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 Block-scheduling strategies.
23 * @author Matthias Braun, Christoph Mallon
27 * The goals of the greedy (and ILP) algorithm here works by assuming that
28 * we want to change as many jumps to fallthroughs as possible (executed jumps
29 * actually, we have to look at the execution frequencies). The algorithms
30 * do this by collecting execution frequencies of all branches (which is easily
31 * possible when all critical edges are split) then removes critical edges where
32 * possible as we don't need and want them anymore now. The algorithms then try
33 * to change as many edges to fallthroughs as possible, this is done by setting
34 * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
35 * execution frequencies and tries to transform them to fallthroughs in this order
41 #include "beblocksched.h"
51 #include "irgraph_t.h"
63 #include "lc_opts_enum.h"
67 #include <lpp/lpp_net.h>
70 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
72 typedef enum _blocksched_algos_t {
73 BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
76 static int algo = BLOCKSCHED_GREEDY;
78 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
79 { "naiv", BLOCKSCHED_NAIV },
80 { "extbb", BLOCKSCHED_EXTBB },
81 { "greedy", BLOCKSCHED_GREEDY },
83 { "ilp", BLOCKSCHED_ILP },
88 static lc_opt_enum_int_var_t algo_var = {
89 &algo, blockschedalgo_items
92 static const lc_opt_table_entry_t be_blocksched_options[] = {
93 LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
99 * / ___|_ __ ___ ___ __| |_ _
100 * | | _| '__/ _ \/ _ \/ _` | | | |
101 * | |_| | | | __/ __/ (_| | |_| |
102 * \____|_| \___|\___|\__,_|\__, |
106 typedef struct _blocksched_entry_t {
108 struct _blocksched_entry_t *next;
109 struct _blocksched_entry_t *prev;
110 } blocksched_entry_t;
112 typedef struct _edge_t {
113 ir_node *block; /**< source block */
114 int pos; /**< number of cfg predecessor (target) */
115 double execfreq; /**< the frequency */
116 int highest_execfreq; /**< flag that indicates whether this edge is the edge with the highest
117 execfreq pointing away from this block */
120 typedef struct _blocksched_env_t {
122 struct obstack *obst;
123 ir_exec_freq *execfreqs;
130 * Collect cfg frequencies of all edges between blocks.
131 * Also determines edge with highest frequency.
133 static void collect_egde_frequency(ir_node *block, void *data)
135 blocksched_env_t *env = data;
138 blocksched_entry_t *entry;
140 entry = obstack_alloc(env->obst, sizeof(entry[0]));
141 entry->block = block;
144 set_irn_link(block, entry);
146 arity = get_Block_n_cfgpreds(block);
149 assert(block == get_irg_start_block(env->irg)
150 || block == get_irg_end_block(env->irg));
151 /* must be the start block (or end-block for endless loops), nothing to
154 } else if (arity == 1) {
157 edge.execfreq = get_block_execfreq(env->execfreqs, block);
158 edge.highest_execfreq = 1;
159 ARR_APP1(edge_t, env->edges, edge);
162 double highest_execfreq = -1.0;
163 int highest_edge_num = -1;
166 for (i = 0; i < arity; ++i) {
168 ir_node *pred_block = get_Block_cfgpred_block(block, i);
170 execfreq = get_block_execfreq(env->execfreqs, pred_block);
173 edge.execfreq = execfreq;
174 edge.highest_execfreq = 0;
175 ARR_APP1(edge_t, env->edges, edge);
177 if (execfreq > highest_execfreq) {
178 highest_execfreq = execfreq;
179 highest_edge_num = ARR_LEN(env->edges) - 1;
183 if(highest_edge_num >= 0)
184 env->edges[highest_edge_num].highest_execfreq = 1;
188 static int cmp_edges(const void *d1, const void *d2)
190 const edge_t *e1 = d1;
191 const edge_t *e2 = d2;
193 return QSORT_CMP(e2->execfreq, e1->execfreq);
196 static void coalesce_blocks(blocksched_env_t *env)
199 int edge_count = ARR_LEN(env->edges);
201 /* run1: only look at jumps */
202 for (i = 0; i < edge_count; ++i) {
203 const edge_t *edge = &env->edges[i];
204 ir_node *block = edge->block;
207 blocksched_entry_t *entry, *pred_entry;
209 /* only check edge with highest frequency */
