2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Spillslot coalescer.
23 * @author Matthias Braun
36 #include "unionfind.h"
42 #include "bespillslots.h"
43 #include "bechordal_t.h"
44 #include "bestatevent.h"
46 #include "beintlive_t.h"
50 #define DBG_COALESCING 1
51 #define DBG_INTERFERENCES 2
53 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
55 typedef struct spill_t {
57 const ir_mode *mode; /**< mode of the spilled value */
58 int alignment; /**< alignment for the spilled value */
62 typedef struct affinity_edge_t {
74 affinity_edge_t **affinity_edges;
76 set_frame_entity_func set_frame_entity;
77 bool at_begin; /**< frame entities should be allocate at
78 the beginning of the stackframe */
81 /** Compare 2 affinity edges (used in quicksort) */
82 static int cmp_affinity(const void *d1, const void *d2)
84 const affinity_edge_t * const *e1 = (const affinity_edge_t**)d1;
85 const affinity_edge_t * const *e2 = (const affinity_edge_t**)d2;
86 double aff1 = (*e1)->affinity;
87 double aff2 = (*e2)->affinity;
89 /* sort in descending order */
92 } else if (aff1 > aff2) {
95 int slot11 = (*e1)->slot1;
96 int slot21 = (*e2)->slot1;
97 if (slot11 < slot21) {
99 } else if (slot11 > slot21) {
102 int slot12 = (*e1)->slot2;
103 int slot22 = (*e2)->slot2;
104 return (slot12<slot22) - (slot12<slot22);
109 static spill_t *get_spill(be_fec_env_t *env, ir_node *node)
111 assert(rbitset_is_set(env->spills_set, get_irn_idx(node)));
112 return (spill_t*)get_irn_link(node);
115 static inline ir_node *get_memory_edge(const ir_node *node)
119 arity = get_irn_arity(node);
120 for (i = arity - 1; i >= 0; --i) {
121 ir_node *arg = get_irn_n(node, i);
122 if (get_irn_mode(arg) == mode_M)
129 static spill_t *collect_spill(be_fec_env_t *env, ir_node *node,
130 const ir_mode *mode, int align)
134 /* already in spill set? */
135 unsigned idx = get_irn_idx(node);
136 if (rbitset_is_set(env->spills_set, idx)) {
137 spill_t *spill = get_spill(env, node);
138 assert(spill->mode == mode);
139 assert(spill->alignment == align);
142 rbitset_set(env->spills_set, idx);
144 spill = OALLOC(&env->obst, spill_t);
145 /* insert into set of spills if not already there */
148 spill->alignment = align;
149 spill->spillslot = (int)ARR_LEN(env->spills);
150 ARR_APP1(spill_t*, env->spills, spill);
151 set_irn_link(node, spill);
152 DB((dbg, DBG_COALESCING, "Slot %d: %+F\n", spill->spillslot, node));
155 int arity = get_irn_arity(node);
157 for (i = 0; i < arity; ++i) {
158 affinity_edge_t *affinty_edge;
159 ir_node *arg = get_irn_n(node, i);
160 spill_t *arg_spill = collect_spill(env, arg, mode, align);
161 ir_node *block = get_nodes_block(arg);
163 /* add an affinity edge */
164 affinty_edge = OALLOC(&env->obst, affinity_edge_t);
165 affinty_edge->affinity = get_block_execfreq(block);
166 affinty_edge->slot1 = spill->spillslot;
167 affinty_edge->slot2 = arg_spill->spillslot;
168 ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
175 void be_node_needs_frame_entity(be_fec_env_t *env, ir_node *node,
176 const ir_mode *mode, int align)
178 ir_node *spillnode = get_memory_edge(node);
179 assert(spillnode != NULL);
181 /* walk upwards and collect all phis and spills on this way */
182 collect_spill(env, spillnode, mode, align);
184 ARR_APP1(ir_node *, env->reloads, node);
187 static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
188 int* spillslot_unionfind, int s1, int s2)
194 /* merge spillslots and interferences */
195 res = uf_union(spillslot_unionfind, s1, s2);
196 /* we assume that we always merge s2 to s1 so swap s1, s2 if necessary */
203 bitset_or(interferences[s1], interferences[s2]);
205 /* update other interferences */
206 spillcount = ARR_LEN(env->spills);
207 for (i = 0; i < spillcount; ++i) {
208 bitset_t *intfs = interferences[i];
209 if (bitset_is_set(intfs, s2))
210 bitset_set(intfs, s1);
216 static int my_values_interfere2(ir_graph *irg, const ir_node *a,
219 be_lv_t *lv = be_get_irg_liveness(irg);
221 int a2b = _value_dominates(a, b);
222 int b2a = _value_dominates(b, a);
224 /* If there is no dominance relation, they do not interfere. */
225 if ((a2b | b2a) > 0) {
229 * Adjust a and b so, that a dominates b if
230 * a dominates b or vice versa.
