2 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
3 * examined. A qnode has a 'target color', nodes out of the opt unit and
4 * a 'conflict graph'. A 'max indep set' is determined form these. We try to
5 * color this mis using a color-exchanging mechanism. Occuring conflicts are
6 * modeled with 'conflict edges' and the qnode is reinserted in the queue. The
7 * first qnode colored without conflicts is the best one.
12 #include "becopyopt.h"
13 #include "becopystat.h"
15 #define DEBUG_LVL SET_LEVEL_1
16 static firm_dbg_module_t *dbg = NULL;
18 #define SLOTS_PINNED_GLOBAL 256
19 #define SLOTS_CONFLICTS 8
20 #define SLOTS_CHANGED_NODES 32
22 #define MIN(a,b) ((a<b)?a:b)
23 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
24 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
27 * Modeling additional conflicts between nodes. NOT live range interference
29 typedef struct _conflict_t {
30 const ir_node *n1, *n2;
34 * If an irn is changed, the changes first get stored in a node_stat_t,
35 * to allow undo of changes (=drop new data) in case of conflicts.
37 typedef struct _node_stat_t {
44 * Represents a node in the optimization queue.
46 typedef struct _qnode_t {
47 struct list_head queue; /**< chaining of unit_t->queue */
48 int color; /**< target color */
49 set *conflicts; /**< contains conflict_t's. All internal conflicts */
50 int mis_size; /**< number of nodes in the mis. */
51 const ir_node **mis; /**< the nodes of unit_t->nodes[] beeing part of the max idependent set */
52 set *changed_nodes; /**< contains node_stat_t's. */
55 pset *pinned_global; /**< optimized nodes should not be altered any more */
57 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
58 const conflict_t *xx = x;
59 const conflict_t *yy = y;
60 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
64 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
65 * The next maximum independent set build, will regard it.
67 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
69 DBG((dbg, LEVEL_4, "\t %n -- %n\n", n1, n2));
71 if ((int)n1 < (int)n2) {
78 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
82 * Checks if two nodes are in a conflict.
84 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
86 /* search for live range interference */
87 if (n1!=n2 && values_interfere(n1, n2))
89 /* search for recoloring conflicts */
90 if ((int)n1 < (int)n2) {
97 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
100 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
101 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
105 * Finds a node status entry of a node if existent. Otherwise return NULL
107 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, const ir_node *irn) {
110 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
114 * Finds a node status entry of a node if existent. Otherwise it will return
115 * an initialized new entry for this node.
117 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, const ir_node *irn) {
120 find.new_color = NO_COLOR;
121 find.pinned_local = 0;
122 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
126 * Returns the virtual color of a node if set before, else returns the real color.
128 static INLINE int qnode_get_new_color(const qnode_t *qn, const ir_node *irn) {
129 node_stat_t *found = qnode_find_node(qn, irn);
131 return found->new_color;
133 return get_irn_color(irn);
137 * Sets the virtual color of a node.
139 static INLINE void qnode_set_new_color(const qnode_t *qn, const ir_node *irn, int color) {
140 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
141 found->new_color = color;
145 * Checks if a node is local pinned. A node is local pinned, iff it belongs
146 * to the same optimization unit and has been optimized before the current
149 static INLINE int qnode_is_pinned_local(const qnode_t *qn, const ir_node *irn) {
150 node_stat_t *found = qnode_find_node(qn, irn);
152 return found->pinned_local;
158 * Local-pins a node, so optimizations of further nodes of the same opt unit
159 * can handle situations in which a color change would undo prior optimizations.
161 static INLINE void qnode_pin_local(const qnode_t *qn, const ir_node *irn) {
162 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
163 found->pinned_local = 1;
167 * Possible return values of qnode_color_irn()
169 #define CHANGE_SAVE NULL
170 #define CHANGE_IMPOSSIBLE (ir_node *)1
171 #define is_conflicting_node(n) (((int)n) > 1)
174 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
175 * other nodes are changed too, as required to preserve correctness. Function is
176 * aware of local and global pinning. Recursive.
177 * @param irn The node to set the color for
178 * @param col The color to set
179 * @param trigger The irn that caused the wish to change the color of the irn
180 * @return CHANGE_SAVE iff setting the color is possible, with all transiteve effects.
181 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
182 * Else the first conflicting ir_node encountered is returned.
184 * ASSUMPTION: Assumes that a life range of a single value can't be spilt into
185 * several smaller intervals where other values can live in between.
