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
  19 #include "xmalloc.h"
  20 #include "bechordal_t.h"
  21 #include "becopyopt.h"
  22 #include "becopystat.h"
  23
  24 #define DEBUG_LVL 0 //SET_LEVEL_1
  25 static firm_dbg_module_t *dbg = NULL;
  26
  27 #define is_curr_reg_class(irn) (arch_get_irn_reg_class(co->chordal_env->arch_env, irn, arch_pos_make_out(0)) == co->chordal_env->cls)
  28
  29 #define MIN(a,b) ((a<b)?(a):(b))
  30
  31 /**
  32  * Computes a 'max independent set' wrt. ifg-edges only (no coloring conflicts, no register constraints)
  33  * @return The size of such a mis
  34  * NOTE: Code adapted from becopyheur
  35  * BETTER: Here we can be sure having a chordal graph to work on, so for 'larger' opt-units we
  36  *         could use a special algorithm.
  37  */
  38 static int get_ifg_mis_size(unit_t *ou) {
  39         int all_size, curr_size, i, o;
  40         int *which;
  41         ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
  42
  43         /* all contains all nodes */
  44         all_size = 0;
  45         for (i=0; i<ou->node_count; ++i)
  46                 all[all_size++] = ou->nodes[i];
  47
  48         /* which[i] says which element to take out of all[] and put into curr[i] */
  49         which = alloca(all_size*sizeof(*which));
  50         for (curr_size=0; curr_size<all_size; ++curr_size)
  51                 which[curr_size] = curr_size;
  52
  53         /* stores the currently examined set */
  54         curr = alloca(all_size*sizeof(*curr));
  55
  56         while (1) { /* this loop will terminate because at least a single node will be a max indep. set */
  57                 /* build current set */
  58                 for (i=0; i<curr_size; ++i)
  59                         curr[i] = all[which[all_size-curr_size+i]];
  60
  61                 /* check current set */
  62                 for (i=0; i<curr_size; ++i)
  63                         for (o=i+1; o<curr_size; ++o)
  64                                 if (nodes_interfere(ou->co->chordal_env, curr[i], curr[o]))
  65                                         goto conflict_found;
  66
  67                 /* We had no conflict. This is the (one) max indep. set */
  68                 return curr_size;
  69
  70 conflict_found:
  71                 /* We had a conflict. Generate next set */
  72                 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
  73                         curr_size--;
  74                         for (i=0; i<curr_size; ++i)
  75                                 which[all_size-curr_size+i] = i;
  76                 } else {
  77                         int redo = 1;
  78                         while (redo) {
  79                                 int pos = all_size;
  80                                 do {
  81                                         pos--;
  82                                 } while (!(which[pos] = (which[pos]+1) % all_size));
  83
  84                                 for (i=pos+1; i<all_size; ++i)
  85                                         which[i] = MIN(which[i-1]+1, all_size-1);
  86
  87                                 redo = 0;
  88                                 for (i=all_size-curr_size; i<all_size-1; ++i)
  89                                         if (which[i]>=which[i+1]) {
  90                                                 redo = 1;
  91                                                 break;
  92                                         }
  93                         }
  94                 }
  95         }
  96         assert(0 && "How did you get here?");
  97 }
  98
  99 /**
 100  * Builds an optimization unit for a given optimizable irn (root).
 101  * This opt-unit is inserted in the main structure co.
 102  * If an arg of root itself is optimizable process this arg before with a
 103  * recursive call. For handling this situation and loops co->root is used
 104  * to remember all roots.
