2 * Copyright (C) 1995-2007 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
33 #include "iredges_t.h"
35 #include "besched_t.h"
36 #include "belistsched.h"
39 /* we need a special mark */
43 typedef struct _trace_irn {
44 sched_timestep_t delay; /**< The delay for this node if already calculated, else 0. */
45 sched_timestep_t etime; /**< The earliest time of this node. */
46 unsigned num_user; /**< The number real users (mode datab) of this node */
47 int reg_diff; /**< The difference of num(out registers) - num(in registers) */
48 int preorder; /**< The pre-order position */
49 unsigned critical_path_len; /**< The weighted length of the longest critical path */
50 unsigned is_root : 1; /**< is a root node of a block */
53 typedef struct _trace_env {
54 trace_irn_t *sched_info; /**< trace scheduling information about the nodes */
55 const arch_env_t *arch_env; /**< the arch environment */
56 sched_timestep_t curr_time; /**< current time of the scheduler */
57 void *selector_env; /**< the backend selector environment */
58 const list_sched_selector_t *selector; /**< the actual backend selector */
59 be_lv_t *liveness; /**< The liveness for the irg */
60 DEBUG_ONLY(firm_dbg_module_t *dbg;)
64 * Returns a random node from a nodeset
66 static ir_node *get_nodeset_node(const ir_nodeset_t *nodeset)
68 ir_nodeset_iterator_t iter;
70 ir_nodeset_iterator_init(&iter, nodeset);
71 return ir_nodeset_iterator_next(&iter);
75 * Returns non-zero if the node is a root node
77 static INLINE unsigned is_root_node(trace_env_t *env, ir_node *n)
79 int idx = get_irn_idx(n);
81 assert(idx < ARR_LEN(env->sched_info));
82 return env->sched_info[idx].is_root;
86 * Mark a node as root node
88 static INLINE void mark_root_node(trace_env_t *env, ir_node *n)
90 int idx = get_irn_idx(n);
92 assert(idx < ARR_LEN(env->sched_info));
93 env->sched_info[idx].is_root = 1;
97 * Get the current delay.
99 static INLINE sched_timestep_t get_irn_delay(trace_env_t *env, ir_node *n) {
100 int idx = get_irn_idx(n);
102 assert(idx < ARR_LEN(env->sched_info));
103 return env->sched_info[idx].delay;
107 * Set the current delay.
109 static INLINE void set_irn_delay(trace_env_t *env, ir_node *n, sched_timestep_t delay) {
110 int idx = get_irn_idx(n);
112 assert(idx < ARR_LEN(env->sched_info));
113 env->sched_info[idx].delay = delay;
117 * Get the current etime.
119 static INLINE sched_timestep_t get_irn_etime(trace_env_t *env, ir_node *n) {
120 int idx = get_irn_idx(n);
122 assert(idx < ARR_LEN(env->sched_info));
123 return env->sched_info[idx].etime;
127 * Set the current etime.
129 static INLINE void set_irn_etime(trace_env_t *env, ir_node *n, sched_timestep_t etime) {
130 int idx = get_irn_idx(n);
132 assert(idx < ARR_LEN(env->sched_info));
133 env->sched_info[idx].etime = etime;
137 * Get the number of users.
139 static INLINE unsigned get_irn_num_user(trace_env_t *env, ir_node *n) {
140 int idx = get_irn_idx(n);
142 assert(idx < ARR_LEN(env->sched_info));
143 return env->sched_info[idx].num_user;
147 * Set the number of users.
149 static INLINE void set_irn_num_user(trace_env_t *env, ir_node *n, unsigned num_user) {
150 int idx = get_irn_idx(n);
152 assert(idx < ARR_LEN(env->sched_info));
153 env->sched_info[idx].num_user = num_user;
157 * Get the register difference.
159 static INLINE int get_irn_reg_diff(trace_env_t *env, ir_node *n) {
160 int 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) {
170 int idx = get_irn_idx(n);
172 assert(idx < ARR_LEN(env->sched_info));
173 env->sched_info[idx].reg_diff = reg_diff;
177 * Get the pre-order position.
179 static INLINE int get_irn_preorder(trace_env_t *env, ir_node *n) {
180 int idx = get_irn_idx(n);
182 assert(idx < ARR_LEN(env->sched_info));
183 return env->sched_info[idx].preorder;
187 * Set the pre-order position.
189 static INLINE void set_irn_preorder(trace_env_t *env, ir_node *n, int pos) {
190 int idx = get_irn_idx(n);
192 assert(idx < ARR_LEN(env->sched_info));
193 env->sched_info[idx].preorder = pos;
197 * Get the pre-order position.
199 static INLINE unsigned get_irn_critical_path_len(trace_env_t *env, ir_node *n) {
200 int idx = get_irn_idx(n);
202 assert(idx < ARR_LEN(env->sched_info));
203 return env->sched_info[idx].critical_path_len;
207 * Set the pre-order position.
209 static INLINE void set_irn_critical_path_len(trace_env_t *env, ir_node *n, unsigned len) {
210 int idx = get_irn_idx(n);
212 assert(idx < ARR_LEN(env->sched_info));
213 env->sched_info[idx].critical_path_len = len;
217 * returns the exec-time for node n.
219 static sched_timestep_t exectime(trace_env_t *env, ir_node *n) {
220 if (be_is_Keep(n) || is_Proj(n))
222 if (env->selector->exectime)
223 return env->selector->exectime(env->selector_env, n);
228 * Calculates the latency for between two ops
230 static sched_timestep_t latency(trace_env_t *env, ir_node *pred, int pred_cycle, ir_node *curr, int curr_cycle) {
231 /* a Keep hides a root */
232 if (be_is_Keep(curr))
233 return exectime(env, pred);
235 /* Proj's are executed immediately */
239 /* predecessors Proj's must be skipped */
241 pred = get_Proj_pred(pred);
243 if (env->selector->latency)
244 return env->selector->latency(env->selector_env, pred, pred_cycle, curr, curr_cycle);
249 * Returns the number of users of a node having mode datab.
251 static int get_num_successors(ir_node *irn) {
253 const ir_edge_t *edge;
255 if (get_irn_mode(irn) == mode_T) {
256 /* for mode_T nodes: count the users of all Projs */
257 foreach_out_edge(irn, edge) {
258 ir_node *proj = get_edge_src_irn(edge);
259 ir_mode *mode = get_irn_mode(proj);
262 sum += get_num_successors(proj);
263 else if (mode_is_datab(mode))
264 sum += get_irn_n_edges(proj);
268 /* do not count keep-alive edges */
269 foreach_out_edge(irn, edge) {
270 if (get_irn_opcode(get_edge_src_irn(edge)) != iro_End)
279 * Returns the difference of regs_output - regs_input;
281 static int get_reg_difference(trace_env_t *env, ir_node *irn) {
285 ir_node *block = get_nodes_block(irn);
287 if (be_is_Call(irn)) {
288 /* we want calls prefered */
292 if (get_irn_mode(irn) == mode_T) {
293 /* mode_T nodes: num out regs == num Projs with mode datab */
294 const ir_edge_t *edge;
295 foreach_out_edge(irn, edge) {
296 ir_node *proj = get_edge_src_irn(edge);
297 if (mode_is_datab(get_irn_mode(proj)))
304 /* num in regs: number of ins with mode datab and not ignore */
305 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
306 ir_node *in = get_irn_n(irn, i);
308 if (mode_is_datab(get_irn_mode(in)) && /* must be data node */
309 ! arch_irn_is(env->arch_env, in, ignore) && /* ignore "ignore" nodes :) */
310 ! be_is_live_end(env->liveness, block, in) /* if the value lives outside of block: do not count */
317 return num_out - num_in;
321 * descent into a dag and create a pre-order list.
323 static void descent(ir_node *root, ir_node *block, ir_node **list, trace_env_t *env, unsigned path_len) {
326 if (! is_Phi(root)) {
327 path_len += exectime(env, root);
328 if (get_irn_critical_path_len(env, root) < path_len) {
329 set_irn_critical_path_len(env, root, path_len);
331 /* calculate number of users (needed for heuristic) */
332 set_irn_num_user(env, root, get_num_successors(root));
334 /* calculate register difference (needed for heuristic) */
335 set_irn_reg_diff(env, root, get_reg_difference(env, root));
337 /* Phi nodes always leave the block */
338 for (i = get_irn_arity(root) - 1; i >= 0; --i) {
339 ir_node *pred = get_irn_n(root, i);
341 DBG((env->dbg, LEVEL_3, " node %+F\n", pred));
343 /* Blocks may happen as predecessors of End nodes */
347 /* already seen nodes are not marked */
348 if (get_irn_link(pred) != MARK)
351 /* don't leave our block */
352 if (get_nodes_block(pred) != block)
355 set_irn_link(pred, NULL);
357 descent(pred, block, list, env, path_len);
360 set_irn_link(root, *list);
365 * Returns non-zero if root is a root in the block block.
367 static int is_root(ir_node *root, ir_node *block) {
368 const ir_edge_t *edge;
370 foreach_out_edge(root, edge) {
371 ir_node *succ = get_edge_src_irn(edge);
375 /* Phi nodes are always in "another block */
378 if (get_nodes_block(succ) == block)
385 * Performs initial block calculations for trace scheduling.
387 static void trace_preprocess_block(trace_env_t *env, ir_node *block) {
388 ir_node *root = NULL, *preord = NULL;
391 const ir_edge_t *edge;
393 /* First step: Find the root set. */
394 foreach_out_edge(block, edge) {
395 ir_node *succ = get_edge_src_irn(edge);
399 if (is_root(succ, block)) {
400 mark_root_node(env, succ);
401 set_irn_link(succ, root);
405 set_irn_link(succ, MARK);
408 /* Second step: calculate the pre-order list. */
410 for (curr = root; curr; curr = irn) {
411 irn = get_irn_link(curr);
412 DBG((env->dbg, LEVEL_2, " DAG root %+F\n", curr));
413 descent(curr, block, &preord, env, 0);
417 /* Third step: calculate the Delay. Note that our
418 * list is now in pre-order, starting at root
420 for (cur_pos = 0, curr = root; curr; curr = get_irn_link(curr), cur_pos++) {
423 if (arch_irn_class_is(env->arch_env, curr, branch)) {
424 /* assure, that branches can be executed last */
428 if (is_root_node(env, curr))
429 d = exectime(env, curr);
432 foreach_out_edge(curr, edge) {
433 ir_node *n = get_edge_src_irn(edge);
435 if (get_nodes_block(n) == block) {
438 ld = latency(env, curr, 1, n, 0) + get_irn_delay(env, n);
444 set_irn_delay(env, curr, d);
445 DB((env->dbg, LEVEL_2, "\t%+F delay %u\n", curr, d));
447 /* set the etime of all nodes to 0 */
448 set_irn_etime(env, curr, 0);
450 set_irn_preorder(env, curr, cur_pos);
455 * This functions gets called after a node finally has been made ready.
457 static void trace_node_ready(void *data, ir_node *irn, ir_node *pred) {
458 trace_env_t *env = data;
459 sched_timestep_t etime_p, etime;
461 etime = env->curr_time;
463 etime_p = get_irn_etime(env, pred);
464 etime += latency(env, pred, 1, irn, 0);
465 etime = etime_p > etime ? etime_p : etime;
468 set_irn_etime(env, irn, etime);
469 DB((env->dbg, LEVEL_2, "\tset etime of %+F to %u\n", irn, etime));
473 * Update the current time after irn has been selected.
475 static void trace_update_time(void *data, ir_node *irn) {
476 trace_env_t *env = data;
477 if (is_Phi(irn) || get_irn_opcode(irn) == iro_Start) {
478 env->curr_time += get_irn_etime(env, irn);
481 env->curr_time += exectime(env, irn);
486 * Allocates memory and initializes trace scheduling environment.
487 * @param birg The backend irg object
488 * @return The environment
490 static trace_env_t *trace_init(const be_irg_t *birg) {
491 trace_env_t *env = xcalloc(1, sizeof(*env));
492 ir_graph *irg = be_get_birg_irg(birg);
493 int nn = get_irg_last_idx(irg);
495 env->arch_env = be_get_birg_arch_env(birg);
497 env->sched_info = NEW_ARR_F(trace_irn_t, nn);
498 env->liveness = be_liveness(birg);
499 FIRM_DBG_REGISTER(env->dbg, "firm.be.sched.trace");
501 be_liveness_assure_chk(env->liveness);
502 memset(env->sched_info, 0, nn * sizeof(*(env->sched_info)));
508 * Frees all memory allocated for trace scheduling environment.
509 * @param env The environment
511 static void trace_free(void *data) {
512 trace_env_t *env = data;
513 be_liveness_free(env->liveness);
514 DEL_ARR_F(env->sched_info);
519 * Simple selector. Just assure that jumps are scheduled last.
521 static ir_node *basic_selection(const arch_env_t *arch_env, ir_nodeset_t *ready_set) {
523 ir_nodeset_iterator_t iter;
525 /* assure that branches and constants are executed last */
526 foreach_ir_nodeset(ready_set, irn, iter) {
527 if (! arch_irn_class_is(arch_env, irn, branch)) {
532 /* at last: schedule branches */
533 irn = get_nodeset_node(ready_set);
539 * The muchnik selector.
541 static ir_node *muchnik_select(void *block_env, ir_nodeset_t *ready_set, ir_nodeset_t *live_set)
543 trace_env_t *env = block_env;
544 ir_nodeset_t mcands, ecands;
545 ir_nodeset_iterator_t iter;
546 sched_timestep_t max_delay = 0;
550 /* calculate the max delay of all candidates */
551 foreach_ir_nodeset(ready_set, irn, iter) {
552 sched_timestep_t d = get_irn_delay(env, irn);
554 max_delay = d > max_delay ? d : max_delay;
557 ir_nodeset_init_size(&mcands, 8);
558 ir_nodeset_init_size(&ecands, 8);
560 /* build mcands and ecands */
561 foreach_ir_nodeset(ready_set, irn, iter) {
562 if (get_irn_delay(env, irn) == max_delay) {
563 ir_nodeset_insert(&mcands, irn);
564 if (get_irn_etime(env, irn) <= env->curr_time)
565 ir_nodeset_insert(&ecands, irn);
570 if (ir_nodeset_size(&mcands) == 1) {
571 irn = get_nodeset_node(&mcands);
572 DB((env->dbg, LEVEL_3, "\tirn = %+F, mcand = 1, max_delay = %u\n", irn, max_delay));
575 int cnt = ir_nodeset_size(&ecands);
577 irn = get_nodeset_node(&ecands);
579 if (arch_irn_class_is(env->arch_env, irn, branch)) {
580 /* BEWARE: don't select a JUMP if others are still possible */
583 DB((env->dbg, LEVEL_3, "\tirn = %+F, ecand = 1, max_delay = %u\n", irn, max_delay));
586 DB((env->dbg, LEVEL_3, "\tecand = %d, max_delay = %u\n", cnt, max_delay));
587 irn = basic_selection(env->arch_env, &ecands);
591 DB((env->dbg, LEVEL_3, "\tmcand = %d\n", ir_nodeset_size(&mcands)));
592 irn = basic_selection(env->arch_env, &mcands);
599 static void *muchnik_init_graph(const list_sched_selector_t *vtab, const be_irg_t *birg)
601 trace_env_t *env = trace_init(birg);
602 env->selector = vtab;
603 env->selector_env = (void*) be_get_birg_arch_env(birg);
607 static void *muchnik_init_block(void *graph_env, ir_node *bl)
609 trace_preprocess_block(graph_env, bl);
613 const list_sched_selector_t muchnik_selector = {
617 NULL, /* to_appear_in_schedule */
618 trace_node_ready, /* node_ready */
619 trace_update_time, /* node_selected */
622 NULL, /* finish_block */
623 trace_free /* finish_graph */
627 * Execute the heuristic function.
629 static ir_node *heuristic_select(void *block_env, ir_nodeset_t *ns, ir_nodeset_t *lv)
631 trace_env_t *trace_env = block_env;
632 ir_node *irn, *cand = NULL;
633 int max_prio = INT_MIN;
634 int cur_prio = INT_MIN;
635 int cur_pressure = ir_nodeset_size(lv);
636 int reg_fact, cand_reg_fact;
637 ir_nodeset_iterator_t iter;
639 /* prefer instructions which can be scheduled early */
641 /* prefer instructions with lots of successors */
642 #define PRIO_NUMSUCCS 8
643 /* prefer instructions with long critical path */
644 #define PRIO_LEVEL 12
645 /* prefer instructions coming early in preorder */
646 #define PRIO_PREORD 8
647 /* weight of current register pressure */
648 #define PRIO_CUR_PRESS 20
649 /* weight of register pressure difference */
650 #define PRIO_CHG_PRESS 8
652 /* priority based selection, heuristic inspired by mueller diss */
653 foreach_ir_nodeset(ns, irn, iter) {
654 /* make sure that branches are scheduled last */
655 if (! arch_irn_class_is(trace_env->arch_env, irn, branch)) {
656 int rdiff = get_irn_reg_diff(trace_env, irn);
657 int sign = rdiff < 0;
658 int chg = (rdiff < 0 ? -rdiff : rdiff) << PRIO_CHG_PRESS;
660 /* reg_fact = chg << cur_pressure; */
661 reg_fact = chg * cur_pressure;
663 reg_fact = INT_MAX - 2;
664 reg_fact = sign ? -reg_fact : reg_fact;
666 cur_prio = (get_irn_critical_path_len(trace_env, irn) << PRIO_LEVEL)
667 //- (get_irn_delay(trace_env, irn) << PRIO_LEVEL)
668 + (get_irn_num_user(trace_env, irn) << PRIO_NUMSUCCS)
669 - (get_irn_etime(trace_env, irn) << PRIO_TIME)
670 //- ((get_irn_reg_diff(trace_env, irn) >> PRIO_CHG_PRESS) << ((cur_pressure >> PRIO_CUR_PRESS) - 3))
672 + (get_irn_preorder(trace_env, irn) << PRIO_PREORD); /* high preorder means early schedule */
673 if (cur_prio > max_prio) {
676 cand_reg_fact = reg_fact;
679 DBG((trace_env->dbg, LEVEL_4, "checked NODE %+F\n", irn));
680 DBG((trace_env->dbg, LEVEL_4, "\tpriority: %d\n", cur_prio));
681 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));
682 DBG((trace_env->dbg, LEVEL_4, "\tdelay: %d (%d)\n", get_irn_delay(trace_env, irn), get_irn_delay(trace_env, irn) << PRIO_LEVEL));
683 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));
684 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)));
685 DBG((trace_env->dbg, LEVEL_4, "\tpreorder: %d (%d)\n", get_irn_preorder(trace_env, irn), get_irn_preorder(trace_env, irn) << PRIO_PREORD));
686 DBG((trace_env->dbg, LEVEL_4, "\treg diff: %d (%d)\n", get_irn_reg_diff(trace_env, irn), 0 - reg_fact));
687 DBG((trace_env->dbg, LEVEL_4, "\tpressure: %d\n", cur_pressure));
692 DBG((trace_env->dbg, LEVEL_4, "heuristic selected %+F:\n", cand));
695 cand = basic_selection(trace_env->arch_env, ns);
701 const list_sched_selector_t heuristic_selector = {
705 NULL, /* to_appear_in_schedule */
706 trace_node_ready, /* node_ready */
707 trace_update_time, /* node_selected */
710 NULL, /* finish_block */
711 trace_free /* finish_graph */