2 * Scheduling algorithms.
3 * An ILP scheduler based on
4 * "ILP-based Instruction Scheduling for IA-64"
5 * by Daniel Kaestner and Sebastian Winkel
8 * @author Christian Wuerdig
26 #include "irphase_t.h"
36 #include <lpp/lpp_net.h>
38 #include <libcore/lc_opts.h>
39 #include <libcore/lc_opts_enum.h>
40 #include <libcore/lc_timing.h>
44 #include "besched_t.h"
45 #include "beilpsched.h"
48 typedef struct _ilpsched_options_t {
55 typedef struct _unit_type_info_t {
57 const be_execution_unit_type_t *tp;
61 * holding the ILP variables of the different types
63 typedef struct _ilp_var_types_t {
64 int *x; /* x_{nt}^k variables */
65 int *a; /* a_{nt}^k variables */
66 int *d; /* d_{nt}^k variables */
67 int *y; /* y_{nt}^k variables */
71 * Holds alive variables for a node live-in to a block.
73 typedef struct _ilp_livein_node_t {
75 unsigned max_alive_steps;
79 /* attributes for a node */
80 typedef struct _ilpsched_node_attr_t {
81 unsigned asap; /**< The ASAP scheduling control step */
82 unsigned alap; /**< The ALAP scheduling control step */
83 unsigned latency; /**< Latency of this node (needed for sorting) */
84 unsigned sched_point; /**< the step in which the node is finally scheduled */
85 unsigned visit_idx; /**< Index of the node having visited this node last */
86 unsigned consumer_idx; /**< Index of the node having counted this node as consumer last */
87 unsigned n_consumer; /**< Number of consumers */
88 ir_node **block_consumer; /**< List of consumer being in the same block */
89 waitq *projkeeps; /**< A List of Projs and Keeps belonging to this node */
90 unsigned block_idx : 30; /**< A unique per block index */
91 unsigned alap_changed : 1; /**< the current ALAP has changed, revisit preds */
92 unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */
93 bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors
94 for ASAP, successors for ALAP */
95 unsigned n_unit_types; /**< number of allowed execution unit types */
96 unit_type_info_t *type_info; /**< list of allowed execution unit types */
97 ilp_var_types_t ilp_vars; /**< the different ILP variables */
98 } ilpsched_node_attr_t;
100 /* attributes for a block */
101 typedef struct _ilpsched_block_attr_t {
102 unsigned block_last_idx; /**< The highest node index in block so far */
103 unsigned n_interesting_nodes; /**< The number of nodes interesting for scheduling */
104 unsigned max_steps; /**< Upper bound for block execution */
105 plist_t *root_nodes; /**< A list of nodes having no user in current block */
106 ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */
107 pset *livein_nodes; /**< A set of nodes which are live-in to this block */
108 } ilpsched_block_attr_t;
110 typedef union _ilpsched_attr_ {
111 ilpsched_node_attr_t node_attr;
112 ilpsched_block_attr_t block_attr;
115 /* A irn for the phase and it's attributes (either node or block) */
118 ilpsched_attr_t attr;
121 /* The ILP scheduling environment */
123 phase_t ph; /**< The phase */
124 ir_graph *irg; /**< The current irg */
125 heights_t *height; /**< The heights object of the irg */
126 void *irg_env; /**< An environment for the irg scheduling, provided by the backend */
127 void *block_env; /**< An environment for scheduling a block, provided by the backend */
128 const arch_env_t *arch_env;
129 const arch_isa_t *isa; /**< The ISA */
130 const be_main_env_t *main_env;
131 const be_machine_t *cpu; /**< the current abstract machine */
132 ilpsched_options_t *opts; /**< the ilp options for current irg */
133 const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */
134 DEBUG_ONLY(firm_dbg_module_t *dbg);
137 /* convenience macros to handle phase irn data */
138 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
139 #define is_ilpsched_block(node) (is_Block((node)->irn))
140 #define get_ilpsched_block_attr(block) (&(block)->attr.block_attr)
141 #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr)
143 /* check if node is considered for ILP scheduling */
144 #define consider_for_sched(isa, irn) \
145 (! (is_Block(irn) || \
146 is_normal_Proj(isa, irn) || \
153 /* gives the valid scheduling time step interval for a node */
154 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
156 /* gives the valid interval where a node can die */
157 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
159 /* gives the corresponding ILP variable for given node, unit and time step */
160 #define ILPVAR_IDX(na, unit, control_step) \
161 ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
163 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
164 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
165 ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
167 /* check if a double value is within an epsilon environment of 0 */
168 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
170 #define ilp_timer_push(t) lc_timer_push((t))
171 #define ilp_timer_pop() lc_timer_pop()
172 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
174 /* option variable */
175 static ilpsched_options_t ilp_opts = {
176 1, /* default is with register pressure constraints */
177 120, /* if we have more than 70 nodes: use alive nodes constraint */
178 300, /* 300 sec per block time limit */
183 static const lc_opt_table_entry_t ilpsched_option_table[] = {
184 LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress),
185 LC_OPT_ENT_INT("limit_dead", "Upto how many nodes the dead node constraint should be used", &ilp_opts.limit_dead),
186 LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
187 LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
192 We need this global variable as we compare nodes dependent on heights,
193 but we cannot pass any information to the qsort compare function.
195 static heights_t *glob_heights;
198 * Check if irn is a Proj, which has no execution units assigned.
199 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
201 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
202 return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
206 * Skips normal Projs.
207 * @return predecessor if irn is a normal Proj, otherwise irn.
209 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
210 if (is_normal_Proj(isa, irn))
211 return get_Proj_pred(irn);
215 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
216 unsigned lat = be_ilp_sched_latency(sel, irn, env);
217 if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
222 static int cmp_live_in_nodes(const void *a, const void *b) {
223 const ilp_livein_node_t *n1 = a;
224 const ilp_livein_node_t *n2 = b;
226 return n1->irn != n2->irn;
230 * Compare scheduling time steps of two be_ilpsched_irn's.
