31c52422f0fb4fad3dc1b06e4a1a77efb764e106
[libfirm] / ir / be / beilpsched.c
1 /**
2  * Scheduling algorithms.
3  * An ILP scheduler based on
4  * "ILP-based Instruction Scheduling for IA-64"
5  * by Daniel Kaestner and Sebastian Winkel
6  * extended with register pressure constraints by Christian Wuerdig
7  *
8  * @date   22.10.2005
9  * @author Christian Wuerdig
10  * @cvs-id $Id$
11  */
12 #ifdef HAVE_CONFIG_H
13 #include "config.h"
14 #endif
15
16 #ifdef WITH_ILP
17
18 #include <math.h>
19
20 #ifndef _WIN32
21 #include <strings.h>
22 #endif /* _WIN32 */
23
24 #include "irnode_t.h"
25 #include "irgwalk.h"
26 #include "irbitset.h"
27 #include "irphase_t.h"
28 #include "height.h"
29 #include "iredges.h"
30 #include "pdeq.h"
31 #include "debug.h"
32 #include "irtools.h"
33 #include "irdump.h"
34 #include "plist.h"
35 #include "irprintf.h"
36
37 #include <lpp/lpp.h>
38 #include <lpp/lpp_net.h>
39
40 #include <libcore/lc_opts.h>
41 #include <libcore/lc_opts_enum.h>
42 #include <libcore/lc_timing.h>
43
44 #include "be.h"
45 #include "benode_t.h"
46 #include "besched_t.h"
47 #include "beilpsched.h"
48 #include "beutil.h"
49 #include "bestat.h"
50
51 typedef struct _ilpsched_options_t {
52         unsigned regpress;
53         unsigned time_limit;
54         char     log_file[1024];
55 } ilpsched_options_t;
56
57 typedef struct _unit_type_info_t {
58         int                            n_units;
59         const be_execution_unit_type_t *tp;
60 } unit_type_info_t;
61
62 /**
63  * holding the ILP variables of the different types
64  */
65 typedef struct _ilp_var_types_t {
66         int *x;   /* x_{nt}^k variables */
67         int *a;   /* a_{nt}^k variables */
68         int *y;   /* y_{nt}^k variables */
69 } ilp_var_types_t;
70
71 /**
72  * Holds alive variables for a node live-in to a block.
73  */
74 typedef struct _ilp_livein_node_t {
75         ir_node  *irn;
76         unsigned max_alive_steps;
77         int      *a;
78 } ilp_livein_node_t;
79
80 /* attributes for a node */
81 typedef struct _ilpsched_node_attr_t {
82         unsigned asap;                     /**< The ASAP scheduling control step */
83         unsigned alap;                     /**< The ALAP scheduling control step */
84         unsigned latency;                  /**< Latency of this node (needed for sorting) */
85         unsigned sched_point;              /**< the step in which the node is finally scheduled */
86         unsigned visit_idx;                /**< Index of the node having visited this node last */
87         unsigned consumer_idx;             /**< Index of the node having counted this node as consumer last */
88         unsigned n_consumer;               /**< Number of consumers */
89         ir_node  **block_consumer;         /**< List of consumer being in the same block */
90         waitq    *projkeeps;               /**< A List of Projs and Keeps belonging to this node */
91         unsigned block_idx     : 30;       /**< A unique per block index */
92         unsigned alap_changed  : 1;        /**< the current ALAP has changed, revisit preds */
93         unsigned is_dummy_node : 1;        /**< this node is assigned to DUMMY unit */
94         bitset_t *transitive_block_nodes;  /**< Set of transitive block nodes (predecessors
95                                                                                         for ASAP, successors for ALAP */
96         unsigned         n_unit_types;     /**< number of allowed execution unit types */
97         unit_type_info_t *type_info;       /**< list of allowed execution unit types */
98         ilp_var_types_t  ilp_vars;         /**< the different ILP variables */
99 } ilpsched_node_attr_t;
100
101 /* attributes for a block */
102 typedef struct _ilpsched_block_attr_t {
103         unsigned block_last_idx;        /**< The highest node index in block so far */
104         unsigned n_interesting_nodes;   /**< The number of nodes interesting for scheduling */
105         unsigned max_steps;             /**< Upper bound for block execution */
106         plist_t  *root_nodes;           /**< A list of nodes having no user in current block */
107         ir_node  *head_ilp_nodes;       /**< A linked list of nodes which will contribute to ILP */
108         pset     *livein_nodes;         /**< A set of nodes which are live-in to this block */
109 } ilpsched_block_attr_t;
110
111 typedef union _ilpsched_attr_ {
112         ilpsched_node_attr_t  node_attr;
113         ilpsched_block_attr_t block_attr;
114 } ilpsched_attr_t;
115
116 /* A irn for the phase and it's attributes (either node or block) */
117 typedef struct {
118         ir_node         *irn;
119         ilpsched_attr_t attr;
120 } be_ilpsched_irn_t;
121
122 /* The ILP scheduling environment */
123 typedef struct {
124         phase_t              ph;            /**< The phase */
125         ir_graph             *irg;          /**< The current irg */
126         heights_t            *height;       /**< The heights object of the irg */
127         void                 *irg_env;      /**< An environment for the irg scheduling, provided by the backend */
128         void                 *block_env;    /**< An environment for scheduling a block, provided by the backend */
129         const arch_env_t     *arch_env;
130         const arch_isa_t     *isa;          /**< The ISA */
131         const be_main_env_t  *main_env;
132         const be_machine_t   *cpu;          /**< the current abstract machine */
133         ilpsched_options_t   *opts;         /**< the ilp options for current irg */
134         const be_irg_t       *birg;         /**< The birg object */
135         be_options_t         *be_opts;      /**< backend options */
136         const ilp_sched_selector_t *sel;    /**< The ILP sched selector provided by the backend */
137         DEBUG_ONLY(firm_dbg_module_t *dbg);
138 } be_ilpsched_env_t;
139
140 /* convenience macros to handle phase irn data */
141 #define get_ilpsched_irn(ilpsched_env, irn) (phase_get_or_set_irn_data(&(ilpsched_env)->ph, (irn)))
142 #define is_ilpsched_block(node)             (is_Block((node)->irn))
143 #define get_ilpsched_block_attr(block)      (&(block)->attr.block_attr)
144 #define get_ilpsched_node_attr(node)        (&(node)->attr.node_attr)
145
146 /* check if node is considered for ILP scheduling */
147 #define consider_for_sched(isa, irn) \
148         (! (is_Block(irn)            ||  \
149                 is_normal_Proj(isa, irn) ||  \
150                 is_Phi(irn)              ||  \
151                 is_NoMem(irn)            ||  \
152                 is_Unknown(irn)          ||  \
153                 is_End(irn)                  \
154                 ))
155
156 /* gives the valid scheduling time step interval for a node */
157 #define VALID_SCHED_INTERVAL(na) ((na)->alap - (na)->asap + 1)
158
159 /* gives the valid interval where a node can die */
160 #define VALID_KILL_INTERVAL(ba, na) ((ba)->max_steps - (na)->asap + 1)
161
162 /* gives the corresponding ILP variable for given node, unit and time step */
163 #define ILPVAR_IDX(na, unit, control_step) \
164         ((unit) * VALID_SCHED_INTERVAL((na)) + (control_step) - (na)->asap + 1)
165
166 /* gives the corresponding dead nodes ILP variable for given node, unit and time step */
167 #define ILPVAR_IDX_DEAD(ba, na, unit, control_step) \
168         ((unit) * VALID_KILL_INTERVAL((ba), (na)) + (control_step) - (na)->asap + 1)
169
170 /* check if a double value is within an epsilon environment of 0 */
171 #define LPP_VALUE_IS_0(dbl) (fabs((dbl)) <= 1e-10)
172
173 #define ilp_timer_push(t)         lc_timer_push((t))
174 #define ilp_timer_pop()           lc_timer_pop()
175 #define ilp_timer_elapsed_usec(t) lc_timer_elapsed_usec((t))
176
177 /* option variable */
178 static ilpsched_options_t ilp_opts = {
179         1,     /* default is with register pressure constraints */
180         300,   /* 300 sec per block time limit */
181         ""     /* no log file */
182 };
183
184 /* ILP options */
185 static const lc_opt_table_entry_t ilpsched_option_table[] = {
186         LC_OPT_ENT_BOOL("regpress",  "Use register pressure constraints", &ilp_opts.regpress),
187         LC_OPT_ENT_INT("time_limit", "ILP time limit per block", &ilp_opts.time_limit),
188         LC_OPT_ENT_STR("lpp_log",    "LPP logfile (stderr and stdout are supported)", ilp_opts.log_file, sizeof(ilp_opts.log_file)),
189         { NULL }
190 };
191
192 /*
193         We need this global variable as we compare nodes dependent on heights,
194         but we cannot pass any information to the qsort compare function.
195 */
196 static heights_t *glob_heights;
197
198 /**
199  * Check if irn is a Proj, which has no execution units assigned.
200  * @return 1 if irn is a Proj having no execution units assigned, 0 otherwise
201  */
202 static INLINE int is_normal_Proj(const arch_isa_t *isa, const ir_node *irn) {
203         return is_Proj(irn) && (arch_isa_get_allowed_execution_units(isa, irn) == NULL);
204 }
205
206 /**
207  * Skips normal Projs.
208  * @return predecessor if irn is a normal Proj, otherwise irn.
209  */
210 static INLINE ir_node *skip_normal_Proj(const arch_isa_t *isa, ir_node *irn) {
211         if (is_normal_Proj(isa, irn))
212                 return get_Proj_pred(irn);
213         return irn;
214 }
215
216 static INLINE int fixed_latency(const ilp_sched_selector_t *sel, ir_node *irn, void *env) {
217         unsigned lat = be_ilp_sched_latency(sel, irn, env);
218         if (lat == 0 && ! is_Proj(irn) && ! be_is_Keep(irn))
219                 lat = 1;
220         return lat;
221 }
222
223 static int cmp_live_in_nodes(const void *a, const void *b) {
224         const ilp_livein_node_t *n1 = a;
225         const ilp_livein_node_t *n2 = b;
226
227         return n1->irn != n2->irn;
228 }
229
230 /**
231  * Compare scheduling time steps of two be_ilpsched_irn's.
232  */
233 static int cmp_ilpsched_irn(const void *a, const void *b) {
234         be_ilpsched_irn_t    *n1   = *(be_ilpsched_irn_t **)a;
235         be_ilpsched_irn_t    *n2   = *(be_ilpsched_irn_t **)b;
236         ilpsched_node_attr_t *n1_a = get_ilpsched_node_attr(n1);
237         ilpsched_node_attr_t *n2_a = get_ilpsched_node_attr(n2);
238
239         if (n1_a->sched_point == n2_a->sched_point) {
240                 ir_node *irn_a = n1->irn;
241                 ir_node *irn_b = n2->irn;
242
243                 if (heights_reachable_in_block(glob_heights, irn_a, irn_b))
244                         return 1;
245                 if (heights_reachable_in_block(glob_heights, irn_b, irn_a))
246                         return -1;
247
248                 /*
249                         Ok, timestep is equal and the nodes are parallel,
250                         so check latency and schedule high latency first.
