Fix non-debug build of FIRM.
[libfirm] / ir / be / beilpsched.c
1 /*
2  * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
3  *
4  * This file is part of libFirm.
5  *
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.
10  *
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.
14  *
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
17  * PURPOSE.
18  */
19
20 /**
21  * @file
22  * @brief       ILP based instruction scheduling.
23  * @author      Christian Wuerdig
24  * @date        22.10.2006
25  * @version     $Id$
26  *
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
31  */
32 #ifdef HAVE_CONFIG_H
33 #include "config.h"
34 #endif
35
36 #ifdef WITH_ILP
37
38 #include <math.h>
39
40 #ifndef _WIN32
41 #include <strings.h>
42 #endif /* _WIN32 */
43
44 #include "irnode_t.h"
45 #include "irgwalk.h"
46 #include "irbitset.h"
47 #include "irphase_t.h"
48 #include "height.h"
49 #include "iredges.h"
50 #include "pdeq.h"
51 #include "debug.h"
52 #include "irtools.h"
53 #include "irdump.h"
54 #include "irprintf.h"
55 #include "plist.h"
56 #include "irprintf.h"
57
58 #include <lpp/lpp.h>
59 #include <lpp/lpp_net.h>
60
61 #include <libcore/lc_opts.h>
62 #include <libcore/lc_opts_enum.h>
63 #include <libcore/lc_timing.h>
64
65 #include "be.h"
66 #include "benode_t.h"
67 #include "besched_t.h"
68 #include "beilpsched.h"
69 #include "beutil.h"
70 #include "bestat.h"
71 #include "beirg_t.h"
72 #include "benodesets.h"
73
74 typedef struct _ilpsched_options_t {
75         unsigned regpress;
76         unsigned time_limit;
77         char     log_file[1024];
78 } ilpsched_options_t;
79
80 typedef struct _unit_type_info_t {
81         int                            n_units;
82         const be_execution_unit_type_t *tp;
83 } unit_type_info_t;
84
85 /**
86  * holding the ILP variables of the different types
87  */
88 typedef struct _ilp_var_types_t {
89         int *x;   /* x_{nt}^k variables */
90         int *a;   /* a_{nt}^k variables */
91         int *y;   /* y_{nt}^k variables */
92 } ilp_var_types_t;
93
94 /**
95  * Holds alive variables for a node live-in to a block.
96  */
97 typedef struct _ilp_livein_node_t {
98         ir_node  *irn;
99         unsigned max_alive_steps;
100         int      *a;
101 } ilp_livein_node_t;
102
103 /* attributes for a node */
104 typedef struct _ilpsched_node_attr_t {
105         unsigned asap;                     /**< The ASAP scheduling control step */
106         unsigned alap;                     /**< The ALAP scheduling control step */
107         unsigned latency;                  /**< Latency of this node (needed for sorting) */
108         unsigned sched_point;              /**< the step in which the node is finally scheduled */
109         unsigned visit_idx;                /**< Index of the node having visited this node last */
110         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
111         unsigned n_consumer;               /**< Number of consumers */
112         ir_node  **block_consumer;         /**< List of consumer being in the same block */
113         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
114         unsigned block_idx     : 30;       /**< A unique per block index */
115         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
116         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
117         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
118                                                                                         for ASAP, successors for ALAP */
119         unsigned         n_unit_types;     /**< number of allowed execution unit types */
120         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
121         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
122 } ilpsched_node_attr_t;
123
124 /* attributes for a block */
125 typedef struct _ilpsched_block_attr_t {
126         unsigned block_last_idx;        /**< The highest node index in block so far */
127         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
128         unsigned max_steps;             /**< Upper bound for block execution */
129         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
130         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
131         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
132 } ilpsched_block_attr_t;
133
134 typedef union _ilpsched_attr_ {
135         ilpsched_node_attr_t  node_attr;
136         ilpsched_block_attr_t block_attr;
137 } ilpsched_attr_t;
138
139 /* A irn for the phase and it's attributes (either node or block) */
140 typedef struct {
141         ir_node         *irn;
142         ilpsched_attr_t attr;
143 } be_ilpsched_irn_t;
144
145 /* The ILP scheduling environment */
146 typedef struct {
147         ir_phase             ph;            /**< The phase */
148         ir_graph             *irg;          /**< The current irg */
149         heights_t            *height;       /**< The heights object of the irg */
150         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
151         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
152         const arch_env_t     *arch_env;
153         const arch_isa_t     *isa;          /**< The ISA */
154         const be_main_env_t  *main_env;
155         const be_machine_t   *cpu;          /**< the current abstract machine */
156         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
157         const be_irg_t       *birg;         /**< The birg object */
158         be_options_t         *be_opts;      /**< backend options */
159         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
160         DEBUG_ONLY(firm_dbg_module_t *dbg);
161 } be_ilpsched_env_t;
162
163 /* convenience macros to handle phase irn data */
164 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
165 #define is_ilpsched_block(node)             (is_Block((node)->irn))
166 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
167 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
168
169 /* check if node is considered for ILP scheduling */
170 #define consider_for_sched(isa, irn) \
171         (! (is_Block(irn)            ||  \
172                 is_normal_Proj(isa, irn) ||  \
173                 is_Phi(irn)              ||  \
174                 is_NoMem(irn)            ||  \
175                 is_Unknown(irn)          ||  \
176                 is_End(irn)                  \
177                 ))
178
179 /* gives the valid scheduling time step interval for a node */
180 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
181
182 /* gives the valid interval where a node can die */
183 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
184
185 /* gives the corresponding ILP variable for given node, unit and time step */
186 #define ILPVAR_IDX(na, unit, control_step) \
187         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
188
189 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
190 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
191         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
192
193 /* check if a double value is within an epsilon environment of 0 */
194 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
195
196 #define ilp_timer_push(t)         lc_timer_push((t))
197 #define ilp_timer_pop()           lc_timer_pop()
198 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
199
200 /* option variable */
201 static ilpsched_options_t ilp_opts = {
202         1,     /* default is with register pressure constraints */
203         300,   /* 300 sec per block time limit */
204         ""     /* no log file */
205 };
206
207 /* ILP options */
208 static const lc_opt_table_entry_t ilpsched_option_table[] = {
209         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
210         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
211         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
212         LC_OPT_LAST
213 };
214
215 /*
216         We need this global variable as we compare nodes dependent on heights,
217         but we cannot pass any information to the qsort compare function.
218 */
219 static heights_t *glob_heights;
220
221 /**
222  * Check if irn is a Proj, which has no execution units assigned.
223  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
224  */
225 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
226         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
227 }
228
229 /**
230  * Skips normal Projs.
231  * @return predecessor if irn is a normal Proj, otherwise irn.
232  */
233 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
234         if (is_normal_Proj(isa, irn))
235                 return get_Proj_pred(irn);
236         return irn;
237 }
238
239 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
240         unsigned lat = be_ilp_sched_latency(sel, irn, env);
241         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
242                 lat = 1;
243         return lat;
244 }
245
246 static int cmp_live_in_nodes(const void *a, const void *b) {
247         const ilp_livein_node_t *n1 = a;
248         const ilp_livein_node_t *n2 = b;
249
250         return n1->irn != n2->irn;
251 }
252
253 /**
254  * Compare scheduling time steps of two be_ilpsched_irn's.