232 static int cmp_ilpsched_irn(const void *a, const void *b) {
233 be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a;
234 be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b;
235 ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
236 ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
238 if (n1_a->sched_point == n2_a->sched_point) {
239 ir_node *irn_a = n1->irn;
240 ir_node *irn_b = n2->irn;
242 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
244 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
248 Ok, timestep is equal and the nodes are parallel,
249 so check latency and schedule high latency first.
251 return QSORT_CMP(n2_a->latency, n1_a->latency);
254 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
258 * In case there is no phase information for irn, initialize it.
260 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
261 be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
264 /* if we have already some data: check for reinitialization */
266 if (! is_Block(irn)) {
267 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
269 if (! na->transitive_block_nodes) {
270 ir_node *block = get_nodes_block(irn);
271 be_ilpsched_irn_t *block_node = phase_get_or_set_irn_data(ph, block);
272 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
274 /* we are called after the block indices have been build: create bitset */
275 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
278 /* we are called from reinit block data: clear the bitset */
279 bitset_clear_all(na->transitive_block_nodes);
281 na->alap_changed = 1;
289 /* set ilpsched irn attributes (either block or irn) */
291 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
293 ba->n_interesting_nodes = 0;
294 ba->block_last_idx = 0;
295 ba->root_nodes = plist_new();
296 ba->head_ilp_nodes = NULL;
297 ba->livein_nodes = new_pset(cmp_live_in_nodes, 16);
301 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
302 memset(na, 0, sizeof(*na));
309 * Assign a per block unique number to each node.
311 static void build_block_idx(ir_node *irn, void *walk_env) {
312 be_ilpsched_env_t *env = walk_env;
313 be_ilpsched_irn_t *node, *block_node;
314 ilpsched_node_attr_t *na;
315 ilpsched_block_attr_t *ba;
317 if (! consider_for_sched(env->arch_env->isa, irn))
320 node = get_ilpsched_irn(env, irn);
321 na = get_ilpsched_node_attr(node);
322 block_node = get_ilpsched_irn(env, get_nodes_block(irn));
323 ba = get_ilpsched_block_attr(block_node);
325 na->block_idx = ba->block_last_idx++;
328 /********************************************************
331 * __ _ ___ __ _ _ __ / / __ _| | __ _ _ __
332 * / _` / __|/ _` | '_ \ / / / _` | |/ _` | '_ \
333 * | (_| \__ \ (_| | |_) | / / | (_| | | (_| | |_) |
334 * \__,_|___/\__,_| .__/ /_/ \__,_|_|\__,_| .__/
337 ********************************************************/
340 * Add all nodes having no user in current block to last_nodes list.
342 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
344 const ir_edge_t *edge;
345 be_ilpsched_irn_t *block_node, *node;
346 ilpsched_block_attr_t *ba;
347 ilpsched_node_attr_t *na;
349 be_ilpsched_env_t *env = walk_env;
350 int has_block_user = 0;
351 unsigned n_consumer = 0;
352 ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
356 if (! consider_for_sched(env->arch_env->isa, irn))
359 block = get_nodes_block(irn);
360 idx = get_irn_idx(irn);
361 consumer = NEW_ARR_F(ir_node *, 0);
363 DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
365 /* check data and dependency out edges */
366 for (i = 0; i < 2 && ! has_block_user; ++i) {
367 foreach_out_edge_kind(irn, edge, ekind[i]) {
368 ir_node *user = get_edge_src_irn(edge);
370 if (is_normal_Proj(env->arch_env->isa, user)) {
371 const ir_edge_t *user_edge;
373 if (get_irn_mode(user) == mode_X)
376 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
377 for (j = 0; j < 2; ++j) {
378 foreach_out_edge_kind(user, user_edge, ekind[j]) {
379 ir_node *real_user = get_edge_src_irn(user_edge);
381 if (! is_Phi(real_user) && ! is_Block(real_user)) {
382 be_ilpsched_irn_t *node = get_ilpsched_irn(env, real_user);
383 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
385 /* skip already visited nodes */
386 if (ua->consumer_idx == idx)
389 /* check if node has user in this block and collect the user if it's a data user */
390 if (get_nodes_block(real_user) == block) {
391 if (i == 0 && j == 0)
392 ARR_APP1(ir_node *, consumer, real_user);
396 /* only count data consumer */
400 /* mark user as visited by this node */
401 ua->consumer_idx = idx;
406 else if (is_Block(user)) {
409 else if (! is_Phi(user)) {
410 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
411 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
413 /* skip already visited nodes */
414 if (ua->consumer_idx == idx)
417 /* check if node has user in this block and collect the user if it's a data user */
418 if (get_nodes_block(user) == block) {
420 ARR_APP1(ir_node *, consumer, user);
424 /* only count data consumer */
428 /* mark user visited by this node */
429 ua->consumer_idx = idx;
434 block_node = get_ilpsched_irn(env, block);
435 ba = get_ilpsched_block_attr(block_node);
437 ba->n_interesting_nodes++;
439 /* current irn has no user inside this block, add to queue */
440 if (! has_block_user) {
441 DB((env->dbg, LEVEL_3, "root node\n"));
442 plist_insert_back(ba->root_nodes, irn);
445 DB((env->dbg, LEVEL_3, "normal node\n"));
448 /* record number of all consumer and the consumer within the same block */
449 node = get_ilpsched_irn(env, irn);
450 na = get_ilpsched_node_attr(node);
451 na->n_consumer = n_consumer;
452 na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
453 memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
458 * Calculate the ASAP scheduling step for current irn.
