[libfirm] / ir / be / belistsched.c

/**
 * Scheduling algorithms.
 * Just a simple list scheduling algorithm is here.
 * @date 20.10.2004
 * @author Sebastian Hack
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include <stdio.h>
#include <stdarg.h>
#include <string.h>
#include <limits.h>

#include "benode_t.h"

#include "obst.h"
#include "list.h"
#include "iterator.h"

#include "iredges_t.h"
#include "irgwalk.h"
#include "irnode_t.h"
#include "irmode_t.h"
#include "irdump.h"
#include "irprintf_t.h"
#include "debug.h"

#include "besched_t.h"
#include "beutil.h"
#include "belive_t.h"
#include "belistsched.h"
#include "bearch.h"
#include "bestat.h"

#define MAX(x,y) ((x) > (y) ? (x) : (y))
#define MIN(x,y) ((x) < (y) ? (x) : (y))

/**
 * Scheduling environment for the whole graph.
 */
typedef struct _sched_env_t {
        const list_sched_selector_t *selector;      /**< The node selector. */
        const arch_env_t *arch_env;                 /**< The architecture enviromnent. */
        const ir_graph *irg;                        /**< The graph to schedule. */
        void *selector_env;                         /**< A pointer to give to the selector. */
} sched_env_t;

#if 0
/*
 * Ugly global variable for the compare function
 * since qsort(3) does not pass an extra pointer.
 */
static ir_node *curr_bl = NULL;

static int cmp_usage(const void *a, const void *b)
{
        struct trivial_sched_env *env;
        const ir_node *p = a;
        const ir_node *q = b;
        int res = 0;

        res = is_live_end(env->curr_bl, a) - is_live_end(env->curr_bl, b);

        /*
         * One of them is live at the end of the block.
         * Then, that one shall be scheduled at after the other
         */
        if(res != 0)
                return res;


        return res;
}
#endif

/**
 * The trivial selector:
 * Just assure that branches are executed last, otherwise select
 * the first node ready.
 */
static ir_node *trivial_select(void *block_env, nodeset *ready_set)
{
        const arch_env_t *arch_env = block_env;
        ir_node *irn = NULL;
        int const_last = 0;

        /* assure that branches and constants are executed last */
        for (irn = nodeset_first(ready_set); irn; irn = nodeset_next(ready_set)) {
                arch_irn_class_t irn_class = arch_irn_classify(arch_env, irn);

                if (irn_class != arch_irn_class_branch && (const_last ? (irn_class != arch_irn_class_const) : 1)) {
                        nodeset_break(ready_set);
                        return irn;
                }
        }

        /* assure that constants are executed before branches */
        if (const_last) {
                for (irn = nodeset_first(ready_set); irn; irn = nodeset_next(ready_set)) {
                        if (arch_irn_classify(arch_env, irn) != arch_irn_class_branch) {
                                nodeset_break(ready_set);
                                return irn;
                        }
                }
        }


        /* at last: schedule branches */
        irn = nodeset_first(ready_set);
        nodeset_break(ready_set);

        return irn;
}

static void *trivial_init_graph(const list_sched_selector_t *vtab, const arch_env_t *arch_env, ir_graph *irg)
{
        return (void *) arch_env;
}

static void *trivial_init_block(void *graph_env, ir_node *bl)
{
        return graph_env;
}

static INLINE int must_appear_in_schedule(const list_sched_selector_t *sel, void *block_env, const ir_node *irn)
{
        int res = 0;

        if(sel->to_appear_in_schedule)
                res = sel->to_appear_in_schedule(block_env, irn);

        return res || to_appear_in_schedule(irn) || be_is_Keep(irn) || be_is_RegParams(irn);
}

static const list_sched_selector_t trivial_selector_struct = {
        trivial_init_graph,
        trivial_init_block,
        trivial_select,
        NULL,
        NULL,
        NULL
};

const list_sched_selector_t *trivial_selector = &trivial_selector_struct;

typedef struct _usage_stats_t {
        ir_node *irn;
        struct _usage_stats_t *next;
        int max_hops;
        int uses_in_block;      /**< Number of uses inside the current block. */
        int already_consumed;   /**< Number of insns using this value already
                                                          scheduled. */
} usage_stats_t;