251                 */
252                 return QSORT_CMP(n2_a->latency, n1_a->latency);
253         }
254         else
255                 return QSORT_CMP(n1_a->sched_point, n2_a->sched_point);
256 }
257
258 /**
259  * In case there is no phase information for irn, initialize it.
260  */
261 static void *init_ilpsched_irn(phase_t *ph, ir_node *irn, void *old) {
262         be_ilpsched_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
263
264         if (res == old) {
265                 /* if we have already some data: check for reinitialization */
266
267                 if (! is_Block(irn)) {
268                         ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
269
270                         if (! na->transitive_block_nodes) {
271                                 ir_node               *block      = get_nodes_block(irn);
272                                 be_ilpsched_irn_t     *block_node = phase_get_or_set_irn_data(ph, block);
273                                 ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
274
275                                 /* we are called after the block indices have been build: create bitset */
276                                 na->transitive_block_nodes = bitset_obstack_alloc(phase_obst(ph), ba->block_last_idx);
277                         }
278                         else {
279                                 /* we are called from reinit block data: clear the bitset */
280                                 bitset_clear_all(na->transitive_block_nodes);
281                                 na->visit_idx    = 0;
282                                 na->alap_changed = 1;
283                         }
284                 }
285                 return old;
286         }
287
288         res->irn = irn;
289
290         /* set ilpsched irn attributes (either block or irn) */
291         if (is_Block(irn)) {
292                 ilpsched_block_attr_t *ba = get_ilpsched_block_attr(res);
293
294                 ba->n_interesting_nodes = 0;
295                 ba->block_last_idx      = 0;
296                 ba->root_nodes          = plist_new();
297                 ba->head_ilp_nodes      = NULL;
298                 ba->livein_nodes        = new_pset(cmp_live_in_nodes, 16);
299                 ba->max_steps           = 0;
300         }
301         else {
302                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(res);
303                 memset(na, 0, sizeof(*na));
304         }
305
306         return res;
307 }
308
309 /**
310  * Assign a per block unique number to each node.
311  */
312 static void build_block_idx(ir_node *irn, void *walk_env) {
313         be_ilpsched_env_t     *env = walk_env;
314         be_ilpsched_irn_t     *node, *block_node;
315         ilpsched_node_attr_t  *na;
316         ilpsched_block_attr_t *ba;
317
318         set_irn_link(irn, NULL);
319         if (! consider_for_sched(env->arch_env->isa, irn))
320                 return;
321
322         node       = get_ilpsched_irn(env, irn);
323         na         = get_ilpsched_node_attr(node);
324         block_node = get_ilpsched_irn(env, get_nodes_block(irn));
325         ba         = get_ilpsched_block_attr(block_node);
326
327         na->block_idx = ba->block_last_idx++;
328 }
329
330 /********************************************************
331  *                              __        _
332  *                             / /       | |
333  *   __ _ ___  __ _ _ __      / /    __ _| | __ _ _ __
334  *  / _` / __|/ _` | '_ \    / /    / _` | |/ _` | '_ \
335  * | (_| \__ \ (_| | |_) |  / /    | (_| | | (_| | |_) |
336  *  \__,_|___/\__,_| .__/  /_/      \__,_|_|\__,_| .__/
337  *                 | |                           | |
338  *                 |_|                           |_|
339  ********************************************************/
340
341 /**
342  * Add all nodes having no user in current block to last_nodes list.
343  */
344 static void collect_alap_root_nodes(ir_node *irn, void *walk_env) {
345         ir_node               *block;
346         const ir_edge_t       *edge;
347         be_ilpsched_irn_t     *block_node, *node;
348         ilpsched_block_attr_t *ba;
349         ilpsched_node_attr_t  *na;
350         int                   i, j;
351         be_ilpsched_env_t     *env           = walk_env;
352         int                   has_block_user = 0;
353         unsigned              n_consumer     = 0;
354         ir_edge_kind_t        ekind[2]       = { EDGE_KIND_NORMAL, EDGE_KIND_DEP };
355         ir_node               **consumer;
356         int                   idx;
357
358         if (! consider_for_sched(env->arch_env->isa, irn))
359                 return;
360
361         block    = get_nodes_block(irn);
362     idx      = get_irn_idx(irn);
363         consumer = NEW_ARR_F(ir_node *, 0);
364
365         DBG((env->dbg, LEVEL_3, "%+F (%+F) is interesting, examining ... ", irn, block));
366
367         /* check data and dependency out edges */
368         for (i = 0; i < 2 && ! has_block_user; ++i) {
369                 foreach_out_edge_kind(irn, edge, ekind[i]) {
370                         ir_node *user = get_edge_src_irn(edge);
371
372                         if (is_normal_Proj(env->arch_env->isa, user)) {
373                                 const ir_edge_t *user_edge;
374
375                                 if (get_irn_mode(user) == mode_X)
376                                         continue;
377
378                                 /* The ABI ensures, that there will be no ProjT nodes in the graph. */
379                                 for (j = 0; j < 2; ++j) {
380                                         foreach_out_edge_kind(user, user_edge, ekind[j]) {
381                                                 ir_node *real_user = get_edge_src_irn(user_edge);
382
383                                                 if (! is_Phi(real_user) && ! is_Block(real_user)) {
384                                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, real_user);
385                                                         ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
386
387                                                         /* skip already visited nodes */
388                                                         if (ua->consumer_idx == idx)
389                                                                 continue;
390
391                                                         /* check if node has user in this block and collect the user if it's a data user */
392                                                         if (get_nodes_block(real_user) == block) {
393                                                                 if (i == 0 && j == 0)
394                                                                         ARR_APP1(ir_node *, consumer, real_user);
395                                                                 has_block_user = 1;
396                                                         }
397
398                                                         /* only count data consumer */
399                                                         if (i == 0)
400                                                                 n_consumer++;
401
402                                                         /* mark user as visited by this node */
403                                                         ua->consumer_idx = idx;
404                                                 }
405                                         }
406                                 }
407                         }
408                         else if (is_Block(user)) {
409                                 continue;
410                         }
411                         else if (! is_Phi(user)) {
412                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
413                                 ilpsched_node_attr_t *ua   = get_ilpsched_node_attr(node);
414
415                                 /* skip already visited nodes */
416                                 if (ua->consumer_idx == idx)
417                                         continue;
418
419                                 /* check if node has user in this block and collect the user if it's a data user */
420                                 if (get_nodes_block(user) == block) {
421                                         if (i == 0)
422                                                 ARR_APP1(ir_node *, consumer, user);
423                                         has_block_user = 1;
424                                 }
425
426                                 /* only count data consumer */
427                                 if (i == 0)
428                                         n_consumer++;
429
430                                 /* mark user visited by this node */
431                                 ua->consumer_idx = idx;
432                         }
433                         else if (get_nodes_block(user) != block) {
434                                 n_consumer++;
435                         }
436                 }
437         }
438
439         block_node = get_ilpsched_irn(env, block);
440         ba         = get_ilpsched_block_attr(block_node);
441
442         ba->n_interesting_nodes++;
443
444         /* current irn has no user inside this block, add to queue */
445         if (! has_block_user) {
446                 DB((env->dbg, LEVEL_3, "root node\n"));
447                 plist_insert_back(ba->root_nodes, irn);
448         }
449         else {
450                 DB((env->dbg, LEVEL_3, "normal node\n"));
451         }
452
453         /* record number of all consumer and the consumer within the same block */
454         node = get_ilpsched_irn(env, irn);
455         na   = get_ilpsched_node_attr(node);
456         na->n_consumer     = n_consumer;
457         na->block_consumer = NEW_ARR_D(ir_node *, phase_obst(&env->ph), ARR_LEN(consumer));
458         memcpy(na->block_consumer, consumer, ARR_LEN(consumer) * sizeof(na->block_consumer[0]));
459         DEL_ARR_F(consumer);
460 }
461
462 /**
463  * Calculate the ASAP scheduling step for current irn.
464  */
465 static void calculate_irn_asap(ir_node *irn, void *walk_env) {
466         be_ilpsched_env_t     *env = walk_env;
467         int                   i;
468         ir_node               *block;
469         be_ilpsched_irn_t     *node, *block_node;
470         ilpsched_node_attr_t  *na;
471         ilpsched_block_attr_t *ba;
472
473         /* These nodes are handled separate */
474         if (! consider_for_sched(env->arch_env->isa, irn))
475                 return;
476
477         DBG((env->dbg, LEVEL_2, "Calculating ASAP of node %+F ... ", irn));
478
479         block    = get_nodes_block(irn);
480         node     = get_ilpsched_irn(env, irn);
481         na       = get_ilpsched_node_attr(node);
482         na->asap = 1;
483
484         for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
485                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
486
487                 /* check for greatest distance to top */
488                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
489                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
490                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
491                         unsigned             lat;
492
493                         lat         = fixed_latency(env->sel, pred, env->block_env);
494                         na->latency = lat;
495                         na->asap    = MAX(na->asap, pna->asap + lat);
496                 }
497         }
498
499         /* add node to ILP node list and update max_steps */
500         block_node = get_ilpsched_irn(env, block);
501         ba         = get_ilpsched_block_attr(block_node);
502
503         set_irn_link(irn, ba->head_ilp_nodes);
504         ba->head_ilp_nodes = irn;
505         ba->max_steps     += fixed_latency(env->sel, irn, env->block_env);
506
507         DB((env->dbg, LEVEL_2, "%u\n", na->asap));
508 }
509
510 /**
511  * Calculate the ALAP scheduling step of all irns in current block.
512  * Depends on max_steps being calculated.