460 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
461 be_ilpsched_env_t *env = walk_env;
464 be_ilpsched_irn_t *node, *block_node;
465 ilpsched_node_attr_t *na;
466 ilpsched_block_attr_t *ba;
468 /* These nodes are handled separate */
469 if (! consider_for_sched(env->arch_env->isa, irn))
472 DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
474 block = get_nodes_block(irn);
475 node = get_ilpsched_irn(env, irn);
476 na = get_ilpsched_node_attr(node);
479 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
480 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
482 /* check for greatest distance to top */
483 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
484 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
485 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
488 lat = fixed_latency(env->sel, pred, env->block_env);
490 na->asap = MAX(na->asap, pna->asap + lat);
494 /* add node to ILP node list and update max_steps */
495 block_node = get_ilpsched_irn(env, block);
496 ba = get_ilpsched_block_attr(block_node);
498 set_irn_link(irn, ba->head_ilp_nodes);
499 ba->head_ilp_nodes = irn;
500 ba->max_steps += fixed_latency(env->sel, irn, env->block_env);
502 DB((env->dbg, LEVEL_2, "%u\n", na->asap));
506 * Calculate the ALAP scheduling step of all irns in current block.
507 * Depends on max_steps being calculated.
509 static void calculate_block_alap(ir_node *block, void *walk_env) {
510 be_ilpsched_env_t *env = walk_env;
511 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
512 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
513 waitq *cur_queue = new_waitq();
516 assert(is_Block(block));
518 DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
519 block, ba->n_interesting_nodes, ba->max_steps));
521 /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
522 //phase_reinit_block_irn_data(&env->ph, block);
524 /* init start queue */
525 foreach_plist(ba->root_nodes, el) {
526 waitq_put(cur_queue, plist_element_get_value(el));
529 /* repeat until all nodes are processed */
530 while (! waitq_empty(cur_queue)) {
531 waitq *next_queue = new_waitq();
533 /* process all nodes in current step */
534 while (! waitq_empty(cur_queue)) {
535 ir_node *cur_irn = waitq_get(cur_queue);
536 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
537 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
540 /* cur_node has no alap set -> it's a root node, set to max alap */
542 na->alap = ba->max_steps;
543 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
547 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
551 /* set the alap's of all predecessors */
552 for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
553 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
555 /* check for greatest distance to bottom */
556 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
557 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
558 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
561 /* mark the predecessor as visited by current irn */
562 if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
564 pna->visit_idx = get_irn_idx(cur_irn);
566 lat = fixed_latency(env->sel, pred, env->block_env);
568 /* set ALAP of current pred */
569 if (pna->alap == 0) {
570 /* current ALAP is 0: node has not yet been visited */
571 pna->alap_changed = 1;
572 pna->alap = na->alap - lat;
574 else if (pna->alap > na->alap - lat) {
575 /* we found a longer path to root node: change ALAP */
576 pna->alap = na->alap - lat;
577 pna->alap_changed = 1;
580 /* current ALAP is best found so far: keep it */
581 pna->alap_changed = 0;
584 DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
586 /* enqueue node for next iteration */
587 if (get_irn_ins_or_deps(pred) > 0)
588 waitq_put(next_queue, pred);
593 /* prepare for next iteration */
594 del_waitq(cur_queue);
595 cur_queue = next_queue;
600 * Free list of root nodes and the set of live-in nodes.
602 static void clear_unwanted_data(ir_node *block, void *walk_env) {
603 be_ilpsched_env_t *env = walk_env;
604 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
605 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
607 plist_free(ba->root_nodes);
608 ba->root_nodes = NULL;
609 del_pset(ba->livein_nodes);
610 ba->livein_nodes = NULL;
614 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
615 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
617 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
618 be_ilpsched_env_t *env = walk_env;
620 be_ilpsched_irn_t *node, *pred_node;
621 ilpsched_node_attr_t *na, *pna;
623 if (! consider_for_sched(env->arch_env->isa, irn))
626 if (! is_Proj(irn) && ! be_is_Keep(irn))
629 /* go to the ancestor */
631 pred = get_irn_n(irn, 0);
632 pred = skip_Proj(pred);
634 node = get_ilpsched_irn(env, irn);
635 pred_node = get_ilpsched_irn(env, pred);
636 na = get_ilpsched_node_attr(node);
637 pna = get_ilpsched_node_attr(pred_node);
639 na->asap = pna->asap;
640 na->alap = pna->alap;
642 /* record all Projs and Keeps for this node */
643 if (! pna->projkeeps)
644 pna->projkeeps = new_waitq();
645 waitq_put(pna->projkeeps, irn);
647 DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
650 /*******************************************
653 * ___ ___| |__ ___ __| |_ _| | ___
654 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
655 * \__ \ (__| | | | __/ (_| | |_| | | __/
656 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
658 *******************************************/
660 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
661 const ir_edge_t *edge;
663 foreach_out_edge(irn, edge) {
664 ir_node *user = get_edge_src_irn(edge);
666 if (be_is_Keep(user)) {
667 assert(get_nodes_block(user) == block && "Keep must not be in different block.");
668 waitq_put(keeps, user);
674 * Inserts @p irn before @p before into schedule and notifies backend.
676 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
677 ir_node *before, ir_node *irn, unsigned cycle)
679 be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
680 sched_add_before(before, irn);
684 * Adds a node, it's Projs (in case of mode_T nodes) and
685 * it's Keeps to schedule.
687 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
688 const ir_edge_t *edge;
689 waitq *keeps = new_waitq();
691 /* mode_M nodes are not scheduled */
692 if (get_irn_mode(irn) == mode_M)
695 if (! sched_is_scheduled(irn))
696 notified_sched_add_before(env, block, irn, cycle);
699 if (get_irn_mode(irn) == mode_T) {
700 foreach_out_edge(irn, edge) {
701 ir_node *user = get_edge_src_irn(edge);
703 if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
704 get_irn_n_edges(user) > 0)
706 notified_sched_add_before(env, block, user, cycle);
709 check_for_keeps(keeps, block, user);
713 check_for_keeps(keeps, block, irn);
717 while (! waitq_empty(keeps)) {
718 ir_node *keep = waitq_get(keeps);
719 if (! sched_is_scheduled(keep))
720 notified_sched_add_before(env, block, keep, cycle);
727 * Schedule all nodes in the given block, according to the ILP solution.
