Remove the unused parameter const arch_env_t *env from arch_get_sp_bias().
[libfirm] / ir / be / beabi.c
index 57f1674..58ecd1e 100644 (file)
@@ -1,11 +1,28 @@
-/**
- * ABI lowering.
+/*
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
  *
- * @author Sebastian Hack
- * @date   7.3.2005
- * @cvsid  $Id$
+ * 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       Backend ABI implementation.
+ * @author      Sebastian Hack, Michael Beck
+ * @version     $Id$
+ */
 #ifdef HAVE_CONFIG_H
 # include "config.h"
 #endif
@@ -13,7 +30,6 @@
 #include "obst.h"
 #include "offset.h"
 
-#include "type.h"
 #include "irgopt.h"
 
 #include "irgraph_t.h"
 #include "height.h"
 #include "pdeq.h"
 #include "irtools.h"
+#include "raw_bitset.h"
+#include "error.h"
 
 #include "be.h"
 #include "beabi.h"
-#include "bearch.h"
+#include "bearch_t.h"
 #include "benode_t.h"
 #include "belive_t.h"
 #include "besched_t.h"
+#include "beirg_t.h"
+#include "bessaconstr.h"
 
 typedef struct _be_abi_call_arg_t {
        unsigned is_res   : 1;  /**< 1: the call argument is a return value. 0: it's a call parameter. */
        unsigned in_reg   : 1;  /**< 1: this argument is transmitted in registers. */
        unsigned on_stack : 1;  /**< 1: this argument is transmitted on the stack. */
 
-       int pos;
+       int                    pos;
        const arch_register_t *reg;
-       entity *stack_ent;
-       unsigned alignment;
-       unsigned space_before;
-       unsigned space_after;
+       ir_entity             *stack_ent;
+       ir_mode               *load_mode;
+       unsigned               alignment;    /**< stack alignment */
+       unsigned               space_before; /**< allocate space before */
+       unsigned               space_after;  /**< allocate space after */
 } be_abi_call_arg_t;
 
 struct _be_abi_call_t {
-       be_abi_call_flags_t flags;
-       const be_abi_callbacks_t *cb;
-       ir_type *between_type;
-       set *params;
+       be_abi_call_flags_t          flags;
+       int                          pop;
+       const be_abi_callbacks_t    *cb;
+       ir_type                     *between_type;
+       set                         *params;
+       const arch_register_class_t *cls_addr;
 };
 
 struct _be_abi_irg_t {
        struct obstack       obst;
-       be_stack_layout_t    *frame;        /**< The stack frame model. */
-       const be_irg_t       *birg;         /**< The back end IRG. */
-       const arch_isa_t     *isa;          /**< The isa. */
+       be_irg_t             *birg;         /**< The back end IRG. */
+       const arch_env_t     *arch_env;
        survive_dce_t        *dce_survivor;
 
        be_abi_call_t        *call;         /**< The ABI call information. */
        ir_type              *method_type;  /**< The type of the method of the IRG. */
 
        ir_node              *init_sp;      /**< The node representing the stack pointer
-                                                                            at the start of the function. */
-
-       ir_node              *start_barrier; /**< The barrier of the start block */
+                                                at the start of the function. */
 
        ir_node              *reg_params;   /**< The reg params node. */
        pmap                 *regs;         /**< A map of all callee-save and ignore regs to
-                                                                                       their Projs to the RegParams node. */
-
-       pset                 *stack_phis;   /**< The set of all Phi nodes inserted due to
-                                                                                       stack pointer modifying nodes. */
+                                                their Projs to the RegParams node. */
 
-       int                  start_block_bias;  /**< The stack bias at the end of the start block. */
+       int                  start_block_bias; /**< The stack bias at the end of the start block. */
 
        void                 *cb;           /**< ABI Callback self pointer. */
 
        pmap                 *keep_map;     /**< mapping blocks to keep nodes. */
        pset                 *ignore_regs;  /**< Additional registers which shall be ignored. */
 
-       arch_irn_handler_t irn_handler;
-       arch_irn_ops_t     irn_ops;
-       DEBUG_ONLY(firm_dbg_module_t    *dbg;)          /**< The debugging module. */
-};
+       ir_node              **calls;       /**< flexible array containing all be_Call nodes */
+
+       arch_register_req_t  sp_req;
+       arch_register_req_t  sp_cls_req;
 
-#define get_abi_from_handler(ptr) firm_container_of(ptr, be_abi_irg_t, irn_handler)
-#define get_abi_from_ops(ptr)     firm_container_of(ptr, be_abi_irg_t, irn_ops)
+       be_stack_layout_t    frame;         /**< The stack frame model. */
+
+       DEBUG_ONLY(firm_dbg_module_t    *dbg;)  /**< The debugging module. */
+};
 
-/* Forward, since be need it in be_abi_introduce(). */
-static const arch_irn_ops_if_t abi_irn_ops;
-static const arch_irn_handler_t abi_irn_handler;
 static heights_t *ir_heights;
 
-/* Flag: if set, try to omit the frame pointer if called by the backend */
+/** Flag: if set, try to omit the frame pointer in all routines. */
 static int be_omit_fp = 1;
 
+/** Flag: if set, try to omit the frame pointer in leaf routines only. */
+static int be_omit_leaf_fp = 1;
+
 /*
      _    ____ ___    ____      _ _ _                _
     / \  | __ )_ _|  / ___|__ _| | | |__   __ _  ___| | _____
@@ -118,6 +137,7 @@ static int be_omit_fp = 1;
 static int cmp_call_arg(const void *a, const void *b, size_t n)
 {
        const be_abi_call_arg_t *p = a, *q = b;
+       (void) n;
        return !(p->is_res == q->is_res && p->pos == q->pos);
 }
 
@@ -160,14 +180,28 @@ static INLINE be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, i
 /* Set the flags for a call. */
 void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
 {
-       call->flags        = flags;
-       call->cb           = cb;
+       call->flags = flags;
+       call->cb    = cb;
+}
+
+void be_abi_call_set_pop(be_abi_call_t *call, int pop)
+{
+       assert(pop >= 0);
+       call->pop = pop;
 }
 
-void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos, unsigned alignment, unsigned space_before, unsigned space_after)
+/* Set register class for call address */
+void be_abi_call_set_call_address_reg_class(be_abi_call_t *call, const arch_register_class_t *cls)
+{
+       call->cls_addr = cls;
+}
+
+
+void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos, ir_mode *load_mode, unsigned alignment, unsigned space_before, unsigned space_after)
 {
        be_abi_call_arg_t *arg = get_or_set_call_arg(call, 0, arg_pos, 1);
        arg->on_stack     = 1;
+       arg->load_mode    = load_mode;
        arg->alignment    = alignment;
        arg->space_before = space_before;
        arg->space_after  = space_after;
@@ -199,14 +233,17 @@ be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
  *
  * @return the new ABI call object
  */
-static be_abi_call_t *be_abi_call_new(void)
+static be_abi_call_t *be_abi_call_new(const arch_register_class_t *cls_addr)
 {
-       be_abi_call_t *call = xmalloc(sizeof(call[0]));
+       be_abi_call_t *call = XMALLOCZ(be_abi_call_t);
+
        call->flags.val  = 0;
        call->params     = new_set(cmp_call_arg, 16);
        call->cb         = NULL;
+       call->cls_addr   = cls_addr;
+
+       call->flags.bits.try_omit_fp = be_omit_fp | be_omit_leaf_fp;
 
-       call->flags.bits.try_omit_fp = be_omit_fp;
        return call;
 }
 
@@ -236,23 +273,22 @@ static void be_abi_call_free(be_abi_call_t *call)
      and the spills.
 */
 
-static int get_stack_entity_offset(be_stack_layout_t *frame, entity *ent, int bias)
+static int get_stack_entity_offset(be_stack_layout_t *frame, ir_entity *ent,
+                                   int bias)
 {
        ir_type *t = get_entity_owner(ent);
-       int ofs    = get_entity_offset_bytes(ent);
+       int ofs    = get_entity_offset(ent);
 
-       int i, index;
+       int index;
 
        /* Find the type the entity is contained in. */
-       for(index = 0; index < N_FRAME_TYPES; ++index) {
-               if(frame->order[index] == t)
+       for (index = 0; index < N_FRAME_TYPES; ++index) {
+               if (frame->order[index] == t)
                        break;
+               /* Add the size of all the types below the one of the entity to the entity's offset */
+               ofs += get_type_size_bytes(frame->order[index]);
        }
 
-       /* Add the size of all the types below the one of the entity to the entity's offset */
-       for(i = 0; i < index; ++i)
-               ofs += get_type_size_bytes(frame->order[i]);
-
        /* correct the offset by the initial position of the frame pointer */
        ofs -= frame->initial_offset;
 
@@ -265,13 +301,13 @@ static int get_stack_entity_offset(be_stack_layout_t *frame, entity *ent, int bi
 /**
  * Retrieve the entity with given offset from a frame type.
  */
-static entity *search_ent_with_offset(ir_type *t, int offset)
+static ir_entity *search_ent_with_offset(ir_type *t, int offset)
 {
        int i, n;
 
        for(i = 0, n = get_compound_n_members(t); i < n; ++i) {
-               entity *ent = get_compound_member(t, i);
-               if(get_entity_offset_bytes(ent) == offset)
+               ir_entity *ent = get_compound_member(t, i);
+               if (get_entity_offset(ent) == offset)
                        return ent;
        }
 
@@ -280,10 +316,11 @@ static entity *search_ent_with_offset(ir_type *t, int offset)
 
 static int stack_frame_compute_initial_offset(be_stack_layout_t *frame)
 {
-       ir_type *base = frame->stack_dir < 0 ? frame->between_type : frame->frame_type;
-       entity *ent   = search_ent_with_offset(base, 0);
-       frame->initial_offset = 0;
-       frame->initial_offset = get_stack_entity_offset(frame, ent, 0);
+       ir_type  *base = frame->stack_dir < 0 ? frame->between_type : frame->frame_type;
+       ir_entity *ent = search_ent_with_offset(base, 0);
+
+       frame->initial_offset = ent ? get_stack_entity_offset(frame, ent, 0) : 0;
+
        return frame->initial_offset;
 }
 
@@ -301,12 +338,13 @@ static int stack_frame_compute_initial_offset(be_stack_layout_t *frame)
  */
 static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
                                            ir_type *between, ir_type *locals, int stack_dir,
-                                           entity *param_map[])
+                                           ir_entity *param_map[])
 {
        frame->arg_type       = args;
        frame->between_type   = between;
        frame->frame_type     = locals;
        frame->initial_offset = 0;
+       frame->initial_bias   = 0;
        frame->stack_dir      = stack_dir;
        frame->order[1]       = between;
        frame->param_map      = param_map;
@@ -334,7 +372,7 @@ static void stack_layout_dump(FILE *file, be_stack_layout_t *frame)
 
                ir_fprintf(file, "type %d: %F size: %d\n", j, t, get_type_size_bytes(t));
                for (i = 0, n = get_compound_n_members(t); i < n; ++i) {
-                       entity *ent = get_compound_member(t, i);
+                       ir_entity *ent = get_compound_member(t, i);
                        ir_fprintf(file, "\t%F int ofs: %d glob ofs: %d\n", ent, get_entity_offset_bytes(ent), get_stack_entity_offset(frame, ent, 0));
                }
        }
@@ -363,122 +401,123 @@ static INLINE int is_on_stack(be_abi_call_t *call, int pos)
 */
 
 /**
- * Transform a call node.
+ * Transform a call node into a be_Call node.
+ *
  * @param env The ABI environment for the current irg.
  * @param irn The call node.
  * @param curr_sp The stack pointer node to use.
  * @return The stack pointer after the call.
  */
-static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, ir_node *alloca_copy)
+static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
 {
-       ir_graph *irg             = env->birg->irg;
-       const arch_isa_t *isa     = env->birg->main_env->arch_env->isa;
-       be_abi_call_t *call       = be_abi_call_new();
-       ir_type *mt               = get_Call_type(irn);
-       ir_node *call_ptr         = get_Call_ptr(irn);
-       int n_params              = get_method_n_params(mt);
-       ir_node *curr_mem         = get_Call_mem(irn);
-       ir_node *bl               = get_nodes_block(irn);
-       pset *results             = pset_new_ptr(8);
-       pset *caller_save         = pset_new_ptr(8);
-       int stack_size            = 0;
-       int stack_dir             = arch_isa_stack_dir(isa);
-       const arch_register_t *sp = arch_isa_sp(isa);
-       ir_mode *mach_mode        = sp->reg_class->mode;
-       struct obstack *obst      = &env->obst;
-       int no_alloc              = call->flags.bits.frame_is_setup_on_call;
-
-       ir_node *res_proj = NULL;
-       int curr_res_proj = pn_Call_max;
-       int n_low_args    = 0;
-       int n_pos         = 0;
+       ir_graph *irg              = env->birg->irg;
+       const arch_env_t *arch_env = env->birg->main_env->arch_env;
+       ir_type *call_tp           = get_Call_type(irn);
+       ir_node *call_ptr          = get_Call_ptr(irn);
+       int n_params               = get_method_n_params(call_tp);
+       ir_node *curr_mem          = get_Call_mem(irn);
+       ir_node *bl                = get_nodes_block(irn);
+       pset *results              = pset_new_ptr(8);
+       pset *caller_save          = pset_new_ptr(8);
+       pset *states               = pset_new_ptr(2);
+       int stack_size             = 0;
+       int stack_dir              = arch_env_stack_dir(arch_env);
+       const arch_register_t *sp  = arch_env_sp(arch_env);
+       be_abi_call_t *call        = be_abi_call_new(sp->reg_class);
+       ir_mode *mach_mode         = sp->reg_class->mode;
+       struct obstack *obst       = &env->obst;
+       int no_alloc               = call->flags.bits.frame_is_setup_on_call;
+       int n_res                  = get_method_n_ress(call_tp);
+       int do_seq                 = call->flags.bits.store_args_sequential && !no_alloc;
+
+       ir_node *res_proj  = NULL;
+       int n_reg_params   = 0;
+       int n_stack_params = 0;
+       int n_ins;
 
        ir_node *low_call;
        ir_node **in;
        ir_node **res_projs;
+       int      n_reg_results = 0;
+       const arch_register_t *reg;
        const ir_edge_t *edge;
-       int *low_args;
-       int *pos;
+       int *reg_param_idxs;
+       int *stack_param_idx;
        int i, n;
 
