Just made a warning go away when building optimized version of the compiler
[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 #include "firm_config.h"
37
38 #ifdef WITH_ILP
39
40 #include <math.h>
41
42 #ifndef _WIN32
43 #include <strings.h>
44 #endif /* _WIN32 */
45
46 #include "irnode_t.h"
47 #include "irgwalk.h"
48 #include "irbitset.h"
49 #include "irphase_t.h"
50 #include "height.h"
51 #include "iredges.h"
52 #include "pdeq.h"
53 #include "debug.h"
54 #include "irtools.h"
55 #include "irdump.h"
56 #include "irprintf.h"
57 #include "plist.h"
58 #include "irprintf.h"
59
60 #include <lpp/lpp.h>
61 #include <lpp/lpp_net.h>
62
63 #include <libcore/lc_opts.h>
64 #include <libcore/lc_opts_enum.h>
65 #include <libcore/lc_timing.h>
66
67 #include "be.h"
68 #include "benode_t.h"
69 #include "besched_t.h"
70 #include "beilpsched.h"
71 #include "beutil.h"
72 #include "bestat.h"
73 #include "beirg_t.h"
74 #include "benodesets.h"
75
76 typedef struct _ilpsched_options_t {
77         unsigned regpress;
78         unsigned time_limit;
79         char     log_file[1024];
80 } ilpsched_options_t;
81
82 typedef struct _unit_type_info_t {
83         int                            n_units;
84         const be_execution_unit_type_t *tp;
85 } unit_type_info_t;
86
87 /**
88  * holding the ILP variables of the different types
89  */
90 typedef struct _ilp_var_types_t {
91         int *x;   /* x_{nt}^k variables */
92         int *a;   /* a_{nt}^k variables */
93         int *y;   /* y_{nt}^k variables */
94 } ilp_var_types_t;
95
96 /**
97  * Holds alive variables for a node live-in to a block.
98  */
99 typedef struct _ilp_livein_node_t {
100         ir_node  *irn;
101         unsigned max_alive_steps;
102         int      *a;
103 } ilp_livein_node_t;
104
105 /* attributes for a node */
106 typedef struct _ilpsched_node_attr_t {
107         unsigned asap;                     /**< The ASAP scheduling control step */
108         unsigned alap;                     /**< The ALAP scheduling control step */
109         unsigned latency;                  /**< Latency of this node (needed for sorting) */
110         unsigned sched_point;              /**< the step in which the node is finally scheduled */
111         unsigned visit_idx;                /**< Index of the node having visited this node last */
112         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
113         unsigned n_consumer;               /**< Number of consumers */
114         ir_node  **block_consumer;         /**< List of consumer being in the same block */
115         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
116         unsigned block_idx     : 30;       /**< A unique per block index */
117         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
118         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
119         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
120                                                                                         for ASAP, successors for ALAP */
121         unsigned         n_unit_types;     /**< number of allowed execution unit types */
122         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
123         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
124 } ilpsched_node_attr_t;
125
126 /* attributes for a block */
127 typedef struct _ilpsched_block_attr_t {
128         unsigned block_last_idx;        /**< The highest node index in block so far */
129         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
130         unsigned max_steps;             /**< Upper bound for block execution */
131         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
132         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
133         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
134 } ilpsched_block_attr_t;
135
136 typedef union _ilpsched_attr_ {
137         ilpsched_node_attr_t  node_attr;
138         ilpsched_block_attr_t block_attr;
139 } ilpsched_attr_t;
140
141 /* A irn for the phase and it's attributes (either node or block) */
142 typedef struct {
143         ir_node         *irn;
144         ilpsched_attr_t attr;
145 } be_ilpsched_irn_t;
146
147 /* The ILP scheduling environment */
148 typedef struct {
149         ir_phase             ph;            /**< The phase */
150         ir_graph             *irg;          /**< The current irg */
151         heights_t            *height;       /**< The heights object of the irg */
152         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
153         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
154         const arch_env_t     *arch_env;
155         const arch_isa_t     *isa;          /**< The ISA */
156         const be_main_env_t  *main_env;
157         const be_machine_t   *cpu;          /**< the current abstract machine */
158         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
159         const be_irg_t       *birg;         /**< The birg object */
160         be_options_t         *be_opts;      /**< backend options */
161         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
162         DEBUG_ONLY(firm_dbg_module_t *dbg);
163 } be_ilpsched_env_t;
164
165 /* convenience macros to handle phase irn data */
166 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
167 #define is_ilpsched_block(node)             (is_Block((node)->irn))
168 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
169 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
170
171 /* check if node is considered for ILP scheduling */
172 #define consider_for_sched(isa, irn) \
173         (! (is_Block(irn)            ||  \
174                 is_normal_Proj(isa, irn) ||  \
175                 is_Phi(irn)              ||  \
176                 is_NoMem(irn)            ||  \
177                 is_Unknown(irn)          ||  \
178                 is_End(irn)                  \
179                 ))
180
181 /* gives the valid scheduling time step interval for a node */
182 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
183
184 /* gives the valid interval where a node can die */
185 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
186
187 /* gives the corresponding ILP variable for given node, unit and time step */
188 #define ILPVAR_IDX(na, unit, control_step) \
189         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
190
191 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
192 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
193         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
194
195 /* check if a double value is within an epsilon environment of 0 */
196 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
197
198 #define ilp_timer_push(t)         lc_timer_push((t))
199 #define ilp_timer_pop()           lc_timer_pop()
200 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
201
202 /* option variable */
203 static ilpsched_options_t ilp_opts = {
204         1,     /* default is with register pressure constraints */
205         300,   /* 300 sec per block time limit */
206         ""     /* no log file */
207 };
208
209 /* ILP options */
210 static const lc_opt_table_entry_t ilpsched_option_table[] = {
211         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
212         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
213         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
214         LC_OPT_LAST
215 };
216
217 /*
218         We need this global variable as we compare nodes dependent on heights,
219         but we cannot pass any information to the qsort compare function.
220 */
221 static heights_t *glob_heights;
222
223 /**
224  * Check if irn is a Proj, which has no execution units assigned.
225  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
226  */
227 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
228         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
229 }
230
231 /**
232  * Skips normal Projs.
233  * @return predecessor if irn is a normal Proj, otherwise irn.
234  */
235 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
236         if (is_normal_Proj(isa, irn))
237                 return get_Proj_pred(irn);
238         return irn;
239 }
240
241 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
242         unsigned lat = be_ilp_sched_latency(sel, irn, env);
243         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
244                 lat = 1;
245         return lat;
246 }
247
248 static int cmp_live_in_nodes(const void *a, const void *b) {
249         const ilp_livein_node_t *n1 = a;
250         const ilp_livein_node_t *n2 = b;
251
252         return n1->irn != n2->irn;
253 }
254
255 /**
256  * Compare scheduling time steps of two be_ilpsched_irn's.
