ia32: we cannot fold ia32_mode_E reloads
[libfirm] / ir / be / benode.c
index 40e5b12..3916616 100644 (file)
+/*
+ * This file is part of libFirm.
+ * Copyright (C) 2012 University of Karlsruhe.
+ */
+
 /**
- * @file   benode.c
- * @date   17.05.2005
- * @author Sebastian Hack
- *
- * Backend node support.
+ * @file
+ * @brief       Backend node support for generic backend nodes.
+ * @author      Sebastian Hack
+ * @date        17.05.2005
  *
- * This file provdies Perm, Copy, Spill and Reload nodes.
- *
- * Copyright (C) 2005 Universitaet Karlsruhe
- * Released under the GPL
+ * Backend node support for generic backend nodes.
+ * This file provides Perm, Copy, Spill and Reload nodes.
  */
-
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include <stdlib.h>
 
+#include "beirg.h"
 #include "obst.h"
 #include "set.h"
 #include "pmap.h"
 #include "util.h"
 #include "debug.h"
 #include "fourcc.h"
+#include "bitfiddle.h"
+#include "raw_bitset.h"
+#include "error.h"
+#include "array_t.h"
 
 #include "irop_t.h"
 #include "irmode_t.h"
 #include "irnode_t.h"
 #include "ircons_t.h"
 #include "irprintf.h"
+#include "irgwalk.h"
+#include "iropt_t.h"
+#include "irbackedge_t.h"
+#include "irverify_t.h"
 
 #include "be_t.h"
 #include "belive_t.h"
-#include "besched_t.h"
-#include "benode_t.h"
+#include "besched.h"
+#include "benode.h"
+#include "bearch.h"
 
 #include "beirgmod.h"
 
-/* Sometimes we want to put const nodes into get_irn_generic_attr ... */
-#define get_irn_attr(irn) get_irn_generic_attr((ir_node *) (irn))
-
-static unsigned be_node_tag = FOURCC('B', 'E', 'N', 'O');
-
-typedef enum _node_kind_t {
-       node_kind_spill,
-       node_kind_reload,
-       node_kind_perm,
-       node_kind_copy,
-       node_kind_kill,
-       node_kind_last
-} node_kind_t;
+typedef struct be_node_attr_t {
+       except_attr  exc;
+} be_node_attr_t;
 
+/** The be_Return nodes attribute type. */
 typedef struct {
-       node_kind_t kind;
-       const arch_register_class_t *cls;
-       ir_op *op;
-       int n_pos;
-       int *pos;
-} be_op_t;
-
-typedef enum {
-       be_req_kind_old_limited,
-       be_req_kind_negate_old_limited,
-       be_req_kind_single_reg
-} be_req_kind_t;
+       be_node_attr_t base;
+       int            num_ret_vals; /**< number of return values */
+       unsigned       pop;          /**< number of bytes that should be popped */
+       int            emit_pop;     /**< if set, emit pop bytes, even if pop = 0 */
+} be_return_attr_t;
 
+/** The be_IncSP attribute type. */
 typedef struct {
-       arch_register_req_t req;
-       be_req_kind_t       kind;
-       union {
-               struct {
-                       void (*old_limited)(void *ptr, bitset_t *bs);
-                       void *old_limited_env;
-               } old_limited;
-
-               const arch_register_t *single_reg;
-       } x;
-} be_req_t;
-
+       be_node_attr_t base;
+       int            offset;    /**< The offset by which the stack shall be
+                                      expanded/shrinked. */
+       int            align;     /**< whether stack should be aligned after the
+                                      IncSP */
+} be_incsp_attr_t;
+
+/** The be_Frame attribute type. */
 typedef struct {
-       const arch_register_t *reg;
-       be_req_t              req;
-       be_req_t              in_req;
-} be_reg_data_t;
+       be_node_attr_t  base;
+       ir_entity      *ent;
+       int             offset;
+} be_frame_attr_t;
 
+/** The be_Call attribute type. */
 typedef struct {
-       int                         n_outs;
-       const arch_register_class_t *cls;
-       be_reg_data_t               *reg_data;
-} be_node_attr_t;
+       be_node_attr_t  base;
+       ir_entity      *ent;        /**< called entity if this is a static call. */
+       unsigned        pop;
+       ir_type        *call_tp;    /**< call type, copied from the original Call */
+} be_call_attr_t;
 
 typedef struct {
-       be_node_attr_t node_attr;
-       ir_node *spill_ctx;  /**< The node in whose context this spill was introduced. */
-       entity *ent;     /**< The entity in the stack frame the spill writes to. */
-} be_spill_attr_t;
-
-static ir_op *op_Spill;
-static ir_op *op_Reload;
-static ir_op *op_Perm;
-static ir_op *op_Copy;
-static ir_op *op_Keep;
-static ir_op *op_Call;
-static ir_op *op_IncSP;
-static ir_op *op_AddSP;
-
-static int beo_base = -1;
-
-static const ir_op_ops be_node_op_ops;
-
-#define N   irop_flag_none
-#define L   irop_flag_labeled
-#define C   irop_flag_commutative
-#define X   irop_flag_cfopcode
-#define I   irop_flag_ip_cfopcode
-#define F   irop_flag_fragile
-#define Y   irop_flag_forking
-#define H   irop_flag_highlevel
-#define c   irop_flag_constlike
-#define K   irop_flag_keep
-
-void be_node_init(void) {
-       static int inited = 0;
-
-       if(inited)
-               return;
+       be_node_attr_t base;
+       ir_entity    **in_entities;
+       ir_entity    **out_entities;
+} be_memperm_attr_t;
+
+ir_op *op_be_Spill;
+ir_op *op_be_Reload;
+ir_op *op_be_Perm;
+ir_op *op_be_MemPerm;
+ir_op *op_be_Copy;
+ir_op *op_be_Keep;
+ir_op *op_be_CopyKeep;
+ir_op *op_be_Call;
+ir_op *op_be_Return;
+ir_op *op_be_IncSP;
+ir_op *op_be_AddSP;
+ir_op *op_be_SubSP;
+ir_op *op_be_Start;
+ir_op *op_be_FrameAddr;
 
-       inited = 1;
+/**
+ * Compare the attributes of two be_FrameAddr nodes.
+ *
+ * @return zero if both nodes have identically attributes
+ */
+static int FrameAddr_cmp_attr(const ir_node *a, const ir_node *b)
+{
+       const be_frame_attr_t *a_attr = (const be_frame_attr_t*)get_irn_generic_attr_const(a);
+       const be_frame_attr_t *b_attr = (const be_frame_attr_t*)get_irn_generic_attr_const(b);
 
-       /* Acquire all needed opcodes. */
-       beo_base = get_next_ir_opcodes(beo_Last - 1);
+       if (a_attr->ent != b_attr->ent || a_attr->offset != b_attr->offset)
+               return 1;
 
-       op_Spill  = new_ir_op(beo_base + beo_Spill,  "Spill",  op_pin_state_mem_pinned, N, oparity_unary,    0, sizeof(be_spill_attr_t), &be_node_op_ops);
-       op_Reload = new_ir_op(beo_base + beo_Reload, "Reload", op_pin_state_mem_pinned, N, oparity_zero,     0, sizeof(be_node_attr_t),  &be_node_op_ops);
-       op_Perm   = new_ir_op(beo_base + beo_Perm,   "Perm",   op_pin_state_pinned,     N, oparity_variable, 0, sizeof(be_node_attr_t),  &be_node_op_ops);
-       op_Copy   = new_ir_op(beo_base + beo_Copy,   "Copy",   op_pin_state_pinned,     N, oparity_unary,    0, sizeof(be_node_attr_t),  &be_node_op_ops);
-       op_Keep   = new_ir_op(beo_base + beo_Keep,   "Keep",   op_pin_state_pinned,     K, oparity_variable, 0, sizeof(be_node_attr_t),  &be_node_op_ops);
+       return be_nodes_equal(a, b);
+}
 
-       set_op_tag(op_Spill,  &be_node_tag);
-       set_op_tag(op_Reload, &be_node_tag);
-       set_op_tag(op_Perm,   &be_node_tag);
-       set_op_tag(op_Copy,   &be_node_tag);
-       set_op_tag(op_Keep,   &be_node_tag);
+/**
+ * Compare the attributes of two be_Return nodes.
+ *
+ * @return zero if both nodes have identically attributes
+ */
+static int Return_cmp_attr(const ir_node *a, const ir_node *b)
+{
+       const be_return_attr_t *a_attr = (const be_return_attr_t*)get_irn_generic_attr_const(a);
+       const be_return_attr_t *b_attr = (const be_return_attr_t*)get_irn_generic_attr_const(b);
+
+       if (a_attr->num_ret_vals != b_attr->num_ret_vals)
+               return 1;
+       if (a_attr->pop != b_attr->pop)
+               return 1;
+       if (a_attr->emit_pop != b_attr->emit_pop)
+               return 1;
+
+       return be_nodes_equal(a, b);
 }
 
