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 "bestatevent.h"
48 #include "bespilloptions.h"
50 #include "beintlive_t.h"
54 #define DBG_COALESCING 1
55 #define DBG_INTERFERENCES 2
58 static firm_dbg_module_t *dbg = NULL;
61 typedef struct _spill_t {
63 /** mode of the spilled value */
65 /** alignment for the spilled value */
67 /** index into spillslot_unionfind unionfind structure */
71 typedef struct _affinity_edge_t {
76 struct _be_fec_env_t {
78 const arch_env_t *arch_env;
82 affinity_edge_t **affinity_edges;
86 /** Compare 2 affinity edges (used in quicksort) */
87 static int cmp_affinity(const void *d1, const void *d2)
89 const affinity_edge_t * const *e1 = d1;
90 const affinity_edge_t * const *e2 = d2;
92 // sort in descending order
93 return (*e1)->affinity < (*e2)->affinity ? 1 : -1;
96 static int cmp_spill(const void* d1, const void* d2, size_t size)
98 const spill_t* s1 = d1;
99 const spill_t* s2 = d2;
102 return s1->spill != s2->spill;
105 static spill_t *get_spill(be_fec_env_t *env, ir_node *node)
108 int hash = hash_irn(node);
111 res = set_find(env->spills, &spill, sizeof(spill), hash);
117 * ____ _ _ _ ____ _ _ _
118 * / ___|___ | | | ___ ___| |_ / ___| _ __ (_) | |___
119 * | | / _ \| | |/ _ \/ __| __| \___ \| '_ \| | | / __|
120 * | |__| (_) | | | __/ (__| |_ ___) | |_) | | | \__ \
121 * \____\___/|_|_|\___|\___|\__| |____/| .__/|_|_|_|___/
125 static INLINE ir_node *get_memory_edge(const ir_node *node)
129 arity = get_irn_arity(node);
130 for(i = arity - 1; i >= 0; --i) {
131 ir_node *arg = get_irn_n(node, i);
132 if(get_irn_mode(arg) == mode_M)
139 static spill_t *collect_spill(be_fec_env_t *env, ir_node *node,
140 const ir_mode *mode, int align)
143 int hash = hash_irn(node);
145 /* insert into set of spills if not already there */
147 res = set_find(env->spills, &spill, sizeof(spill), hash);
150 spill.spillslot = set_count(env->spills);
152 spill.alignment = align;
153 res = set_insert(env->spills, &spill, sizeof(spill), hash);
155 assert(res->mode == mode);
156 assert(res->alignment == align);
162 static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
163 const ir_mode *mode, int align)
167 int hash = hash_irn(node);
168 const ir_exec_freq *exec_freq = be_get_birg_exec_freq(env->birg);
170 assert(is_Phi(node));
173 res = set_find(env->spills, &spill, sizeof(spill), hash);
175 assert(res->mode == mode);
176 assert(res->alignment == align);
180 spill.spillslot = set_count(env->spills);
182 spill.alignment = align;
183 res = set_insert(env->spills, &spill, sizeof(spill), hash);
185 // collect attached spills and mem-phis
186 arity = get_irn_arity(node);
187 for(i = 0; i < arity; ++i) {
188 affinity_edge_t *affinty_edge;
189 ir_node *arg = get_irn_n(node, i);
193 arg_spill = collect_memphi(env, arg, mode, align);
195 arg_spill = collect_spill(env, arg, mode, align);
198 // add an affinity edge
199 affinty_edge = obstack_alloc(&env->obst, sizeof(affinty_edge[0]));
200 affinty_edge->affinity = get_block_execfreq(exec_freq, get_nodes_block(arg));
201 affinty_edge->slot1 = res->spillslot;
202 affinty_edge->slot2 = arg_spill->spillslot;
203 ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
209 void be_node_needs_frame_entity(be_fec_env_t *env, ir_node *node,
210 const ir_mode *mode, int align)
212 ir_node *spillnode = get_memory_edge(node);
215 assert(spillnode != NULL);
217 if (is_Phi(spillnode)) {
218 spill = collect_memphi(env, spillnode, mode, align);
220 spill = collect_spill(env, spillnode, mode, align);
223 ARR_APP1(ir_node *, env->reloads, node);
228 * / ___|___ __ _| | ___ ___ ___ ___ / ___|| | ___ | |_ ___
229 * | | / _ \ / _` | |/ _ \/ __|/ __/ _ \ \___ \| |/ _ \| __/ __|
