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
36 #include "firm_config.h"
49 #include "irphase_t.h"
62 #include <lpp/lpp_net.h>
65 #include "lc_opts_enum.h"
69 #include "besched_t.h"
70 #include "beilpsched.h"
75 typedef struct _ilpsched_options_t {
81 typedef struct _unit_type_info_t {
83 const be_execution_unit_type_t *tp;
87 * holding the ILP variables of the different types
89 typedef struct _ilp_var_types_t {
90 int *x; /* x_{nt}^k variables */
91 int *a; /* a_{nt}^k variables */
92 int *y; /* y_{nt}^k variables */
96 * Holds alive variables for a node live-in to a block.
98 typedef struct _ilp_livein_node_t {
100 unsigned max_alive_steps;
104 /* attributes for a node */
105 typedef struct _ilpsched_node_attr_t {
106 unsigned asap; /**< The ASAP scheduling control step */
107 unsigned alap; /**< The ALAP scheduling control step */
108 unsigned latency; /**< Latency of this node (needed for sorting) */
109 unsigned sched_point; /**< the step in which the node is finally scheduled */
110 unsigned visit_idx; /**< Index of the node having visited this node last */
111 unsigned consumer_idx; /**< Index of the node having counted this node as consumer last */
112 unsigned n_consumer; /**< Number of consumers */
113 ir_node **block_consumer; /**< List of consumer being in the same block */
114 waitq *projkeeps; /**< A List of Projs and Keeps belonging to this node */
115 unsigned block_idx : 30; /**< A unique per block index */
116 unsigned alap_changed : 1; /**< the current ALAP has changed, revisit preds */
117 unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */
118 bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors
119 for ASAP, successors for ALAP */
120 unsigned n_unit_types; /**< number of allowed execution unit types */
121 unit_type_info_t *type_info; /**< list of allowed execution unit types */
122 ilp_var_types_t ilp_vars; /**< the different ILP variables */
123 } ilpsched_node_attr_t;
125 /* attributes for a block */
126 typedef struct _ilpsched_block_attr_t {
127 unsigned block_last_idx; /**< The highest node index in block so far */
128 unsigned n_interesting_nodes; /**< The number of nodes interesting for scheduling */
129 unsigned max_steps; /**< Upper bound for block execution */
130 plist_t *root_nodes; /**< A list of nodes having no user in current block */
131 ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */
132 pset *livein_nodes; /**< A set of nodes which are live-in to this block */
133 } ilpsched_block_attr_t;
135 typedef union _ilpsched_attr_ {
136 ilpsched_node_attr_t node_attr;
137 ilpsched_block_attr_t block_attr;
140 /* A irn for the phase and it's attributes (either node or block) */
143 ilpsched_attr_t attr;
146 /* The ILP scheduling environment */
148 ir_phase ph; /**< The phase */
149 ir_graph *irg; /**< The current irg */
150 heights_t *height; /**< The heights object of the irg */
151 void *irg_env; /**< An environment for the irg scheduling, provided by the backend */
152 void *block_env; /**< An environment for scheduling a block, provided by the backend */
153 const arch_env_t *arch_env;
154 const be_main_env_t *main_env;
155 const be_machine_t *cpu; /**< the current abstract machine */
156 ilpsched_options_t *opts; /**< the ilp options for current irg */
157 const be_irg_t *birg; /**< The birg object */
158 be_options_t *be_opts; /**< backend options */
159 const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */
160 DEBUG_ONLY(firm_dbg_module_t *dbg);
163 /* convenience macros to handle phase irn data */
164 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
165 #define is_ilpsched_block(node) (is_Block((node)->irn))
166 #define get_ilpsched_block_attr(block) (&(block)->attr.block_attr)
167 #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr)
169 /* check if node is considered for ILP scheduling */
170 #define consider_for_sched(env, irn) \
171 (! (is_Block(irn) || \
172 is_normal_Proj(env, irn) || \
179 /* gives the valid scheduling time step interval for a node */
180 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
182 /* gives the valid interval where a node can die */
183 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
185 /* gives the corresponding ILP variable for given node, unit and time step */
186 #define ILPVAR_IDX(na, unit, control_step) \
187 ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
189 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
190 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
191 ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
193 /* check if a double value is within an epsilon environment of 0 */
194 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
196 #define ilp_timer_push(t) ir_timer_push((t))
197 #define ilp_timer_pop() ir_timer_pop()
198 #define ilp_timer_elapsed_usec(t) ir_timer_elapsed_usec((t))
200 /* option variable */
201 static ilpsched_options_t ilp_opts = {
202 1, /* default is with register pressure constraints */
203 300, /* 300 sec per block time limit */
208 static const lc_opt_table_entry_t ilpsched_option_table[] = {
209 LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress),
210 LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
211 LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
216 We need this global variable as we compare nodes dependent on heights,
217 but we cannot pass any information to the qsort compare function.
219 static heights_t *glob_heights;
222 * Check if irn is a Proj, which has no execution units assigned.
223 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
225 static INLINE int is_normal_Proj(const arch_env_t *env, const ir_node *irn) {
226 return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
230 * Skips normal Projs.
231 * @return predecessor if irn is a normal Proj, otherwise irn.
233 static INLINE ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn) {
234 if (is_normal_Proj(env, irn))
235 return get_Proj_pred(irn);
239 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
240 unsigned lat = be_ilp_sched_latency(sel, irn, env);
241 if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
246 static int cmp_live_in_nodes(const void *a, const void *b) {
247 const ilp_livein_node_t *n1 = a;
248 const ilp_livein_node_t *n2 = b;
250 return n1->irn != n2->irn;
254 * Compare scheduling time steps of two be_ilpsched_irn's.
256 static int cmp_ilpsched_irn(const void *a, const void *b) {
257 be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a;
258 be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b;
259 ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
260 ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
262 if (n1_a->sched_point == n2_a->sched_point) {
263 const ir_node *irn_a = n1->irn;
264 const ir_node *irn_b = n2->irn;
266 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
268 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
272 Ok, timestep is equal and the nodes are parallel,
273 so check latency and schedule high latency first.
275 return QSORT_CMP(n2_a->latency, n1_a->latency);
278 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
282 * In case there is no phase information for irn, initialize it.
284 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old) {
285 be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
288 /* if we have already some data: check for reinitialization */
290 if (! is_Block(irn)) {
291 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
293 if (! na->transitive_block_nodes) {
294 ir_node *block = get_nodes_block(irn);
295 be_ilpsched_irn_t *block_node = phase_get_or_set_irn_data(ph, block);
296 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
298 /* we are called after the block indices have been build: create bitset */
299 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
302 /* we are called from reinit block data: clear the bitset */
303 bitset_clear_all(na->transitive_block_nodes);
305 na->alap_changed = 1;
313 /* set ilpsched irn attributes (either block or irn) */
315 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
317 ba->n_interesting_nodes = 0;
318 ba->block_last_idx = 0;
319 ba->root_nodes = plist_new();
320 ba->head_ilp_nodes = NULL;
321 ba->livein_nodes = new_pset(cmp_live_in_nodes, 16);
325 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
326 memset(na, 0, sizeof(*na));
333 * Assign a per block unique number to each node.
335 static void build_block_idx(ir_node *irn, void *walk_env) {
336 be_ilpsched_env_t *env = walk_env;
337 be_ilpsched_irn_t *node, *block_node;
338 ilpsched_node_attr_t *na;
339 ilpsched_block_attr_t *ba;
341 set_irn_link(irn, NULL);
342 if (! consider_for_sched(env->arch_env, irn))
345 node = get_ilpsched_irn(env, irn);
346 na = get_ilpsched_node_attr(node);
347 block_node = get_ilpsched_irn(env, get_nodes_block(irn));
348 ba = get_ilpsched_block_attr(block_node);
350 na->block_idx = ba->block_last_idx++;
353 /********************************************************
356 * __ _ ___ __ _ _ __ / / __ _| | __ _ _ __
357 * / _` / __|/ _` | '_ \ / / / _` | |/ _` | '_ \
358 * | (_| \__ \ (_| | |_) | / / | (_| | | (_| | |_) |
359 * \__,_|___/\__,_| .__/ /_/ \__,_|_|\__,_| .__/
362 ********************************************************/
365 * Add all nodes having no user in current block to last_nodes list.
367 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
369 const ir_edge_t *edge;
370 be_ilpsched_irn_t *block_node, *node;
371 ilpsched_block_attr_t *ba;
372 ilpsched_node_attr_t *na;
374 be_ilpsched_env_t *env = walk_env;
375 int has_block_user = 0;
376 unsigned n_consumer = 0;
377 ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
381 if (! consider_for_sched(env->arch_env, irn))
384 block = get_nodes_block(irn);
385 idx = get_irn_idx(irn);
386 consumer = NEW_ARR_F(ir_node *, 0);
388 DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
390 /* check data and dependency out edges */
391 for (i = 0; i < 2 && ! has_block_user; ++i) {
392 foreach_out_edge_kind(irn, edge, ekind[i]) {
393 ir_node *user = get_edge_src_irn(edge);
395 if (is_normal_Proj(env->arch_env, user)) {
396 const ir_edge_t *user_edge;
398 if (get_irn_mode(user) == mode_X)
401 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
402 for (j = 0; j < 2; ++j) {
403 foreach_out_edge_kind(user, user_edge, ekind[j]) {
404 ir_node *real_user = get_edge_src_irn(user_edge);
406 if (! is_Phi(real_user) && ! is_Block(real_user)) {
407 be_ilpsched_irn_t *node = get_ilpsched_irn(env, real_user);
408 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
410 /* skip already visited nodes */
411 if (ua->consumer_idx == idx)
414 /* check if node has user in this block and collect the user if it's a data user */
415 if (get_nodes_block(real_user) == block) {
416 if (i == 0 && j == 0)
417 ARR_APP1(ir_node *, consumer, real_user);
421 /* only count data consumer */
425 /* mark user as visited by this node */
426 ua->consumer_idx = idx;
431 else if (is_Block(user)) {
434 else if (! is_Phi(user)) {
435 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
436 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
438 /* skip already visited nodes */
439 if (ua->consumer_idx == idx)
442 /* check if node has user in this block and collect the user if it's a data user */
443 if (get_nodes_block(user) == block) {
445 ARR_APP1(ir_node *, consumer, user);
449 /* only count data consumer */
453 /* mark user visited by this node */
454 ua->consumer_idx = idx;
456 else if (get_nodes_block(user) != block) {
462 block_node = get_ilpsched_irn(env, block);
463 ba = get_ilpsched_block_attr(block_node);
465 ba->n_interesting_nodes++;
467 /* current irn has no user inside this block, add to queue */
468 if (! has_block_user) {
469 DB((env->dbg, LEVEL_3, "root node\n"));
470 plist_insert_back(ba->root_nodes, irn);
473 DB((env->dbg, LEVEL_3, "normal node\n"));
476 /* record number of all consumer and the consumer within the same block */
477 node = get_ilpsched_irn(env, irn);
478 na = get_ilpsched_node_attr(node);
479 na->n_consumer = n_consumer;
480 na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
481 memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
486 * Calculate the ASAP scheduling step for current irn.
488 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
489 be_ilpsched_env_t *env = walk_env;
492 be_ilpsched_irn_t *node, *block_node;
493 ilpsched_node_attr_t *na;
494 ilpsched_block_attr_t *ba;
496 /* These nodes are handled separate */
497 if (! consider_for_sched(env->arch_env, irn))
500 DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
502 block = get_nodes_block(irn);
503 node = get_ilpsched_irn(env, irn);
504 na = get_ilpsched_node_attr(node);
507 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
508 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
510 /* check for greatest distance to top */
511 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
512 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
513 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
516 lat = fixed_latency(env->sel, pred, env->block_env);
518 na->asap = MAX(na->asap, pna->asap + lat);
522 /* add node to ILP node list and update max_steps */
523 block_node = get_ilpsched_irn(env, block);
524 ba = get_ilpsched_block_attr(block_node);
526 set_irn_link(irn, ba->head_ilp_nodes);
527 ba->head_ilp_nodes = irn;
528 ba->max_steps += fixed_latency(env->sel, irn, env->block_env);
530 DB((env->dbg, LEVEL_2, "%u\n", na->asap));
534 * Calculate the ALAP scheduling step of all irns in current block.
535 * Depends on max_steps being calculated.
537 static void calculate_block_alap(ir_node *block, void *walk_env) {
538 be_ilpsched_env_t *env = walk_env;
539 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
540 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
541 waitq *cur_queue = new_waitq();
544 assert(is_Block(block));
546 DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
547 block, ba->n_interesting_nodes, ba->max_steps));
549 /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
550 //phase_reinit_block_irn_data(&env->ph, block);
552 /* init start queue */
553 foreach_plist(ba->root_nodes, el) {
554 waitq_put(cur_queue, plist_element_get_value(el));
557 /* repeat until all nodes are processed */
558 while (! waitq_empty(cur_queue)) {
559 waitq *next_queue = new_waitq();
561 /* process all nodes in current step */
562 while (! waitq_empty(cur_queue)) {
563 ir_node *cur_irn = waitq_get(cur_queue);
564 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
565 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
568 /* cur_node has no alap set -> it's a root node, set to max alap */
570 na->alap = ba->max_steps;
571 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
575 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
579 /* set the alap's of all predecessors */
580 for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
581 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
583 /* check for greatest distance to bottom */
584 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
585 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
586 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
589 /* mark the predecessor as visited by current irn */
590 if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
592 pna->visit_idx = get_irn_idx(cur_irn);
594 lat = fixed_latency(env->sel, pred, env->block_env);
596 /* set ALAP of current pred */
597 if (pna->alap == 0) {
598 /* current ALAP is 0: node has not yet been visited */
599 pna->alap_changed = 1;
600 pna->alap = na->alap - lat;
602 else if (pna->alap > na->alap - lat) {
603 /* we found a longer path to root node: change ALAP */
604 pna->alap = na->alap - lat;
605 pna->alap_changed = 1;
608 /* current ALAP is best found so far: keep it */
609 pna->alap_changed = 0;
612 DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
614 /* enqueue node for next iteration */
615 if (get_irn_ins_or_deps(pred) > 0)
616 waitq_put(next_queue, pred);
621 /* prepare for next iteration */
622 del_waitq(cur_queue);
623 cur_queue = next_queue;
628 * Free list of root nodes and the set of live-in nodes.
630 static void clear_unwanted_data(ir_node *block, void *walk_env) {
631 be_ilpsched_env_t *env = walk_env;
632 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
633 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
635 plist_free(ba->root_nodes);
636 ba->root_nodes = NULL;
637 del_pset(ba->livein_nodes);
638 ba->livein_nodes = NULL;
642 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
643 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
645 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
646 be_ilpsched_env_t *env = walk_env;
648 be_ilpsched_irn_t *node, *pred_node;
649 ilpsched_node_attr_t *na, *pna;
651 if (! consider_for_sched(env->arch_env, irn))
654 if (! is_Proj(irn) && ! be_is_Keep(irn))
657 /* go to the ancestor */
659 pred = get_irn_n(irn, 0);
660 pred = skip_Proj(pred);
662 node = get_ilpsched_irn(env, irn);
663 pred_node = get_ilpsched_irn(env, pred);
664 na = get_ilpsched_node_attr(node);
665 pna = get_ilpsched_node_attr(pred_node);
667 na->asap = pna->asap;
668 na->alap = pna->alap;
670 /* record all Projs and Keeps for this node */
671 if (! pna->projkeeps)
672 pna->projkeeps = new_waitq();
673 waitq_put(pna->projkeeps, irn);
675 DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
678 /*******************************************
681 * ___ ___| |__ ___ __| |_ _| | ___
682 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
683 * \__ \ (__| | | | __/ (_| | |_| | | __/
684 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
686 *******************************************/
688 static INLINE void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn) {
689 const ir_edge_t *edge;
692 foreach_out_edge(irn, edge) {
693 ir_node *user = get_edge_src_irn(edge);
695 if (be_is_Keep(user)) {
696 assert(get_nodes_block(user) == block && "Keep must not be in different block.");
697 waitq_put(keeps, user);
703 * Inserts @p irn before @p before into schedule and notifies backend.
705 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
706 const ir_node *before, const ir_node *irn, unsigned cycle)
708 be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
709 sched_add_before(before, irn);
713 * Adds a node, it's Projs (in case of mode_T nodes) and
714 * it's Keeps to schedule.
