[libfirm] / ir / be / beabi.c

/*
 * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
 *
 * This file is part of libFirm.
 *
 * This file may be distributed and/or modified under the terms of the
 * GNU General Public License version 2 as published by the Free Software
 * Foundation and appearing in the file LICENSE.GPL included in the
 * packaging of this file.
 *
 * Licensees holding valid libFirm Professional Edition licenses may use
 * this file in accordance with the libFirm Commercial License.
 * Agreement provided with the Software.
 *
 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE.
 */

/**
 * @file
 * @brief       Backend ABI implementation.
 * @author      Sebastian Hack, Michael Beck
 */
#include "config.h"

#include "obst.h"

#include "irgopt.h"

#include "irgraph_t.h"
#include "irnode_t.h"
#include "ircons_t.h"
#include "iredges_t.h"
#include "irgmod.h"
#include "irgwalk.h"
#include "irprintf.h"
#include "irgopt.h"
#include "iropt_t.h"
#include "irtools.h"
#include "heights.h"
#include "pdeq.h"
#include "util.h"
#include "raw_bitset.h"
#include "error.h"
#include "pset_new.h"

#include "be.h"
#include "beabi.h"
#include "beabihelper.h"
#include "bearch.h"
#include "benode.h"
#include "belive_t.h"
#include "besched.h"
#include "beirg.h"
#include "bessaconstr.h"
#include "bemodule.h"
#include "betranshlp.h"

DEBUG_ONLY(static firm_dbg_module_t *dbg;)

typedef struct be_abi_call_arg_t {
        unsigned is_res   : 1;  /**< 1: the call argument is a return value. 0: it's a call parameter. */
        unsigned in_reg   : 1;  /**< 1: this argument is transmitted 1: in registers, 0: on stack. */
        unsigned callee   : 1;  /**< 1: someone called us. 0: We call another function */

        int                    pos;
        const arch_register_t *reg;
        ir_entity             *stack_ent;
        ir_mode               *load_mode;
        unsigned               alignment;    /**< stack alignment */
        unsigned               space_before; /**< allocate space before */
        unsigned               space_after;  /**< allocate space after */
} be_abi_call_arg_t;

struct be_abi_call_t {
        be_abi_call_flags_t       flags;  /**< Flags describing the ABI behavior on calls */
        int                       pop;    /**< number of bytes the stack frame is shrinked by the callee on return. */
        const be_abi_callbacks_t *cb;
        set                      *params;
};

/**
 * The ABI information for the current graph.
 */
struct be_abi_irg_t {
        be_abi_call_t        *call;         /**< The ABI call information. */

        ir_node              *init_sp;      /**< The node representing the stack pointer
                                                 at the start of the function. */

        pmap                 *regs;         /**< A map of all callee-save and ignore regs to
                                                 their Projs to the RegParams node. */
        pmap                 *keep_map;     /**< mapping blocks to keep nodes. */

        ir_node              **calls;       /**< flexible array containing all be_Call nodes */
};

static ir_heights_t *ir_heights;

static ir_node *be_abi_reg_map_get(pmap *map, const arch_register_t *reg)
{
        return pmap_get(ir_node, map, reg);
}

static void be_abi_reg_map_set(pmap *map, const arch_register_t* reg,
                               ir_node *node)
{
        pmap_insert(map, reg, node);
}

/**
 * Check if the given register is callee save, i.e. will be saved by the callee.
 */
static bool arch_register_is_callee_save(
        const arch_env_t      *arch_env,
        const arch_register_t *reg)
{
        if (arch_env->impl->register_saved_by)
                return arch_env->impl->register_saved_by(reg, /*callee=*/1);
        return false;
}

/**
 * Check if the given register is caller save, i.e. must be saved by the caller.
 */
static bool arch_register_is_caller_save(
        const arch_env_t      *arch_env,
        const arch_register_t *reg)
{
        if (arch_env->impl->register_saved_by)
                return arch_env->impl->register_saved_by(reg, /*callee=*/0);
        return false;
}



/*
     _    ____ ___    ____      _ _ _                _
    / \  | __ )_ _|  / ___|__ _| | | |__   __ _  ___| | _____
   / _ \ |  _ \| |  | |   / _` | | | '_ \ / _` |/ __| |/ / __|
  / ___ \| |_) | |  | |__| (_| | | | |_) | (_| | (__|   <\__ \
 /_/   \_\____/___|  \____\__,_|_|_|_.__/ \__,_|\___|_|\_\___/

  These callbacks are used by the backend to set the parameters
  for a specific call type.
*/

/**
 * Set compare function: compares two ABI call object arguments.
 */
static int cmp_call_arg(const void *a, const void *b, size_t n)
{
        const be_abi_call_arg_t *p = (const be_abi_call_arg_t*)a;
        const be_abi_call_arg_t *q = (const be_abi_call_arg_t*)b;
        (void) n;
        return !(p->is_res == q->is_res && p->pos == q->pos && p->callee == q->callee);
}

/**
 * Get  an ABI call object argument.
 *
 * @param call      the abi call
 * @param is_res    true for call results, false for call arguments
 * @param pos       position of the argument
 * @param callee    context type - if we are callee or caller
 */
static be_abi_call_arg_t *get_call_arg(be_abi_call_t *call, int is_res, int pos, int callee)
{
        be_abi_call_arg_t arg;
        unsigned hash;

        memset(&arg, 0, sizeof(arg));
        arg.is_res = is_res;
        arg.pos    = pos;
        arg.callee = callee;

        hash = is_res * 128 + pos;

        return set_find(be_abi_call_arg_t, call->params, &arg, sizeof(arg), hash);
}

/**
 * Set an ABI call object argument.
 */
static void remember_call_arg(be_abi_call_arg_t *arg, be_abi_call_t *call, be_abi_context_t context)
{
        unsigned hash = arg->is_res * 128 + arg->pos;
        if (context & ABI_CONTEXT_CALLEE) {
                arg->callee = 1;
                (void)set_insert(be_abi_call_arg_t, call->params, arg, sizeof(*arg), hash);
        }
        if (context & ABI_CONTEXT_CALLER) {
                arg->callee = 0;
                (void)set_insert(be_abi_call_arg_t, call->params, arg, sizeof(*arg), hash);
        }
}

/* Set the flags for a call. */
void be_abi_call_set_flags(be_abi_call_t *call, be_abi_call_flags_t flags, const be_abi_callbacks_t *cb)
{
        call->flags = flags;
        call->cb    = cb;
}

/* Sets the number of bytes the stackframe is shrinked by the callee on return */
void be_abi_call_set_pop(be_abi_call_t *call, int pop)
{
        assert(pop >= 0);
        call->pop = pop;
}

void be_abi_call_param_stack(be_abi_call_t *call, int arg_pos,
                             ir_mode *load_mode, unsigned alignment,
                             unsigned space_before, unsigned space_after,
                             be_abi_context_t context)
{
        be_abi_call_arg_t arg;
        memset(&arg, 0, sizeof(arg));
        assert(alignment > 0 && "Alignment must be greater than 0");
        arg.load_mode    = load_mode;
        arg.alignment    = alignment;
        arg.space_before = space_before;
        arg.space_after  = space_after;
        arg.is_res       = 0;
        arg.pos          = arg_pos;

        remember_call_arg(&arg, call, context);
}

void be_abi_call_param_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
{
        be_abi_call_arg_t arg;
        memset(&arg, 0, sizeof(arg));

        arg.in_reg = 1;
        arg.reg    = reg;
        arg.is_res = 0;
        arg.pos    = arg_pos;

        remember_call_arg(&arg, call, context);
}

void be_abi_call_res_reg(be_abi_call_t *call, int arg_pos, const arch_register_t *reg, be_abi_context_t context)
{
        be_abi_call_arg_t arg;
        memset(&arg, 0, sizeof(arg));

        arg.in_reg = 1;
        arg.reg    = reg;
        arg.is_res = 1;
        arg.pos    = arg_pos;

        remember_call_arg(&arg, call, context);
}

/* Get the flags of a ABI call object. */
be_abi_call_flags_t be_abi_call_get_flags(const be_abi_call_t *call)
{
        return call->flags;
}

