2 * Scheduling algorithms.
3 * An ILP scheduler based on
4 * "ILP-based Instruction Scheduling for IA-64"
5 * by Daniel Kaestner and Sebastian Winkel
6 * extended with register pressure constraints by Christian Wuerdig
9 * @author Christian Wuerdig
27 #include "irphase_t.h"
39 #include <lpp/lpp_net.h>
41 #include <libcore/lc_opts.h>
42 #include <libcore/lc_opts_enum.h>
43 #include <libcore/lc_timing.h>
47 #include "besched_t.h"
48 #include "beilpsched.h"
52 typedef struct _ilpsched_options_t {
58 typedef struct _unit_type_info_t {
60 const be_execution_unit_type_t *tp;
64 * holding the ILP variables of the different types
66 typedef struct _ilp_var_types_t {
67 int *x; /* x_{nt}^k variables */
68 int *a; /* a_{nt}^k variables */
69 int *y; /* y_{nt}^k variables */
73 * Holds alive variables for a node live-in to a block.
75 typedef struct _ilp_livein_node_t {
77 unsigned max_alive_steps;
81 /* attributes for a node */
82 typedef struct _ilpsched_node_attr_t {
83 unsigned asap; /**< The ASAP scheduling control step */
84 unsigned alap; /**< The ALAP scheduling control step */
85 unsigned latency; /**< Latency of this node (needed for sorting) */
86 unsigned sched_point; /**< the step in which the node is finally scheduled */
87 unsigned visit_idx; /**< Index of the node having visited this node last */
88 unsigned consumer_idx; /**< Index of the node having counted this node as consumer last */
89 unsigned n_consumer; /**< Number of consumers */
90 ir_node **block_consumer; /**< List of consumer being in the same block */
91 waitq *projkeeps; /**< A List of Projs and Keeps belonging to this node */
92 unsigned block_idx : 30; /**< A unique per block index */
93 unsigned alap_changed : 1; /**< the current ALAP has changed, revisit preds */
94 unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */
95 bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors
96 for ASAP, successors for ALAP */
97 unsigned n_unit_types; /**< number of allowed execution unit types */
98 unit_type_info_t *type_info; /**< list of allowed execution unit types */
99 ilp_var_types_t ilp_vars; /**< the different ILP variables */
100 } ilpsched_node_attr_t;
102 /* attributes for a block */
103 typedef struct _ilpsched_block_attr_t {
104 unsigned block_last_idx; /**< The highest node index in block so far */
105 unsigned n_interesting_nodes; /**< The number of nodes interesting for scheduling */
106 unsigned max_steps; /**< Upper bound for block execution */
107 plist_t *root_nodes; /**< A list of nodes having no user in current block */
108 ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */
109 pset *livein_nodes; /**< A set of nodes which are live-in to this block */
110 } ilpsched_block_attr_t;
112 typedef union _ilpsched_attr_ {
113 ilpsched_node_attr_t node_attr;
114 ilpsched_block_attr_t block_attr;
117 /* A irn for the phase and it's attributes (either node or block) */
120 ilpsched_attr_t attr;
123 /* The ILP scheduling environment */
125 phase_t ph; /**< The phase */
126 ir_graph *irg; /**< The current irg */
127 heights_t *height; /**< The heights object of the irg */
128 void *irg_env; /**< An environment for the irg scheduling, provided by the backend */
129 void *block_env; /**< An environment for scheduling a block, provided by the backend */
130 const arch_env_t *arch_env;
131 const arch_isa_t *isa; /**< The ISA */
132 const be_main_env_t *main_env;
133 const be_machine_t *cpu; /**< the current abstract machine */
134 ilpsched_options_t *opts; /**< the ilp options for current irg */
135 const be_irg_t *birg; /**< The birg object */
136 be_options_t *be_opts; /**< backend options */
137 const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */
138 DEBUG_ONLY(firm_dbg_module_t *dbg);
141 /* convenience macros to handle phase irn data */
142 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
143 #define is_ilpsched_block(node) (is_Block((node)->irn))
144 #define get_ilpsched_block_attr(block) (&(block)->attr.block_attr)
145 #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr)
147 /* check if node is considered for ILP scheduling */
148 #define consider_for_sched(isa, irn) \
149 (! (is_Block(irn) || \
150 is_normal_Proj(isa, irn) || \
157 /* gives the valid scheduling time step interval for a node */
158 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
160 /* gives the valid interval where a node can die */
161 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
163 /* gives the corresponding ILP variable for given node, unit and time step */
164 #define ILPVAR_IDX(na, unit, control_step) \
165 ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
167 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
168 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
169 ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
171 /* check if a double value is within an epsilon environment of 0 */
172 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
174 #define ilp_timer_push(t) lc_timer_push((t))
175 #define ilp_timer_pop() lc_timer_pop()
176 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
178 /* option variable */
179 static ilpsched_options_t ilp_opts = {
180 1, /* default is with register pressure constraints */
181 300, /* 300 sec per block time limit */
186 static const lc_opt_table_entry_t ilpsched_option_table[] = {
187 LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress),
188 LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
189 LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
194 We need this global variable as we compare nodes dependent on heights,
195 but we cannot pass any information to the qsort compare function.
197 static heights_t *glob_heights;
200 * Check if irn is a Proj, which has no execution units assigned.
201 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
203 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
204 return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
208 * Skips normal Projs.
209 * @return predecessor if irn is a normal Proj, otherwise irn.
211 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
212 if (is_normal_Proj(isa, irn))
213 return get_Proj_pred(irn);
217 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
218 unsigned lat = be_ilp_sched_latency(sel, irn, env);
219 if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
224 static int cmp_live_in_nodes(const void *a, const void *b) {
225 const ilp_livein_node_t *n1 = a;
226 const ilp_livein_node_t *n2 = b;
228 return n1->irn != n2->irn;
232 * Compare scheduling time steps of two be_ilpsched_irn's.
