Added if conversion

[libfirm] / ir / opt / ifconv.c

/**
 * If conversion.
 * Make Mux nodes from Conds where it its possible.
 * @author Sebastian Hack
 * @date 4.2.2005
 */

#include <stdlib.h>
#include <alloca.h>

#include "irgraph_t.h"
#include "irnode_t.h"
#include "irmode_t.h"
#include "ircons_t.h"
#include "irdom_t.h"

#include "ifconv.h"
#include "irflag_t.h"

#include "debug.h"
#include "set.h"

#define MAX_DEPTH 4

/*
 * Mux optimization routines.
 */

#if 0
static ir_node *local_optimize_mux(ir_node *mux)
{
        int i, n;
        ir_node *res = mux;
  ir_node *sel = get_Mux_sel(mux);
        ir_node *cmp = skip_Proj(sel);

        /* Optimize the children  */
        for(i = 1, n = get_irn_arity(mux); i < n; ++i) {
                ir_node *operand = get_irn_n(mux, i);
                if(get_irn_op(operand) == op_Mux)
                        optimize_mux(operand);
        }

        /* If we have no cmp above the mux, get out. */
        if(is_Proj(sel) && get_irn_mode(sel) == mode_b && get_irn_opcode(cmp) == iro_Cmp) {

                pnc_number cc = get_Proj_proj(sel);
                ir_mode *mode = get_irn_mode(mux);
                ir_node *block = get_nodes_block(n);
                ir_node *cmp_left = get_Cmp_left(cmp);
                ir_node *cmp_right = get_Cmp_right(cmp);
                ir_node *mux_true = get_Mux_true(mux);
                ir_node *mux_false = get_Mux_false(mux);

                /*
                 * Check for comparisons with signed integers.
                 */
                if(mode_is_int(mode)                                    /* We need an integral mode */
                                && mode_is_signed(mode)   /* which is signed */
                                && cc == Lt) {                                          /* and have to compare for < */

                        /*
                         * Mux(x:T < 0, -1, 0) -> Shrs(x, sizeof_bits(T) - 1)
                         * Conditions:
                         * T must be signed.
                         */
                        if(classify_Const(cmp_right) == CNST_NULL
                                && classify_Const(mux_true) == CNST_ALL_ONE
                                && classify_Const(mux_false) == CNST_NULL) {

                                ir_mode *u_mode = find_unsigned_mode(mode);

                                res = new_r_Shrs(current_ir_graph, block, cmp_left,
                                                new_r_Const_long(current_ir_graph, block, u_mode,
                                                        get_mode_size_bits(mode) - 1),
                                                mode);
                        }

                        /*
                         * Mux(0 < x:T, 1, 0) -> Shr(-x, sizeof_bits(T) - 1)
                         * Conditions:
                         * T must be signed.
                         */
                        else if(classify_Const(cmp_left) == CNST_NULL
                                && classify_Const(mux_true) == CNST_ONE
                                && classify_Const(mux_false) == CNST_NULL) {

                                ir_mode *u_mode = find_unsigned_mode(mode);

                                res = new_r_Shr(current_ir_graph, block,

                                                /* -x goes to 0 - x in Firm (cmp_left is 0, see the if) */
                                                new_r_Sub(current_ir_graph, block, cmp_left, cmp_right, mode),

                                                /* This is sizeof_bits(T) - 1 */
                                                new_r_Const_long(current_ir_graph, block, u_mode,
                                                        get_mode_size_bits(mode) - 1),
                                                mode);
                        }
                }
        }

        return res;
}
#endif

static tarval *get_value_or(ir_node *cnst, tarval *or)
{
        return get_irn_op(cnst) == op_Const ? get_Const_tarval(cnst) : or;
}


static ir_node *optimize_mux_chain(ir_node *mux)
{
        int i;
        ir_node *res;
        ir_node *ops[2];
        ir_mode *mode;
        tarval *null;
        tarval *minus_one;

        if(get_irn_op(mux) != op_Mux)
                return mux;

        res = mux;
        mode = get_irn_mode(mux);
        null = get_tarval_null(mode);
        minus_one = tarval_sub(null, get_tarval_one(mode));

        ops[0] = get_Mux_false(mux);
        ops[1] = get_Mux_true(mux);

        for(i = 0; i < 2; ++i) {
                ir_node *a, *b, *d;
                tarval *tva, *tvb, *tvd;
                ir_node *child_mux;

