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
45 #include "irphase_t.h"
58 #include <lpp/lpp_net.h>
61 #include "lc_opts_enum.h"
65 #include "besched_t.h"
66 #include "beilpsched.h"
71 typedef struct _ilpsched_options_t {
77 typedef struct _unit_type_info_t {
79 const be_execution_unit_type_t *tp;
83 * holding the ILP variables of the different types
85 typedef struct _ilp_var_types_t {
86 int *x; /* x_{nt}^k variables */
87 int *a; /* a_{nt}^k variables */
88 int *y; /* y_{nt}^k variables */
92 * Holds alive variables for a node live-in to a block.
94 typedef struct _ilp_livein_node_t {
96 unsigned max_alive_steps;
100 /* attributes for a node */
101 typedef struct _ilpsched_node_attr_t {
102 unsigned asap; /**< The ASAP scheduling control step */
103 unsigned alap; /**< The ALAP scheduling control step */
104 unsigned latency; /**< Latency of this node (needed for sorting) */
105 unsigned sched_point; /**< the step in which the node is finally scheduled */
106 unsigned visit_idx; /**< Index of the node having visited this node last */
107 unsigned consumer_idx; /**< Index of the node having counted this node as consumer last */
108 unsigned n_consumer; /**< Number of consumers */
109 ir_node **block_consumer; /**< List of consumer being in the same block */
110 waitq *projkeeps; /**< A List of Projs and Keeps belonging to this node */
111 unsigned block_idx : 30; /**< A unique per block index */
112 unsigned alap_changed : 1; /**< the current ALAP has changed, revisit preds */
113 unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */
114 bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors
115 for ASAP, successors for ALAP */
116 unsigned n_unit_types; /**< number of allowed execution unit types */
117 unit_type_info_t *type_info; /**< list of allowed execution unit types */
118 ilp_var_types_t ilp_vars; /**< the different ILP variables */
119 } ilpsched_node_attr_t;
121 /* attributes for a block */
122 typedef struct _ilpsched_block_attr_t {
123 unsigned block_last_idx; /**< The highest node index in block so far */
124 unsigned n_interesting_nodes; /**< The number of nodes interesting for scheduling */
125 unsigned max_steps; /**< Upper bound for block execution */
126 plist_t *root_nodes; /**< A list of nodes having no user in current block */
127 ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */
128 pset *livein_nodes; /**< A set of nodes which are live-in to this block */
129 } ilpsched_block_attr_t;
131 typedef union _ilpsched_attr_ {
132 ilpsched_node_attr_t node_attr;
133 ilpsched_block_attr_t block_attr;
136 /* A irn for the phase and it's attributes (either node or block) */
139 ilpsched_attr_t attr;
142 /* The ILP scheduling environment */
144 ir_phase ph; /**< The phase */
145 ir_graph *irg; /**< The current irg */
146 heights_t *height; /**< The heights object of the irg */
147 void *irg_env; /**< An environment for the irg scheduling, provided by the backend */
148 void *block_env; /**< An environment for scheduling a block, provided by the backend */
149 const arch_env_t *arch_env;
150 const be_main_env_t *main_env;
151 const be_machine_t *cpu; /**< the current abstract machine */
152 ilpsched_options_t *opts; /**< the ilp options for current irg */
153 const be_irg_t *birg; /**< The birg object */
154 be_options_t *be_opts; /**< backend options */
155 const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */
156 DEBUG_ONLY(firm_dbg_module_t *dbg);
159 /* convenience macros to handle phase irn data */
160 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
161 #define is_ilpsched_block(node) (is_Block((node)->irn))
162 #define get_ilpsched_block_attr(block) (&(block)->attr.block_attr)
163 #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr)
165 /* check if node is considered for ILP scheduling */
166 #define consider_for_sched(env, irn) \
167 (! (is_Block(irn) || \
168 is_normal_Proj(env, irn) || \
175 /* gives the valid scheduling time step interval for a node */
176 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
178 /* gives the valid interval where a node can die */
179 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
181 /* gives the corresponding ILP variable for given node, unit and time step */
182 #define ILPVAR_IDX(na, unit, control_step) \
183 ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
185 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
186 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
187 ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
189 /* check if a double value is within an epsilon environment of 0 */
190 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
192 #define ilp_timer_push(t) ir_timer_push((t))
193 #define ilp_timer_pop() ir_timer_pop()
194 #define ilp_timer_elapsed_usec(t) ir_timer_elapsed_usec((t))
196 /* option variable */
197 static ilpsched_options_t ilp_opts = {
198 1, /* default is with register pressure constraints */
199 300, /* 300 sec per block time limit */
204 static const lc_opt_table_entry_t ilpsched_option_table[] = {
205 LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress),
206 LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
207 LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
212 We need this global variable as we compare nodes dependent on heights,
213 but we cannot pass any information to the qsort compare function.
215 static heights_t *glob_heights;
218 * Check if irn is a Proj, which has no execution units assigned.
219 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
221 static inline int is_normal_Proj(const arch_env_t *env, const ir_node *irn) {
222 return is_Proj(irn) && (arch_env_get_allowed_execution_units(env, irn) == NULL);
226 * Skips normal Projs.
227 * @return predecessor if irn is a normal Proj, otherwise irn.
229 static inline ir_node *skip_normal_Proj(const arch_env_t *env, ir_node *irn) {
230 if (is_normal_Proj(env, irn))
231 return get_Proj_pred(irn);
235 static inline int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
236 unsigned lat = be_ilp_sched_latency(sel, irn, env);
237 if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
242 static int cmp_live_in_nodes(const void *a, const void *b) {
243 const ilp_livein_node_t *n1 = a;
244 const ilp_livein_node_t *n2 = b;
246 return n1->irn != n2->irn;
250 * Compare scheduling time steps of two be_ilpsched_irn's.
252 static int cmp_ilpsched_irn(const void *a, const void *b) {
253 be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a;
254 be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b;
255 ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
256 ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
258 if (n1_a->sched_point == n2_a->sched_point) {
259 const ir_node *irn_a = n1->irn;
260 const ir_node *irn_b = n2->irn;
262 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
264 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
268 Ok, timestep is equal and the nodes are parallel,
269 so check latency and schedule high latency first.
271 return QSORT_CMP(n2_a->latency, n1_a->latency);
274 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
278 * In case there is no phase information for irn, initialize it.
280 static void *init_ilpsched_irn(ir_phase *ph, const ir_node *irn, void *old) {
281 be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
284 /* if we have already some data: check for reinitialization */
286 if (! is_Block(irn)) {
287 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
289 if (! na->transitive_block_nodes) {
290 ir_node *block = get_nodes_block(irn);
291 be_ilpsched_irn_t *block_node = phase_get_or_set_irn_data(ph, block);
292 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
294 /* we are called after the block indices have been build: create bitset */
295 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
298 /* we are called from reinit block data: clear the bitset */
299 bitset_clear_all(na->transitive_block_nodes);
301 na->alap_changed = 1;
309 /* set ilpsched irn attributes (either block or irn) */
311 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
313 ba->n_interesting_nodes = 0;
314 ba->block_last_idx = 0;
315 ba->root_nodes = plist_new();
316 ba->head_ilp_nodes = NULL;
317 ba->livein_nodes = new_pset(cmp_live_in_nodes, 16);
321 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
322 memset(na, 0, sizeof(*na));
329 * Assign a per block unique number to each node.
331 static void build_block_idx(ir_node *irn, void *walk_env) {
332 be_ilpsched_env_t *env = walk_env;
333 be_ilpsched_irn_t *node, *block_node;
334 ilpsched_node_attr_t *na;
335 ilpsched_block_attr_t *ba;
337 set_irn_link(irn, NULL);
338 if (! consider_for_sched(env->arch_env, irn))
341 node = get_ilpsched_irn(env, irn);
342 na = get_ilpsched_node_attr(node);
343 block_node = get_ilpsched_irn(env, get_nodes_block(irn));
344 ba = get_ilpsched_block_attr(block_node);
346 na->block_idx = ba->block_last_idx++;
349 /********************************************************
352 * __ _ ___ __ _ _ __ / / __ _| | __ _ _ __
353 * / _` / __|/ _` | '_ \ / / / _` | |/ _` | '_ \
354 * | (_| \__ \ (_| | |_) | / / | (_| | | (_| | |_) |
355 * \__,_|___/\__,_| .__/ /_/ \__,_|_|\__,_| .__/
358 ********************************************************/
361 * Add all nodes having no user in current block to last_nodes list.
363 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
365 const ir_edge_t *edge;
366 be_ilpsched_irn_t *block_node, *node;
367 ilpsched_block_attr_t *ba;
368 ilpsched_node_attr_t *na;
370 be_ilpsched_env_t *env = walk_env;
371 int has_block_user = 0;
372 unsigned n_consumer = 0;
373 ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
377 if (! consider_for_sched(env->arch_env, irn))
380 block = get_nodes_block(irn);
381 idx = get_irn_idx(irn);
382 consumer = NEW_ARR_F(ir_node *, 0);
384 DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
386 /* check data and dependency out edges */
387 for (i = 0; i < 2 && ! has_block_user; ++i) {
388 foreach_out_edge_kind(irn, edge, ekind[i]) {
389 ir_node *user = get_edge_src_irn(edge);
391 if (is_normal_Proj(env->arch_env, user)) {
392 const ir_edge_t *user_edge;
394 if (get_irn_mode(user) == mode_X)
397 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
398 for (j = 0; j < 2; ++j) {
399 foreach_out_edge_kind(user, user_edge, ekind[j]) {
400 ir_node *real_user = get_edge_src_irn(user_edge);
402 if (! is_Phi(real_user) && ! is_Block(real_user)) {
403 be_ilpsched_irn_t *node = get_ilpsched_irn(env, real_user);
404 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
406 /* skip already visited nodes */
407 if (ua->consumer_idx == idx)
410 /* check if node has user in this block and collect the user if it's a data user */
411 if (get_nodes_block(real_user) == block) {
412 if (i == 0 && j == 0)
413 ARR_APP1(ir_node *, consumer, real_user);
417 /* only count data consumer */
421 /* mark user as visited by this node */
422 ua->consumer_idx = idx;
427 else if (is_Block(user)) {
430 else if (! is_Phi(user)) {
431 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
432 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
434 /* skip already visited nodes */
435 if (ua->consumer_idx == idx)
438 /* check if node has user in this block and collect the user if it's a data user */
439 if (get_nodes_block(user) == block) {
441 ARR_APP1(ir_node *, consumer, user);
445 /* only count data consumer */
449 /* mark user visited by this node */
450 ua->consumer_idx = idx;
452 else if (get_nodes_block(user) != block) {
458 block_node = get_ilpsched_irn(env, block);
459 ba = get_ilpsched_block_attr(block_node);
461 ba->n_interesting_nodes++;
463 /* current irn has no user inside this block, add to queue */
464 if (! has_block_user) {
465 DB((env->dbg, LEVEL_3, "root node\n"));
466 plist_insert_back(ba->root_nodes, irn);
469 DB((env->dbg, LEVEL_3, "normal node\n"));
472 /* record number of all consumer and the consumer within the same block */
473 node = get_ilpsched_irn(env, irn);
474 na = get_ilpsched_node_attr(node);
475 na->n_consumer = n_consumer;
476 na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
477 memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
482 * Calculate the ASAP scheduling step for current irn.
