2 * Author: Matthias Braun, Christoph Mallon
4 * Copyright: (c) Universitaet Karlsruhe
5 * License: This file is protected by GPL - GNU GENERAL PUBLIC LICENSE.
10 #endif /* HAVE_CONFIG_H */
12 #include "beblocksched.h"
21 #include "irgraph_t.h"
30 #include <libcore/lc_opts.h>
31 #include <libcore/lc_opts_enum.h>
32 #include <libcore/lc_timing.h>
33 #endif /* WITH_LIBCORE */
37 #include <lpp/lpp_net.h>
40 typedef enum _blocksched_algos_t {
41 BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
44 static int algo = BLOCKSCHED_GREEDY;
46 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
47 { "naiv", BLOCKSCHED_NAIV },
48 { "extbb", BLOCKSCHED_EXTBB },
49 { "greedy", BLOCKSCHED_GREEDY },
51 { "ilp", BLOCKSCHED_ILP },
56 static lc_opt_enum_int_var_t algo_var = {
57 &algo, blockschedalgo_items
60 static const lc_opt_table_entry_t be_blocksched_options[] = {
61 LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
67 * / ___|_ __ ___ ___ __| |_ _
68 * | | _| '__/ _ \/ _ \/ _` | | | |
69 * | |_| | | | __/ __/ (_| | |_| |
70 * \____|_| \___|\___|\__,_|\__, |
74 typedef struct _blocksched_entry_t {
76 struct _blocksched_entry_t *next;
77 struct _blocksched_entry_t *prev;
80 typedef struct _edge_t {
81 ir_node *block; /**< source block */
82 int pos; /**< number of cfg predecessor (target) */
83 double execfreq; /**< the frequency */
84 int highest_execfreq; /**< flag that indicates wether this edge is the edge with the highest
85 execfreq pointing away from this block */
88 typedef struct _blocksched_env_t {
91 ir_exec_freq *execfreqs;
95 DEBUG_ONLY(firm_dbg_module_t *dbg;)
99 * Collect cfg frequencies of all edges between blocks.
100 * Also determines edge with highest frequency.
102 static void collect_egde_frequency(ir_node *block, void *data)
104 blocksched_env_t *env = data;
105 ir_graph *irg = env->irg;
106 ir_node *startblock = get_irg_start_block(irg);
109 blocksched_entry_t *entry;
111 entry = obstack_alloc(env->obst, sizeof(entry[0]));
112 entry->block = block;
115 set_irn_link(block, entry);
117 if (block == startblock)
120 arity = get_irn_arity(block);
125 edge.execfreq = get_block_execfreq(env->execfreqs, block);
126 edge.highest_execfreq = 1;
127 ARR_APP1(edge_t, env->edges, edge);
130 double highest_execfreq = -1;
131 int highest_edge_num;
134 for (i = 0; i < arity; ++i) {
136 ir_node *pred_block = get_Block_cfgpred_block(block, i);
138 execfreq = get_block_execfreq(env->execfreqs, pred_block);
141 edge.execfreq = execfreq;
142 edge.highest_execfreq = 0;
143 ARR_APP1(edge_t, env->edges, edge);
145 if (execfreq > highest_execfreq) {
146 highest_execfreq = execfreq;
147 highest_edge_num = ARR_LEN(env->edges) - 1;
151 env->edges[highest_edge_num].highest_execfreq = 1;
155 static int cmp_edges(const void *d1, const void *d2)
157 const edge_t *e1 = d1;
158 const edge_t *e2 = d2;
160 return QSORT_CMP(e2->execfreq, e1->execfreq);
163 static void coalesce_blocks(blocksched_env_t *env)
166 int edge_count = ARR_LEN(env->edges);
168 /* run1: only look at jumps */
169 for (i = 0; i < edge_count; ++i) {
170 const edge_t *edge = &env->edges[i];
171 ir_node *block = edge->block;
173 blocksched_entry_t *entry, *pred_entry;
175 /* the block might have been removed already... */
176 if (is_Bad(get_Block_cfgpred(block, 0)))
179 /* only check edge with highest frequency */
180 if (! edge->highest_execfreq)
183 pred_block = get_Block_cfgpred_block(block, edge->pos);
184 entry = get_irn_link(block);
185 pred_entry = get_irn_link(pred_block);
187 if (pred_entry->next != NULL || entry->prev != NULL)
190 /* only coalesce jumps */
191 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
194 /* schedule the 2 blocks behind each other */
195 DBG((env->dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