255  */
256 static int cmp_ilpsched_irn(const void *a, const void *b) {
257         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
258         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
259         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
260         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
261
262         if (n1_a->sched_point == n2_a->sched_point) {
263                 ir_node *irn_a = n1->irn;
264                 ir_node *irn_b = n2->irn;
265
266                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
267                         return 1;
268                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
269                         return -1;
270
271                 /*
272                         Ok, timestep is equal and the nodes are parallel,
273                         so check latency and schedule high latency first.
274                 */
275                 return QSORT_CMP(n2_a->latency, n1_a->latency);
276         }
277         else
278                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
279 }
280
281 /**
282  * In case there is no phase information for irn, initialize it.
283  */
284 static void *init_ilpsched_irn(ir_phase *ph, ir_node *irn, void *old) {
285         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
286
287         if (res == old) {
288                 /* if we have already some data: check for reinitialization */
289
290                 if (! is_Block(irn)) {
291                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
292
293                         if (! na->transitive_block_nodes) {
294                                 ir_node               *block      = get_nodes_block(irn);
295                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
296                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
297
298                                 /* we are called after the block indices have been build: create bitset */
299                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
300                         }
301                         else {
302                                 /* we are called from reinit block data: clear the bitset */
303                                 bitset_clear_all(na->transitive_block_nodes);
304                                 na->visit_idx    = 0;
305                                 na->alap_changed = 1;
306                         }
307                 }
308                 return old;
309         }
310
311         res->irn = irn;
312
313         /* set ilpsched irn attributes (either block or irn) */
314         if (is_Block(irn)) {
315                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
316
317                 ba->n_interesting_nodes = 0;
318                 ba->block_last_idx      = 0;
319                 ba->root_nodes          = plist_new();
320                 ba->head_ilp_nodes      = NULL;
321                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
322                 ba->max_steps           = 0;
323         }
324         else {
325                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
326                 memset(na, 0, sizeof(*na));
327         }
328
329         return res;
330 }
331
332 /**
333  * Assign a per block unique number to each node.
334  */
335 static void build_block_idx(ir_node *irn, void *walk_env) {
336         be_ilpsched_env_t     *env = walk_env;
337         be_ilpsched_irn_t     *node, *block_node;
338         ilpsched_node_attr_t  *na;
339         ilpsched_block_attr_t *ba;
340
341         set_irn_link(irn, NULL);
342         if (! consider_for_sched(env->arch_env->isa, irn))
343                 return;
344
345         node       = get_ilpsched_irn(env, irn);
346         na         = get_ilpsched_node_attr(node);
347         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
348         ba         = get_ilpsched_block_attr(block_node);
349
350         na->block_idx = ba->block_last_idx++;
351 }
352
353 /********************************************************
354  *                              __        _
355  *                             / /       | |
356  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
357  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
358  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
359  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
360  *                 | |                           | |
361  *                 |_|                           |_|
362  ********************************************************/
363
364 /**
365  * Add all nodes having no user in current block to last_nodes list.
366  */
367 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
368         ir_node               *block;
369         const ir_edge_t       *edge;
370         be_ilpsched_irn_t     *block_node, *node;
371         ilpsched_block_attr_t *ba;
372         ilpsched_node_attr_t  *na;
373         int                   i, j;
374         be_ilpsched_env_t     *env           = walk_env;
375         int                   has_block_user = 0;
376         unsigned              n_consumer     = 0;
377         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
378         ir_node               **consumer;
379         unsigned              idx;
380
381         if (! consider_for_sched(env->arch_env->isa, irn))
382                 return;
383
384         block    = get_nodes_block(irn);
385     idx      = get_irn_idx(irn);
386         consumer = NEW_ARR_F(ir_node *, 0);
387
388         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
389
390         /* check data and dependency out edges */
391         for (i = 0; i < 2 && ! has_block_user; ++i) {
392                 foreach_out_edge_kind(irn, edge, ekind[i]) {
393                         ir_node *user = get_edge_src_irn(edge);
394
395                         if (is_normal_Proj(env->arch_env->isa, user)) {
396                                 const ir_edge_t *user_edge;
397
398                                 if (get_irn_mode(user) == mode_X)
399                                         continue;
400
401                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
402                                 for (j = 0; j < 2; ++j) {
403                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
404                                                 ir_node *real_user = get_edge_src_irn(user_edge);
405
406                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
407                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
408                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
409
410                                                         /* skip already visited nodes */
411                                                         if (ua->consumer_idx == idx)
412                                                                 continue;
413
414                                                         /* check if node has user in this block and collect the user if it's a data user */
415                                                         if (get_nodes_block(real_user) == block) {
416                                                                 if (i == 0 && j == 0)
417                                                                         ARR_APP1(ir_node *, consumer, real_user);
418                                                                 has_block_user = 1;
419                                                         }
420
421                                                         /* only count data consumer */
422                                                         if (i == 0)
423                                                                 n_consumer++;
424
425                                                         /* mark user as visited by this node */
426                                                         ua->consumer_idx = idx;
427                                                 }
428                                         }
429                                 }
430                         }
431                         else if (is_Block(user)) {
432                                 continue;
433                         }
434                         else if (! is_Phi(user)) {
435                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
436                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
437
438                                 /* skip already visited nodes */
439                                 if (ua->consumer_idx == idx)
440                                         continue;
441
442                                 /* check if node has user in this block and collect the user if it's a data user */
443                                 if (get_nodes_block(user) == block) {
444                                         if (i == 0)
445                                                 ARR_APP1(ir_node *, consumer, user);
446                                         has_block_user = 1;
447                                 }
448
449                                 /* only count data consumer */
450                                 if (i == 0)
451                                         n_consumer++;
452
453                                 /* mark user visited by this node */
454                                 ua->consumer_idx = idx;
455                         }
456                         else if (get_nodes_block(user) != block) {
457                                 n_consumer++;
458                         }
459                 }
460         }
461
462         block_node = get_ilpsched_irn(env, block);
463         ba         = get_ilpsched_block_attr(block_node);
464
465         ba->n_interesting_nodes++;
466
467         /* current irn has no user inside this block, add to queue */
468         if (! has_block_user) {
469                 DB((env->dbg, LEVEL_3, "root node\n"));
470                 plist_insert_back(ba->root_nodes, irn);
471         }
472         else {
473                 DB((env->dbg, LEVEL_3, "normal node\n"));
474         }
475
476         /* record number of all consumer and the consumer within the same block */
477         node = get_ilpsched_irn(env, irn);
478         na   = get_ilpsched_node_attr(node);
479         na->n_consumer     = n_consumer;
480         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
481         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
482         DEL_ARR_F(consumer);
483 }
484
485 /**
486  * Calculate the ASAP scheduling step for current irn.
