[libfirm] / ir / be / belower.c

/*
 * Copyright (C) 1995-2008 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.
 *
 * 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     $Id$
 */
#include "config.h"

#include <stdlib.h>

#include "ircons.h"
#include "debug.h"
#include "irhooks.h"
#include "xmalloc.h"
#include "irnodeset.h"
#include "irgmod.h"
#include "iredges_t.h"
#include "irgwalk.h"
#include "array_t.h"

#include "bearch_t.h"
#include "belower.h"
#include "benode_t.h"
#include "besched_t.h"
#include "bestat.h"
#include "bessaconstr.h"
#include "beintlive_t.h"

#undef KEEP_ALIVE_COPYKEEP_HACK

/** Associates an 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_irg_t         *birg;
        unsigned          do_copy : 1;
        DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
} lower_env_t;

/** Holds a Perm register pair. */
typedef struct _reg_pair_t {
        const arch_register_t *in_reg;    /**< a perm IN register */
        ir_node               *in_node;   /**< the in node to which the register belongs */

        const arch_register_t *out_reg;   /**< a perm OUT register */
        ir_node               *out_node;  /**< the out node to which the register belongs */

        int                    checked;   /**< indicates whether the pair was check for cycle or not */
} reg_pair_t;

typedef enum _perm_type_t {
        PERM_CYCLE,
        PERM_CHAIN,
        PERM_SWAP,
        PERM_COPY
} perm_type_t;

/** Structure to represent cycles or chains in a Perm. */
typedef struct _perm_cycle_t {
        const arch_register_t **elems;       /**< the registers in the cycle */
        int                     n_elems;     /**< number of elements in the cycle */
        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;
        const reg_pair_t *pair_b = b;

        if (pair_a->in_reg->index > pair_b->in_reg->index)
                return 1;
        else
                return -1;
}

/** returns the number register pairs marked as checked. */
static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
        int i, n_checked = 0;

        for (i = 0; i < n; i++) {
                if (pairs[i].checked)
                        n_checked++;
        }

        return n_checked;
}

/**
 * Gets the node corresponding to a register from an array of register pairs.
 * NOTE: The given registers pairs and the register to look for must belong
 *       to the same register class.
 *
 * @param pairs  The array of register pairs
 * @param n      The number of pairs
 * @param reg    The register to look for
 * @param in_out 0 == look for IN register, 1 == look for OUT register
 * @return The corresponding node or NULL if not found
 */
static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
        int i;

        if (in_out) {
                for (i = 0; i < n; i++) {
                        /* out register matches */
                        if (pairs[i].out_reg->index == reg->index)
                                return pairs[i].out_node;
                }
        }
        else {
                for (i = 0; i < n; i++) {
                        /* in register matches */
                        if (pairs[i].in_reg->index == reg->index)
                                return pairs[i].in_node;
                }
        }

        return NULL;
}

/**
 * Gets the index in the register pair array where the in/out register
 * corresponds to reg_idx.
 *
 * @param pairs  The array of register pairs
 * @param n      The number of pairs
 * @param reg    The register index to look for
 * @param in_out 0 == look for IN register, 1 == look for OUT register
 * @return The corresponding index in pairs or -1 if not found
 */
static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
        int i;

        if (in_out) {
                for (i = 0; i < n; i++) {
                        /* out register matches */
                        if ((int) pairs[i].out_reg->index == reg_idx)
                                return i;
                }
        }
        else {
                for (i = 0; i < n; i++) {
                        /* in register matches */
                        if ((int) pairs[i].in_reg->index == reg_idx)
                                return i;
                }
        }

        return -1;
}

/**
 * Gets an array of register pairs and tries to identify a cycle or chain starting
 * at position start.
 *
 * @param cycle Variable to hold the cycle
 * @param pairs Array of register pairs
 * @param start Index to start
 * @return The cycle or chain
 */
static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
        int head         = pairs[start].in_reg->index;
        int cur_idx      = pairs[start].out_reg->index;
        int cur_pair_idx = start;
        int n_pairs_done = get_n_checked_pairs(pairs, n);
        int idx;
        perm_type_t cycle_tp = PERM_CYCLE;