257  */
258 static int cmp_ilpsched_irn(const void *a, const void *b) {
259         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
260         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
261         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
262         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
263
264         if (n1_a->sched_point == n2_a->sched_point) {
265                 ir_node *irn_a = n1->irn;
266                 ir_node *irn_b = n2->irn;
267
268                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
269                         return 1;
270                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
271                         return -1;
272
273                 /*
274                         Ok, timestep is equal and the nodes are parallel,
275                         so check latency and schedule high latency first.
276                 */
277                 return QSORT_CMP(n2_a->latency, n1_a->latency);
278         }
279         else
280                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
281 }
282
283 /**
284  * In case there is no phase information for irn, initialize it.
285  */
286 static void *init_ilpsched_irn(ir_phase *ph, ir_node *irn, void *old) {
287         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
288
289         if (res == old) {
290                 /* if we have already some data: check for reinitialization */
291
292                 if (! is_Block(irn)) {
293                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
294
295                         if (! na->transitive_block_nodes) {
296                                 ir_node               *block      = get_nodes_block(irn);
297                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
298                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
299
300                                 /* we are called after the block indices have been build: create bitset */
301                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
302                         }
303                         else {
304                                 /* we are called from reinit block data: clear the bitset */
305                                 bitset_clear_all(na->transitive_block_nodes);
306                                 na->visit_idx    = 0;
307                                 na->alap_changed = 1;
308                         }
309                 }
310                 return old;
311         }
312
313         res->irn = irn;
314
315         /* set ilpsched irn attributes (either block or irn) */
316         if (is_Block(irn)) {
317                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
318
319                 ba->n_interesting_nodes = 0;
320                 ba->block_last_idx      = 0;
321                 ba->root_nodes          = plist_new();
322                 ba->head_ilp_nodes      = NULL;
323                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
324                 ba->max_steps           = 0;
325         }
326         else {
327                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
328                 memset(na, 0, sizeof(*na));
329         }
330
331         return res;
332 }
333
334 /**
335  * Assign a per block unique number to each node.
336  */
337 static void build_block_idx(ir_node *irn, void *walk_env) {
338         be_ilpsched_env_t     *env = walk_env;
339         be_ilpsched_irn_t     *node, *block_node;
340         ilpsched_node_attr_t  *na;
341         ilpsched_block_attr_t *ba;
342
343         set_irn_link(irn, NULL);
344         if (! consider_for_sched(env->arch_env->isa, irn))
345                 return;
346
347         node       = get_ilpsched_irn(env, irn);
348         na         = get_ilpsched_node_attr(node);
349         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
350         ba         = get_ilpsched_block_attr(block_node);
351
352         na->block_idx = ba->block_last_idx++;
353 }
354
355 /********************************************************
356  *                              __        _
357  *                             / /       | |
358  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
359  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
360  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
361  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
362  *                 | |                           | |
363  *                 |_|                           |_|
364  ********************************************************/
365
366 /**
367  * Add all nodes having no user in current block to last_nodes list.
368  */
369 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
370         ir_node               *block;
371         const ir_edge_t       *edge;
372         be_ilpsched_irn_t     *block_node, *node;
373         ilpsched_block_attr_t *ba;
374         ilpsched_node_attr_t  *na;
375         int                   i, j;
376         be_ilpsched_env_t     *env           = walk_env;
377         int                   has_block_user = 0;
378         unsigned              n_consumer     = 0;
379         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
380         ir_node               **consumer;
381         unsigned              idx;
382
383         if (! consider_for_sched(env->arch_env->isa, irn))
384                 return;
385
386         block    = get_nodes_block(irn);
387     idx      = get_irn_idx(irn);
388         consumer = NEW_ARR_F(ir_node *, 0);
389
390         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
391
392         /* check data and dependency out edges */
393         for (i = 0; i < 2 && ! has_block_user; ++i) {
394                 foreach_out_edge_kind(irn, edge, ekind[i]) {
395                         ir_node *user = get_edge_src_irn(edge);
396
397                         if (is_normal_Proj(env->arch_env->isa, user)) {
398                                 const ir_edge_t *user_edge;
399
400                                 if (get_irn_mode(user) == mode_X)
401                                         continue;
402
403                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
404                                 for (j = 0; j < 2; ++j) {
405                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
406                                                 ir_node *real_user = get_edge_src_irn(user_edge);
407
408                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
409                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
410                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
411
412                                                         /* skip already visited nodes */
413                                                         if (ua->consumer_idx == idx)
414                                                                 continue;
415
416                                                         /* check if node has user in this block and collect the user if it's a data user */
417                                                         if (get_nodes_block(real_user) == block) {
418                                                                 if (i == 0 && j == 0)
419                                                                         ARR_APP1(ir_node *, consumer, real_user);
420                                                                 has_block_user = 1;
421                                                         }
422
423                                                         /* only count data consumer */
424                                                         if (i == 0)
425                                                                 n_consumer++;
426
427                                                         /* mark user as visited by this node */
428                                                         ua->consumer_idx = idx;
429                                                 }
430                                         }
431                                 }
432                         }
433                         else if (is_Block(user)) {
434                                 continue;
435                         }
436                         else if (! is_Phi(user)) {
437                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
438                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
439
440                                 /* skip already visited nodes */
441                                 if (ua->consumer_idx == idx)
442                                         continue;
443
444                                 /* check if node has user in this block and collect the user if it's a data user */
445                                 if (get_nodes_block(user) == block) {
446                                         if (i == 0)
447                                                 ARR_APP1(ir_node *, consumer, user);
448                                         has_block_user = 1;
449                                 }
450
451                                 /* only count data consumer */
452                                 if (i == 0)
453                                         n_consumer++;
454
455                                 /* mark user visited by this node */
456                                 ua->consumer_idx = idx;
457                         }
458                         else if (get_nodes_block(user) != block) {
459                                 n_consumer++;
460                         }
461                 }
462         }
463
464         block_node = get_ilpsched_irn(env, block);
465         ba         = get_ilpsched_block_attr(block_node);
466
467         ba->n_interesting_nodes++;
468
469         /* current irn has no user inside this block, add to queue */
470         if (! has_block_user) {
471                 DB((env->dbg, LEVEL_3, "root node\n"));
472                 plist_insert_back(ba->root_nodes, irn);
473         }
474         else {
475                 DB((env->dbg, LEVEL_3, "normal node\n"));
476         }
477
478         /* record number of all consumer and the consumer within the same block */
479         node = get_ilpsched_irn(env, irn);
480         na   = get_ilpsched_node_attr(node);
481         na->n_consumer     = n_consumer;
482         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
483         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
484         DEL_ARR_F(consumer);
485 }
486
487 /**
488  * Calculate the ASAP scheduling step for current irn.