513  */
514 static void calculate_block_alap(ir_node *block, void *walk_env) {
515         be_ilpsched_env_t     *env        = walk_env;
516         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
517         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
518         waitq                 *cur_queue  = new_waitq();
519         plist_element_t       *el;
520
521         assert(is_Block(block));
522
523         DBG((env->dbg, LEVEL_2, "Calculating ALAP for nodes in %+F (%u nodes, %u max steps)\n",
524                 block, ba->n_interesting_nodes, ba->max_steps));
525
526         /* TODO: Might be faster to use out edges and call phase_reinit_single_irn_data */
527         //phase_reinit_block_irn_data(&env->ph, block);
528
529         /* init start queue */
530         foreach_plist(ba->root_nodes, el) {
531                 waitq_put(cur_queue, plist_element_get_value(el));
532         }
533
534         /* repeat until all nodes are processed */
535         while (! waitq_empty(cur_queue)) {
536                 waitq *next_queue = new_waitq();
537
538                 /* process all nodes in current step */
539                 while (! waitq_empty(cur_queue)) {
540                         ir_node              *cur_irn = waitq_get(cur_queue);
541                         be_ilpsched_irn_t    *node    = get_ilpsched_irn(env, cur_irn);
542                         ilpsched_node_attr_t *na      = get_ilpsched_node_attr(node);
543                         int                  i;
544
545                         /* cur_node has no alap set -> it's a root node, set to max alap */
546                         if (na->alap == 0) {
547                                 na->alap = ba->max_steps;
548                                 DBG((env->dbg, LEVEL_2, "setting ALAP of node %+F to %u, handling preds:\n",
549                                         cur_irn, na->alap));
550                         }
551                         else {
552                                 DBG((env->dbg, LEVEL_2, "ALAP of node %+F is %u, handling preds:\n",
553                                         cur_irn, na->alap));
554                         }
555
556                         /* set the alap's of all predecessors */
557                         for (i = get_irn_ins_or_deps(cur_irn) - 1; i >= 0; --i) {
558                                 ir_node *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(cur_irn, i));
559
560                                 /* check for greatest distance to bottom */
561                                 if (! is_Phi(pred) && ! is_NoMem(pred) && get_nodes_block(pred) == block) {
562                                         be_ilpsched_irn_t    *pred_node = get_ilpsched_irn(env, pred);
563                                         ilpsched_node_attr_t *pna       = get_ilpsched_node_attr(pred_node);
564                                         unsigned             lat;
565
566                                         /* mark the predecessor as visited by current irn */
567                                         if (pna->visit_idx == get_irn_idx(cur_irn) && ! na->alap_changed)
568                                                 continue;
569                                         pna->visit_idx = get_irn_idx(cur_irn);
570
571                                         lat = fixed_latency(env->sel, pred, env->block_env);
572
573                                         /* set ALAP of current pred */
574                                         if (pna->alap == 0) {
575                                                 /* current ALAP is 0: node has not yet been visited */
576                                                 pna->alap_changed = 1;
577                                                 pna->alap         = na->alap - lat;
578                                         }
579                                         else if (pna->alap > na->alap - lat) {
580                                                 /* we found a longer path to root node: change ALAP */
581                                                 pna->alap         = na->alap - lat;
582                                                 pna->alap_changed = 1;
583                                         }
584                                         else {
585                                                 /* current ALAP is best found so far: keep it */
586                                                 pna->alap_changed = 0;
587                                         }
588
589                                         DBG((env->dbg, LEVEL_2, "\tsetting ALAP of node %+F to %u\n", pred, pna->alap));
590
591                                         /* enqueue node for next iteration */
592                                         if (get_irn_ins_or_deps(pred) > 0)
593                                                 waitq_put(next_queue, pred);
594                                 }
595                         }
596                 }
597
598                 /* prepare for next iteration */
599                 del_waitq(cur_queue);
600                 cur_queue = next_queue;
601         }
602 }
603
604 /**
605  * Free list of root nodes and the set of live-in nodes.
606  */
607 static void clear_unwanted_data(ir_node *block, void *walk_env) {
608         be_ilpsched_env_t     *env        = walk_env;
609         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
610         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
611
612         plist_free(ba->root_nodes);
613         ba->root_nodes = NULL;
614         del_pset(ba->livein_nodes);
615         ba->livein_nodes = NULL;
616 }
617
618 /**
619  * Refine the {ASAP(n), ALAP(n)} interval for the nodes.
620  * Set the ASAP/ALAP times of Projs and Keeps to their ancestor ones.
621  */
622 static void refine_asap_alap_times(ir_node *irn, void *walk_env) {
623         be_ilpsched_env_t    *env  = walk_env;
624         ir_node              *pred = irn;
625         be_ilpsched_irn_t    *node, *pred_node;
626         ilpsched_node_attr_t *na, *pna;
627
628         if (! consider_for_sched(env->arch_env->isa, irn))
629                 return;
630
631         if (! is_Proj(irn) && ! be_is_Keep(irn))
632                 return;
633
634         /* go to the ancestor */
635         if (be_is_Keep(irn))
636                 pred = get_irn_n(irn, 0);
637         pred = skip_Proj(pred);
638
639         node      = get_ilpsched_irn(env, irn);
640         pred_node = get_ilpsched_irn(env, pred);
641         na        = get_ilpsched_node_attr(node);
642         pna       = get_ilpsched_node_attr(pred_node);
643
644         na->asap = pna->asap;
645         na->alap = pna->alap;
646
647         /* record all Projs and Keeps for this node */
648         if (! pna->projkeeps)
649                 pna->projkeeps = new_waitq();
650         waitq_put(pna->projkeeps, irn);
651
652         DBG((env->dbg, LEVEL_2, "fixing ASAP/ALAP of %+F to %u/%u\n", irn, na->asap, na->alap));
653 }
654
655 /*******************************************
656  *           _              _       _
657  *          | |            | |     | |
658  *  ___  ___| |__   ___  __| |_   _| | ___
659  * / __|/ __| '_ \ / _ \/ _` | | | | |/ _ \
660  * \__ \ (__| | | |  __/ (_| | |_| | |  __/
661  * |___/\___|_| |_|\___|\__,_|\__,_|_|\___|
662  *
663  *******************************************/
664
665 static INLINE void check_for_keeps(waitq *keeps, ir_node *block, ir_node *irn) {
666         const ir_edge_t *edge;
667
668         foreach_out_edge(irn, edge) {
669                 ir_node *user = get_edge_src_irn(edge);
670
671                 if (be_is_Keep(user)) {
672                         assert(get_nodes_block(user) == block && "Keep must not be in different block.");
673                         waitq_put(keeps, user);
674                 }
675         }
676 }
677
678 /**
679  * Inserts @p irn before @p before into schedule and notifies backend.
680  */
681 static INLINE void notified_sched_add_before(be_ilpsched_env_t *env,
682         ir_node *before, ir_node *irn, unsigned cycle)
683 {
684         be_ilp_sched_node_scheduled(env->sel, irn, cycle, env->block_env);
685         sched_add_before(before, irn);
686 }
687
688 /**
689  * Adds a node, it's Projs (in case of mode_T nodes) and
690  * it's Keeps to schedule.
691  */
692 static void add_to_sched(be_ilpsched_env_t *env, ir_node *block, ir_node *irn, unsigned cycle) {
693         const ir_edge_t *edge;
694         waitq           *keeps = new_waitq();
695
696         /* mode_M nodes are not scheduled */
697         if (get_irn_mode(irn) == mode_M)
698                 return;
699
700         if (! sched_is_scheduled(irn))
701                 notified_sched_add_before(env, block, irn, cycle);
702
703         /* add Projs */
704         if (get_irn_mode(irn) == mode_T) {
705                 foreach_out_edge(irn, edge) {
706                         ir_node *user = get_edge_src_irn(edge);
707
708                         if ((to_appear_in_schedule(user) || get_irn_mode(user) == mode_b) &&
709                                 get_irn_n_edges(user) > 0)
710                         {
711                                 notified_sched_add_before(env, block, user, cycle);
712                         }
713
714                         check_for_keeps(keeps, block, user);
715                 }
716         }
717         else {
718                 check_for_keeps(keeps, block, irn);
719         }
720
721         /* add Keeps */
722         while (! waitq_empty(keeps)) {
723                 ir_node *keep = waitq_get(keeps);
724                 if (! sched_is_scheduled(keep))
725                         notified_sched_add_before(env, block, keep, cycle);
726         }
727
728         del_waitq(keeps);
729 }
730
731 /**
732  * Schedule all nodes in the given block, according to the ILP solution.
733  */
734 static void apply_solution(be_ilpsched_env_t *env, lpp_t *lpp, ir_node *block) {
735         be_ilpsched_irn_t     *block_node = get_ilpsched_irn(env, block);
736         ilpsched_block_attr_t *ba         = get_ilpsched_block_attr(block_node);
737         sched_info_t          *info       = get_irn_sched_info(block);
738         be_ilpsched_irn_t     **sched_nodes;
739         unsigned              i, l;
740         ir_node               *cfop, *irn;
741         const ir_edge_t       *edge;
742
743         /* init block schedule list */
744         INIT_LIST_HEAD(&info->list);
745         info->scheduled = 1;
746
747         /* collect nodes and their scheduling time step */
748         sched_nodes = NEW_ARR_F(be_ilpsched_irn_t *, 0);
749         if (ba->n_interesting_nodes == 0) {
750                 /* ignore */
751         }
752         else if (ba->n_interesting_nodes == 1) {
753                 be_ilpsched_irn_t *node = get_ilpsched_irn(env, ba->head_ilp_nodes);
754
755                 /* add the single node */
756                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
757         }
758         else {
759                 /* check all nodes for their positive solution */
760                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
761                         be_ilpsched_irn_t    *node;
762                         ilpsched_node_attr_t *na;
763                         int                  tp_idx, found;
764                         unsigned             cur_var, t;
765
766                         node    = get_ilpsched_irn(env, irn);
767                         na      = get_ilpsched_node_attr(node);
768                         cur_var = 0;
769                         found   = 0;
770
771                         if (! na->is_dummy_node) {
772                                 for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
773                                         for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
774                                                 double cost = lpp_get_var_sol(lpp, na->ilp_vars.y[cur_var]);
775
776                                                 if (! LPP_VALUE_IS_0(cost)) {
777                                                         DBG((env->dbg, LEVEL_3, "%+F has additional regpressure costs of %f\n", irn, cost));
778                                                         found = 1;
779                                                 }
780                                         }
781                                 }
782                         }
783
784                         found = 0;
785                         /* go over all variables of a node until the non-zero one is found */
786                         for (tp_idx = na->n_unit_types - 1; ! found && tp_idx >= 0; --tp_idx) {
787                                 for (t = na->asap - 1; ! found && t <= na->alap - 1; ++t) {
788                                         double val = lpp_get_var_sol(lpp, na->ilp_vars.x[cur_var++]);
789
790                                         /* check, if variable is set to one (it's not zero then :) */
791                                         if (! LPP_VALUE_IS_0(val)) {
792                                                 na->sched_point = t;
793                                                 ARR_APP1(be_ilpsched_irn_t *, sched_nodes, node);
794                                                 DBG((env->dbg, LEVEL_2, "Schedpoint of %+F is %u at unit type %s\n",
795                                                         irn, t, na->type_info[tp_idx].tp->name));
796                                                 found = 1;
797                                         }
798                                 }
799                         }
800                 }
801
802                 glob_heights = env->height;
803                 /* sort nodes ascending by scheduling time step */
804                 qsort(sched_nodes, ARR_LEN(sched_nodes), sizeof(sched_nodes[0]), cmp_ilpsched_irn);
805         }
806
807         /* make all Phis ready and remember the single cf op */
808         cfop = NULL;
809         foreach_out_edge(block, edge) {
810                 irn = get_edge_src_irn(edge);
811
812                 switch (get_irn_opcode(irn)) {
813                         case iro_Phi:
814                                 add_to_sched(env, block, irn, 0);
815                                 break;
816                         case iro_Start:
817                         case iro_End:
818                         case iro_Proj:
819                         case iro_Bad:
820                         case iro_Unknown:
821                                 break;
822                         default:
823                                 if (is_cfop(irn)) {
824                                         assert(cfop == NULL && "Highlander - there can be only one");
825                                         cfop = irn;
826                                 }
827                         break;
828                 }
829         }
830
831         /* add all nodes from list */
832         for (i = 0, l = ARR_LEN(sched_nodes); i < l; ++i) {
833                 ilpsched_node_attr_t *na = get_ilpsched_node_attr(sched_nodes[i]);
834                 if (sched_nodes[i]->irn != cfop)
835                         add_to_sched(env, block, sched_nodes[i]->irn, na->sched_point);
836         }
837
838         /* schedule control flow node if not already done */
839         if (cfop && ! sched_is_scheduled(cfop))
840                 add_to_sched(env, block, cfop, 0);
841
842         DEL_ARR_F(sched_nodes);
843 }
844
845 /***************************************************************
846  *   _____ _      _____     _____           _   _
847  *  |_   _| |    |  __ \   / ____|         | | (_)
848  *    | | | |    | |__) | | (___   ___  ___| |_ _  ___  _ __
849  *    | | | |    |  ___/   \___ \ / _ \/ __| __| |/ _ \| '_ \
850  *   _| |_| |____| |       ____) |  __/ (__| |_| | (_) | | | |
851  *  |_____|______|_|      |_____/ \___|\___|\__|_|\___/|_| |_|
852  *
853  ***************************************************************/
854
855 /**
856  * Check if node can be executed on given unit type.