487  */
488 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
489         be_ilpsched_env_t     *env = walk_env;
490         int                   i;
491         ir_node               *block;
492         be_ilpsched_irn_t     *node, *block_node;
493         ilpsched_node_attr_t  *na;
494         ilpsched_block_attr_t *ba;
495
496         /* These nodes are handled separate */
497         if (! consider_for_sched(env->arch_env->isa, irn))
498                 return;
499
500         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
501
502         block    = get_nodes_block(irn);
503         node     = get_ilpsched_irn(env, irn);
504         na       = get_ilpsched_node_attr(node);
505         na->asap = 1;
506
507         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
508                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
509
510                 /* check for greatest distance to top */
511                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
512                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
513                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
514                         unsigned             lat;
515
516                         lat         = fixed_latency(env->sel, pred, env->block_env);
517                         na->latency = lat;
518                         na->asap    = MAX(na->asap, pna->asap + lat);
519                 }
520         }
521
522         /* add node to ILP node list and update max_steps */
523         block_node = get_ilpsched_irn(env, block);
524         ba         = get_ilpsched_block_attr(block_node);
525
526         set_irn_link(irn, ba->head_ilp_nodes);
527         ba->head_ilp_nodes = irn;
528         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
529
530         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
531 }
532
533 /**
534  * Calculate the ALAP scheduling step of all irns in current block.
535  * Depends on max_steps being calculated.
536  */
537 static void calculate_block_alap(ir_node *block, void *walk_env) {
538         be_ilpsched_env_t     *env        = walk_env;
539         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
540         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
541         waitq                 *cur_queue  = new_waitq();
542         plist_element_t       *el;
543
544         assert(is_Block(block));
545
546         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
547                 block, ba->n_interesting_nodes, ba->max_steps));
548
549         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
550         //phase_reinit_block_irn_data(&env->ph, block);
551
552         /* init start queue */
553         foreach_plist(ba->root_nodes, el) {
554                 waitq_put(cur_queue, plist_element_get_value(el));
555         }
556
557         /* repeat until all nodes are processed */
558         while (! waitq_empty(cur_queue)) {
559                 waitq *next_queue = new_waitq();
560
561                 /* process all nodes in current step */
562                 while (! waitq_empty(cur_queue)) {
563                         ir_node              *cur_irn = waitq_get(cur_queue);
564                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
565                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
566                         int                  i;
567
568                         /* cur_node has no alap set -> it's a root node, set to max alap */
569                         if (na->alap == 0) {
570                                 na->alap = ba->max_steps;
571                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
572                                         cur_irn, na->alap));
573                         }
574                         else {
575                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
576                                         cur_irn, na->alap));
577                         }
578
579                         /* set the alap's of all predecessors */
580                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
581                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
582
583                                 /* check for greatest distance to bottom */
584                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
585                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
586                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
587                                         unsigned             lat;
588
589                                         /* mark the predecessor as visited by current irn */
590                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
591                                                 continue;
592                                         pna->visit_idx = get_irn_idx(cur_irn);
593
594                                         lat = fixed_latency(env->sel, pred, env->block_env);
595
596                                         /* set ALAP of current pred */
597                                         if (pna->alap == 0) {
598                                                 /* current ALAP is 0: node has not yet been visited */
599                                                 pna->alap_changed = 1;
600                                                 pna->alap         = na->alap - lat;
601                                         }
602                                         else if (pna->alap > na->alap - lat) {
603                                                 /* we found a longer path to root node: change ALAP */
604                                                 pna->alap         = na->alap - lat;
605                                                 pna->alap_changed = 1;
606                                         }
607                                         else {
608                                                 /* current ALAP is best found so far: keep it */
609                                                 pna->alap_changed = 0;
610                                         }
611
612                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
613
614                                         /* enqueue node for next iteration */
615                                         if (get_irn_ins_or_deps(pred) > 0)
616                                                 waitq_put(next_queue, pred);
617                                 }
618                         }
619                 }
620
621                 /* prepare for next iteration */
622                 del_waitq(cur_queue);
623                 cur_queue = next_queue;
624         }
625 }
626
627 /**
628  * Free list of root nodes and the set of live-in nodes.
629  */
630 static void clear_unwanted_data(ir_node *block, void *walk_env) {
631         be_ilpsched_env_t     *env        = walk_env;
632         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
633         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
634
635         plist_free(ba->root_nodes);
636         ba->root_nodes = NULL;
637         del_pset(ba->livein_nodes);
638         ba->livein_nodes = NULL;
639 }
640
641 /**
642  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
643  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
644  */
645 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
646         be_ilpsched_env_t    *env  = walk_env;
647         ir_node              *pred = irn;
648         be_ilpsched_irn_t    *node, *pred_node;
649         ilpsched_node_attr_t *na, *pna;
650
651         if (! consider_for_sched(env->arch_env->isa, irn))
652                 return;
653
654         if (! is_Proj(irn) && ! be_is_Keep(irn))
655                 return;
656
657         /* go to the ancestor */
658         if (be_is_Keep(irn))
659                 pred = get_irn_n(irn, 0);
660         pred = skip_Proj(pred);
661
662         node      = get_ilpsched_irn(env, irn);
663         pred_node = get_ilpsched_irn(env, pred);
664         na        = get_ilpsched_node_attr(node);
665         pna       = get_ilpsched_node_attr(pred_node);
666
667         na->asap = pna->asap;
668         na->alap = pna->alap;
669
670         /* record all Projs and Keeps for this node */
671         if (! pna->projkeeps)
672                 pna->projkeeps = new_waitq();
673         waitq_put(pna->projkeeps, irn);
674
675         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
676 }
677
678 /*******************************************
679  *           _              _       _
680  *          | |            | |     | |
681  *  ___  ___| |__   ___  __| |_   _| | ___
682  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
683  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
684  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
685  *
686  *******************************************/
687
688 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
689         const ir_edge_t *edge;
690
691         foreach_out_edge(irn, edge) {
692                 ir_node *user = get_edge_src_irn(edge);
693
694                 if (be_is_Keep(user)) {
695                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
696                         waitq_put(keeps, user);
697                 }
698         }
699 }
700
701 /**
702  * Inserts @p irn before @p before into schedule and notifies backend.
703  */
704 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
705         ir_node *before, ir_node *irn, unsigned cycle)
706 {
707         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
708         sched_add_before(before, irn);
709 }
710
711 /**
712  * Adds a node, it's Projs (in case of mode_T nodes) and
713  * it's Keeps to schedule.
