  */
-void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy) {
+void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
        lower_env_t env;
+       ir_graph    *irg = be_get_birg_irg(birg);
 
-       env.chord_env  = chord_env;
-       env.do_copy    = do_copy;
-       env.dbg_module = firm_dbg_register("firm.be.lower");
+       env.birg     = birg;
+       env.arch_env = be_get_birg_arch_env(birg);
+       env.do_copy  = do_copy;
+       FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
 
-       irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);
-}
+       /* we will need interference */
+       be_liveness_assure_chk(be_get_birg_liveness(birg));
 
-#undef is_Perm
-#undef is_Call
+       irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);
+}