-static void *init_node_attr(ir_node* irn, const arch_register_class_t *cls, ir_graph *irg, int n_outs)
+/**
+ * Compare the attributes of two be_IncSP nodes.
+ *
+ * @return zero if both nodes have identically attributes
+ */
+static int IncSP_cmp_attr(const ir_node *a, const ir_node *b)
 {
-       be_node_attr_t *a = get_irn_attr(irn);
+       const be_incsp_attr_t *a_attr = (const be_incsp_attr_t*)get_irn_generic_attr_const(a);
+       const be_incsp_attr_t *b_attr = (const be_incsp_attr_t*)get_irn_generic_attr_const(b);
 
-       a->n_outs   = n_outs;
-       a->cls      = cls;
-       a->reg_data = NULL;
+       if (a_attr->offset != b_attr->offset)
+               return 1;
 
-       if(n_outs > 0) {
-               int i;
+       return be_nodes_equal(a, b);
+}
+
+/**
+ * Compare the attributes of two be_Call nodes.
+ *
+ * @return zero if both nodes have identically attributes
+ */
+static int Call_cmp_attr(const ir_node *a, const ir_node *b)
+{
+       const be_call_attr_t *a_attr = (const be_call_attr_t*)get_irn_generic_attr_const(a);
+       const be_call_attr_t *b_attr = (const be_call_attr_t*)get_irn_generic_attr_const(b);
+
+       if (a_attr->ent != b_attr->ent ||
+               a_attr->call_tp != b_attr->call_tp)
+               return 1;
+
+       return be_nodes_equal(a, b);
+}
+
+static arch_register_req_t *allocate_reg_req(ir_graph *const irg)
+{
+       struct obstack *obst = be_get_be_obst(irg);
+
+       arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
+       return req;
+}
+
+void be_set_constr_in(ir_node *node, int pos, const arch_register_req_t *req)
+{
+       backend_info_t *info = be_get_info(node);
+       assert(pos < get_irn_arity(node));
+       info->in_reqs[pos] = req;
+}
+
+void be_set_constr_out(ir_node *node, int pos, const arch_register_req_t *req)
+{
+       backend_info_t *info = be_get_info(node);
+       info->out_infos[pos].req = req;
+}
+
+/**
+ * Initializes the generic attribute of all be nodes and return it.
+ */
+static void init_node_attr(ir_node *node, int n_inputs, int n_outputs)
+{
+       assert(n_outputs >= 0);
+
+       ir_graph       *irg  = get_irn_irg(node);
+       struct obstack *obst = be_get_be_obst(irg);
+       backend_info_t *info = be_get_info(node);
+       const arch_register_req_t **in_reqs;
 
-               a->reg_data = NEW_ARR_D(be_reg_data_t, get_irg_obstack(irg), n_outs);
-               memset(a->reg_data, 0, n_outs * sizeof(a->reg_data[0]));
-               for(i = 0; i < n_outs; ++i) {
-                       a->reg_data[i].req.req.cls  = cls;
-                       a->reg_data[i].req.req.type = arch_register_req_type_normal;
+       if (n_inputs >= 0) {
+               int i;
+               assert(n_inputs == get_irn_arity(node));
+               in_reqs = OALLOCN(obst, const arch_register_req_t*, n_inputs);
+               for (i = 0; i < n_inputs; ++i) {
+                       in_reqs[i] = arch_no_register_req;
                }
+       } else {
+               in_reqs = NEW_ARR_F(const arch_register_req_t*, 0);
        }
+       info->in_reqs = in_reqs;
 
-       return a;
+       info->out_infos = NEW_ARR_DZ(reg_out_info_t, obst, n_outputs);
+       for (int i = 0; i < n_outputs; ++i) {
+               info->out_infos[i].req = arch_no_register_req;
+       }
 }
 
-static INLINE int is_be_node(const ir_node *irn)
+static void add_register_req_in(ir_node *node, const arch_register_req_t *req)
 {
-       return get_op_tag(get_irn_op(irn)) == &be_node_tag;
+       backend_info_t *info = be_get_info(node);
+       ARR_APP1(const arch_register_req_t*, info->in_reqs, req);
 }
 
-be_opcode_t get_irn_be_opcode(const ir_node *irn)
+ir_node *be_new_Spill(const arch_register_class_t *cls,
+               const arch_register_class_t *cls_frame, ir_node *bl,
+               ir_node *frame, ir_node *to_spill)
 {
-       return is_be_node(irn) ? get_irn_opcode(irn) - beo_base : beo_NoBeOp;
+       be_frame_attr_t *a;
+       ir_node         *in[2];
+       ir_node         *res;
+       ir_graph        *irg = get_Block_irg(bl);
+
+       in[0]     = frame;
+       in[1]     = to_spill;
+       res       = new_ir_node(NULL, irg, bl, op_be_Spill, mode_M, 2, in);
+       init_node_attr(res, 2, 1);
+       a         = (be_frame_attr_t*) get_irn_generic_attr(res);
+       a->ent    = NULL;
+       a->offset = 0;
+       a->base.exc.pin_state = op_pin_state_pinned;
+
+       be_node_set_reg_class_in(res, n_be_Spill_frame, cls_frame);
+       be_node_set_reg_class_in(res, n_be_Spill_val, cls);
+       arch_set_irn_register_req_out(res, 0, arch_no_register_req);
+       arch_add_irn_flags(res, arch_irn_flags_spill);
+
+       return res;
 }
 
-ir_node *be_new_Spill(const arch_register_class_t *cls, ir_graph *irg, ir_node *bl, ir_node *to_spill, ir_node *ctx)
+ir_node *be_new_Reload(const arch_register_class_t *cls,
+               const arch_register_class_t *cls_frame, ir_node *block,
+               ir_node *frame, ir_node *mem, ir_mode *mode)
 {
-       be_spill_attr_t *a;
-       ir_node *in[1];
-       ir_node *res;
+       ir_node  *in[2];
+       ir_node  *res;
+       ir_graph *irg = get_Block_irg(block);
+       be_frame_attr_t *a;
+
+       in[0] = frame;
+       in[1] = mem;
+       res   = new_ir_node(NULL, irg, block, op_be_Reload, mode, 2, in);
+
+       init_node_attr(res, 2, 1);
+       be_node_set_reg_class_out(res, 0, cls);
+
+       be_node_set_reg_class_in(res, n_be_Reload_frame, cls_frame);
+       arch_set_irn_flags(res, arch_irn_flags_rematerializable);
+
+       a         = (be_frame_attr_t*) get_irn_generic_attr(res);
+       a->ent    = NULL;
+       a->offset = 0;
+       a->base.exc.pin_state = op_pin_state_pinned;
 
-       in[0] = to_spill;
-       res   = new_ir_node(NULL, irg, bl, op_Spill, mode_M, 1, in);
-       a     = init_node_attr(res, cls, irg, 0);
-       a->ent       = NULL;
-       a->spill_ctx = ctx;
        return res;
 }
 
-ir_node *be_new_Reload(const arch_register_class_t *cls, ir_graph *irg, ir_node *bl, ir_mode *mode, ir_node *mem)
+ir_node *be_get_Reload_mem(const ir_node *irn)
 {
-       ir_node *in[1];
-       ir_node *res;
+       assert(be_is_Reload(irn));
+       return get_irn_n(irn, n_be_Reload_mem);
+}
 
-       in[0] = mem;
-       res   = new_ir_node(NULL, irg, bl, op_Reload, mode, 1, in);
-       init_node_attr(res, cls, irg, 1);
-       return res;
+ir_node *be_get_Reload_frame(const ir_node *irn)
+{
+       assert(be_is_Reload(irn));
+       return get_irn_n(irn, n_be_Reload_frame);
 }
 
