4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
20 #include "bechordal_t.h"
21 #include "becopyopt.h"
22 #include "becopystat.h"
24 #define DEBUG_LVL 0 //SET_LEVEL_1
25 static firm_dbg_module_t *dbg = NULL;
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)
29 #define MIN(a,b) ((a<b)?(a):(b))
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.
38 static int get_ifg_mis_size(unit_t *ou) {
39 int all_size, curr_size, i, o;
41 ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
43 /* all contains all nodes */
45 for (i=0; i<ou->node_count; ++i)
46 all[all_size++] = ou->nodes[i];
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;
53 /* stores the currently examined set */
54 curr = alloca(all_size*sizeof(*curr));
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]];
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]))
67 /* We had no conflict. This is the (one) max indep. set */
71 /* We had a conflict. Generate next set */
72 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
74 for (i=0; i<curr_size; ++i)
75 which[all_size-curr_size+i] = i;
82 } while (!(which[pos] = (which[pos]+1) % all_size));
84 for (i=pos+1; i<all_size; ++i)
85 which[i] = MIN(which[i-1]+1, all_size-1);
88 for (i=all_size-curr_size; i<all_size-1; ++i)
89 if (which[i]>=which[i+1]) {
96 assert(0 && "How did you get here?");
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.
106 static void co_append_unit(copy_opt_t *co, ir_node *root) {
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))
113 pset_insert_ptr(co->roots, root);
115 assert(is_curr_reg_class(root) && "node is in wrong register class!");
118 arity = get_irn_arity(root);
119 unit = xcalloc(1, sizeof(*unit));
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);
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.");
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;
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));
148 assert(0 && "This is not an optimizable node!");
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);
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);
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);
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);
172 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env) {
173 const char *s1, *s2, *s3;
177 dbg = firm_dbg_register("ir.be.copyopt");
178 firm_dbg_set_mask(dbg, DEBUG_LVL);
180 co = xcalloc(1, sizeof(*co));
181 co->chordal_env = chordal_env;
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);
192 firm_dbg_set_mask(dbg, DEBUG_LVL);
194 INIT_LIST_HEAD(&co->units);
195 co_collect_units(co);
199 void free_copy_opt(copy_opt_t *co) {
202 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
208 int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
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)))
218 int co_get_copy_count(const copy_opt_t *co) {
221 list_for_each_entry(unit_t, curr, &co->units, units) {
222 int root_col = get_irn_col(co, curr->nodes[0]);
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]));
234 int co_get_lower_bound(const copy_opt_t *co) {
237 list_for_each_entry(unit_t, curr, &co->units, units)
238 res += curr->interf + curr->node_count - curr->ifg_mis_size;
242 int co_get_interferer_count(const copy_opt_t *co) {
245 list_for_each_entry(unit_t, curr, &co->units, units)
251 * Needed for result checking
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);
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);
264 * This O(n^2) checker checks if
265 * two ifg-connected nodes have the same color
266 * register constraint are satisfied
268 void co_check_allocation(copy_opt_t *co) {
269 ir_node **nodes, *n1, *n2;
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);
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!");
287 obstack_free(&co->ob, NULL);
288 DBG((dbg, 2, "The checker seems to be happy!\n"));