234 static int cmp_ilpsched_irn(const void *a, const void *b) {
235 be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a;
236 be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b;
237 ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
238 ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
240 if (n1_a->sched_point == n2_a->sched_point) {
241 ir_node *irn_a = n1->irn;
242 ir_node *irn_b = n2->irn;
244 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
246 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
250 Ok, timestep is equal and the nodes are parallel,
251 so check latency and schedule high latency first.
253 return QSORT_CMP(n2_a->latency, n1_a->latency);
256 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
260 * In case there is no phase information for irn, initialize it.
262 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
263 be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
266 /* if we have already some data: check for reinitialization */
268 if (! is_Block(irn)) {
269 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
271 if (! na->transitive_block_nodes) {
272 ir_node *block = get_nodes_block(irn);
273 be_ilpsched_irn_t *block_node = phase_get_or_set_irn_data(ph, block);
274 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
276 /* we are called after the block indices have been build: create bitset */
277 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
280 /* we are called from reinit block data: clear the bitset */
281 bitset_clear_all(na->transitive_block_nodes);
283 na->alap_changed = 1;
291 /* set ilpsched irn attributes (either block or irn) */
293 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
295 ba->n_interesting_nodes = 0;
296 ba->block_last_idx = 0;
297 ba->root_nodes = plist_new();
298 ba->head_ilp_nodes = NULL;
299 ba->livein_nodes = new_pset(cmp_live_in_nodes, 16);
303 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
304 memset(na, 0, sizeof(*na));
311 * Assign a per block unique number to each node.
313 static void build_block_idx(ir_node *irn, void *walk_env) {
314 be_ilpsched_env_t *env = walk_env;
315 be_ilpsched_irn_t *node, *block_node;
316 ilpsched_node_attr_t *na;
317 ilpsched_block_attr_t *ba;
319 set_irn_link(irn, NULL);
320 if (! consider_for_sched(env->arch_env->isa, irn))
323 node = get_ilpsched_irn(env, irn);
324 na = get_ilpsched_node_attr(node);
325 block_node = get_ilpsched_irn(env, get_nodes_block(irn));
326 ba = get_ilpsched_block_attr(block_node);
328 na->block_idx = ba->block_last_idx++;
331 /********************************************************
334 * __ _ ___ __ _ _ __ / / __ _| | __ _ _ __
335 * / _` / __|/ _` | '_ \ / / / _` | |/ _` | '_ \
336 * | (_| \__ \ (_| | |_) | / / | (_| | | (_| | |_) |
337 * \__,_|___/\__,_| .__/ /_/ \__,_|_|\__,_| .__/
340 ********************************************************/
343 * Add all nodes having no user in current block to last_nodes list.
345 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
347 const ir_edge_t *edge;
348 be_ilpsched_irn_t *block_node, *node;
349 ilpsched_block_attr_t *ba;
350 ilpsched_node_attr_t *na;
352 be_ilpsched_env_t *env = walk_env;
353 int has_block_user = 0;
354 unsigned n_consumer = 0;
355 ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
359 if (! consider_for_sched(env->arch_env->isa, irn))
362 block = get_nodes_block(irn);
363 idx = get_irn_idx(irn);
364 consumer = NEW_ARR_F(ir_node *, 0);
366 DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
368 /* check data and dependency out edges */
369 for (i = 0; i < 2 && ! has_block_user; ++i) {
370 foreach_out_edge_kind(irn, edge, ekind[i]) {
371 ir_node *user = get_edge_src_irn(edge);
373 if (is_normal_Proj(env->arch_env->isa, user)) {
374 const ir_edge_t *user_edge;
376 if (get_irn_mode(user) == mode_X)
379 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
380 for (j = 0; j < 2; ++j) {
381 foreach_out_edge_kind(user, user_edge, ekind[j]) {
382 ir_node *real_user = get_edge_src_irn(user_edge);
384 if (! is_Phi(real_user) && ! is_Block(real_user)) {
385 be_ilpsched_irn_t *node = get_ilpsched_irn(env, real_user);
386 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
388 /* skip already visited nodes */
389 if (ua->consumer_idx == idx)
392 /* check if node has user in this block and collect the user if it's a data user */
393 if (get_nodes_block(real_user) == block) {
394 if (i == 0 && j == 0)
395 ARR_APP1(ir_node *, consumer, real_user);
399 /* only count data consumer */
403 /* mark user as visited by this node */
404 ua->consumer_idx = idx;
409 else if (is_Block(user)) {
412 else if (! is_Phi(user)) {
413 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
414 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
416 /* skip already visited nodes */
417 if (ua->consumer_idx == idx)
420 /* check if node has user in this block and collect the user if it's a data user */
421 if (get_nodes_block(user) == block) {
423 ARR_APP1(ir_node *, consumer, user);
427 /* only count data consumer */
431 /* mark user visited by this node */
432 ua->consumer_idx = idx;
434 else if (get_nodes_block(user) != block) {
440 block_node = get_ilpsched_irn(env, block);
441 ba = get_ilpsched_block_attr(block_node);
443 ba->n_interesting_nodes++;
445 /* current irn has no user inside this block, add to queue */
446 if (! has_block_user) {
447 DB((env->dbg, LEVEL_3, "root node\n"));
448 plist_insert_back(ba->root_nodes, irn);
451 DB((env->dbg, LEVEL_3, "normal node\n"));
454 /* record number of all consumer and the consumer within the same block */
455 node = get_ilpsched_irn(env, irn);
456 na = get_ilpsched_node_attr(node);
457 na->n_consumer = n_consumer;
458 na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
459 memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
464 * Calculate the ASAP scheduling step for current irn.