489  */
490 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
491         be_ilpsched_env_t     *env = walk_env;
492         int                   i;
493         ir_node               *block;
494         be_ilpsched_irn_t     *node, *block_node;
495         ilpsched_node_attr_t  *na;
496         ilpsched_block_attr_t *ba;
497
498         /* These nodes are handled separate */
499         if (! consider_for_sched(env->arch_env->isa, irn))
500                 return;
501
502         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
503
504         block    = get_nodes_block(irn);
505         node     = get_ilpsched_irn(env, irn);
506         na       = get_ilpsched_node_attr(node);
507         na->asap = 1;
508
509         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
510                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
511
512                 /* check for greatest distance to top */
513                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
514                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
515                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
516                         unsigned             lat;
517
518                         lat         = fixed_latency(env->sel, pred, env->block_env);
519                         na->latency = lat;
520                         na->asap    = MAX(na->asap, pna->asap + lat);
521                 }
522         }
523
524         /* add node to ILP node list and update max_steps */
525         block_node = get_ilpsched_irn(env, block);
526         ba         = get_ilpsched_block_attr(block_node);
527
528         set_irn_link(irn, ba->head_ilp_nodes);
529         ba->head_ilp_nodes = irn;
530         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
531
532         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
533 }
534
535 /**
536  * Calculate the ALAP scheduling step of all irns in current block.
537  * Depends on max_steps being calculated.
538  */
539 static void calculate_block_alap(ir_node *block, void *walk_env) {
540         be_ilpsched_env_t     *env        = walk_env;
541         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
542         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
543         waitq                 *cur_queue  = new_waitq();
544         plist_element_t       *el;
545
546         assert(is_Block(block));
547
548         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
549                 block, ba->n_interesting_nodes, ba->max_steps));
550
551         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
552         //phase_reinit_block_irn_data(&env->ph, block);
553
554         /* init start queue */
555         foreach_plist(ba->root_nodes, el) {
556                 waitq_put(cur_queue, plist_element_get_value(el));
557         }
558
559         /* repeat until all nodes are processed */
560         while (! waitq_empty(cur_queue)) {
561                 waitq *next_queue = new_waitq();
562
563                 /* process all nodes in current step */
564                 while (! waitq_empty(cur_queue)) {
565                         ir_node              *cur_irn = waitq_get(cur_queue);
566                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
567                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
568                         int                  i;
569
570                         /* cur_node has no alap set -> it's a root node, set to max alap */
571                         if (na->alap == 0) {
572                                 na->alap = ba->max_steps;
573                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
574                                         cur_irn, na->alap));
575                         }
576                         else {
577                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
578                                         cur_irn, na->alap));
579                         }
580
581                         /* set the alap's of all predecessors */
582                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
583                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
584
585                                 /* check for greatest distance to bottom */
586                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
587                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
588                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
589                                         unsigned             lat;
590
591                                         /* mark the predecessor as visited by current irn */
592                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
593                                                 continue;
594                                         pna->visit_idx = get_irn_idx(cur_irn);
595
596                                         lat = fixed_latency(env->sel, pred, env->block_env);
597
598                                         /* set ALAP of current pred */
599                                         if (pna->alap == 0) {
600                                                 /* current ALAP is 0: node has not yet been visited */
601                                                 pna->alap_changed = 1;
602                                                 pna->alap         = na->alap - lat;
603                                         }
604                                         else if (pna->alap > na->alap - lat) {
605                                                 /* we found a longer path to root node: change ALAP */
606                                                 pna->alap         = na->alap - lat;
607                                                 pna->alap_changed = 1;
608                                         }
609                                         else {
610                                                 /* current ALAP is best found so far: keep it */
611                                                 pna->alap_changed = 0;
612                                         }
613
614                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
615
616                                         /* enqueue node for next iteration */
617                                         if (get_irn_ins_or_deps(pred) > 0)
618                                                 waitq_put(next_queue, pred);
619                                 }
620                         }
621                 }
622
623                 /* prepare for next iteration */
624                 del_waitq(cur_queue);
625                 cur_queue = next_queue;
626         }
627 }
628
629 /**
630  * Free list of root nodes and the set of live-in nodes.
631  */
632 static void clear_unwanted_data(ir_node *block, void *walk_env) {
633         be_ilpsched_env_t     *env        = walk_env;
634         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
635         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
636
637         plist_free(ba->root_nodes);
638         ba->root_nodes = NULL;
639         del_pset(ba->livein_nodes);
640         ba->livein_nodes = NULL;
641 }
642
643 /**
644  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
645  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
646  */
647 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
648         be_ilpsched_env_t    *env  = walk_env;
649         ir_node              *pred = irn;
650         be_ilpsched_irn_t    *node, *pred_node;
651         ilpsched_node_attr_t *na, *pna;
652
653         if (! consider_for_sched(env->arch_env->isa, irn))
654                 return;
655
656         if (! is_Proj(irn) && ! be_is_Keep(irn))
657                 return;
658
659         /* go to the ancestor */
660         if (be_is_Keep(irn))
661                 pred = get_irn_n(irn, 0);
662         pred = skip_Proj(pred);
663
664         node      = get_ilpsched_irn(env, irn);
665         pred_node = get_ilpsched_irn(env, pred);
666         na        = get_ilpsched_node_attr(node);
667         pna       = get_ilpsched_node_attr(pred_node);
668
669         na->asap = pna->asap;
670         na->alap = pna->alap;
671
672         /* record all Projs and Keeps for this node */
673         if (! pna->projkeeps)
674                 pna->projkeeps = new_waitq();
675         waitq_put(pna->projkeeps, irn);
676
677         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
678 }
679
680 /*******************************************
681  *           _              _       _
682  *          | |            | |     | |
683  *  ___  ___| |__   ___  __| |_   _| | ___
684  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
685  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
686  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
687  *
688  *******************************************/
689
690 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
691         const ir_edge_t *edge;
692
693         foreach_out_edge(irn, edge) {
694                 ir_node *user = get_edge_src_irn(edge);
695
696                 if (be_is_Keep(user)) {
697                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
698                         waitq_put(keeps, user);
699                 }
700         }
701 }
702
703 /**
704  * Inserts @p irn before @p before into schedule and notifies backend.
705  */
706 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
707         ir_node *before, ir_node *irn, unsigned cycle)
708 {
709         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
710         sched_add_before(before, irn);
711 }
712
713 /**
714  * Adds a node, it's Projs (in case of mode_T nodes) and
715  * it's Keeps to schedule.
716  */
717 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
718         const ir_edge_t *edge;
719         waitq           *keeps = new_waitq();
720
721         /* mode_M nodes are not scheduled */
722         if (get_irn_mode(irn) == mode_M)
723                 return;
724
725         if (! sched_is_scheduled(irn))
726                 notified_sched_add_before(env, block, irn, cycle);
727
728         /* add Projs */
729         if (get_irn_mode(irn) == mode_T) {
730                 foreach_out_edge(irn, edge) {
731                         ir_node *user = get_edge_src_irn(edge);
732
733                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
734                                 get_irn_n_edges(user) > 0)
735                         {
736                                 notified_sched_add_before(env, block, user, cycle);
737                         }
738
739                         check_for_keeps(keeps, block, user);
740                 }
741         }
742         else {
743                 check_for_keeps(keeps, block, irn);
744         }
745
746         /* add Keeps */
747         while (! waitq_empty(keeps)) {
748                 ir_node *keep = waitq_get(keeps);
749                 if (! sched_is_scheduled(keep))
750                         notified_sched_add_before(env, block, keep, cycle);
751         }
752
753         del_waitq(keeps);
754 }
755
756 /**
757  * Schedule all nodes in the given block, according to the ILP solution.
