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 Implements a trace scheduler as presented in Muchnik[TM].
23 * @author Michael Beck
31 #include "iredges_t.h"
34 #include "belistsched.h"
38 /* we need a special mark */
42 typedef struct _trace_irn {
43 sched_timestep_t delay; /**< The delay for this node if already calculated, else 0. */
44 sched_timestep_t etime; /**< The earliest time of this node. */
45 unsigned num_user; /**< The number real users (mode datab) of this node */
46 int reg_diff; /**< The difference of num(out registers) - num(in registers) */
47 int preorder; /**< The pre-order position */
48 unsigned critical_path_len; /**< The weighted length of the longest critical path */
49 unsigned is_root : 1; /**< is a root node of a block */
52 typedef struct _trace_env {
53 trace_irn_t *sched_info; /**< trace scheduling information about the nodes */
54 sched_timestep_t curr_time; /**< current time of the scheduler */
55 void *selector_env; /**< the backend selector environment */
56 const list_sched_selector_t *selector; /**< the actual backend selector */
57 be_lv_t *liveness; /**< The liveness for the irg */
58 DEBUG_ONLY(firm_dbg_module_t *dbg;)
62 * Returns a random node from a nodeset
64 static ir_node *get_nodeset_node(const ir_nodeset_t *nodeset)
66 ir_nodeset_iterator_t iter;
68 ir_nodeset_iterator_init(&iter, nodeset);
69 return ir_nodeset_iterator_next(&iter);
73 * Returns non-zero if the node is a root node
75 static inline unsigned is_root_node(trace_env_t *env, ir_node *n)
77 int idx = get_irn_idx(n);
79 assert(idx < ARR_LEN(env->sched_info));
80 return env->sched_info[idx].is_root;
84 * Mark a node as root node
86 static inline void mark_root_node(trace_env_t *env, ir_node *n)
88 int idx = get_irn_idx(n);
90 assert(idx < ARR_LEN(env->sched_info));
91 env->sched_info[idx].is_root = 1;
95 * Get the current delay.
97 static inline sched_timestep_t get_irn_delay(trace_env_t *env, ir_node *n) {
98 int idx = get_irn_idx(n);
100 assert(idx < ARR_LEN(env->sched_info));
101 return env->sched_info[idx].delay;
105 * Set the current delay.
107 static inline void set_irn_delay(trace_env_t *env, ir_node *n, sched_timestep_t delay) {
108 int idx = get_irn_idx(n);
110 assert(idx < ARR_LEN(env->sched_info));
111 env->sched_info[idx].delay = delay;
115 * Get the current etime.
117 static inline sched_timestep_t get_irn_etime(trace_env_t *env, ir_node *n) {
118 int idx = get_irn_idx(n);
120 assert(idx < ARR_LEN(env->sched_info));
121 return env->sched_info[idx].etime;
125 * Set the current etime.
127 static inline void set_irn_etime(trace_env_t *env, ir_node *n, sched_timestep_t etime) {
128 int idx = get_irn_idx(n);
130 assert(idx < ARR_LEN(env->sched_info));
131 env->sched_info[idx].etime = etime;
135 * Get the number of users.
137 static inline unsigned get_irn_num_user(trace_env_t *env, ir_node *n) {
138 int idx = get_irn_idx(n);
140 assert(idx < ARR_LEN(env->sched_info));
141 return env->sched_info[idx].num_user;
145 * Set the number of users.
147 static inline void set_irn_num_user(trace_env_t *env, ir_node *n, unsigned num_user) {
148 int idx = get_irn_idx(n);
150 assert(idx < ARR_LEN(env->sched_info));
151 env->sched_info[idx].num_user = num_user;
155 * Get the register difference.
157 static inline int get_irn_reg_diff(trace_env_t *env, ir_node *n) {
158 int idx = get_irn_idx(n);
160 assert(idx < ARR_LEN(env->sched_info));
161 return env->sched_info[idx].reg_diff;
165 * Set the register difference.
167 static inline void set_irn_reg_diff(trace_env_t *env, ir_node *n, int reg_diff) {
168 int idx = get_irn_idx(n);
170 assert(idx < ARR_LEN(env->sched_info));
171 env->sched_info[idx].reg_diff = reg_diff;
175 * Get the pre-order position.
