- Split bearch.h correctly into bearch.h and bearch_t.h
[libfirm] / ir / be / beabi.c
index 5fb5897..0f62ac7 100644 (file)
@@ -2,9 +2,9 @@
  * ABI lowering.
  *
  * @author Sebastian Hack
- * @date 7.3.2005
+ * @date   7.3.2005
+ * @cvsid  $Id$
  */
-
 #ifdef HAVE_CONFIG_H
 # include "config.h"
 #endif
 #include "irgwalk.h"
 #include "irprintf_t.h"
 #include "irgopt.h"
+#include "irbitset.h"
+#include "height.h"
+#include "pdeq.h"
+#include "irtools.h"
+#include "raw_bitset.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"
-
-#define MAX(x, y) ((x) > (y) ? (x) : (y))
-#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#include "beirg_t.h"
+#include "bessaconstr.h"
 
 typedef struct _be_abi_call_arg_t {
-       unsigned is_res   : 1;
-       unsigned in_reg   : 1;
-       unsigned on_stack : 1;
+       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;
        const arch_register_t *reg;
-       entity *stack_ent;
+       ir_entity *stack_ent;
        unsigned alignment;
        unsigned space_before;
        unsigned space_after;
 } be_abi_call_arg_t;
 
 struct _be_abi_call_t {
-       be_abi_call_flags_t flags;
-       const be_abi_callbacks_t *cb;
-       type *between_type;
-       set *params;
-};
-
-#define N_FRAME_TYPES 3
-
-typedef struct _be_stack_frame_t {
-       type *arg_type;
-       type *between_type;
-       type *frame_type;
-
-       type *order[N_FRAME_TYPES];        /**< arg, between and frame types ordered. */
-
-       int initial_offset;
-       int stack_dir;
-} be_stack_frame_t;
-
-struct _be_stack_slot_t {
-       struct _be_stack_frame_t *frame;
-       entity *ent;
+       be_abi_call_flags_t         flags;
+       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_frame_t     *frame;        /**< The stack frame model. */
-       const be_irg_t       *birg;         /**< The back end IRG. */
+       be_stack_layout_t    *frame;        /**< The stack frame model. */
+       be_irg_t             *birg;         /**< The back end IRG. */
        const arch_isa_t     *isa;          /**< The isa. */
        survive_dce_t        *dce_survivor;
 
        be_abi_call_t        *call;         /**< The ABI call information. */
-       type                 *method_type;  /**< The type of the method of the IRG. */
+       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 */
+
        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. */
-
        int                  start_block_bias;  /**< The stack bias at the end of the start block. */
 
        void                 *cb;           /**< ABI Callback self pointer. */
 
-       arch_irn_handler_t irn_handler;
-       arch_irn_ops_t     irn_ops;
+       pmap                 *keep_map;     /**< mapping blocks to keep nodes. */
+       pset                 *ignore_regs;  /**< Additional registers which shall be ignored. */
+
+       ir_node             **calls;        /**< flexible array containing all be_Call nodes */
+
+       arch_register_req_t sp_req;
+       arch_register_req_t sp_cls_req;
+
        DEBUG_ONLY(firm_dbg_module_t    *dbg;)          /**< The debugging module. */
 };
 
-#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)
+static heights_t *ir_heights;
 
-/* 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;
-
-/* Flag: if set, try to omit the frame pointer if alled by the backend */
-int be_omit_fp = 0;
+/* Flag: if set, try to omit the frame pointer if called by the backend */
+static int be_omit_fp = 1;
 
 /*
      _    ____ ___    ____      _ _ _                _
@@ -170,10 +157,18 @@ 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;
+}
+
+
+/* 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, unsigned alignment, unsigned space_before, unsigned space_after)
 {
        be_abi_call_arg_t *arg = get_or_set_call_arg(call, 0, arg_pos, 1);
@@ -209,14 +204,17 @@ 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()
+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]));
+
        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;
+
        return call;
 }
 
@@ -246,10 +244,10 @@ static void be_abi_call_free(be_abi_call_t *call)
      and the spills.
 */
 
-static int get_stack_entity_offset(be_stack_frame_t *frame, entity *ent, int bias)
+static int get_stack_entity_offset(be_stack_layout_t *frame, ir_entity *ent, int bias)
 {
-       type *t = get_entity_owner(ent);
-       int ofs = get_entity_offset_bytes(ent);
+       ir_type *t = get_entity_owner(ent);
+       int ofs    = get_entity_offset(ent);
 
        int i, index;
 
@@ -275,29 +273,44 @@ static int get_stack_entity_offset(be_stack_frame_t *frame, entity *ent, int bia
 /**
  * Retrieve the entity with given offset from a frame type.
  */
-static entity *search_ent_with_offset(type *t, int offset)
+static ir_entity *search_ent_with_offset(ir_type *t, int offset)
 {
        int i, n;
 
-       for(i = 0, n = get_class_n_members(t); i < n; ++i) {
-               entity *ent = get_class_member(t, i);
-               if(get_entity_offset_bytes(ent) == offset)
+       for(i = 0, n = get_compound_n_members(t); i < n; ++i) {
+               ir_entity *ent = get_compound_member(t, i);
+               if(get_entity_offset(ent) == offset)
                        return ent;
        }
 
        return NULL;
 }
 
-static int stack_frame_compute_initial_offset(be_stack_frame_t *frame)
+static int stack_frame_compute_initial_offset(be_stack_layout_t *frame)
 {
-       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;
 }
 
-static be_stack_frame_t *stack_frame_init(be_stack_frame_t *frame, type *args, type *between, type *locals, int stack_dir)
+/**
+ * Initializes the frame layout from parts
+ *
+ * @param frame     the stack layout that will be initialized
+ * @param args      the stack argument layout type
+ * @param between   the between layout type
+ * @param locals    the method frame type
+ * @param stack_dir the stack direction
+ * @param param_map an array mapping method argument positions to the stack argument type
+ *
+ * @return the initialized stack layout
+ */
+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,
+                                           ir_entity *param_map[])
 {
        frame->arg_type       = args;
        frame->between_type   = between;
@@ -305,68 +318,37 @@ static be_stack_frame_t *stack_frame_init(be_stack_frame_t *frame, type *args, t
        frame->initial_offset = 0;
        frame->stack_dir      = stack_dir;
        frame->order[1]       = between;
+       frame->param_map      = param_map;
 
        if(stack_dir > 0) {
                frame->order[0] = args;
                frame->order[2] = locals;
        }
-
        else {
                frame->order[0] = locals;
                frame->order[2] = args;
        }
-
        return frame;
 }
 
-static void stack_frame_dump(FILE *file, be_stack_frame_t *frame)
+#if 0
+/** Dumps the stack layout to file. */
+static void stack_layout_dump(FILE *file, be_stack_layout_t *frame)
 {
        int i, j, n;
 
        ir_fprintf(file, "initial offset: %d\n", frame->initial_offset);
-       for(j = 0; j < N_FRAME_TYPES; ++j) {
-               type *t = frame->order[j];
+       for (j = 0; j < N_FRAME_TYPES; ++j) {
+               ir_type *t = frame->order[j];
 
