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
39 #include "beblocksched.h"
49 #include "irgraph_t.h"
63 #include "lc_opts_enum.h"
68 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
70 typedef enum blocksched_algos_t {
71 BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
74 static int algo = BLOCKSCHED_GREEDY;
76 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
77 { "naiv", BLOCKSCHED_NAIV },
78 { "greedy", BLOCKSCHED_GREEDY },
79 { "ilp", BLOCKSCHED_ILP },
83 static lc_opt_enum_int_var_t algo_var = {
84 &algo, blockschedalgo_items
87 static const lc_opt_table_entry_t be_blocksched_options[] = {
88 LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var),
94 * / ___|_ __ ___ ___ __| |_ _
95 * | | _| '__/ _ \/ _ \/ _` | | | |
96 * | |_| | | | __/ __/ (_| | |_| |
97 * \____|_| \___|\___|\__,_|\__, |
101 typedef struct blocksched_entry_t blocksched_entry_t;
102 struct blocksched_entry_t {
104 blocksched_entry_t *next;
105 blocksched_entry_t *prev;
108 typedef struct edge_t edge_t;
110 ir_node *block; /**< source block */
111 int pos; /**< number of cfg predecessor (target) */
112 double execfreq; /**< the frequency */
113 double outedge_penalty_freq; /**< for edges leaving the loop this is the
114 penality when we make them a
116 int highest_execfreq; /**< flag that indicates whether this edge is
117 the edge with the highest execfreq pointing
118 away from this block */
121 typedef struct blocksched_env_t blocksched_env_t;
122 struct blocksched_env_t {
124 struct obstack *obst;
125 ir_exec_freq *execfreqs;
132 * Collect cfg frequencies of all edges between blocks.
133 * Also determines edge with highest frequency.
135 static void collect_egde_frequency(ir_node *block, void *data)
137 blocksched_env_t *env = (blocksched_env_t*)data;
140 blocksched_entry_t *entry;
143 memset(&edge, 0, sizeof(edge));
145 entry = OALLOCZ(env->obst, blocksched_entry_t);
146 entry->block = block;
147 set_irn_link(block, entry);
149 loop = get_irn_loop(block);
151 arity = get_Block_n_cfgpreds(block);
154 /* must be the start block (or end-block for endless loops),
155 * everything else is dead code and should be removed by now */
156 assert(block == get_irg_start_block(env->irg)
157 || block == get_irg_end_block(env->irg));
158 /* nothing to do here */
160 } else if (arity == 1) {
161 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
162 ir_loop *pred_loop = get_irn_loop(pred_block);
163 float freq = (float)get_block_execfreq(env->execfreqs, block);
165 /* is it an edge leaving a loop */
166 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
167 float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
168 edge.outedge_penalty_freq = -(pred_freq - freq);
173 edge.execfreq = freq;
174 edge.highest_execfreq = 1;
175 ARR_APP1(edge_t, env->edges, edge);
178 double highest_execfreq = -1.0;
179 int highest_edge_num = -1;
182 for (i = 0; i < arity; ++i) {
184 ir_node *pred_block = get_Block_cfgpred_block(block, i);
186 execfreq = get_block_execfreq(env->execfreqs, pred_block);
189 edge.execfreq = execfreq;
190 edge.highest_execfreq = 0;
191 ARR_APP1(edge_t, env->edges, edge);
193 if (execfreq > highest_execfreq) {
194 highest_execfreq = execfreq;
195 highest_edge_num = ARR_LEN(env->edges) - 1;
199 if (highest_edge_num >= 0)
200 env->edges[highest_edge_num].highest_execfreq = 1;
204 static int cmp_edges(const void *d1, const void *d2)
206 const edge_t *e1 = (const edge_t*)d1;
207 const edge_t *e2 = (const edge_t*)d2;
209 return QSORT_CMP(e2->execfreq, e1->execfreq);
212 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
214 const edge_t *e1 = (const edge_t*)d1;
215 const edge_t *e2 = (const edge_t*)d2;
216 /* reverse sorting as penalties are negative */
217 return QSORT_CMP(e1->outedge_penalty_freq, e2->outedge_penalty_freq);
220 static void clear_loop_links(ir_loop *loop)
224 set_loop_link(loop, NULL);
225 n = get_loop_n_elements(loop);
226 for (i = 0; i < n; ++i) {
227 loop_element elem = get_loop_element(loop, i);
228 if (*elem.kind == k_ir_loop) {
229 clear_loop_links(elem.son);
234 static void coalesce_blocks(blocksched_env_t *env)
237 int edge_count = ARR_LEN(env->edges);
238 edge_t *edges = env->edges;
240 /* sort interblock edges by execution frequency */
241 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
243 /* run1: only look at jumps */
244 for (i = 0; i < edge_count; ++i) {
245 const edge_t *edge = &edges[i];
246 ir_node *block = edge->block;
249 blocksched_entry_t *entry, *pred_entry;
251 /* only check edge with highest frequency */
252 if (! edge->highest_execfreq)
255 /* the block might have been removed already... */
256 if (is_Bad(get_Block_cfgpred(block, 0)))
259 pred_block = get_Block_cfgpred_block(block, pos);
260 entry = (blocksched_entry_t*)get_irn_link(block);
261 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
263 if (pred_entry->next != NULL || entry->prev != NULL)
266 /* only coalesce jumps */
267 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
270 /* schedule the 2 blocks behind each other */
271 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
272 pred_entry->block, entry->block, edge->execfreq));
273 pred_entry->next = entry;
274 entry->prev = pred_entry;
277 /* run2: pick loop fallthroughs */
278 clear_loop_links(get_irg_loop(env->irg));
280 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
281 for (i = 0; i < edge_count; ++i) {
282 const edge_t *edge = &edges[i];
283 ir_node *block = edge->block;
286 blocksched_entry_t *entry, *pred_entry;
290 /* already seen all loop outedges? */
291 if (edge->outedge_penalty_freq == 0)
294 /* the block might have been removed already... */
295 if (is_Bad(get_Block_cfgpred(block, pos)))
298 pred_block = get_Block_cfgpred_block(block, pos);
299 entry = (blocksched_entry_t*)get_irn_link(block);
300 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
302 if (pred_entry->next != NULL || entry->prev != NULL)
305 /* we want at most 1 outedge fallthrough per loop */
306 loop = get_irn_loop(pred_block);
307 if (get_loop_link(loop) != NULL)
310 /* schedule the 2 blocks behind each other */
311 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
312 pred_entry->block, entry->block, edge->execfreq));
313 pred_entry->next = entry;
314 entry->prev = pred_entry;
316 /* all loops left have an outedge now */
317 outer_loop = get_irn_loop(block);
319 /* we set loop link to loop to mark it */
320 set_loop_link(loop, loop);
321 loop = get_loop_outer_loop(loop);
322 } while (loop != outer_loop);
325 /* sort interblock edges by execution frequency */
326 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
328 /* run3: remaining edges */
329 for (i = 0; i < edge_count; ++i) {
330 const edge_t *edge = &edges[i];
331 ir_node *block = edge->block;
334 blocksched_entry_t *entry, *pred_entry;
336 /* the block might have been removed already... */
337 if (is_Bad(get_Block_cfgpred(block, pos)))
340 pred_block = get_Block_cfgpred_block(block, pos);
341 entry = (blocksched_entry_t*)get_irn_link(block);
342 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
344 /* is 1 of the blocks already attached to another block? */
345 if (pred_entry->next != NULL || entry->prev != NULL)
348 /* schedule the 2 blocks behind each other */
349 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
350 pred_entry->block, entry->block, edge->execfreq));
351 pred_entry->next = entry;
352 entry->prev = pred_entry;
356 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
358 ir_node *block = entry->block;
359 ir_node *succ = NULL;
360 blocksched_entry_t *succ_entry;
361 const ir_edge_t *edge;
362 double best_succ_execfreq;
364 if (irn_visited_else_mark(block))
369 DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
371 /* put all successors into the worklist */
372 foreach_block_succ(block, edge) {