typedef struct {
        const list_sched_selector_t *vtab;
        const arch_env_t *arch_env;
} reg_pressure_main_env_t;

typedef struct {
        struct obstack obst;
        const reg_pressure_main_env_t *main_env;
        usage_stats_t *root;
        nodeset *already_scheduled;
} reg_pressure_selector_env_t;

static INLINE usage_stats_t *get_or_set_usage_stats(reg_pressure_selector_env_t *env, ir_node *irn)
{
        usage_stats_t *us = get_irn_link(irn);

        if(!us) {
                us                   = obstack_alloc(&env->obst, sizeof(us[0]));
                us->irn              = irn;
                us->already_consumed = 0;
                us->max_hops         = INT_MAX;
                us->next             = env->root;
                env->root            = us;
                set_irn_link(irn, us);
        }

        return us;
}

static INLINE usage_stats_t *get_usage_stats(ir_node *irn)
{
        usage_stats_t *us = get_irn_link(irn);
        assert(us && "This node must have usage stats");
        return us;
}

static int max_hops_walker(reg_pressure_selector_env_t *env, ir_node *irn, ir_node *curr_bl, int depth, unsigned visited_nr)
{
        ir_node *bl = get_nodes_block(irn);
        /*
         * If the reached node is not in the block desired,
         * return the value passed for this situation.
         */
        if(get_nodes_block(irn) != bl)
                return block_dominates(bl, curr_bl) ? 0 : INT_MAX;

        /*
         * If the node is in the current block but not
         * yet scheduled, we keep on searching from that node.
         */
        if(!nodeset_find(env->already_scheduled, irn)) {
                int i, n;
                int res = 0;
                for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
                        ir_node *operand = get_irn_n(irn, i);

                        if(get_irn_visited(operand) < visited_nr) {
                                int tmp;

                                set_irn_visited(operand, visited_nr);
                                tmp = max_hops_walker(env, operand, bl, depth + 1, visited_nr);
                                res = MAX(tmp, res);
                        }
                }

                return res;
        }

        /*
         * If the node is in the current block and scheduled, return
         * the depth which indicates the number of steps to the
         * region of scheduled nodes.
         */
        return depth;
}

static int compute_max_hops(reg_pressure_selector_env_t *env, ir_node *irn)
{
        ir_node *bl   = get_nodes_block(irn);
        ir_graph *irg = get_irn_irg(bl);
        int res       = 0;

        const ir_edge_t *edge;

        foreach_out_edge(irn, edge) {
                ir_node *user       = get_edge_src_irn(edge);
                unsigned visited_nr = get_irg_visited(irg) + 1;
                int max_hops;

                set_irg_visited(irg, visited_nr);
                max_hops = max_hops_walker(env, user, irn, 0, visited_nr);
                res      = MAX(res, max_hops);
        }

        return res;
}

static void *reg_pressure_graph_init(const list_sched_selector_t *vtab, const arch_env_t *arch_env, ir_graph *irg)
{
        reg_pressure_main_env_t *main_env = xmalloc(sizeof(main_env[0]));

        main_env->arch_env = arch_env;
        main_env->vtab     = vtab;
        irg_walk_graph(irg, firm_clear_link, NULL, NULL);

        return main_env;
}

static void *reg_pressure_block_init(void *graph_env, ir_node *bl)
{
        ir_node *irn;
        reg_pressure_selector_env_t *env  = xmalloc(sizeof(env[0]));

        obstack_init(&env->obst);
        env->already_scheduled = new_nodeset(32);
        env->root              = NULL;
        env->main_env          = graph_env;

        /*
         * Collect usage statistics.
         */
        sched_foreach(bl, irn) {
                if(must_appear_in_schedule(env->main_env->vtab, env, irn)) {
                        int i, n;

