2 * Implements a trace scheduler as presented in Muchnik[TM].
3 * Originally implemented by Michael Beck.
4 * @author Christian Wuerdig
11 #include "iredges_t.h"
13 #include "besched_t.h"
14 #include "belistsched.h"
17 /* we need a special mark */
21 typedef struct _trace_irn {
22 sched_timestep_t delay; /**< The delay for this node if already calculated, else 0. */
23 sched_timestep_t etime; /**< The earliest time of this node. */
24 unsigned num_user; /**< The number real users (mode datab) of this node */
25 int reg_diff; /**< The difference of num(out registers) - num(in registers) */
26 int preorder; /**< The pre-order position */
27 unsigned critical_path_len; /**< The weighted length of the longest critical path */
28 unsigned is_root : 1; /**< is a root node of a block */
31 typedef struct _trace_env {
32 trace_irn_t *sched_info; /**< trace scheduling information about the nodes */
33 const arch_env_t *arch_env; /**< the arch environment */
34 sched_timestep_t curr_time; /**< current time of the scheduler */
35 void *selector_env; /**< the backend selector environment */
36 const list_sched_selector_t *selector; /**< the actual backend selector */
37 be_lv_t *liveness; /**< The liveness for the irg */
38 DEBUG_ONLY(firm_dbg_module_t *dbg;)
42 * Returns non-zero if the node is a root node
44 static INLINE unsigned is_root_node(trace_env_t *env, ir_node *n)
46 int idx = get_irn_idx(n);
48 assert(idx < ARR_LEN(env->sched_info));
49 return env->sched_info[idx].is_root;
53 * Mark a node as root node
55 static INLINE void mark_root_node(trace_env_t *env, ir_node *n)
57 int idx = get_irn_idx(n);
59 assert(idx < ARR_LEN(env->sched_info));
60 env->sched_info[idx].is_root = 1;
64 * Get the current delay.
66 static INLINE sched_timestep_t get_irn_delay(trace_env_t *env, ir_node *n) {
67 int idx = get_irn_idx(n);
69 assert(idx < ARR_LEN(env->sched_info));
70 return env->sched_info[idx].delay;
74 * Set the current delay.
76 static INLINE void set_irn_delay(trace_env_t *env, ir_node *n, sched_timestep_t delay) {
77 int idx = get_irn_idx(n);
79 assert(idx < ARR_LEN(env->sched_info));
80 env->sched_info[idx].delay = delay;
84 * Get the current etime.
86 static INLINE sched_timestep_t get_irn_etime(trace_env_t *env, ir_node *n) {
87 int idx = get_irn_idx(n);
89 assert(idx < ARR_LEN(env->sched_info));
90 return env->sched_info[idx].etime;
94 * Set the current etime.
96 static INLINE void set_irn_etime(trace_env_t *env, ir_node *n, sched_timestep_t etime) {
97 int idx = get_irn_idx(n);
99 assert(idx < ARR_LEN(env->sched_info));
100 env->sched_info[idx].etime = etime;
104 * Get the number of users.
106 static INLINE unsigned get_irn_num_user(trace_env_t *env, ir_node *n) {
107 int idx = get_irn_idx(n);
109 assert(idx < ARR_LEN(env->sched_info));
110 return env->sched_info[idx].num_user;
114 * Set the number of users.
116 static INLINE void set_irn_num_user(trace_env_t *env, ir_node *n, unsigned num_user) {
117 int idx = get_irn_idx(n);
119 assert(idx < ARR_LEN(env->sched_info));
120 env->sched_info[idx].num_user = num_user;
124 * Get the register difference.
126 static INLINE int get_irn_reg_diff(trace_env_t *env, ir_node *n) {
127 int idx = get_irn_idx(n);
129 assert(idx < ARR_LEN(env->sched_info));
130 return env->sched_info[idx].reg_diff;
134 * Set the register difference.
