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 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"
62 #include <libcore/lc_opts.h>
63 #include <libcore/lc_opts_enum.h>
64 #include <libcore/lc_timing.h>
68 #include <lpp/lpp_net.h>
71 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
73 typedef enum _blocksched_algos_t {
74 BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
77 static int algo = BLOCKSCHED_GREEDY;
79 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
80 { "naiv", BLOCKSCHED_NAIV },
81 { "extbb", BLOCKSCHED_EXTBB },
82 { "greedy", BLOCKSCHED_GREEDY },
84 { "ilp", BLOCKSCHED_ILP },
89 static lc_opt_enum_int_var_t algo_var = {
90 &algo, blockschedalgo_items
93 static const lc_opt_table_entry_t be_blocksched_options[] = {
94 LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
100 * / ___|_ __ ___ ___ __| |_ _
101 * | | _| '__/ _ \/ _ \/ _` | | | |
102 * | |_| | | | __/ __/ (_| | |_| |
103 * \____|_| \___|\___|\__,_|\__, |
107 typedef struct _blocksched_entry_t {
109 struct _blocksched_entry_t *next;
110 struct _blocksched_entry_t *prev;
111 } blocksched_entry_t;
113 typedef struct _edge_t {
114 ir_node *block; /**< source block */
115 int pos; /**< number of cfg predecessor (target) */
116 double execfreq; /**< the frequency */
117 int highest_execfreq; /**< flag that indicates whether this edge is the edge with the highest
118 execfreq pointing away from this block */
121 typedef struct _blocksched_env_t {
123 struct obstack *obst;
124 ir_exec_freq *execfreqs;
131 * Collect cfg frequencies of all edges between blocks.
132 * Also determines edge with highest frequency.
134 static void collect_egde_frequency(ir_node *block, void *data)
136 blocksched_env_t *env = data;
139 blocksched_entry_t *entry;
141 entry = obstack_alloc(env->obst, sizeof(entry[0]));
142 entry->block = block;
145 set_irn_link(block, entry);
147 arity = get_Block_n_cfgpreds(block);
150 assert(block == get_irg_start_block(env->irg));
151 /* must be the start block, nothing to do here */
153 } else if (arity == 1) {
156 edge.execfreq = get_block_execfreq(env->execfreqs, block);
157 edge.highest_execfreq = 1;
158 ARR_APP1(edge_t, env->edges, edge);
161 double highest_execfreq = -1.0;
162 int highest_edge_num = -1;
165 for (i = 0; i < arity; ++i) {
167 ir_node *pred_block = get_Block_cfgpred_block(block, i);
169 execfreq = get_block_execfreq(env->execfreqs, pred_block);
172 edge.execfreq = execfreq;
173 edge.highest_execfreq = 0;
174 ARR_APP1(edge_t, env->edges, edge);
176 if (execfreq > highest_execfreq) {
177 highest_execfreq = execfreq;
178 highest_edge_num = ARR_LEN(env->edges) - 1;
182 if(highest_edge_num >= 0)
183 env->edges[highest_edge_num].highest_execfreq = 1;
187 static int cmp_edges(const void *d1, const void *d2)
189 const edge_t *e1 = d1;
190 const edge_t *e2 = d2;
192 return QSORT_CMP(e2->execfreq, e1->execfreq);
195 static void coalesce_blocks(blocksched_env_t *env)
198 int edge_count = ARR_LEN(env->edges);
200 /* run1: only look at jumps */
201 for (i = 0; i < edge_count; ++i) {
202 const edge_t *edge = &env->edges[i];
203 ir_node *block = edge->block;
206 blocksched_entry_t *entry, *pred_entry;
208 /* only check edge with highest frequency */
209 if (! edge->highest_execfreq)
212 /* the block might have been removed already... */
213 if (is_Bad(get_Block_cfgpred(block, 0)))
216 pred_block = get_Block_cfgpred_block(block, pos);
217 entry = get_irn_link(block);
218 pred_entry = get_irn_link(pred_block);
220 if (pred_entry->next != NULL || entry->prev != NULL)
223 /* only coalesce jumps */
224 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
227 /* schedule the 2 blocks behind each other */
228 DBG((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
229 pred_entry->block, entry->block, edge->execfreq));
230 pred_entry->next = entry;
231 entry->prev = pred_entry;
234 /* run2: remaining edges */
235 for (i = 0; i < edge_count; ++i) {
236 const edge_t *edge = &env->edges[i];
237 ir_node *block = edge->block;
240 blocksched_entry_t *entry, *pred_entry;
242 /* the block might have been removed already... */
243 if (is_Bad(get_Block_cfgpred(block, 0)))
246 /* we can't do fallthroughs in backedges */
247 if (is_backedge(block, pos))
250 pred_block = get_Block_cfgpred_block(block, pos);
251 entry = get_irn_link(block);
252 pred_entry = get_irn_link(pred_block);
254 /* is 1 of the blocks already attached to another block? */
255 if (pred_entry->next != NULL || entry->prev != NULL)
258 /* schedule the 2 blocks behind each other */
259 DBG((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
260 pred_entry->block, entry->block, edge->execfreq));
261 pred_entry->next = entry;
262 entry->prev = pred_entry;
266 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
268 ir_node *block = entry->block;
269 ir_node *succ = NULL;
270 blocksched_entry_t *succ_entry;
271 const ir_edge_t *edge;
272 double best_succ_execfreq;
274 if (irn_visited(block))
278 mark_irn_visited(block);
280 DBG((dbg, LEVEL_1, "Pick succ of %+F\n", block));
282 /* put all successors into the worklist */
283 foreach_block_succ(block, edge) {
284 ir_node *succ_block = get_edge_src_irn(edge);
286 if (irn_visited(succ_block))
289 /* we only need to put the first of a series of already connected
290 * blocks into the worklist */
291 succ_entry = get_irn_link(succ_block);
292 while (succ_entry->prev != NULL) {
293 /* break cycles... */
294 if (succ_entry->prev->block == succ_block) {
295 succ_entry->prev->next = NULL;
296 succ_entry->prev = NULL;
299 succ_entry = succ_entry->prev;
302 if (irn_visited(succ_entry->block))
305 DBG((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
306 pdeq_putr(env->worklist, succ_entry->block);
309 if (entry->next != NULL) {
310 pick_block_successor(entry->next, env);
314 DBG((dbg, LEVEL_1, "deciding...\n"));
315 best_succ_execfreq = -1;
317 /* no successor yet: pick the successor block with the highest execution
318 * frequency which has no predecessor yet */
320 foreach_block_succ(block, edge) {
321 ir_node *succ_block = get_edge_src_irn(edge);
324 if (irn_visited(succ_block))
327 succ_entry = get_irn_link(succ_block);
328 if (succ_entry->prev != NULL)
331 execfreq = get_block_execfreq(env->execfreqs, succ_block);
332 if (execfreq > best_succ_execfreq) {
333 best_succ_execfreq = execfreq;
339 DBG((dbg, LEVEL_1, "pick from worklist\n"));
342 if (pdeq_empty(env->worklist)) {
343 DBG((dbg, LEVEL_1, "worklist empty\n"));
346 succ = pdeq_getl(env->worklist);
347 } while (irn_visited(succ));
350 succ_entry = get_irn_link(succ);
351 entry->next = succ_entry;
352 succ_entry->prev = entry;
354 pick_block_successor(succ_entry, env);
357 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
359 ir_graph *irg = env->irg;
360 ir_node *startblock = get_irg_start_block(irg);
361 blocksched_entry_t *entry = get_irn_link(startblock);
363 set_using_visited(irg);
364 inc_irg_visited(irg);
366 env->worklist = new_pdeq();
367 pick_block_successor(entry, env);
368 assert(pdeq_empty(env->worklist));
369 del_pdeq(env->worklist);
371 clear_using_visited(irg);
376 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
377 int count, struct obstack* obst)
380 ir_node **block_list;
381 blocksched_entry_t *entry;
384 block_list = NEW_ARR_D(ir_node *, obst, count);
385 DBG((dbg, LEVEL_1, "Blockschedule:\n"));
387 for (entry = first; entry != NULL; entry = entry->next) {
389 block_list[i++] = entry->block;
390 DBG((dbg, LEVEL_1, "\t%+F\n", entry->block));
397 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