233 const ir_node *t = a;
238 bb = get_nodes_block(b);
241 * If a is live end in b's block it is
242 * live at b's definition (a dominates b)
244 if (be_is_live_end(lv, bb, a))
248 * Look at all usages of a.
249 * If there's one usage of a in the block of b, then
250 * we check, if this use is dominated by b, if that's true
251 * a and b interfere. Note that b must strictly dominate the user,
252 * since if b is the last user of in the block, b and a do not
254 * Uses of a not in b's block can be disobeyed, because the
255 * check for a being live at the end of b's block is already
258 foreach_out_edge(a, edge) {
259 const ir_node *user = get_edge_src_irn(edge);
261 foreach_out_edge(user, edge2) {
262 const ir_node *user2 = get_edge_src_irn(edge2);
263 assert(!is_Sync(user2));
264 if (get_nodes_block(user2) == bb && !is_Phi(user2) &&
265 _value_strictly_dominates(b, user2))
269 if (get_nodes_block(user) == bb && !is_Phi(user) &&
270 _value_strictly_dominates(b, user))
280 * same as values_interfere but with special handling for Syncs
282 static int my_values_interfere(ir_graph *irg, ir_node *a, ir_node *b)
285 int i, arity = get_irn_arity(a);
286 for (i = 0; i < arity; ++i) {
287 ir_node *in = get_irn_n(a, i);
288 if (my_values_interfere(irg, in, b))
292 } else if (is_Sync(b)) {
293 int i, arity = get_irn_arity(b);
294 for (i = 0; i < arity; ++i) {
295 ir_node *in = get_irn_n(b, i);
296 /* a is not a sync, so no need for my_values_interfere */
297 if (my_values_interfere2(irg, a, in))
303 return my_values_interfere2(irg, a, b);
307 * A greedy coalescing algorithm for spillslots:
308 * 1. Sort the list of affinity edges
309 * 2. Try to merge slots with affinity edges (most expensive slots first)
310 * 3. Try to merge everything else that is possible
312 static void do_greedy_coalescing(be_fec_env_t *env)
314 spill_t **spills = env->spills;
315 size_t spillcount = ARR_LEN(spills);
317 size_t affinity_edge_count;
318 bitset_t **interferences;
319 int* spillslot_unionfind;
327 DB((dbg, DBG_COALESCING, "Coalescing %d spillslots\n", spillcount));
329 interferences = OALLOCN(&data, bitset_t*, spillcount);
330 spillslot_unionfind = OALLOCN(&data, int, spillcount);
332 uf_init(spillslot_unionfind, spillcount);
334 for (i = 0; i < spillcount; ++i) {
335 interferences[i] = bitset_obstack_alloc(&data, spillcount);
338 /* construct interferences */
339 for (i = 0; i < spillcount; ++i) {
341 ir_node *spill1 = spills[i]->spill;
342 if (is_NoMem(spill1))
345 for (i2 = i+1; i2 < spillcount; ++i2) {
346 ir_node *spill2 = spills[i2]->spill;
347 if (is_NoMem(spill2))
350 if (my_values_interfere(env->irg, spill1, spill2)) {
351 DB((dbg, DBG_INTERFERENCES,
352 "Slot %d and %d interfere\n", i, i2));
354 bitset_set(interferences[i], i2);
355 bitset_set(interferences[i2], i);
360 /* sort affinity edges */
361 affinity_edge_count = ARR_LEN(env->affinity_edges);
362 qsort(env->affinity_edges, affinity_edge_count,
363 sizeof(env->affinity_edges[0]), cmp_affinity);
365 /* try to merge affine nodes */
366 for (i = 0; i < affinity_edge_count; ++i) {
367 const affinity_edge_t *edge = env->affinity_edges[i];
368 int s1 = uf_find(spillslot_unionfind, edge->slot1);
369 int s2 = uf_find(spillslot_unionfind, edge->slot2);
371 /* test if values interfere */
372 if (bitset_is_set(interferences[s1], s2)) {
373 assert(bitset_is_set(interferences[s2], s1));
377 DB((dbg, DBG_COALESCING,
378 "Merging %d and %d because of affinity edge\n", s1, s2));
380 merge_interferences(env, interferences, spillslot_unionfind, s1, s2);
383 /* try to merge as much remaining spillslots as possible */
384 for (i = 0; i < spillcount; ++i) {
386 int s1 = uf_find(spillslot_unionfind, i);
390 for (i2 = i+1; i2 < spillcount; ++i2) {
391 int s2 = uf_find(spillslot_unionfind, i2);
395 /* test if values interfere