196 pred_entry->block, entry->block, edge->execfreq));
197 pred_entry->next = entry;
198 entry->prev = pred_entry;
201 /* run2: remaining edges */
202 for (i = 0; i < edge_count; ++i) {
203 const edge_t *edge = &env->edges[i];
204 ir_node *block = edge->block;
206 blocksched_entry_t *entry, *pred_entry;
208 /* the block might have been removed already... */
209 if (is_Bad(get_Block_cfgpred(block, 0)))
212 pred_block = get_Block_cfgpred_block(block, edge->pos);
213 entry = get_irn_link(block);
214 pred_entry = get_irn_link(pred_block);
216 /* TODO: what's this check for? */
217 if (pred_entry->next != NULL || entry->prev != NULL)
220 /* schedule the 2 blocks behind each other */
221 DBG((env->dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
222 pred_entry->block, entry->block, edge->execfreq));
223 pred_entry->next = entry;
224 entry->prev = pred_entry;
228 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
230 ir_node *block = entry->block;
231 ir_node *succ = NULL;
232 blocksched_entry_t *succ_entry;
233 const ir_edge_t *edge;
234 double best_succ_execfreq;
236 if (irn_visited(block))
240 mark_irn_visited(block);
242 DBG((env->dbg, LEVEL_1, "Pick succ of %+F\n", block));
244 /* put all successors into the worklist */
245 foreach_block_succ(block, edge) {
246 ir_node *succ_block = get_edge_src_irn(edge);
248 if (irn_visited(succ_block))
251 /* we only need to put the first of a series of already connected
252 * blocks into the worklist */
253 succ_entry = get_irn_link(succ_block);
254 while (succ_entry->prev != NULL) {
255 /* break cycles... */
256 if (succ_entry->prev->block == succ_block) {
257 succ_entry->prev->next = NULL;
258 succ_entry->prev = NULL;
261 succ_entry = succ_entry->prev;
264 if (irn_visited(succ_entry->block))
267 DBG((env->dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
268 pdeq_putr(env->worklist, succ_entry->block);
271 if (entry->next != NULL) {
272 pick_block_successor(entry->next, env);
276 DBG((env->dbg, LEVEL_1, "deciding...\n"));
277 best_succ_execfreq = -1;
279 /* no successor yet: pick the successor block with the highest execution
280 * frequency which has no predecessor yet */
282 foreach_block_succ(block, edge) {
283 ir_node *succ_block = get_edge_src_irn(edge);
286 if (irn_visited(succ_block))
289 succ_entry = get_irn_link(succ_block);
290 if (succ_entry->prev != NULL)
293 execfreq = get_block_execfreq(env->execfreqs, succ_block);
294 if (execfreq > best_succ_execfreq) {
295 best_succ_execfreq = execfreq;
301 DBG((env->dbg, LEVEL_1, "pick from worklist\n"));
304 if (pdeq_empty(env->worklist)) {
305 DBG((env->dbg, LEVEL_1, "worklist empty\n"));
308 succ = pdeq_getl(env->worklist);
309 } while (irn_visited(succ));
312 succ_entry = get_irn_link(succ);
313 entry->next = succ_entry;
314 succ_entry->prev = entry;
316 pick_block_successor(succ_entry, env);
319 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
321 ir_graph *irg = env->irg;
322 ir_node *startblock = get_irg_start_block(irg);
323 blocksched_entry_t *entry = get_irn_link(startblock);
325 inc_irg_visited(irg);
327 env->worklist = new_pdeq();
328 pick_block_successor(entry, env);
329 assert(pdeq_empty(env->worklist));
330 del_pdeq(env->worklist);
335 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
336 int count, struct obstack* obst)
339 ir_node **block_list;
340 blocksched_entry_t *entry;
342 block_list = NEW_ARR_D(ir_node *, obst, count);
343 DBG((env->dbg, LEVEL_1, "Blockschedule:\n"));
345 for (entry = first; entry != NULL; entry = entry->next) {
347 block_list[i++] = entry->block;
348 DBG((env->dbg, LEVEL_1, "\t%+F\n", entry->block));
355 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
357 blocksched_env_t env;