136 static INLINE void set_irn_reg_diff(trace_env_t *env, ir_node *n, int reg_diff) {
137 int idx = get_irn_idx(n);
139 assert(idx < ARR_LEN(env->sched_info));
140 env->sched_info[idx].reg_diff = reg_diff;
144 * Get the pre-order position.
146 static INLINE int get_irn_preorder(trace_env_t *env, ir_node *n) {
147 int idx = get_irn_idx(n);
149 assert(idx < ARR_LEN(env->sched_info));
150 return env->sched_info[idx].preorder;
154 * Set the pre-order position.
156 static INLINE void set_irn_preorder(trace_env_t *env, ir_node *n, int pos) {
157 int idx = get_irn_idx(n);
159 assert(idx < ARR_LEN(env->sched_info));
160 env->sched_info[idx].preorder = pos;
164 * Get the pre-order position.
166 static INLINE unsigned get_irn_critical_path_len(trace_env_t *env, ir_node *n) {
167 int idx = get_irn_idx(n);
169 assert(idx < ARR_LEN(env->sched_info));
170 return env->sched_info[idx].critical_path_len;
174 * Set the pre-order position.
176 static INLINE void set_irn_critical_path_len(trace_env_t *env, ir_node *n, unsigned len) {
177 int idx = get_irn_idx(n);
179 assert(idx < ARR_LEN(env->sched_info));
180 env->sched_info[idx].critical_path_len = len;
184 * returns the exec-time for node n.
186 static sched_timestep_t exectime(trace_env_t *env, ir_node *n) {
187 if (be_is_Keep(n) || is_Proj(n))
189 if (env->selector->exectime)
190 return env->selector->exectime(env->selector_env, n);
195 * Calculates the latency for between two ops
197 static sched_timestep_t latency(trace_env_t *env, ir_node *pred, int pred_cycle, ir_node *curr, int curr_cycle) {
198 /* a Keep hides a root */
199 if (be_is_Keep(curr))
200 return exectime(env, pred);
202 /* Proj's are executed immediately */
206 /* predecessors Proj's must be skipped */
208 pred = get_Proj_pred(pred);
210 if (env->selector->latency)
211 return env->selector->latency(env->selector_env, pred, pred_cycle, curr, curr_cycle);
216 * Returns the number of users of a node having mode datab.
218 static int get_num_successors(ir_node *irn) {
220 const ir_edge_t *edge;
222 if (get_irn_mode(irn) == mode_T) {
223 /* for mode_T nodes: count the users of all Projs */
224 foreach_out_edge(irn, edge) {
225 ir_node *proj = get_edge_src_irn(edge);
226 ir_mode *mode = get_irn_mode(proj);
229 sum += get_num_successors(proj);
230 else if (mode_is_datab(mode))
231 sum += get_irn_n_edges(proj);
235 /* do not count keep-alive edges */
236 foreach_out_edge(irn, edge) {
237 if (get_irn_opcode(get_edge_src_irn(edge)) != iro_End)
246 * Returns the difference of regs_output - regs_input;
248 static int get_reg_difference(trace_env_t *env, ir_node *irn) {
252 ir_node *block = get_nodes_block(irn);
254 if (be_is_Call(irn)) {
255 /* we want calls prefered */
259 if (get_irn_mode(irn) == mode_T) {
260 /* mode_T nodes: num out regs == num Projs with mode datab */
261 const ir_edge_t *edge;
262 foreach_out_edge(irn, edge) {
263 ir_node *proj = get_edge_src_irn(edge);
264 if (mode_is_datab(get_irn_mode(proj)))
271 /* num in regs: number of ins with mode datab and not ignore */
272 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
273 ir_node *in = get_irn_n(irn, i);
275 if (! be_is_live_end(env->liveness, block, in) && /* if the value lives outside of block: do not count */
276 mode_is_datab(get_irn_mode(in)) && /* must be data node */
277 ! arch_irn_is(env->arch_env, in, ignore)) /* ignore "ignore" nodes :) */
281 return num_out - num_in;
285 * descent into a dag and create a pre-order list.