186 * This should be true in SSA.
188 static const ir_node *qnode_color_irn(const qnode_t *qn, const ir_node *irn, int col, const ir_node *trigger) {
190 struct obstack confl_ob;
191 ir_node **confl, *cn;
194 DBG((dbg, LEVEL_3, "\t %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
195 obstack_init(&confl_ob);
196 irn_col = qnode_get_new_color(qn, irn);
200 if (!is_possible_color(irn, col))
202 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
207 /* get all nodes which would conflict with this change */
213 irn_bl = get_nodes_block(irn);
215 /* first check for a conflicting node which is 'living in' the irns block */
218 pset *live_ins = get_live_in(irn_bl);
219 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
220 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
221 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
222 obstack_ptr_grow(&confl_ob, n);
223 pset_break(live_ins);
228 /* setup the queue of blocks. */
230 obstack_ptr_grow(&q, irn_bl);
234 /* process the queue. The code below checks for every block dominated
235 * by the irns one, and in which the irn is live, if there are
236 * conflicting nodes */
238 ir_node *curr_bl, *sub_bl;
241 curr_bl = ((ir_node **)obstack_base(&q))[out++];
243 /* Add to the result all nodes in the block, which have
244 * the target color and interfere with the irn */
245 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
246 ir_node *n = get_irn_out(curr_bl, i);
247 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
248 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
249 obstack_ptr_grow(&confl_ob, n);
253 /* If irn lives out check i-dominated blocks where the irn lives in */
255 if (is_live_out(curr_bl, irn)) {
256 dominates_for_each(curr_bl, sub_bl)
257 if (is_live_in(sub_bl, irn)) {
258 obstack_ptr_grow(&q, sub_bl);
263 obstack_free(&q, NULL);
264 obstack_ptr_grow(&confl_ob, NULL);
265 confl = (ir_node **) obstack_finish(&confl_ob);
268 /* process all nodes which would conflict with this change */
269 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
270 const ir_node *sub_res;
272 /* try to color the conflicting node cn with the color of the irn itself */
273 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
274 if (sub_res != CHANGE_SAVE) {
279 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
282 DBG((dbg, LEVEL_3, "\t %n save\n", irn));
283 obstack_free(&confl_ob, NULL);
284 qnode_set_new_color(qn, irn, col);
288 DBG((dbg, LEVEL_3, "\t %n impossible\n", irn));
289 obstack_free(&confl_ob, NULL);
293 DBG((dbg, LEVEL_3, "\t %n conflicting\n", irn));
294 obstack_free(&confl_ob, NULL);
299 * Tries to set the colors for all members of this queue node;
300 * to the target color qn->color
301 * @returns 1 iff all members colors could be set
304 static int qnode_try_color(const qnode_t *qn) {
306 for (i=0; i<qn->mis_size; ++i) {
307 const ir_node *test_node, *confl_node;
309 test_node = qn->mis[i];
310 DBG((dbg, LEVEL_3, "\t Testing %n\n", test_node));
311 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
313 if (confl_node == CHANGE_SAVE) {
314 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
315 qnode_pin_local(qn, test_node);
316 } else if (confl_node == CHANGE_IMPOSSIBLE) {
317 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
318 qnode_add_conflict(qn, test_node, test_node);
320 if (qnode_is_pinned_local(qn, confl_node)) {
321 /* changing test_node would change back a node of current ou */
322 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
323 qnode_add_conflict(qn, confl_node, test_node);
325 if (pset_find_ptr(pinned_global, confl_node)) {
326 /* changing test_node would change back a node of a prior ou */
327 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
328 qnode_add_conflict(qn, test_node, test_node);
332 if (confl_node != CHANGE_SAVE)
339 * Determines a maximum independent set with respect to the interference and
340 * conflict edges of all nodes in a qnode.