466 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
467 be_ilpsched_env_t *env = walk_env;
470 be_ilpsched_irn_t *node, *block_node;
471 ilpsched_node_attr_t *na;
472 ilpsched_block_attr_t *ba;
474 /* These nodes are handled separate */
475 if (! consider_for_sched(env->arch_env->isa, irn))
478 DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
480 block = get_nodes_block(irn);
481 node = get_ilpsched_irn(env, irn);
482 na = get_ilpsched_node_attr(node);
485 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
486 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
488 /* check for greatest distance to top */
489 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
490 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
491 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
494 lat = fixed_latency(env->sel, pred, env->block_env);
496 na->asap = MAX(na->asap, pna->asap + lat);
500 /* add node to ILP node list and update max_steps */
501 block_node = get_ilpsched_irn(env, block);
502 ba = get_ilpsched_block_attr(block_node);
504 set_irn_link(irn, ba->head_ilp_nodes);
505 ba->head_ilp_nodes = irn;
506 ba->max_steps += fixed_latency(env->sel, irn, env->block_env);
508 DB((env->dbg, LEVEL_2, "%u\n", na->asap));
512 * Calculate the ALAP scheduling step of all irns in current block.
513 * Depends on max_steps being calculated.
515 static void calculate_block_alap(ir_node *block, void *walk_env) {
516 be_ilpsched_env_t *env = walk_env;
517 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
518 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
519 waitq *cur_queue = new_waitq();
522 assert(is_Block(block));
524 DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
525 block, ba->n_interesting_nodes, ba->max_steps));
527 /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
528 //phase_reinit_block_irn_data(&env->ph, block);
530 /* init start queue */
531 foreach_plist(ba->root_nodes, el) {
532 waitq_put(cur_queue, plist_element_get_value(el));
535 /* repeat until all nodes are processed */
536 while (! waitq_empty(cur_queue)) {
537 waitq *next_queue = new_waitq();
539 /* process all nodes in current step */
540 while (! waitq_empty(cur_queue)) {
541 ir_node *cur_irn = waitq_get(cur_queue);
542 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
543 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
546 /* cur_node has no alap set -> it's a root node, set to max alap */
548 na->alap = ba->max_steps;
549 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
553 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
557 /* set the alap's of all predecessors */
558 for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
559 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
561 /* check for greatest distance to bottom */
562 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
563 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
564 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
567 /* mark the predecessor as visited by current irn */
568 if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
570 pna->visit_idx = get_irn_idx(cur_irn);
572 lat = fixed_latency(env->sel, pred, env->block_env);
574 /* set ALAP of current pred */
575 if (pna->alap == 0) {
576 /* current ALAP is 0: node has not yet been visited */
577 pna->alap_changed = 1;
578 pna->alap = na->alap - lat;
580 else if (pna->alap > na->alap - lat) {
581 /* we found a longer path to root node: change ALAP */
582 pna->alap = na->alap - lat;
583 pna->alap_changed = 1;
586 /* current ALAP is best found so far: keep it */
587 pna->alap_changed = 0;
590 DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
592 /* enqueue node for next iteration */
593 if (get_irn_ins_or_deps(pred) > 0)
594 waitq_put(next_queue, pred);
599 /* prepare for next iteration */
600 del_waitq(cur_queue);
601 cur_queue = next_queue;
606 * Free list of root nodes and the set of live-in nodes.
608 static void clear_unwanted_data(ir_node *block, void *walk_env) {
609 be_ilpsched_env_t *env = walk_env;
610 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
611 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
613 plist_free(ba->root_nodes);
614 ba->root_nodes = NULL;
615 del_pset(ba->livein_nodes);
616 ba->livein_nodes = NULL;
620 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
621 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
623 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
624 be_ilpsched_env_t *env = walk_env;
626 be_ilpsched_irn_t *node, *pred_node;
627 ilpsched_node_attr_t *na, *pna;
629 if (! consider_for_sched(env->arch_env->isa, irn))
632 if (! is_Proj(irn) && ! be_is_Keep(irn))
635 /* go to the ancestor */
637 pred = get_irn_n(irn, 0);
638 pred = skip_Proj(pred);
640 node = get_ilpsched_irn(env, irn);
641 pred_node = get_ilpsched_irn(env, pred);
642 na = get_ilpsched_node_attr(node);
643 pna = get_ilpsched_node_attr(pred_node);
645 na->asap = pna->asap;
646 na->alap = pna->alap;
648 /* record all Projs and Keeps for this node */
649 if (! pna->projkeeps)
650 pna->projkeeps = new_waitq();
651 waitq_put(pna->projkeeps, irn);
653 DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
656 /*******************************************
659 * ___ ___| |__ ___ __| |_ _| | ___
660 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
661 * \__ \ (__| | | | __/ (_| | |_| | | __/
662 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
664 *******************************************/
666 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
667 const ir_edge_t *edge;
669 foreach_out_edge(irn, edge) {
670 ir_node *user = get_edge_src_irn(edge);
672 if (be_is_Keep(user)) {
673 assert(get_nodes_block(user) == block && "Keep must not be in different block.");
674 waitq_put(keeps, user);
680 * Inserts @p irn before @p before into schedule and notifies backend.
682 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
683 ir_node *before, ir_node *irn, unsigned cycle)
685 be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
686 sched_add_before(before, irn);
690 * Adds a node, it's Projs (in case of mode_T nodes) and
691 * it's Keeps to schedule.