-               ir_fprintf(file, "type %d: %Fm size: %d\n", j, t, get_type_size_bytes(t));
-               for(i = 0, n = get_class_n_members(t); i < n; ++i) {
-                       entity *ent = get_class_member(t, i);
+               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) {
+                       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));
                }
        }
 }
-
-/**
- * If irn is a Sel node computes the address of an entity
- * on the frame type return the entity, else NULL.
- */
-static INLINE entity *get_sel_ent(ir_node *irn)
-{
-       if(is_Sel(irn) && get_Sel_ptr(irn) == get_irg_frame(get_irn_irg(irn))) {
-               return get_Sel_entity(irn);
-       }
-
-       return NULL;
-}
-
-/**
- * Walker: Replaces Loads, Stores and Sels of frame type entities
- * by FrameLoad, FrameStore and FrameAdress.
- */
-static void lower_frame_sels_walker(ir_node *irn, void *data)
-{
-       ir_node *nw  = NULL;
-       entity *ent = get_sel_ent(irn);
-
-       if(ent != NULL) {
-               be_abi_irg_t *env = data;
-               ir_node *bl       = get_nodes_block(irn);
-               ir_graph *irg     = get_irn_irg(bl);
-               ir_node *frame    = get_irg_frame(irg);
-
-               nw = be_new_FrameAddr(env->isa->sp->reg_class, irg, bl, frame, ent);
-               exchange(irn, nw);
-       }
-}
+#endif
 
 /**
  * Returns non-zero if the call argument at given position
@@ -396,11 +378,11 @@ static INLINE int is_on_stack(be_abi_call_t *call, int pos)
  * @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)
+static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp, ir_node *alloca_copy)
 {
        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();
+       const arch_env_t *arch_env = env->birg->main_env->arch_env;
+       const arch_isa_t *isa     = arch_env->isa;
        ir_type *mt               = get_Call_type(irn);
        ir_node *call_ptr         = get_Call_ptr(irn);
        int n_params              = get_method_n_params(mt);
@@ -408,22 +390,25 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
        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_isa_stack_dir(isa);
        const arch_register_t *sp = arch_isa_sp(isa);
+       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;
-       ir_node *no_mem           = get_irg_no_mem(irg);
        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;
+       int count;
 
        ir_node *low_call;
        ir_node **in;
        ir_node **res_projs;
+       const arch_register_t *reg;
        const ir_edge_t *edge;
        int *low_args;
        int *pos;
@@ -437,11 +422,12 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
        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) {
-                       stack_size += arg->space_before;
-                       stack_size =  round_up2(stack_size, arg->alignment);
-                       stack_size += get_type_size_bytes(get_method_param_type(mt, i));
-                       stack_size += arg->space_after;
+               if (arg->on_stack) {
+                       int arg_size = get_type_size_bytes(get_method_param_type(mt, 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++;
                }
@@ -465,11 +451,11 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
 
                /*
                 * 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) {
+                       for (i = 0; i < n_pos >> 1; ++i) {
                                int other  = n_pos - i - 1;
                                int tmp    = pos[i];
                                pos[i]     = pos[other];
@@ -484,62 +470,92 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
                 * 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, no_mem, stack_size, be_stack_dir_expand);
+                       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;
+                       }
+               }
+
+               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];
                        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;
-                       type *param_type       = get_method_param_type(mt, p);
+                       ir_type *param_type    = get_method_param_type(mt, p);
                        int param_size         = get_type_size_bytes(param_type) + arg->space_after;
 
-                       curr_ofs += arg->space_before;
-                       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);
-                               addr = new_r_Add(irg, bl, curr_sp, addr, mach_mode);
+                       /*
+                        * If we wanted to build the arguments sequentially,
+                        * the stack pointer for the next must be incremented,
+                        * and the memory value propagated.
+                        */
+                       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;
+                               }
+                               add_irn_dep(curr_sp, curr_mem);
+                       }
+                       else {
+                               curr_ofs += arg->space_before;
+                               curr_ofs =  round_up2(curr_ofs, arg->alignment);
+
+                               /* Make the expression to compute the argument's offset. */
+                               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);
+                               }
                        }
 
                        /* Insert a store for primitive arguments. */
-                       if(is_atomic_type(param_type)) {
-                               mem = new_r_Store(irg, bl, curr_mem, addr, param);
-                               mem = new_r_Proj(irg, bl, mem, mode_M, pn_Store_M);
+                       if (is_atomic_type(param_type)) {
+                               ir_node *store;
+                               store = new_r_Store(irg, bl, curr_mem, addr, param);
+                               mem = new_r_Proj(irg, bl, store, mode_M, pn_Store_M);
                        }
 
                        /* Make a mem copy for compound arguments. */
                        else {
+                               ir_node *copy;
+
                                assert(mode_is_reference(get_irn_mode(param)));
-                               mem = new_r_CopyB(irg, bl, curr_mem, addr, param, param_type);
-                               mem = new_r_Proj(irg, bl, mem, mode_M, pn_CopyB_M_regular);
+                               copy = new_r_CopyB(irg, bl, curr_mem, addr, param, param_type);
+                               mem = new_r_Proj(irg, bl, copy, mode_M, pn_CopyB_M_regular);
                        }
 
-                       obstack_ptr_grow(obst, mem);
-
                        curr_ofs += param_size;
 
-                       /*
-                        * If we wanted to build the arguments sequentially,
-                        * the stack pointer for the next must be incremented,
-                        * and the memory value propagated.
-                        */
-                       if(do_seq) {
-                               curr_ofs = 0;
-                               curr_sp  = be_new_IncSP(sp, irg, bl, curr_sp, curr_mem, param_size, be_stack_dir_expand);
+                       if (do_seq)
                                curr_mem = mem;
-                       }
+                       else
+                               obstack_ptr_grow(obst, mem);
                }
 
                in = (ir_node **) obstack_finish(obst);
 
                /* We need the sync only, if we didn't build the stores sequentially. */
-               if(!do_seq)
-                       curr_mem = new_r_Sync(irg, bl, n_pos, in);
+               if(!do_seq) {
+                       if(n_pos >= 1) {
+                               curr_mem = new_r_Sync(irg, bl, n_pos + 1, in);
+                       } else {
+                               curr_mem = get_Call_mem(irn);
+                       }
+               }
                obstack_free(obst, in);
        }
 
@@ -549,8 +565,13 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
                const arch_register_class_t *cls = arch_isa_get_reg_class(isa, 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);
+                       }
                }
        }
 
@@ -604,29 +625,43 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
        res_projs = obstack_finish(obst);
 
        /* make the back end call node and set its register requirements. */
-       for(i = 0; i < n_low_args; ++i)
+       for(i = 0; i < n_low_args; ++i) {
                obstack_ptr_grow(obst, get_Call_param(irn, low_args[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);
+       }
+       count = n_low_args + 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));
+               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), count,
+                                      in, get_Call_type(irn));
                be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
+       } 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),
+                                      count, in, get_Call_type(irn));
        }
-
-  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));
+       ARR_APP1(ir_node*, env->calls, low_call);
 
