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.
23 #include "becopyopt.h"
24 #include "becopystat.h"
26 #define DEBUG_LVL 0 //SET_LEVEL_1
27 static firm_dbg_module_t *dbg = NULL;
29 #define SLOTS_PINNED_GLOBAL 256
30 #define SLOTS_CONFLICTS 8
31 #define SLOTS_CHANGED_NODES 32
33 #define MIN(a,b) ((a<b)?a:b)
34 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
35 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
38 * Modeling additional conflicts between nodes. NOT live range interference
40 typedef struct _conflict_t {
41 const ir_node *n1, *n2;
45 * If an irn is changed, the changes first get stored in a node_stat_t,
46 * to allow undo of changes (=drop new data) in case of conflicts.
48 typedef struct _node_stat_t {
55 * Represents a node in the optimization queue.
57 typedef struct _qnode_t {
58 struct list_head queue; /**< chaining of unit_t->queue */
59 const unit_t *ou; /**< the opt unit this qnode belongs to */
60 int color; /**< target color */
61 set *conflicts; /**< contains conflict_t's. All internal conflicts */
62 int mis_size; /**< number of nodes in the mis. */
63 const ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
64 set *changed_nodes; /**< contains node_stat_t's. */
67 pset *pinned_global; /**< optimized nodes should not be altered any more */
69 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
70 const conflict_t *xx = x;
71 const conflict_t *yy = y;
72 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
76 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
77 * The next maximum independent set build, will regard it.
79 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
81 DBG((dbg, LEVEL_4, "\t %n -- %n\n", n1, n2));
83 if ((int)n1 < (int)n2) {
90 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
94 * Checks if two nodes are in a conflict.
96 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
98 /* search for live range interference */
99 if (n1!=n2 && values_interfere(n1, n2))
101 /* search for recoloring conflicts */
102 if ((int)n1 < (int)n2) {
109 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
112 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
113 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
117 * Finds a node status entry of a node if existent. Otherwise return NULL
119 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, const ir_node *irn) {
122 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
126 * Finds a node status entry of a node if existent. Otherwise it will return
127 * an initialized new entry for this node.
129 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, const ir_node *irn) {
132 find.new_color = NO_COLOR;
133 find.pinned_local = 0;
134 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
138 * Returns the virtual color of a node if set before, else returns the real color.
140 static INLINE int qnode_get_new_color(const qnode_t *qn, const ir_node *irn) {
141 node_stat_t *found = qnode_find_node(qn, irn);
143 return found->new_color;
145 return get_irn_color(irn);
149 * Sets the virtual color of a node.
151 static INLINE void qnode_set_new_color(const qnode_t *qn, const ir_node *irn, int color) {
152 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
153 found->new_color = color;
157 * Checks if a node is local pinned. A node is local pinned, iff it belongs
158 * to the same optimization unit and has been optimized before the current
161 static INLINE int qnode_is_pinned_local(const qnode_t *qn, const ir_node *irn) {
162 node_stat_t *found = qnode_find_node(qn, irn);
164 return found->pinned_local;
170 * Local-pins a node, so optimizations of further nodes of the same opt unit
171 * can handle situations in which a color change would undo prior optimizations.
173 static INLINE void qnode_pin_local(const qnode_t *qn, const ir_node *irn) {
174 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
175 found->pinned_local = 1;
179 * Possible return values of qnode_color_irn()
181 #define CHANGE_SAVE NULL
182 #define CHANGE_IMPOSSIBLE (ir_node *)1
183 #define is_conflicting_node(n) (((int)n) > 1)
186 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
187 * other nodes are changed too, as required to preserve correctness. Function is
188 * aware of local and global pinning. Recursive.
189 * @param irn The node to set the color for
190 * @param col The color to set
191 * @param trigger The irn that caused the wish to change the color of the irn
192 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
193 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
194 * Else the first conflicting ir_node encountered is returned.
196 * ASSUMPTION: Assumes that a life range of a single value can't be split into
197 * several smaller intervals where other values can live in between.
198 * This should be true in SSA.
