Corrected setting register classes of Return and Barrier nodes
[libfirm] / ir / be / beabi.c
index 9fff1b8..5f88bd2 100644 (file)
@@ -1,14 +1,17 @@
 /**
  * ABI lowering.
  *
- *
- *
+ * @author Sebastian Hack
+ * @date 7.3.2005
  */
 
-#include "firm_config.h"
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
 #include "obst.h"
 
 #include "type.h"
+#include "irgopt.h"
 
 #include "irgraph_t.h"
 #include "irnode_t.h"
 #include "iredges_t.h"
 #include "irgmod.h"
 #include "irgwalk.h"
+#include "irprintf_t.h"
+#include "irgopt.h"
 
 #include "be.h"
 #include "beabi.h"
 #include "bearch.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))
 
 typedef struct _be_abi_call_arg_t {
-       unsigned is_res : 1;
-       unsigned in_reg : 1;
+       unsigned is_res   : 1;
+       unsigned in_reg   : 1;
+       unsigned on_stack : 1;
 
        int pos;
        const arch_register_t *reg;
+       entity *stack_ent;
 } 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;
+};
+
 struct _be_abi_irg_t {
-       struct obstack      obst;
-       be_irg_t            *birg;
-       ir_node             *init_sp;      /**< The node representing the stack pointer
+       struct obstack       obst;
+       firm_dbg_module_t    *dbg;          /**< The debugging module. */
+       be_stack_frame_t     *frame;        /**< The stack frame model. */
+       const 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_node              *init_sp;      /**< The node representing the stack pointer
                                                                             at the start of the function. */
 
-       pset                *stack_ops;    /**< Contains all nodes modifying the stack pointer. */
-       pset                *ignore_regs;  /**< Contains all registers which shall be ignored
-                                                                            during register allocation. */
+       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. */
 
-       unsigned omit_framepointer: 1;     /**< If one, the frame(base-)pointer can be used
-                                                                            as an ordinary register. */
+       arch_irn_handler_t irn_handler;
+       arch_irn_ops_t     irn_ops;
 };
 
+#define abi_offset_of(type,member) ((char *) &(((type *) 0)->member) - (char *) 0)
+#define abi_get_relative(ptr, member) ((void *) ((char *) (ptr) - abi_offset_of(be_abi_irg_t, member)))
+#define get_abi_from_handler(ptr) abi_get_relative(ptr, irn_handler)
+#define get_abi_from_ops(ptr)     abi_get_relative(ptr, irn_ops)
+
+/* 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;
+
+/*
+     _    ____ ___    ____      _ _ _                _
+    / \  | __ )_ _|  / ___|__ _| | | |__   __ _  ___| | _____
+   / _ \ |  _ \| |  | |   / _` | | | '_ \ / _` |/ __| |/ / __|
+  / ___ \| |_) | |  | |__| (_| | | | |_) | (_| | (__|   <\__ \
+ /_/   \_\____/___|  \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/
+
+  These callbacks are used by the backend to set the parameters
+  for a specific call type.
+*/
+
 static int cmp_call_arg(const void *a, const void *b, size_t n)
 {
        const be_abi_call_arg_t *p = a, *q = b;
@@ -63,6 +127,7 @@ static be_abi_call_arg_t *get_or_set_call_arg(be_abi_call_t *call, int is_res, i
        be_abi_call_arg_t arg;
        unsigned hash;
 
+       memset(&arg, 0, sizeof(arg));
        arg.is_res = is_res;
        arg.pos    = pos;
 
@@ -78,33 +143,43 @@ static INLINE be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, i
        return get_or_set_call_arg(call, is_res, pos, 0);
 }
 
-void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags)
+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->flags        = flags;
+       call->cb           = cb;
 }
 
 void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos)
 {
        be_abi_call_arg_t *arg = get_or_set_call_arg(call, 0, arg_pos, 1);
+       arg->on_stack = 1;
 }
 
 void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg)
 {
        be_abi_call_arg_t *arg = get_or_set_call_arg(call, 0, arg_pos, 1);
+       arg->in_reg = 1;
        arg->reg = reg;
 }
 
 void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg)
 {
        be_abi_call_arg_t *arg = get_or_set_call_arg(call, 1, arg_pos, 1);
+       arg->in_reg = 1;
        arg->reg = reg;
 }
 
+be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
+{
+       return call->flags;
+}
+
 be_abi_call_t *be_abi_call_new(void)
 {
        be_abi_call_t *call = malloc(sizeof(call[0]));
-       call->flags  = BE_ABI_NONE;
-       call->params = new_set(cmp_call_arg, 16);
+       call->flags.val  = 0;
+       call->params     = new_set(cmp_call_arg, 16);
+       call->cb         = NULL;
        return call;
 }
 
@@ -114,20 +189,178 @@ void be_abi_call_free(be_abi_call_t *call)
        free(call);
 }
 