-ir_node *be_new_Perm(const arch_register_class_t *cls, ir_graph *irg, ir_node *bl, int n, ir_node *in[])
+ir_node *be_get_Spill_val(const ir_node *irn)
 {
-       ir_node *irn = new_ir_node(NULL, irg, bl, op_Perm, mode_T, n, in);
-       init_node_attr(irn, cls, irg, n);
+       assert(be_is_Spill(irn));
+       return get_irn_n(irn, n_be_Spill_val);
+}
+
+ir_node *be_get_Spill_frame(const ir_node *irn)
+{
+       assert(be_is_Spill(irn));
+       return get_irn_n(irn, n_be_Spill_frame);
+}
+
+ir_node *be_new_Perm(const arch_register_class_t *cls, ir_node *block,
+                     int n, ir_node *in[])
+{
+       int            i;
+       ir_graph       *irg = get_Block_irg(block);
+       be_node_attr_t *attr;
+
+       ir_node *irn = new_ir_node(NULL, irg, block, op_be_Perm, mode_T, n, in);
+       init_node_attr(irn, n, n);
+       attr                = (be_node_attr_t*) get_irn_generic_attr(irn);
+       attr->exc.pin_state = op_pin_state_pinned;
+       for (i = 0; i < n; ++i) {
+               const ir_node             *input = in[i];
+               const arch_register_req_t *req   = arch_get_irn_register_req(input);
+               if (req->width == 1) {
+                       be_set_constr_in(irn, i, cls->class_req);
+                       be_set_constr_out(irn, i, cls->class_req);
+               } else {
+                       arch_register_req_t *const new_req = allocate_reg_req(irg);
+                       new_req->cls   = cls;
+                       new_req->type  = (req->type & arch_register_req_type_aligned);
+                       new_req->width = req->width;
+                       be_set_constr_in(irn, i, new_req);
+                       be_set_constr_out(irn, i, new_req);
+               }
+       }
+
+       return irn;
+}
+
+void be_Perm_reduce(ir_node *perm, int new_size, int *map)
+{
+       int             arity      = get_irn_arity(perm);
+       const arch_register_req_t **old_in_reqs
+               = ALLOCAN(const arch_register_req_t*, arity);
+       reg_out_info_t  *old_infos = ALLOCAN(reg_out_info_t, arity);
+       backend_info_t  *info      = be_get_info(perm);
+       ir_node        **new_in;
+       int              i;
+
+       assert(be_is_Perm(perm));
+       assert(new_size <= arity);
+
+       new_in = ALLOCAN(ir_node*, new_size);
+
+       /* save the old register data */
+       memcpy(old_in_reqs, info->in_reqs, arity * sizeof(old_in_reqs[0]));
+       memcpy(old_infos, info->out_infos, arity * sizeof(old_infos[0]));
+
+       /* compose the new in array and set the new register data directly */
+       for (i = 0; i < new_size; ++i) {
+               int idx = map[i];
+               new_in[i]          = get_irn_n(perm, idx);
+               info->in_reqs[i]   = old_in_reqs[idx];
+               info->out_infos[i] = old_infos[idx];
+       }
+
+       set_irn_in(perm, new_size, new_in);
+}
+
+ir_node *be_new_MemPerm(ir_node *block, int n, ir_node *in[])
+{
+       ir_graph                     *irg       = get_Block_irg(block);
+       const arch_env_t             *arch_env  = be_get_irg_arch_env(irg);
+       ir_node                      *frame     = get_irg_frame(irg);
+       const arch_register_t        *sp        = arch_env->sp;
+       ir_node                      *irn;
+       be_memperm_attr_t            *attr;
+       ir_node                     **real_in;
+
+       real_in = ALLOCAN(ir_node*, n + 1);
+       real_in[0] = frame;
+       memcpy(&real_in[1], in, n * sizeof(real_in[0]));
+
+       irn = new_ir_node(NULL, irg, block, op_be_MemPerm, mode_T, n+1, real_in);
+
+       init_node_attr(irn, n + 1, n);
+       be_node_set_reg_class_in(irn, 0, sp->reg_class);
+
+       attr               = (be_memperm_attr_t*)get_irn_generic_attr(irn);
+       attr->in_entities  = OALLOCNZ(get_irg_obstack(irg), ir_entity*, n);
+       attr->out_entities = OALLOCNZ(get_irg_obstack(irg), ir_entity*, n);
+
        return irn;
 }
 
-ir_node *be_new_Copy(const arch_register_class_t *cls, ir_graph *irg, ir_node *bl, ir_node *op)
+ir_node *be_new_Copy(ir_node *bl, ir_node *op)
 {
        ir_node *in[1];
        ir_node *res;
+       be_node_attr_t *attr;
+       ir_graph *irg = get_Block_irg(bl);
+       const arch_register_req_t   *in_req = arch_get_irn_register_req(op);
+       const arch_register_class_t *cls    = in_req->cls;
 
        in[0] = op;
-       res   = new_ir_node(NULL, irg, bl, op_Copy, get_irn_mode(op), 1, in);
-       init_node_attr(res, cls, irg, 1);
+       res   = new_ir_node(NULL, irg, bl, op_be_Copy, get_irn_mode(op), 1, in);
+       init_node_attr(res, 1, 1);
+       attr = (be_node_attr_t*) get_irn_generic_attr(res);
+       attr->exc.pin_state = op_pin_state_floats;
+       be_node_set_reg_class_in(res, 0, cls);
+       be_node_set_reg_class_out(res, 0, cls);
+
+       arch_register_req_t *const req = allocate_reg_req(irg);
+       req->cls        = cls;
+       req->type       = arch_register_req_type_should_be_same
+               | (in_req->type & arch_register_req_type_aligned);
+       req->other_same = 1U << 0;
+       req->width      = in_req->width;
+       be_set_constr_out(res, 0, req);
+
        return res;
 }
 
-ir_node *be_new_Keep(const arch_register_class_t *cls, ir_graph *irg, ir_node *bl, int n, ir_node *in[])
+ir_node *be_get_Copy_op(const ir_node *cpy)
 {
-       ir_node *irn;
+       return get_irn_n(cpy, n_be_Copy_op);
+}
 
-       irn = new_ir_node(NULL, irg, bl, op_Keep, mode_ANY, n, in);
-       init_node_attr(irn, cls, irg, 0);
-       keep_alive(irn);
+ir_node *be_new_Keep(ir_node *block, int n, ir_node *in[])
+{
+       int i;
+       ir_node *res;
+       ir_graph *irg = get_Block_irg(block);
+       be_node_attr_t *attr;
+
+       res = new_ir_node(NULL, irg, block, op_be_Keep, mode_ANY, -1, NULL);
+       init_node_attr(res, -1, 1);
+       attr = (be_node_attr_t*) get_irn_generic_attr(res);
+       attr->exc.pin_state = op_pin_state_pinned;
+
+       for (i = 0; i < n; ++i) {
+               ir_node *pred = in[i];
+               add_irn_n(res, pred);
+               const arch_register_req_t *req = arch_get_irn_register_req(pred);
+               req = req->cls != NULL ? req->cls->class_req : arch_no_register_req;
+               add_register_req_in(res, req);
+       }
+       keep_alive(res);
+
+       return res;
+}
+
+void be_Keep_add_node(ir_node *keep, const arch_register_class_t *cls, ir_node *node)
+{
+       assert(be_is_Keep(keep));
+       add_irn_n(keep, node);
+       add_register_req_in(keep, cls->class_req);
+}
+
+ir_node *be_new_Call(dbg_info *dbg, ir_node *bl, ir_node *mem,
+               const arch_register_req_t *sp_req, ir_node *sp,
+               const arch_register_req_t *ptr_req, ir_node *ptr,
+               int n_outs, int n, ir_node *in[], ir_type *call_tp)
+{
+       be_call_attr_t *a;
+       int real_n = n_be_Call_first_arg + n;
+       ir_node **real_in;
+
+       NEW_ARR_A(ir_node *, real_in, real_n);
+       real_in[n_be_Call_mem] = mem;
+       real_in[n_be_Call_sp]  = sp;
+       real_in[n_be_Call_ptr] = ptr;
+       memcpy(&real_in[n_be_Call_first_arg], in, n * sizeof(in[0]));
+
+       ir_graph *const irg = get_Block_irg(bl);
+       ir_node  *const irn = new_ir_node(dbg, irg, bl, op_be_Call, mode_T, real_n, real_in);
+       init_node_attr(irn, real_n, n_outs);
+       a                     = (be_call_attr_t*)get_irn_generic_attr(irn);
+       a->ent                = NULL;
+       a->call_tp            = call_tp;
+       a->pop                = 0;
+       a->base.exc.pin_state = op_pin_state_pinned;
+       be_set_constr_in(irn, n_be_Call_sp, sp_req);
+       be_set_constr_in(irn, n_be_Call_ptr, ptr_req);
        return irn;
 }
 
-int be_is_Spill(const ir_node *irn)
+ir_entity *be_Call_get_entity(const ir_node *call)
 {
-       return get_irn_be_opcode(irn) == beo_Spill;
+       const be_call_attr_t *a = (const be_call_attr_t*)get_irn_generic_attr_const(call);
+       assert(be_is_Call(call));
+       return a->ent;
 }
 
-int be_is_Reload(const ir_node *irn)
+void be_Call_set_entity(ir_node *call, ir_entity *ent)
 {
-       return get_irn_be_opcode(irn) == beo_Reload;
+       be_call_attr_t *a = (be_call_attr_t*)get_irn_generic_attr(call);
+       assert(be_is_Call(call));
+       a->ent = ent;
 }
 
-int be_is_Copy(const ir_node *irn)
+ir_type *be_Call_get_type(ir_node *call)
 {
-       return get_irn_be_opcode(irn) == beo_Copy;
+       const be_call_attr_t *a = (const be_call_attr_t*)get_irn_generic_attr_const(call);
+       assert(be_is_Call(call));
+       return a->call_tp;
 }
 
-int be_is_Perm(const ir_node *irn)
+void be_Call_set_type(ir_node *call, ir_type *call_tp)
 {
-       return get_irn_be_opcode(irn) == beo_Perm;
+       be_call_attr_t *a = (be_call_attr_t*)get_irn_generic_attr(call);
+       assert(be_is_Call(call));
+       a->call_tp = call_tp;
 }
 
-int be_is_Keep(const ir_node *irn)
+void be_Call_set_pop(ir_node *call, unsigned pop)
 {
-       return get_irn_be_opcode(irn) == beo_Keep;
+       be_call_attr_t *a = (be_call_attr_t*)get_irn_generic_attr(call);
+       a->pop = pop;
 }
 
-static void be_limited(void *data, bitset_t *bs)
+unsigned be_Call_get_pop(const ir_node *call)
 {
-       be_req_t *req = data;
+       const be_call_attr_t *a = (const be_call_attr_t*)get_irn_generic_attr_const(call);
+       return a->pop;
+}
 