177 static inline int get_irn_preorder(trace_env_t *env, ir_node *n) {
178 int idx = get_irn_idx(n);
180 assert(idx < ARR_LEN(env->sched_info));
181 return env->sched_info[idx].preorder;
185 * Set the pre-order position.
187 static inline void set_irn_preorder(trace_env_t *env, ir_node *n, int pos) {
188 int idx = get_irn_idx(n);
190 assert(idx < ARR_LEN(env->sched_info));
191 env->sched_info[idx].preorder = pos;
195 * Get the pre-order position.
197 static inline unsigned get_irn_critical_path_len(trace_env_t *env, ir_node *n) {
198 int idx = get_irn_idx(n);
200 assert(idx < ARR_LEN(env->sched_info));
201 return env->sched_info[idx].critical_path_len;
205 * Set the pre-order position.
207 static inline void set_irn_critical_path_len(trace_env_t *env, ir_node *n, unsigned len) {
208 int idx = get_irn_idx(n);
210 assert(idx < ARR_LEN(env->sched_info));
211 env->sched_info[idx].critical_path_len = len;
215 * returns the exec-time for node n.
217 static sched_timestep_t exectime(trace_env_t *env, ir_node *n) {
218 if (be_is_Keep(n) || is_Proj(n))
220 if (env->selector->exectime)
221 return env->selector->exectime(env->selector_env, n);
226 * Calculates the latency for between two ops
228 static sched_timestep_t latency(trace_env_t *env, ir_node *pred, int pred_cycle, ir_node *curr, int curr_cycle) {
229 /* a Keep hides a root */
230 if (be_is_Keep(curr))
231 return exectime(env, pred);
233 /* Proj's are executed immediately */
237 /* predecessors Proj's must be skipped */
239 pred = get_Proj_pred(pred);
241 if (env->selector->latency)
242 return env->selector->latency(env->selector_env, pred, pred_cycle, curr, curr_cycle);
247 * Returns the number of users of a node having mode datab.
249 static int get_num_successors(ir_node *irn) {
251 const ir_edge_t *edge;
253 if (get_irn_mode(irn) == mode_T) {
254 /* for mode_T nodes: count the users of all Projs */
255 foreach_out_edge(irn, edge) {
256 ir_node *proj = get_edge_src_irn(edge);
257 ir_mode *mode = get_irn_mode(proj);
260 sum += get_num_successors(proj);
261 else if (mode_is_datab(mode))
262 sum += get_irn_n_edges(proj);
266 /* do not count keep-alive edges */
267 foreach_out_edge(irn, edge) {
268 if (get_irn_opcode(get_edge_src_irn(edge)) != iro_End)
277 * Returns the difference of regs_output - regs_input;
279 static int get_reg_difference(trace_env_t *env, ir_node *irn) {
283 ir_node *block = get_nodes_block(irn);
285 if (be_is_Call(irn)) {
286 /* we want calls preferred */
290 if (get_irn_mode(irn) == mode_T) {
291 /* mode_T nodes: num out regs == num Projs with mode datab */
292 const ir_edge_t *edge;
293 foreach_out_edge(irn, edge) {
294 ir_node *proj = get_edge_src_irn(edge);
295 if (mode_is_datab(get_irn_mode(proj)))
302 /* num in regs: number of ins with mode datab and not ignore */
303 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
304 ir_node *in = get_irn_n(irn, i);
306 if (!mode_is_datab(get_irn_mode(in)))
309 if (arch_irn_is_ignore(in))
312 if (be_is_live_end(env->liveness, block, in))
318 return num_out - num_in;
322 * descent into a dag and create a pre-order list.
