2 * Copyright (C) 1995-2007 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"
61 #include <lpp/lpp_net.h>
63 #include <libcore/lc_opts.h>
64 #include <libcore/lc_opts_enum.h>
65 #include <libcore/lc_timing.h>
69 #include "besched_t.h"
70 #include "beilpsched.h"
74 #include "benodesets.h"
76 typedef struct _ilpsched_options_t {
82 typedef struct _unit_type_info_t {
84 const be_execution_unit_type_t *tp;
88 * holding the ILP variables of the different types
90 typedef struct _ilp_var_types_t {
91 int *x; /* x_{nt}^k variables */
92 int *a; /* a_{nt}^k variables */
93 int *y; /* y_{nt}^k variables */
97 * Holds alive variables for a node live-in to a block.
99 typedef struct _ilp_livein_node_t {
101 unsigned max_alive_steps;
105 /* attributes for a node */
106 typedef struct _ilpsched_node_attr_t {
107 unsigned asap; /**< The ASAP scheduling control step */
108 unsigned alap; /**< The ALAP scheduling control step */
109 unsigned latency; /**< Latency of this node (needed for sorting) */
110 unsigned sched_point; /**< the step in which the node is finally scheduled */
111 unsigned visit_idx; /**< Index of the node having visited this node last */
112 unsigned consumer_idx; /**< Index of the node having counted this node as consumer last */
113 unsigned n_consumer; /**< Number of consumers */
114 ir_node **block_consumer; /**< List of consumer being in the same block */
115 waitq *projkeeps; /**< A List of Projs and Keeps belonging to this node */
116 unsigned block_idx : 30; /**< A unique per block index */
117 unsigned alap_changed : 1; /**< the current ALAP has changed, revisit preds */
118 unsigned is_dummy_node : 1; /**< this node is assigned to DUMMY unit */
119 bitset_t *transitive_block_nodes; /**< Set of transitive block nodes (predecessors
120 for ASAP, successors for ALAP */
121 unsigned n_unit_types; /**< number of allowed execution unit types */
122 unit_type_info_t *type_info; /**< list of allowed execution unit types */
123 ilp_var_types_t ilp_vars; /**< the different ILP variables */
124 } ilpsched_node_attr_t;
126 /* attributes for a block */
127 typedef struct _ilpsched_block_attr_t {
128 unsigned block_last_idx; /**< The highest node index in block so far */
129 unsigned n_interesting_nodes; /**< The number of nodes interesting for scheduling */
130 unsigned max_steps; /**< Upper bound for block execution */
131 plist_t *root_nodes; /**< A list of nodes having no user in current block */
132 ir_node *head_ilp_nodes; /**< A linked list of nodes which will contribute to ILP */
133 pset *livein_nodes; /**< A set of nodes which are live-in to this block */
134 } ilpsched_block_attr_t;
136 typedef union _ilpsched_attr_ {
137 ilpsched_node_attr_t node_attr;
138 ilpsched_block_attr_t block_attr;
141 /* A irn for the phase and it's attributes (either node or block) */
144 ilpsched_attr_t attr;
147 /* The ILP scheduling environment */
149 ir_phase ph; /**< The phase */
150 ir_graph *irg; /**< The current irg */
151 heights_t *height; /**< The heights object of the irg */
152 void *irg_env; /**< An environment for the irg scheduling, provided by the backend */
153 void *block_env; /**< An environment for scheduling a block, provided by the backend */
154 const arch_env_t *arch_env;
155 const arch_isa_t *isa; /**< The ISA */
156 const be_main_env_t *main_env;
157 const be_machine_t *cpu; /**< the current abstract machine */
158 ilpsched_options_t *opts; /**< the ilp options for current irg */
159 const be_irg_t *birg; /**< The birg object */
160 be_options_t *be_opts; /**< backend options */
161 const ilp_sched_selector_t *sel; /**< The ILP sched selector provided by the backend */
162 DEBUG_ONLY(firm_dbg_module_t *dbg);
165 /* convenience macros to handle phase irn data */
166 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
167 #define is_ilpsched_block(node) (is_Block((node)->irn))
168 #define get_ilpsched_block_attr(block) (&(block)->attr.block_attr)
169 #define get_ilpsched_node_attr(node) (&(node)->attr.node_attr)
171 /* check if node is considered for ILP scheduling */
172 #define consider_for_sched(isa, irn) \
173 (! (is_Block(irn) || \
174 is_normal_Proj(isa, irn) || \
181 /* gives the valid scheduling time step interval for a node */
182 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
184 /* gives the valid interval where a node can die */
185 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
187 /* gives the corresponding ILP variable for given node, unit and time step */
188 #define ILPVAR_IDX(na, unit, control_step) \
189 ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
191 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
192 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
193 ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
195 /* check if a double value is within an epsilon environment of 0 */
196 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
198 #define ilp_timer_push(t) lc_timer_push((t))
199 #define ilp_timer_pop() lc_timer_pop()
200 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
202 /* option variable */
203 static ilpsched_options_t ilp_opts = {
204 1, /* default is with register pressure constraints */
205 300, /* 300 sec per block time limit */
210 static const lc_opt_table_entry_t ilpsched_option_table[] = {
211 LC_OPT_ENT_BOOL("regpress", "Use register pressure constraints", &ilp_opts.regpress),
212 LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
213 LC_OPT_ENT_STR("lpp_log", "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
218 We need this global variable as we compare nodes dependent on heights,
219 but we cannot pass any information to the qsort compare function.
221 static heights_t *glob_heights;
224 * Check if irn is a Proj, which has no execution units assigned.
225 * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
227 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
228 return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
232 * Skips normal Projs.
233 * @return predecessor if irn is a normal Proj, otherwise irn.