-       switch(req->kind) {
-       case be_req_kind_negate_old_limited:
-       case be_req_kind_old_limited:
-               req->x.old_limited.old_limited(req->x.old_limited.old_limited_env, bs);
-               if(req->kind == be_req_kind_negate_old_limited)
-                       bitset_flip_all(bs);
-               break;
-       case be_req_kind_single_reg:
-               bitset_clear_all(bs);
-               bitset_set(bs, req->x.single_reg->index);
-               break;
-       }
+ir_node *be_new_Return(dbg_info *const dbg, ir_node *const block, int const n_res, unsigned const pop, int const n, ir_node **const in)
+{
+       ir_graph *const irg = get_Block_irg(block);
+       ir_node  *const res = new_ir_node(dbg, irg, block, op_be_Return, mode_X, n, in);
+       init_node_attr(res, n, 1);
+       be_set_constr_out(res, 0, arch_no_register_req);
+
+       be_return_attr_t *const a = (be_return_attr_t*)get_irn_generic_attr(res);
+       a->num_ret_vals       = n_res;
+       a->pop                = pop;
+       a->emit_pop           = 0;
+       a->base.exc.pin_state = op_pin_state_pinned;
+
+       return res;
 }
 
-void be_set_constr_single_reg(ir_node *irn, int pos, const arch_register_t *reg)
+int be_Return_get_n_rets(const ir_node *ret)
 {
-       int idx           = pos < 0 ? -(pos - 1) : pos;
-       be_node_attr_t *a = get_irn_attr(irn);
-       be_reg_data_t *rd = &a->reg_data[idx];
-       be_req_t       *r = pos < 0 ? &rd->req : &rd->in_req;
+       const be_return_attr_t *a = (const be_return_attr_t*)get_irn_generic_attr_const(ret);
+       return a->num_ret_vals;
+}
 
-       assert(is_be_node(irn));
-       assert(!(pos >= 0) || pos < get_irn_arity(irn));
-       assert(!(pos < 0)  || -(pos + 1) <= a->n_outs);
+unsigned be_Return_get_pop(const ir_node *ret)
+{
+       const be_return_attr_t *a = (const be_return_attr_t*)get_irn_generic_attr_const(ret);
+       return a->pop;
+}
 
-       r->kind            = be_req_kind_single_reg;
-       r->x.single_reg    = reg;
-       r->req.limited     = be_limited;
-       r->req.limited_env = r;
-       r->req.type        = arch_register_req_type_limited;
-       r->req.cls         = reg->reg_class;
+int be_Return_get_emit_pop(const ir_node *ret)
+{
+       const be_return_attr_t *a = (const be_return_attr_t*)get_irn_generic_attr_const(ret);
+       return a->emit_pop;
 }
 
-void be_set_constr_limited(ir_node *irn, int pos, const arch_register_req_t *req)
+void be_Return_set_emit_pop(ir_node *ret, int emit_pop)
 {
-       int idx           = pos < 0 ? -(pos - 1) : pos;
-       be_node_attr_t *a = get_irn_attr(irn);
-       be_reg_data_t *rd = &a->reg_data[idx];
-       be_req_t       *r = pos < 0 ? &rd->req : &rd->in_req;
+       be_return_attr_t *a = (be_return_attr_t*)get_irn_generic_attr(ret);
+       a->emit_pop = emit_pop;
+}
 
-       assert(is_be_node(irn));
-       assert(!(pos >= 0) || pos < get_irn_arity(irn));
-       assert(!(pos < 0)  || -(pos + 1) <= a->n_outs);
-       assert(arch_register_req_is(req, limited));
+ir_node *be_new_IncSP(const arch_register_t *sp, ir_node *bl,
+                      ir_node *old_sp, int offset, int align)
+{
+       be_incsp_attr_t *a;
+       ir_node *irn;
+       ir_node *in[1];
+       ir_graph *irg = get_Block_irg(bl);
+
+       in[0]     = old_sp;
+       irn       = new_ir_node(NULL, irg, bl, op_be_IncSP, sp->reg_class->mode,
+                               ARRAY_SIZE(in), in);
+       init_node_attr(irn, 1, 1);
+       a                     = (be_incsp_attr_t*)get_irn_generic_attr(irn);
+       a->offset             = offset;
+       a->align              = align;
+       a->base.exc.pin_state = op_pin_state_pinned;
 
-       r->kind            = be_req_kind_old_limited;
-       r->req.limited     = be_limited;
-       r->req.limited_env = r;
-       r->req.type        = arch_register_req_type_limited;
-       r->req.cls         = req->cls;
+       /* Set output constraint to stack register. */
+       be_node_set_reg_class_in(irn, 0, sp->reg_class);
+       be_set_constr_single_reg_out(irn, 0, sp, arch_register_req_type_produces_sp);
 
-       r->x.old_limited.old_limited     = req->limited;
-       r->x.old_limited.old_limited_env = req->limited_env;
+       return irn;
 }
 
+ir_node *be_new_AddSP(const arch_register_t *sp, ir_node *bl, ir_node *old_sp,
+               ir_node *sz)
+{
+       ir_node *irn;
+       ir_node *in[n_be_AddSP_last];
+       ir_graph *irg;
+       be_node_attr_t *attr;
+
+       in[n_be_AddSP_old_sp] = old_sp;
+       in[n_be_AddSP_size]   = sz;
+
+       irg = get_Block_irg(bl);
+       irn = new_ir_node(NULL, irg, bl, op_be_AddSP, mode_T, n_be_AddSP_last, in);
+       init_node_attr(irn, n_be_AddSP_last, pn_be_AddSP_last);
+       attr = (be_node_attr_t*) get_irn_generic_attr(irn);
+       attr->exc.pin_state = op_pin_state_pinned;
+
+       /* Set output constraint to stack register. */
+       be_set_constr_single_reg_in(irn, n_be_AddSP_old_sp, sp,
+                                   arch_register_req_type_none);
+       be_node_set_reg_class_in(irn, n_be_AddSP_size, sp->reg_class);
+       be_set_constr_single_reg_out(irn, pn_be_AddSP_sp, sp,
+                                    arch_register_req_type_produces_sp);
 
-void be_set_IncSP_offset(ir_node *irn, int offset)
+       return irn;
+}
+
+ir_node *be_new_SubSP(const arch_register_t *sp, ir_node *bl, ir_node *old_sp, ir_node *sz)
 {
+       ir_node *irn;
+       ir_node *in[n_be_SubSP_last];
+       ir_graph *irg;
+       be_node_attr_t *attr;
+
+       in[n_be_SubSP_old_sp] = old_sp;
+       in[n_be_SubSP_size]   = sz;
+
+       irg = get_Block_irg(bl);
+       irn = new_ir_node(NULL, irg, bl, op_be_SubSP, mode_T, n_be_SubSP_last, in);
+       init_node_attr(irn, n_be_SubSP_last, pn_be_SubSP_last);
+       attr = (be_node_attr_t*) get_irn_generic_attr(irn);
+       attr->exc.pin_state = op_pin_state_pinned;
 
+       /* Set output constraint to stack register. */
+       be_set_constr_single_reg_in(irn, n_be_SubSP_old_sp, sp,
+                                   arch_register_req_type_none);
+       be_node_set_reg_class_in(irn, n_be_SubSP_size, sp->reg_class);
+       be_set_constr_single_reg_out(irn, pn_be_SubSP_sp, sp, arch_register_req_type_produces_sp);
+
+       return irn;
 }
 
-int be_get_IncSP_offset(ir_node *irn)
+ir_node *be_new_Start(dbg_info *dbgi, ir_node *bl, int n_outs)
 {
-       return -1;
+       ir_node *res;
+       ir_graph *irg = get_Block_irg(bl);
+       be_node_attr_t *attr;
+
+       res = new_ir_node(dbgi, irg, bl, op_be_Start, mode_T, 0, NULL);
+       init_node_attr(res, 0, n_outs);
+       attr = (be_node_attr_t*) get_irn_generic_attr(res);
+       attr->exc.pin_state = op_pin_state_pinned;
+
+       return res;
 }
 
-void be_set_Spill_entity(ir_node *irn, entity *ent)
+ir_node *be_new_FrameAddr(const arch_register_class_t *cls_frame, ir_node *bl, ir_node *frame, ir_entity *ent)
 {
-       be_spill_attr_t *a = get_irn_attr(irn);
-       assert(be_is_Spill(irn));
-       a->ent = ent;
+       be_frame_attr_t *a;
+       ir_node *irn;
+       ir_node *in[1];
+       ir_graph *irg = get_Block_irg(bl);
+
+       in[0]  = frame;
+       irn    = new_ir_node(NULL, irg, bl, op_be_FrameAddr, get_irn_mode(frame), 1, in);
+       init_node_attr(irn, 1, 1);
+       a                     = (be_frame_attr_t*)get_irn_generic_attr(irn);
+       a->ent                = ent;
+       a->offset             = 0;
+       a->base.exc.pin_state = op_pin_state_floats;
+       be_node_set_reg_class_in(irn, 0, cls_frame);
+       be_node_set_reg_class_out(irn, 0, cls_frame);
+
+       return optimize_node(irn);
 }
 
-static ir_node *find_a_spill_walker(ir_node *irn, unsigned visited_nr)
+ir_node *be_get_FrameAddr_frame(const ir_node *node)
 {
-       if(get_irn_visited(irn) < visited_nr) {
-               set_irn_visited(irn, visited_nr);
+       assert(be_is_FrameAddr(node));
+       return get_irn_n(node, n_be_FrameAddr_ptr);
+}
 