729 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
730 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
731 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
732 sched_info_t *info = get_irn_sched_info(block);
733 be_ilpsched_irn_t **sched_nodes;
736 const ir_edge_t *edge;
738 /* init block schedule list */
739 INIT_LIST_HEAD(&info->list);
742 /* collect nodes and their scheduling time step */
743 sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
744 if (ba->n_interesting_nodes == 0) {
747 else if (ba->n_interesting_nodes == 1) {
748 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
750 /* add the single node */
751 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
754 /* check all nodes for their positive solution */
755 foreach_linked_irns(ba->head_ilp_nodes, irn) {
756 be_ilpsched_irn_t *node;
757 ilpsched_node_attr_t *na;
761 node = get_ilpsched_irn(env, irn);
762 na = get_ilpsched_node_attr(node);
766 if (! na->is_dummy_node) {
767 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
768 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
769 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
771 if (! LPP_VALUE_IS_0(cost)) {
772 ir_fprintf(stderr, "\t\t%+F has additional regpressure costs of %f\n", irn, cost);
780 /* go over all variables of a node until the non-zero one is found */
781 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
782 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
783 double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
785 /* check, if variable is set to one (it's not zero then :) */
786 if (! LPP_VALUE_IS_0(val)) {
788 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
789 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
790 irn, t, na->type_info[tp_idx].tp->name));
797 glob_heights = env->height;
798 /* sort nodes ascending by scheduling time step */
799 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
802 /* make all Phis ready and remember the single cf op */
804 foreach_out_edge(block, edge) {
805 irn = get_edge_src_irn(edge);
807 switch (get_irn_opcode(irn)) {
809 add_to_sched(env, block, irn, 0);
819 assert(cfop == NULL && "Highlander - there can be only one");
826 /* add all nodes from list */
827 for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
828 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
829 if (sched_nodes[i]->irn != cfop)
830 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
833 /* schedule control flow node if not already done */
834 if (cfop && ! sched_is_scheduled(cfop))
835 add_to_sched(env, block, cfop, 0);
837 DEL_ARR_F(sched_nodes);
840 /***************************************************************
841 * _____ _ _____ _____ _ _
842 * |_ _| | | __ \ / ____| | | (_)
843 * | | | | | |__) | | (___ ___ ___| |_ _ ___ _ __
844 * | | | | | ___/ \___ \ / _ \/ __| __| |/ _ \| '_ \
845 * _| |_| |____| | ____) | __/ (__| |_| | (_) | | | |
846 * |_____|______|_| |_____/ \___|\___|\__|_|\___/|_| |_|
848 ***************************************************************/
851 * Check if node can be executed on given unit type.
853 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
855 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
857 for (i = na->n_unit_types - 1; i >= 0; --i) {
858 if (na->type_info[i].tp == tp)
865 /************************************************
868 * __ ____ _ _ __ _ __ _| |__ | | ___ ___
869 * \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
870 * \ V / (_| | | | | (_| | |_) | | __/\__ \
871 * \_/ \__,_|_| |_|\__,_|_.__/|_|\___||___/
873 ************************************************/
875 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
876 const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
877 unsigned n_unit_types = 0;
878 be_ilpsched_irn_t *node;
879 ilpsched_node_attr_t *na;
880 unsigned unit_idx, tp_idx;
882 /* count number of available unit types for this node */
883 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
886 node = get_ilpsched_irn(env, irn);
887 na = get_ilpsched_node_attr(node);
889 if (! na->type_info) {
890 na->n_unit_types = n_unit_types;
891 na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
893 /* fill the type info array */
894 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
895 for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
896 /* beware: we also count number of available units here */
897 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
898 na->is_dummy_node = 1;
901 na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp;
902 na->type_info[tp_idx].n_units = unit_idx;
910 * Returns the largest alap time of a user of @p irn.
911 * The user must be in block @p block.
913 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
914 const ir_edge_t *edge;
915 unsigned max_alap = 0;
917 foreach_out_edge(irn, edge) {
918 ir_node *user = get_edge_src_irn(edge);
920 if (get_nodes_block(user) == block) {
921 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
922 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
924 max_alap = MAX(max_alap, na->alap);
928 assert(max_alap > 0);
933 * Create the following variables:
934 * - x_{nt}^k binary weigthed with: t
935 * node n is scheduled at time step t to unit type k
936 * ==>> These variables represent the schedule
938 * - d_{nt}^k binary weighted with: t
939 * node n dies at time step t on unit type k
940 * - a_{nt}^k binary weighted with num_nodes
941 * node n is alive at time step t on unit type k
943 * - y_{nt}^k binary weighted with: num_nodes^2
944 * node n is scheduled at time step t to unit type k
945 * although all units of this type are occupied
946 * ==>> These variables represent the register pressure
949 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
952 unsigned num_block_var, num_nodes;
953 ilp_livein_node_t *livein;
954 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
955 unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
956 lc_timer_t *t_var = lc_timer_register("beilpsched_var", "create ilp variables");
958 ilp_timer_push(t_var);
959 num_block_var = num_nodes = 0;
960 foreach_linked_irns(ba->head_ilp_nodes, irn) {
961 be_ilpsched_irn_t *node;
962 ilpsched_node_attr_t *na;
963 unsigned n_unit_types, tp_idx, n_var, cur_unit;
964 unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad;
967 node = get_ilpsched_irn(env, irn);
968 na = get_ilpsched_node_attr(node);
969 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
971 /* allocate space for ilp variables */
972 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
973 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
975 /* we need these variables only for "real" nodes */
976 if (! na->is_dummy_node) {
977 na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
978 memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
980 num_ad = ba->max_steps - na->asap + 1;
982 if (ba->n_interesting_nodes > env->opts->limit_dead) {
983 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
984 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
987 na->ilp_vars.d = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
988 memset(na->ilp_vars.d, -1, ARR_LEN(na->ilp_vars.d) * sizeof(na->ilp_vars.d[0]));
992 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
993 irn, na->asap, na->alap, na->n_unit_types));
995 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
996 /* create variables */
997 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1000 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1001 /* x_{nt}^k variables */
1002 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1003 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1004 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1005 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1006 /* variable counter */
1010 if (! na->is_dummy_node) {
1011 /* y_{nt}^k variables */
1012 snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1013 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1014 na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1015 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1017 /* variable counter */
1023 /* a node can die at any step t: asap(n) <= t <= U */
1024 if (! na->is_dummy_node) {
1025 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1027 if (ba->n_interesting_nodes > env->opts->limit_dead) {
1028 /* a_{nt}^k variables */
1029 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1030 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1031 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1034 /* d_{nt}^k variables */