        /* Let the isa fill out the abi description for that call node. */
-       arch_isa_get_call_abi(isa, mt, call);
+       arch_env_get_call_abi(arch_env, call_tp, call);
 
        /* Insert code to put the stack arguments on the stack. */
        assert(get_Call_n_params(irn) == n_params);
-       for(i = 0; i < n_params; ++i) {
+       for (i = 0; i < n_params; ++i) {
                be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
                assert(arg);
                if (arg->on_stack) {
-                       int arg_size = get_type_size_bytes(get_method_param_type(mt, i));
+                       int arg_size = get_type_size_bytes(get_method_param_type(call_tp, i));
 
                        stack_size += round_up2(arg->space_before, arg->alignment);
                        stack_size += round_up2(arg_size, arg->alignment);
                        stack_size += round_up2(arg->space_after, arg->alignment);
                        obstack_int_grow(obst, i);
-                       n_pos++;
+                       ++n_stack_params;
                }
        }
-       pos = obstack_finish(obst);
+       stack_param_idx = obstack_finish(obst);
 
        /* Collect all arguments which are passed in registers. */
-       for(i = 0, n = get_Call_n_params(irn); i < n; ++i) {
+       for (i = 0; i < n_params; ++i) {
                be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
-               if(arg && arg->in_reg) {
+               if (arg && arg->in_reg) {
                        obstack_int_grow(obst, i);
-                       n_low_args++;
+                       ++n_reg_params;
                }
        }
-       low_args = obstack_finish(obst);
+       reg_param_idxs = obstack_finish(obst);
+
+       /*
+        * If the stack is decreasing and we do not want to store sequentially,
+        * or someone else allocated the call frame
+        * we allocate as much space on the stack all parameters need, by
+        * moving the stack pointer along the stack's direction.
+        *
+        * Note: we also have to do this for stack_size == 0, because we may have
+        * to adjust stack alignment for the call.
+        */
+       if (stack_dir < 0 && !do_seq && !no_alloc) {
+               curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, stack_size, 1);
+       }
 
        /* If there are some parameters which shall be passed on the stack. */
-       if(n_pos > 0) {
+       if (n_stack_params > 0) {
                int curr_ofs      = 0;
-               int do_seq        = call->flags.bits.store_args_sequential && !no_alloc;
 
                /*
                 * Reverse list of stack parameters if call arguments are from left to right.
-                * We must them reverse again in they are pushed (not stored) and the stack
+                * We must them reverse again if they are pushed (not stored) and the stack
                 * direction is downwards.
                 */
                if (call->flags.bits.left_to_right ^ (do_seq && stack_dir < 0)) {
-                       for(i = 0; i < n_pos >> 1; ++i) {
-                               int other  = n_pos - i - 1;
-                               int tmp    = pos[i];
-                               pos[i]     = pos[other];
-                               pos[other] = tmp;
+                       for (i = 0; i < n_stack_params >> 1; ++i) {
+                               int other  = n_stack_params - i - 1;
+                               int tmp    = stack_param_idx[i];
+                               stack_param_idx[i]     = stack_param_idx[other];
+                               stack_param_idx[other] = tmp;
                        }
                }
 
-               /*
-                * If the stack is decreasing and we do not want to store sequentially,
-                * or someone else allocated the call frame
-                * we allocate as much space on the stack all parameters need, by
-                * moving the stack pointer along the stack's direction.
-                */
-               if(stack_dir < 0 && !do_seq && !no_alloc) {
-                       curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, stack_size);
-                       if(alloca_copy) {
-                               add_irn_dep(curr_sp, alloca_copy);
-                               alloca_copy = NULL;
-                       }
+               curr_mem = get_Call_mem(irn);
+               if (! do_seq) {
+                       obstack_ptr_grow(obst, curr_mem);
                }
 
-               if(!do_seq) {
-                       obstack_ptr_grow(obst, get_Call_mem(irn));
-                       curr_mem = new_NoMem();
-               } else {
-                       curr_mem = get_Call_mem(irn);
-               }
-
-               assert(mode_is_reference(mach_mode) && "machine mode must be pointer");
-               for(i = 0; i < n_pos; ++i) {
-                       int p                  = pos[i];
+               for (i = 0; i < n_stack_params; ++i) {
+                       int p                  = stack_param_idx[i];
                        be_abi_call_arg_t *arg = get_call_arg(call, 0, p);
                        ir_node *param         = get_Call_param(irn, p);
                        ir_node *addr          = curr_sp;
                        ir_node *mem           = NULL;
-                       ir_type *param_type    = get_method_param_type(mt, p);
+                       ir_type *param_type    = get_method_param_type(call_tp, p);
                        int param_size         = get_type_size_bytes(param_type) + arg->space_after;
 
                        /*
@@ -488,11 +527,7 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
                         */
                        if (do_seq) {
                                curr_ofs = 0;
-                               addr = curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, param_size + arg->space_before);
-                               if(alloca_copy) {
-                                       add_irn_dep(curr_sp, alloca_copy);
-                                       alloca_copy = NULL;
-                               }
+                               addr = curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, param_size + arg->space_before, 0);
                                add_irn_dep(curr_sp, curr_mem);
                        }
                        else {
@@ -500,8 +535,12 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
                                curr_ofs =  round_up2(curr_ofs, arg->alignment);
 
                                /* Make the expression to compute the argument's offset. */
-                               if(curr_ofs > 0) {
-                                       addr = new_r_Const_long(irg, bl, mode_Is, curr_ofs);
+                               if (curr_ofs > 0) {
+                                       ir_mode *constmode = mach_mode;
+                                       if(mode_is_reference(mach_mode)) {
+                                               constmode = mode_Is;
+                                       }
+                                       addr = new_r_Const_long(irg, bl, constmode, curr_ofs);
                                        addr = new_r_Add(irg, bl, curr_sp, addr, mach_mode);
                                }
                        }
@@ -509,7 +548,8 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
                        /* Insert a store for primitive arguments. */
                        if (is_atomic_type(param_type)) {
                                ir_node *store;
-                               store = new_r_Store(irg, bl, curr_mem, addr, param);
+                               ir_node *mem_input = do_seq ? curr_mem : new_NoMem();
+                               store = new_r_Store(irg, bl, mem_input, addr, param);
                                mem = new_r_Proj(irg, bl, store, mode_M, pn_Store_M);
                        }
 
@@ -533,9 +573,9 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
                in = (ir_node **) obstack_finish(obst);
 
                /* We need the sync only, if we didn't build the stores sequentially. */
-               if(!do_seq) {
-                       if(n_pos >= 1) {
-                               curr_mem = new_r_Sync(irg, bl, n_pos + 1, in);
+               if (! do_seq) {
+                       if (n_stack_params >= 1) {
+                               curr_mem = new_r_Sync(irg, bl, n_stack_params + 1, in);
                        } else {
                                curr_mem = get_Call_mem(irn);
                        }
@@ -544,139 +584,206 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
        }
 
        /* Collect caller save registers */
-       for(i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
-               int j;
-               const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
-               for(j = 0; j < cls->n_regs; ++j) {
+       for (i = 0, n = arch_env_get_n_reg_class(arch_env); i < n; ++i) {
+               unsigned j;
+               const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
+               for (j = 0; j < cls->n_regs; ++j) {
                        const arch_register_t *reg = arch_register_for_index(cls, j);
-                       if(arch_register_type_is(reg, caller_save))
+                       if (arch_register_type_is(reg, caller_save)) {
                                pset_insert_ptr(caller_save, (void *) reg);
+                       }
+                       if (arch_register_type_is(reg, state)) {
+                               pset_insert_ptr(caller_save, (void*) reg);
+                               pset_insert_ptr(states, (void*) reg);
+                       }
                }
        }
 
        /* search the greatest result proj number */
 
-       /* TODO: what if the result is NOT used? Currently there is
-        * no way to detect this later, especially there is no way to
-        * see this in the proj numbers.
-        * While this is ok for the register allocator, it is bad for
-        * backends which need to change the be_Call further (x87 simulator
-        * for instance. However for this particular case the call_type is
-        * sufficient.).
-        */
+       res_projs = alloca(n_res * sizeof(res_projs[0]));
+       memset(res_projs, 0, n_res * sizeof(res_projs[0]));
+
        foreach_out_edge(irn, edge) {
                const ir_edge_t *res_edge;
-               ir_node *irn = get_edge_src_irn(edge);
+               ir_node         *irn = get_edge_src_irn(edge);
 
-               if(is_Proj(irn) && get_Proj_proj(irn) == pn_Call_T_result) {
-                       res_proj = irn;
-                       foreach_out_edge(irn, res_edge) {
-                               int proj;
-                               be_abi_call_arg_t *arg;
-                               ir_node *res = get_edge_src_irn(res_edge);
-
-                               assert(is_Proj(res));
-
-                               proj = get_Proj_proj(res);
-                               arg = get_call_arg(call, 1, proj);
-
-                               /*
-                                       shift the proj number to the right, since we will drop the
-                                       unspeakable Proj_T from the Call. Therefore, all real argument
-                                       Proj numbers must be increased by pn_be_Call_first_res
-                               */
-                               proj += pn_be_Call_first_res;
-                               set_Proj_proj(res, proj);
-                               obstack_ptr_grow(obst, res);
-
-                               if(proj > curr_res_proj)
-                                       curr_res_proj = proj;
-                               if(arg->in_reg) {
-                                       pset_remove_ptr(caller_save, arg->reg);
-                                       //pmap_insert(arg_regs, arg->reg, INT_TO_PTR(proj + 1))
-                               }
-                       }
+               if(!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
+                       continue;
+
+               foreach_out_edge(irn, res_edge) {
+                       int proj;
+                       ir_node *res = get_edge_src_irn(res_edge);
+
+                       assert(is_Proj(res));
+
+                       proj = get_Proj_proj(res);
+                       assert(proj < n_res);
+                       assert(res_projs[proj] == NULL);
+                       res_projs[proj] = res;
                }
+               res_proj = irn;
+               break;
        }
 
-       curr_res_proj++;
-       obstack_ptr_grow(obst, NULL);
-       res_projs = obstack_finish(obst);
+       /** TODO: this is not correct for cases where return values are passed
+        * on the stack, but no known ABI does this currently...
+        */
+       n_reg_results = n_res;
 
        /* make the back end call node and set its register requirements. */
-       for(i = 0; i < n_low_args; ++i)
-               obstack_ptr_grow(obst, get_Call_param(irn, low_args[i]));
+       for (i = 0; i < n_reg_params; ++i) {
+               obstack_ptr_grow(obst, get_Call_param(irn, reg_param_idxs[i]));
+       }
+       foreach_pset(states, reg) {
+               const arch_register_class_t *cls = arch_register_get_class(reg);
+#if 0
+               ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
+               ir_fprintf(stderr, "Adding %+F\n", regnode);
+#endif
+               ir_node *regnode = new_rd_Unknown(irg, arch_register_class_mode(cls));
+               obstack_ptr_grow(obst, regnode);
+       }
+       n_ins = n_reg_params + pset_count(states);
 
        in = obstack_finish(obst);
 
-       if(env->call->flags.bits.call_has_imm && get_irn_opcode(call_ptr) == iro_SymConst) {
-               low_call = be_new_Call(get_irn_dbg_info(irn), irg, bl, curr_mem, curr_sp, curr_sp,
-                                      curr_res_proj + pset_count(caller_save), n_low_args, in,
-                                      get_Call_type(irn));
+       if (env->call->flags.bits.call_has_imm && is_SymConst(call_ptr)) {
+               /* direct call */
+               low_call = be_new_Call(get_irn_dbg_info(irn), irg, bl, curr_mem,
+                                      curr_sp, curr_sp,
+                                      n_reg_results + pn_be_Call_first_res + pset_count(caller_save),
+                                      n_ins, in, get_Call_type(irn));
                be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
+       } else {
+               /* indirect call */
+               low_call = be_new_Call(get_irn_dbg_info(irn), irg, bl, curr_mem,
+                                      curr_sp, call_ptr,
+                                      n_reg_results + pn_be_Call_first_res + pset_count(caller_save),
+                                      n_ins, in, get_Call_type(irn));
        }
+       be_Call_set_pop(low_call, call->pop);
+       ARR_APP1(ir_node *, env->calls, low_call);
 