373 ir_node *succ_block = get_edge_src_irn(edge);
375 if (irn_visited(succ_block))
378 /* we only need to put the first of a series of already connected
379 * blocks into the worklist */
380 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
381 while (succ_entry->prev != NULL) {
382 /* break cycles... */
383 if (succ_entry->prev->block == succ_block) {
384 succ_entry->prev->next = NULL;
385 succ_entry->prev = NULL;
388 succ_entry = succ_entry->prev;
391 if (irn_visited(succ_entry->block))
394 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
395 pdeq_putr(env->worklist, succ_entry->block);
398 if (entry->next != NULL) {
399 pick_block_successor(entry->next, env);
403 DB((dbg, LEVEL_1, "deciding...\n"));
404 best_succ_execfreq = -1;
406 /* no successor yet: pick the successor block with the highest execution
407 * frequency which has no predecessor yet */
409 foreach_block_succ(block, edge) {
410 ir_node *succ_block = get_edge_src_irn(edge);
413 if (irn_visited(succ_block))
416 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
417 if (succ_entry->prev != NULL)
420 execfreq = get_block_execfreq(env->execfreqs, succ_block);
421 if (execfreq > best_succ_execfreq) {
422 best_succ_execfreq = execfreq;
428 DB((dbg, LEVEL_1, "pick from worklist\n"));
431 if (pdeq_empty(env->worklist)) {
432 DB((dbg, LEVEL_1, "worklist empty\n"));
435 succ = (ir_node*)pdeq_getl(env->worklist);
436 } while (irn_visited(succ));
439 succ_entry = (blocksched_entry_t*)get_irn_link(succ);
440 entry->next = succ_entry;
441 succ_entry->prev = entry;
443 pick_block_successor(succ_entry, env);
446 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
448 ir_graph *irg = env->irg;
449 ir_node *startblock = get_irg_start_block(irg);
450 blocksched_entry_t *entry = (blocksched_entry_t*)get_irn_link(startblock);
452 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
453 inc_irg_visited(irg);
455 env->worklist = new_pdeq();
456 pick_block_successor(entry, env);
457 assert(pdeq_empty(env->worklist));
458 del_pdeq(env->worklist);
460 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
465 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
466 int count, struct obstack* obst)
469 ir_node **block_list;
470 blocksched_entry_t *entry;
473 block_list = NEW_ARR_D(ir_node *, obst, count);
474 DB((dbg, LEVEL_1, "Blockschedule:\n"));
476 for (entry = first; entry != NULL; entry = entry->next) {
478 block_list[i++] = entry->block;
479 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
486 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
488 blocksched_env_t env;
490 blocksched_entry_t *start_entry;
491 ir_node **block_list;
497 env.execfreqs = execfreqs;
498 env.edges = NEW_ARR_F(edge_t, 0);
502 assure_loopinfo(irg);
504 // collect edge execution frequencies
505 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
507 (void)be_remove_empty_blocks(irg);
509 if (algo != BLOCKSCHED_NAIV)
510 coalesce_blocks(&env);
512 start_entry = finish_block_schedule(&env);
513 block_list = create_blocksched_array(&env, start_entry, env.blockcount,
514 be_get_be_obst(irg));
516 DEL_ARR_F(env.edges);
517 obstack_free(&obst, NULL);
531 typedef struct ilp_edge_t {
532 ir_node *block; /**< source block */
533 int pos; /**< number of cfg predecessor (target) */
537 typedef struct blocksched_ilp_env_t {
538 blocksched_env_t env;
539 ilp_edge_t *ilpedges;
541 } blocksched_ilp_env_t;
543 typedef struct blocksched_ilp_entry_t {
545 struct blocksched_entry_t *next;
546 struct blocksched_entry_t *prev;
549 } blocksched_ilp_entry_t;
551 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
555 int edgeidx = ARR_LEN(env->ilpedges);
557 snprintf(name, sizeof(name), "edge%d", edgeidx);
561 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
563 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
567 static void collect_egde_frequency_ilp(ir_node *block, void *data)