/**
 * Constructor for a new ABI call object.
 *
 * @return the new ABI call object
 */
static be_abi_call_t *be_abi_call_new(void)
{
        be_abi_call_t *call = XMALLOCZ(be_abi_call_t);

        call->params            = new_set(cmp_call_arg, 16);
        call->cb                = NULL;
        call->flags.try_omit_fp = be_options.omit_fp;

        return call;
}

/**
 * Destructor for an ABI call object.
 */
static void be_abi_call_free(be_abi_call_t *call)
{
        del_set(call->params);
        free(call);
}

/**
 * Initializes the frame layout from parts
 *
 * @param frame     the stack layout that will be initialized
 * @param args      the stack argument layout type
 * @param between   the between layout type
 * @param locals    the method frame type
 *
 * @return the initialized stack layout
 */
static be_stack_layout_t *stack_frame_init(be_stack_layout_t *frame, ir_type *args,
                                           ir_type *between, ir_type *locals)
{
        frame->arg_type       = args;
        frame->between_type   = between;
        frame->frame_type     = locals;
        frame->initial_offset = 0;
        frame->initial_bias   = 0;
        frame->order[1]       = between;

        /* typical decreasing stack: locals have the
         * lowest addresses, arguments the highest */
        frame->order[0] = locals;
        frame->order[2] = args;
        return frame;
}

/*
   ____      _ _
  / ___|__ _| | |___
 | |   / _` | | / __|
 | |__| (_| | | \__ \
  \____\__,_|_|_|___/

  Adjustment of the calls inside a graph.

*/

/**
 * Transform a call node into a be_Call node.
 *
 * @param env The ABI environment for the current irg.
 * @param irn The call node.
 * @param curr_sp The stack pointer node to use.
 * @return The stack pointer after the call.
 */
static ir_node *adjust_call(be_abi_irg_t *env, ir_node *irn, ir_node *curr_sp)
{
        ir_graph *irg              = get_irn_irg(irn);
        const arch_env_t *arch_env = be_get_irg_arch_env(irg);
        ir_type *call_tp           = get_Call_type(irn);
        ir_node *call_ptr          = get_Call_ptr(irn);
        size_t   n_params          = get_method_n_params(call_tp);
        ir_node *curr_mem          = get_Call_mem(irn);
        ir_node *bl                = get_nodes_block(irn);
        int stack_size             = 0;
        const arch_register_t *sp  = arch_env->sp;
        be_abi_call_t *call        = be_abi_call_new();
        ir_mode *mach_mode         = sp->reg_class->mode;
        int n_res                  = get_method_n_ress(call_tp);

        ir_node *res_proj  = NULL;
        int n_reg_params   = 0;
        int n_stack_params = 0;
        int n_ins;

        const arch_register_t **states = NEW_ARR_F(const arch_register_t*, 0);
        const arch_register_t **destroyed_regs = NEW_ARR_F(const arch_register_t*, 0);
        ir_node                *low_call;
        ir_node               **in;
        ir_node               **res_projs;
        int                     n_reg_results = 0;
        int                    *reg_param_idxs;
        int                    *stack_param_idx;
        int                     i, n;
        int                     throws_exception;
        size_t                  s;
        size_t                  p;
        dbg_info               *dbgi;

        /* Let the isa fill out the abi description for that call node. */
        arch_env_get_call_abi(arch_env, call_tp, call);

        /* Insert code to put the stack arguments on the stack. */
        assert((size_t)get_Call_n_params(irn) == n_params);
        stack_param_idx = ALLOCAN(int, n_params);
        for (p = 0; p < n_params; ++p) {
                be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
                assert(arg);
                if (!arg->in_reg) {
                        int arg_size = get_type_size_bytes(get_method_param_type(call_tp, p));

                        stack_size += round_up2(arg->space_before, arg->alignment);
                        stack_size += round_up2(arg_size, arg->alignment);
                        stack_size += round_up2(arg->space_after, arg->alignment);

                        stack_param_idx[n_stack_params++] = p;
                }
        }

        /* Collect all arguments which are passed in registers. */
        reg_param_idxs = ALLOCAN(int, n_params);
        for (p = 0; p < n_params; ++p) {
                be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
                if (arg && arg->in_reg) {
                        reg_param_idxs[n_reg_params++] = p;
                }
        }

        /*
         * If the stack is decreasing and we do not want to store sequentially,
         * or someone else allocated the call frame
         * we allocate as much space on the stack all parameters need, by
         * moving the stack pointer along the stack's direction.
         *
         * Note: we also have to do this for stack_size == 0, because we may have
         * to adjust stack alignment for the call.
         */
        curr_sp = be_new_IncSP(sp, bl, curr_sp, stack_size, 1);

        dbgi = get_irn_dbg_info(irn);
        /* If there are some parameters which shall be passed on the stack. */
        if (n_stack_params > 0) {
                int       curr_ofs = 0;
                ir_node **in       = ALLOCAN(ir_node*, n_stack_params+1);
                unsigned  n_in     = 0;

                curr_mem = get_Call_mem(irn);
                in[n_in++] = curr_mem;

                for (i = 0; i < n_stack_params; ++i) {
                        int p                  = stack_param_idx[i];
                        be_abi_call_arg_t *arg = get_call_arg(call, 0, p, 0);
                        ir_node *param         = get_Call_param(irn, p);
                        ir_node *addr          = curr_sp;
                        ir_node *mem           = NULL;
                        ir_type *param_type    = get_method_param_type(call_tp, p);
                        int param_size         = get_type_size_bytes(param_type) + arg->space_after;

                        /*
                         * If we wanted to build the arguments sequentially,
                         * the stack pointer for the next must be incremented,
                         * and the memory value propagated.
                         */
                        curr_ofs += arg->space_before;
                        curr_ofs =  round_up2(curr_ofs, arg->alignment);

                        /* Make the expression to compute the argument's offset. */
                        if (curr_ofs > 0) {
                                ir_mode *constmode = mach_mode;
                                if (mode_is_reference(mach_mode)) {
                                        constmode = mode_Is;
                                }
                                addr = new_r_Const_long(irg, constmode, curr_ofs);
                                addr = new_r_Add(bl, curr_sp, addr, mach_mode);
                        }

                        /* Insert a store for primitive arguments. */
                        if (is_atomic_type(param_type)) {
                                ir_node *nomem     = get_irg_no_mem(irg);
                                ir_node *mem_input = nomem;
                                ir_node *store     = new_rd_Store(dbgi, bl, mem_input, addr, param, cons_none);
                                mem   = new_r_Proj(store, mode_M, pn_Store_M);
                        } else {
                                /* Make a mem copy for compound arguments. */
                                ir_node *copy;

                                assert(mode_is_reference(get_irn_mode(param)));
                                copy = new_rd_CopyB(dbgi, bl, curr_mem, addr, param, param_type);
                                mem = new_r_Proj(copy, mode_M, pn_CopyB_M);
                        }

                        curr_ofs += param_size;

                        in[n_in++] = mem;
                }

                /* We need the sync only, if we didn't build the stores sequentially. */
                if (n_stack_params >= 1) {
                        curr_mem = new_r_Sync(bl, n_in, in);
                } else {
                        curr_mem = get_Call_mem(irn);
                }
        }

        /* Put caller save into the destroyed set and state registers in the states
         * set */
        for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
                unsigned j;
                const arch_register_class_t *cls = &arch_env->register_classes[i];
                for (j = 0; j < cls->n_regs; ++j) {
                        const arch_register_t *reg = arch_register_for_index(cls, j);