-       else
-               low_call = be_new_Call(get_irn_dbg_info(irn), irg, bl, curr_mem, curr_sp, call_ptr,
-                                      curr_res_proj + pset_count(caller_save), n_low_args, in,
-                                      get_Call_type(irn));
+       /* create new stack pointer */
+       curr_sp = new_r_Proj(irg, bl, low_call, get_irn_mode(curr_sp),
+                            pn_be_Call_sp);
+       be_set_constr_single_reg(low_call, BE_OUT_POS(pn_be_Call_sp), sp);
+       arch_set_irn_register(arch_env, curr_sp, sp);
+       be_node_set_flags(low_call, BE_OUT_POS(pn_be_Call_sp),
+                       arch_irn_flags_ignore | arch_irn_flags_modify_sp);
+
+       for(i = 0; i < n_res; ++i) {
+               int pn;
+               ir_node           *proj = res_projs[i];
+               be_abi_call_arg_t *arg  = get_call_arg(call, 1, i);
+
+               /* returns values on stack not supported yet */
+               assert(arg->in_reg);
+
+               /*
+                       shift the proj number to the right, since we will drop the
+                       unspeakable Proj_T from the Call. Therefore, all real argument
+                       Proj numbers must be increased by pn_be_Call_first_res
+               */
+               pn = i + pn_be_Call_first_res;
+
+               if(proj == NULL) {
+                       ir_type *res_type = get_method_res_type(call_tp, i);
+                       ir_mode *mode     = get_type_mode(res_type);
+                       proj              = new_r_Proj(irg, bl, low_call, mode, pn);
+                       res_projs[i]      = proj;
+               } else {
+                       set_Proj_pred(proj, low_call);
+                       set_Proj_proj(proj, pn);
+               }
+
+               if (arg->in_reg) {
+                       pset_remove_ptr(caller_save, arg->reg);
+               }
+       }
 
        /*
-               TODO:
-               Set the register class of the call address to the same as the stack pointer's.
-               That' probably buggy for some architectures.
+               Set the register class of the call address to
+               the backend provided class (default: stack pointer class)
        */
-       be_node_set_reg_class(low_call, be_pos_Call_ptr, sp->reg_class);
+       be_node_set_reg_class(low_call, be_pos_Call_ptr, call->cls_addr);
 
        DBG((env->dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));
 
        /* Set the register classes and constraints of the Call parameters. */
-       for(i = 0; i < n_low_args; ++i) {
-               int index = low_args[i];
+       for (i = 0; i < n_reg_params; ++i) {
+               int index = reg_param_idxs[i];
                be_abi_call_arg_t *arg = get_call_arg(call, 0, index);
                assert(arg->reg != NULL);
 
-               be_set_constr_single_reg(low_call, be_pos_Call_first_arg + index, arg->reg);
+               be_set_constr_single_reg(low_call, be_pos_Call_first_arg + i, arg->reg);
        }
 
        /* Set the register constraints of the results. */
-       for(i = 0; res_projs[i]; ++i) {
-               ir_node *irn                 = res_projs[i];
-               int proj                     = get_Proj_proj(irn);
-
-               /* Correct Proj number since it has been adjusted! (see above) */
-               const be_abi_call_arg_t *arg = get_call_arg(call, 1, proj - pn_Call_max);
+       for (i = 0; i < n_res; ++i) {
+               ir_node                 *proj = res_projs[i];
+               const be_abi_call_arg_t *arg  = get_call_arg(call, 1, i);
+               int                      pn   = get_Proj_proj(proj);
 
                assert(arg->in_reg);
-               be_set_constr_single_reg(low_call, BE_OUT_POS(proj), arg->reg);
+               be_set_constr_single_reg(low_call, BE_OUT_POS(pn), arg->reg);
+               arch_set_irn_register(arch_env, proj, arg->reg);
        }
        obstack_free(obst, in);
        exchange(irn, low_call);
 
-       /* redirect the result projs to the lowered call instead of the Proj_T */
-       for(i = 0; res_projs[i]; ++i)
-               set_Proj_pred(res_projs[i], low_call);
+       /* kill the ProjT node */
+       if (res_proj != NULL) {
+               kill_node(res_proj);
+       }
 
        /* Make additional projs for the caller save registers
           and the Keep node which keeps them alive. */
-       if(pset_count(caller_save) > 0) {
+       if (1 || pset_count(caller_save) + n_reg_results > 0) {
                const arch_register_t *reg;
-               ir_node **in, *keep;
-               int i, n;
+               ir_node               **in, *keep;
+               int                   i;
+               int                   n = 0;
+               int                   curr_res_proj
+                       = pn_be_Call_first_res + n_reg_results;
+
+               /* also keep the stack pointer */
+               ++n;
+               set_irn_link(curr_sp, (void*) sp);
+               obstack_ptr_grow(obst, curr_sp);
 
-               for(reg = pset_first(caller_save), n = 0; reg; reg = pset_next(caller_save), ++n) {
-                       ir_node *proj = new_r_Proj(irg, bl, low_call, reg->reg_class->mode, curr_res_proj);
+               for (reg = pset_first(caller_save); reg; reg = pset_next(caller_save), ++n) {
+                       ir_node *proj = new_r_Proj(irg, bl, low_call, reg->reg_class->mode,
+                                                  curr_res_proj);
 
                        /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
                        be_set_constr_single_reg(low_call, BE_OUT_POS(curr_res_proj), reg);
-                       set_irn_link(proj, (void *) reg);
+                       arch_set_irn_register(arch_env, proj, reg);
+
+                       /* a call can produce ignore registers, in this case set the flag and register for the Proj */
+                       if (arch_register_type_is(reg, ignore)) {
+                               be_node_set_flags(low_call, BE_OUT_POS(curr_res_proj),
+                                                 arch_irn_flags_ignore);
+                       }
+
+                       set_irn_link(proj, (void*) reg);
                        obstack_ptr_grow(obst, proj);
                        curr_res_proj++;
                }
 
+               for(i = 0; i < n_reg_results; ++i) {
+                       ir_node *proj = res_projs[i];
+                       const arch_register_t *reg = arch_get_irn_register(arch_env, proj);
+                       set_irn_link(proj, (void*) reg);
+                       obstack_ptr_grow(obst, proj);
+               }
+               n += n_reg_results;
+
+               /* create the Keep for the caller save registers */
                in   = (ir_node **) obstack_finish(obst);
                keep = be_new_Keep(NULL, irg, bl, n, in);
-               for(i = 0; i < n; ++i) {
+               for (i = 0; i < n; ++i) {
                        const arch_register_t *reg = get_irn_link(in[i]);
                        be_node_set_reg_class(keep, i, reg->reg_class);
                }
@@ -684,7 +791,10 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
        }
 
        /* Clean up the stack. */
-       if(stack_size > 0) {
+       assert(stack_size >= call->pop);
+       stack_size -= call->pop;
+
+       if (stack_size > 0) {
                ir_node *mem_proj = NULL;
 
                foreach_out_edge(low_call, edge) {
@@ -695,111 +805,153 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, i
                        }
                }
 
-               if(!mem_proj) {
-                       mem_proj = new_r_Proj(irg, bl, low_call, mode_M, pn_Call_M);
+               if (! mem_proj) {
+                       mem_proj = new_r_Proj(irg, bl, low_call, mode_M, pn_be_Call_M_regular);
                        keep_alive(mem_proj);
                }
-
-                /* Clean up the stack frame if we allocated it */
-               if(!no_alloc) {
-                       curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, -stack_size);
-                       add_irn_dep(curr_sp, mem_proj);
-                       if(alloca_copy) {
-                               add_irn_dep(curr_sp, alloca_copy);
-                               alloca_copy = NULL;
-                       }
-               }
+       }
+       /* Clean up the stack frame or revert alignment fixes if we allocated it */
+       if (! no_alloc) {
+               curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, -stack_size, 0);
        }
 
        be_abi_call_free(call);
-       obstack_free(obst, pos);
+       obstack_free(obst, stack_param_idx);
        del_pset(results);
+       del_pset(states);
        del_pset(caller_save);
 
        return curr_sp;
 }
 
+/**
+ * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
+ *
+ * @param alignment  the minimum stack alignment
+ * @param size       the node containing the non-aligned size
+ * @param irg        the irg where new nodes are allocated on
+ * @param irg        the block where new nodes are allocated on
+ * @param dbg        debug info for new nodes
+ *
+ * @return a node representing the aligned size
+ */
+static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
+                                  ir_graph *irg, ir_node *block, dbg_info *dbg)
+{
+       if (stack_alignment > 1) {
+               ir_mode *mode;
+               tarval  *tv;
+               ir_node *mask;
+
+               assert(is_po2(stack_alignment));
+
+               mode = get_irn_mode(size);
+               tv   = new_tarval_from_long(stack_alignment-1, mode);
+               mask = new_r_Const(irg, block, mode, tv);
+               size = new_rd_Add(dbg, irg, block, size, mask, mode);
+
+               tv   = new_tarval_from_long(-(long)stack_alignment, mode);
+               mask = new_r_Const(irg, block, mode, tv);
+               size = new_rd_And(dbg, irg, block, size, mask, mode);
+       }
+       return size;
+}
 /**
  * Adjust an alloca.
  * The alloca is transformed into a back end alloca node and connected to the stack nodes.
  */
-static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp, ir_node **result_copy)
+static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
 {
-       if (get_Alloc_where(alloc) == stack_alloc) {
-               ir_node *bl        = get_nodes_block(alloc);
-               ir_graph *irg      = get_irn_irg(bl);
-               ir_node *alloc_mem = NULL;
-               ir_node *alloc_res = NULL;
+       ir_node *block;
+       ir_graph *irg;
+       ir_node *alloc_mem;
+       ir_node *alloc_res;
+       ir_type *type;
+       dbg_info *dbg;
 
-               const ir_edge_t *edge;
-               ir_node *new_alloc;
-               ir_node *addr;
-               ir_node *copy;
-               ir_node *ins[2];
+       const ir_edge_t *edge;
+       ir_node *new_alloc, *size, *addr, *ins[2];
+       unsigned stack_alignment;
 
-               foreach_out_edge(alloc, edge) {
-                       ir_node *irn = get_edge_src_irn(edge);
+       assert(get_Alloc_where(alloc) == stack_alloc);
 
-                       assert(is_Proj(irn));
-                       switch(get_Proj_proj(irn)) {
-                       case pn_Alloc_M:
-                               alloc_mem = irn;
-                               break;
-                       case pn_Alloc_res:
-                               alloc_res = irn;
-                               break;
-                       default:
-                               break;
-                       }
-               }
+       block = get_nodes_block(alloc);
+       irg = get_irn_irg(block);
+       alloc_mem = NULL;
+       alloc_res = NULL;
+       type = get_Alloc_type(alloc);
 
-               /* Beware: currently Alloc nodes without a result might happen,
-                  only escape analysis kills them and this phase runs only for object
-                  oriented source. We kill the Alloc here. */
-               if (alloc_res == NULL && alloc_mem) {
-                       exchange(alloc_mem, get_Alloc_mem(alloc));
-                       return curr_sp;
+       foreach_out_edge(alloc, edge) {
+               ir_node *irn = get_edge_src_irn(edge);
+
+               assert(is_Proj(irn));
+               switch (get_Proj_proj(irn)) {
+               case pn_Alloc_M:
+                       alloc_mem = irn;
+                       break;
+               case pn_Alloc_res:
+                       alloc_res = irn;
+                       break;
+               default:
+                       break;
                }
+       }
 
-               /* The stack pointer will be modified in an unknown manner.
-                  We cannot omit it. */
-               env->call->flags.bits.try_omit_fp = 0;
-               new_alloc = be_new_AddSP(env->isa->sp, irg, bl, curr_sp, get_Alloc_size(alloc));
+       /* Beware: currently Alloc nodes without a result might happen,
+          only escape analysis kills them and this phase runs only for object
+          oriented source. We kill the Alloc here. */
+       if (alloc_res == NULL && alloc_mem) {
+               exchange(alloc_mem, get_Alloc_mem(alloc));
+               return curr_sp;
+       }
 
-               if(alloc_mem != NULL) {
-                       ir_node *addsp_mem;
-                       ir_node *sync;
+       dbg = get_irn_dbg_info(alloc);
+
+       /* we might need to multiply the size with the element size */
+       if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
+               tarval *tv    = new_tarval_from_long(get_type_size_bytes(type),
+                                                    mode_Iu);
+               ir_node *cnst = new_rd_Const(dbg, irg, block, mode_Iu, tv);
+               ir_node *mul  = new_rd_Mul(dbg, irg, block, get_Alloc_size(alloc),
+                                          cnst, mode_Iu);
+               size = mul;
+       } else {
+               size = get_Alloc_size(alloc);
+       }
 
-                       addsp_mem = new_r_Proj(irg, bl, new_alloc, mode_M, pn_be_AddSP_M);
+       /* The stack pointer will be modified in an unknown manner.
+          We cannot omit it. */
+       env->call->flags.bits.try_omit_fp = 0;
 
-                       // We need to sync the output mem of the AddSP with the input mem
-                       // edge into the alloc node
-                       ins[0] = get_Alloc_mem(alloc);
-                       ins[1] = addsp_mem;
-                       sync = new_r_Sync(irg, bl, 2, ins);
+       stack_alignment = 1 << env->arch_env->stack_alignment;
+       size            = adjust_alloc_size(stack_alignment, size, irg, block, dbg);
+       new_alloc       = be_new_AddSP(env->arch_env->sp, irg, block, curr_sp, size);
+       set_irn_dbg_info(new_alloc, dbg);
 
-                       exchange(alloc_mem, sync);
-               }
+       if(alloc_mem != NULL) {
+               ir_node *addsp_mem;
+               ir_node *sync;
 
-               exchange(alloc, new_alloc);
+               addsp_mem = new_r_Proj(irg, block, new_alloc, mode_M, pn_be_AddSP_M);
 
-               /* fix projnum of alloca res */
-               set_Proj_proj(alloc_res, pn_be_AddSP_res);
+               /* We need to sync the output mem of the AddSP with the input mem
+                  edge into the alloc node. */
+               ins[0] = get_Alloc_mem(alloc);
+               ins[1] = addsp_mem;
+               sync = new_r_Sync(irg, block, 2, ins);
 