210 if (! edge->highest_execfreq)
213 /* the block might have been removed already... */
214 if (is_Bad(get_Block_cfgpred(block, 0)))
217 pred_block = get_Block_cfgpred_block(block, pos);
218 entry = get_irn_link(block);
219 pred_entry = get_irn_link(pred_block);
221 if (pred_entry->next != NULL || entry->prev != NULL)
224 /* only coalesce jumps */
225 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
228 /* schedule the 2 blocks behind each other */
229 DBG((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
230 pred_entry->block, entry->block, edge->execfreq));
231 pred_entry->next = entry;
232 entry->prev = pred_entry;
235 /* run2: remaining edges */
236 for (i = 0; i < edge_count; ++i) {
237 const edge_t *edge = &env->edges[i];
238 ir_node *block = edge->block;
241 blocksched_entry_t *entry, *pred_entry;
243 /* the block might have been removed already... */
244 if (is_Bad(get_Block_cfgpred(block, 0)))
247 /* we can't do fallthroughs in backedges */
248 if (is_backedge(block, pos))
251 pred_block = get_Block_cfgpred_block(block, pos);
252 entry = get_irn_link(block);
253 pred_entry = get_irn_link(pred_block);
255 /* is 1 of the blocks already attached to another block? */
256 if (pred_entry->next != NULL || entry->prev != NULL)
259 /* schedule the 2 blocks behind each other */
260 DBG((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
261 pred_entry->block, entry->block, edge->execfreq));
262 pred_entry->next = entry;
263 entry->prev = pred_entry;
267 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
269 ir_node *block = entry->block;
270 ir_node *succ = NULL;
271 blocksched_entry_t *succ_entry;
272 const ir_edge_t *edge;
273 double best_succ_execfreq;
275 if (irn_visited(block))
279 mark_irn_visited(block);
281 DBG((dbg, LEVEL_1, "Pick succ of %+F\n", block));
283 /* put all successors into the worklist */
284 foreach_block_succ(block, edge) {
285 ir_node *succ_block = get_edge_src_irn(edge);
287 if (irn_visited(succ_block))
290 /* we only need to put the first of a series of already connected
291 * blocks into the worklist */
292 succ_entry = get_irn_link(succ_block);
293 while (succ_entry->prev != NULL) {
294 /* break cycles... */
295 if (succ_entry->prev->block == succ_block) {
296 succ_entry->prev->next = NULL;
297 succ_entry->prev = NULL;
300 succ_entry = succ_entry->prev;
303 if (irn_visited(succ_entry->block))
306 DBG((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
307 pdeq_putr(env->worklist, succ_entry->block);
310 if (entry->next != NULL) {
311 pick_block_successor(entry->next, env);
315 DBG((dbg, LEVEL_1, "deciding...\n"));
316 best_succ_execfreq = -1;
318 /* no successor yet: pick the successor block with the highest execution
319 * frequency which has no predecessor yet */
321 foreach_block_succ(block, edge) {
322 ir_node *succ_block = get_edge_src_irn(edge);
325 if (irn_visited(succ_block))
328 succ_entry = get_irn_link(succ_block);
329 if (succ_entry->prev != NULL)
332 execfreq = get_block_execfreq(env->execfreqs, succ_block);
333 if (execfreq > best_succ_execfreq) {
334 best_succ_execfreq = execfreq;
340 DBG((dbg, LEVEL_1, "pick from worklist\n"));
343 if (pdeq_empty(env->worklist)) {
344 DBG((dbg, LEVEL_1, "worklist empty\n"));
347 succ = pdeq_getl(env->worklist);
348 } while (irn_visited(succ));
351 succ_entry = get_irn_link(succ);
352 entry->next = succ_entry;
353 succ_entry->prev = entry;
355 pick_block_successor(succ_entry, env);
358 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
360 ir_graph *irg = env->irg;
361 ir_node *startblock = get_irg_start_block(irg);
362 blocksched_entry_t *entry = get_irn_link(startblock);
364 set_using_irn_visited(irg);
365 inc_irg_visited(irg);
367 env->worklist = new_pdeq();
368 pick_block_successor(entry, env);
369 assert(pdeq_empty(env->worklist));
370 del_pdeq(env->worklist);
372 clear_using_irn_visited(irg);
377 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
378 int count, struct obstack* obst)