287 static void descent(ir_node *root, ir_node *block, ir_node **list, trace_env_t *env, unsigned path_len) {
290 if (! is_Phi(root)) {
291 path_len += exectime(env, root);
292 if (get_irn_critical_path_len(env, root) < path_len) {
293 set_irn_critical_path_len(env, root, path_len);
295 /* calculate number of users (needed for heuristic) */
296 set_irn_num_user(env, root, get_num_successors(root));
298 /* calculate register difference (needed for heuristic) */
299 set_irn_reg_diff(env, root, get_reg_difference(env, root));
301 /* Phi nodes always leave the block */
302 for (i = get_irn_arity(root) - 1; i >= 0; --i) {
303 ir_node *pred = get_irn_n(root, i);
305 DBG((env->dbg, LEVEL_3, " node %+F\n", pred));
307 /* Blocks may happen as predecessors of End nodes */
311 /* already seen nodes are not marked */
312 if (get_irn_link(pred) != MARK)
315 /* don't leave our block */
316 if (get_nodes_block(pred) != block)
319 set_irn_link(pred, NULL);
321 descent(pred, block, list, env, path_len);
324 set_irn_link(root, *list);
329 * Returns non-zero if root is a root in the block block.
331 static int is_root(ir_node *root, ir_node *block) {
332 const ir_edge_t *edge;
334 foreach_out_edge(root, edge) {
335 ir_node *succ = get_edge_src_irn(edge);
339 /* Phi nodes are always in "another block */
342 if (get_nodes_block(succ) == block)
349 * Performs initial block calculations for trace scheduling.
351 static void trace_preprocess_block(trace_env_t *env, ir_node *block) {
352 ir_node *root = NULL, *preord = NULL;
355 const ir_edge_t *edge;
357 /* First step: Find the root set. */
358 foreach_out_edge(block, edge) {
359 ir_node *succ = get_edge_src_irn(edge);
361 if (is_root(succ, block)) {
362 mark_root_node(env, succ);
363 set_irn_link(succ, root);
367 set_irn_link(succ, MARK);
370 /* Second step: calculate the pre-order list. */
372 for (curr = root; curr; curr = irn) {
373 irn = get_irn_link(curr);
374 DBG((env->dbg, LEVEL_2, " DAG root %+F\n", curr));
375 descent(curr, block, &preord, env, 0);
379 /* Third step: calculate the Delay. Note that our
380 * list is now in pre-order, starting at root
382 for (cur_pos = 0, curr = root; curr; curr = get_irn_link(curr), cur_pos++) {
385 if (arch_irn_class_is(env->arch_env, curr, branch)) {
386 /* assure, that branches can be executed last */
390 if (is_root_node(env, curr))
391 d = exectime(env, curr);
394 foreach_out_edge(curr, edge) {
395 ir_node *n = get_edge_src_irn(edge);
397 if (get_nodes_block(n) == block) {
400 ld = latency(env, curr, 1, n, 0) + get_irn_delay(env, n);
406 set_irn_delay(env, curr, d);
407 DB((env->dbg, LEVEL_2, "\t%+F delay %u\n", curr, d));
409 /* set the etime of all nodes to 0 */
410 set_irn_etime(env, curr, 0);
412 set_irn_preorder(env, curr, cur_pos);
417 * This functions gets called after a node finally has been made ready.
419 static void trace_node_ready(void *data, ir_node *irn, ir_node *pred) {
420 trace_env_t *env = data;
421 sched_timestep_t etime_p, etime;
423 etime = env->curr_time;
425 etime_p = get_irn_etime(env, pred);
426 etime += latency(env, pred, 1, irn, 0);
427 etime = etime_p > etime ? etime_p : etime;
430 set_irn_etime(env, irn, etime);
431 DB((env->dbg, LEVEL_2, "\tset etime of %+F to %u\n", irn, etime));
435 * Update the current time after irn has been selected.