230 * | |__| (_) | (_| | | __/\__ \ (_| __/ ___) | | (_) | |_\__ \
231 * \____\___/ \__,_|_|\___||___/\___\___| |____/|_|\___/ \__|___/
234 static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
235 int* spillslot_unionfind, int s1, int s2)
241 // merge spillslots and interferences
242 res = uf_union(spillslot_unionfind, s1, s2);
243 // we assume that we always merge s2 to s1 so swap s1, s2 if necessary
250 bitset_or(interferences[s1], interferences[s2]);
252 // update other interferences
253 spillcount = set_count(env->spills);
254 for(i = 0; i < spillcount; ++i) {
255 bitset_t *intfs = interferences[i];
256 if(bitset_is_set(intfs, s2))
257 bitset_set(intfs, s1);
263 static int my_values_interfere2(be_irg_t *birg, const ir_node *a,
266 be_lv_t *lv = be_get_birg_liveness(birg);
268 int a2b = _value_dominates(a, b);
269 int b2a = _value_dominates(b, a);
271 /* If there is no dominance relation, they do not interfere. */
272 if((a2b | b2a) > 0) {
273 const ir_edge_t *edge;
277 * Adjust a and b so, that a dominates b if
278 * a dominates b or vice versa.
281 const ir_node *t = a;
286 bb = get_nodes_block(b);
289 * If a is live end in b's block it is
290 * live at b's definition (a dominates b)
292 if(be_is_live_end(lv, bb, a))
296 * Look at all usages of a.
297 * If there's one usage of a in the block of b, then
298 * we check, if this use is dominated by b, if that's true
299 * a and b interfere. Note that b must strictly dominate the user,
300 * since if b is the last user of in the block, b and a do not
302 * Uses of a not in b's block can be disobeyed, because the
303 * check for a being live at the end of b's block is already
306 foreach_out_edge(a, edge) {
307 const ir_node *user = get_edge_src_irn(edge);
309 const ir_edge_t *edge2;
310 foreach_out_edge(user, edge2) {
311 const ir_node *user2 = get_edge_src_irn(edge2);
312 assert(!is_Sync(user2));
313 if(get_nodes_block(user2) == bb && !is_Phi(user2) &&
314 _value_strictly_dominates(b, user2))
318 if(get_nodes_block(user) == bb && !is_Phi(user) &&
319 _value_strictly_dominates(b, user))
329 * same as values_interfere but with special handling for Syncs
331 static int my_values_interfere(be_irg_t *birg, ir_node *a, ir_node *b)
334 int i, arity = get_irn_arity(a);
335 for(i = 0; i < arity; ++i) {
336 ir_node *in = get_irn_n(a, i);
337 if(my_values_interfere(birg, in, b))
341 } else if(is_Sync(b)) {
342 int i, arity = get_irn_arity(b);
343 for(i = 0; i < arity; ++i) {
344 ir_node *in = get_irn_n(b, i);
345 /* a is not a sync, so no need for my_values_interfere */
346 if(my_values_interfere2(birg, a, in))
352 return my_values_interfere2(birg, a, b);
356 * A greedy coalescing algorithm for spillslots:
357 * 1. Sort the list of affinity edges
358 * 2. Try to merge slots with affinity edges (most expensive slots first)
359 * 3. Try to merge everything else that is possible
361 static void do_greedy_coalescing(be_fec_env_t *env)
367 int affinity_edge_count;
368 bitset_t **interferences;
369 int* spillslot_unionfind;
371 spillcount = set_count(env->spills);
375 DBG((dbg, DBG_COALESCING, "Coalescing %d spillslots\n", spillcount));
377 interferences = alloca(spillcount * sizeof(interferences[0]));
378 spillslot_unionfind = alloca(spillcount * sizeof(spillslot_unionfind[0]));
379 spilllist = alloca(spillcount * sizeof(spilllist[0]));
381 uf_init(spillslot_unionfind, 0, spillcount);
384 memset(spilllist, 0, spillcount * sizeof(spilllist[0]));
387 for(spill = set_first(env->spills), i = 0; spill != NULL; spill = set_next(env->spills), ++i) {
388 assert(spill->spillslot < spillcount);
389 spilllist[spill->spillslot] = spill;
392 for(i = 0; i < spillcount; ++i) {
