remove #ifdef HAVE_CONFIG_Hs
[libfirm] / ir / be / belower.c
index 16bafc8..2934508 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -24,9 +24,7 @@
  * @date        14.12.2005
  * @version     $Id$
  */
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include <stdlib.h>
 
@@ -38,6 +36,7 @@
 #include "irgmod.h"
 #include "iredges_t.h"
 #include "irgwalk.h"
+#include "array_t.h"
 
 #include "bearch_t.h"
 #include "belower.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 */
+/** 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 */
+       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 */
+/** Environment for constraints. */
 typedef struct {
        be_irg_t       *birg;
        pset           *op_set;
@@ -65,15 +63,14 @@ typedef struct {
        DEBUG_ONLY(firm_dbg_module_t *dbg;)
 } constraint_env_t;
 
-/* lowering walker environment */
+/** Lowering walker environment. */
 typedef struct _lower_env_t {
        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 */
+/** 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 */
@@ -91,15 +88,14 @@ typedef enum _perm_type_t {
        PERM_COPY
 } perm_type_t;
 
-/* structure to represent cycles or chains in a perm */
+/** 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 */
+/** 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;
@@ -107,7 +103,7 @@ static int cmp_op_copy_assoc(const void *a, const void *b) {
        return op1->op != op2->op;
 }
 
-/* Compare the in registers of two register pairs */
+/** 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;
@@ -118,7 +114,7 @@ static int compare_reg_pair(const void *a, const void *b) {
                return -1;
 }
 
-/* returns the number register pairs marked as checked */
+/** 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;
 
@@ -178,14 +174,14 @@ static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_
        if (in_out) {
                for (i = 0; i < n; i++) {
                        /* out register matches */
-                       if (pairs[i].out_reg->index == reg_idx)
+                       if ((int) pairs[i].out_reg->index == reg_idx)
                                return i;
                }
        }
        else {
                for (i = 0; i < n; i++) {
                        /* in register matches */
-                       if (pairs[i].in_reg->index == reg_idx)
+                       if ((int) pairs[i].in_reg->index == reg_idx)
                                return i;
                }
        }
@@ -226,7 +222,7 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int
        }
 
        /* assume worst case: all remaining pairs build a cycle or chain */
-       cycle->elems    = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
+       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;
@@ -284,20 +280,17 @@ static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int
 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;
        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;
        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->arch_env;
        do_copy  = env->do_copy;
        DEBUG_ONLY(mod = env->dbg_module;)
        block    = get_nodes_block(irn);
@@ -309,13 +302,14 @@ 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");
 
        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(arch_env, get_irn_n(irn, 0))->reg_class;
+       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) */
@@ -325,8 +319,8 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                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(arch_env, pairs[i].in_node);
-               pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
+               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++;
@@ -349,11 +343,6 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                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 */
                        exchange(pairs[i].out_node, pairs[i].in_node);
                        //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
@@ -370,15 +359,15 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
 
        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) {
+               perm_cycle_t *cycle;
+
                i = n_ops = 0;
 
                /* go to the first not-checked pair */
                while (pairs[i].checked) i++;
-               cycle = xcalloc(1, sizeof(*cycle));
+               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"));
@@ -465,27 +454,14 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                        /* 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]);
+                               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);
@@ -500,13 +476,9 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
                                        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(arch_env, cpyxchg, cycle->elems[i + 1]);
+                               arch_set_irn_register(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);
 
@@ -518,8 +490,6 @@ 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);
        }
@@ -527,39 +497,35 @@ static void lower_perm_node(ir_node *irn, void *walk_env) {
        /* remove the perm from schedule */
        if (! keep_perm) {
                sched_remove(irn);
-               be_kill_node(irn);
+               kill_node(irn);
        }
 }
 
 
 
-static int get_n_out_edges(const ir_node *irn) {
-       const ir_edge_t *edge;
-       int cnt = 0;
-
-       foreach_out_edge(irn, edge) {
-               cnt++;
-       }
-
-       return cnt;
+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(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;
+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(arch_env, cur_node, dont_spill))
+               if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
                        return cur_node;
        }
 
@@ -570,14 +536,14 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
        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);
+       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(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
+       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;
        }
@@ -587,8 +553,8 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
        /* in block far far away                             */
        /* The copy is optimized later if not needed         */
 
-       /* check if already exists such a copy in the schedule immediatly before */
-       cpy = find_copy(env, belower_skip_proj(irn), other_different);
+       /* 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);
@@ -603,13 +569,13 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
 
        /* 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, irg, block, 2, in);
-       }
-       else {
+       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));
 