                        /* even if destroyed all is specified, neither SP nor FP are
                         * destroyed (else bad things will happen) */
                        if (reg == arch_env->sp || reg == arch_env->bp)
                                continue;

                        if (reg->type & arch_register_type_state) {
                                ARR_APP1(const arch_register_t*, destroyed_regs, reg);
                                ARR_APP1(const arch_register_t*, states, reg);
                                /* we're already in the destroyed set so no need for further
                                 * checking */
                                continue;
                        }
                        if (arch_register_is_caller_save(arch_env, reg))
                                ARR_APP1(const arch_register_t*, destroyed_regs, reg);
                }
        }

        /* search the largest result proj number */
        res_projs = ALLOCANZ(ir_node*, n_res);

        foreach_out_edge(irn, edge) {
                ir_node *irn = get_edge_src_irn(edge);

                if (!is_Proj(irn) || get_Proj_proj(irn) != pn_Call_T_result)
                        continue;

                foreach_out_edge(irn, res_edge) {
                        int proj;
                        ir_node *res = get_edge_src_irn(res_edge);

                        assert(is_Proj(res));

                        proj = get_Proj_proj(res);
                        assert(proj < n_res);
                        assert(res_projs[proj] == NULL);
                        res_projs[proj] = res;
                }
                res_proj = irn;
                break;
        }

        /** TODO: this is not correct for cases where return values are passed
         * on the stack, but no known ABI does this currently...
         */
        n_reg_results = n_res;

        n_ins = 0;
        in    = ALLOCAN(ir_node*, n_reg_params + ARR_LEN(states));

        /* make the back end call node and set its register requirements. */
        for (i = 0; i < n_reg_params; ++i) {
                in[n_ins++] = get_Call_param(irn, reg_param_idxs[i]);
        }

        /* add state registers ins */
        for (s = 0; s < ARR_LEN(states); ++s) {
                const arch_register_t       *reg = states[s];
                const arch_register_class_t *cls = reg->reg_class;
                ir_node *regnode = new_r_Unknown(irg, cls->mode);
                in[n_ins++]      = regnode;
        }
        assert(n_ins == (int) (n_reg_params + ARR_LEN(states)));

        /* ins collected, build the call */
        throws_exception = ir_throws_exception(irn);
        if (env->call->flags.call_has_imm && is_SymConst(call_ptr)) {
                /* direct call */
                low_call = be_new_Call(dbgi, irg, bl, curr_mem, sp->single_req, curr_sp,
                                       sp->single_req, curr_sp,
                                       n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
                                       n_ins, in, get_Call_type(irn));
                be_Call_set_entity(low_call, get_SymConst_entity(call_ptr));
        } else {
                /* indirect call */
                low_call = be_new_Call(dbgi, irg, bl, curr_mem, sp->single_req, curr_sp,
                                       sp->reg_class->class_req, call_ptr,
                                       n_reg_results + pn_be_Call_first_res + ARR_LEN(destroyed_regs),
                                       n_ins, in, get_Call_type(irn));
        }
        ir_set_throws_exception(low_call, throws_exception);
        be_Call_set_pop(low_call, call->pop);

        /* put the call into the list of all calls for later processing */
        ARR_APP1(ir_node *, env->calls, low_call);

        /* create new stack pointer */
        curr_sp = new_r_Proj(low_call, get_irn_mode(curr_sp), pn_be_Call_sp);
        be_set_constr_single_reg_out(low_call, pn_be_Call_sp, sp,
                        arch_register_req_type_ignore | arch_register_req_type_produces_sp);
        arch_set_irn_register(curr_sp, sp);

        /* now handle results */
        for (i = 0; i < n_res; ++i) {
                ir_node           *proj = res_projs[i];
                be_abi_call_arg_t *arg  = get_call_arg(call, 1, i, 0);

                /* returns values on stack not supported yet */
                assert(arg->in_reg);

                /*
                        shift the proj number to the right, since we will drop the
                        unspeakable Proj_T from the Call. Therefore, all real argument
                        Proj numbers must be increased by pn_be_Call_first_res
                */
                long pn = i + pn_be_Call_first_res;

                if (proj == NULL) {
                        ir_type *res_type = get_method_res_type(call_tp, i);
                        ir_mode *mode     = get_type_mode(res_type);
                        proj              = new_r_Proj(low_call, mode, pn);
                        res_projs[i]      = proj;
                } else {
                        set_Proj_pred(proj, low_call);
                        set_Proj_proj(proj, pn);
                }

                if (arg->in_reg) {
                        /* remove register from destroyed regs */
                        size_t j;
                        size_t n = ARR_LEN(destroyed_regs);
                        for (j = 0; j < n; ++j) {
                                if (destroyed_regs[j] == arg->reg) {
                                        destroyed_regs[j] = destroyed_regs[n-1];
                                        ARR_SHRINKLEN(destroyed_regs,n-1);
                                        break;
                                }
                        }
                }
        }

        DBG((dbg, LEVEL_3, "\tcreated backend call %+F\n", low_call));

        /* Set the register classes and constraints of the Call parameters. */
        for (i = 0; i < n_reg_params; ++i) {
                int index = reg_param_idxs[i];
                be_abi_call_arg_t *arg = get_call_arg(call, 0, index, 0);
                assert(arg->reg != NULL);

                be_set_constr_single_reg_in(low_call, n_be_Call_first_arg + i,
                                            arg->reg, arch_register_req_type_none);
        }

        /* Set the register constraints of the results. */
        for (i = 0; i < n_res; ++i) {
                ir_node                 *proj = res_projs[i];
                const be_abi_call_arg_t *arg  = get_call_arg(call, 1, i, 0);
                int                      pn   = get_Proj_proj(proj);

                assert(arg->in_reg);
                be_set_constr_single_reg_out(low_call, pn, arg->reg,
                                             arch_register_req_type_none);
                arch_set_irn_register(proj, arg->reg);
        }
        exchange(irn, low_call);

        /* kill the ProjT node */
        if (res_proj != NULL) {
                kill_node(res_proj);
        }

        /* Make additional projs for the caller save registers
           and the Keep node which keeps them alive. */
        {
                ir_node               **in, *keep;
                int                   i;
                size_t                d;
                int                   n = 0;
                int                   curr_res_proj = pn_be_Call_first_res + n_reg_results;
                int                   n_ins;

                n_ins = ARR_LEN(destroyed_regs) + n_reg_results + 1;
                in    = ALLOCAN(ir_node *, n_ins);

                /* also keep the stack pointer */
                set_irn_link(curr_sp, (void*) sp);
                in[n++] = curr_sp;

                for (d = 0; d < ARR_LEN(destroyed_regs); ++d) {
                        const arch_register_t *reg = destroyed_regs[d];
                        ir_node *proj = new_r_Proj(low_call, reg->reg_class->mode, curr_res_proj);

                        /* memorize the register in the link field. we need afterwards to set the register class of the keep correctly. */
                        be_set_constr_single_reg_out(low_call, curr_res_proj, reg,
                                                     arch_register_req_type_none);
                        arch_set_irn_register(proj, reg);

                        set_irn_link(proj, (void*) reg);
                        in[n++] = proj;
                        ++curr_res_proj;
                }

                for (i = 0; i < n_reg_results; ++i) {
                        ir_node *proj = res_projs[i];
                        const arch_register_t *reg = arch_get_irn_register(proj);
                        set_irn_link(proj, (void*) reg);
                        in[n++] = proj;
                }
                assert(n <= n_ins);

                /* create the Keep for the caller save registers */
                keep = be_new_Keep(bl, n, in);
                for (i = 0; i < n; ++i) {
                        const arch_register_t *reg = (const arch_register_t*)get_irn_link(in[i]);
                        be_node_set_reg_class_in(keep, i, reg->reg_class);
                }
        }

