make basic block scheduler more deterministic
[libfirm] / ir / be / beblocksched.c
index ecf3884..b9333e5 100644 (file)
@@ -22,7 +22,6 @@
  * @brief       Block-scheduling strategies.
  * @author      Matthias Braun, Christoph Mallon
  * @date        27.09.2006
- * @version     $Id$
  *
  * The goals of the greedy (and ILP) algorithm here works by assuming that
  * we want to change as many jumps to fallthroughs as possible (executed jumps
@@ -52,6 +51,7 @@
 #include "execfreq.h"
 #include "irdump_t.h"
 #include "irtools.h"
+#include "util.h"
 #include "debug.h"
 #include "beirgmod.h"
 #include "bemodule.h"
 #include "lc_opts.h"
 #include "lc_opts_enum.h"
 
-#ifdef WITH_ILP
-#include <lpp/lpp.h>
-#include <lpp/lpp_net.h>
-#endif /* WITH_ILP */
+#include "lpp.h"
+#include "lpp_net.h"
 
 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
-typedef enum _blocksched_algos_t {
+typedef enum blocksched_algos_t {
        BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
 } blocksched_algos_t;
 
 static int algo = BLOCKSCHED_GREEDY;
 
 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
-       { "naiv",       BLOCKSCHED_NAIV },
+       { "naiv",   BLOCKSCHED_NAIV },
        { "greedy", BLOCKSCHED_GREEDY },
-#ifdef WITH_ILP
        { "ilp",    BLOCKSCHED_ILP },
-#endif /* WITH_ILP */
        { NULL,     0 }
 };
 
@@ -88,7 +84,7 @@ static lc_opt_enum_int_var_t algo_var = {
 };
 
 static const lc_opt_table_entry_t be_blocksched_options[] = {
-       LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
+       LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var),
        LC_OPT_LAST
 };
 
@@ -137,7 +133,7 @@ struct blocksched_env_t {
  */
 static void collect_egde_frequency(ir_node *block, void *data)
 {
-       blocksched_env_t   *env = data;
+       blocksched_env_t   *env = (blocksched_env_t*)data;
        int                arity;
        edge_t             edge;
        blocksched_entry_t *entry;
@@ -204,20 +200,53 @@ static void collect_egde_frequency(ir_node *block, void *data)
        }
 }
 
-static int cmp_edges(const void *d1, const void *d2)
+static int cmp_edges_base(const edge_t *e1, const edge_t *e2)
 {
-       const edge_t *e1 = d1;
-       const edge_t *e2 = d2;
+       long nr1 = get_irn_node_nr(e1->block);
+       long nr2 = get_irn_node_nr(e2->block);
+       if (nr1 < nr2) {
+               return 1;
+       } else if (nr1 > nr2) {
+               return -1;
+       } else {
+               if (e1->pos < e2->pos) {
+                       return 1;
+               } else if (e1->pos > e2->pos) {
+                       return -1;
+               } else {
+                       return 0;
+               }
+       }
+}
 
-       return QSORT_CMP(e2->execfreq, e1->execfreq);
+static int cmp_edges(const void *d1, const void *d2)
+{
+       const edge_t *e1 = (const edge_t*)d1;
+       const edge_t *e2 = (const edge_t*)d2;
+       double        freq1 = e1->execfreq;
+       double        freq2 = e2->execfreq;
+       if (freq1 < freq2) {
+               return 1;
+       } else if (freq1 > freq2) {
+               return -1;
+       } else {
+               return cmp_edges_base(e1, e2);
+       }
 }
 
