[libfirm] / ir / opt / return.c

/*
 * Copyright (C) 1995-2008 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   Normalize returns.
 * @author  Michael Beck
 * @version $Id$
 */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include "iroptimize.h"
#include "irgraph_t.h"
#include "ircons_t.h"
#include "irnode_t.h"
#include "irgmod.h"

#define set_bit(n)      (returns[(n) >> 3] |= 1 << ((n) & 7))
#define get_bit(n)      (returns[(n) >> 3] & (1 << ((n) & 7)))

#undef IMAX
#define IMAX(a, b)       ((a) > (b) ? (a) : (b))

/*
 * Normalize the Returns of a graph by creating a new End block
 * with One Return(Phi).
 * This is the preferred input for the if-conversion.
 *
 * In pseudocode, it means:
 *
 * if (a)
 *   return b;
 * else
 *   return c;
 *
 * is transformed into
 *
 * if (a)
 *   res = b;
 * else
 *   res = c;
 * return res;
 */
void normalize_one_return(ir_graph *irg) {
        ir_node *endbl = get_irg_end_block(irg);
        int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
        unsigned char *returns;
        ir_node **in, **retvals, **endbl_in;

        ir_node *block;

        /* look, if we have more than one return */
        n = get_Block_n_cfgpreds(endbl);
        if (n <= 0) {
                /* The end block has no predecessors, we have an endless
                   loop. In that case, no returns exists. */
                return;
        }

        returns = alloca((n + 7) >> 3);
        memset(returns, 0, (n + 7) >> 3);

        for (n_rets = i = 0; i < n; ++i) {
                ir_node *node = get_Block_cfgpred(endbl, i);

                if (is_Return(node)) {
                        ++n_rets;

                        set_bit(i);

                        if (n_ret_vals < 0)
                                n_ret_vals = get_irn_arity(node);
                }
        }

        /* there should be at least one Return node in Firm */
        if (n_rets <= 1)
                return;

        in       = alloca(sizeof(*in)       * IMAX(n_rets, n_ret_vals));
        retvals  = alloca(sizeof(*retvals)  * n_rets * n_ret_vals);
        endbl_in = alloca(sizeof(*endbl_in) * n);

        last_idx = 0;
        for (j = i = 0; i < n; ++i) {
                ir_node *ret = get_Block_cfgpred(endbl, i);

                if (get_bit(i)) {
                        ir_node *block = get_nodes_block(ret);

                        /* create a new Jmp for every Ret and place the in in */
                        in[j] = new_r_Jmp(irg, block);

                        /* save the return values and shuffle them */
                        for (k = 0; k < n_ret_vals; ++k)
                                retvals[j + k*n_rets] = get_irn_n(ret, k);

                        ++j;
                } else
                        endbl_in[last_idx++] = ret;
        }

        /* ok, create a new block with all created in's */
        block = new_r_Block(irg, n_rets, in);

        /* now create the Phi nodes */
        for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
                int k;
                ir_node *first;
                /* the return values are already shuffled */

                /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
                   but NOT, if a is Unknown. Here, we known that this case can be optimize also,
                   so do it here */
                first = retvals[j + 0];
                for (k = 1; k < n_rets; ++k) {
                        if (retvals[j + k] != first) {
                                first = NULL;
                                break;
                        }
                }
                if (first)
                        in[i] = first;
                else
                        in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
        }

        endbl_in[last_idx++] = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);

        set_irn_in(endbl, last_idx, endbl_in);

        /* invalidate analysis information:
         * a new Block was added, so dominator, outs and loop are inconsistent,
         * trouts and callee-state should be still valid
         */
        set_irg_doms_inconsistent(irg);
        set_irg_outs_inconsistent(irg);
        set_irg_extblk_inconsistent(irg);
        set_irg_loopinfo_inconsistent(irg);
}

/**
 * Check, whether a Return can be moved on block upwards.
 *
 * In a block with a Return, all live nodes must be linked
 * with the Return, otherwise they are dead (because the Return leaves
 * the graph, so no more users of the other nodes can exists.
 *
 * We can move a Return, if it's predecessors are Phi nodes or
 * comes from another block. In the later case, it is always possible
 * to move the Return one block up, because the predecessor block must
 * dominate the Return block (SSA) and then it dominates the predecessor
 * block of the Return block as well.
 *
 * All predecessors of the Return block must be Jmp's of course, or we
 * cannot move it up, so we add blocks if needed.
 */
static int can_move_ret(ir_node *ret) {
        ir_node *retbl = get_nodes_block(ret);
        int i, n = get_irn_arity(ret);

        for (i = 0; i < n; ++i) {
                ir_node *pred = get_irn_n(ret, i);

                if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
                        /* first condition failed, found a non-Phi predecessor
                         * then is in the Return block */
                        return 0;
                }
        }

