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"
32 #include <libcore/lc_opts.h>
33 #include <libcore/lc_opts_enum.h>
34 #include <libcore/lc_timing.h>
35 #endif /* WITH_LIBCORE */
39 #include <lpp/lpp_net.h>
42 typedef enum _blocksched_algos_t {
43 BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
46 static int algo = BLOCKSCHED_GREEDY;
49 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
50 { "naiv", BLOCKSCHED_NAIV },
51 { "extbb", BLOCKSCHED_EXTBB },
52 { "greedy", BLOCKSCHED_GREEDY },
54 { "ilp", BLOCKSCHED_ILP },
59 static lc_opt_enum_int_var_t algo_var = {
60 &algo, blockschedalgo_items
63 static const lc_opt_table_entry_t be_blocksched_options[] = {
64 LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
71 * / ___|_ __ ___ ___ __| |_ _
72 * | | _| '__/ _ \/ _ \/ _` | | | |
73 * | |_| | | | __/ __/ (_| | |_| |
74 * \____|_| \___|\___|\__,_|\__, |
78 typedef struct _blocksched_entry_t {
80 struct _blocksched_entry_t *next;
81 struct _blocksched_entry_t *prev;
84 typedef struct _edge_t {
85 ir_node *block; /**< source block */
86 int pos; /**< number of cfg predecessor (target) */
87 double execfreq; /**< the frequency */
88 int highest_execfreq; /**< flag that indicates wether this edge is the edge with the highest
89 execfreq pointing away from this block */
92 typedef struct _blocksched_env_t {
95 ir_exec_freq *execfreqs;
99 DEBUG_ONLY(firm_dbg_module_t *dbg;)
103 * Collect cfg frequencies of all edges between blocks.
104 * Also determines edge with highest frequency.
106 static void collect_egde_frequency(ir_node *block, void *data)
108 blocksched_env_t *env = data;
111 blocksched_entry_t *entry;
113 entry = obstack_alloc(env->obst, sizeof(entry[0]));
114 entry->block = block;
117 set_irn_link(block, entry);
119 if (block == get_irg_start_block(env->irg))
122 arity = get_irn_arity(block);
127 edge.execfreq = get_block_execfreq(env->execfreqs, block);
128 edge.highest_execfreq = 1;
129 ARR_APP1(edge_t, env->edges, edge);
132 double highest_execfreq = -1.0;
133 int highest_edge_num = -1;
136 for (i = 0; i < arity; ++i) {
138 ir_node *pred_block = get_Block_cfgpred_block(block, i);
140 execfreq = get_block_execfreq(env->execfreqs, pred_block);
143 edge.execfreq = execfreq;
144 edge.highest_execfreq = 0;
145 ARR_APP1(edge_t, env->edges, edge);
147 if (execfreq > highest_execfreq) {
148 highest_execfreq = execfreq;
149 highest_edge_num = ARR_LEN(env->edges) - 1;
153 if(highest_edge_num >= 0)
154 env->edges[highest_edge_num].highest_execfreq = 1;
158 static int cmp_edges(const void *d1, const void *d2)
160 const edge_t *e1 = d1;
161 const edge_t *e2 = d2;
163 return QSORT_CMP(e2->execfreq, e1->execfreq);
166 static void coalesce_blocks(blocksched_env_t *env)
169 int edge_count = ARR_LEN(env->edges);
171 /* run1: only look at jumps */
172 for (i = 0; i < edge_count; ++i) {
173 const edge_t *edge = &env->edges[i];
174 ir_node *block = edge->block;
176 blocksched_entry_t *entry, *pred_entry;
178 /* the block might have been removed already... */
179 if (is_Bad(get_Block_cfgpred(block, 0)))
182 /* only check edge with highest frequency */
183 if (! edge->highest_execfreq)
186 pred_block = get_Block_cfgpred_block(block, edge->pos);
187 entry = get_irn_link(block);
188 pred_entry = get_irn_link(pred_block);
190 if (pred_entry->next != NULL || entry->prev != NULL)
193 /* only coalesce jumps */
194 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
197 /* schedule the 2 blocks behind each other */
198 DBG((env->dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
199 pred_entry->block, entry->block, edge->execfreq));
200 pred_entry->next = entry;
201 entry->prev = pred_entry;
204 /* run2: remaining edges */
