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 = -1;
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 if(highest_edge_num >= 0)
152 env->edges[highest_edge_num].highest_execfreq = 1;
156 static int cmp_edges(const void *d1, const void *d2)
158 const edge_t *e1 = d1;
159 const edge_t *e2 = d2;
161 return QSORT_CMP(e2->execfreq, e1->execfreq);
164 static void coalesce_blocks(blocksched_env_t *env)
167 int edge_count = ARR_LEN(env->edges);
169 /* run1: only look at jumps */
170 for (i = 0; i < edge_count; ++i) {
171 const edge_t *edge = &env->edges[i];
172 ir_node *block = edge->block;
174 blocksched_entry_t *entry, *pred_entry;
176 /* the block might have been removed already... */
177 if (is_Bad(get_Block_cfgpred(block, 0)))
180 /* only check edge with highest frequency */
181 if (! edge->highest_execfreq)
184 pred_block = get_Block_cfgpred_block(block, edge->pos);
185 entry = get_irn_link(block);
186 pred_entry = get_irn_link(pred_block);
188 if (pred_entry->next != NULL || entry->prev != NULL)
191 /* only coalesce jumps */
192 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
195 /* schedule the 2 blocks behind each other */
196 DBG((env->dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
197 pred_entry->block, entry->block, edge->execfreq));
198 pred_entry->next = entry;
199 entry->prev = pred_entry;
202 /* run2: remaining edges */
203 for (i = 0; i < edge_count; ++i) {
204 const edge_t *edge = &env->edges[i];
205 ir_node *block = edge->block;
207 blocksched_entry_t *entry, *pred_entry;
209 /* the block might have been removed already... */
210 if (is_Bad(get_Block_cfgpred(block, 0)))
213 pred_block = get_Block_cfgpred_block(block, edge->pos);
214 entry = get_irn_link(block);
215 pred_entry = get_irn_link(pred_block);
217 /* TODO: what's this check for? */
218 if (pred_entry->next != NULL || entry->prev != NULL)
221 /* schedule the 2 blocks behind each other */
222 DBG((env->dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
223 pred_entry->block, entry->block, edge->execfreq));
224 pred_entry->next = entry;
225 entry->prev = pred_entry;
229 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
231 ir_node *block = entry->block;
232 ir_node *succ = NULL;
233 blocksched_entry_t *succ_entry;
234 const ir_edge_t *edge;
235 double best_succ_execfreq;
237 if (irn_visited(block))
241 mark_irn_visited(block);
243 DBG((env->dbg, LEVEL_1, "Pick succ of %+F\n", block));
245 /* put all successors into the worklist */
246 foreach_block_succ(block, edge) {
247 ir_node *succ_block = get_edge_src_irn(edge);
249 if (irn_visited(succ_block))
252 /* we only need to put the first of a series of already connected
253 * blocks into the worklist */
254 succ_entry = get_irn_link(succ_block);
255 while (succ_entry->prev != NULL) {
256 /* break cycles... */
257 if (succ_entry->prev->block == succ_block) {
258 succ_entry->prev->next = NULL;
259 succ_entry->prev = NULL;
262 succ_entry = succ_entry->prev;
265 if (irn_visited(succ_entry->block))
268 DBG((env->dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
269 pdeq_putr(env->worklist, succ_entry->block);
272 if (entry->next != NULL) {
273 pick_block_successor(entry->next, env);
277 DBG((env->dbg, LEVEL_1, "deciding...\n"));
278 best_succ_execfreq = -1;
280 /* no successor yet: pick the successor block with the highest execution
281 * frequency which has no predecessor yet */
283 foreach_block_succ(block, edge) {
284 ir_node *succ_block = get_edge_src_irn(edge);
287 if (irn_visited(succ_block))
290 succ_entry = get_irn_link(succ_block);
291 if (succ_entry->prev != NULL)
294 execfreq = get_block_execfreq(env->execfreqs, succ_block);
295 if (execfreq > best_succ_execfreq) {
296 best_succ_execfreq = execfreq;
302 DBG((env->dbg, LEVEL_1, "pick from worklist\n"));
305 if (pdeq_empty(env->worklist)) {