 105  */
 106 static void co_append_unit(copy_opt_t *co, ir_node *root) {
 107         int i, arity;
 108         unit_t *unit;
 109         DBG((dbg, LEVEL_1, "\t  Root: %n %N\n", root, root));
 110         /* check if we encountered this root earlier */
 111         if (pset_find_ptr(co->roots, root))
 112                 return;
 113         pset_insert_ptr(co->roots, root);
 114
 115         assert(is_curr_reg_class(root) && "node is in wrong register class!");
 116
 117         /* init unit */
 118         arity = get_irn_arity(root);
 119         unit = xcalloc(1, sizeof(*unit));
 120         unit->co = co;
 121         unit->interf = 0;
 122         unit->node_count = 1;
 123         unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
 124         unit->nodes[0] = root;
 125         INIT_LIST_HEAD(&unit->queue);
 126
 127         /* check all args */
 128         if (is_Phi(root)) {
 129                 for (i=0; i<arity; ++i) {
 130                         ir_node *arg = get_irn_n(root, i);
 131                         assert(is_curr_reg_class(arg) && "Argument not in same register class.");
 132                         if (arg != root) {
 133                                 if (!nodes_interfere(co->chordal_env, root, arg)) {
 134                                         DBG((dbg, LEVEL_1, "\t   Member: %n %N\n", arg, arg));
 135                                         if (is_optimizable(co->chordal_env->arch_env, arg))
 136                                                 co_append_unit(co, arg);
 137                                         unit->nodes[unit->node_count++] = arg;
 138                                 } else
 139                                         unit->interf++;
 140                         }
 141                 }
 142                 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
 143         } else if (is_Copy(co->chordal_env->arch_env, root)) {
 144                 assert(!nodes_interfere(co->chordal_env, root, get_Copy_src(root)));
 145                 unit->nodes[unit->node_count++] = get_Copy_src(root);
 146                 unit->nodes = xrealloc(unit->nodes, 2 * sizeof(*unit->nodes));
 147         } else
 148                 assert(0 && "This is not an optimizable node!");
 149
 150         list_add_tail(&unit->units, &co->units);
 151         /* Init ifg_mis_size to node_count. So get_lower_bound returns correct results. */
 152         unit->ifg_mis_size = get_ifg_mis_size(unit);
 153 }
 154
 155 static void co_collect_in_block(ir_node *block, void *env) {
 156         copy_opt_t *co = env;
 157         struct list_head *head = get_block_border_head(co->chordal_env, block);
 158         border_t *curr;
 159
 160         list_for_each_entry_reverse(border_t, curr, head, list)
 161                 if (curr->is_def && curr->is_real && is_optimizable(co->chordal_env->arch_env, curr->irn))
 162                         co_append_unit(co, curr->irn);
 163 }
 164
 165 static void co_collect_units(copy_opt_t *co) {
 166         DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
 167         co->roots = pset_new_ptr(64);
 168         dom_tree_walk_irg(co->chordal_env->irg, co_collect_in_block, NULL, co);
 169         del_pset(co->roots);
 170 }
 171
 172 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env) {
 173         const char *s1, *s2, *s3;
 174         int len;
 175         copy_opt_t *co;
 176
 177         dbg = firm_dbg_register("ir.be.copyopt");
 178         firm_dbg_set_mask(dbg, DEBUG_LVL);
 179
 180         co = xcalloc(1, sizeof(*co));
 181         co->chordal_env = chordal_env;
 182
 183         s1 = get_irp_prog_name();
 184         s2 = get_entity_name(get_irg_entity(co->chordal_env->irg));
 185         s3 = chordal_env->cls->name;
 186         len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
 187         co->name = xmalloc(len);
 188         snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
 189         if (!strcmp(co->name, DEBUG_IRG))
 190                 firm_dbg_set_mask(dbg, DEBUG_LVL_CO);
 191         else
 192                 firm_dbg_set_mask(dbg, DEBUG_LVL);