                /*
                 * This is the or case, the child mux is the false operand
                 * of the parent mux.
                 *
                 * mux(c1, mux(c2, a, b), d)
                 *
                 * This can be made into:
                 * 1) mux(c1, 0, d) | mux(c2, a, b)
                 *    if a | d == d and b | d == d
                 *
                 * 2) mux(c1, -1, d) & mux(c2, a, b)
                 *    if a & d == d and a & b == b
                 */
                if(get_irn_op(ops[i]) == op_Mux) {

                        child_mux = ops[i];
                        a = get_Mux_false(child_mux);
                        b = get_Mux_true(child_mux);
                        d = ops[1 - i];

                        /* Try the or stuff */
                        tva = get_value_or(a, minus_one);
                        tvb = get_value_or(b, minus_one);
                        tvd = get_value_or(d, null);

                        if(tarval_cmp(tarval_or(tva, tvd), tvd) == Eq
                                        && tarval_cmp(tarval_or(tvb, tvd), tvd) == Eq) {

                                ops[i] = new_Const(mode, null);
                                res = new_r_Or(current_ir_graph, get_nodes_block(mux),
                                                mux, child_mux, mode);
                                break;
                        }

                        /* If the or didn't go, try the and stuff */
                        tva = get_value_or(a, null);
                        tvb = get_value_or(b, null);
                        tvd = get_value_or(d, minus_one);

                        if(tarval_cmp(tarval_and(tva, tvd), tvd) == Eq
                                        && tarval_cmp(tarval_and(tvb, tvd), tvd) == Eq) {

                                ops[i] = new_Const(mode, minus_one);
                                res = new_r_And(current_ir_graph, get_nodes_block(mux),
                                                mux, child_mux, mode);
                                break;
                        }
                }
        }

        set_irn_n(mux, 1, optimize_mux_chain(ops[0]));
        set_irn_n(mux, 2, optimize_mux_chain(ops[1]));

        return res;
}


/***********************************************************
 * The If conversion itself.
 ***********************************************************/

/**
 * Default options.
 */
static opt_if_conv_info_t default_info = {
        4
};

/** THe debugging module. */
static firm_dbg_module_t *dbg;

/**
 * A simple check for sde effects upton an opcode of a ir node.
 * @param irn The ir node to check,
 * @return 1 if the opcode itself may produce side effects, 0 if not.
 */
static INLINE int has_side_effects(const ir_node *irn)
{
        opcode opc = get_irn_opcode(irn);

        if(opc == iro_Cmp)
                return 0;

        return !mode_is_datab(get_irn_mode(irn));
}

/**
 * Decdies, if a given expression and its subexpressions
 * (to certain, also given extent) can be moved to a block.
 * @param expr The expression to examine.
 * @param block The block where the expression should go.
 * @param depth The current depth, passed recursively. Use 0 for
 * non-recursive calls.
 * @param max_depth The maximum depth to which the expression should be
 * examined.
 */
static int _can_move_to(ir_node *expr, ir_node *dest_block, int depth, int max_depth)
{
        int i, n;
        int res = 1;
        ir_node *expr_block = get_nodes_block(expr);


        /*
         * If we are forced to look too deep into the expression,
         * treat it like it could not be moved.
         */
        if(depth >= max_depth) {
                res = 0;
                goto end;
        }

        /*
         * If the block of the expression dominates the specified
         * destination block, it does not matter if the expression
         * has side effects or anything else. It is executed on each
         * path the destination block is reached.
         */
        if(block_dominates(expr_block, dest_block))
                goto end;

        /*
         * This should be superflous and could be converted into a assertion.
         * The destination block _must_ dominate the block of the expression,
         * else the expression could be used without its definition.
         */
        if(!block_dominates(dest_block, expr_block)) {
                res = 0;
                goto end;
        }

        /*
         * Surely, if the expression does not have a data mode, it is not
         * movable. Perhaps onw should also test the floating property of
         * the opcode/node.
         */
        if(has_side_effects(expr)) {
                res = 0;
                goto end;
        }