437 static void trace_update_time(void *data, ir_node *irn) {
438 trace_env_t *env = data;
439 if (is_Phi(irn) || get_irn_opcode(irn) == iro_Start) {
440 env->curr_time += get_irn_etime(env, irn);
443 env->curr_time += exectime(env, irn);
448 * Allocates memory and initializes trace scheduling environment.
449 * @param birg The backend irg object
450 * @return The environment
452 static trace_env_t *trace_init(const arch_env_t *arch_env, ir_graph *irg) {
453 trace_env_t *env = xcalloc(1, sizeof(*env));
454 int nn = get_irg_last_idx(irg);
456 env->arch_env = arch_env;
458 env->sched_info = NEW_ARR_F(trace_irn_t, nn);
459 env->liveness = be_liveness(irg);
460 FIRM_DBG_REGISTER(env->dbg, "firm.be.sched.trace");
462 memset(env->sched_info, 0, nn * sizeof(*(env->sched_info)));
468 * Frees all memory allocated for trace scheduling environment.
469 * @param env The environment
471 static void trace_free(void *data) {
472 trace_env_t *env = data;
473 be_liveness_free(env->liveness);
474 DEL_ARR_F(env->sched_info);
479 * Simple selector. Just assure that jumps are scheduled last.
481 static ir_node *basic_selection(const arch_env_t *arch_env, nodeset *ready_set) {
484 /* assure that branches and constants are executed last */
485 for (irn = nodeset_first(ready_set); irn; irn = nodeset_next(ready_set)) {
486 if (! arch_irn_class_is(arch_env, irn, branch)) {
487 nodeset_break(ready_set);
492 /* at last: schedule branches */
493 irn = nodeset_first(ready_set);
494 nodeset_break(ready_set);
500 * The muchnik selector.
502 static ir_node *muchnik_select(void *block_env, nodeset *ready_set, nodeset *live_set)
504 trace_env_t *env = block_env;
505 nodeset *mcands, *ecands;
506 sched_timestep_t max_delay = 0;
509 /* calculate the max delay of all candidates */
510 foreach_nodeset(ready_set, irn) {
511 sched_timestep_t d = get_irn_delay(env, irn);
513 max_delay = d > max_delay ? d : max_delay;
516 mcands = new_nodeset(8);
517 ecands = new_nodeset(8);
519 /* build mcands and ecands */
520 foreach_nodeset(ready_set, irn) {
521 if (get_irn_delay(env, irn) == max_delay) {
522 nodeset_insert(mcands, irn);
523 if (get_irn_etime(env, irn) <= env->curr_time)
524 nodeset_insert(ecands, irn);
529 if (nodeset_count(mcands) == 1) {
530 irn = nodeset_first(mcands);
531 DB((env->dbg, LEVEL_3, "\tirn = %+F, mcand = 1, max_delay = %u\n", irn, max_delay));
534 int cnt = nodeset_count(ecands);
536 irn = nodeset_first(ecands);
538 if (arch_irn_class_is(env->arch_env, irn, branch)) {
539 /* BEWARE: don't select a JUMP if others are still possible */
542 DB((env->dbg, LEVEL_3, "\tirn = %+F, ecand = 1, max_delay = %u\n", irn, max_delay));
545 DB((env->dbg, LEVEL_3, "\tecand = %d, max_delay = %u\n", cnt, max_delay));
546 irn = basic_selection(env->arch_env, ecands);
550 DB((env->dbg, LEVEL_3, "\tmcand = %d\n", nodeset_count(mcands)));
551 irn = basic_selection(env->arch_env, mcands);
558 static void *muchnik_init_graph(const list_sched_selector_t *vtab, const arch_env_t *arch_env, ir_graph *irg)
560 trace_env_t *env = trace_init(arch_env, irg);
561 env->selector = vtab;
562 env->selector_env = (void*) arch_env;
566 static void *muchnik_init_block(void *graph_env, ir_node *bl)
568 trace_preprocess_block(graph_env, bl);
572 static const list_sched_selector_t muchnik_selector_struct = {
576 NULL, /* to_appear_in_schedule */
577 trace_node_ready, /* node_ready */
578 trace_update_time, /* node_selected */
581 NULL, /* finish_block */
582 trace_free /* finish_graph */
585 const list_sched_selector_t *muchnik_selector = &muchnik_selector_struct;
588 * Execute the heuristic function.