399 blocksched_env_t env;
401 blocksched_entry_t *start_entry;
402 ir_node **block_list;
408 env.execfreqs = execfreqs;
409 env.edges = NEW_ARR_F(edge_t, 0);
413 // collect edge execution frequencies
414 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
416 // sort interblock edges by execution frequency
417 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
419 (void)be_remove_empty_blocks(irg);
421 if (algo != BLOCKSCHED_NAIV)
422 coalesce_blocks(&env);
424 start_entry = finish_block_schedule(&env);
425 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
427 DEL_ARR_F(env.edges);
428 obstack_free(&obst, NULL);
443 typedef struct _ilp_edge_t {
444 ir_node *block; /**< source block */
445 int pos; /**< number of cfg predecessor (target) */
449 typedef struct _blocksched_ilp_env_t {
450 blocksched_env_t env;
451 ilp_edge_t *ilpedges;
453 } blocksched_ilp_env_t;
455 typedef struct _blocksched_ilp_entry_t {
457 struct _blocksched_entry_t *next;
458 struct _blocksched_entry_t *prev;
461 } blocksched_ilp_entry_t;
463 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
467 int edgeidx = ARR_LEN(env->ilpedges);
469 snprintf(name, sizeof(name), "edge%d", edgeidx);
473 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
475 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
479 static void collect_egde_frequency_ilp(ir_node *block, void *data)
481 blocksched_ilp_env_t *env = data;
482 ir_graph *irg = env->env.irg;
483 ir_node *startblock = get_irg_start_block(irg);
488 blocksched_ilp_entry_t *entry;
490 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
491 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
493 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
494 entry->block = block;
497 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
498 set_irn_link(block, entry);
500 if (block == startblock)
503 arity = get_irn_arity(block);
505 double execfreq = get_block_execfreq(env->env.execfreqs, block);
506 add_ilp_edge(block, 0, execfreq, env);
510 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
512 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
513 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
515 for (i = 0; i < arity; ++i) {
519 ir_node *pred_block = get_Block_cfgpred_block(block, i);
521 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
522 edgenum = add_ilp_edge(block, i, execfreq, env);
523 edge = &env->ilpedges[edgenum];
524 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
530 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
533 int edge_count = ARR_LEN(env->ilpedges);
535 /* complete out constraints */
536 for(i = 0; i < edge_count; ++i) {
537 const ilp_edge_t *edge = &env->ilpedges[i];
538 ir_node *block = edge->block;
540 blocksched_ilp_entry_t *entry;
542 /* the block might have been removed already... */
543 if (is_Bad(get_Block_cfgpred(block, 0)))
546 pred = get_Block_cfgpred_block(block, edge->pos);
547 entry = get_irn_link(pred);
549 DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
550 pred, block, edge->pos));
551 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
558 lpp_dump(env->lpp, "lpp.out");
559 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
560 f = fopen(fname, "w");
561 lpp_dump_plain(env->lpp, f);
566 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
567 lpp_solve_net(env->lpp, "i44pc52", "cplex");
568 assert(lpp_is_sol_valid(env->lpp));
570 /* Apply results to edges */
571 for (i = 0; i < edge_count; ++i) {
572 const ilp_edge_t *edge = &env->ilpedges[i];
573 ir_node *block = edge->block;
576 blocksched_entry_t *entry;
577 blocksched_entry_t *pred_entry;
579 /* the block might have been removed already... */
580 if (is_Bad(get_Block_cfgpred(block, 0)))
583 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