        /*
         * If the node looks alright so far, look at its operands and
         * check them out. If one of them cannot be moved, this one
         * cannot be moved either.
         */
        for(i = 0, n = get_irn_arity(expr); i < n; ++i) {
                ir_node *op = get_irn_n(expr, i);
                int new_depth = is_Proj(op) ? depth : depth + 1;
                if(!_can_move_to(op, dest_block, new_depth, max_depth)) {
                        res = 0;
                        goto end;
                }
        }

end:
        DBG((dbg, LEVEL_5, "\t\t\tcan move to(%d) %n: %d\n", depth, expr, res));

        return res;
}

/**
 * Convenience function for _can_move_to.
 * Checks, if an expression can be moved to another block. The check can
 * be limited to a expression depth meaning if we need to crawl in
 * deeper into an expression than a given threshold to examine if
 * it can be moved, the expression is rejected and the test returns
 * false.
 * @param expr The expression to check for.
 * @param dest_block The destination block you want @p expr to be.
 * @param max_depth The maximum depth @p expr should be investigated.
 * @return 1, if the expression can be moved to the destination block,
 * 0 if not.
 */
static INLINE int can_move_to(ir_node *expr, ir_node *dest_block, int max_depth)
{
        return _can_move_to(expr, dest_block, 0, max_depth);
}

static void move_to(ir_node *expr, ir_node *dest_block)
{
        int i, n;
        ir_node *expr_block = get_nodes_block(expr);

        /*
         * If we reached the dominator, we are done.
         * We will never put code through the dominator
         */
        if(block_dominates(expr_block, dest_block))
                return;

        for(i = 0, n = get_irn_arity(expr); i < n; ++i)
                move_to(get_irn_n(expr, i), dest_block);

        set_nodes_block(expr, dest_block);
}

/**
 * Information about a cond node.
 */
typedef struct _cond_t {
        ir_node *cond;                          /**< The cond node. */
        ir_node *mux;                                   /**< The mux node, that will be generated for this cond. */

        /**
         * Information about the both 'branches'
         * (true and false), the cond creates.
         */
        struct {
                int pos;                                                /**< Number of the predecessor of the
                                                                                                        phi block by which this branch is
                                                                                                        reached. It is -1, if this branch is
                                                                                                        only reached through another cond. */

                ir_node *masked_by;     /**< If this cond's branch is only reached
                                                                                                        through another cond, we store this
                                                                                                        cond ir_node here. */
        } cases[2];
} cond_t;

/**
 * Compare two conds for use in a firm set.
 * Two cond_t's are equal, if they designate the same cond node.
 * @param a A cond_t
 * @param b Another one.
 * @param size Not used.
 * @return 0 (!) if they are equal, != 0 otherwise.
 */
static int cond_cmp(const void *a, const void *b, size_t size)
{
        const cond_t *x = a;
        const cond_t *y = b;
        return x->cond != y->cond;
}

/**
 * @see find_conds.
 */
static void _find_conds(ir_node *irn, ir_node *start_block,
                ir_node *dominator, ir_node *masked_by, int pos, int depth, set *conds)
{
        ir_node *block;

        block = get_nodes_block(irn);

        if(block_dominates(dominator, block)) {
                ir_node *cond = NULL;
                int i, n;

                /* We ran into a loop, since we saw the start block twice. */
                if(block == start_block && depth > 0)
                        return;

                /* check, if we're on a ProjX */
                if(is_Proj(irn) && get_irn_mode(irn) == mode_X) {

                        int proj = get_Proj_proj(irn);
                        cond = get_Proj_pred(irn);

                        /* Check, if the pred of the proj is a Cond
                         * with a Projb as selector. */
                        if(get_irn_opcode(cond) == iro_Cond
                                        && get_irn_mode(get_Cond_selector(cond)) == mode_b) {

                                cond_t *res, c;

                                c.cond = cond;
                                c.mux = NULL;
                                c.cases[0].pos = -1;
                                c.cases[1].pos = -1;

                                /* get or insert the cond info into the set. */
                                res = set_insert(conds, &c, sizeof(c), HASH_PTR(cond));

                                /*
                                 * Link it to the cond ir_node. We need that later, since
                                 * one cond masks the other we want to retreive the cond_t
                                 * data from the masking cond ir_node.
                                 */
                                set_irn_link(cond, res);

                                /*
                                 * Set masked by (either NULL or another cond node.
                                 * If this cond is truly masked by another one, set
                                 * the position of the actually investigated branch
                                 * to -1. Since the cond is masked by another one,
                                 * there could be more ways from the start block
                                 * to this branch, so we choose -1.
                                 */
                                res->cases[proj].masked_by = masked_by;
                                if(!masked_by)
                                        res->cases[proj].pos = pos;

                                DBG((dbg, LEVEL_5, "found cond %n (%s branch) for pos %d in block %n reached by %n\n",
                                                        cond, get_Proj_proj(irn) ? "true" : "false", pos, block, masked_by));
                        }
                }

