4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
21 #include "bechordal_t.h"
22 #include "becopyopt.h"
23 #include "becopystat.h"
25 static firm_dbg_module_t *dbg = NULL;
27 #define is_curr_reg_class(irn) \
28 (arch_get_irn_reg_class(get_arch_env(co), \
29 irn, arch_pos_make_out(0)) == co->chordal_env->cls)
31 #define MIN(a,b) ((a<b)?(a):(b))
32 #define MAX(a,b) ((a<b)?(b):(a))
36 * Determines a maximum weighted independent set with respect to
37 * the interference and conflict edges of all nodes in a qnode.
38 * TODO: This runs in 2^n in worst case. Use a heuristic iff n>???
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;
45 int max, pos, curr_weight, best_weight = 0;
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
51 safe = alloca((ou->node_count-1) * sizeof(*safe));
54 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
55 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
57 for(i=1; i<ou->node_count; ++i) {
59 for(o=1; o<ou->node_count; ++o) {
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];
71 safe_costs += ou->costs[i];
72 safe[safe_count++] = ou->nodes[i];
78 /* now brute force the best set out of the unsafe nodes*/
79 curr = bitset_alloca(unsafe_count);
82 while ((max = bitset_popcnt(curr)) != 0) {
83 /* check if curr is a stable set */
84 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
85 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) */
86 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
89 /* if we arrive here, we have a stable set */
90 /* compute the weigth of the stable set*/
92 bitset_foreach(curr, pos)
93 curr_weight += unsafe_costs[pos];
96 if (curr_weight > best_weight) {
97 best_weight = curr_weight;
104 return safe_costs+best_weight;
108 * Builds an optimization unit for a given optimizable irn (root).
109 * This opt-unit is inserted in the main structure co.
110 * If an arg of root itself is optimizable process this arg before with a
111 * recursive call. For handling this situation and loops co->root is used
112 * to remember all roots.
114 static void co_append_unit(copy_opt_t *co, ir_node *root) {
117 struct list_head *tmp;
119 DBG((dbg, LEVEL_1, "\t Root: %n %N\n", root, root));
120 /* check if we encountered this root earlier */
121 if (pset_find_ptr(co->roots, root))
123 pset_insert_ptr(co->roots, root);
125 assert(is_curr_reg_class(root) && "node is in wrong register class!");
128 arity = get_irn_arity(root);
129 unit = xcalloc(1, sizeof(*unit));
131 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
132 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
133 unit->node_count = 1;
134 unit->nodes[0] = root;
135 INIT_LIST_HEAD(&unit->queue);
138 if (is_Phi(root) && mode_is_datab(get_irn_mode(root))) {
139 for (i=0; i<arity; ++i) {
141 ir_node *arg = get_irn_n(root, i);
143 assert(is_curr_reg_class(arg) && "Argument not in same register class.");
146 if (nodes_interfere(co->chordal_env, root, arg)) {
147 unit->inevitable_costs += co->get_costs(root, arg, i);
151 /* Else insert the argument of the phi to the members of this ou */
152 DBG((dbg, LEVEL_1, "\t Member: %n %N\n", arg, arg));
154 /* Check if arg has occurred at a prior position in the arg/list */
155 for (o=0; o<unit->node_count; ++o)
156 if (unit->nodes[o] == arg) {
161 if (!arg_pos) { /* a new argument */
162 /* insert node, set costs */
163 unit->nodes[unit->node_count] = arg;
164 unit->costs[unit->node_count] = co->get_costs(root, arg, i);
166 } else { /* arg has occured before in same phi */
167 /* increase costs for existing arg */
168 unit->costs[arg_pos] += co->get_costs(root, arg, i);
171 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
172 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
173 } else if (is_Copy(get_arch_env(co), root)) {
174 assert(!nodes_interfere(co->chordal_env, root, get_Copy_src(root)));
175 unit->nodes[1] = get_Copy_src(root);
176 unit->costs[1] = co->get_costs(root, unit->nodes[1], -1);