758  */
759 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
760         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
761         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
762         sched_info_t          *info       = get_irn_sched_info(block);
763         be_ilpsched_irn_t     **sched_nodes;
764         unsigned              i, l;
765         ir_node               *cfop, *irn;
766         const ir_edge_t       *edge;
767
768         /* init block schedule list */
769         INIT_LIST_HEAD(&info->list);
770         info->scheduled = 1;
771
772         /* collect nodes and their scheduling time step */
773         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
774         if (ba->n_interesting_nodes == 0) {
775                 /* ignore */
776         }
777         else if (ba->n_interesting_nodes == 1) {
778                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
779
780                 /* add the single node */
781                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
782         }
783         else {
784                 /* check all nodes for their positive solution */
785                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
786                         be_ilpsched_irn_t    *node;
787                         ilpsched_node_attr_t *na;
788                         int                  tp_idx, found;
789                         unsigned             cur_var, t;
790
791                         node    = get_ilpsched_irn(env, irn);
792                         na      = get_ilpsched_node_attr(node);
793                         cur_var = 0;
794                         found   = 0;
795
796                         if (! na->is_dummy_node) {
797                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
798                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
799                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
800
801                                                 if (! LPP_VALUE_IS_0(cost)) {
802                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
803                                                         found = 1;
804                                                 }
805                                         }
806                                 }
807                         }
808
809                         found = 0;
810                         /* go over all variables of a node until the non-zero one is found */
811                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
812                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
813                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
814
815                                         /* check, if variable is set to one (it's not zero then :) */
816                                         if (! LPP_VALUE_IS_0(val)) {
817                                                 na->sched_point = t;
818                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
819                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
820                                                         irn, t, na->type_info[tp_idx].tp->name));
821                                                 found = 1;
822                                         }
823                                 }
824                         }
825                 }
826
827                 glob_heights = env->height;
828                 /* sort nodes ascending by scheduling time step */
829                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
830         }
831
832         /* make all Phis ready and remember the single cf op */
833         cfop = NULL;
834         foreach_out_edge(block, edge) {
835                 irn = get_edge_src_irn(edge);
836
837                 switch (get_irn_opcode(irn)) {
838                         case iro_Phi:
839                                 add_to_sched(env, block, irn, 0);
840                                 break;
841                         case iro_Start:
842                         case iro_End:
843                         case iro_Proj:
844                         case iro_Bad:
845                         case iro_Unknown:
846                                 break;
847                         default:
848                                 if (is_cfop(irn)) {
849                                         assert(cfop == NULL && "Highlander - there can be only one");
850                                         cfop = irn;
851                                 }
852                         break;
853                 }
854         }
855
856         /* add all nodes from list */
857         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
858                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
859                 if (sched_nodes[i]->irn != cfop)
860                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
861         }
862
863         /* schedule control flow node if not already done */
864         if (cfop && ! sched_is_scheduled(cfop))
865                 add_to_sched(env, block, cfop, 0);
866
867         DEL_ARR_F(sched_nodes);
868 }
869
870 /***************************************************************
871  *   _____ _      _____     _____           _   _
872  *  |_   _| |    |  __ \   / ____|         | | (_)
873  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
874  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
875  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
876  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
877  *
878  ***************************************************************/
879
880 /**
881  * Check if node can be executed on given unit type.
882  */
883 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
884         int                  i;
885         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
886
887         for (i = na->n_unit_types - 1; i >= 0; --i) {
888                 if (na->type_info[i].tp == tp)
889                         return i;
890         }
891
892         return -1;
893 }
894
895 /************************************************
896  *                   _       _     _
897  *                  (_)     | |   | |
898  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
899  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
900  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
901  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
902  *
903  ************************************************/
904
905 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
906         const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
907         unsigned                  n_unit_types = 0;
908         be_ilpsched_irn_t         *node;
909         ilpsched_node_attr_t      *na;
910         unsigned                  unit_idx, tp_idx;
911
912         /* count number of available unit types for this node */
913         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
914                 /* just count */ ;
915
916         node = get_ilpsched_irn(env, irn);
917         na   = get_ilpsched_node_attr(node);
918
919         if (! na->type_info) {
920                 na->n_unit_types = n_unit_types;
921                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
922
923                 /* fill the type info array */
924                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
925                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
926                                 /* beware: we also count number of available units here */
927                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
928                                         na->is_dummy_node = 1;
929                         }
930
931                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
932                         na->type_info[tp_idx].n_units = unit_idx;
933                 }
934         }
935
936         return n_unit_types;
937 }
938
939 /**
940  * Returns the largest alap time of a user of @p irn.
941  * The user must be in block @p block.
