2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Block-scheduling strategies.
23 * @author Matthias Braun, Christoph Mallon
27 * The goals of the greedy (and ILP) algorithm here works by assuming that
28 * we want to change as many jumps to fallthroughs as possible (executed jumps
29 * actually, we have to look at the execution frequencies). The algorithms
30 * do this by collecting execution frequencies of all branches (which is easily
31 * possible when all critical edges are split) then removes critical edges where
32 * possible as we don't need and want them anymore now. The algorithms then try
33 * to change as many edges to fallthroughs as possible, this is done by setting
34 * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
35 * execution frequencies and tries to transform them to fallthroughs in this order
39 #include "beblocksched.h"
49 #include "irgraph_t.h"
62 #include "lc_opts_enum.h"
67 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
69 typedef enum blocksched_algos_t {
70 BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
73 static int algo = BLOCKSCHED_GREEDY;
75 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
76 { "naiv", BLOCKSCHED_NAIV },
77 { "greedy", BLOCKSCHED_GREEDY },
78 { "ilp", BLOCKSCHED_ILP },
82 static lc_opt_enum_int_var_t algo_var = {
83 &algo, blockschedalgo_items
86 static const lc_opt_table_entry_t be_blocksched_options[] = {
87 LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var),
93 * / ___|_ __ ___ ___ __| |_ _
94 * | | _| '__/ _ \/ _ \/ _` | | | |
95 * | |_| | | | __/ __/ (_| | |_| |
96 * \____|_| \___|\___|\__,_|\__, |
100 typedef struct blocksched_entry_t blocksched_entry_t;
101 struct blocksched_entry_t {
103 blocksched_entry_t *next;
104 blocksched_entry_t *prev;
107 typedef struct edge_t edge_t;
109 ir_node *block; /**< source block */
110 int pos; /**< number of cfg predecessor (target) */
111 double execfreq; /**< the frequency */
112 double outedge_penalty_freq; /**< for edges leaving the loop this is the
113 penality when we make them a
115 int highest_execfreq; /**< flag that indicates whether this edge is
116 the edge with the highest execfreq pointing
117 away from this block */
120 typedef struct blocksched_env_t blocksched_env_t;
121 struct blocksched_env_t {
123 struct obstack *obst;
124 ir_exec_freq *execfreqs;
131 * Collect cfg frequencies of all edges between blocks.
132 * Also determines edge with highest frequency.
134 static void collect_egde_frequency(ir_node *block, void *data)
136 blocksched_env_t *env = (blocksched_env_t*)data;
139 blocksched_entry_t *entry;
142 memset(&edge, 0, sizeof(edge));
144 entry = OALLOCZ(env->obst, blocksched_entry_t);
145 entry->block = block;
146 set_irn_link(block, entry);
148 loop = get_irn_loop(block);
150 arity = get_Block_n_cfgpreds(block);
153 /* must be the start block (or end-block for endless loops),
154 * everything else is dead code and should be removed by now */
155 assert(block == get_irg_start_block(env->irg)
156 || block == get_irg_end_block(env->irg));
157 /* nothing to do here */
159 } else if (arity == 1) {
160 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
161 ir_loop *pred_loop = get_irn_loop(pred_block);
162 float freq = (float)get_block_execfreq(env->execfreqs, block);
164 /* is it an edge leaving a loop */
165 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
166 float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
167 edge.outedge_penalty_freq = -(pred_freq - freq);
172 edge.execfreq = freq;
173 edge.highest_execfreq = 1;
174 ARR_APP1(edge_t, env->edges, edge);
177 double highest_execfreq = -1.0;
178 int highest_edge_num = -1;
181 for (i = 0; i < arity; ++i) {
183 ir_node *pred_block = get_Block_cfgpred_block(block, i);
185 execfreq = get_block_execfreq(env->execfreqs, pred_block);
188 edge.execfreq = execfreq;
189 edge.highest_execfreq = 0;
190 ARR_APP1(edge_t, env->edges, edge);
192 if (execfreq > highest_execfreq) {
193 highest_execfreq = execfreq;
194 highest_edge_num = ARR_LEN(env->edges) - 1;
198 if (highest_edge_num >= 0)