-               addr = env->isa->stack_dir < 0 ? alloc_res : curr_sp;
+               exchange(alloc_mem, sync);
+       }
 
-               /* copy the address away, since it could be used after further stack pointer modifications. */
-               /* Let it point curr_sp just for the moment, I'll reroute it in a second. */
-               *result_copy = copy = be_new_Copy(env->isa->sp->reg_class, irg, bl, curr_sp);
+       exchange(alloc, new_alloc);
 
-               /* Let all users of the Alloc() result now point to the copy. */
-               edges_reroute(alloc_res, copy, irg);
+       /* fix projnum of alloca res */
+       set_Proj_proj(alloc_res, pn_be_AddSP_res);
 
-               /* Rewire the copy appropriately. */
-               set_irn_n(copy, be_pos_Copy_op, addr);
+       addr    = alloc_res;
+       curr_sp = new_r_Proj(irg, block, new_alloc,  get_irn_mode(curr_sp),
+                            pn_be_AddSP_sp);
 
-               curr_sp = alloc_res;
-       }
        return curr_sp;
 }  /* adjust_alloc */
 
@@ -809,22 +961,58 @@ static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp
  */
 static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
 {
-       if (get_Free_where(free) == stack_alloc) {
-               ir_node *bl        = get_nodes_block(free);
-               ir_graph *irg      = get_irn_irg(bl);
-               ir_node *addsp, *mem, *res;
+       ir_node *block;
+       ir_graph *irg;
+       ir_node *subsp, *mem, *res, *size, *sync;
+       ir_type *type;
+       ir_node *in[2];
+       ir_mode *sp_mode;
+       unsigned stack_alignment;
+       dbg_info *dbg;
+
+       assert(get_Free_where(free) == stack_alloc);
+
+       block = get_nodes_block(free);
+       irg = get_irn_irg(block);
+       type = get_Free_type(free);
+       sp_mode = env->arch_env->sp->reg_class->mode;
+       dbg = get_irn_dbg_info(free);
+
+       /* we might need to multiply the size with the element size */
+       if (type != firm_unknown_type && get_type_size_bytes(type) != 1) {
+               tarval *tv = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
+               ir_node *cnst = new_rd_Const(dbg, irg, block, mode_Iu, tv);
+               ir_node *mul = new_rd_Mul(dbg, irg, block, get_Free_size(free),
+                                         cnst, mode_Iu);
+               size = mul;
+       } else {
+               size = get_Free_size(free);
+       }
 
-               /* The stack pointer will be modified in an unknown manner.
-                  We cannot omit it. */
-               env->call->flags.bits.try_omit_fp = 0;
-               addsp = be_new_SubSP(env->isa->sp, irg, bl, curr_sp, get_Free_size(free));
+       stack_alignment = 1 << env->arch_env->stack_alignment;
+       size            = adjust_alloc_size(stack_alignment, size, irg, block, dbg);
 
-               mem = new_r_Proj(irg, bl, addsp, mode_M, pn_be_SubSP_M);
-               res = new_r_Proj(irg, bl, addsp, mode_P_data, pn_be_SubSP_res);
+       /* The stack pointer will be modified in an unknown manner.
+          We cannot omit it. */
+       env->call->flags.bits.try_omit_fp = 0;
+       subsp = be_new_SubSP(env->arch_env->sp, irg, block, curr_sp, size);
+       set_irn_dbg_info(subsp, dbg);
+
+       mem = new_r_Proj(irg, block, subsp, mode_M, pn_be_SubSP_M);
+       res = new_r_Proj(irg, block, subsp, sp_mode, pn_be_SubSP_sp);
+
+       /* we need to sync the memory */
+       in[0] = get_Free_mem(free);
+       in[1] = mem;
+       sync = new_r_Sync(irg, block, 2, in);
+
+       /* and make the AddSP dependent on the former memory */
+       add_irn_dep(subsp, get_Free_mem(free));
+
+       /* kill the free */
+       exchange(free, sync);
+       curr_sp = res;
 
-               exchange(free, mem);
-               curr_sp = res;
-       }
        return curr_sp;
 }  /* adjust_free */
 
@@ -867,15 +1055,12 @@ static int check_dependence(ir_node *curr, ir_node *tgt, ir_node *bl)
  */
 static int dependent_on(ir_node *n1, ir_node *n2)
 {
-       ir_node *bl   = get_nodes_block(n1);
-
-       assert(bl == get_nodes_block(n2));
+       assert(get_nodes_block(n1) == get_nodes_block(n2));
 
        return heights_reachable_in_block(ir_heights, n1, n2);
-       //return check_dependence(n1, n2, bl);
 }
 
-static int cmp_call_dependecy(const void *c1, const void *c2)
+static int cmp_call_dependency(const void *c1, const void *c2)
 {
        ir_node *n1 = *(ir_node **) c1;
        ir_node *n2 = *(ir_node **) c2;
@@ -896,15 +1081,16 @@ static int cmp_call_dependecy(const void *c1, const void *c2)
 }
 
 /**
- * Walker: links all Call/alloc/Free nodes to the Block they are contained.
+ * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
+ * Clears the irg_is_leaf flag if a Call is detected.
  */
-static void link_calls_in_block_walker(ir_node *irn, void *data)
+static void link_ops_in_block_walker(ir_node *irn, void *data)
 {
-       opcode code = get_irn_opcode(irn);
+       ir_opcode code = get_irn_opcode(irn);
 
        if (code == iro_Call ||
-               (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
-               (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
+          (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
+          (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
                be_abi_irg_t *env = data;
                ir_node *bl       = get_nodes_block(irn);
                void *save        = get_irn_link(bl);
@@ -919,58 +1105,69 @@ static void link_calls_in_block_walker(ir_node *irn, void *data)
 
 /**
  * Block-walker:
- * Process all Call nodes inside a basic block.
+ * Process all Call/Alloc/Free nodes inside a basic block.
  * Note that the link field of the block must contain a linked list of all
  * Call nodes inside the Block. We first order this list according to data dependency
  * and that connect the calls together.
  */
-static void process_calls_in_block(ir_node *bl, void *data)
+static void process_ops_in_block(ir_node *bl, void *data)
 {
        be_abi_irg_t *env = data;
        ir_node *curr_sp  = env->init_sp;
        ir_node *irn;
        int n;
 
-       for(irn = get_irn_link(bl), n = 0; irn; irn = get_irn_link(irn), ++n)
+       for (irn = get_irn_link(bl), n = 0; irn; irn = get_irn_link(irn), ++n)
                obstack_ptr_grow(&env->obst, irn);
 
        /* If there were call nodes in the block. */
-       if(n > 0) {
+       if (n > 0) {
                ir_node *keep;
                ir_node **nodes;
-               ir_node *copy = NULL;
                int i;
 
                nodes = obstack_finish(&env->obst);
 
                /* order the call nodes according to data dependency */
-               qsort(nodes, n, sizeof(nodes[0]), cmp_call_dependecy);
+               qsort(nodes, n, sizeof(nodes[0]), cmp_call_dependency);
 
-               for(i = n - 1; i >= 0; --i) {
+               for (i = n - 1; i >= 0; --i) {
                        ir_node *irn = nodes[i];
 
                        DBG((env->dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
-                       switch(get_irn_opcode(irn)) {
+                       switch (get_irn_opcode(irn)) {
                        case iro_Call:
-                               curr_sp = adjust_call(env, irn, curr_sp, copy);
+                               if (! be_omit_fp) {
+                                       /* The stack pointer will be modified due to a call. */
+                                       env->call->flags.bits.try_omit_fp = 0;
+                               }
+                               curr_sp = adjust_call(env, irn, curr_sp);
                                break;
                        case iro_Alloc:
-                               curr_sp = adjust_alloc(env, irn, curr_sp, &copy);
+                               if (get_Alloc_where(irn) == stack_alloc)
+                                       curr_sp = adjust_alloc(env, irn, curr_sp);
                                break;
                        case iro_Free:
-                               curr_sp = adjust_free(env, irn, curr_sp);
+                               if (get_Free_where(irn) == stack_alloc)
+                                       curr_sp = adjust_free(env, irn, curr_sp);
                                break;
                        default:
+                               panic("invalid call");
                                break;
                        }
                }
 
                obstack_free(&env->obst, nodes);
 
-               /* Keep the last stack state in the block by tying it to Keep node */
-               nodes[0] = curr_sp;
-               keep     = be_new_Keep(env->isa->sp->reg_class, get_irn_irg(bl), bl, 1, nodes);
-               pmap_insert(env->keep_map, bl, keep);
+               /* Keep the last stack state in the block by tying it to Keep node,
+                * the proj from calls is already kept */
+               if (curr_sp != env->init_sp &&
+                   !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
+                       nodes[0] = curr_sp;
+                       keep     = be_new_Keep(env->arch_env->sp->reg_class,
+                                              get_irn_irg(bl), bl, 1, nodes);
+                       pmap_insert(env->keep_map, bl, keep);
+               }
        }
 
        set_irn_link(bl, curr_sp);
@@ -984,93 +1181,13 @@ static void process_calls(be_abi_irg_t *env)
        ir_graph *irg = env->birg->irg;
 
        env->call->flags.bits.irg_is_leaf = 1;
-       irg_walk_graph(irg, firm_clear_link, link_calls_in_block_walker, env);
+       irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, env);
 
        ir_heights = heights_new(env->birg->irg);
-       irg_block_walk_graph(irg, NULL, process_calls_in_block, env);
+       irg_block_walk_graph(irg, NULL, process_ops_in_block, env);
        heights_free(ir_heights);
 }
 
-#if 0 /*
-static ir_node *setup_frame(be_abi_irg_t *env)
-{
-       const arch_isa_t *isa = env->birg->main_env->arch_env->isa;
-       const arch_register_t *sp = isa->sp;
-       const arch_register_t *bp = isa->bp;
-       be_abi_call_flags_bits_t flags = env->call->flags.bits;
-       ir_graph *irg      = env->birg->irg;
-       ir_node *bl        = get_irg_start_block(irg);
-       ir_node *no_mem    = get_irg_no_mem(irg);
-       ir_node *old_frame = get_irg_frame(irg);
-       ir_node *stack     = pmap_get(env->regs, (void *) sp);
-       ir_node *frame     = pmap_get(env->regs, (void *) bp);
-
-       int stack_nr       = get_Proj_proj(stack);
-
-       if(flags.try_omit_fp) {
-               stack = be_new_IncSP(sp, irg, bl, stack, no_mem, BE_STACK_FRAME_SIZE_EXPAND);
-               frame = stack;
-       }
-
-       else {
-               frame = be_new_Copy(bp->reg_class, irg, bl, stack);
-
-               be_node_set_flags(frame, -1, arch_irn_flags_dont_spill);
-               if(!flags.fp_free) {
-                       be_set_constr_single_reg(frame, -1, bp);
-                       be_node_set_flags(frame, -1, arch_irn_flags_ignore);
-                       arch_set_irn_register(env->birg->main_env->arch_env, frame, bp);
-               }
-
-               stack = be_new_IncSP(sp, irg, bl, stack, frame, BE_STACK_FRAME_SIZE_EXPAND);
-       }
-
-       be_node_set_flags(env->reg_params, -(stack_nr + 1), arch_irn_flags_ignore);
-       env->init_sp = stack;
-       set_irg_frame(irg, frame);
-       edges_reroute(old_frame, frame, irg);
-
-       return frame;
-}
-
-static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct obstack *obst)
-{
-       const arch_isa_t *isa = env->birg->main_env->arch_env->isa;
-       const arch_register_t *sp = isa->sp;
-       const arch_register_t *bp = isa->bp;
-       ir_graph *irg      = env->birg->irg;
-       ir_node *ret_mem   = get_Return_mem(ret);
-       ir_node *frame     = get_irg_frame(irg);
-       ir_node *bl        = get_nodes_block(ret);
-       ir_node *stack     = get_irn_link(bl);
-
-       pmap_entry *ent;
-
-       if(env->call->flags.bits.try_omit_fp) {
-               stack = be_new_IncSP(sp, irg, bl, stack, ret_mem, -BE_STACK_FRAME_SIZE_SHRINK);
-       }
-
-       else {
-               stack = be_new_SetSP(sp, irg, bl, stack, frame, ret_mem);
-               be_set_constr_single_reg(stack, -1, sp);
-               be_node_set_flags(stack, -1, arch_irn_flags_ignore);
-       }
-
-       pmap_foreach(env->regs, ent) {
-               const arch_register_t *reg = ent->key;
-               ir_node *irn               = ent->value;
-
-               if(reg == sp)
-                       obstack_ptr_grow(&env->obst, stack);
-               else if(reg == bp)
-                       obstack_ptr_grow(&env->obst, frame);
-               else if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
-                       obstack_ptr_grow(obst, irn);
-       }
-}
-*/
-#endif
-
 /**
  * Computes the stack argument layout type.
  * Changes a possibly allocated value param type by moving
@@ -1083,10 +1200,10 @@ static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct
  *
  * @return the stack argument layout type
  */
-static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type *method_type, entity ***param_map)
+static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type *method_type, ir_entity ***param_map)
 {
        int dir  = env->call->flags.bits.left_to_right ? 1 : -1;
-       int inc  = env->birg->main_env->arch_env->isa->stack_dir * dir;
+       int inc  = env->birg->main_env->arch_env->stack_dir * dir;
        int n    = get_method_n_params(method_type);
        int curr = inc > 0 ? 0 : n - 1;
        int ofs  = 0;
@@ -1096,9 +1213,9 @@ static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type
        int i;
        ir_type *val_param_tp = get_method_value_param_type(method_type);
        ident *id = get_entity_ident(get_irg_entity(env->birg->irg));
-       entity **map;
+       ir_entity **map;
 