235 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
236 if (is_normal_Proj(isa, irn))
237 return get_Proj_pred(irn);
241 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
242 unsigned lat = be_ilp_sched_latency(sel, irn, env);
243 if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
248 static int cmp_live_in_nodes(const void *a, const void *b) {
249 const ilp_livein_node_t *n1 = a;
250 const ilp_livein_node_t *n2 = b;
252 return n1->irn != n2->irn;
256 * Compare scheduling time steps of two be_ilpsched_irn's.
258 static int cmp_ilpsched_irn(const void *a, const void *b) {
259 be_ilpsched_irn_t *n1 = *(be_ilpsched_irn_t **)a;
260 be_ilpsched_irn_t *n2 = *(be_ilpsched_irn_t **)b;
261 ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
262 ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
264 if (n1_a->sched_point == n2_a->sched_point) {
265 ir_node *irn_a = n1->irn;
266 ir_node *irn_b = n2->irn;
268 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
270 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
274 Ok, timestep is equal and the nodes are parallel,
275 so check latency and schedule high latency first.
277 return QSORT_CMP(n2_a->latency, n1_a->latency);
280 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
284 * In case there is no phase information for irn, initialize it.
286 static void *init_ilpsched_irn(ir_phase *ph, ir_node *irn, void *old) {
287 be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
290 /* if we have already some data: check for reinitialization */
292 if (! is_Block(irn)) {
293 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
295 if (! na->transitive_block_nodes) {
296 ir_node *block = get_nodes_block(irn);
297 be_ilpsched_irn_t *block_node = phase_get_or_set_irn_data(ph, block);
298 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
300 /* we are called after the block indices have been build: create bitset */
301 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
304 /* we are called from reinit block data: clear the bitset */
305 bitset_clear_all(na->transitive_block_nodes);
307 na->alap_changed = 1;
315 /* set ilpsched irn attributes (either block or irn) */
317 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
319 ba->n_interesting_nodes = 0;
320 ba->block_last_idx = 0;
321 ba->root_nodes = plist_new();
322 ba->head_ilp_nodes = NULL;
323 ba->livein_nodes = new_pset(cmp_live_in_nodes, 16);
327 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
328 memset(na, 0, sizeof(*na));
335 * Assign a per block unique number to each node.
337 static void build_block_idx(ir_node *irn, void *walk_env) {
338 be_ilpsched_env_t *env = walk_env;
339 be_ilpsched_irn_t *node, *block_node;
340 ilpsched_node_attr_t *na;
341 ilpsched_block_attr_t *ba;
343 set_irn_link(irn, NULL);
344 if (! consider_for_sched(env->arch_env->isa, irn))
347 node = get_ilpsched_irn(env, irn);
348 na = get_ilpsched_node_attr(node);
349 block_node = get_ilpsched_irn(env, get_nodes_block(irn));
350 ba = get_ilpsched_block_attr(block_node);
352 na->block_idx = ba->block_last_idx++;
355 /********************************************************
358 * __ _ ___ __ _ _ __ / / __ _| | __ _ _ __
359 * / _` / __|/ _` | '_ \ / / / _` | |/ _` | '_ \
360 * | (_| \__ \ (_| | |_) | / / | (_| | | (_| | |_) |
361 * \__,_|___/\__,_| .__/ /_/ \__,_|_|\__,_| .__/
364 ********************************************************/
367 * Add all nodes having no user in current block to last_nodes list.
369 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
371 const ir_edge_t *edge;
372 be_ilpsched_irn_t *block_node, *node;
373 ilpsched_block_attr_t *ba;
374 ilpsched_node_attr_t *na;
376 be_ilpsched_env_t *env = walk_env;
377 int has_block_user = 0;
378 unsigned n_consumer = 0;
379 ir_edge_kind_t ekind[2] = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
383 if (! consider_for_sched(env->arch_env->isa, irn))
386 block = get_nodes_block(irn);
387 idx = get_irn_idx(irn);
388 consumer = NEW_ARR_F(ir_node *, 0);
390 DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
392 /* check data and dependency out edges */
393 for (i = 0; i < 2 && ! has_block_user; ++i) {
394 foreach_out_edge_kind(irn, edge, ekind[i]) {
395 ir_node *user = get_edge_src_irn(edge);
397 if (is_normal_Proj(env->arch_env->isa, user)) {
398 const ir_edge_t *user_edge;
400 if (get_irn_mode(user) == mode_X)
403 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
404 for (j = 0; j < 2; ++j) {
405 foreach_out_edge_kind(user, user_edge, ekind[j]) {
406 ir_node *real_user = get_edge_src_irn(user_edge);
408 if (! is_Phi(real_user) && ! is_Block(real_user)) {
409 be_ilpsched_irn_t *node = get_ilpsched_irn(env, real_user);
410 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
412 /* skip already visited nodes */
413 if (ua->consumer_idx == idx)
416 /* check if node has user in this block and collect the user if it's a data user */
417 if (get_nodes_block(real_user) == block) {
418 if (i == 0 && j == 0)
419 ARR_APP1(ir_node *, consumer, real_user);
423 /* only count data consumer */
427 /* mark user as visited by this node */
428 ua->consumer_idx = idx;
433 else if (is_Block(user)) {
436 else if (! is_Phi(user)) {
437 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
438 ilpsched_node_attr_t *ua = get_ilpsched_node_attr(node);
440 /* skip already visited nodes */
441 if (ua->consumer_idx == idx)
444 /* check if node has user in this block and collect the user if it's a data user */
445 if (get_nodes_block(user) == block) {
447 ARR_APP1(ir_node *, consumer, user);
451 /* only count data consumer */
455 /* mark user visited by this node */
456 ua->consumer_idx = idx;
458 else if (get_nodes_block(user) != block) {
464 block_node = get_ilpsched_irn(env, block);
465 ba = get_ilpsched_block_attr(block_node);
467 ba->n_interesting_nodes++;
469 /* current irn has no user inside this block, add to queue */
470 if (! has_block_user) {
471 DB((env->dbg, LEVEL_3, "root node\n"));
472 plist_insert_back(ba->root_nodes, irn);
475 DB((env->dbg, LEVEL_3, "normal node\n"));
478 /* record number of all consumer and the consumer within the same block */
479 node = get_ilpsched_irn(env, irn);
480 na = get_ilpsched_node_attr(node);
481 na->n_consumer = n_consumer;
482 na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
483 memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
488 * Calculate the ASAP scheduling step for current irn.
