2 * Copyright (C) 1995-2007 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
39 #include "unionfind.h"
44 #include "bespillslots.h"
45 #include "bechordal_t.h"
46 #include "bejavacoal.h"
47 #include "benodesets.h"
48 #include "bestatevent.h"
49 #include "bespilloptions.h"
51 #include "beintlive_t.h"
55 #define DBG_COALESCING 1
56 #define DBG_INTERFERENCES 2
59 static firm_dbg_module_t *dbg = NULL;
62 typedef struct _spill_t {
64 /** mode of the spilled value */
66 /** alignment for the spilled value */
68 /** index into spillslot_unionfind unionfind structure */
72 typedef struct _affinity_edge_t {
77 struct _be_fec_env_t {
79 const arch_env_t *arch_env;
83 affinity_edge_t **affinity_edges;
87 /** Compare 2 affinity edges (used in quicksort) */
88 static int cmp_affinity(const void *d1, const void *d2)
90 const affinity_edge_t * const *e1 = d1;
91 const affinity_edge_t * const *e2 = d2;
93 // sort in descending order
94 return (*e1)->affinity < (*e2)->affinity ? 1 : -1;
97 static int cmp_spill(const void* d1, const void* d2, size_t size)
99 const spill_t* s1 = d1;
100 const spill_t* s2 = d2;
103 return s1->spill != s2->spill;
106 static spill_t *get_spill(be_fec_env_t *env, ir_node *node)
109 int hash = nodeset_hash(node);
112 res = set_find(env->spills, &spill, sizeof(spill), hash);
118 * ____ _ _ _ ____ _ _ _
119 * / ___|___ | | | ___ ___| |_ / ___| _ __ (_) | |___
120 * | | / _ \| | |/ _ \/ __| __| \___ \| '_ \| | | / __|
121 * | |__| (_) | | | __/ (__| |_ ___) | |_) | | | \__ \
122 * \____\___/|_|_|\___|\___|\__| |____/| .__/|_|_|_|___/
126 static INLINE ir_node *get_memory_edge(const ir_node *node)
130 arity = get_irn_arity(node);
131 for(i = arity - 1; i >= 0; --i) {
132 ir_node *arg = get_irn_n(node, i);
133 if(get_irn_mode(arg) == mode_M)
140 static spill_t *collect_spill(be_fec_env_t *env, ir_node *node,
141 const ir_mode *mode, int align)
144 int hash = nodeset_hash(node);
146 /* insert into set of spills if not already there */
148 res = set_find(env->spills, &spill, sizeof(spill), hash);
151 spill.spillslot = set_count(env->spills);
153 spill.alignment = align;
154 res = set_insert(env->spills, &spill, sizeof(spill), hash);
156 assert(res->mode == mode);
157 assert(res->alignment == align);
163 static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
164 const ir_mode *mode, int align)
168 int hash = nodeset_hash(node);
169 const ir_exec_freq *exec_freq = be_get_birg_exec_freq(env->birg);
171 assert(is_Phi(node));
174 res = set_find(env->spills, &spill, sizeof(spill), hash);
176 assert(res->mode == mode);
177 assert(res->alignment == align);
181 spill.spillslot = set_count(env->spills);
183 spill.alignment = align;
184 res = set_insert(env->spills, &spill, sizeof(spill), hash);
186 // collect attached spills and mem-phis
187 arity = get_irn_arity(node);
188 for(i = 0; i < arity; ++i) {
189 affinity_edge_t *affinty_edge;
190 ir_node *arg = get_irn_n(node, i);
194 arg_spill = collect_memphi(env, arg, mode, align);
196 arg_spill = collect_spill(env, arg, mode, align);
199 // add an affinity edge
200 affinty_edge = obstack_alloc(&env->obst, sizeof(affinty_edge[0]));
201 affinty_edge->affinity = get_block_execfreq(exec_freq, get_nodes_block(arg));
202 affinty_edge->slot1 = res->spillslot;
203 affinty_edge->slot2 = arg_spill->spillslot;
204 ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
210 void be_node_needs_frame_entity(be_fec_env_t *env, ir_node *node,
211 const ir_mode *mode, int align)
213 ir_node *spillnode = get_memory_edge(node);
216 assert(spillnode != NULL);
218 if (is_Phi(spillnode)) {
219 spill = collect_memphi(env, spillnode, mode, align);
221 spill = collect_spill(env, spillnode, mode, align);
224 ARR_APP1(ir_node *, env->reloads, node);