199 env->edges[highest_edge_num].highest_execfreq = 1;
203 static int cmp_edges(const void *d1, const void *d2)
205 const edge_t *e1 = (const edge_t*)d1;
206 const edge_t *e2 = (const edge_t*)d2;
208 return QSORT_CMP(e2->execfreq, e1->execfreq);
211 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
213 const edge_t *e1 = (const edge_t*)d1;
214 const edge_t *e2 = (const edge_t*)d2;
215 /* reverse sorting as penalties are negative */
216 return QSORT_CMP(e1->outedge_penalty_freq, e2->outedge_penalty_freq);
219 static void clear_loop_links(ir_loop *loop)
223 set_loop_link(loop, NULL);
224 n = get_loop_n_elements(loop);
225 for (i = 0; i < n; ++i) {
226 loop_element elem = get_loop_element(loop, i);
227 if (*elem.kind == k_ir_loop) {
228 clear_loop_links(elem.son);
233 static void coalesce_blocks(blocksched_env_t *env)
236 int edge_count = ARR_LEN(env->edges);
237 edge_t *edges = env->edges;
239 /* sort interblock edges by execution frequency */
240 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
242 /* run1: only look at jumps */
243 for (i = 0; i < edge_count; ++i) {
244 const edge_t *edge = &edges[i];
245 ir_node *block = edge->block;
248 blocksched_entry_t *entry, *pred_entry;
250 /* only check edge with highest frequency */
251 if (! edge->highest_execfreq)
254 /* the block might have been removed already... */
255 if (is_Bad(get_Block_cfgpred(block, 0)))
258 pred_block = get_Block_cfgpred_block(block, pos);
259 entry = (blocksched_entry_t*)get_irn_link(block);
260 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
262 if (pred_entry->next != NULL || entry->prev != NULL)
265 /* only coalesce jumps */
266 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
269 /* schedule the 2 blocks behind each other */
270 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
271 pred_entry->block, entry->block, edge->execfreq));
272 pred_entry->next = entry;
273 entry->prev = pred_entry;
276 /* run2: pick loop fallthroughs */
277 clear_loop_links(get_irg_loop(env->irg));
279 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
280 for (i = 0; i < edge_count; ++i) {
281 const edge_t *edge = &edges[i];
282 ir_node *block = edge->block;
285 blocksched_entry_t *entry, *pred_entry;
289 /* already seen all loop outedges? */
290 if (edge->outedge_penalty_freq == 0)
293 /* the block might have been removed already... */
294 if (is_Bad(get_Block_cfgpred(block, pos)))
297 pred_block = get_Block_cfgpred_block(block, pos);
298 entry = (blocksched_entry_t*)get_irn_link(block);
299 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
301 if (pred_entry->next != NULL || entry->prev != NULL)
304 /* we want at most 1 outedge fallthrough per loop */
305 loop = get_irn_loop(pred_block);
306 if (get_loop_link(loop) != NULL)
309 /* schedule the 2 blocks behind each other */
310 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
311 pred_entry->block, entry->block, edge->execfreq));
312 pred_entry->next = entry;
313 entry->prev = pred_entry;
315 /* all loops left have an outedge now */
316 outer_loop = get_irn_loop(block);
318 /* we set loop link to loop to mark it */
319 set_loop_link(loop, loop);
320 loop = get_loop_outer_loop(loop);
321 } while (loop != outer_loop);
324 /* sort interblock edges by execution frequency */
325 qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
327 /* run3: remaining edges */
328 for (i = 0; i < edge_count; ++i) {
329 const edge_t *edge = &edges[i];
330 ir_node *block = edge->block;
333 blocksched_entry_t *entry, *pred_entry;
335 /* the block might have been removed already... */
336 if (is_Bad(get_Block_cfgpred(block, pos)))
339 pred_block = get_Block_cfgpred_block(block, pos);
340 entry = (blocksched_entry_t*)get_irn_link(block);
341 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
343 /* is 1 of the blocks already attached to another block? */
344 if (pred_entry->next != NULL || entry->prev != NULL)
347 /* schedule the 2 blocks behind each other */
348 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
349 pred_entry->block, entry->block, edge->execfreq));
350 pred_entry->next = entry;
351 entry->prev = pred_entry;