200 static const ir_node *qnode_color_irn(const qnode_t *qn, const ir_node *irn, int col, const ir_node *trigger) {
202 struct obstack confl_ob;
203 ir_node **confl, *cn;
206 DBG((dbg, LEVEL_3, "\t %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
207 obstack_init(&confl_ob);
208 irn_col = qnode_get_new_color(qn, irn);
212 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
216 if (!qn->ou->co->isa->is_reg_allocatable(irn, arch_register_for_index(qn->ou->co->cls, col)))
219 /* get all nodes which would conflict with this change */
225 irn_bl = get_nodes_block(irn);
227 /* first check for a conflicting node which is 'living in' the irns block */
230 pset *live_ins = get_live_in(irn_bl);
231 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
232 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
233 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
234 obstack_ptr_grow(&confl_ob, n);
235 pset_break(live_ins);
240 /* setup the queue of blocks. */
242 obstack_ptr_grow(&q, irn_bl);
246 /* process the queue. The code below checks for every block dominated
247 * by the irns one, and in which the irn is live, if there are
248 * conflicting nodes */
250 ir_node *curr_bl, *sub_bl;
253 curr_bl = ((ir_node **)obstack_base(&q))[out++];
255 /* Add to the result all nodes in the block, which have
256 * the target color and interfere with the irn */
257 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
258 ir_node *n = get_irn_out(curr_bl, i);
259 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
260 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
261 obstack_ptr_grow(&confl_ob, n);
265 /* If irn lives out check i-dominated blocks where the irn lives in */
267 if (is_live_out(curr_bl, irn)) {
268 dominates_for_each(curr_bl, sub_bl)
269 if (is_live_in(sub_bl, irn)) {
270 obstack_ptr_grow(&q, sub_bl);
275 obstack_free(&q, NULL);
276 obstack_ptr_grow(&confl_ob, NULL);
277 confl = (ir_node **) obstack_finish(&confl_ob);
280 /* process all nodes which would conflict with this change */
281 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
282 const ir_node *sub_res;
284 /* try to color the conflicting node cn with the color of the irn itself */
285 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
286 if (sub_res != CHANGE_SAVE) {
291 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
294 DBG((dbg, LEVEL_3, "\t %n save\n", irn));
295 obstack_free(&confl_ob, NULL);
296 qnode_set_new_color(qn, irn, col);
300 DBG((dbg, LEVEL_3, "\t %n impossible\n", irn));
301 obstack_free(&confl_ob, NULL);
302 return CHANGE_IMPOSSIBLE;
305 DBG((dbg, LEVEL_3, "\t %n conflicting\n", irn));
306 obstack_free(&confl_ob, NULL);
311 * Tries to set the colors for all members of this queue node;
312 * to the target color qn->color
313 * @returns 1 iff all members colors could be set
316 static int qnode_try_color(const qnode_t *qn) {
318 for (i=0; i<qn->mis_size; ++i) {
319 const ir_node *test_node, *confl_node;
321 test_node = qn->mis[i];
322 DBG((dbg, LEVEL_3, "\t Testing %n\n", test_node));
323 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
325 if (confl_node == CHANGE_SAVE) {
326 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
327 qnode_pin_local(qn, test_node);
328 } else if (confl_node == CHANGE_IMPOSSIBLE) {
329 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
330 qnode_add_conflict(qn, test_node, test_node);
332 if (qnode_is_pinned_local(qn, confl_node)) {
333 /* changing test_node would change back a node of current ou */
334 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
335 qnode_add_conflict(qn, confl_node, test_node);
337 if (pset_find_ptr(pinned_global, confl_node)) {
338 /* changing test_node would change back a node of a prior ou */
339 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
340 qnode_add_conflict(qn, test_node, test_node);
344 if (confl_node != CHANGE_SAVE)
351 * Determines a maximum independent set with respect to the interference and
352 * conflict edges of all nodes in a qnode.