229 * / ___|___ __ _| | ___ ___ ___ ___ / ___|| | ___ | |_ ___
230 * | | / _ \ / _` | |/ _ \/ __|/ __/ _ \ \___ \| |/ _ \| __/ __|
231 * | |__| (_) | (_| | | __/\__ \ (_| __/ ___) | | (_) | |_\__ \
232 * \____\___/ \__,_|_|\___||___/\___\___| |____/|_|\___/ \__|___/
235 static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
236 int* spillslot_unionfind, int s1, int s2)
242 // merge spillslots and interferences
243 res = uf_union(spillslot_unionfind, s1, s2);
244 // we assume that we always merge s2 to s1 so swap s1, s2 if necessary
251 bitset_or(interferences[s1], interferences[s2]);
253 // update other interferences
254 spillcount = set_count(env->spills);
255 for(i = 0; i < spillcount; ++i) {
256 bitset_t *intfs = interferences[i];
257 if(bitset_is_set(intfs, s2))
258 bitset_set(intfs, s1);
264 static int my_values_interfere2(be_irg_t *birg, const ir_node *a,
267 be_lv_t *lv = be_get_birg_liveness(birg);
269 int a2b = _value_dominates(a, b);
270 int b2a = _value_dominates(b, a);
272 /* If there is no dominance relation, they do not interfere. */
273 if((a2b | b2a) > 0) {
274 const ir_edge_t *edge;
278 * Adjust a and b so, that a dominates b if
279 * a dominates b or vice versa.
282 const ir_node *t = a;
287 bb = get_nodes_block(b);
290 * If a is live end in b's block it is
291 * live at b's definition (a dominates b)
293 if(be_is_live_end(lv, bb, a))
297 * Look at all usages of a.
298 * If there's one usage of a in the block of b, then
299 * we check, if this use is dominated by b, if that's true
300 * a and b interfere. Note that b must strictly dominate the user,
301 * since if b is the last user of in the block, b and a do not
303 * Uses of a not in b's block can be disobeyed, because the
304 * check for a being live at the end of b's block is already
307 foreach_out_edge(a, edge) {
308 const ir_node *user = get_edge_src_irn(edge);
310 const ir_edge_t *edge2;
311 foreach_out_edge(user, edge2) {
312 const ir_node *user2 = get_edge_src_irn(edge2);
313 assert(!is_Sync(user2));
314 if(get_nodes_block(user2) == bb && !is_Phi(user2) &&
315 _value_strictly_dominates(b, user2))
319 if(get_nodes_block(user) == bb && !is_Phi(user) &&
320 _value_strictly_dominates(b, user))
330 * same as values_interfere but with special handling for Syncs
332 static int my_values_interfere(be_irg_t *birg, ir_node *a, ir_node *b)
335 int i, arity = get_irn_arity(a);
336 for(i = 0; i < arity; ++i) {
337 ir_node *in = get_irn_n(a, i);
338 if(my_values_interfere(birg, in, b))
342 } else if(is_Sync(b)) {
343 int i, arity = get_irn_arity(b);
344 for(i = 0; i < arity; ++i) {
345 ir_node *in = get_irn_n(b, i);
346 /* a is not a sync, so no need for my_values_interfere */
347 if(my_values_interfere2(birg, a, in))
353 return my_values_interfere2(birg, a, b);
357 * A greedy coalescing algorithm for spillslots:
358 * 1. Sort the list of affinity edges
359 * 2. Try to merge slots with affinity edges (most expensive slots first)
360 * 3. Try to merge everything else that is possible
362 static void do_greedy_coalescing(be_fec_env_t *env)
368 int affinity_edge_count;
369 bitset_t **interferences;
370 int* spillslot_unionfind;
372 spillcount = set_count(env->spills);
376 DBG((dbg, DBG_COALESCING, "Coalescing %d spillslots\n", spillcount));
378 interferences = alloca(spillcount * sizeof(interferences[0]));
379 spillslot_unionfind = alloca(spillcount * sizeof(spillslot_unionfind[0]));
380 spilllist = alloca(spillcount * sizeof(spilllist[0]));
382 uf_init(spillslot_unionfind, 0, spillcount);
385 memset(spilllist, 0, spillcount * sizeof(spilllist[0]));
388 for(spill = set_first(env->spills), i = 0; spill != NULL; spill = set_next(env->spills), ++i) {
389 assert(spill->spillslot < spillcount);
390 spilllist[spill->spillslot] = spill;