                        for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
                                ir_node *op = get_irn_n(irn, i);
                                if(must_appear_in_schedule(env->main_env->vtab, env, irn)) {
                                        usage_stats_t *us = get_or_set_usage_stats(env, irn);
                                        if(is_live_end(bl, op))
                                                us->uses_in_block = 99999;
                                        else
                                                us->uses_in_block++;
                                }
                        }
                }
        }

        return env;
}

static void reg_pressure_block_free(void *block_env)
{
        reg_pressure_selector_env_t *env = block_env;
        usage_stats_t *us;

        for(us = env->root; us; us = us->next)
                set_irn_link(us->irn, NULL);

        obstack_free(&env->obst, NULL);
        del_nodeset(env->already_scheduled);
        free(env);
}

static int get_result_hops_sum(reg_pressure_selector_env_t *env, ir_node *irn)
{
        int res = 0;
        if(get_irn_mode(irn) == mode_T) {
                const ir_edge_t *edge;

                foreach_out_edge(irn, edge)
                        res += get_result_hops_sum(env, get_edge_src_irn(edge));
        }

        else if(mode_is_data(get_irn_mode(irn)))
                res = compute_max_hops(env, irn);


        return res;
}

static INLINE int reg_pr_costs(reg_pressure_selector_env_t *env, ir_node *irn)
{
        int i, n;
        int sum = 0;

        for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
                ir_node *op = get_irn_n(irn, i);

                if(must_appear_in_schedule(env->main_env->vtab, env, op))
                        sum += compute_max_hops(env, op);
        }

        sum += get_result_hops_sum(env, irn);

        return sum;
}

static ir_node *reg_pressure_select(void *block_env, nodeset *ready_set)
{
        reg_pressure_selector_env_t *env = block_env;
        ir_node *irn, *res     = NULL;
        int curr_cost          = INT_MAX;

        assert(nodeset_count(ready_set) > 0);

        for (irn = nodeset_first(ready_set); irn; irn = nodeset_next(ready_set)) {
                /*
                        Ignore branch instructions for the time being.
                        They should only be scheduled if there is nothing else.
                */
                if (arch_irn_classify(env->main_env->arch_env, irn) != arch_irn_class_branch) {
                        int costs = reg_pr_costs(env, irn);
                        if (costs <= curr_cost) {
                                res       = irn;
                                curr_cost = costs;
                        }
                }
        }

        /*
                There was no result so we only saw a branch.
                Take it and finish.
        */

        if(!res) {
                res = nodeset_first(ready_set);
                nodeset_break(ready_set);

                assert(res && "There must be a node scheduled.");
        }

        nodeset_insert(env->already_scheduled, res);
        return res;
}

static const list_sched_selector_t reg_pressure_selector_struct = {
        reg_pressure_graph_init,
        reg_pressure_block_init,
        reg_pressure_select,
        NULL,
        reg_pressure_block_free,
        free
};

const list_sched_selector_t *reg_pressure_selector = &reg_pressure_selector_struct;

static void list_sched_block(ir_node *block, void *env_ptr);

void list_sched(const arch_env_t *arch_env, ir_graph *irg)
{
        sched_env_t env;

        memset(&env, 0, sizeof(env));
        env.selector = arch_env->isa->impl->get_list_sched_selector(arch_env->isa);
        env.arch_env = arch_env;
        env.irg      = irg;

        if(env.selector->init_graph)
                env.selector_env = env.selector->init_graph(env.selector, arch_env, irg);

        /* Assure, that the out edges are computed */
        edges_assure(irg);

        /* Schedule each single block. */
        irg_block_walk_graph(irg, list_sched_block, NULL, &env);

        if(env.selector->finish_graph)
                env.selector->finish_graph(env.selector_env);
}


/**
 * Environment for a block scheduler.
 */
typedef struct _block_sched_env_t {
        int curr_time;
        nodeset *ready_set;
        nodeset *already_scheduled;
        ir_node *block;
        const list_sched_selector_t *selector;
        void *selector_block_env;
        DEBUG_ONLY(firm_dbg_module_t *dbg;)
} block_sched_env_t;

/**
 * Try to put a node in the ready set.
 * @param env The block scheduler environment.
 * @param irn The node to make ready.
 * @return 1, if the node could be made ready, 0 else.
 */
static INLINE int make_ready(block_sched_env_t *env, ir_node *irn)
{
    int i, n;

    /* Blocks cannot be scheduled. */
    if(is_Block(irn))
        return 0;

    /*
     * Check, if the given ir node is in a different block as the
     * currently scheduled one. If that is so, don't make the node ready.
     */
    if(env->block != get_nodes_block(irn))
        return 0;

    for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
        ir_node *op = get_irn_n(irn, i);

