1 /*
2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
3 *
4 * This file is part of libFirm.
5 *
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
10 *
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
14 *
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
17 * PURPOSE.
18 */
19
20 /**
21 * @file
22 * @brief Preference Guided Register Assignment
23 * @author Matthias Braun
24 * @date 14.2.2009
25 *
26 * The idea is to allocate registers in 2 passes:
27 * 1. A first pass to determine "preferred" registers for live-ranges. This
28 * calculates for each register and each live-range a value indicating
29 * the usefulness. (You can roughly think of the value as the negative
30 * costs needed for copies when the value is in the specific registers...)
31 *
32 * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
33 * registers with high preferences. When register constraints are not met,
34 * add copies and split live-ranges.
35 *
36 * TODO:
37 * - make use of free registers in the permute_values code
38 */
39 #include "config.h"
40
41 #include <float.h>
42 #include <stdbool.h>
43 #include <math.h>
44 #include "lpp.h"
45
46 #include "error.h"
47 #include "execfreq.h"
48 #include "ircons.h"
49 #include "irdom.h"
50 #include "iredges_t.h"
51 #include "irgraph_t.h"
52 #include "irgwalk.h"
53 #include "irnode_t.h"
54 #include "irprintf.h"
55 #include "irdump.h"
56 #include "irtools.h"
57 #include "util.h"
58 #include "obst.h"
59 #include "raw_bitset.h"
60 #include "unionfind.h"
61 #include "pdeq.h"
62 #include "hungarian.h"
63
64 #include "beabi.h"
65 #include "bechordal_t.h"
66 #include "be.h"
67 #include "beirg.h"
68 #include "belive_t.h"
69 #include "bemodule.h"
70 #include "benode.h"
71 #include "bera.h"
72 #include "besched.h"
73 #include "bespill.h"
74 #include "bespillutil.h"
75 #include "beverify.h"
76 #include "beutil.h"
77 #include "bestack.h"
78
79 #define USE_FACTOR 1.0f
80 #define DEF_FACTOR 1.0f
81 #define NEIGHBOR_FACTOR 0.2f
82 #define AFF_SHOULD_BE_SAME 0.5f
83 #define AFF_PHI 1.0f
84 #define SPLIT_DELTA 1.0f
85 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
86
87 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
88
89 static struct obstack obst;
90 static ir_graph *irg;
91 static const arch_register_class_t *cls;
92 static be_lv_t *lv;
93 static unsigned n_regs;
94 static unsigned *normal_regs;
95 static int *congruence_classes;
96 static ir_node **block_order;
97 static size_t n_block_order;
98
99 /** currently active assignments (while processing a basic block)
100 * maps registers to values(their current copies) */
101 static ir_node **assignments;
102
103 /**
104 * allocation information: last_uses, register preferences
105 * the information is per firm-node.
106 */
107 struct allocation_info_t {
108 unsigned last_uses[2]; /**< bitset indicating last uses (input pos) */
109 ir_node *current_value; /**< copy of the value that should be used */
110 ir_node *original_value; /**< for copies point to original value */
111 float prefs[]; /**< register preferences */
112 };
113 typedef struct allocation_info_t allocation_info_t;
114
115 /** helper datastructure used when sorting register preferences */
116 struct reg_pref_t {
117 unsigned num;
118 float pref;
119 };
120 typedef struct reg_pref_t reg_pref_t;
121
122 /** per basic-block information */
123 struct block_info_t {
124 bool processed; /**< indicate whether block is processed */
125 ir_node *assignments[]; /**< register assignments at end of block */
126 };
127 typedef struct block_info_t block_info_t;
128
129 /**
130 * Get the allocation info for a node.
131 * The info is allocated on the first visit of a node.
132 */
get_allocation_info(ir_node * node)133 static allocation_info_t *get_allocation_info(ir_node *node)
134 {
135 allocation_info_t *info = (allocation_info_t*)get_irn_link(node);
136 if (info == NULL) {
137 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
138 info->current_value = node;
139 info->original_value = node;
140 set_irn_link(node, info);
141 }
142
143 return info;
144 }
145
try_get_allocation_info(const ir_node * node)146 static allocation_info_t *try_get_allocation_info(const ir_node *node)
147 {
148 return (allocation_info_t*) get_irn_link(node);
149 }
150
151 /**
152 * Get allocation information for a basic block
153 */
get_block_info(ir_node * block)154 static block_info_t *get_block_info(ir_node *block)
155 {
156 block_info_t *info = (block_info_t*)get_irn_link(block);
157
158 assert(is_Block(block));
159 if (info == NULL) {
160 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
161 set_irn_link(block, info);
162 }
163
164 return info;
165 }
166
167 /**
168 * Link the allocation info of a node to a copy.