-       *param_map = map = obstack_alloc(&env->obst, n * sizeof(entity *));
+       *param_map = map = obstack_alloc(&env->obst, n * sizeof(ir_entity *));
        res = new_type_struct(mangle_u(id, new_id_from_chars("arg_type", 8)));
        for (i = 0; i < n; ++i, curr += inc) {
                ir_type *param_type    = get_method_param_type(method_type, curr);
@@ -1121,7 +1238,7 @@ static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type
                        }
                        ofs += arg->space_before;
                        ofs = round_up2(ofs, arg->alignment);
-                       set_entity_offset_bytes(arg->stack_ent, ofs);
+                       set_entity_offset(arg->stack_ent, ofs);
                        ofs += arg->space_after;
                        ofs += get_type_size_bytes(param_type);
                        map[i] = arg->stack_ent;
@@ -1141,13 +1258,13 @@ static void create_register_perms(const arch_isa_t *isa, ir_graph *irg, ir_node
        obstack_init(&obst);
 
        /* Create a Perm after the RegParams node to delimit it. */
-       for(i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
+       for (i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
                const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
                ir_node *perm;
                ir_node **in;
                int n_regs;
 
-               for(n_regs = 0, j = 0; j < cls->n_regs; ++j) {
+               for (n_regs = 0, j = 0; j < cls->n_regs; ++j) {
                        const arch_register_t *reg = &cls->regs[j];
                        ir_node *irn = pmap_get(regs, (void *) reg);
 
@@ -1160,9 +1277,9 @@ static void create_register_perms(const arch_isa_t *isa, ir_graph *irg, ir_node
 
                obstack_ptr_grow(&obst, NULL);
                in = obstack_finish(&obst);
-               if(n_regs > 0) {
+               if (n_regs > 0) {
                        perm = be_new_Perm(cls, irg, bl, n_regs, in);
-                       for(j = 0; j < n_regs; ++j) {
+                       for (j = 0; j < n_regs; ++j) {
                                ir_node *arg = in[j];
                                arch_register_t *reg = get_irn_link(arg);
                                pmap_insert(regs, reg, arg);
@@ -1199,7 +1316,7 @@ static reg_node_map_t *reg_map_to_arr(struct obstack *obst, pmap *reg_map)
        int i = 0;
        reg_node_map_t *res = obstack_alloc(obst, n * sizeof(res[0]));
 
-       pmap_foreach(reg_map, ent) {
+       foreach_pmap(reg_map, ent) {
                res[i].reg = ent->key;
                res[i].irn = ent->value;
                i++;
@@ -1223,10 +1340,10 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
 
        rm = reg_map_to_arr(&env->obst, regs);
 
-       for(n = 0; n < n_regs; ++n)
+       for (n = 0; n < n_regs; ++n)
                obstack_ptr_grow(&env->obst, rm[n].irn);
 
-       if(mem) {
+       if (mem) {
                obstack_ptr_grow(&env->obst, *mem);
                n++;
        }
@@ -1243,17 +1360,17 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
 
                proj = new_r_Proj(irg, bl, irn, get_irn_mode(rm[n].irn), n);
                be_node_set_reg_class(irn, n, reg->reg_class);
-               if(in_req)
+               if (in_req)
                        be_set_constr_single_reg(irn, n, reg);
                be_set_constr_single_reg(irn, pos, reg);
                be_node_set_reg_class(irn, pos, reg->reg_class);
                arch_set_irn_register(env->birg->main_env->arch_env, proj, reg);
 
                /* if the proj projects a ignore register or a node which is set to ignore, propagate this property. */
-               if(arch_register_type_is(reg, ignore) || arch_irn_is(env->birg->main_env->arch_env, in[n], ignore))
+               if (arch_register_type_is(reg, ignore) || arch_irn_is(env->birg->main_env->arch_env, in[n], ignore))
                        flags |= arch_irn_flags_ignore;
 
-               if(arch_irn_is(env->birg->main_env->arch_env, in[n], modify_sp))
+               if (arch_irn_is(env->birg->main_env->arch_env, in[n], modify_sp))
                        flags |= arch_irn_flags_modify_sp;
 
                be_node_set_flags(irn, pos, flags);
@@ -1261,7 +1378,7 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
                pmap_insert(regs, (void *) reg, proj);
        }
 
-       if(mem) {
+       if (mem) {
                *mem = new_r_Proj(irg, bl, irn, mode_M, n);
        }
 
@@ -1278,15 +1395,18 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
  * @param mem     the current memory
  * @param n_res   number of return results
  */
-static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, ir_node *mem, int n_res) {
-       be_abi_call_t *call = env->call;
-       const arch_isa_t *isa = env->birg->main_env->arch_env->isa;
-
+static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl,
+               ir_node *mem, int n_res)
+{
+       be_abi_call_t    *call     = env->call;
+       const arch_env_t *arch_env = env->birg->main_env->arch_env;
+       dbg_info *dbgi;
        pmap *reg_map  = pmap_create();
        ir_node *keep  = pmap_get(env->keep_map, bl);
        int in_max;
        ir_node *ret;
        int i, n;
+       unsigned pop;
        ir_node **in;
        ir_node *stack;
        const arch_register_t **regs;
@@ -1299,17 +1419,15 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i
                it then. Else we use the stack from the start block and let
                the ssa construction fix the usage.
        */
-       stack = be_abi_reg_map_get(env->regs, isa->sp);
+       stack = be_abi_reg_map_get(env->regs, arch_env->sp);
        if (keep) {
-               ir_node *bad = new_r_Bad(env->birg->irg);
                stack = get_irn_n(keep, 0);
-               set_nodes_block(keep, bad);
-               set_irn_n(keep, 0, bad);
-               // exchange(keep, new_r_Bad(env->birg->irg));
+               kill_node(keep);
+               remove_End_keepalive(get_irg_end(env->birg->irg), keep);
        }
 
        /* Insert results for Return into the register map. */
-       for(i = 0; i < n_res; ++i) {
+       for (i = 0; i < n_res; ++i) {
                ir_node *res           = get_Return_res(irn, i);
                be_abi_call_arg_t *arg = get_call_arg(call, 1, i);
                assert(arg->in_reg && "return value must be passed in register");
@@ -1317,13 +1435,13 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i
        }
 
        /* Add uses of the callee save registers. */
-       pmap_foreach(env->regs, ent) {
+       foreach_pmap(env->regs, ent) {
                const arch_register_t *reg = ent->key;
                if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
                        pmap_insert(reg_map, ent->key, ent->value);
        }
 
-       be_abi_reg_map_set(reg_map, isa->sp, stack);
+       be_abi_reg_map_set(reg_map, arch_env->sp, stack);
 
        /* Make the Epilogue node and call the arch's epilogue maker. */
        create_barrier(env, bl, &mem, reg_map, 1);
@@ -1339,14 +1457,14 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i
        regs = obstack_alloc(&env->obst, in_max * sizeof(regs[0]));
 
        in[0]   = mem;
-       in[1]   = be_abi_reg_map_get(reg_map, isa->sp);
+       in[1]   = be_abi_reg_map_get(reg_map, arch_env->sp);
        regs[0] = NULL;
-       regs[1] = isa->sp;
+       regs[1] = arch_env->sp;
        n       = 2;
 
        /* clear SP entry, since it has already been grown. */
-       pmap_insert(reg_map, (void *) isa->sp, NULL);
-       for(i = 0; i < n_res; ++i) {
+       pmap_insert(reg_map, (void *) arch_env->sp, NULL);
+       for (i = 0; i < n_res; ++i) {
                be_abi_call_arg_t *arg = get_call_arg(call, 1, i);
 
                in[n]     = be_abi_reg_map_get(reg_map, arg->reg);
@@ -1357,19 +1475,26 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i
        }
 
        /* grow the rest of the stuff. */
-       pmap_foreach(reg_map, ent) {
-               if(ent->value) {
+       foreach_pmap(reg_map, ent) {
+               if (ent->value) {
                        in[n]     = ent->value;
                        regs[n++] = ent->key;
                }
        }
 
        /* The in array for the new back end return is now ready. */
-       ret = be_new_Return(irn ? get_irn_dbg_info(irn) : NULL, env->birg->irg, bl, n_res, n, in);
+       if (irn != NULL) {
+               dbgi = get_irn_dbg_info(irn);
+       } else {
+               dbgi = NULL;
+       }
+       /* we have to pop the shadow parameter in in case of struct returns */
+       pop = call->pop;
+       ret = be_new_Return(dbgi, env->birg->irg, bl, n_res, pop, n, in);
 
        /* Set the register classes of the return's parameter accordingly. */
-       for(i = 0; i < n; ++i)
-               if(regs[i])
+       for (i = 0; i < n; ++i)
+               if (regs[i])
                        be_node_set_reg_class(ret, i, regs[i]->reg_class);
 
        /* Free the space of the Epilog's in array and the register <-> proj map. */
@@ -1381,15 +1506,16 @@ static ir_node *create_be_return(be_abi_irg_t *env, ir_node *irn, ir_node *bl, i
 
 typedef struct lower_frame_sels_env_t {
        be_abi_irg_t *env;
-       entity       *value_param_list;  /**< the list of all value param entities */
+       ir_entity    *value_param_list;  /**< the list of all value param entities */
+       ir_entity    *value_param_tail;  /**< the tail of the list of all value param entities */
 } lower_frame_sels_env_t;
 
 /**
  * Walker: Replaces Sels of frame type and
  * value param type entities by FrameAddress.
+ * Links all used entities.
  */
-static void lower_frame_sels_walker(ir_node *irn, void *data)
-{
+static void lower_frame_sels_walker(ir_node *irn, void *data) {
        lower_frame_sels_env_t *ctx = data;
 
        if (is_Sel(irn)) {
@@ -1400,17 +1526,20 @@ static void lower_frame_sels_walker(ir_node *irn, void *data)
 
                if (ptr == frame || ptr == param_base) {
                        be_abi_irg_t *env = ctx->env;
-                       entity       *ent = get_Sel_entity(irn);
+                       ir_entity    *ent = get_Sel_entity(irn);
                        ir_node      *bl  = get_nodes_block(irn);
                        ir_node      *nw;
 
-                       nw = be_new_FrameAddr(env->isa->sp->reg_class, irg, bl, frame, ent);
+                       nw = be_new_FrameAddr(env->arch_env->sp->reg_class, irg, bl, frame, ent);
                        exchange(irn, nw);
 
-                       /* check, if it's a param sel and if have not seen this entity immediatly before */
-                       if (ptr == param_base && ctx->value_param_list != ent) {
+                       /* check, if it's a param sel and if have not seen this entity before */
+                       if (ptr == param_base &&
+                                       ent != ctx->value_param_tail &&
+                                       get_entity_link(ent) == NULL) {
                                set_entity_link(ent, ctx->value_param_list);
                                ctx->value_param_list = ent;
+                               if (ctx->value_param_tail == NULL) ctx->value_param_tail = ent;
                        }
                }
        }
@@ -1427,10 +1556,10 @@ static void lower_frame_sels_walker(ir_node *irn, void *data)
  * In the default case we move the entity to the frame type and create
  * a backing store into the first block.
  */
-static void fix_address_of_parameter_access(be_abi_irg_t *env, entity *value_param_list) {
+static void fix_address_of_parameter_access(be_abi_irg_t *env, ir_entity *value_param_list) {
        be_abi_call_t *call = env->call;
        ir_graph *irg       = env->birg->irg;
-       entity *ent, *next_ent, *new_list;
+       ir_entity *ent, *next_ent, *new_list;
        ir_type *frame_tp;
        DEBUG_ONLY(firm_dbg_module_t *dbg = env->dbg;)
 
@@ -1453,7 +1582,7 @@ static void fix_address_of_parameter_access(be_abi_irg_t *env, entity *value_par
                ir_node *frame, *imem, *nmem, *store, *mem, *args, *args_bl;
                const ir_edge_t *edge;
                optimization_state_t state;
-               int offset;
+               unsigned offset;
 
                foreach_block_succ(start_bl, edge) {
                        ir_node *succ = get_edge_src_irn(edge);
@@ -1490,7 +1619,7 @@ static void fix_address_of_parameter_access(be_abi_irg_t *env, entity *value_par
                        ir_node *addr;
 
                        /* address for the backing store */
-                       addr = be_new_FrameAddr(env->isa->sp->reg_class, irg, first_bl, frame, ent);
+                       addr = be_new_FrameAddr(env->arch_env->sp->reg_class, irg, first_bl, frame, ent);
 
                        if (store)
                                mem = new_r_Proj(irg, first_bl, store, mode_M, pn_Store_M);
@@ -1506,64 +1635,122 @@ static void fix_address_of_parameter_access(be_abi_irg_t *env, entity *value_par
                /* move all entities to the frame type */
                frame_tp = get_irg_frame_type(irg);
                offset   = get_type_size_bytes(frame_tp);
+
+               /* we will add new entities: set the layout to undefined */
+               assert(get_type_state(frame_tp) == layout_fixed);
+               set_type_state(frame_tp, layout_undefined);
                for (ent = new_list; ent; ent = get_entity_link(ent)) {
-                       ir_type *tp = get_entity_type(ent);
-                       int align = get_type_alignment_bytes(tp);
+                       ir_type  *tp   = get_entity_type(ent);
+                       unsigned align = get_type_alignment_bytes(tp);
 
                        offset += align - 1;
-                       offset &= -align;
+                       offset &= ~(align - 1);
                        set_entity_owner(ent, frame_tp);
                        add_class_member(frame_tp, ent);
                        /* must be automatic to set a fixed layout */
                        set_entity_allocation(ent, allocation_automatic);
-                       set_entity_offset_bytes(ent, offset);
+                       set_entity_offset(ent, offset);
                        offset += get_type_size_bytes(tp);
                }
                set_type_size_bytes(frame_tp, offset);
+               /* fix the layout again */
+               set_type_state(frame_tp, layout_fixed);
        }
 }
 