324 static void descent(ir_node *root, ir_node *block, ir_node **list, trace_env_t *env, unsigned path_len) {
327 if (! is_Phi(root)) {
328 path_len += exectime(env, root);
329 if (get_irn_critical_path_len(env, root) < path_len) {
330 set_irn_critical_path_len(env, root, path_len);
332 /* calculate number of users (needed for heuristic) */
333 set_irn_num_user(env, root, get_num_successors(root));
335 /* calculate register difference (needed for heuristic) */
336 set_irn_reg_diff(env, root, get_reg_difference(env, root));
338 /* Phi nodes always leave the block */
339 for (i = get_irn_arity(root) - 1; i >= 0; --i) {
340 ir_node *pred = get_irn_n(root, i);
342 DBG((env->dbg, LEVEL_3, " node %+F\n", pred));
344 /* Blocks may happen as predecessors of End nodes */
348 /* already seen nodes are not marked */
349 if (get_irn_link(pred) != MARK)
352 /* don't leave our block */
353 if (get_nodes_block(pred) != block)
356 set_irn_link(pred, NULL);
358 descent(pred, block, list, env, path_len);
361 set_irn_link(root, *list);
366 * Returns non-zero if root is a root in the block block.
368 static int is_root(ir_node *root, ir_node *block) {
369 const ir_edge_t *edge;
371 foreach_out_edge(root, edge) {
372 ir_node *succ = get_edge_src_irn(edge);
376 /* Phi nodes are always in "another block */
379 if (get_nodes_block(succ) == block)
386 * Performs initial block calculations for trace scheduling.
388 static void trace_preprocess_block(trace_env_t *env, ir_node *block) {
389 ir_node *root = NULL, *preord = NULL;
392 const ir_edge_t *edge;
394 /* First step: Find the root set. */
395 foreach_out_edge(block, edge) {
396 ir_node *succ = get_edge_src_irn(edge);
398 if (is_Block(succ)) {
399 /* A Block-Block edge. This should be the MacroBlock
400 * edge, ignore it. */
401 assert(get_Block_MacroBlock(succ) == block && "Block-Block edge found");
404 if (is_Anchor(succ)) {
405 /* ignore a keep alive edge */
408 if (is_root(succ, block)) {
409 mark_root_node(env, succ);
410 set_irn_link(succ, root);
414 set_irn_link(succ, MARK);
417 /* Second step: calculate the pre-order list. */
419 for (curr = root; curr; curr = irn) {
420 irn = get_irn_link(curr);
421 DBG((env->dbg, LEVEL_2, " DAG root %+F\n", curr));
422 descent(curr, block, &preord, env, 0);
426 /* Third step: calculate the Delay. Note that our
427 * list is now in pre-order, starting at root
429 for (cur_pos = 0, curr = root; curr; curr = get_irn_link(curr), cur_pos++) {
432 if (arch_irn_class_is(curr, branch)) {
433 /* assure, that branches can be executed last */
437 if (is_root_node(env, curr))
438 d = exectime(env, curr);
441 foreach_out_edge(curr, edge) {
442 ir_node *n = get_edge_src_irn(edge);
444 if (get_nodes_block(n) == block) {
447 ld = latency(env, curr, 1, n, 0) + get_irn_delay(env, n);
453 set_irn_delay(env, curr, d);
454 DB((env->dbg, LEVEL_2, "\t%+F delay %u\n", curr, d));
456 /* set the etime of all nodes to 0 */
457 set_irn_etime(env, curr, 0);
459 set_irn_preorder(env, curr, cur_pos);
464 * This functions gets called after a node finally has been made ready.
466 static void trace_node_ready(void *data, ir_node *irn, ir_node *pred) {
467 trace_env_t *env = data;
468 sched_timestep_t etime_p, etime;
470 etime = env->curr_time;
472 etime_p = get_irn_etime(env, pred);
473 etime += latency(env, pred, 1, irn, 0);
474 etime = etime_p > etime ? etime_p : etime;
477 set_irn_etime(env, irn, etime);
478 DB((env->dbg, LEVEL_2, "\tset etime of %+F to %u\n", irn, etime));
482 * Update the current time after irn has been selected.
484 static void trace_update_time(void *data, ir_node *irn) {
485 trace_env_t *env = data;
486 if (is_Phi(irn) || get_irn_opcode(irn) == iro_Start) {
487 env->curr_time += get_irn_etime(env, irn);
490 env->curr_time += exectime(env, irn);
495 * Allocates memory and initializes trace scheduling environment.