        /*
-               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) {
@@ -638,43 +673,66 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
        }
 
        /* 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);
+       for (i = 0; res_projs[i]; ++i) {
+               int pn = get_Proj_proj(res_projs[i]);
 
                /* 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);
+               const be_abi_call_arg_t *arg = get_call_arg(call, 1, pn - pn_Call_max);
+
+               /* Matze: we need the information about the real mode for later
+                * transforms (signed/unsigend compares, stores...), so leave the fixup
+                * for the backend transform phase... */
+#if 0
+               /* correct mode */
+               const arch_register_class_t *cls = arch_register_get_class(arg->reg);
+               ir_mode *mode = arch_register_class_mode(cls);
+               set_irn_mode(irn, mode);
+#endif
 
                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, res_projs[i], 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)
+       for (i = 0; res_projs[i]; ++i)
                set_Proj_pred(res_projs[i], low_call);
 
+       /* set the now unnecessary projT to bad */
+       if(res_proj != NULL) {
+               be_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 (pset_count(caller_save) > 0) {
                const arch_register_t *reg;
-               ir_node **in, *keep;
-               int i, n;
+               ir_node               **in, *keep;
+               int                   i, n;
 
-               for(reg = pset_first(caller_save), n = 0; reg; reg = pset_next(caller_save), ++n) {
+               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);
 
                        /* 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);
+
+                       /* a call can produce ignore registers, in this case set the flag and register for the Proj */
+                       if (arch_register_type_is(reg, ignore)) {
+                               arch_set_irn_register(env->birg->main_env->arch_env, proj, reg);
+                               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++;
                }
 
+               /* 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);
                }
@@ -693,17 +751,26 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
                        }
                }
 
-               if(!mem_proj)
+               if(!mem_proj) {
                        mem_proj = new_r_Proj(irg, bl, low_call, mode_M, pn_Call_M);
+                       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, mem_proj, stack_size, be_stack_dir_shrink);
+               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;
+                       }
+               }
        }
 
        be_abi_call_free(call);
        obstack_free(obst, pos);
        del_pset(results);
+       del_pset(states);
        del_pset(caller_save);
 
        return curr_sp;
@@ -713,84 +780,195 @@ static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
  * 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)
+static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp, ir_node **result_copy)
 {
-       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;
 
-               const ir_edge_t *edge;
-               ir_node *new_alloc;
+       const ir_edge_t *edge;
+       ir_node *new_alloc;
+       ir_node *size;
+       ir_node *addr;
+       ir_node *copy;
+       ir_node *ins[2];
+
+       if (get_Alloc_where(alloc) != stack_alloc) {
+               assert(0);
+               return alloc;
+       }
 
-               foreach_out_edge(alloc, edge) {
-                       ir_node *irn = get_edge_src_irn(edge);
+       block = get_nodes_block(alloc);
+       irg = get_irn_irg(block);
+       alloc_mem = NULL;
+       alloc_res = NULL;
+       type = get_Alloc_type(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;
-                       }
-               }
+       foreach_out_edge(alloc, edge) {
+               ir_node *irn = get_edge_src_irn(edge);
 
-               /* 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) {
-                       exchange(alloc_mem, get_Alloc_mem(alloc));
-                       return curr_sp;
+               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;
+       }
 
-               exchange(alloc_res, env->isa->stack_dir < 0 ? new_alloc : curr_sp);
+       /* we might need to multiply the size with the element size */
+       if(type != get_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(NULL, irg, block, mode_Iu, tv);
+               ir_node *mul = new_rd_Mul(NULL, irg, block, get_Alloc_size(alloc),
+                                         cnst, mode_Iu);
+               size = mul;
+       } else {
+               size = get_Alloc_size(alloc);
+       }
 
-               if(alloc_mem != NULL)
-                       exchange(alloc_mem, new_r_NoMem(irg));
+       /* 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, block, curr_sp, size);
 
-               curr_sp = new_alloc;
+       if(alloc_mem != NULL) {
+               ir_node *addsp_mem;
+               ir_node *sync;
+
+               addsp_mem = new_r_Proj(irg, block, new_alloc, mode_M, pn_be_AddSP_M);
+
+               // 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);
+
+               exchange(alloc_mem, sync);
        }
 
+       exchange(alloc, new_alloc);
+
+       /* fix projnum of alloca res */
+       set_Proj_proj(alloc_res, pn_be_AddSP_res);
+
+       addr = env->isa->stack_dir < 0 ? alloc_res : curr_sp;
+
+       /* 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, block, curr_sp);
+
+       /* Let all users of the Alloc() result now point to the copy. */
+       edges_reroute(alloc_res, copy, irg);
+
+       /* Rewire the copy appropriately. */
+       set_irn_n(copy, be_pos_Copy_op, addr);
+
+       curr_sp = alloc_res;
+
        return curr_sp;
-}
+}  /* adjust_alloc */
+
+/**
+ * Adjust a Free.
+ * The Free is transformed into a back end free node and connected to the stack nodes.
+ */
+static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
+{
+       ir_node *block;
+       ir_graph *irg;
+       ir_node *subsp, *mem, *res, *size, *sync;
+       ir_type *type;
+       ir_node *in[2];
+       ir_mode *sp_mode;
+
+       if (get_Free_where(free) != stack_alloc) {
+               assert(0);
+               return free;
+       }
+
+       block = get_nodes_block(free);
+       irg = get_irn_irg(block);
+       type = get_Free_type(free);
+       sp_mode = env->isa->sp->reg_class->mode;
+
+       /* we might need to multiply the size with the element size */
+       if(type != get_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(NULL, irg, block, mode_Iu, tv);
+               ir_node *mul = new_rd_Mul(NULL, 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;
+       subsp = be_new_SubSP(env->isa->sp, irg, block, curr_sp, size);
+
+       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_res);
+
+       /* 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;
+
+       return curr_sp;
+}  /* adjust_free */
 
+/* the following function is replaced by the usage of the heights module */
+#if 0
 /**
  * Walker for dependent_on().
  * This function searches a node tgt recursively from a given node
  * but is restricted to the given block.
  * @return 1 if tgt was reachable from curr, 0 if not.
  */
-static int check_dependence(ir_node *curr, ir_node *tgt, ir_node *bl, unsigned long visited_nr)
+static int check_dependence(ir_node *curr, ir_node *tgt, ir_node *bl)
 {
        int n, i;
 
-       if(get_irn_visited(curr) >= visited_nr)
+       if (get_nodes_block(curr) != bl)
                return 0;
 
-       set_irn_visited(curr, visited_nr);
-       if(get_nodes_block(curr) != bl)
-               return 0;
-
-       if(curr == tgt)
+       if (curr == tgt)
                return 1;
 