714  */
715 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
716         const ir_edge_t *edge;
717         waitq           *keeps = new_waitq();
718
719         /* mode_M nodes are not scheduled */
720         if (get_irn_mode(irn) == mode_M)
721                 return;
722
723         if (! sched_is_scheduled(irn))
724                 notified_sched_add_before(env, block, irn, cycle);
725
726         /* add Projs */
727         if (get_irn_mode(irn) == mode_T) {
728                 foreach_out_edge(irn, edge) {
729                         ir_node *user = get_edge_src_irn(edge);
730
731                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
732                                 get_irn_n_edges(user) > 0)
733                         {
734                                 notified_sched_add_before(env, block, user, cycle);
735                         }
736
737                         check_for_keeps(keeps, block, user);
738                 }
739         }
740         else {
741                 check_for_keeps(keeps, block, irn);
742         }
743
744         /* add Keeps */
745         while (! waitq_empty(keeps)) {
746                 ir_node *keep = waitq_get(keeps);
747                 if (! sched_is_scheduled(keep))
748                         notified_sched_add_before(env, block, keep, cycle);
749         }
750
751         del_waitq(keeps);
752 }
753
754 /**
755  * Schedule all nodes in the given block, according to the ILP solution.
756  */
757 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
758         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
759         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
760         sched_info_t          *info       = get_irn_sched_info(block);
761         be_ilpsched_irn_t     **sched_nodes;
762         unsigned              i, l;
763         ir_node               *cfop, *irn;
764         const ir_edge_t       *edge;
765
766         /* init block schedule list */
767         INIT_LIST_HEAD(&info->list);
768         info->scheduled = 1;
769
770         /* collect nodes and their scheduling time step */
771         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
772         if (ba->n_interesting_nodes == 0) {
773                 /* ignore */
774         }
775         else if (ba->n_interesting_nodes == 1) {
776                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
777
778                 /* add the single node */
779                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
780         }
781         else {
782                 /* check all nodes for their positive solution */
783                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
784                         be_ilpsched_irn_t    *node;
785                         ilpsched_node_attr_t *na;
786                         int                  tp_idx, found;
787                         unsigned             cur_var, t;
788
789                         node    = get_ilpsched_irn(env, irn);
790                         na      = get_ilpsched_node_attr(node);
791                         cur_var = 0;
792                         found   = 0;
793
794                         if (! na->is_dummy_node) {
795                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
796                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
797                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
798
799                                                 if (! LPP_VALUE_IS_0(cost)) {
800                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
801                                                         found = 1;
802                                                 }
803                                         }
804                                 }
805                         }
806
807                         found = 0;
808                         /* go over all variables of a node until the non-zero one is found */
809                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
810                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
811                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
812
813                                         /* check, if variable is set to one (it's not zero then :) */
814                                         if (! LPP_VALUE_IS_0(val)) {
815                                                 na->sched_point = t;
816                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
817                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
818                                                         irn, t, na->type_info[tp_idx].tp->name));
819                                                 found = 1;
820                                         }
821                                 }
822                         }
823                 }
824
825                 glob_heights = env->height;
826                 /* sort nodes ascending by scheduling time step */
827                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
828         }
829
830         /* make all Phis ready and remember the single cf op */
831         cfop = NULL;
832         foreach_out_edge(block, edge) {
833                 irn = get_edge_src_irn(edge);
834
835                 switch (get_irn_opcode(irn)) {
836                         case iro_Phi:
837                                 add_to_sched(env, block, irn, 0);
838                                 break;
839                         case iro_Start:
840                         case iro_End:
841                         case iro_Proj:
842                         case iro_Bad:
843                         case iro_Unknown:
844                                 break;
845                         default:
846                                 if (is_cfop(irn)) {
847                                         assert(cfop == NULL && "Highlander - there can be only one");
848                                         cfop = irn;
849                                 }
850                         break;
851                 }
852         }
853
854         /* add all nodes from list */
855         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
856                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
857                 if (sched_nodes[i]->irn != cfop)
858                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
859         }
860
861         /* schedule control flow node if not already done */
862         if (cfop && ! sched_is_scheduled(cfop))
863                 add_to_sched(env, block, cfop, 0);
864
865         DEL_ARR_F(sched_nodes);
866 }
867
868 /***************************************************************
869  *   _____ _      _____     _____           _   _
870  *  |_   _| |    |  __ \   / ____|         | | (_)
871  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
872  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
873  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
874  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
875  *
876  ***************************************************************/
877
878 /**
879  * Check if node can be executed on given unit type.
880  */
881 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
882         int                  i;
883         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
884
885         for (i = na->n_unit_types - 1; i >= 0; --i) {
886                 if (na->type_info[i].tp == tp)
887                         return i;
888         }
889
890         return -1;
891 }
892
893 /************************************************
894  *                   _       _     _
895  *                  (_)     | |   | |
896  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
897  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
898  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
899  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
900  *
901  ************************************************/
902
903 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
904         const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
905         unsigned                  n_unit_types = 0;
906         be_ilpsched_irn_t         *node;
907         ilpsched_node_attr_t      *na;
908         unsigned                  unit_idx, tp_idx;
909
910         /* count number of available unit types for this node */
911         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
912                 /* just count */ ;
913
914         node = get_ilpsched_irn(env, irn);
915         na   = get_ilpsched_node_attr(node);
916
917         if (! na->type_info) {
918                 na->n_unit_types = n_unit_types;
919                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
920
921                 /* fill the type info array */
922                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
923                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
924                                 /* beware: we also count number of available units here */
925                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
926                                         na->is_dummy_node = 1;
927                         }
928
929                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
930                         na->type_info[tp_idx].n_units = unit_idx;
931                 }
932         }
933
934         return n_unit_types;
935 }
936
937 /**
938  * Returns the largest alap time of a user of @p irn.
939  * The user must be in block @p block.
940  */
941 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
942         const ir_edge_t *edge;
943         unsigned        max_alap = 0;
944
945         foreach_out_edge(irn, edge) {
946                 ir_node *user = get_edge_src_irn(edge);
947
948                 if (get_nodes_block(user) == block) {
949                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
950                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
951
952                         max_alap = MAX(max_alap, na->alap);
953                 }
954         }
955
956         assert(max_alap > 0);
957         return max_alap;
958 }
959
960 /**
961  * Create the following variables:
962  * - x_{nt}^k    binary     weigthed with: t
963  *      node n is scheduled at time step t to unit type k
964  * ==>> These variables represent the schedule
965  *
966  * - a_{nt}^k    binary     weighted with num_nodes
967  *      node n is alive at time step t on unit type k
968  *
969  * - y_{nt}^k    continuous  weighted with: num_nodes^2
970  *      register pressure over limit for unit type k
971  * ==>> These variables represent the register pressure
972  *
973  */
974 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
975         char                  buf[1024];
976         ir_node               *irn;
977         unsigned              num_block_var, num_nodes;
978         ilp_livein_node_t     *livein;
979         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
980         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
981         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
982
983         ilp_timer_push(t_var);
984         num_block_var = num_nodes = 0;
985         foreach_linked_irns(ba->head_ilp_nodes, irn) {
986                 be_ilpsched_irn_t    *node;
987                 ilpsched_node_attr_t *na;
988                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
989                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
990                 int                  i;
991
992                 node         = get_ilpsched_irn(env, irn);
993                 na           = get_ilpsched_node_attr(node);
994                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
995
996                 /* allocate space for ilp variables */
997                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
998                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
999
1000                 /* we need these variables only for "real" nodes */
1001                 if (! na->is_dummy_node) {
1002                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1003                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1004
1005                         num_ad         = ba->max_steps - na->asap + 1;
1006                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1007                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1008                 }
1009
1010                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1011                         irn, na->asap, na->alap, na->n_unit_types));
1012
1013                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1014                 /* create variables */
1015                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1016                         unsigned t;
1017
1018                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1019                                 /* x_{nt}^k variables */
1020                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1021                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1022                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1023                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1024                                 /* variable counter */
1025                                 n_var++;
1026                                 num_block_var++;
1027
1028                                 if (! na->is_dummy_node) {
1029                                         /* y_{nt}^k variables */
1030                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1031                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1032                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1033                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1034
1035                                         /* variable counter */
1036                                         n_var++;
1037                                         num_block_var++;
1038                                 }
1039                         }
1040
1041                         /* a node can die at any step t: asap(n) <= t <= U */
1042                         if (! na->is_dummy_node) {
1043                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1044
1045                                         /* a_{nt}^k variables */
1046                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1047                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1048                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1049                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1050
1051                                         /* variable counter */
1052                                         n_var++;
1053                                         num_block_var++;
1054                                 }
1055                         }
1056
1057                         /* collect live-in nodes */
1058                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1059                                 ir_node *pred = get_irn_n(irn, i);
1060
1061                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1062                                         be_ilpsched_set_type_info(env, pred, var_obst);
1063                                         if (! na->is_dummy_node) {
1064                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1065                                                 entry->irn = pred;
1066                                                 entry->a   = NULL;
1067                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1068                                         }
1069                                 }
1070                         }
1071                 }
1072
1073                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1074                 num_nodes++;
1075         }
1076
1077         /* create alive variables a_{nt}^k for live-ins */
1078         foreach_pset(ba->livein_nodes, livein) {
1079                 be_ilpsched_irn_t    *node;
1080                 ilpsched_node_attr_t *na;
1081                 unsigned             tp_idx, var_idx;
1082                 ir_node              *irn;
1083
1084                 irn  = livein->irn;
1085                 node = get_ilpsched_irn(env, irn);
1086                 na   = get_ilpsched_node_attr(node);
1087
1088                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1089
1090                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1091                 var_idx   = 0;
1092
1093                 /* create variables */
1094                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1095                         unsigned t;
1096
1097                         for (t = 0; t < livein->max_alive_steps; ++t) {
1098                                 /* a_{nt}^k variables */
1099                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1100                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1101                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1102                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1103                                 num_block_var++;
1104                         }
1105                 }
1106         }
1107
1108         ilp_timer_pop();
1109         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1110                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1111 }
1112
1113 /*******************************************************
1114  *                      _             _       _
1115  *                     | |           (_)     | |
1116  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1117  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1118  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1119  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1120  *
1121  *******************************************************/
1122
1123 /**
1124  * Collect all operands and nodes @p irn depends on.
1125  * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1126  */
1127 static nodeset *sta_collect_in_deps(ir_node *irn, nodeset *deps) {
1128         int i;
1129
1130         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1131                 ir_node *p = get_irn_in_or_dep(irn, i);
1132
1133                 if (is_Proj(p)) {
1134                         const ir_edge_t *edge;
1135
1136                         p = get_Proj_pred(p);
1137                         foreach_out_edge(p, edge) {
1138                                 ir_node *src = get_edge_src_irn(edge);
1139                                 nodeset_insert(deps, src);
1140                         }
1141                 }
1142                 else {
1143                         nodeset_insert(deps, p);
1144                 }
1145         }
1146
1147         return deps;
1148 }
1149
1150 /**
1151  * Create following ILP constraints:
1152  * - the assignment constraints:
1153  *     assure each node is executed once by exactly one (allowed) execution unit
1154  * - the dead node assignment constraints:
1155  *     assure a node can only die at most once
1156  * - the precedence constraints:
1157  *     assure that no data dependencies are violated
1158  */
1159 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1160         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1161         char                  buf[1024];
1162         ir_node               *irn;
1163         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1164         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1165         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1166         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1167
1168         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1169         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1170                 int                  cst, tp_idx;
1171                 unsigned             cur_var;
1172                 be_ilpsched_irn_t    *node;
1173                 ilpsched_node_attr_t *na;
1174                 ir_node              *pred;
1175                 nodeset              *deps = new_nodeset(16);
1176
1177                 node    = get_ilpsched_irn(env, irn);
1178                 na      = get_ilpsched_node_attr(node);
1179                 cur_var = 0;
1180
1181                 /* the assignment constraint */
1182                 ilp_timer_push(t_cst_assign);
1183                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1184                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1185                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1186                 num_cst_assign++;
1187
1188                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1189                 ilp_timer_pop();
1190
1191                 /* We have separate constraints for Projs and Keeps */
1192                 // ILP becomes infeasible ?!?
1193 //              if (is_Proj(irn) || be_is_Keep(irn))
1194 //                      continue;
1195
1196                 /* the precedence constraints */
1197                 ilp_timer_push(t_cst_prec);
1198                 bs_block_irns = bitset_clear_all(bs_block_irns);
1199
1200                 deps = sta_collect_in_deps(irn, deps);
1201                 foreach_nodeset(deps, pred) {
1202                         unsigned             t_low, t_high, t;
1203                         be_ilpsched_irn_t    *pred_node;
1204                         ilpsched_node_attr_t *pna;
1205                         unsigned             delay;
1206
1207                         pred = skip_normal_Proj(env->arch_env->isa, pred);
1208                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1209                                 continue;
1210
1211                         pred_node = get_ilpsched_irn(env, pred);
1212                         pna       = get_ilpsched_node_attr(pred_node);
1213
1214                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1215
1216                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1217                                 bitset_set(bs_block_irns, pna->block_idx);
1218                         else
1219                                 continue;
1220
1221                         /* irn = n, pred = m */
1222                         delay  = fixed_latency(env->sel, pred, env->block_env);
1223                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1224                         t_high = MIN(na->alap, pna->alap + delay - 1);
1225                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1226                                 unsigned tn, tm;
1227                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1228
1229                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1230                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1231                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1232                                 num_cst_prec++;
1233
1234                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1235                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1236                                         /* node variables have smaller index than pred variables */
1237                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1238                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1239                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1240                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1241                                                 }
1242                                         }
1243
1244                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1245                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1246                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1247                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1248                                                 }
1249                                         }
1250                                 }
1251                                 else {
1252                                         /* pred variables have smaller index than node variables */
1253                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1254                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1255                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1256                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1257                                                 }
1258                                         }
1259
1260                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1261                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1262                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1263                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1264                                                 }
1265                                         }
1266                                 }
1267
1268                                 if (ARR_LEN(tmp_var_idx) > 0)
1269                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1270
1271                                 DEL_ARR_F(tmp_var_idx);
1272                         }
1273                 }
1274                 del_nodeset(deps);
1275                 ilp_timer_pop();
1276         }
1277         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1278                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1279         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1280                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1281 }
1282
1283 /**
1284  * Create ILP resource constraints:
1285  * - assure that for each time step not more instructions are scheduled
1286  *   to the same unit types as units of this type are available
1287  */
1288 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1289         int                   glob_type_idx;
1290         char                  buf[1024];
1291         unsigned              num_cst_resrc = 0;
1292         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1293         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1294
1295         ilp_timer_push(t_cst_rsrc);
1296         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1297                 unsigned                 t;
1298                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1299
1300                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1301
1302                 /* check each time step */
1303                 for (t = 0; t < ba->max_steps; ++t) {
1304                         ir_node *irn;
1305                         int     cst;
1306                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1307
1308                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1309                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1310                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1311                         num_cst_resrc++;
1312
1313                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1314                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1315                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1316                                 int                  tp_idx;
1317
1318                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1319
1320                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1321                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1322                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1323                                 }
1324                         }
1325
1326                         /* set constraints if we have some */
1327                         if (ARR_LEN(tmp_var_idx) > 0)
1328                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1329
1330                         DEL_ARR_F(tmp_var_idx);
1331                 }
1332         }
1333         ilp_timer_pop();
1334         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1335                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1336 }
1337
1338 /**
1339  * Create ILP bundle constraints:
1340  * - assure, at most bundle_size * bundles_per_cycle instructions
1341  *   can be started at a certain point.