496 * @param birg The backend irg object
497 * @return The environment
499 static trace_env_t *trace_init(const be_irg_t *birg) {
500 trace_env_t *env = XMALLOCZ(trace_env_t);
501 ir_graph *irg = be_get_birg_irg(birg);
502 int nn = get_irg_last_idx(irg);
505 env->sched_info = NEW_ARR_F(trace_irn_t, nn);
506 env->liveness = be_liveness(irg);
507 FIRM_DBG_REGISTER(env->dbg, "firm.be.sched.trace");
509 be_liveness_assure_chk(env->liveness);
510 memset(env->sched_info, 0, nn * sizeof(*(env->sched_info)));
516 * Frees all memory allocated for trace scheduling environment.
517 * @param env The environment
519 static void trace_free(void *data) {
520 trace_env_t *env = data;
521 be_liveness_free(env->liveness);
522 DEL_ARR_F(env->sched_info);
527 * Simple selector. Just assure that jumps are scheduled last.
529 static ir_node *basic_selection(ir_nodeset_t *ready_set)
532 ir_nodeset_iterator_t iter;
534 /* assure that branches and constants are executed last */
535 foreach_ir_nodeset(ready_set, irn, iter) {
536 if (!arch_irn_class_is(irn, branch)) {
541 /* at last: schedule branches */
542 irn = get_nodeset_node(ready_set);
548 * The muchnik selector.
550 static ir_node *muchnik_select(void *block_env, ir_nodeset_t *ready_set, ir_nodeset_t *live_set)
552 trace_env_t *env = block_env;
553 ir_nodeset_t mcands, ecands;
554 ir_nodeset_iterator_t iter;
555 sched_timestep_t max_delay = 0;
559 /* calculate the max delay of all candidates */
560 foreach_ir_nodeset(ready_set, irn, iter) {
561 sched_timestep_t d = get_irn_delay(env, irn);
563 max_delay = d > max_delay ? d : max_delay;
566 ir_nodeset_init_size(&mcands, 8);
567 ir_nodeset_init_size(&ecands, 8);
569 /* build mcands and ecands */
570 foreach_ir_nodeset(ready_set, irn, iter) {
571 if (get_irn_delay(env, irn) == max_delay) {
572 ir_nodeset_insert(&mcands, irn);
573 if (get_irn_etime(env, irn) <= env->curr_time)
574 ir_nodeset_insert(&ecands, irn);
579 if (ir_nodeset_size(&mcands) == 1) {
580 irn = get_nodeset_node(&mcands);
581 DB((env->dbg, LEVEL_3, "\tirn = %+F, mcand = 1, max_delay = %u\n", irn, max_delay));
584 int cnt = ir_nodeset_size(&ecands);
586 irn = get_nodeset_node(&ecands);
588 if (arch_irn_class_is(irn, branch)) {
589 /* BEWARE: don't select a JUMP if others are still possible */
592 DB((env->dbg, LEVEL_3, "\tirn = %+F, ecand = 1, max_delay = %u\n", irn, max_delay));
595 DB((env->dbg, LEVEL_3, "\tecand = %d, max_delay = %u\n", cnt, max_delay));
596 irn = basic_selection(&ecands);
600 DB((env->dbg, LEVEL_3, "\tmcand = %d\n", ir_nodeset_size(&mcands)));
601 irn = basic_selection(&mcands);
608 static void *muchnik_init_graph(const list_sched_selector_t *vtab, const be_irg_t *birg)
610 trace_env_t *env = trace_init(birg);
611 env->selector = vtab;
612 env->selector_env = (void*) be_get_birg_arch_env(birg);
616 static void *muchnik_init_block(void *graph_env, ir_node *bl)
618 trace_preprocess_block(graph_env, bl);
622 const list_sched_selector_t muchnik_selector = {
626 NULL, /* to_appear_in_schedule */
627 trace_node_ready, /* node_ready */
628 trace_update_time, /* node_selected */
631 NULL, /* finish_block */
632 trace_free /* finish_graph */
636 * Execute the heuristic function.