359 blocksched_entry_t *start_entry;
360 ir_node **block_list;
366 env.execfreqs = execfreqs;
367 env.edges = NEW_ARR_F(edge_t, 0);
370 FIRM_DBG_REGISTER(env.dbg, "firm.be.blocksched");
372 // collect edge execution frequencies
373 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
375 // sort interblock edges by execution frequency
376 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
378 be_remove_empty_blocks(irg);
380 if (algo != BLOCKSCHED_NAIV)
381 coalesce_blocks(&env);
383 start_entry = finish_block_schedule(&env);
384 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
386 DEL_ARR_F(env.edges);
387 obstack_free(&obst, NULL);
402 typedef struct _ilp_edge_t {
403 ir_node *block; /**< source block */
404 int pos; /**< number of cfg predecessor (target) */
408 typedef struct _blocksched_ilp_env_t {
409 blocksched_env_t env;
410 ilp_edge_t *ilpedges;
412 } blocksched_ilp_env_t;
414 typedef struct _blocksched_ilp_entry_t {
416 struct _blocksched_entry_t *next;
417 struct _blocksched_entry_t *prev;
420 } blocksched_ilp_entry_t;
422 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
426 int edgeidx = ARR_LEN(env->ilpedges);
428 snprintf(name, sizeof(name), "edge%d", edgeidx);
432 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
434 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
438 static void collect_egde_frequency_ilp(ir_node *block, void *data)
440 blocksched_ilp_env_t *env = data;
441 ir_graph *irg = env->env.irg;
442 ir_node *startblock = get_irg_start_block(irg);
447 blocksched_ilp_entry_t *entry;
449 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
450 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
452 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
453 entry->block = block;
456 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
457 set_irn_link(block, entry);
459 if (block == startblock)
462 arity = get_irn_arity(block);
464 double execfreq = get_block_execfreq(env->env.execfreqs, block);
465 add_ilp_edge(block, 0, execfreq, env);
469 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
471 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
472 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
474 for (i = 0; i < arity; ++i) {
478 ir_node *pred_block = get_Block_cfgpred_block(block, i);
480 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
481 edgenum = add_ilp_edge(block, i, execfreq, env);
482 edge = &env->ilpedges[edgenum];
483 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
489 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
492 int edge_count = ARR_LEN(env->ilpedges);
496 /* complete out constraints */
497 for(i = 0; i < edge_count; ++i) {
498 const ilp_edge_t *edge = &env->ilpedges[i];
499 ir_node *block = edge->block;
501 blocksched_ilp_entry_t *entry;
503 /* the block might have been removed already... */
504 if (is_Bad(get_Block_cfgpred(block, 0)))
507 pred = get_Block_cfgpred_block(block, edge->pos);
508 entry = get_irn_link(pred);
510 DBG((env->env.dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
511 pred, block, edge->pos));
512 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
515 lpp_dump(env->lpp, "lpp.out");
516 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
517 f = fopen(fname, "w");
518 lpp_dump_plain(env->lpp, f);
520 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
521 lpp_solve_net(env->lpp, "i44pc52", "cplex");
522 assert(lpp_is_sol_valid(env->lpp));
524 /* Apply results to edges */
525 for (i = 0; i < edge_count; ++i) {
526 const ilp_edge_t *edge = &env->ilpedges[i];
527 ir_node *block = edge->block;
530 blocksched_entry_t *entry;
531 blocksched_entry_t *pred_entry;
533 /* the block might have been removed already... */
534 if (is_Bad(get_Block_cfgpred(block, 0)))