-       for(i = 0, n = get_irn_arity(curr); i < n; ++i) {
-               if(check_dependence(get_irn_n(curr, i), tgt, bl, visited_nr))
-                       return 1;
+       /* Phi functions stop the recursion inside a basic block */
+       if (! is_Phi(curr)) {
+               for(i = 0, n = get_irn_arity(curr); i < n; ++i) {
+                       if (check_dependence(get_irn_n(curr, i), tgt, bl))
+                               return 1;
+               }
        }
 
        return 0;
 }
+#endif /* if 0 */
 
 /**
  * Check if a node is somehow data dependent on another one.
@@ -801,13 +979,9 @@ static int check_dependence(ir_node *curr, ir_node *tgt, ir_node *bl, unsigned l
  */
 static int dependent_on(ir_node *n1, ir_node *n2)
 {
-       ir_node *bl   = get_nodes_block(n1);
-       ir_graph *irg = get_irn_irg(bl);
-       long vis_nr   = get_irg_visited(irg) + 1;
+       assert(get_nodes_block(n1) == get_nodes_block(n2));
 
-       assert(bl == get_nodes_block(n2));
-       set_irg_visited(irg, vis_nr);
-       return check_dependence(n1, n2, bl, vis_nr);
+       return heights_reachable_in_block(ir_heights, n1, n2);
 }
 
 static int cmp_call_dependecy(const void *c1, const void *c2)
@@ -821,17 +995,31 @@ static int cmp_call_dependecy(const void *c1, const void *c2)
                1  if second is "smaller" that first
                -1 if first is "smaller" that second
        */
-       return n1 == n2 ? 0 : (dependent_on(n1, n2) ? -1 : 1);
+       if (dependent_on(n1, n2))
+               return -1;
+
+       if (dependent_on(n2, n1))
+               return 1;
+
+       return 0;
 }
 
+/**
+ * Walker: links all Call/alloc/Free nodes to the Block they are contained.
+ */
 static void link_calls_in_block_walker(ir_node *irn, void *data)
 {
-       if(is_Call(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)) {
                be_abi_irg_t *env = data;
                ir_node *bl       = get_nodes_block(irn);
                void *save        = get_irn_link(bl);
 
-               env->call->flags.bits.irg_is_leaf = 0;
+               if (code == iro_Call)
+                       env->call->flags.bits.irg_is_leaf = 0;
 
                set_irn_link(irn, save);
                set_irn_link(bl, irn);
@@ -839,9 +1027,10 @@ static void link_calls_in_block_walker(ir_node *irn, void *data)
 }
 
 /**
- * Process all call nodes inside a basic block.
+ * Block-walker:
+ * Process all Call 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
+ * 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)
@@ -856,7 +1045,9 @@ static void process_calls_in_block(ir_node *bl, void *data)
 
        /* If there were call nodes in the block. */
        if(n > 0) {
+               ir_node *keep;
                ir_node **nodes;
+               ir_node *copy = NULL;
                int i;
 
                nodes = obstack_finish(&env->obst);
@@ -867,12 +1058,16 @@ static void process_calls_in_block(ir_node *bl, void *data)
                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)) {
                        case iro_Call:
-                               curr_sp = adjust_call(env, irn, curr_sp);
+                               curr_sp = adjust_call(env, irn, curr_sp, copy);
                                break;
                        case iro_Alloc:
-                               curr_sp = adjust_alloc(env, irn, curr_sp);
+                               curr_sp = adjust_alloc(env, irn, curr_sp, &copy);
+                               break;
+                       case iro_Free:
+                               curr_sp = adjust_free(env, irn, curr_sp);
                                break;
                        default:
                                break;
@@ -883,11 +1078,12 @@ static void process_calls_in_block(ir_node *bl, void *data)
 
                /* Keep the last stack state in the block by tying it to Keep node */
                nodes[0] = curr_sp;
-               be_new_Keep(env->isa->sp->reg_class, get_irn_irg(bl), bl, 1, nodes);
+               keep     = be_new_Keep(env->isa->sp->reg_class, get_irn_irg(bl), bl, 1, nodes);
+               pmap_insert(env->keep_map, bl, keep);
        }
 
        set_irn_link(bl, curr_sp);
-}
+}  /* process_calls_in_block */
 
 /**
  * Adjust all call nodes in the graph to the ABI conventions.
@@ -898,15 +1094,10 @@ static void process_calls(be_abi_irg_t *env)
 
        env->call->flags.bits.irg_is_leaf = 1;
        irg_walk_graph(irg, firm_clear_link, link_calls_in_block_walker, env);
-       irg_block_walk_graph(irg, NULL, process_calls_in_block, env);
-}
 
-static void collect_return_walker(ir_node *irn, void *data)
-{
-       if(get_irn_opcode(irn) == iro_Return) {
-               struct obstack *obst = data;
-               obstack_ptr_grow(obst, irn);
-       }
+       ir_heights = heights_new(env->birg->irg);
+       irg_block_walk_graph(irg, NULL, process_calls_in_block, env);
+       heights_free(ir_heights);
 }
 
 #if 0 /*
@@ -926,7 +1117,7 @@ static ir_node *setup_frame(be_abi_irg_t *env)
        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, be_stack_dir_expand);
+               stack = be_new_IncSP(sp, irg, bl, stack, no_mem, BE_STACK_FRAME_SIZE_EXPAND);
                frame = stack;
        }
 
@@ -940,7 +1131,7 @@ static ir_node *setup_frame(be_abi_irg_t *env)
                        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, be_stack_dir_expand);
+               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);
@@ -965,7 +1156,7 @@ static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct
        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, be_stack_dir_shrink);
+               stack = be_new_IncSP(sp, irg, bl, stack, ret_mem, -BE_STACK_FRAME_SIZE_SHRINK);
        }
 
        else {
@@ -989,7 +1180,19 @@ static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct
 */
 #endif
 
-static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type *method_type)
+/**
+ * Computes the stack argument layout type.
+ * Changes a possibly allocated value param type by moving
+ * entities to the stack layout type.
+ *
+ * @param env          the ABI environment
+ * @param call         the current call ABI
+ * @param method_type  the method type
+ * @param param_map    an array mapping method arguments to the stack layout type
+ *
+ * @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, 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;
@@ -1000,29 +1203,45 @@ static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type
        char buf[128];
        ir_type *res;
        int i;
-
-       snprintf(buf, sizeof(buf), "%s_arg_type", get_entity_name(get_irg_entity(env->birg->irg)));
-       res = new_type_class(new_id_from_str(buf));
-
-       for(i = 0; i < n; ++i, curr += inc) {
-               type *param_type       = get_method_param_type(method_type, curr);
+       ir_type *val_param_tp = get_method_value_param_type(method_type);
+       ident *id = get_entity_ident(get_irg_entity(env->birg->irg));
+       ir_entity **map;
+
+       *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);
                be_abi_call_arg_t *arg = get_call_arg(call, 0, curr);
 
-               if(arg->on_stack) {
-                       snprintf(buf, sizeof(buf), "param_%d", i);
-                       arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
+               map[i] = NULL;
+               if (arg->on_stack) {
+                       if (val_param_tp) {
+                               /* the entity was already created, move it to the param type */
+                               arg->stack_ent = get_method_value_param_ent(method_type, i);
+                               remove_struct_member(val_param_tp, arg->stack_ent);
+                               set_entity_owner(arg->stack_ent, res);
+                               add_struct_member(res, arg->stack_ent);
+                               /* must be automatic to set a fixed layout */
+                               set_entity_allocation(arg->stack_ent, allocation_automatic);
+                       }
+                       else {
+                               snprintf(buf, sizeof(buf), "param_%d", i);
+                               arg->stack_ent = new_entity(res, new_id_from_str(buf), param_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;
                }
        }
-
        set_type_size_bytes(res, ofs);
+       set_type_state(res, layout_fixed);
        return res;
 }
 