638 static ir_node *heuristic_select(void *block_env, ir_nodeset_t *ns, ir_nodeset_t *lv)
640 trace_env_t *trace_env = block_env;
641 ir_node *irn, *cand = NULL;
642 int max_prio = INT_MIN;
643 int cur_prio = INT_MIN;
644 int cur_pressure = ir_nodeset_size(lv);
645 int reg_fact, cand_reg_fact;
646 ir_nodeset_iterator_t iter;
648 /* prefer instructions which can be scheduled early */
650 /* prefer instructions with lots of successors */
651 #define PRIO_NUMSUCCS 8
652 /* prefer instructions with long critical path */
653 #define PRIO_LEVEL 12
654 /* prefer instructions coming early in preorder */
655 #define PRIO_PREORD 8
656 /* weight of current register pressure */
657 #define PRIO_CUR_PRESS 20
658 /* weight of register pressure difference */
659 #define PRIO_CHG_PRESS 8
661 /* priority based selection, heuristic inspired by mueller diss */
662 foreach_ir_nodeset(ns, irn, iter) {
663 /* make sure that branches are scheduled last */
664 if (!arch_irn_class_is(irn, branch)) {
665 int rdiff = get_irn_reg_diff(trace_env, irn);
666 int sign = rdiff < 0;
667 int chg = (rdiff < 0 ? -rdiff : rdiff) << PRIO_CHG_PRESS;
669 /* reg_fact = chg << cur_pressure; */
670 reg_fact = chg * cur_pressure;
672 reg_fact = INT_MAX - 2;
673 reg_fact = sign ? -reg_fact : reg_fact;
675 cur_prio = (get_irn_critical_path_len(trace_env, irn) << PRIO_LEVEL)
676 //- (get_irn_delay(trace_env, irn) << PRIO_LEVEL)
677 + (get_irn_num_user(trace_env, irn) << PRIO_NUMSUCCS)
678 - (get_irn_etime(trace_env, irn) << PRIO_TIME)
679 //- ((get_irn_reg_diff(trace_env, irn) >> PRIO_CHG_PRESS) << ((cur_pressure >> PRIO_CUR_PRESS) - 3))
681 + (get_irn_preorder(trace_env, irn) << PRIO_PREORD); /* high preorder means early schedule */
682 if (cur_prio > max_prio) {
685 cand_reg_fact = reg_fact;
688 DBG((trace_env->dbg, LEVEL_4, "checked NODE %+F\n", irn));
689 DBG((trace_env->dbg, LEVEL_4, "\tpriority: %d\n", cur_prio));
690 DBG((trace_env->dbg, LEVEL_4, "\tpath len: %d (%d)\n", get_irn_critical_path_len(trace_env, irn), get_irn_critical_path_len(trace_env, irn) << PRIO_LEVEL));
691 DBG((trace_env->dbg, LEVEL_4, "\tdelay: %d (%d)\n", get_irn_delay(trace_env, irn), get_irn_delay(trace_env, irn) << PRIO_LEVEL));
692 DBG((trace_env->dbg, LEVEL_4, "\t#user: %d (%d)\n", get_irn_num_user(trace_env, irn), get_irn_num_user(trace_env, irn) << PRIO_NUMSUCCS));
693 DBG((trace_env->dbg, LEVEL_4, "\tetime: %d (%d)\n", get_irn_etime(trace_env, irn), 0 - (get_irn_etime(trace_env, irn) << PRIO_TIME)));
694 DBG((trace_env->dbg, LEVEL_4, "\tpreorder: %d (%d)\n", get_irn_preorder(trace_env, irn), get_irn_preorder(trace_env, irn) << PRIO_PREORD));
695 DBG((trace_env->dbg, LEVEL_4, "\treg diff: %d (%d)\n", get_irn_reg_diff(trace_env, irn), 0 - reg_fact));
696 DBG((trace_env->dbg, LEVEL_4, "\tpressure: %d\n", cur_pressure));
701 DBG((trace_env->dbg, LEVEL_4, "heuristic selected %+F:\n", cand));
704 cand = basic_selection(ns);
710 const list_sched_selector_t heuristic_selector = {
714 NULL, /* to_appear_in_schedule */
715 trace_node_ready, /* node_ready */
716 trace_update_time, /* node_selected */
719 NULL, /* finish_block */
720 trace_free /* finish_graph */