                /* Search recursively from this cond. */
                for(i = 0, n = get_irn_arity(block); i < n; ++i) {
                        ir_node *pred = get_irn_n(block, i);

                        /*
                         * If the depth is 0 (the first recursion), we set the pos to
                         * the current viewed predecessor, else we adopt the position
                         * as given by the caller. We also increase the depth for the
                         * recursively called functions.
                         */
                        _find_conds(pred, start_block, dominator, cond, depth == 0 ? i : pos, depth + 1, conds);
                }
        }
}

/**
 * A convenience function for _find_conds.
 * It sets some parameters needed for recursion to appropriate start
 * values. Always use this function.
 * @param irn The node to start looking for conds from. This might
 *      be the phi node we are investigating.
 * @param dominator The dominator up to which we want to look for conds.
 * @param conds The set to record the found conds in.
 */
static INLINE void find_conds(ir_node *irn, ir_node *dominator, set *conds)
{
        _find_conds(irn, get_nodes_block(irn), dominator, NULL, 0, 0, conds);
}


/**
 * Make the mux for a given cond.
 * @param phi The phi node which shall be replaced by a mux.
 * @param dom The block where the muxes shall be placed.
 * @param cond The cond information.
 * @return The mux node made for this cond.
 */
static ir_node *make_mux_on_demand(ir_node *phi, ir_node *dom, cond_t *cond)
{
        int i;
        ir_node *projb = get_Cond_selector(cond->cond);
        ir_node *operands[2];

        for(i = 0; i < 2; ++i) {

                /*
                 * If this cond branch is masked by another cond, make the mux
                 * for that cond first, since the mux for this cond takes
                 * it as an operand.
                 */
                if(cond->cases[i].masked_by) {
                        cond_t *masking_cond = get_irn_link(cond->cases[i].masked_by);
                        operands[i] = make_mux_on_demand(phi, dom, masking_cond);
                }

                /*
                 * If this cond branch is not masked by another cond, take
                 * the corresponding phi operand as an operand to the mux.
                 */
                else {
                        assert(cond->cases[i].pos >= 0);
                        operands[i] = get_irn_n(phi, cond->cases[i].pos);
                }

                /* Move the selected operand to the dominator block. */
                move_to(operands[i], dom);
        }

        /* Move the comparison expression of the cond to the dominator. */
        move_to(projb, dom);

        /* Make the mux. */
        cond->mux = new_r_Mux(current_ir_graph, dom, projb,
                        operands[0], operands[1], get_irn_mode(operands[0]));

        return cond->mux;
}

/**
 * Examine a phi node if it can be replaced by some muxes.
 * @param irn A phi node.
 * @param info Parameters for the if conversion algorithm.
 */
static void check_out_phi(ir_node *irn, opt_if_conv_info_t *info)
{
        int max_depth = info->max_depth;
        int i;
        ir_node *block;
        int arity;
        ir_node *idom;
        ir_node *mux = NULL;

        cond_t **conds;
        cond_t *cond;
        cond_t *largest_cond;
        set *cond_set;
        int n_conds = 0;

        if(!is_Phi(irn))
                return;

        block = get_nodes_block(irn);
        arity = get_irn_arity(irn);
        idom = get_Block_idom(block);

        assert(is_Phi(irn));
        assert(get_irn_arity(irn) == get_irn_arity(block));
        assert(arity > 0);

        cond_set = get_irn_link(block);
        assert(conds && "no cond set for this phi");

        DBG((dbg, LEVEL_5, "phi candidate: %n\n", irn));

        /*
         * Check, if we can move all operands of the
         * phi node to the dominator. Else exit.
         */
        for(i = 0; i < arity; ++i) {
                if(!can_move_to(get_irn_n(irn, i), idom, max_depth)) {
                        DBG((dbg, LEVEL_5, "cannot move operand %d of %n to %n\n", i, irn, idom));
                        return;
                }
        }

        n_conds = set_count(cond_set);

        /* This should never happen and can be turned into an assertion */
        if(n_conds == 0) {
                DBG((dbg, LEVEL_5, "no conds found. how can this be?"));
                return;
        }

        /*
         * Put all cond information structures into an array.
         * This is just done for convenience. It's not neccessary.
         */
        conds = alloca(n_conds * sizeof(conds[0]));
        for(i = 0, cond = set_first(cond_set); cond; cond = set_next(cond_set))
                conds[i++] = cond;