-               if(is_Phi(irn)) {
-                       int i, n;
-                       for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
-                               ir_node *n = find_a_spill_walker(get_irn_n(irn, i), visited_nr);
-                               if(n != NULL)
-                                       return n;
-                       }
-               }
+ir_entity *be_get_FrameAddr_entity(const ir_node *node)
+{
+       const be_frame_attr_t *attr = (const be_frame_attr_t*)get_irn_generic_attr_const(node);
+       return attr->ent;
+}
 
-               else if(get_irn_be_opcode(irn) == beo_Spill)
-                       return irn;
+ir_node *be_new_CopyKeep(ir_node *bl, ir_node *src, int n, ir_node *in_keep[])
+{
+       ir_node  *irn;
+       ir_node **in = ALLOCAN(ir_node*, n + 1);
+       ir_graph *irg = get_Block_irg(bl);
+       const arch_register_req_t   *req  = arch_get_irn_register_req(src);
+       const arch_register_class_t *cls  = req->cls;
+       ir_mode                     *mode = get_irn_mode(src);
+       be_node_attr_t *attr;
+
+       in[0] = src;
+       memcpy(&in[1], in_keep, n * sizeof(in[0]));
+       irn   = new_ir_node(NULL, irg, bl, op_be_CopyKeep, mode, n + 1, in);
+       init_node_attr(irn, n + 1, 1);
+       attr = (be_node_attr_t*) get_irn_generic_attr(irn);
+       attr->exc.pin_state = op_pin_state_floats;
+       be_node_set_reg_class_in(irn, 0, cls);
+       be_node_set_reg_class_out(irn, 0, cls);
+       for (int i = 0; i < n; ++i) {
+               ir_node *pred = in_keep[i];
+               const arch_register_req_t *req = arch_get_irn_register_req(pred);
+               req = req->cls != NULL ? req->cls->class_req : arch_no_register_req;
+               be_set_constr_in(irn, i+1, req);
        }
 
-       return NULL;
+       return irn;
 }
 
-ir_node *be_get_Spill_context(const ir_node *irn) {
-       const be_spill_attr_t *a = get_irn_attr(irn);
-       assert(be_is_Spill(irn));
-       return a->spill_ctx;
+ir_node *be_new_CopyKeep_single(ir_node *bl, ir_node *src, ir_node *keep)
+{
+       return be_new_CopyKeep(bl, src, 1, &keep);
 }
 
-/**
- * Finds a spill for a reload.
- * If the reload is directly using the spill, this is simple,
- * else we perform DFS from the reload (over all PhiMs) and return
- * the first spill node we find.
- */
-static INLINE ir_node *find_a_spill(ir_node *irn)
+ir_node *be_get_CopyKeep_op(const ir_node *cpy)
 {
-       ir_graph *irg       = get_irn_irg(irn);
-       unsigned visited_nr = get_irg_visited(irg) + 1;
-
-       assert(be_is_Reload(irn));
-       set_irg_visited(irg, visited_nr);
-       return find_a_spill_walker(irn, visited_nr);
+       return get_irn_n(cpy, n_be_CopyKeep_op);
 }
 
-entity *be_get_spill_entity(ir_node *irn)
+void be_set_CopyKeep_op(ir_node *cpy, ir_node *op)
 {
-       int opc           = get_irn_opcode(irn);
+       set_irn_n(cpy, n_be_CopyKeep_op, op);
+}
 
-       switch(get_irn_be_opcode(irn)) {
-       case beo_Reload:
-               return be_get_spill_entity(find_a_spill(irn));
+static bool be_has_frame_entity(const ir_node *irn)
+{
+       switch (get_irn_opcode(irn)) {
        case beo_Spill:
-               {
-                       be_spill_attr_t *a = get_irn_attr(irn);
-                       return a->ent;
-               }
+       case beo_Reload:
+       case beo_FrameAddr:
+               return true;
        default:
-               assert(0 && "Must give spill/reload node");
+               return false;
        }
+}
 
+ir_entity *be_get_frame_entity(const ir_node *irn)
+{
+       if (be_has_frame_entity(irn)) {
+               const be_frame_attr_t *a = (const be_frame_attr_t*)get_irn_generic_attr_const(irn);
+               return a->ent;
+       }
        return NULL;
 }
 
-ir_node *be_spill(const arch_env_t *arch_env, ir_node *irn, ir_node *ctx)
+int be_get_frame_offset(const ir_node *irn)
 {
-       const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
+       assert(is_be_node(irn));
+       if (be_has_frame_entity(irn)) {
+               const be_frame_attr_t *a = (const be_frame_attr_t*)get_irn_generic_attr_const(irn);
+               return a->offset;
+       }
+       return 0;
+}
 
-       ir_node *bl    = get_nodes_block(irn);
-       ir_graph *irg  = get_irn_irg(bl);
-       ir_node *spill = be_new_Spill(cls, irg, bl, irn, ctx);
-       ir_node *insert;
+void be_set_MemPerm_in_entity(const ir_node *irn, int n, ir_entity *ent)
+{
+       const be_memperm_attr_t *attr = (const be_memperm_attr_t*)get_irn_generic_attr_const(irn);
 
-       /*
-        * search the right insertion point. a spill of a phi cannot be put
-        * directly after the phi, if there are some phis behind the one which
-        * is spilled.
-        */
-       insert = sched_next(irn);
-       while(is_Phi(insert) && !sched_is_end(insert))
-               insert = sched_next(insert);
+       assert(be_is_MemPerm(irn));
+       assert(n < be_get_MemPerm_entity_arity(irn));
 
-       sched_add_before(insert, spill);
-       return spill;
+       attr->in_entities[n] = ent;
 }
 
-ir_node *be_reload(const arch_env_t *arch_env,
-                                  const arch_register_class_t *cls,
-                                  ir_node *irn, int pos, ir_mode *mode, ir_node *spill)
+ir_entity* be_get_MemPerm_in_entity(const ir_node* irn, int n)
 {
-       ir_node *reload;
-
-       ir_node *bl   = get_nodes_block(irn);
-       ir_graph *irg = get_irn_irg(bl);
-
-       assert(be_is_Spill(spill) || (is_Phi(spill) && get_irn_mode(spill) == mode_M));
+       const be_memperm_attr_t *attr = (const be_memperm_attr_t*)get_irn_generic_attr_const(irn);
 
-       reload = be_new_Reload(cls, irg, bl, mode, spill);
+       assert(be_is_MemPerm(irn));
+       assert(n < be_get_MemPerm_entity_arity(irn));
 
-       set_irn_n(irn, pos, reload);
-       sched_add_before(irn, reload);
-       return reload;
+       return attr->in_entities[n];
 }
 
-static int redir_proj(const ir_node **node, int pos)
+void be_set_MemPerm_out_entity(const ir_node *irn, int n, ir_entity *ent)
 {
-       const ir_node *n = *node;
+       const be_memperm_attr_t *attr = (const be_memperm_attr_t*)get_irn_generic_attr_const(irn);
 
-       if(is_Proj(n)) {
-               assert(pos == -1 && "Illegal pos for a Proj");
-               *node = get_Proj_pred(n);
-               return get_Proj_proj(n);
-       }
+       assert(be_is_MemPerm(irn));
+       assert(n < be_get_MemPerm_entity_arity(irn));
 
-       return 0;
+       attr->out_entities[n] = ent;
 }
 
-static void *put_out_reg_req(arch_register_req_t *req, const ir_node *irn, int out_pos)
+ir_entity* be_get_MemPerm_out_entity(const ir_node* irn, int n)
 {
-       const be_node_attr_t *a = get_irn_attr(irn);
+       const be_memperm_attr_t *attr = (const be_memperm_attr_t*)get_irn_generic_attr_const(irn);
 
-       if(out_pos < a->n_outs)
-               memcpy(req, &a->reg_data[out_pos].req, sizeof(req[0]));
-       else {
-               req->type = arch_register_req_type_none;
-               req->cls  = NULL;
-       }
+       assert(be_is_MemPerm(irn));
+       assert(n < be_get_MemPerm_entity_arity(irn));
 
-       return req;
+       return attr->out_entities[n];
 }
 
-static void *put_in_reg_req(arch_register_req_t *req, const ir_node *irn, int pos)
+int be_get_MemPerm_entity_arity(const ir_node *irn)
 {
-       const be_node_attr_t *a = get_irn_attr(irn);
-       int n                   = get_irn_arity(irn);
+       return get_irn_arity(irn) - 1;
+}
 
-       if(pos < get_irn_arity(irn))
-               memcpy(req, &a->reg_data[pos].in_req, sizeof(req[0]));
-       else {
-               req->type = arch_register_req_type_none;
-               req->cls  = NULL;
-       }
+const arch_register_req_t *be_create_reg_req(struct obstack *obst,
+               const arch_register_t *reg, arch_register_req_type_t additional_types)
+{
+       arch_register_req_t         *req = OALLOC(obst, arch_register_req_t);
+       const arch_register_class_t *cls = reg->reg_class;
+       unsigned                    *limited_bitset;
 
