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;
130 static blocksched_entry_t* get_blocksched_entry(const ir_node *block)
132 return (blocksched_entry_t*)get_irn_link(block);
136 * Collect cfg frequencies of all edges between blocks.
137 * Also determines edge with highest frequency.
139 static void collect_egde_frequency(ir_node *block, void *data)
141 blocksched_env_t *env = (blocksched_env_t*)data;
144 blocksched_entry_t *entry;
147 memset(&edge, 0, sizeof(edge));
149 entry = OALLOCZ(env->obst, blocksched_entry_t);
150 entry->block = block;
151 set_irn_link(block, entry);
153 loop = get_irn_loop(block);
155 arity = get_Block_n_cfgpreds(block);
158 /* must be the start block (or end-block for endless loops),
159 * everything else is dead code and should be removed by now */
160 assert(block == get_irg_start_block(env->irg)
161 || block == get_irg_end_block(env->irg));
162 /* nothing to do here */
164 } else if (arity == 1) {
165 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
166 ir_loop *pred_loop = get_irn_loop(pred_block);
167 float freq = (float)get_block_execfreq(env->execfreqs, block);
169 /* is it an edge leaving a loop */
170 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
171 float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
172 edge.outedge_penalty_freq = -(pred_freq - freq);
177 edge.execfreq = freq;
178 edge.highest_execfreq = 1;
179 ARR_APP1(edge_t, env->edges, edge);
182 double highest_execfreq = -1.0;
183 int highest_edge_num = -1;
186 for (i = 0; i < arity; ++i) {
188 ir_node *pred_block = get_Block_cfgpred_block(block, i);
190 execfreq = get_block_execfreq(env->execfreqs, pred_block);
193 edge.execfreq = execfreq;
194 edge.highest_execfreq = 0;
195 ARR_APP1(edge_t, env->edges, edge);
197 if (execfreq > highest_execfreq) {
198 highest_execfreq = execfreq;
199 highest_edge_num = ARR_LEN(env->edges) - 1;
203 if (highest_edge_num >= 0)
204 env->edges[highest_edge_num].highest_execfreq = 1;
208 static int cmp_edges_base(const edge_t *e1, const edge_t *e2)
210 long nr1 = get_irn_node_nr(e1->block);
211 long nr2 = get_irn_node_nr(e2->block);
214 } else if (nr1 > nr2) {
217 if (e1->pos < e2->pos) {
219 } else if (e1->pos > e2->pos) {
227 static int cmp_edges(const void *d1, const void *d2)
229 const edge_t *e1 = (const edge_t*)d1;
230 const edge_t *e2 = (const edge_t*)d2;
231 double freq1 = e1->execfreq;
232 double freq2 = e2->execfreq;
235 } else if (freq1 > freq2) {
238 return cmp_edges_base(e1, e2);
242 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
244 const edge_t *e1 = (const edge_t*)d1;
245 const edge_t *e2 = (const edge_t*)d2;
246 double pen1 = e1->outedge_penalty_freq;
247 double pen2 = e2->outedge_penalty_freq;
250 } else if (pen1 < pen2) {
253 return cmp_edges_base(e1, e2);
257 static void clear_loop_links(ir_loop *loop)
261 set_loop_link(loop, NULL);
262 n = get_loop_n_elements(loop);
263 for (i = 0; i < n; ++i) {
264 loop_element elem = get_loop_element(loop, i);
265 if (*elem.kind == k_ir_loop) {
266 clear_loop_links(elem.son);
271 static void coalesce_blocks(blocksched_env_t *env)
274 int edge_count = ARR_LEN(env->edges);
275 edge_t *edges = env->edges;
277 /* sort interblock edges by execution frequency */
278 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
280 /* run1: only look at jumps */
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;
288 /* only check edge with highest frequency */
289 if (! edge->highest_execfreq)
292 /* the block might have been removed already... */
293 if (is_Bad(get_Block_cfgpred(block, 0)))
296 pred_block = get_Block_cfgpred_block(block, pos);
297 entry = (blocksched_entry_t*)get_irn_link(block);
298 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
300 if (pred_entry->next != NULL || entry->prev != NULL)
303 /* only coalesce jumps */
304 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
307 /* schedule the 2 blocks behind each other */
308 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
309 pred_entry->block, entry->block, edge->execfreq));
310 pred_entry->next = entry;
311 entry->prev = pred_entry;
314 /* run2: pick loop fallthroughs */
315 clear_loop_links(get_irg_loop(env->irg));
317 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
318 for (i = 0; i < edge_count; ++i) {
319 const edge_t *edge = &edges[i];
320 ir_node *block = edge->block;
323 blocksched_entry_t *entry, *pred_entry;
327 /* already seen all loop outedges? */
328 if (edge->outedge_penalty_freq == 0)
331 /* the block might have been removed already... */
332 if (is_Bad(get_Block_cfgpred(block, pos)))
335 pred_block = get_Block_cfgpred_block(block, pos);
336 entry = (blocksched_entry_t*)get_irn_link(block);
337 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
339 if (pred_entry->next != NULL || entry->prev != NULL)
342 /* we want at most 1 outedge fallthrough per loop */