+#if 0
 static void create_register_perms(const arch_isa_t *isa, ir_graph *irg, ir_node *bl, pmap *regs)
 {
        int i, j, n;
@@ -1064,6 +1283,7 @@ static void create_register_perms(const arch_isa_t *isa, ir_graph *irg, ir_node
 
        obstack_free(&obst, NULL);
 }
+#endif
 
 typedef struct {
        const arch_register_t *reg;
@@ -1098,11 +1318,14 @@ static reg_node_map_t *reg_map_to_arr(struct obstack *obst, pmap *reg_map)
        return res;
 }
 
+/**
+ * Creates a barrier.
+ */
 static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pmap *regs, int in_req)
 {
        ir_graph *irg = env->birg->irg;
+       int n_regs    = pmap_count(regs);
        int n;
-       int n_regs = pmap_count(regs);
        ir_node *irn;
        ir_node **in;
        reg_node_map_t *rm;
@@ -1122,9 +1345,10 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
        obstack_free(&env->obst, in);
 
        for(n = 0; n < n_regs; ++n) {
-               int pos = BE_OUT_POS(n);
-               ir_node *proj;
                const arch_register_t *reg = rm[n].reg;
+               int flags                  = 0;
+               int pos                    = BE_OUT_POS(n);
+               ir_node *proj;
 
                proj = new_r_Proj(irg, bl, irn, get_irn_mode(rm[n].irn), n);
                be_node_set_reg_class(irn, n, reg->reg_class);
@@ -1133,8 +1357,15 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
                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(arch_register_type_is(reg, ignore))
-                       be_node_set_flags(irn, pos, arch_irn_flags_ignore);
+
+               /* 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))
+                       flags |= arch_irn_flags_ignore;
+
+               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);
 
                pmap_insert(regs, (void *) reg, proj);
        }
@@ -1147,6 +1378,291 @@ static ir_node *create_barrier(be_abi_irg_t *env, ir_node *bl, ir_node **mem, pm
        return irn;
 }
 
+/**
+ * Creates a be_Return for a Return node.
+ *
+ * @param @env    the abi environment
+ * @param irn     the Return node or NULL if there was none
+ * @param bl      the block where the be_Retun should be placed
+ * @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;
+
+       pmap *reg_map  = pmap_create();
+       ir_node *keep  = pmap_get(env->keep_map, bl);
+       int in_max;
+       ir_node *ret;
+       int i, n;
+       ir_node **in;
+       ir_node *stack;
+       const arch_register_t **regs;
+       pmap_entry *ent ;
+
+       /*
+               get the valid stack node in this block.
+               If we had a call in that block there is a Keep constructed by process_calls()
+               which points to the last stack modification in that block. we'll use
+               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);
+       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));
+       }
+
+       /* Insert results for Return into the register map. */
+       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");
+               pmap_insert(reg_map, (void *) arg->reg, res);
+       }
+
+       /* Add uses of the callee save registers. */
+       pmap_foreach(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);
+
+       /* Make the Epilogue node and call the arch's epilogue maker. */
+       create_barrier(env, bl, &mem, reg_map, 1);
+       call->cb->epilogue(env->cb, bl, &mem, reg_map);
+
+       /*
+               Maximum size of the in array for Return nodes is
+               return args + callee save/ignore registers + memory + stack pointer
+       */
+       in_max = pmap_count(reg_map) + n_res + 2;
+
+       in   = obstack_alloc(&env->obst, in_max * sizeof(in[0]));
+       regs = obstack_alloc(&env->obst, in_max * sizeof(regs[0]));
+
+       in[0]   = mem;
+       in[1]   = be_abi_reg_map_get(reg_map, isa->sp);
+       regs[0] = NULL;
+       regs[1] = isa->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) {
+               be_abi_call_arg_t *arg = get_call_arg(call, 1, i);
+
+               in[n]     = be_abi_reg_map_get(reg_map, arg->reg);
+               regs[n++] = arg->reg;
+
+               /* Clear the map entry to mark the register as processed. */
+               be_abi_reg_map_set(reg_map, arg->reg, NULL);
+       }
+
+       /* grow the rest of the stuff. */
+       pmap_foreach(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);
+
+       /* Set the register classes of the return's parameter accordingly. */
+       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. */
+       obstack_free(&env->obst, in);
+       pmap_destroy(reg_map);
+
+       return ret;
+}
+
+typedef struct lower_frame_sels_env_t {
+       be_abi_irg_t *env;
+       ir_entity    *value_param_list;  /**< 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.
+ */
+static void lower_frame_sels_walker(ir_node *irn, void *data)
+{
+       lower_frame_sels_env_t *ctx = data;
+
+       if (is_Sel(irn)) {
+               ir_graph *irg        = current_ir_graph;
+               ir_node  *frame      = get_irg_frame(irg);
+               ir_node  *param_base = get_irg_value_param_base(irg);
+               ir_node  *ptr        = get_Sel_ptr(irn);
+
+               if (ptr == frame || ptr == param_base) {
+                       be_abi_irg_t *env = ctx->env;
+                       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);
+                       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) {
+                               set_entity_link(ent, ctx->value_param_list);
+                               ctx->value_param_list = ent;
+                       }
+               }
+       }
+}
+
+/**
+ * Check if a value parameter is transmitted as a register.
+ * This might happen if the address of an parameter is taken which is
+ * transmitted in registers.
+ *
+ * Note that on some architectures this case must be handled specially
+ * because the place of the backing store is determined by their ABI.
+ *
+ * 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, ir_entity *value_param_list) {
+       be_abi_call_t *call = env->call;
+       ir_graph *irg       = env->birg->irg;
+       ir_entity *ent, *next_ent, *new_list;
+       ir_type *frame_tp;
+       DEBUG_ONLY(firm_dbg_module_t *dbg = env->dbg;)
+
+       new_list = NULL;
+       for (ent = value_param_list; ent; ent = next_ent) {
+               int i = get_struct_member_index(get_entity_owner(ent), ent);
+               be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
+
+               next_ent = get_entity_link(ent);
+               if (arg->in_reg) {
+                       DBG((dbg, LEVEL_2, "\targ #%d need backing store\n", i));
+                       set_entity_link(ent, new_list);
+                       new_list = ent;
+               }
+       }
+       if (new_list) {
+               /* ok, change the graph */
+               ir_node *start_bl = get_irg_start_block(irg);
+               ir_node *first_bl = NULL;
+               ir_node *frame, *imem, *nmem, *store, *mem, *args, *args_bl;
+               const ir_edge_t *edge;
+               optimization_state_t state;
+               int offset;
+
+               foreach_block_succ(start_bl, edge) {
+                       ir_node *succ = get_edge_src_irn(edge);
+                       if (start_bl != succ) {
+                               first_bl = succ;
+                               break;
+                       }
+               }
+               assert(first_bl);
+               /* we had already removed critical edges, so the following
+                  assertion should be always true. */
+               assert(get_Block_n_cfgpreds(first_bl) == 1);
+
+               /* now create backing stores */
+               frame = get_irg_frame(irg);
+               imem = get_irg_initial_mem(irg);
+
+               save_optimization_state(&state);
+               set_optimize(0);
+               nmem = new_r_Proj(irg, first_bl, get_irg_start(irg), mode_M, pn_Start_M);
+               restore_optimization_state(&state);
+
+               /* reroute all edges to the new memory source */
+               edges_reroute(imem, nmem, irg);
+
+               store   = NULL;
+               mem     = imem;
+               args    = get_irg_args(irg);
+               args_bl = get_nodes_block(args);
+               for (ent = new_list; ent; ent = get_entity_link(ent)) {
+                       int     i     = get_struct_member_index(get_entity_owner(ent), ent);
+                       ir_type *tp   = get_entity_type(ent);
+                       ir_mode *mode = get_type_mode(tp);
+                       ir_node *addr;
+
+                       /* address for the backing store */
+                       addr = be_new_FrameAddr(env->isa->sp->reg_class, irg, first_bl, frame, ent);
+
+                       if (store)
+                               mem = new_r_Proj(irg, first_bl, store, mode_M, pn_Store_M);
+
+                       /* the backing store itself */
+                       store = new_r_Store(irg, first_bl, mem, addr,
+                                           new_r_Proj(irg, args_bl, args, mode, i));
+               }
+               /* the new memory Proj gets the last Proj from store */
+               set_Proj_pred(nmem, store);
+               set_Proj_proj(nmem, pn_Store_M);
+
+               /* move all entities to the frame type */
+               frame_tp = get_irg_frame_type(irg);
+               offset   = get_type_size_bytes(frame_tp);
+               for (ent = new_list; ent; ent = get_entity_link(ent)) {
+                       ir_type *tp = get_entity_type(ent);
+                       int align = get_type_alignment_bytes(tp);
+
+                       offset += align - 1;
+                       offset &= -align;
+                       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(ent, offset);
+                       offset += get_type_size_bytes(tp);
+               }
+               set_type_size_bytes(frame_tp, offset);
+       }
+}
+
+/**
+ * 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;
+               }
+       }
+}
+
 /**
  * Modify the irg itself and the frame type.
  */
@@ -1158,15 +1674,14 @@ static void modify_irg(be_abi_irg_t *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 *arg_tuple        = get_irg_args(irg);
-       ir_node *no_mem           = get_irg_no_mem(irg);
-       ir_node *mem              = get_irg_initial_mem(irg);
-       type *method_type         = get_entity_type(get_irg_entity(irg));
+       ir_node *old_mem          = get_irg_initial_mem(irg);
+       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              = get_method_n_params(method_type);
-       int max_arg               = 0;
 