490 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
491 be_ilpsched_env_t *env = walk_env;
494 be_ilpsched_irn_t *node, *block_node;
495 ilpsched_node_attr_t *na;
496 ilpsched_block_attr_t *ba;
498 /* These nodes are handled separate */
499 if (! consider_for_sched(env->arch_env->isa, irn))
502 DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
504 block = get_nodes_block(irn);
505 node = get_ilpsched_irn(env, irn);
506 na = get_ilpsched_node_attr(node);
509 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
510 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
512 /* check for greatest distance to top */
513 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
514 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
515 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
518 lat = fixed_latency(env->sel, pred, env->block_env);
520 na->asap = MAX(na->asap, pna->asap + lat);
524 /* add node to ILP node list and update max_steps */
525 block_node = get_ilpsched_irn(env, block);
526 ba = get_ilpsched_block_attr(block_node);
528 set_irn_link(irn, ba->head_ilp_nodes);
529 ba->head_ilp_nodes = irn;
530 ba->max_steps += fixed_latency(env->sel, irn, env->block_env);
532 DB((env->dbg, LEVEL_2, "%u\n", na->asap));
536 * Calculate the ALAP scheduling step of all irns in current block.
537 * Depends on max_steps being calculated.
539 static void calculate_block_alap(ir_node *block, void *walk_env) {
540 be_ilpsched_env_t *env = walk_env;
541 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
542 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
543 waitq *cur_queue = new_waitq();
546 assert(is_Block(block));
548 DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
549 block, ba->n_interesting_nodes, ba->max_steps));
551 /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
552 //phase_reinit_block_irn_data(&env->ph, block);
554 /* init start queue */
555 foreach_plist(ba->root_nodes, el) {
556 waitq_put(cur_queue, plist_element_get_value(el));
559 /* repeat until all nodes are processed */
560 while (! waitq_empty(cur_queue)) {
561 waitq *next_queue = new_waitq();
563 /* process all nodes in current step */
564 while (! waitq_empty(cur_queue)) {
565 ir_node *cur_irn = waitq_get(cur_queue);
566 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
567 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
570 /* cur_node has no alap set -> it's a root node, set to max alap */
572 na->alap = ba->max_steps;
573 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
577 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
581 /* set the alap's of all predecessors */
582 for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
583 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
585 /* check for greatest distance to bottom */
586 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
587 be_ilpsched_irn_t *pred_node = get_ilpsched_irn(env, pred);
588 ilpsched_node_attr_t *pna = get_ilpsched_node_attr(pred_node);
591 /* mark the predecessor as visited by current irn */
592 if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
594 pna->visit_idx = get_irn_idx(cur_irn);
596 lat = fixed_latency(env->sel, pred, env->block_env);
598 /* set ALAP of current pred */
599 if (pna->alap == 0) {
600 /* current ALAP is 0: node has not yet been visited */
601 pna->alap_changed = 1;
602 pna->alap = na->alap - lat;
604 else if (pna->alap > na->alap - lat) {
605 /* we found a longer path to root node: change ALAP */
606 pna->alap = na->alap - lat;
607 pna->alap_changed = 1;
610 /* current ALAP is best found so far: keep it */
611 pna->alap_changed = 0;
614 DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
616 /* enqueue node for next iteration */
617 if (get_irn_ins_or_deps(pred) > 0)
618 waitq_put(next_queue, pred);
623 /* prepare for next iteration */
624 del_waitq(cur_queue);
625 cur_queue = next_queue;
630 * Free list of root nodes and the set of live-in nodes.
632 static void clear_unwanted_data(ir_node *block, void *walk_env) {
633 be_ilpsched_env_t *env = walk_env;
634 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
635 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
637 plist_free(ba->root_nodes);
638 ba->root_nodes = NULL;
639 del_pset(ba->livein_nodes);
640 ba->livein_nodes = NULL;
644 * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
645 * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
647 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
648 be_ilpsched_env_t *env = walk_env;
650 be_ilpsched_irn_t *node, *pred_node;
651 ilpsched_node_attr_t *na, *pna;
653 if (! consider_for_sched(env->arch_env->isa, irn))
656 if (! is_Proj(irn) && ! be_is_Keep(irn))
659 /* go to the ancestor */
661 pred = get_irn_n(irn, 0);
662 pred = skip_Proj(pred);
664 node = get_ilpsched_irn(env, irn);
665 pred_node = get_ilpsched_irn(env, pred);
666 na = get_ilpsched_node_attr(node);
667 pna = get_ilpsched_node_attr(pred_node);
669 na->asap = pna->asap;
670 na->alap = pna->alap;
672 /* record all Projs and Keeps for this node */
673 if (! pna->projkeeps)
674 pna->projkeeps = new_waitq();
675 waitq_put(pna->projkeeps, irn);
677 DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
680 /*******************************************
683 * ___ ___| |__ ___ __| |_ _| | ___
684 * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
685 * \__ \ (__| | | | __/ (_| | |_| | | __/
686 * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
688 *******************************************/
690 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
691 const ir_edge_t *edge;
694 foreach_out_edge(irn, edge) {
695 ir_node *user = get_edge_src_irn(edge);
697 if (be_is_Keep(user)) {
698 assert(get_nodes_block(user) == block && "Keep must not be in different block.");
699 waitq_put(keeps, user);
705 * Inserts @p irn before @p before into schedule and notifies backend.