+#if 1
+/**
+ * The start block has no jump, instead it has an initial exec Proj.
+ * The backend wants to handle all blocks the same way, so we replace
+ * the out cfg edge with a real jump.
+ */
+static void fix_start_block(ir_node *block, void *env) {
+       int      *done = env;
+       int      i;
+       ir_node  *start_block;
+       ir_graph *irg;
+
+       /* we processed the start block, return */
+       if (*done)
+               return;
+
+       irg         = get_irn_irg(block);
+       start_block = get_irg_start_block(irg);
+
+       for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
+               ir_node *pred       = get_Block_cfgpred(block, i);
+               ir_node *pred_block = get_nodes_block(pred);
+
+               /* ok, we are in the block, having start as cfg predecessor */
+               if (pred_block == start_block) {
+                       ir_node *jump = new_r_Jmp(irg, pred_block);
+                       set_Block_cfgpred(block, i, jump);
+                       *done = 1;
+               }
+       }
+}
+#endif
+
 /**
  * Modify the irg itself and the frame type.
  */
 static void modify_irg(be_abi_irg_t *env)
 {
        be_abi_call_t *call       = env->call;
-       const arch_isa_t *isa     = env->birg->main_env->arch_env->isa;
-       const arch_register_t *sp = arch_isa_sp(isa);
+       const arch_env_t *arch_env= env->birg->main_env->arch_env;
+       const arch_register_t *sp = arch_env_sp(arch_env);
        ir_graph *irg             = env->birg->irg;
-       ir_node *bl               = get_irg_start_block(irg);
-       ir_node *end              = get_irg_end_block(irg);
-       ir_node *mem              = get_irg_initial_mem(irg);
+       ir_node *start_bl;
+       ir_node *end;
+       ir_node *old_mem;
+       ir_node *new_mem_proj;
+       ir_node *mem;
        ir_type *method_type      = get_entity_type(get_irg_entity(irg));
-       pset *dont_save           = pset_new_ptr(8);
 
        int n_params;
-       int i, j, n;
+       int i, n;
+       unsigned j;
 
        reg_node_map_t *rm;
        const arch_register_t *fp_reg;
        ir_node *frame_pointer;
-       ir_node *barrier;
        ir_node *reg_params_bl;
        ir_node **args;
        ir_node *arg_tuple;
+       ir_node *value_param_base;
        const ir_edge_t *edge;
-       ir_type *arg_type, *bet_type;
+       ir_type *arg_type, *bet_type, *tp;
        lower_frame_sels_env_t ctx;
-       entity **param_map;
+       ir_entity **param_map;
 
        bitset_t *used_proj_nr;
        DEBUG_ONLY(firm_dbg_module_t *dbg = env->dbg;)
 
        DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
 
+       /* Must fetch memory here, otherwise the start Barrier gets the wrong
+        * memory, which leads to loops in the DAG. */
+       old_mem = get_irg_initial_mem(irg);
+
+       /* set the links of all frame entities to NULL, we use it
+          to detect if an entity is already linked in the value_param_list */
+       tp = get_method_value_param_type(method_type);
+       if (tp != NULL) {
+               for (i = get_struct_n_members(tp) - 1; i >= 0; --i)
+                       set_entity_link(get_struct_member(tp, i), NULL);
+       }
+
        /* Convert the Sel nodes in the irg to frame load/store/addr nodes. */
        ctx.env              = env;
        ctx.value_param_list = NULL;
+       ctx.value_param_tail = NULL;
        irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
 
-       env->frame = obstack_alloc(&env->obst, sizeof(env->frame[0]));
+       /* value_param_base anchor is not needed anymore now */
+       value_param_base = get_irg_value_param_base(irg);
+       kill_node(value_param_base);
+       set_irg_value_param_base(irg, new_r_Bad(irg));
+
        env->regs  = pmap_create();
 
        used_proj_nr = bitset_alloca(1024);
@@ -1587,19 +1774,21 @@ static void modify_irg(be_abi_irg_t *env)
        arg_tuple = get_irg_args(irg);
        foreach_out_edge(arg_tuple, edge) {
                ir_node *irn = get_edge_src_irn(edge);
-               int nr       = get_Proj_proj(irn);
-               args[nr]     = irn;
-               DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
+               if (! is_Anchor(irn)) {
+                       int nr       = get_Proj_proj(irn);
+                       args[nr]     = irn;
+                       DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
+               }
        }
 
        arg_type = compute_arg_type(env, call, method_type, &param_map);
        bet_type = call->cb->get_between_type(env->cb);
-       stack_frame_init(env->frame, arg_type, bet_type, get_irg_frame_type(irg), isa->stack_dir, param_map);
+       stack_frame_init(&env->frame, arg_type, bet_type, get_irg_frame_type(irg), arch_env->stack_dir, param_map);
 
        /* Count the register params and add them to the number of Projs for the RegParams node */
-       for(i = 0; i < n_params; ++i) {
+       for (i = 0; i < n_params; ++i) {
                be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
-               if(arg->in_reg && args[i]) {
+               if (arg->in_reg && args[i]) {
                        assert(arg->reg != sp && "cannot use stack pointer as parameter register");
                        assert(i == get_Proj_proj(args[i]));
 
@@ -1611,17 +1800,19 @@ static void modify_irg(be_abi_irg_t *env)
        }
 
        /* Collect all callee-save registers */
-       for(i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
-               const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
-               for(j = 0; j < cls->n_regs; ++j) {
+       for (i = 0, n = arch_env_get_n_reg_class(arch_env); i < n; ++i) {
+               const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
+               for (j = 0; j < cls->n_regs; ++j) {
                        const arch_register_t *reg = &cls->regs[j];
-                       if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
+                       if (arch_register_type_is(reg, callee_save) ||
+                                       arch_register_type_is(reg, state)) {
                                pmap_insert(env->regs, (void *) reg, NULL);
+                       }
                }
        }
 
        pmap_insert(env->regs, (void *) sp, NULL);
-       pmap_insert(env->regs, (void *) isa->bp, NULL);
+       pmap_insert(env->regs, (void *) arch_env->bp, NULL);
        reg_params_bl   = get_irg_start_block(irg);
        env->reg_params = be_new_RegParams(irg, reg_params_bl, pmap_count(env->regs));
        add_irn_dep(env->reg_params, get_irg_start(irg));
@@ -1635,10 +1826,9 @@ static void modify_irg(be_abi_irg_t *env)
         */
 
        rm = reg_map_to_arr(&env->obst, env->regs);
-       for(i = 0, n = pmap_count(env->regs); i < n; ++i) {
+       for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
                arch_register_t *reg = (void *) rm[i].reg;
-               ir_node *arg_proj    = rm[i].irn;
-               ir_mode *mode        = arg_proj ? get_irn_mode(arg_proj) : reg->reg_class->mode;
+               ir_mode *mode        = reg->reg_class->mode;
                long nr              = i;
                int pos              = BE_OUT_POS((int) nr);
                int flags            = 0;
@@ -1656,10 +1846,10 @@ static void modify_irg(be_abi_irg_t *env)
                 * If the register is an ignore register,
                 * The Proj for that register shall also be ignored during register allocation.
                 */
-               if(arch_register_type_is(reg, ignore))
+               if (arch_register_type_is(reg, ignore))
                        flags |= arch_irn_flags_ignore;
 
-               if(reg == sp)
+               if (reg == sp)
                        flags |= arch_irn_flags_modify_sp;
 
                be_node_set_flags(env->reg_params, pos, flags);
@@ -1668,113 +1858,356 @@ static void modify_irg(be_abi_irg_t *env)
        }
        obstack_free(&env->obst, rm);
 
+       /* create a new initial memory proj */
+       assert(is_Proj(old_mem));
+       new_mem_proj = new_r_Proj(irg, get_nodes_block(old_mem),
+                                 new_r_Unknown(irg, mode_T), mode_M,
+                                 get_Proj_proj(old_mem));
+       mem = new_mem_proj;
+
        /* Generate the Prologue */
-       fp_reg  = call->cb->prologue(env->cb, &mem, env->regs);
+       fp_reg  = call->cb->prologue(env->cb, &mem, env->regs, &env->frame.initial_bias);
 
        /* do the stack allocation BEFORE the barrier, or spill code
           might be added before it */
-       env->init_sp  = be_abi_reg_map_get(env->regs, sp);
-       env->init_sp = be_new_IncSP(sp, irg, bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND);
+       env->init_sp = be_abi_reg_map_get(env->regs, sp);
+       start_bl     = get_irg_start_block(irg);
+       env->init_sp = be_new_IncSP(sp, irg, start_bl, env->init_sp, BE_STACK_FRAME_SIZE_EXPAND, 0);
        be_abi_reg_map_set(env->regs, sp, env->init_sp);
 
-       env->start_barrier = barrier = create_barrier(env, bl, &mem, env->regs, 0);
+       create_barrier(env, start_bl, &mem, env->regs, 0);
 
-       env->init_sp  = be_abi_reg_map_get(env->regs, sp);
+       env->init_sp = be_abi_reg_map_get(env->regs, sp);
        arch_set_irn_register(env->birg->main_env->arch_env, env->init_sp, sp);
 
        frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
        set_irg_frame(irg, frame_pointer);
        pset_insert_ptr(env->ignore_regs, fp_reg);
 
+       /* rewire old mem users to new mem */
+       set_Proj_pred(new_mem_proj, get_Proj_pred(old_mem));
+       exchange(old_mem, mem);
+
+       set_irg_initial_mem(irg, mem);
+
        /* Now, introduce stack param nodes for all parameters passed on the stack */
-       for(i = 0; i < n_params; ++i) {
+       for (i = 0; i < n_params; ++i) {
                ir_node *arg_proj = args[i];
                ir_node *repl     = NULL;
 
-               if(arg_proj != NULL) {
+               if (arg_proj != NULL) {
                        be_abi_call_arg_t *arg;
                        ir_type *param_type;
-                       int nr = get_Proj_proj(arg_proj);
+                       int     nr = get_Proj_proj(arg_proj);
+                       ir_mode *mode;
 
                        nr         = MIN(nr, n_params);
                        arg        = get_call_arg(call, 0, nr);
                        param_type = get_method_param_type(method_type, nr);
 
-                       if(arg->in_reg) {
+                       if (arg->in_reg) {
                                repl = pmap_get(env->regs, (void *) arg->reg);
-                       }
+                       } else if (arg->on_stack) {
+                               ir_node *addr = be_new_FrameAddr(sp->reg_class, irg, reg_params_bl, frame_pointer, arg->stack_ent);
 
-                       else if(arg->on_stack) {
-                               /* For atomic parameters which are actually used, we create a StackParam node. */
+                               /* For atomic parameters which are actually used, we create a Load node. */
                                if(is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
-                                       ir_mode *mode                    = get_type_mode(param_type);
-                                       const arch_register_class_t *cls = arch_isa_get_reg_class_for_mode(isa, mode);
-                                       repl = be_new_StackParam(cls, isa->bp->reg_class, irg, reg_params_bl, mode, frame_pointer, arg->stack_ent);
-                               }
-
-                               /* The stack parameter is not primitive (it is a struct or array),
-                               we thus will create a node representing the parameter's address
-                               on the stack. */
-                               else {
-                                       repl = be_new_FrameAddr(sp->reg_class, irg, reg_params_bl, frame_pointer, arg->stack_ent);
+                                       ir_mode *mode      = get_type_mode(param_type);
+                                       ir_mode *load_mode = arg->load_mode;
+
+                                       ir_node *load = new_r_Load(irg, reg_params_bl, new_NoMem(), addr, load_mode);
+                                       set_irn_pinned(load, op_pin_state_floats);
+                                       repl = new_r_Proj(irg, reg_params_bl, load, load_mode, pn_Load_res);
+
+                                       if (mode != load_mode) {
+                                               repl = new_r_Conv(irg, reg_params_bl, repl, mode);
+                                       }
+                               } else {
+                                       /* The stack parameter is not primitive (it is a struct or array),
+                                        * we thus will create a node representing the parameter's address
+                                        * on the stack. */
+                                       repl = addr;
                                }
                        }
 
                        assert(repl != NULL);
-                       edges_reroute(args[i], repl, irg);
+
+                       /* Beware: the mode of the register parameters is always the mode of the register class
+                          which may be wrong. Add Conv's then. */
+                       mode = get_irn_mode(args[i]);
+                       if (mode != get_irn_mode(repl)) {
+                               repl = new_r_Conv(irg, get_irn_n(repl, -1), repl, mode);
+                       }
+                       exchange(args[i], repl);
                }
        }
 
+       /* the arg proj is not needed anymore now and should be only used by the anchor */
+       assert(get_irn_n_edges(arg_tuple) == 1);
+       kill_node(arg_tuple);
+       set_irg_args(irg, new_rd_Bad(irg));
+
        /* All Return nodes hang on the End node, so look for them there. */
+       end = get_irg_end_block(irg);
        for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
                ir_node *irn = get_Block_cfgpred(end, i);
 
                if (is_Return(irn)) {
-                       ir_node *ret = create_be_return(env, irn, get_nodes_block(irn), get_Return_mem(irn), get_Return_n_ress(irn));
+                       ir_node *blk = get_nodes_block(irn);
+                       ir_node *mem = get_Return_mem(irn);
+                       ir_node *ret = create_be_return(env, irn, blk, mem, get_Return_n_ress(irn));
                        exchange(irn, ret);
                }
        }
-       /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return than,
+       /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
           the code is dead and will never be executed. */
 