355 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
357 ir_node *block = entry->block;
358 ir_node *succ = NULL;
359 blocksched_entry_t *succ_entry;
360 const ir_edge_t *edge;
361 double best_succ_execfreq;
363 if (irn_visited_else_mark(block))
368 DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
370 /* put all successors into the worklist */
371 foreach_block_succ(block, edge) {
372 ir_node *succ_block = get_edge_src_irn(edge);
374 if (irn_visited(succ_block))
377 /* we only need to put the first of a series of already connected
378 * blocks into the worklist */
379 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
380 while (succ_entry->prev != NULL) {
381 /* break cycles... */
382 if (succ_entry->prev->block == succ_block) {
383 succ_entry->prev->next = NULL;
384 succ_entry->prev = NULL;
387 succ_entry = succ_entry->prev;
390 if (irn_visited(succ_entry->block))
393 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
394 pdeq_putr(env->worklist, succ_entry->block);
397 if (entry->next != NULL) {
398 pick_block_successor(entry->next, env);
402 DB((dbg, LEVEL_1, "deciding...\n"));
403 best_succ_execfreq = -1;
405 /* no successor yet: pick the successor block with the highest execution
406 * frequency which has no predecessor yet */
408 foreach_block_succ(block, edge) {
409 ir_node *succ_block = get_edge_src_irn(edge);
412 if (irn_visited(succ_block))
415 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
416 if (succ_entry->prev != NULL)
419 execfreq = get_block_execfreq(env->execfreqs, succ_block);
420 if (execfreq > best_succ_execfreq) {
421 best_succ_execfreq = execfreq;
427 DB((dbg, LEVEL_1, "pick from worklist\n"));
430 if (pdeq_empty(env->worklist)) {
431 DB((dbg, LEVEL_1, "worklist empty\n"));
434 succ = (ir_node*)pdeq_getl(env->worklist);
435 } while (irn_visited(succ));
438 succ_entry = (blocksched_entry_t*)get_irn_link(succ);
439 entry->next = succ_entry;
440 succ_entry->prev = entry;
442 pick_block_successor(succ_entry, env);
445 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
447 ir_graph *irg = env->irg;
448 ir_node *startblock = get_irg_start_block(irg);
449 blocksched_entry_t *entry = (blocksched_entry_t*)get_irn_link(startblock);
451 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
452 inc_irg_visited(irg);
454 env->worklist = new_pdeq();
455 pick_block_successor(entry, env);
456 assert(pdeq_empty(env->worklist));
457 del_pdeq(env->worklist);
459 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
464 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
465 int count, struct obstack* obst)
468 ir_node **block_list;
469 blocksched_entry_t *entry;
472 block_list = NEW_ARR_D(ir_node *, obst, count);
473 DB((dbg, LEVEL_1, "Blockschedule:\n"));
475 for (entry = first; entry != NULL; entry = entry->next) {
477 block_list[i++] = entry->block;
478 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
485 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
487 blocksched_env_t env;
489 blocksched_entry_t *start_entry;
490 ir_node **block_list;
496 env.execfreqs = execfreqs;
497 env.edges = NEW_ARR_F(edge_t, 0);
501 assure_loopinfo(irg);
503 // collect edge execution frequencies
504 irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
506 (void)be_remove_empty_blocks(irg);
508 if (algo != BLOCKSCHED_NAIV)
509 coalesce_blocks(&env);
511 start_entry = finish_block_schedule(&env);
512 block_list = create_blocksched_array(&env, start_entry, env.blockcount,
513 be_get_be_obst(irg));
515 DEL_ARR_F(env.edges);
516 obstack_free(&obst, NULL);
530 typedef struct ilp_edge_t {
531 ir_node *block; /**< source block */
532 int pos; /**< number of cfg predecessor (target) */
536 typedef struct blocksched_ilp_env_t {
537 blocksched_env_t env;
538 ilp_edge_t *ilpedges;
540 } blocksched_ilp_env_t;
542 typedef struct blocksched_ilp_entry_t {
544 struct blocksched_entry_t *next;
545 struct blocksched_entry_t *prev;
548 } blocksched_ilp_entry_t;
550 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
554 int edgeidx = ARR_LEN(env->ilpedges);
556 snprintf(name, sizeof(name), "edge%d", edgeidx);