        /* Clean up the stack. */
        assert(stack_size >= call->pop);
        stack_size -= call->pop;

        if (stack_size > 0) {
                ir_node *mem_proj = NULL;

                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(low_call, mode_M, pn_be_Call_M);
                        keep_alive(mem_proj);
                }
        }
        /* Clean up the stack frame or revert alignment fixes if we allocated it */
        curr_sp = be_new_IncSP(sp, bl, curr_sp, -stack_size, 0);

        be_abi_call_free(call);

        DEL_ARR_F(states);
        DEL_ARR_F(destroyed_regs);

        return curr_sp;
}

/**
 * Adjust the size of a node representing a stack alloc or free for the minimum stack alignment.
 *
 * @param alignment  the minimum stack alignment
 * @param size       the node containing the non-aligned size
 * @param block      the block where new nodes are allocated on
 * @param dbg        debug info for new nodes
 *
 * @return a node representing the aligned size
 */
static ir_node *adjust_alloc_size(unsigned stack_alignment, ir_node *size,
                                  ir_node *block, dbg_info *dbg)
{
        if (stack_alignment > 1) {
                ir_mode   *mode;
                ir_tarval *tv;
                ir_node   *mask;
                ir_graph  *irg;

                assert(is_po2(stack_alignment));

                mode = get_irn_mode(size);
                tv   = new_tarval_from_long(stack_alignment-1, mode);
                irg  = get_Block_irg(block);
                mask = new_r_Const(irg, tv);
                size = new_rd_Add(dbg, block, size, mask, mode);

                tv   = new_tarval_from_long(-(long)stack_alignment, mode);
                mask = new_r_Const(irg, tv);
                size = new_rd_And(dbg, block, size, mask, mode);
        }
        return size;
}
/**
 * Adjust an alloca.
 * The alloca is transformed into a back end alloca node and connected to the stack nodes.
 */
static ir_node *adjust_alloc(be_abi_irg_t *env, ir_node *alloc, ir_node *curr_sp)
{
        ir_node          *block     = get_nodes_block(alloc);
        ir_graph         *irg       = get_Block_irg(block);
        const arch_env_t *arch_env  = be_get_irg_arch_env(irg);
        ir_node          *alloc_mem = NULL;
        ir_node          *alloc_res = NULL;
        ir_type          *type      = get_Alloc_type(alloc);
        dbg_info         *dbg;

        ir_node *new_alloc;
        ir_node *count;
        ir_node *size;
        ir_node *ins[2];
        unsigned stack_alignment;

        /* all non-stack Alloc nodes should already be lowered before the backend */
        assert(get_Alloc_where(alloc) == stack_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;
                }
        }

        /* Beware: currently Alloc nodes without a result might happen,
           only escape analysis kills them and this phase runs only for object
           oriented source. We kill the Alloc here. */
        if (alloc_res == NULL && alloc_mem) {
                exchange(alloc_mem, get_Alloc_mem(alloc));
                return curr_sp;
        }

        dbg   = get_irn_dbg_info(alloc);
        count = get_Alloc_count(alloc);

        /* we might need to multiply the count with the element size */
        if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
                ir_mode   *mode  = get_irn_mode(count);
                ir_tarval *tv    = new_tarval_from_long(get_type_size_bytes(type),
                                                        mode);
                ir_node   *cnst = new_rd_Const(dbg, irg, tv);
                size            = new_rd_Mul(dbg, block, count, cnst, mode);
        } else {
                size = count;
        }

        /* The stack pointer will be modified in an unknown manner.
           We cannot omit it. */
        env->call->flags.try_omit_fp = 0;

        stack_alignment = 1 << arch_env->stack_alignment;
        size            = adjust_alloc_size(stack_alignment, size, block, dbg);
        new_alloc       = be_new_AddSP(arch_env->sp, block, curr_sp, size);
        set_irn_dbg_info(new_alloc, dbg);

        if (alloc_mem != NULL) {
                ir_node *addsp_mem;
                ir_node *sync;

                addsp_mem = new_r_Proj(new_alloc, mode_M, pn_be_AddSP_M);

                /* We need to sync the output mem of the AddSP with the input mem
                   edge into the alloc node. */
                ins[0] = get_Alloc_mem(alloc);
                ins[1] = addsp_mem;
                sync = new_r_Sync(block, 2, ins);

                exchange(alloc_mem, sync);
        }

        exchange(alloc, new_alloc);

        /* fix projnum of alloca res */
        set_Proj_proj(alloc_res, pn_be_AddSP_res);

        curr_sp = new_r_Proj(new_alloc,  get_irn_mode(curr_sp), pn_be_AddSP_sp);

        return curr_sp;
}

/**
 * Adjust a Free.
 * The Free is transformed into a back end free node and connected to the stack nodes.
 */
static ir_node *adjust_free(be_abi_irg_t *env, ir_node *free, ir_node *curr_sp)
{
        ir_node          *block    = get_nodes_block(free);
        ir_graph         *irg      = get_irn_irg(free);
        ir_type          *type     = get_Free_type(free);
        const arch_env_t *arch_env = be_get_irg_arch_env(irg);
        ir_mode          *sp_mode  = arch_env->sp->reg_class->mode;
        dbg_info         *dbg      = get_irn_dbg_info(free);
        ir_node  *subsp, *mem, *res, *size, *sync;
        ir_node *in[2];
        unsigned stack_alignment;

        /* all non-stack-alloc Free nodes should already be lowered before the
         * backend phase */
        assert(get_Free_where(free) == stack_alloc);

        /* we might need to multiply the size with the element size */
        if (!is_unknown_type(type) && get_type_size_bytes(type) != 1) {
                ir_tarval *tv   = new_tarval_from_long(get_type_size_bytes(type), mode_Iu);
                ir_node   *cnst = new_rd_Const(dbg, irg, tv);
                ir_node   *mul  = new_rd_Mul(dbg, block, get_Free_count(free),
                                             cnst, mode_Iu);
                size = mul;
        } else {
                size = get_Free_count(free);
        }

        stack_alignment = 1 << arch_env->stack_alignment;
        size            = adjust_alloc_size(stack_alignment, size, block, dbg);

        /* The stack pointer will be modified in an unknown manner.
           We cannot omit it. */
        env->call->flags.try_omit_fp = 0;
        subsp = be_new_SubSP(arch_env->sp, block, curr_sp, size);
        set_irn_dbg_info(subsp, dbg);

        mem = new_r_Proj(subsp, mode_M, pn_be_SubSP_M);
        res = new_r_Proj(subsp, sp_mode, pn_be_SubSP_sp);

        /* we need to sync the memory */
        in[0] = get_Free_mem(free);
        in[1] = mem;
        sync = new_r_Sync(block, 2, in);

        /* and make the AddSP dependent on the former memory */
        add_irn_dep(subsp, get_Free_mem(free));

        /* kill the free */
        exchange(free, sync);
        curr_sp = res;

        return curr_sp;
}

/**
 * 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 int dependent_on(ir_node *n1, ir_node *n2)
{
        assert(get_nodes_block(n1) == get_nodes_block(n2));

        return heights_reachable_in_block(ir_heights, n1, n2);
}

static int cmp_call_dependency(const void *c1, const void *c2)
{
        ir_node *n1 = *(ir_node **) c1;
        ir_node *n2 = *(ir_node **) c2;
        unsigned h1, h2;

        /*
                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
        */
        if (dependent_on(n1, n2))
                return -1;

        if (dependent_on(n2, n1))
                return 1;

        /* The nodes have no depth order, but we need a total order because qsort()
         * is not stable.
         *
         * Additionally, we need to respect transitive dependencies. Consider a
         * Call a depending on Call b and an independent Call c.
         * We MUST NOT order c > a and b > c. */
        h1 = get_irn_height(ir_heights, n1);
        h2 = get_irn_height(ir_heights, n2);
        if (h1 < h2) return -1;
        if (h1 > h2) return  1;
        /* Same height, so use a random (but stable) order */
        return get_irn_idx(n1) - get_irn_idx(n2);
}