+/*
+  _____                           _   _                 _ _ _
+ |  ___| __ __ _ _ __ ___   ___  | | | | __ _ _ __   __| | (_)_ __   __ _
+ | |_ | '__/ _` | '_ ` _ \ / _ \ | |_| |/ _` | '_ \ / _` | | | '_ \ / _` |
+ |  _|| | | (_| | | | | | |  __/ |  _  | (_| | | | | (_| | | | | | | (_| |
+ |_|  |_|  \__,_|_| |_| |_|\___| |_| |_|\__,_|_| |_|\__,_|_|_|_| |_|\__, |
+                                                                    |___/
+
+  Handling of the stack frame. It is composed of three types:
+  1) The type of the arguments which are pushed on the stack.
+  2) The "between type" which consists of stuff the call of the
+     function pushes on the stack (like the return address and
+        the old base pointer for ia32).
+  3) The Firm frame type which consists of all local variables
+     and the spills.
+*/
+
+static int get_stack_entity_offset(be_stack_frame_t *frame, entity *ent, int bias)
+{
+       type *t = get_entity_owner(ent);
+       int ofs = get_entity_offset_bytes(ent);
+
+       int i, index;
+
+       /* Find the type the entity is contained in. */
+       for(index = 0; index < N_FRAME_TYPES; ++index) {
+               if(frame->order[index] == t)
+                       break;
+       }
+
+       /* Add the size of all the types below the one of the entity to the entity's offset */
+       for(i = 0; i < index; ++i)
+               ofs += get_type_size_bytes(frame->order[i]);
+
+       /* correct the offset by the initial position of the frame pointer */
+       ofs -= frame->initial_offset;
+
+       /* correct the offset with the current bias. */
+       ofs += bias;
+
+       return ofs;
+}
+
+static entity *search_ent_with_offset(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)
+                       return ent;
+       }
+
+       return NULL;
+}
+
+static int stack_frame_compute_initial_offset(be_stack_frame_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);
+       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)
+{
+       frame->arg_type       = args;
+       frame->between_type   = between;
+       frame->frame_type     = locals;
+       frame->initial_offset = 0;
+       frame->stack_dir      = stack_dir;
+       frame->order[1]       = between;
+
+       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)
+{
+       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];
+
+               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, "\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 computing the address of an entity
+ * on the frame type return the entity, else NULL.
+ */
+static INLINE entity *get_sel_ent(ir_node *irn)
+{
+       if(get_irn_opcode(irn) == iro_Sel
+               && 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);
+       }
+
+       if(nw != NULL)
+               exchange(irn, nw);
+}
+
 static INLINE int is_on_stack(be_abi_call_t *call, int pos)
 {
        be_abi_call_arg_t *arg = get_call_arg(call, 0, pos);
        return arg && !arg->in_reg;
 }
 
-static void adjust_call(be_abi_irg_t *env, ir_node *irn)
+/*
+   ____      _ _
+  / ___|__ _| | |___
+ | |   / _` | | / __|
+ | |__| (_| | | \__ \
+  \____\__,_|_|_|___/
+
+  Adjustment of the calls inside a graph.
+
+*/
+
+/**
+ * Transform a call node.
+ * @param env The ABI environment for the current irg.
+ * @param irn The call node.
+ * @param curr_sp The stack pointer node to use.
+ * @return The stack pointer after the call.
+ */
+static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
 {
        ir_graph *irg             = env->birg->irg;
        const arch_isa_t *isa     = env->birg->main_env->arch_env->isa;
        be_abi_call_t *call       = be_abi_call_new();
        ir_type *mt               = get_Call_type(irn);
+       ir_node *call_ptr         = get_Call_ptr(irn);
        int n_params              = get_method_n_params(mt);
-       ir_node *curr_sp          = get_irg_frame(irg);
        ir_node *curr_mem         = get_Call_mem(irn);
        ir_node *bl               = get_nodes_block(irn);
        pset *results             = pset_new_ptr(8);
@@ -137,15 +370,16 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
        const arch_register_t *sp = arch_isa_sp(isa);
        ir_mode *mach_mode        = sp->reg_class->mode;
        struct obstack *obst      = &env->obst;
+       ir_node *no_mem           = get_irg_no_mem(irg);
 
        ir_node *res_proj = NULL;
-       int curr_res_proj = -1;
+       int curr_res_proj = pn_Call_max;
        int n_low_args    = 0;
        int n_pos         = 0;
 
        ir_node *low_call;
        ir_node **in;
-       ir_node *sp_proj;
+       ir_node **res_projs;
        const ir_edge_t *edge;
        int *low_args;
        int *pos;
@@ -155,8 +389,11 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
        arch_isa_get_call_abi(isa, mt, call);
 
        /* Insert code to put the stack arguments on the stack. */
-       for(i = get_irn_arity(irn); i >= 0; --i) {
-               if(is_on_stack(call, i)) {
+       assert(get_Call_n_params(irn) == n_params);
+       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 += get_type_size_bytes(get_method_param_type(mt, i));
                        obstack_int_grow(obst, i);
                        n_pos++;
@@ -165,7 +402,7 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
        pos = obstack_finish(obst);
 
        /* Collect all arguments which are passed in registers. */
-       for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
+       for(i = 0, n = get_Call_n_params(irn); i < n; ++i) {
                be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
                if(arg && arg->in_reg) {
                        obstack_int_grow(obst, i);
@@ -174,13 +411,13 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
        }
        low_args = obstack_finish(obst);
 
-       /* If there are some parameters shich shall be passed on the stack. */
+       /* If there are some parameters which shall be passed on the stack. */
        if(n_pos > 0) {
                int curr_ofs      = 0;
-               int do_seq        = (call->flags & BE_ABI_USE_PUSH);
+               int do_seq        = call->flags.bits.store_args_sequential;
 
                /* Reverse list of stack parameters if call arguments are from left to right */
-               if(call->flags & BE_ABI_LEFT_TO_RIGHT) {
+               if(call->flags.bits.left_to_right) {
                        for(i = 0; i < n_pos / 2; ++i) {
                                int other  = n_pos - i - 1;
                                int tmp    = pos[i];
@@ -195,13 +432,13 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
                 * moving the stack pointer along the stack's direction.
                 */
                if(stack_dir < 0 && !do_seq) {
-                       curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, stack_size, be_stack_dir_along);
-                       pset_insert_ptr(env->stack_ops, curr_sp);
+                       curr_sp = be_new_IncSP(sp, irg, bl, curr_sp, no_mem, stack_size, be_stack_dir_along);
                }
 
+               assert(mode_is_reference(mach_mode) && "machine mode must be pointer");
                for(i = 0; i < n_pos; ++i) {
                        int p            = pos[i];
-                       ir_node *param   = get_irn_n(irn, 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);
@@ -209,12 +446,12 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
 
                        /* Make the expression to compute the argument's offset. */
                        if(curr_ofs > 0) {
-                               addr = new_r_Const_long(irg, bl, mach_mode, curr_ofs);
+                               addr = new_r_Const_long(irg, bl, mode_Is, curr_ofs);
                                addr = new_r_Add(irg, bl, curr_sp, addr, mach_mode);
                        }
 