393 for(i = 0; i < spillcount; ++i) {
394 interferences[i] = bitset_alloca(spillcount);
397 /* construct interferences */
398 for (i = 0; i < spillcount; ++i) {
399 ir_node *spill1 = spilllist[i]->spill;
401 if (is_NoMem(spill1))
404 for(i2 = i+1; i2 < spillcount; ++i2) {
405 ir_node *spill2 = spilllist[i2]->spill;
407 if (is_NoMem(spill2))
410 if (my_values_interfere(env->birg, spill1, spill2)) {
411 DBG((dbg, DBG_INTERFERENCES, "Slot %d and %d interfere\n", i, i2));
412 bitset_set(interferences[i], i2);
413 bitset_set(interferences[i2], i);
418 /* sort affinity edges */
419 affinity_edge_count = ARR_LEN(env->affinity_edges);
420 qsort(env->affinity_edges, affinity_edge_count, sizeof(env->affinity_edges[0]), cmp_affinity);
422 //dump_interference_graph(env, interferences, "before");
424 /* try to merge affine nodes */
425 for(i = 0; i < affinity_edge_count; ++i) {
426 const affinity_edge_t *edge = env->affinity_edges[i];
427 int s1 = uf_find(spillslot_unionfind, edge->slot1);
428 int s2 = uf_find(spillslot_unionfind, edge->slot2);
430 /* test if values interfere */
431 if (bitset_is_set(interferences[s1], s2)) {
432 assert(bitset_is_set(interferences[s2], s1));
436 DBG((dbg, DBG_COALESCING, "Merging %d and %d because of affinity edge\n", s1, s2));
438 merge_interferences(env, interferences, spillslot_unionfind, s1, s2);
441 // try to merge as much remaining spillslots as possible
442 for(i = 0; i < spillcount; ++i) {
443 int s1 = uf_find(spillslot_unionfind, i);
447 for(i2 = i+1; i2 < spillcount; ++i2) {
448 int s2 = uf_find(spillslot_unionfind, i2);
452 /* test if values interfere
453 * we have to test n1-n2 and n2-n1, because only 1 side gets updated
454 * when node merging occurs
456 if(bitset_is_set(interferences[s1], s2)) {
457 assert(bitset_is_set(interferences[s2], s1));
461 DBG((dbg, DBG_COALESCING, "Merging %d and %d because it is possible\n", s1, s2));
463 if(merge_interferences(env, interferences, spillslot_unionfind, s1, s2) != 0) {
464 // we can break the loop here, because s2 is the new supernode now
465 // and we'll test s2 again later anyway
471 // assign spillslots to spills
472 for(i = 0; i < spillcount; ++i) {
473 spill_t *spill = spilllist[i];
475 spill->spillslot = uf_find(spillslot_unionfind, i);
478 //dump_interference_graph(env, interferences, "after");
483 * / \ ___ ___(_) __ _ _ __ | ____|_ __ | |_(_) |_(_) ___ ___
484 * / _ \ / __/ __| |/ _` | '_ \ | _| | '_ \| __| | __| |/ _ \/ __|
485 * / ___ \\__ \__ \ | (_| | | | | | |___| | | | |_| | |_| | __/\__ \
486 * /_/ \_\___/___/_|\__, |_| |_| |_____|_| |_|\__|_|\__|_|\___||___/
490 typedef struct _spill_slot_t {
496 typedef struct _memperm_entry_t {
501 struct _memperm_entry_t *next;
504 typedef struct _memperm_t {
507 memperm_entry_t *entries;
510 static int cmp_memperm(const void* d1, const void* d2, size_t size)
512 const memperm_t* e1 = d1;
513 const memperm_t* e2 = d2;
516 return e1->block != e2->block;
519 static memperm_t *get_memperm(be_fec_env_t *env, ir_node *block)
521 memperm_t entry, *res;
525 hash = nodeset_hash(block);
527 res = set_find(env->memperms, &entry, sizeof(entry), hash);
530 entry.entrycount = 0;
531 entry.entries = NULL;
532 res = set_insert(env->memperms, &entry, sizeof(entry), hash);
538 static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
540 ir_graph *irg = be_get_birg_irg(env->birg);
541 ir_type *frame = get_irg_frame_type(irg);
542 ir_entity *res = frame_alloc_area(frame, slot->size, slot->align, 0);
544 /* adjust size of the entity type... */
545 ir_type *enttype = get_entity_type(res);
546 set_type_size_bytes(enttype, slot->size);
554 * Enlarges a spillslot (if necessary) so that it can carry a value of size
555 * @p othersize and alignment @p otheralign.