177 unit->node_count = 2;
178 unit->nodes = xrealloc(unit->nodes, 2 * sizeof(*unit->nodes));
179 unit->costs = xrealloc(unit->costs, 2 * sizeof(*unit->costs));
181 assert(0 && "This is not an optimizable node!");
182 /* TODO add ou's for 2-addr-code instructions */
185 /* Determine the maximum costs this unit can cause: all_nodes_cost */
186 for(i=1; i<unit->node_count; ++i) {
187 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
188 unit->all_nodes_costs += unit->costs[i];
191 /* Determine the minimal costs this unit will cause: min_nodes_costs */
192 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
194 /* Insert the new ou according to its sort_key */
196 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
198 list_add(&unit->units, tmp);
201 static void co_collect_in_block(ir_node *block, void *env) {
202 copy_opt_t *co = env;
203 struct list_head *head = get_block_border_head(co->chordal_env, block);
206 list_for_each_entry_reverse(border_t, curr, head, list)
207 if (curr->is_def && curr->is_real && is_optimizable(get_arch_env(co), curr->irn))
208 co_append_unit(co, curr->irn);
211 static void co_collect_units(copy_opt_t *co) {
212 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
213 co->roots = pset_new_ptr(64);
214 dom_tree_walk_irg(get_irg(co), co_collect_in_block, NULL, co);
218 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)) {
219 const char *s1, *s2, *s3;
223 dbg = firm_dbg_register("ir.be.copyopt");
224 firm_dbg_set_mask(dbg, DEBUG_LVL_CO);
226 co = xcalloc(1, sizeof(*co));
227 co->chordal_env = chordal_env;
228 co->get_costs = get_costs;
230 s1 = get_irp_prog_name();
231 s2 = get_entity_name(get_irg_entity(get_irg(co)));
232 s3 = chordal_env->cls->name;
233 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
234 co->name = xmalloc(len);
235 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
236 if (!strcmp(co->name, DEBUG_IRG))
237 firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_CO);
239 firm_dbg_set_mask(dbg, DEBUG_LVL_CO);
241 INIT_LIST_HEAD(&co->units);
242 co_collect_units(co);
246 void free_copy_opt(copy_opt_t *co) {
249 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
255 int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
257 for(i=0, max=get_irn_n_outs(irn); i<max; ++i) {
258 ir_node *n = get_irn_out(irn, i);
259 if (((is_Phi(n) && mode_is_datab(get_irn_mode(n))) ||
260 is_Perm(get_arch_env(co), n)) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
266 int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos) {
269 ir_node *root_block = get_nodes_block(root);
271 assert(pos==-1 || is_Phi(root));
273 /* a perm places the copy in the same block as it resides */
274 loop = get_irn_loop(root_block);
276 /* for phis the copies are placed in the corresponding pred-block */
277 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
280 int d = get_loop_depth(loop);
286 int get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
290 int co_get_max_copy_costs(const copy_opt_t *co) {
294 list_for_each_entry(unit_t, curr, &co->units, units) {
295 res += curr->inevitable_costs;
296 for (i=1; i<curr->node_count; ++i)
297 res += curr->costs[i];
302 int co_get_inevit_copy_costs(const copy_opt_t *co) {
306 list_for_each_entry(unit_t, curr, &co->units, units)
307 res += curr->inevitable_costs;
311 int co_get_copy_costs(const copy_opt_t *co) {
315 list_for_each_entry(unit_t, curr, &co->units, units) {
316 int root_col = get_irn_col(co, curr->nodes[0]);
317 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
318 res += curr->inevitable_costs;
319 for (i=1; i<curr->node_count; ++i) {
320 int arg_col = get_irn_col(co, curr->nodes[i]);
321 if (root_col != arg_col) {
322 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
323 res += curr->costs[i];
330 int co_get_lower_bound(const copy_opt_t *co) {
333 list_for_each_entry(unit_t, curr, &co->units, units)
334 res += curr->inevitable_costs + curr->min_nodes_costs;