        /* We could be right in the middle of a chain, so we need to find the start */
        while (head != cur_idx) {
                /* goto previous register in cycle or chain */
                cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);

                if (cur_pair_idx < 0) {
                        cycle_tp = PERM_CHAIN;
                        break;
                }
                else {
                        head  = pairs[cur_pair_idx].in_reg->index;
                        start = cur_pair_idx;
                }
        }

        /* assume worst case: all remaining pairs build a cycle or chain */
        cycle->elems    = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
        cycle->n_elems  = 2;  /* initial number of elements is 2 */
        cycle->elems[0] = pairs[start].in_reg;
        cycle->elems[1] = pairs[start].out_reg;
        cycle->type     = cycle_tp;
        cur_idx         = pairs[start].out_reg->index;

        idx = 2;
        /* check for cycle or end of a chain */
        while (cur_idx != head) {
                /* goto next register in cycle or chain */
                cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);

                if (cur_pair_idx < 0)
                        break;

                cur_idx = pairs[cur_pair_idx].out_reg->index;

                /* it's not the first element: insert it */
                if (cur_idx != head) {
                        cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
                        cycle->n_elems++;
                }
                else {
                        /* we are there where we started -> CYCLE */
                        cycle->type = PERM_CYCLE;
                }
        }

        /* mark all pairs having one in/out register with cycle in common as checked */
        for (idx = 0; idx < cycle->n_elems; idx++) {
                cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);

                if (cur_pair_idx >= 0)
                        pairs[cur_pair_idx].checked = 1;

                cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);

                if (cur_pair_idx >= 0)
                        pairs[cur_pair_idx].checked = 1;
        }

        return cycle;
}

/**
 * Lowers a perm node.  Resolves cycles and creates a bunch of
 * copy and swap operations to permute registers.
 * Note: The caller of this function has to make sure, that irn
 *       is a Perm node.
 *
 * @param irn      The perm node
 * @param block    The block the perm node belongs to
 * @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;
        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;
        ir_node         *sched_point, *block, *in[2];
        ir_node         *arg1, *arg2, *res1, *res2;
        ir_node         *cpyxchg = NULL;
        DEBUG_ONLY(firm_dbg_module_t *mod;)

        do_copy  = env->do_copy;
        DEBUG_ONLY(mod = env->dbg_module;)
        block    = get_nodes_block(irn);

        /*
                Get the schedule predecessor node to the perm
                NOTE: This works with auto-magic. If we insert the
                        new copy/exchange nodes after this node, everything
                        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");

        n = get_irn_arity(irn);
        assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");

        reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
        pairs     = alloca(n * sizeof(pairs[0]));

        /* build the list of register pairs (in, out) */
        i = 0;
        foreach_out_edge(irn, edge) {
                pairs[i].out_node = get_edge_src_irn(edge);
                pn                = get_Proj_proj(pairs[i].out_node);
                pairs[i].in_node  = get_irn_n(irn, pn);

                pairs[i].in_reg  = arch_get_irn_register(pairs[i].in_node);
                pairs[i].out_reg = arch_get_irn_register(pairs[i].out_node);

                pairs[i].checked = 0;
                i++;
        }

        /* sort the register pairs by the indices of the in registers */
        qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);

        /* Mark all equal pairs as checked, and exchange the OUT proj with
                the IN node. */
        for (i = 0; i < n; i++) {
                if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
                        DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
                                irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));

                        /* 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 (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));
                        }

                        /* reroute the edges from the proj to the argument */
                        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;
                }
        }

        /* Set do_copy to 0 if it's on but we have no free register */
        if (do_copy) {
                do_copy = 0;
        }

        real_size = n - get_n_checked_pairs(pairs, n);

        /* check for cycles and chains */
        while (get_n_checked_pairs(pairs, n) < n) {
                perm_cycle_t *cycle;

                i = n_ops = 0;