-       int i, j, n;
+       int n_params;
+       int i, j, n, temp;
 
        reg_node_map_t *rm;
        const arch_register_t *fp_reg;
@@ -1174,34 +1689,50 @@ static void modify_irg(be_abi_irg_t *env)
        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;
+       lower_frame_sels_env_t ctx;
+       ir_entity **param_map;
 
-       pmap_entry *ent;
        bitset_t *used_proj_nr;
        DEBUG_ONLY(firm_dbg_module_t *dbg = env->dbg;)
 
        DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));
 
        /* Convert the Sel nodes in the irg to frame load/store/addr nodes. */
-       irg_walk_graph(irg, lower_frame_sels_walker, NULL, env);
+       ctx.env              = env;
+       ctx.value_param_list = NULL;
+       irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);
+
+       /* value_param_base anchor is not needed anymore now */
+       value_param_base = get_irg_value_param_base(irg);
+       be_kill_node(value_param_base);
+       set_irg_value_param_base(irg, new_r_Bad(irg));
 
        env->frame = obstack_alloc(&env->obst, sizeof(env->frame[0]));
        env->regs  = pmap_create();
 
-       /* Find the maximum proj number of the argument tuple proj */
-       foreach_out_edge(arg_tuple, edge)  {
-               ir_node *irn = get_edge_src_irn(edge);
-               int nr       = get_Proj_proj(irn);
-               max_arg      = MAX(max_arg, nr);
-       }
-
        used_proj_nr = bitset_alloca(1024);
-       max_arg      = MAX(max_arg + 1, n_params);
-       args         = obstack_alloc(&env->obst, max_arg * sizeof(args[0]));
-       memset(args, 0, max_arg * sizeof(args[0]));
+       n_params     = get_method_n_params(method_type);
+       args         = obstack_alloc(&env->obst, n_params * sizeof(args[0]));
+       memset(args, 0, n_params * sizeof(args[0]));
+
+       /* Check if a value parameter is transmitted as a register.
+        * This might happen if the address of an parameter is taken which is
+        * transmitted in registers.
+        *
+        * Note that on some architectures this case must be handled specially
+        * because the place of the backing store is determined by their ABI.
+        *
+        * In the default case we move the entity to the frame type and create
+        * a backing store into the first block.
+        */
+       fix_address_of_parameter_access(env, ctx.value_param_list);
 
        /* Fill the argument vector */
+       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);
@@ -1209,9 +1740,9 @@ static void modify_irg(be_abi_irg_t *env)
                DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
        }
 
-       arg_type = compute_arg_type(env, call, method_type);
+       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);
+       stack_frame_init(env->frame, arg_type, bet_type, get_irg_frame_type(irg), isa->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) {
@@ -1232,8 +1763,10 @@ static void modify_irg(be_abi_irg_t *env)
                const arch_register_class_t *cls = arch_isa_get_reg_class(isa, 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);
+                       }
                }
        }
 
@@ -1241,6 +1774,7 @@ static void modify_irg(be_abi_irg_t *env)
        pmap_insert(env->regs, (void *) isa->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));
 
        /*
         * make proj nodes for the callee save registers.
@@ -1253,11 +1787,12 @@ 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) {
                arch_register_t *reg = (void *) rm[i].reg;
-               ir_node *arg_proj    = rm[i].irn;
-               ir_node *proj;
-               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;
+
+               ir_node *proj;
 
                assert(nr >= 0);
                bitset_set(used_proj_nr, nr);
@@ -1271,28 +1806,50 @@ static void modify_irg(be_abi_irg_t *env)
                 * The Proj for that register shall also be ignored during register allocation.
                 */
                if(arch_register_type_is(reg, ignore))
-                       be_node_set_flags(env->reg_params, pos, arch_irn_flags_ignore);
+                       flags |= arch_irn_flags_ignore;
+
+               if(reg == sp)
+                       flags |= arch_irn_flags_modify_sp;
+
+               be_node_set_flags(env->reg_params, pos, flags);
 
                DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
        }
        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);
-       barrier = create_barrier(env, bl, &mem, env->regs, 0);
 