942  */
943 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
944         const ir_edge_t *edge;
945         unsigned        max_alap = 0;
946
947         foreach_out_edge(irn, edge) {
948                 ir_node *user = get_edge_src_irn(edge);
949
950                 if (get_nodes_block(user) == block) {
951                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
952                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
953
954                         max_alap = MAX(max_alap, na->alap);
955                 }
956         }
957
958         assert(max_alap > 0);
959         return max_alap;
960 }
961
962 /**
963  * Create the following variables:
964  * - x_{nt}^k    binary     weigthed with: t
965  *      node n is scheduled at time step t to unit type k
966  * ==>> These variables represent the schedule
967  *
968  * - a_{nt}^k    binary     weighted with num_nodes
969  *      node n is alive at time step t on unit type k
970  *
971  * - y_{nt}^k    continuous  weighted with: num_nodes^2
972  *      register pressure over limit for unit type k
973  * ==>> These variables represent the register pressure
974  *
975  */
976 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
977         char                  buf[1024];
978         ir_node               *irn;
979         unsigned              num_block_var, num_nodes;
980         ilp_livein_node_t     *livein;
981         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
982         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
983         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
984
985         ilp_timer_push(t_var);
986         num_block_var = num_nodes = 0;
987         foreach_linked_irns(ba->head_ilp_nodes, irn) {
988                 be_ilpsched_irn_t    *node;
989                 ilpsched_node_attr_t *na;
990                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
991                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
992                 int                  i;
993
994                 node         = get_ilpsched_irn(env, irn);
995                 na           = get_ilpsched_node_attr(node);
996                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
997
998                 /* allocate space for ilp variables */
999                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1000                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
1001
1002                 /* we need these variables only for "real" nodes */
1003                 if (! na->is_dummy_node) {
1004                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
1005                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
1006
1007                         num_ad         = ba->max_steps - na->asap + 1;
1008                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
1009                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
1010                 }
1011
1012                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
1013                         irn, na->asap, na->alap, na->n_unit_types));
1014
1015                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
1016                 /* create variables */
1017                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
1018                         unsigned t;
1019
1020                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1021                                 /* x_{nt}^k variables */
1022                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
1023                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1024                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1025                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1026                                 /* variable counter */
1027                                 n_var++;
1028                                 num_block_var++;
1029
1030                                 if (! na->is_dummy_node) {
1031                                         /* y_{nt}^k variables */
1032                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1033                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1034                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1035                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1036
1037                                         /* variable counter */
1038                                         n_var++;
1039                                         num_block_var++;
1040                                 }
1041                         }
1042
1043                         /* a node can die at any step t: asap(n) <= t <= U */
1044                         if (! na->is_dummy_node) {
1045                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1046
1047                                         /* a_{nt}^k variables */
1048                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1049                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1050                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1051                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1052
1053                                         /* variable counter */
1054                                         n_var++;
1055                                         num_block_var++;
1056                                 }
1057                         }
1058
1059                         /* collect live-in nodes */
1060                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1061                                 ir_node *pred = get_irn_n(irn, i);
1062
1063                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1064                                         be_ilpsched_set_type_info(env, pred, var_obst);
1065                                         if (! na->is_dummy_node) {
1066                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1067                                                 entry->irn = pred;
1068                                                 entry->a   = NULL;
1069                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1070                                         }
1071                                 }
1072                         }
1073                 }
1074
1075                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1076                 num_nodes++;
1077         }
1078
1079         /* create alive variables a_{nt}^k for live-ins */
1080         foreach_pset(ba->livein_nodes, livein) {
1081                 be_ilpsched_irn_t    *node;
1082                 ilpsched_node_attr_t *na;
1083                 unsigned             tp_idx, var_idx;
1084                 ir_node              *irn;
1085
1086                 irn  = livein->irn;
1087                 node = get_ilpsched_irn(env, irn);
1088                 na   = get_ilpsched_node_attr(node);
1089
1090                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1091
1092                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1093                 var_idx   = 0;
1094
1095                 /* create variables */
1096                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1097                         unsigned t;
1098
1099                         for (t = 0; t < livein->max_alive_steps; ++t) {
1100                                 /* a_{nt}^k variables */
1101                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1102                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1103                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1104                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1105                                 num_block_var++;
1106                         }
1107                 }
1108         }
1109
1110         ilp_timer_pop();
1111         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1112                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1113 }
1114
1115 /*******************************************************
1116  *                      _             _       _
1117  *                     | |           (_)     | |
1118  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1119  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1120  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1121  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1122  *
1123  *******************************************************/
1124
1125 /**
1126  * Collect all operands and nodes @p irn depends on.
1127  * If there is a Proj within the dependencies, all other Projs of the parent node are added as well.
1128  */
1129 static nodeset *sta_collect_in_deps(ir_node *irn, nodeset *deps) {
1130         int i;
1131
1132         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1133                 ir_node *p = get_irn_in_or_dep(irn, i);
1134
1135                 if (is_Proj(p)) {
1136                         const ir_edge_t *edge;
1137
1138                         p = get_Proj_pred(p);
1139                         foreach_out_edge(p, edge) {
1140                                 ir_node *src = get_edge_src_irn(edge);
1141                                 nodeset_insert(deps, src);
1142                         }
1143                 }
1144                 else {
1145                         nodeset_insert(deps, p);
1146                 }
1147         }
1148
1149         return deps;
1150 }
1151
1152 /**
1153  * Create following ILP constraints:
1154  * - the assignment constraints:
1155  *     assure each node is executed once by exactly one (allowed) execution unit
1156  * - the dead node assignment constraints:
1157  *     assure a node can only die at most once
1158  * - the precedence constraints:
1159  *     assure that no data dependencies are violated
1160  */
1161 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1162         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1163         char                  buf[1024];
1164         ir_node               *irn;
1165         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1166         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1167         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1168         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1169
1170         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1171         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1172                 int                  cst, tp_idx;
1173                 unsigned             cur_var;
1174                 be_ilpsched_irn_t    *node;
1175                 ilpsched_node_attr_t *na;
1176                 ir_node              *pred;
1177                 nodeset              *deps = new_nodeset(16);
1178
1179                 node    = get_ilpsched_irn(env, irn);
1180                 na      = get_ilpsched_node_attr(node);
1181                 cur_var = 0;
1182
1183                 /* the assignment constraint */
1184                 ilp_timer_push(t_cst_assign);
1185                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1186                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1187                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1188                 num_cst_assign++;
1189
1190                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1191                 ilp_timer_pop();
1192
1193                 /* We have separate constraints for Projs and Keeps */
1194                 // ILP becomes infeasible ?!?
1195 //              if (is_Proj(irn) || be_is_Keep(irn))
1196 //                      continue;
1197
1198                 /* the precedence constraints */
1199                 ilp_timer_push(t_cst_prec);
1200                 bs_block_irns = bitset_clear_all(bs_block_irns);
1201
1202                 deps = sta_collect_in_deps(irn, deps);
1203                 foreach_nodeset(deps, pred) {
1204                         unsigned             t_low, t_high, t;
1205                         be_ilpsched_irn_t    *pred_node;
1206                         ilpsched_node_attr_t *pna;
1207                         unsigned             delay;
1208
1209                         pred = skip_normal_Proj(env->arch_env->isa, pred);
1210                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1211                                 continue;
1212
1213                         pred_node = get_ilpsched_irn(env, pred);
1214                         pna       = get_ilpsched_node_attr(pred_node);
1215
1216                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1217
1218                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1219                                 bitset_set(bs_block_irns, pna->block_idx);
1220                         else
1221                                 continue;
1222
1223                         /* irn = n, pred = m */
1224                         delay  = fixed_latency(env->sel, pred, env->block_env);
1225                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1226                         t_high = MIN(na->alap, pna->alap + delay - 1);
1227                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1228                                 unsigned tn, tm;
1229                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1230
1231                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1232                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1233                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1234                                 num_cst_prec++;
1235
1236                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1237                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1238                                         /* node variables have smaller index than pred variables */
1239                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1240                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1241                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1242                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1243                                                 }
1244                                         }
1245
1246                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1247                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1248                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1249                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1250                                                 }
1251                                         }
1252                                 }
1253                                 else {
1254                                         /* pred variables have smaller index than node variables */
1255                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1256                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1257                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1258                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1259                                                 }
1260                                         }
1261
1262                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1263                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1264                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1265                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1266                                                 }
1267                                         }
1268                                 }
1269
1270                                 if (ARR_LEN(tmp_var_idx) > 0)
1271                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1272
1273                                 DEL_ARR_F(tmp_var_idx);
1274                         }
1275                 }
1276                 del_nodeset(deps);
1277                 ilp_timer_pop();
1278         }
1279         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1280                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1281         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1282                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1283 }
1284
1285 /**
1286  * Create ILP resource constraints:
1287  * - assure that for each time step not more instructions are scheduled
1288  *   to the same unit types as units of this type are available
1289  */
1290 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1291         int                   glob_type_idx;
1292         char                  buf[1024];
1293         unsigned              num_cst_resrc = 0;
1294         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1295         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1296
1297         ilp_timer_push(t_cst_rsrc);
1298         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1299                 unsigned                 t;
1300                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1301
1302                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1303
1304                 /* check each time step */
1305                 for (t = 0; t < ba->max_steps; ++t) {
1306                         ir_node *irn;
1307                         int     cst;
1308                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1309
1310                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1311                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1312                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1313                         num_cst_resrc++;
1314
1315                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1316                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1317                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1318                                 int                  tp_idx;
1319
1320                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1321
1322                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1323                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1324                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1325                                 }
1326                         }
1327
1328                         /* set constraints if we have some */
1329                         if (ARR_LEN(tmp_var_idx) > 0)
1330                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1331
1332                         DEL_ARR_F(tmp_var_idx);
1333                 }
1334         }
1335         ilp_timer_pop();
1336         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1337                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1338 }
1339
1340 /**
1341  * Create ILP bundle constraints:
1342  * - assure, at most bundle_size * bundles_per_cycle instructions
1343  *   can be started at a certain point.
