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"
14 #define DEBUG_LVL SET_LEVEL_1
15 static firm_dbg_module_t *dbg = NULL;
17 #define SLOTS_PINNED_GLOBAL 256
18 #define SLOTS_CONFLICTS 8
19 #define SLOTS_CHANGED_NODES 32
21 #define MIN(a,b) ((a<b)?a:b)
22 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
23 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
26 * Modeling additional conflicts between nodes. NOT live range interference
28 typedef struct _conflict_t {
29 const ir_node *n1, *n2;
33 * If an irn is changed, the changes first get stored in a node_stat_t,
34 * to allow undo of changes (=drop new data) in case of conflicts.
36 typedef struct _node_stat_t {
43 * Represents a node in the optimization queue.
45 typedef struct _qnode_t {
46 struct list_head queue; /**< chaining of unit_t->queue */
47 int color; /**< target color */
48 bitset_t *nodes; /**< marks the nodes of unit_t->nodes[i] beeing part of this qnode */
49 set *conflicts; /**< contains conflict_t's. All internal conflicts */
50 int size; /**< number of nodes in the mis. equals number of set bits in mis */
51 bitset_t *mis; /**< marks the nodes of unit_t->nodes[i] beeing part of the max idependent set */
52 set *changed_nodes; /**< contains node_stat_t's. */
56 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
57 const conflict_t *xx = x;
58 const conflict_t *yy = y;
59 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
63 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
64 * The next maximum independent set build, will regard it.
66 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
68 DBG((dbg, LEVEL_3, "\t %n -- %n\n", n1, n2));
70 if ((int)n1 < (int)n2) {
77 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
81 * Checks if two nodes are in a conflict.
83 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
85 /* search for live range interference */
86 if (values_interfere(n1, n2))
88 /* search for recoloring conflicts */
89 if ((int)n1 < (int)n2) {
96 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
99 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
100 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
104 * Finds a node status entry of a node if existent. Otherwise return NULL
106 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, const ir_node *irn) {
109 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
113 * Finds a node status entry of a node if existent. Otherwise it will return
114 * an initialized new entry for this node.
116 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, const ir_node *irn) {
119 find.new_color = NO_COLOR;
120 find.pinned_local = 0;
121 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
125 * Returns the virtual color of a node if set before, else returns the real color.
127 static INLINE int qnode_get_new_color(const qnode_t *qn, const ir_node *irn) {
128 node_stat_t *found = qnode_find_node(qn, irn);
130 return found->new_color;
132 return get_irn_color(irn);
136 * Sets the virtual color of a node.
138 static INLINE void qnode_set_new_color(const qnode_t *qn, const ir_node *irn, int color) {
139 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
140 found->new_color = color;
144 * Checks if a node is local pinned. A node is local pinned, iff it belongs
145 * to the same optimization unit and has been optimized before the current
148 static INLINE int qnode_is_pinned_local(const qnode_t *qn, const ir_node *irn) {
149 node_stat_t *found = qnode_find_node(qn, irn);
151 return found->pinned_local;
157 * Local-pins a node, so optimizations of further nodes of the same opt unit
158 * can handle situations in which a color change would undo prior optimizations.
160 static INLINE void qnode_pin_local(const qnode_t *qn, const ir_node *irn) {
161 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
162 found->pinned_local = 1;
166 * Possible return values of qnode_color_irn()
168 #define CHANGE_SAVE NULL
169 #define CHANGE_IMPOSSIBLE (ir_node *)1
170 #define is_conflicting_node(n) (((int)n) > 1)
173 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
174 * other nodes are changed too, as required to preserve correctness. Function is
175 * aware of local and global pinning. Recursive.
176 * @param irn The node to set the color for
177 * @param col The color to set
178 * @param trigger The irn that caused the wish to change the color of the irn
179 * @return CHANGE_SAVE iff setting the color is possible, with all transiteve effects.
180 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
181 * Else the first conflicting ir_node encountered is returned.
183 * ASSUMPTION: Assumes that a life range of a single value can't be spilt into
184 * several smaller intervals where other values can live in between.
185 * This should be true in SSA.