569 blocksched_ilp_env_t *env = data;
570 ir_graph *irg = env->env.irg;
571 ir_node *startblock = get_irg_start_block(irg);
576 blocksched_ilp_entry_t *entry;
578 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
579 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
581 entry = OALLOC(env->env.obst, blocksched_ilp_entry_t);
582 entry->block = block;
585 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
586 set_irn_link(block, entry);
588 if (block == startblock)
591 arity = get_irn_arity(block);
593 double execfreq = get_block_execfreq(env->env.execfreqs, block);
594 add_ilp_edge(block, 0, execfreq, env);
599 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
600 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
602 for (i = 0; i < arity; ++i) {
606 ir_node *pred_block = get_Block_cfgpred_block(block, i);
608 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
609 edgenum = add_ilp_edge(block, i, execfreq, env);
610 edge = &env->ilpedges[edgenum];
611 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
617 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
619 int edge_count = ARR_LEN(env->ilpedges);
620 be_options_t *options = be_get_irg_options(env->env.irg);
623 /* complete out constraints */
624 for (i = 0; i < edge_count; ++i) {
625 const ilp_edge_t *edge = &env->ilpedges[i];
626 ir_node *block = edge->block;
628 blocksched_ilp_entry_t *entry;
630 /* the block might have been removed already... */
631 if (is_Bad(get_Block_cfgpred(block, 0)))
634 pred = get_Block_cfgpred_block(block, edge->pos);
635 entry = get_irn_link(pred);
637 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
638 pred, block, edge->pos));
639 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
642 lpp_solve_net(env->lpp, options->ilp_server, options->ilp_solver);
643 assert(lpp_is_sol_valid(env->lpp));
645 /* Apply results to edges */
646 for (i = 0; i < edge_count; ++i) {
647 const ilp_edge_t *edge = &env->ilpedges[i];
648 ir_node *block = edge->block;
651 blocksched_entry_t *entry;
652 blocksched_entry_t *pred_entry;
654 /* the block might have been removed already... */
655 if (is_Bad(get_Block_cfgpred(block, 0)))
658 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
662 pred = get_Block_cfgpred_block(block, edge->pos);
663 entry = get_irn_link(block);
664 pred_entry = get_irn_link(pred);
666 assert(entry->prev == NULL && pred_entry->next == NULL);
667 entry->prev = pred_entry;
668 pred_entry->next = entry;
672 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
674 blocksched_ilp_env_t env;
676 blocksched_entry_t *start_entry;
677 ir_node **block_list;
682 env.env.obst = &obst;
683 env.env.execfreqs = execfreqs;
684 env.env.worklist = NULL;
685 env.env.blockcount = 0;
686 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
688 env.lpp = lpp_new("blockschedule", lpp_minimize);
689 lpp_set_time_limit(env.lpp, 20);
690 lpp_set_log(env.lpp, stdout);
692 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
694 (void)be_remove_empty_blocks(irg);
695 coalesce_blocks_ilp(&env);
697 start_entry = finish_block_schedule(&env.env);
698 block_list = create_blocksched_array(&env.env, start_entry,
700 be_get_be_obst(irg));
702 DEL_ARR_F(env.ilpedges);
704 obstack_free(&obst, NULL);
712 * | |\/| |/ _` | | '_ \
713 * | | | | (_| | | | | |
714 * |_| |_|\__,_|_|_| |_|
717 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
718 void be_init_blocksched(void)
720 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
722 lc_opt_add_table(be_grp, be_blocksched_options);
724 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
727 ir_node **be_create_block_schedule(ir_graph *irg)
729 ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
732 case BLOCKSCHED_GREEDY:
733 case BLOCKSCHED_NAIV:
734 return create_block_schedule_greedy(irg, execfreqs);
736 return create_block_schedule_ilp(irg, execfreqs);
739 panic("unknown blocksched algo");