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 static 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;
159 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, 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;
182 if (found->new_color == NO_COLOR)
183 found->new_color = get_irn_col(qn->ou->co, irn);
187 * Possible return values of qnode_color_irn()
189 #define CHANGE_SAVE NULL
190 #define CHANGE_IMPOSSIBLE (ir_node *)1
191 #define is_conflicting_node(n) (((int)n) > 1)
194 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
195 * other nodes are changed too, as required to preserve correctness. Function is
196 * aware of local and global pinning. Recursive.
197 * @param irn The node to set the color for
198 * @param col The color to set
199 * @param trigger The irn that caused the wish to change the color of the irn
200 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
201 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
202 * Else the first conflicting ir_node encountered is returned.
204 * ASSUMPTION: Assumes that a life range of a single value can't be split into
205 * several smaller intervals where other values can live in between.
206 * This should be true in SSA.
208 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
209 const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env;
210 const arch_register_class_t *cls = chordal_env->cls;
211 const arch_env_t *arch_env = chordal_env->main_env->arch_env;
212 int irn_col = qnode_get_new_color(qn, irn);
214 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
216 if (irn_col == col) {
217 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
220 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
221 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
224 if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
225 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
226 return CHANGE_IMPOSSIBLE;
230 * Process all nodes which would conflict with this change
233 be_ifg_t *ifg = chordal_env->ifg;
234 void *iter = be_ifg_neighbours_iter_alloca(ifg);
235 ir_node *sub_res, *curr;
238 * Try to color all conflicting nodes 'curr'
239 * with the color of the irn itself.
241 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
242 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
243 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
244 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
245 if (sub_res != CHANGE_SAVE) {
246 be_ifg_neighbours_break(ifg, iter);
254 * If we arrive here, all sub changes have been applied.
255 * So it's save to change this irn
257 qnode_set_new_color(qn, irn, col);
262 * Tries to set the colors for all members of this queue node;
263 * to the target color qn->color
264 * @returns 1 iff all members colors could be set
267 static int qnode_try_color(const qnode_t *qn) {
269 for (i=0; i<qn->mis_size; ++i) {
270 ir_node *test_node, *confl_node;
272 test_node = qn->mis[i];
273 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
274 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
276 if (confl_node == CHANGE_SAVE) {
277 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
278 qnode_pin_local(qn, test_node);
279 } else if (confl_node == CHANGE_IMPOSSIBLE) {
280 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
281 qnode_add_conflict(qn, test_node, test_node);
283 if (qnode_is_pinned_local(qn, confl_node)) {
284 /* changing test_node would change back a node of current ou */
285 if (confl_node == qn->ou->nodes[0]) {
286 /* Adding a conflict edge between testnode and conflnode
287 * would introduce a root -- arg interference.
288 * So remove the arg of the qn */
289 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
290 qnode_add_conflict(qn, test_node, test_node);
292 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
293 qnode_add_conflict(qn, confl_node, test_node);
296 if (pset_find_ptr(pinned_global, confl_node)) {
297 /* changing test_node would change back a node of a prior ou */
298 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
299 qnode_add_conflict(qn, test_node, test_node);
303 if (confl_node != CHANGE_SAVE)
310 * Determines a maximum weighted independent set with respect to
311 * the interference and conflict edges of all nodes in a qnode.
313 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
314 ir_node **safe, **unsafe;
315 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
316 bitset_t *curr, *best;
317 int max, next, pos, curr_weight, best_weight = 0;
319 /* assign the nodes into two groups.
320 * safe: node has no interference, hence it is in every max stable set.