169 * Afterwards, both nodes uses the same allocation info.
170 * Copy must not have an allocation info assigned yet.
171 *
172 * @param copy the node that gets the allocation info assigned
173 * @param value the original node
174 */
mark_as_copy_of(ir_node * copy,ir_node * value)175 static void mark_as_copy_of(ir_node *copy, ir_node *value)
176 {
177 allocation_info_t *info = get_allocation_info(value);
178 allocation_info_t *copy_info = get_allocation_info(copy);
179
180 /* find original value */
181 ir_node *original = info->original_value;
182 if (original != value) {
183 info = get_allocation_info(original);
184 }
185
186 assert(info->original_value == original);
187 info->current_value = copy;
188
189 /* the copy should not be linked to something else yet */
190 assert(copy_info->original_value == copy);
191 copy_info->original_value = original;
192
193 /* copy over allocation preferences */
194 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
195 }
196
197 /**
198 * Calculate the penalties for every register on a node and its live neighbors.
199 *
200 * @param live_nodes the set of live nodes at the current position, may be NULL
201 * @param penalty the penalty to subtract from
202 * @param limited a raw bitset containing the limited set for the node
203 * @param node the node
204 */
give_penalties_for_limits(const ir_nodeset_t * live_nodes,float penalty,const unsigned * limited,ir_node * node)205 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
206 float penalty, const unsigned* limited,
207 ir_node *node)
208 {
209 allocation_info_t *info = get_allocation_info(node);
210
211 /* give penalty for all forbidden regs */
212 for (unsigned r = 0; r < n_regs; ++r) {
213 if (rbitset_is_set(limited, r))
214 continue;
215
216 info->prefs[r] -= penalty;
217 }
218
219 /* all other live values should get a penalty for allowed regs */
220 if (live_nodes == NULL)
221 return;
222
223 penalty *= NEIGHBOR_FACTOR;
224 size_t n_allowed = rbitset_popcount(limited, n_regs);
225 if (n_allowed > 1) {
226 /* only create a very weak penalty if multiple regs are allowed */
227 penalty = (penalty * 0.8f) / n_allowed;
228 }
229 foreach_ir_nodeset(live_nodes, neighbor, iter) {
230 allocation_info_t *neighbor_info;
231
232 /* TODO: if op is used on multiple inputs we might not do a
233 * continue here */
234 if (neighbor == node)
235 continue;
236
237 neighbor_info = get_allocation_info(neighbor);
238 for (unsigned r = 0; r < n_regs; ++r) {
239 if (!rbitset_is_set(limited, r))
240 continue;
241
242 neighbor_info->prefs[r] -= penalty;
243 }
244 }
245 }
246
247 /**
248 * Calculate the preferences of a definition for the current register class.
249 * If the definition uses a limited set of registers, reduce the preferences
250 * for the limited register on the node and its neighbors.
251 *
252 * @param live_nodes the set of live nodes at the current node
253 * @param weight the weight
254 * @param node the current node
255 */
check_defs(const ir_nodeset_t * live_nodes,float weight,ir_node * node)256 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
257 ir_node *node)
258 {
259 const arch_register_req_t *req = arch_get_irn_register_req(node);
260 if (req->type & arch_register_req_type_limited) {
261 const unsigned *limited = req->limited;
262 float penalty = weight * DEF_FACTOR;
263 give_penalties_for_limits(live_nodes, penalty, limited, node);
264 }
265
266 if (req->type & arch_register_req_type_should_be_same) {
267 ir_node *insn = skip_Proj(node);
268 allocation_info_t *info = get_allocation_info(node);
269 int arity = get_irn_arity(insn);
270
271 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
272 for (int i = 0; i < arity; ++i) {
273 if (!rbitset_is_set(&req->other_same, i))
274 continue;
275
276 ir_node *op = get_irn_n(insn, i);
277
278 /* if we the value at the should_be_same input doesn't die at the
279 * node, then it is no use to propagate the constraints (since a
280 * copy will emerge anyway) */
281 if (ir_nodeset_contains(live_nodes, op))
282 continue;
283
284 allocation_info_t *op_info = get_allocation_info(op);
285 for (unsigned r = 0; r < n_regs; ++r) {
286 op_info->prefs[r] += info->prefs[r] * factor;
287 }
288 }
289 }
290 }
291
292 /**
293 * Walker: Runs an a block calculates the preferences for any
294 * node and every register from the considered register class.