306 DBG((env->dbg, LEVEL_1, "worklist empty\n"));
309 succ = pdeq_getl(env->worklist);
310 } while (irn_visited(succ));
313 succ_entry = get_irn_link(succ);
314 entry->next = succ_entry;
315 succ_entry->prev = entry;
317 pick_block_successor(succ_entry, env);
320 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
322 ir_graph *irg = env->irg;
323 ir_node *startblock = get_irg_start_block(irg);
324 blocksched_entry_t *entry = get_irn_link(startblock);
326 inc_irg_visited(irg);
328 env->worklist = new_pdeq();
329 pick_block_successor(entry, env);
330 assert(pdeq_empty(env->worklist));
331 del_pdeq(env->worklist);
336 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
337 int count, struct obstack* obst)
340 ir_node **block_list;
341 blocksched_entry_t *entry;
343 block_list = NEW_ARR_D(ir_node *, obst, count);
344 DBG((env->dbg, LEVEL_1, "Blockschedule:\n"));
346 for (entry = first; entry != NULL; entry = entry->next) {
348 block_list[i++] = entry->block;
349 DBG((env->dbg, LEVEL_1, "\t%+F\n", entry->block));
356 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
358 blocksched_env_t env;
360 blocksched_entry_t *start_entry;
361 ir_node **block_list;
367 env.execfreqs = execfreqs;
368 env.edges = NEW_ARR_F(edge_t, 0);
371 FIRM_DBG_REGISTER(env.dbg, "firm.be.blocksched");
373 // collect edge execution frequencies
374 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
376 // sort interblock edges by execution frequency
377 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
379 be_remove_empty_blocks(irg);
381 if (algo != BLOCKSCHED_NAIV)
382 coalesce_blocks(&env);
384 start_entry = finish_block_schedule(&env);
385 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
387 DEL_ARR_F(env.edges);
388 obstack_free(&obst, NULL);
403 typedef struct _ilp_edge_t {
404 ir_node *block; /**< source block */
405 int pos; /**< number of cfg predecessor (target) */
409 typedef struct _blocksched_ilp_env_t {
410 blocksched_env_t env;
411 ilp_edge_t *ilpedges;
413 } blocksched_ilp_env_t;
415 typedef struct _blocksched_ilp_entry_t {
417 struct _blocksched_entry_t *next;
418 struct _blocksched_entry_t *prev;
421 } blocksched_ilp_entry_t;
423 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
427 int edgeidx = ARR_LEN(env->ilpedges);
429 snprintf(name, sizeof(name), "edge%d", edgeidx);
433 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
435 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
439 static void collect_egde_frequency_ilp(ir_node *block, void *data)
441 blocksched_ilp_env_t *env = data;
442 ir_graph *irg = env->env.irg;
443 ir_node *startblock = get_irg_start_block(irg);
448 blocksched_ilp_entry_t *entry;
450 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
451 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
453 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
454 entry->block = block;
457 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
458 set_irn_link(block, entry);
460 if (block == startblock)
463 arity = get_irn_arity(block);
465 double execfreq = get_block_execfreq(env->env.execfreqs, block);
466 add_ilp_edge(block, 0, execfreq, env);
470 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
472 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
473 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
475 for (i = 0; i < arity; ++i) {
479 ir_node *pred_block = get_Block_cfgpred_block(block, i);
481 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
482 edgenum = add_ilp_edge(block, i, execfreq, env);
483 edge = &env->ilpedges[edgenum];
484 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
490 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
493 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);
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);
521 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