1342  */
1343 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1344         char                  buf[1024];
1345         unsigned              t;
1346         unsigned              num_cst_bundle = 0;
1347         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1348         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1349         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1350
1351         ilp_timer_push(t_cst_bundle);
1352         for (t = 0; t < ba->max_steps; ++t) {
1353                 ir_node *irn;
1354                 int     cst;
1355                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1356
1357                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1358                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1359                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1360                 num_cst_bundle++;
1361
1362                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1363                         be_ilpsched_irn_t    *node;
1364                         ilpsched_node_attr_t *na;
1365                         int                  tp_idx;
1366
1367                         /* Projs and Keeps do not contribute to bundle size */
1368                         if (is_Proj(irn) || be_is_Keep(irn))
1369                                 continue;
1370
1371                         node = get_ilpsched_irn(env, irn);
1372                         na   = get_ilpsched_node_attr(node);
1373
1374                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1375                         if (na->is_dummy_node)
1376                                 continue;
1377
1378                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1379                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1380                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1381                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1382                                 }
1383                         }
1384                 }
1385
1386                 if (ARR_LEN(tmp_var_idx) > 0)
1387                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1388
1389                 DEL_ARR_F(tmp_var_idx);
1390         }
1391         ilp_timer_pop();
1392         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1393                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1394 }
1395
1396 /**
1397  * Create ILP alive nodes constraints:
1398  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1399  */
1400 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1401         char                  buf[1024];
1402         ir_node               *irn;
1403         unsigned              num_cst = 0;
1404         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1405         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1406
1407         ilp_timer_push(t_cst);
1408         /* for each node */
1409         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1410                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1411                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1412                 unsigned             t;
1413
1414                 /* we ignore nodes assigned to dummy unit here */
1415                 if (na->is_dummy_node)
1416                         continue;
1417
1418                 /* check check all time steps: asap(n) <= t <= U */
1419                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1420                         int node_tp_idx;
1421
1422                         /* for all unit types available for this node */
1423                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1424                                 unsigned tn, tn_max, idx;
1425                                 int      cst, i;
1426                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1427                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1428
1429                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1430                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1431                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1432                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1433                                 num_cst++;
1434
1435                                 tn_max = MIN(na->alap - 1, t);
1436                                 /* check if the node has been scheduled so far */
1437                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1438                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1439                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1440                                 }
1441
1442                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1443                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1444                                 DEL_ARR_F(tmp_var_idx_n);
1445
1446                                 /* subtract the number of consumer scheduled so far */
1447                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1448                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1449                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1450                                         int                  tp_idx;
1451                                         unsigned             tm, tm_max;
1452
1453                                         tm_max = MIN(ca->alap - 1, t);
1454                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1455                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1456                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1457                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1458                                                 }
1459                                         }
1460                                 }
1461
1462                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1463                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1464                                 DEL_ARR_F(tmp_var_idx_m);
1465
1466                                 /* -c * a_{nt}^k */
1467                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1468                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1469
1470                         }
1471                 }
1472         }
1473         ilp_timer_pop();
1474         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1475                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1476 }
1477
1478 /**
1479  * Create ILP alive nodes constraints for live-in nodes:
1480  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1481  */
1482 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1483         char                  buf[1024];
1484         ilp_livein_node_t     *livein;
1485         unsigned              num_cst = 0;
1486         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1487         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1488
1489         ilp_timer_push(t_cst);
1490         /* for each node */
1491         foreach_pset(ba->livein_nodes, livein) {
1492                 ir_node              *irn  = livein->irn;
1493                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1494                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1495                 unsigned             t;
1496
1497                 /* check check all time steps: 0 <= t < max_alive_steps */
1498                 for (t = 0; t < livein->max_alive_steps; ++t) {
1499                         int node_tp_idx;
1500
1501                         /* for all unit types available for this node */
1502                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1503                                 const ir_edge_t *edge;
1504                                 unsigned idx;
1505                                 int      cst, num_block_user;
1506                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1507
1508                                 /* check the number of consumer scheduled so far */
1509                                 num_block_user = 0;
1510                                 foreach_out_edge(irn, edge) {
1511                                         ir_node              *user = get_edge_src_irn(edge);
1512                                         be_ilpsched_irn_t    *cons;
1513                                         ilpsched_node_attr_t *ca;
1514                                         int                  tp_idx;
1515                                         unsigned             tm, tm_max;
1516
1517                                         /* check only users within current block */
1518                                         if (get_nodes_block(user) != block_node->irn)
1519                                                 continue;
1520
1521                                         num_block_user++;
1522                                         cons = get_ilpsched_irn(env, user);
1523                                         ca   = get_ilpsched_node_attr(cons);
1524
1525                                         tm_max = MIN(ca->alap - 1, t);
1526                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1527                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1528                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1529                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1530                                                 }
1531                                         }
1532                                 }
1533
1534                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1535                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1536                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1537                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1538                                 num_cst++;
1539
1540                                 /* sum(scheduled users) */
1541                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1542                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1543                                 DEL_ARR_F(tmp_var_idx_m);
1544
1545                                 /* + c * a_{nt}^k */
1546                                 idx = node_tp_idx * livein->max_alive_steps + t;
1547                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1548                         }
1549                 }
1550         }
1551         ilp_timer_pop();
1552         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1553                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1554 }
1555
1556 /**
1557  * Create ILP pressure constraints, based on alive nodes:
1558  * - add additional costs to objective function if a node is scheduled
1559  *   on a unit although all units of this type are currently occupied
1560  */
1561 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1562         char                  buf[1024];
1563         ir_node               *cur_irn;
1564         unsigned              num_cst = 0;
1565         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1566         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1567
1568         ilp_timer_push(t_cst);
1569         /* y_{nt}^k is set for each node and timestep and unit type */
1570         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1571                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1572                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1573                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1574                 int                  glob_type_idx;
1575
1576                 /* we ignore nodes assigned to DUMMY unit here */
1577                 if (cur_na->is_dummy_node)
1578                         continue;
1579
1580                 /* for all types */
