4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
7 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
8 * examined. A qnode has a 'target color', nodes out of the opt unit and
9 * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges'
10 * A 'max indep set' is determined form these. We try to color this mis using a
11 * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges'
12 * and the qnode is reinserted in the queue. The first qnode colored without
13 * conflicts is the best one.
27 #include "becopyopt.h"
28 #include "becopystat.h"
31 #define DEBUG_LVL 0 //SET_LEVEL_1
32 static firm_dbg_module_t *dbg = NULL;
34 #define SLOTS_PINNED_GLOBAL 256
35 #define SLOTS_CONFLICTS 8
36 #define SLOTS_CHANGED_NODES 32
38 #define MIN(a,b) ((a<b)?(a):(b))
39 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
40 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
43 * Modeling additional conflicts between nodes. NOT live range interference
45 typedef struct _conflict_t {
46 const ir_node *n1, *n2;
50 * If an irn is changed, the changes first get stored in a node_stat_t,
51 * to allow undo of changes (=drop new data) in case of conflicts.
53 typedef struct _node_stat_t {
60 * Represents a node in the optimization queue.
62 typedef struct _qnode_t {
63 struct list_head queue; /**< chaining of unit_t->queue */
64 const unit_t *ou; /**< the opt unit this qnode belongs to */
65 int color; /**< target color */
66 set *conflicts; /**< contains conflict_t's. All internal conflicts */
67 int mis_costs; /**< costs of nodes/copies in the mis. */
68 int mis_size; /**< size of the array below */
69 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
70 set *changed_nodes; /**< contains node_stat_t's. */
73 pset *pinned_global; /**< optimized nodes should not be altered any more */
75 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
76 const conflict_t *xx = x;
77 const conflict_t *yy = y;
78 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
82 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
83 * The next maximum independent set build, will regard it.
85 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
87 DBG((dbg, LEVEL_4, "\t %n -- %n\n", n1, n2));
89 if ((int)n1 < (int)n2) {
96 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
100 * Checks if two nodes are in a conflict.
102 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
104 /* search for live range interference */
105 if (n1!=n2 && nodes_interfere(qn->ou->co->chordal_env, n1, n2))
107 /* search for recoloring conflicts */
108 if ((int)n1 < (int)n2) {
115 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
118 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
119 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
123 * Finds a node status entry of a node if existent. Otherwise return NULL
125 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
128 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
132 * Finds a node status entry of a node if existent. Otherwise it will return
133 * an initialized new entry for this node.
135 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
138 find.new_color = NO_COLOR;
139 find.pinned_local = 0;
140 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
144 * Returns the virtual color of a node if set before, else returns the real color.
146 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
147 node_stat_t *found = qnode_find_node(qn, irn);
149 return found->new_color;
151 return get_irn_col(qn->ou->co, irn);
155 * Sets the virtual color of a node.
157 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
158 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
159 found->new_color = color;
163 * Checks if a node is local pinned. A node is local pinned, iff it belongs
164 * to the same optimization unit and has been optimized before the current
167 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
168 node_stat_t *found = qnode_find_node(qn, irn);
170 return found->pinned_local;
176 * Local-pins a node, so optimizations of further nodes of the same opt unit
177 * can handle situations in which a color change would undo prior optimizations.
179 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
180 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
181 found->pinned_local = 1;
185 * Possible return values of qnode_color_irn()
187 #define CHANGE_SAVE NULL
188 #define CHANGE_IMPOSSIBLE (ir_node *)1
189 #define is_conflicting_node(n) (((int)n) > 1)
192 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
193 * other nodes are changed too, as required to preserve correctness. Function is
194 * aware of local and global pinning. Recursive.
195 * @param irn The node to set the color for
196 * @param col The color to set
197 * @param trigger The irn that caused the wish to change the color of the irn
198 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
199 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
200 * Else the first conflicting ir_node encountered is returned.
202 * ASSUMPTION: Assumes that a life range of a single value can't be split into
203 * several smaller intervals where other values can live in between.
204 * This should be true in SSA.
206 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
208 struct obstack confl_ob;
209 ir_node **confl, *cn;
211 const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env;
212 const arch_env_t *arch_env = chordal_env->arch_env;
213 const arch_register_class_t *cls = chordal_env->cls;
215 DBG((dbg, LEVEL_3, "\t %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
216 obstack_init(&confl_ob);
217 irn_col = qnode_get_new_color(qn, irn);
221 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
225 if (!arch_reg_is_allocatable(arch_env,
227 arch_pos_make_out(0),
228 arch_register_for_index(cls, col)))
231 /* get all nodes which would conflict with this change */
237 irn_bl = get_nodes_block(irn);
239 /* first check for a conflicting node which is 'living in' the irns block */
242 pset *live_ins = put_live_in(irn_bl, pset_new_ptr_default());
243 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
244 if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
245 && n != trigger && qnode_get_new_color(qn, n) == col
246 && nodes_interfere(chordal_env, irn, n)) {
248 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
249 obstack_ptr_grow(&confl_ob, n);
250 pset_break(live_ins);
256 /* setup the queue of blocks. */
258 obstack_ptr_grow(&q, irn_bl);
262 /* process the queue. The code below checks for every block dominated
263 * by the irns one, and in which the irn is live, if there are
264 * conflicting nodes */
266 ir_node *curr_bl, *sub_bl;
269 curr_bl = ((ir_node **)obstack_base(&q))[out++];
271 /* Add to the result all nodes in the block, which have
272 * the target color and interfere with the irn */
273 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
274 ir_node *n = get_irn_out(curr_bl, i);
275 if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
276 && n != trigger && qnode_get_new_color(qn, n) == col
277 && nodes_interfere(chordal_env, irn, n)) {
279 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
280 obstack_ptr_grow(&confl_ob, n);
284 /* If irn lives out check i-dominated blocks where the irn lives in */
286 if (is_live_out(curr_bl, irn)) {
287 dominates_for_each(curr_bl, sub_bl)
288 if (is_live_in(sub_bl, irn)) {
289 obstack_ptr_grow(&q, sub_bl);
294 obstack_free(&q, NULL);
295 obstack_ptr_grow(&confl_ob, NULL);
296 confl = (ir_node **) obstack_finish(&confl_ob);
299 /* process all nodes which would conflict with this change */
300 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
303 /* try to color the conflicting node cn with the color of the irn itself */
304 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
305 if (sub_res != CHANGE_SAVE) {
310 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
313 DBG((dbg, LEVEL_3, "\t %n save\n", irn));
314 obstack_free(&confl_ob, NULL);
315 qnode_set_new_color(qn, irn, col);
319 DBG((dbg, LEVEL_3, "\t %n impossible\n", irn));
320 obstack_free(&confl_ob, NULL);
321 return CHANGE_IMPOSSIBLE;
324 DBG((dbg, LEVEL_3, "\t %n conflicting\n", irn));
325 obstack_free(&confl_ob, NULL);
330 * Tries to set the colors for all members of this queue node;
331 * to the target color qn->color
332 * @returns 1 iff all members colors could be set
335 static int qnode_try_color(const qnode_t *qn) {
337 for (i=0; i<qn->mis_size; ++i) {
338 ir_node *test_node, *confl_node;
340 test_node = qn->mis[i];
341 DBG((dbg, LEVEL_3, "\t Testing %n\n", test_node));
342 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
344 if (confl_node == CHANGE_SAVE) {
345 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
346 qnode_pin_local(qn, test_node);
347 } else if (confl_node == CHANGE_IMPOSSIBLE) {
348 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
349 qnode_add_conflict(qn, test_node, test_node);
351 if (qnode_is_pinned_local(qn, confl_node)) {
352 /* changing test_node would change back a node of current ou */
353 if (confl_node == qn->ou->nodes[0]) {
354 /* Adding a conflict edge between testnode and conflnode
355 * would introduce a root -- arg interference.
356 * So remove the arg of the qn */
357 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
358 qnode_add_conflict(qn, test_node, test_node);
360 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
361 qnode_add_conflict(qn, confl_node, test_node);
364 if (pset_find_ptr(pinned_global, confl_node)) {
365 /* changing test_node would change back a node of a prior ou */
366 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
367 qnode_add_conflict(qn, test_node, test_node);
371 if (confl_node != CHANGE_SAVE)
378 * Determines a maximum weighted independent set with respect to
379 * the interference and conflict edges of all nodes in a qnode.
381 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
383 int max, next, pos, curr_weight, best_weight = 0;
384 bitset_t *best, *curr;
386 irns = alloca((ou->node_count-1) * sizeof(*irns));
387 best = bitset_alloca(ou->node_count-1);
388 curr = bitset_alloca(ou->node_count-1);
390 /* brute force the best set */
391 bitset_set_all(curr);
392 while ((max = bitset_popcnt(curr)) != 0) {
393 /* check if curr is a stable set */
394 int i, o, is_stable_set = 1;
398 bitset_foreach(curr, pos)
399 irns[i++] = ou->nodes[1+pos];
403 for(o=i; o<max; ++o) /* !!!!! difference to ou_max_ind_set_costs(): NOT o=i+1 */
404 if (qnode_are_conflicting(qn, irns[i], irns[o])) {
410 /* calc current weigth */
412 bitset_foreach(curr, pos)
413 curr_weight += ou->costs[1+pos];
416 if (curr_weight > best_weight) {
417 best_weight = curr_weight;
418 bitset_copy(best, curr);
425 /* transfer the best set into the qn */
426 qn->mis_size = bitset_popcnt(best);
427 qn->mis_costs = best_weight;
429 bitset_foreach(best, pos)
430 qn->mis[next++] = ou->nodes[1+pos];
434 * Creates a new qnode
436 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
437 qnode_t *qn = xmalloc(sizeof(*qn));
440 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
441 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
442 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
447 * Frees space used by a queue node
449 static INLINE void free_qnode(qnode_t *qn) {
450 del_set(qn->conflicts);
451 del_set(qn->changed_nodes);
457 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
458 * ordered by field 'size' (the size of the mis) in decreasing order.
460 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
461 struct list_head *lh;
463 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
464 /* root node is not in qnode */
469 qnode_max_ind_set(qn, ou);
470 /* do the insertion */
471 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
473 while (lh->next != &ou->queue) {
474 qnode_t *curr = list_entry_queue(lh->next);
475 if (curr->mis_costs <= qn->mis_costs)
479 list_add(&qn->queue, lh);
483 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
484 * costs of copy instructions placed during SSA-destruction and lowering.
485 * Works only for opt units with exactly 1 root node, which is the
486 * case for approximately 80% of all phi classes and 100% of register constrained
487 * nodes. (All other phi classes are reduced to this case.)
489 static void ou_optimize(unit_t *ou) {
492 bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs);
494 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
495 for (i=0; i<ou->node_count; ++i)
496 DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
499 INIT_LIST_HEAD(&ou->queue);
500 arch_get_allocatable_regs(ou->co->chordal_env->arch_env, ou->nodes[0], arch_pos_make_out(0), ou->co->chordal_env->cls, pos_regs);
501 bitset_foreach(pos_regs, i)
502 ou_insert_qnode(ou, new_qnode(ou, i));
505 while (!list_empty(&ou->queue)) {
506 /* get head of queue */
507 curr = list_entry_queue(ou->queue.next);
508 list_del(&curr->queue);
509 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
512 if (qnode_try_color(curr))
514 /* no success, so re-insert */
515 del_set(curr->changed_nodes);
516 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
517 ou_insert_qnode(ou, curr);
520 /* apply the best found qnode */
521 if (curr->mis_size >= 2) {
524 DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d/%d\n", curr->color, ou->complete_costs - curr->mis_costs, ou->complete_costs));
525 /* globally pin root and eventually others */
526 pset_insert_ptr(pinned_global, ou->nodes[0]);
527 for (i=1; i<ou->node_count; ++i) {
528 ir_node *irn = ou->nodes[i];
529 int nc = qnode_get_new_color(curr, irn);
530 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
531 pset_insert_ptr(pinned_global, irn);
534 /* set color of all changed nodes */
535 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
536 /* NO_COLOR is possible, if we had an undo */
537 if (ns->new_color != NO_COLOR) {
538 DBG((dbg, LEVEL_2, "\t color(%n) := %d\n", ns->irn, ns->new_color));
539 set_irn_col(ou->co, ns->irn, ns->new_color);
544 /* free best qnode (curr) and queue */
546 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
550 void co_heur_opt(copy_opt_t *co) {
552 dbg = firm_dbg_register("ir.be.copyoptheur");
553 firm_dbg_set_mask(dbg, DEBUG_LVL);
554 if (!strcmp(co->name, DEBUG_IRG))
555 firm_dbg_set_mask(dbg, DEBUG_LVL_HEUR);
557 firm_dbg_set_mask(dbg, DEBUG_LVL);
559 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
560 list_for_each_entry(unit_t, curr, &co->units, units)
561 if (curr->node_count > 1)
564 del_pset(pinned_global);