-       env->init_sp  = be_abi_reg_map_get(env->regs, sp);
-       env->init_sp  = be_new_IncSP(sp, irg, bl, env->init_sp, no_mem, BE_STACK_FRAME_SIZE, be_stack_dir_expand);
-       arch_set_irn_register(env->birg->main_env->arch_env, env->init_sp, sp);
+       /* 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);
        be_abi_reg_map_set(env->regs, sp, env->init_sp);
+
+       env->start_barrier = barrier = create_barrier(env, bl, &mem, env->regs, 0);
+
+       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);
 
-       if (is_Proj(frame_pointer) && get_Proj_pred(frame_pointer) == barrier)
-               be_node_set_flags(barrier, BE_OUT_POS(get_Proj_proj(frame_pointer)), arch_irn_flags_ignore);
+       set_irg_initial_mem(irg, mem);
 
        /* Now, introduce stack param nodes for all parameters passed on the stack */
-       for(i = 0; i < max_arg; ++i) {
+       for(i = 0; i < n_params; ++i) {
                ir_node *arg_proj = args[i];
                ir_node *repl     = NULL;
 
@@ -1326,109 +1883,73 @@ static void modify_irg(be_abi_irg_t *env)
                        }
 
                        assert(repl != NULL);
-                       edges_reroute(args[i], repl, irg);
+                       exchange(args[i], repl);
                }
        }
 
-       /* All Return nodes hang on the End node, so look for them there. */
-       for(i = 0, n = get_irn_arity(end); i < n; ++i) {
-               ir_node *irn = get_irn_n(end, i);
-
-               if(get_irn_opcode(irn) == iro_Return) {
-                       ir_node *bl    = get_nodes_block(irn);
-                       int n_res      = get_Return_n_ress(irn);
-                       pmap *reg_map  = pmap_create();
-                       ir_node *mem   = get_Return_mem(irn);
-                       int in_max;
-                       ir_node *ret;
-                       int i, n;
-                       ir_node **in;
-                       const arch_register_t **regs;
-
-                       pmap_insert(reg_map, (void *) sp, pmap_get(env->regs, (void *) sp));
-
-                       /* Insert results for Return into the register map. */
-                       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");
-                               pmap_insert(reg_map, (void *) arg->reg, res);
-                       }
-
-                       /* Add uses of the callee save registers. */
-                       pmap_foreach(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);
-                       }
-
-                       /* Make the Epilogue node and call the arch's epilogue maker. */
-                       create_barrier(env, bl, &mem, reg_map, 1);
-                       call->cb->epilogue(env->cb, bl, &mem, reg_map);
-
-                       /*
-                               Maximum size of the in array for Return nodes is
-                               return args + callee save/ignore registers + memory + stack pointer
-                       */
-                       in_max = pmap_count(reg_map) + get_Return_n_ress(irn) + 2;
-
-                       in   = obstack_alloc(&env->obst, in_max * sizeof(in[0]));
-                       regs = obstack_alloc(&env->obst, in_max * sizeof(regs[0]));
-
-                       in[0]   = mem;
-                       in[1]   = be_abi_reg_map_get(reg_map, sp);
-                       regs[0] = NULL;
-                       regs[1] = sp;
-                       n       = 2;
-
-                       /* clear SP entry, since it has already been grown. */
-                       pmap_insert(reg_map, (void *) sp, NULL);
-                       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);
-
-                               in[n]     = be_abi_reg_map_get(reg_map, arg->reg);
-                               regs[n++] = arg->reg;
-
-                               /* Clear the map entry to mark the register as processed. */
-                               be_abi_reg_map_set(reg_map, arg->reg, NULL);
-                       }
-
-                       /* grow the rest of the stuff. */
-                       pmap_foreach(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(get_irn_dbg_info(irn), irg, bl, n, in);
+       /* the arg proj is not needed anymore now */
+       assert(get_irn_n_edges(arg_tuple) == 0);
+       be_kill_node(arg_tuple);
+       set_irg_args(irg, new_rd_Bad(irg));
 
-                       /* Set the register classes of the return's parameter accordingly. */
-                       for(i = 0; i < n; ++i)
-                               if(regs[i])
-                                       be_node_set_reg_class(ret, i, regs[i]->reg_class);
+       /* All Return nodes hang on the End node, so look for them there. */
+       for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
+               ir_node *irn = get_Block_cfgpred(end, i);
 
-                       /* Free the space of the Epilog's in array and the register <-> proj map. */
-                       obstack_free(&env->obst, in);
+               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));
                        exchange(irn, ret);
-                       pmap_destroy(reg_map);
                }
        }
+       /* 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) */
+       temp = 0;
+       irg_block_walk_graph(irg, fix_start_block, NULL, &temp);
 }
 
-/**
- * Walker: puts all Alloc(stack_alloc) on a obstack
- */
-static void collect_alloca_walker(ir_node *irn, void *data)
+/** Fix the state inputs of calls that still hang on unknowns */
+static
+void fix_call_state_inputs(be_abi_irg_t *env)
 {
-       be_abi_irg_t *env = data;
-       if(get_irn_opcode(irn) == iro_Alloc && get_Alloc_where(irn) == stack_alloc)
-               obstack_ptr_grow(&env->obst, irn);
+       const arch_isa_t *isa = env->isa;
+       int i, n, n_states;
+       arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);
+
+       /* Collect caller save registers */
+       n = arch_isa_get_n_reg_class(isa);
+       for(i = 0; 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) {
+                       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 statereg 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);
+               }
+       }
 }
 
 be_abi_irg_t *be_abi_introduce(be_irg_t *birg)
@@ -1439,57 +1960,100 @@ be_abi_irg_t *be_abi_introduce(be_irg_t *birg)
 
        pmap_entry *ent;
        ir_node *dummy;
+       optimization_state_t state;
+       unsigned *limited_bitset;
+
+       be_omit_fp = birg->main_env->options->omit_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();
+       env->isa         = birg->main_env->arch_env->isa;
+       env->method_type = get_entity_type(get_irg_entity(irg));
+       env->call        = be_abi_call_new(env->isa->sp->reg_class);
        arch_isa_get_call_abi(env->isa, env->method_type, env->call);
 
-       env->dce_survivor     = new_survive_dce();
-       env->birg             = birg;
-       env->stack_phis       = pset_new_ptr(16);
-       env->init_sp = dummy  = new_r_Unknown(irg, env->isa->sp->reg_class->mode);
-       FIRM_DBG_REGISTER(env->dbg, "firm.be.abi");
+       env->ignore_regs  = pset_new_ptr_default();
+       env->keep_map     = pmap_create();
+       env->dce_survivor = new_survive_dce();
+       env->birg         = birg;
 
-       env->cb = env->call->cb->init(env->call, birg->main_env->arch_env, irg);
+       env->sp_req.type    = arch_register_req_type_limited;
+       env->sp_req.cls     = arch_register_get_class(env->isa->sp);
+       limited_bitset      = rbitset_obstack_alloc(&env->obst, env->sp_req.cls->n_regs);
+       rbitset_set(limited_bitset, arch_register_get_index(env->isa->sp));
+       env->sp_req.limited = limited_bitset;
 