 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
 {
-       const edge_t *e1 = d1;
-       const edge_t *e2 = d2;
-       /* reverse sorting as penalties are negative */
-       return QSORT_CMP(e1->outedge_penalty_freq, e2->outedge_penalty_freq);
+       const edge_t *e1   = (const edge_t*)d1;
+       const edge_t *e2   = (const edge_t*)d2;
+       double        pen1 = e1->outedge_penalty_freq;
+       double        pen2 = e2->outedge_penalty_freq;
+       if (pen1 > pen2) {
+               return 1;
+       } else if (pen1 < pen2) {
+               return -1;
+       } else {
+               return cmp_edges_base(e1, e2);
+       }
 }
 
 static void clear_loop_links(ir_loop *loop)
@@ -260,8 +289,8 @@ static void coalesce_blocks(blocksched_env_t *env)
                        continue;
 
                pred_block = get_Block_cfgpred_block(block, pos);
-               entry      = get_irn_link(block);
-               pred_entry = get_irn_link(pred_block);
+               entry      = (blocksched_entry_t*)get_irn_link(block);
+               pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 
                if (pred_entry->next != NULL || entry->prev != NULL)
                        continue;
@@ -299,8 +328,8 @@ static void coalesce_blocks(blocksched_env_t *env)
                        continue;
 
                pred_block = get_Block_cfgpred_block(block, pos);
-               entry      = get_irn_link(block);
-               pred_entry = get_irn_link(pred_block);
+               entry      = (blocksched_entry_t*)get_irn_link(block);
+               pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 
                if (pred_entry->next != NULL || entry->prev != NULL)
                        continue;
@@ -341,8 +370,8 @@ static void coalesce_blocks(blocksched_env_t *env)
                        continue;
 
                pred_block = get_Block_cfgpred_block(block, pos);
-               entry      = get_irn_link(block);
-               pred_entry = get_irn_link(pred_block);
+               entry      = (blocksched_entry_t*)get_irn_link(block);
+               pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 
                /* is 1 of the blocks already attached to another block? */
                if (pred_entry->next != NULL || entry->prev != NULL)
@@ -380,7 +409,7 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en
 
                /* we only need to put the first of a series of already connected
                 * blocks into the worklist */
-               succ_entry = get_irn_link(succ_block);
+               succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
                while (succ_entry->prev != NULL) {
                        /* break cycles... */
                        if (succ_entry->prev->block == succ_block) {
@@ -389,7 +418,7 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en
                                break;
                        }
                        succ_entry = succ_entry->prev;
-               };
+               }
 
                if (irn_visited(succ_entry->block))
                        continue;
@@ -416,7 +445,7 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en
                if (irn_visited(succ_block))
                        continue;
 
-               succ_entry = get_irn_link(succ_block);
+               succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
                if (succ_entry->prev != NULL)
                        continue;
 
@@ -435,11 +464,11 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en
                                DB((dbg, LEVEL_1, "worklist empty\n"));
                                return;
                        }
-                       succ = pdeq_getl(env->worklist);
+                       succ = (ir_node*)pdeq_getl(env->worklist);
                } while (irn_visited(succ));
        }
 
-       succ_entry       = get_irn_link(succ);
+       succ_entry       = (blocksched_entry_t*)get_irn_link(succ);
        entry->next      = succ_entry;
        succ_entry->prev = entry;
 
@@ -450,7 +479,7 @@ static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 {
        ir_graph           *irg        = env->irg;
        ir_node            *startblock = get_irg_start_block(irg);
-       blocksched_entry_t *entry      = get_irn_link(startblock);
+       blocksched_entry_t *entry      = (blocksched_entry_t*)get_irn_link(startblock);
 
        ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
        inc_irg_visited(irg);
@@ -502,10 +531,7 @@ static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execf
        env.worklist   = NULL;
        env.blockcount = 0;
 
-       /* make sure loopinfo is up-to-date */
-       if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
-               construct_cf_backedges(irg);
-       }
+       assure_loopinfo(irg);
 
        // collect edge execution frequencies
        irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
@@ -517,7 +543,7 @@ static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execf
 
        start_entry = finish_block_schedule(&env);
        block_list  = create_blocksched_array(&env, start_entry, env.blockcount,
-                                             be_get_birg_obst(irg));
+                                             be_get_be_obst(irg));
 
        DEL_ARR_F(env.edges);
        obstack_free(&obst, NULL);
@@ -534,23 +560,22 @@ static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execf
  *
  */
 