381 ir_node **block_list;
382 blocksched_entry_t *entry;
385 block_list = NEW_ARR_D(ir_node *, obst, count);
386 DBG((dbg, LEVEL_1, "Blockschedule:\n"));
388 for (entry = first; entry != NULL; entry = entry->next) {
390 block_list[i++] = entry->block;
391 DBG((dbg, LEVEL_1, "\t%+F\n", entry->block));
398 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
400 blocksched_env_t env;
402 blocksched_entry_t *start_entry;
403 ir_node **block_list;
409 env.execfreqs = execfreqs;
410 env.edges = NEW_ARR_F(edge_t, 0);
414 // collect edge execution frequencies
415 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
417 // sort interblock edges by execution frequency
418 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
420 (void)be_remove_empty_blocks(irg);
422 if (algo != BLOCKSCHED_NAIV)
423 coalesce_blocks(&env);
425 start_entry = finish_block_schedule(&env);
426 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
428 DEL_ARR_F(env.edges);
429 obstack_free(&obst, NULL);
444 typedef struct _ilp_edge_t {
445 ir_node *block; /**< source block */
446 int pos; /**< number of cfg predecessor (target) */
450 typedef struct _blocksched_ilp_env_t {
451 blocksched_env_t env;
452 ilp_edge_t *ilpedges;
454 } blocksched_ilp_env_t;
456 typedef struct _blocksched_ilp_entry_t {
458 struct _blocksched_entry_t *next;
459 struct _blocksched_entry_t *prev;
462 } blocksched_ilp_entry_t;
464 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
468 int edgeidx = ARR_LEN(env->ilpedges);
470 snprintf(name, sizeof(name), "edge%d", edgeidx);
474 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
476 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
480 static void collect_egde_frequency_ilp(ir_node *block, void *data)
482 blocksched_ilp_env_t *env = data;
483 ir_graph *irg = env->env.irg;
484 ir_node *startblock = get_irg_start_block(irg);
489 blocksched_ilp_entry_t *entry;
491 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
492 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
494 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
495 entry->block = block;
498 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
499 set_irn_link(block, entry);
501 if (block == startblock)
504 arity = get_irn_arity(block);
506 double execfreq = get_block_execfreq(env->env.execfreqs, block);
507 add_ilp_edge(block, 0, execfreq, env);
511 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
513 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
514 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
516 for (i = 0; i < arity; ++i) {
520 ir_node *pred_block = get_Block_cfgpred_block(block, i);
522 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
523 edgenum = add_ilp_edge(block, i, execfreq, env);
524 edge = &env->ilpedges[edgenum];
525 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
531 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
534 int edge_count = ARR_LEN(env->ilpedges);
536 /* complete out constraints */
537 for(i = 0; i < edge_count; ++i) {
538 const ilp_edge_t *edge = &env->ilpedges[i];
539 ir_node *block = edge->block;
541 blocksched_ilp_entry_t *entry;
543 /* the block might have been removed already... */
544 if (is_Bad(get_Block_cfgpred(block, 0)))
547 pred = get_Block_cfgpred_block(block, edge->pos);
548 entry = get_irn_link(pred);
550 DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
551 pred, block, edge->pos));
552 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
559 lpp_dump(env->lpp, "lpp.out");
560 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
561 f = fopen(fname, "w");
562 lpp_dump_plain(env->lpp, f);
567 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
568 lpp_solve_net(env->lpp, "i44pc52", "cplex");
569 assert(lpp_is_sol_valid(env->lpp));
571 /* Apply results to edges */
572 for (i = 0; i < edge_count; ++i) {
573 const ilp_edge_t *edge = &env->ilpedges[i];
574 ir_node *block = edge->block;
577 blocksched_entry_t *entry;