/**
 * Walker: links all Call/Alloc/Free nodes to the Block they are contained.
 */
static void link_ops_in_block_walker(ir_node *irn, void *data)
{
        be_abi_irg_t *env  = (be_abi_irg_t*)data;
        unsigned      code = get_irn_opcode(irn);

        if (code == iro_Call ||
           (code == iro_Alloc && get_Alloc_where(irn) == stack_alloc) ||
           (code == iro_Free && get_Free_where(irn) == stack_alloc)) {
                ir_node *bl       = get_nodes_block(irn);
                void *save        = get_irn_link(bl);

                set_irn_link(irn, save);
                set_irn_link(bl, irn);
        }

        if (code == iro_Builtin && get_Builtin_kind(irn) == ir_bk_return_address) {
                ir_node       *param = get_Builtin_param(irn, 0);
                ir_tarval     *tv    = get_Const_tarval(param);
                unsigned long  value = get_tarval_long(tv);
                /* use ebp, so the climbframe algo works... */
                if (value > 0) {
                        env->call->flags.try_omit_fp = 0;
                }
        }
}

/**
 * Block-walker:
 * Process all Call/Alloc/Free nodes inside a basic block.
 * Note that the link field of the block must contain a linked list of all
 * nodes inside the Block. We first order this list according to data dependency
 * and that connect the nodes together.
 */
static void process_ops_in_block(ir_node *bl, void *data)
{
        be_abi_irg_t   *env     = (be_abi_irg_t*)data;
        ir_node        *curr_sp = env->init_sp;
        ir_node        *irn;
        ir_node       **nodes;
        int             n;
        int             n_nodes;

        n_nodes = 0;
        for (irn = (ir_node*)get_irn_link(bl); irn != NULL;
             irn = (ir_node*)get_irn_link(irn)) {
                ++n_nodes;
        }

        nodes = ALLOCAN(ir_node*, n_nodes);
        for (irn = (ir_node*)get_irn_link(bl), n = 0; irn != NULL;
             irn = (ir_node*)get_irn_link(irn), ++n) {
                nodes[n] = irn;
        }

        /* If there were call nodes in the block. */
        if (n > 0) {
                ir_node *keep;
                int i;

                /* order the call nodes according to data dependency */
                qsort(nodes, n_nodes, sizeof(nodes[0]), cmp_call_dependency);

                for (i = n_nodes - 1; i >= 0; --i) {
                        ir_node *irn = nodes[i];

                        DBG((dbg, LEVEL_3, "\tprocessing call %+F\n", irn));
                        switch (get_irn_opcode(irn)) {
                        case iro_Call:
                                curr_sp = adjust_call(env, irn, curr_sp);
                                break;
                        case iro_Alloc:
                                if (get_Alloc_where(irn) == stack_alloc)
                                        curr_sp = adjust_alloc(env, irn, curr_sp);
                                break;
                        case iro_Free:
                                if (get_Free_where(irn) == stack_alloc)
                                        curr_sp = adjust_free(env, irn, curr_sp);
                                break;
                        default:
                                panic("invalid call");
                        }
                }

                /* Keep the last stack state in the block by tying it to Keep node,
                 * the proj from calls is already kept */
                if (curr_sp != env->init_sp &&
                    !(is_Proj(curr_sp) && be_is_Call(get_Proj_pred(curr_sp)))) {
                        nodes[0] = curr_sp;
                        keep     = be_new_Keep(bl, 1, nodes);
                        pmap_insert(env->keep_map, bl, keep);
                }
        }

        set_irn_link(bl, curr_sp);
}

/**
 * Adjust all call nodes in the graph to the ABI conventions.
 */
static void process_calls(ir_graph *const irg, be_abi_irg_t *const abi)
{
        irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, abi);

        ir_heights = heights_new(irg);
        irg_block_walk_graph(irg, NULL, process_ops_in_block, abi);
        heights_free(ir_heights);
}

/**
 * Computes the stack argument layout type.
 * Changes a possibly allocated value param type by moving
 * entities to the stack layout type.
 *
 * @param call          the current call ABI
 * @param method_type   the method type
 *
 * @return the stack argument layout type
 */
static ir_type *compute_arg_type(ir_graph *irg, be_abi_call_t *call,
                                                                 ir_type *method_type)
{
        struct obstack *obst = be_get_be_obst(irg);
        ir_type   *frame_type      = get_irg_frame_type(irg);
        size_t     n_params        = get_method_n_params(method_type);
        size_t     n_frame_members = get_compound_n_members(frame_type);
        ir_entity *va_start_entity = NULL;
        size_t   f;
        int      ofs  = 0;

        ir_type *res;
        size_t i;
        ir_entity **map = OALLOCNZ(obst, ir_entity*, n_params);
        res = new_type_struct(new_id_from_chars("arg_type", 8));

        /* collect existing entities for value_param_types */
        for (f = n_frame_members; f > 0; ) {
                ir_entity *entity = get_compound_member(frame_type, --f);
                size_t     num;

                set_entity_link(entity, NULL);
                if (!is_parameter_entity(entity))
                        continue;
                num = get_entity_parameter_number(entity);
                if (num == IR_VA_START_PARAMETER_NUMBER) {
                        /* move entity to new arg_type */
                        set_entity_owner(entity, res);
                        va_start_entity = entity;
                        continue;
                }
                assert(num < n_params);
                if (map[num] != NULL)
                        panic("multiple entities for parameter %u in %+F found", f, irg);

                if (num != n_params && get_call_arg(call, 0, num, 1)->in_reg) {
                        /* don't move this entity */
                        continue;
                }

                map[num] = entity;
                /* move entity to new arg_type */
                set_entity_owner(entity, res);
        }

        for (i = 0; i < n_params; ++i) {
                be_abi_call_arg_t *arg        = get_call_arg(call, 0, i, 1);
                ir_type           *param_type = get_method_param_type(method_type, i);
                ir_entity         *entity;

                if (arg->in_reg)
                        continue;
                entity = map[i];
                if (entity == NULL) {
                        /* create a new entity */
                        entity = new_parameter_entity(res, i, param_type);
                }
                ofs += arg->space_before;
                ofs = round_up2(ofs, arg->alignment);
                set_entity_offset(entity, ofs);
                ofs += arg->space_after;
                ofs += get_type_size_bytes(param_type);
                arg->stack_ent = entity;
        }
        if (va_start_entity != NULL) {
                set_entity_offset(va_start_entity, ofs);
        }
        set_type_size_bytes(res, ofs);
        set_type_state(res, layout_fixed);

        return res;
}

static int cmp_regs(const void *a, const void *b)
{
        arch_register_t const *const p = *(arch_register_t const**)a;
        arch_register_t const *const q = *(arch_register_t const**)b;

        if (p->reg_class == q->reg_class)
                return p->index - q->index;
        else
                return p->reg_class < q->reg_class ? -1 : +1;
}

static void reg_map_to_arr(arch_register_t const **const res, pmap *const reg_map)
{
        pmap_entry *ent;
        size_t n = pmap_count(reg_map);
        size_t i = 0;

        foreach_pmap(reg_map, ent) {
                res[i++] = (arch_register_t const*)ent->key;
        }

        qsort(res, n, sizeof(res[0]), cmp_regs);
}

/**
 * Creates a be_Return for a Return node.
 *
 * @param @env  the abi environment
 * @param irn   the Return node
 */
static ir_node *create_be_return(be_abi_irg_t *const env, ir_node *const irn)
{
        ir_node          *const bl = get_nodes_block(irn);
        be_abi_call_t    *call     = env->call;
        ir_graph         *irg      = get_Block_irg(bl);
        const arch_env_t *arch_env = be_get_irg_arch_env(irg);
        pmap *reg_map  = pmap_create();
        ir_node *keep  = pmap_get(ir_node, env->keep_map, bl);
        size_t in_max;
        int i, n;
        ir_node **in;
        ir_node *stack;
        const arch_register_t **regs;
        pmap_entry *ent;