                /* go to the first not-checked pair */
                while (pairs[i].checked) i++;
                cycle = XMALLOCZ(perm_cycle_t);
                cycle = get_perm_cycle(cycle, pairs, n, i);

                DB((mod, LEVEL_1, "%+F: following %s created:\n  ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
                for (j = 0; j < cycle->n_elems; j++) {
                        DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
                }
                DB((mod, LEVEL_1, "\n"));

                /*
                        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
                */
                if (n == 2 && cycle->type == PERM_CYCLE) {
                        free(cycle);
                        keep_perm = 1;
                        continue;
                }

//TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
//        the copy cascade (this reduces the cycle into a chain)

                /* build copy/swap nodes from back to front */
                for (i = cycle->n_elems - 2; i >= 0; i--) {
                        arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
                        arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);

                        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
                                IN_1  = in node with register i
                                IN_2  = in node with register i + 1
                                OUT_1 = out node with register i + 1
                                OUT_2 = out node with register i
                        */
                        if (cycle->type == PERM_CYCLE && !do_copy) {
                                in[0] = arg1;
                                in[1] = arg2;

                                /* At this point we have to handle the following problem:     */
                                /*                                                            */
                                /* If we have a cycle with more than two elements, then       */
                                /* this could correspond to the following Perm node:          */
                                /*                                                            */
                                /*   +----+   +----+   +----+                                 */
                                /*   | r1 |   | r2 |   | r3 |                                 */
                                /*   +-+--+   +-+--+   +--+-+                                 */
                                /*     |        |         |                                   */
                                /*     |        |         |                                   */
                                /*   +-+--------+---------+-+                                 */
                                /*   |         Perm         |                                 */
                                /*   +-+--------+---------+-+                                 */
                                /*     |        |         |                                   */
                                /*     |        |         |                                   */
                                /*   +-+--+   +-+--+   +--+-+                                 */
                                /*   |Proj|   |Proj|   |Proj|                                 */
                                /*   | r2 |   | r3 |   | r1 |                                 */
                                /*   +----+   +----+   +----+                                 */
                                /*                                                            */
                                /* This node is about to be split up into two 2x Perm's       */
                                /* for which we need 4 Proj's and the one additional Proj     */
                                /* of the first Perm has to be one IN of the second. So in    */
                                /* general we need to create one additional Proj for each     */
                                /* "middle" Perm and set this to one in node of the successor */
                                /* Perm.                                                      */

                                DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
                                        irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
                                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, 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(irg, block, cpyxchg, get_irn_mode(res1), 0);

                                        /* set as in for next Perm */
                                        pairs[pidx].in_node = res1;
                                }

                                set_Proj_pred(res2, cpyxchg);
                                set_Proj_proj(res2, 0);
                                set_Proj_pred(res1, cpyxchg);
                                set_Proj_proj(res1, 1);

                                arch_set_irn_register(res2, cycle->elems[i + 1]);
                                arch_set_irn_register(res1, cycle->elems[i]);

                                /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
                                sched_add_after(sched_point, cpyxchg);

                                DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));

                                /* set the new scheduling point */
                                sched_point = res1;
                        }
                        else {
                                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, irg, block, arg1);
                                arch_set_irn_register(cpyxchg, cycle->elems[i + 1]);
                                n_ops++;

                                /* exchange copy node and proj */
                                exchange(res2, cpyxchg);

                                /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
                                sched_add_after(sched_point, cpyxchg);

                                /* set the new scheduling point */
                                sched_point = cpyxchg;
                        }
                }

                free((void *) cycle->elems);
                free(cycle);
        }

        /* remove the perm from schedule */
        if (! keep_perm) {
                sched_remove(irn);
                kill_node(irn);
        }
}



static int has_irn_users(const ir_node *irn) {
        return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
}