1035 snprintf(buf, sizeof(buf), "d_n%u_%s_%u",
1036 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1037 na->ilp_vars.d[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1039 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1041 /* variable counter */
1047 /* collect live-in nodes */
1048 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1049 ir_node *pred = get_irn_n(irn, i);
1051 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1052 be_ilpsched_set_type_info(env, pred, var_obst);
1053 if (! na->is_dummy_node) {
1054 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1057 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1063 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1067 /* create alive variables a_{nt}^k for live-ins */
1068 foreach_pset(ba->livein_nodes, livein) {
1069 be_ilpsched_irn_t *node;
1070 ilpsched_node_attr_t *na;
1071 unsigned tp_idx, var_idx, max_alap;
1075 node = get_ilpsched_irn(env, irn);
1076 na = get_ilpsched_node_attr(node);
1078 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1080 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1083 /* create variables */
1084 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1087 for (t = 0; t < livein->max_alive_steps; ++t) {
1088 /* a_{nt}^k variables */
1089 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1090 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1091 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1092 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1099 DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1100 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1103 /*******************************************************
1106 * ___ ___ _ __ ___| |_ _ __ __ _ _ _ __ | |_ ___
1107 * / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1108 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1109 * \___\___/|_| |_|___/\__|_| \__,_|_|_| |_|\__|___/
1111 *******************************************************/
1114 * Create following ILP constraints:
1115 * - the assignment constraints:
1116 * assure each node is executed once by exactly one (allowed) execution unit
1117 * - the dead node assignment constraints:
1118 * assure a node can only die at most once
1119 * - the precedence constraints:
1120 * assure that no data dependencies are violated
1122 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1123 unsigned num_cst_assign, num_cst_prec, num_cst_dead;
1126 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1127 bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx);
1128 lc_timer_t *t_cst_assign = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1129 lc_timer_t *t_cst_dead = lc_timer_register("beilpsched_cst_assign_dead", "create dead node assignment constraints");
1130 lc_timer_t *t_cst_prec = lc_timer_register("beilpsched_cst_prec", "create precedence constraints");
1132 num_cst_assign = num_cst_prec = num_cst_dead = 0;
1133 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1136 be_ilpsched_irn_t *node;
1137 ilpsched_node_attr_t *na;
1139 node = get_ilpsched_irn(env, irn);
1140 na = get_ilpsched_node_attr(node);
1143 /* the assignment constraint */
1144 ilp_timer_push(t_cst_assign);
1145 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1146 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1147 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1150 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1153 /* the dead node assignment constraint */
1154 if (! na->is_dummy_node && ba->n_interesting_nodes <= env->opts->limit_dead) {
1155 ilp_timer_push(t_cst_dead);
1156 snprintf(buf, sizeof(buf), "dead_node_assign_cst_n%u", get_irn_idx(irn));
1157 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1158 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1160 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.d, ARR_LEN(na->ilp_vars.d), 1.0);
1164 /* We have separate constraints for Projs and Keeps */
1165 // ILP becomes infeasible ?!?
1166 // if (is_Proj(irn) || be_is_Keep(irn))
1169 /* the precedence constraints */
1170 ilp_timer_push(t_cst_prec);
1171 bs_block_irns = bitset_clear_all(bs_block_irns);
1172 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1173 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1174 unsigned t_low, t_high, t;
1175 be_ilpsched_irn_t *pred_node;
1176 ilpsched_node_attr_t *pna;
1179 if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1182 pred_node = get_ilpsched_irn(env, pred);
1183 pna = get_ilpsched_node_attr(pred_node);
1185 assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1187 if (! bitset_is_set(bs_block_irns, pna->block_idx))
1188 bitset_set(bs_block_irns, pna->block_idx);
1192 /* irn = n, pred = m */
1193 delay = fixed_latency(env->sel, pred, env->block_env);
1194 t_low = MAX(na->asap, pna->asap + delay - 1);
1195 t_high = MIN(na->alap, pna->alap + delay - 1);
1196 for (t = t_low - 1; t <= t_high - 1; ++t) {
1198 int *tmp_var_idx = NEW_ARR_F(int, 0);
1200 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1201 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1202 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1205 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1206 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1207 /* node variables have smaller index than pred variables */
1208 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1209 for (tn = na->asap - 1; tn <= t; ++tn) {
1210 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1211 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1215 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1216 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1217 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1218 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1223 /* pred variables have smaller index than node variables */
1224 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1225 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1226 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1227 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1231 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1232 for (tn = na->asap - 1; tn <= t; ++tn) {
1233 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1234 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1239 if (ARR_LEN(tmp_var_idx) > 0)
1240 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1242 DEL_ARR_F(tmp_var_idx);
1247 DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1248 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1249 DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1250 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1254 * Create ILP resource constraints:
1255 * - assure that for each time step not more instructions are scheduled
1256 * to the same unit types as units of this type are available
1258 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1261 unsigned num_cst_resrc = 0;
1262 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1263 lc_timer_t *t_cst_rsrc = lc_timer_register("beilpsched_cst_rsrc", "create resource constraints");
1265 ilp_timer_push(t_cst_rsrc);
1266 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1268 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1270 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1272 /* check each time step */
1273 for (t = 0; t < ba->max_steps; ++t) {
1276 int *tmp_var_idx = NEW_ARR_F(int, 0);
1278 snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1279 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1280 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1283 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1284 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1285 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1288 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1290 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1291 int cur_var = ILPVAR_IDX(na, tp_idx, t);
1292 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1296 /* set constraints if we have some */
1297 if (ARR_LEN(tmp_var_idx) > 0)
1298 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1300 DEL_ARR_F(tmp_var_idx);
1304 DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1305 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1309 * Create ILP bundle constraints:
1310 * - assure, at most bundle_size * bundles_per_cycle instructions
1311 * can be started at a certain point.