-       del_pset(dont_save);
        obstack_free(&env->obst, args);
+
+       /* handle start block here (place a jump in the block) */
+       i = 0;
+       irg_block_walk_graph(irg, fix_start_block, NULL, &i);
+}
+
+/** Fix the state inputs of calls that still hang on unknowns */
+static
+void fix_call_state_inputs(be_abi_irg_t *env)
+{
+       const arch_env_t *arch_env = env->arch_env;
+       int i, n, n_states;
+       arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
+
+       /* Collect caller save registers */
+       n = arch_env_get_n_reg_class(arch_env);
+       for (i = 0; i < n; ++i) {
+               unsigned j;
+               const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
+               for (j = 0; j < cls->n_regs; ++j) {
+                       const arch_register_t *reg = arch_register_for_index(cls, j);
+                       if (arch_register_type_is(reg, state)) {
+                               ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
+                       }
+               }
+       }
+
+       n = ARR_LEN(env->calls);
+       n_states = ARR_LEN(stateregs);
+       for (i = 0; i < n; ++i) {
+               int s, arity;
+               ir_node *call = env->calls[i];
+
+               arity = get_irn_arity(call);
+
+               /* the state reg inputs are the last n inputs of the calls */
+               for (s = 0; s < n_states; ++s) {
+                       int inp = arity - n_states + s;
+                       const arch_register_t *reg = stateregs[s];
+                       ir_node *regnode = be_abi_reg_map_get(env->regs, reg);
+
+                       set_irn_n(call, inp, regnode);
+               }
+       }
+
+       DEL_ARR_F(stateregs);
+}
+
+/**
+ * Create a trampoline entity for the given method.
+ */
+static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
+{
+       ir_type   *type   = get_entity_type(method);
+       ident     *old_id = get_entity_ld_ident(method);
+       ident     *id     = mangle3("L", old_id, "$stub");
+       ir_type   *parent = be->pic_trampolines_type;
+       ir_entity *ent    = new_entity(parent, old_id, type);
+       set_entity_ld_ident(ent, id);
+       set_entity_visibility(ent, visibility_local);
+       set_entity_variability(ent, variability_uninitialized);
+
+       return ent;
+}
+
+/**
+ * Returns the trampoline entity for the given method.
+ */
+static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
+{
+       ir_entity *result = pmap_get(env->ent_trampoline_map, method);
+       if (result == NULL) {
+               result = create_trampoline(env, method);
+               pmap_insert(env->ent_trampoline_map, method, result);
+       }
+
+       return result;
+}
+
+static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
+{
+       ident     *old_id = get_entity_ld_ident(entity);
+       ident     *id     = mangle3("L", old_id, "$non_lazy_ptr");
+       ir_type   *e_type = get_entity_type(entity);
+       ir_type   *type   = new_type_pointer(id, e_type, mode_P_data);
+       ir_type   *parent = be->pic_symbols_type;
+       ir_entity *ent    = new_entity(parent, old_id, type);
+       set_entity_ld_ident(ent, id);
+       set_entity_visibility(ent, visibility_local);
+       set_entity_variability(ent, variability_uninitialized);
+
+       return ent;
+}
+
+static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
+{
+       ir_entity *result = pmap_get(env->ent_pic_symbol_map, entity);
+       if (result == NULL) {
+               result = create_pic_symbol(env, entity);
+               pmap_insert(env->ent_pic_symbol_map, entity, result);
+       }
+
+       return result;
+}
+
+
+
+/**
+ * Returns non-zero if a given entity can be accessed using a relative address.
+ */
+static int can_address_relative(ir_entity *entity)
+{
+       return get_entity_variability(entity) == variability_initialized
+               || get_entity_visibility(entity) == visibility_local;
+}
+
+/** patches SymConsts to work in position independent code */
+static void fix_pic_symconsts(ir_node *node, void *data)
+{
+       ir_graph     *irg;
+       ir_node      *pic_base;
+       ir_node      *add;
+       ir_node      *block;
+       ir_node      *unknown;
+       ir_mode      *mode;
+       ir_node      *load;
+       ir_node      *load_res;
+       be_abi_irg_t *env = data;
+       int           arity, i;
+       be_main_env_t *be = env->birg->main_env;
+
+       arity = get_irn_arity(node);
+       for (i = 0; i < arity; ++i) {
+               dbg_info  *dbgi;
+               ir_node   *pred = get_irn_n(node, i);
+               ir_entity *entity;
+               ir_entity *pic_symbol;
+               ir_node   *pic_symconst;
+
+               if (!is_SymConst(pred))
+                       continue;
+
+               entity = get_SymConst_entity(pred);
+               block  = get_nodes_block(pred);
+               irg    = get_irn_irg(pred);
+
+               /* calls can jump to relative addresses, so we can directly jump to
+                  the (relatively) known call address or the trampoline */
+               if (is_Call(node) && i == 1) {
+                       ir_entity *trampoline;
+                       ir_node   *trampoline_const;
+
+                       if (can_address_relative(entity))
+                               continue;
+
+                       dbgi             = get_irn_dbg_info(pred);
+                       trampoline       = get_trampoline(be, entity);
+                       trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
+                                                                   trampoline, NULL);
+                       set_irn_n(node, i, trampoline_const);
+                       continue;
+               }
+
+               /* everything else is accessed relative to EIP */
+               mode     = get_irn_mode(pred);
+               unknown  = new_r_Unknown(irg, mode);
+               pic_base = arch_code_generator_get_pic_base(env->birg->cg);
+
+               /* all ok now for locally constructed stuff */
+               if (can_address_relative(entity)) {
+                       ir_node *add = new_r_Add(irg, block, pic_base, pred, mode);
+
+                       /* make sure the walker doesn't visit this add again */
+                       mark_irn_visited(add);
+                       set_irn_n(node, i, add);
+                       continue;
+               }
+
+               /* get entry from pic symbol segment */
+               dbgi         = get_irn_dbg_info(pred);
+               pic_symbol   = get_pic_symbol(be, entity);
+               pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
+                                                       pic_symbol, NULL);
+               add = new_r_Add(irg, block, pic_base, pic_symconst, mode);
+               mark_irn_visited(add);
+
+               /* we need an extra indirection for global data outside our current
+                  module. The loads are always safe and can therefore float
+                  and need no memory input */
+               load     = new_r_Load(irg, block, new_NoMem(), add, mode);
+               load_res = new_r_Proj(irg, block, load, mode, pn_Load_res);
+               set_irn_pinned(load, op_pin_state_floats);
+
+               set_irn_n(node, i, load_res);
+       }
 }
 
 be_abi_irg_t *be_abi_introduce(be_irg_t *birg)
 {
-       be_abi_irg_t *env  = xmalloc(sizeof(env[0]));
+       be_abi_irg_t *env  = XMALLOC(be_abi_irg_t);
        ir_node *old_frame = get_irg_frame(birg->irg);
        ir_graph *irg      = birg->irg;
 
        pmap_entry *ent;
        ir_node *dummy;
        optimization_state_t state;
+       unsigned *limited_bitset;
 
-       be_omit_fp = birg->main_env->options->omit_fp;
+       be_omit_fp      = birg->main_env->options->omit_fp;
+       be_omit_leaf_fp = birg->main_env->options->omit_leaf_fp;
 
        obstack_init(&env->obst);
 
-       env->isa           = birg->main_env->arch_env->isa;
-       env->method_type   = get_entity_type(get_irg_entity(irg));
-       env->call          = be_abi_call_new();
-       arch_isa_get_call_abi(env->isa, env->method_type, env->call);
+       env->arch_env    = birg->main_env->arch_env;
+       env->method_type = get_entity_type(get_irg_entity(irg));
+       env->call        = be_abi_call_new(env->arch_env->sp->reg_class);
+       arch_env_get_call_abi(env->arch_env, env->method_type, env->call);
+
+       env->ignore_regs  = pset_new_ptr_default();
+       env->keep_map     = pmap_create();
+       env->dce_survivor = new_survive_dce();
+       env->birg         = birg;
+
+       env->sp_req.type    = arch_register_req_type_limited;
+       env->sp_req.cls     = arch_register_get_class(env->arch_env->sp);
+       limited_bitset      = rbitset_obstack_alloc(&env->obst, env->sp_req.cls->n_regs);
+       rbitset_set(limited_bitset, arch_register_get_index(env->arch_env->sp));
+       env->sp_req.limited = limited_bitset;
+
+       env->sp_cls_req.type  = arch_register_req_type_normal;
+       env->sp_cls_req.cls   = arch_register_get_class(env->arch_env->sp);
 
-       env->ignore_regs      = pset_new_ptr_default();
-       env->keep_map         = pmap_create();
-       env->dce_survivor     = new_survive_dce();
-       env->birg             = birg;
-       env->stack_phis       = pset_new_ptr(16);
        /* Beware: later we replace this node by the real one, ensure it is not CSE'd
           to another Unknown or the stack pointer gets used */
        save_optimization_state(&state);
        set_optimize(0);
-       env->init_sp = dummy  = new_r_Unknown(irg, env->isa->sp->reg_class->mode);
+       env->init_sp = dummy  = new_r_Unknown(irg, env->arch_env->sp->reg_class->mode);
        restore_optimization_state(&state);
        FIRM_DBG_REGISTER(env->dbg, "firm.be.abi");
 
-       memcpy(&env->irn_handler, &abi_irn_handler, sizeof(abi_irn_handler));
-       env->irn_ops.impl = &abi_irn_ops;
+       env->calls = NEW_ARR_F(ir_node*, 0);
+
+       if (birg->main_env->options->pic) {
+               irg_walk_graph(irg, fix_pic_symconsts, NULL, env);
+       }
 
        /* Lower all call nodes in the IRG. */
        process_calls(env);
@@ -1788,19 +2221,26 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg)
        /* Process the IRG */
        modify_irg(env);
 
+       /* fix call inputs for state registers */
+       fix_call_state_inputs(env);
+
        /* We don't need the keep map anymore. */
        pmap_destroy(env->keep_map);
+       env->keep_map = NULL;
+
+       /* calls array is not needed anymore */
+       DEL_ARR_F(env->calls);
+       env->calls = NULL;
 
        /* reroute the stack origin of the calls to the true stack origin. */
-       edges_reroute(dummy, env->init_sp, irg);
-       edges_reroute(old_frame, get_irg_frame(irg), irg);
+       exchange(dummy, env->init_sp);
+       exchange(old_frame, get_irg_frame(irg));
 
        /* Make some important node pointers survive the dead node elimination. */
        survive_dce_register_irn(env->dce_survivor, &env->init_sp);
-       pmap_foreach(env->regs, ent)
+       foreach_pmap(env->regs, ent) {
                survive_dce_register_irn(env->dce_survivor, (ir_node **) &ent->value);
-
-       arch_env_push_irn_handler(env->birg->main_env->arch_env, &env->irn_handler);
+       }
 
        env->call->cb->done(env->cb);
        env->cb = NULL;
@@ -1809,12 +2249,11 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg)
 
 void be_abi_free(be_abi_irg_t *env)
 {
+       be_abi_call_free(env->call);
        free_survive_dce(env->dce_survivor);
-       del_pset(env->stack_phis);
        del_pset(env->ignore_regs);
        pmap_destroy(env->regs);
        obstack_free(&env->obst, NULL);
-       arch_env_pop_irn_handler(env->birg->main_env->arch_env);
        free(env);
 }
 
@@ -1827,9 +2266,30 @@ void be_abi_put_ignore_regs(be_abi_irg_t *abi, const arch_register_class_t *cls,
                        bitset_set(bs, reg->index);
 }
 
+void be_abi_set_non_ignore_regs(be_abi_irg_t *abi, const arch_register_class_t *cls, unsigned *raw_bitset)
+{
+       unsigned         i;
+       arch_register_t *reg;
+
+       for (i = 0; i < cls->n_regs; ++i) {
+               if (arch_register_type_is(&cls->regs[i], ignore))
+                       continue;
+
+               rbitset_set(raw_bitset, i);
+       }
+
+       for (reg = pset_first(abi->ignore_regs); reg != NULL;
+            reg = pset_next(abi->ignore_regs)) {
+               if (reg->reg_class != cls)
+                       continue;
+
+               rbitset_clear(raw_bitset, reg->index);
+       }
+}
+
 /* Returns the stack layout from a abi environment. */
 const be_stack_layout_t *be_abi_get_stack_layout(const be_abi_irg_t *abi) {
-       return abi->frame;
+       return &abi->frame;
 }
 
 /*
@@ -1842,90 +2302,157 @@ const be_stack_layout_t *be_abi_get_stack_layout(const be_abi_irg_t *abi) {
 
 */
 
-struct fix_stack_walker_info {
-       nodeset *nodes;
-       const arch_env_t *aenv;
-};
+typedef ir_node **node_array;
+
+typedef struct fix_stack_walker_env_t {
+       node_array sp_nodes;
+       const arch_env_t *arch_env;
+} fix_stack_walker_env_t;
 
 /**
  * Walker. Collect all stack modifying nodes.
  */
-static void collect_stack_nodes_walker(ir_node *irn, void *data)
+static void collect_stack_nodes_walker(ir_node *node, void *data)
 {
-       struct fix_stack_walker_info *info = data;
+       fix_stack_walker_env_t *env = data;
 
-       if (is_Block(irn))
-               return;
-
-       if (arch_irn_is(info->aenv, irn, modify_sp)) {
-               assert(get_irn_mode(irn) != mode_M && get_irn_mode(irn) != mode_T);
-               pset_insert_ptr(info->nodes, irn);
+       if (arch_irn_is(env->arch_env, node, modify_sp)) {
+               assert(get_irn_mode(node) != mode_M && get_irn_mode(node) != mode_T);
+               ARR_APP1(ir_node*, env->sp_nodes, node);
        }
 }
 