1344  */
1345 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1346         char                  buf[1024];
1347         unsigned              t;
1348         unsigned              num_cst_bundle = 0;
1349         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1350         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1351         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1352
1353         ilp_timer_push(t_cst_bundle);
1354         for (t = 0; t < ba->max_steps; ++t) {
1355                 ir_node *irn;
1356                 int     cst;
1357                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1358
1359                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1360                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1361                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1362                 num_cst_bundle++;
1363
1364                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1365                         be_ilpsched_irn_t    *node;
1366                         ilpsched_node_attr_t *na;
1367                         int                  tp_idx;
1368
1369                         /* Projs and Keeps do not contribute to bundle size */
1370                         if (is_Proj(irn) || be_is_Keep(irn))
1371                                 continue;
1372
1373                         node = get_ilpsched_irn(env, irn);
1374                         na   = get_ilpsched_node_attr(node);
1375
1376                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1377                         if (na->is_dummy_node)
1378                                 continue;
1379
1380                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1381                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1382                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1383                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1384                                 }
1385                         }
1386                 }
1387
1388                 if (ARR_LEN(tmp_var_idx) > 0)
1389                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1390
1391                 DEL_ARR_F(tmp_var_idx);
1392         }
1393         ilp_timer_pop();
1394         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1395                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1396 }
1397
1398 /**
1399  * Create ILP alive nodes constraints:
1400  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1401  */
1402 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1403         char                  buf[1024];
1404         ir_node               *irn;
1405         unsigned              num_cst = 0;
1406         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1407         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1408
1409         ilp_timer_push(t_cst);
1410         /* for each node */
1411         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1412                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1413                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1414                 unsigned             t;
1415
1416                 /* we ignore nodes assigned to dummy unit here */
1417                 if (na->is_dummy_node)
1418                         continue;
1419
1420                 /* check check all time steps: asap(n) <= t <= U */
1421                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1422                         int node_tp_idx;
1423
1424                         /* for all unit types available for this node */
1425                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1426                                 unsigned tn, tn_max, idx;
1427                                 int      cst, i;
1428                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1429                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1430
1431                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1432                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1433                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1434                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1435                                 num_cst++;
1436
1437                                 tn_max = MIN(na->alap - 1, t);
1438                                 /* check if the node has been scheduled so far */
1439                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1440                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1441                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1442                                 }
1443
1444                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1445                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1446                                 DEL_ARR_F(tmp_var_idx_n);
1447
1448                                 /* subtract the number of consumer scheduled so far */
1449                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1450                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1451                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1452                                         int                  tp_idx;
1453                                         unsigned             tm, tm_max;
1454
1455                                         tm_max = MIN(ca->alap - 1, t);
1456                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1457                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1458                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1459                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1460                                                 }
1461                                         }
1462                                 }
1463
1464                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1465                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1466                                 DEL_ARR_F(tmp_var_idx_m);
1467
1468                                 /* -c * a_{nt}^k */
1469                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1470                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1471
1472                         }
1473                 }
1474         }
1475         ilp_timer_pop();
1476         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1477                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1478 }
1479
1480 /**
1481  * Create ILP alive nodes constraints for live-in nodes:
1482  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1483  */
1484 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1485         char                  buf[1024];
1486         ilp_livein_node_t     *livein;
1487         unsigned              num_cst = 0;
1488         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1489         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1490
1491         ilp_timer_push(t_cst);
1492         /* for each node */
1493         foreach_pset(ba->livein_nodes, livein) {
1494                 ir_node              *irn  = livein->irn;
1495                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1496                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1497                 unsigned             t;
1498
1499                 /* check check all time steps: 0 <= t < max_alive_steps */
1500                 for (t = 0; t < livein->max_alive_steps; ++t) {
1501                         int node_tp_idx;
1502
1503                         /* for all unit types available for this node */
1504                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1505                                 const ir_edge_t *edge;
1506                                 unsigned idx;
1507                                 int      cst, num_block_user;
1508                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1509
1510                                 /* check the number of consumer scheduled so far */
1511                                 num_block_user = 0;
1512                                 foreach_out_edge(irn, edge) {
1513                                         ir_node              *user = get_edge_src_irn(edge);
1514                                         be_ilpsched_irn_t    *cons;
1515                                         ilpsched_node_attr_t *ca;
1516                                         int                  tp_idx;
1517                                         unsigned             tm, tm_max;
1518
1519                                         /* check only users within current block */
1520                                         if (get_nodes_block(user) != block_node->irn)
1521                                                 continue;
1522
1523                                         num_block_user++;
1524                                         cons = get_ilpsched_irn(env, user);
1525                                         ca   = get_ilpsched_node_attr(cons);
1526
1527                                         tm_max = MIN(ca->alap - 1, t);
1528                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1529                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1530                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1531                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1532                                                 }
1533                                         }
1534                                 }
1535
1536                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1537                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1538                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1539                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1540                                 num_cst++;
1541
1542                                 /* sum(scheduled users) */
1543                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1544                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1545                                 DEL_ARR_F(tmp_var_idx_m);
1546
1547                                 /* + c * a_{nt}^k */
1548                                 idx = node_tp_idx * livein->max_alive_steps + t;
1549                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1550                         }
1551                 }
1552         }
1553         ilp_timer_pop();
1554         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1555                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1556 }
1557
1558 /**
1559  * Create ILP pressure constraints, based on alive nodes:
1560  * - add additional costs to objective function if a node is scheduled
1561  *   on a unit although all units of this type are currently occupied
1562  */
1563 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1564         char                  buf[1024];
1565         ir_node               *cur_irn;
1566         unsigned              num_cst = 0;
1567         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1568         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1569
1570         ilp_timer_push(t_cst);
1571         /* y_{nt}^k is set for each node and timestep and unit type */
1572         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1573                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1574                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1575                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1576                 int                  glob_type_idx;
1577
1578                 /* we ignore nodes assigned to DUMMY unit here */
1579                 if (cur_na->is_dummy_node)
1580                         continue;
1581
1582                 /* for all types */