578 blocksched_entry_t *pred_entry;
580 /* the block might have been removed already... */
581 if (is_Bad(get_Block_cfgpred(block, 0)))
584 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
588 pred = get_Block_cfgpred_block(block, edge->pos);
589 entry = get_irn_link(block);
590 pred_entry = get_irn_link(pred);
592 assert(entry->prev == NULL && pred_entry->next == NULL);
593 entry->prev = pred_entry;
594 pred_entry->next = entry;
598 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
600 blocksched_ilp_env_t env;
602 blocksched_entry_t *start_entry;
603 ir_node **block_list;
608 env.env.obst = &obst;
609 env.env.execfreqs = execfreqs;
610 env.env.worklist = NULL;
611 env.env.blockcount = 0;
612 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
614 env.lpp = new_lpp("blockschedule", lpp_minimize);
615 lpp_set_time_limit(env.lpp, 20);
616 lpp_set_log(env.lpp, stdout);
618 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
620 (void)be_remove_empty_blocks(irg);
621 coalesce_blocks_ilp(&env);
623 start_entry = finish_block_schedule(&env.env);
624 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
626 DEL_ARR_F(env.ilpedges);
628 obstack_free(&obst, NULL);
632 #endif /* WITH_ILP */
636 * | ____|_ _| |_| __ )| __ )
637 * | _| \ \/ / __| _ \| _ \
638 * | |___ > <| |_| |_) | |_) |
639 * |_____/_/\_\\__|____/|____/
643 /** A simple forward single linked list. */
645 ir_node *start; /**< start of the list */
646 ir_node *end; /**< last block in the list */
647 unsigned n_blks; /**< number of blocks in the list */
650 static void add_block(anchor *list, ir_node *block) {
651 if (list->start == NULL) {
655 set_irn_link(list->end, block);
662 static void create_block_list(ir_node *leader_block, anchor *list) {
664 const ir_edge_t *edge;
665 ir_node *block = NULL;
666 ir_extblk *extbb = get_Block_extbb(leader_block);
668 if (extbb_visited(extbb))
670 mark_extbb_visited(extbb);
672 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
673 block = get_extbb_block(extbb, i);
674 add_block(list, block);
677 assert(block != NULL);
679 /* pick successor extbbs */
680 foreach_block_succ(block, edge) {
681 ir_node *succ = get_edge_src_irn(edge);
682 create_block_list(succ, list);
685 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
686 block = get_extbb_block(extbb, i);
688 foreach_block_succ(block, edge) {
689 ir_node *succ = get_edge_src_irn(edge);
690 create_block_list(succ, list);
695 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
698 * Calculates a block schedule. The schedule is stored as a linked
699 * list starting at the start_block of the irg.
701 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
704 ir_node **blk_list, *b, *n;
707 /* schedule extended basic blocks */
708 compute_extbb_execfreqs(irg, execfreqs);
709 //compute_extbb(irg);
715 set_using_irn_link(irg);
716 set_using_irn_visited(irg);
717 inc_irg_block_visited(irg);
719 create_block_list(get_irg_start_block(irg), &list);
721 /** create an array, so we can go forward and backward */
722 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
724 for (i = 0, b = list.start; b; b = n, ++i) {
729 clear_using_irn_link(irg);
730 clear_using_irn_visited(irg);
738 * | |\/| |/ _` | | '_ \
739 * | | | | (_| | | | | |
740 * |_| |_|\__,_|_|_| |_|
743 void be_init_blocksched(void)
745 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
746 lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
748 lc_opt_add_table(blocksched_grp, be_blocksched_options);
750 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
753 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
755 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
758 case BLOCKSCHED_GREEDY:
759 case BLOCKSCHED_NAIV:
760 return create_block_schedule_greedy(irg, execfreqs);
761 case BLOCKSCHED_EXTBB:
762 return create_extbb_block_schedule(irg, execfreqs);
765 return create_block_schedule_ilp(irg, execfreqs);
766 #endif /* WITH_ILP */
769 assert(0 && "unknown blocksched algo");