342 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
343 int all_size, curr_size, i, o;
345 const ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
347 /* all contains all nodes not removed in this qn */
349 for (i=0; i<ou->node_count; ++i)
350 if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i]))
351 all[all_size++] = ou->nodes[i];
353 /* which[i] says which element to take out of all[] and put into curr[i] */
354 which = alloca(all_size*sizeof(*which));
355 for (curr_size=0; curr_size<all_size; ++curr_size)
356 which[curr_size] = curr_size;
358 /* stores the currently examined set */
359 curr = alloca(all_size*sizeof(*curr));
361 while (1) { /* this loop will terminate because at least a single node'll be a max indep. set */
362 /* build current set */
363 for (i=0; i<curr_size; ++i)
364 curr[i] = all[which[all_size-curr_size+i]];
366 /* check current set */
367 for (i=0; i<curr_size; ++i)
368 for (o=i+1; o<curr_size; ++o)
369 if (qnode_are_conflicting(qn, curr[i], curr[o]))
372 /* We had no conflict. This is the max indep. set */
373 qn->mis_size = curr_size;
374 for (i=0; i<curr_size; ++i)
375 qn->mis[i] = curr[i];
379 /* We had a conflict. Generate next set */
380 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
382 for (i=0; i<curr_size; ++i)
383 which[all_size-curr_size+i] = i;
390 } while (!(which[pos] = (which[pos]+1) % all_size));
392 for (i=pos+1; i<all_size; ++i)
393 which[i] = MIN(which[i-1]+1, all_size-1);
396 for (i=all_size-curr_size; i<all_size-1; ++i)
397 if (which[i]>=which[i+1]) {
407 * Creates a new qnode
409 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
410 qnode_t *qn = malloc(sizeof(*qn));
412 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
413 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
414 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
419 * Frees space used by a queue node
421 static INLINE void free_qnode(qnode_t *qn) {
422 del_set(qn->conflicts);
423 del_set(qn->changed_nodes);
429 * Inserts a qnode in the sorted queue of the outimization unit. Queue is
430 * ordered by field 'size' (the size of the mis) in decreasing order.
432 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
433 struct list_head *lh;
435 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
436 /* root node is not in qnode */
441 qnode_max_ind_set(qn, ou);
443 /* set ou->mis_size for lower bound compution */
444 if (ou->mis_size < qn->mis_size)
445 ou->mis_size = qn->mis_size;
447 /* do the insertion */
448 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with size %d\n", qn->color, qn->mis_size));
450 while (lh->next != &ou->queue) {
451 qnode_t *curr = list_entry_queue(lh->next);
452 if (curr->mis_size <= qn->mis_size)
456 list_add(&qn->queue, lh);
460 * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
461 * number of copy instructions placed during SSA-destruction and lowering.
462 * Works only for opt units with exactly 1 root node, which is the
463 * case for approximately 80% of all phi classes and all register constrained
464 * nodes. (All other phi classes are reduced to this case.)
466 static void ou_optimize(unit_t *ou) {
470 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
471 for (i=0; i<ou->node_count; ++i)
472 DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
475 INIT_LIST_HEAD(&ou->queue);
476 for (i=0; i<MAX_COLORS; ++i)
477 if (is_possible_color(ou->nodes[0], i))
478 ou_insert_qnode(ou, new_qnode(ou, i));
481 while (!list_empty(&ou->queue)) {
482 /* get head of queue */
483 curr = list_entry_queue(ou->queue.next);
484 list_del(&curr->queue);
485 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with size %d\n", curr->color, curr->mis_size));
488 if (qnode_try_color(curr))
490 /* no success, so re-insert */
491 del_set(curr->changed_nodes);
492 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
493 ou_insert_qnode(ou, curr);
496 /* apply the best found qnode */
497 if (curr->mis_size >= 2) {
498 DBG((dbg, LEVEL_1, "\t Best color: %d Copies: %d/%d\n", curr->color, ou->interf+ou->node_count-curr->mis_size, ou->interf+ou->node_count-1));
499 /* globally pin root and eventually others */
500 pset_insert_ptr(pinned_global, ou->nodes[0]);
501 for (i=1; i<ou->node_count; ++i) {
502 const ir_node *irn = ou->nodes[i];
503 int nc = qnode_get_new_color(curr, irn);
504 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
505 pset_insert_ptr(pinned_global, irn);
508 /* set color of all changed nodes */
510 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
511 /* NO_COLOR is possible, if we had an undo */
512 if (ns->new_color != NO_COLOR) {
513 DBG((dbg, LEVEL_2, "\t color(%n) := %d\n", ns->irn, ns->new_color));
514 set_irn_color(ns->irn, ns->new_color);
519 /* free best qnode (curr) and queue */
521 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
525 void co_heur_opt(copy_opt_t *co) {
527 dbg = firm_dbg_register("ir.be.copyoptheur");
528 firm_dbg_set_mask(dbg, DEBUG_LVL);
530 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
531 list_for_each_entry(unit_t, curr, &co->units, units)
532 if (curr->node_count > 1)
535 del_pset(pinned_global);