2 * Copyright (C) 1995-2011 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
30 #include "iredges_t.h"
33 #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 be_lv_t *liveness; /**< The liveness for the irg */
56 DEBUG_ONLY(firm_dbg_module_t *dbg;)
60 * Returns a random node from a nodeset
62 static ir_node *get_nodeset_node(const ir_nodeset_t *nodeset)
64 return ir_nodeset_first(nodeset);
68 * Returns non-zero if the node is a root node
70 static inline unsigned is_root_node(trace_env_t *env, ir_node *n)
72 unsigned const idx = get_irn_idx(n);
74 assert(idx < ARR_LEN(env->sched_info));
75 return env->sched_info[idx].is_root;
79 * Mark a node as root node
81 static inline void mark_root_node(trace_env_t *env, ir_node *n)
83 unsigned const idx = get_irn_idx(n);
85 assert(idx < ARR_LEN(env->sched_info));
86 env->sched_info[idx].is_root = 1;
90 * Get the current delay.
92 static inline sched_timestep_t get_irn_delay(trace_env_t *env, ir_node *n)
94 unsigned const idx = get_irn_idx(n);
96 assert(idx < ARR_LEN(env->sched_info));
97 return env->sched_info[idx].delay;
101 * Set the current delay.
103 static inline void set_irn_delay(trace_env_t *env, ir_node *n, sched_timestep_t delay)
105 unsigned const idx = get_irn_idx(n);
107 assert(idx < ARR_LEN(env->sched_info));
108 env->sched_info[idx].delay = delay;
112 * Get the current etime.
114 static inline sched_timestep_t get_irn_etime(trace_env_t *env, ir_node *n)
116 unsigned const idx = get_irn_idx(n);
118 assert(idx < ARR_LEN(env->sched_info));
119 return env->sched_info[idx].etime;
123 * Set the current etime.
125 static inline void set_irn_etime(trace_env_t *env, ir_node *n, sched_timestep_t etime)
127 unsigned const idx = get_irn_idx(n);
129 assert(idx < ARR_LEN(env->sched_info));
130 env->sched_info[idx].etime = etime;
134 * Get the number of users.
136 static inline unsigned get_irn_num_user(trace_env_t *env, ir_node *n)
138 unsigned const 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)
149 unsigned const idx = get_irn_idx(n);
151 assert(idx < ARR_LEN(env->sched_info));
152 env->sched_info[idx].num_user = num_user;
156 * Get the register difference.
158 static inline int get_irn_reg_diff(trace_env_t *env, ir_node *n)
160 unsigned const idx = get_irn_idx(n);
162 assert(idx < ARR_LEN(env->sched_info));
163 return env->sched_info[idx].reg_diff;
167 * Set the register difference.
169 static inline void set_irn_reg_diff(trace_env_t *env, ir_node *n, int reg_diff)
171 unsigned const idx = get_irn_idx(n);
173 assert(idx < ARR_LEN(env->sched_info));
174 env->sched_info[idx].reg_diff = reg_diff;
178 * Get the pre-order position.
180 static inline int get_irn_preorder(trace_env_t *env, ir_node *n)
182 unsigned const idx = get_irn_idx(n);
184 assert(idx < ARR_LEN(env->sched_info));
185 return env->sched_info[idx].preorder;
189 * Set the pre-order position.
191 static inline void set_irn_preorder(trace_env_t *env, ir_node *n, int pos)
193 unsigned const idx = get_irn_idx(n);
195 assert(idx < ARR_LEN(env->sched_info));
196 env->sched_info[idx].preorder = pos;
200 * Get the pre-order position.
202 static inline unsigned get_irn_critical_path_len(trace_env_t *env, ir_node *n)
204 unsigned const idx = get_irn_idx(n);
206 assert(idx < ARR_LEN(env->sched_info));
207 return env->sched_info[idx].critical_path_len;
211 * Set the pre-order position.
213 static inline void set_irn_critical_path_len(trace_env_t *env, ir_node *n, unsigned len)
215 unsigned const idx = get_irn_idx(n);
217 assert(idx < ARR_LEN(env->sched_info));
218 env->sched_info[idx].critical_path_len = len;
222 * returns the exec-time for node n.