1581                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1582                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1583                         int                      cur_tp_idx;
1584                         unsigned                 t;
1585
1586                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1587
1588                         /* check if node can be executed on this unit type */
1589                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1590                         if (cur_tp_idx < 0)
1591                                 continue;
1592
1593                         /* check all time_steps at which the current node can be scheduled */
1594                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1595                                 int     cst, y_idx;
1596                                 ir_node *irn;
1597                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1598                                 ilp_livein_node_t *livein;
1599
1600                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1601                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1602                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1603                                 num_cst++;
1604
1605                                 /* - accumulate all nodes alive at point t on unit type k */
1606                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1607                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1608                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1609                                         int                  a_idx, tp_idx;
1610
1611                                         /* check if node can be alive here */
1612                                         if (t < na->asap - 1)
1613                                                 continue;
1614
1615                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1616
1617                                         /* current type is not suitable */
1618                                         if (tp_idx < 0)
1619                                                 continue;
1620
1621                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1622                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1623                                 }
1624                                 /* do the same for livein nodes */
1625                                 foreach_pset(ba->livein_nodes, livein) {
1626                                         ir_node              *irn  = livein->irn;
1627                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1628                                         int                  a_idx, tp_idx;
1629
1630                                         /* check if node can be alive here */
1631                                         if (t >= livein->max_alive_steps)
1632                                                 continue;
1633
1634                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1635
1636                                         /* current type is not suitable */
1637                                         if (tp_idx < 0)
1638                                                 continue;
1639
1640                                         a_idx = tp_idx * livein->max_alive_steps + t;
1641                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1642                                 }
1643
1644                                 if (ARR_LEN(tmp_var_idx) > 0)
1645                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1646                                 DEL_ARR_F(tmp_var_idx);
1647
1648                                 /* - num_nodes * y_{nt}^k */
1649                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1650                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1651                         }
1652                 }
1653         }
1654         ilp_timer_pop();
1655         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1656                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1657 }
1658
1659 /**
1660  * Create ILP branch constraints:
1661  * Assure, alle nodes are scheduled prior to cfg op.
1662  */
1663 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1664         char                  buf[1024];
1665         ir_node               *cur_irn, *cfop;
1666         unsigned              num_cst          = 0;
1667         unsigned              num_non_branches = 0;
1668         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1669         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1670
1671         ilp_timer_push(t_cst);
1672         cfop = NULL;
1673         /* determine number of non-branch nodes and the one and only branch node */
1674         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1675                 switch (get_irn_opcode(cur_irn)) {
1676                         case iro_Phi:
1677                         case iro_Start:
1678                         case iro_End:
1679                         case iro_Proj:
1680                         case iro_Bad:
1681                         case iro_Unknown:
1682                                 num_non_branches++;
1683                                 break;
1684                         default:
1685                                 if (is_cfop(cur_irn)) {
1686                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1687                                         cfop = cur_irn;
1688                                 }
1689                                 else {
1690                                         num_non_branches++;
1691                                 }
1692                                 break;
1693                 }
1694         }
1695
1696         if (cfop) {
1697                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1698                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1699                 unsigned t;
1700
1701                 /* for each time step */
1702                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1703                         int *non_branch_vars, *branch_vars;
1704                         int cst;
1705
1706                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1707                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1708                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1709                         num_cst++;
1710
1711                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1712                         non_branch_vars = NEW_ARR_F(int, 0);
1713                         branch_vars     = NEW_ARR_F(int, 0);
1714                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1715                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1716                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1717                                 int                  tp_idx;
1718
1719                                 if (cur_irn == cfop) {
1720                                         /* for all unit types available for this node */
1721                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1722                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1723                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1724                                         }
1725                                 }
1726                                 else {
1727                                         /* sum up all possible schedule points for this node upto current timestep */
1728                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1729                                                 unsigned tn;
1730                                                 unsigned tmax = MIN(t, na->alap - 1);
1731
1732                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1733                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1734                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1735                                                 }
1736                                         }
1737                                 }
1738
1739                         }
1740
1741                         if (ARR_LEN(non_branch_vars) > 0)
1742                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1743                         if (ARR_LEN(branch_vars) > 0)
1744                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1745
1746                         DEL_ARR_F(branch_vars);
1747                         DEL_ARR_F(non_branch_vars);
1748                 }
1749         }
1750         ilp_timer_pop();
1751         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1752                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1753 }
1754
1755 #if 0
1756 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1757         char                  buf[1024];
1758         ir_node               *irn;
1759         unsigned              num_cst = 0;
1760         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1761         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1762
1763         ilp_timer_push(t_cst);
1764         /* check all nodes */
1765         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1766                 be_ilpsched_irn_t    *node;
1767                 ilpsched_node_attr_t *na;
1768                 unsigned             t;
1769                 ir_node              **pk;
1770
1771                 /* only mode_T nodes can have Projs and Keeps assigned */
1772                 if (get_irn_mode(irn) != mode_T)
1773                         continue;
1774
1775                 node = get_ilpsched_irn(env, irn);
1776                 na   = get_ilpsched_node_attr(node);
1777
1778                 /* check if has some Projs and Keeps assigned */
1779                 if (! na->projkeeps)
1780                         continue;
1781
1782                 /* we can run only once over the queue, so preserve the nodes */
1783                 pk = NEW_ARR_F(ir_node *, 0);
1784                 while (! waitq_empty(na->projkeeps))
1785                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1786                 del_waitq(na->projkeeps);
1787                 na->projkeeps = NULL;
1788
1789                 /* for all time steps at which this node can be scheduled */
1790                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1791                         int cst, tp_idx, i;
1792                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1793
1794                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1795                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1796                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1797                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1798                         num_cst++;
1799
1800                         /* sum up scheduling variables for this time step */
1801                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1802                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1803                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1804                         }
1805
1806                         if (ARR_LEN(tmp_var_idx_n) > 0)
1807                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1808                         DEL_ARR_F(tmp_var_idx_n);
1809
1810                         /* subtract all Proj and Keep variables for this step */
1811                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1812                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1813                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1814                                 int                  pk_tp_idx;
1815
1816                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1817                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1818                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1819                                 }