+       limited_bitset = rbitset_obstack_alloc(obst, arch_register_class_n_regs(cls));
+       rbitset_set(limited_bitset, reg->index);
+
+       req->type    = arch_register_req_type_limited | additional_types;
+       req->cls     = cls;
+       req->limited = limited_bitset;
+       req->width   = 1;
        return req;
 }
 
-static const arch_register_req_t *
-be_node_get_irn_reg_req(const void *self, arch_register_req_t *req, const ir_node *irn, int pos)
+void be_set_constr_single_reg_in(ir_node *node, int pos,
+               const arch_register_t *reg, arch_register_req_type_t additional_types)
 {
-       int out_pos = pos;
+       const arch_register_req_t *req;
+
+       if (additional_types == 0) {
+               req = reg->single_req;
+       } else {
+               ir_graph       *irg  = get_irn_irg(node);
+               struct obstack *obst = be_get_be_obst(irg);
+               req = be_create_reg_req(obst, reg, additional_types);
+       }
+       be_set_constr_in(node, pos, req);
+}
 
-       if(pos < 0) {
-               if(get_irn_mode(irn) == mode_T)
-                       return NULL;
+void be_set_constr_single_reg_out(ir_node *node, int pos,
+               const arch_register_t *reg, arch_register_req_type_t additional_types)
+{
+       ir_graph                  *irg  = get_irn_irg(node);
+       be_irg_t                  *birg = be_birg_from_irg(irg);
+       const arch_register_req_t *req;
 
-               out_pos = redir_proj((const ir_node **) &irn, pos);
-               assert(is_be_node(irn));
-               return put_out_reg_req(req, irn, out_pos);
+       /* if we have an ignore register, add ignore flag and just assign it */
+       if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
+               additional_types |= arch_register_req_type_ignore;
        }
 
-       else {
-               return is_be_node(irn) ? put_in_reg_req(req, irn, pos) : NULL;
+       if (additional_types == 0) {
+               req = reg->single_req;
+       } else {
+               struct obstack *obst = be_get_be_obst(irg);
+               req = be_create_reg_req(obst, reg, additional_types);
        }
 
-       return req;
+       arch_set_irn_register_out(node, pos, reg);
+       be_set_constr_out(node, pos, req);
 }
 
-static void
-be_node_set_irn_reg(const void *_self, ir_node *irn, const arch_register_t *reg)
+void be_node_set_reg_class_in(ir_node *irn, int pos,
+                              const arch_register_class_t *cls)
 {
-       int out_pos;
-       be_node_attr_t *a;
+       be_set_constr_in(irn, pos, cls->class_req);
+}
 
-       out_pos = redir_proj((const ir_node **) &irn, -1);
-       a       = get_irn_attr(irn);
+void be_node_set_reg_class_out(ir_node *irn, int pos,
+                               const arch_register_class_t *cls)
+{
+       be_set_constr_out(irn, pos, cls->class_req);
+}
 
-       assert(is_be_node(irn));
-       assert(out_pos < a->n_outs && "position too high");
-       a->reg_data[out_pos].reg = reg;
+ir_node *be_get_IncSP_pred(ir_node *irn)
+{
+       assert(be_is_IncSP(irn));
+       return get_irn_n(irn, 0);
 }
 
-const arch_register_t *
-be_node_get_irn_reg(const void *_self, const ir_node *irn)
+void be_set_IncSP_pred(ir_node *incsp, ir_node *pred)
+{
+       assert(be_is_IncSP(incsp));
+       set_irn_n(incsp, 0, pred);
+}
+
+void be_set_IncSP_offset(ir_node *irn, int offset)
 {
-       int out_pos;
-       be_node_attr_t *a;
+       be_incsp_attr_t *a = (be_incsp_attr_t*)get_irn_generic_attr(irn);
+       assert(be_is_IncSP(irn));
+       a->offset = offset;
+}
 
-       out_pos = redir_proj((const ir_node **) &irn, -1);
-       a       = get_irn_attr(irn);
+int be_get_IncSP_offset(const ir_node *irn)
+{
+       const be_incsp_attr_t *a = (const be_incsp_attr_t*)get_irn_generic_attr_const(irn);
+       assert(be_is_IncSP(irn));
+       return a->offset;
+}
 
-       assert(is_be_node(irn));
-       assert(out_pos < a->n_outs && "position too high");
+int be_get_IncSP_align(const ir_node *irn)
+{
+       const be_incsp_attr_t *a = (const be_incsp_attr_t*)get_irn_generic_attr_const(irn);
+       assert(be_is_IncSP(irn));
+       return a->align;
+}
 
-       return a->reg_data[out_pos].reg;
+static ir_entity *be_node_get_frame_entity(const ir_node *irn)
+{
+       return be_get_frame_entity(irn);
 }
 
-arch_irn_class_t be_node_classify(const void *_self, const ir_node *irn)
+void be_node_set_frame_entity(ir_node *irn, ir_entity *ent)
 {
-       redir_proj((const ir_node **) &irn, -1);
+       be_frame_attr_t *a;
 
-       switch(get_irn_be_opcode(irn)) {
-#define XXX(a,b) case beo_ ## a: return arch_irn_class_ ## b;
-               XXX(Spill, spill)
-               XXX(Reload, reload)
-               XXX(Perm, perm)
-               XXX(Copy, copy)
-#undef XXX
-               default:
-               return 0;
-       }
+       assert(be_has_frame_entity(irn));
 
-       return 0;
+       a = (be_frame_attr_t*)get_irn_generic_attr(irn);
+       a->ent = ent;
 }
 
-arch_irn_class_t be_node_get_flags(const void *_self, const ir_node *irn)
+static void be_node_set_frame_offset(ir_node *irn, int offset)
 {
+       be_frame_attr_t *a;
+
+       if (!be_has_frame_entity(irn))
+               return;
+
+       a = (be_frame_attr_t*)get_irn_generic_attr(irn);
+       a->offset = offset;
+}
+
+static int be_node_get_sp_bias(const ir_node *irn)
+{
+       if (be_is_IncSP(irn))
+               return be_get_IncSP_offset(irn);
+       if (be_is_Call(irn))
+               return -(int)be_Call_get_pop(irn);
+
        return 0;
 }
 
-static const arch_irn_ops_if_t be_node_irn_ops_if = {
-       be_node_get_irn_reg_req,
-       be_node_set_irn_reg,
-       be_node_get_irn_reg,
-       be_node_classify,
-       be_node_get_flags,
-};
 
+
+/* for be nodes */
 static const arch_irn_ops_t be_node_irn_ops = {
-       &be_node_irn_ops_if
+       be_node_get_frame_entity,
+       be_node_set_frame_offset,
+       be_node_get_sp_bias,
+       NULL,    /* get_op_estimated_cost   */
+       NULL,    /* possible_memory_operand */
+       NULL,    /* perform_memory_operand  */
 };
 
-const void *be_node_get_arch_ops(const arch_irn_handler_t *self, const ir_node *irn)
+static int get_start_reg_index(ir_graph *irg, const arch_register_t *reg)
 {
-       redir_proj((const ir_node **) &irn, -1);
-       return is_be_node(irn) ? &be_node_irn_ops : NULL;
+       ir_node *start  = get_irg_start(irg);
+
+       /* do a naive linear search... */
+       be_foreach_out(start, i) {
+               arch_register_req_t const *const out_req = arch_get_irn_register_req_out(start, i);
+               if (!arch_register_req_is(out_req, limited))
+                       continue;
+               if (out_req->cls != reg->reg_class)
+                       continue;
+               if (!rbitset_is_set(out_req->limited, reg->index))
+                       continue;
+               return i;
+       }
+       panic("Tried querying undefined register '%s' at Start", reg->name);
 }
 
-const arch_irn_handler_t be_node_irn_handler = {
-       be_node_get_arch_ops
-};
-
-static int dump_node(ir_node *irn, FILE *f, dump_reason_t reason)
+ir_node *be_get_initial_reg_value(ir_graph *irg, const arch_register_t *reg)
 {
-       be_node_attr_t *at = get_irn_attr(irn);
-       int i;
-
-       assert(is_be_node(irn));
-
-       switch(reason) {
-               case dump_node_opcode_txt:
-                       fprintf(f, get_op_name(get_irn_op(irn)));
-                       break;
-               case dump_node_mode_txt:
-                       fprintf(f, get_mode_name(get_irn_mode(irn)));
-                       break;
-               case dump_node_nodeattr_txt:
-                       break;
-               case dump_node_info_txt:
-                       fprintf(f, "reg class: %s\n", at->cls->name);
-                       for(i = 0; i < at->n_outs; ++i) {
-                               const arch_register_t *reg = at->reg_data[i].reg;
-                               fprintf(f, "reg #%d: %s\n", i, reg ? reg->name : "n/a");
-                       }
+       int      i     = get_start_reg_index(irg, reg);
+       ir_node *start = get_irg_start(irg);
+       ir_mode *mode  = arch_register_class_mode(reg->reg_class);
+
+       foreach_out_edge(start, edge) {
+               ir_node *proj = get_edge_src_irn(edge);
+               if (!is_Proj(proj)) // maybe End/Anchor
+                       continue;
+               if (get_Proj_proj(proj) == i) {
+                       return proj;
+               }
+       }
+       return new_r_Proj(start, mode, i);
+}
 