                        /* Insert a store for primitive arguments. */
-                       if(is_Primitive_type(param_type)) {
+                       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);
                        }
@@ -237,16 +474,8 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
                        */
                        if(do_seq) {
                                curr_ofs = 0;
-                               curr_sp  = be_new_IncSP(sp, irg, bl, curr_sp, param_size, be_stack_dir_along);
+                               curr_sp  = be_new_IncSP(sp, irg, bl, curr_sp, no_mem, param_size, be_stack_dir_along);
                                curr_mem = mem;
-
-                               /*
-                                * only put the first IncSP to the stack fixup set since the other
-                                * ones are correctly connected to other nodes and do not need
-                                * to be fixed.
-                                */
-                               if(i == 0)
-                                       pset_insert_ptr(env->stack_ops, curr_sp);
                        }
                }
 
@@ -259,8 +488,15 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
        }
 
        /* Collect caller save registers */
-       for(i = 0; env->birg->main_env->caller_save[i]; ++i)
-               pset_insert_ptr(caller_save, env->birg->main_env->caller_save[i]);
+       for(i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
+               int j;
+               const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
+               for(j = 0; j < cls->n_regs; ++j) {
+                       const arch_register_t *reg = arch_register_for_index(cls, j);
+                       if(arch_register_type_is(reg, caller_save))
+                               pset_insert_ptr(caller_save, (void *) reg);
+               }
+       }
 
        /* search the greatest result proj number */
        foreach_out_edge(irn, edge) {
@@ -275,93 +511,552 @@ static void adjust_call(be_abi_irg_t *env, ir_node *irn)
                                ir_node *res = get_edge_src_irn(res_edge);
 
                                assert(is_Proj(res));
+
                                proj = get_Proj_proj(res);
                                arg = get_call_arg(call, 1, proj);
+
+                               /*
+                                       shift the proj number to the right, since we will drop the
+                                       unspeakable Proj_T from the Call. Therefore, all real argument
+                                       Proj numbers must be increased by pn_Call_max
+                               */
+                               proj += pn_Call_max;
+                               set_Proj_proj(res, proj);
+                               obstack_ptr_grow(obst, res);
+
                                if(proj > curr_res_proj)
                                        curr_res_proj = proj;
-                               if(arg->in_reg)
+                               if(arg->in_reg) {
                                        pset_remove_ptr(caller_save, arg->reg);
+                                       //pmap_insert(arg_regs, arg->reg, INT_TO_PTR(proj + 1))
+                               }
                        }
                }
        }
        curr_res_proj++;
+       obstack_ptr_grow(obst, NULL);
+       res_projs = obstack_finish(obst);
+
+       /* make the back end call node and set its register requirements. */
+       for(i = 0; i < n_low_args; ++i)
+               obstack_ptr_grow(obst, get_Call_param(irn, low_args[i]));
+
+       in = obstack_finish(obst);
+       if(env->call->flags.bits.call_has_imm && get_irn_opcode(call_ptr) == iro_SymConst) {
+               low_call = be_new_Call(irg, bl, curr_mem, curr_sp, curr_sp, curr_res_proj, n_low_args, in);
+               be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
+       }
+
+       else
+               low_call = be_new_Call(irg, bl, curr_mem, curr_sp, call_ptr, curr_res_proj, n_low_args, in);
+
+       /* Set the register classes and constraints of the Call parameters. */
+       for(i = 0; i < n_low_args; ++i) {
+               int index = low_args[i];
+               be_abi_call_arg_t *arg = get_call_arg(call, 0, index);
+               assert(arg->reg != NULL);
+
+#if 0
+               if(pmap_contains(arg_regs, (void *) arg->reg)) {
+                       int out_proj_num = PTR_TO_INT(be_abi_reg_map_get(arg_regs, arg->reg)) - 1;
+                       be_node_set_must_be_same(low_call, out_proj_num, low_args[i]);
+               }
+
+               else
+#endif
+                       be_set_constr_single_reg(low_call, index, arg->reg);
+       }
+
+       /* Set the register constraints of the results. */
+       for(i = 0; res_projs[i]; ++i) {
+               ir_node *irn                 = res_projs[i];
+               int proj                     = get_Proj_proj(irn);
+
+               /* Correct Proj number since it has been adjusted! (see above) */
+               const be_abi_call_arg_t *arg = get_call_arg(call, 1, proj - pn_Call_max);
+
+               assert(arg->in_reg);
+               be_set_constr_single_reg(low_call, BE_OUT_POS(proj), arg->reg);
+       }
+       obstack_free(obst, in);
+       exchange(irn, low_call);
+
+       /* redirect the result projs to the lowered call instead of the Proj_T */
+       for(i = 0; res_projs[i]; ++i)
+               set_Proj_pred(res_projs[i], low_call);
 
        /* Make additional projs for the caller save registers
           and the Keep node which keeps them alive. */
        if(pset_count(caller_save) > 0) {
                const arch_register_t *reg;
-               ir_node **in;
+               ir_node **in, *keep;
+               int i, n;
 
-               if(!res_proj)
-                       res_proj = new_r_Proj(irg, bl, irn, mode_T, pn_Call_T_result);
+               for(reg = pset_first(caller_save), n = 0; reg; reg = pset_next(caller_save), ++n) {
+                       ir_node *proj = new_r_Proj(irg, bl, low_call, reg->reg_class->mode, curr_res_proj++);
 
-               for(reg = pset_first(caller_save); reg; reg = pset_next(caller_save))
-                       obstack_ptr_grow(obst, new_r_Proj(irg, bl, res_proj, 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. */
+                       set_irn_link(proj, (void *) reg);
+                       obstack_ptr_grow(obst, proj);
+               }
 
-               in = (ir_node **) obstack_finish(obst);
-               be_new_Keep(NULL, irg, bl, pset_count(caller_save), in);
+               in   = (ir_node **) obstack_finish(obst);
+               keep = be_new_Keep(NULL, irg, bl, n, in);
+               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);
+               }
                obstack_free(obst, in);
        }
 
        /* Clean up the stack. */
        if(stack_size > 0) {
-               ir_node *last_inc_sp;
+               ir_node *mem_proj = NULL;
 
-               /* Get the result ProjT */
-               if(!res_proj)
-                       res_proj = new_r_Proj(irg, bl, irn, mode_T, pn_Call_T_result);
+               foreach_out_edge(low_call, edge) {
+                       ir_node *irn = get_edge_src_irn(edge);
+                       if(is_Proj(irn) && get_Proj_proj(irn) == pn_Call_M) {
+                               mem_proj = irn;
+                               break;
+                       }
+               }
+
+               if(!mem_proj)
+                       mem_proj = new_r_Proj(irg, bl, low_call, mode_M, pn_Call_M);
 