354 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
355 int all_size, curr_size, i, o;
357 const ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
359 /* all contains all nodes not removed in this qn */
361 for (i=0; i<ou->node_count; ++i)
362 if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i]))
363 all[all_size++] = ou->nodes[i];
365 /* which[i] says which element to take out of all[] and put into curr[i] */
366 which = alloca(all_size*sizeof(*which));
367 for (curr_size=0; curr_size<all_size; ++curr_size)
368 which[curr_size] = curr_size;
370 /* stores the currently examined set */
371 curr = alloca(all_size*sizeof(*curr));
373 while (1) { /* this loop will terminate because at least a single node will be a max indep. set */
374 /* build current set */
375 for (i=0; i<curr_size; ++i)
376 curr[i] = all[which[all_size-curr_size+i]];
378 /* check current set */
379 for (i=0; i<curr_size; ++i)
380 for (o=i+1; o<curr_size; ++o)
381 if (qnode_are_conflicting(qn, curr[i], curr[o]))
384 /* We had no conflict. This is the max indep. set */
385 qn->mis_size = curr_size;
386 for (i=0; i<curr_size; ++i)
387 qn->mis[i] = curr[i];
391 /* We had a conflict. Generate next set */
392 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
394 for (i=0; i<curr_size; ++i)
395 which[all_size-curr_size+i] = i;
402 } while (!(which[pos] = (which[pos]+1) % all_size));
404 for (i=pos+1; i<all_size; ++i)
405 which[i] = MIN(which[i-1]+1, all_size-1);
408 for (i=all_size-curr_size; i<all_size-1; ++i)
409 if (which[i]>=which[i+1]) {
419 * Creates a new qnode
421 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
422 qnode_t *qn = xmalloc(sizeof(*qn));
425 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
426 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
427 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
432 * Frees space used by a queue node
434 static INLINE void free_qnode(qnode_t *qn) {
435 del_set(qn->conflicts);
436 del_set(qn->changed_nodes);
442 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
443 * ordered by field 'size' (the size of the mis) in decreasing order.
445 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
446 struct list_head *lh;
448 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
449 /* root node is not in qnode */
454 qnode_max_ind_set(qn, ou);
456 /* set ou->mis_size for lower bound compution */
457 if (ou->mis_size < qn->mis_size)
458 ou->mis_size = qn->mis_size;
460 /* do the insertion */
461 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with size %d\n", qn->color, qn->mis_size));
463 while (lh->next != &ou->queue) {
464 qnode_t *curr = list_entry_queue(lh->next);
465 if (curr->mis_size <= qn->mis_size)
469 list_add(&qn->queue, lh);
473 * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
474 * number of copy instructions placed during SSA-destruction and lowering.
475 * Works only for opt units with exactly 1 root node, which is the
476 * case for approximately 80% of all phi classes and all register constrained
477 * nodes. (All other phi classes are reduced to this case.)
479 static void ou_optimize(unit_t *ou) {
482 bitset_t *pos_regs = bitset_alloca(ou->co->cls->n_regs);
484 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
485 for (i=0; i<ou->node_count; ++i)
486 DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
489 INIT_LIST_HEAD(&ou->queue);
490 ou->co->isa->get_allocatable_regs(ou->nodes[0], ou->co->cls, pos_regs);
491 bitset_foreach(pos_regs, i)
492 ou_insert_qnode(ou, new_qnode(ou, i));
495 while (!list_empty(&ou->queue)) {
496 /* get head of queue */
497 curr = list_entry_queue(ou->queue.next);
498 list_del(&curr->queue);
499 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with size %d\n", curr->color, curr->mis_size));
502 if (qnode_try_color(curr))
504 /* no success, so re-insert */
505 del_set(curr->changed_nodes);
506 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
507 ou_insert_qnode(ou, curr);
510 /* apply the best found qnode */
511 if (curr->mis_size >= 2) {
514 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));
515 /* globally pin root and eventually others */
516 pset_insert_ptr(pinned_global, ou->nodes[0]);
517 for (i=1; i<ou->node_count; ++i) {
518 const ir_node *irn = ou->nodes[i];
519 int nc = qnode_get_new_color(curr, irn);
520 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
521 pset_insert_ptr(pinned_global, irn);
524 /* set color of all changed nodes */
525 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
526 /* NO_COLOR is possible, if we had an undo */
527 if (ns->new_color != NO_COLOR) {
528 DBG((dbg, LEVEL_2, "\t color(%n) := %d\n", ns->irn, ns->new_color));
529 set_irn_color(ns->irn, ns->new_color);
534 /* free best qnode (curr) and queue */
536 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
540 void co_heur_opt(copy_opt_t *co) {
542 dbg = firm_dbg_register("ir.be.copyoptheur");
543 firm_dbg_set_mask(dbg, DEBUG_LVL);
544 if (!strcmp(co->name, DEBUG_IRG))
545 firm_dbg_set_mask(dbg, -1);
547 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
548 list_for_each_entry(unit_t, curr, &co->units, units)
549 if (curr->node_count > 1)
552 del_pset(pinned_global);