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
26 * The goals of the greedy (and ILP) algorithm here works by assuming that
27 * we want to change as many jumps to fallthroughs as possible (executed jumps
28 * actually, we have to look at the execution frequencies). The algorithms
29 * do this by collecting execution frequencies of all branches (which is easily
30 * possible when all critical edges are split) then removes critical edges where
31 * possible as we don't need and want them anymore now. The algorithms then try
32 * to change as many edges to fallthroughs as possible, this is done by setting
33 * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
34 * execution frequencies and tries to transform them to fallthroughs in this order
38 #include "beblocksched.h"
48 #include "irgraph_t.h"
62 #include "lc_opts_enum.h"
67 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
69 typedef enum blocksched_algos_t {
70 BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
73 static int algo = BLOCKSCHED_GREEDY;
75 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
76 { "naiv", BLOCKSCHED_NAIV },
77 { "greedy", BLOCKSCHED_GREEDY },
78 { "ilp", BLOCKSCHED_ILP },
82 static lc_opt_enum_int_var_t algo_var = {
83 &algo, blockschedalgo_items
86 static const lc_opt_table_entry_t be_blocksched_options[] = {
87 LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var),
93 * / ___|_ __ ___ ___ __| |_ _
94 * | | _| '__/ _ \/ _ \/ _` | | | |
95 * | |_| | | | __/ __/ (_| | |_| |
96 * \____|_| \___|\___|\__,_|\__, |
100 typedef struct blocksched_entry_t blocksched_entry_t;
101 struct blocksched_entry_t {
103 blocksched_entry_t *next;
104 blocksched_entry_t *prev;
107 typedef struct edge_t edge_t;
109 ir_node *block; /**< source block */
110 int pos; /**< number of cfg predecessor (target) */
111 double execfreq; /**< the frequency */
112 double outedge_penalty_freq; /**< for edges leaving the loop this is the
113 penality when we make them a
115 int highest_execfreq; /**< flag that indicates whether this edge is
116 the edge with the highest execfreq pointing
117 away from this block */
120 typedef struct blocksched_env_t blocksched_env_t;
121 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 = (blocksched_env_t*)data;
139 blocksched_entry_t *entry;
142 memset(&edge, 0, sizeof(edge));
144 entry = OALLOCZ(env->obst, blocksched_entry_t);
145 entry->block = block;
146 set_irn_link(block, entry);
148 loop = get_irn_loop(block);
150 arity = get_Block_n_cfgpreds(block);
153 /* must be the start block (or end-block for endless loops),
154 * everything else is dead code and should be removed by now */
155 assert(block == get_irg_start_block(env->irg)
156 || block == get_irg_end_block(env->irg));
157 /* nothing to do here */
159 } else if (arity == 1) {
160 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
161 ir_loop *pred_loop = get_irn_loop(pred_block);
162 float freq = (float)get_block_execfreq(env->execfreqs, block);
164 /* is it an edge leaving a loop */
165 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
166 float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
167 edge.outedge_penalty_freq = -(pred_freq - freq);
172 edge.execfreq = freq;
173 edge.highest_execfreq = 1;
174 ARR_APP1(edge_t, env->edges, edge);
177 double highest_execfreq = -1.0;
178 int highest_edge_num = -1;
181 for (i = 0; i < arity; ++i) {
183 ir_node *pred_block = get_Block_cfgpred_block(block, i);
185 execfreq = get_block_execfreq(env->execfreqs, pred_block);
188 edge.execfreq = execfreq;
189 edge.highest_execfreq = 0;
190 ARR_APP1(edge_t, env->edges, edge);
192 if (execfreq > highest_execfreq) {
193 highest_execfreq = execfreq;
194 highest_edge_num = ARR_LEN(env->edges) - 1;
198 if (highest_edge_num >= 0)
199 env->edges[highest_edge_num].highest_execfreq = 1;
203 static int cmp_edges_base(const edge_t *e1, const edge_t *e2)
205 long nr1 = get_irn_node_nr(e1->block);
206 long nr2 = get_irn_node_nr(e2->block);
209 } else if (nr1 > nr2) {
212 if (e1->pos < e2->pos) {
214 } else if (e1->pos > e2->pos) {