587 pred = get_Block_cfgpred_block(block, edge->pos);
588 entry = get_irn_link(block);
589 pred_entry = get_irn_link(pred);
591 assert(entry->prev == NULL && pred_entry->next == NULL);
592 entry->prev = pred_entry;
593 pred_entry->next = entry;
597 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
599 blocksched_ilp_env_t env;
601 blocksched_entry_t *start_entry;
602 ir_node **block_list;
607 env.env.obst = &obst;
608 env.env.execfreqs = execfreqs;
609 env.env.worklist = NULL;
610 env.env.blockcount = 0;
611 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
613 env.lpp = new_lpp("blockschedule", lpp_minimize);
614 lpp_set_time_limit(env.lpp, 20);
615 lpp_set_log(env.lpp, stdout);
617 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
619 (void)be_remove_empty_blocks(irg);
620 coalesce_blocks_ilp(&env);
622 start_entry = finish_block_schedule(&env.env);
623 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
625 DEL_ARR_F(env.ilpedges);
627 obstack_free(&obst, NULL);
631 #endif /* WITH_ILP */
635 * | ____|_ _| |_| __ )| __ )
636 * | _| \ \/ / __| _ \| _ \
637 * | |___ > <| |_| |_) | |_) |
638 * |_____/_/\_\\__|____/|____/
642 /** A simple forward single linked list. */
644 ir_node *start; /**< start of the list */
645 ir_node *end; /**< last block in the list */
646 unsigned n_blks; /**< number of blocks in the list */
649 static void add_block(anchor *list, ir_node *block) {
650 if (list->start == NULL) {
654 set_irn_link(list->end, block);
661 static void create_block_list(ir_node *leader_block, anchor *list) {
663 const ir_edge_t *edge;
664 ir_node *block = NULL;
665 ir_extblk *extbb = get_Block_extbb(leader_block);
667 if (extbb_visited(extbb))
669 mark_extbb_visited(extbb);
671 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
672 block = get_extbb_block(extbb, i);
673 add_block(list, block);
676 assert(block != NULL);
678 /* pick successor extbbs */
679 foreach_block_succ(block, edge) {
680 ir_node *succ = get_edge_src_irn(edge);
681 create_block_list(succ, list);
684 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
685 block = get_extbb_block(extbb, i);
687 foreach_block_succ(block, edge) {
688 ir_node *succ = get_edge_src_irn(edge);
689 create_block_list(succ, list);
694 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
697 * Calculates a block schedule. The schedule is stored as a linked
698 * list starting at the start_block of the irg.
700 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
703 ir_node **blk_list, *b, *n;
706 /* schedule extended basic blocks */
707 compute_extbb_execfreqs(irg, execfreqs);
708 //compute_extbb(irg);
714 set_using_irn_link(irg);
715 set_using_visited(irg);
716 inc_irg_block_visited(irg);
718 create_block_list(get_irg_start_block(irg), &list);
720 /** create an array, so we can go forward and backward */
721 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
723 for (i = 0, b = list.start; b; b = n, ++i) {
728 clear_using_irn_link(irg);
729 clear_using_visited(irg);
737 * | |\/| |/ _` | | '_ \
738 * | | | | (_| | | | | |
739 * |_| |_|\__,_|_|_| |_|
742 void be_init_blocksched(void)
744 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
745 lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
747 lc_opt_add_table(blocksched_grp, be_blocksched_options);
749 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
752 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
754 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
757 case BLOCKSCHED_GREEDY:
758 case BLOCKSCHED_NAIV:
759 return create_block_schedule_greedy(irg, execfreqs);
760 case BLOCKSCHED_EXTBB:
761 return create_extbb_block_schedule(irg, execfreqs);
764 return create_block_schedule_ilp(irg, execfreqs);
765 #endif /* WITH_ILP */
768 assert(0 && "unknown blocksched algo");