187 static const ir_node *qnode_color_irn(const qnode_t *qn, const ir_node *irn, int col, const ir_node *trigger, const copy_opt_t *co) {
189 struct obstack confl_ob;
190 ir_node **confl, *cn;
193 DBG((dbg, LEVEL_3, "\t\t%n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
194 obstack_init(&confl_ob);
195 irn_col = qnode_get_new_color(qn, irn);
199 if (pset_find_ptr(co->pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
204 /* get all nodes which would conflict with this change */
210 irn_bl = get_nodes_block(irn);
212 /* first check for a conflicting node which is 'living in' the irns block */
215 pset *live_ins = get_live_in(irn_bl);
216 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
217 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
218 DBG((dbg, LEVEL_4, "\t\t %n\ttroubles\n", n));
219 obstack_ptr_grow(&confl_ob, n);
220 pset_break(live_ins);
225 /* setup the queue of blocks. */
227 obstack_ptr_grow(&q, irn_bl);
231 /* process the queue. The code below checks for every block dominated
232 * by the irns one, and in which the irn is live, if there are
233 * conflicting nodes */
235 ir_node *curr_bl, *sub_bl;
238 curr_bl = ((ir_node **)obstack_base(&q))[out++];
240 /* Add to the result all nodes in the block, which have
241 * the target color and interfere with the irn */
242 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
243 ir_node *n = get_irn_out(curr_bl, i);
244 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
245 DBG((dbg, LEVEL_4, "\t\t %n\ttroubles\n", n));
246 obstack_ptr_grow(&confl_ob, n);
250 /* If irn lives out check i-dominated blocks where the irn lives in */
252 if (is_live_out(curr_bl, irn)) {
253 dominates_for_each(curr_bl, sub_bl)
254 if (is_live_in(sub_bl, irn)) {
255 obstack_ptr_grow(&q, sub_bl);
260 obstack_free(&q, NULL);
261 obstack_ptr_grow(&confl_ob, NULL);
262 confl = (ir_node **) obstack_finish(&confl_ob);
265 /* process all nodes which would conflict with this change */
266 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
267 const ir_node *sub_res;
269 /* try to color the conflicting node cn with the color of the irn itself */
270 sub_res = qnode_color_irn(qn, cn, irn_col, irn, co);
271 if (sub_res != CHANGE_SAVE) {
276 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
279 DBG((dbg, LEVEL_3, "\t\t%n save\n", irn));
280 obstack_free(&confl_ob, NULL);
281 qnode_set_new_color(qn, irn, col);
285 DBG((dbg, LEVEL_3, "\t\t%n conflicting\n", irn));
286 obstack_free(&confl_ob, NULL);
291 * Tries to set the colors for all members of this queue node;
292 * to the target color qn->color
293 * @returns 1 iff all members colors could be set
296 static int qnode_try_color(const qnode_t *qn, const unit_t *ou, const copy_opt_t *co) {
298 bitset_foreach(qn->mis, i) {
299 const ir_node *test_node, *confl_node;
301 test_node = ou->nodes[i];
302 DBG((dbg, LEVEL_2, "\t Testing %n\n", test_node));
303 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node, co);
305 if (confl_node == CHANGE_SAVE) {
306 DBG((dbg, LEVEL_2, "\t Save\n"));
307 qnode_pin_local(qn, test_node);
308 } else if (confl_node == CHANGE_IMPOSSIBLE) {
309 DBG((dbg, LEVEL_2, "\t Impossible\n"));
310 bitset_clear(qn->nodes, i);
312 DBG((dbg, LEVEL_2, "\t Conflicting\n"));
313 if (qnode_is_pinned_local(qn, confl_node)) {
314 /* changing test_node would change back a node of current ou */
315 qnode_add_conflict(qn, confl_node, test_node);
317 if (pset_find_ptr(co->pinned_global, confl_node)) {
318 /* changing test_node would change back a node of a prior ou */
319 bitset_clear(qn->nodes, i);
323 if (confl_node != CHANGE_SAVE)
330 * Creates a new qnode
332 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
333 qnode_t *qn = malloc(sizeof(*qn));
335 qn->mis = bitset_malloc(ou->node_count-1);
336 qn->nodes = bitset_malloc(ou->node_count-1);
337 bitset_set_all(qn->nodes);
338 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
339 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
344 * Frees space used by a queue node
346 static INLINE void free_qnode(qnode_t *qn) {
347 del_set(qn->conflicts);
348 del_set(qn->changed_nodes);
349 bitset_free(qn->nodes);
350 bitset_free(qn->mis);
355 * Determines a maximum independent set with respect to the interference and
356 * conflict edges of all nodes in a qnode.