707 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
708 ir_node *before, ir_node *irn, unsigned cycle)
710 be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
711 sched_add_before(before, irn);
715 * Adds a node, it's Projs (in case of mode_T nodes) and
716 * it's Keeps to schedule.
718 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
719 const ir_edge_t *edge;
720 waitq *keeps = new_waitq();
722 /* mode_M nodes are not scheduled */
723 if (get_irn_mode(irn) == mode_M)
726 if (! sched_is_scheduled(irn))
727 notified_sched_add_before(env, block, irn, cycle);
730 if (get_irn_mode(irn) == mode_T) {
731 foreach_out_edge(irn, edge) {
732 ir_node *user = get_edge_src_irn(edge);
734 if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
735 get_irn_n_edges(user) > 0)
737 notified_sched_add_before(env, block, user, cycle);
740 check_for_keeps(keeps, block, user);
744 check_for_keeps(keeps, block, irn);
748 while (! waitq_empty(keeps)) {
749 ir_node *keep = waitq_get(keeps);
750 if (! sched_is_scheduled(keep))
751 notified_sched_add_before(env, block, keep, cycle);
758 * Schedule all nodes in the given block, according to the ILP solution.
760 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
761 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
762 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
763 sched_info_t *info = get_irn_sched_info(block);
764 be_ilpsched_irn_t **sched_nodes;
767 const ir_edge_t *edge;
769 /* init block schedule list */
770 INIT_LIST_HEAD(&info->list);
773 /* collect nodes and their scheduling time step */
774 sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
775 if (ba->n_interesting_nodes == 0) {
778 else if (ba->n_interesting_nodes == 1) {
779 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
781 /* add the single node */
782 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
785 /* check all nodes for their positive solution */
786 foreach_linked_irns(ba->head_ilp_nodes, irn) {
787 be_ilpsched_irn_t *node;
788 ilpsched_node_attr_t *na;
792 node = get_ilpsched_irn(env, irn);
793 na = get_ilpsched_node_attr(node);
797 if (! na->is_dummy_node) {
798 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
799 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
800 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
802 if (! LPP_VALUE_IS_0(cost)) {
803 DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
811 /* go over all variables of a node until the non-zero one is found */
812 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
813 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
814 double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
816 /* check, if variable is set to one (it's not zero then :) */
817 if (! LPP_VALUE_IS_0(val)) {
819 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
820 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
821 irn, t, na->type_info[tp_idx].tp->name));
828 glob_heights = env->height;
829 /* sort nodes ascending by scheduling time step */
830 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
833 /* make all Phis ready and remember the single cf op */
835 foreach_out_edge(block, edge) {
836 irn = get_edge_src_irn(edge);
838 switch (get_irn_opcode(irn)) {
840 add_to_sched(env, block, irn, 0);
850 assert(cfop == NULL && "Highlander - there can be only one");
857 /* add all nodes from list */
858 for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
859 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
860 if (sched_nodes[i]->irn != cfop)
861 add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
864 /* schedule control flow node if not already done */
865 if (cfop && ! sched_is_scheduled(cfop))
866 add_to_sched(env, block, cfop, 0);
868 DEL_ARR_F(sched_nodes);
871 /***************************************************************
872 * _____ _ _____ _____ _ _
873 * |_ _| | | __ \ / ____| | | (_)
874 * | | | | | |__) | | (___ ___ ___| |_ _ ___ _ __
875 * | | | | | ___/ \___ \ / _ \/ __| __| |/ _ \| '_ \
876 * _| |_| |____| | ____) | __/ (__| |_| | (_) | | | |
877 * |_____|______|_| |_____/ \___|\___|\__|_|\___/|_| |_|
879 ***************************************************************/
882 * Check if node can be executed on given unit type.
884 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
886 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
888 for (i = na->n_unit_types - 1; i >= 0; --i) {
889 if (na->type_info[i].tp == tp)
896 /************************************************
899 * __ ____ _ _ __ _ __ _| |__ | | ___ ___
900 * \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
901 * \ V / (_| | | | | (_| | |_) | | __/\__ \
902 * \_/ \__,_|_| |_|\__,_|_.__/|_|\___||___/
904 ************************************************/
906 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
907 const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
908 unsigned n_unit_types = 0;
909 be_ilpsched_irn_t *node;
910 ilpsched_node_attr_t *na;
911 unsigned unit_idx, tp_idx;
913 /* count number of available unit types for this node */
914 for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
917 node = get_ilpsched_irn(env, irn);
918 na = get_ilpsched_node_attr(node);
920 if (! na->type_info) {
921 na->n_unit_types = n_unit_types;
922 na->type_info = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
924 /* fill the type info array */
925 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
926 for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
927 /* beware: we also count number of available units here */
928 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
929 na->is_dummy_node = 1;
932 na->type_info[tp_idx].tp = execunits[tp_idx][0]->tp;
933 na->type_info[tp_idx].n_units = unit_idx;
941 * Returns the largest alap time of a user of @p irn.
942 * The user must be in block @p block.