@@ -620,8 +586,8 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
        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));
+       key.op = other_different;
+       entry  = pset_find(op_set, &key, hash_irn(other_different));
 
        if (! entry) {
                entry         = obstack_alloc(&env->obst, sizeof(*entry));
@@ -638,32 +604,46 @@ static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different,
                ir_nodeset_insert(&entry->copies, keep);
        }
 
-       pset_insert(op_set, entry, nodeset_hash(other_different));
+       pset_insert(op_set, entry, hash_irn(other_different));
 }
 
 /**
- * Checks if node has a should_be_different constraint in output
- * and adds a Keep then to assure the constraint.
+ * 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) {
-       const arch_register_req_t *req;
-       const arch_env_t          *arch_env = be_get_birg_arch_env(env->birg);
-
-       req = arch_get_register_req(arch_env, irn, -1);
-
-       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);
+       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.
  *
@@ -676,8 +656,6 @@ static void assure_constraints_walker(ir_node *irn, void *walk_env) {
 
        if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
                assure_different_constraints(irn, walk_env);
-
-       return;
 }
 
 /**
@@ -765,7 +743,7 @@ static void melt_copykeeps(constraint_env_t *cenv) {
                                        /* 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]);
+                                               kill_node(melt_arr[j]);
                                }
 
 #ifdef KEEP_ALIVE_COPYKEEP_HACK
@@ -775,27 +753,24 @@ static void melt_copykeeps(constraint_env_t *cenv) {
                                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 */
+                               /* 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 */
-                               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_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 */
-                               be_kill_node(ref);
+                               kill_node(ref);
                        }
                }
 
@@ -810,14 +785,11 @@ static void melt_copykeeps(constraint_env_t *cenv) {
  */
 void assure_constraints(be_irg_t *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);
@@ -866,13 +838,13 @@ void assure_constraints(be_irg_t *birg) {
                                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]);
+                               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);
-                               be_kill_node(cp);
+                               kill_node(cp);
                        }
                }
 
@@ -898,44 +870,47 @@ void assure_constraints(be_irg_t *birg) {
  */
 static int push_through_perm(ir_node *perm, void *data)
 {
-       lower_env_t *env       = data;
-       const arch_env_t *aenv = env->arch_env;
+       lower_env_t *env = data;
 
        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);
+       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, new_size, n_keep;
+       int i, n;
        const ir_edge_t *edge;
-       ir_node *last_proj, *irn;
-       const arch_register_class_t *cls;
+       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 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;
-       }
+       /* 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
+       /* 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.  */
-
+        * 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) {
+               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);
+                       if (arch_irn_consider_in_reg_alloc(cls, op) &&
+                           !values_interfere(env->birg, op, one_proj)) {
+                               frontier = irn;
                                goto found_front;
                        }
                }
@@ -944,84 +919,101 @@ 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;
-                               }
+       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;
                        }
                }
-       }
-
-       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]));
+               /* 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;
 
-               else {
-                       ir_node *move = get_irn_n(perm, i);
+               DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
 
-                       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(node);
+               sched_add_after(perm, node);
 
-                       /* 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(node, arch_get_irn_register(proj));
 
-                       /* 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, node, irg);
 
-                       /* reroute all users of the proj to the moved node. */
-                       edges_reroute(proj, move, irg);
+               /* and kill it */
+               set_Proj_pred(proj, new_Bad());
+               kill_node(proj);
 
-#ifdef SCHEDULE_PROJS
-                       /* remove the proj from the schedule. */
-                       sched_remove(proj);
-#endif
+               bitset_set(moved, input);
+               n_moved++;
 
-                       /* 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));
+               node = sched_prev(node);
+       }
 
-                       map[i] = -1;
-               }
+       /* 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;
        }
 
-       if (n_keep > 0)
-               be_Perm_reduce(perm, new_size, map);
+       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);
+       }
 
-       return n_keep > 0;
+       be_Perm_reduce(perm, new_size, map);
+       return 1;
 }
 
 /**
@@ -1031,15 +1023,18 @@ found_front:
  * @param walk_env The walker environment
  */
 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
-       if (! is_Block(irn) && ! is_Proj(irn)) {
-               if (be_is_Perm(irn)) {
-                       int perm_stayed = push_through_perm(irn, walk_env);
-                       if (perm_stayed)
-                               lower_perm_node(irn, 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;
 
-       return;
+       lower_perm_node(irn, walk_env);
 }
 
 /**
@@ -1053,9 +1048,8 @@ 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.arch_env = be_get_birg_arch_env(birg);
-       env.do_copy  = do_copy;
+       env.birg    = birg;
+       env.do_copy = do_copy;
        FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
 
        /* we will need interference */