2 * Author: Matthias Braun, Christoph Mallon
4 * Copyright: (c) Universitaet Karlsruhe
5 * License: This file is protected by GPL - GNU GENERAL PUBLIC LICENSE.
12 #include "beblocksched.h"
21 #include "irgraph_t.h"
31 #include <libcore/lc_opts.h>
32 #include <libcore/lc_opts_enum.h>
33 #include <libcore/lc_timing.h>
37 #include <lpp/lpp_net.h>
40 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
42 typedef enum _blocksched_algos_t {
43 BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
46 static int algo = BLOCKSCHED_GREEDY;
48 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
49 { "naiv", BLOCKSCHED_NAIV },
50 { "extbb", BLOCKSCHED_EXTBB },
51 { "greedy", BLOCKSCHED_GREEDY },
53 { "ilp", BLOCKSCHED_ILP },
58 static lc_opt_enum_int_var_t algo_var = {
59 &algo, blockschedalgo_items
62 static const lc_opt_table_entry_t be_blocksched_options[] = {
63 LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
69 * / ___|_ __ ___ ___ __| |_ _
70 * | | _| '__/ _ \/ _ \/ _` | | | |
71 * | |_| | | | __/ __/ (_| | |_| |
72 * \____|_| \___|\___|\__,_|\__, |
76 typedef struct _blocksched_entry_t {
78 struct _blocksched_entry_t *next;
79 struct _blocksched_entry_t *prev;
82 typedef struct _edge_t {
83 ir_node *block; /**< source block */
84 int pos; /**< number of cfg predecessor (target) */
85 double execfreq; /**< the frequency */
86 int highest_execfreq; /**< flag that indicates wether this edge is the edge with the highest
87 execfreq pointing away from this block */
90 typedef struct _blocksched_env_t {
93 ir_exec_freq *execfreqs;
100 * Collect cfg frequencies of all edges between blocks.
101 * Also determines edge with highest frequency.
103 static void collect_egde_frequency(ir_node *block, void *data)
105 blocksched_env_t *env = data;
108 blocksched_entry_t *entry;
110 entry = obstack_alloc(env->obst, sizeof(entry[0]));
111 entry->block = block;
114 set_irn_link(block, entry);
116 if (block == get_irg_start_block(env->irg))
119 arity = get_irn_arity(block);
124 edge.execfreq = get_block_execfreq(env->execfreqs, block);
125 edge.highest_execfreq = 1;
126 ARR_APP1(edge_t, env->edges, edge);
129 double highest_execfreq = -1.0;
130 int highest_edge_num = -1;
133 for (i = 0; i < arity; ++i) {
135 ir_node *pred_block = get_Block_cfgpred_block(block, i);
137 execfreq = get_block_execfreq(env->execfreqs, pred_block);
140 edge.execfreq = execfreq;
141 edge.highest_execfreq = 0;
142 ARR_APP1(edge_t, env->edges, edge);
144 if (execfreq > highest_execfreq) {
145 highest_execfreq = execfreq;
146 highest_edge_num = ARR_LEN(env->edges) - 1;
150 if(highest_edge_num >= 0)
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((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 /* is 1 of the blocks already attached to another block? */
217 if (pred_entry->next != NULL || entry->prev != NULL)
220 /* schedule the 2 blocks behind each other */
221 DBG((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((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((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((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((dbg, LEVEL_1, "pick from worklist\n"));
304 if (pdeq_empty(env->worklist)) {
305 DBG((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((dbg, LEVEL_1, "Blockschedule:\n"));
345 for (entry = first; entry != NULL; entry = entry->next) {
347 block_list[i++] = entry->block;
348 DBG((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);
371 // collect edge execution frequencies
372 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
374 // sort interblock edges by execution frequency
375 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
377 (void)be_remove_empty_blocks(irg);
379 if (algo != BLOCKSCHED_NAIV)
380 coalesce_blocks(&env);
382 start_entry = finish_block_schedule(&env);
383 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
385 DEL_ARR_F(env.edges);
386 obstack_free(&obst, NULL);
401 typedef struct _ilp_edge_t {
402 ir_node *block; /**< source block */
403 int pos; /**< number of cfg predecessor (target) */
407 typedef struct _blocksched_ilp_env_t {
408 blocksched_env_t env;
409 ilp_edge_t *ilpedges;
411 } blocksched_ilp_env_t;
413 typedef struct _blocksched_ilp_entry_t {
415 struct _blocksched_entry_t *next;
416 struct _blocksched_entry_t *prev;
419 } blocksched_ilp_entry_t;
421 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
425 int edgeidx = ARR_LEN(env->ilpedges);
427 snprintf(name, sizeof(name), "edge%d", edgeidx);
431 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
433 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
437 static void collect_egde_frequency_ilp(ir_node *block, void *data)
439 blocksched_ilp_env_t *env = data;
440 ir_graph *irg = env->env.irg;
441 ir_node *startblock = get_irg_start_block(irg);
446 blocksched_ilp_entry_t *entry;
448 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
449 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
451 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
452 entry->block = block;
455 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
456 set_irn_link(block, entry);
458 if (block == startblock)
461 arity = get_irn_arity(block);
463 double execfreq = get_block_execfreq(env->env.execfreqs, block);
464 add_ilp_edge(block, 0, execfreq, env);
468 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