                /* Make a Proj for the stack pointer. */
-               sp_proj     = new_r_Proj(irg, bl, res_proj, sp->reg_class->mode, curr_res_proj++);
-               last_inc_sp = be_new_IncSP(sp, irg, bl, sp_proj, stack_size, be_stack_dir_against);
-               pset_insert_ptr(env->stack_ops, last_inc_sp);
+               curr_sp     = be_new_IncSP(sp, irg, bl, curr_sp, mem_proj, stack_size, be_stack_dir_against);
        }
 
-       /* at last make the backend call node and set its register requirements. */
-       for(i = 0; i < n_low_args; ++i)
-               obstack_ptr_grow(obst, get_irn_n(irn, low_args[i]));
-       in = obstack_finish(obst);
-       low_call = be_new_Call(irg, bl, curr_mem, curr_sp, get_Call_ptr(irn), curr_res_proj, n_low_args, in);
-       obstack_free(obst, in);
-
-       exchange(irn, low_call);
-
        be_abi_call_free(call);
        obstack_free(obst, pos);
        del_pset(results);
        del_pset(caller_save);
+
+       return curr_sp;
 }
 
-static void adjust_call_walker(ir_node *irn, void *data)
+/**
+ * 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)
 {
-       if(get_irn_opcode(irn) == iro_Call)
-               adjust_call(data, irn);
+       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;
+
+               const ir_edge_t *edge;
+               ir_node *new_alloc;
+
+               env->call->flags.bits.try_omit_fp = 0;
+
+               new_alloc = be_new_AddSP(env->isa->sp, irg, bl, curr_sp, get_Alloc_size(alloc));
+
+               foreach_out_edge(alloc, edge) {
+                       ir_node *irn = get_edge_src_irn(edge);
+
+                       assert(is_Proj(irn));
+                       switch(get_Proj_proj(irn)) {
+                       case pn_Alloc_M:
+                               alloc_mem = irn;
+                               break;
+                       case pn_Alloc_res:
+                               alloc_res = irn;
+                               break;
+                       default:
+                               break;
+                       }
+               }
+
+               assert(alloc_res != NULL);
+               exchange(alloc_res, env->isa->stack_dir < 0 ? new_alloc : curr_sp);
+
+               if(alloc_mem != NULL)
+                       exchange(alloc_mem, new_r_NoMem(irg));
+
+               curr_sp = new_alloc;
+       }
+
+       return curr_sp;
+}
+
+/**
+ * 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)
+{
+       int n, i;
+
+       if(get_irn_visited(curr) >= visited_nr)
+               return 0;
+
+       set_irn_visited(curr, visited_nr);
+       if(get_nodes_block(curr) != bl)
+               return 0;
+
+       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;
+       }
+
+       return 0;
 }
 
 /**
- * Walker to implement alloca-style allocations.
- * They are implemented using an add to the stack pointer
- * and a copy instruction.
+ * Check if a node is somehow data dependent on another one.
+ * both nodes must be in the same basic block.
+ * @param n1 The first node.
+ * @param n2 The second node.
+ * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
  */
-static void implement_stack_alloc(be_abi_irg_t *env, ir_node *irn)
+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(bl == get_nodes_block(n2));
+       set_irg_visited(irg, vis_nr);
+       return check_dependence(n1, n2, bl, vis_nr);
+}
+
+static int cmp_call_dependecy(const void *c1, const void *c2)
+{
+       ir_node *n1 = *(ir_node **) c1;
+       ir_node *n2 = *(ir_node **) c2;
+
+       /*
+               Classical qsort() comparison function behavior:
+               0  if both elements are equal
+               1  if second is "smaller" that first
+               -1 if first is "smaller" that second
+       */
+       return n1 == n2 ? 0 : (dependent_on(n1, n2) ? -1 : 1);
+}
+
+static void link_calls_in_block_walker(ir_node *irn, void *data)
+{
+       if(is_Call(irn)) {
+               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;
+
+               set_irn_link(irn, save);
+               set_irn_link(bl, irn);
+       }
+}
+
+/**
+ * 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
+ * and that connect the calls together.
+ */
+static void process_calls_in_block(ir_node *bl, void *data)
+{
+       be_abi_irg_t *env = data;
+       ir_node *curr_sp  = env->init_sp;
+       ir_node *irn;
+       int n;
+
+       for(irn = get_irn_link(bl), n = 0; irn; irn = get_irn_link(irn), ++n)
+               obstack_ptr_grow(&env->obst, irn);
+
+       /* If there were call nodes in the block. */
+       if(n > 0) {
+               ir_node **nodes;
+               int i;
+
+               nodes = obstack_finish(&env->obst);
+
+               /* order the call nodes according to data dependency */
+               qsort(nodes, n, sizeof(nodes[0]), cmp_call_dependecy);
+
+               for(i = n - 1; i >= 0; --i) {
+                       ir_node *irn = nodes[i];
+
+                       switch(get_irn_opcode(irn)) {
+                       case iro_Call:
+                               curr_sp = adjust_call(env, irn, curr_sp);
+                               break;
+                       case iro_Alloc:
+                               curr_sp = adjust_alloc(env, irn, curr_sp);
+                               break;
+                       default:
+                               break;
+                       }
+               }
+
+               obstack_free(&env->obst, nodes);
+
+               /* 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);
+       }
+
+       set_irn_link(bl, curr_sp);
+}
+
+/**
+ * Adjust all call nodes in the graph to the ABI conventions.
+ */
+static void process_calls(be_abi_irg_t *env)
+{
+       ir_graph *irg = env->birg->irg;
+
+       env->call->flags.bits.irg_is_leaf = 1;
+       irg_walk_graph(irg, firm_clear_link, link_calls_in_block_walker, env);
+       irg_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);
+       }
+}
+
+static ir_node *setup_frame(be_abi_irg_t *env)
+{
+       const arch_isa_t *isa = env->birg->main_env->arch_env->isa;
+       const arch_register_t *sp = isa->sp;
+       const arch_register_t *bp = isa->bp;
+       be_abi_call_flags_bits_t flags = env->call->flags.bits;
+       ir_graph *irg      = env->birg->irg;
+       ir_node *bl        = get_irg_start_block(irg);
+       ir_node *no_mem    = get_irg_no_mem(irg);
+       ir_node *old_frame = get_irg_frame(irg);
+       ir_node *stack     = pmap_get(env->regs, (void *) sp);
+       ir_node *frame     = pmap_get(env->regs, (void *) bp);
+
+       int stack_nr       = get_Proj_proj(stack);
+
+       if(flags.try_omit_fp) {
+               stack = be_new_IncSP(sp, irg, bl, stack, no_mem, BE_STACK_FRAME_SIZE, be_stack_dir_along);
+               frame = stack;
+       }
+
+       else {
+               frame = be_new_Copy(bp->reg_class, irg, bl, stack);
+
+               be_node_set_flags(frame, -1, arch_irn_flags_dont_spill);
+               if(!flags.fp_free) {
+                       be_set_constr_single_reg(frame, -1, bp);
+                       be_node_set_flags(frame, -1, arch_irn_flags_ignore);
+                       arch_set_irn_register(env->birg->main_env->arch_env, frame, bp);
+               }
+
+               stack = be_new_IncSP(sp, irg, bl, stack, frame, BE_STACK_FRAME_SIZE, be_stack_dir_along);
+       }
+
+       be_node_set_flags(env->reg_params, -(stack_nr + 1), arch_irn_flags_ignore);
+       env->init_sp = stack;
+       set_irg_frame(irg, frame);
+       edges_reroute(old_frame, frame, irg);
+
+       return frame;
+}
+
+static void clearup_frame(be_abi_irg_t *env, ir_node *ret, pmap *reg_map, struct obstack *obst)
 {
        const arch_isa_t *isa = env->birg->main_env->arch_env->isa;
-       ir_node *bl           = get_nodes_block(irn);
-       ir_node *res          = env->init_sp;
-       ir_node *size;
+       const arch_register_t *sp = isa->sp;
+       const arch_register_t *bp = isa->bp;
+       ir_graph *irg      = env->birg->irg;
+       ir_node *ret_mem   = get_Return_mem(ret);
+       ir_node *frame     = get_irg_frame(irg);
+       ir_node *bl        = get_nodes_block(ret);
+       ir_node *stack     = get_irn_link(bl);
+
+       pmap_entry *ent;
+
+       if(env->call->flags.bits.try_omit_fp) {
+               stack = be_new_IncSP(sp, irg, bl, stack, ret_mem, BE_STACK_FRAME_SIZE, be_stack_dir_against);
+       }
+
+       else {
+               stack = be_new_SetSP(sp, irg, bl, stack, frame, ret_mem);
+               be_set_constr_single_reg(stack, -1, sp);
+               be_node_set_flags(stack, -1, arch_irn_flags_ignore);
+       }
+
+       pmap_foreach(env->regs, ent) {
+               const arch_register_t *reg = ent->key;
+               ir_node *irn               = ent->value;
 