-       obstack_init(&env->obst);
+       env->sp_cls_req.type  = arch_register_req_type_normal;
+       env->sp_cls_req.cls   = arch_register_get_class(env->isa->sp);
+
+       /* 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);
+       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);
 
        /* Lower all call nodes in the IRG. */
        process_calls(env);
 
+       /*
+               Beware: init backend abi call object after processing calls,
+               otherwise some information might be not yet available.
+       */
+       env->cb = env->call->cb->init(env->call, birg->main_env->arch_env, irg);
+
        /* 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);
+
+       /* calls array is not needed anymore */
+       DEL_ARR_F(env->calls);
+
        /* 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)
+       pmap_foreach(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);
-       be_liveness(irg);
+       env->cb = NULL;
        return env;
 }
 
 void be_abi_free(be_abi_irg_t *env)
 {
        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);
 }
 
+void be_abi_put_ignore_regs(be_abi_irg_t *abi, const arch_register_class_t *cls, bitset_t *bs)
+{
+       arch_register_t *reg;
+
+       for(reg = pset_first(abi->ignore_regs); reg; reg = pset_next(abi->ignore_regs))
+               if(reg->reg_class == cls)
+                       bitset_set(bs, 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;
+}
 
 /*
 
@@ -1501,85 +2065,120 @@ void be_abi_free(be_abi_irg_t *env)
 
 */
 
+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)
 {
-       pset *s = data;
+       fix_stack_walker_env_t *env = data;
 
-       if(be_is_AddSP(irn)     || be_is_IncSP(irn)     || be_is_SetSP(irn))
-               pset_insert_ptr(s, 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)
 {
-       dom_front_info_t *df;
-       pset *stack_nodes;
-
-       /* We need dominance frontiers for fix up */
-       df = be_compute_dominance_frontiers(env->birg->irg);
-       stack_nodes = pset_new_ptr(16);
-       pset_insert_ptr(stack_nodes, env->init_sp);
-       irg_walk_graph(env->birg->irg, collect_stack_nodes_walker, NULL, stack_nodes);
-       be_ssa_constr_set_phis(df, stack_nodes, env->stack_phis);
-       del_pset(stack_nodes);
-
-       /* Liveness could have changed due to Phi nodes. */
-       be_liveness(env->birg->irg);
-
-       /* 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;
+       arch_isa_t *isa;
+
+       walker_env.sp_nodes = NEW_ARR_F(ir_node*, 0);
+       walker_env.arch_env = birg->main_env->arch_env;
+       isa = walker_env.arch_env->isa;
+
+       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;
+       }
 
-/**
- * Translates a direction of an IncSP node (either be_stack_dir_shrink, or ...expand)
- * into -1 or 1, respectively.
- * @param irn The node.
- * @return 1, if the direction of the IncSP was along, -1 if against.
- */
-static int get_dir(ir_node *irn)
-{
-       return 1 - 2 * (be_get_IncSP_direction(irn) == be_stack_dir_shrink);
+       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->isa->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)
 {
-       const arch_env_t *aenv = env->birg->main_env->arch_env;
+       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;
 
        sched_foreach(bl, irn) {
 
                /*
-                       If the node modifies the stack pointer by a constant offset,
-                       record that in the bias.
-               */
-               if(be_is_IncSP(irn)) {
-                       int ofs = be_get_IncSP_offset(irn);
-                       int dir = get_dir(irn);
-
-                       if(ofs == BE_STACK_FRAME_SIZE) {
-                               ofs = get_type_size_bytes(get_irg_frame_type(env->birg->irg));
-                               be_set_IncSP_offset(irn, ofs);
-                       }
-
-                       if(omit_fp)
-                               bias += dir * 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.
+                */
+               ir_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));
                }
 
                /*
-                       Else check, if the node relates to an entity on the stack frame.
-                       If so, set the true offset (including the bias) for that
-                       node.
-               */
-               else {
-                       entity *ent = arch_get_frame_entity(aenv, irn);
-                       if(ent) {
-                               int offset = get_stack_entity_offset(env->frame, ent, bias);
-                               arch_set_frame_offset(aenv, irn, offset);
-                               DBG((env->dbg, LEVEL_2, "%F has offset %d\n", ent, offset));
+                  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);
+                               }
                        }
+
+                       if(omit_fp)
+                               bias += ofs;
                }
        }
 
@@ -1592,6 +2191,7 @@ 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. */
 };
 
 /**
@@ -1599,8 +2199,8 @@ struct bias_walk {
  */
 static void stack_bias_walker(ir_node *bl, void *data)
 {
-       if(bl != get_irg_start_block(get_irn_irg(bl))) {
-               struct bias_walk *bw = data;
+       struct bias_walk *bw = data;
+       if (bl != bw->start_block) {
                process_stack_bias(bw->env, bl, bw->start_block_bias);
        }
 }
@@ -1611,13 +2211,14 @@ void be_abi_fix_stack_bias(be_abi_irg_t *env)
        struct bias_walk bw;
 
        stack_frame_compute_initial_offset(env->frame);
-       // stack_frame_dump(stdout, env->frame);
+       // stack_layout_dump(stdout, env->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);
 
        /* fix the bias is all other blocks */
        bw.env = env;
+       bw.start_block = get_irg_start_block(irg);
        irg_block_walk_graph(irg, stack_bias_walker, NULL, &bw);
 }
 
@@ -1628,100 +2229,22 @@ ir_node *be_abi_get_callee_save_irn(be_abi_irg_t *abi, const arch_register_t *re
        return pmap_get(abi->regs, (void *) reg);
 }
 
-/*
-  _____ _____  _   _   _    _                 _ _
- |_   _|  __ \| \ | | | |  | |               | | |
-   | | | |__) |  \| | | |__| | __ _ _ __   __| | | ___ _ __
-   | | |  _  /| . ` | |  __  |/ _` | '_ \ / _` | |/ _ \ '__|
-  _| |_| | \ \| |\  | | |  | | (_| | | | | (_| | |  __/ |
- |_____|_|  \_\_| \_| |_|  |_|\__,_|_| |_|\__,_|_|\___|_|
-
-  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)
+ir_node *be_abi_get_ignore_irn(be_abi_irg_t *abi, const arch_register_t *reg)
 {
-       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;
+       assert(arch_register_type_is(reg, ignore));
+       assert(pmap_contains(abi->regs, (void *) reg));
+       return pmap_get(abi->regs, (void *) reg);
 }
 
-static entity *abi_get_frame_entity(const void *_self, const ir_node *irn)
+ir_node *be_abi_get_start_barrier(be_abi_irg_t *abi)
 {
-       return NULL;
+       return abi->start_barrier;
 }
 
-static void abi_set_stack_bias(const void *_self, ir_node *irn, int bias)
-{
+/**
+ * Returns non-zero if the ABI has omitted the frame pointer in
+ * the current graph.
+ */
+int be_abi_omit_fp(const be_abi_irg_t *abi) {
+       return abi->call->flags.bits.try_omit_fp;
 }
-
-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_stack_bias
-};
-
-static const arch_irn_handler_t abi_irn_handler = {
-       abi_get_irn_ops
-};