396 * we have to test n1-n2 and n2-n1, because only 1 side gets updated
397 * when node merging occurs
399 if (bitset_is_set(interferences[s1], s2)) {
400 assert(bitset_is_set(interferences[s2], s1));
404 DB((dbg, DBG_COALESCING,
405 "Merging %d and %d because it is possible\n", s1, s2));
407 if (merge_interferences(env, interferences, spillslot_unionfind, s1, s2) != 0) {
408 /* we can break the loop here, because s2 is the new supernode
409 * now and we'll test s2 again later anyway */
415 /* assign spillslots to spills */
416 for (i = 0; i < spillcount; ++i) {
417 spills[i]->spillslot = uf_find(spillslot_unionfind, i);
420 obstack_free(&data, 0);
423 typedef struct spill_slot_t {
429 typedef struct memperm_entry_t {
434 struct memperm_entry_t *next;
437 typedef struct memperm_t {
440 memperm_entry_t *entries;
443 static int cmp_memperm(const void* d1, const void* d2, size_t size)
445 const memperm_t* e1 = (const memperm_t*)d1;
446 const memperm_t* e2 = (const memperm_t*)d2;
449 return e1->block != e2->block;
452 static memperm_t *get_memperm(be_fec_env_t *env, ir_node *block)
454 memperm_t entry, *res;
458 hash = hash_irn(block);
460 res = set_find(memperm_t, env->memperms, &entry, sizeof(entry), hash);
463 entry.entrycount = 0;
464 entry.entries = NULL;
465 res = set_insert(memperm_t, env->memperms, &entry, sizeof(entry), hash);
471 static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
473 ir_graph *irg = env->irg;
474 ir_type *frame = get_irg_frame_type(irg);
475 ir_entity *res = frame_alloc_area(frame, slot->size, slot->align,
483 * Enlarges a spillslot (if necessary) so that it can carry a value of size
484 * @p othersize and alignment @p otheralign.
486 static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
488 if (othersize > slot->size) {
489 slot->size = othersize;
491 if (otheralign > slot->align) {
492 if (otheralign % slot->align != 0)
493 slot->align *= otheralign;
495 slot->align = otheralign;
496 } else if (slot->align % otheralign != 0) {
497 slot->align *= otheralign;
501 static void assign_spill_entity(be_fec_env_t *env,
502 ir_node *node, ir_entity *entity)
509 arity = get_irn_arity(node);
510 for (i = 0; i < arity; ++i) {
511 ir_node *in = get_irn_n(node, i);
514 assign_spill_entity(env, in, entity);
519 /* beware: we might have Stores with Memory Proj's, ia32 fisttp for
521 node = skip_Proj(node);
522 assert(arch_get_frame_entity(node) == NULL);
523 env->set_frame_entity(node, entity);
527 * Create stack entities for the spillslots and assign them to the spill and
530 static void assign_spillslots(be_fec_env_t *env)
532 spill_t **spills = env->spills;
533 size_t spillcount = ARR_LEN(spills);
534 spill_slot_t *spillslots = ALLOCANZ(spill_slot_t, spillcount);
537 /* construct spillslots */
538 for (s = 0; s < spillcount; ++s) {
539 const spill_t *spill = spills[s];
540 int slotid = spill->spillslot;
541 const ir_mode *mode = spill->mode;
542 spill_slot_t *slot = & (spillslots[slotid]);
543 int size = get_mode_size_bytes(mode);
544 int align = spill->alignment;
546 if (slot->align == 0 && slot->size == 0) {
550 enlarge_spillslot(slot, align, size);
554 for (s = 0; s < spillcount; ++s) {
555 const spill_t *spill = spills[s];
556 ir_node *node = spill->spill;
557 int slotid = spill->spillslot;
558 spill_slot_t *slot = &spillslots[slotid];
560 if (slot->entity == NULL) {
561 create_stack_entity(env, slot);
565 int arity = get_irn_arity(node);
567 ir_node *block = get_nodes_block(node);
569 /* should be a PhiM */
570 assert(get_irn_mode(node) == mode_M);
572 for (i = 0; i < arity; ++i) {
573 ir_node *arg = get_irn_n(node, i);
574 ir_node *predblock = get_Block_cfgpred_block(block, i);
575 spill_t *argspill = get_spill(env, arg);
576 int argslotid = argspill->spillslot;
578 if (slotid != argslotid) {
580 memperm_entry_t *entry;
581 spill_slot_t *argslot = &spillslots[argslotid];