222 static int cmp_edges(const void *d1, const void *d2)
224 const edge_t *e1 = (const edge_t*)d1;
225 const edge_t *e2 = (const edge_t*)d2;
226 double freq1 = e1->execfreq;
227 double freq2 = e2->execfreq;
230 } else if (freq1 > freq2) {
233 return cmp_edges_base(e1, e2);
237 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
239 const edge_t *e1 = (const edge_t*)d1;
240 const edge_t *e2 = (const edge_t*)d2;
241 double pen1 = e1->outedge_penalty_freq;
242 double pen2 = e2->outedge_penalty_freq;
245 } else if (pen1 < pen2) {
248 return cmp_edges_base(e1, e2);
252 static void clear_loop_links(ir_loop *loop)
256 set_loop_link(loop, NULL);
257 n = get_loop_n_elements(loop);
258 for (i = 0; i < n; ++i) {
259 loop_element elem = get_loop_element(loop, i);
260 if (*elem.kind == k_ir_loop) {
261 clear_loop_links(elem.son);
266 static void coalesce_blocks(blocksched_env_t *env)
269 int edge_count = ARR_LEN(env->edges);
270 edge_t *edges = env->edges;
272 /* sort interblock edges by execution frequency */
273 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
275 /* run1: only look at jumps */
276 for (i = 0; i < edge_count; ++i) {
277 const edge_t *edge = &edges[i];
278 ir_node *block = edge->block;
281 blocksched_entry_t *entry, *pred_entry;
283 /* only check edge with highest frequency */
284 if (! edge->highest_execfreq)
287 /* the block might have been removed already... */
288 if (is_Bad(get_Block_cfgpred(block, 0)))
291 pred_block = get_Block_cfgpred_block(block, pos);
292 entry = (blocksched_entry_t*)get_irn_link(block);
293 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
295 if (pred_entry->next != NULL || entry->prev != NULL)
298 /* only coalesce jumps */
299 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
302 /* schedule the 2 blocks behind each other */
303 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
304 pred_entry->block, entry->block, edge->execfreq));
305 pred_entry->next = entry;
306 entry->prev = pred_entry;
309 /* run2: pick loop fallthroughs */
310 clear_loop_links(get_irg_loop(env->irg));
312 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
313 for (i = 0; i < edge_count; ++i) {
314 const edge_t *edge = &edges[i];
315 ir_node *block = edge->block;
318 blocksched_entry_t *entry, *pred_entry;
322 /* already seen all loop outedges? */
323 if (edge->outedge_penalty_freq == 0)
326 /* the block might have been removed already... */
327 if (is_Bad(get_Block_cfgpred(block, pos)))
330 pred_block = get_Block_cfgpred_block(block, pos);
331 entry = (blocksched_entry_t*)get_irn_link(block);
332 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
334 if (pred_entry->next != NULL || entry->prev != NULL)
337 /* we want at most 1 outedge fallthrough per loop */
338 loop = get_irn_loop(pred_block);
339 if (get_loop_link(loop) != NULL)
342 /* schedule the 2 blocks behind each other */
343 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
344 pred_entry->block, entry->block, edge->execfreq));
345 pred_entry->next = entry;
346 entry->prev = pred_entry;
348 /* all loops left have an outedge now */
349 outer_loop = get_irn_loop(block);
351 /* we set loop link to loop to mark it */
352 set_loop_link(loop, loop);
353 loop = get_loop_outer_loop(loop);
354 } while (loop != outer_loop);
357 /* sort interblock edges by execution frequency */
358 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
360 /* run3: remaining edges */
361 for (i = 0; i < edge_count; ++i) {
362 const edge_t *edge = &edges[i];
363 ir_node *block = edge->block;
366 blocksched_entry_t *entry, *pred_entry;
368 /* the block might have been removed already... */
369 if (is_Bad(get_Block_cfgpred(block, pos)))
372 pred_block = get_Block_cfgpred_block(block, pos);
373 entry = (blocksched_entry_t*)get_irn_link(block);
374 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
376 /* is 1 of the blocks already attached to another block? */
377 if (pred_entry->next != NULL || entry->prev != NULL)
380 /* schedule the 2 blocks behind each other */
381 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
382 pred_entry->block, entry->block, edge->execfreq));