693 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
694 const ir_edge_t *edge;
695 waitq *keeps = new_waitq();
697 /* mode_M nodes are not scheduled */
698 if (get_irn_mode(irn) == mode_M)
701 if (! sched_is_scheduled(irn))
702 notified_sched_add_before(env, block, irn, cycle);
705 if (get_irn_mode(irn) == mode_T) {
706 foreach_out_edge(irn, edge) {
707 ir_node *user = get_edge_src_irn(edge);
709 if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
710 get_irn_n_edges(user) > 0)
712 notified_sched_add_before(env, block, user, cycle);
715 check_for_keeps(keeps, block, user);
719 check_for_keeps(keeps, block, irn);
723 while (! waitq_empty(keeps)) {
724 ir_node *keep = waitq_get(keeps);
725 if (! sched_is_scheduled(keep))
726 notified_sched_add_before(env, block, keep, cycle);
733 * Schedule all nodes in the given block, according to the ILP solution.
735 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
736 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
737 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
738 sched_info_t *info = get_irn_sched_info(block);
739 be_ilpsched_irn_t **sched_nodes;
742 const ir_edge_t *edge;
744 /* init block schedule list */
745 INIT_LIST_HEAD(&info->list);
748 /* collect nodes and their scheduling time step */
749 sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
750 if (ba->n_interesting_nodes == 0) {
753 else if (ba->n_interesting_nodes == 1) {
754 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
756 /* add the single node */
757 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
760 /* check all nodes for their positive solution */
761 foreach_linked_irns(ba->head_ilp_nodes, irn) {
762 be_ilpsched_irn_t *node;
763 ilpsched_node_attr_t *na;
767 node = get_ilpsched_irn(env, irn);
768 na = get_ilpsched_node_attr(node);
772 if (! na->is_dummy_node) {
773 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
774 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
775 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
777 if (! LPP_VALUE_IS_0(cost)) {
778 DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
786 /* go over all variables of a node until the non-zero one is found */
787 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
788 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
789 double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
791 /* check, if variable is set to one (it's not zero then :) */
792 if (! LPP_VALUE_IS_0(val)) {
794 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
795 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
796 irn, t, na->type_info[tp_idx].tp->name));
803 glob_heights = env->height;
804 /* sort nodes ascending by scheduling time step */
805 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
808 /* make all Phis ready and remember the single cf op */
810 foreach_out_edge(block, edge) {
811 irn = get_edge_src_irn(edge);
813 switch (get_irn_opcode(irn)) {
815 add_to_sched(env, block, irn, 0);
825 assert(cfop == NULL && "Highlander - there can be only one");
832 /* add all nodes from list */
833 for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
834 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
835 if (sched_nodes[i]->irn != cfop)
836 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
839 /* schedule control flow node if not already done */
840 if (cfop && ! sched_is_scheduled(cfop))
841 add_to_sched(env, block, cfop, 0);
843 DEL_ARR_F(sched_nodes);
846 /***************************************************************
847 * _____ _ _____ _____ _ _
848 * |_ _| | | __ \ / ____| | | (_)
849 * | | | | | |__) | | (___ ___ ___| |_ _ ___ _ __
850 * | | | | | ___/ \___ \ / _ \/ __| __| |/ _ \| '_ \
851 * _| |_| |____| | ____) | __/ (__| |_| | (_) | | | |
852 * |_____|______|_| |_____/ \___|\___|\__|_|\___/|_| |_|
854 ***************************************************************/
857 * Check if node can be executed on given unit type.
859 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
861 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
863 for (i = na->n_unit_types - 1; i >= 0; --i) {
864 if (na->type_info[i].tp == tp)
871 /************************************************
874 * __ ____ _ _ __ _ __ _| |__ | | ___ ___
875 * \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
876 * \ V / (_| | | | | (_| | |_) | | __/\__ \
877 * \_/ \__,_|_| |_|\__,_|_.__/|_|\___||___/
879 ************************************************/
881 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
882 const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
883 unsigned n_unit_types = 0;
884 be_ilpsched_irn_t *node;
885 ilpsched_node_attr_t *na;
886 unsigned unit_idx, tp_idx;
888 /* count number of available unit types for this node */
889 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
892 node = get_ilpsched_irn(env, irn);
893 na = get_ilpsched_node_attr(node);
895 if (! na->type_info) {
896 na->n_unit_types = n_unit_types;
897 na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
899 /* fill the type info array */
900 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
901 for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
902 /* beware: we also count number of available units here */
903 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
904 na->is_dummy_node = 1;
907 na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp;
908 na->type_info[tp_idx].n_units = unit_idx;
916 * Returns the largest alap time of a user of @p irn.
917 * The user must be in block @p block.