582 if (argslot->entity == NULL) {
583 create_stack_entity(env, argslot);
586 memperm = get_memperm(env, predblock);
588 entry = OALLOC(&env->obst, memperm_entry_t);
591 entry->in = argslot->entity;
592 entry->out = slot->entity;
593 entry->next = memperm->entries;
594 memperm->entrycount++;
595 memperm->entries = entry;
599 assign_spill_entity(env, node, slot->entity);
603 for (s = 0; s < ARR_LEN(env->reloads); ++s) {
604 ir_node *reload = env->reloads[s];
605 ir_node *spillnode = get_memory_edge(reload);
606 const spill_t *spill = get_spill(env, spillnode);
607 const spill_slot_t *slot = &spillslots[spill->spillslot];
609 assert(slot->entity != NULL);
611 env->set_frame_entity(reload, slot->entity);
616 * Returns the last node in a block which is no control flow changing node
618 static ir_node *get_end_of_block_insertion_point(ir_node* block)
620 ir_node* ins = sched_last(block);
621 while (is_Proj(ins) && get_irn_mode(ins) == mode_X) {
622 ins = sched_prev(ins);
628 ir_node *prev = sched_prev(ins);
638 static void create_memperms(be_fec_env_t *env)
640 foreach_set(env->memperms, memperm_t, memperm) {
641 ir_node **nodes = ALLOCAN(ir_node*, memperm->entrycount);
642 memperm_entry_t *entry;
644 ir_node *mempermnode;
647 assert(memperm->entrycount > 0);
649 for (entry = memperm->entries, i = 0; entry != NULL; entry = entry->next, ++i) {
650 ir_node* arg = get_irn_n(entry->node, entry->pos);
654 mempermnode = be_new_MemPerm(memperm->block, memperm->entrycount,
657 /* insert node into schedule */
658 blockend = get_end_of_block_insertion_point(memperm->block);
659 sched_add_before(blockend, mempermnode);
660 stat_ev_dbl("mem_perm", memperm->entrycount);
663 for (entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
665 ir_node* arg = get_irn_n(entry->node, entry->pos);
667 be_set_MemPerm_in_entity(mempermnode, i, entry->in);
668 be_set_MemPerm_out_entity(mempermnode, i, entry->out);
669 proj = new_r_Proj(mempermnode, get_irn_mode(arg), i);
671 set_irn_n(entry->node, entry->pos, proj);
676 static unsigned count_spillslots(const be_fec_env_t *env)
678 size_t spillcount = ARR_LEN(env->spills);
679 unsigned slotcount = 0;
683 rbitset_alloca(counted, spillcount);
685 for (s = 0; s < spillcount; ++s) {
686 spill_t *spill = env->spills[s];
687 int spillslot = spill->spillslot;
688 if (!rbitset_is_set(counted, spillslot)) {
690 rbitset_set(counted, spillslot);
697 be_fec_env_t *be_new_frame_entity_coalescer(ir_graph *irg)
699 be_fec_env_t *env = XMALLOCZ(be_fec_env_t);
701 be_assure_live_chk(irg);
703 obstack_init(&env->obst);
705 env->spills = NEW_ARR_F(spill_t*, 0);
706 env->spills_set = rbitset_malloc(get_irg_last_idx(irg));
707 env->reloads = NEW_ARR_F(ir_node*, 0);
708 env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
709 env->memperms = new_set(cmp_memperm, 10);
711 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
716 void be_free_frame_entity_coalescer(be_fec_env_t *env)
718 ir_free_resources(env->irg, IR_RESOURCE_IRN_LINK);
720 del_set(env->memperms);
721 DEL_ARR_F(env->reloads);
722 DEL_ARR_F(env->affinity_edges);
723 DEL_ARR_F(env->spills);
724 xfree(env->spills_set);
725 obstack_free(&env->obst, NULL);
730 void be_assign_entities(be_fec_env_t *env,
731 set_frame_entity_func set_frame_entity,
732 bool alloc_entities_at_begin)
734 env->set_frame_entity = set_frame_entity;
735 env->at_begin = alloc_entities_at_begin;
737 if (stat_ev_enabled) {
738 stat_ev_dbl("spillslots", ARR_LEN(env->spills));
741 if (be_coalesce_spill_slots) {
742 do_greedy_coalescing(env);
745 if (stat_ev_enabled) {
746 stat_ev_dbl("spillslots_after_coalescing", count_spillslots(env));
749 assign_spillslots(env);
751 create_memperms(env);
754 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots)
755 void be_init_spillslots(void)
757 FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");