857  */
858 static INLINE int is_valid_unit_type_for_node(const be_execution_unit_type_t *tp, be_ilpsched_irn_t *node) {
859         int                  i;
860         ilpsched_node_attr_t *na = get_ilpsched_node_attr(node);
861
862         for (i = na->n_unit_types - 1; i >= 0; --i) {
863                 if (na->type_info[i].tp == tp)
864                         return i;
865         }
866
867         return -1;
868 }
869
870 /************************************************
871  *                   _       _     _
872  *                  (_)     | |   | |
873  *  __   ____ _ _ __ _  __ _| |__ | | ___  ___
874  *  \ \ / / _` | '__| |/ _` | '_ \| |/ _ \/ __|
875  *   \ V / (_| | |  | | (_| | |_) | |  __/\__ \
876  *    \_/ \__,_|_|  |_|\__,_|_.__/|_|\___||___/
877  *
878  ************************************************/
879
880 static int be_ilpsched_set_type_info(be_ilpsched_env_t *env, ir_node *irn, struct obstack *obst) {
881         const be_execution_unit_t ***execunits = arch_isa_get_allowed_execution_units(env->arch_env->isa, irn);
882         unsigned                  n_unit_types = 0;
883         be_ilpsched_irn_t         *node;
884         ilpsched_node_attr_t      *na;
885         unsigned                  unit_idx, tp_idx;
886
887         /* count number of available unit types for this node */
888         for (n_unit_types = 0; execunits[n_unit_types]; ++n_unit_types)
889                 /* just count */ ;
890
891         node = get_ilpsched_irn(env, irn);
892         na   = get_ilpsched_node_attr(node);
893
894         if (! na->type_info) {
895                 na->n_unit_types = n_unit_types;
896                 na->type_info    = NEW_ARR_D(unit_type_info_t, obst, n_unit_types);
897
898                 /* fill the type info array */
899                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
900                         for (unit_idx = 0; execunits[tp_idx][unit_idx]; ++unit_idx) {
901                                 /* beware: we also count number of available units here */
902                                 if (be_machine_is_dummy_unit(execunits[tp_idx][unit_idx]))
903                                         na->is_dummy_node = 1;
904                         }
905
906                         na->type_info[tp_idx].tp      = execunits[tp_idx][0]->tp;
907                         na->type_info[tp_idx].n_units = unit_idx;
908                 }
909         }
910
911         return n_unit_types;
912 }
913
914 /**
915  * Returns the largest alap time of a user of @p irn.
916  * The user must be in block @p block.
917  */
918 static unsigned be_ilpsched_get_max_alap_user(be_ilpsched_env_t *env, ir_node *irn, ir_node *block) {
919         const ir_edge_t *edge;
920         unsigned        max_alap = 0;
921
922         foreach_out_edge(irn, edge) {
923                 ir_node *user = get_edge_src_irn(edge);
924
925                 if (get_nodes_block(user) == block) {
926                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, user);
927                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
928
929                         max_alap = MAX(max_alap, na->alap);
930                 }
931         }
932
933         assert(max_alap > 0);
934         return max_alap;
935 }
936
937 /**
938  * Create the following variables:
939  * - x_{nt}^k    binary     weigthed with: t
940  *      node n is scheduled at time step t to unit type k
941  * ==>> These variables represent the schedule
942  *
943  * - a_{nt}^k    binary     weighted with num_nodes
944  *      node n is alive at time step t on unit type k
945  *
946  * - y_{nt}^k    continuous  weighted with: num_nodes^2
947  *      register pressure over limit for unit type k
948  * ==>> These variables represent the register pressure
949  *
950  */
951 static void create_variables(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node, struct obstack *var_obst) {
952         char                  buf[1024];
953         ir_node               *irn;
954         unsigned              num_block_var, num_nodes;
955         ilp_livein_node_t     *livein;
956         ilpsched_block_attr_t *ba      = get_ilpsched_block_attr(block_node);
957         unsigned              weigth_y = ba->n_interesting_nodes * ba->n_interesting_nodes;
958         lc_timer_t            *t_var   = lc_timer_register("beilpsched_var", "create ilp variables");
959
960         ilp_timer_push(t_var);
961         num_block_var = num_nodes = 0;
962         foreach_linked_irns(ba->head_ilp_nodes, irn) {
963                 be_ilpsched_irn_t    *node;
964                 ilpsched_node_attr_t *na;
965                 unsigned             n_unit_types, tp_idx, n_var, cur_unit;
966                 unsigned             cur_var_ad, cur_var_x, cur_var_y, num_ad;
967                 int                  i;
968
969                 node         = get_ilpsched_irn(env, irn);
970                 na           = get_ilpsched_node_attr(node);
971                 n_unit_types = be_ilpsched_set_type_info(env, irn, var_obst);
972
973                 /* allocate space for ilp variables */
974                 na->ilp_vars.x = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
975                 memset(na->ilp_vars.x, -1, ARR_LEN(na->ilp_vars.x) * sizeof(na->ilp_vars.x[0]));
976
977                 /* we need these variables only for "real" nodes */
978                 if (! na->is_dummy_node) {
979                         na->ilp_vars.y = NEW_ARR_D(int, var_obst, n_unit_types * VALID_SCHED_INTERVAL(na));
980                         memset(na->ilp_vars.y, -1, ARR_LEN(na->ilp_vars.y) * sizeof(na->ilp_vars.y[0]));
981
982                         num_ad         = ba->max_steps - na->asap + 1;
983                         na->ilp_vars.a = NEW_ARR_D(int, var_obst, n_unit_types * num_ad);
984                         memset(na->ilp_vars.a, -1, ARR_LEN(na->ilp_vars.a) * sizeof(na->ilp_vars.a[0]));
985                 }
986
987                 DBG((env->dbg, LEVEL_3, "\thandling %+F (asap %u, alap %u, unit types %u):\n",
988                         irn, na->asap, na->alap, na->n_unit_types));
989
990                 cur_var_x = cur_var_ad = cur_var_y = cur_unit = n_var = 0;
991                 /* create variables */
992                 for (tp_idx = 0; tp_idx < n_unit_types; ++tp_idx) {
993                         unsigned t;
994
995                         for (t = na->asap - 1; t <= na->alap - 1; ++t) {
996                                 /* x_{nt}^k variables */
997                                 snprintf(buf, sizeof(buf), "x_n%u_%s_%u",
998                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
999                                 na->ilp_vars.x[cur_var_x++] = lpp_add_var(lpp, buf, lpp_binary, (double)(t + 1));
1000                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1001                                 /* variable counter */
1002                                 n_var++;
1003                                 num_block_var++;
1004
1005                                 if (! na->is_dummy_node) {
1006                                         /* y_{nt}^k variables */
1007                                         snprintf(buf, sizeof(buf), "y_n%u_%s_%u",
1008                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1009                                         na->ilp_vars.y[cur_var_y++] = lpp_add_var(lpp, buf, lpp_continous, (double)(weigth_y));
1010                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1011
1012                                         /* variable counter */
1013                                         n_var++;
1014                                         num_block_var++;
1015                                 }
1016                         }
1017
1018                         /* a node can die at any step t: asap(n) <= t <= U */
1019                         if (! na->is_dummy_node) {
1020                                 for (t = na->asap - 1; t <= ba->max_steps; ++t) {
1021
1022                                         /* a_{nt}^k variables */
1023                                         snprintf(buf, sizeof(buf), "a_n%u_%s_%u",
1024                                                 get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1025                                         na->ilp_vars.a[cur_var_ad++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1026                                         DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1027
1028                                         /* variable counter */
1029                                         n_var++;
1030                                         num_block_var++;
1031                                 }
1032                         }
1033
1034                         /* collect live-in nodes */
1035                         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1036                                 ir_node *pred = get_irn_n(irn, i);
1037
1038                                 if (get_nodes_block(pred) != block_node->irn && consider_for_sched(env->arch_env->isa, pred)) {
1039                                         be_ilpsched_set_type_info(env, pred, var_obst);
1040                                         if (! na->is_dummy_node) {
1041                                                 ilp_livein_node_t *entry = obstack_alloc(var_obst, sizeof(*entry));
1042                                                 entry->irn = pred;
1043                                                 entry->a   = NULL;
1044                                                 pset_insert(ba->livein_nodes, entry, (unsigned)get_irn_idx(pred));
1045                                         }
1046                                 }
1047                         }
1048                 }
1049
1050                 DB((env->dbg, LEVEL_3, "%u variables created\n", n_var));
1051                 num_nodes++;
1052         }
1053
1054         /* create alive variables a_{nt}^k for live-ins */
1055         foreach_pset(ba->livein_nodes, livein) {
1056                 be_ilpsched_irn_t    *node;
1057                 ilpsched_node_attr_t *na;
1058                 unsigned             tp_idx, var_idx;
1059                 ir_node              *irn;
1060
1061                 irn  = livein->irn;
1062                 node = get_ilpsched_irn(env, irn);
1063                 na   = get_ilpsched_node_attr(node);
1064
1065                 livein->max_alive_steps = be_ilpsched_get_max_alap_user(env, irn, block_node->irn);
1066
1067                 livein->a = NEW_ARR_D(int, var_obst, na->n_unit_types * livein->max_alive_steps);
1068                 var_idx   = 0;
1069
1070                 /* create variables */
1071                 for (tp_idx = 0; tp_idx < na->n_unit_types; ++tp_idx) {
1072                         unsigned t;
1073
1074                         for (t = 0; t < livein->max_alive_steps; ++t) {
1075                                 /* a_{nt}^k variables */
1076                                 snprintf(buf, sizeof(buf), "al_n%u_%s_%u",
1077                                         get_irn_idx(irn), na->type_info[tp_idx].tp->name, t);
1078                                 livein->a[var_idx++] = lpp_add_var(lpp, buf, lpp_binary, (double)(ba->n_interesting_nodes));
1079                                 DBG((env->dbg, LEVEL_4, "\t\tcreated ILP variable %s\n", buf));
1080                                 num_block_var++;
1081                         }
1082                 }
1083         }
1084
1085         ilp_timer_pop();
1086         DBG((env->dbg, LEVEL_1, "... %u variables for %u nodes created (%g sec)\n",
1087                 num_block_var, num_nodes, ilp_timer_elapsed_usec(t_var) / 1000000.0));
1088 }
1089
1090 /*******************************************************
1091  *                      _             _       _
1092  *                     | |           (_)     | |
1093  *   ___ ___  _ __  ___| |_ _ __ __ _ _ _ __ | |_ ___
1094  *  / __/ _ \| '_ \/ __| __| '__/ _` | | '_ \| __/ __|
1095  * | (_| (_) | | | \__ \ |_| | | (_| | | | | | |_\__ \
1096  *  \___\___/|_| |_|___/\__|_|  \__,_|_|_| |_|\__|___/
1097  *
1098  *******************************************************/
1099
1100 /**
1101  * Create following ILP constraints:
1102  * - the assignment constraints:
1103  *     assure each node is executed once by exactly one (allowed) execution unit
1104  * - the dead node assignment constraints:
1105  *     assure a node can only die at most once
1106  * - the precedence constraints:
1107  *     assure that no data dependencies are violated
1108  */
1109 static void create_assignment_and_precedence_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1110         unsigned              num_cst_assign, num_cst_prec, num_cst_dead;
1111         char                  buf[1024];
1112         ir_node               *irn;
1113         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1114         bitset_t              *bs_block_irns = bitset_alloca(ba->block_last_idx);
1115         lc_timer_t            *t_cst_assign  = lc_timer_register("beilpsched_cst_assign", "create assignment constraints");
1116         lc_timer_t            *t_cst_prec    = lc_timer_register("beilpsched_cst_prec",   "create precedence constraints");
1117
1118         num_cst_assign = num_cst_prec = num_cst_dead = 0;
1119         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1120                 int                  cst, tp_idx, i;
1121                 unsigned             cur_var;
1122                 be_ilpsched_irn_t    *node;
1123                 ilpsched_node_attr_t *na;
1124
1125                 node    = get_ilpsched_irn(env, irn);
1126                 na      = get_ilpsched_node_attr(node);
1127                 cur_var = 0;
1128
1129                 /* the assignment constraint */
1130                 ilp_timer_push(t_cst_assign);
1131                 snprintf(buf, sizeof(buf), "assignment_cst_n%u", get_irn_idx(irn));
1132                 cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 1.0);
1133                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1134                 num_cst_assign++;
1135
1136                 lpp_set_factor_fast_bulk(lpp, cst, na->ilp_vars.x, ARR_LEN(na->ilp_vars.x), 1.0);
1137                 ilp_timer_pop();
1138
1139                 /* We have separate constraints for Projs and Keeps */
1140                 // ILP becomes infeasible ?!?