224 static sched_timestep_t exectime(trace_env_t *env, ir_node *n)
227 if (be_is_Keep(n) || is_Proj(n))
230 if (env->selector->exectime)
231 return env->selector->exectime(env->selector_env, n);
237 * Calculates the latency for between two ops
239 static sched_timestep_t latency(trace_env_t *env, ir_node *pred, int pred_cycle, ir_node *curr, int curr_cycle)
243 /* a Keep hides a root */
244 if (be_is_Keep(curr))
245 return exectime(env, pred);
247 /* Proj's are executed immediately */
251 /* predecessors Proj's must be skipped */
253 pred = get_Proj_pred(pred);
256 if (env->selector->latency)
257 return env->selector->latency(env->selector_env, pred, pred_cycle, curr, curr_cycle);
264 * Returns the number of users of a node having mode datab.
266 static int get_num_successors(ir_node *irn)
270 if (get_irn_mode(irn) == mode_T) {
271 /* for mode_T nodes: count the users of all Projs */
272 foreach_out_edge(irn, edge) {
273 ir_node *proj = get_edge_src_irn(edge);
274 ir_mode *mode = get_irn_mode(proj);
277 sum += get_num_successors(proj);
278 else if (mode_is_datab(mode))
279 sum += get_irn_n_edges(proj);
283 /* do not count keep-alive edges */
284 foreach_out_edge(irn, edge) {
285 if (get_irn_opcode(get_edge_src_irn(edge)) != iro_End)
294 * Returns the difference of regs_output - regs_input;
296 static int get_reg_difference(trace_env_t *env, ir_node *irn)
301 ir_node *block = get_nodes_block(irn);
303 if (be_is_Call(irn)) {
304 /* we want calls preferred */
308 if (get_irn_mode(irn) == mode_T) {
309 /* mode_T nodes: num out regs == num Projs with mode datab */
310 foreach_out_edge(irn, edge) {
311 ir_node *proj = get_edge_src_irn(edge);
312 if (mode_is_datab(get_irn_mode(proj)))
319 /* num in regs: number of ins with mode datab and not ignore */
320 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
321 ir_node *in = get_irn_n(irn, i);
323 if (!mode_is_datab(get_irn_mode(in)))
326 if (arch_irn_is_ignore(in))
329 if (be_is_live_end(env->liveness, block, in))
335 return num_out - num_in;
339 * descent into a dag and create a pre-order list.
341 static void descent(ir_node *root, ir_node *block, ir_node **list, trace_env_t *env, unsigned path_len)
345 if (! is_Phi(root)) {
346 path_len += exectime(env, root);
347 if (get_irn_critical_path_len(env, root) < path_len) {
348 set_irn_critical_path_len(env, root, path_len);
350 /* calculate number of users (needed for heuristic) */
351 set_irn_num_user(env, root, get_num_successors(root));
353 /* calculate register difference (needed for heuristic) */
354 set_irn_reg_diff(env, root, get_reg_difference(env, root));
356 /* Phi nodes always leave the block */
357 for (i = get_irn_arity(root) - 1; i >= 0; --i) {
358 ir_node *pred = get_irn_n(root, i);
360 DBG((env->dbg, LEVEL_3, " node %+F\n", pred));
362 /* Blocks may happen as predecessors of End nodes */
366 /* already seen nodes are not marked */
367 if (get_irn_link(pred) != MARK)
370 /* don't leave our block */
371 if (get_nodes_block(pred) != block)
374 set_irn_link(pred, NULL);
376 descent(pred, block, list, env, path_len);
379 set_irn_link(root, *list);
384 * Returns non-zero if root is a root in the block block.
386 static int is_root(ir_node *root, ir_node *block)
388 foreach_out_edge(root, edge) {
389 ir_node *succ = get_edge_src_irn(edge);
393 /* Phi nodes are always in "another block */
396 if (get_nodes_block(succ) == block)
403 * Performs initial block calculations for trace scheduling.
405 static void trace_preprocess_block(trace_env_t *env, ir_node *block)
407 ir_node *root = NULL, *preord = NULL;
411 /* First step: Find the root set. */
412 foreach_out_edge(block, edge) {
413 ir_node *succ = get_edge_src_irn(edge);
415 if (is_Anchor(succ)) {
416 /* ignore a keep alive edge */
419 if (is_root(succ, block)) {
420 mark_root_node(env, succ);
421 set_irn_link(succ, root);
425 set_irn_link(succ, MARK);
428 /* Second step: calculate the pre-order list. */
430 for (curr = root; curr; curr = irn) {
431 irn = (ir_node*)get_irn_link(curr);
432 DBG((env->dbg, LEVEL_2, " DAG root %+F\n", curr));
433 descent(curr, block, &preord, env, 0);
437 /* Third step: calculate the Delay. Note that our
438 * list is now in pre-order, starting at root
440 for (cur_pos = 0, curr = root; curr; curr = (ir_node*)get_irn_link(curr), cur_pos++) {
444 /* assure, that branches can be executed last */
448 if (is_root_node(env, curr))
449 d = exectime(env, curr);
452 foreach_out_edge(curr, edge) {
453 ir_node *n = get_edge_src_irn(edge);
455 if (get_nodes_block(n) == block) {
458 ld = latency(env, curr, 1, n, 0) + get_irn_delay(env, n);
464 set_irn_delay(env, curr, d);
465 DB((env->dbg, LEVEL_2, "\t%+F delay %u\n", curr, d));
467 /* set the etime of all nodes to 0 */
468 set_irn_etime(env, curr, 0);
470 set_irn_preorder(env, curr, cur_pos);
475 * This functions gets called after a node finally has been made ready.
477 static void trace_node_ready(void *data, ir_node *irn, ir_node *pred)
479 trace_env_t *env = (trace_env_t*)data;
480 sched_timestep_t etime_p, etime;
482 etime = env->curr_time;
484 etime_p = get_irn_etime(env, pred);
485 etime += latency(env, pred, 1, irn, 0);
486 etime = etime_p > etime ? etime_p : etime;
489 set_irn_etime(env, irn, etime);
490 DB((env->dbg, LEVEL_2, "\tset etime of %+F to %u\n", irn, etime));
494 * Update the current time after irn has been selected.
496 static void trace_update_time(void *data, ir_node *irn)
498 trace_env_t *env = (trace_env_t*)data;
499 if (is_Phi(irn) || get_irn_opcode(irn) == beo_Start) {
500 env->curr_time += get_irn_etime(env, irn);
503 env->curr_time += exectime(env, irn);
508 * Allocates memory and initializes trace scheduling environment.
509 * @param irg The backend irg object
510 * @return The environment
512 static trace_env_t *trace_init(ir_graph *irg)
514 trace_env_t *env = XMALLOCZ(trace_env_t);
515 int nn = get_irg_last_idx(irg);
518 env->sched_info = NEW_ARR_F(trace_irn_t, nn);
519 env->liveness = be_get_irg_liveness(irg);
520 FIRM_DBG_REGISTER(env->dbg, "firm.be.sched.trace");
522 be_assure_live_chk(irg);
523 memset(env->sched_info, 0, nn * sizeof(*(env->sched_info)));
529 * Frees all memory allocated for trace scheduling environment.
530 * @param env The environment
532 static void trace_free(void *data)
534 trace_env_t *env = (trace_env_t*)data;
535 DEL_ARR_F(env->sched_info);
540 * Simple selector. Just assure that jumps are scheduled last.
542 static ir_node *basic_selection(ir_nodeset_t *ready_set)
544 /* assure that branches and constants are executed last */
545 foreach_ir_nodeset(ready_set, irn, iter) {
551 /* at last: schedule branches */
552 return get_nodeset_node(ready_set);
556 * The muchnik selector.
558 static ir_node *muchnik_select(void *block_env, ir_nodeset_t *ready_set)
560 trace_env_t *env = (trace_env_t*)block_env;
561 ir_nodeset_t mcands, ecands;
562 sched_timestep_t max_delay = 0;
564 /* calculate the max delay of all candidates */
565 foreach_ir_nodeset(ready_set, irn, iter) {
566 sched_timestep_t d = get_irn_delay(env, irn);
568 max_delay = d > max_delay ? d : max_delay;
571 ir_nodeset_init_size(&mcands, 8);
572 ir_nodeset_init_size(&ecands, 8);
574 /* build mcands and ecands */
575 foreach_ir_nodeset(ready_set, irn, iter) {
576 if (get_irn_delay(env, irn) == max_delay) {
577 ir_nodeset_insert(&mcands, irn);
578 if (get_irn_etime(env, irn) <= env->curr_time)
579 ir_nodeset_insert(&ecands, irn);
585 if (ir_nodeset_size(&mcands) == 1) {
586 irn = get_nodeset_node(&mcands);
587 DB((env->dbg, LEVEL_3, "\tirn = %+F, mcand = 1, max_delay = %u\n", irn, max_delay));
590 size_t cnt = ir_nodeset_size(&ecands);
592 irn = get_nodeset_node(&ecands);
595 /* BEWARE: don't select a JUMP if others are still possible */
598 DB((env->dbg, LEVEL_3, "\tirn = %+F, ecand = 1, max_delay = %u\n", irn, max_delay));
601 DB((env->dbg, LEVEL_3, "\tecand = %zu, max_delay = %u\n", cnt, max_delay));
602 irn = basic_selection(&ecands);
606 DB((env->dbg, LEVEL_3, "\tmcand = %zu\n", ir_nodeset_size(&mcands)));
607 irn = basic_selection(&mcands);
614 static void *muchnik_init_graph(ir_graph *irg)
616 trace_env_t *env = trace_init(irg);
620 static void *muchnik_init_block(void *graph_env, ir_node *bl)
622 trace_env_t *env = (trace_env_t*) graph_env;
623 trace_preprocess_block(env, bl);
627 static void sched_muchnik(ir_graph *irg)
629 static const list_sched_selector_t muchnik_selector = {
633 trace_node_ready, /* node_ready */
634 trace_update_time, /* node_selected */
635 NULL, /* finish_block */
636 trace_free /* finish_graph */
638 be_list_sched_graph(irg, &muchnik_selector);
642 * Execute the heuristic function.
644 static ir_node *heuristic_select(void *block_env, ir_nodeset_t *ns)
646 trace_env_t *trace_env = (trace_env_t*)block_env;
647 ir_node *cand = NULL;
648 int max_prio = INT_MIN;
649 int cur_prio = INT_MIN;
651 /* Note: register pressure calculation needs an overhaul, you need correct
652 * tracking for each register class indidually and weight by each class
653 int cur_pressure = ir_nodeset_size(lv); */
654 int cur_pressure = 1;
656 /* prefer instructions which can be scheduled early */
658 /* prefer instructions with lots of successors */
659 #define PRIO_NUMSUCCS 8
660 /* prefer instructions with long critical path */
661 #define PRIO_LEVEL 12
662 /* prefer instructions coming early in preorder */
663 #define PRIO_PREORD 8
664 /* weight of current register pressure */
665 #define PRIO_CUR_PRESS 20
666 /* weight of register pressure difference */
667 #define PRIO_CHG_PRESS 8
669 /* priority based selection, heuristic inspired by mueller diss */
670 foreach_ir_nodeset(ns, irn, iter) {
671 /* make sure that branches are scheduled last */
673 int rdiff = get_irn_reg_diff(trace_env, irn);
674 int sign = rdiff < 0;
675 int chg = (rdiff < 0 ? -rdiff : rdiff) << PRIO_CHG_PRESS;
677 reg_fact = chg * cur_pressure;
679 reg_fact = INT_MAX - 2;
680 reg_fact = sign ? -reg_fact : reg_fact;
682 cur_prio = (get_irn_critical_path_len(trace_env, irn) << PRIO_LEVEL)
683 //- (get_irn_delay(trace_env, irn) << PRIO_LEVEL)
684 + (get_irn_num_user(trace_env, irn) << PRIO_NUMSUCCS)
685 - (get_irn_etime(trace_env, irn) << PRIO_TIME)
686 //- ((get_irn_reg_diff(trace_env, irn) >> PRIO_CHG_PRESS) << ((cur_pressure >> PRIO_CUR_PRESS) - 3))
688 + (get_irn_preorder(trace_env, irn) << PRIO_PREORD); /* high preorder means early schedule */
689 if (cur_prio > max_prio) {
694 DBG((trace_env->dbg, LEVEL_4, "checked NODE %+F\n", irn));
695 DBG((trace_env->dbg, LEVEL_4, "\tpriority: %d\n", cur_prio));
696 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));
697 DBG((trace_env->dbg, LEVEL_4, "\tdelay: %d (%d)\n", get_irn_delay(trace_env, irn), get_irn_delay(trace_env, irn) << PRIO_LEVEL));
698 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));
699 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)));
700 DBG((trace_env->dbg, LEVEL_4, "\tpreorder: %d (%d)\n", get_irn_preorder(trace_env, irn), get_irn_preorder(trace_env, irn) << PRIO_PREORD));
701 DBG((trace_env->dbg, LEVEL_4, "\treg diff: %d (%d)\n", get_irn_reg_diff(trace_env, irn), 0 - reg_fact));
702 DBG((trace_env->dbg, LEVEL_4, "\tpressure: %d\n", cur_pressure));
707 DBG((trace_env->dbg, LEVEL_4, "heuristic selected %+F:\n", cand));
710 cand = basic_selection(ns);
716 static void sched_heuristic(ir_graph *irg)
718 static const list_sched_selector_t heuristic_selector = {
722 trace_node_ready, /* node_ready */
723 trace_update_time, /* node_selected */
724 NULL, /* finish_block */
725 trace_free /* finish_graph */
727 be_list_sched_graph(irg, &heuristic_selector);
730 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_sched_trace)
731 void be_init_sched_trace(void)
733 be_register_scheduler("heur", sched_heuristic);
734 be_register_scheduler("muchnik", sched_muchnik);