919 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
920 const ir_edge_t *edge;
921 unsigned max_alap = 0;
923 foreach_out_edge(irn, edge) {
924 ir_node *user = get_edge_src_irn(edge);
926 if (get_nodes_block(user) == block) {
927 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
928 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
930 max_alap = MAX(max_alap, na->alap);
934 assert(max_alap > 0);
939 * Create the following variables:
940 * - x_{nt}^k binary weigthed with: t
941 * node n is scheduled at time step t to unit type k
942 * ==>> These variables represent the schedule
944 * - a_{nt}^k binary weighted with num_nodes
945 * node n is alive at time step t on unit type k
947 * - y_{nt}^k continuous weighted with: num_nodes^2
948 * register pressure over limit for unit type k
949 * ==>> These variables represent the register pressure
952 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
955 unsigned num_block_var, num_nodes;
956 ilp_livein_node_t *livein;
957 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
958 unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
959 lc_timer_t *t_var = lc_timer_register("beilpsched_var", "create ilp variables");
961 ilp_timer_push(t_var);
962 num_block_var = num_nodes = 0;
963 foreach_linked_irns(ba->head_ilp_nodes, irn) {
964 be_ilpsched_irn_t *node;
965 ilpsched_node_attr_t *na;
966 unsigned n_unit_types, tp_idx, n_var, cur_unit;
967 unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad;
970 node = get_ilpsched_irn(env, irn);
971 na = get_ilpsched_node_attr(node);
972 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
974 /* allocate space for ilp variables */
975 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
976 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
978 /* we need these variables only for "real" nodes */
979 if (! na->is_dummy_node) {
980 na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
981 memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
983 num_ad = ba->max_steps - na->asap + 1;
984 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
985 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
988 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
989 irn, na->asap, na->alap, na->n_unit_types));
991 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
992 /* create variables */
993 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
996 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
997 /* x_{nt}^k variables */
998 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
999 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1000 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1001 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1002 /* variable counter */
1006 if (! na->is_dummy_node) {
1007 /* y_{nt}^k variables */
1008 snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1009 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1010 na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1011 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1013 /* variable counter */
1019 /* a node can die at any step t: asap(n) <= t <= U */
1020 if (! na->is_dummy_node) {
1021 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1023 /* a_{nt}^k variables */
1024 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1025 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1026 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1027 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1029 /* variable counter */
1035 /* collect live-in nodes */
1036 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1037 ir_node *pred = get_irn_n(irn, i);
1039 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1040 be_ilpsched_set_type_info(env, pred, var_obst);
1041 if (! na->is_dummy_node) {
1042 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1045 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1051 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1055 /* create alive variables a_{nt}^k for live-ins */
1056 foreach_pset(ba->livein_nodes, livein) {
1057 be_ilpsched_irn_t *node;
1058 ilpsched_node_attr_t *na;
1059 unsigned tp_idx, var_idx;
1063 node = get_ilpsched_irn(env, irn);
1064 na = get_ilpsched_node_attr(node);
1066 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1068 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1071 /* create variables */
1072 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1075 for (t = 0; t < livein->max_alive_steps; ++t) {
1076 /* a_{nt}^k variables */
1077 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1078 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1079 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1080 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1087 DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1088 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1091 /*******************************************************
1094 * ___ ___ _ __ ___| |_ _ __ __ _ _ _ __ | |_ ___
1095 * / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1096 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1097 * \___\___/|_| |_|___/\__|_| \__,_|_|_| |_|\__|___/
1099 *******************************************************/
1102 * Collect all operands and nodes @p irn depends on.
1103 * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1105 static nodeset *sta_collect_in_deps(ir_node *irn, nodeset *deps) {
1108 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1109 ir_node *p = get_irn_in_or_dep(irn, i);
1112 const ir_edge_t *edge;
1114 p = get_Proj_pred(p);
1115 foreach_out_edge(p, edge) {
1116 ir_node *src = get_edge_src_irn(edge);
1117 nodeset_insert(deps, src);
1121 nodeset_insert(deps, p);
1129 * Create following ILP constraints:
1130 * - the assignment constraints:
1131 * assure each node is executed once by exactly one (allowed) execution unit
1132 * - the dead node assignment constraints:
1133 * assure a node can only die at most once
1134 * - the precedence constraints:
1135 * assure that no data dependencies are violated
1137 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1138 unsigned num_cst_assign, num_cst_prec, num_cst_dead;
1141 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1142 bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx);
1143 lc_timer_t *t_cst_assign = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1144 lc_timer_t *t_cst_prec = lc_timer_register("beilpsched_cst_prec", "create precedence constraints");
1146 num_cst_assign = num_cst_prec = num_cst_dead = 0;
1147 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1150 be_ilpsched_irn_t *node;
1151 ilpsched_node_attr_t *na;
1153 nodeset *deps = new_nodeset(16);
1155 node = get_ilpsched_irn(env, irn);
1156 na = get_ilpsched_node_attr(node);
1159 /* the assignment constraint */
1160 ilp_timer_push(t_cst_assign);
1161 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1162 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1163 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1166 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1169 /* We have separate constraints for Projs and Keeps */
1170 // ILP becomes infeasible ?!?
1171 // if (is_Proj(irn) || be_is_Keep(irn))
1174 /* the precedence constraints */
1175 ilp_timer_push(t_cst_prec);
1176 bs_block_irns = bitset_clear_all(bs_block_irns);
1178 deps = sta_collect_in_deps(irn, deps);
1179 foreach_nodeset(deps, pred) {
1180 unsigned t_low, t_high, t;
1181 be_ilpsched_irn_t *pred_node;
1182 ilpsched_node_attr_t *pna;
1185 pred = skip_normal_Proj(env->arch_env->isa, pred);
1186 if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1189 pred_node = get_ilpsched_irn(env, pred);
1190 pna = get_ilpsched_node_attr(pred_node);
1192 assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1194 if (! bitset_is_set(bs_block_irns, pna->block_idx))
1195 bitset_set(bs_block_irns, pna->block_idx);
1199 /* irn = n, pred = m */
1200 delay = fixed_latency(env->sel, pred, env->block_env);
1201 t_low = MAX(na->asap, pna->asap + delay - 1);
1202 t_high = MIN(na->alap, pna->alap + delay - 1);
1203 for (t = t_low - 1; t <= t_high - 1; ++t) {
1205 int *tmp_var_idx = NEW_ARR_F(int, 0);
1207 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1208 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1209 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1212 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1213 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1214 /* node variables have smaller index than pred variables */
1215 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1216 for (tn = na->asap - 1; tn <= t; ++tn) {
1217 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1218 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1222 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1223 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1224 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1225 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1230 /* pred variables have smaller index than node variables */
1231 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1232 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1233 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1234 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1238 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1239 for (tn = na->asap - 1; tn <= t; ++tn) {
1240 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1241 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1246 if (ARR_LEN(tmp_var_idx) > 0)
1247 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1249 DEL_ARR_F(tmp_var_idx);
1255 DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1256 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1257 DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1258 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1262 * Create ILP resource constraints:
1263 * - assure that for each time step not more instructions are scheduled
1264 * to the same unit types as units of this type are available
1266 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1269 unsigned num_cst_resrc = 0;
1270 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1271 lc_timer_t *t_cst_rsrc = lc_timer_register("beilpsched_cst_rsrc", "create resource constraints");
1273 ilp_timer_push(t_cst_rsrc);
1274 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1276 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1278 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1280 /* check each time step */
1281 for (t = 0; t < ba->max_steps; ++t) {
1284 int *tmp_var_idx = NEW_ARR_F(int, 0);
1286 snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1287 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1288 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1291 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1292 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1293 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1296 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1298 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1299 int cur_var = ILPVAR_IDX(na, tp_idx, t);
1300 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1304 /* set constraints if we have some */
1305 if (ARR_LEN(tmp_var_idx) > 0)
1306 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1308 DEL_ARR_F(tmp_var_idx);
1312 DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1313 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1317 * Create ILP bundle constraints:
1318 * - assure, at most bundle_size * bundles_per_cycle instructions
1319 * can be started at a certain point.