393 interferences[i] = bitset_alloca(spillcount);
396 /* construct interferences */
397 for (i = 0; i < spillcount; ++i) {
398 ir_node *spill1 = spilllist[i]->spill;
400 if (is_NoMem(spill1))
403 for(i2 = i+1; i2 < spillcount; ++i2) {
404 ir_node *spill2 = spilllist[i2]->spill;
406 if (is_NoMem(spill2))
409 if (my_values_interfere(env->birg, spill1, spill2)) {
410 DBG((dbg, DBG_INTERFERENCES, "Slot %d and %d interfere\n", i, i2));
411 bitset_set(interferences[i], i2);
412 bitset_set(interferences[i2], i);
417 /* sort affinity edges */
418 affinity_edge_count = ARR_LEN(env->affinity_edges);
419 qsort(env->affinity_edges, affinity_edge_count, sizeof(env->affinity_edges[0]), cmp_affinity);
421 //dump_interference_graph(env, interferences, "before");
423 /* try to merge affine nodes */
424 for(i = 0; i < affinity_edge_count; ++i) {
425 const affinity_edge_t *edge = env->affinity_edges[i];
426 int s1 = uf_find(spillslot_unionfind, edge->slot1);
427 int s2 = uf_find(spillslot_unionfind, edge->slot2);
429 /* test if values interfere */
430 if (bitset_is_set(interferences[s1], s2)) {
431 assert(bitset_is_set(interferences[s2], s1));
435 DBG((dbg, DBG_COALESCING, "Merging %d and %d because of affinity edge\n", s1, s2));
437 merge_interferences(env, interferences, spillslot_unionfind, s1, s2);
440 // try to merge as much remaining spillslots as possible
441 for(i = 0; i < spillcount; ++i) {
442 int s1 = uf_find(spillslot_unionfind, i);
446 for(i2 = i+1; i2 < spillcount; ++i2) {
447 int s2 = uf_find(spillslot_unionfind, i2);
451 /* test if values interfere
452 * we have to test n1-n2 and n2-n1, because only 1 side gets updated
453 * when node merging occurs
455 if(bitset_is_set(interferences[s1], s2)) {
456 assert(bitset_is_set(interferences[s2], s1));
460 DBG((dbg, DBG_COALESCING, "Merging %d and %d because it is possible\n", s1, s2));
462 if(merge_interferences(env, interferences, spillslot_unionfind, s1, s2) != 0) {
463 // we can break the loop here, because s2 is the new supernode now
464 // and we'll test s2 again later anyway
470 // assign spillslots to spills
471 for(i = 0; i < spillcount; ++i) {
472 spill_t *spill = spilllist[i];
474 spill->spillslot = uf_find(spillslot_unionfind, i);
477 //dump_interference_graph(env, interferences, "after");
482 * / \ ___ ___(_) __ _ _ __ | ____|_ __ | |_(_) |_(_) ___ ___
483 * / _ \ / __/ __| |/ _` | '_ \ | _| | '_ \| __| | __| |/ _ \/ __|
484 * / ___ \\__ \__ \ | (_| | | | | | |___| | | | |_| | |_| | __/\__ \
485 * /_/ \_\___/___/_|\__, |_| |_| |_____|_| |_|\__|_|\__|_|\___||___/
489 typedef struct _spill_slot_t {
495 typedef struct _memperm_entry_t {
500 struct _memperm_entry_t *next;
503 typedef struct _memperm_t {
506 memperm_entry_t *entries;
509 static int cmp_memperm(const void* d1, const void* d2, size_t size)
511 const memperm_t* e1 = d1;
512 const memperm_t* e2 = d2;
515 return e1->block != e2->block;
518 static memperm_t *get_memperm(be_fec_env_t *env, ir_node *block)
520 memperm_t entry, *res;
524 hash = hash_irn(block);
526 res = set_find(env->memperms, &entry, sizeof(entry), hash);
529 entry.entrycount = 0;
530 entry.entries = NULL;
531 res = set_insert(env->memperms, &entry, sizeof(entry), hash);
537 static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
539 ir_graph *irg = be_get_birg_irg(env->birg);
540 ir_type *frame = get_irg_frame_type(irg);
541 ir_entity *res = frame_alloc_area(frame, slot->size, slot->align, 0);
543 /* adjust size of the entity type... */
544 ir_type *enttype = get_entity_type(res);
545 set_type_size_bytes(enttype, slot->size);
553 * Enlarges a spillslot (if necessary) so that it can carry a value of size
554 * @p othersize and alignment @p otheralign.