1313 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1316 unsigned num_cst_bundle = 0;
1317 unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1318 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1319 lc_timer_t *t_cst_bundle = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1321 ilp_timer_push(t_cst_bundle);
1322 for (t = 0; t < ba->max_steps; ++t) {
1325 int *tmp_var_idx = NEW_ARR_F(int, 0);
1327 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1328 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1329 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1332 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1333 be_ilpsched_irn_t *node;
1334 ilpsched_node_attr_t *na;
1337 /* Projs and Keeps do not contribute to bundle size */
1338 if (is_Proj(irn) || be_is_Keep(irn))
1341 node = get_ilpsched_irn(env, irn);
1342 na = get_ilpsched_node_attr(node);
1344 /* nodes assigned to DUMMY unit do not contribute to bundle size */
1345 if (na->is_dummy_node)
1348 if (t >= na->asap - 1 && t <= na->alap - 1) {
1349 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1350 int idx = ILPVAR_IDX(na, tp_idx, t);
1351 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1356 if (ARR_LEN(tmp_var_idx) > 0)
1357 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1359 DEL_ARR_F(tmp_var_idx);
1362 DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1363 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1367 * Create ILP dying nodes constraints:
1368 * - set variable d_{nt}^k to 1 if nodes n dies at step t on unit k
1370 static void create_dying_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1373 unsigned num_cst = 0;
1374 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1375 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_dying_nodes", "create dying nodes constraints");
1377 ilp_timer_push(t_cst);
1378 /* check all time_steps */
1379 for (t = 0; t < ba->max_steps; ++t) {
1383 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1384 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1385 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1387 /* if node has no consumer within current block, it cannot die here */
1388 /* we also ignore nodes assigned to dummy unit */
1389 if (ARR_LEN(na->block_consumer) < 1 || na->is_dummy_node)
1392 /* node can only die here if t at least asap(n) */
1393 if (t >= na->asap - 1) {
1396 /* for all unit types */
1397 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1399 int *tmp_var_idx = NEW_ARR_F(int, 0);
1401 snprintf(buf, sizeof(buf), "dying_node_cst_%u_n%u", t, get_irn_idx(irn));
1402 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(na->n_consumer - 1));
1403 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1406 /* number of consumer scheduled till t */
1407 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1408 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1409 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1411 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1414 for (tm = ca->asap - 1; tm <= t && tm <= ca->alap - 1; ++tm) {
1415 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1416 ARR_APP1(int, tmp_var_idx, ca->ilp_vars.x[idx]);
1421 /* could be that no consumer can be scheduled at this point */
1422 if (ARR_LEN(tmp_var_idx)) {
1426 /* subtract possible prior kill points */
1427 for (tn = na->asap - 1; tn < t; ++tn) {
1428 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, tn);
1429 lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], -1.0);
1432 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1433 lpp_set_factor_fast(lpp, cst, na->ilp_vars.d[idx], 0.0 - (double)(na->n_consumer));
1434 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1437 DEL_ARR_F(tmp_var_idx);
1444 DBG((env->dbg, LEVEL_1, "\t%u dying nodes constraints (%g sec)\n",
1445 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1449 * Create ILP alive nodes constraints:
1450 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1452 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1455 unsigned num_cst = 0;
1456 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1457 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1459 ilp_timer_push(t_cst);
1461 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1462 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1463 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1466 /* we ignore nodes assigned to dummy unit here */
1467 if (na->is_dummy_node)
1470 /* check check all time steps: asap(n) <= t <= U */
1471 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1474 /* for all unit types available for this node */
1475 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1476 unsigned tn, tn_max, idx;
1478 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1479 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1481 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1482 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1483 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1484 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1487 tn_max = MIN(na->alap - 1, t);
1488 /* check if the node has been scheduled so far */
1489 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1490 int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1491 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1494 if (ARR_LEN(tmp_var_idx_n) > 0)
1495 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1496 DEL_ARR_F(tmp_var_idx_n);
1498 /* subtract the number of consumer scheduled so far */
1499 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1500 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1501 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1503 unsigned tm, tm_max;
1505 tm_max = MIN(ca->alap - 1, t);
1506 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1507 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1508 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1509 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1514 if (ARR_LEN(tmp_var_idx_m) > 0)
1515 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1516 DEL_ARR_F(tmp_var_idx_m);
1519 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1520 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1526 DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1527 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1531 * Create ILP alive nodes constraints for live-in nodes:
1532 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1534 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1536 ilp_livein_node_t *livein;
1537 unsigned num_cst = 0;
1538 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1539 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1541 ilp_timer_push(t_cst);
1543 foreach_pset(ba->livein_nodes, livein) {
1544 ir_node *irn = livein->irn;
1545 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1546 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1549 /* check check all time steps: 0 <= t < max_alive_steps */
1550 for (t = 0; t < livein->max_alive_steps; ++t) {
1553 /* for all unit types available for this node */
1554 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1555 const ir_edge_t *edge;
1556 unsigned tn, tn_max, idx;
1557 int cst, i, num_block_user;
1558 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1560 /* check the number of consumer scheduled so far */
1562 foreach_out_edge(irn, edge) {
1563 ir_node *user = get_edge_src_irn(edge);
1564 be_ilpsched_irn_t *cons;
1565 ilpsched_node_attr_t *ca;
1567 unsigned tm, tm_max;
1569 /* check only users within current block */
1570 if (get_nodes_block(user) != block_node->irn)
1574 cons = get_ilpsched_irn(env, user);
1575 ca = get_ilpsched_node_attr(cons);
1577 tm_max = MIN(ca->alap - 1, t);