205 for (i = 0; i < edge_count; ++i) {
206 const edge_t *edge = &env->edges[i];
207 ir_node *block = edge->block;
209 blocksched_entry_t *entry, *pred_entry;
211 /* the block might have been removed already... */
212 if (is_Bad(get_Block_cfgpred(block, 0)))
215 pred_block = get_Block_cfgpred_block(block, edge->pos);
216 entry = get_irn_link(block);
217 pred_entry = get_irn_link(pred_block);
219 /* is 1 of the blocks already attached to another block? */
220 if (pred_entry->next != NULL || entry->prev != NULL)
223 /* schedule the 2 blocks behind each other */
224 DBG((env->dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
225 pred_entry->block, entry->block, edge->execfreq));
226 pred_entry->next = entry;
227 entry->prev = pred_entry;
231 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
233 ir_node *block = entry->block;
234 ir_node *succ = NULL;
235 blocksched_entry_t *succ_entry;
236 const ir_edge_t *edge;
237 double best_succ_execfreq;
239 if (irn_visited(block))
243 mark_irn_visited(block);
245 DBG((env->dbg, LEVEL_1, "Pick succ of %+F\n", block));
247 /* put all successors into the worklist */
248 foreach_block_succ(block, edge) {
249 ir_node *succ_block = get_edge_src_irn(edge);
251 if (irn_visited(succ_block))
254 /* we only need to put the first of a series of already connected
255 * blocks into the worklist */
256 succ_entry = get_irn_link(succ_block);
257 while (succ_entry->prev != NULL) {
258 /* break cycles... */
259 if (succ_entry->prev->block == succ_block) {
260 succ_entry->prev->next = NULL;
261 succ_entry->prev = NULL;
264 succ_entry = succ_entry->prev;
267 if (irn_visited(succ_entry->block))
270 DBG((env->dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
271 pdeq_putr(env->worklist, succ_entry->block);
274 if (entry->next != NULL) {
275 pick_block_successor(entry->next, env);
279 DBG((env->dbg, LEVEL_1, "deciding...\n"));
280 best_succ_execfreq = -1;
282 /* no successor yet: pick the successor block with the highest execution
283 * frequency which has no predecessor yet */
285 foreach_block_succ(block, edge) {
286 ir_node *succ_block = get_edge_src_irn(edge);
289 if (irn_visited(succ_block))
292 succ_entry = get_irn_link(succ_block);
293 if (succ_entry->prev != NULL)
296 execfreq = get_block_execfreq(env->execfreqs, succ_block);
297 if (execfreq > best_succ_execfreq) {
298 best_succ_execfreq = execfreq;
304 DBG((env->dbg, LEVEL_1, "pick from worklist\n"));
307 if (pdeq_empty(env->worklist)) {
308 DBG((env->dbg, LEVEL_1, "worklist empty\n"));
311 succ = pdeq_getl(env->worklist);
312 } while (irn_visited(succ));
315 succ_entry = get_irn_link(succ);
316 entry->next = succ_entry;
317 succ_entry->prev = entry;
319 pick_block_successor(succ_entry, env);
322 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
324 ir_graph *irg = env->irg;
325 ir_node *startblock = get_irg_start_block(irg);
326 blocksched_entry_t *entry = get_irn_link(startblock);
328 inc_irg_visited(irg);
330 env->worklist = new_pdeq();
331 pick_block_successor(entry, env);
332 assert(pdeq_empty(env->worklist));
333 del_pdeq(env->worklist);
338 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
339 int count, struct obstack* obst)
342 ir_node **block_list;
343 blocksched_entry_t *entry;
345 block_list = NEW_ARR_D(ir_node *, obst, count);
346 DBG((env->dbg, LEVEL_1, "Blockschedule:\n"));
348 for (entry = first; entry != NULL; entry = entry->next) {
350 block_list[i++] = entry->block;
351 DBG((env->dbg, LEVEL_1, "\t%+F\n", entry->block));
358 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
360 blocksched_env_t env;
362 blocksched_entry_t *start_entry;
363 ir_node **block_list;
369 env.execfreqs = execfreqs;