        /* if irn is an End we have keep-alives and op might be a block, skip that */
        if (is_Block(op)) {
          assert(get_irn_op(irn) == op_End);
          continue;
        }

        /* If the operand is local to the scheduled block and not yet
         * scheduled, this nodes cannot be made ready, so exit. */
        if(!nodeset_find(env->already_scheduled, op) && get_nodes_block(op) == env->block)
            return 0;
    }

    DBG((env->dbg, LEVEL_2, "\tmaking ready: %+F\n", irn));
    nodeset_insert(env->ready_set, irn);

    return 1;
}

/**
 * Check, if a node is ready in a block schedule.
 * @param env The block schedule environment.
 * @param irn The node to check for.
 * @return 1 if the node was ready, 0 if not.
 */
#define is_ready(env,irn) \
  (nodeset_find((env)->ready_set, irn) != NULL)

/**
 * Check, if a node has already been schedules.
 * @param env The block schedule environment.
 * @param irn The node to check for.
 * @return 1 if the node was already scheduled, 0 if not.
 */
#define is_scheduled(env,irn) \
  (nodeset_find((env)->already_scheduled, irn) != NULL)

/**
 * Try, to make all users of a node ready.
 * In fact, a usage node can only be made ready, if all its operands
 * have already been scheduled yet. This is checked my make_ready().
 * @param env The block schedule environment.
 * @param irn The node, which usages (successors) are to be made ready.
 */
static INLINE void make_users_ready(block_sched_env_t *env, ir_node *irn)
{
        const ir_edge_t *edge;

        foreach_out_edge(irn, edge) {
                ir_node *user = edge->src;
                if(!is_Phi(user))
                        make_ready(env, user);
        }
}

/**
 * Compare to nodes using pointer equality.
 * @param p1 Node one.
 * @param p2 Node two.
 * @return 0 if they are identical.
 */
static int node_cmp_func(const void *p1, const void *p2)
{
    return p1 != p2;
}

/**
 * Append an instruction to a schedule.
 * @param env The block scheduling environment.
 * @param irn The node to add to the schedule.
 * @return    The given node.
 */
static ir_node *add_to_sched(block_sched_env_t *env, ir_node *irn)
{
    /* If the node consumes/produces data, it is appended to the schedule
     * list, otherwise, it is not put into the list */
    if(must_appear_in_schedule(env->selector, env->selector_block_env, irn)) {
        sched_info_t *info = get_irn_sched_info(irn);
        INIT_LIST_HEAD(&info->list);
        info->scheduled = 1;
        sched_add_before(env->block, irn);

        DBG((env->dbg, LEVEL_2, "\tadding %+F\n", irn));
    }

    /* Insert the node in the set of all already scheduled nodes. */
    nodeset_insert(env->already_scheduled, irn);

    /* Remove the node from the ready set */
    if(nodeset_find(env->ready_set, irn))
        nodeset_remove(env->ready_set, irn);

    return irn;
}

/**
 * Add the proj nodes of a tuple-mode irn to the schedule immediately
 * after the tuple-moded irn. By pinning the projs after the irn, no
 * other nodes can create a new lifetime between the tuple-moded irn and
 * one of its projs. This should render a realistic image of a
 * tuple-moded irn, which in fact models a node which defines multiple
 * values.
 *
 * @param irn The tuple-moded irn.
 * @param list The schedule list to append all the projs.
 * @param time The time step to which the irn and all its projs are
 * related to.
 * @param obst The obstack the scheduling data structures shall be
 * created upon.
 * @param ready_set The ready set of the list scheduler.
 * @param already_scheduled A set containing all nodes already
 * scheduled.
 */
static void add_tuple_projs(block_sched_env_t *env, ir_node *irn)
{
        const ir_edge_t *edge;

        assert(get_irn_mode(irn) == mode_T && "Mode of node must be tuple");

        foreach_out_edge(irn, edge) {
                ir_node *out = edge->src;

                assert(is_Proj(out) && "successor of a modeT node must be a proj");

                if(get_irn_mode(out) == mode_T)
                        add_tuple_projs(env, out);
                else {
                        add_to_sched(env, out);
                        make_users_ready(env, out);
                }
        }
}