295 */
analyze_block(ir_node * block,void * data)296 static void analyze_block(ir_node *block, void *data)
297 {
298 float weight = (float)get_block_execfreq(block);
299 ir_nodeset_t live_nodes;
300 (void) data;
301
302 ir_nodeset_init(&live_nodes);
303 be_liveness_end_of_block(lv, cls, block, &live_nodes);
304
305 sched_foreach_reverse(block, node) {
306 if (is_Phi(node))
307 break;
308
309 be_foreach_definition(node, cls, value,
310 check_defs(&live_nodes, weight, value);
311 );
312
313 /* mark last uses */
314 int arity = get_irn_arity(node);
315
316 /* the allocation info node currently only uses 1 unsigned value
317 to mark last used inputs. So we will fail for a node with more than
318 32 inputs. */
319 allocation_info_t *info = get_allocation_info(node);
320 if (arity >= (int) sizeof(info->last_uses) * 8) {
321 panic("Node with more than %d inputs not supported yet",
322 (int) sizeof(info->last_uses) * 8);
323 }
324
325 for (int i = 0; i < arity; ++i) {
326 ir_node *op = get_irn_n(node, i);
327 const arch_register_req_t *req = arch_get_irn_register_req(op);
328 if (req->cls != cls)
329 continue;
330
331 /* last usage of a value? */
332 if (!ir_nodeset_contains(&live_nodes, op)) {
333 rbitset_set(info->last_uses, i);
334 }
335 }
336
337 be_liveness_transfer(cls, node, &live_nodes);
338
339 /* update weights based on usage constraints */
340 for (int i = 0; i < arity; ++i) {
341 ir_node *op = get_irn_n(node, i);
342 if (!arch_irn_consider_in_reg_alloc(cls, op))
343 continue;
344
345 const arch_register_req_t *req
346 = arch_get_irn_register_req_in(node, i);
347 if (!(req->type & arch_register_req_type_limited))
348 continue;
349
350 const unsigned *limited = req->limited;
351 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
352 limited, op);
353 }
354 }
355
356 ir_nodeset_destroy(&live_nodes);
357 }
358
congruence_def(ir_nodeset_t * live_nodes,const ir_node * node)359 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
360 {
361 const arch_register_req_t *req = arch_get_irn_register_req(node);
362
363 /* should be same constraint? */
364 if (req->type & arch_register_req_type_should_be_same) {
365 const ir_node *insn = skip_Proj_const(node);
366 int arity = get_irn_arity(insn);
367 unsigned node_idx = get_irn_idx(node);
368 node_idx = uf_find(congruence_classes, node_idx);
369
370 for (int i = 0; i < arity; ++i) {
371 if (!rbitset_is_set(&req->other_same, i))
372 continue;
373
374 ir_node *op = get_irn_n(insn, i);
375 int op_idx = get_irn_idx(op);
376 op_idx = uf_find(congruence_classes, op_idx);
377
378 /* do we interfere with the value */
379 bool interferes = false;
380 foreach_ir_nodeset(live_nodes, live, iter) {
381 int lv_idx = get_irn_idx(live);
382 lv_idx = uf_find(congruence_classes, lv_idx);
383 if (lv_idx == op_idx) {
384 interferes = true;
385 break;
386 }
387 }
388 /* don't put in same affinity class if we interfere */
389 if (interferes)
390 continue;
391
392 uf_union(congruence_classes, node_idx, op_idx);
393 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
394 node, op));
395 /* one should_be_same is enough... */
396 break;
397 }
398 }
399 }
400
create_congruence_class(ir_node * block,void * data)401 static void create_congruence_class(ir_node *block, void *data)
402 {
403 ir_nodeset_t live_nodes;
404
405 (void) data;
406 ir_nodeset_init(&live_nodes);
407 be_liveness_end_of_block(lv, cls, block, &live_nodes);
408
409 /* check should be same constraints */
