nsz Git - libfirm/blob - ir/be/becopyopt.c

   1 /**
   2  * Author:      Daniel Grund
   3  * Date:                12.04.2005
   4  * Copyright:   (c) Universitaet Karlsruhe
   5  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
   6  */
   7 #ifdef HAVE_CONFIG_H
   8 #include "config.h"
   9 #endif
  10 #ifdef HAVE_ALLOCA_H
  11 #include <alloca.h>
  12 #endif
  13 #ifdef HAVE_MALLOC_H
  14 #include <malloc.h>
  15 #endif
  16
  17 #include "irprog.h"
  18 #include "irloop_t.h"
  19
  20 #include "xmalloc.h"
  21 #include "beutil.h"
  22 #include "bechordal_t.h"
  23 #include "becopyopt.h"
  24 #include "becopystat.h"
  25
  26 static firm_dbg_module_t *dbg = NULL;
  27
  28 #define is_curr_reg_class(irn) \
  29   (arch_get_irn_reg_class(get_arch_env(co), \
  30                           irn, -1) == co->chordal_env->cls)
  31
  32 #define MIN(a,b) ((a<b)?(a):(b))
  33 #define MAX(a,b) ((a<b)?(b):(a))
  34
  35
  36 /**
  37  * Determines a maximum weighted independent set with respect to
  38  * the interference and conflict edges of all nodes in a qnode.
  39  */
  40 static int ou_max_ind_set_costs(unit_t *ou) {
  41         be_chordal_env_t *chordal_env = ou->co->chordal_env;
  42         ir_node **safe, **unsafe;
  43         int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
  44         bitset_t *curr;
  45         int max, pos, curr_weight, best_weight = 0;
  46
  47         /* assign the nodes into two groups.
  48          * safe: node has no interference, hence it is in every max stable set.
  49          * unsafe: node has an interference
  50          */
  51         safe = alloca((ou->node_count-1) * sizeof(*safe));
  52         safe_costs = 0;
  53         safe_count = 0;
  54         unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
  55         unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
  56         unsafe_count = 0;
  57         for(i=1; i<ou->node_count; ++i) {
  58                 int is_safe = 1;
  59                 for(o=1; o<ou->node_count; ++o) {
  60                         if (i==o)
  61                                 continue;
  62                         if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
  63                                 unsafe_costs[unsafe_count] = ou->costs[i];
  64                                 unsafe[unsafe_count] = ou->nodes[i];
  65                                 ++unsafe_count;
  66                                 is_safe = 0;
  67                                 break;
  68                         }
  69                 }
  70                 if (is_safe) {
  71                         safe_costs += ou->costs[i];
  72                         safe[safe_count++] = ou->nodes[i];
  73                 }
  74         }
  75
  76
  77         /* now compute the best set out of the unsafe nodes*/
  78         if (unsafe_count > MIS_HEUR_TRIGGER) {
  79                 bitset_t *best = bitset_alloca(unsafe_count);
  80                 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
  81                 for (i=0; i<unsafe_count; ++i) {
  82                         bitset_set(best, i);
  83                         /* check if it is a stable set */
  84                         for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
  85                                 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
  86                                         bitset_clear(best, i); /* clear the bit and try next one */
  87                                         break;
  88                                 }
  89                 }
  90                 /* compute the weight */
  91                 bitset_foreach(best, pos)
  92                         best_weight += unsafe_costs[pos];
  93         } else {
  94                 /* Exact Algorithm: Brute force */
  95                 curr = bitset_alloca(unsafe_count);
  96                 bitset_set_all(curr);
  97                 while ((max = bitset_popcnt(curr)) != 0) {
  98                         /* check if curr is a stable set */
  99                         for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
 100                                 for (o=bitset_next_set(curr, i+1); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to qnode_max_ind_set(): NOT (curr, i) */
 101                                                 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
 102                                                         goto no_stable_set;
 103
 104                         /* if we arrive here, we have a stable set */
 105                         /* compute the weigth of the stable set*/
 106                         curr_weight = 0;
 107                         bitset_foreach(curr, pos)
 108                                 curr_weight += unsafe_costs[pos];
 109
 110                         /* any better ? */
 111                         if (curr_weight > best_weight) {
 112                                 best_weight = curr_weight;
 113                         }
 114
 115         no_stable_set:
 116                         bitset_minus1(curr);
 117                 }
 118         }