1321 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1324 unsigned num_cst_bundle = 0;
1325 unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1326 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1327 lc_timer_t *t_cst_bundle = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1329 ilp_timer_push(t_cst_bundle);
1330 for (t = 0; t < ba->max_steps; ++t) {
1333 int *tmp_var_idx = NEW_ARR_F(int, 0);
1335 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1336 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1337 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1340 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1341 be_ilpsched_irn_t *node;
1342 ilpsched_node_attr_t *na;
1345 /* Projs and Keeps do not contribute to bundle size */
1346 if (is_Proj(irn) || be_is_Keep(irn))
1349 node = get_ilpsched_irn(env, irn);
1350 na = get_ilpsched_node_attr(node);
1352 /* nodes assigned to DUMMY unit do not contribute to bundle size */
1353 if (na->is_dummy_node)
1356 if (t >= na->asap - 1 && t <= na->alap - 1) {
1357 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1358 int idx = ILPVAR_IDX(na, tp_idx, t);
1359 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1364 if (ARR_LEN(tmp_var_idx) > 0)
1365 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1367 DEL_ARR_F(tmp_var_idx);
1370 DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1371 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1375 * Create ILP alive nodes constraints:
1376 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1378 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1381 unsigned num_cst = 0;
1382 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1383 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1385 ilp_timer_push(t_cst);
1387 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1388 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1389 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1392 /* we ignore nodes assigned to dummy unit here */
1393 if (na->is_dummy_node)
1396 /* check check all time steps: asap(n) <= t <= U */
1397 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1400 /* for all unit types available for this node */
1401 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1402 unsigned tn, tn_max, idx;
1404 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1405 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1407 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1408 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1409 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1410 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1413 tn_max = MIN(na->alap - 1, t);
1414 /* check if the node has been scheduled so far */
1415 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1416 int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1417 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1420 if (ARR_LEN(tmp_var_idx_n) > 0)
1421 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1422 DEL_ARR_F(tmp_var_idx_n);
1424 /* subtract the number of consumer scheduled so far */
1425 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1426 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1427 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1429 unsigned tm, tm_max;
1431 tm_max = MIN(ca->alap - 1, t);
1432 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1433 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1434 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1435 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1440 if (ARR_LEN(tmp_var_idx_m) > 0)
1441 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1442 DEL_ARR_F(tmp_var_idx_m);
1445 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1446 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1452 DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1453 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1457 * Create ILP alive nodes constraints for live-in nodes:
1458 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1460 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1462 ilp_livein_node_t *livein;
1463 unsigned num_cst = 0;
1464 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1465 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1467 ilp_timer_push(t_cst);
1469 foreach_pset(ba->livein_nodes, livein) {
1470 ir_node *irn = livein->irn;
1471 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1472 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1475 /* check check all time steps: 0 <= t < max_alive_steps */
1476 for (t = 0; t < livein->max_alive_steps; ++t) {
1479 /* for all unit types available for this node */
1480 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1481 const ir_edge_t *edge;
1483 int cst, num_block_user;
1484 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1486 /* check the number of consumer scheduled so far */
1488 foreach_out_edge(irn, edge) {
1489 ir_node *user = get_edge_src_irn(edge);
1490 be_ilpsched_irn_t *cons;
1491 ilpsched_node_attr_t *ca;
1493 unsigned tm, tm_max;
1495 /* check only users within current block */
1496 if (get_nodes_block(user) != block_node->irn)
1500 cons = get_ilpsched_irn(env, user);
1501 ca = get_ilpsched_node_attr(cons);
1503 tm_max = MIN(ca->alap - 1, t);
1504 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1505 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1506 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1507 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1512 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1513 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1514 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1515 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1518 /* sum(scheduled users) */
1519 if (ARR_LEN(tmp_var_idx_m) > 0)
1520 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1521 DEL_ARR_F(tmp_var_idx_m);
1523 /* + c * a_{nt}^k */
1524 idx = node_tp_idx * livein->max_alive_steps + t;
1525 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1530 DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1531 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1535 * Create ILP pressure constraints, based on alive nodes:
1536 * - add additional costs to objective function if a node is scheduled
1537 * on a unit although all units of this type are currently occupied
1539 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1542 unsigned num_cst = 0;
1543 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1544 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1546 ilp_timer_push(t_cst);
1547 /* y_{nt}^k is set for each node and timestep and unit type */
1548 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1549 unsigned cur_idx = get_irn_idx(cur_irn);
1550 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1551 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1554 /* we ignore nodes assigned to DUMMY unit here */
1555 if (cur_na->is_dummy_node)
1559 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1560 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1564 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1566 /* check if node can be executed on this unit type */
1567 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1571 /* check all time_steps at which the current node can be scheduled */
1572 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1575 int *tmp_var_idx = NEW_ARR_F(int, 0);
1576 ilp_livein_node_t *livein;
1578 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1579 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1580 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1583 /* - accumulate all nodes alive at point t on unit type k */
1584 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1585 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1586 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1589 /* check if node can be alive here */
1590 if (t < na->asap - 1)
1593 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1595 /* current type is not suitable */
1599 a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1600 ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1602 /* do the same for livein nodes */
1603 foreach_pset(ba->livein_nodes, livein) {
1604 ir_node *irn = livein->irn;
1605 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1608 /* check if node can be alive here */
1609 if (t >= livein->max_alive_steps)
1612 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1614 /* current type is not suitable */
1618 a_idx = tp_idx * livein->max_alive_steps + t;
1619 ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1622 if (ARR_LEN(tmp_var_idx) > 0)
1623 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1624 DEL_ARR_F(tmp_var_idx);
1626 /* - num_nodes * y_{nt}^k */
1627 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1628 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1633 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1634 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1638 * Create ILP branch constraints:
1639 * Assure, alle nodes are scheduled prior to cfg op.