484 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
485 be_ilpsched_env_t *env = walk_env;
488 be_ilpsched_irn_t *node, *block_node;
489 ilpsched_node_attr_t *na;
490 ilpsched_block_attr_t *ba;
492 /* These nodes are handled separate */
493 if (! consider_for_sched(env->arch_env, irn))
496 DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
498 block = get_nodes_block(irn);
499 node = get_ilpsched_irn(env, irn);
500 na = get_ilpsched_node_attr(node);
503 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
504 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(irn, i));
506 /* check for greatest distance to top */
507 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
508 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
509 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
512 lat = fixed_latency(env->sel, pred, env->block_env);
514 na->asap = MAX(na->asap, pna->asap + lat);
518 /* add node to ILP node list and update max_steps */
519 block_node = get_ilpsched_irn(env, block);
520 ba = get_ilpsched_block_attr(block_node);
522 set_irn_link(irn, ba->head_ilp_nodes);
523 ba->head_ilp_nodes = irn;
524 ba->max_steps += fixed_latency(env->sel, irn, env->block_env);
526 DB((env->dbg, LEVEL_2, "%u\n", na->asap));
530 * Calculate the ALAP scheduling step of all irns in current block.
531 * Depends on max_steps being calculated.
533 static void calculate_block_alap(ir_node *block, void *walk_env) {
534 be_ilpsched_env_t *env = walk_env;
535 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
536 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
537 waitq *cur_queue = new_waitq();
540 assert(is_Block(block));
542 DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
543 block, ba->n_interesting_nodes, ba->max_steps));
545 /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
546 //phase_reinit_block_irn_data(&env->ph, block);
548 /* init start queue */
549 foreach_plist(ba->root_nodes, el) {
550 waitq_put(cur_queue, plist_element_get_value(el));
553 /* repeat until all nodes are processed */
554 while (! waitq_empty(cur_queue)) {
555 waitq *next_queue = new_waitq();
557 /* process all nodes in current step */
558 while (! waitq_empty(cur_queue)) {
559 ir_node *cur_irn = waitq_get(cur_queue);
560 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
561 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
564 /* cur_node has no alap set -> it's a root node, set to max alap */
566 na->alap = ba->max_steps;
567 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
571 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
575 /* set the alap's of all predecessors */
576 for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
577 ir_node *pred = skip_normal_Proj(env->arch_env, get_irn_in_or_dep(cur_irn, i));
579 /* check for greatest distance to bottom */
580 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
581 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
582 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
585 /* mark the predecessor as visited by current irn */
586 if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
588 pna->visit_idx = get_irn_idx(cur_irn);
590 lat = fixed_latency(env->sel, pred, env->block_env);
592 /* set ALAP of current pred */
593 if (pna->alap == 0) {
594 /* current ALAP is 0: node has not yet been visited */
595 pna->alap_changed = 1;
596 pna->alap = na->alap - lat;
598 else if (pna->alap > na->alap - lat) {
599 /* we found a longer path to root node: change ALAP */
600 pna->alap = na->alap - lat;
601 pna->alap_changed = 1;
604 /* current ALAP is best found so far: keep it */
605 pna->alap_changed = 0;
608 DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
610 /* enqueue node for next iteration */
611 if (get_irn_ins_or_deps(pred) > 0)
612 waitq_put(next_queue, pred);
617 /* prepare for next iteration */
618 del_waitq(cur_queue);
619 cur_queue = next_queue;
624 * Free list of root nodes and the set of live-in nodes.
626 static void clear_unwanted_data(ir_node *block, void *walk_env) {
627 be_ilpsched_env_t *env = walk_env;
628 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
629 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
631 plist_free(ba->root_nodes);
632 ba->root_nodes = NULL;
633 del_pset(ba->livein_nodes);
634 ba->livein_nodes = NULL;
638 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
639 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
641 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
642 be_ilpsched_env_t *env = walk_env;
644 be_ilpsched_irn_t *node, *pred_node;
645 ilpsched_node_attr_t *na, *pna;
647 if (! consider_for_sched(env->arch_env, irn))
650 if (! is_Proj(irn) && ! be_is_Keep(irn))
653 /* go to the ancestor */
655 pred = get_irn_n(irn, 0);
656 pred = skip_Proj(pred);
658 node = get_ilpsched_irn(env, irn);
659 pred_node = get_ilpsched_irn(env, pred);
660 na = get_ilpsched_node_attr(node);
661 pna = get_ilpsched_node_attr(pred_node);
663 na->asap = pna->asap;
664 na->alap = pna->alap;
666 /* record all Projs and Keeps for this node */
667 if (! pna->projkeeps)
668 pna->projkeeps = new_waitq();
669 waitq_put(pna->projkeeps, irn);
671 DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
674 /*******************************************
677 * ___ ___| |__ ___ __| |_ _| | ___
678 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
679 * \__ \ (__| | | | __/ (_| | |_| | | __/
680 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
682 *******************************************/
684 static inline void check_for_keeps(waitq *keeps, const ir_node *block, const ir_node *irn) {
685 const ir_edge_t *edge;
688 foreach_out_edge(irn, edge) {
689 ir_node *user = get_edge_src_irn(edge);
691 if (be_is_Keep(user)) {
692 assert(get_nodes_block(user) == block && "Keep must not be in different block.");
693 waitq_put(keeps, user);
699 * Inserts @p irn before @p before into schedule and notifies backend.
701 static inline void notified_sched_add_before(be_ilpsched_env_t *env,
702 const ir_node *before, const ir_node *irn, unsigned cycle)
704 be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
705 sched_add_before((ir_node*) before, (ir_node*) irn);
709 * Adds a node, it's Projs (in case of mode_T nodes) and
710 * it's Keeps to schedule.
712 static void add_to_sched(be_ilpsched_env_t *env, const ir_node *block, const ir_node *irn, unsigned cycle) {
713 const ir_edge_t *edge;
714 waitq *keeps = new_waitq();
716 /* mode_M nodes are not scheduled */
717 if (get_irn_mode(irn) == mode_M)
720 if (! sched_is_scheduled(irn))
721 notified_sched_add_before(env, block, irn, cycle);
724 if (get_irn_mode(irn) == mode_T) {
725 foreach_out_edge(irn, edge) {
726 ir_node *user = get_edge_src_irn(edge);
728 if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
729 get_irn_n_edges(user) > 0)
731 notified_sched_add_before(env, block, user, cycle);
734 check_for_keeps(keeps, block, user);
738 check_for_keeps(keeps, block, irn);
742 while (! waitq_empty(keeps)) {
743 ir_node *keep = waitq_get(keeps);
744 if (! sched_is_scheduled(keep))
745 notified_sched_add_before(env, block, keep, cycle);
752 * Schedule all nodes in the given block, according to the ILP solution.
754 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
755 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
756 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
757 be_ilpsched_irn_t **sched_nodes;
760 const ir_edge_t *edge;
762 /* init block schedule list */
763 sched_init_block(block);
765 /* collect nodes and their scheduling time step */
766 sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
767 if (ba->n_interesting_nodes == 0) {
770 else if (ba->n_interesting_nodes == 1) {
771 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
773 /* add the single node */
774 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
777 /* check all nodes for their positive solution */
778 foreach_linked_irns(ba->head_ilp_nodes, irn) {
779 be_ilpsched_irn_t *node;
780 ilpsched_node_attr_t *na;
784 node = get_ilpsched_irn(env, irn);
785 na = get_ilpsched_node_attr(node);
789 if (! na->is_dummy_node) {
790 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
791 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
792 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
794 if (! LPP_VALUE_IS_0(cost)) {
795 DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
803 /* go over all variables of a node until the non-zero one is found */
804 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
805 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
806 double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
808 /* check, if variable is set to one (it's not zero then :) */
809 if (! LPP_VALUE_IS_0(val)) {
811 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
812 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
813 irn, t, na->type_info[tp_idx].tp->name));
820 glob_heights = env->height;
821 /* sort nodes ascending by scheduling time step */
822 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
825 /* make all Phis ready and remember the single cf op */
827 foreach_out_edge(block, edge) {
828 irn = get_edge_src_irn(edge);
830 switch (get_irn_opcode(irn)) {
832 add_to_sched(env, block, irn, 0);
842 assert(cfop == NULL && "Highlander - there can be only one");
849 /* add all nodes from list */
850 for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
851 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
852 if (sched_nodes[i]->irn != cfop)
853 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
856 /* schedule control flow node if not already done */
857 if (cfop && ! sched_is_scheduled(cfop))
858 add_to_sched(env, block, cfop, 0);
860 DEL_ARR_F(sched_nodes);
863 /***************************************************************
864 * _____ _ _____ _____ _ _
865 * |_ _| | | __ \ / ____| | | (_)
866 * | | | | | |__) | | (___ ___ ___| |_ _ ___ _ __
867 * | | | | | ___/ \___ \ / _ \/ __| __| |/ _ \| '_ \
868 * _| |_| |____| | ____) | __/ (__| |_| | (_) | | | |
869 * |_____|______|_| |_____/ \___|\___|\__|_|\___/|_| |_|
871 ***************************************************************/
874 * Check if node can be executed on given unit type.
876 static inline int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
878 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
880 for (i = na->n_unit_types - 1; i >= 0; --i) {
881 if (na->type_info[i].tp == tp)
888 /************************************************
891 * __ ____ _ _ __ _ __ _| |__ | | ___ ___
892 * \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
893 * \ V / (_| | | | | (_| | |_) | | __/\__ \
894 * \_/ \__,_|_| |_|\__,_|_.__/|_|\___||___/
896 ************************************************/
898 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
899 const be_execution_unit_t ***execunits = arch_env_get_allowed_execution_units(env->arch_env, irn);
900 unsigned n_unit_types = 0;
901 be_ilpsched_irn_t *node;
902 ilpsched_node_attr_t *na;
903 unsigned unit_idx, tp_idx;
905 /* count number of available unit types for this node */
906 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
909 node = get_ilpsched_irn(env, irn);
910 na = get_ilpsched_node_attr(node);
912 if (! na->type_info) {
913 na->n_unit_types = n_unit_types;
914 na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
916 /* fill the type info array */
917 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
918 for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
919 /* beware: we also count number of available units here */
920 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
921 na->is_dummy_node = 1;
924 na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp;
925 na->type_info[tp_idx].n_units = unit_idx;
933 * Returns the largest alap time of a user of @p irn.
934 * The user must be in block @p block.