1578 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1579 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1580 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1581 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1586 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1587 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1588 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1589 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1592 /* sum(scheduled users) */
1593 if (ARR_LEN(tmp_var_idx_m) > 0)
1594 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1595 DEL_ARR_F(tmp_var_idx_m);
1597 /* + c * a_{nt}^k */
1598 idx = node_tp_idx * livein->max_alive_steps + t;
1599 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1604 DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1605 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1609 * Create ILP pressure constraints, based on dead nodes:
1610 * - add additional costs to objective function if a node is scheduled
1611 * on a unit although all units of this type are currently occupied
1613 static void create_pressure_dead_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1616 unsigned num_cst = 0;
1617 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1618 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1620 ilp_timer_push(t_cst);
1621 /* y_{nt}^k is set for each node and timestep and unit type */
1622 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1623 unsigned cur_idx = get_irn_idx(cur_irn);
1624 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1625 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1628 /* we ignore nodes assigned to DUMMY unit here */
1629 if (cur_na->is_dummy_node)
1633 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1634 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1638 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1640 /* check if node can be executed on this unit type */
1641 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1645 /* check all time_steps */
1646 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1649 int *tmp_idx_1 = NEW_ARR_F(int, 0);
1650 int *tmp_idx_m1 = NEW_ARR_F(int, 0);
1652 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1653 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1654 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1658 - accumulate all nodes scheduled on unit type k till t
1659 - subtract all nodes died on unit type k till t
1661 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1662 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1663 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1667 tmax = MIN(t, na->alap - 1);
1668 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1670 /* current unit type is not suitable for current node */
1674 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1677 /* node scheduled */
1678 idx = ILPVAR_IDX(na, tp_idx, tn);
1679 ARR_APP1(int, tmp_idx_1, na->ilp_vars.x[idx]);
1682 idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, tn);
1683 ARR_APP1(int, tmp_idx_m1, na->ilp_vars.d[idx]);
1687 if (ARR_LEN(tmp_idx_1) > 0)
1688 lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_1, ARR_LEN(tmp_idx_1), 1.0);
1690 if (ARR_LEN(tmp_idx_m1) > 0)
1691 lpp_set_factor_fast_bulk(lpp, cst, tmp_idx_m1, ARR_LEN(tmp_idx_m1), -1.0);
1693 /* BEWARE: t is unsigned, so (double)(-t) won't work */
1694 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1695 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1697 DEL_ARR_F(tmp_idx_1);
1698 DEL_ARR_F(tmp_idx_m1);
1703 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1704 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1708 * Create ILP pressure constraints, based on alive nodes:
1709 * - add additional costs to objective function if a node is scheduled
1710 * on a unit although all units of this type are currently occupied
1712 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1715 unsigned num_cst = 0;
1716 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1717 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1719 ilp_timer_push(t_cst);
1720 /* y_{nt}^k is set for each node and timestep and unit type */
1721 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1722 unsigned cur_idx = get_irn_idx(cur_irn);
1723 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1724 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1727 /* we ignore nodes assigned to DUMMY unit here */
1728 if (cur_na->is_dummy_node)
1732 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1733 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1737 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1739 /* check if node can be executed on this unit type */
1740 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1744 /* check all time_steps at which the current node can be scheduled */
1745 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1748 int *tmp_var_idx = NEW_ARR_F(int, 0);
1749 ilp_livein_node_t *livein;
1751 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1752 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1753 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1756 /* - accumulate all nodes alive at point t on unit type k */
1757 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1758 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1759 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1762 /* check if node can be alive here */
1763 if (t < na->asap - 1)
1766 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1768 /* current type is not suitable */
1772 a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1773 ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1775 /* do the same for livein nodes */
1776 foreach_pset(ba->livein_nodes, livein) {
1777 ir_node *irn = livein->irn;
1778 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1781 /* check if node can be alive here */
1782 if (t >= livein->max_alive_steps)
1785 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1787 /* current type is not suitable */
1791 a_idx = tp_idx * livein->max_alive_steps + t;
1792 ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1795 if (ARR_LEN(tmp_var_idx) > 0)
1796 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1797 DEL_ARR_F(tmp_var_idx);
1799 /* - num_nodes * y_{nt}^k */
1800 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1801 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1806 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1807 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1811 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1814 unsigned num_cst = 0;
1815 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1816 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1818 ilp_timer_push(t_cst);
1819 /* check all nodes */
1820 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1821 be_ilpsched_irn_t *node;
1822 ilpsched_node_attr_t *na;
1826 /* only mode_T nodes can have Projs and Keeps assigned */
1827 if (get_irn_mode(irn) != mode_T)
1830 node = get_ilpsched_irn(env, irn);
1831 na = get_ilpsched_node_attr(node);
1833 /* check if has some Projs and Keeps assigned */
1834 if (! na->projkeeps)
1837 /* we can run only once over the queue, so preserve the nodes */
1838 pk = NEW_ARR_F(ir_node *, 0);
1839 while (! waitq_empty(na->projkeeps))
1840 ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1841 del_waitq(na->projkeeps);
1842 na->projkeeps = NULL;
1844 /* for all time steps at which this node can be scheduled */
1845 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1847 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1849 /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1850 snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1851 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1852 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1855 /* sum up scheduling variables for this time step */