        /*
                get the valid stack node in this block.
                If we had a call in that block there is a Keep constructed by process_calls()
                which points to the last stack modification in that block. we'll use
                it then. Else we use the stack from the start block and let
                the ssa construction fix the usage.
        */
        stack = be_abi_reg_map_get(env->regs, arch_env->sp);
        if (keep) {
                stack = get_irn_n(keep, 0);
                kill_node(keep);
                remove_End_keepalive(get_irg_end(irg), keep);
        }

        int const n_res = get_Return_n_ress(irn);
        /* 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, 1);
                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. */
        foreach_pmap(env->regs, ent) {
                const arch_register_t *reg = (const arch_register_t*)ent->key;
                if (arch_register_is_callee_save(arch_env, reg))
                        pmap_insert(reg_map, ent->key, ent->value);
        }

        be_abi_reg_map_set(reg_map, arch_env->sp, stack);

        /*
                Maximum size of the in array for Return nodes is
                return args + callee save/ignore registers + memory + stack pointer
        */
        in_max = pmap_count(reg_map) + n_res + 2;

        in   = ALLOCAN(ir_node*,               in_max);
        regs = ALLOCAN(arch_register_t const*, in_max);

        in[0]   = get_Return_mem(irn);
        in[1]   = be_abi_reg_map_get(reg_map, arch_env->sp);
        regs[0] = NULL;
        regs[1] = arch_env->sp;
        n       = 2;

        /* clear SP entry, since it has already been grown. */
        pmap_insert(reg_map, (void *) arch_env->sp, NULL);
        for (i = 0; i < n_res; ++i) {
                be_abi_call_arg_t *arg = get_call_arg(call, 1, i, 1);

                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. */
        foreach_pmap(reg_map, ent) {
                if (ent->value) {
                        in[n]     = (ir_node*)ent->value;
                        regs[n++] = (const arch_register_t*)ent->key;
                }
        }

        /* The in array for the new back end return is now ready. */
        dbg_info *const dbgi = get_irn_dbg_info(irn);
        ir_node  *const ret  = be_new_Return(dbgi, irg, bl, n_res, call->pop, n, in);

        /* Set the register classes of the return's parameter accordingly. */
        for (i = 0; i < n; ++i) {
                if (regs[i] == NULL)
                        continue;

                be_set_constr_single_reg_in(ret, i, regs[i], arch_register_req_type_none);
        }

        /* Free the space of the Epilog's in array and the register <-> proj map. */
        pmap_destroy(reg_map);

        return ret;
}

typedef struct lower_frame_sels_env_t {
        ir_node      *frame;                     /**< the current frame */
        const arch_register_class_t *sp_class;   /**< register class of the stack pointer */
} lower_frame_sels_env_t;

/**
 * Walker: Replaces Sels of frame type and
 * value param type entities by FrameAddress.
 * Links all used entities.
 */
static void lower_frame_sels_walker(ir_node *irn, void *data)
{
        lower_frame_sels_env_t *ctx = (lower_frame_sels_env_t*)data;

        if (is_Sel(irn)) {
                ir_node *ptr = get_Sel_ptr(irn);

                if (ptr == ctx->frame) {
                        ir_entity    *ent = get_Sel_entity(irn);
                        ir_node      *bl  = get_nodes_block(irn);
                        ir_node      *nw;

                        nw = be_new_FrameAddr(ctx->sp_class, bl, ctx->frame, ent);
                        exchange(irn, nw);
                }
        }
}

/**
 * The start block has no jump, instead it has an initial exec Proj.
 * The backend wants to handle all blocks the same way, so we replace
 * the out cfg edge with a real jump.
 */
static void fix_start_block(ir_graph *irg)
{
        ir_node *initial_X   = get_irg_initial_exec(irg);
        ir_node *start_block = get_irg_start_block(irg);
        ir_node *jmp         = new_r_Jmp(start_block);

        assert(is_Proj(initial_X));
        exchange(initial_X, jmp);
        set_irg_initial_exec(irg, new_r_Bad(irg, mode_X));

        /* merge start block with successor if possible */
        {
                foreach_out_edge(jmp, edge) {
                        ir_node *succ = get_edge_src_irn(edge);
                        if (!is_Block(succ))
                                continue;

                        if (get_irn_arity(succ) == 1) {
                                exchange(succ, start_block);
                        }
                        break;
                }
        }
}

/**
 * Modify the irg itself and the frame type.
 */
static void modify_irg(ir_graph *const irg, be_abi_irg_t *const env)
{
        be_abi_call_t         *call         = env->call;
        const arch_env_t      *arch_env     = be_get_irg_arch_env(irg);
        const arch_register_t *sp           = arch_env->sp;
        ir_type               *method_type  = get_entity_type(get_irg_entity(irg));
        be_irg_t              *birg         = be_birg_from_irg(irg);
        struct obstack        *obst         = be_get_be_obst(irg);
        be_stack_layout_t     *stack_layout = be_get_irg_stack_layout(irg);
        ir_node *end;
        ir_node *old_mem;
        ir_node *new_mem_proj;
        ir_node *mem;

        int n_params;
        int i, n;
        unsigned j;

        const arch_register_t *fp_reg;
        ir_node *frame_pointer;
        ir_node *start_bl;
        ir_node **args;
        ir_node *arg_tuple;
        ir_type *arg_type, *bet_type;
        lower_frame_sels_env_t ctx;

        DBG((dbg, LEVEL_1, "introducing abi on %+F\n", irg));

        old_mem = get_irg_initial_mem(irg);

        irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);

        arg_type = compute_arg_type(irg, call, method_type);

        /* Convert the Sel nodes in the irg to frame addr nodes: */
        ctx.frame    = get_irg_frame(irg);
        ctx.sp_class = arch_env->sp->reg_class;

        ir_type *const frame_tp = get_irg_frame_type(irg);
        /* layout the stackframe now */
        if (get_type_state(frame_tp) == layout_undefined) {
                default_layout_compound_type(frame_tp);
        }

        /* align stackframe */
        unsigned const alignment  = 1U << arch_env->stack_alignment;
        unsigned const frame_size = round_up2(get_type_size_bytes(frame_tp), alignment);
        set_type_size_bytes(frame_tp, frame_size);

        env->regs  = pmap_create();

        n_params = get_method_n_params(method_type);
        args     = OALLOCNZ(obst, ir_node*, n_params);

        be_add_parameter_entity_stores(irg);

        irg_walk_graph(irg, lower_frame_sels_walker, NULL, &ctx);

        irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);

        /* Fill the argument vector */
        arg_tuple = get_irg_args(irg);
        foreach_out_edge(arg_tuple, edge) {
                ir_node *irn = get_edge_src_irn(edge);
                if (! is_Anchor(irn)) {
                        int nr       = get_Proj_proj(irn);
                        args[nr]     = irn;
                        DBG((dbg, LEVEL_2, "\treading arg: %d -> %+F\n", nr, irn));
                }
        }

        stack_layout->sp_relative = call->flags.try_omit_fp;
        bet_type = call->cb->get_between_type(irg);
        stack_frame_init(stack_layout, arg_type, bet_type,
                         get_irg_frame_type(irg));

        /* 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, 1);
                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]);
                        DBG((dbg, LEVEL_2, "\targ #%d -> reg %s\n", i, arg->reg->name));
                }
        }