-                       if(get_irn_be_opcode(irn) == beo_Spill) {
-                               be_spill_attr_t *a = (be_spill_attr_t *) at;
+static ir_entity* dummy_get_frame_entity(const ir_node *node)
+{
+       (void) node;
+       return NULL;
+}
 
-                               ir_fprintf(f, "spill context: %+F\n", a->spill_ctx);
-                               if (a->ent) {
-                                       unsigned ofs = get_entity_offset_bytes(a->ent);
-                                       ir_fprintf(f, "spill entity: %+F offset %x (%d)\n", a->ent, ofs, ofs);
-                               }
-                               else {
-                                       ir_fprintf(f, "spill entity: n/a\n");
-                               }
-                       }
-                       break;
-       }
+static void dummy_set_frame_offset(ir_node *node, int bias)
+{
+       (void) node;
+       (void) bias;
+       panic("should not be called");
+}
 
+static int dummy_get_sp_bias(const ir_node *node)
+{
+       (void) node;
        return 0;
 }
 
-static const ir_op_ops be_node_op_ops = {
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       NULL,
-       dump_node,
-       NULL
+/* for "middleend" nodes */
+static const arch_irn_ops_t dummy_be_irn_ops = {
+       dummy_get_frame_entity,
+       dummy_set_frame_offset,
+       dummy_get_sp_bias,
+       NULL,      /* get_op_estimated_cost */
+       NULL,      /* possible_memory_operand */
+       NULL,      /* perform_memory_operand */
 };
 
-pset *nodes_live_at(const arch_env_t *arch_env, const arch_register_class_t *cls, const ir_node *pos, pset *live)
-{
-       firm_dbg_module_t *dbg = firm_dbg_register("firm.be.node");
-       const ir_node *bl      = is_Block(pos) ? pos : get_nodes_block(pos);
-       ir_node *irn;
-       irn_live_t *li;
 
-       live_foreach(bl, li) {
-               ir_node *irn = (ir_node *) li->irn;
-               if(live_is_end(li) && arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
-                       pset_insert_ptr(live, irn);
+
+ir_node *be_new_Phi(ir_node *block, int n_ins, ir_node **ins, ir_mode *mode,
+                    const arch_register_req_t *req)
+{
+       ir_graph       *irg  = get_irn_irg(block);
+       struct obstack *obst = be_get_be_obst(irg);
+       backend_info_t *info;
+       int             i;
+
+       ir_node *phi = new_ir_node(NULL, irg, block, op_Phi, mode, n_ins, ins);
+       phi->attr.phi.u.backedge = new_backedge_arr(get_irg_obstack(irg), n_ins);
+       info = be_get_info(phi);
+       info->out_infos = NEW_ARR_DZ(reg_out_info_t, obst, 1);
+       info->in_reqs   = OALLOCN(obst, const arch_register_req_t*, n_ins);
+
+       info->out_infos[0].req = req;
+       for (i = 0; i < n_ins; ++i) {
+               info->in_reqs[i] = req;
        }
+       irn_verify_irg(phi, irg);
+       phi = optimize_node(phi);
+       return phi;
+}
 
-       sched_foreach_reverse(bl, irn) {
-               int i, n;
-               ir_node *x;
+void be_set_phi_reg_req(ir_node *node, const arch_register_req_t *req)
+{
+       int arity = get_irn_arity(node);
+       int i;
 
-               /*
-                * If we encounter the node we want to insert the Perm after,
-               * exit immediately, so that this node is still live
-               */
-               if(irn == pos)
-                       return live;
+       backend_info_t *info = be_get_info(node);
+       info->out_infos[0].req = req;
+       for (i = 0; i < arity; ++i) {
+               info->in_reqs[i] = req;
+       }
 
-               DBG((dbg, LEVEL_1, "%+F\n", irn));
-               for(x = pset_first(live); x; x = pset_next(live))
-                       DBG((dbg, LEVEL_1, "\tlive: %+F\n", x));
+       assert(mode_is_datab(get_irn_mode(node)));
+}
 
-               if(arch_irn_has_reg_class(arch_env, irn, -1, cls))
-                       pset_remove_ptr(live, irn);
+void be_dump_phi_reg_reqs(FILE *F, const ir_node *node, dump_reason_t reason)
+{
+       ir_graph *irg = get_irn_irg(node);
+       if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_BACKEND))
+               return;
 
-               for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
-                       ir_node *op = get_irn_n(irn, i);
+       switch (reason) {
+       case dump_node_opcode_txt:
+               fputs(get_op_name(get_irn_op(node)), F);
+               break;
+       case dump_node_mode_txt:
+               fprintf(F, "%s", get_mode_name(get_irn_mode(node)));
+               break;
+       case dump_node_nodeattr_txt:
+               break;
+       case dump_node_info_txt:
+               arch_dump_reqs_and_registers(F, node);
+               break;
 
-                       if(arch_irn_consider_in_reg_alloc(arch_env, cls, op))
-                               pset_insert_ptr(live, op);
-               }
+       default:
+               break;
        }
-
-       return live;
 }
 
-ir_node *insert_Perm_after(const arch_env_t *arch_env,
-                                                  const arch_register_class_t *cls,
-                                                  dom_front_info_t *dom_front,
-                                                  ir_node *pos)
+static const arch_irn_ops_t phi_irn_ops = {
+       dummy_get_frame_entity,
+       dummy_set_frame_offset,
+       dummy_get_sp_bias,
+       NULL,    /* get_op_estimated_cost   */
+       NULL,    /* possible_memory_operand */
+       NULL,    /* perform_memory_operand  */
+};
+
+
+
+/**
+ * ir_op-Operation: dump a be node to file
+ */
+static void dump_node(FILE *f, const ir_node *irn, dump_reason_t reason)
 {
-       ir_node *bl                 = is_Block(pos) ? pos : get_nodes_block(pos);
-       ir_graph *irg               = get_irn_irg(bl);
-       pset *live                  = pset_new_ptr_default();
-       firm_dbg_module_t *dbg      = firm_dbg_register("be.node");
+       assert(is_be_node(irn));
 
-       ir_node *curr, *irn, *perm, **nodes;
-       int i, n;
+       switch (reason) {
+               case dump_node_opcode_txt:
+                       fputs(get_op_name(get_irn_op(irn)), f);
+                       break;
+               case dump_node_mode_txt:
+                       if (be_is_Copy(irn) || be_is_CopyKeep(irn)) {
+                               fprintf(f, "%s", get_mode_name(get_irn_mode(irn)));
+                       }
+                       break;
+               case dump_node_nodeattr_txt:
+                       if (be_is_Call(irn)) {
+                               const be_call_attr_t *a = (const be_call_attr_t*)get_irn_generic_attr_const(irn);
+                               if (a->ent)
+                                       fprintf(f, " [%s] ", get_entity_name(a->ent));
+                       }
+                       if (be_is_IncSP(irn)) {
+                               const be_incsp_attr_t *attr = (const be_incsp_attr_t*)get_irn_generic_attr_const(irn);
+                               fprintf(f, " [%d] ", attr->offset);
+                       }
+                       break;
+               case dump_node_info_txt:
+                       arch_dump_reqs_and_registers(f, irn);
 
-       DBG((dbg, LEVEL_1, "Insert Perm after: %+F\n", pos));
+                       if (be_has_frame_entity(irn)) {
+                               const be_frame_attr_t *a = (const be_frame_attr_t*)get_irn_generic_attr_const(irn);
+                               if (a->ent) {
+                                       unsigned size = get_type_size_bytes(get_entity_type(a->ent));
+                                       ir_fprintf(f, "frame entity: %+F, offset 0x%x (%d), size 0x%x (%d) bytes\n",
+                                         a->ent, a->offset, a->offset, size, size);
+                               }
 
-       if(!nodes_live_at(arch_env, cls, pos, live));
+                       }
 
-       n = pset_count(live);
+                       switch (get_irn_opcode(irn)) {
+                       case beo_IncSP: {
+                               const be_incsp_attr_t *a = (const be_incsp_attr_t*)get_irn_generic_attr_const(irn);
+                               fprintf(f, "align: %d\n", a->align);
+                               fprintf(f, "offset: %d\n", a->offset);
+                               break;
+                       }
+                       case beo_Call: {
+                               const be_call_attr_t *a = (const be_call_attr_t*)get_irn_generic_attr_const(irn);
 
-       if(n == 0)
-               return NULL;
+                               if (a->ent)
+                                       fprintf(f, "\ncalling: %s\n", get_entity_name(a->ent));
+                               break;
+                       }
+                       case beo_MemPerm: {
+                               int i;
+                               for (i = 0; i < be_get_MemPerm_entity_arity(irn); ++i) {
+                                       ir_entity *in, *out;
+                                       in = be_get_MemPerm_in_entity(irn, i);
+                                       out = be_get_MemPerm_out_entity(irn, i);
+                                       if (in) {
+                                               fprintf(f, "\nin[%d]: %s\n", i, get_entity_name(in));
+                                       }
+                                       if (out) {
+                                               fprintf(f, "\nout[%d]: %s\n", i, get_entity_name(out));
+                                       }
+                               }
+                               break;
+                       }
 
-       nodes = malloc(n * sizeof(nodes[0]));
+                       default:
+                               break;
+                       }
+       }
+}
 