 193
 194         INIT_LIST_HEAD(&co->units);
 195         co_collect_units(co);
 196         return co;
 197 }
 198
 199 void free_copy_opt(copy_opt_t *co) {
 200         unit_t *curr, *tmp;
 201         xfree(co->name);
 202         list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
 203                 xfree(curr->nodes);
 204                 xfree(curr);
 205         }
 206 }
 207
 208 int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
 209         int i, max;
 210         for(i=0, max=get_irn_n_outs(irn); i<max; ++i) {
 211                 ir_node *n = get_irn_out(irn, i);
 212                 if ((is_Phi(n) || is_Perm(co->chordal_env->arch_env, n)) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
 213                         return 1;
 214         }
 215         return 0;
 216 }
 217
 218 int co_get_copy_count(const copy_opt_t *co) {
 219         int i, res = 0;
 220         unit_t *curr;
 221         list_for_each_entry(unit_t, curr, &co->units, units) {
 222                 int root_col = get_irn_col(co, curr->nodes[0]);
 223                 res += curr->interf;
 224                 DBG((dbg, LEVEL_1, "%n %N has %d intf\n", curr->nodes[0], curr->nodes[0], curr->interf));
 225                 for (i=1; i<curr->node_count; ++i)
 226                         if (root_col != get_irn_col(co, curr->nodes[i])) {
 227                                 DBG((dbg, LEVEL_1, "  %n %N\n", curr->nodes[i], curr->nodes[i]));
 228                                 res++;
 229                         }
 230         }
 231         return res;
 232 }
 233
 234 int co_get_lower_bound(const copy_opt_t *co) {
 235         int res = 0;
 236         unit_t *curr;
 237         list_for_each_entry(unit_t, curr, &co->units, units)
 238                 res += curr->interf + curr->node_count - curr->ifg_mis_size;
 239         return res;
 240 }
 241
 242 int co_get_interferer_count(const copy_opt_t *co) {
 243         int res = 0;
 244         unit_t *curr;
 245         list_for_each_entry(unit_t, curr, &co->units, units)
 246                 res += curr->interf;
 247         return res;
 248 }
 249
 250 /**
 251  * Needed for result checking
 252  */
 253 static void co_collect_for_checker(ir_node *block, void *env) {
 254         copy_opt_t *co = env;
 255         struct list_head *head = get_block_border_head(co->chordal_env, block);
 256         border_t *curr;
 257
 258         list_for_each_entry_reverse(border_t, curr, head, list)
 259                 if (curr->is_def && curr->is_real && is_curr_reg_class(curr->irn))
 260                         obstack_ptr_grow(&co->ob, curr->irn);
 261 }
 262
 263 /**
 264  * This O(n^2) checker checks if
 265  *      two ifg-connected nodes have the same color
 266  *      register constraint are satisfied
 267  */
 268 void co_check_allocation(copy_opt_t *co) {
 269         ir_node **nodes, *n1, *n2;
 270         int i, o;
 271
 272         obstack_init(&co->ob);
 273         dom_tree_walk_irg(co->chordal_env->irg, co_collect_for_checker, NULL, co);
 274         obstack_ptr_grow(&co->ob, NULL);
 275
 276         nodes = (ir_node **) obstack_finish(&co->ob);
 277         for (i = 0, n1 = nodes[i]; n1; n1 = nodes[++i]) {
 278                 assert(arch_reg_is_allocatable(co->chordal_env->arch_env, n1, arch_pos_make_out(0),
 279                         arch_get_irn_register(co->chordal_env->arch_env, n1, 0)) && "Constraint does not hold");
 280                 for (o = i+1, n2 = nodes[o]; n2; n2 = nodes[++o])
 281                         if (nodes_interfere(co->chordal_env, n1, n2)
 282           && get_irn_col(co, n1) == get_irn_col(co, n2)) {
 283                                 DBG((dbg, 0, "Error in graph %s: %n %d  and  %n %d have the same color %d.\n", co->name, n1, get_irn_graph_nr(n1), n2, get_irn_graph_nr(n2), get_irn_col(co, n1)));
 284                                 assert(0 && "Interfering values have the same color!");
 285                         }
 286         }
 287         obstack_free(&co->ob, NULL);
 288         DBG((dbg, 2, "The checker seems to be happy!\n"));
 289 }