/**
 * Skip all Proj nodes.
 */
static INLINE ir_node *belower_skip_proj(ir_node *irn) {
        while(is_Proj(irn))
                irn = get_Proj_pred(irn);
        return irn;
}

static ir_node *find_copy(ir_node *irn, ir_node *op)
{
        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(cur_node, dont_spill))
                        return cur_node;
        }

        return NULL;
}

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;
        ir_node                     *block    = get_nodes_block(irn);
        const arch_register_class_t *cls      = arch_get_irn_reg_class(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(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));
                return;
        }

        /* Make a not spillable copy of the different node   */
        /* this is needed because the different irn could be */
        /* in block far far away                             */
        /* The copy is optimized later if not needed         */

        /* check if already exists such a copy in the schedule immediately before */
        cpy = find_copy(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;

        /* Add the Keep resp. CopyKeep and reroute the users */
        /* of the other_different irn in case of CopyKeep.   */
        if (has_irn_users(other_different)) {
                keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
                be_node_set_reg_class(keep, 1, cls);
        }
        else {
                keep = be_new_Keep(cls, irg, block, 2, in);
        }

        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, hash_irn(other_different));

        if (! entry) {
                entry         = obstack_alloc(&env->obst, sizeof(*entry));
                ir_nodeset_init(&entry->copies);
                entry->op     = other_different;
                entry->cls    = cls;
        }

        /* 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, hash_irn(other_different));
}

/**
 * Checks if node has a must_be_different constraint in output and adds a Keep
 * then to assure the constraint.
 */
static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
        ir_node                   *skipped_irn = belower_skip_proj(irn);
        const arch_register_req_t *req         = arch_get_register_req(irn, -1);

        if (arch_register_req_is(req, must_be_different)) {
                const unsigned other = req->other_different;
                int i;

                if (arch_register_req_is(req, should_be_same)) {
                        const unsigned same = req->other_same;

                        if (is_po2(other) && is_po2(same)) {
                                int idx_other = ntz(other);
                                int idx_same  = ntz(same);

                                /*
                                 * We can safely ignore a should_be_same x must_be_different y
                                 * IFF both inputs are equal!
                                 */
                                if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
                                        return;
                                }
                        }
                }
                for (i = 0; 1U << i <= other; ++i) {
                        if (other & (1U << i)) {
                                ir_node *different_from = get_irn_n(skipped_irn, i);
                                gen_assure_different_pattern(irn, different_from, env);
                        }
                }
        }
}

/**
 * Calls the functions to assure register constraints.
 *
 * @param irn      The node to be checked for lowering
 * @param walk_env The walker environment
 */
static void assure_constraints_walker(ir_node *irn, void *walk_env) {
        if (is_Block(irn))
                return;

        if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
                assure_different_constraints(irn, walk_env);
}

/**
 * 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)
                                                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 kept 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 */
                                sched_pt = ref_mode_T;
                                do {
                                        /* just walk along the schedule until a non-Keep/CopyKeep node is found */
                                        sched_pt = sched_next(sched_pt);
                                } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));

                                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 */
                                kill_node(ref);
                        }
                }

                obstack_free(&obst, NULL);
        }
}

/**
 * Walks over all nodes to assure register constraints.
 *
 * @param birg  The birg structure containing the irg
 */
void assure_constraints(be_irg_t *birg) {
        ir_graph         *irg      = be_get_birg_irg(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");

        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(cp, -1),
                                        irg, get_nodes_block(cp), n, 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);
                                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;

        ir_graph *irg     = get_irn_irg(perm);
        ir_node *bl       = get_nodes_block(perm);
        ir_node *node;
        int  arity        = get_irn_arity(perm);
        int *map;
        int *proj_map;
        bitset_t *moved   = bitset_alloca(arity);
        int n_moved;
        int new_size;
        ir_node *frontier = bl;
        FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");

        int i, n;
        const ir_edge_t *edge;
        ir_node *one_proj = NULL, *irn;
        const arch_register_class_t *cls = NULL;

        DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));

        /* get some Proj and find out the register class of that Proj. */
        edge     = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
        one_proj = get_edge_src_irn(edge);
        assert(is_Proj(one_proj));
        cls      = arch_get_irn_reg_class(one_proj, -1);

