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 set_using_visited(irg);
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);
333 clear_using_visited(irg);
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((dbg, LEVEL_1, "Blockschedule:\n"));
348 for (entry = first; entry != NULL; entry = entry->next) {
350 block_list[i++] = entry->block;
351 DBG((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);
374 // collect edge execution frequencies
375 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
377 // sort interblock edges by execution frequency
378 qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
380 (void)be_remove_empty_blocks(irg);
382 if (algo != BLOCKSCHED_NAIV)
383 coalesce_blocks(&env);
385 start_entry = finish_block_schedule(&env);
386 block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
388 DEL_ARR_F(env.edges);
389 obstack_free(&obst, NULL);
404 typedef struct _ilp_edge_t {
405 ir_node *block; /**< source block */
406 int pos; /**< number of cfg predecessor (target) */
410 typedef struct _blocksched_ilp_env_t {
411 blocksched_env_t env;
412 ilp_edge_t *ilpedges;
414 } blocksched_ilp_env_t;
416 typedef struct _blocksched_ilp_entry_t {
418 struct _blocksched_entry_t *next;
419 struct _blocksched_entry_t *prev;
422 } blocksched_ilp_entry_t;
424 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
428 int edgeidx = ARR_LEN(env->ilpedges);
430 snprintf(name, sizeof(name), "edge%d", edgeidx);
434 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
436 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
440 static void collect_egde_frequency_ilp(ir_node *block, void *data)
442 blocksched_ilp_env_t *env = data;
443 ir_graph *irg = env->env.irg;
444 ir_node *startblock = get_irg_start_block(irg);
449 blocksched_ilp_entry_t *entry;
451 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
452 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
454 entry = obstack_alloc(env->env.obst, sizeof(entry[0]));
455 entry->block = block;
458 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
459 set_irn_link(block, entry);
461 if (block == startblock)
464 arity = get_irn_arity(block);
466 double execfreq = get_block_execfreq(env->env.execfreqs, block);
467 add_ilp_edge(block, 0, execfreq, env);
471 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
473 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
474 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
476 for (i = 0; i < arity; ++i) {
480 ir_node *pred_block = get_Block_cfgpred_block(block, i);
482 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
483 edgenum = add_ilp_edge(block, i, execfreq, env);
484 edge = &env->ilpedges[edgenum];
485 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
491 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
494 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((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);
519 lpp_dump(env->lpp, "lpp.out");
520 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
521 f = fopen(fname, "w");
522 lpp_dump_plain(env->lpp, f);
527 //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
528 lpp_solve_net(env->lpp, "i44pc52", "cplex");
529 assert(lpp_is_sol_valid(env->lpp));
531 /* Apply results to edges */
532 for (i = 0; i < edge_count; ++i) {
533 const ilp_edge_t *edge = &env->ilpedges[i];
534 ir_node *block = edge->block;
537 blocksched_entry_t *entry;
538 blocksched_entry_t *pred_entry;
540 /* the block might have been removed already... */
541 if (is_Bad(get_Block_cfgpred(block, 0)))
544 is_jump = lpp_get_var_sol(env->lpp, edge->ilpvar);
548 pred = get_Block_cfgpred_block(block, edge->pos);
549 entry = get_irn_link(block);
550 pred_entry = get_irn_link(pred);
552 assert(entry->prev == NULL && pred_entry->next == NULL);
553 entry->prev = pred_entry;
554 pred_entry->next = entry;
558 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
560 blocksched_ilp_env_t env;
562 blocksched_entry_t *start_entry;
563 ir_node **block_list;
568 env.env.obst = &obst;
569 env.env.execfreqs = execfreqs;
570 env.env.worklist = NULL;
571 env.env.blockcount = 0;
572 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
574 env.lpp = new_lpp("blockschedule", lpp_minimize);
575 lpp_set_time_limit(env.lpp, 20);
576 lpp_set_log(env.lpp, stdout);
578 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
580 (void)be_remove_empty_blocks(irg);
581 coalesce_blocks_ilp(&env);
583 start_entry = finish_block_schedule(&env.env);
584 block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
586 DEL_ARR_F(env.ilpedges);
588 obstack_free(&obst, NULL);
592 #endif /* WITH_ILP */
596 * | ____|_ _| |_| __ )| __ )
597 * | _| \ \/ / __| _ \| _ \
598 * | |___ > <| |_| |_) | |_) |
599 * |_____/_/\_\\__|____/|____/
603 /** A simple forward single linked list. */
605 ir_node *start; /**< start of the list */
606 ir_node *end; /**< last block in the list */
607 unsigned n_blks; /**< number of blocks in the list */
610 static void add_block(anchor *list, ir_node *block) {
611 if (list->start == NULL) {
615 set_irn_link(list->end, block);
622 static void create_block_list(ir_node *leader_block, anchor *list) {
624 const ir_edge_t *edge;
625 ir_node *block = NULL;
626 ir_extblk *extbb = get_Block_extbb(leader_block);
628 if (extbb_visited(extbb))
630 mark_extbb_visited(extbb);
632 for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
633 block = get_extbb_block(extbb, i);
634 add_block(list, block);
637 assert(block != NULL);
639 /* pick successor extbbs */
640 foreach_block_succ(block, edge) {
641 ir_node *succ = get_edge_src_irn(edge);
642 create_block_list(succ, list);
645 for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
646 block = get_extbb_block(extbb, i);
648 foreach_block_succ(block, edge) {
649 ir_node *succ = get_edge_src_irn(edge);
650 create_block_list(succ, list);
655 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
658 * Calculates a block schedule. The schedule is stored as a linked
659 * list starting at the start_block of the irg.
661 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
664 ir_node **blk_list, *b, *n;
667 /* schedule extended basic blocks */
668 compute_extbb_execfreqs(irg, execfreqs);
669 //compute_extbb(irg);
675 set_using_irn_link(irg);
676 set_using_visited(irg);
677 inc_irg_block_visited(irg);
679 create_block_list(get_irg_start_block(irg), &list);
681 /** create an array, so we can go forward and backward */
682 blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
684 for (i = 0, b = list.start; b; b = n, ++i) {
689 clear_using_irn_link(irg);
690 clear_using_visited(irg);
698 * | |\/| |/ _` | | '_ \
699 * | | | | (_| | | | | |
700 * |_| |_|\__,_|_|_| |_|
703 void be_init_blocksched(void)
705 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
706 lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
708 lc_opt_add_table(blocksched_grp, be_blocksched_options);
710 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
713 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
715 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
718 case BLOCKSCHED_GREEDY:
719 case BLOCKSCHED_NAIV:
720 return create_block_schedule_greedy(irg, execfreqs);
721 case BLOCKSCHED_EXTBB:
722 return create_extbb_block_schedule(irg, execfreqs);
725 return create_block_schedule_ilp(irg, execfreqs);
726 #endif /* WITH_ILP */
729 assert(0 && "unknown blocksched algo");