556 static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
558 if(othersize > slot->size) {
559 slot->size = othersize;
561 if(otheralign > slot->align) {
562 if(otheralign % slot->align != 0)
563 slot->align *= otheralign;
565 slot->align = otheralign;
566 } else if(slot->align % otheralign != 0) {
567 slot->align *= otheralign;
572 static void assign_spill_entity(const arch_env_t *arch_env, ir_node *node, ir_entity *entity)
579 arity = get_irn_arity(node);
580 for(i = 0; i < arity; ++i) {
581 ir_node *in = get_irn_n(node, i);
584 assign_spill_entity(arch_env, in, entity);
589 assert(arch_get_frame_entity(arch_env, node) == NULL);
590 arch_set_frame_entity(arch_env, node, entity);
594 * Create stack entities for the spillslots and assign them to the spill and
597 static void assign_spillslots(be_fec_env_t *env)
599 const arch_env_t *arch_env = env->arch_env;
603 spill_slot_t* spillslots;
605 spillcount = set_count(env->spills);
606 spillslots = alloca(spillcount * sizeof(spillslots[0]));
608 memset(spillslots, 0, spillcount * sizeof(spillslots[0]));
610 // construct spillslots
611 for(spill = set_first(env->spills); spill != NULL; spill = set_next(env->spills)) {
612 int slotid = spill->spillslot;
613 const ir_mode *mode = spill->mode;
614 spill_slot_t *slot = & (spillslots[slotid]);
615 int size = get_mode_size_bytes(mode);
616 int align = spill->alignment;
618 if(slot->align == 0 && slot->size == 0) {
622 enlarge_spillslot(slot, align, size);
626 for(spill = set_first(env->spills); spill != NULL; spill = set_next(env->spills)) {
628 ir_node *node = spill->spill;
629 int slotid = spill->spillslot;
631 slot = &spillslots[slotid];
632 if(slot->entity == NULL) {
633 create_stack_entity(env, slot);
638 ir_node *block = get_nodes_block(node);
641 assert(is_Phi(node));
643 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
644 ir_node *arg = get_irn_n(node, i);
645 ir_node *predblock = get_Block_cfgpred_block(block, i);
649 argspill = get_spill(env, arg);
650 assert(argspill != NULL);
652 argslotid = argspill->spillslot;
653 if(slotid != argslotid) {
655 memperm_entry_t *entry;
656 spill_slot_t *argslot = &spillslots[argslotid];
657 if(argslot->entity == NULL) {
658 create_stack_entity(env, argslot);
661 memperm = get_memperm(env, predblock);
663 entry = obstack_alloc(&env->obst, sizeof(entry[0]));
666 entry->in = argslot->entity;
667 entry->out = slot->entity;
668 entry->next = memperm->entries;
669 memperm->entrycount++;
670 memperm->entries = entry;
674 assign_spill_entity(arch_env, node, slot->entity);
678 for(i = 0; i < ARR_LEN(env->reloads); ++i) {
679 ir_node* reload = env->reloads[i];
680 ir_node* spillnode = get_memory_edge(reload);
681 spill_t *spill = get_spill(env, spillnode);
682 const spill_slot_t *slot = & spillslots[spill->spillslot];
684 assert(slot->entity != NULL);
686 arch_set_frame_entity(arch_env, reload, slot->entity);
691 * Returns the last node in a block which is no control flow changing node
693 static ir_node *get_end_of_block_insertion_point(ir_node* block)
695 ir_node* ins = sched_last(block);
696 while(is_Proj(ins) && get_irn_mode(ins) == mode_X) {
697 ins = sched_prev(ins);
703 ir_node *prev = sched_prev(ins);
713 static void create_memperms(be_fec_env_t *env)
715 const arch_env_t *arch_env = env->arch_env;
716 ir_graph *irg = be_get_birg_irg(env->birg);
719 for(memperm = set_first(env->memperms); memperm != NULL; memperm = set_next(env->memperms)) {