-       assert(get_irn_opcode(irn) == iro_Alloc && get_Alloc_where(irn) == stack_alloc);
+               if(reg == sp)
+                       obstack_ptr_grow(&env->obst, stack);
+               else if(reg == bp)
+                       obstack_ptr_grow(&env->obst, frame);
+               else if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
+                       obstack_ptr_grow(obst, irn);
+       }
+}
+
+static ir_type *compute_arg_type(be_abi_irg_t *env, be_abi_call_t *call, ir_type *method_type)
+{
+       int dir  = env->call->flags.bits.left_to_right ? 1 : -1;
+       int inc  = env->birg->main_env->arch_env->isa->stack_dir * dir;
+       int n    = get_method_n_params(method_type);
+       int curr = inc > 0 ? 0 : n - 1;
+       int ofs  = 0;
+
+       char buf[128];
+       ir_type *res;
+       int i;
 
-       size = get_Alloc_size(irn);
-       if(isa->stack_dir > 0)
-               res = be_new_Copy(isa->sp->reg_class, env->birg->irg, bl, res);
+       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));
 
-       res = be_new_AddSP(isa->sp, env->birg->irg, bl, res, size);
-       pset_insert_ptr(env->stack_ops, res);
+       for(i = 0; i < n; ++i, curr += inc) {
+               type *param_type       = get_method_param_type(method_type, curr);
+               be_abi_call_arg_t *arg = get_call_arg(call, 0, curr);
 
-       if(isa->stack_dir < 0)
-               res = be_new_Copy(isa->sp->reg_class, env->birg->irg, bl, res);
+               if(arg->on_stack) {
+                       snprintf(buf, sizeof(buf), "param_%d", i);
+                       arg->stack_ent = new_entity(res, new_id_from_str(buf), param_type);
+                       set_entity_offset_bytes(arg->stack_ent, ofs);
+                       ofs += get_type_size_bytes(param_type);
+               }
+       }
+
+       set_type_size_bytes(res, ofs);
+       return res;
+}
+
+static void create_register_perms(const arch_isa_t *isa, ir_graph *irg, ir_node *bl, pmap *regs)
+{
+       int i, j, n;
+       struct obstack obst;
+
+       obstack_init(&obst);
+
+       /* Create a Perm after the RegParams node to delimit it. */
+       for(i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
+               const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
+               ir_node *perm;
+               ir_node **in;
+               int n_regs;
+
+               for(n_regs = 0, j = 0; j < cls->n_regs; ++j) {
+                       const arch_register_t *reg = &cls->regs[j];
+                       ir_node *irn = pmap_get(regs, (void *) reg);
+
+                       if(irn && !arch_register_type_is(reg, ignore)) {
+                               n_regs++;
+                               obstack_ptr_grow(&obst, irn);
+                               set_irn_link(irn, (void *) reg);
+                       }
+               }
+
+               obstack_ptr_grow(&obst, NULL);
+               in = obstack_finish(&obst);
+               if(n_regs > 0) {
+                       perm = be_new_Perm(cls, irg, bl, n_regs, in);
+                       for(j = 0; j < n_regs; ++j) {
+                               ir_node *arg = in[j];
+                               arch_register_t *reg = get_irn_link(arg);
+                               pmap_insert(regs, reg, arg);
+                               be_set_constr_single_reg(perm, BE_OUT_POS(j), reg);
+                       }
+               }
+               obstack_free(&obst, in);
+       }
 
+       obstack_free(&obst, NULL);
+}
+
+typedef struct {
+       const arch_register_t *reg;
+       ir_node *irn;
+} reg_node_map_t;
+
+static int cmp_regs(const void *a, const void *b)
+{
+       const reg_node_map_t *p = a;
+       const reg_node_map_t *q = b;
+
+       if(p->reg->reg_class == q->reg->reg_class)
+               return p->reg->index - q->reg->index;
+       else
+               return p->reg->reg_class - q->reg->reg_class;
+}
+
+static reg_node_map_t *reg_map_to_arr(struct obstack *obst, pmap *reg_map)
+{
+       pmap_entry *ent;
+       int n = pmap_count(reg_map);
+       int i = 0;
+       reg_node_map_t *res = obstack_alloc(obst, n * sizeof(res[0]));
+
+       pmap_foreach(reg_map, ent) {
+               res[i].reg = ent->key;
+               res[i].irn = ent->value;
+               i++;
+       }
+
+       qsort(res, n, sizeof(res[0]), cmp_regs);
+       return res;
+}
+
+static void 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 i, n;
+       int n_regs = pmap_count(regs);
+       ir_node *irn;
+       ir_node **in;
+       reg_node_map_t *rm;
+
+       rm = reg_map_to_arr(&env->obst, regs);
+
+       for(i = 0, n = 0; i < n_regs; ++i, ++n)
+               obstack_ptr_grow(&env->obst, rm[i].irn);
+
+       if(mem) {
+               obstack_ptr_grow(&env->obst, *mem);
+               n++;
+       }
+
+       in = (ir_node **) obstack_finish(&env->obst);
+       irn = be_new_Barrier(env->birg->irg, bl, n, in);
+       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;
+
+               proj = new_r_Proj(env->birg->irg, bl, irn, get_irn_mode(rm[i].irn), n);
+               be_node_set_reg_class(irn, n, reg->reg_class);
+               if(in_req)
+                       be_set_constr_single_reg(irn, n, reg);
+               be_set_constr_single_reg(irn, pos, reg);
+               be_node_set_reg_class(irn, pos, reg->reg_class);
+               arch_set_irn_register(env->birg->main_env->arch_env, proj, reg);
+               if(arch_register_type_is(reg, ignore))
+                       be_node_set_flags(irn, pos, arch_irn_flags_ignore);
+
+               pmap_insert(regs, (void *) reg, proj);
+       }
+
+       if(mem) {
+               *mem = new_r_Proj(env->birg->irg, bl, irn, mode_M, n);
+       }
+
+       obstack_free(&env->obst, rm);
 }
 
 /**
@@ -369,26 +1064,43 @@ static void implement_stack_alloc(be_abi_irg_t *env, ir_node *irn)
  */
 static void modify_irg(be_abi_irg_t *env)
 {
-       be_abi_call_t *call       = be_abi_call_new();
+       firm_dbg_module_t *dbg    = env->dbg;
+       be_abi_call_t *call       = env->call;
        const arch_isa_t *isa     = env->birg->main_env->arch_env->isa;
+       const arch_register_t *sp = arch_isa_sp(isa);
        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));
+       pset *dont_save           = pset_new_ptr(8);
+       pmap *reg_proj_map        = pmap_create();
        int n_params              = get_method_n_params(method_type);
-       const arch_register_t *sp = arch_isa_sp(isa);
-
-       int max_arg = 0;
-       int reg_params_nr = 0;
-       int n_reg_param   = 0;
+       int max_arg               = 0;
+       int arg_offset            = 0;
 
        int i, j, n;
-       ir_node *proj_sp;
+
+       reg_node_map_t *rm;
+       const arch_register_t *fp_reg;
        ir_node *frame_pointer;
-       ir_node *reg_params, *reg_params_bl;
-       ir_node **args, **args_repl, **return_params;
+       ir_node *reg_params_bl;
+       ir_node **args;
        const ir_edge_t *edge;
+       ir_type *arg_type, *bet_type;
 
+       pmap_entry *ent;
+       bitset_t *used_proj_nr;
+
+       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);
+
+       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)  {
@@ -396,71 +1108,103 @@ static void modify_irg(be_abi_irg_t *env)
                int nr       = get_Proj_proj(irn);
                max_arg      = MAX(max_arg, nr);
        }
-       max_arg += 1;
-       args      = obstack_alloc(&env->obst, max_arg * sizeof(args[0]));
-       args_repl = obstack_alloc(&env->obst, max_arg * sizeof(args[0]));
+       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]));
-       memset(args_repl, 0, max_arg * sizeof(args[0]));
+       used_proj_nr = bitset_alloca(1024);
 
        /* Fill the argument vector */
        foreach_out_edge(arg_tuple, edge) {
                ir_node *irn = get_edge_src_irn(edge);
                int nr       = get_Proj_proj(irn);
                args[nr]     = irn;
+               DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
        }
 
-       /* Get the ABI constraints from the ISA */
-       arch_isa_get_call_abi(isa, method_type, call);
+       arg_type = compute_arg_type(env, call, method_type);
+       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);
 
-       {
-               const arch_register_t **callee_save = env->birg->main_env->callee_save;
-               int n_reg_param = 0;
-
-               /* Count the register params and add them to the number of projs for the RegParams node */
-               for(i = 0; i < n_params; ++i) {
-                       be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
-                       n_reg_param += arg->in_reg;
+       /* Count the register params and add them to the number of Projs for the RegParams node */
+       for(i = 0; i < n_params; ++i) {
+               be_abi_call_arg_t *arg = get_call_arg(call, 0, i);
+               if(arg->in_reg && args[i]) {
+                       assert(arg->reg != sp && "cannot use stack pointer as parameter register");
+                       assert(i == get_Proj_proj(args[i]));
+
+                       /* For now, associate the register with the old Proj from Start representing that argument. */
+                       pmap_insert(env->regs, (void *) arg->reg, args[i]);
+                       bitset_set(used_proj_nr, i);
+                       DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
                }
+       }
 