716 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle) {
717 const ir_edge_t *edge;
718 waitq *keeps = new_waitq();
720 /* mode_M nodes are not scheduled */
721 if (get_irn_mode(irn) == mode_M)
724 if (! sched_is_scheduled(irn))
725 notified_sched_add_before(env, block, irn, cycle);
728 if (get_irn_mode(irn) == mode_T) {
729 foreach_out_edge(irn, edge) {
730 ir_node *user = get_edge_src_irn(edge);
732 if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
733 get_irn_n_edges(user) > 0)
735 notified_sched_add_before(env, block, user, cycle);
738 check_for_keeps(keeps, block, user);
742 check_for_keeps(keeps, block, irn);
746 while (! waitq_empty(keeps)) {
747 ir_node *keep = waitq_get(keeps);
748 if (! sched_is_scheduled(keep))
749 notified_sched_add_before(env, block, keep, cycle);
756 * Schedule all nodes in the given block, according to the ILP solution.
758 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
759 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
760 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
761 sched_info_t *info = get_irn_sched_info(block);
762 be_ilpsched_irn_t **sched_nodes;
765 const ir_edge_t *edge;
767 /* init block schedule list */
768 INIT_LIST_HEAD(&info->list);
771 /* collect nodes and their scheduling time step */
772 sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
773 if (ba->n_interesting_nodes == 0) {
776 else if (ba->n_interesting_nodes == 1) {
777 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
779 /* add the single node */
780 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
783 /* check all nodes for their positive solution */
784 foreach_linked_irns(ba->head_ilp_nodes, irn) {
785 be_ilpsched_irn_t *node;
786 ilpsched_node_attr_t *na;
790 node = get_ilpsched_irn(env, irn);
791 na = get_ilpsched_node_attr(node);
795 if (! na->is_dummy_node) {
796 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
797 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
798 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
800 if (! LPP_VALUE_IS_0(cost)) {
801 DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
809 /* go over all variables of a node until the non-zero one is found */
810 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
811 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
812 double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
814 /* check, if variable is set to one (it's not zero then :) */
815 if (! LPP_VALUE_IS_0(val)) {
817 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
818 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
819 irn, t, na->type_info[tp_idx].tp->name));
826 glob_heights = env->height;
827 /* sort nodes ascending by scheduling time step */
828 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
831 /* make all Phis ready and remember the single cf op */
833 foreach_out_edge(block, edge) {
834 irn = get_edge_src_irn(edge);
836 switch (get_irn_opcode(irn)) {
838 add_to_sched(env, block, irn, 0);
848 assert(cfop == NULL && "Highlander - there can be only one");
855 /* add all nodes from list */
856 for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
857 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
858 if (sched_nodes[i]->irn != cfop)
859 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
862 /* schedule control flow node if not already done */
863 if (cfop && ! sched_is_scheduled(cfop))
864 add_to_sched(env, block, cfop, 0);
866 DEL_ARR_F(sched_nodes);
869 /***************************************************************
870 * _____ _ _____ _____ _ _
871 * |_ _| | | __ \ / ____| | | (_)
872 * | | | | | |__) | | (___ ___ ___| |_ _ ___ _ __
873 * | | | | | ___/ \___ \ / _ \/ __| __| |/ _ \| '_ \
874 * _| |_| |____| | ____) | __/ (__| |_| | (_) | | | |
875 * |_____|______|_| |_____/ \___|\___|\__|_|\___/|_| |_|
877 ***************************************************************/
880 * Check if node can be executed on given unit type.
882 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
884 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
886 for (i = na->n_unit_types - 1; i >= 0; --i) {
887 if (na->type_info[i].tp == tp)
894 /************************************************
897 * __ ____ _ _ __ _ __ _| |__ | | ___ ___
898 * \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
899 * \ V / (_| | | | | (_| | |_) | | __/\__ \
900 * \_/ \__,_|_| |_|\__,_|_.__/|_|\___||___/
902 ************************************************/
904 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
905 const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
906 unsigned n_unit_types = 0;
907 be_ilpsched_irn_t *node;
908 ilpsched_node_attr_t *na;
909 unsigned unit_idx, tp_idx;
911 /* count number of available unit types for this node */
912 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
915 node = get_ilpsched_irn(env, irn);
916 na = get_ilpsched_node_attr(node);
918 if (! na->type_info) {
919 na->n_unit_types = n_unit_types;
920 na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
922 /* fill the type info array */
923 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
924 for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
925 /* beware: we also count number of available units here */
926 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
927 na->is_dummy_node = 1;
930 na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp;
931 na->type_info[tp_idx].n_units = unit_idx;
939 * Returns the largest alap time of a user of @p irn.
940 * The user must be in block @p block.
942 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block) {
943 const ir_edge_t *edge;
944 unsigned max_alap = 0;
946 foreach_out_edge(irn, edge) {
947 ir_node *user = get_edge_src_irn(edge);
949 if (get_nodes_block(user) == block) {
950 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
951 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
953 max_alap = MAX(max_alap, na->alap);
957 assert(max_alap > 0);
962 * Create the following variables:
963 * - x_{nt}^k binary weigthed with: t
964 * node n is scheduled at time step t to unit type k
965 * ==>> These variables represent the schedule
967 * - a_{nt}^k binary weighted with num_nodes
968 * node n is alive at time step t on unit type k
970 * - y_{nt}^k continuous weighted with: num_nodes^2
971 * register pressure over limit for unit type k
972 * ==>> These variables represent the register pressure
975 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
978 unsigned num_block_var, num_nodes;
979 ilp_livein_node_t *livein;
980 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
981 unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
982 ir_timer_t *t_var = ir_timer_register("beilpsched_var", "create ilp variables");
984 ilp_timer_push(t_var);
985 num_block_var = num_nodes = 0;
986 foreach_linked_irns(ba->head_ilp_nodes, irn) {
987 be_ilpsched_irn_t *node;
988 ilpsched_node_attr_t *na;
989 unsigned n_unit_types, tp_idx, n_var, cur_unit;
990 unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad;
993 node = get_ilpsched_irn(env, irn);
994 na = get_ilpsched_node_attr(node);
995 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
997 /* allocate space for ilp variables */
998 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
999 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
1001 /* we need these variables only for "real" nodes */
1002 if (! na->is_dummy_node) {
1003 na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1004 memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1006 num_ad = ba->max_steps - na->asap + 1;
1007 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1008 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1011 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1012 irn, na->asap, na->alap, na->n_unit_types));
1014 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1015 /* create variables */
1016 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1019 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1020 /* x_{nt}^k variables */
1021 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1022 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1023 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1024 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1025 /* variable counter */
1029 if (! na->is_dummy_node) {
1030 /* y_{nt}^k variables */
1031 snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1032 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1033 na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1034 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1036 /* variable counter */
1042 /* a node can die at any step t: asap(n) <= t <= U */
1043 if (! na->is_dummy_node) {
1044 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1046 /* a_{nt}^k variables */
1047 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1048 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1049 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1050 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1052 /* variable counter */
1058 /* collect live-in nodes */
1059 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1060 ir_node *pred = get_irn_n(irn, i);
1062 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1063 be_ilpsched_set_type_info(env, pred, var_obst);
1064 if (! na->is_dummy_node) {
1065 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1068 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1074 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1078 /* create alive variables a_{nt}^k for live-ins */
1079 foreach_pset(ba->livein_nodes, livein) {
1080 be_ilpsched_irn_t *node;
1081 ilpsched_node_attr_t *na;
1082 unsigned tp_idx, var_idx;
1086 node = get_ilpsched_irn(env, irn);
1087 na = get_ilpsched_node_attr(node);
1089 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1091 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1094 /* create variables */
1095 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1098 for (t = 0; t < livein->max_alive_steps; ++t) {
1099 /* a_{nt}^k variables */
1100 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1101 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1102 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1103 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1110 DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1111 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1114 /*******************************************************
1117 * ___ ___ _ __ ___| |_ _ __ __ _ _ _ __ | |_ ___
1118 * / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1119 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1120 * \___\___/|_| |_|___/\__|_| \__,_|_|_| |_|\__|___/
1122 *******************************************************/
1125 * Collect all operands and nodes @p irn depends on.
1126 * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1128 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps) {
1131 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1132 ir_node *p = get_irn_in_or_dep(irn, i);
1135 const ir_edge_t *edge;
1137 p = get_Proj_pred(p);
1138 foreach_out_edge(p, edge) {
1139 ir_node *src = get_edge_src_irn(edge);
1140 ir_nodeset_insert(deps, src);
1144 ir_nodeset_insert(deps, p);
1150 * Create following ILP constraints:
1151 * - the assignment constraints:
1152 * assure each node is executed once by exactly one (allowed) execution unit
1153 * - the dead node assignment constraints:
1154 * assure a node can only die at most once
1155 * - the precedence constraints:
1156 * assure that no data dependencies are violated
1158 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1159 unsigned num_cst_assign, num_cst_prec, num_cst_dead;
1162 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1163 bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx);
1164 ir_timer_t *t_cst_assign = ir_timer_register("beilpsched_cst_assign", "create assignment constraints");
1165 ir_timer_t *t_cst_prec = ir_timer_register("beilpsched_cst_prec", "create precedence constraints");
1167 num_cst_assign = num_cst_prec = num_cst_dead = 0;
1168 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1171 be_ilpsched_irn_t *node;
1172 ilpsched_node_attr_t *na;
1175 ir_nodeset_iterator_t iter;
1177 ir_nodeset_init(&deps);
1179 node = get_ilpsched_irn(env, irn);
1180 na = get_ilpsched_node_attr(node);
1183 /* the assignment constraint */
1184 ilp_timer_push(t_cst_assign);
1185 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1186 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1187 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1190 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1193 /* We have separate constraints for Projs and Keeps */
1194 // ILP becomes infeasible ?!?