410 ir_node *last_phi = NULL;
411 sched_foreach_reverse(block, node) {
412 if (is_Phi(node)) {
413 last_phi = node;
414 break;
415 }
416
417 be_foreach_definition(node, cls, value,
418 congruence_def(&live_nodes, value);
419 );
420 be_liveness_transfer(cls, node, &live_nodes);
421 }
422 if (!last_phi) {
423 ir_nodeset_destroy(&live_nodes);
424 return;
425 }
426
427 /* check phi congruence classes */
428 sched_foreach_reverse_from(last_phi, phi) {
429 assert(is_Phi(phi));
430
431 if (!arch_irn_consider_in_reg_alloc(cls, phi))
432 continue;
433
434 int node_idx = get_irn_idx(phi);
435 node_idx = uf_find(congruence_classes, node_idx);
436
437 int arity = get_irn_arity(phi);
438 for (int i = 0; i < arity; ++i) {
439 ir_node *op = get_Phi_pred(phi, i);
440 int op_idx = get_irn_idx(op);
441 op_idx = uf_find(congruence_classes, op_idx);
442
443 /* do we interfere with the value */
444 bool interferes = false;
445 foreach_ir_nodeset(&live_nodes, live, iter) {
446 int lv_idx = get_irn_idx(live);
447 lv_idx = uf_find(congruence_classes, lv_idx);
448 if (lv_idx == op_idx) {
449 interferes = true;
450 break;
451 }
452 }
453 /* don't put in same affinity class if we interfere */
454 if (interferes)
455 continue;
456 /* any other phi has the same input? */
457 sched_foreach(block, phi) {
458 ir_node *oop;
459 int oop_idx;
460 if (!is_Phi(phi))
461 break;
462 if (!arch_irn_consider_in_reg_alloc(cls, phi))
463 continue;
464 oop = get_Phi_pred(phi, i);
465 if (oop == op)
466 continue;
467 oop_idx = get_irn_idx(oop);
468 oop_idx = uf_find(congruence_classes, oop_idx);
469 if (oop_idx == op_idx) {
470 interferes = true;
471 break;
472 }
473 }
474 if (interferes)
475 continue;
476
477 /* merge the 2 congruence classes and sum up their preferences */
478 int old_node_idx = node_idx;
479 node_idx = uf_union(congruence_classes, node_idx, op_idx);
480 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
481 phi, op));
482
483 old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
484 allocation_info_t *head_info
485 = get_allocation_info(get_idx_irn(irg, node_idx));
486 allocation_info_t *other_info
487 = get_allocation_info(get_idx_irn(irg, old_node_idx));
488 for (unsigned r = 0; r < n_regs; ++r) {
489 head_info->prefs[r] += other_info->prefs[r];
490 }
491 }
492 }
493 ir_nodeset_destroy(&live_nodes);
494 }
495
set_congruence_prefs(ir_node * node,void * data)496 static void set_congruence_prefs(ir_node *node, void *data)
497 {
498 (void) data;
499 unsigned node_idx = get_irn_idx(node);
500 unsigned node_set = uf_find(congruence_classes, node_idx);
501
502 /* head of congruence class or not in any class */
503 if (node_set == node_idx)
504 return;
505
506 if (!arch_irn_consider_in_reg_alloc(cls, node))
507 return;
508
509 ir_node *head = get_idx_irn(irg, node_set);
510 allocation_info_t *head_info = get_allocation_info(head);
511 allocation_info_t *info = get_allocation_info(node);
512
513 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
514 }
515
combine_congruence_classes(void)516 static void combine_congruence_classes(void)
517 {
518 size_t n = get_irg_last_idx(irg);
519 congruence_classes = XMALLOCN(int, n);
520 uf_init(congruence_classes, n);
521
522 /* create congruence classes */
523 irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
524 /* merge preferences */
525 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
526 free(congruence_classes);
527 }
528
529
530
531 /**
532 * Assign register reg to the given node.
