2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief ILP based instruction scheduling.
23 * @author Christian Wuerdig
27 * An ILP scheduler based on
28 * "ILP-based Instruction Scheduling for IA-64"
29 * by Daniel Kaestner and Sebastian Winkel
30 * extended with register pressure constraints by Christian Wuerdig
47 #include "irphase_t.h"
60 #include <lpp/lpp_net.h>
63 #include "lc_opts_enum.h"
67 #include "besched_t.h"
68 #include "beilpsched.h"
73 typedef struct _ilpsched_options_t {
79 typedef struct _unit_type_info_t {
81 const be_execution_unit_type_t *tp;
85 * holding the ILP variables of the different types
87 typedef struct _ilp_var_types_t {
88 int *x; /* x_{nt}^k variables */
89 int *a; /* a_{nt}^k variables */
90 int *y; /* y_{nt}^k variables */
94 * Holds alive variables for a node live-in to a block.
96 typedef struct _ilp_livein_node_t {
98 unsigned max_alive_steps;
102 /* attributes for a node */
103 typedef struct _ilpsched_node_attr_t {
104 unsigned asap; /**< The ASAP scheduling control step */
105 unsigned alap; /**< The ALAP scheduling control step */
106 unsigned latency; /**< Latency of this node (needed for sorting) */
107 unsigned sched_point; /**< the step in which the node is finally scheduled */
108 unsigned visit_idx; /**< Index of the node having visited this node last */
109 unsigned consumer_idx; /**< Index of the node having counted this node as consumer last */
110 unsigned n_consumer; /**< Number of consumers */
111 ir_node **block_consumer; /**< List of consumer being in the same block */
112 waitq *projkeeps; /**< A List of Projs and Keeps belonging to this node */
113 unsigned block_idx : 30; /**< A unique per block index */
114 unsigned alap_changed : 1; /**< the current ALAP has changed, revisit preds */
115 unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */
116 bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors
117 for ASAP, successors for ALAP */
118 unsigned n_unit_types; /**< number of allowed execution unit types */
119 unit_type_info_t *type_info; /**< list of allowed execution unit types */
120 ilp_var_types_t ilp_vars; /**< the different ILP variables */
121 } ilpsched_node_attr_t;
123 /* attributes for a block */
124 typedef struct _ilpsched_block_attr_t {
125 unsigned block_last_idx; /**< The highest node index in block so far */
126 unsigned n_interesting_nodes; /**< The number of nodes interesting for scheduling */
127 unsigned max_steps; /**< Upper bound for block execution */
128 plist_t *root_nodes; /**< A list of nodes having no user in current block */
129 ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */
130 pset *livein_nodes; /**< A set of nodes which are live-in to this block */
131 } ilpsched_block_attr_t;
133 typedef union _ilpsched_attr_ {
134 ilpsched_node_attr_t node_attr;
135 ilpsched_block_attr_t block_attr;
138 /* A irn for the phase and it's attributes (either node or block) */
141 ilpsched_attr_t attr;
144 /* The ILP scheduling environment */
146 ir_phase ph; /**< The phase */
147 ir_graph *irg; /**< The current irg */
148 heights_t *height; /**< The heights object of the irg */
149 void *irg_env; /**< An environment for the irg scheduling, provided by the backend */
150 void *block_env; /**< An environment for scheduling a block, provided by the backend */
151 const arch_env_t *arch_env;
152 const be_main_env_t *main_env;
153 const be_machine_t *cpu; /**< the current abstract machine */
154 ilpsched_options_t *opts; /**< the ilp options for current irg */
155 const be_irg_t *birg; /**< The birg object */
156 be_options_t *be_opts; /**< backend options */
157 const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */
158 DEBUG_ONLY(firm_dbg_module_t *dbg);
161 /* convenience macros to handle phase irn data */
162 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
163 #define is_ilpsched_block(node) (is_Block((node)->irn))
164 #define get_ilpsched_block_attr(block) (&(block)->attr.block_attr)
165 #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr)
167 /* check if node is considered for ILP scheduling */
168 #define consider_for_sched(env, irn) \
169 (! (is_Block(irn) || \
170 is_normal_Proj(env, irn) || \
177 /* gives the valid scheduling time step interval for a node */
178 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
180 /* gives the valid interval where a node can die */
181 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
183 /* gives the corresponding ILP variable for given node, unit and time step */
184 #define ILPVAR_IDX(na, unit, control_step) \
185 ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
187 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
188 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
189 ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
191 /* check if a double value is within an epsilon environment of 0 */
192 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
194 #define ilp_timer_push(t) ir_timer_push((t))
195 #define ilp_timer_pop() ir_timer_pop()
196 #define ilp_timer_elapsed_usec(t) ir_timer_elapsed_usec((t))
198 /* option variable */
199 static ilpsched_options_t ilp_opts = {
200 1, /* default is with register pressure constraints */
201 300, /* 300 sec per block time limit */
206 static const lc_opt_table_entry_t ilpsched_option_table[] = {
207 LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress),
208 LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
209 LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
214 We need this global variable as we compare nodes dependent on heights,
215 but we cannot pass any information to the qsort compare function.
217 static heights_t *glob_heights;
220 * Check if irn is a Proj, which has no execution units assigned.
221 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
223 static INLINE int is_normal_Proj(const arch_env_t *env, const ir_node *irn) {
224 return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
228 * Skips normal Projs.
229 * @return predecessor if irn is a normal Proj, otherwise irn.