936 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, const ir_node *irn, const ir_node *block) {
937 const ir_edge_t *edge;
938 unsigned max_alap = 0;
940 foreach_out_edge(irn, edge) {
941 ir_node *user = get_edge_src_irn(edge);
943 if (get_nodes_block(user) == block) {
944 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
945 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
947 max_alap = MAX(max_alap, na->alap);
951 assert(max_alap > 0);
956 * Create the following variables:
957 * - x_{nt}^k binary weigthed with: t
958 * node n is scheduled at time step t to unit type k
959 * ==>> These variables represent the schedule
961 * - a_{nt}^k binary weighted with num_nodes
962 * node n is alive at time step t on unit type k
964 * - y_{nt}^k continuous weighted with: num_nodes^2
965 * register pressure over limit for unit type k
966 * ==>> These variables represent the register pressure
969 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
972 unsigned num_block_var, num_nodes;
973 ilp_livein_node_t *livein;
974 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
975 unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
976 ir_timer_t *t_var = ir_timer_register("beilpsched_var", "create ilp variables");
978 ilp_timer_push(t_var);
979 num_block_var = num_nodes = 0;
980 foreach_linked_irns(ba->head_ilp_nodes, irn) {
981 be_ilpsched_irn_t *node;
982 ilpsched_node_attr_t *na;
983 unsigned n_unit_types, tp_idx, n_var, cur_unit;
984 unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad;
987 node = get_ilpsched_irn(env, irn);
988 na = get_ilpsched_node_attr(node);
989 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
991 /* allocate space for ilp variables */
992 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
993 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
995 /* we need these variables only for "real" nodes */
996 if (! na->is_dummy_node) {
997 na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
998 memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1000 num_ad = ba->max_steps - na->asap + 1;
1001 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1002 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1005 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1006 irn, na->asap, na->alap, na->n_unit_types));
1008 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1009 /* create variables */
1010 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1013 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1014 /* x_{nt}^k variables */
1015 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1016 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1017 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1018 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1019 /* variable counter */
1023 if (! na->is_dummy_node) {
1024 /* y_{nt}^k variables */
1025 snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1026 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1027 na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1028 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1030 /* variable counter */
1036 /* a node can die at any step t: asap(n) <= t <= U */
1037 if (! na->is_dummy_node) {
1038 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1040 /* a_{nt}^k variables */
1041 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1042 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1043 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1044 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1046 /* variable counter */
1052 /* collect live-in nodes */
1053 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1054 ir_node *pred = get_irn_n(irn, i);
1056 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env, pred)) {
1057 be_ilpsched_set_type_info(env, pred, var_obst);
1058 if (! na->is_dummy_node) {
1059 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1062 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1068 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1072 /* create alive variables a_{nt}^k for live-ins */
1073 foreach_pset(ba->livein_nodes, livein) {
1074 be_ilpsched_irn_t *node;
1075 ilpsched_node_attr_t *na;
1076 unsigned tp_idx, var_idx;
1080 node = get_ilpsched_irn(env, irn);
1081 na = get_ilpsched_node_attr(node);
1083 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1085 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1088 /* create variables */
1089 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1092 for (t = 0; t < livein->max_alive_steps; ++t) {
1093 /* a_{nt}^k variables */
1094 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1095 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1096 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1097 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1104 DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1105 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1108 /*******************************************************
1111 * ___ ___ _ __ ___| |_ _ __ __ _ _ _ __ | |_ ___
1112 * / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1113 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1114 * \___\___/|_| |_|___/\__|_| \__,_|_|_| |_|\__|___/
1116 *******************************************************/
1119 * Collect all operands and nodes @p irn depends on.
1120 * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1122 static void sta_collect_in_deps(ir_node *irn, ir_nodeset_t *deps) {
1125 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1126 ir_node *p = get_irn_in_or_dep(irn, i);
1129 const ir_edge_t *edge;
1131 p = get_Proj_pred(p);
1132 foreach_out_edge(p, edge) {
1133 ir_node *src = get_edge_src_irn(edge);
1134 ir_nodeset_insert(deps, src);
1138 ir_nodeset_insert(deps, p);
1144 * Create following ILP constraints:
1145 * - the assignment constraints:
1146 * assure each node is executed once by exactly one (allowed) execution unit
1147 * - the dead node assignment constraints:
1148 * assure a node can only die at most once
1149 * - the precedence constraints:
1150 * assure that no data dependencies are violated
1152 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1153 unsigned num_cst_assign, num_cst_prec, num_cst_dead;
1156 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1157 bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx);
1158 ir_timer_t *t_cst_assign = ir_timer_register("beilpsched_cst_assign", "create assignment constraints");
1159 ir_timer_t *t_cst_prec = ir_timer_register("beilpsched_cst_prec", "create precedence constraints");
1161 num_cst_assign = num_cst_prec = num_cst_dead = 0;
1162 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1165 be_ilpsched_irn_t *node;
1166 ilpsched_node_attr_t *na;
1169 ir_nodeset_iterator_t iter;
1171 ir_nodeset_init(&deps);
1173 node = get_ilpsched_irn(env, irn);
1174 na = get_ilpsched_node_attr(node);
1177 /* the assignment constraint */
1178 ilp_timer_push(t_cst_assign);
1179 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1180 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1181 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1184 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1187 /* We have separate constraints for Projs and Keeps */
1188 // ILP becomes infeasible ?!?