383 pred_entry->next = entry;
384 entry->prev = pred_entry;
388 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
390 ir_node *block = entry->block;
391 ir_node *succ = NULL;
392 blocksched_entry_t *succ_entry;
393 const ir_edge_t *edge;
394 double best_succ_execfreq;
396 if (irn_visited_else_mark(block))
401 DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
403 /* put all successors into the worklist */
404 foreach_block_succ(block, edge) {
405 ir_node *succ_block = get_edge_src_irn(edge);
407 if (irn_visited(succ_block))
410 /* we only need to put the first of a series of already connected
411 * blocks into the worklist */
412 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
413 while (succ_entry->prev != NULL) {
414 /* break cycles... */
415 if (succ_entry->prev->block == succ_block) {
416 succ_entry->prev->next = NULL;
417 succ_entry->prev = NULL;
420 succ_entry = succ_entry->prev;
423 if (irn_visited(succ_entry->block))
426 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
427 pdeq_putr(env->worklist, succ_entry->block);
430 if (entry->next != NULL) {
431 pick_block_successor(entry->next, env);
435 DB((dbg, LEVEL_1, "deciding...\n"));
436 best_succ_execfreq = -1;
438 /* no successor yet: pick the successor block with the highest execution
439 * frequency which has no predecessor yet */
441 foreach_block_succ(block, edge) {
442 ir_node *succ_block = get_edge_src_irn(edge);
445 if (irn_visited(succ_block))
448 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
449 if (succ_entry->prev != NULL)
452 execfreq = get_block_execfreq(env->execfreqs, succ_block);
453 if (execfreq > best_succ_execfreq) {
454 best_succ_execfreq = execfreq;
460 DB((dbg, LEVEL_1, "pick from worklist\n"));
463 if (pdeq_empty(env->worklist)) {
464 DB((dbg, LEVEL_1, "worklist empty\n"));
467 succ = (ir_node*)pdeq_getl(env->worklist);
468 } while (irn_visited(succ));
471 succ_entry = (blocksched_entry_t*)get_irn_link(succ);
472 entry->next = succ_entry;
473 succ_entry->prev = entry;
475 pick_block_successor(succ_entry, env);
478 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
480 ir_graph *irg = env->irg;
481 ir_node *startblock = get_irg_start_block(irg);
482 blocksched_entry_t *entry = (blocksched_entry_t*)get_irn_link(startblock);
484 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
485 inc_irg_visited(irg);
487 env->worklist = new_pdeq();
488 pick_block_successor(entry, env);
489 assert(pdeq_empty(env->worklist));
490 del_pdeq(env->worklist);
492 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
497 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
498 int count, struct obstack* obst)
501 ir_node **block_list;
502 blocksched_entry_t *entry;
505 block_list = NEW_ARR_D(ir_node *, obst, count);
506 DB((dbg, LEVEL_1, "Blockschedule:\n"));
508 for (entry = first; entry != NULL; entry = entry->next) {
510 block_list[i++] = entry->block;
511 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
518 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
520 blocksched_env_t env;
522 blocksched_entry_t *start_entry;
523 ir_node **block_list;
529 env.execfreqs = execfreqs;
530 env.edges = NEW_ARR_F(edge_t, 0);
534 assure_loopinfo(irg);
536 // collect edge execution frequencies
537 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
539 (void)be_remove_empty_blocks(irg);
541 if (algo != BLOCKSCHED_NAIV)
542 coalesce_blocks(&env);
544 start_entry = finish_block_schedule(&env);
545 block_list = create_blocksched_array(&env, start_entry, env.blockcount,
546 be_get_be_obst(irg));
548 DEL_ARR_F(env.edges);
549 obstack_free(&obst, NULL);
563 typedef struct ilp_edge_t {
564 ir_node *block; /**< source block */
565 int pos; /**< number of cfg predecessor (target) */
569 typedef struct blocksched_ilp_env_t {
570 blocksched_env_t env;
571 ilp_edge_t *ilpedges;
573 } blocksched_ilp_env_t;
575 typedef struct blocksched_ilp_entry_t {
577 struct blocksched_entry_t *next;
578 struct blocksched_entry_t *prev;
581 } blocksched_ilp_entry_t;
583 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
587 int edgeidx = ARR_LEN(env->ilpedges);