1195 // if (is_Proj(irn) || be_is_Keep(irn))
1198 /* the precedence constraints */
1199 ilp_timer_push(t_cst_prec);
1200 bs_block_irns = bitset_clear_all(bs_block_irns);
1202 sta_collect_in_deps(irn, &deps);
1203 foreach_ir_nodeset(&deps, pred, iter) {
1204 unsigned t_low, t_high, t;
1205 be_ilpsched_irn_t *pred_node;
1206 ilpsched_node_attr_t *pna;
1209 pred = skip_normal_Proj(env->arch_env, pred);
1210 if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1213 pred_node = get_ilpsched_irn(env, pred);
1214 pna = get_ilpsched_node_attr(pred_node);
1216 assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1218 if (! bitset_is_set(bs_block_irns, pna->block_idx))
1219 bitset_set(bs_block_irns, pna->block_idx);
1223 /* irn = n, pred = m */
1224 delay = fixed_latency(env->sel, pred, env->block_env);
1225 t_low = MAX(na->asap, pna->asap + delay - 1);
1226 t_high = MIN(na->alap, pna->alap + delay - 1);
1227 for (t = t_low - 1; t <= t_high - 1; ++t) {
1229 int *tmp_var_idx = NEW_ARR_F(int, 0);
1231 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1232 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1233 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1236 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1237 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1238 /* node variables have smaller index than pred variables */
1239 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1240 for (tn = na->asap - 1; tn <= t; ++tn) {
1241 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1242 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1246 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1247 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1248 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1249 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1254 /* pred variables have smaller index than node variables */
1255 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1256 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1257 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1258 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1262 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1263 for (tn = na->asap - 1; tn <= t; ++tn) {
1264 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1265 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1270 if (ARR_LEN(tmp_var_idx) > 0)
1271 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1273 DEL_ARR_F(tmp_var_idx);
1276 ir_nodeset_destroy(&deps);
1279 DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1280 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1281 DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1282 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1286 * Create ILP resource constraints:
1287 * - assure that for each time step not more instructions are scheduled
1288 * to the same unit types as units of this type are available
1290 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1293 unsigned num_cst_resrc = 0;
1294 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1295 ir_timer_t *t_cst_rsrc = ir_timer_register("beilpsched_cst_rsrc", "create resource constraints");
1297 ilp_timer_push(t_cst_rsrc);
1298 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1300 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1302 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1304 /* check each time step */
1305 for (t = 0; t < ba->max_steps; ++t) {
1308 int *tmp_var_idx = NEW_ARR_F(int, 0);
1310 snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1311 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1312 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1315 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1316 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1317 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1320 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1322 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1323 int cur_var = ILPVAR_IDX(na, tp_idx, t);
1324 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1328 /* set constraints if we have some */
1329 if (ARR_LEN(tmp_var_idx) > 0)
1330 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1332 DEL_ARR_F(tmp_var_idx);
1336 DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1337 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1341 * Create ILP bundle constraints:
1342 * - assure, at most bundle_size * bundles_per_cycle instructions
1343 * can be started at a certain point.
1345 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1348 unsigned num_cst_bundle = 0;
1349 unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1350 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1351 ir_timer_t *t_cst_bundle = ir_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1353 ilp_timer_push(t_cst_bundle);
1354 for (t = 0; t < ba->max_steps; ++t) {
1357 int *tmp_var_idx = NEW_ARR_F(int, 0);
1359 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1360 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1361 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1364 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1365 be_ilpsched_irn_t *node;
1366 ilpsched_node_attr_t *na;
1369 /* Projs and Keeps do not contribute to bundle size */
1370 if (is_Proj(irn) || be_is_Keep(irn))
1373 node = get_ilpsched_irn(env, irn);
1374 na = get_ilpsched_node_attr(node);
1376 /* nodes assigned to DUMMY unit do not contribute to bundle size */
1377 if (na->is_dummy_node)
1380 if (t >= na->asap - 1 && t <= na->alap - 1) {
1381 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1382 int idx = ILPVAR_IDX(na, tp_idx, t);
1383 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1388 if (ARR_LEN(tmp_var_idx) > 0)
1389 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1391 DEL_ARR_F(tmp_var_idx);
1394 DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1395 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1399 * Create ILP alive nodes constraints:
1400 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1402 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1405 unsigned num_cst = 0;
1406 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1407 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1409 ilp_timer_push(t_cst);
1411 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1412 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1413 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1416 /* we ignore nodes assigned to dummy unit here */
1417 if (na->is_dummy_node)
1420 /* check check all time steps: asap(n) <= t <= U */
1421 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1424 /* for all unit types available for this node */
1425 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1426 unsigned tn, tn_max, idx;
1428 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1429 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1431 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1432 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1433 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1434 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1437 tn_max = MIN(na->alap - 1, t);
1438 /* check if the node has been scheduled so far */
1439 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1440 int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1441 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1444 if (ARR_LEN(tmp_var_idx_n) > 0)
1445 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1446 DEL_ARR_F(tmp_var_idx_n);
1448 /* subtract the number of consumer scheduled so far */
1449 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1450 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1451 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1453 unsigned tm, tm_max;
1455 tm_max = MIN(ca->alap - 1, t);
1456 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1457 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1458 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1459 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1464 if (ARR_LEN(tmp_var_idx_m) > 0)
1465 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1466 DEL_ARR_F(tmp_var_idx_m);
1469 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1470 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1476 DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1477 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1481 * Create ILP alive nodes constraints for live-in nodes:
1482 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1484 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1486 ilp_livein_node_t *livein;
1487 unsigned num_cst = 0;
1488 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1489 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1491 ilp_timer_push(t_cst);
1493 foreach_pset(ba->livein_nodes, livein) {
1494 ir_node *irn = livein->irn;
1495 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1496 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1499 /* check check all time steps: 0 <= t < max_alive_steps */
1500 for (t = 0; t < livein->max_alive_steps; ++t) {
1503 /* for all unit types available for this node */
1504 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1505 const ir_edge_t *edge;
1507 int cst, num_block_user;
1508 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1510 /* check the number of consumer scheduled so far */
1512 foreach_out_edge(irn, edge) {
1513 ir_node *user = get_edge_src_irn(edge);
1514 be_ilpsched_irn_t *cons;
1515 ilpsched_node_attr_t *ca;
1517 unsigned tm, tm_max;
1519 /* check only users within current block */
1520 if (get_nodes_block(user) != block_node->irn)
1524 cons = get_ilpsched_irn(env, user);
1525 ca = get_ilpsched_node_attr(cons);
1527 tm_max = MIN(ca->alap - 1, t);
1528 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1529 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1530 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1531 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1536 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1537 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1538 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1539 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1542 /* sum(scheduled users) */
1543 if (ARR_LEN(tmp_var_idx_m) > 0)
1544 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1545 DEL_ARR_F(tmp_var_idx_m);
1547 /* + c * a_{nt}^k */
1548 idx = node_tp_idx * livein->max_alive_steps + t;
1549 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1554 DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1555 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1559 * Create ILP pressure constraints, based on alive nodes:
1560 * - add additional costs to objective function if a node is scheduled
1561 * on a unit although all units of this type are currently occupied
1563 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1566 unsigned num_cst = 0;
1567 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1568 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1570 ilp_timer_push(t_cst);
1571 /* y_{nt}^k is set for each node and timestep and unit type */
1572 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1573 unsigned cur_idx = get_irn_idx(cur_irn);
1574 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1575 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1578 /* we ignore nodes assigned to DUMMY unit here */
1579 if (cur_na->is_dummy_node)
1583 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1584 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1588 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1590 /* check if node can be executed on this unit type */
1591 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1595 /* check all time_steps at which the current node can be scheduled */
1596 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1599 int *tmp_var_idx = NEW_ARR_F(int, 0);
1600 ilp_livein_node_t *livein;
1602 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1603 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1604 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1607 /* - accumulate all nodes alive at point t on unit type k */
1608 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1609 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1610 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1613 /* check if node can be alive here */
1614 if (t < na->asap - 1)
1617 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1619 /* current type is not suitable */
1623 a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1624 ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1626 /* do the same for livein nodes */
1627 foreach_pset(ba->livein_nodes, livein) {
1628 ir_node *irn = livein->irn;
1629 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1632 /* check if node can be alive here */
1633 if (t >= livein->max_alive_steps)
1636 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1638 /* current type is not suitable */
1642 a_idx = tp_idx * livein->max_alive_steps + t;
1643 ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1646 if (ARR_LEN(tmp_var_idx) > 0)
1647 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1648 DEL_ARR_F(tmp_var_idx);
1650 /* - num_nodes * y_{nt}^k */
1651 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1652 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1657 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1658 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1662 * Create ILP branch constraints:
1663 * Assure, alle nodes are scheduled prior to cfg op.
1665 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1667 ir_node *cur_irn, *cfop;
1668 unsigned num_cst = 0;
1669 unsigned num_non_branches = 0;
1670 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1671 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_branch", "create branch constraints");
1673 ilp_timer_push(t_cst);
1675 /* determine number of non-branch nodes and the one and only branch node */
1676 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1677 switch (get_irn_opcode(cur_irn)) {
1687 if (is_cfop(cur_irn)) {
1688 assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1699 be_ilpsched_irn_t *cf_node = get_ilpsched_irn(env, cfop);
1700 ilpsched_node_attr_t *cf_na = get_ilpsched_node_attr(cf_node);
1703 /* for each time step */
1704 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1705 int *non_branch_vars, *branch_vars;
1708 snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1709 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1710 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1713 /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1714 non_branch_vars = NEW_ARR_F(int, 0);
1715 branch_vars = NEW_ARR_F(int, 0);
1716 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1717 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
1718 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1721 if (cur_irn == cfop) {
1722 /* for all unit types available for this node */
1723 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1724 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1725 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1729 /* sum up all possible schedule points for this node upto current timestep */
1730 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1732 unsigned tmax = MIN(t, na->alap - 1);
1734 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1735 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1736 ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1743 if (ARR_LEN(non_branch_vars) > 0)
1744 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1745 if (ARR_LEN(branch_vars) > 0)
1746 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1748 DEL_ARR_F(branch_vars);
1749 DEL_ARR_F(non_branch_vars);
1753 DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1754 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1758 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1761 unsigned num_cst = 0;
1762 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1763 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1765 ilp_timer_push(t_cst);
1766 /* check all nodes */
1767 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1768 be_ilpsched_irn_t *node;
1769 ilpsched_node_attr_t *na;
1773 /* only mode_T nodes can have Projs and Keeps assigned */
1774 if (get_irn_mode(irn) != mode_T)
1777 node = get_ilpsched_irn(env, irn);
1778 na = get_ilpsched_node_attr(node);
1780 /* check if has some Projs and Keeps assigned */
1781 if (! na->projkeeps)
1784 /* we can run only once over the queue, so preserve the nodes */
1785 pk = NEW_ARR_F(ir_node *, 0);
1786 while (! waitq_empty(na->projkeeps))
1787 ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1788 del_waitq(na->projkeeps);
1789 na->projkeeps = NULL;
1791 /* for all time steps at which this node can be scheduled */