        /* check, that predecessors are Jmps */
        n = get_Block_n_cfgpreds(retbl);
        if (n <= 1)
                return 0;
        for (i = 0; i < n; ++i) {
                ir_node *pred = get_Block_cfgpred(retbl, i);

                pred = skip_Tuple(pred);
                if (! is_Jmp(pred) && !is_Bad(pred)) {
                        /* simply place a new block here */
                        ir_graph *irg  = get_irn_irg(retbl);
                        ir_node *block = new_r_Block(irg, 1, &pred);
                        ir_node *jmp   = new_r_Jmp(irg, block);
                        set_Block_cfgpred(retbl, i, jmp);
                }
        }
        return 1;
}

/*
 * Normalize the Returns of a graph by moving
 * the Returns upwards as much as possible.
 * This might be preferred for code generation.
 *
 * In pseudocode, it means:
 *
 * if (a)
 *   res = b;
 * else
 *   res = c;
 * return res;
 *
 * is transformed into
 *
 * if (a)
 *   return b;
 * else
 *   return c;
 */
void normalize_n_returns(ir_graph *irg) {
        int i, j, n, n_rets, n_finals, n_ret_vals;
        ir_node *list  = NULL;
        ir_node *final = NULL;
        ir_node **in;
        ir_node *endbl = get_irg_end_block(irg);
        ir_node *end;

        /*
         * First, link all returns:
         * These must be predecessors of the endblock.
         * Place Returns that can be moved on list, all others
         * on final.
         */
        n = get_Block_n_cfgpreds(endbl);
        for (n_finals = n_rets = i = 0; i < n; ++i) {
                ir_node *ret = get_Block_cfgpred(endbl, i);

                if (is_Return(ret) && can_move_ret(ret)) {
                        /*
                         * Ok, all conditions met, we can move this Return, put it
                         * on our work list.
                         */
                        set_irn_link(ret, list);
                        list = ret;
                        ++n_rets;
                } else {
                        /* Put all nodes that are not changed on the final list. */
                        set_irn_link(ret, final);
                        final = ret;
                        ++n_finals;
                }
        }

        if (n_rets <= 0)
                return;

        /*
         * Now move the Returns upwards. We move always one block up (and create n
         * new Returns), than we check if a newly created Return can be moved even further.
         * If yes, we simply add it to our work list, else to the final list.
         */
        end        = get_irg_end(irg);
        n_ret_vals = get_irn_arity(list);
        in         = alloca(sizeof(*in) * n_ret_vals);
        while (list) {
                ir_node *ret   = list;
                ir_node *block = get_nodes_block(ret);
                ir_node *phiM;

                list = get_irn_link(ret);
                --n_rets;

                n = get_Block_n_cfgpreds(block);
                for (i = 0; i < n; ++i) {
                        ir_node *jmp = get_Block_cfgpred(block, i);
                        ir_node *new_bl, *new_ret;

                        if (is_Bad(jmp))
                                continue;
                        assert(is_Jmp(jmp));

                        new_bl = get_nodes_block(jmp);

                        /* create the in-array for the new Return */
                        for (j = 0; j < n_ret_vals; ++j) {
                                ir_node *pred = get_irn_n(ret, j);

                                in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
                        }

                        new_ret = new_r_Return(irg, new_bl, in[0], n_ret_vals - 1, &in[1]);

                        if (! is_Bad(new_ret)) {
                                /*
                                 * The newly created node might be bad, if we
                                 * create it in a block with only Bad predecessors.
                                 * In that case ignore this block.
                                 *
                                 * We could even kill the jmp then ...
                                 */
                                if (can_move_ret(new_ret)) {
                                        set_irn_link(new_ret, list);
                                        list = new_ret;
                                        ++n_rets;
                                } else {
                                        set_irn_link(new_ret, final);
                                        final = new_ret;
                                        ++n_finals;
                                }
                        }

                        /* remove the Jmp, we have placed a Return here */
                        exchange(jmp, new_r_Bad(irg));
                }

                /*
                 * if the memory of the old Return is a PhiM, remove it
                 * from the keep-alives, or it will keep the block which
                 * will crash the dominator algorithm.
                 */
                phiM = get_Return_mem(ret);
                if (is_Phi(phiM)) {
                        n = get_End_n_keepalives(end);
                        for (i = 0; i < n; ++i) {
                                if (get_End_keepalive(end, i) == phiM) {
                                        set_End_keepalive(end, i, new_r_Bad(irg));
                                        break;
                                }
                        }
                }
        }

        /*
         * Last step: Create a new endblock, with all nodes on the final
         * list as predecessors.
         */
        in = alloca(sizeof(*in) * n_finals);

        for (i = 0; final; ++i, final = get_irn_link(final))
                in[i] = final;

        exchange(endbl, new_r_Block(irg, n_finals, in));

        /* the end block is not automatically skipped, so do it here */
        set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));

        /* Invalidate analysis information:
         * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
         * trouts and callee-state should be still valid
         */
        set_irg_doms_inconsistent(irg);
        set_irg_extblk_inconsistent(irg);  /* may not be needed */
        set_irg_outs_inconsistent(irg);
        set_irg_loopinfo_inconsistent(current_ir_graph);
}