533 *
534 * @param node the node
535 * @param reg the register
536 */
use_reg(ir_node * node,const arch_register_t * reg,unsigned width)537 static void use_reg(ir_node *node, const arch_register_t *reg, unsigned width)
538 {
539 unsigned r = reg->index;
540 for (unsigned r0 = r; r0 < r + width; ++r0)
541 assignments[r0] = node;
542 arch_set_irn_register(node, reg);
543 }
544
free_reg_of_value(ir_node * node)545 static void free_reg_of_value(ir_node *node)
546 {
547 if (!arch_irn_consider_in_reg_alloc(cls, node))
548 return;
549
550 const arch_register_t *reg = arch_get_irn_register(node);
551 const arch_register_req_t *req = arch_get_irn_register_req(node);
552 unsigned r = reg->index;
553 /* assignment->value may be NULL if a value is used at 2 inputs
554 * so it gets freed twice. */
555 for (unsigned r0 = r; r0 < r + req->width; ++r0) {
556 assert(assignments[r0] == node || assignments[r0] == NULL);
557 assignments[r0] = NULL;
558 }
559 }
560
561 /**
562 * Compare two register preferences in decreasing order.
563 */
compare_reg_pref(const void * e1,const void * e2)564 static int compare_reg_pref(const void *e1, const void *e2)
565 {
566 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
567 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
568 if (rp1->pref < rp2->pref)
569 return 1;
570 if (rp1->pref > rp2->pref)
571 return -1;
572 return 0;
573 }
574
fill_sort_candidates(reg_pref_t * regprefs,const allocation_info_t * info)575 static void fill_sort_candidates(reg_pref_t *regprefs,
576 const allocation_info_t *info)
577 {
578 for (unsigned r = 0; r < n_regs; ++r) {
579 float pref = info->prefs[r];
580 regprefs[r].num = r;
581 regprefs[r].pref = pref;
582 }
583 /* TODO: use a stable sort here to avoid unnecessary register jumping */
584 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
585 }
586
try_optimistic_split(ir_node * to_split,ir_node * before,float pref,float pref_delta,unsigned * forbidden_regs,int recursion)587 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
588 float pref, float pref_delta,
589 unsigned *forbidden_regs, int recursion)
590 {
591 (void) pref;
592 unsigned r = 0;
593 allocation_info_t *info = get_allocation_info(to_split);
594 float delta = 0;
595
596 /* stupid hack: don't optimisticallt split don't spill nodes...
597 * (so we don't split away the values produced because of
598 * must_be_different constraints) */
599 ir_node *original_insn = skip_Proj(info->original_value);
600 if (arch_get_irn_flags(original_insn) & arch_irn_flags_dont_spill)
601 return false;
602
603 const arch_register_t *from_reg = arch_get_irn_register(to_split);
604 unsigned from_r = from_reg->index;
605 ir_node *block = get_nodes_block(before);
606 float split_threshold = (float)get_block_execfreq(block) * SPLIT_DELTA;
607
608 if (pref_delta < split_threshold*0.5)
609 return false;
610
611 /* find the best free position where we could move to */
612 reg_pref_t *prefs = ALLOCAN(reg_pref_t, n_regs);
613 fill_sort_candidates(prefs, info);
614 unsigned i;
615 for (i = 0; i < n_regs; ++i) {
616 /* we need a normal register which is not an output register
617 an different from the current register of to_split */
618 r = prefs[i].num;
619 if (!rbitset_is_set(normal_regs, r))
620 continue;
621 if (rbitset_is_set(forbidden_regs, r))
622 continue;
623 if (r == from_r)
624 continue;
625
626 /* is the split worth it? */
627 delta = pref_delta + prefs[i].pref;
628 if (delta < split_threshold) {
629 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
630 to_split, recursion, delta));
631 return false;
632 }
633
634 /* if the register is free then we can do the split */
635 if (assignments[r] == NULL)
636 break;
637
638 /* otherwise we might try recursively calling optimistic_split */
639 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
640 continue;
641
642 float apref = prefs[i].pref;
643 float apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
644 apref_delta += pref_delta - split_threshold;
645
646 /* our source register isn't a useful destination for recursive
647 splits */
648 bool old_source_state = rbitset_is_set(forbidden_regs, from_r);
649 rbitset_set(forbidden_regs, from_r);
650 /* try recursive split */
651 bool res = try_optimistic_split(assignments[r], before, apref,
652 apref_delta, forbidden_regs, recursion+1);
653 /* restore our destination */
654 if (old_source_state) {