 119
 120         return safe_costs+best_weight;
 121 }
 122
 123 static void co_collect_units(ir_node *irn, void *env) {
 124         copy_opt_t *co = env;
 125         unit_t *unit;
 126         arch_register_req_t req;
 127         arch_env_t *aenv = co->chordal_env->main_env->arch_env;
 128
 129         if (!is_curr_reg_class(irn))
 130                 return;
 131         if (!is_optimizable(get_arch_env(co), irn, &req))
 132                 return;
 133
 134         /* Init a new unit */
 135         unit = xcalloc(1, sizeof(*unit));
 136         unit->co = co;
 137         unit->node_count = 1;
 138         INIT_LIST_HEAD(&unit->queue);
 139
 140         /* Phi with some/all of its arguments */
 141         if (is_Reg_Phi(irn)) {
 142                 int i, arity;
 143
 144                 /* init */
 145                 arity = get_irn_arity(irn);
 146                 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
 147                 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
 148                 unit->nodes[0] = irn;
 149
 150                 /* fill */
 151                 for (i=0; i<arity; ++i) {
 152                         int o, arg_pos;
 153                         ir_node *arg = get_irn_n(irn, i);
 154
 155                         assert(is_curr_reg_class(arg) && "Argument not in same register class.");
 156                         if (arg == irn)
 157                                 continue;
 158                         if (nodes_interfere(co->chordal_env, irn, arg)) {
 159                                 unit->inevitable_costs += co->get_costs(irn, arg, i);
 160                                 continue;
 161                         }
 162
 163                         /* Else insert the argument of the phi to the members of this ou */
 164                         DBG((dbg, LEVEL_1, "\t   Member: %+F\n", arg));
 165
 166                         /* Check if arg has occurred at a prior position in the arg/list */
 167                         arg_pos = 0;
 168                         for (o=0; o<unit->node_count; ++o)
 169                                 if (unit->nodes[o] == arg) {
 170                                         arg_pos = o;
 171                                         break;
 172                                 }
 173
 174                         if (!arg_pos) { /* a new argument */
 175                                 /* insert node, set costs */
 176                                 unit->nodes[unit->node_count] = arg;
 177                                 unit->costs[unit->node_count] = co->get_costs(irn, arg, i);
 178                                 unit->node_count++;
 179                         } else { /* arg has occured before in same phi */
 180                                 /* increase costs for existing arg */
 181                                 unit->costs[arg_pos] += co->get_costs(irn, arg, i);
 182                         }
 183                 }
 184                 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
 185                 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
 186         } else
 187
 188         /* Proj of a perm with corresponding arg */
 189         if (is_Perm_Proj(get_arch_env(co), irn)) {
 190                 assert(!nodes_interfere(co->chordal_env, irn, get_Copy_src(irn)));
 191                 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
 192                 unit->costs = xmalloc(2 * sizeof(*unit->costs));
 193                 unit->node_count = 2;
 194                 unit->nodes[0] = irn;
 195                 unit->nodes[1] = get_Copy_src(irn);
 196                 unit->costs[1] = co->get_costs(irn, unit->nodes[1], -1);
 197         } else
 198
 199         /* Src == Tgt of a 2-addr-code instruction */
 200         if (is_2addr_code(get_arch_env(co), irn, &req)) {
 201                 ir_node *other = req.other;
 202                 if (!nodes_interfere(co->chordal_env, irn, other)) {
 203                         unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
 204                         unit->costs = xmalloc(2 * sizeof(*unit->costs));
 205                         unit->node_count = 2;
 206                         unit->nodes[0] = irn;
 207                         unit->nodes[1] = other;
 208                         unit->costs[1] = co->get_costs(irn, other, -120480);
 209                 }
 210         } else
 211                 assert(0 && "This is not an optimizable node!");
 212
 213         /* Insert the new unit at a position according to its costs */
 214         if (unit->node_count > 1) {
 215                 int i;
 216                 struct list_head *tmp;
 217
 218                 /* Determine the maximum costs this unit can cause: all_nodes_cost */
 219                 for(i=1; i<unit->node_count; ++i) {
 220                         unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
 221                         unit->all_nodes_costs += unit->costs[i];
 222                 }
 223
 224                 /* Determine the minimal costs this unit will cause: min_nodes_costs */
 225                 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