358 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
359 int all_size, curr_size, i, o;
361 const ir_node **curr, **all = alloca(bitset_popcnt(qn->nodes) * sizeof(*all));
363 /* all contains all nodes not removed in this qn */
364 assert(bitset_is_set(qn->nodes, 0) && "root is not element of this queue node");
366 bitset_foreach(qn->nodes, i)
367 all[all_size++] = ou->nodes[i];
369 /* which[i] says which element to take out of all[] and put into curr[i] */
370 which = alloca(all_size*sizeof(*which));
371 for (curr_size=0; curr_size<all_size; ++curr_size)
372 which[curr_size] = curr_size;
374 /* stores the currently examined set */
375 curr = alloca(all_size*sizeof(*curr));
377 while (1) { /* this loop will terminate because at least a single node'll be a max indep. set */
378 /* build current set */
379 for (i=0; i<curr_size; ++i)
380 curr[i] = all[which[all_size-curr_size+i]];
382 /* check current set */
383 for (i=0; i<curr_size; ++i)
384 for (o=i+1; o<curr_size; ++o)
385 if (qnode_are_conflicting(qn, curr[i], curr[o]))
388 /* We had no conflict. This is the max indep. set */
389 bitset_clear_all(qn->mis);
390 qn->size = curr_size;
391 for (i=0; i<curr_size; ++i)
392 for (o=0; o<ou->node_count; ++o)
393 if (curr[i] == ou->nodes[o])
394 bitset_set(qn->mis, o);
398 /* We had a conflict. Generate next set */
399 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
401 for (i=0; i<curr_size; ++i)
402 which[all_size-curr_size+i] = i;
409 } while (!(which[pos] = (which[pos]+1) % all_size));
411 for (i=pos+1; i<all_size; ++i)
412 which[i] = MIN(which[i-1]+1, all_size-1);
415 for (i=all_size-curr_size; i<all_size-1; ++i)
416 if (which[i]>=which[i+1]) {
426 * Inserts a qnode in the sorted queue of the outimization unit. Queue is
427 * ordered by field 'size' (the size of the mis) in decreasing order.
429 static INLINE void qnode_insert(qnode_t *qn, unit_t *ou) {
430 struct list_head *lh;
431 /* root node is not in qnode */
432 if (!bitset_is_set(qn->nodes, 0)) {
437 qnode_max_ind_set(qn, ou);
438 if (ou->mis_size < qn->size)
439 ou->mis_size = qn->size;
442 while (lh->next != &ou->queue) {
443 qnode_t *curr = list_entry_queue(lh->next);
444 if (curr->size <= qn->size)
448 list_add(&qn->queue, lh);
452 * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
453 * number of copy instructions placed during SSA-destruction and lowering.
454 * Works only for opt units with exactly 1 root node, which is the
455 * case for approximately 80% of all phi classes and all register constrained
456 * nodes. (All other phi classes are reduced to this case.)
458 static void co_heur_opt_unit(const copy_opt_t *co, unit_t *ou) {
463 INIT_LIST_HEAD(&ou->queue);
464 for (i=MAX_COLORS-1; i>=0; --i)
465 if (is_possible_color(ou->nodes[0], i))
466 qnode_insert(new_qnode(ou, i), ou);
469 while (!list_empty(&ou->queue)) {
470 /* get head of queue */
471 curr = list_entry_queue(ou->queue.next);
472 list_del(&curr->queue);
474 if (qnode_try_color(curr, ou, co))
476 /* no success, so re-insert */
477 del_set(curr->changed_nodes);
478 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
479 qnode_insert(curr, ou);
482 /* apply the best found qnode */
483 if (curr->size >= 2) {
484 DBG((dbg, 1, "\tBest color: %d Copies: %d/%d\n", curr->color, ou->interf+ou->node_count-1-curr->size));
485 /* globally pin root and eventually others */
486 pset_insert_ptr(co->pinned_global, ou->nodes[0]);
487 for (i=1; i<ou->node_count; ++i) {
488 const ir_node *irn = ou->nodes[i];
489 int nc = qnode_get_new_color(curr, irn);
490 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
491 pset_insert_ptr(co->pinned_global, irn);
494 /* set color of all changed nodes */
496 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
497 /* NO_COLOR is possible, if we had an undo */
498 if (ns->new_color != NO_COLOR) {
499 DBG((dbg, 1, "\t color(%n) := %d\n", ns->irn, ns->new_color));
500 set_irn_color(ns->irn, ns->new_color);
505 /* free best qnode (curr) and queue */
507 list_for_each_entry(qnode_t, curr, &ou->queue, queue)
511 void co_heur_opt(copy_opt_t *co) {
513 dbg = firm_dbg_register("ir.be.copyoptheur");
514 firm_dbg_set_mask(dbg, DEBUG_LVL);
516 co->pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
518 list_for_each_entry(unit_t, curr, &co->units, units)
519 co_heur_opt_unit(co, curr);
521 del_pset(co->pinned_global);