655 rbitset_set(forbidden_regs, from_r);
656 } else {
657 rbitset_clear(forbidden_regs, from_r);
658 }
659
660 if (res)
661 break;
662 }
663 if (i >= n_regs)
664 return false;
665
666 const arch_register_t *reg = arch_register_for_index(cls, r);
667 ir_node *copy = be_new_Copy(block, to_split);
668 unsigned width = 1;
669 mark_as_copy_of(copy, to_split);
670 /* hacky, but correct here */
671 if (assignments[from_reg->index] == to_split)
672 free_reg_of_value(to_split);
673 use_reg(copy, reg, width);
674 sched_add_before(before, copy);
675
676 DB((dbg, LEVEL_3,
677 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
678 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
679 return true;
680 }
681
682 /**
683 * Determine and assign a register for node @p node
684 */
assign_reg(const ir_node * block,ir_node * node,unsigned * forbidden_regs)685 static void assign_reg(const ir_node *block, ir_node *node,
686 unsigned *forbidden_regs)
687 {
688 assert(!is_Phi(node));
689 /* preassigned register? */
690 const arch_register_t *final_reg = arch_get_irn_register(node);
691 const arch_register_req_t *req = arch_get_irn_register_req(node);
692 unsigned width = req->width;
693 if (final_reg != NULL) {
694 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, final_reg->name));
695 use_reg(node, final_reg, width);
696 return;
697 }
698
699 /* ignore reqs must be preassigned */
700 assert (! (req->type & arch_register_req_type_ignore));
701
702 /* give should_be_same boni */
703 allocation_info_t *info = get_allocation_info(node);
704 ir_node *in_node = skip_Proj(node);
705 if (req->type & arch_register_req_type_should_be_same) {
706 float weight = (float)get_block_execfreq(block);
707 int arity = get_irn_arity(in_node);
708
709 assert(arity <= (int) sizeof(req->other_same) * 8);
710 for (int i = 0; i < arity; ++i) {
711 if (!rbitset_is_set(&req->other_same, i))
712 continue;
713
714 ir_node *in = get_irn_n(in_node, i);
715 const arch_register_t *reg = arch_get_irn_register(in);
716 unsigned reg_index = reg->index;
717
718 /* if the value didn't die here then we should not propagate the
719 * should_be_same info */
720 if (assignments[reg_index] == in)
721 continue;
722
723 info->prefs[reg_index] += weight * AFF_SHOULD_BE_SAME;
724 }
725 }
726
727 /* create list of register candidates and sort by their preference */
728 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
729 reg_pref_t *reg_prefs = ALLOCAN(reg_pref_t, n_regs);
730 fill_sort_candidates(reg_prefs, info);
731 for (unsigned r = 0; r < n_regs; ++r) {
732 unsigned num = reg_prefs[r].num;
733 if (!rbitset_is_set(normal_regs, num))
734 continue;
735 const arch_register_t *reg = arch_register_for_index(cls, num);
736 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[r].pref));
737 }
738 DB((dbg, LEVEL_2, "\n"));
739
740 const unsigned *allowed_regs = normal_regs;
741 if (req->type & arch_register_req_type_limited) {
742 allowed_regs = req->limited;
743 }
744
745 unsigned final_reg_index = 0;
746 unsigned r;
747 for (r = 0; r < n_regs; ++r) {
748 final_reg_index = reg_prefs[r].num;
749 if (!rbitset_is_set(allowed_regs, final_reg_index))
750 continue;
751 /* alignment constraint? */
752 if (width > 1) {
753 if ((req->type & arch_register_req_type_aligned)
754 && (final_reg_index % width) != 0)
755 continue;
756 bool fine = true;
757 for (unsigned r0 = r+1; r0 < r+width; ++r0) {
758 if (assignments[r0] != NULL)
759 fine = false;
760 }
761 /* TODO: attempt optimistic split here */
762 if (!fine)
763 continue;
764 }
765
766 if (assignments[final_reg_index] == NULL)
767 break;
768 float pref = reg_prefs[r].pref;
769 float delta = r+1 < n_regs ? pref - reg_prefs[r+1].pref : 0;
770 ir_node *before = skip_Proj(node);
771 bool res
772 = try_optimistic_split(assignments[final_reg_index], before, pref,
773 delta, forbidden_regs, 0);
774 if (res)
775 break;
776 }
777 if (r >= n_regs) {
778 /* the common reason to hit this panic is when 1 of your nodes is not
779 * register pressure faithful */
780 panic("No register left for %+F\n", node);
781 }
782
783 final_reg = arch_register_for_index(cls, final_reg_index);
784 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, final_reg->name));
785 use_reg(node, final_reg, width);
786 }
787
788 /**
789 * Add an permutation in front of a node and change the assignments
790 * due to this permutation.