589 snprintf(name, sizeof(name), "edge%d", edgeidx);
593 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
595 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
599 static void collect_egde_frequency_ilp(ir_node *block, void *data)
601 blocksched_ilp_env_t *env = data;
602 ir_graph *irg = env->env.irg;
603 ir_node *startblock = get_irg_start_block(irg);
608 blocksched_ilp_entry_t *entry;
610 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
611 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
613 entry = OALLOC(env->env.obst, blocksched_ilp_entry_t);
614 entry->block = block;
617 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
618 set_irn_link(block, entry);
620 if (block == startblock)
623 arity = get_irn_arity(block);
625 double execfreq = get_block_execfreq(env->env.execfreqs, block);
626 add_ilp_edge(block, 0, execfreq, env);
631 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
632 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
634 for (i = 0; i < arity; ++i) {
638 ir_node *pred_block = get_Block_cfgpred_block(block, i);
640 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
641 edgenum = add_ilp_edge(block, i, execfreq, env);
642 edge = &env->ilpedges[edgenum];
643 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
649 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
651 int edge_count = ARR_LEN(env->ilpedges);
654 /* complete out constraints */
655 for (i = 0; i < edge_count; ++i) {
656 const ilp_edge_t *edge = &env->ilpedges[i];
657 ir_node *block = edge->block;
659 blocksched_ilp_entry_t *entry;
661 /* the block might have been removed already... */
662 if (is_Bad(get_Block_cfgpred(block, 0)))
665 pred = get_Block_cfgpred_block(block, edge->pos);
666 entry = get_irn_link(pred);
668 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
669 pred, block, edge->pos));
670 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
673 lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver);
674 assert(lpp_is_sol_valid(env->lpp));
676 /* Apply results to edges */
677 for (i = 0; i < edge_count; ++i) {
678 const ilp_edge_t *edge = &env->ilpedges[i];
679 ir_node *block = edge->block;
682 blocksched_entry_t *entry;
683 blocksched_entry_t *pred_entry;
685 /* the block might have been removed already... */
686 if (is_Bad(get_Block_cfgpred(block, 0)))
689 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
693 pred = get_Block_cfgpred_block(block, edge->pos);
694 entry = get_irn_link(block);
695 pred_entry = get_irn_link(pred);
697 assert(entry->prev == NULL && pred_entry->next == NULL);
698 entry->prev = pred_entry;
699 pred_entry->next = entry;
703 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
705 blocksched_ilp_env_t env;
707 blocksched_entry_t *start_entry;
708 ir_node **block_list;
713 env.env.obst = &obst;
714 env.env.execfreqs = execfreqs;
715 env.env.worklist = NULL;
716 env.env.blockcount = 0;
717 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
719 env.lpp = lpp_new("blockschedule", lpp_minimize);
720 lpp_set_time_limit(env.lpp, 20);
721 lpp_set_log(env.lpp, stdout);
723 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
725 (void)be_remove_empty_blocks(irg);
726 coalesce_blocks_ilp(&env);
728 start_entry = finish_block_schedule(&env.env);
729 block_list = create_blocksched_array(&env.env, start_entry,
731 be_get_be_obst(irg));
733 DEL_ARR_F(env.ilpedges);
735 obstack_free(&obst, NULL);
743 * | |\/| |/ _` | | '_ \
744 * | | | | (_| | | | | |
745 * |_| |_|\__,_|_|_| |_|
748 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
749 void be_init_blocksched(void)
751 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
753 lc_opt_add_table(be_grp, be_blocksched_options);
755 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
758 ir_node **be_create_block_schedule(ir_graph *irg)
760 ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
763 case BLOCKSCHED_GREEDY:
764 case BLOCKSCHED_NAIV:
765 return create_block_schedule_greedy(irg, execfreqs);
767 return create_block_schedule_ilp(irg, execfreqs);
770 panic("unknown blocksched algo");