1583                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1584                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1585                         int                      cur_tp_idx;
1586                         unsigned                 t;
1587
1588                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1589
1590                         /* check if node can be executed on this unit type */
1591                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1592                         if (cur_tp_idx < 0)
1593                                 continue;
1594
1595                         /* check all time_steps at which the current node can be scheduled */
1596                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1597                                 int     cst, y_idx;
1598                                 ir_node *irn;
1599                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1600                                 ilp_livein_node_t *livein;
1601
1602                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1603                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1604                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1605                                 num_cst++;
1606
1607                                 /* - accumulate all nodes alive at point t on unit type k */
1608                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1609                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1610                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1611                                         int                  a_idx, tp_idx;
1612
1613                                         /* check if node can be alive here */
1614                                         if (t < na->asap - 1)
1615                                                 continue;
1616
1617                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1618
1619                                         /* current type is not suitable */
1620                                         if (tp_idx < 0)
1621                                                 continue;
1622
1623                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1624                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1625                                 }
1626                                 /* do the same for livein nodes */
1627                                 foreach_pset(ba->livein_nodes, livein) {
1628                                         ir_node              *irn  = livein->irn;
1629                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1630                                         int                  a_idx, tp_idx;
1631
1632                                         /* check if node can be alive here */
1633                                         if (t >= livein->max_alive_steps)
1634                                                 continue;
1635
1636                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1637
1638                                         /* current type is not suitable */
1639                                         if (tp_idx < 0)
1640                                                 continue;
1641
1642                                         a_idx = tp_idx * livein->max_alive_steps + t;
1643                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1644                                 }
1645
1646                                 if (ARR_LEN(tmp_var_idx) > 0)
1647                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1648                                 DEL_ARR_F(tmp_var_idx);
1649
1650                                 /* - num_nodes * y_{nt}^k */
1651                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1652                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1653                         }
1654                 }
1655         }
1656         ilp_timer_pop();
1657         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1658                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1659 }
1660
1661 /**
1662  * Create ILP branch constraints:
1663  * Assure, alle nodes are scheduled prior to cfg op.
1664  */
1665 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1666         char                  buf[1024];
1667         ir_node               *cur_irn, *cfop;
1668         unsigned              num_cst          = 0;
1669         unsigned              num_non_branches = 0;
1670         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1671         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1672
1673         ilp_timer_push(t_cst);
1674         cfop = NULL;
1675         /* determine number of non-branch nodes and the one and only branch node */
1676         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1677                 switch (get_irn_opcode(cur_irn)) {
1678                         case iro_Phi:
1679                         case iro_Start:
1680                         case iro_End:
1681                         case iro_Proj:
1682                         case iro_Bad:
1683                         case iro_Unknown:
1684                                 num_non_branches++;
1685                                 break;
1686                         default:
1687                                 if (is_cfop(cur_irn)) {
1688                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1689                                         cfop = cur_irn;
1690                                 }
1691                                 else {
1692                                         num_non_branches++;
1693                                 }
1694                                 break;
1695                 }
1696         }
1697
1698         if (cfop) {
1699                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1700                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1701                 unsigned t;
1702
1703                 /* for each time step */
1704                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1705                         int *non_branch_vars, *branch_vars;
1706                         int cst;
1707
1708                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1709                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1710                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1711                         num_cst++;
1712
1713                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1714                         non_branch_vars = NEW_ARR_F(int, 0);
1715                         branch_vars     = NEW_ARR_F(int, 0);
1716                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1717                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1718                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1719                                 int                  tp_idx;
1720
1721                                 if (cur_irn == cfop) {
1722                                         /* for all unit types available for this node */
1723                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1724                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1725                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1726                                         }
1727                                 }
1728                                 else {
1729                                         /* sum up all possible schedule points for this node upto current timestep */
1730                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1731                                                 unsigned tn;
1732                                                 unsigned tmax = MIN(t, na->alap - 1);
1733
1734                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1735                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1736                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1737                                                 }
1738                                         }
1739                                 }
1740
1741                         }
1742
1743                         if (ARR_LEN(non_branch_vars) > 0)
1744                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1745                         if (ARR_LEN(branch_vars) > 0)
1746                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1747
1748                         DEL_ARR_F(branch_vars);
1749                         DEL_ARR_F(non_branch_vars);
1750                 }
1751         }
1752         ilp_timer_pop();
1753         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1754                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1755 }
1756
1757 #if 0
1758 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1759         char                  buf[1024];
1760         ir_node               *irn;
1761         unsigned              num_cst = 0;
1762         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1763         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1764
1765         ilp_timer_push(t_cst);
1766         /* check all nodes */
1767         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1768                 be_ilpsched_irn_t    *node;
1769                 ilpsched_node_attr_t *na;
1770                 unsigned             t;
1771                 ir_node              **pk;
1772
1773                 /* only mode_T nodes can have Projs and Keeps assigned */
1774                 if (get_irn_mode(irn) != mode_T)
1775                         continue;
1776
1777                 node = get_ilpsched_irn(env, irn);
1778                 na   = get_ilpsched_node_attr(node);
1779
1780                 /* check if has some Projs and Keeps assigned */
1781                 if (! na->projkeeps)
1782                         continue;
1783
1784                 /* we can run only once over the queue, so preserve the nodes */
1785                 pk = NEW_ARR_F(ir_node *, 0);
1786                 while (! waitq_empty(na->projkeeps))
1787                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1788                 del_waitq(na->projkeeps);
1789                 na->projkeeps = NULL;
1790
1791                 /* for all time steps at which this node can be scheduled */
1792                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1793                         int cst, tp_idx, i;
1794                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1795
1796                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1797                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1798                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1799                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1800                         num_cst++;
1801
1802                         /* sum up scheduling variables for this time step */
1803                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1804                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1805                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1806                         }
1807
1808                         if (ARR_LEN(tmp_var_idx_n) > 0)
1809                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1810                         DEL_ARR_F(tmp_var_idx_n);
1811
1812                         /* subtract all Proj and Keep variables for this step */
1813                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1814                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1815                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1816                                 int                  pk_tp_idx;
1817
1818                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1819                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1820                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1821                                 }