1856 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1857 int idx = ILPVAR_IDX(na, tp_idx, t);
1858 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1861 if (ARR_LEN(tmp_var_idx_n) > 0)
1862 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1863 DEL_ARR_F(tmp_var_idx_n);
1865 /* subtract all Proj and Keep variables for this step */
1866 for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1867 be_ilpsched_irn_t *pk_node = get_ilpsched_irn(env, pk[i]);
1868 ilpsched_node_attr_t *pk_na = get_ilpsched_node_attr(pk_node);
1871 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1872 int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1873 lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1879 DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1880 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1884 /***************************************************
1886 * |_ _| | | __ \ (_)
1887 * | | | | | |__) | _ __ ___ __ _ _ _ __
1888 * | | | | | ___/ | '_ ` _ \ / _` | | '_ \
1889 * _| |_| |____| | | | | | | | (_| | | | | |
1890 * |_____|______|_| |_| |_| |_|\__,_|_|_| |_|
1892 ***************************************************/
1895 * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1897 static void create_ilp(ir_node *block, void *walk_env) {
1898 be_ilpsched_env_t *env = walk_env;
1899 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
1900 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1901 FILE *logfile = NULL;
1903 struct obstack var_obst;
1906 DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1907 DBG((env->dbg, 255, " ILP Scheduling for %+F\n", block));
1908 DBG((env->dbg, 255, "=========================================\n\n"));
1910 DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1911 block, ba->n_interesting_nodes, ba->max_steps));
1913 /* notify backend and get block environment */
1914 env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1916 /* if we have less than two interesting nodes, there is no need to create the ILP */
1917 if (ba->n_interesting_nodes > 1) {
1918 double fact_var = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1919 double fact_cst = ba->n_interesting_nodes > 25 ? 3 : 4.5;
1920 int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes;
1921 int estimated_n_var = (int)((double)base_num * fact_var);
1922 int estimated_n_cst = (int)((double)base_num * fact_cst);
1924 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1925 estimated_n_var, estimated_n_cst));
1927 /* set up the LPP object */
1928 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1930 lpp = new_lpp_userdef(
1933 estimated_n_cst, /* num vars */
1934 estimated_n_cst + 1, /* num cst */
1935 1.3); /* grow factor */
1936 obstack_init(&var_obst);
1938 /* create ILP variables */
1939 create_variables(env, lpp, block_node, &var_obst);
1941 /* create ILP constraints */
1942 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1943 create_assignment_and_precedence_constraints(env, lpp, block_node);
1944 create_ressource_constraints(env, lpp, block_node);
1945 create_bundle_constraints(env, lpp, block_node);
1946 //create_proj_keep_constraints(env, lpp, block_node);
1948 if (env->opts->regpress) {
1949 if (ba->n_interesting_nodes > env->opts->limit_dead) {
1950 create_alive_nodes_constraint(env, lpp, block_node);
1951 create_alive_livein_nodes_constraint(env, lpp, block_node);
1952 create_pressure_alive_constraint(env, lpp, block_node);
1954 create_dying_nodes_constraint(env, lpp, block_node);
1955 create_pressure_dead_constraint(env, lpp, block_node);
1959 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1961 /* debug stuff, dump lpp when debugging is on */
1963 if (firm_dbg_get_mask(env->dbg) > 1) {
1967 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1968 f = fopen(buf, "w");
1969 lpp_dump_plain(lpp, f);
1971 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1976 /* set solve time limit */
1977 lpp_set_time_limit(lpp, env->opts->time_limit);
1979 /* set logfile if requested */
1980 if (strlen(env->opts->log_file) > 0) {
1981 if (strcasecmp(env->opts->log_file, "stdout") == 0)
1982 lpp_set_log(lpp, stdout);
1983 else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1984 lpp_set_log(lpp, stderr);
1986 logfile = fopen(env->opts->log_file, "w");
1988 fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1990 lpp_set_log(lpp, logfile);
1995 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
2000 /* check for valid solution */
2001 if (! lpp_is_sol_valid(lpp)) {
2005 snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.txt", get_irn_node_nr(block));
2006 f = fopen(buf, "w");
2007 lpp_dump_plain(lpp, f);
2009 snprintf(buf, sizeof(buf), "lpp_block_%lu.assert.mps", get_irn_node_nr(block));
2011 dump_ir_block_graph(env->irg, "-assert");
2013 assert(0 && "ILP solution is not feasible!");
2016 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
2017 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
2018 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
2019 DBG((env->dbg, LEVEL_1, "\tmatrix: %u elements, density %.2f%%, size %.2fMB\n", lpp->n_elems, lpp->density, (double)lpp->matrix_mem / 1024.0 / 1024.0));
2020 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
2021 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
2022 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
2023 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
2025 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
2028 /* apply solution */
2029 apply_solution(env, lpp, block);
2034 /* notify backend */
2035 be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2039 * Perform ILP scheduling on the given irg.
2041 void be_ilp_sched(const be_irg_t *birg) {
2042 be_ilpsched_env_t env;
2043 const char *name = "be ilp scheduling";
2044 arch_isa_t *isa = birg->main_env->arch_env->isa;
2045 const ilp_sched_selector_t *sel = isa->impl->get_ilp_sched_selector(isa);
2047 FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2049 // firm_dbg_set_mask(env.dbg, 1);
2051 env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
2053 env.irg = birg->irg;
2054 env.height = heights_new(birg->irg);
2055 env.main_env = birg->main_env;
2056 env.arch_env = birg->main_env->arch_env;
2057 env.cpu = arch_isa_get_machine(birg->main_env->arch_env->isa);
2058 env.opts = &ilp_opts;
2059 phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
2061 /* assign a unique per block number to all interesting nodes */
2062 irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2065 The block indices are completely build after the walk,
2066 now we can allocate the bitsets (size depends on block indices)
2069 phase_reinit_irn_data(&env.ph);
2071 /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2072 irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2074 /* Calculate ALAP of all irns */
2075 irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2077 /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2078 irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2080 /* perform ILP scheduling */
2081 irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2084 if (firm_dbg_get_mask(env.dbg)) {
2086 phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2088 fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2092 /* free data allocated dynamically */
2093 irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2095 /* free all allocated object */
2096 phase_free(&env.ph);
2097 heights_free(env.height);
2099 /* notify backend */
2100 be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2104 * Register ILP scheduler options.
2106 void be_init_ilpsched(void)
2108 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2109 lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2111 lc_opt_add_table(sched_grp, ilpsched_option_table);
2114 #else /* WITH_ILP */
2116 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2119 #endif /* WITH_ILP */