721 memperm_entry_t *entry;
723 ir_node** nodes = alloca(memperm->entrycount * sizeof(nodes[0]));
724 ir_node* mempermnode;
726 assert(memperm->entrycount > 0);
728 for(entry = memperm->entries, i = 0; entry != NULL; entry = entry->next, ++i) {
729 ir_node* arg = get_irn_n(entry->node, entry->pos);
733 mempermnode = be_new_MemPerm(arch_env, irg, memperm->block,
734 memperm->entrycount, nodes);
736 // insert node into schedule
737 blockend = get_end_of_block_insertion_point(memperm->block);
738 sched_add_before(blockend, mempermnode);
739 stat_ev_dbl("mem_perm", memperm->entrycount);
742 for(entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
744 ir_node* arg = get_irn_n(entry->node, entry->pos);
746 be_set_MemPerm_in_entity(mempermnode, i, entry->in);
747 be_set_MemPerm_out_entity(mempermnode, i, entry->out);
748 set_irg_current_block(irg, memperm->block);
749 proj = new_Proj(mempermnode, get_irn_mode(arg), i);
751 set_irn_n(entry->node, entry->pos, proj);
756 static int count_spillslots(const be_fec_env_t *env)
758 const spill_t *spill;
759 int spillcount = set_count(env->spills);
760 bitset_t *counted = bitset_alloca(spillcount);
764 for(spill = set_first(env->spills); spill != NULL;
765 spill = set_next(env->spills)) {
766 int spillslot = spill->spillslot;
767 if(!bitset_is_set(counted, spillslot)) {
769 bitset_set(counted, spillslot);
776 be_fec_env_t *be_new_frame_entity_coalescer(be_irg_t *birg)
778 const arch_env_t *arch_env = birg->main_env->arch_env;
779 be_fec_env_t *env = xmalloc(sizeof(env[0]));
781 be_liveness_assure_chk(be_assure_liveness(birg));
783 obstack_init(&env->obst);
784 env->arch_env = arch_env;
786 env->spills = new_set(cmp_spill, 10);
787 env->reloads = NEW_ARR_F(ir_node*, 0);
788 env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
789 env->memperms = new_set(cmp_memperm, 10);
794 void be_free_frame_entity_coalescer(be_fec_env_t *env)
796 del_set(env->memperms);
797 DEL_ARR_F(env->reloads);
798 DEL_ARR_F(env->affinity_edges);
799 del_set(env->spills);
800 obstack_free(&env->obst, NULL);
805 void be_assign_entities(be_fec_env_t *env)
807 stat_ev_dbl("spillslots", set_count(env->spills));
809 if(be_coalesce_spill_slots) {
810 do_greedy_coalescing(env);
813 stat_ev_dbl("spillslots_after_coalescing", count_spillslots(env));
815 assign_spillslots(env);
817 create_memperms(env);
821 * This walker function searches for reloads and collects all the spills
822 * and memphis attached to them.
824 static void collect_spills_walker(ir_node *node, void *data)
826 be_fec_env_t *env = data;
827 const arch_env_t *arch_env = env->arch_env;
829 const arch_register_class_t *cls;
832 /* classify returns classification of the irn the proj is attached to */
836 if (!arch_irn_class_is(arch_env, node, reload))
839 mode = get_irn_mode(node);
840 cls = arch_get_irn_reg_class(arch_env, node, -1);
841 align = arch_isa_get_reg_class_alignment(arch_env_get_isa(arch_env), cls);
843 be_node_needs_frame_entity(env, node, mode, align);
846 void be_coalesce_spillslots(be_irg_t *birg)
848 be_fec_env_t *env = be_new_frame_entity_coalescer(birg);
850 /* collect reloads */
851 irg_walk_graph(birg->irg, NULL, collect_spills_walker, env);
853 be_assign_entities(env);
855 be_free_frame_entity_coalescer(env);
858 void be_init_spillslots(void)
860 FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");
863 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots);