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 static firm_dbg_module_t *dbg = NULL;
33 #define SLOTS_PINNED_GLOBAL 256
34 #define SLOTS_CONFLICTS 8
35 #define SLOTS_CHANGED_NODES 32
37 #define MIN(a,b) ((a<b)?(a):(b))
38 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
39 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
42 * Modeling additional conflicts between nodes. NOT live range interference
44 typedef struct _conflict_t {
45 const ir_node *n1, *n2;
49 * If an irn is changed, the changes first get stored in a node_stat_t,
50 * to allow undo of changes (=drop new data) in case of conflicts.
52 typedef struct _node_stat_t {
59 * Represents a node in the optimization queue.
61 typedef struct _qnode_t {
62 struct list_head queue; /**< chaining of unit_t->queue */
63 const unit_t *ou; /**< the opt unit this qnode belongs to */
64 int color; /**< target color */
65 set *conflicts; /**< contains conflict_t's. All internal conflicts */
66 int mis_costs; /**< costs of nodes/copies in the mis. */
67 int mis_size; /**< size of the array below */
68 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
69 set *changed_nodes; /**< contains node_stat_t's. */
72 pset *pinned_global; /**< optimized nodes should not be altered any more */
74 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
75 const conflict_t *xx = x;
76 const conflict_t *yy = y;
77 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
81 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
82 * The next maximum independent set build, will regard it.
84 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
86 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
88 if ((int)n1 < (int)n2) {
95 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
99 * Checks if two nodes are in a conflict.
101 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
103 /* search for live range interference */
104 if (n1!=n2 && nodes_interfere(qn->ou->co->chordal_env, n1, n2))
106 /* search for recoloring conflicts */
107 if ((int)n1 < (int)n2) {
114 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
117 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
118 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
122 * Finds a node status entry of a node if existent. Otherwise return NULL
124 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
127 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
131 * Finds a node status entry of a node if existent. Otherwise it will return
132 * an initialized new entry for this node.
134 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
137 find.new_color = NO_COLOR;
138 find.pinned_local = 0;
139 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
143 * Returns the virtual color of a node if set before, else returns the real color.
145 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
146 node_stat_t *found = qnode_find_node(qn, irn);
148 return found->new_color;
150 return get_irn_col(qn->ou->co, irn);
154 * Sets the virtual color of a node.
156 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
157 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
158 found->new_color = color;
162 * Checks if a node is local pinned. A node is local pinned, iff it belongs
163 * to the same optimization unit and has been optimized before the current
166 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
167 node_stat_t *found = qnode_find_node(qn, irn);
169 return found->pinned_local;
175 * Local-pins a node, so optimizations of further nodes of the same opt unit
176 * can handle situations in which a color change would undo prior optimizations.
178 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
179 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
180 found->pinned_local = 1;
184 * Possible return values of qnode_color_irn()
186 #define CHANGE_SAVE NULL
187 #define CHANGE_IMPOSSIBLE (ir_node *)1
188 #define is_conflicting_node(n) (((int)n) > 1)
191 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
192 * other nodes are changed too, as required to preserve correctness. Function is
193 * aware of local and global pinning. Recursive.
194 * @param irn The node to set the color for
195 * @param col The color to set
196 * @param trigger The irn that caused the wish to change the color of the irn
197 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
198 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
199 * Else the first conflicting ir_node encountered is returned.
201 * ASSUMPTION: Assumes that a life range of a single value can't be split into
202 * several smaller intervals where other values can live in between.
203 * This should be true in SSA.
205 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
207 struct obstack confl_ob;
208 ir_node **confl, *cn;
210 const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env;
211 const arch_env_t *arch_env = get_arch_env(qn->ou->co);
212 const arch_register_class_t *cls = chordal_env->cls;
214 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
215 obstack_init(&confl_ob);
216 irn_col = qnode_get_new_color(qn, irn);
218 if (irn_col == col) {
219 DBG((dbg, LEVEL_4, "\t Already same color.\n"));
222 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
226 if (!arch_reg_is_allocatable(arch_env,
228 arch_pos_make_out(0),
229 arch_register_for_index(cls, col)))
232 /* get all nodes which would conflict with this change */
238 irn_bl = get_nodes_block(irn);
240 /* first check for a conflicting node which is 'living in' the irns block */
243 pset *live_ins = put_live_in(irn_bl, pset_new_ptr_default());
244 for (n = pset_first(live_ins); n; n = pset_next(live_ins)) {
245 DBG((dbg, LEVEL_4, "Checking %+F which is live-in at the block\n", n));
246 if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
247 && n != trigger && qnode_get_new_color(qn, n) == col
248 && nodes_interfere(chordal_env, irn, n)) {
250 DBG((dbg, LEVEL_4, "\t %+F\ttroubles\n", n));
251 obstack_ptr_grow(&confl_ob, n);
252 pset_break(live_ins);
259 /* setup the queue of blocks. */
261 obstack_ptr_grow(&q, irn_bl);
265 /* process the queue. The code below checks for every block dominated
266 * by the irns one, and in which the irn is live, if there are
267 * conflicting nodes */
269 ir_node *curr_bl, *sub_bl;
272 curr_bl = ((ir_node **)obstack_base(&q))[out++];
274 /* Add to the result all nodes in the block, which have
275 * the target color and interfere with the irn */
276 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
277 ir_node *n = get_irn_out(curr_bl, i);
278 DBG((dbg, LEVEL_4, "Checking %+F defined in same block\n", n));
279 if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
280 && n != trigger && qnode_get_new_color(qn, n) == col
281 && nodes_interfere(chordal_env, irn, n)) {
282 DBG((dbg, LEVEL_4, "\t %+F\ttroubles\n", n));
283 obstack_ptr_grow(&confl_ob, n);
287 /* If irn lives out check i-dominated blocks where the irn lives in */
289 if (is_live_out(curr_bl, irn)) {
290 dominates_for_each(curr_bl, sub_bl)
291 if (is_live_in(sub_bl, irn)) {
292 obstack_ptr_grow(&q, sub_bl);
297 obstack_free(&q, NULL);
298 obstack_ptr_grow(&confl_ob, NULL);
299 confl = (ir_node **) obstack_finish(&confl_ob);
302 /* process all nodes which would conflict with this change */
303 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
306 /* try to color the conflicting node cn with the color of the irn itself */
307 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
308 if (sub_res != CHANGE_SAVE) {
313 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
316 DBG((dbg, LEVEL_3, "\t %+F save\n", irn));
317 obstack_free(&confl_ob, NULL);
318 qnode_set_new_color(qn, irn, col);
322 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
323 obstack_free(&confl_ob, NULL);
324 return CHANGE_IMPOSSIBLE;
327 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
328 obstack_free(&confl_ob, NULL);
333 * Tries to set the colors for all members of this queue node;
334 * to the target color qn->color
335 * @returns 1 iff all members colors could be set
338 static int qnode_try_color(const qnode_t *qn) {
340 for (i=0; i<qn->mis_size; ++i) {
341 ir_node *test_node, *confl_node;
343 test_node = qn->mis[i];
344 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
345 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
347 if (confl_node == CHANGE_SAVE) {
348 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
349 qnode_pin_local(qn, test_node);
350 } else if (confl_node == CHANGE_IMPOSSIBLE) {
351 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
352 qnode_add_conflict(qn, test_node, test_node);
354 if (qnode_is_pinned_local(qn, confl_node)) {
355 /* changing test_node would change back a node of current ou */
356 if (confl_node == qn->ou->nodes[0]) {
357 /* Adding a conflict edge between testnode and conflnode
358 * would introduce a root -- arg interference.
359 * So remove the arg of the qn */
360 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
361 qnode_add_conflict(qn, test_node, test_node);
363 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
364 qnode_add_conflict(qn, confl_node, test_node);
367 if (pset_find_ptr(pinned_global, confl_node)) {
368 /* changing test_node would change back a node of a prior ou */
369 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
370 qnode_add_conflict(qn, test_node, test_node);
374 if (confl_node != CHANGE_SAVE)
381 * Determines a maximum weighted independent set with respect to
382 * the interference and conflict edges of all nodes in a qnode.
384 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
385 ir_node **safe, **unsafe;
386 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
387 bitset_t *curr, *best;
388 int max, next, pos, curr_weight, best_weight = 0;
390 /* assign the nodes into two groups.
391 * safe: node has no interference, hence it is in every max stable set.
392 * unsafe: node has an interference
394 safe = alloca((ou->node_count-1) * sizeof(*safe));
397 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
398 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
400 for(i=1; i<ou->node_count; ++i) {
402 for(o=1; o<ou->node_count; ++o) {
403 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
405 unsafe_costs[unsafe_count] = ou->costs[i];
406 unsafe[unsafe_count] = ou->nodes[i];
414 safe_costs += ou->costs[i];
415 safe[safe_count++] = ou->nodes[i];
421 /* now brute force the best set out of the unsafe nodes*/
422 best = bitset_alloca(unsafe_count);
423 curr = bitset_alloca(unsafe_count);
425 bitset_set_all(curr);
426 while ((max = bitset_popcnt(curr)) != 0) {
427 /* check if curr is a stable set */
428 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
429 for (o=bitset_next_set(curr, i); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to ou_max_ind_set_costs(): NOT (curr, i+1) */
430 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
433 /* if we arrive here, we have a stable set */
434 /* compute the weigth of the stable set*/
436 bitset_foreach(curr, pos)
437 curr_weight += unsafe_costs[pos];
440 if (curr_weight > best_weight) {
441 best_weight = curr_weight;
442 bitset_copy(best, curr);
449 /* transfer the best set into the qn */
450 qn->mis_size = 1+safe_count+bitset_popcnt(best);
451 qn->mis_costs = safe_costs+best_weight;
452 qn->mis[0] = ou->nodes[0]; /* the root is alwazs in a max stable set */
454 for (i=0; i<safe_count; ++i)
455 qn->mis[next++] = safe[i];
456 bitset_foreach(best, pos)
457 qn->mis[next++] = unsafe[pos];
461 * Creates a new qnode
463 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
464 qnode_t *qn = xmalloc(sizeof(*qn));
467 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
468 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
469 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
474 * Frees space used by a queue node
476 static INLINE void free_qnode(qnode_t *qn) {
477 del_set(qn->conflicts);
478 del_set(qn->changed_nodes);
484 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
485 * ordered by field 'size' (the size of the mis) in decreasing order.
487 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
488 struct list_head *lh;
490 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
491 /* root node is not in qnode */
496 qnode_max_ind_set(qn, ou);
497 /* do the insertion */
498 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
500 while (lh->next != &ou->queue) {
501 qnode_t *curr = list_entry_queue(lh->next);
502 if (curr->mis_costs <= qn->mis_costs)
506 list_add(&qn->queue, lh);
510 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
511 * costs of copy instructions placed during SSA-destruction and lowering.
512 * Works only for opt units with exactly 1 root node, which is the
513 * case for approximately 80% of all phi classes and 100% of register constrained
514 * nodes. (All other phi classes are reduced to this case.)
516 static void ou_optimize(unit_t *ou) {
519 bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs);
521 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
522 for (i=0; i<ou->node_count; ++i)
523 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
526 INIT_LIST_HEAD(&ou->queue);
527 arch_get_allocatable_regs(get_arch_env(ou->co), ou->nodes[0], arch_pos_make_out(0), ou->co->chordal_env->cls, pos_regs);
528 bitset_foreach(pos_regs, i)
529 ou_insert_qnode(ou, new_qnode(ou, i));
532 while (!list_empty(&ou->queue)) {
533 /* get head of queue */
534 curr = list_entry_queue(ou->queue.next);
535 list_del(&curr->queue);
536 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
539 if (qnode_try_color(curr))
541 /* no success, so re-insert */
542 del_set(curr->changed_nodes);
543 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
544 ou_insert_qnode(ou, curr);
547 /* apply the best found qnode */
548 if (curr->mis_size >= 2) {
550 DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d << %d << %d\n", curr->color, ou->min_nodes_costs, ou->all_nodes_costs - curr->mis_costs, ou->all_nodes_costs));
551 /* globally pin root and all args which have the same color */
552 pset_insert_ptr(pinned_global, ou->nodes[0]);
553 for (i=1; i<ou->node_count; ++i) {
554 ir_node *irn = ou->nodes[i];
555 int nc = qnode_get_new_color(curr, irn);
556 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
557 pset_insert_ptr(pinned_global, irn);
560 /* set color of all changed nodes */
561 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
562 /* NO_COLOR is possible, if we had an undo */
563 if (ns->new_color != NO_COLOR) {
564 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
565 set_irn_col(ou->co, ns->irn, ns->new_color);
569 * Enable for checking register allocation after each ou
570 * be_ra_chordal_check(ou->co->chordal_env);
574 /* free best qnode (curr) and queue */
576 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
580 void co_heur_opt(copy_opt_t *co) {
582 dbg = firm_dbg_register("ir.be.copyoptheur");
583 if (!strcmp(co->name, DEBUG_IRG))
584 firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_HEUR);
586 firm_dbg_set_mask(dbg, DEBUG_LVL_HEUR);
588 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
589 list_for_each_entry(unit_t, curr, &co->units, units)
590 if (curr->node_count > 1)
593 del_pset(pinned_global);