590 static ir_node *heuristic_select(void *block_env, nodeset *ns, nodeset *lv)
592 trace_env_t *trace_env = block_env;
593 ir_node *irn, *cand = NULL;
594 int max_prio = INT_MIN;
595 int cur_prio = INT_MIN;
596 int cur_pressure = nodeset_count(lv);
597 int reg_fact, cand_reg_fact;
599 /* prefer instructions which can be scheduled early */
601 /* prefer instructions with lots of successors */
602 #define PRIO_NUMSUCCS 8
603 /* prefer instructions with long critical path */
604 #define PRIO_LEVEL 12
605 /* prefer instructions coming early in preorder */
606 #define PRIO_PREORD 8
607 /* weight of current register pressure */
608 #define PRIO_CUR_PRESS 20
609 /* weight of register pressure difference */
610 #define PRIO_CHG_PRESS 8
612 /* priority based selection, heuristic inspired by mueller diss */
613 foreach_nodeset(ns, irn) {
614 /* make sure that branches are scheduled last */
615 if (! arch_irn_class_is(trace_env->arch_env, irn, branch)) {
616 int rdiff = get_irn_reg_diff(trace_env, irn);
617 int sign = rdiff < 0;
618 int chg = (rdiff < 0 ? -rdiff : rdiff) << PRIO_CHG_PRESS;
620 reg_fact = chg << cur_pressure;
622 reg_fact = INT_MAX - 2;
623 reg_fact = sign ? -reg_fact : reg_fact;
625 cur_prio = (get_irn_critical_path_len(trace_env, irn) << PRIO_LEVEL)
626 //- (get_irn_delay(trace_env, irn) << PRIO_LEVEL)
627 + (get_irn_num_user(trace_env, irn) << PRIO_NUMSUCCS)
628 - (get_irn_etime(trace_env, irn) << PRIO_TIME)
629 //- ((get_irn_reg_diff(trace_env, irn) >> PRIO_CHG_PRESS) << ((cur_pressure >> PRIO_CUR_PRESS) - 3))
631 + (get_irn_preorder(trace_env, irn) << PRIO_PREORD); /* high preorder means early schedule */
632 if (cur_prio > max_prio) {
635 cand_reg_fact = reg_fact;
638 DBG((trace_env->dbg, LEVEL_4, "checked NODE %+F\n", irn));
639 DBG((trace_env->dbg, LEVEL_4, "\tpriority: %d\n", cur_prio));
640 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));
641 DBG((trace_env->dbg, LEVEL_4, "\tdelay: %d (%d)\n", get_irn_delay(trace_env, irn), get_irn_delay(trace_env, irn) << PRIO_LEVEL));
642 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));
643 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)));
644 DBG((trace_env->dbg, LEVEL_4, "\tpreorder: %d (%d)\n", get_irn_preorder(trace_env, irn), get_irn_preorder(trace_env, irn) << PRIO_PREORD));
645 DBG((trace_env->dbg, LEVEL_4, "\treg diff: %d (%d)\n", get_irn_reg_diff(trace_env, irn), 0 - reg_fact));
646 DBG((trace_env->dbg, LEVEL_4, "\tpressure: %d\n", cur_pressure));
651 DBG((trace_env->dbg, LEVEL_4, "heuristic selected %+F:\n", cand));
654 cand = basic_selection(trace_env->arch_env, ns);
660 static const list_sched_selector_t heuristic_selector_struct = {
664 NULL, /* to_appear_in_schedule */
665 trace_node_ready, /* node_ready */
666 trace_update_time, /* node_selected */
669 NULL, /* finish_block */
670 trace_free /* finish_graph */
673 const list_sched_selector_t *heuristic_selector = &heuristic_selector_struct;