1641 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1643 ir_node *cur_irn, *cfop;
1644 unsigned num_cst = 0;
1645 unsigned num_non_branches = 0;
1646 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1647 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1649 ilp_timer_push(t_cst);
1651 /* determine number of non-branch nodes and the one and only branch node */
1652 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1653 switch (get_irn_opcode(cur_irn)) {
1663 if (is_cfop(cur_irn)) {
1664 assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1675 be_ilpsched_irn_t *cf_node = get_ilpsched_irn(env, cfop);
1676 ilpsched_node_attr_t *cf_na = get_ilpsched_node_attr(cf_node);
1679 /* for each time step */
1680 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1681 int *non_branch_vars, *branch_vars;
1684 snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1685 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1686 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1689 /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1690 non_branch_vars = NEW_ARR_F(int, 0);
1691 branch_vars = NEW_ARR_F(int, 0);
1692 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1693 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
1694 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1697 if (cur_irn == cfop) {
1698 /* for all unit types available for this node */
1699 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1700 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1701 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1705 /* sum up all possible schedule points for this node upto current timestep */
1706 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1708 unsigned tmax = MIN(t, na->alap - 1);
1710 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1711 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1712 ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1719 if (ARR_LEN(non_branch_vars) > 0)
1720 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1721 if (ARR_LEN(branch_vars) > 0)
1722 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1724 DEL_ARR_F(branch_vars);
1725 DEL_ARR_F(non_branch_vars);
1729 DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1730 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1734 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1737 unsigned num_cst = 0;
1738 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1739 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1741 ilp_timer_push(t_cst);
1742 /* check all nodes */
1743 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1744 be_ilpsched_irn_t *node;
1745 ilpsched_node_attr_t *na;
1749 /* only mode_T nodes can have Projs and Keeps assigned */
1750 if (get_irn_mode(irn) != mode_T)
1753 node = get_ilpsched_irn(env, irn);
1754 na = get_ilpsched_node_attr(node);
1756 /* check if has some Projs and Keeps assigned */
1757 if (! na->projkeeps)
1760 /* we can run only once over the queue, so preserve the nodes */
1761 pk = NEW_ARR_F(ir_node *, 0);
1762 while (! waitq_empty(na->projkeeps))
1763 ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1764 del_waitq(na->projkeeps);
1765 na->projkeeps = NULL;
1767 /* for all time steps at which this node can be scheduled */
1768 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1770 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1772 /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1773 snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1774 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1775 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1778 /* sum up scheduling variables for this time step */
1779 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1780 int idx = ILPVAR_IDX(na, tp_idx, t);
1781 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1784 if (ARR_LEN(tmp_var_idx_n) > 0)
1785 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1786 DEL_ARR_F(tmp_var_idx_n);
1788 /* subtract all Proj and Keep variables for this step */
1789 for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1790 be_ilpsched_irn_t *pk_node = get_ilpsched_irn(env, pk[i]);
1791 ilpsched_node_attr_t *pk_na = get_ilpsched_node_attr(pk_node);
1794 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1795 int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1796 lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1802 DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1803 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1807 /***************************************************
1809 * |_ _| | | __ \ (_)
1810 * | | | | | |__) | _ __ ___ __ _ _ _ __
1811 * | | | | | ___/ | '_ ` _ \ / _` | | '_ \
1812 * _| |_| |____| | | | | | | | (_| | | | | |
1813 * |_____|______|_| |_| |_| |_|\__,_|_|_| |_|
1815 ***************************************************/
1818 * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1820 static void create_ilp(ir_node *block, void *walk_env) {
1821 be_ilpsched_env_t *env = walk_env;
1822 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
1823 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1824 FILE *logfile = NULL;
1827 struct obstack var_obst;
1830 DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1831 DBG((env->dbg, 255, " ILP Scheduling for %+F\n", block));
1832 DBG((env->dbg, 255, "=========================================\n\n"));
1834 DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1835 block, ba->n_interesting_nodes, ba->max_steps));
1837 /* notify backend and get block environment */
1838 env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1840 /* if we have less than two interesting nodes, there is no need to create the ILP */
1841 if (ba->n_interesting_nodes > 1) {
1842 double fact_var = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1843 double fact_cst = ba->n_interesting_nodes > 25 ? 3 : 4.5;
1844 int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes;
1845 int estimated_n_var = (int)((double)base_num * fact_var);
1846 int estimated_n_cst = (int)((double)base_num * fact_cst);
1848 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1849 estimated_n_var, estimated_n_cst));
1851 /* set up the LPP object */