557 static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
559 if(othersize > slot->size) {
560 slot->size = othersize;
562 if(otheralign > slot->align) {
563 if(otheralign % slot->align != 0)
564 slot->align *= otheralign;
566 slot->align = otheralign;
567 } else if(slot->align % otheralign != 0) {
568 slot->align *= otheralign;
573 static void assign_spill_entity(const arch_env_t *arch_env, ir_node *node, ir_entity *entity)
580 arity = get_irn_arity(node);
581 for(i = 0; i < arity; ++i) {
582 ir_node *in = get_irn_n(node, i);
585 assign_spill_entity(arch_env, in, entity);
590 assert(arch_get_frame_entity(arch_env, node) == NULL);
591 arch_set_frame_entity(arch_env, node, entity);
595 * Create stack entities for the spillslots and assign them to the spill and
598 static void assign_spillslots(be_fec_env_t *env)
600 const arch_env_t *arch_env = env->arch_env;
604 spill_slot_t* spillslots;
606 spillcount = set_count(env->spills);
607 spillslots = alloca(spillcount * sizeof(spillslots[0]));
609 memset(spillslots, 0, spillcount * sizeof(spillslots[0]));
611 // construct spillslots
612 for(spill = set_first(env->spills); spill != NULL; spill = set_next(env->spills)) {
613 int slotid = spill->spillslot;
614 const ir_mode *mode = spill->mode;
615 spill_slot_t *slot = & (spillslots[slotid]);
616 int size = get_mode_size_bytes(mode);
617 int align = spill->alignment;
619 if(slot->align == 0 && slot->size == 0) {
623 enlarge_spillslot(slot, align, size);
627 for(spill = set_first(env->spills); spill != NULL; spill = set_next(env->spills)) {
629 ir_node *node = spill->spill;
630 int slotid = spill->spillslot;
632 slot = &spillslots[slotid];
633 if(slot->entity == NULL) {
634 create_stack_entity(env, slot);
639 ir_node *block = get_nodes_block(node);
642 assert(is_Phi(node));
644 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
645 ir_node *arg = get_irn_n(node, i);
646 ir_node *predblock = get_Block_cfgpred_block(block, i);
650 argspill = get_spill(env, arg);
651 assert(argspill != NULL);
653 argslotid = argspill->spillslot;
654 if(slotid != argslotid) {
656 memperm_entry_t *entry;
657 spill_slot_t *argslot = &spillslots[argslotid];
658 if(argslot->entity == NULL) {
659 create_stack_entity(env, argslot);
662 memperm = get_memperm(env, predblock);
664 entry = obstack_alloc(&env->obst, sizeof(entry[0]));
667 entry->in = argslot->entity;
668 entry->out = slot->entity;
669 entry->next = memperm->entries;
670 memperm->entrycount++;
671 memperm->entries = entry;
675 assign_spill_entity(arch_env, node, slot->entity);
679 for(i = 0; i < ARR_LEN(env->reloads); ++i) {
680 ir_node* reload = env->reloads[i];
681 ir_node* spillnode = get_memory_edge(reload);
682 spill_t *spill = get_spill(env, spillnode);
683 const spill_slot_t *slot = & spillslots[spill->spillslot];
685 assert(slot->entity != NULL);
687 arch_set_frame_entity(arch_env, reload, slot->entity);
692 * Returns the last node in a block which is no control flow changing node
694 static ir_node *get_end_of_block_insertion_point(ir_node* block)
696 ir_node* ins = sched_last(block);
697 while(is_Proj(ins) && get_irn_mode(ins) == mode_X) {
698 ins = sched_prev(ins);
704 ir_node *prev = sched_prev(ins);
714 static void create_memperms(be_fec_env_t *env)
716 const arch_env_t *arch_env = env->arch_env;
717 ir_graph *irg = be_get_birg_irg(env->birg);
720 for(memperm = set_first(env->memperms); memperm != NULL; memperm = set_next(env->memperms)) {