-void be_abi_fix_stack_nodes(be_abi_irg_t *env, be_lv_t *lv)
+void be_abi_fix_stack_nodes(be_abi_irg_t *env)
 {
-       dom_front_info_t *df;
-       pset *stack_nodes = pset_new_ptr(16);
-       struct fix_stack_walker_info info;
-
-       info.nodes = stack_nodes;
-       info.aenv  = env->birg->main_env->arch_env;
-
-       /* We need dominance frontiers for fix up */
-       df = be_compute_dominance_frontiers(env->birg->irg);
-       irg_walk_graph(env->birg->irg, collect_stack_nodes_walker, NULL, &info);
-       pset_insert_ptr(stack_nodes, env->init_sp);
-       be_ssa_constr_set_phis(df, lv, stack_nodes, env->stack_phis);
-       del_pset(stack_nodes);
-
-       /* free these dominance frontiers */
-       be_free_dominance_frontiers(df);
+       be_ssa_construction_env_t senv;
+       int i, len;
+       ir_node **phis;
+       be_irg_t *birg = env->birg;
+       be_lv_t *lv = be_get_birg_liveness(birg);
+       fix_stack_walker_env_t walker_env;
+
+       walker_env.sp_nodes = NEW_ARR_F(ir_node*, 0);
+       walker_env.arch_env = birg->main_env->arch_env;
+
+       irg_walk_graph(birg->irg, collect_stack_nodes_walker, NULL, &walker_env);
+
+       /* nothing to be done if we didn't find any node, in fact we mustn't
+        * continue, as for endless loops incsp might have had no users and is bad
+        * now.
+        */
+       len = ARR_LEN(walker_env.sp_nodes);
+       if (len == 0) {
+               DEL_ARR_F(walker_env.sp_nodes);
+               return;
+       }
+
+       be_ssa_construction_init(&senv, birg);
+       be_ssa_construction_add_copies(&senv, walker_env.sp_nodes,
+                                   ARR_LEN(walker_env.sp_nodes));
+       be_ssa_construction_fix_users_array(&senv, walker_env.sp_nodes,
+                                     ARR_LEN(walker_env.sp_nodes));
+
+       if(lv != NULL) {
+               len = ARR_LEN(walker_env.sp_nodes);
+               for(i = 0; i < len; ++i) {
+                       be_liveness_update(lv, walker_env.sp_nodes[i]);
+               }
+               be_ssa_construction_update_liveness_phis(&senv, lv);
+       }
+
+       phis = be_ssa_construction_get_new_phis(&senv);
+
+       /* set register requirements for stack phis */
+       len = ARR_LEN(phis);
+       for(i = 0; i < len; ++i) {
+               ir_node *phi = phis[i];
+               be_set_phi_reg_req(walker_env.arch_env, phi, &env->sp_req);
+               be_set_phi_flags(walker_env.arch_env, phi, arch_irn_flags_ignore | arch_irn_flags_modify_sp);
+               arch_set_irn_register(walker_env.arch_env, phi, env->arch_env->sp);
+       }
+       be_ssa_construction_destroy(&senv);
+
+       DEL_ARR_F(walker_env.sp_nodes);
 }
 
-static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int bias)
+/**
+ * Fix all stack accessing operations in the block bl.
+ *
+ * @param env        the abi environment
+ * @param bl         the block to process
+ * @param real_bias  the bias value
+ *
+ * @return the bias at the end of this block
+ */
+static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int real_bias)
 {
-       const arch_env_t *arch_env = env->birg->main_env->arch_env;
-       int omit_fp            = env->call->flags.bits.try_omit_fp;
-       ir_node *irn;
+       int               omit_fp  = env->call->flags.bits.try_omit_fp;
+       ir_node          *irn;
+       int               wanted_bias = real_bias;
 
        sched_foreach(bl, irn) {
+               int ofs;
 
                /*
                   Check, if the node relates to an entity on the stack frame.
                   If so, set the true offset (including the bias) for that
                   node.
                 */
-               entity *ent = arch_get_frame_entity(arch_env, irn);
-               if(ent) {
-                       int offset = get_stack_entity_offset(env->frame, ent, bias);
-                       arch_set_frame_offset(arch_env, irn, offset);
-                       DBG((env->dbg, LEVEL_2, "%F has offset %d (including bias %d)\n", ent, offset, bias));
+               ir_entity *ent = arch_get_frame_entity(irn);
+               if (ent) {
+                       int bias   = omit_fp ? real_bias : 0;
+                       int offset = get_stack_entity_offset(&env->frame, ent, bias);
+                       arch_set_frame_offset(irn, offset);
+                       DBG((env->dbg, LEVEL_2, "%F has offset %d (including bias %d)\n",
+                            ent, offset, bias));
                }
 
                /*
-                  If the node modifies the stack pointer by a constant offset,
-                  record that in the bias.
+                * If the node modifies the stack pointer by a constant offset,
+                * record that in the bias.
                 */
-               if(arch_irn_is(arch_env, irn, modify_sp)) {
-                       int ofs = arch_get_sp_bias(arch_env, irn);
-
-                       if(be_is_IncSP(irn)) {
-                               if(ofs == BE_STACK_FRAME_SIZE_EXPAND) {
-                                       ofs = get_type_size_bytes(get_irg_frame_type(env->birg->irg));
-                                       be_set_IncSP_offset(irn, ofs);
-                               } else if(ofs == BE_STACK_FRAME_SIZE_SHRINK) {
-                                       ofs = - get_type_size_bytes(get_irg_frame_type(env->birg->irg));
-                                       be_set_IncSP_offset(irn, ofs);
+               ofs = arch_get_sp_bias(irn);
+
+               if (be_is_IncSP(irn)) {
+                       /* fill in real stack frame size */
+                       if (ofs == BE_STACK_FRAME_SIZE_EXPAND) {
+                               ir_type *frame_type = get_irg_frame_type(env->birg->irg);
+                               ofs = (int) get_type_size_bytes(frame_type);
+                               be_set_IncSP_offset(irn, ofs);
+                       } else if (ofs == BE_STACK_FRAME_SIZE_SHRINK) {
+                               ir_type *frame_type = get_irg_frame_type(env->birg->irg);
+                               ofs = - (int)get_type_size_bytes(frame_type);
+                               be_set_IncSP_offset(irn, ofs);
+                       } else {
+                               if (be_get_IncSP_align(irn)) {
+                                       /* patch IncSP to produce an aligned stack pointer */
+                                       ir_type *between_type = env->frame.between_type;
+                                       int      between_size = get_type_size_bytes(between_type);
+                                       int      alignment    = 1 << env->arch_env->stack_alignment;
+                                       int      delta        = (real_bias + ofs + between_size) & (alignment - 1);
+                                       assert(ofs >= 0);
+                                       if (delta > 0) {
+                                               be_set_IncSP_offset(irn, ofs + alignment - delta);
+                                               real_bias += alignment - delta;
+                                       }
+                               } else {
+                                       /* adjust so real_bias corresponds with wanted_bias */
+                                       int delta = wanted_bias - real_bias;
+                                       assert(delta <= 0);
+                                       if (delta != 0) {
+                                               be_set_IncSP_offset(irn, ofs + delta);
+                                               real_bias += delta;
+                                       }
                                }
                        }
-
-                       if(omit_fp)
-                               bias += ofs;
                }
+
+               real_bias   += ofs;
+               wanted_bias += ofs;
        }
 
-       return bias;
+       assert(real_bias == wanted_bias);
+       return real_bias;
 }
 
 /**
@@ -1933,12 +2460,14 @@ static int process_stack_bias(be_abi_irg_t *env, ir_node *bl, int bias)
  */
 struct bias_walk {
        be_abi_irg_t *env;     /**< The ABI irg environment. */
-       int start_block_bias;  /**< The bias at the end of the start block. */
-       ir_node *start_block;  /**< The start block of the current graph. */
+       int           start_block_bias;  /**< The bias at the end of the start block. */
+       int           between_size;
+       ir_node      *start_block;  /**< The start block of the current graph. */
 };
 
 /**
- * Block-Walker: fix all stack offsets
+ * Block-Walker: fix all stack offsets for all blocks
+ * except the start block
  */
 static void stack_bias_walker(ir_node *bl, void *data)
 {
@@ -1950,14 +2479,15 @@ static void stack_bias_walker(ir_node *bl, void *data)
 
 void be_abi_fix_stack_bias(be_abi_irg_t *env)
 {
-       ir_graph *irg  = env->birg->irg;
-       struct bias_walk bw;
+       ir_graph          *irg   = env->birg->irg;
+       struct bias_walk  bw;
 
-       stack_frame_compute_initial_offset(env->frame);
-       // stack_layout_dump(stdout, env->frame);
+       stack_frame_compute_initial_offset(&env->frame);
+       // stack_layout_dump(stdout, frame);
 
        /* Determine the stack bias at the end of the start block. */
-       bw.start_block_bias = process_stack_bias(env, get_irg_start_block(irg), 0);
+       bw.start_block_bias = process_stack_bias(env, get_irg_start_block(irg), env->frame.initial_bias);
+       bw.between_size     = get_type_size_bytes(env->frame.between_type);
 
        /* fix the bias is all other blocks */
        bw.env = env;
@@ -1979,124 +2509,6 @@ ir_node *be_abi_get_ignore_irn(be_abi_irg_t *abi, const arch_register_t *reg)
        return pmap_get(abi->regs, (void *) reg);
 }
 
-ir_node *be_abi_get_start_barrier(be_abi_irg_t *abi)
-{
-       return abi->start_barrier;
-}
-
-/*
-  _____ _____  _   _   _    _                 _ _
- |_   _|  __ \| \ | | | |  | |               | | |
-   | | | |__) |  \| | | |__| | __ _ _ __   __| | | ___ _ __
-   | | |  _  /| . ` | |  __  |/ _` | '_ \ / _` | |/ _ \ '__|
-  _| |_| | \ \| |\  | | |  | | (_| | | | | (_| | |  __/ |
- |_____|_|  \_\_| \_| |_|  |_|\__,_|_| |_|\__,_|_|\___|_|
-
-  for Phi nodes which are created due to stack modifying nodes
-  such as IncSP, AddSP and SetSP.
-
-  These Phis are always to be ignored by the reg alloc and are
-  fixed on the SP register of the ISA.
-*/
-
-static const void *abi_get_irn_ops(const arch_irn_handler_t *handler, const ir_node *irn)
-{
-       const be_abi_irg_t *abi = get_abi_from_handler(handler);
-       const void *res = NULL;
-
-       if(is_Phi(irn) && pset_find_ptr(abi->stack_phis, (void *) irn))
-               res = &abi->irn_ops;
-
-       return res;
-}
-
-static void be_abi_limited(void *data, bitset_t *bs)
-{
-       be_abi_irg_t *abi = data;
-       bitset_clear_all(bs);
-       bitset_set(bs, abi->isa->sp->index);
-}
-
-static const arch_register_req_t *abi_get_irn_reg_req(const void *self, arch_register_req_t *req, const ir_node *irn, int pos)
-{
-       be_abi_irg_t *abi          = get_abi_from_ops(self);
-       const arch_register_t *reg = abi->isa->sp;
-
-       memset(req, 0, sizeof(req[0]));
-
-       if(pos == BE_OUT_POS(0)) {
-               req->cls         = reg->reg_class;
-               req->type        = arch_register_req_type_limited;
-               req->limited     = be_abi_limited;
-               req->limited_env = abi;
-       }
-
-       else if(pos >= 0 && pos < get_irn_arity(irn)) {
-               req->cls  = reg->reg_class;
-               req->type = arch_register_req_type_normal;
-       }
-
-       return req;
-}
-
-static void abi_set_irn_reg(const void *self, ir_node *irn, const arch_register_t *reg)
-{
-}
-
-static const arch_register_t *abi_get_irn_reg(const void *self, const ir_node *irn)
-{
-       const be_abi_irg_t *abi = get_abi_from_ops(self);
-       return abi->isa->sp;
-}
-
-static arch_irn_class_t abi_classify(const void *_self, const ir_node *irn)
-{
-       return arch_irn_class_normal;
-}
-
-static arch_irn_flags_t abi_get_flags(const void *_self, const ir_node *irn)
-{
-       return arch_irn_flags_ignore | arch_irn_flags_modify_sp;
-}
-
-static entity *abi_get_frame_entity(const void *_self, const ir_node *irn)
-{
-       return NULL;
-}
-
-static void abi_set_frame_entity(const void *_self, ir_node *irn, entity *ent)
-{
-}
-
-static void abi_set_frame_offset(const void *_self, ir_node *irn, int bias)
-{
-}
-
-static int abi_get_sp_bias(const void *self, const ir_node *irn)
-{
-       return 0;
-}
-
-static const arch_irn_ops_if_t abi_irn_ops = {
-       abi_get_irn_reg_req,
-       abi_set_irn_reg,
-       abi_get_irn_reg,
-       abi_classify,
-       abi_get_flags,
-       abi_get_frame_entity,
-       abi_set_frame_entity,
-       abi_set_frame_offset,
-       abi_get_sp_bias,
-       NULL,    /* get_inverse             */
-       NULL,    /* get_op_estimated_cost   */
-       NULL,    /* possible_memory_operand */
-       NULL,    /* perform_memory_operand  */
-};
-
-static const arch_irn_handler_t abi_irn_handler = {
-       abi_get_irn_ops
-};
-
 /**
  * Returns non-zero if the ABI has omitted the frame pointer in
  * the current graph.