343 loop = get_irn_loop(pred_block);
344 if (get_loop_link(loop) != NULL)
347 /* schedule the 2 blocks behind each other */
348 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
349 pred_entry->block, entry->block, edge->execfreq));
350 pred_entry->next = entry;
351 entry->prev = pred_entry;
353 /* all loops left have an outedge now */
354 outer_loop = get_irn_loop(block);
356 /* we set loop link to loop to mark it */
357 set_loop_link(loop, loop);
358 loop = get_loop_outer_loop(loop);
359 } while (loop != outer_loop);
362 /* sort interblock edges by execution frequency */
363 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
365 /* run3: remaining edges */
366 for (i = 0; i < edge_count; ++i) {
367 const edge_t *edge = &edges[i];
368 ir_node *block = edge->block;
371 blocksched_entry_t *entry, *pred_entry;
373 /* the block might have been removed already... */
374 if (is_Bad(get_Block_cfgpred(block, pos)))
377 pred_block = get_Block_cfgpred_block(block, pos);
378 entry = (blocksched_entry_t*)get_irn_link(block);
379 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
381 /* is 1 of the blocks already attached to another block? */
382 if (pred_entry->next != NULL || entry->prev != NULL)
385 /* schedule the 2 blocks behind each other */
386 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
387 pred_entry->block, entry->block, edge->execfreq));
388 pred_entry->next = entry;
389 entry->prev = pred_entry;
393 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
395 ir_node *block = entry->block;
396 ir_node *succ = NULL;
397 blocksched_entry_t *succ_entry;
398 const ir_edge_t *edge;
399 double best_succ_execfreq;
401 if (irn_visited_else_mark(block))
406 DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
408 /* put all successors into the worklist */
409 foreach_block_succ(block, edge) {
410 ir_node *succ_block = get_edge_src_irn(edge);
412 if (irn_visited(succ_block))
415 /* we only need to put the first of a series of already connected
416 * blocks into the worklist */
417 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
418 while (succ_entry->prev != NULL) {
419 /* break cycles... */
420 if (succ_entry->prev->block == succ_block) {
421 succ_entry->prev->next = NULL;
422 succ_entry->prev = NULL;
425 succ_entry = succ_entry->prev;
428 if (irn_visited(succ_entry->block))
431 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
432 pdeq_putr(env->worklist, succ_entry->block);
435 if (entry->next != NULL) {
436 pick_block_successor(entry->next, env);
440 DB((dbg, LEVEL_1, "deciding...\n"));
441 best_succ_execfreq = -1;
443 /* no successor yet: pick the successor block with the highest execution
444 * frequency which has no predecessor yet */
446 foreach_block_succ(block, edge) {
447 ir_node *succ_block = get_edge_src_irn(edge);
450 if (irn_visited(succ_block))
453 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
454 if (succ_entry->prev != NULL)
457 execfreq = get_block_execfreq(env->execfreqs, succ_block);
458 if (execfreq > best_succ_execfreq) {
459 best_succ_execfreq = execfreq;
465 DB((dbg, LEVEL_1, "pick from worklist\n"));
468 if (pdeq_empty(env->worklist)) {
469 DB((dbg, LEVEL_1, "worklist empty\n"));
472 succ = (ir_node*)pdeq_getl(env->worklist);
473 } while (irn_visited(succ));
476 succ_entry = (blocksched_entry_t*)get_irn_link(succ);
477 entry->next = succ_entry;
478 succ_entry->prev = entry;
480 pick_block_successor(succ_entry, env);
483 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
485 ir_graph *irg = env->irg;
486 ir_node *startblock = get_irg_start_block(irg);
487 blocksched_entry_t *entry = (blocksched_entry_t*)get_irn_link(startblock);
489 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
490 inc_irg_visited(irg);
492 env->worklist = new_pdeq();
493 pick_block_successor(entry, env);
494 assert(pdeq_empty(env->worklist));
495 del_pdeq(env->worklist);
497 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
502 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
503 int count, struct obstack* obst)
506 ir_node **block_list;
507 blocksched_entry_t *entry;
510 block_list = NEW_ARR_D(ir_node *, obst, count);
511 DB((dbg, LEVEL_1, "Blockschedule:\n"));
513 for (entry = first; entry != NULL; entry = entry->next) {
515 block_list[i++] = entry->block;
516 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
523 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
525 blocksched_env_t env;
527 blocksched_entry_t *start_entry;
528 ir_node **block_list;
534 env.execfreqs = execfreqs;
535 env.edges = NEW_ARR_F(edge_t, 0);
539 assure_loopinfo(irg);
541 // collect edge execution frequencies
542 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
544 (void)be_remove_empty_blocks(irg);
546 if (algo != BLOCKSCHED_NAIV)
547 coalesce_blocks(&env);
549 start_entry = finish_block_schedule(&env);
550 block_list = create_blocksched_array(&env, start_entry, env.blockcount,