-               reg_params_bl = get_irg_start_block(irg);
-               reg_params    = be_new_RegParams(irg, reg_params_bl, n_reg_param);
-               reg_params_nr = 0;
-
-               proj_sp       = new_r_Proj(irg, reg_params_bl, reg_params, sp->reg_class->mode, reg_params_nr);
-               be_set_constr_single_reg(reg_params, -(reg_params_nr + 1), sp);
-               reg_params_nr++;
+       /* Collect all callee-save registers */
+       for(i = 0, n = arch_isa_get_n_reg_class(isa); i < n; ++i) {
+               const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
+               for(j = 0; j < cls->n_regs; ++j) {
+                       const arch_register_t *reg = &cls->regs[j];
+                       if(arch_register_type_is(reg, callee_save) || arch_register_type_is(reg, ignore))
+                               pmap_insert(env->regs, (void *) reg, NULL);
+               }
+       }
 
-               /* Make Projs for the callee save registers. */
-               for(i = 0; callee_save[i] != NULL; ++i, ++reg_params_nr) {
-                       const arch_register_t *reg = callee_save[i];
-                       ir_node *irn = new_r_Proj(irg, reg_params_bl, reg_params, reg->reg_class->mode, reg_params_nr);
-                       be_set_constr_single_reg(reg_params, -(reg_params_nr + 1), reg);
+       pmap_insert(env->regs, (void *) sp, NULL);
+       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));
+
+       /*
+        * make proj nodes for the callee save registers.
+        * memorize them, since Return nodes get those as inputs.
+        *
+        * Note, that if a register corresponds to an argument, the regs map contains
+        * the old Proj from start for that argument.
+        */
+
+       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;
+               long nr              = i;
+               int pos              = BE_OUT_POS((int) nr);
+
+               assert(nr >= 0);
+               bitset_set(used_proj_nr, nr);
+               proj = new_r_Proj(irg, reg_params_bl, env->reg_params, mode, nr);
+               pmap_insert(env->regs, (void *) reg, proj);
+               be_set_constr_single_reg(env->reg_params, pos, reg);
+               arch_set_irn_register(env->birg->main_env->arch_env, proj, reg);
 
-                       /* memorize the callee save proj since they will be arguments to the Return nodes. */
-                       obstack_ptr_grow(&env->obst, irn);
-               }
+               /*
+                * If the register is an ignore register,
+                * The Proj for that register shall also be ignored during register allocation.
+                */
+               if(arch_register_type_is(reg, ignore))
+                       be_node_set_flags(env->reg_params, pos, arch_irn_flags_ignore);
 
-               obstack_ptr_grow(&env->obst, NULL);
-               return_params = obstack_finish(&env->obst);
+               DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
        }
+       obstack_free(&env->obst, rm);
 
-       /* If we can omit the frame pointer, the stack pointer will become the frame pointer */
-       env->init_sp = be_new_IncSP(sp, irg, reg_params_bl, proj_sp, 0, be_stack_dir_along);
-       /* memorize this node for later fix up */
-       frame_pointer = env->omit_framepointer ? env->init_sp : be_new_Copy(sp->reg_class, irg, reg_params_bl, proj_sp);
-       set_irg_frame(irg, frame_pointer);
+       /* Generate the Prologue */
+       fp_reg = call->cb->prologue(env->cb, &mem, env->regs);
+       create_barrier(env, bl, &mem, env->regs, 0);
 
-       /* reroute the edges from the original argument projs to the RegParam ones. */
-       for(i = 0; i < max_arg; ++i) {
-               if(args[i]) {
-                       assert(args_repl != NULL);
-                       edges_reroute(args[i], args_repl[i], irg);
-               }
-       }
+       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_along);
+       arch_set_irn_register(env->birg->main_env->arch_env, env->init_sp, sp);
+       be_abi_reg_map_set(env->regs, sp, env->init_sp);
+       frame_pointer = be_abi_reg_map_get(env->regs, sp);
+       set_irg_frame(irg, frame_pointer);
 
        /* Now, introduce stack param nodes for all parameters passed on the stack */
        for(i = 0; i < max_arg; ++i) {
                ir_node *arg_proj = args[i];
+               ir_node *repl     = NULL;
+
                if(arg_proj != NULL) {
                        be_abi_call_arg_t *arg;
                        ir_type *param_type;
@@ -471,33 +1215,123 @@ static void modify_irg(be_abi_irg_t *env)
                        param_type = get_method_param_type(method_type, nr);
 
                        if(arg->in_reg) {
-                               args_repl[i] = new_r_Proj(irg, reg_params_bl, reg_params, get_irn_mode(arg_proj), reg_params_nr);
-                               be_set_constr_single_reg(reg_params, -(reg_params_nr + 1), arg->reg);
-                               reg_params_nr++;
+                               repl = pmap_get(env->regs, (void *) arg->reg);
                        }
 
-                       /* when the (stack) parameter is primitive, we insert a StackParam
-                          node representing the load of that parameter */
-                       else if(is_Primitive_type(param_type)) {
-                               ir_mode *mode                    = get_type_mode(param_type);
-                               const arch_register_class_t *cls = arch_isa_get_reg_class_for_mode(isa, mode);
-                               // TODO: Correct offset computation!
-                               args_repl[i] = be_new_StackParam(cls, irg, reg_params_bl, mode, frame_pointer, 0);
+                       else if(arg->on_stack) {
+                               /* For atomic parameters which are actually used, we create a StackParam node. */
+                               if(is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
+                                       ir_mode *mode                    = get_type_mode(param_type);
+                                       const arch_register_class_t *cls = arch_isa_get_reg_class_for_mode(isa, mode);
+                                       repl = be_new_StackParam(cls, isa->bp->reg_class, irg, reg_params_bl, mode, frame_pointer, arg->stack_ent);
+                               }
+
+                               /* The stack parameter is not primitive (it is a struct or array),
+                               we thus will create a node representing the parameter's address
+                               on the stack. */
+                               else {
+                                       repl = be_new_FrameAddr(sp->reg_class, irg, reg_params_bl, frame_pointer, arg->stack_ent);
+                               }
                        }
 