1189 // if (is_Proj(irn) || be_is_Keep(irn))
1192 /* the precedence constraints */
1193 ilp_timer_push(t_cst_prec);
1194 bs_block_irns = bitset_clear_all(bs_block_irns);
1196 sta_collect_in_deps(irn, &deps);
1197 foreach_ir_nodeset(&deps, pred, iter) {
1198 unsigned t_low, t_high, t;
1199 be_ilpsched_irn_t *pred_node;
1200 ilpsched_node_attr_t *pna;
1203 pred = skip_normal_Proj(env->arch_env, pred);
1204 if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1207 pred_node = get_ilpsched_irn(env, pred);
1208 pna = get_ilpsched_node_attr(pred_node);
1210 assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1212 if (! bitset_is_set(bs_block_irns, pna->block_idx))
1213 bitset_set(bs_block_irns, pna->block_idx);
1217 /* irn = n, pred = m */
1218 delay = fixed_latency(env->sel, pred, env->block_env);
1219 t_low = MAX(na->asap, pna->asap + delay - 1);
1220 t_high = MIN(na->alap, pna->alap + delay - 1);
1221 for (t = t_low - 1; t <= t_high - 1; ++t) {
1223 int *tmp_var_idx = NEW_ARR_F(int, 0);
1225 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1226 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1227 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1230 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1231 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1232 /* node variables have smaller index than pred variables */
1233 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1234 for (tn = na->asap - 1; tn <= t; ++tn) {
1235 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1236 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1240 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1241 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1242 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1243 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1248 /* pred variables have smaller index than node variables */
1249 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1250 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1251 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1252 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1256 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1257 for (tn = na->asap - 1; tn <= t; ++tn) {
1258 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1259 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1264 if (ARR_LEN(tmp_var_idx) > 0)
1265 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1267 DEL_ARR_F(tmp_var_idx);
1270 ir_nodeset_destroy(&deps);
1273 DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1274 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1275 DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1276 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1280 * Create ILP resource constraints:
1281 * - assure that for each time step not more instructions are scheduled
1282 * to the same unit types as units of this type are available
1284 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1287 unsigned num_cst_resrc = 0;
1288 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1289 ir_timer_t *t_cst_rsrc = ir_timer_register("beilpsched_cst_rsrc", "create resource constraints");
1291 ilp_timer_push(t_cst_rsrc);
1292 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1294 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1296 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1298 /* check each time step */
1299 for (t = 0; t < ba->max_steps; ++t) {
1302 int *tmp_var_idx = NEW_ARR_F(int, 0);
1304 snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1305 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1306 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1309 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1310 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1311 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1314 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1316 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1317 int cur_var = ILPVAR_IDX(na, tp_idx, t);
1318 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1322 /* set constraints if we have some */
1323 if (ARR_LEN(tmp_var_idx) > 0)
1324 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1326 DEL_ARR_F(tmp_var_idx);
1330 DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1331 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1335 * Create ILP bundle constraints:
1336 * - assure, at most bundle_size * bundles_per_cycle instructions
1337 * can be started at a certain point.
1339 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1342 unsigned num_cst_bundle = 0;
1343 unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1344 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1345 ir_timer_t *t_cst_bundle = ir_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1347 ilp_timer_push(t_cst_bundle);
1348 for (t = 0; t < ba->max_steps; ++t) {
1351 int *tmp_var_idx = NEW_ARR_F(int, 0);
1353 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1354 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1355 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1358 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1359 be_ilpsched_irn_t *node;
1360 ilpsched_node_attr_t *na;
1363 /* Projs and Keeps do not contribute to bundle size */
1364 if (is_Proj(irn) || be_is_Keep(irn))
1367 node = get_ilpsched_irn(env, irn);
1368 na = get_ilpsched_node_attr(node);
1370 /* nodes assigned to DUMMY unit do not contribute to bundle size */
1371 if (na->is_dummy_node)
1374 if (t >= na->asap - 1 && t <= na->alap - 1) {
1375 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1376 int idx = ILPVAR_IDX(na, tp_idx, t);
1377 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1382 if (ARR_LEN(tmp_var_idx) > 0)
1383 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1385 DEL_ARR_F(tmp_var_idx);
1388 DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1389 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1393 * Create ILP alive nodes constraints:
1394 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1396 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1399 unsigned num_cst = 0;
1400 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1401 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1403 ilp_timer_push(t_cst);
1405 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1406 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1407 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1410 /* we ignore nodes assigned to dummy unit here */
1411 if (na->is_dummy_node)
1414 /* check check all time steps: asap(n) <= t <= U */
1415 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1418 /* for all unit types available for this node */
1419 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1420 unsigned tn, tn_max, idx;
1422 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1423 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1425 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1426 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1427 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1428 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1431 tn_max = MIN(na->alap - 1, t);
1432 /* check if the node has been scheduled so far */
1433 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1434 int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1435 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1438 if (ARR_LEN(tmp_var_idx_n) > 0)
1439 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1440 DEL_ARR_F(tmp_var_idx_n);
1442 /* subtract the number of consumer scheduled so far */
1443 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1444 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1445 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1447 unsigned tm, tm_max;
1449 tm_max = MIN(ca->alap - 1, t);
1450 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1451 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1452 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1453 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1458 if (ARR_LEN(tmp_var_idx_m) > 0)
1459 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1460 DEL_ARR_F(tmp_var_idx_m);
1463 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1464 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1470 DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1471 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1475 * Create ILP alive nodes constraints for live-in nodes:
1476 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1478 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1480 ilp_livein_node_t *livein;