470 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
471 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
473 for (i = 0; i < arity; ++i) {
477 ir_node *pred_block = get_Block_cfgpred_block(block, i);
479 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
480 edgenum = add_ilp_edge(block, i, execfreq, env);
481 edge = &env->ilpedges[edgenum];
482 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
488 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
491 int edge_count = ARR_LEN(env->ilpedges);
493 /* complete out constraints */
494 for(i = 0; i < edge_count; ++i) {
495 const ilp_edge_t *edge = &env->ilpedges[i];
496 ir_node *block = edge->block;
498 blocksched_ilp_entry_t *entry;
500 /* the block might have been removed already... */
501 if (is_Bad(get_Block_cfgpred(block, 0)))
504 pred = get_Block_cfgpred_block(block, edge->pos);
505 entry = get_irn_link(pred);
507 DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
508 pred, block, edge->pos));
509 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
516 lpp_dump(env->lpp, "lpp.out");
517 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
518 f = fopen(fname, "w");
519 lpp_dump_plain(env->lpp, f);
524 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
525 lpp_solve_net(env->lpp, "i44pc52", "cplex");
526 assert(lpp_is_sol_valid(env->lpp));
528 /* Apply results to edges */
529 for (i = 0; i < edge_count; ++i) {
530 const ilp_edge_t *edge = &env->ilpedges[i];
531 ir_node *block = edge->block;
534 blocksched_entry_t *entry;
535 blocksched_entry_t *pred_entry;
537 /* the block might have been removed already... */
538 if (is_Bad(get_Block_cfgpred(block, 0)))
541 is_jump = lpp_get_var_sol(env->lpp, edge->ilpvar);
545 pred = get_Block_cfgpred_block(block, edge->pos);
546 entry = get_irn_link(block);
547 pred_entry = get_irn_link(pred);
549 assert(entry->prev == NULL && pred_entry->next == NULL);
550 entry->prev = pred_entry;
551 pred_entry->next = entry;
555 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
557 blocksched_ilp_env_t env;
559 blocksched_entry_t *start_entry;
560 ir_node **block_list;
565 env.env.obst = &obst;
566 env.env.execfreqs = execfreqs;
567 env.env.worklist = NULL;
568 env.env.blockcount = 0;
569 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
571 env.lpp = new_lpp("blockschedule", lpp_minimize);
572 lpp_set_time_limit(env.lpp, 20);
573 lpp_set_log(env.lpp, stdout);
575 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
577 (void)be_remove_empty_blocks(irg);
578 coalesce_blocks_ilp(&env);
580 start_entry = finish_block_schedule(&env.env);
581 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
583 DEL_ARR_F(env.ilpedges);
585 obstack_free(&obst, NULL);
589 #endif /* WITH_ILP */
593 * | ____|_ _| |_| __ )| __ )
594 * | _| \ \/ / __| _ \| _ \
595 * | |___ > <| |_| |_) | |_) |
596 * |_____/_/\_\\__|____/|____/
600 /** A simple forward single linked list. */
602 ir_node *start; /**< start of the list */
603 ir_node *end; /**< last block in the list */
604 unsigned n_blks; /**< number of blocks in the list */
607 static void add_block(anchor *list, ir_node *block) {
608 if (list->start == NULL) {
613 set_irn_link(list->end, block);
620 static void create_block_list(ir_node *leader_block, anchor *list) {
622 const ir_edge_t *edge;
623 ir_node *block = NULL;
624 ir_extblk *extbb = get_Block_extbb(leader_block);
626 if (extbb_visited(extbb))
628 mark_extbb_visited(extbb);
630 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
631 block = get_extbb_block(extbb, i);
632 add_block(list, block);
635 assert(block != NULL);
637 /* pick successor extbbs */
638 foreach_block_succ(block, edge) {
639 ir_node *succ = get_edge_src_irn(edge);
640 create_block_list(succ, list);
643 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
644 block = get_extbb_block(extbb, i);
646 foreach_block_succ(block, edge) {
647 ir_node *succ = get_edge_src_irn(edge);
648 create_block_list(succ, list);
653 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
656 * Calculates a block schedule. The schedule is stored as a linked
657 * list starting at the start_block of the irg.
659 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
662 ir_node **blk_list, *b, *n;
665 /* schedule extended basic blocks */
666 compute_extbb_execfreqs(irg, execfreqs);
667 //compute_extbb(irg);
672 inc_irg_block_visited(irg);
673 create_block_list(get_irg_start_block(irg), &list);
675 /** create an array, so we can go forward and backward */
676 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
678 for (i = 0, b = list.start; b; b = n, ++i) {
689 * | |\/| |/ _` | | '_ \
690 * | | | | (_| | | | | |
691 * |_| |_|\__,_|_|_| |_|
694 void be_init_blocksched(void)
696 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
697 lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
699 lc_opt_add_table(blocksched_grp, be_blocksched_options);
701 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
704 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
706 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
709 case BLOCKSCHED_GREEDY:
710 case BLOCKSCHED_NAIV:
711 return create_block_schedule_greedy(irg, execfreqs);
712 case BLOCKSCHED_EXTBB:
713 return create_extbb_block_schedule(irg, execfreqs);
716 return create_block_schedule_ilp(irg, execfreqs);
717 #endif /* WITH_ILP */
720 assert(0 && "unknown blocksched algo");