551 be_get_be_obst(irg));
553 DEL_ARR_F(env.edges);
554 obstack_free(&obst, NULL);
568 typedef struct ilp_edge_t {
569 ir_node *block; /**< source block */
570 int pos; /**< number of cfg predecessor (target) */
574 typedef struct blocksched_ilp_env_t {
575 blocksched_env_t env;
576 ilp_edge_t *ilpedges;
578 } blocksched_ilp_env_t;
580 typedef struct blocksched_ilp_entry_t {
582 struct blocksched_entry_t *next;
583 struct blocksched_entry_t *prev;
586 } blocksched_ilp_entry_t;
588 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
592 int edgeidx = ARR_LEN(env->ilpedges);
594 snprintf(name, sizeof(name), "edge%d", edgeidx);
598 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
600 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
604 static void collect_egde_frequency_ilp(ir_node *block, void *data)
606 blocksched_ilp_env_t *env = (blocksched_ilp_env_t*)data;
607 ir_graph *irg = env->env.irg;
608 ir_node *startblock = get_irg_start_block(irg);
612 blocksched_ilp_entry_t *entry;
614 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
615 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
617 entry = OALLOC(env->env.obst, blocksched_ilp_entry_t);
618 entry->block = block;
621 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
622 set_irn_link(block, entry);
624 if (block == startblock)
627 arity = get_irn_arity(block);
629 double execfreq = get_block_execfreq(env->env.execfreqs, block);
630 add_ilp_edge(block, 0, execfreq, env);
636 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
637 cst_idx = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
639 for (i = 0; i < arity; ++i) {
643 ir_node *pred_block = get_Block_cfgpred_block(block, i);
645 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
646 edgenum = add_ilp_edge(block, i, execfreq, env);
647 edge = &env->ilpedges[edgenum];
648 lpp_set_factor_fast(env->lpp, cst_idx, edge->ilpvar, 1.0);
653 static blocksched_ilp_entry_t *get_blocksched_ilp_entry(const ir_node *block)
655 return (blocksched_ilp_entry_t*)get_irn_link(block);
658 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
660 int edge_count = ARR_LEN(env->ilpedges);
663 /* complete out constraints */
664 for (i = 0; i < edge_count; ++i) {
665 const ilp_edge_t *edge = &env->ilpedges[i];
666 ir_node *block = edge->block;
668 blocksched_ilp_entry_t *entry;
670 /* the block might have been removed already... */
671 if (is_Bad(get_Block_cfgpred(block, 0)))
674 pred = get_Block_cfgpred_block(block, edge->pos);
675 entry = get_blocksched_ilp_entry(pred);
677 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
678 pred, block, edge->pos));
679 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
682 lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver);
683 assert(lpp_is_sol_valid(env->lpp));
685 /* Apply results to edges */
686 for (i = 0; i < edge_count; ++i) {
687 const ilp_edge_t *edge = &env->ilpedges[i];
688 ir_node *block = edge->block;
691 blocksched_entry_t *entry;
692 blocksched_entry_t *pred_entry;
694 /* the block might have been removed already... */
695 if (is_Bad(get_Block_cfgpred(block, 0)))
698 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
702 pred = get_Block_cfgpred_block(block, edge->pos);
703 entry = get_blocksched_entry(block);
704 pred_entry = get_blocksched_entry(pred);
706 assert(entry->prev == NULL && pred_entry->next == NULL);
707 entry->prev = pred_entry;
708 pred_entry->next = entry;
712 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
714 blocksched_ilp_env_t env;
716 blocksched_entry_t *start_entry;
717 ir_node **block_list;
722 env.env.obst = &obst;
723 env.env.execfreqs = execfreqs;
724 env.env.worklist = NULL;
725 env.env.blockcount = 0;
726 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
728 env.lpp = lpp_new("blockschedule", lpp_minimize);
729 lpp_set_time_limit(env.lpp, 20);
730 lpp_set_log(env.lpp, stdout);
732 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
734 (void)be_remove_empty_blocks(irg);
735 coalesce_blocks_ilp(&env);
737 start_entry = finish_block_schedule(&env.env);
738 block_list = create_blocksched_array(&env.env, start_entry,
740 be_get_be_obst(irg));
742 DEL_ARR_F(env.ilpedges);
744 obstack_free(&obst, NULL);
752 * | |\/| |/ _` | | '_ \
753 * | | | | (_| | | | | |
754 * |_| |_|\__,_|_|_| |_|
757 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
758 void be_init_blocksched(void)
760 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
762 lc_opt_add_table(be_grp, be_blocksched_options);
764 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
767 ir_node **be_create_block_schedule(ir_graph *irg)
769 ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
772 case BLOCKSCHED_GREEDY:
773 case BLOCKSCHED_NAIV:
774 return create_block_schedule_greedy(irg, execfreqs);
776 return create_block_schedule_ilp(irg, execfreqs);
779 panic("unknown blocksched algo");