1792 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1794 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1796 /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1797 snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1798 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1799 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1802 /* sum up scheduling variables for this time step */
1803 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1804 int idx = ILPVAR_IDX(na, tp_idx, t);
1805 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1808 if (ARR_LEN(tmp_var_idx_n) > 0)
1809 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1810 DEL_ARR_F(tmp_var_idx_n);
1812 /* subtract all Proj and Keep variables for this step */
1813 for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1814 be_ilpsched_irn_t *pk_node = get_ilpsched_irn(env, pk[i]);
1815 ilpsched_node_attr_t *pk_na = get_ilpsched_node_attr(pk_node);
1818 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1819 int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1820 lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1826 DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1827 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1831 /***************************************************
1833 * |_ _| | | __ \ (_)
1834 * | | | | | |__) | _ __ ___ __ _ _ _ __
1835 * | | | | | ___/ | '_ ` _ \ / _` | | '_ \
1836 * _| |_| |____| | | | | | | | (_| | | | | |
1837 * |_____|______|_| |_| |_| |_|\__,_|_|_| |_|
1839 ***************************************************/
1842 * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1844 static void create_ilp(ir_node *block, void *walk_env) {
1845 be_ilpsched_env_t *env = walk_env;
1846 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
1847 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1848 FILE *logfile = NULL;
1851 struct obstack var_obst;
1854 DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1855 DBG((env->dbg, 255, " ILP Scheduling for %+F\n", block));
1856 DBG((env->dbg, 255, "=========================================\n\n"));
1858 DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1859 block, ba->n_interesting_nodes, ba->max_steps));
1861 /* notify backend and get block environment */
1862 env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1864 /* if we have less than two interesting nodes, there is no need to create the ILP */
1865 if (ba->n_interesting_nodes > 1) {
1866 double fact_var = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1867 double fact_cst = ba->n_interesting_nodes > 25 ? 3 : 4.5;
1868 int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes;
1869 int estimated_n_var = (int)((double)base_num * fact_var);
1870 int estimated_n_cst = (int)((double)base_num * fact_cst);
1872 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1873 estimated_n_var, estimated_n_cst));
1874 (void) estimated_n_var;
1876 /* set up the LPP object */
1877 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1879 lpp = new_lpp_userdef(
1882 estimated_n_cst, /* num vars */
1883 estimated_n_cst + 1, /* num cst */
1884 1.3); /* grow factor */
1885 obstack_init(&var_obst);
1887 /* create ILP variables */
1888 create_variables(env, lpp, block_node, &var_obst);
1890 /* create ILP constraints */
1891 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1892 create_assignment_and_precedence_constraints(env, lpp, block_node);
1893 create_ressource_constraints(env, lpp, block_node);
1894 create_bundle_constraints(env, lpp, block_node);
1895 create_branch_constraint(env, lpp, block_node);
1896 //create_proj_keep_constraints(env, lpp, block_node);
1898 if (env->opts->regpress) {
1899 create_alive_nodes_constraint(env, lpp, block_node);
1900 create_alive_livein_nodes_constraint(env, lpp, block_node);
1901 create_pressure_alive_constraint(env, lpp, block_node);
1904 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1906 /* debug stuff, dump lpp when debugging is on */
1908 if (firm_dbg_get_mask(env->dbg) > 1) {
1912 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1913 f = fopen(buf, "w");
1914 lpp_dump_plain(lpp, f);
1916 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1921 /* set solve time limit */
1922 lpp_set_time_limit(lpp, env->opts->time_limit);
1924 /* set logfile if requested */
1925 if (strlen(env->opts->log_file) > 0) {
1926 if (strcasecmp(env->opts->log_file, "stdout") == 0)
1927 lpp_set_log(lpp, stdout);
1928 else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1929 lpp_set_log(lpp, stderr);
1931 logfile = fopen(env->opts->log_file, "w");
1933 fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1935 lpp_set_log(lpp, logfile);
1940 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1945 /* check for valid solution */
1946 if (! lpp_is_sol_valid(lpp)) {
1951 if (firm_dbg_get_mask(env->dbg) >= 2) {
1952 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1953 f = fopen(buf, "w");
1954 lpp_dump_plain(lpp, f);
1956 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1958 dump_ir_block_graph(env->irg, "-infeasible");
1962 ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1966 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1967 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1968 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1969 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));
1970 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1971 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1972 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1973 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1975 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1978 /* apply solution */
1979 be_stat_ev("nodes", ba->block_last_idx);
1980 be_stat_ev("vars", lpp ? lpp->var_next : 0);
1981 be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1983 be_stat_ev("time", -1);
1984 be_stat_ev_dbl("opt", 0.0);
1985 list_sched_single_block(env->birg, block, env->be_opts);
1989 double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1990 be_stat_ev_dbl("time", lpp->sol_time);
1991 be_stat_ev_dbl("opt", opt);
1994 be_stat_ev_dbl("time", 0.0);
1995 be_stat_ev_dbl("opt", 100.0);
1997 apply_solution(env, lpp, block);
2003 /* notify backend */
2004 be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2008 * Perform ILP scheduling on the given irg.
2010 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2011 be_ilpsched_env_t env;
2012 const char *name = "be ilp scheduling";
2013 ir_graph *irg = be_get_birg_irg(birg);
2014 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
2015 const ilp_sched_selector_t *sel = arch_env->impl->get_ilp_sched_selector(arch_env);
2017 FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2019 stat_ev_ctx_push("ilpsched");
2021 // firm_dbg_set_mask(env.dbg, 1);
2023 env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2026 env.height = heights_new(irg);
2027 env.main_env = birg->main_env;
2028 env.arch_env = arch_env;
2029 env.cpu = arch_env_get_machine(arch_env);
2030 env.opts = &ilp_opts;
2032 env.be_opts = be_opts;
2033 phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2035 /* assign a unique per block number to all interesting nodes */
2036 irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2039 The block indices are completely build after the walk,
2040 now we can allocate the bitsets (size depends on block indices)
2043 phase_reinit_irn_data(&env.ph);
2045 /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2046 irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2048 /* Calculate ALAP of all irns */
2049 irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2051 /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2052 irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2054 /* perform ILP scheduling */
2055 irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2058 if (firm_dbg_get_mask(env.dbg)) {
2060 phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2062 fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2066 /* free data allocated dynamically */
2067 irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2069 /* free all allocated object */
2070 phase_free(&env.ph);
2071 heights_free(env.height);
2073 /* notify backend */
2074 be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2076 stat_ev_ctx_pop("ilpsched");
2080 * Register ILP scheduler options.
2082 void be_init_ilpsched(void)
2084 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2085 lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2087 lc_opt_add_table(sched_grp, ilpsched_option_table);
2090 #else /* WITH_ILP */
2092 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2095 #endif /* WITH_ILP */