-#ifdef WITH_ILP
-typedef struct _ilp_edge_t {
+typedef struct ilp_edge_t {
        ir_node *block;   /**< source block */
        int     pos;      /**< number of cfg predecessor (target) */
        int     ilpvar;
 } ilp_edge_t;
 
-typedef struct _blocksched_ilp_env_t {
+typedef struct blocksched_ilp_env_t {
        blocksched_env_t env;
        ilp_edge_t       *ilpedges;
        lpp_t            *lpp;
 } blocksched_ilp_env_t;
 
-typedef struct _blocksched_ilp_entry_t {
+typedef struct blocksched_ilp_entry_t {
        ir_node *block;
-       struct _blocksched_entry_t *next;
-       struct _blocksched_entry_t *prev;
+       struct blocksched_entry_t *next;
+       struct blocksched_entry_t *prev;
 
        int out_cst;
 } blocksched_ilp_entry_t;
@@ -589,7 +614,7 @@ static void collect_egde_frequency_ilp(ir_node *block, void *data)
        entry->block   = block;
        entry->next    = NULL;
        entry->prev    = NULL;
-       entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
+       entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
        set_irn_link(block, entry);
 
        if (block == startblock)
@@ -604,7 +629,7 @@ static void collect_egde_frequency_ilp(ir_node *block, void *data)
                int i;
 
                snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
-               cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
+               cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
 
                for (i = 0; i < arity; ++i) {
                        double     execfreq;
@@ -623,8 +648,8 @@ static void collect_egde_frequency_ilp(ir_node *block, void *data)
 
 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 {
-       int  i;
-       int  edge_count = ARR_LEN(env->ilpedges);
+       int           edge_count = ARR_LEN(env->ilpedges);
+       int           i;
 
        /* complete out constraints */
        for (i = 0; i < edge_count; ++i) {
@@ -645,20 +670,7 @@ static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
                lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
        }
 
-#if 0
-       {
-               FILE *f;
-               char fname[256];
-               lpp_dump(env->lpp, "lpp.out");
-               snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
-               f = fopen(fname, "w");
-               lpp_dump_plain(env->lpp, f);
-               fclose(f);
-       }
-#endif
-
-       //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
-       lpp_solve_net(env->lpp, "i44pc52", "cplex");
+       lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver);
        assert(lpp_is_sol_valid(env->lpp));
 
        /* Apply results to edges */
@@ -704,7 +716,7 @@ static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreq
        env.env.blockcount = 0;
        env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 
-       env.lpp = new_lpp("blockschedule", lpp_minimize);
+       env.lpp = lpp_new("blockschedule", lpp_minimize);
        lpp_set_time_limit(env.lpp, 20);
        lpp_set_log(env.lpp, stdout);
 
@@ -716,15 +728,14 @@ static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreq
        start_entry = finish_block_schedule(&env.env);
        block_list  = create_blocksched_array(&env.env, start_entry,
                                              env.env.blockcount,
-                                             be_get_birg_obst(irg));
+                                             be_get_be_obst(irg));
 
        DEL_ARR_F(env.ilpedges);
-       free_lpp(env.lpp);
+       lpp_free(env.lpp);
        obstack_free(&obst, NULL);
 
        return block_list;
 }
-#endif /* WITH_ILP */
 
 /*
  *  __  __       _
@@ -734,27 +745,26 @@ static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreq
  * |_|  |_|\__,_|_|_| |_|
  *
  */
-BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
 void be_init_blocksched(void)
 {
        lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
-       lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
 
-       lc_opt_add_table(blocksched_grp, be_blocksched_options);
+       lc_opt_add_table(be_grp, be_blocksched_options);
 
        FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 }
 
-ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
+ir_node **be_create_block_schedule(ir_graph *irg)
 {
+       ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
+
        switch (algo) {
        case BLOCKSCHED_GREEDY:
        case BLOCKSCHED_NAIV:
                return create_block_schedule_greedy(irg, execfreqs);
-#ifdef WITH_ILP
        case BLOCKSCHED_ILP:
                return create_block_schedule_ilp(irg, execfreqs);
-#endif /* WITH_ILP */
        }
 
        panic("unknown blocksched algo");