        /*
         * Check, if we can move the compare nodes of the conds to
         * the dominator.
         */
        for(i = 0; i < n_conds; ++i) {
                ir_node *projb = get_Cond_selector(conds[i]->cond);
                if(!can_move_to(projb, idom, max_depth)) {
                        DBG((dbg, LEVEL_5, "cannot move Projb %d of %n to %n\n", i, projb, idom));
                        return;
                }
        }

        /*
         * Find the largest cond (the one that dominates all others)
         * and start the mux generation from there.
         */
        largest_cond = conds[0];
        DBG((dbg, LEVEL_5, "\tlargest cond %n\n", largest_cond->cond));
        for(i = 1; i < n_conds; ++i) {
                ir_node *curr_largest_block = get_nodes_block(largest_cond->cond);
                ir_node *bl = get_nodes_block(conds[i]->cond);

                if(block_dominates(bl, curr_largest_block)) {
                        DBG((dbg, LEVEL_5, "\tnew largest cond %n\n", largest_cond->cond));
                        largest_cond = conds[i];
                }
        }

#if 0
        for(i = 0; i < n_conds; ++i) {
                cond_t *c = conds[i];
                DBG((dbg, LEVEL_5, "\tcond %n (t: (%d,%n), f: (%d,%n))\n", c->cond,
                                        c->cases[1].pos, c->cases[1].masked_by,
                                        c->cases[0].pos, c->cases[0].masked_by));
        }
#endif

        /*
         * Make the mux for the 'largest' cond. This will also
         * produce all other muxes.
         * @see make_mux_on_demand.
         */
        mux = make_mux_on_demand(irn, idom, largest_cond);

        mux = optimize_mux_chain(mux);
        /*
         * Set all preds of the phi node to the mux
         * for the 'largest' cond.
         */
        for(i = 0; i < arity; ++i)
                set_irn_n(irn, i, mux);
}

static void annotate_cond_info_pre(ir_node *irn, void *data)
{
        set_irn_link(irn, NULL);
}

static void annotate_cond_info_post(ir_node *irn, void *data)
{
        /*
         * Check, if the node is a phi
         * we then compute a set of conds which are reachable from this
         * phi's block up to its dominator.
         * The set is attached to the blocks link field.
         */
        if(is_Phi(irn)) {
                ir_node *block = get_nodes_block(irn);
                ir_node **phi_list_head = (ir_node **) data;

                set *conds = get_irn_link(block);

                /* If the set is not yet computed, do it now. */
                if(!conds) {
                        ir_node *idom = get_Block_idom(block);
                        conds = new_set(cond_cmp, 8);

                        /*
                         * Fill the set with conds we find on the way from
                         * the block to its dominator.
                         */
                        find_conds(irn, idom, conds);

                        /*
                         * If there where no suitable conds, delete the set
                         * immediately and reset the set pointer to NULL
                         */
                        if(set_count(conds) == 0) {
                                del_set(conds);
                                conds = NULL;
                        }
                }

                set_irn_link(block, conds);

                /*
                 * If this phi node has a set of conds reachable, enqueue
                 * the phi node in a list with its link field.
                 * Then, we do not have to walk the graph again. We can
                 * use the list to reach all phi nodes for which if conversion
                 * can be tested.
                 */
                if(conds) {
                        ir_node *old = *phi_list_head;
                        set_irn_link(irn, old);
                        *phi_list_head = irn;
                }

        }
}

static void free_sets(ir_node *irn, void *data)
{
        if(is_Block(irn) && get_irn_link(irn)) {
                set *conds = get_irn_link(irn);
                del_set(conds);
        }
}

void opt_if_conv(ir_graph *irg, opt_if_conv_info_t *params)
{
        opt_if_conv_info_t *p = params ? params : &default_info;
        ir_node *list_head = NULL;

        if(!get_opt_if_conversion())
                return;

        dbg = firm_dbg_register("firm.opt.ifconv");
        firm_dbg_set_mask(dbg, -1);

        compute_doms(irg);
        DBG((dbg, LEVEL_4, "if conversion for irg %s(%p)\n",
                                get_entity_name(get_irg_entity(irg)), irg));

        irg_walk_graph(irg, annotate_cond_info_pre, annotate_cond_info_post, &list_head);

        /* traverse the list of linked phis */
        while(list_head) {
                check_out_phi(list_head, p);
                list_head = get_irn_link(list_head);
        }

        irg_walk_graph(irg, free_sets, NULL, NULL);
}