-       DBG((dbg, LEVEL_1, "live:\n"));
-       for(irn = pset_first(live), i = 0; irn; irn = pset_next(live), i++) {
-               DBG((dbg, LEVEL_1, "\t%+F\n", irn));
-               nodes[i] = irn;
+/**
+ * ir_op-Operation:
+ * Copies the backend specific attributes from old node to new node.
+ */
+static void copy_attr(ir_graph *irg, const ir_node *old_node, ir_node *new_node)
+{
+       const void     *old_attr = get_irn_generic_attr_const(old_node);
+       void           *new_attr = get_irn_generic_attr(new_node);
+       struct obstack *obst     = be_get_be_obst(irg);
+       backend_info_t *old_info = be_get_info(old_node);
+       backend_info_t *new_info = be_get_info(new_node);
+
+       assert(is_be_node(old_node));
+       assert(is_be_node(new_node));
+
+       memcpy(new_attr, old_attr, get_op_attr_size(get_irn_op(old_node)));
+
+       new_info->flags     = old_info->flags;
+       new_info->out_infos = old_info->out_infos ? DUP_ARR_D(reg_out_info_t, obst, old_info->out_infos) : NULL;
+
+       /* input infos */
+       if (old_info->in_reqs != NULL) {
+               unsigned n_ins = get_irn_arity(old_node);
+               /* need dynamic in infos? */
+               if (get_irn_op(old_node)->opar == oparity_dynamic) {
+                       new_info->in_reqs = NEW_ARR_F(const arch_register_req_t*, n_ins);
+               } else {
+                       new_info->in_reqs = OALLOCN(obst,const arch_register_req_t*, n_ins);
+               }
+               memcpy(new_info->in_reqs, old_info->in_reqs,
+                      n_ins * sizeof(new_info->in_reqs[0]));
+       } else {
+               new_info->in_reqs = NULL;
        }
+}
 
-       perm = be_new_Perm(cls, irg, bl, n, nodes);
-       sched_add_after(pos, perm);
-       free(nodes);
+int is_be_node(const ir_node *irn)
+{
+       return get_op_ops(get_irn_op(irn))->be_ops == &be_node_irn_ops;
+}
 
-       curr = perm;
-       for(i = 0; i < n; ++i) {
-               ir_node *copies[1];
-               ir_node *perm_op = get_irn_n(perm, i);
-               const arch_register_t *reg = arch_get_irn_register(arch_env, perm_op);
+static ir_op *new_be_op(unsigned code, const char *name, op_pin_state p,
+                        irop_flags flags, op_arity opar, size_t attr_size)
+{
+       ir_op *res = new_ir_op(code, name, p, flags, opar, 0, attr_size);
+       res->ops.dump_node = dump_node;
+       res->ops.copy_attr = copy_attr;
+       res->ops.be_ops    = &be_node_irn_ops;
+       return res;
+}
 
-               ir_mode *mode = get_irn_mode(perm_op);
-               ir_node *proj = new_r_Proj(irg, bl, perm, mode, i);
-               arch_set_irn_register(arch_env, proj, reg);
+void be_init_op(void)
+{
+       unsigned opc;
 
-               sched_add_after(curr, proj);
-               curr = proj;
+       assert(op_be_Spill == NULL);
 
-               copies[0] = proj;
-               be_ssa_constr_single(dom_front, perm_op, 1, copies);
+       /* Acquire all needed opcodes. */
+       op_be_Spill     = new_be_op(beo_Spill,     "be_Spill",     op_pin_state_exc_pinned, irop_flag_none,                          oparity_unary,    sizeof(be_frame_attr_t));
+       op_be_Reload    = new_be_op(beo_Reload,    "be_Reload",    op_pin_state_exc_pinned, irop_flag_none,                          oparity_zero,     sizeof(be_frame_attr_t));
+       op_be_Perm      = new_be_op(beo_Perm,      "be_Perm",      op_pin_state_exc_pinned, irop_flag_none,                          oparity_variable, sizeof(be_node_attr_t));
+       op_be_MemPerm   = new_be_op(beo_MemPerm,   "be_MemPerm",   op_pin_state_exc_pinned, irop_flag_none,                          oparity_variable, sizeof(be_memperm_attr_t));
+       op_be_Copy      = new_be_op(beo_Copy,      "be_Copy",      op_pin_state_exc_pinned, irop_flag_none,                          oparity_unary,    sizeof(be_node_attr_t));
+       op_be_Keep      = new_be_op(beo_Keep,      "be_Keep",      op_pin_state_exc_pinned, irop_flag_keep,                          oparity_dynamic,  sizeof(be_node_attr_t));
+       op_be_CopyKeep  = new_be_op(beo_CopyKeep,  "be_CopyKeep",  op_pin_state_exc_pinned, irop_flag_keep,                          oparity_variable, sizeof(be_node_attr_t));
+       op_be_Call      = new_be_op(beo_Call,      "be_Call",      op_pin_state_exc_pinned, irop_flag_fragile|irop_flag_uses_memory, oparity_variable, sizeof(be_call_attr_t));
+       ir_op_set_memory_index(op_be_Call, n_be_Call_mem);
+       ir_op_set_fragile_indices(op_be_Call, pn_be_Call_X_regular, pn_be_Call_X_except);
+       op_be_Return    = new_be_op(beo_Return,    "be_Return",    op_pin_state_exc_pinned, irop_flag_cfopcode,                      oparity_variable, sizeof(be_return_attr_t));
+       op_be_AddSP     = new_be_op(beo_AddSP,     "be_AddSP",     op_pin_state_exc_pinned, irop_flag_none,                          oparity_unary,    sizeof(be_node_attr_t));
+       op_be_SubSP     = new_be_op(beo_SubSP,     "be_SubSP",     op_pin_state_exc_pinned, irop_flag_none,                          oparity_unary,    sizeof(be_node_attr_t));
+       op_be_IncSP     = new_be_op(beo_IncSP,     "be_IncSP",     op_pin_state_exc_pinned, irop_flag_none,                          oparity_unary,    sizeof(be_incsp_attr_t));
+       op_be_Start     = new_be_op(beo_Start,     "be_Start",     op_pin_state_exc_pinned, irop_flag_none,                          oparity_zero,     sizeof(be_node_attr_t));
+       op_be_FrameAddr = new_be_op(beo_FrameAddr, "be_FrameAddr", op_pin_state_exc_pinned, irop_flag_none,                          oparity_unary,    sizeof(be_frame_attr_t));
+
+       op_be_Spill->ops.node_cmp_attr     = FrameAddr_cmp_attr;
+       op_be_Reload->ops.node_cmp_attr    = FrameAddr_cmp_attr;
+       op_be_Perm->ops.node_cmp_attr      = be_nodes_equal;
+       op_be_MemPerm->ops.node_cmp_attr   = be_nodes_equal;
+       op_be_Copy->ops.node_cmp_attr      = be_nodes_equal;
+       op_be_Keep->ops.node_cmp_attr      = be_nodes_equal;
+       op_be_CopyKeep->ops.node_cmp_attr  = be_nodes_equal;
+       op_be_Call->ops.node_cmp_attr      = Call_cmp_attr;
+       op_be_Return->ops.node_cmp_attr    = Return_cmp_attr;
+       op_be_AddSP->ops.node_cmp_attr     = be_nodes_equal;
+       op_be_SubSP->ops.node_cmp_attr     = be_nodes_equal;
+       op_be_IncSP->ops.node_cmp_attr     = IncSP_cmp_attr;
+       op_be_Start->ops.node_cmp_attr     = be_nodes_equal;
+       op_be_FrameAddr->ops.node_cmp_attr = FrameAddr_cmp_attr;
+
+       /* attach out dummy_ops to middle end nodes */
+       for (opc = iro_First; opc <= iro_Last; ++opc) {
+               ir_op *op = ir_get_opcode(opc);
+               assert(op->ops.be_ops == NULL);
+               op->ops.be_ops = &dummy_be_irn_ops;
        }
-       return perm;
+
+       op_Phi->ops.be_ops = &phi_irn_ops;
+}
+
+void be_finish_op(void)
+{
+       free_ir_op(op_be_Spill);     op_be_Spill     = NULL;
+       free_ir_op(op_be_Reload);    op_be_Reload    = NULL;
+       free_ir_op(op_be_Perm);      op_be_Perm      = NULL;
+       free_ir_op(op_be_MemPerm);   op_be_MemPerm   = NULL;
+       free_ir_op(op_be_Copy);      op_be_Copy      = NULL;
+       free_ir_op(op_be_Keep);      op_be_Keep      = NULL;
+       free_ir_op(op_be_CopyKeep);  op_be_CopyKeep  = NULL;
+       free_ir_op(op_be_Call);      op_be_Call      = NULL;
+       free_ir_op(op_be_Return);    op_be_Return    = NULL;
+       free_ir_op(op_be_IncSP);     op_be_IncSP     = NULL;
+       free_ir_op(op_be_AddSP);     op_be_AddSP     = NULL;
+       free_ir_op(op_be_SubSP);     op_be_SubSP     = NULL;
+       free_ir_op(op_be_Start);     op_be_Start     = NULL;
+       free_ir_op(op_be_FrameAddr); op_be_FrameAddr = NULL;
 }