1141 //              if (is_Proj(irn) || be_is_Keep(irn))
1142 //                      continue;
1143
1144                 /* the precedence constraints */
1145                 ilp_timer_push(t_cst_prec);
1146                 bs_block_irns = bitset_clear_all(bs_block_irns);
1147                 for (i = get_irn_ins_or_deps(irn) - 1; i >= 0; --i) {
1148                         ir_node              *pred = skip_normal_Proj(env->arch_env->isa, get_irn_in_or_dep(irn, i));
1149                         unsigned             t_low, t_high, t;
1150                         be_ilpsched_irn_t    *pred_node;
1151                         ilpsched_node_attr_t *pna;
1152                         unsigned             delay;
1153
1154                         if (is_Phi(pred) || block_node->irn != get_nodes_block(pred) || is_NoMem(pred))
1155                                 continue;
1156
1157                         pred_node = get_ilpsched_irn(env, pred);
1158                         pna       = get_ilpsched_node_attr(pred_node);
1159
1160                         assert(pna->asap > 0 && pna->alap >= pna->asap && "Invalid scheduling interval.");
1161
1162                         if (! bitset_is_set(bs_block_irns, pna->block_idx))
1163                                 bitset_set(bs_block_irns, pna->block_idx);
1164                         else
1165                                 continue;
1166
1167                         /* irn = n, pred = m */
1168                         delay  = fixed_latency(env->sel, pred, env->block_env);
1169                         t_low  = MAX(na->asap, pna->asap + delay - 1);
1170                         t_high = MIN(na->alap, pna->alap + delay - 1);
1171                         for (t = t_low - 1; t <= t_high - 1; ++t) {
1172                                 unsigned tn, tm;
1173                                 int      *tmp_var_idx = NEW_ARR_F(int, 0);
1174
1175                                 snprintf(buf, sizeof(buf), "precedence_n%u_n%u_%u", get_irn_idx(pred), get_irn_idx(irn), t);
1176                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 1.0);
1177                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1178                                 num_cst_prec++;
1179
1180                                 /* lpp_set_factor_fast_bulk needs variables sorted ascending by index */
1181                                 if (na->ilp_vars.x[0] < pna->ilp_vars.x[0]) {
1182                                         /* node variables have smaller index than pred variables */
1183                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1184                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1185                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1186                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1187                                                 }
1188                                         }
1189
1190                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1191                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1192                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1193                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1194                                                 }
1195                                         }
1196                                 }
1197                                 else {
1198                                         /* pred variables have smaller index than node variables */
1199                                         for (tp_idx = pna->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1200                                                 for (tm = t - delay + 1; tm < pna->alap; ++tm) {
1201                                                         unsigned idx = ILPVAR_IDX(pna, tp_idx, tm);
1202                                                         ARR_APP1(int, tmp_var_idx, pna->ilp_vars.x[idx]);
1203                                                 }
1204                                         }
1205
1206                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1207                                                 for (tn = na->asap - 1; tn <= t; ++tn) {
1208                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1209                                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1210                                                 }
1211                                         }
1212                                 }
1213
1214                                 if (ARR_LEN(tmp_var_idx) > 0)
1215                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1216
1217                                 DEL_ARR_F(tmp_var_idx);
1218                         }
1219                 }
1220                 ilp_timer_pop();
1221         }
1222         DBG((env->dbg, LEVEL_1, "\t%u assignement constraints (%g sec)\n",
1223                 num_cst_assign, ilp_timer_elapsed_usec(t_cst_assign) / 1000000.0));
1224         DBG((env->dbg, LEVEL_1, "\t%u precedence constraints (%g sec)\n",
1225                 num_cst_prec, ilp_timer_elapsed_usec(t_cst_prec) / 1000000.0));
1226 }
1227
1228 /**
1229  * Create ILP resource constraints:
1230  * - assure that for each time step not more instructions are scheduled
1231  *   to the same unit types as units of this type are available
1232  */
1233 static void create_ressource_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1234         int                   glob_type_idx;
1235         char                  buf[1024];
1236         unsigned              num_cst_resrc = 0;
1237         ilpsched_block_attr_t *ba           = get_ilpsched_block_attr(block_node);
1238         lc_timer_t            *t_cst_rsrc   = lc_timer_register("beilpsched_cst_rsrc",   "create resource constraints");
1239
1240         ilp_timer_push(t_cst_rsrc);
1241         for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1242                 unsigned                 t;
1243                 be_execution_unit_type_t *cur_tp = &env->cpu->unit_types[glob_type_idx];
1244
1245                 /* BEWARE: the DUMMY unit type is not in CPU, so it's skipped automatically */
1246
1247                 /* check each time step */
1248                 for (t = 0; t < ba->max_steps; ++t) {
1249                         ir_node *irn;
1250                         int     cst;
1251                         int     *tmp_var_idx = NEW_ARR_F(int, 0);
1252
1253                         snprintf(buf, sizeof(buf), "resource_cst_%s_%u", cur_tp->name, t);
1254                         cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)cur_tp->n_units);
1255                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1256                         num_cst_resrc++;
1257
1258                         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1259                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1260                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1261                                 int                  tp_idx;
1262
1263                                 tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1264
1265                                 if (tp_idx >= 0 && t >= na->asap - 1 && t <= na->alap - 1) {
1266                                         int cur_var = ILPVAR_IDX(na, tp_idx, t);
1267                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[cur_var]);
1268                                 }
1269                         }
1270
1271                         /* set constraints if we have some */
1272                         if (ARR_LEN(tmp_var_idx) > 0)
1273                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1274
1275                         DEL_ARR_F(tmp_var_idx);
1276                 }
1277         }
1278         ilp_timer_pop();
1279         DBG((env->dbg, LEVEL_1, "\t%u resource constraints (%g sec)\n",
1280                 num_cst_resrc, ilp_timer_elapsed_usec(t_cst_rsrc) / 1000000.0));
1281 }
1282
1283 /**
1284  * Create ILP bundle constraints:
1285  * - assure, at most bundle_size * bundles_per_cycle instructions
1286  *   can be started at a certain point.