722 memperm_entry_t *entry;
724 ir_node** nodes = alloca(memperm->entrycount * sizeof(nodes[0]));
725 ir_node* mempermnode;
727 assert(memperm->entrycount > 0);
729 for(entry = memperm->entries, i = 0; entry != NULL; entry = entry->next, ++i) {
730 ir_node* arg = get_irn_n(entry->node, entry->pos);
734 mempermnode = be_new_MemPerm(arch_env, irg, memperm->block,
735 memperm->entrycount, nodes);
737 // insert node into schedule
738 blockend = get_end_of_block_insertion_point(memperm->block);
739 sched_add_before(blockend, mempermnode);
740 stat_ev_dbl("mem_perm", memperm->entrycount);
743 for(entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
745 ir_node* arg = get_irn_n(entry->node, entry->pos);
747 be_set_MemPerm_in_entity(mempermnode, i, entry->in);
748 be_set_MemPerm_out_entity(mempermnode, i, entry->out);
749 set_irg_current_block(irg, memperm->block);
750 proj = new_Proj(mempermnode, get_irn_mode(arg), i);
752 set_irn_n(entry->node, entry->pos, proj);
757 static int count_spillslots(const be_fec_env_t *env)
759 const spill_t *spill;
760 int spillcount = set_count(env->spills);
761 bitset_t *counted = bitset_alloca(spillcount);
765 for(spill = set_first(env->spills); spill != NULL;
766 spill = set_next(env->spills)) {
767 int spillslot = spill->spillslot;
768 if(!bitset_is_set(counted, spillslot)) {
770 bitset_set(counted, spillslot);
777 be_fec_env_t *be_new_frame_entity_coalescer(be_irg_t *birg)
779 const arch_env_t *arch_env = birg->main_env->arch_env;
780 be_fec_env_t *env = xmalloc(sizeof(env[0]));
782 be_liveness_assure_chk(be_assure_liveness(birg));
784 obstack_init(&env->obst);
785 env->arch_env = arch_env;
787 env->spills = new_set(cmp_spill, 10);
788 env->reloads = NEW_ARR_F(ir_node*, 0);
789 env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
790 env->memperms = new_set(cmp_memperm, 10);
795 void be_free_frame_entity_coalescer(be_fec_env_t *env)
797 del_set(env->memperms);
798 DEL_ARR_F(env->reloads);
799 DEL_ARR_F(env->affinity_edges);
800 del_set(env->spills);
801 obstack_free(&env->obst, NULL);
806 void be_assign_entities(be_fec_env_t *env)
808 stat_ev_dbl("spillslots", set_count(env->spills));
810 if(be_coalesce_spill_slots) {
811 do_greedy_coalescing(env);
814 stat_ev_dbl("spillslots_after_coalescing", count_spillslots(env));
816 assign_spillslots(env);
818 create_memperms(env);
822 * This walker function searches for reloads and collects all the spills
823 * and memphis attached to them.
825 static void collect_spills_walker(ir_node *node, void *data)
827 be_fec_env_t *env = data;
828 const arch_env_t *arch_env = env->arch_env;
830 const arch_register_class_t *cls;
833 /* classify returns classification of the irn the proj is attached to */
837 if (!arch_irn_class_is(arch_env, node, reload))
840 mode = get_irn_mode(node);
841 cls = arch_get_irn_reg_class(arch_env, node, -1);
842 align = arch_isa_get_reg_class_alignment(arch_env_get_isa(arch_env), cls);
844 be_node_needs_frame_entity(env, node, mode, align);
847 void be_coalesce_spillslots(be_irg_t *birg)
849 be_fec_env_t *env = be_new_frame_entity_coalescer(birg);
851 /* collect reloads */
852 irg_walk_graph(birg->irg, NULL, collect_spills_walker, env);
854 be_assign_entities(env);
856 be_free_frame_entity_coalescer(env);
859 void be_init_spillslots(void)
861 FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");
864 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots);