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"
33 #include "besched_t.h"
34 #include "belistsched.h"
37 /* we need a special mark */
41 typedef struct _trace_irn {
42 sched_timestep_t delay; /**< The delay for this node if already calculated, else 0. */
43 sched_timestep_t etime; /**< The earliest time of this node. */
44 unsigned num_user; /**< The number real users (mode datab) of this node */
45 int reg_diff; /**< The difference of num(out registers) - num(in registers) */
46 int preorder; /**< The pre-order position */
47 unsigned critical_path_len; /**< The weighted length of the longest critical path */
48 unsigned is_root : 1; /**< is a root node of a block */
51 typedef struct _trace_env {
52 trace_irn_t *sched_info; /**< trace scheduling information about the nodes */
53 sched_timestep_t curr_time; /**< current time of the scheduler */
54 void *selector_env; /**< the backend selector environment */
55 const list_sched_selector_t *selector; /**< the actual backend selector */
56 be_lv_t *liveness; /**< The liveness for the irg */
57 DEBUG_ONLY(firm_dbg_module_t *dbg;)
61 * Returns a random node from a nodeset
63 static ir_node *get_nodeset_node(const ir_nodeset_t *nodeset)
65 ir_nodeset_iterator_t iter;
67 ir_nodeset_iterator_init(&iter, nodeset);
68 return ir_nodeset_iterator_next(&iter);
72 * Returns non-zero if the node is a root node
74 static INLINE unsigned is_root_node(trace_env_t *env, ir_node *n)
76 int idx = get_irn_idx(n);
78 assert(idx < ARR_LEN(env->sched_info));
79 return env->sched_info[idx].is_root;
83 * Mark a node as root node
85 static INLINE void mark_root_node(trace_env_t *env, ir_node *n)
87 int idx = get_irn_idx(n);
89 assert(idx < ARR_LEN(env->sched_info));
90 env->sched_info[idx].is_root = 1;
94 * Get the current delay.
96 static INLINE sched_timestep_t get_irn_delay(trace_env_t *env, ir_node *n) {
97 int idx = get_irn_idx(n);
99 assert(idx < ARR_LEN(env->sched_info));
100 return env->sched_info[idx].delay;
104 * Set the current delay.
106 static INLINE void set_irn_delay(trace_env_t *env, ir_node *n, sched_timestep_t delay) {
107 int idx = get_irn_idx(n);
109 assert(idx < ARR_LEN(env->sched_info));
110 env->sched_info[idx].delay = delay;
114 * Get the current etime.
116 static INLINE sched_timestep_t get_irn_etime(trace_env_t *env, ir_node *n) {
117 int idx = get_irn_idx(n);
119 assert(idx < ARR_LEN(env->sched_info));
120 return env->sched_info[idx].etime;
124 * Set the current etime.
126 static INLINE void set_irn_etime(trace_env_t *env, ir_node *n, sched_timestep_t etime) {
127 int 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) {
137 int idx = get_irn_idx(n);
139 assert(idx < ARR_LEN(env->sched_info));
140 return env->sched_info[idx].num_user;
144 * Set the number of users.
146 static INLINE void set_irn_num_user(trace_env_t *env, ir_node *n, unsigned num_user) {
147 int idx = get_irn_idx(n);
149 assert(idx < ARR_LEN(env->sched_info));
150 env->sched_info[idx].num_user = num_user;
154 * Get the register difference.
156 static INLINE int get_irn_reg_diff(trace_env_t *env, ir_node *n) {
157 int idx = get_irn_idx(n);
159 assert(idx < ARR_LEN(env->sched_info));
160 return env->sched_info[idx].reg_diff;
164 * Set the register difference.
166 static INLINE void set_irn_reg_diff(trace_env_t *env, ir_node *n, int reg_diff) {
167 int idx = get_irn_idx(n);
169 assert(idx < ARR_LEN(env->sched_info));
170 env->sched_info[idx].reg_diff = reg_diff;
174 * Get the pre-order position.
176 static INLINE int get_irn_preorder(trace_env_t *env, ir_node *n) {
177 int idx = get_irn_idx(n);
179 assert(idx < ARR_LEN(env->sched_info));
180 return env->sched_info[idx].preorder;
184 * Set the pre-order position.
186 static INLINE void set_irn_preorder(trace_env_t *env, ir_node *n, int pos) {
187 int idx = get_irn_idx(n);
189 assert(idx < ARR_LEN(env->sched_info));
190 env->sched_info[idx].preorder = pos;
194 * Get the pre-order position.
196 static INLINE unsigned get_irn_critical_path_len(trace_env_t *env, ir_node *n) {
197 int idx = get_irn_idx(n);
199 assert(idx < ARR_LEN(env->sched_info));
200 return env->sched_info[idx].critical_path_len;
204 * Set the pre-order position.
206 static INLINE void set_irn_critical_path_len(trace_env_t *env, ir_node *n, unsigned len) {
207 int idx = get_irn_idx(n);
209 assert(idx < ARR_LEN(env->sched_info));
210 env->sched_info[idx].critical_path_len = len;
214 * returns the exec-time for node n.
216 static sched_timestep_t exectime(trace_env_t *env, ir_node *n) {
217 if (be_is_Keep(n) || is_Proj(n))
219 if (env->selector->exectime)
220 return env->selector->exectime(env->selector_env, n);
225 * Calculates the latency for between two ops
227 static sched_timestep_t latency(trace_env_t *env, ir_node *pred, int pred_cycle, ir_node *curr, int curr_cycle) {
228 /* a Keep hides a root */
229 if (be_is_Keep(curr))
230 return exectime(env, pred);
232 /* Proj's are executed immediately */
236 /* predecessors Proj's must be skipped */
238 pred = get_Proj_pred(pred);
240 if (env->selector->latency)
241 return env->selector->latency(env->selector_env, pred, pred_cycle, curr, curr_cycle);
246 * Returns the number of users of a node having mode datab.
248 static int get_num_successors(ir_node *irn) {
250 const ir_edge_t *edge;
252 if (get_irn_mode(irn) == mode_T) {
253 /* for mode_T nodes: count the users of all Projs */
254 foreach_out_edge(irn, edge) {
255 ir_node *proj = get_edge_src_irn(edge);
256 ir_mode *mode = get_irn_mode(proj);
259 sum += get_num_successors(proj);
260 else if (mode_is_datab(mode))
261 sum += get_irn_n_edges(proj);
265 /* do not count keep-alive edges */
266 foreach_out_edge(irn, edge) {
267 if (get_irn_opcode(get_edge_src_irn(edge)) != iro_End)
276 * Returns the difference of regs_output - regs_input;
278 static int get_reg_difference(trace_env_t *env, ir_node *irn) {
282 ir_node *block = get_nodes_block(irn);
284 if (be_is_Call(irn)) {
285 /* we want calls prefered */
289 if (get_irn_mode(irn) == mode_T) {
290 /* mode_T nodes: num out regs == num Projs with mode datab */
291 const ir_edge_t *edge;
292 foreach_out_edge(irn, edge) {
293 ir_node *proj = get_edge_src_irn(edge);
294 if (mode_is_datab(get_irn_mode(proj)))
301 /* num in regs: number of ins with mode datab and not ignore */
302 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
303 ir_node *in = get_irn_n(irn, i);
305 if (mode_is_datab(get_irn_mode(in)) && /* must be data node */
306 !arch_irn_is(in, ignore) && /* ignore "ignore" nodes :) */
307 !be_is_live_end(env->liveness, block, in)) { /* if the value lives outside of block: do not count */
312 return num_out - num_in;
316 * descent into a dag and create a pre-order list.
318 static void descent(ir_node *root, ir_node *block, ir_node **list, trace_env_t *env, unsigned path_len) {
321 if (! is_Phi(root)) {
322 path_len += exectime(env, root);
323 if (get_irn_critical_path_len(env, root) < path_len) {
324 set_irn_critical_path_len(env, root, path_len);
326 /* calculate number of users (needed for heuristic) */
327 set_irn_num_user(env, root, get_num_successors(root));
329 /* calculate register difference (needed for heuristic) */
330 set_irn_reg_diff(env, root, get_reg_difference(env, root));
332 /* Phi nodes always leave the block */
333 for (i = get_irn_arity(root) - 1; i >= 0; --i) {
334 ir_node *pred = get_irn_n(root, i);
336 DBG((env->dbg, LEVEL_3, " node %+F\n", pred));
338 /* Blocks may happen as predecessors of End nodes */
342 /* already seen nodes are not marked */
343 if (get_irn_link(pred) != MARK)
346 /* don't leave our block */
347 if (get_nodes_block(pred) != block)
350 set_irn_link(pred, NULL);
352 descent(pred, block, list, env, path_len);
355 set_irn_link(root, *list);
360 * Returns non-zero if root is a root in the block block.
362 static int is_root(ir_node *root, ir_node *block) {
363 const ir_edge_t *edge;
365 foreach_out_edge(root, edge) {
366 ir_node *succ = get_edge_src_irn(edge);
370 /* Phi nodes are always in "another block */
373 if (get_nodes_block(succ) == block)
380 * Performs initial block calculations for trace scheduling.
382 static void trace_preprocess_block(trace_env_t *env, ir_node *block) {
383 ir_node *root = NULL, *preord = NULL;
386 const ir_edge_t *edge;
388 /* First step: Find the root set. */
389 foreach_out_edge(block, edge) {
390 ir_node *succ = get_edge_src_irn(edge);
392 if (is_Block(succ)) {
393 /* A Block-Block edge. This should be the MacroBlock
394 * edge, ignore it. */
395 assert(get_Block_MacroBlock(succ) == block && "Block-Block edge found");
398 if (is_Anchor(succ)) {
399 /* ignore a keep alive edge */
402 if (is_root(succ, block)) {
403 mark_root_node(env, succ);
404 set_irn_link(succ, root);
408 set_irn_link(succ, MARK);
411 /* Second step: calculate the pre-order list. */
413 for (curr = root; curr; curr = irn) {
414 irn = get_irn_link(curr);
415 DBG((env->dbg, LEVEL_2, " DAG root %+F\n", curr));
416 descent(curr, block, &preord, env, 0);
420 /* Third step: calculate the Delay. Note that our
421 * list is now in pre-order, starting at root
423 for (cur_pos = 0, curr = root; curr; curr = get_irn_link(curr), cur_pos++) {
426 if (arch_irn_class_is(curr, branch)) {
427 /* assure, that branches can be executed last */
431 if (is_root_node(env, curr))
432 d = exectime(env, curr);
435 foreach_out_edge(curr, edge) {
436 ir_node *n = get_edge_src_irn(edge);
438 if (get_nodes_block(n) == block) {
441 ld = latency(env, curr, 1, n, 0) + get_irn_delay(env, n);
447 set_irn_delay(env, curr, d);
448 DB((env->dbg, LEVEL_2, "\t%+F delay %u\n", curr, d));
450 /* set the etime of all nodes to 0 */
451 set_irn_etime(env, curr, 0);
453 set_irn_preorder(env, curr, cur_pos);
458 * This functions gets called after a node finally has been made ready.
460 static void trace_node_ready(void *data, ir_node *irn, ir_node *pred) {
461 trace_env_t *env = data;
462 sched_timestep_t etime_p, etime;
464 etime = env->curr_time;
466 etime_p = get_irn_etime(env, pred);
467 etime += latency(env, pred, 1, irn, 0);
468 etime = etime_p > etime ? etime_p : etime;
471 set_irn_etime(env, irn, etime);
472 DB((env->dbg, LEVEL_2, "\tset etime of %+F to %u\n", irn, etime));
476 * Update the current time after irn has been selected.