1822                         }
1823                 }
1824         }
1825         ilp_timer_pop();
1826         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1827                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1828 }
1829 #endif /* if 0 */
1830
1831 /***************************************************
1832  *  _____ _      _____                    _
1833  * |_   _| |    |  __ \                  (_)
1834  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1835  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1836  *  _| |_| |____| |      | | | | | | (_| | | | | |
1837  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1838  *
1839  ***************************************************/
1840
1841 /**
1842  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1843  */
1844 static void create_ilp(ir_node *block, void *walk_env) {
1845         be_ilpsched_env_t     *env           = walk_env;
1846         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1847         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1848         FILE                  *logfile       = NULL;
1849         lpp_t                 *lpp           = NULL;
1850         int                   need_heur      = 0;
1851         struct obstack        var_obst;
1852         char                  name[1024];
1853
1854         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1855         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1856         DBG((env->dbg, 255, "=========================================\n\n"));
1857
1858         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1859                 block, ba->n_interesting_nodes, ba->max_steps));
1860
1861         /* notify backend and get block environment */
1862         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1863
1864         /* if we have less than two interesting nodes, there is no need to create the ILP */
1865         if (ba->n_interesting_nodes > 1) {
1866                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1867                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1868                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1869                 int    estimated_n_var = (int)((double)base_num * fact_var);
1870                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1871
1872                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1873                         estimated_n_var, estimated_n_cst));
1874
1875                 /* set up the LPP object */
1876                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1877
1878                 lpp = new_lpp_userdef(
1879                         (const char *)name,
1880                         lpp_minimize,
1881                         estimated_n_cst,     /* num vars */
1882                         estimated_n_cst + 1, /* num cst */
1883                         1.3);                /* grow factor */
1884                 obstack_init(&var_obst);
1885
1886                 /* create ILP variables */
1887                 create_variables(env, lpp, block_node, &var_obst);
1888
1889                 /* create ILP constraints */
1890                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1891                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1892                 create_ressource_constraints(env, lpp, block_node);
1893                 create_bundle_constraints(env, lpp, block_node);
1894                 create_branch_constraint(env, lpp, block_node);
1895                 //create_proj_keep_constraints(env, lpp, block_node);
1896
1897                 if (env->opts->regpress) {
1898                         create_alive_nodes_constraint(env, lpp, block_node);
1899                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1900                         create_pressure_alive_constraint(env, lpp, block_node);
1901                 }
1902
1903                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1904
1905                 /* debug stuff, dump lpp when debugging is on  */
1906                 DEBUG_ONLY({
1907                         if (firm_dbg_get_mask(env->dbg) > 1) {
1908                                 char buf[1024];
1909                                 FILE *f;
1910
1911                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1912                                 f = fopen(buf, "w");
1913                                 lpp_dump_plain(lpp, f);
1914                                 fclose(f);
1915                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1916                                 lpp_dump(lpp, buf);
1917                         }
1918                 })
1919
1920                 /* set solve time limit */
1921                 lpp_set_time_limit(lpp, env->opts->time_limit);
1922
1923                 /* set logfile if requested */
1924                 if (strlen(env->opts->log_file) > 0) {
1925                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1926                                 lpp_set_log(lpp, stdout);
1927                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1928                                 lpp_set_log(lpp, stderr);
1929                         else {
1930                                 logfile = fopen(env->opts->log_file, "w");
1931                                 if (! logfile)
1932                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1933                                 else
1934                                         lpp_set_log(lpp, logfile);
1935                         }
1936                 }
1937
1938                 /* solve the ILP */
1939                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1940
1941                 if (logfile)
1942                         fclose(logfile);
1943
1944                 /* check for valid solution */
1945                 if (! lpp_is_sol_valid(lpp)) {
1946                         char buf[1024];
1947                         FILE *f;
1948
1949                         DEBUG_ONLY({
1950                                 if (firm_dbg_get_mask(env->dbg) >= 2) {
1951                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1952                                         f = fopen(buf, "w");
1953                                         lpp_dump_plain(lpp, f);
1954                                         fclose(f);
1955                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1956                                         lpp_dump(lpp, buf);
1957                                         dump_ir_block_graph(env->irg, "-infeasible");
1958                                 }
1959                         })
1960
1961                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1962                         need_heur = 1;
1963                 }
1964
1965                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1966                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1967                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1968                 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));
1969                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1970                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1971                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1972                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1973
1974                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1975         }
1976
1977         /* apply solution */
1978         be_stat_ev("nodes", ba->block_last_idx);
1979         be_stat_ev("vars", lpp ? lpp->var_next : 0);
1980         be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1981         if (need_heur) {
1982                         be_stat_ev("time", -1);
1983                         be_stat_ev_dbl("opt", 0.0);
1984                 list_sched_single_block(env->birg, block, env->be_opts);
1985         }
1986         else {
1987                 if (lpp) {
1988                         double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1989                         be_stat_ev_dbl("time", lpp->sol_time);
1990                         be_stat_ev_dbl("opt", opt);
1991                 }
1992                 else {
1993                         be_stat_ev_dbl("time", 0.0);
1994                         be_stat_ev_dbl("opt", 100.0);
1995                 }
1996                 apply_solution(env, lpp, block);
1997         }
1998
1999         if (lpp)
2000                 free_lpp(lpp);
2001
2002         /* notify backend */
2003         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
2004 }
2005
2006 /**
2007  * Perform ILP scheduling on the given irg.
2008  */
2009 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
2010         be_ilpsched_env_t          env;
2011         const char                 *name     = "be ilp scheduling";
2012         ir_graph                   *irg      = be_get_birg_irg(birg);
2013         const arch_env_t           *arch_env = be_get_birg_arch_env(birg);
2014         const arch_isa_t           *isa      = arch_env->isa;
2015         const ilp_sched_selector_t *sel      = isa->impl->get_ilp_sched_selector(isa);
2016
2017         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
2018
2019         stat_ev_ctx_push("phase", "ilpsched");
2020
2021 //      firm_dbg_set_mask(env.dbg, 1);
2022
2023         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, irg);
2024         env.sel        = sel;
2025         env.irg        = irg;
2026         env.height     = heights_new(irg);
2027         env.main_env   = birg->main_env;
2028         env.arch_env   = arch_env;
2029         env.cpu        = arch_isa_get_machine(arch_env->isa);
2030         env.opts       = &ilp_opts;
2031         env.birg       = birg;
2032         env.be_opts    = be_opts;
2033         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn, NULL);
2034
2035         /* assign a unique per block number to all interesting nodes */
2036         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
2037
2038         /*
2039                 The block indices are completely build after the walk,
2040                 now we can allocate the bitsets (size depends on block indices)
2041                 for all nodes.
2042         */
2043         phase_reinit_irn_data(&env.ph);
2044
2045         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2046         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2047
2048         /* Calculate ALAP of all irns */
2049         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2050
2051         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2052         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2053
2054         /* perform ILP scheduling */
2055         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2056
2057         DEBUG_ONLY(
2058                 if (firm_dbg_get_mask(env.dbg)) {
2059                         phase_stat_t stat;
2060                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2061
2062                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2063                 }
2064         );
2065
2066         /* free data allocated dynamically */
2067         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2068
2069         /* free all allocated object */
2070         phase_free(&env.ph);
2071         heights_free(env.height);
2072
2073         /* notify backend */
2074         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2075
2076         stat_ev_ctx_pop();
2077 }
2078
2079 /**
2080  * Register ILP scheduler options.
2081  */
2082 void be_init_ilpsched(void)
2083 {
2084         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2085         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2086
2087         lc_opt_add_table(sched_grp, ilpsched_option_table);
2088 }
2089
2090 #else /* WITH_ILP */
2091
2092 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2093 {}
2094
2095 #endif /* WITH_ILP */