944 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
945 const ir_edge_t *edge;
946 unsigned max_alap = 0;
948 foreach_out_edge(irn, edge) {
949 ir_node *user = get_edge_src_irn(edge);
951 if (get_nodes_block(user) == block) {
952 be_ilpsched_irn_t *node = get_ilpsched_irn(env, user);
953 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
955 max_alap = MAX(max_alap, na->alap);
959 assert(max_alap > 0);
964 * Create the following variables:
965 * - x_{nt}^k binary weigthed with: t
966 * node n is scheduled at time step t to unit type k
967 * ==>> These variables represent the schedule
969 * - a_{nt}^k binary weighted with num_nodes
970 * node n is alive at time step t on unit type k
972 * - y_{nt}^k continuous weighted with: num_nodes^2
973 * register pressure over limit for unit type k
974 * ==>> These variables represent the register pressure
977 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
980 unsigned num_block_var, num_nodes;
981 ilp_livein_node_t *livein;
982 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
983 unsigned weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
984 lc_timer_t *t_var = lc_timer_register("beilpsched_var", "create ilp variables");
986 ilp_timer_push(t_var);
987 num_block_var = num_nodes = 0;
988 foreach_linked_irns(ba->head_ilp_nodes, irn) {
989 be_ilpsched_irn_t *node;
990 ilpsched_node_attr_t *na;
991 unsigned n_unit_types, tp_idx, n_var, cur_unit;
992 unsigned cur_var_ad, cur_var_x, cur_var_y, num_ad;
995 node = get_ilpsched_irn(env, irn);
996 na = get_ilpsched_node_attr(node);
997 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
999 /* allocate space for ilp variables */
1000 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1001 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
1003 /* we need these variables only for "real" nodes */
1004 if (! na->is_dummy_node) {
1005 na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1006 memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1008 num_ad = ba->max_steps - na->asap + 1;
1009 na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1010 memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1013 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1014 irn, na->asap, na->alap, na->n_unit_types));
1016 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1017 /* create variables */
1018 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1021 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1022 /* x_{nt}^k variables */
1023 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1024 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1025 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1026 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1027 /* variable counter */
1031 if (! na->is_dummy_node) {
1032 /* y_{nt}^k variables */
1033 snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1034 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1035 na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1036 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1038 /* variable counter */
1044 /* a node can die at any step t: asap(n) <= t <= U */
1045 if (! na->is_dummy_node) {
1046 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1048 /* a_{nt}^k variables */
1049 snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1050 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1051 na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1052 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1054 /* variable counter */
1060 /* collect live-in nodes */
1061 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1062 ir_node *pred = get_irn_n(irn, i);
1064 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1065 be_ilpsched_set_type_info(env, pred, var_obst);
1066 if (! na->is_dummy_node) {
1067 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1070 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1076 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1080 /* create alive variables a_{nt}^k for live-ins */
1081 foreach_pset(ba->livein_nodes, livein) {
1082 be_ilpsched_irn_t *node;
1083 ilpsched_node_attr_t *na;
1084 unsigned tp_idx, var_idx;
1088 node = get_ilpsched_irn(env, irn);
1089 na = get_ilpsched_node_attr(node);
1091 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1093 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1096 /* create variables */
1097 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1100 for (t = 0; t < livein->max_alive_steps; ++t) {
1101 /* a_{nt}^k variables */
1102 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1103 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1104 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1105 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1112 DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1113 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1116 /*******************************************************
1119 * ___ ___ _ __ ___| |_ _ __ __ _ _ _ __ | |_ ___
1120 * / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1121 * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1122 * \___\___/|_| |_|___/\__|_| \__,_|_|_| |_|\__|___/
1124 *******************************************************/
1127 * Collect all operands and nodes @p irn depends on.
1128 * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1130 static nodeset *sta_collect_in_deps(ir_node *irn, nodeset *deps) {
1133 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1134 ir_node *p = get_irn_in_or_dep(irn, i);
1137 const ir_edge_t *edge;
1139 p = get_Proj_pred(p);
1140 foreach_out_edge(p, edge) {
1141 ir_node *src = get_edge_src_irn(edge);
1142 nodeset_insert(deps, src);
1146 nodeset_insert(deps, p);
1154 * Create following ILP constraints:
1155 * - the assignment constraints:
1156 * assure each node is executed once by exactly one (allowed) execution unit
1157 * - the dead node assignment constraints:
1158 * assure a node can only die at most once
1159 * - the precedence constraints:
1160 * assure that no data dependencies are violated
1162 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1163 unsigned num_cst_assign, num_cst_prec, num_cst_dead;
1166 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1167 bitset_t *bs_block_irns = bitset_alloca(ba->block_last_idx);
1168 lc_timer_t *t_cst_assign = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1169 lc_timer_t *t_cst_prec = lc_timer_register("beilpsched_cst_prec", "create precedence constraints");
1171 num_cst_assign = num_cst_prec = num_cst_dead = 0;
1172 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1175 be_ilpsched_irn_t *node;
1176 ilpsched_node_attr_t *na;
1178 nodeset *deps = new_nodeset(16);
1180 node = get_ilpsched_irn(env, irn);
1181 na = get_ilpsched_node_attr(node);
1184 /* the assignment constraint */
1185 ilp_timer_push(t_cst_assign);
1186 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1187 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1188 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1191 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1194 /* We have separate constraints for Projs and Keeps */
1195 // ILP becomes infeasible ?!?