1852 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1854 lpp = new_lpp_userdef(
1857 estimated_n_cst, /* num vars */
1858 estimated_n_cst + 1, /* num cst */
1859 1.3); /* grow factor */
1860 obstack_init(&var_obst);
1862 /* create ILP variables */
1863 create_variables(env, lpp, block_node, &var_obst);
1865 /* create ILP constraints */
1866 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1867 create_assignment_and_precedence_constraints(env, lpp, block_node);
1868 create_ressource_constraints(env, lpp, block_node);
1869 create_bundle_constraints(env, lpp, block_node);
1870 create_branch_constraint(env, lpp, block_node);
1871 //create_proj_keep_constraints(env, lpp, block_node);
1873 if (env->opts->regpress) {
1874 create_alive_nodes_constraint(env, lpp, block_node);
1875 create_alive_livein_nodes_constraint(env, lpp, block_node);
1876 create_pressure_alive_constraint(env, lpp, block_node);
1879 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1881 /* debug stuff, dump lpp when debugging is on */
1883 if (firm_dbg_get_mask(env->dbg) > 1) {
1887 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1888 f = fopen(buf, "w");
1889 lpp_dump_plain(lpp, f);
1891 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1896 /* set solve time limit */
1897 lpp_set_time_limit(lpp, env->opts->time_limit);
1899 /* set logfile if requested */
1900 if (strlen(env->opts->log_file) > 0) {
1901 if (strcasecmp(env->opts->log_file, "stdout") == 0)
1902 lpp_set_log(lpp, stdout);
1903 else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1904 lpp_set_log(lpp, stderr);
1906 logfile = fopen(env->opts->log_file, "w");
1908 fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1910 lpp_set_log(lpp, logfile);
1915 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1920 /* check for valid solution */
1921 if (! lpp_is_sol_valid(lpp)) {
1926 if (firm_dbg_get_mask(env->dbg) >= 2) {
1927 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1928 f = fopen(buf, "w");
1929 lpp_dump_plain(lpp, f);
1931 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1933 dump_ir_block_graph(env->irg, "-infeasible");
1937 ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1941 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1942 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1943 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1944 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));
1945 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1946 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1947 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1948 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1950 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1953 /* apply solution */
1954 #ifdef FIRM_STATISTICS
1955 if (be_stat_ev_is_active()) {
1956 be_stat_ev("nodes", ba->block_last_idx);
1957 be_stat_ev("vars", lpp ? lpp->var_next : 0);
1958 be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1960 #endif /* FIRM_STATISTICS */
1962 #ifdef FIRM_STATISTICS
1963 if (be_stat_ev_is_active()) {
1964 be_stat_ev("time", -1);
1965 be_stat_ev_dbl("opt", 0.0);
1967 #endif /* FIRM_STATISTICS */
1968 list_sched_single_block(env->birg, block, env->be_opts);
1971 #ifdef FIRM_STATISTICS
1972 if (be_stat_ev_is_active()) {
1974 double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1975 be_stat_ev_dbl("time", lpp->sol_time);
1976 be_stat_ev_dbl("opt", opt);
1979 be_stat_ev_dbl("time", 0.0);
1980 be_stat_ev_dbl("opt", 100.0);
1983 #endif /* FIRM_STATISTICS */
1984 apply_solution(env, lpp, block);
1990 /* notify backend */
1991 be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1995 * Perform ILP scheduling on the given irg.
1997 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
1998 be_ilpsched_env_t env;
1999 const char *name = "be ilp scheduling";
2000 arch_isa_t *isa = birg->main_env->arch_env->isa;
2001 const ilp_sched_selector_t *sel = isa->impl->get_ilp_sched_selector(isa);
2003 FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2005 #ifdef FIRM_STATISTICS
2006 if (be_stat_ev_is_active()) {
2007 be_stat_tags[STAT_TAG_CLS] = "ilpsched";
2008 be_stat_ev_push(be_stat_tags, STAT_TAG_LAST, be_stat_file);
2010 #endif /* FIRM_STATISTICS */
2012 // firm_dbg_set_mask(env.dbg, 1);
2014 env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
2016 env.irg = birg->irg;
2017 env.height = heights_new(birg->irg);
2018 env.main_env = birg->main_env;
2019 env.arch_env = birg->main_env->arch_env;
2020 env.cpu = arch_isa_get_machine(birg->main_env->arch_env->isa);
2021 env.opts = &ilp_opts;
2023 env.be_opts = be_opts;
2024 phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
2026 /* assign a unique per block number to all interesting nodes */
2027 irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2030 The block indices are completely build after the walk,
2031 now we can allocate the bitsets (size depends on block indices)
2034 phase_reinit_irn_data(&env.ph);
2036 /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2037 irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2039 /* Calculate ALAP of all irns */
2040 irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2042 /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2043 irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2045 /* perform ILP scheduling */
2046 irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2049 if (firm_dbg_get_mask(env.dbg)) {
2051 phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2053 fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2057 /* free data allocated dynamically */
2058 irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2060 /* free all allocated object */
2061 phase_free(&env.ph);
2062 heights_free(env.height);
2064 /* notify backend */
2065 be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2067 #ifdef FIRM_STATISTICS
2068 if (be_stat_ev_is_active()) {
2071 #endif /* FIRM_STATISTICS */
2075 * Register ILP scheduler options.
2077 void be_init_ilpsched(void)
2079 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2080 lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2082 lc_opt_add_table(sched_grp, ilpsched_option_table);
2085 #else /* WITH_ILP */
2087 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2090 #endif /* WITH_ILP */