791 *
792 * To understand this imagine a permutation like this:
793 *
794 * 1 -> 2
795 * 2 -> 3
796 * 3 -> 1, 5
797 * 4 -> 6
798 * 5
799 * 6
800 * 7 -> 7
801 *
802 * First we count how many destinations a single value has. At the same time
803 * we can be sure that each destination register has at most 1 source register
804 * (it can have 0 which means we don't care what value is in it).
805 * We ignore all fulfilled permuations (like 7->7)
806 * In a first pass we create as much copy instructions as possible as they
807 * are generally cheaper than exchanges. We do this by counting into how many
808 * destinations a register has to be copied (in the example it's 2 for register
809 * 3, or 1 for the registers 1,2,4 and 7).
810 * We can then create a copy into every destination register when the usecount
811 * of that register is 0 (= noone else needs the value in the register).
812 *
813 * After this step we should only have cycles left. We implement a cyclic
814 * permutation of n registers with n-1 transpositions.
815 *
816 * @param live_nodes the set of live nodes, updated due to live range split
817 * @param before the node before we add the permutation
818 * @param permutation the permutation array indices are the destination
819 * registers, the values in the array are the source
820 * registers.
821 */
permute_values(ir_nodeset_t * live_nodes,ir_node * before,unsigned * permutation)822 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
823 unsigned *permutation)
824 {
825 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
826
827 /* determine how often each source register needs to be read */
828 for (unsigned r = 0; r < n_regs; ++r) {
829 unsigned old_reg = permutation[r];
830 ir_node *value;
831
832 value = assignments[old_reg];
833 if (value == NULL) {
834 /* nothing to do here, reg is not live. Mark it as fixpoint
835 * so we ignore it in the next steps */
836 permutation[r] = r;
837 continue;
838 }
839
840 ++n_used[old_reg];
841 }
842
843 ir_node *block = get_nodes_block(before);
844
845 /* step1: create copies where immediately possible */
846 for (unsigned r = 0; r < n_regs; /* empty */) {
847 unsigned old_r = permutation[r];
848
849 /* - no need to do anything for fixed points.
850 - we can't copy if the value in the dest reg is still needed */
851 if (old_r == r || n_used[r] > 0) {
852 ++r;
853 continue;
854 }
855
856 /* create a copy */
857 ir_node *src = assignments[old_r];
858 ir_node *copy = be_new_Copy(block, src);
859 sched_add_before(before, copy);
860 const arch_register_t *reg = arch_register_for_index(cls, r);
861 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
862 copy, src, before, reg->name));
863 mark_as_copy_of(copy, src);
864 unsigned width = 1; /* TODO */
865 use_reg(copy, reg, width);
866
867 if (live_nodes != NULL) {
868 ir_nodeset_insert(live_nodes, copy);
869 }
870
871 /* old register has 1 user less, permutation is resolved */
872 assert(arch_get_irn_register(src)->index == old_r);
873 permutation[r] = r;
874
875 assert(n_used[old_r] > 0);
876 --n_used[old_r];
877 if (n_used[old_r] == 0) {
878 if (live_nodes != NULL) {
879 ir_nodeset_remove(live_nodes, src);
880 }
881 free_reg_of_value(src);
882 }
883
884 /* advance or jump back (if this copy enabled another copy) */
885 if (old_r < r && n_used[old_r] == 0) {
886 r = old_r;
887 } else {
888 ++r;
889 }
890 }
891
892 /* at this point we only have "cycles" left which we have to resolve with
893 * perm instructions
894 * TODO: if we have free registers left, then we should really use copy
895 * instructions for any cycle longer than 2 registers...