1196 // if (is_Proj(irn) || be_is_Keep(irn))
1199 /* the precedence constraints */
1200 ilp_timer_push(t_cst_prec);
1201 bs_block_irns = bitset_clear_all(bs_block_irns);
1203 deps = sta_collect_in_deps(irn, deps);
1204 foreach_nodeset(deps, pred) {
1205 unsigned t_low, t_high, t;
1206 be_ilpsched_irn_t *pred_node;
1207 ilpsched_node_attr_t *pna;
1210 pred = skip_normal_Proj(env->arch_env->isa, pred);
1211 if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1214 pred_node = get_ilpsched_irn(env, pred);
1215 pna = get_ilpsched_node_attr(pred_node);
1217 assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1219 if (! bitset_is_set(bs_block_irns, pna->block_idx))
1220 bitset_set(bs_block_irns, pna->block_idx);
1224 /* irn = n, pred = m */
1225 delay = fixed_latency(env->sel, pred, env->block_env);
1226 t_low = MAX(na->asap, pna->asap + delay - 1);
1227 t_high = MIN(na->alap, pna->alap + delay - 1);
1228 for (t = t_low - 1; t <= t_high - 1; ++t) {
1230 int *tmp_var_idx = NEW_ARR_F(int, 0);
1232 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1233 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1234 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1237 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1238 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1239 /* node variables have smaller index than pred variables */
1240 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1241 for (tn = na->asap - 1; tn <= t; ++tn) {
1242 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1243 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1247 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1248 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1249 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1250 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1255 /* pred variables have smaller index than node variables */
1256 for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1257 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1258 unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1259 ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1263 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1264 for (tn = na->asap - 1; tn <= t; ++tn) {
1265 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1266 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1271 if (ARR_LEN(tmp_var_idx) > 0)
1272 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1274 DEL_ARR_F(tmp_var_idx);
1280 DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1281 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1282 DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1283 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1287 * Create ILP resource constraints:
1288 * - assure that for each time step not more instructions are scheduled
1289 * to the same unit types as units of this type are available
1291 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1294 unsigned num_cst_resrc = 0;
1295 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1296 lc_timer_t *t_cst_rsrc = lc_timer_register("beilpsched_cst_rsrc", "create resource constraints");
1298 ilp_timer_push(t_cst_rsrc);
1299 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1301 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1303 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1305 /* check each time step */
1306 for (t = 0; t < ba->max_steps; ++t) {
1309 int *tmp_var_idx = NEW_ARR_F(int, 0);
1311 snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1312 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1313 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1316 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1317 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1318 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1321 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1323 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1324 int cur_var = ILPVAR_IDX(na, tp_idx, t);
1325 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1329 /* set constraints if we have some */
1330 if (ARR_LEN(tmp_var_idx) > 0)
1331 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1333 DEL_ARR_F(tmp_var_idx);
1337 DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1338 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1342 * Create ILP bundle constraints:
1343 * - assure, at most bundle_size * bundles_per_cycle instructions
1344 * can be started at a certain point.
1346 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1349 unsigned num_cst_bundle = 0;
1350 unsigned n_instr_max = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1351 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1352 lc_timer_t *t_cst_bundle = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1354 ilp_timer_push(t_cst_bundle);
1355 for (t = 0; t < ba->max_steps; ++t) {
1358 int *tmp_var_idx = NEW_ARR_F(int, 0);
1360 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1361 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1362 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1365 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1366 be_ilpsched_irn_t *node;
1367 ilpsched_node_attr_t *na;
1370 /* Projs and Keeps do not contribute to bundle size */
1371 if (is_Proj(irn) || be_is_Keep(irn))
1374 node = get_ilpsched_irn(env, irn);
1375 na = get_ilpsched_node_attr(node);
1377 /* nodes assigned to DUMMY unit do not contribute to bundle size */
1378 if (na->is_dummy_node)
1381 if (t >= na->asap - 1 && t <= na->alap - 1) {
1382 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1383 int idx = ILPVAR_IDX(na, tp_idx, t);
1384 ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1389 if (ARR_LEN(tmp_var_idx) > 0)
1390 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1392 DEL_ARR_F(tmp_var_idx);
1395 DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1396 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1400 * Create ILP alive nodes constraints:
1401 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1403 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1406 unsigned num_cst = 0;
1407 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1408 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1410 ilp_timer_push(t_cst);
1412 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1413 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1414 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1417 /* we ignore nodes assigned to dummy unit here */
1418 if (na->is_dummy_node)
1421 /* check check all time steps: asap(n) <= t <= U */
1422 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1425 /* for all unit types available for this node */
1426 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1427 unsigned tn, tn_max, idx;
1429 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1430 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1432 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1433 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1434 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1435 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1438 tn_max = MIN(na->alap - 1, t);
1439 /* check if the node has been scheduled so far */
1440 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1441 int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1442 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1445 if (ARR_LEN(tmp_var_idx_n) > 0)
1446 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1447 DEL_ARR_F(tmp_var_idx_n);
1449 /* subtract the number of consumer scheduled so far */
1450 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1451 be_ilpsched_irn_t *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1452 ilpsched_node_attr_t *ca = get_ilpsched_node_attr(cons);
1454 unsigned tm, tm_max;
1456 tm_max = MIN(ca->alap - 1, t);
1457 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1458 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1459 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1460 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1465 if (ARR_LEN(tmp_var_idx_m) > 0)
1466 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1467 DEL_ARR_F(tmp_var_idx_m);
1470 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1471 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1477 DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1478 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1482 * Create ILP alive nodes constraints for live-in nodes:
1483 * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1485 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1487 ilp_livein_node_t *livein;