static ir_node *select_node(block_sched_env_t *be)
{
        ir_node *irn;

        for (irn = nodeset_first(be->ready_set); irn; irn = nodeset_next(be->ready_set)) {
                if (be_is_Keep(irn)) {
                        nodeset_break(be->ready_set);
                        return irn;
                }
        }

        return be->selector->select(be->selector_block_env, be->ready_set);
}

/**
 * Perform list scheduling on a block.
 *
 * Note, that the caller must compute a linked list of nodes in the block
 * using the link field before calling this function.
 *
 * Also the outs must have been computed.
 *
 * @param block The block node.
 * @param env Scheduling environment.
 */
static void list_sched_block(ir_node *block, void *env_ptr)
{
        sched_env_t *env                      = env_ptr;
        const list_sched_selector_t *selector = env->selector;
        ir_node *start_node                   = get_irg_start(get_irn_irg(block));
        int phi_seen                          = 0;
        sched_info_t *info                    = get_irn_sched_info(block);

        block_sched_env_t be;
        const ir_edge_t *edge;
        ir_node *irn;
        int j, m;

        /* Initialize the block's list head that will hold the schedule. */
        INIT_LIST_HEAD(&info->list);

        /* Initialize the block scheduling environment */
        be.block             = block;
        be.curr_time         = 0;
        be.ready_set         = new_nodeset(get_irn_n_edges(block));
        be.already_scheduled = new_nodeset(get_irn_n_edges(block));
        be.selector          = selector;
        FIRM_DBG_REGISTER(be.dbg, "firm.be.sched");

        if(selector->init_block)
                be.selector_block_env = selector->init_block(env->selector_env, block);

        DBG((be.dbg, LEVEL_1, "scheduling %+F\n", block));

        /* Then one can add all nodes are ready to the set. */
        foreach_out_edge(block, edge) {
                ir_node *irn = get_edge_src_irn(edge);

                /* Skip the end node because of keepalive edges. */
                if(get_irn_opcode(irn) == iro_End)
                        continue;

                /* Phi functions are scheduled immediately, since they only transfer
                 * data flow from the predecessors to this block. */
                if(is_Phi(irn)) {
                        add_to_sched(&be, irn);
                        make_users_ready(&be, irn);
                        phi_seen = 1;
                }

                /* The start block will be scheduled as the first node */
                else if(irn == start_node) {
                        add_to_sched(&be, irn);
                        add_tuple_projs(&be, irn);
                }


                /* Other nodes must have all operands in other blocks to be made
                 * ready */
                else {
                        int ready = 1;

                        /* Check, if the operands of a node are not local to this block */
                        for(j = 0, m = get_irn_arity(irn); j < m; ++j) {
                                ir_node *operand = get_irn_n(irn, j);

                                if(get_nodes_block(operand) == block) {
                                        ready = 0;
                                        break;
                                }
                        }

                        /* Make the node ready, if all operands live in a foreign block */
                        if(ready) {
                                DBG((be.dbg, LEVEL_2, "\timmediately ready: %+F\n", irn));
                                make_ready(&be, irn);
                        }
                }
        }

        /* Increase the time, if some phi functions have been scheduled */
        be.curr_time += phi_seen;

        while (nodeset_count(be.ready_set) > 0) {
                /* collect statitics about amount of ready nodes */
                be_do_stat_sched_ready(block, be.ready_set);

                /* select a node to be scheduled and check if it was ready */
                irn = select_node(&be);

                DBG((be.dbg, LEVEL_3, "\tpicked node %+F\n", irn));

                /* Add the node to the schedule. */
                add_to_sched(&be, irn);

                if(get_irn_mode(irn) == mode_T)
                        add_tuple_projs(&be, irn);
                else
                        make_users_ready(&be, irn);

                /* Increase the time step. */
                be.curr_time += 1;

                /* remove the scheduled node from the ready list. */
                if (nodeset_find(be.ready_set, irn))
                        nodeset_remove(be.ready_set, irn);
        }

        if(selector->finish_block)
                selector->finish_block(be.selector_block_env);

        del_nodeset(be.ready_set);
        del_nodeset(be.already_scheduled);
}