478 static void trace_update_time(void *data, ir_node *irn) {
479 trace_env_t *env = data;
480 if (is_Phi(irn) || get_irn_opcode(irn) == iro_Start) {
481 env->curr_time += get_irn_etime(env, irn);
484 env->curr_time += exectime(env, irn);
489 * Allocates memory and initializes trace scheduling environment.
490 * @param birg The backend irg object
491 * @return The environment
493 static trace_env_t *trace_init(const be_irg_t *birg) {
494 trace_env_t *env = XMALLOCZ(trace_env_t);
495 ir_graph *irg = be_get_birg_irg(birg);
496 int nn = get_irg_last_idx(irg);
499 env->sched_info = NEW_ARR_F(trace_irn_t, nn);
500 env->liveness = be_liveness(birg);
501 FIRM_DBG_REGISTER(env->dbg, "firm.be.sched.trace");
503 be_liveness_assure_chk(env->liveness);
504 memset(env->sched_info, 0, nn * sizeof(*(env->sched_info)));
510 * Frees all memory allocated for trace scheduling environment.
511 * @param env The environment
513 static void trace_free(void *data) {
514 trace_env_t *env = data;
515 be_liveness_free(env->liveness);
516 DEL_ARR_F(env->sched_info);
521 * Simple selector. Just assure that jumps are scheduled last.
523 static ir_node *basic_selection(ir_nodeset_t *ready_set)
526 ir_nodeset_iterator_t iter;
528 /* assure that branches and constants are executed last */
529 foreach_ir_nodeset(ready_set, irn, iter) {
530 if (!arch_irn_class_is(irn, branch)) {
535 /* at last: schedule branches */
536 irn = get_nodeset_node(ready_set);
542 * The muchnik selector.
544 static ir_node *muchnik_select(void *block_env, ir_nodeset_t *ready_set, ir_nodeset_t *live_set)
546 trace_env_t *env = block_env;
547 ir_nodeset_t mcands, ecands;
548 ir_nodeset_iterator_t iter;
549 sched_timestep_t max_delay = 0;
553 /* calculate the max delay of all candidates */
554 foreach_ir_nodeset(ready_set, irn, iter) {
555 sched_timestep_t d = get_irn_delay(env, irn);
557 max_delay = d > max_delay ? d : max_delay;
560 ir_nodeset_init_size(&mcands, 8);
561 ir_nodeset_init_size(&ecands, 8);
563 /* build mcands and ecands */
564 foreach_ir_nodeset(ready_set, irn, iter) {
565 if (get_irn_delay(env, irn) == max_delay) {
566 ir_nodeset_insert(&mcands, irn);
567 if (get_irn_etime(env, irn) <= env->curr_time)
568 ir_nodeset_insert(&ecands, irn);
573 if (ir_nodeset_size(&mcands) == 1) {
574 irn = get_nodeset_node(&mcands);
575 DB((env->dbg, LEVEL_3, "\tirn = %+F, mcand = 1, max_delay = %u\n", irn, max_delay));
578 int cnt = ir_nodeset_size(&ecands);
580 irn = get_nodeset_node(&ecands);
582 if (arch_irn_class_is(irn, branch)) {
583 /* BEWARE: don't select a JUMP if others are still possible */
586 DB((env->dbg, LEVEL_3, "\tirn = %+F, ecand = 1, max_delay = %u\n", irn, max_delay));
589 DB((env->dbg, LEVEL_3, "\tecand = %d, max_delay = %u\n", cnt, max_delay));
590 irn = basic_selection(&ecands);
594 DB((env->dbg, LEVEL_3, "\tmcand = %d\n", ir_nodeset_size(&mcands)));
595 irn = basic_selection(&mcands);
602 static void *muchnik_init_graph(const list_sched_selector_t *vtab, const be_irg_t *birg)
604 trace_env_t *env = trace_init(birg);
605 env->selector = vtab;
606 env->selector_env = (void*) be_get_birg_arch_env(birg);
610 static void *muchnik_init_block(void *graph_env, ir_node *bl)
612 trace_preprocess_block(graph_env, bl);
616 const list_sched_selector_t muchnik_selector = {
620 NULL, /* to_appear_in_schedule */
621 trace_node_ready, /* node_ready */
622 trace_update_time, /* node_selected */
625 NULL, /* finish_block */
626 trace_free /* finish_graph */
630 * Execute the heuristic function.
632 static ir_node *heuristic_select(void *block_env, ir_nodeset_t *ns, ir_nodeset_t *lv)
634 trace_env_t *trace_env = block_env;
635 ir_node *irn, *cand = NULL;
636 int max_prio = INT_MIN;
637 int cur_prio = INT_MIN;
638 int cur_pressure = ir_nodeset_size(lv);
639 int reg_fact, cand_reg_fact;
640 ir_nodeset_iterator_t iter;
642 /* prefer instructions which can be scheduled early */
644 /* prefer instructions with lots of successors */
645 #define PRIO_NUMSUCCS 8
646 /* prefer instructions with long critical path */
647 #define PRIO_LEVEL 12
648 /* prefer instructions coming early in preorder */
649 #define PRIO_PREORD 8
650 /* weight of current register pressure */
651 #define PRIO_CUR_PRESS 20
652 /* weight of register pressure difference */
653 #define PRIO_CHG_PRESS 8
655 /* priority based selection, heuristic inspired by mueller diss */
656 foreach_ir_nodeset(ns, irn, iter) {
657 /* make sure that branches are scheduled last */
658 if (!arch_irn_class_is(irn, branch)) {
659 int rdiff = get_irn_reg_diff(trace_env, irn);
660 int sign = rdiff < 0;
661 int chg = (rdiff < 0 ? -rdiff : rdiff) << PRIO_CHG_PRESS;
663 /* reg_fact = chg << cur_pressure; */
664 reg_fact = chg * cur_pressure;
666 reg_fact = INT_MAX - 2;
667 reg_fact = sign ? -reg_fact : reg_fact;
669 cur_prio = (get_irn_critical_path_len(trace_env, irn) << PRIO_LEVEL)
670 //- (get_irn_delay(trace_env, irn) << PRIO_LEVEL)
671 + (get_irn_num_user(trace_env, irn) << PRIO_NUMSUCCS)
672 - (get_irn_etime(trace_env, irn) << PRIO_TIME)
673 //- ((get_irn_reg_diff(trace_env, irn) >> PRIO_CHG_PRESS) << ((cur_pressure >> PRIO_CUR_PRESS) - 3))
675 + (get_irn_preorder(trace_env, irn) << PRIO_PREORD); /* high preorder means early schedule */
676 if (cur_prio > max_prio) {
679 cand_reg_fact = reg_fact;
682 DBG((trace_env->dbg, LEVEL_4, "checked NODE %+F\n", irn));
683 DBG((trace_env->dbg, LEVEL_4, "\tpriority: %d\n", cur_prio));
684 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));
685 DBG((trace_env->dbg, LEVEL_4, "\tdelay: %d (%d)\n", get_irn_delay(trace_env, irn), get_irn_delay(trace_env, irn) << PRIO_LEVEL));
686 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));
687 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)));
688 DBG((trace_env->dbg, LEVEL_4, "\tpreorder: %d (%d)\n", get_irn_preorder(trace_env, irn), get_irn_preorder(trace_env, irn) << PRIO_PREORD));
689 DBG((trace_env->dbg, LEVEL_4, "\treg diff: %d (%d)\n", get_irn_reg_diff(trace_env, irn), 0 - reg_fact));
690 DBG((trace_env->dbg, LEVEL_4, "\tpressure: %d\n", cur_pressure));
695 DBG((trace_env->dbg, LEVEL_4, "heuristic selected %+F:\n", cand));
698 cand = basic_selection(ns);
704 const list_sched_selector_t heuristic_selector = {
708 NULL, /* to_appear_in_schedule */
709 trace_node_ready, /* node_ready */
710 trace_update_time, /* node_selected */
713 NULL, /* finish_block */
714 trace_free /* finish_graph */