1820                         }
1821                 }
1822         }
1823         ilp_timer_pop();
1824         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1825                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1826 }
1827 #endif /* if 0 */
1828
1829 /***************************************************
1830  *  _____ _      _____                    _
1831  * |_   _| |    |  __ \                  (_)
1832  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1833  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1834  *  _| |_| |____| |      | | | | | | (_| | | | | |
1835  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1836  *
1837  ***************************************************/
1838
1839 /**
1840  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1841  */
1842 static void create_ilp(ir_node *block, void *walk_env) {
1843         be_ilpsched_env_t     *env           = walk_env;
1844         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1845         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1846         FILE                  *logfile       = NULL;
1847         lpp_t                 *lpp           = NULL;
1848         int                   need_heur      = 0;
1849         struct obstack        var_obst;
1850         char                  name[1024];
1851
1852         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1853         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1854         DBG((env->dbg, 255, "=========================================\n\n"));
1855
1856         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1857                 block, ba->n_interesting_nodes, ba->max_steps));
1858
1859         /* notify backend and get block environment */
1860         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1861
1862         /* if we have less than two interesting nodes, there is no need to create the ILP */
1863         if (ba->n_interesting_nodes > 1) {
1864                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1865                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1866                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1867                 int    estimated_n_var = (int)((double)base_num * fact_var);
1868                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1869
1870                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1871                         estimated_n_var, estimated_n_cst));
1872
1873                 /* set up the LPP object */
1874                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1875
1876                 lpp = new_lpp_userdef(
1877                         (const char *)name,
1878                         lpp_minimize,
1879                         estimated_n_cst,     /* num vars */
1880                         estimated_n_cst + 1, /* num cst */
1881                         1.3);                /* grow factor */
1882                 obstack_init(&var_obst);
1883
1884                 /* create ILP variables */
1885                 create_variables(env, lpp, block_node, &var_obst);
1886
1887                 /* create ILP constraints */
1888                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1889                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1890                 create_ressource_constraints(env, lpp, block_node);
1891                 create_bundle_constraints(env, lpp, block_node);
1892                 create_branch_constraint(env, lpp, block_node);
1893                 //create_proj_keep_constraints(env, lpp, block_node);
1894
1895                 if (env->opts->regpress) {
1896                         create_alive_nodes_constraint(env, lpp, block_node);
1897                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1898                         create_pressure_alive_constraint(env, lpp, block_node);
1899                 }
1900
1901                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1902
1903                 /* debug stuff, dump lpp when debugging is on  */
1904                 DEBUG_ONLY({
1905                         if (firm_dbg_get_mask(env->dbg) > 1) {
1906                                 char buf[1024];
1907                                 FILE *f;
1908
1909                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1910                                 f = fopen(buf, "w");
1911                                 lpp_dump_plain(lpp, f);
1912                                 fclose(f);
1913                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1914                                 lpp_dump(lpp, buf);
1915                         }
1916                 })
1917
1918                 /* set solve time limit */
1919                 lpp_set_time_limit(lpp, env->opts->time_limit);
1920
1921                 /* set logfile if requested */
1922                 if (strlen(env->opts->log_file) > 0) {
1923                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1924                                 lpp_set_log(lpp, stdout);
1925                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1926                                 lpp_set_log(lpp, stderr);
1927                         else {
1928                                 logfile = fopen(env->opts->log_file, "w");
1929                                 if (! logfile)
1930                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1931                                 else
1932                                         lpp_set_log(lpp, logfile);
1933                         }
1934                 }
1935
1936                 /* solve the ILP */
1937                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1938
1939                 if (logfile)
1940                         fclose(logfile);
1941
1942                 /* check for valid solution */
1943                 if (! lpp_is_sol_valid(lpp)) {
1944                         char buf[1024];
1945                         FILE *f;
1946
1947                         DEBUG_ONLY({
1948                                 if (firm_dbg_get_mask(env->dbg) >= 2) {
1949                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1950                                         f = fopen(buf, "w");
1951                                         lpp_dump_plain(lpp, f);
1952                                         fclose(f);
1953                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1954                                         lpp_dump(lpp, buf);
1955                                         dump_ir_block_graph(env->irg, "-infeasible");
1956                                 }
1957                         })
1958
1959                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1960                         need_heur = 1;
1961                 }
1962
1963                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1964                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1965                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1966                 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));
1967                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1968                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1969                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1970                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1971
1972                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1973         }
1974
1975         /* apply solution */
1976         be_stat_ev("nodes", ba->block_last_idx);
1977         be_stat_ev("vars", lpp ? lpp->var_next : 0);
1978         be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1979         if (need_heur) {
1980                         be_stat_ev("time", -1);
1981                         be_stat_ev_dbl("opt", 0.0);
1982                 list_sched_single_block(env->birg, block, env->be_opts);
1983         }
1984         else {
1985                 if (lpp) {
1986                         double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1987                         be_stat_ev_dbl("time", lpp->sol_time);
1988                         be_stat_ev_dbl("opt", opt);
1989                 }
1990                 else {
1991                         be_stat_ev_dbl("time", 0.0);
1992                         be_stat_ev_dbl("opt", 100.0);
1993                 }
1994                 apply_solution(env, lpp, block);
1995         }
1996
1997         if (lpp)
1998                 free_lpp(lpp);
1999
2000         /* notify backend */
2001         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2002 }
2003
2004 /**
2005  * Perform ILP scheduling on the given irg.
2006  */
2007 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2008         be_ilpsched_env_t          env;
2009         const char                 *name     = "be ilp scheduling";
2010         ir_graph                   *irg      = be_get_birg_irg(birg);
2011         const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
2012         const arch_isa_t           *isa      = arch_env->isa;
2013         const ilp_sched_selector_t *sel      = isa->impl->get_ilp_sched_selector(isa);
2014
2015         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2016
2017         stat_ev_ctx_push("phase", "ilpsched");
2018
2019 //      firm_dbg_set_mask(env.dbg, 1);
2020
2021         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2022         env.sel        = sel;
2023         env.irg        = irg;
2024         env.height     = heights_new(irg);
2025         env.main_env   = birg->main_env;
2026         env.arch_env   = arch_env;
2027         env.cpu        = arch_isa_get_machine(arch_env->isa);
2028         env.opts       = &ilp_opts;
2029         env.birg       = birg;
2030         env.be_opts    = be_opts;
2031         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2032
2033         /* assign a unique per block number to all interesting nodes */
2034         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2035
2036         /*
2037                 The block indices are completely build after the walk,
2038                 now we can allocate the bitsets (size depends on block indices)
2039                 for all nodes.
2040         */
2041         phase_reinit_irn_data(&env.ph);
2042
2043         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2044         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2045
2046         /* Calculate ALAP of all irns */
2047         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2048
2049         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2050         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2051
2052         /* perform ILP scheduling */
2053         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2054
2055         DEBUG_ONLY(
2056                 if (firm_dbg_get_mask(env.dbg)) {
2057                         phase_stat_t stat;
2058                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2059
2060                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2061                 }
2062         );
2063
2064         /* free data allocated dynamically */
2065         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2066
2067         /* free all allocated object */
2068         phase_free(&env.ph);
2069         heights_free(env.height);
2070
2071         /* notify backend */
2072         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2073
2074         stat_ev_ctx_pop();
2075 }
2076
2077 /**
2078  * Register ILP scheduler options.
2079  */
2080 void be_init_ilpsched(void)
2081 {
2082         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2083         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2084
2085         lc_opt_add_table(sched_grp, ilpsched_option_table);
2086 }
2087
2088 #else /* WITH_ILP */
2089
2090 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2091 {}
2092
2093 #endif /* WITH_ILP */