1488 unsigned num_cst = 0;
1489 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1490 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1492 ilp_timer_push(t_cst);
1494 foreach_pset(ba->livein_nodes, livein) {
1495 ir_node *irn = livein->irn;
1496 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1497 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1500 /* check check all time steps: 0 <= t < max_alive_steps */
1501 for (t = 0; t < livein->max_alive_steps; ++t) {
1504 /* for all unit types available for this node */
1505 for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1506 const ir_edge_t *edge;
1508 int cst, num_block_user;
1509 int *tmp_var_idx_m = NEW_ARR_F(int, 0);
1511 /* check the number of consumer scheduled so far */
1513 foreach_out_edge(irn, edge) {
1514 ir_node *user = get_edge_src_irn(edge);
1515 be_ilpsched_irn_t *cons;
1516 ilpsched_node_attr_t *ca;
1518 unsigned tm, tm_max;
1520 /* check only users within current block */
1521 if (get_nodes_block(user) != block_node->irn)
1525 cons = get_ilpsched_irn(env, user);
1526 ca = get_ilpsched_node_attr(cons);
1528 tm_max = MIN(ca->alap - 1, t);
1529 for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1530 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1531 int idx = ILPVAR_IDX(ca, tp_idx, tm);
1532 ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1537 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1538 t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1539 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1540 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1543 /* sum(scheduled users) */
1544 if (ARR_LEN(tmp_var_idx_m) > 0)
1545 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1546 DEL_ARR_F(tmp_var_idx_m);
1548 /* + c * a_{nt}^k */
1549 idx = node_tp_idx * livein->max_alive_steps + t;
1550 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1555 DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1556 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1560 * Create ILP pressure constraints, based on alive nodes:
1561 * - add additional costs to objective function if a node is scheduled
1562 * on a unit although all units of this type are currently occupied
1564 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1567 unsigned num_cst = 0;
1568 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1569 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1571 ilp_timer_push(t_cst);
1572 /* y_{nt}^k is set for each node and timestep and unit type */
1573 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1574 unsigned cur_idx = get_irn_idx(cur_irn);
1575 be_ilpsched_irn_t *cur_node = get_ilpsched_irn(env, cur_irn);
1576 ilpsched_node_attr_t *cur_na = get_ilpsched_node_attr(cur_node);
1579 /* we ignore nodes assigned to DUMMY unit here */
1580 if (cur_na->is_dummy_node)
1584 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1585 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1589 /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1591 /* check if node can be executed on this unit type */
1592 cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1596 /* check all time_steps at which the current node can be scheduled */
1597 for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1600 int *tmp_var_idx = NEW_ARR_F(int, 0);
1601 ilp_livein_node_t *livein;
1603 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1604 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1605 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1608 /* - accumulate all nodes alive at point t on unit type k */
1609 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1610 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1611 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1614 /* check if node can be alive here */
1615 if (t < na->asap - 1)
1618 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1620 /* current type is not suitable */
1624 a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1625 ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1627 /* do the same for livein nodes */
1628 foreach_pset(ba->livein_nodes, livein) {
1629 ir_node *irn = livein->irn;
1630 be_ilpsched_irn_t *node = get_ilpsched_irn(env, irn);
1633 /* check if node can be alive here */
1634 if (t >= livein->max_alive_steps)
1637 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1639 /* current type is not suitable */
1643 a_idx = tp_idx * livein->max_alive_steps + t;
1644 ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1647 if (ARR_LEN(tmp_var_idx) > 0)
1648 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1649 DEL_ARR_F(tmp_var_idx);
1651 /* - num_nodes * y_{nt}^k */
1652 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1653 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1658 DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1659 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1663 * Create ILP branch constraints:
1664 * Assure, alle nodes are scheduled prior to cfg op.
1666 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1668 ir_node *cur_irn, *cfop;
1669 unsigned num_cst = 0;
1670 unsigned num_non_branches = 0;
1671 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1672 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1674 ilp_timer_push(t_cst);
1676 /* determine number of non-branch nodes and the one and only branch node */
1677 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1678 switch (get_irn_opcode(cur_irn)) {
1688 if (is_cfop(cur_irn)) {
1689 assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1700 be_ilpsched_irn_t *cf_node = get_ilpsched_irn(env, cfop);
1701 ilpsched_node_attr_t *cf_na = get_ilpsched_node_attr(cf_node);
1704 /* for each time step */
1705 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1706 int *non_branch_vars, *branch_vars;
1709 snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1710 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1711 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1714 /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1715 non_branch_vars = NEW_ARR_F(int, 0);
1716 branch_vars = NEW_ARR_F(int, 0);
1717 foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1718 be_ilpsched_irn_t *node = get_ilpsched_irn(env, cur_irn);
1719 ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
1722 if (cur_irn == cfop) {
1723 /* for all unit types available for this node */
1724 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1725 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1726 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1730 /* sum up all possible schedule points for this node upto current timestep */
1731 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1733 unsigned tmax = MIN(t, na->alap - 1);
1735 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1736 unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1737 ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1744 if (ARR_LEN(non_branch_vars) > 0)
1745 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1746 if (ARR_LEN(branch_vars) > 0)
1747 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1749 DEL_ARR_F(branch_vars);
1750 DEL_ARR_F(non_branch_vars);
1754 DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1755 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1759 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1762 unsigned num_cst = 0;
1763 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1764 lc_timer_t *t_cst = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1766 ilp_timer_push(t_cst);
1767 /* check all nodes */
1768 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1769 be_ilpsched_irn_t *node;
1770 ilpsched_node_attr_t *na;
1774 /* only mode_T nodes can have Projs and Keeps assigned */
1775 if (get_irn_mode(irn) != mode_T)
1778 node = get_ilpsched_irn(env, irn);
1779 na = get_ilpsched_node_attr(node);
1781 /* check if has some Projs and Keeps assigned */
1782 if (! na->projkeeps)
1785 /* we can run only once over the queue, so preserve the nodes */
1786 pk = NEW_ARR_F(ir_node *, 0);
1787 while (! waitq_empty(na->projkeeps))
1788 ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1789 del_waitq(na->projkeeps);
1790 na->projkeeps = NULL;
1792 /* for all time steps at which this node can be scheduled */
1793 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1795 int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1797 /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1798 snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1799 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1800 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1803 /* sum up scheduling variables for this time step */