        /* Collect all callee-save registers */
        for (i = 0, n = arch_env->n_register_classes; i < n; ++i) {
                const arch_register_class_t *cls = &arch_env->register_classes[i];
                for (j = 0; j < cls->n_regs; ++j) {
                        const arch_register_t *reg = &cls->regs[j];
                        if ((reg->type & arch_register_type_state) || arch_register_is_callee_save(arch_env, reg)) {
                                pmap_insert(env->regs, (void *) reg, NULL);
                        }
                }
        }

        fp_reg = call->flags.try_omit_fp ? arch_env->sp : arch_env->bp;
        rbitset_clear(birg->allocatable_regs, fp_reg->global_index);

        /* handle start block here (place a jump in the block) */
        fix_start_block(irg);

        pmap_insert(env->regs, (void *) sp, NULL);
        pmap_insert(env->regs, (void *) arch_env->bp, NULL);
        start_bl   = get_irg_start_block(irg);
        ir_node *const start = be_new_Start(NULL, start_bl, pmap_count(env->regs) + 1);
        set_irg_start(irg, start);

        /*
         * 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.
         */
        arch_register_t const **const regs = ALLOCAN(arch_register_t const*, pmap_count(env->regs));
        reg_map_to_arr(regs, env->regs);
        for (i = 0, n = pmap_count(env->regs); i < n; ++i) {
                const arch_register_t    *reg      = regs[i];
                ir_mode                  *mode     = reg->reg_class->mode;
                long                      nr       = i;
                arch_register_req_type_t  add_type = arch_register_req_type_none;
                ir_node                  *proj;

                if (reg == sp)
                        add_type |= arch_register_req_type_produces_sp;
                if (!rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
                        add_type |= arch_register_req_type_ignore;
                }

                assert(nr >= 0);
                proj = new_r_Proj(start, mode, nr + 1);
                pmap_insert(env->regs, (void *) reg, proj);
                be_set_constr_single_reg_out(start, nr + 1, reg, add_type);
                arch_set_irn_register(proj, reg);

                DBG((dbg, LEVEL_2, "\tregister save proj #%d -> reg %s\n", nr, reg->name));
        }

        /* create a new initial memory proj */
        assert(is_Proj(old_mem));
        arch_set_irn_register_req_out(start, 0, arch_no_register_req);
        new_mem_proj = new_r_Proj(start, mode_M, 0);
        mem = new_mem_proj;
        set_irg_initial_mem(irg, mem);

        env->init_sp = be_abi_reg_map_get(env->regs, sp);

        /* set new frame_pointer */
        frame_pointer = be_abi_reg_map_get(env->regs, fp_reg);
        set_irg_frame(irg, frame_pointer);

        /* rewire old mem users to new mem */
        exchange(old_mem, mem);

        /* keep the mem (for functions with an endless loop = no return) */
        keep_alive(mem);

        set_irg_initial_mem(irg, mem);

        /* Now, introduce stack param nodes for all parameters passed on the stack */
        for (i = 0; i < n_params; ++i) {
                ir_node *arg_proj = args[i];
                ir_node *repl     = NULL;

                if (arg_proj != NULL) {
                        be_abi_call_arg_t *arg;
                        ir_type *param_type;
                        int     nr = get_Proj_proj(arg_proj);
                        ir_mode *mode;

                        nr         = MIN(nr, n_params);
                        arg        = get_call_arg(call, 0, nr, 1);
                        param_type = get_method_param_type(method_type, nr);

                        if (arg->in_reg) {
                                repl = pmap_get(ir_node, env->regs, arg->reg);
                        } else {
                                ir_node *addr = be_new_FrameAddr(sp->reg_class, start_bl, frame_pointer, arg->stack_ent);

                                /* For atomic parameters which are actually used, we create a Load node. */
                                if (is_atomic_type(param_type) && get_irn_n_edges(args[i]) > 0) {
                                        ir_mode *mode      = get_type_mode(param_type);
                                        ir_mode *load_mode = arg->load_mode;
                                        ir_node *nomem     = get_irg_no_mem(irg);

                                        ir_node *load = new_r_Load(start_bl, nomem, addr, load_mode, cons_floats);
                                        repl = new_r_Proj(load, load_mode, pn_Load_res);

                                        if (mode != load_mode) {
                                                repl = new_r_Conv(start_bl, repl, mode);
                                        }
                                } else {
                                        /* The stack parameter is not primitive (it is a struct or array),
                                         * we thus will create a node representing the parameter's address
                                         * on the stack. */
                                        repl = addr;
                                }
                        }

                        assert(repl != NULL);

                        /* Beware: the mode of the register parameters is always the mode of the register class
                           which may be wrong. Add Conv's then. */
                        mode = get_irn_mode(args[i]);
                        if (mode != get_irn_mode(repl)) {
                                repl = new_r_Conv(get_nodes_block(repl), repl, mode);
                        }
                        exchange(args[i], repl);
                }
        }

        /* the arg proj is not needed anymore now and should be only used by the anchor */
        assert(get_irn_n_edges(arg_tuple) == 1);
        kill_node(arg_tuple);
        set_irg_args(irg, new_r_Bad(irg, mode_T));

        /* All Return nodes hang on the End node, so look for them there. */
        end = get_irg_end_block(irg);
        for (i = 0, n = get_Block_n_cfgpreds(end); i < n; ++i) {
                ir_node *irn = get_Block_cfgpred(end, i);

                if (is_Return(irn)) {
                        ir_node *const ret = create_be_return(env, irn);
                        exchange(irn, ret);
                }
        }

        /* if we have endless loops here, n might be <= 0. Do NOT create a be_Return then,
           the code is dead and will never be executed. */
}

/** Fix the state inputs of calls that still hang on unknowns */
static void fix_call_state_inputs(ir_graph *const irg, be_abi_irg_t *const env)
{
        const arch_env_t *arch_env = be_get_irg_arch_env(irg);
        int i, n, n_states;
        arch_register_t **stateregs = NEW_ARR_F(arch_register_t*, 0);

        /* Collect caller save registers */
        n = arch_env->n_register_classes;
        for (i = 0; i < n; ++i) {
                unsigned j;
                const arch_register_class_t *cls = &arch_env->register_classes[i];
                for (j = 0; j < cls->n_regs; ++j) {
                        const arch_register_t *reg = arch_register_for_index(cls, j);
                        if (reg->type & arch_register_type_state) {
                                ARR_APP1(arch_register_t*, stateregs, (arch_register_t *)reg);
                        }
                }
        }

        n = ARR_LEN(env->calls);
        n_states = ARR_LEN(stateregs);
        for (i = 0; i < n; ++i) {
                int s, arity;
                ir_node *call = env->calls[i];

                arity = get_irn_arity(call);

                /* the state reg inputs are the last n inputs of the calls */
                for (s = 0; s < n_states; ++s) {
                        int inp = arity - n_states + s;
                        const arch_register_t *reg = stateregs[s];
                        ir_node *regnode = be_abi_reg_map_get(env->regs, reg);