 226
 227                 /* Insert the new ou according to its sort_key */
 228                 tmp = &co->units;
 229                 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
 230                         tmp = tmp->next;
 231                 list_add(&unit->units, tmp);
 232         } else {
 233                 free(unit);
 234         }
 235 }
 236
 237 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)) {
 238         const char *s1, *s2, *s3;
 239         int len;
 240         copy_opt_t *co;
 241
 242         dbg = firm_dbg_register("ir.be.copyopt");
 243
 244         co = xcalloc(1, sizeof(*co));
 245         co->chordal_env = chordal_env;
 246         co->get_costs = get_costs;
 247
 248         s1 = get_irp_prog_name();
 249         s2 = get_entity_name(get_irg_entity(get_irg(co)));
 250         s3 = chordal_env->cls->name;
 251         len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
 252         co->name = xmalloc(len);
 253         snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
 254
 255         DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
 256         INIT_LIST_HEAD(&co->units);
 257         irg_walk_graph(get_irg(co), co_collect_units, NULL, co);
 258         return co;
 259 }
 260
 261 void free_copy_opt(copy_opt_t *co) {
 262         unit_t *curr, *tmp;
 263         xfree(co->name);
 264         list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
 265                 xfree(curr->nodes);
 266                 xfree(curr->costs);
 267                 xfree(curr);
 268         }
 269 }
 270
 271 int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
 272         arch_env_t *aenv = get_arch_env(co);
 273         const ir_edge_t *edge;
 274
 275         if (arch_irn_is_ignore(aenv, irn))
 276                 return 0;
 277
 278         foreach_out_edge(irn, edge) {
 279                 ir_node *n = edge->src;
 280
 281                 if (!nodes_interfere(co->chordal_env, irn, n) || irn == n) {
 282                         arch_register_req_t req;
 283                         arch_get_register_req(aenv, &req, n, -1);
 284
 285                         if(is_Reg_Phi(n) ||
 286                            is_Perm(aenv, n) ||
 287                            (arch_register_req_is(&req, should_be_same) && req.other == irn)
 288                           )
 289                                 return 1;
 290                 }
 291         }
 292
 293         return 0;
 294 }
 295
 296 int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos) {
 297         int cost = 0;
 298         ir_loop *loop;
 299         ir_node *root_block = get_nodes_block(root);
 300
 301         if (is_Phi(root)) {
 302                 /* for phis the copies are placed in the corresponding pred-block */
 303                 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
 304         } else {
 305                 /* a perm places the copy in the same block as it resides */
 306                 loop = get_irn_loop(root_block);
 307         }
 308         if (loop) {
 309                 int d = get_loop_depth(loop);
 310                 cost = d*d;
 311         }
 312         return cost+1;
 313 }
 314
 315 int get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
 316         return 1;
 317 }
 318
 319 int co_get_max_copy_costs(const copy_opt_t *co) {
 320         int i, res = 0;
 321         unit_t *curr;
 322
 323         list_for_each_entry(unit_t, curr, &co->units, units) {
 324                 res += curr->inevitable_costs;
 325                 for (i=1; i<curr->node_count; ++i)
 326                         res += curr->costs[i];
 327         }
 328         return res;
 329 }
 330
 331 int co_get_inevit_copy_costs(const copy_opt_t *co) {
 332         int res = 0;
 333         unit_t *curr;
 334
 335         list_for_each_entry(unit_t, curr, &co->units, units)
 336                 res += curr->inevitable_costs;
 337         return res;
 338 }
 339
 340 int co_get_copy_costs(const copy_opt_t *co) {
 341         int i, res = 0;
 342         unit_t *curr;
 343
 344         list_for_each_entry(unit_t, curr, &co->units, units) {
 345                 int root_col = get_irn_col(co, curr->nodes[0]);
 346                 DBG((dbg, LEVEL_1, "  %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
 347                 res += curr->inevitable_costs;
 348                 for (i=1; i<curr->node_count; ++i) {
 349                         int arg_col = get_irn_col(co, curr->nodes[i]);
 350                         if (root_col != arg_col) {
 351                                 DBG((dbg, LEVEL_1, "  %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
 352                                 res += curr->costs[i];
 353                         }
 354                 }
 355         }
 356         return res;
 357 }
 358
 359 int co_get_lower_bound(const copy_opt_t *co) {
 360         int res = 0;
 361         unit_t *curr;
 362         list_for_each_entry(unit_t, curr, &co->units, units)
 363                 res += curr->inevitable_costs + curr->min_nodes_costs;
 364         return res;
 365 }