231 static INLINE ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn) {
232 if (is_normal_Proj(env, irn))
233 return get_Proj_pred(irn);
237 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
238 unsigned lat = be_ilp_sched_latency(sel, irn, env);
239 if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
244 static int cmp_live_in_nodes(const void *a, const void *b) {
245 const ilp_livein_node_t *n1 = a;
246 const ilp_livein_node_t *n2 = b;
248 return n1->irn != n2->irn;
252 * Compare scheduling time steps of two be_ilpsched_irn's.
254 static int cmp_ilpsched_irn(const void *a, const void *b) {
255 be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a;
256 be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b;
257 ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
258 ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
260 if (n1_a->sched_point == n2_a->sched_point) {
261 const ir_node *irn_a = n1->irn;
262 const ir_node *irn_b = n2->irn;
264 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
266 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
270 Ok, timestep is equal and the nodes are parallel,
271 so check latency and schedule high latency first.
273 return QSORT_CMP(n2_a->latency, n1_a->latency);
276 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
280 * In case there is no phase information for irn, initialize it.
282 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old) {
283 be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
286 /* if we have already some data: check for reinitialization */
288 if (! is_Block(irn)) {
289 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
291 if (! na->transitive_block_nodes) {
292 ir_node *block = get_nodes_block(irn);
293 be_ilpsched_irn_t *block_node = phase_get_or_set_irn_data(ph, block);
294 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
296 /* we are called after the block indices have been build: create bitset */
297 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
300 /* we are called from reinit block data: clear the bitset */
301 bitset_clear_all(na->transitive_block_nodes);
303 na->alap_changed = 1;
311 /* set ilpsched irn attributes (either block or irn) */
313 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
315 ba->n_interesting_nodes = 0;
316 ba->block_last_idx = 0;
317 ba->root_nodes = plist_new();
318 ba->head_ilp_nodes = NULL;
319 ba->livein_nodes = new_pset(cmp_live_in_nodes, 16);
323 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
324 memset(na, 0, sizeof(*na));
331 * Assign a per block unique number to each node.
333 static void build_block_idx(ir_node *irn, void *walk_env) {
334 be_ilpsched_env_t *env = walk_env;
335 be_ilpsched_irn_t *node, *block_node;
336 ilpsched_node_attr_t *na;
337 ilpsched_block_attr_t *ba;
339 set_irn_link(irn, NULL);
340 if (! consider_for_sched(env->arch_env, irn))
343 node = get_ilpsched_irn(env, irn);
344 na = get_ilpsched_node_attr(node);
345 block_node = get_ilpsched_irn(env, get_nodes_block(irn));
346 ba = get_ilpsched_block_attr(block_node);
348 na->block_idx = ba->block_last_idx++;
351 /********************************************************
354 * __ _ ___ __ _ _ __ / / __ _| | __ _ _ __
355 * / _` / __|/ _` | '_ \ / / / _` | |/ _` | '_ \
356 * | (_| \__ \ (_| | |_) | / / | (_| | | (_| | |_) |
357 * \__,_|___/\__,_| .__/ /_/ \__,_|_|\__,_| .__/
360 ********************************************************/
363 * Add all nodes having no user in current block to last_nodes list.
365 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
367 const ir_edge_t *edge;
368 be_ilpsched_irn_t *block_node, *node;
369 ilpsched_block_attr_t *ba;
370 ilpsched_node_attr_t *na;
372 be_ilpsched_env_t *env = walk_env;
373 int has_block_user = 0;
374 unsigned n_consumer = 0;
375 ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
379 if (! consider_for_sched(env->arch_env, irn))
382 block = get_nodes_block(irn);
383 idx = get_irn_idx(irn);
384 consumer = NEW_ARR_F(ir_node *, 0);
386 DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
388 /* check data and dependency out edges */
389 for (i = 0; i < 2 && ! has_block_user; ++i) {
390 foreach_out_edge_kind(irn, edge, ekind[i]) {
391 ir_node *user = get_edge_src_irn(edge);
393 if (is_normal_Proj(env->arch_env, user)) {
394 const ir_edge_t *user_edge;
396 if (get_irn_mode(user) == mode_X)
399 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
400 for (j = 0; j < 2; ++j) {
401 foreach_out_edge_kind(user, user_edge, ekind[j]) {
402 ir_node *real_user = get_edge_src_irn(user_edge);
404 if (! is_Phi(real_user) && ! is_Block(real_user)) {
405 be_ilpsched_irn_t *node = get_ilpsched_irn(env, real_user);
406 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
408 /* skip already visited nodes */
409 if (ua->consumer_idx == idx)
412 /* check if node has user in this block and collect the user if it's a data user */
413 if (get_nodes_block(real_user) == block) {
414 if (i == 0 && j == 0)
415 ARR_APP1(ir_node *, consumer, real_user);
419 /* only count data consumer */
423 /* mark user as visited by this node */
424 ua->consumer_idx = idx;
429 else if (is_Block(user)) {
432 else if (! is_Phi(user)) {
433 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
434 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
436 /* skip already visited nodes */
437 if (ua->consumer_idx == idx)
440 /* check if node has user in this block and collect the user if it's a data user */
441 if (get_nodes_block(user) == block) {
443 ARR_APP1(ir_node *, consumer, user);
447 /* only count data consumer */
451 /* mark user visited by this node */
452 ua->consumer_idx = idx;
454 else if (get_nodes_block(user) != block) {
460 block_node = get_ilpsched_irn(env, block);
461 ba = get_ilpsched_block_attr(block_node);
463 ba->n_interesting_nodes++;
465 /* current irn has no user inside this block, add to queue */
466 if (! has_block_user) {
467 DB((env->dbg, LEVEL_3, "root node\n"));
468 plist_insert_back(ba->root_nodes, irn);
471 DB((env->dbg, LEVEL_3, "normal node\n"));
474 /* record number of all consumer and the consumer within the same block */
475 node = get_ilpsched_irn(env, irn);
476 na = get_ilpsched_node_attr(node);
477 na->n_consumer = n_consumer;
478 na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
479 memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
484 * Calculate the ASAP scheduling step for current irn.
486 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
487 be_ilpsched_env_t *env = walk_env;
490 be_ilpsched_irn_t *node, *block_node;
491 ilpsched_node_attr_t *na;
492 ilpsched_block_attr_t *ba;
494 /* These nodes are handled separate */
495 if (! consider_for_sched(env->arch_env, irn))
498 DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
500 block = get_nodes_block(irn);
501 node = get_ilpsched_irn(env, irn);
502 na = get_ilpsched_node_attr(node);
505 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
506 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
508 /* check for greatest distance to top */
509 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
510 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
511 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
514 lat = fixed_latency(env->sel, pred, env->block_env);
516 na->asap = MAX(na->asap, pna->asap + lat);
520 /* add node to ILP node list and update max_steps */
521 block_node = get_ilpsched_irn(env, block);
522 ba = get_ilpsched_block_attr(block_node);
524 set_irn_link(irn, ba->head_ilp_nodes);
525 ba->head_ilp_nodes = irn;
526 ba->max_steps += fixed_latency(env->sel, irn, env->block_env);
528 DB((env->dbg, LEVEL_2, "%u\n", na->asap));
532 * Calculate the ALAP scheduling step of all irns in current block.
533 * Depends on max_steps being calculated.
535 static void calculate_block_alap(ir_node *block, void *walk_env) {
536 be_ilpsched_env_t *env = walk_env;
537 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
538 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
539 waitq *cur_queue = new_waitq();
542 assert(is_Block(block));
544 DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
545 block, ba->n_interesting_nodes, ba->max_steps));
547 /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
548 //phase_reinit_block_irn_data(&env->ph, block);
550 /* init start queue */
551 foreach_plist(ba->root_nodes, el) {
552 waitq_put(cur_queue, plist_element_get_value(el));
555 /* repeat until all nodes are processed */
556 while (! waitq_empty(cur_queue)) {
557 waitq *next_queue = new_waitq();
559 /* process all nodes in current step */
560 while (! waitq_empty(cur_queue)) {
561 ir_node *cur_irn = waitq_get(cur_queue);
562 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
563 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
566 /* cur_node has no alap set -> it's a root node, set to max alap */
568 na->alap = ba->max_steps;
569 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
573 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
577 /* set the alap's of all predecessors */
578 for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
579 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
581 /* check for greatest distance to bottom */
582 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
583 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
584 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
587 /* mark the predecessor as visited by current irn */
588 if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
590 pna->visit_idx = get_irn_idx(cur_irn);
592 lat = fixed_latency(env->sel, pred, env->block_env);
594 /* set ALAP of current pred */
595 if (pna->alap == 0) {
596 /* current ALAP is 0: node has not yet been visited */
597 pna->alap_changed = 1;
598 pna->alap = na->alap - lat;
600 else if (pna->alap > na->alap - lat) {
601 /* we found a longer path to root node: change ALAP */
602 pna->alap = na->alap - lat;
603 pna->alap_changed = 1;
606 /* current ALAP is best found so far: keep it */
607 pna->alap_changed = 0;
610 DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
612 /* enqueue node for next iteration */
613 if (get_irn_ins_or_deps(pred) > 0)
614 waitq_put(next_queue, pred);
619 /* prepare for next iteration */
620 del_waitq(cur_queue);
621 cur_queue = next_queue;
626 * Free list of root nodes and the set of live-in nodes.
628 static void clear_unwanted_data(ir_node *block, void *walk_env) {
629 be_ilpsched_env_t *env = walk_env;
630 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
631 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
633 plist_free(ba->root_nodes);
634 ba->root_nodes = NULL;
635 del_pset(ba->livein_nodes);
636 ba->livein_nodes = NULL;
640 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
641 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
643 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
644 be_ilpsched_env_t *env = walk_env;
646 be_ilpsched_irn_t *node, *pred_node;
647 ilpsched_node_attr_t *na, *pna;
649 if (! consider_for_sched(env->arch_env, irn))
652 if (! is_Proj(irn) && ! be_is_Keep(irn))
655 /* go to the ancestor */
657 pred = get_irn_n(irn, 0);
658 pred = skip_Proj(pred);
660 node = get_ilpsched_irn(env, irn);
661 pred_node = get_ilpsched_irn(env, pred);
662 na = get_ilpsched_node_attr(node);
663 pna = get_ilpsched_node_attr(pred_node);
665 na->asap = pna->asap;
666 na->alap = pna->alap;
668 /* record all Projs and Keeps for this node */
669 if (! pna->projkeeps)
670 pna->projkeeps = new_waitq();
671 waitq_put(pna->projkeeps, irn);
673 DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
676 /*******************************************
679 * ___ ___| |__ ___ __| |_ _| | ___
680 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
681 * \__ \ (__| | | | __/ (_| | |_| | | __/
682 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
684 *******************************************/
686 static INLINE void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn) {
687 const ir_edge_t *edge;
690 foreach_out_edge(irn, edge) {
691 ir_node *user = get_edge_src_irn(edge);
693 if (be_is_Keep(user)) {
694 assert(get_nodes_block(user) == block && "Keep must not be in different block.");
695 waitq_put(keeps, user);
701 * Inserts @p irn before @p before into schedule and notifies backend.
703 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
704 const ir_node *before, const ir_node *irn, unsigned cycle)
706 be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
707 sched_add_before(before, irn);
711 * Adds a node, it's Projs (in case of mode_T nodes) and
712 * it's Keeps to schedule.
714 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle) {
715 const ir_edge_t *edge;
716 waitq *keeps = new_waitq();
718 /* mode_M nodes are not scheduled */
719 if (get_irn_mode(irn) == mode_M)
722 if (! sched_is_scheduled(irn))
723 notified_sched_add_before(env, block, irn, cycle);
726 if (get_irn_mode(irn) == mode_T) {
727 foreach_out_edge(irn, edge) {
728 ir_node *user = get_edge_src_irn(edge);
730 if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
731 get_irn_n_edges(user) > 0)
733 notified_sched_add_before(env, block, user, cycle);
736 check_for_keeps(keeps, block, user);
740 check_for_keeps(keeps, block, irn);
744 while (! waitq_empty(keeps)) {
745 ir_node *keep = waitq_get(keeps);
746 if (! sched_is_scheduled(keep))
747 notified_sched_add_before(env, block, keep, cycle);
754 * Schedule all nodes in the given block, according to the ILP solution.
756 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
757 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
758 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
759 sched_info_t *info = get_irn_sched_info(block);
760 be_ilpsched_irn_t **sched_nodes;
763 const ir_edge_t *edge;
765 /* init block schedule list */
766 INIT_LIST_HEAD(&info->list);
769 /* collect nodes and their scheduling time step */
770 sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
771 if (ba->n_interesting_nodes == 0) {
774 else if (ba->n_interesting_nodes == 1) {
775 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
777 /* add the single node */
778 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
781 /* check all nodes for their positive solution */
782 foreach_linked_irns(ba->head_ilp_nodes, irn) {
783 be_ilpsched_irn_t *node;
784 ilpsched_node_attr_t *na;
788 node = get_ilpsched_irn(env, irn);
789 na = get_ilpsched_node_attr(node);
793 if (! na->is_dummy_node) {
794 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
795 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
796 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
798 if (! LPP_VALUE_IS_0(cost)) {
799 DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
807 /* go over all variables of a node until the non-zero one is found */
808 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
809 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
810 double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
812 /* check, if variable is set to one (it's not zero then :) */
813 if (! LPP_VALUE_IS_0(val)) {
815 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
816 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
817 irn, t, na->type_info[tp_idx].tp->name));
824 glob_heights = env->height;
825 /* sort nodes ascending by scheduling time step */
826 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
829 /* make all Phis ready and remember the single cf op */
831 foreach_out_edge(block, edge) {
832 irn = get_edge_src_irn(edge);
834 switch (get_irn_opcode(irn)) {
836 add_to_sched(env, block, irn, 0);
846 assert(cfop == NULL && "Highlander - there can be only one");
853 /* add all nodes from list */
854 for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
855 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
856 if (sched_nodes[i]->irn != cfop)
857 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
860 /* schedule control flow node if not already done */
861 if (cfop && ! sched_is_scheduled(cfop))
862 add_to_sched(env, block, cfop, 0);
864 DEL_ARR_F(sched_nodes);
867 /***************************************************************
868 * _____ _ _____ _____ _ _
869 * |_ _| | | __ \ / ____| | | (_)
870 * | | | | | |__) | | (___ ___ ___| |_ _ ___ _ __
871 * | | | | | ___/ \___ \ / _ \/ __| __| |/ _ \| '_ \
872 * _| |_| |____| | ____) | __/ (__| |_| | (_) | | | |
873 * |_____|______|_| |_____/ \___|\___|\__|_|\___/|_| |_|
875 ***************************************************************/
878 * Check if node can be executed on given unit type.
880 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
882 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
884 for (i = na->n_unit_types - 1; i >= 0; --i) {
885 if (na->type_info[i].tp == tp)
892 /************************************************
895 * __ ____ _ _ __ _ __ _| |__ | | ___ ___
896 * \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
897 * \ V / (_| | | | | (_| | |_) | | __/\__ \
898 * \_/ \__,_|_| |_|\__,_|_.__/|_|\___||___/
900 ************************************************/
902 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
903 const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
904 unsigned n_unit_types = 0;
905 be_ilpsched_irn_t *node;
906 ilpsched_node_attr_t *na;
907 unsigned unit_idx, tp_idx;
909 /* count number of available unit types for this node */
910 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
913 node = get_ilpsched_irn(env, irn);
914 na = get_ilpsched_node_attr(node);
916 if (! na->type_info) {
917 na->n_unit_types = n_unit_types;
918 na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
920 /* fill the type info array */
921 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
922 for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
923 /* beware: we also count number of available units here */
924 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
925 na->is_dummy_node = 1;
928 na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp;
929 na->type_info[tp_idx].n_units = unit_idx;
937 * Returns the largest alap time of a user of @p irn.
938 * The user must be in block @p block.
940 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block) {
941 const ir_edge_t *edge;
942 unsigned max_alap = 0;
944 foreach_out_edge(irn, edge) {
945 ir_node *user = get_edge_src_irn(edge);
947 if (get_nodes_block(user) == block) {
948 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
949 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
951 max_alap = MAX(max_alap, na->alap);
955 assert(max_alap > 0);
960 * Create the following variables:
961 * - x_{nt}^k binary weigthed with: t
962 * node n is scheduled at time step t to unit type k
963 * ==>> These variables represent the schedule
965 * - a_{nt}^k binary weighted with num_nodes
966 * node n is alive at time step t on unit type k
968 * - y_{nt}^k continuous weighted with: num_nodes^2
969 * register pressure over limit for unit type k
970 * ==>> These variables represent the register pressure
973 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
976 unsigned num_block_var, num_nodes;
977 ilp_livein_node_t *livein;
978 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
979 unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
980 ir_timer_t *t_var = ir_timer_register("beilpsched_var", "create ilp variables");
982 ilp_timer_push(t_var);
983 num_block_var = num_nodes = 0;
984 foreach_linked_irns(ba->head_ilp_nodes, irn) {
985 be_ilpsched_irn_t *node;
986 ilpsched_node_attr_t *na;
987 unsigned n_unit_types, tp_idx, n_var, cur_unit;
988 unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad;
991 node = get_ilpsched_irn(env, irn);
992 na = get_ilpsched_node_attr(node);
993 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
995 /* allocate space for ilp variables */
996 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
997 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
999 /* we need these variables only for "real" nodes */
1000 if (! na->is_dummy_node) {
1001 na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1002 memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1004 num_ad = ba->max_steps - na->asap + 1;
1005 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1006 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1009 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1010 irn, na->asap, na->alap, na->n_unit_types));
1012 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1013 /* create variables */
1014 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1017 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1018 /* x_{nt}^k variables */
1019 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1020 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1021 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1022 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1023 /* variable counter */
1027 if (! na->is_dummy_node) {
1028 /* y_{nt}^k variables */
1029 snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1030 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1031 na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1032 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1034 /* variable counter */
1040 /* a node can die at any step t: asap(n) <= t <= U */
1041 if (! na->is_dummy_node) {
1042 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1044 /* a_{nt}^k variables */
1045 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1046 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1047 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1048 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1050 /* variable counter */
1056 /* collect live-in nodes */
1057 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1058 ir_node *pred = get_irn_n(irn, i);
1060 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1061 be_ilpsched_set_type_info(env, pred, var_obst);
1062 if (! na->is_dummy_node) {
1063 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1066 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1072 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1076 /* create alive variables a_{nt}^k for live-ins */
1077 foreach_pset(ba->livein_nodes, livein) {
1078 be_ilpsched_irn_t *node;
1079 ilpsched_node_attr_t *na;
1080 unsigned tp_idx, var_idx;
1084 node = get_ilpsched_irn(env, irn);
1085 na = get_ilpsched_node_attr(node);
1087 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1089 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1092 /* create variables */
1093 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1096 for (t = 0; t < livein->max_alive_steps; ++t) {
1097 /* a_{nt}^k variables */
1098 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1099 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1100 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1101 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1108 DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1109 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1112 /*******************************************************
1115 * ___ ___ _ __ ___| |_ _ __ __ _ _ _ __ | |_ ___
1116 * / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1117 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1118 * \___\___/|_| |_|___/\__|_| \__,_|_|_| |_|\__|___/
1120 *******************************************************/
1123 * Collect all operands and nodes @p irn depends on.
1124 * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1126 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps) {
1129 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1130 ir_node *p = get_irn_in_or_dep(irn, i);
1133 const ir_edge_t *edge;
1135 p = get_Proj_pred(p);
1136 foreach_out_edge(p, edge) {
1137 ir_node *src = get_edge_src_irn(edge);
1138 ir_nodeset_insert(deps, src);
1142 ir_nodeset_insert(deps, p);
1148 * Create following ILP constraints:
1149 * - the assignment constraints:
1150 * assure each node is executed once by exactly one (allowed) execution unit
1151 * - the dead node assignment constraints:
1152 * assure a node can only die at most once
1153 * - the precedence constraints:
1154 * assure that no data dependencies are violated
1156 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1157 unsigned num_cst_assign, num_cst_prec, num_cst_dead;
1160 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1161 bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx);
1162 ir_timer_t *t_cst_assign = ir_timer_register("beilpsched_cst_assign", "create assignment constraints");
1163 ir_timer_t *t_cst_prec = ir_timer_register("beilpsched_cst_prec", "create precedence constraints");
1165 num_cst_assign = num_cst_prec = num_cst_dead = 0;
1166 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1169 be_ilpsched_irn_t *node;
1170 ilpsched_node_attr_t *na;
1173 ir_nodeset_iterator_t iter;
1175 ir_nodeset_init(&deps);
1177 node = get_ilpsched_irn(env, irn);
1178 na = get_ilpsched_node_attr(node);
1181 /* the assignment constraint */
1182 ilp_timer_push(t_cst_assign);
1183 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1184 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1185 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1188 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1191 /* We have separate constraints for Projs and Keeps */
1192 // ILP becomes infeasible ?!?
1193 // if (is_Proj(irn) || be_is_Keep(irn))
1196 /* the precedence constraints */
1197 ilp_timer_push(t_cst_prec);
1198 bs_block_irns = bitset_clear_all(bs_block_irns);
1200 sta_collect_in_deps(irn, &deps);
1201 foreach_ir_nodeset(&deps, pred, iter) {
1202 unsigned t_low, t_high, t;
1203 be_ilpsched_irn_t *pred_node;
1204 ilpsched_node_attr_t *pna;
1207 pred = skip_normal_Proj(env->arch_env, pred);
1208 if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1211 pred_node = get_ilpsched_irn(env, pred);
1212 pna = get_ilpsched_node_attr(pred_node);
1214 assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1216 if (! bitset_is_set(bs_block_irns, pna->block_idx))
1217 bitset_set(bs_block_irns, pna->block_idx);
1221 /* irn = n, pred = m */
1222 delay = fixed_latency(env->sel, pred, env->block_env);
1223 t_low = MAX(na->asap, pna->asap + delay - 1);
1224 t_high = MIN(na->alap, pna->alap + delay - 1);
1225 for (t = t_low - 1; t <= t_high - 1; ++t) {
1227 int *tmp_var_idx = NEW_ARR_F(int, 0);
1229 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1230 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1231 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1234 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1235 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1236 /* node variables have smaller index than pred variables */
1237 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1238 for (tn = na->asap - 1; tn <= t; ++tn) {
1239 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1240 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1244 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1245 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1246 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1247 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1252 /* pred variables have smaller index than node variables */
1253 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1254 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1255 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1256 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1260 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1261 for (tn = na->asap - 1; tn <= t; ++tn) {
1262 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1263 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1268 if (ARR_LEN(tmp_var_idx) > 0)
1269 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1271 DEL_ARR_F(tmp_var_idx);
1274 ir_nodeset_destroy(&deps);
1277 DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1278 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1279 DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1280 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1284 * Create ILP resource constraints:
1285 * - assure that for each time step not more instructions are scheduled
1286 * to the same unit types as units of this type are available
1288 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1291 unsigned num_cst_resrc = 0;
1292 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1293 ir_timer_t *t_cst_rsrc = ir_timer_register("beilpsched_cst_rsrc", "create resource constraints");
1295 ilp_timer_push(t_cst_rsrc);
1296 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1298 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1300 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1302 /* check each time step */
1303 for (t = 0; t < ba->max_steps; ++t) {
1306 int *tmp_var_idx = NEW_ARR_F(int, 0);
1308 snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1309 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1310 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1313 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1314 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1315 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1318 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1320 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1321 int cur_var = ILPVAR_IDX(na, tp_idx, t);
1322 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1326 /* set constraints if we have some */
1327 if (ARR_LEN(tmp_var_idx) > 0)
1328 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1330 DEL_ARR_F(tmp_var_idx);
1334 DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1335 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1339 * Create ILP bundle constraints:
1340 * - assure, at most bundle_size * bundles_per_cycle instructions
1341 * can be started at a certain point.
1343 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1346 unsigned num_cst_bundle = 0;
1347 unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1348 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1349 ir_timer_t *t_cst_bundle = ir_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1351 ilp_timer_push(t_cst_bundle);
1352 for (t = 0; t < ba->max_steps; ++t) {
1355 int *tmp_var_idx = NEW_ARR_F(int, 0);
1357 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1358 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1359 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1362 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1363 be_ilpsched_irn_t *node;
1364 ilpsched_node_attr_t *na;
1367 /* Projs and Keeps do not contribute to bundle size */
1368 if (is_Proj(irn) || be_is_Keep(irn))
1371 node = get_ilpsched_irn(env, irn);
1372 na = get_ilpsched_node_attr(node);
1374 /* nodes assigned to DUMMY unit do not contribute to bundle size */
1375 if (na->is_dummy_node)
1378 if (t >= na->asap - 1 && t <= na->alap - 1) {
1379 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1380 int idx = ILPVAR_IDX(na, tp_idx, t);
1381 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1386 if (ARR_LEN(tmp_var_idx) > 0)
1387 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1389 DEL_ARR_F(tmp_var_idx);
1392 DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1393 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1397 * Create ILP alive nodes constraints:
1398 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1400 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1403 unsigned num_cst = 0;
1404 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1405 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1407 ilp_timer_push(t_cst);
1409 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1410 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1411 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1414 /* we ignore nodes assigned to dummy unit here */
1415 if (na->is_dummy_node)
1418 /* check check all time steps: asap(n) <= t <= U */
1419 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1422 /* for all unit types available for this node */
1423 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1424 unsigned tn, tn_max, idx;
1426 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1427 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1429 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1430 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1431 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1432 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1435 tn_max = MIN(na->alap - 1, t);
1436 /* check if the node has been scheduled so far */
1437 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1438 int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1439 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1442 if (ARR_LEN(tmp_var_idx_n) > 0)
1443 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1444 DEL_ARR_F(tmp_var_idx_n);
1446 /* subtract the number of consumer scheduled so far */
1447 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1448 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1449 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1451 unsigned tm, tm_max;
1453 tm_max = MIN(ca->alap - 1, t);
1454 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1455 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1456 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1457 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1462 if (ARR_LEN(tmp_var_idx_m) > 0)
1463 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1464 DEL_ARR_F(tmp_var_idx_m);
1467 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1468 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1474 DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1475 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1479 * Create ILP alive nodes constraints for live-in nodes:
1480 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1482 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1484 ilp_livein_node_t *livein;
1485 unsigned num_cst = 0;
1486 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1487 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1489 ilp_timer_push(t_cst);
1491 foreach_pset(ba->livein_nodes, livein) {
1492 ir_node *irn = livein->irn;
1493 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1494 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1497 /* check check all time steps: 0 <= t < max_alive_steps */
1498 for (t = 0; t < livein->max_alive_steps; ++t) {
1501 /* for all unit types available for this node */
1502 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1503 const ir_edge_t *edge;
1505 int cst, num_block_user;
1506 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1508 /* check the number of consumer scheduled so far */
1510 foreach_out_edge(irn, edge) {
1511 ir_node *user = get_edge_src_irn(edge);
1512 be_ilpsched_irn_t *cons;
1513 ilpsched_node_attr_t *ca;
1515 unsigned tm, tm_max;
1517 /* check only users within current block */
1518 if (get_nodes_block(user) != block_node->irn)
1522 cons = get_ilpsched_irn(env, user);
1523 ca = get_ilpsched_node_attr(cons);
1525 tm_max = MIN(ca->alap - 1, t);
1526 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1527 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1528 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1529 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1534 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1535 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1536 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1537 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1540 /* sum(scheduled users) */
1541 if (ARR_LEN(tmp_var_idx_m) > 0)
1542 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1543 DEL_ARR_F(tmp_var_idx_m);
1545 /* + c * a_{nt}^k */
1546 idx = node_tp_idx * livein->max_alive_steps + t;
1547 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1552 DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1553 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1557 * Create ILP pressure constraints, based on alive nodes:
1558 * - add additional costs to objective function if a node is scheduled
1559 * on a unit although all units of this type are currently occupied
1561 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1564 unsigned num_cst = 0;
1565 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1566 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1568 ilp_timer_push(t_cst);
1569 /* y_{nt}^k is set for each node and timestep and unit type */
1570 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1571 unsigned cur_idx = get_irn_idx(cur_irn);
1572 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1573 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1576 /* we ignore nodes assigned to DUMMY unit here */
1577 if (cur_na->is_dummy_node)
1581 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1582 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1586 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1588 /* check if node can be executed on this unit type */
1589 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1593 /* check all time_steps at which the current node can be scheduled */
1594 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1597 int *tmp_var_idx = NEW_ARR_F(int, 0);
1598 ilp_livein_node_t *livein;
1600 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1601 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1602 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1605 /* - accumulate all nodes alive at point t on unit type k */
1606 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1607 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1608 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1611 /* check if node can be alive here */
1612 if (t < na->asap - 1)
1615 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1617 /* current type is not suitable */
1621 a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1622 ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1624 /* do the same for livein nodes */
1625 foreach_pset(ba->livein_nodes, livein) {
1626 ir_node *irn = livein->irn;
1627 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1630 /* check if node can be alive here */
1631 if (t >= livein->max_alive_steps)
1634 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1636 /* current type is not suitable */
1640 a_idx = tp_idx * livein->max_alive_steps + t;
1641 ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1644 if (ARR_LEN(tmp_var_idx) > 0)
1645 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1646 DEL_ARR_F(tmp_var_idx);
1648 /* - num_nodes * y_{nt}^k */
1649 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1650 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1655 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1656 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1660 * Create ILP branch constraints:
1661 * Assure, alle nodes are scheduled prior to cfg op.
1663 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1665 ir_node *cur_irn, *cfop;
1666 unsigned num_cst = 0;
1667 unsigned num_non_branches = 0;
1668 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1669 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_branch", "create branch constraints");
1671 ilp_timer_push(t_cst);
1673 /* determine number of non-branch nodes and the one and only branch node */
1674 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1675 switch (get_irn_opcode(cur_irn)) {
1685 if (is_cfop(cur_irn)) {
1686 assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1697 be_ilpsched_irn_t *cf_node = get_ilpsched_irn(env, cfop);
1698 ilpsched_node_attr_t *cf_na = get_ilpsched_node_attr(cf_node);
1701 /* for each time step */
1702 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1703 int *non_branch_vars, *branch_vars;
1706 snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1707 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1708 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1711 /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1712 non_branch_vars = NEW_ARR_F(int, 0);
1713 branch_vars = NEW_ARR_F(int, 0);
1714 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1715 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
1716 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1719 if (cur_irn == cfop) {
1720 /* for all unit types available for this node */
1721 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1722 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1723 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1727 /* sum up all possible schedule points for this node upto current timestep */
1728 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1730 unsigned tmax = MIN(t, na->alap - 1);
1732 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1733 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1734 ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1741 if (ARR_LEN(non_branch_vars) > 0)
1742 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1743 if (ARR_LEN(branch_vars) > 0)
1744 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1746 DEL_ARR_F(branch_vars);
1747 DEL_ARR_F(non_branch_vars);
1751 DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1752 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1756 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1759 unsigned num_cst = 0;
1760 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1761 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1763 ilp_timer_push(t_cst);
1764 /* check all nodes */
1765 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1766 be_ilpsched_irn_t *node;
1767 ilpsched_node_attr_t *na;
1771 /* only mode_T nodes can have Projs and Keeps assigned */
1772 if (get_irn_mode(irn) != mode_T)
1775 node = get_ilpsched_irn(env, irn);
1776 na = get_ilpsched_node_attr(node);
1778 /* check if has some Projs and Keeps assigned */
1779 if (! na->projkeeps)
1782 /* we can run only once over the queue, so preserve the nodes */
1783 pk = NEW_ARR_F(ir_node *, 0);
1784 while (! waitq_empty(na->projkeeps))
1785 ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1786 del_waitq(na->projkeeps);
1787 na->projkeeps = NULL;
1789 /* for all time steps at which this node can be scheduled */