                        set_irn_n(call, inp, regnode);
                }
        }

        DEL_ARR_F(stateregs);
}

/**
 * Create a trampoline entity for the given method.
 */
static ir_entity *create_trampoline(be_main_env_t *be, ir_entity *method)
{
        ir_type   *type   = get_entity_type(method);
        ident     *old_id = get_entity_ld_ident(method);
        ident     *id     = id_mangle3("", old_id, "$stub");
        ir_type   *parent = be->pic_trampolines_type;
        ir_entity *ent    = new_entity(parent, old_id, type);
        set_entity_ld_ident(ent, id);
        set_entity_visibility(ent, ir_visibility_private);

        return ent;
}

/**
 * Returns the trampoline entity for the given method.
 */
static ir_entity *get_trampoline(be_main_env_t *env, ir_entity *method)
{
        ir_entity *result = pmap_get(ir_entity, env->ent_trampoline_map, method);
        if (result == NULL) {
                result = create_trampoline(env, method);
                pmap_insert(env->ent_trampoline_map, method, result);
        }

        return result;
}

static ir_entity *create_pic_symbol(be_main_env_t *be, ir_entity *entity)
{
        ident     *old_id = get_entity_ld_ident(entity);
        ident     *id     = id_mangle3("", old_id, "$non_lazy_ptr");
        ir_type   *e_type = get_entity_type(entity);
        ir_type   *type   = new_type_pointer(e_type);
        ir_type   *parent = be->pic_symbols_type;
        ir_entity *ent    = new_entity(parent, old_id, type);
        set_entity_ld_ident(ent, id);
        set_entity_visibility(ent, ir_visibility_private);

        return ent;
}

static ir_entity *get_pic_symbol(be_main_env_t *env, ir_entity *entity)
{
        ir_entity *result = pmap_get(ir_entity, env->ent_pic_symbol_map, entity);
        if (result == NULL) {
                result = create_pic_symbol(env, entity);
                pmap_insert(env->ent_pic_symbol_map, entity, result);
        }

        return result;
}



/**
 * Returns non-zero if a given entity can be accessed using a relative address.
 */
static int can_address_relative(ir_entity *entity)
{
        return entity_has_definition(entity) && !(get_entity_linkage(entity) & IR_LINKAGE_MERGE);
}

static ir_node *get_pic_base(ir_graph *irg)
{
        const arch_env_t *arch_env = be_get_irg_arch_env(irg);
        if (arch_env->impl->get_pic_base == NULL)
                return NULL;
        return arch_env->impl->get_pic_base(irg);
}

/** patches SymConsts to work in position independent code */
static void fix_pic_symconsts(ir_node *node, void *data)
{
        ir_graph         *irg = get_irn_irg(node);
        be_main_env_t    *be  = be_get_irg_main_env(irg);
        ir_node          *pic_base;
        ir_node          *add;
        ir_node          *block;
        ir_mode          *mode;
        ir_node          *load;
        ir_node          *load_res;
        int               arity, i;
        (void) data;

        arity = get_irn_arity(node);
        for (i = 0; i < arity; ++i) {
                dbg_info  *dbgi;
                ir_node   *pred = get_irn_n(node, i);
                ir_entity *entity;
                ir_entity *pic_symbol;
                ir_node   *pic_symconst;

                if (!is_SymConst(pred))
                        continue;

                entity = get_SymConst_entity(pred);
                block  = get_nodes_block(pred);

                /* calls can jump to relative addresses, so we can directly jump to
                   the (relatively) known call address or the trampoline */
                if (i == 1 && is_Call(node)) {
                        ir_entity *trampoline;
                        ir_node   *trampoline_const;

                        if (can_address_relative(entity))
                                continue;

                        dbgi             = get_irn_dbg_info(pred);
                        trampoline       = get_trampoline(be, entity);
                        trampoline_const = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
                                                                    trampoline);
                        set_irn_n(node, i, trampoline_const);
                        continue;
                }

                /* everything else is accessed relative to EIP */
                mode     = get_irn_mode(pred);
                pic_base = get_pic_base(irg);

                /* all ok now for locally constructed stuff */
                if (can_address_relative(entity)) {
                        ir_node *add = new_r_Add(block, pic_base, pred, mode);

                        /* make sure the walker doesn't visit this add again */
                        mark_irn_visited(add);
                        set_irn_n(node, i, add);
                        continue;
                }

                /* get entry from pic symbol segment */
                dbgi         = get_irn_dbg_info(pred);
                pic_symbol   = get_pic_symbol(be, entity);
                pic_symconst = new_rd_SymConst_addr_ent(dbgi, irg, mode_P_code,
                                                        pic_symbol);
                add = new_r_Add(block, pic_base, pic_symconst, mode);
                mark_irn_visited(add);

                /* we need an extra indirection for global data outside our current
                   module. The loads are always safe and can therefore float
                   and need no memory input */
                load     = new_r_Load(block, get_irg_no_mem(irg), add, mode, cons_floats);
                load_res = new_r_Proj(load, mode, pn_Load_res);

                set_irn_n(node, i, load_res);
        }
}

void be_abi_introduce(ir_graph *irg)
{
        ir_node          *old_frame   = get_irg_frame(irg);
        const arch_env_t *arch_env    = be_get_irg_arch_env(irg);
        ir_entity        *entity      = get_irg_entity(irg);
        ir_type          *method_type = get_entity_type(entity);
        be_irg_t         *birg        = be_birg_from_irg(irg);
        struct obstack   *obst        = &birg->obst;
        ir_node          *dummy       = new_r_Dummy(irg,
                                                    arch_env->sp->reg_class->mode);
        unsigned          r;

        /* determine allocatable registers */
        assert(birg->allocatable_regs == NULL);
        birg->allocatable_regs = rbitset_obstack_alloc(obst, arch_env->n_registers);
        for (r = 0; r < arch_env->n_registers; ++r) {
                const arch_register_t *reg = &arch_env->registers[r];
                if ( !(reg->type & arch_register_type_ignore)) {
                        rbitset_set(birg->allocatable_regs, r);
                }
        }

        /* Break here if backend provides a custom API. */

        be_abi_irg_t env;
        env.keep_map     = pmap_create();
        env.call         = be_abi_call_new();
        arch_env_get_call_abi(arch_env, method_type, env.call);

        env.init_sp = dummy;
        env.calls   = NEW_ARR_F(ir_node*, 0);

        assure_edges(irg);

        if (be_options.pic) {
                irg_walk_graph(irg, fix_pic_symconsts, NULL, NULL);
        }

        /* Lower all call nodes in the IRG. */
        process_calls(irg, &env);

        /* Process the IRG */
        modify_irg(irg, &env);

        /* fix call inputs for state registers */
        fix_call_state_inputs(irg, &env);

        be_abi_call_free(env.call);

        /* We don't need the keep map anymore. */
        pmap_destroy(env.keep_map);

        /* calls array is not needed anymore */
        DEL_ARR_F(env.calls);

        /* reroute the stack origin of the calls to the true stack origin. */
        exchange(dummy, env.init_sp);
        exchange(old_frame, get_irg_frame(irg));

        pmap_destroy(env.regs);
}

void be_put_allocatable_regs(const ir_graph *irg,
                             const arch_register_class_t *cls, bitset_t *bs)
{
        be_irg_t *birg             = be_birg_from_irg(irg);
        unsigned *allocatable_regs = birg->allocatable_regs;
        unsigned  i;

        assert(bitset_size(bs) == cls->n_regs);
        bitset_clear_all(bs);
        for (i = 0; i < cls->n_regs; ++i) {
                const arch_register_t *reg = &cls->regs[i];
                if (rbitset_is_set(allocatable_regs, reg->global_index))
                        bitset_set(bs, i);
        }
}

unsigned be_get_n_allocatable_regs(const ir_graph *irg,
                                   const arch_register_class_t *cls)
{
        bitset_t *bs = bitset_alloca(cls->n_regs);
        be_put_allocatable_regs(irg, cls, bs);
        return bitset_popcount(bs);
}

void be_set_allocatable_regs(const ir_graph *irg,
                             const arch_register_class_t *cls,
                             unsigned *raw_bitset)
{
        be_irg_t *birg             = be_birg_from_irg(irg);
        unsigned *allocatable_regs = birg->allocatable_regs;
        unsigned  i;

        rbitset_clear_all(raw_bitset, cls->n_regs);
        for (i = 0; i < cls->n_regs; ++i) {
                const arch_register_t *reg = &cls->regs[i];
                if (rbitset_is_set(allocatable_regs, reg->global_index))
                        rbitset_set(raw_bitset, i);
        }
}

BE_REGISTER_MODULE_CONSTRUCTOR(be_init_abi)
void be_init_abi(void)
{
        FIRM_DBG_REGISTER(dbg, "firm.be.abi");
}