522 lpp_solve_net(env->lpp, "i44pc52", "cplex");
523 assert(lpp_is_sol_valid(env->lpp));
525 /* Apply results to edges */
526 for (i = 0; i < edge_count; ++i) {
527 const ilp_edge_t *edge = &env->ilpedges[i];
528 ir_node *block = edge->block;
531 blocksched_entry_t *entry;
532 blocksched_entry_t *pred_entry;
534 /* the block might have been removed already... */
535 if (is_Bad(get_Block_cfgpred(block, 0)))
538 is_jump = lpp_get_var_sol(env->lpp, edge->ilpvar);
542 pred = get_Block_cfgpred_block(block, edge->pos);
543 entry = get_irn_link(block);
544 pred_entry = get_irn_link(pred);
546 assert(entry->prev == NULL && pred_entry->next == NULL);
547 entry->prev = pred_entry;
548 pred_entry->next = entry;
552 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
554 blocksched_ilp_env_t env;
556 blocksched_entry_t *start_entry;
557 ir_node **block_list;
562 env.env.obst = &obst;
563 env.env.execfreqs = execfreqs;
564 env.env.worklist = NULL;
565 env.env.blockcount = 0;
566 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
567 FIRM_DBG_REGISTER(env.env.dbg, "firm.be.blocksched");
569 env.lpp = new_lpp("blockschedule", lpp_minimize);
570 lpp_set_time_limit(env.lpp, 20);
571 lpp_set_log(env.lpp, stdout);
573 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
575 be_remove_empty_blocks(irg);
576 coalesce_blocks_ilp(&env);
578 start_entry = finish_block_schedule(&env.env);
579 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
581 DEL_ARR_F(env.ilpedges);
583 obstack_free(&obst, NULL);
587 #endif /* WITH_ILP */
591 * | ____|_ _| |_| __ )| __ )
592 * | _| \ \/ / __| _ \| _ \
593 * | |___ > <| |_| |_) | |_) |
594 * |_____/_/\_\\__|____/|____/
598 /** A simple forward single linked list. */
600 ir_node *start; /**< start of the list */
601 ir_node *end; /**< last block in the list */
602 unsigned n_blks; /**< number of blocks in the list */
605 static void add_block(anchor *list, ir_node *block) {
606 if (list->start == NULL) {
611 set_irn_link(list->end, block);
618 static void create_block_list(ir_node *leader_block, anchor *list) {
620 const ir_edge_t *edge;
621 ir_node *block = NULL;
622 ir_extblk *extbb = get_Block_extbb(leader_block);
624 if (extbb_visited(extbb))
626 mark_extbb_visited(extbb);
628 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
629 block = get_extbb_block(extbb, i);
630 add_block(list, block);
633 assert(block != NULL);
635 /* pick successor extbbs */
636 foreach_block_succ(block, edge) {
637 ir_node *succ = get_edge_src_irn(edge);
638 create_block_list(succ, list);
641 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
642 block = get_extbb_block(extbb, i);
644 foreach_block_succ(block, edge) {
645 ir_node *succ = get_edge_src_irn(edge);
646 create_block_list(succ, list);
651 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
654 * Calculates a block schedule. The schedule is stored as a linked
655 * list starting at the start_block of the irg.
657 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
660 ir_node **blk_list, *b, *n;
663 /* schedule extended basic blocks */
664 compute_extbb_execfreqs(irg, execfreqs);
665 //compute_extbb(irg);
670 inc_irg_block_visited(irg);
671 create_block_list(get_irg_start_block(irg), &list);
673 /** create an array, so we can go forward and backward */
674 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
676 for (i = 0, b = list.start; b; b = n, ++i) {
687 * | |\/| |/ _` | | '_ \
688 * | | | | (_| | | | | |
689 * |_| |_|\__,_|_|_| |_|
694 void be_block_schedule_register_options(lc_opt_entry_t *grp)
696 static int run_once = 0;
697 lc_opt_entry_t *blocksched_grp;
703 blocksched_grp = lc_opt_get_grp(grp, "blocksched");
705 lc_opt_add_table(blocksched_grp, be_blocksched_options);
707 #endif /* WITH_LIBCORE */
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");