-                       /* The stack parameter is not primitive (it is a struct or array),
-                          we thus will create a node representing the parameter's address
-                          on the stack. */
-                       else {
-                               assert(0 && "struct parameters are not supported");
+                       assert(repl != NULL);
+                       edges_reroute(args[i], repl, irg);
+               }
+       }
+
+       /* 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(irg, bl, 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);
+                       exchange(irn, ret);
+                       pmap_destroy(reg_map);
                }
        }
 
        obstack_free(&env->obst, args);
-       be_abi_call_free(call);
 }
 
+/**
+ * Walker: puts all Alloc(stack_alloc) on a obstack
+ */
 static void collect_alloca_walker(ir_node *irn, void *data)
 {
        be_abi_irg_t *env = data;
@@ -507,47 +1341,290 @@ static void collect_alloca_walker(ir_node *irn, void *data)
 
 be_abi_irg_t *be_abi_introduce(be_irg_t *birg)
 {
-       be_abi_irg_t *env = malloc(sizeof(env[0]));
+       be_abi_irg_t *env  = xmalloc(sizeof(env[0]));
+       ir_node *old_frame = get_irg_frame(birg->irg);
+       ir_graph *irg      = birg->irg;
 
-       int i;
-       ir_node **stack_allocs;
+       pmap_entry *ent;
+       ir_node *dummy;
+
+       env->isa           = birg->main_env->arch_env->isa;
+       env->method_type   = get_entity_type(get_irg_entity(irg));
+       env->call          = be_abi_call_new();
+       arch_isa_get_call_abi(env->isa, env->method_type, env->call);
+
+       env->dce_survivor     = new_survive_dce();
+       env->birg             = birg;
+       env->dbg              = firm_dbg_register("firm.be.abi");
+       env->stack_phis       = pset_new_ptr(16);
+       env->init_sp = dummy  = new_r_Unknown(irg, env->isa->sp->reg_class->mode);
+
+       env->cb = env->call->cb->init(env->call, env->isa, irg);
 
-       env->birg        = birg;
-       env->stack_ops   = pset_new_ptr(32);
-       env->omit_framepointer = 1;
        obstack_init(&env->obst);
 
-       /* search for stack allocation nodes and record them */
-       irg_walk_graph(env->birg->irg, collect_alloca_walker, NULL, env);
-       obstack_ptr_grow(&env->obst, NULL);
-       stack_allocs = obstack_finish(&env->obst);
+       memcpy(&env->irn_handler, &abi_irn_handler, sizeof(abi_irn_handler));
+       env->irn_ops.impl = &abi_irn_ops;
 
-       /* If there are stack allocations in the irg, we need a frame pointer */
-       if(stack_allocs[0] != NULL)
-               env->omit_framepointer = 0;
+       /* Lower all call nodes in the IRG. */
+       process_calls(env);
 
+       /* Process the IRG */
        modify_irg(env);
 
-       for(i = 0; stack_allocs[i] != NULL; ++i)
-               implement_stack_alloc(env, stack_allocs[i]);
+       /* 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);
 
-       irg_walk_graph(env->birg->irg, NULL, adjust_call_walker, &env);
+       /* 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)
+               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);
        return env;
 }
 
-void be_abi_fix_stack(be_abi_irg_t *env)
+void be_abi_free(be_abi_irg_t *env)
 {
-       pset *origs = pset_new_ptr_default();
+       free_survive_dce(env->dce_survivor);
+       del_pset(env->stack_phis);
+       pmap_destroy(env->regs);
+       obstack_free(&env->obst, NULL);
+       arch_env_pop_irn_handler(env->birg->main_env->arch_env);
+       free(env);
+}
+
+
+/*
+
+  _____ _        ____  _             _
+ |  ___(_)_  __ / ___|| |_ __ _  ___| | __
+ | |_  | \ \/ / \___ \| __/ _` |/ __| |/ /
+ |  _| | |>  <   ___) | || (_| | (__|   <
+ |_|   |_/_/\_\ |____/ \__\__,_|\___|_|\_\
 
-       pset_insert_ptr(origs, env->init_sp);
-       be_ssa_constr_sets(env->birg->dom_front, origs, env->stack_ops);
-       del_pset(origs);
+*/
+
+static void collect_stack_nodes_walker(ir_node *irn, void *data)
+{
+       pset *s = data;
+
+       if(be_is_AddSP(irn)     || be_is_IncSP(irn)     || be_is_SetSP(irn))
+               pset_insert_ptr(s, irn);
 }
 
+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);
+}
 
-void be_abi_free(be_abi_irg_t *env)
+/**
+ * Translates a direction of an IncSP node (either be_stack_dir_against, or ...along)
+ * 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)
 {
-       del_pset(env->stack_ops);
-       obstack_free(&env->obst, NULL);
-       free(env);
+       return 1 - 2 * (be_get_IncSP_direction(irn) == be_stack_dir_against);
 }
+
+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;
+       ir_node *irn;
+       int start_bias = bias;
+       int omit_fp    = env->call->flags.bits.try_omit_fp;
+
+       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;
+               }
+
+               /*
+                       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));
+                       }
+               }
+       }
+
+       return bias;
+}
+
+/**
+ * A helper struct for the bias walker.
+ */
+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. */
+};
+
+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;
+               process_stack_bias(bw->env, bl, bw->start_block_bias);
+       }
+}
+
+void be_abi_fix_stack_bias(be_abi_irg_t *env)
+{
+       ir_graph *irg  = env->birg->irg;
+       struct bias_walk bw;
+
+       stack_frame_compute_initial_offset(env->frame);
+       // stack_frame_dump(stdout, env->frame);
+
+       /* Determine the stack bias at the and 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;
+       irg_block_walk_graph(irg, stack_bias_walker, NULL, &bw);
+}
+
+ir_node *be_abi_get_callee_save_irn(be_abi_irg_t *abi, const arch_register_t *reg)
+{
+       assert(arch_register_type_is(reg, callee_save));
+       assert(pmap_contains(abi->regs, (void *) reg));
+       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)
+{
+       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;
+}
+
+static entity *abi_get_frame_entity(const void *_self, const ir_node *irn)
+{
+       return NULL;
+}
+
+static void abi_set_stack_bias(const void *_self, ir_node *irn, int bias)
+{
+}
+
+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
+};