1287  */
1288 static void create_bundle_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1289         char                  buf[1024];
1290         unsigned              t;
1291         unsigned              num_cst_bundle = 0;
1292         unsigned              n_instr_max    = env->cpu->bundle_size * env->cpu->bundels_per_cycle;
1293         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1294         lc_timer_t            *t_cst_bundle  = lc_timer_register("beilpsched_cst_bundle", "create bundle constraints");
1295
1296         ilp_timer_push(t_cst_bundle);
1297         for (t = 0; t < ba->max_steps; ++t) {
1298                 ir_node *irn;
1299                 int     cst;
1300                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1301
1302                 snprintf(buf, sizeof(buf), "bundle_cst_%u", t);
1303                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)n_instr_max);
1304                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1305                 num_cst_bundle++;
1306
1307                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1308                         be_ilpsched_irn_t    *node;
1309                         ilpsched_node_attr_t *na;
1310                         int                  tp_idx;
1311
1312                         /* Projs and Keeps do not contribute to bundle size */
1313                         if (is_Proj(irn) || be_is_Keep(irn))
1314                                 continue;
1315
1316                         node = get_ilpsched_irn(env, irn);
1317                         na   = get_ilpsched_node_attr(node);
1318
1319                         /* nodes assigned to DUMMY unit do not contribute to bundle size */
1320                         if (na->is_dummy_node)
1321                                 continue;
1322
1323                         if (t >= na->asap - 1 && t <= na->alap - 1) {
1324                                 for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1325                                         int idx = ILPVAR_IDX(na, tp_idx, t);
1326                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.x[idx]);
1327                                 }
1328                         }
1329                 }
1330
1331                 if (ARR_LEN(tmp_var_idx) > 0)
1332                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1333
1334                 DEL_ARR_F(tmp_var_idx);
1335         }
1336         ilp_timer_pop();
1337         DBG((env->dbg, LEVEL_1, "\t%u bundle constraints (%g sec)\n",
1338                 num_cst_bundle, ilp_timer_elapsed_usec(t_cst_bundle) / 1000000.0));
1339 }
1340
1341 /**
1342  * Create ILP alive nodes constraints:
1343  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1344  */
1345 static void create_alive_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1346         char                  buf[1024];
1347         ir_node               *irn;
1348         unsigned              num_cst = 0;
1349         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1350         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_nodes", "create alive nodes constraints");
1351
1352         ilp_timer_push(t_cst);
1353         /* for each node */
1354         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1355                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1356                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1357                 unsigned             t;
1358
1359                 /* we ignore nodes assigned to dummy unit here */
1360                 if (na->is_dummy_node)
1361                         continue;
1362
1363                 /* check check all time steps: asap(n) <= t <= U */
1364                 for (t = na->asap - 1; t < ba->max_steps; ++t) {
1365                         int node_tp_idx;
1366
1367                         /* for all unit types available for this node */
1368                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1369                                 unsigned tn, tn_max, idx;
1370                                 int      cst, i;
1371                                 int      *tmp_var_idx_n = NEW_ARR_F(int, 0);
1372                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1373
1374                                 snprintf(buf, sizeof(buf), "alive_node_cst_%u_n%u_%s",
1375                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1376                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, 0.0);
1377                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1378                                 num_cst++;
1379
1380                                 tn_max = MIN(na->alap - 1, t);
1381                                 /* check if the node has been scheduled so far */
1382                                 for (tn = na->asap - 1; tn <= tn_max; ++tn) {
1383                                         int idx = ILPVAR_IDX(na, node_tp_idx, tn);
1384                                         ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1385                                 }
1386
1387                                 if (ARR_LEN(tmp_var_idx_n) > 0)
1388                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(na->n_consumer));
1389                                 DEL_ARR_F(tmp_var_idx_n);
1390
1391                                 /* subtract the number of consumer scheduled so far */
1392                                 for (i = ARR_LEN(na->block_consumer) - 1; i >= 0; --i) {
1393                                         be_ilpsched_irn_t    *cons = get_ilpsched_irn(env, na->block_consumer[i]);
1394                                         ilpsched_node_attr_t *ca   = get_ilpsched_node_attr(cons);
1395                                         int                  tp_idx;
1396                                         unsigned             tm, tm_max;
1397
1398                                         tm_max = MIN(ca->alap - 1, t);
1399                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1400                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1401                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1402                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1403                                                 }
1404                                         }
1405                                 }
1406
1407                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1408                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), -1.0);
1409                                 DEL_ARR_F(tmp_var_idx_m);
1410
1411                                 /* -c * a_{nt}^k */
1412                                 idx = ILPVAR_IDX_DEAD(ba, na, node_tp_idx, t);
1413                                 lpp_set_factor_fast(lpp, cst, na->ilp_vars.a[idx], 0.0 - (double)(na->n_consumer));
1414
1415                         }
1416                 }
1417         }
1418         ilp_timer_pop();
1419         DBG((env->dbg, LEVEL_1, "\t%u alive nodes constraints (%g sec)\n",
1420                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1421 }
1422
1423 /**
1424  * Create ILP alive nodes constraints for live-in nodes:
1425  * - set variable a_{nt}^k to 1 if nodes n is alive at step t on unit k
1426  */
1427 static void create_alive_livein_nodes_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1428         char                  buf[1024];
1429         ilp_livein_node_t     *livein;
1430         unsigned              num_cst = 0;
1431         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1432         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_alive_livein_nodes", "create alive livein nodes constraints");
1433
1434         ilp_timer_push(t_cst);
1435         /* for each node */
1436         foreach_pset(ba->livein_nodes, livein) {
1437                 ir_node              *irn  = livein->irn;
1438                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1439                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1440                 unsigned             t;
1441
1442                 /* check check all time steps: 0 <= t < max_alive_steps */
1443                 for (t = 0; t < livein->max_alive_steps; ++t) {
1444                         int node_tp_idx;
1445
1446                         /* for all unit types available for this node */
1447                         for (node_tp_idx = na->n_unit_types - 1; node_tp_idx >= 0; --node_tp_idx) {
1448                                 const ir_edge_t *edge;
1449                                 unsigned idx;
1450                                 int      cst, num_block_user;
1451                                 int      *tmp_var_idx_m = NEW_ARR_F(int, 0);
1452
1453                                 /* check the number of consumer scheduled so far */
1454                                 num_block_user = 0;
1455                                 foreach_out_edge(irn, edge) {
1456                                         ir_node              *user = get_edge_src_irn(edge);
1457                                         be_ilpsched_irn_t    *cons;
1458                                         ilpsched_node_attr_t *ca;
1459                                         int                  tp_idx;
1460                                         unsigned             tm, tm_max;
1461
1462                                         /* check only users within current block */
1463                                         if (get_nodes_block(user) != block_node->irn)
1464                                                 continue;
1465
1466                                         num_block_user++;
1467                                         cons = get_ilpsched_irn(env, user);
1468                                         ca   = get_ilpsched_node_attr(cons);
1469
1470                                         tm_max = MIN(ca->alap - 1, t);
1471                                         for (tp_idx = ca->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1472                                                 for (tm = ca->asap - 1; tm <= tm_max; ++tm) {
1473                                                         int idx = ILPVAR_IDX(ca, tp_idx, tm);
1474                                                         ARR_APP1(int, tmp_var_idx_m, ca->ilp_vars.x[idx]);
1475                                                 }
1476                                         }
1477                                 }
1478
1479                                 snprintf(buf, sizeof(buf), "alive_livein_node_cst_%u_n%u_%s",
1480                                         t, get_irn_idx(irn), na->type_info[node_tp_idx].tp->name);
1481                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, (double)num_block_user);
1482                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1483                                 num_cst++;
1484
1485                                 /* sum(scheduled users) */
1486                                 if (ARR_LEN(tmp_var_idx_m) > 0)
1487                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_m, ARR_LEN(tmp_var_idx_m), 1.0);
1488                                 DEL_ARR_F(tmp_var_idx_m);
1489
1490                                 /* + c * a_{nt}^k */
1491                                 idx = node_tp_idx * livein->max_alive_steps + t;
1492                                 lpp_set_factor_fast(lpp, cst, livein->a[idx], (double)(num_block_user));
1493                         }
1494                 }
1495         }
1496         ilp_timer_pop();
1497         DBG((env->dbg, LEVEL_1, "\t%u alive livein nodes constraints (%g sec)\n",
1498                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1499 }
1500
1501 /**
1502  * Create ILP pressure constraints, based on alive nodes:
1503  * - add additional costs to objective function if a node is scheduled
1504  *   on a unit although all units of this type are currently occupied
1505  */
1506 static void create_pressure_alive_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1507         char                  buf[1024];
1508         ir_node               *cur_irn;
1509         unsigned              num_cst = 0;
1510         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1511         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_pressure", "create pressure constraints");
1512
1513         ilp_timer_push(t_cst);
1514         /* y_{nt}^k is set for each node and timestep and unit type */
1515         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1516                 unsigned             cur_idx   = get_irn_idx(cur_irn);
1517                 be_ilpsched_irn_t    *cur_node = get_ilpsched_irn(env, cur_irn);
1518                 ilpsched_node_attr_t *cur_na   = get_ilpsched_node_attr(cur_node);
1519                 int                  glob_type_idx;
1520
1521                 /* we ignore nodes assigned to DUMMY unit here */
1522                 if (cur_na->is_dummy_node)
1523                         continue;
1524
1525                 /* for all types */
1526                 for (glob_type_idx = env->cpu->n_unit_types - 1; glob_type_idx >= 0; --glob_type_idx) {
1527                         be_execution_unit_type_t *cur_tp   = &env->cpu->unit_types[glob_type_idx];
1528                         int                      cur_tp_idx;
1529                         unsigned                 t;
1530
1531                         /* BEWARE: the DUMMY unit types is not in CPU, so it's skipped automatically */
1532
1533                         /* check if node can be executed on this unit type */
1534                         cur_tp_idx = is_valid_unit_type_for_node(cur_tp, cur_node);
1535                         if (cur_tp_idx < 0)
1536                                 continue;
1537
1538                         /* check all time_steps at which the current node can be scheduled */
1539                         for (t = cur_na->asap - 1; t <= cur_na->alap - 1; ++t) {
1540                                 int     cst, y_idx;
1541                                 ir_node *irn;
1542                                 int     *tmp_var_idx = NEW_ARR_F(int, 0);
1543                                 ilp_livein_node_t *livein;
1544
1545                                 snprintf(buf, sizeof(buf), "pressure_cst_n%u_%u_%s", cur_idx, t, cur_tp->name);
1546                                 cst = lpp_add_cst_uniq(lpp, buf, lpp_less, (double)(cur_tp->n_units - 1));
1547                                 DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1548                                 num_cst++;
1549
1550                                 /* - accumulate all nodes alive at point t on unit type k */
1551                                 foreach_linked_irns(ba->head_ilp_nodes, irn) {
1552                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1553                                         ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1554                                         int                  a_idx, tp_idx;
1555
1556                                         /* check if node can be alive here */
1557                                         if (t < na->asap - 1)
1558                                                 continue;
1559
1560                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1561
1562                                         /* current type is not suitable */
1563                                         if (tp_idx < 0)
1564                                                 continue;
1565
1566                                         a_idx = ILPVAR_IDX_DEAD(ba, na, tp_idx, t);
1567                                         ARR_APP1(int, tmp_var_idx, na->ilp_vars.a[a_idx]);
1568                                 }
1569                                 /* do the same for livein nodes */
1570                                 foreach_pset(ba->livein_nodes, livein) {
1571                                         ir_node              *irn  = livein->irn;
1572                                         be_ilpsched_irn_t    *node = get_ilpsched_irn(env, irn);
1573                                         int                  a_idx, tp_idx;
1574
1575                                         /* check if node can be alive here */
1576                                         if (t >= livein->max_alive_steps)
1577                                                 continue;
1578
1579                                         tp_idx = is_valid_unit_type_for_node(cur_tp, node);
1580
1581                                         /* current type is not suitable */
1582                                         if (tp_idx < 0)
1583                                                 continue;
1584
1585                                         a_idx = tp_idx * livein->max_alive_steps + t;
1586                                         ARR_APP1(int, tmp_var_idx, livein->a[a_idx]);
1587                                 }
1588
1589                                 if (ARR_LEN(tmp_var_idx) > 0)
1590                                         lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx, ARR_LEN(tmp_var_idx), 1.0);
1591                                 DEL_ARR_F(tmp_var_idx);
1592
1593                                 /* - num_nodes * y_{nt}^k */
1594                                 y_idx = ILPVAR_IDX(cur_na, cur_tp_idx, t);
1595                                 lpp_set_factor_fast(lpp, cst, cur_na->ilp_vars.y[y_idx], -1.0);
1596                         }
1597                 }
1598         }
1599         ilp_timer_pop();
1600         DBG((env->dbg, LEVEL_1, "\t%u pressure constraints (%g sec)\n",
1601                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1602 }
1603
1604 /**
1605  * Create ILP branch constraints:
1606  * Assure, alle nodes are scheduled prior to cfg op.