        /* 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.
         * If we would allow to move the Perm above this instruction,
         * the former dead operand would be live now at the point of
         * the Perm, increasing the register pressure by one.
         */
        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(cls, op) &&
                            !values_interfere(env->birg, op, one_proj)) {
                                frontier = irn;
                                goto found_front;
                        }
                }
        }
found_front:

        DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));

        node = sched_prev(perm);
        n_moved = 0;
        while(!sched_is_begin(node)) {
                const arch_register_req_t *req;
                int                        input = -1;
                ir_node                   *proj;

                /* search if node is a INPUT of Perm */
                foreach_out_edge(perm, edge) {
                        ir_node *out = get_edge_src_irn(edge);
                        int      pn  = get_Proj_proj(out);
                        ir_node *in  = get_irn_n(perm, pn);
                        if (node == in) {
                                proj  = out;
                                input = pn;
                                break;
                        }
                }
                /* it wasn't an input to the perm, we can't do anything more */
                if(input < 0)
                        break;
                if(!sched_comes_after(frontier, node))
                        break;
                if (arch_irn_is(node, modify_flags))
                        break;
                if(is_Proj(node)) {
                        req = arch_get_register_req(get_Proj_pred(node),
                                                    -1 - get_Proj_proj(node));
                } else {
                        req = arch_get_register_req(node, -1);
                }
                if(req->type != arch_register_req_type_normal)
                        break;
                for(i = get_irn_arity(node) - 1; i >= 0; --i) {
                        ir_node *opop = get_irn_n(node, i);
                        if (arch_irn_consider_in_reg_alloc(cls, opop)) {
                                break;
                        }
                }
                if(i >= 0)
                        break;

                DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));

                /* move the movable node in front of the Perm */
                sched_remove(node);
                sched_add_after(perm, node);

                /* give it the proj's register */
                arch_set_irn_register(node, arch_get_irn_register(proj));

                /* reroute all users of the proj to the moved node. */
                edges_reroute(proj, node, irg);

                /* and kill it */
                set_Proj_pred(proj, new_Bad());
                kill_node(proj);

                bitset_set(moved, input);
                n_moved++;

                node = sched_prev(node);
        }

        /* well, we could not push anything through the perm */
        if(n_moved == 0)
                return 1;

        new_size = arity - n_moved;
        if(new_size == 0) {
                return 0;
        }

        map      = alloca(new_size * sizeof(map[0]));
        proj_map = alloca(arity * sizeof(proj_map[0]));
        memset(proj_map, -1, sizeof(proj_map[0]));
        n   = 0;
        for(i = 0; i < arity; ++i) {
                if(bitset_is_set(moved, i))
                        continue;
                map[n]      = i;
                proj_map[i] = n;
                n++;
        }
        assert(n == new_size);
        foreach_out_edge(perm, edge) {
                ir_node *proj = get_edge_src_irn(edge);
                int      pn   = get_Proj_proj(proj);
                pn = proj_map[pn];
                assert(pn >= 0);
                set_Proj_proj(proj, pn);
        }

        be_Perm_reduce(perm, new_size, map);
        return 1;
}

/**
 * Calls the corresponding lowering function for the node.
 *
 * @param irn      The node to be checked for lowering
 * @param walk_env The walker environment
 */
static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
        int perm_stayed;

        if (is_Block(irn) || is_Proj(irn))
                return;
        if (!be_is_Perm(irn))
                return;

        perm_stayed = push_through_perm(irn, walk_env);
        if (!perm_stayed)
                return;

        lower_perm_node(irn, walk_env);
}

/**
 * Walks over all blocks in an irg and performs lowering need to be
 * done after register allocation (e.g. perm lowering).
 *
 * @param birg      The birg object
 * @param do_copy   1 == resolve cycles with a free reg if available
 */
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.birg    = birg;
        env.do_copy = do_copy;
        FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");

        /* we will need interference */
        be_liveness_assure_chk(be_get_birg_liveness(birg));

        irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);
}