321 * unsafe: node has an interference
323 safe = alloca((ou->node_count-1) * sizeof(*safe));
326 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
327 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
329 for(i=1; i<ou->node_count; ++i) {
331 for(o=1; o<ou->node_count; ++o) {
332 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
334 unsafe_costs[unsafe_count] = ou->costs[i];
335 unsafe[unsafe_count] = ou->nodes[i];
343 safe_costs += ou->costs[i];
344 safe[safe_count++] = ou->nodes[i];
350 /* now compute the best set out of the unsafe nodes*/
351 best = bitset_alloca(unsafe_count);
353 if (unsafe_count > MIS_HEUR_TRIGGER) {
354 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
355 for (i=0; i<unsafe_count; ++i) {
357 /* check if it is a stable set */
358 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
359 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
360 bitset_clear(best, i); /* clear the bit and try next one */
364 /* compute the weight */
365 bitset_foreach(best, pos)
366 best_weight += unsafe_costs[pos];
368 /* Exact Algorithm: Brute force */
369 curr = bitset_alloca(unsafe_count);
370 bitset_set_all(curr);
371 while ((max = bitset_popcnt(curr)) != 0) {
372 /* check if curr is a stable set */
373 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
374 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) */
375 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
378 /* if we arrive here, we have a stable set */
379 /* compute the weigth of the stable set*/
381 bitset_foreach(curr, pos)
382 curr_weight += unsafe_costs[pos];
385 if (curr_weight > best_weight) {
386 best_weight = curr_weight;
387 bitset_copy(best, curr);
395 /* transfer the best set into the qn */
396 qn->mis_size = 1+safe_count+bitset_popcnt(best);
397 qn->mis_costs = safe_costs+best_weight;
398 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
400 for (i=0; i<safe_count; ++i)
401 qn->mis[next++] = safe[i];
402 bitset_foreach(best, pos)
403 qn->mis[next++] = unsafe[pos];
407 * Creates a new qnode
409 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
410 qnode_t *qn = xmalloc(sizeof(*qn));
413 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
414 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
415 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
420 * Frees space used by a queue node
422 static INLINE void free_qnode(qnode_t *qn) {
423 del_set(qn->conflicts);
424 del_set(qn->changed_nodes);
430 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
431 * ordered by field 'size' (the size of the mis) in decreasing order.
433 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
434 struct list_head *lh;
436 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
437 /* root node is not in qnode */
442 qnode_max_ind_set(qn, ou);
443 /* do the insertion */
444 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
446 while (lh->next != &ou->queue) {
447 qnode_t *curr = list_entry_queue(lh->next);
448 if (curr->mis_costs <= qn->mis_costs)
452 list_add(&qn->queue, lh);
456 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
457 * costs of copy instructions placed during SSA-destruction and lowering.
458 * Works only for opt units with exactly 1 root node, which is the
459 * case for approximately 80% of all phi classes and 100% of register constrained
460 * nodes. (All other phi classes are reduced to this case.)
462 static void ou_optimize(unit_t *ou) {
464 qnode_t *curr = NULL, *tmp;
465 bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs);
467 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
468 for (i=0; i<ou->node_count; ++i)
469 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
472 INIT_LIST_HEAD(&ou->queue);
473 arch_get_allocatable_regs(get_arch_env(ou->co), ou->nodes[0], -1, ou->co->chordal_env->cls, pos_regs);
474 bitset_foreach(pos_regs, i)
475 ou_insert_qnode(ou, new_qnode(ou, i));
478 while (!list_empty(&ou->queue)) {
479 /* get head of queue */
480 curr = list_entry_queue(ou->queue.next);
481 list_del(&curr->queue);
482 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
485 if (qnode_try_color(curr))
488 /* no success, so re-insert */
489 del_set(curr->changed_nodes);
490 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
491 ou_insert_qnode(ou, curr);
494 /* apply the best found qnode */
495 if (curr->mis_size >= 2) {
497 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
498 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));
499 /* globally pin root and all args which have the same color */
500 pset_insert_ptr(pinned_global, ou->nodes[0]);
501 for (i=1; i<ou->node_count; ++i) {
502 ir_node *irn = ou->nodes[i];
503 int nc = qnode_get_new_color(curr, irn);
504 if (nc != NO_COLOR && nc == root_col)
505 pset_insert_ptr(pinned_global, irn);
508 /* set color of all changed nodes */
509 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
510 /* NO_COLOR is possible, if we had an undo */
511 if (ns->new_color != NO_COLOR) {
512 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
513 set_irn_col(ou->co, ns->irn, ns->new_color);
518 /* free best qnode (curr) and queue */
520 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
524 void co_heur_opt(copy_opt_t *co) {
526 dbg = firm_dbg_register("ir.be.copyoptheur");
528 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
529 list_for_each_entry(unit_t, curr, &co->units, units)
530 if (curr->node_count > 1)
533 del_pset(pinned_global);