1607  */
1608 static void create_branch_constraint(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1609         char                  buf[1024];
1610         ir_node               *cur_irn, *cfop;
1611         unsigned              num_cst          = 0;
1612         unsigned              num_non_branches = 0;
1613         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1614         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_branch", "create branch constraints");
1615
1616         ilp_timer_push(t_cst);
1617         cfop = NULL;
1618         /* determine number of non-branch nodes and the one and only branch node */
1619         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1620                 switch (get_irn_opcode(cur_irn)) {
1621                         case iro_Phi:
1622                         case iro_Start:
1623                         case iro_End:
1624                         case iro_Proj:
1625                         case iro_Bad:
1626                         case iro_Unknown:
1627                                 num_non_branches++;
1628                                 break;
1629                         default:
1630                                 if (is_cfop(cur_irn)) {
1631                                         assert(cfop == NULL && "Highlander - there can be only one to be constrained");
1632                                         cfop = cur_irn;
1633                                 }
1634                                 else {
1635                                         num_non_branches++;
1636                                 }
1637                                 break;
1638                 }
1639         }
1640
1641         if (cfop) {
1642                 be_ilpsched_irn_t    *cf_node = get_ilpsched_irn(env, cfop);
1643                 ilpsched_node_attr_t *cf_na   = get_ilpsched_node_attr(cf_node);
1644                 unsigned t;
1645
1646                 /* for each time step */
1647                 for (t = cf_na->asap - 1; t <= cf_na->alap - 1; ++t) {
1648                         int *non_branch_vars, *branch_vars;
1649                         int cst;
1650
1651                         snprintf(buf, sizeof(buf), "branch_cst_%u_n%u", t, get_irn_idx(cfop));
1652                         cst = lpp_add_cst_uniq(lpp, buf, lpp_greater, 0.0);
1653                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1654                         num_cst++;
1655
1656                         /* sum(overall non branches: n)x_{nt}^k - sum(overall branches: b)(num_non_branches * x_{bt}^k >= 0) */
1657                         non_branch_vars = NEW_ARR_F(int, 0);
1658                         branch_vars     = NEW_ARR_F(int, 0);
1659                         foreach_linked_irns(ba->head_ilp_nodes, cur_irn) {
1660                                 be_ilpsched_irn_t    *node = get_ilpsched_irn(env, cur_irn);
1661                                 ilpsched_node_attr_t *na   = get_ilpsched_node_attr(node);
1662                                 int                  tp_idx;
1663
1664                                 if (cur_irn == cfop) {
1665                                         /* for all unit types available for this node */
1666                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1667                                                 unsigned idx = ILPVAR_IDX(na, tp_idx, t);
1668                                                 ARR_APP1(int, branch_vars, na->ilp_vars.x[idx]);
1669                                         }
1670                                 }
1671                                 else {
1672                                         /* sum up all possible schedule points for this node upto current timestep */
1673                                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1674                                                 unsigned tn;
1675                                                 unsigned tmax = MIN(t, na->alap - 1);
1676
1677                                                 for (tn = na->asap - 1; tn <= tmax; ++tn) {
1678                                                         unsigned idx = ILPVAR_IDX(na, tp_idx, tn);
1679                                                         ARR_APP1(int, non_branch_vars, na->ilp_vars.x[idx]);
1680                                                 }
1681                                         }
1682                                 }
1683
1684                         }
1685
1686                         if (ARR_LEN(non_branch_vars) > 0)
1687                                 lpp_set_factor_fast_bulk(lpp, cst, non_branch_vars, ARR_LEN(non_branch_vars), 1.0);
1688                         if (ARR_LEN(branch_vars) > 0)
1689                                 lpp_set_factor_fast_bulk(lpp, cst, branch_vars, ARR_LEN(branch_vars), 0.0 - (double)num_non_branches);
1690
1691                         DEL_ARR_F(branch_vars);
1692                         DEL_ARR_F(non_branch_vars);
1693                 }
1694         }
1695         ilp_timer_pop();
1696         DBG((env->dbg, LEVEL_1, "\t%u branch constraints (%g sec)\n",
1697                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1698 }
1699
1700 #if 0
1701 static void create_proj_keep_constraints(be_ilpsched_env_t *env, lpp_t *lpp, be_ilpsched_irn_t *block_node) {
1702         char                  buf[1024];
1703         ir_node               *irn;
1704         unsigned              num_cst = 0;
1705         ilpsched_block_attr_t *ba     = get_ilpsched_block_attr(block_node);
1706         lc_timer_t            *t_cst  = lc_timer_register("beilpsched_cst_projkeep", "create proj and keep constraints");
1707
1708         ilp_timer_push(t_cst);
1709         /* check all nodes */
1710         foreach_linked_irns(ba->head_ilp_nodes, irn) {
1711                 be_ilpsched_irn_t    *node;
1712                 ilpsched_node_attr_t *na;
1713                 unsigned             t;
1714                 ir_node              **pk;
1715
1716                 /* only mode_T nodes can have Projs and Keeps assigned */
1717                 if (get_irn_mode(irn) != mode_T)
1718                         continue;
1719
1720                 node = get_ilpsched_irn(env, irn);
1721                 na   = get_ilpsched_node_attr(node);
1722
1723                 /* check if has some Projs and Keeps assigned */
1724                 if (! na->projkeeps)
1725                         continue;
1726
1727                 /* we can run only once over the queue, so preserve the nodes */
1728                 pk = NEW_ARR_F(ir_node *, 0);
1729                 while (! waitq_empty(na->projkeeps))
1730                         ARR_APP1(ir_node *, pk, waitq_get(na->projkeeps));
1731                 del_waitq(na->projkeeps);
1732                 na->projkeeps = NULL;
1733
1734                 /* for all time steps at which this node can be scheduled */
1735                 for (t = na->asap - 1; t <= na->alap - 1; ++t) {
1736                         int cst, tp_idx, i;
1737                         int *tmp_var_idx_n = NEW_ARR_F(int, 0);
1738
1739                         /* add the constraint, assure, that a node is always scheduled along with it's Projs and Keeps */
1740                         snprintf(buf, sizeof(buf), "projkeep_cst_n%u_%u", get_irn_idx(irn), t);
1741                         cst = lpp_add_cst_uniq(lpp, buf, lpp_equal, 0.0);
1742                         DBG((env->dbg, LEVEL_2, "added constraint %s\n", buf));
1743                         num_cst++;
1744
1745                         /* sum up scheduling variables for this time step */
1746                         for (tp_idx = na->n_unit_types - 1; tp_idx >= 0; --tp_idx) {
1747                                 int idx = ILPVAR_IDX(na, tp_idx, t);
1748                                 ARR_APP1(int, tmp_var_idx_n, na->ilp_vars.x[idx]);
1749                         }
1750
1751                         if (ARR_LEN(tmp_var_idx_n) > 0)
1752                                 lpp_set_factor_fast_bulk(lpp, cst, tmp_var_idx_n, ARR_LEN(tmp_var_idx_n), (double)(ARR_LEN(pk)));
1753                         DEL_ARR_F(tmp_var_idx_n);
1754
1755                         /* subtract all Proj and Keep variables for this step */
1756                         for (i = ARR_LEN(pk) - 1; i >= 0; --i) {
1757                                 be_ilpsched_irn_t    *pk_node = get_ilpsched_irn(env, pk[i]);
1758                                 ilpsched_node_attr_t *pk_na   = get_ilpsched_node_attr(pk_node);
1759                                 int                  pk_tp_idx;
1760
1761                                 for (pk_tp_idx = pk_na->n_unit_types - 1; pk_tp_idx >= 0; --pk_tp_idx) {
1762                                         int idx = ILPVAR_IDX(pk_na, pk_tp_idx, t);
1763                                         lpp_set_factor_fast(lpp, cst, pk_na->ilp_vars.x[idx], -1.0);
1764                                 }
1765                         }
1766                 }
1767         }
1768         ilp_timer_pop();
1769         DBG((env->dbg, LEVEL_1, "\t%u Proj and Keep constraints (%g sec)\n",
1770                 num_cst, ilp_timer_elapsed_usec(t_cst) / 1000000.0));
1771 }
1772 #endif /* if 0 */
1773
1774 /***************************************************
1775  *  _____ _      _____                    _
1776  * |_   _| |    |  __ \                  (_)
1777  *   | | | |    | |__) |  _ __ ___   __ _ _ _ __
1778  *   | | | |    |  ___/  | '_ ` _ \ / _` | | '_ \
1779  *  _| |_| |____| |      | | | | | | (_| | | | | |
1780  * |_____|______|_|      |_| |_| |_|\__,_|_|_| |_|
1781  *
1782  ***************************************************/
1783
1784 /**
1785  * Create the ilp (add variables, build constraints, solve, build schedule from solution).