1790 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1792 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1794 /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1795 snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1796 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1797 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1800 /* sum up scheduling variables for this time step */
1801 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1802 int idx = ILPVAR_IDX(na, tp_idx, t);
1803 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1806 if (ARR_LEN(tmp_var_idx_n) > 0)
1807 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1808 DEL_ARR_F(tmp_var_idx_n);
1810 /* subtract all Proj and Keep variables for this step */
1811 for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1812 be_ilpsched_irn_t *pk_node = get_ilpsched_irn(env, pk[i]);
1813 ilpsched_node_attr_t *pk_na = get_ilpsched_node_attr(pk_node);
1816 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1817 int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1818 lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1824 DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1825 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1829 /***************************************************
1831 * |_ _| | | __ \ (_)
1832 * | | | | | |__) | _ __ ___ __ _ _ _ __
1833 * | | | | | ___/ | '_ ` _ \ / _` | | '_ \
1834 * _| |_| |____| | | | | | | | (_| | | | | |
1835 * |_____|______|_| |_| |_| |_|\__,_|_|_| |_|
1837 ***************************************************/
1840 * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1842 static void create_ilp(ir_node *block, void *walk_env) {
1843 be_ilpsched_env_t *env = walk_env;
1844 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
1845 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1846 FILE *logfile = NULL;
1849 struct obstack var_obst;
1852 DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1853 DBG((env->dbg, 255, " ILP Scheduling for %+F\n", block));
1854 DBG((env->dbg, 255, "=========================================\n\n"));
1856 DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1857 block, ba->n_interesting_nodes, ba->max_steps));
1859 /* notify backend and get block environment */
1860 env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1862 /* if we have less than two interesting nodes, there is no need to create the ILP */
1863 if (ba->n_interesting_nodes > 1) {
1864 double fact_var = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1865 double fact_cst = ba->n_interesting_nodes > 25 ? 3 : 4.5;
1866 int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes;
1867 int estimated_n_var = (int)((double)base_num * fact_var);
1868 int estimated_n_cst = (int)((double)base_num * fact_cst);
1870 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1871 estimated_n_var, estimated_n_cst));
1872 (void) estimated_n_var;
1874 /* set up the LPP object */
1875 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1877 lpp = new_lpp_userdef(
1880 estimated_n_cst, /* num vars */
1881 estimated_n_cst + 1, /* num cst */
1882 1.3); /* grow factor */
1883 obstack_init(&var_obst);
1885 /* create ILP variables */
1886 create_variables(env, lpp, block_node, &var_obst);
1888 /* create ILP constraints */
1889 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1890 create_assignment_and_precedence_constraints(env, lpp, block_node);
1891 create_ressource_constraints(env, lpp, block_node);
1892 create_bundle_constraints(env, lpp, block_node);
1893 create_branch_constraint(env, lpp, block_node);
1894 //create_proj_keep_constraints(env, lpp, block_node);
1896 if (env->opts->regpress) {
1897 create_alive_nodes_constraint(env, lpp, block_node);
1898 create_alive_livein_nodes_constraint(env, lpp, block_node);
1899 create_pressure_alive_constraint(env, lpp, block_node);
1902 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1904 /* debug stuff, dump lpp when debugging is on */
1906 if (firm_dbg_get_mask(env->dbg) > 1) {
1910 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1911 f = fopen(buf, "w");
1912 lpp_dump_plain(lpp, f);
1914 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1919 /* set solve time limit */
1920 lpp_set_time_limit(lpp, env->opts->time_limit);
1922 /* set logfile if requested */
1923 if (strlen(env->opts->log_file) > 0) {
1924 if (strcasecmp(env->opts->log_file, "stdout") == 0)
1925 lpp_set_log(lpp, stdout);
1926 else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1927 lpp_set_log(lpp, stderr);
1929 logfile = fopen(env->opts->log_file, "w");
1931 fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1933 lpp_set_log(lpp, logfile);
1938 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1943 /* check for valid solution */
1944 if (! lpp_is_sol_valid(lpp)) {
1949 if (firm_dbg_get_mask(env->dbg) >= 2) {
1950 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1951 f = fopen(buf, "w");
1952 lpp_dump_plain(lpp, f);
1954 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1956 dump_ir_block_graph(env->irg, "-infeasible");
1960 ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1964 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1965 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1966 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1967 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));
1968 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1969 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1970 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1971 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1973 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1976 /* apply solution */
1977 be_stat_ev("nodes", ba->block_last_idx);
1978 be_stat_ev("vars", lpp ? lpp->var_next : 0);
1979 be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1981 be_stat_ev("time", -1);
1982 be_stat_ev_dbl("opt", 0.0);
1983 list_sched_single_block(env->birg, block, env->be_opts);
1987 double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1988 be_stat_ev_dbl("time", lpp->sol_time);
1989 be_stat_ev_dbl("opt", opt);
1992 be_stat_ev_dbl("time", 0.0);
1993 be_stat_ev_dbl("opt", 100.0);
1995 apply_solution(env, lpp, block);
2001 /* notify backend */
2002 be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2006 * Perform ILP scheduling on the given irg.
2008 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2009 be_ilpsched_env_t env;
2010 const char *name = "be ilp scheduling";
2011 ir_graph *irg = be_get_birg_irg(birg);
2012 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
2013 const ilp_sched_selector_t *sel = arch_env->impl->get_ilp_sched_selector(arch_env);
2015 FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2017 stat_ev_ctx_push("ilpsched");
2019 // firm_dbg_set_mask(env.dbg, 1);
2021 env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2024 env.height = heights_new(irg);
2025 env.main_env = birg->main_env;
2026 env.arch_env = arch_env;
2027 env.cpu = arch_env_get_machine(arch_env);
2028 env.opts = &ilp_opts;
2030 env.be_opts = be_opts;
2031 phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2033 /* assign a unique per block number to all interesting nodes */
2034 irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2037 The block indices are completely build after the walk,
2038 now we can allocate the bitsets (size depends on block indices)
2041 phase_reinit_irn_data(&env.ph);
2043 /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2044 irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2046 /* Calculate ALAP of all irns */
2047 irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2049 /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2050 irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2052 /* perform ILP scheduling */
2053 irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2056 if (firm_dbg_get_mask(env.dbg)) {
2058 phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2060 fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2064 /* free data allocated dynamically */
2065 irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2067 /* free all allocated object */
2068 phase_free(&env.ph);
2069 heights_free(env.height);
2071 /* notify backend */
2072 be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2074 stat_ev_ctx_pop("ilpsched");
2078 * Register ILP scheduler options.
2080 void be_init_ilpsched(void)
2082 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2083 lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2085 lc_opt_add_table(sched_grp, ilpsched_option_table);
2088 #else /* WITH_ILP */
2090 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2093 #endif /* WITH_ILP */