560 edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
562 ARR_APP1(ilp_edge_t, env->ilpedges, edge);
566 static void collect_egde_frequency_ilp(ir_node *block, void *data)
568 blocksched_ilp_env_t *env = data;
569 ir_graph *irg = env->env.irg;
570 ir_node *startblock = get_irg_start_block(irg);
575 blocksched_ilp_entry_t *entry;
577 snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
578 out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
580 entry = OALLOC(env->env.obst, blocksched_ilp_entry_t);
581 entry->block = block;
584 entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
585 set_irn_link(block, entry);
587 if (block == startblock)
590 arity = get_irn_arity(block);
592 double execfreq = get_block_execfreq(env->env.execfreqs, block);
593 add_ilp_edge(block, 0, execfreq, env);
598 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
599 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
601 for (i = 0; i < arity; ++i) {
605 ir_node *pred_block = get_Block_cfgpred_block(block, i);
607 execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
608 edgenum = add_ilp_edge(block, i, execfreq, env);
609 edge = &env->ilpedges[edgenum];
610 lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
616 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
618 int edge_count = ARR_LEN(env->ilpedges);
619 be_options_t *options = be_get_irg_options(env->env.irg);
622 /* complete out constraints */
623 for (i = 0; i < edge_count; ++i) {
624 const ilp_edge_t *edge = &env->ilpedges[i];
625 ir_node *block = edge->block;
627 blocksched_ilp_entry_t *entry;
629 /* the block might have been removed already... */
630 if (is_Bad(get_Block_cfgpred(block, 0)))
633 pred = get_Block_cfgpred_block(block, edge->pos);
634 entry = get_irn_link(pred);
636 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
637 pred, block, edge->pos));
638 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
641 lpp_solve_net(env->lpp, options->ilp_server, options->ilp_solver);
642 assert(lpp_is_sol_valid(env->lpp));
644 /* Apply results to edges */
645 for (i = 0; i < edge_count; ++i) {
646 const ilp_edge_t *edge = &env->ilpedges[i];
647 ir_node *block = edge->block;
650 blocksched_entry_t *entry;
651 blocksched_entry_t *pred_entry;
653 /* the block might have been removed already... */
654 if (is_Bad(get_Block_cfgpred(block, 0)))
657 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
661 pred = get_Block_cfgpred_block(block, edge->pos);
662 entry = get_irn_link(block);
663 pred_entry = get_irn_link(pred);
665 assert(entry->prev == NULL && pred_entry->next == NULL);
666 entry->prev = pred_entry;
667 pred_entry->next = entry;
671 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
673 blocksched_ilp_env_t env;
675 blocksched_entry_t *start_entry;
676 ir_node **block_list;
681 env.env.obst = &obst;
682 env.env.execfreqs = execfreqs;
683 env.env.worklist = NULL;
684 env.env.blockcount = 0;
685 env.ilpedges = NEW_ARR_F(ilp_edge_t, 0);
687 env.lpp = lpp_new("blockschedule", lpp_minimize);
688 lpp_set_time_limit(env.lpp, 20);
689 lpp_set_log(env.lpp, stdout);
691 irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
693 (void)be_remove_empty_blocks(irg);
694 coalesce_blocks_ilp(&env);
696 start_entry = finish_block_schedule(&env.env);
697 block_list = create_blocksched_array(&env.env, start_entry,
699 be_get_be_obst(irg));
701 DEL_ARR_F(env.ilpedges);
703 obstack_free(&obst, NULL);
711 * | |\/| |/ _` | | '_ \
712 * | | | | (_| | | | | |
713 * |_| |_|\__,_|_|_| |_|
716 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
717 void be_init_blocksched(void)
719 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
721 lc_opt_add_table(be_grp, be_blocksched_options);
723 FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
726 ir_node **be_create_block_schedule(ir_graph *irg)
728 ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
731 case BLOCKSCHED_GREEDY:
732 case BLOCKSCHED_NAIV:
733 return create_block_schedule_greedy(irg, execfreqs);
735 return create_block_schedule_ilp(irg, execfreqs);
738 panic("unknown blocksched algo");