1481 unsigned num_cst = 0;
1482 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1483 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1485 ilp_timer_push(t_cst);
1487 foreach_pset(ba->livein_nodes, livein) {
1488 ir_node *irn = livein->irn;
1489 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1490 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1493 /* check check all time steps: 0 <= t < max_alive_steps */
1494 for (t = 0; t < livein->max_alive_steps; ++t) {
1497 /* for all unit types available for this node */
1498 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1499 const ir_edge_t *edge;
1501 int cst, num_block_user;
1502 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1504 /* check the number of consumer scheduled so far */
1506 foreach_out_edge(irn, edge) {
1507 ir_node *user = get_edge_src_irn(edge);
1508 be_ilpsched_irn_t *cons;
1509 ilpsched_node_attr_t *ca;
1511 unsigned tm, tm_max;
1513 /* check only users within current block */
1514 if (get_nodes_block(user) != block_node->irn)
1518 cons = get_ilpsched_irn(env, user);
1519 ca = get_ilpsched_node_attr(cons);
1521 tm_max = MIN(ca->alap - 1, t);
1522 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1523 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1524 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1525 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1530 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1531 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1532 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1533 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1536 /* sum(scheduled users) */
1537 if (ARR_LEN(tmp_var_idx_m) > 0)
1538 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1539 DEL_ARR_F(tmp_var_idx_m);
1541 /* + c * a_{nt}^k */
1542 idx = node_tp_idx * livein->max_alive_steps + t;
1543 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1548 DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1549 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1553 * Create ILP pressure constraints, based on alive nodes:
1554 * - add additional costs to objective function if a node is scheduled
1555 * on a unit although all units of this type are currently occupied
1557 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1560 unsigned num_cst = 0;
1561 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1562 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1564 ilp_timer_push(t_cst);
1565 /* y_{nt}^k is set for each node and timestep and unit type */
1566 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1567 unsigned cur_idx = get_irn_idx(cur_irn);
1568 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1569 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1572 /* we ignore nodes assigned to DUMMY unit here */
1573 if (cur_na->is_dummy_node)
1577 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1578 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1582 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1584 /* check if node can be executed on this unit type */
1585 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1589 /* check all time_steps at which the current node can be scheduled */
1590 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1593 int *tmp_var_idx = NEW_ARR_F(int, 0);
1594 ilp_livein_node_t *livein;
1596 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1597 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1598 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1601 /* - accumulate all nodes alive at point t on unit type k */
1602 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1603 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1604 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1607 /* check if node can be alive here */
1608 if (t < na->asap - 1)
1611 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1613 /* current type is not suitable */
1617 a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1618 ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1620 /* do the same for livein nodes */
1621 foreach_pset(ba->livein_nodes, livein) {
1622 ir_node *irn = livein->irn;
1623 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1626 /* check if node can be alive here */
1627 if (t >= livein->max_alive_steps)
1630 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1632 /* current type is not suitable */
1636 a_idx = tp_idx * livein->max_alive_steps + t;
1637 ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1640 if (ARR_LEN(tmp_var_idx) > 0)
1641 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1642 DEL_ARR_F(tmp_var_idx);
1644 /* - num_nodes * y_{nt}^k */
1645 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1646 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1651 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1652 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1656 * Create ILP branch constraints:
1657 * Assure, alle nodes are scheduled prior to cfg op.
1659 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1661 ir_node *cur_irn, *cfop;
1662 unsigned num_cst = 0;
1663 unsigned num_non_branches = 0;
1664 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1665 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_branch", "create branch constraints");
1667 ilp_timer_push(t_cst);
1669 /* determine number of non-branch nodes and the one and only branch node */
1670 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1671 switch (get_irn_opcode(cur_irn)) {
1681 if (is_cfop(cur_irn)) {
1682 assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1693 be_ilpsched_irn_t *cf_node = get_ilpsched_irn(env, cfop);
1694 ilpsched_node_attr_t *cf_na = get_ilpsched_node_attr(cf_node);
1697 /* for each time step */
1698 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1699 int *non_branch_vars, *branch_vars;
1702 snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1703 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1704 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1707 /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1708 non_branch_vars = NEW_ARR_F(int, 0);
1709 branch_vars = NEW_ARR_F(int, 0);
1710 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1711 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
1712 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1715 if (cur_irn == cfop) {
1716 /* for all unit types available for this node */
1717 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1718 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1719 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1723 /* sum up all possible schedule points for this node upto current timestep */
1724 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1726 unsigned tmax = MIN(t, na->alap - 1);
1728 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1729 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1730 ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1737 if (ARR_LEN(non_branch_vars) > 0)
1738 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1739 if (ARR_LEN(branch_vars) > 0)
1740 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1742 DEL_ARR_F(branch_vars);
1743 DEL_ARR_F(non_branch_vars);
1747 DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1748 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1752 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1755 unsigned num_cst = 0;
1756 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1757 ir_timer_t *t_cst = ir_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1759 ilp_timer_push(t_cst);
1760 /* check all nodes */
1761 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1762 be_ilpsched_irn_t *node;
1763 ilpsched_node_attr_t *na;
1767 /* only mode_T nodes can have Projs and Keeps assigned */
1768 if (get_irn_mode(irn) != mode_T)
1771 node = get_ilpsched_irn(env, irn);
1772 na = get_ilpsched_node_attr(node);
1774 /* check if has some Projs and Keeps assigned */
1775 if (! na->projkeeps)
1778 /* we can run only once over the queue, so preserve the nodes */
1779 pk = NEW_ARR_F(ir_node *, 0);
1780 while (! waitq_empty(na->projkeeps))
1781 ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1782 del_waitq(na->projkeeps);
1783 na->projkeeps = NULL;
1785 /* for all time steps at which this node can be scheduled */