537 is_jump = lpp_get_var_sol(env->lpp, edge->ilpvar);
541 pred = get_Block_cfgpred_block(block, edge->pos);
542 entry = get_irn_link(block);
543 pred_entry = get_irn_link(pred);
545 assert(entry->prev == NULL && pred_entry->next == NULL);
546 entry->prev = pred_entry;
547 pred_entry->next = entry;
551 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
553 blocksched_ilp_env_t env;
555 blocksched_entry_t *start_entry;
556 ir_node **block_list;
561 env.env.obst = &obst;
562 env.env.execfreqs = execfreqs;
563 env.env.worklist = NULL;
564 env.env.blockcount = 0;
565 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
566 FIRM_DBG_REGISTER(env.env.dbg, "firm.be.blocksched");
568 env.lpp = new_lpp("blockschedule", lpp_minimize);
569 lpp_set_time_limit(env.lpp, 20);
570 lpp_set_log(env.lpp, stdout);
572 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
574 be_remove_empty_blocks(irg);
575 coalesce_blocks_ilp(&env);
577 start_entry = finish_block_schedule(&env.env);
578 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
580 DEL_ARR_F(env.ilpedges);
582 obstack_free(&obst, NULL);
586 #endif /* WITH_ILP */
590 * | ____|_ _| |_| __ )| __ )
591 * | _| \ \/ / __| _ \| _ \
592 * | |___ > <| |_| |_) | |_) |
593 * |_____/_/\_\\__|____/|____/
597 /** A simple forward single linked list. */
599 ir_node *start; /**< start of the list */
600 ir_node *end; /**< last block in the list */
601 unsigned n_blks; /**< number of blocks in the list */
604 static void add_block(anchor *list, ir_node *block) {
605 if (list->start == NULL) {
610 set_irn_link(list->end, block);
617 static void create_block_list(ir_node *leader_block, anchor *list) {
619 const ir_edge_t *edge;
620 ir_node *block = NULL;
621 ir_extblk *extbb = get_Block_extbb(leader_block);
623 if (extbb_visited(extbb))
625 mark_extbb_visited(extbb);
627 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
628 block = get_extbb_block(extbb, i);
629 add_block(list, block);
632 assert(block != NULL);
634 /* pick successor extbbs */
635 foreach_block_succ(block, edge) {
636 ir_node *succ = get_edge_src_irn(edge);
637 create_block_list(succ, list);
640 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
641 block = get_extbb_block(extbb, i);
643 foreach_block_succ(block, edge) {
644 ir_node *succ = get_edge_src_irn(edge);
645 create_block_list(succ, list);
650 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
653 * Calculates a block schedule. The schedule is stored as a linked
654 * list starting at the start_block of the irg.
656 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
659 ir_node **blk_list, *b, *n;
662 /* schedule extended basic blocks */
663 compute_extbb_execfreqs(irg, execfreqs);
664 //compute_extbb(irg);
669 inc_irg_block_visited(irg);
670 create_block_list(get_irg_start_block(irg), &list);
672 /** create an array, so we can go forward and backward */
673 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
675 for (i = 0, b = list.start; b; b = n, ++i) {
686 * | |\/| |/ _` | | '_ \
687 * | | | | (_| | | | | |
688 * |_| |_|\__,_|_|_| |_|
693 void be_block_schedule_register_options(lc_opt_entry_t *grp)
695 static int run_once = 0;
696 lc_opt_entry_t *blocksched_grp;
702 blocksched_grp = lc_opt_get_grp(grp, "blocksched");
704 lc_opt_add_table(blocksched_grp, be_blocksched_options);
706 #endif /* WITH_LIBCORE */
708 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
711 case BLOCKSCHED_GREEDY:
712 case BLOCKSCHED_NAIV:
713 return create_block_schedule_greedy(irg, execfreqs);
714 case BLOCKSCHED_EXTBB:
715 return create_extbb_block_schedule(irg, execfreqs);
718 return create_block_schedule_ilp(irg, execfreqs);
719 #endif /* WITH_ILP */
722 assert(0 && "unknown blocksched algo");