370 env.edges = NEW_ARR_F(edge_t, 0);
373 FIRM_DBG_REGISTER(env.dbg, "firm.be.blocksched");
375 // collect edge execution frequencies
376 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
378 // sort interblock edges by execution frequency
379 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
381 (void)be_remove_empty_blocks(irg);
383 if (algo != BLOCKSCHED_NAIV)
384 coalesce_blocks(&env);
386 start_entry = finish_block_schedule(&env);
387 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
389 DEL_ARR_F(env.edges);
390 obstack_free(&obst, NULL);
405 typedef struct _ilp_edge_t {
406 ir_node *block; /**< source block */
407 int pos; /**< number of cfg predecessor (target) */
411 typedef struct _blocksched_ilp_env_t {
412 blocksched_env_t env;
413 ilp_edge_t *ilpedges;
415 } blocksched_ilp_env_t;
417 typedef struct _blocksched_ilp_entry_t {
419 struct _blocksched_entry_t *next;
420 struct _blocksched_entry_t *prev;
423 } blocksched_ilp_entry_t;
425 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
429 int edgeidx = ARR_LEN(env->ilpedges);
431 snprintf(name, sizeof(name), "edge%d", edgeidx);
435 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
437 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
441 static void collect_egde_frequency_ilp(ir_node *block, void *data)
443 blocksched_ilp_env_t *env = data;
444 ir_graph *irg = env->env.irg;
445 ir_node *startblock = get_irg_start_block(irg);
450 blocksched_ilp_entry_t *entry;
452 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
453 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
455 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
456 entry->block = block;
459 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
460 set_irn_link(block, entry);
462 if (block == startblock)
465 arity = get_irn_arity(block);
467 double execfreq = get_block_execfreq(env->env.execfreqs, block);
468 add_ilp_edge(block, 0, execfreq, env);
472 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
474 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
475 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
477 for (i = 0; i < arity; ++i) {
481 ir_node *pred_block = get_Block_cfgpred_block(block, i);
483 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
484 edgenum = add_ilp_edge(block, i, execfreq, env);
485 edge = &env->ilpedges[edgenum];
486 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
492 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
495 int edge_count = ARR_LEN(env->ilpedges);
497 /* complete out constraints */
498 for(i = 0; i < edge_count; ++i) {
499 const ilp_edge_t *edge = &env->ilpedges[i];
500 ir_node *block = edge->block;
502 blocksched_ilp_entry_t *entry;
504 /* the block might have been removed already... */
505 if (is_Bad(get_Block_cfgpred(block, 0)))
508 pred = get_Block_cfgpred_block(block, edge->pos);
509 entry = get_irn_link(pred);
511 DBG((env->env.dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
512 pred, block, edge->pos));
513 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
520 lpp_dump(env->lpp, "lpp.out");
521 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
522 f = fopen(fname, "w");
523 lpp_dump_plain(env->lpp, f);
528 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
529 lpp_solve_net(env->lpp, "i44pc52", "cplex");
530 assert(lpp_is_sol_valid(env->lpp));
532 /* Apply results to edges */
533 for (i = 0; i < edge_count; ++i) {
534 const ilp_edge_t *edge = &env->ilpedges[i];
535 ir_node *block = edge->block;
538 blocksched_entry_t *entry;
539 blocksched_entry_t *pred_entry;
541 /* the block might have been removed already... */
542 if (is_Bad(get_Block_cfgpred(block, 0)))
545 is_jump = lpp_get_var_sol(env->lpp, edge->ilpvar);