1786 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1788 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1790 /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1791 snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1792 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1793 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1796 /* sum up scheduling variables for this time step */
1797 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1798 int idx = ILPVAR_IDX(na, tp_idx, t);
1799 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1802 if (ARR_LEN(tmp_var_idx_n) > 0)
1803 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1804 DEL_ARR_F(tmp_var_idx_n);
1806 /* subtract all Proj and Keep variables for this step */
1807 for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1808 be_ilpsched_irn_t *pk_node = get_ilpsched_irn(env, pk[i]);
1809 ilpsched_node_attr_t *pk_na = get_ilpsched_node_attr(pk_node);
1812 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1813 int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1814 lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1820 DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1821 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1825 /***************************************************
1827 * |_ _| | | __ \ (_)
1828 * | | | | | |__) | _ __ ___ __ _ _ _ __
1829 * | | | | | ___/ | '_ ` _ \ / _` | | '_ \
1830 * _| |_| |____| | | | | | | | (_| | | | | |
1831 * |_____|______|_| |_| |_| |_|\__,_|_|_| |_|
1833 ***************************************************/
1836 * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1838 static void create_ilp(ir_node *block, void *walk_env) {
1839 be_ilpsched_env_t *env = walk_env;
1840 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
1841 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1842 FILE *logfile = NULL;
1845 struct obstack var_obst;
1848 DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1849 DBG((env->dbg, 255, " ILP Scheduling for %+F\n", block));
1850 DBG((env->dbg, 255, "=========================================\n\n"));
1852 DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1853 block, ba->n_interesting_nodes, ba->max_steps));
1855 /* notify backend and get block environment */
1856 env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1858 /* if we have less than two interesting nodes, there is no need to create the ILP */
1859 if (ba->n_interesting_nodes > 1) {
1860 double fact_var = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1861 double fact_cst = ba->n_interesting_nodes > 25 ? 3 : 4.5;
1862 int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes;
1863 int estimated_n_var = (int)((double)base_num * fact_var);
1864 int estimated_n_cst = (int)((double)base_num * fact_cst);
1866 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1867 estimated_n_var, estimated_n_cst));
1868 (void) estimated_n_var;
1870 /* set up the LPP object */
1871 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1873 lpp = new_lpp_userdef(
1876 estimated_n_cst, /* num vars */
1877 estimated_n_cst + 1, /* num cst */
1878 1.3); /* grow factor */
1879 obstack_init(&var_obst);
1881 /* create ILP variables */
1882 create_variables(env, lpp, block_node, &var_obst);
1884 /* create ILP constraints */
1885 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1886 create_assignment_and_precedence_constraints(env, lpp, block_node);
1887 create_ressource_constraints(env, lpp, block_node);
1888 create_bundle_constraints(env, lpp, block_node);
1889 create_branch_constraint(env, lpp, block_node);
1890 //create_proj_keep_constraints(env, lpp, block_node);
1892 if (env->opts->regpress) {
1893 create_alive_nodes_constraint(env, lpp, block_node);
1894 create_alive_livein_nodes_constraint(env, lpp, block_node);
1895 create_pressure_alive_constraint(env, lpp, block_node);
1898 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1900 /* debug stuff, dump lpp when debugging is on */
1902 if (firm_dbg_get_mask(env->dbg) > 1) {
1906 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1907 f = fopen(buf, "w");
1908 lpp_dump_plain(lpp, f);
1910 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1915 /* set solve time limit */
1916 lpp_set_time_limit(lpp, env->opts->time_limit);
1918 /* set logfile if requested */
1919 if (strlen(env->opts->log_file) > 0) {
1920 if (strcasecmp(env->opts->log_file, "stdout") == 0)
1921 lpp_set_log(lpp, stdout);
1922 else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1923 lpp_set_log(lpp, stderr);
1925 logfile = fopen(env->opts->log_file, "w");
1927 fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1929 lpp_set_log(lpp, logfile);
1934 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1939 /* check for valid solution */
1940 if (! lpp_is_sol_valid(lpp)) {
1945 if (firm_dbg_get_mask(env->dbg) >= 2) {
1946 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1947 f = fopen(buf, "w");
1948 lpp_dump_plain(lpp, f);
1950 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1952 dump_ir_block_graph(env->irg, "-infeasible");
1956 ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1960 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1961 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1962 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1963 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));
1964 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1965 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1966 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1967 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1969 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1972 /* apply solution */
1973 be_stat_ev("nodes", ba->block_last_idx);
1974 be_stat_ev("vars", lpp ? lpp->var_next : 0);
1975 be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1977 be_stat_ev("time", -1);
1978 be_stat_ev_dbl("opt", 0.0);
1979 list_sched_single_block(env->birg, block, env->be_opts);
1983 double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1984 be_stat_ev_dbl("time", lpp->sol_time);
1985 be_stat_ev_dbl("opt", opt);
1988 be_stat_ev_dbl("time", 0.0);
1989 be_stat_ev_dbl("opt", 100.0);
1991 apply_solution(env, lpp, block);
1997 /* notify backend */
1998 be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2002 * Perform ILP scheduling on the given irg.
2004 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2005 be_ilpsched_env_t env;
2006 const char *name = "be ilp scheduling";
2007 ir_graph *irg = be_get_birg_irg(birg);
2008 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
2009 const ilp_sched_selector_t *sel = arch_env->impl->get_ilp_sched_selector(arch_env);
2011 FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2013 stat_ev_ctx_push("ilpsched");
2015 // firm_dbg_set_mask(env.dbg, 1);
2017 env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2020 env.height = heights_new(irg);
2021 env.main_env = birg->main_env;
2022 env.arch_env = arch_env;
2023 env.cpu = arch_env_get_machine(arch_env);
2024 env.opts = &ilp_opts;
2026 env.be_opts = be_opts;
2027 phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2029 /* assign a unique per block number to all interesting nodes */
2030 irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2033 The block indices are completely build after the walk,
2034 now we can allocate the bitsets (size depends on block indices)
2037 phase_reinit_irn_data(&env.ph);
2039 /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2040 irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2042 /* Calculate ALAP of all irns */
2043 irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2045 /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2046 irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2048 /* perform ILP scheduling */
2049 irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2052 if (firm_dbg_get_mask(env.dbg)) {
2054 phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2056 fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2060 /* free data allocated dynamically */
2061 irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2063 /* free all allocated object */
2064 phase_free(&env.ph);
2065 heights_free(env.height);
2067 /* notify backend */
2068 be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2070 stat_ev_ctx_pop("ilpsched");
2074 * Register ILP scheduler options.
2076 void be_init_ilpsched(void)
2078 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2079 lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2081 lc_opt_add_table(sched_grp, ilpsched_option_table);
2084 #else /* WITH_ILP */
2086 static inline void some_picky_compiler_do_not_allow_empty_files(void)
2089 #endif /* WITH_ILP */