1804 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1805 int idx = ILPVAR_IDX(na, tp_idx, t);
1806 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1809 if (ARR_LEN(tmp_var_idx_n) > 0)
1810 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1811 DEL_ARR_F(tmp_var_idx_n);
1813 /* subtract all Proj and Keep variables for this step */
1814 for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1815 be_ilpsched_irn_t *pk_node = get_ilpsched_irn(env, pk[i]);
1816 ilpsched_node_attr_t *pk_na = get_ilpsched_node_attr(pk_node);
1819 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1820 int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1821 lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1827 DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1828 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1832 /***************************************************
1834 * |_ _| | | __ \ (_)
1835 * | | | | | |__) | _ __ ___ __ _ _ _ __
1836 * | | | | | ___/ | '_ ` _ \ / _` | | '_ \
1837 * _| |_| |____| | | | | | | | (_| | | | | |
1838 * |_____|______|_| |_| |_| |_|\__,_|_|_| |_|
1840 ***************************************************/
1843 * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1845 static void create_ilp(ir_node *block, void *walk_env) {
1846 be_ilpsched_env_t *env = walk_env;
1847 be_ilpsched_irn_t *block_node = get_ilpsched_irn(env, block);
1848 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(block_node);
1849 FILE *logfile = NULL;
1852 struct obstack var_obst;
1855 DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1856 DBG((env->dbg, 255, " ILP Scheduling for %+F\n", block));
1857 DBG((env->dbg, 255, "=========================================\n\n"));
1859 DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1860 block, ba->n_interesting_nodes, ba->max_steps));
1862 /* notify backend and get block environment */
1863 env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1865 /* if we have less than two interesting nodes, there is no need to create the ILP */
1866 if (ba->n_interesting_nodes > 1) {
1867 double fact_var = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1868 double fact_cst = ba->n_interesting_nodes > 25 ? 3 : 4.5;
1869 int base_num = ba->n_interesting_nodes * ba->n_interesting_nodes;
1870 int estimated_n_var = (int)((double)base_num * fact_var);
1871 int estimated_n_cst = (int)((double)base_num * fact_cst);
1873 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1874 estimated_n_var, estimated_n_cst));
1875 (void) estimated_n_var;
1877 /* set up the LPP object */
1878 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1880 lpp = new_lpp_userdef(
1883 estimated_n_cst, /* num vars */
1884 estimated_n_cst + 1, /* num cst */
1885 1.3); /* grow factor */
1886 obstack_init(&var_obst);
1888 /* create ILP variables */
1889 create_variables(env, lpp, block_node, &var_obst);
1891 /* create ILP constraints */
1892 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1893 create_assignment_and_precedence_constraints(env, lpp, block_node);
1894 create_ressource_constraints(env, lpp, block_node);
1895 create_bundle_constraints(env, lpp, block_node);
1896 create_branch_constraint(env, lpp, block_node);
1897 //create_proj_keep_constraints(env, lpp, block_node);
1899 if (env->opts->regpress) {
1900 create_alive_nodes_constraint(env, lpp, block_node);
1901 create_alive_livein_nodes_constraint(env, lpp, block_node);
1902 create_pressure_alive_constraint(env, lpp, block_node);
1905 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1907 /* debug stuff, dump lpp when debugging is on */
1909 if (firm_dbg_get_mask(env->dbg) > 1) {
1913 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1914 f = fopen(buf, "w");
1915 lpp_dump_plain(lpp, f);
1917 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1922 /* set solve time limit */
1923 lpp_set_time_limit(lpp, env->opts->time_limit);
1925 /* set logfile if requested */
1926 if (strlen(env->opts->log_file) > 0) {
1927 if (strcasecmp(env->opts->log_file, "stdout") == 0)
1928 lpp_set_log(lpp, stdout);
1929 else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1930 lpp_set_log(lpp, stderr);
1932 logfile = fopen(env->opts->log_file, "w");
1934 fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1936 lpp_set_log(lpp, logfile);
1941 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1946 /* check for valid solution */
1947 if (! lpp_is_sol_valid(lpp)) {
1952 if (firm_dbg_get_mask(env->dbg) >= 2) {
1953 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1954 f = fopen(buf, "w");
1955 lpp_dump_plain(lpp, f);
1957 snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1959 dump_ir_block_graph(env->irg, "-infeasible");
1963 ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1967 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1968 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1969 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1970 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));
1971 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1972 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1973 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1974 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1976 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1979 /* apply solution */
1980 be_stat_ev("nodes", ba->block_last_idx);
1981 be_stat_ev("vars", lpp ? lpp->var_next : 0);
1982 be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1984 be_stat_ev("time", -1);
1985 be_stat_ev_dbl("opt", 0.0);
1986 list_sched_single_block(env->birg, block, env->be_opts);
1990 double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1991 be_stat_ev_dbl("time", lpp->sol_time);
1992 be_stat_ev_dbl("opt", opt);
1995 be_stat_ev_dbl("time", 0.0);
1996 be_stat_ev_dbl("opt", 100.0);
1998 apply_solution(env, lpp, block);
2004 /* notify backend */
2005 be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2009 * Perform ILP scheduling on the given irg.
2011 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2012 be_ilpsched_env_t env;
2013 const char *name = "be ilp scheduling";
2014 ir_graph *irg = be_get_birg_irg(birg);
2015 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
2016 const arch_isa_t *isa = arch_env->isa;
2017 const ilp_sched_selector_t *sel = isa->impl->get_ilp_sched_selector(isa);
2019 FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2021 stat_ev_ctx_push("ilpsched");
2023 // firm_dbg_set_mask(env.dbg, 1);
2025 env.irg_env = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2028 env.height = heights_new(irg);
2029 env.main_env = birg->main_env;
2030 env.arch_env = arch_env;
2031 env.cpu = arch_isa_get_machine(arch_env->isa);
2032 env.opts = &ilp_opts;
2034 env.be_opts = be_opts;
2035 phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2037 /* assign a unique per block number to all interesting nodes */
2038 irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2041 The block indices are completely build after the walk,
2042 now we can allocate the bitsets (size depends on block indices)
2045 phase_reinit_irn_data(&env.ph);
2047 /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2048 irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2050 /* Calculate ALAP of all irns */
2051 irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2053 /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2054 irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2056 /* perform ILP scheduling */
2057 irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2060 if (firm_dbg_get_mask(env.dbg)) {
2062 phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2064 fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2068 /* free data allocated dynamically */
2069 irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2071 /* free all allocated object */
2072 phase_free(&env.ph);
2073 heights_free(env.height);
2075 /* notify backend */
2076 be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2078 stat_ev_ctx_pop("ilpsched");
2082 * Register ILP scheduler options.
2084 void be_init_ilpsched(void)
2086 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2087 lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2089 lc_opt_add_table(sched_grp, ilpsched_option_table);
2092 #else /* WITH_ILP */
2094 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2097 #endif /* WITH_ILP */