1786  */
1787 static void create_ilp(ir_node *block, void *walk_env) {
1788         be_ilpsched_env_t     *env           = walk_env;
1789         be_ilpsched_irn_t     *block_node    = get_ilpsched_irn(env, block);
1790         ilpsched_block_attr_t *ba            = get_ilpsched_block_attr(block_node);
1791         FILE                  *logfile       = NULL;
1792         lpp_t                 *lpp           = NULL;
1793         int                   need_heur      = 0;
1794         struct obstack        var_obst;
1795         char                  name[1024];
1796
1797         DBG((env->dbg, 255, "\n\n\n=========================================\n"));
1798         DBG((env->dbg, 255, "  ILP Scheduling for %+F\n", block));
1799         DBG((env->dbg, 255, "=========================================\n\n"));
1800
1801         DBG((env->dbg, LEVEL_1, "Creating ILP Variables for nodes in %+F (%u interesting nodes, %u max steps)\n",
1802                 block, ba->n_interesting_nodes, ba->max_steps));
1803
1804         /* notify backend and get block environment */
1805         env->block_env = be_ilp_sched_init_block_ilp_schedule(env->sel, block);
1806
1807         /* if we have less than two interesting nodes, there is no need to create the ILP */
1808         if (ba->n_interesting_nodes > 1) {
1809                 double fact_var        = ba->n_interesting_nodes > 25 ? 2.3 : 3;
1810                 double fact_cst        = ba->n_interesting_nodes > 25 ? 3   : 4.5;
1811                 int    base_num        = ba->n_interesting_nodes * ba->n_interesting_nodes;
1812                 int    estimated_n_var = (int)((double)base_num * fact_var);
1813                 int    estimated_n_cst = (int)((double)base_num * fact_cst);
1814
1815                 DBG((env->dbg, LEVEL_1, "Creating LPP with estimated numbers: %d vars, %d cst\n",
1816                         estimated_n_var, estimated_n_cst));
1817
1818                 /* set up the LPP object */
1819                 snprintf(name, sizeof(name), "ilp scheduling IRG %s", get_entity_ld_name(get_irg_entity(env->irg)));
1820
1821                 lpp = new_lpp_userdef(
1822                         (const char *)name,
1823                         lpp_minimize,
1824                         estimated_n_cst,     /* num vars */
1825                         estimated_n_cst + 1, /* num cst */
1826                         1.3);                /* grow factor */
1827                 obstack_init(&var_obst);
1828
1829                 /* create ILP variables */
1830                 create_variables(env, lpp, block_node, &var_obst);
1831
1832                 /* create ILP constraints */
1833                 DBG((env->dbg, LEVEL_1, "Creating constraints for nodes in %+F:\n", block));
1834                 create_assignment_and_precedence_constraints(env, lpp, block_node);
1835                 create_ressource_constraints(env, lpp, block_node);
1836                 create_bundle_constraints(env, lpp, block_node);
1837                 create_branch_constraint(env, lpp, block_node);
1838                 //create_proj_keep_constraints(env, lpp, block_node);
1839
1840                 if (env->opts->regpress) {
1841                         create_alive_nodes_constraint(env, lpp, block_node);
1842                         create_alive_livein_nodes_constraint(env, lpp, block_node);
1843                         create_pressure_alive_constraint(env, lpp, block_node);
1844                 }
1845
1846                 DBG((env->dbg, LEVEL_1, "ILP to solve: %u variables, %u constraints\n", lpp->var_next, lpp->cst_next));
1847
1848                 /* debug stuff, dump lpp when debugging is on  */
1849                 DEBUG_ONLY(
1850                         if (firm_dbg_get_mask(env->dbg) > 1) {
1851                                 char buf[1024];
1852                                 FILE *f;
1853
1854                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.txt", get_irn_node_nr(block));
1855                                 f = fopen(buf, "w");
1856                                 lpp_dump_plain(lpp, f);
1857                                 fclose(f);
1858                                 snprintf(buf, sizeof(buf), "lpp_block_%lu.mps", get_irn_node_nr(block));
1859                                 lpp_dump(lpp, buf);
1860                         }
1861                 );
1862
1863                 /* set solve time limit */
1864                 lpp_set_time_limit(lpp, env->opts->time_limit);
1865
1866                 /* set logfile if requested */
1867                 if (strlen(env->opts->log_file) > 0) {
1868                         if (strcasecmp(env->opts->log_file, "stdout") == 0)
1869                                 lpp_set_log(lpp, stdout);
1870                         else if (strcasecmp(env->opts->log_file, "stderr") == 0)
1871                                 lpp_set_log(lpp, stderr);
1872                         else {
1873                                 logfile = fopen(env->opts->log_file, "w");
1874                                 if (! logfile)
1875                                         fprintf(stderr, "Could not open logfile '%s'! Logging disabled.\n", env->opts->log_file);
1876                                 else
1877                                         lpp_set_log(lpp, logfile);
1878                         }
1879                 }
1880
1881                 /* solve the ILP */
1882                 lpp_solve_net(lpp, env->main_env->options->ilp_server, env->main_env->options->ilp_solver);
1883
1884                 if (logfile)
1885                         fclose(logfile);
1886
1887                 /* check for valid solution */
1888                 if (! lpp_is_sol_valid(lpp)) {
1889                         char buf[1024];
1890                         FILE *f;
1891
1892                         DEBUG_ONLY(
1893                                 if (firm_dbg_get_mask(env->dbg) >= 2) {
1894                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.txt", get_irn_node_nr(block));
1895                                         f = fopen(buf, "w");
1896                                         lpp_dump_plain(lpp, f);
1897                                         fclose(f);
1898                                         snprintf(buf, sizeof(buf), "lpp_block_%lu.infeasible.mps", get_irn_node_nr(block));
1899                                         lpp_dump(lpp, buf);
1900                                         dump_ir_block_graph(env->irg, "-infeasible");
1901                                 }
1902                         )
1903
1904                         ir_fprintf(stderr, "ILP found no solution within time (%+F, %+F), falling back to heuristics.\n", block, env->irg);
1905                         need_heur = 1;
1906                 }
1907
1908                 DBG((env->dbg, LEVEL_1, "\nSolution:\n"));
1909                 DBG((env->dbg, LEVEL_1, "\tsend time: %g sec\n", lpp->send_time / 1000000.0));
1910                 DBG((env->dbg, LEVEL_1, "\treceive time: %g sec\n", lpp->recv_time / 1000000.0));
1911                 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));
1912                 DBG((env->dbg, LEVEL_1, "\titerations: %d\n", lpp->iterations));
1913                 DBG((env->dbg, LEVEL_1, "\tsolution time: %g\n", lpp->sol_time));
1914                 DBG((env->dbg, LEVEL_1, "\tobjective function: %g\n", LPP_VALUE_IS_0(lpp->objval) ? 0.0 : lpp->objval));
1915                 DBG((env->dbg, LEVEL_1, "\tbest bound: %g\n", LPP_VALUE_IS_0(lpp->best_bound) ? 0.0 : lpp->best_bound));
1916
1917                 DBG((env->dbg, LEVEL_1, "variables used %u bytes\n", obstack_memory_used(&var_obst)));
1918         }
1919
1920         /* apply solution */
1921 #ifdef FIRM_STATISTICS
1922         if (be_stat_ev_is_active()) {
1923                 be_stat_ev("nodes", ba->block_last_idx);
1924                 be_stat_ev("vars", lpp ? lpp->var_next : 0);
1925                 be_stat_ev("csts", lpp ? lpp->cst_next : 0);
1926         }
1927 #endif /* FIRM_STATISTICS */
1928         if (need_heur) {
1929 #ifdef FIRM_STATISTICS
1930                 if (be_stat_ev_is_active()) {
1931                         be_stat_ev("time", -1);
1932                         be_stat_ev_dbl("opt", 0.0);
1933                 }
1934 #endif /* FIRM_STATISTICS */
1935                 list_sched_single_block(env->birg, block, env->be_opts);
1936         }
1937         else {
1938 #ifdef FIRM_STATISTICS
1939                 if (be_stat_ev_is_active()) {
1940                         if (lpp) {
1941                                 double opt = lpp->sol_state == lpp_optimal ? 100.0 : 100.0 * lpp->best_bound / lpp->objval;
1942                                 be_stat_ev_dbl("time", lpp->sol_time);
1943                                 be_stat_ev_dbl("opt", opt);
1944                         }
1945                         else {
1946                                 be_stat_ev_dbl("time", 0.0);
1947                                 be_stat_ev_dbl("opt", 100.0);
1948                         }
1949                 }
1950 #endif /* FIRM_STATISTICS */
1951                 apply_solution(env, lpp, block);
1952         }
1953
1954         if (lpp)
1955                 free_lpp(lpp);
1956
1957         /* notify backend */
1958         be_ilp_sched_finish_block_ilp_schedule(env->sel, block, env->block_env);
1959 }
1960
1961 /**
1962  * Perform ILP scheduling on the given irg.
1963  */
1964 void be_ilp_sched(const be_irg_t *birg, be_options_t *be_opts) {
1965         be_ilpsched_env_t          env;
1966         const char                 *name = "be ilp scheduling";
1967         arch_isa_t                 *isa  = birg->main_env->arch_env->isa;
1968         const ilp_sched_selector_t *sel  = isa->impl->get_ilp_sched_selector(isa);
1969
1970         FIRM_DBG_REGISTER(env.dbg, "firm.be.sched.ilp");
1971
1972 #ifdef FIRM_STATISTICS
1973         if (be_stat_ev_is_active()) {
1974                 be_stat_tags[STAT_TAG_CLS] = "ilpsched";
1975                 be_stat_ev_push(be_stat_tags, STAT_TAG_LAST, be_stat_file);
1976         }
1977 #endif /* FIRM_STATISTICS */
1978
1979 //      firm_dbg_set_mask(env.dbg, 1);
1980
1981         env.irg_env    = be_ilp_sched_init_irg_ilp_schedule(sel, birg->irg);
1982         env.sel        = sel;
1983         env.irg        = birg->irg;
1984         env.height     = heights_new(birg->irg);
1985         env.main_env   = birg->main_env;
1986         env.arch_env   = birg->main_env->arch_env;
1987         env.cpu        = arch_isa_get_machine(birg->main_env->arch_env->isa);
1988         env.opts       = &ilp_opts;
1989         env.birg       = birg;
1990         env.be_opts    = be_opts;
1991         phase_init(&env.ph, name, env.irg, PHASE_DEFAULT_GROWTH, init_ilpsched_irn);
1992
1993         /* assign a unique per block number to all interesting nodes */
1994         irg_walk_in_or_dep_graph(env.irg, NULL, build_block_idx, &env);
1995
1996         /*
1997                 The block indices are completely build after the walk,
1998                 now we can allocate the bitsets (size depends on block indices)
1999                 for all nodes.
2000         */
2001         phase_reinit_irn_data(&env.ph);
2002
2003         /* Collect all root nodes (having no user in their block) and calculate ASAP. */
2004         irg_walk_in_or_dep_blkwise_graph(env.irg, collect_alap_root_nodes, calculate_irn_asap, &env);
2005
2006         /* Calculate ALAP of all irns */
2007         irg_block_walk_graph(env.irg, NULL, calculate_block_alap, &env);
2008
2009         /* We refine the {ASAP(n), ALAP(n)} interval and fix the time steps for Projs and Keeps */
2010         irg_walk_in_or_dep_blkwise_graph(env.irg, NULL, refine_asap_alap_times, &env);
2011
2012         /* perform ILP scheduling */
2013         irg_block_walk_graph(env.irg, NULL, create_ilp, &env);
2014
2015         DEBUG_ONLY(
2016                 if (firm_dbg_get_mask(env.dbg)) {
2017                         phase_stat_t stat;
2018                         phase_stat_t *stat_ptr = phase_stat(&env.ph, &stat);
2019
2020                         fprintf(stderr, "Phase used: %u bytes\n", stat_ptr->overall_bytes);
2021                 }
2022         );
2023
2024         /* free data allocated dynamically */
2025         irg_block_walk_graph(env.irg, NULL, clear_unwanted_data, &env);
2026
2027         /* free all allocated object */
2028         phase_free(&env.ph);
2029         heights_free(env.height);
2030
2031         /* notify backend */
2032         be_ilp_sched_finish_irg_ilp_schedule(sel, birg->irg, env.irg_env);
2033
2034 #ifdef FIRM_STATISTICS
2035         if (be_stat_ev_is_active()) {
2036                 be_stat_ev_pop();
2037         }
2038 #endif /* FIRM_STATISTICS */
2039 }
2040
2041 /**
2042  * Register ILP scheduler options.
2043  */
2044 void be_init_ilpsched(void)
2045 {
2046         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2047         lc_opt_entry_t *sched_grp = lc_opt_get_grp(be_grp, "ilpsched");
2048
2049         lc_opt_add_table(sched_grp, ilpsched_option_table);
2050 }
2051
2052 #else /* WITH_ILP */
2053
2054 static INLINE void some_picky_compiler_do_not_allow_empty_files(void)
2055 {}
2056
2057 #endif /* WITH_ILP */