549 pred = get_Block_cfgpred_block(block, edge->pos);
550 entry = get_irn_link(block);
551 pred_entry = get_irn_link(pred);
553 assert(entry->prev == NULL && pred_entry->next == NULL);
554 entry->prev = pred_entry;
555 pred_entry->next = entry;
559 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
561 blocksched_ilp_env_t env;
563 blocksched_entry_t *start_entry;
564 ir_node **block_list;
569 env.env.obst = &obst;
570 env.env.execfreqs = execfreqs;
571 env.env.worklist = NULL;
572 env.env.blockcount = 0;
573 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
574 FIRM_DBG_REGISTER(env.env.dbg, "firm.be.blocksched");
576 env.lpp = new_lpp("blockschedule", lpp_minimize);
577 lpp_set_time_limit(env.lpp, 20);
578 lpp_set_log(env.lpp, stdout);
580 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
582 (void)be_remove_empty_blocks(irg);
583 coalesce_blocks_ilp(&env);
585 start_entry = finish_block_schedule(&env.env);
586 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
588 DEL_ARR_F(env.ilpedges);
590 obstack_free(&obst, NULL);
594 #endif /* WITH_ILP */
598 * | ____|_ _| |_| __ )| __ )
599 * | _| \ \/ / __| _ \| _ \
600 * | |___ > <| |_| |_) | |_) |
601 * |_____/_/\_\\__|____/|____/
605 /** A simple forward single linked list. */
607 ir_node *start; /**< start of the list */
608 ir_node *end; /**< last block in the list */
609 unsigned n_blks; /**< number of blocks in the list */
612 static void add_block(anchor *list, ir_node *block) {
613 if (list->start == NULL) {
618 set_irn_link(list->end, block);
625 static void create_block_list(ir_node *leader_block, anchor *list) {
627 const ir_edge_t *edge;
628 ir_node *block = NULL;
629 ir_extblk *extbb = get_Block_extbb(leader_block);
631 if (extbb_visited(extbb))
633 mark_extbb_visited(extbb);
635 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
636 block = get_extbb_block(extbb, i);
637 add_block(list, block);
640 assert(block != NULL);
642 /* pick successor extbbs */
643 foreach_block_succ(block, edge) {
644 ir_node *succ = get_edge_src_irn(edge);
645 create_block_list(succ, list);
648 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
649 block = get_extbb_block(extbb, i);
651 foreach_block_succ(block, edge) {
652 ir_node *succ = get_edge_src_irn(edge);
653 create_block_list(succ, list);
658 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
661 * Calculates a block schedule. The schedule is stored as a linked
662 * list starting at the start_block of the irg.
664 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
667 ir_node **blk_list, *b, *n;
670 /* schedule extended basic blocks */
671 compute_extbb_execfreqs(irg, execfreqs);
672 //compute_extbb(irg);
677 inc_irg_block_visited(irg);
678 create_block_list(get_irg_start_block(irg), &list);
680 /** create an array, so we can go forward and backward */
681 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
683 for (i = 0, b = list.start; b; b = n, ++i) {
694 * | |\/| |/ _` | | '_ \
695 * | | | | (_| | | | | |
696 * |_| |_|\__,_|_|_| |_|
699 void be_init_blocksched(void)
701 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
702 lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
704 lc_opt_add_table(blocksched_grp, be_blocksched_options);
707 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
709 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
712 case BLOCKSCHED_GREEDY:
713 case BLOCKSCHED_NAIV:
714 return create_block_schedule_greedy(irg, execfreqs);
715 case BLOCKSCHED_EXTBB:
716 return create_extbb_block_schedule(irg, execfreqs);
719 return create_block_schedule_ilp(irg, execfreqs);
720 #endif /* WITH_ILP */
723 assert(0 && "unknown blocksched algo");