896 * (this is probably architecture dependent, there might be archs where
897 * copies are preferable even for 2-cycles) */
898
899 /* create perms with the rest */
900 for (unsigned r = 0; r < n_regs; /* empty */) {
901 unsigned old_r = permutation[r];
902
903 if (old_r == r) {
904 ++r;
905 continue;
906 }
907
908 /* we shouldn't have copies from 1 value to multiple destinations left*/
909 assert(n_used[old_r] == 1);
910
911 /* exchange old_r and r2; after that old_r is a fixed point */
912 unsigned r2 = permutation[old_r];
913
914 ir_node *in[2] = { assignments[r2], assignments[old_r] };
915 ir_node *perm = be_new_Perm(cls, block, 2, in);
916 sched_add_before(before, perm);
917 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
918 perm, in[0], in[1], before));
919
920 unsigned width = 1; /* TODO */
921
922 ir_node *proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
923 mark_as_copy_of(proj0, in[0]);
924 const arch_register_t *reg0 = arch_register_for_index(cls, old_r);
925 use_reg(proj0, reg0, width);
926
927 ir_node *proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
928 mark_as_copy_of(proj1, in[1]);
929 const arch_register_t *reg1 = arch_register_for_index(cls, r2);
930 use_reg(proj1, reg1, width);
931
932 /* 1 value is now in the correct register */
933 permutation[old_r] = old_r;
934 /* the source of r changed to r2 */
935 permutation[r] = r2;
936
937 /* if we have reached a fixpoint update data structures */
938 if (live_nodes != NULL) {
939 ir_nodeset_remove(live_nodes, in[0]);
940 ir_nodeset_remove(live_nodes, in[1]);
941 ir_nodeset_remove(live_nodes, proj0);
942 ir_nodeset_insert(live_nodes, proj1);
943 }
944 }
945
946 #ifndef NDEBUG
947 /* now we should only have fixpoints left */
948 for (unsigned r = 0; r < n_regs; ++r) {
949 assert(permutation[r] == r);
950 }
951 #endif
952 }
953
954 /**
955 * Free regs for values last used.
956 *
957 * @param live_nodes set of live nodes, will be updated
958 * @param node the node to consider
959 */
free_last_uses(ir_nodeset_t * live_nodes,ir_node * node)960 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
961 {
962 allocation_info_t *info = get_allocation_info(node);
963 const unsigned *last_uses = info->last_uses;
964 int arity = get_irn_arity(node);
965
966 for (int i = 0; i < arity; ++i) {
967 /* check if one operand is the last use */
968 if (!rbitset_is_set(last_uses, i))
969 continue;
970
971 ir_node *op = get_irn_n(node, i);
972 free_reg_of_value(op);
973 ir_nodeset_remove(live_nodes, op);
974 }
975 }
976
977 /**
978 * change inputs of a node to the current value (copies/perms)
979 */
rewire_inputs(ir_node * node)980 static void rewire_inputs(ir_node *node)
981 {
982 int arity = get_irn_arity(node);
983 for (int i = 0; i < arity; ++i) {
984 ir_node *op = get_irn_n(node, i);
985 allocation_info_t *info = try_get_allocation_info(op);
986
987 if (info == NULL)
988 continue;
989
990 info = get_allocation_info(info->original_value);
991 if (info->current_value != op) {
992 set_irn_n(node, i, info->current_value);
993 }
994 }
995 }
996
997 /**
998 * Create a bitset of registers occupied with value living through an
999 * instruction
1000 */
determine_live_through_regs(unsigned * bitset,ir_node * node)1001 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1002 {
1003 const allocation_info_t *info = get_allocation_info(node);
1004
1005 /* mark all used registers as potentially live-through */
1006 for (unsigned r = 0; r < n_regs; ++r) {
1007 if (assignments[r] == NULL)
1008 continue;
1009 if (!rbitset_is_set(normal_regs, r))
1010 continue;
1011
1012 rbitset_set(bitset, r);
1013 }
1014
1015 /* remove registers of value dying at the instruction */
1016 int arity = get_irn_arity(node);
1017 for (int i = 0; i < arity; ++i) {
1018 if (!rbitset_is_set(info->last_uses, i))
1019 continue;
1020
1021 ir_node *op = get_irn_n(node, i);
1022 const arch_register_t *reg = arch_get_irn_register(op);
1023 rbitset_clear(bitset, reg->index);
1024 }
1025 }
1026
solve_lpp(ir_nodeset_t * live_nodes,ir_node * node,unsigned * forbidden_regs,unsigned * live_through_regs)1027 static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
1028 unsigned *forbidden_regs, unsigned *live_through_regs)
1029 {
1030 unsigned *forbidden_edges = rbitset_malloc(n_regs * n_regs);
1031 int *lpp_vars = XMALLOCNZ(int, n_regs*n_regs);
1032
1033 lpp_t *lpp = lpp_new("prefalloc", lpp_minimize);
1034 //lpp_set_time_limit(lpp, 20);
1035 lpp_set_log(lpp, stdout);
1036
1037 /** mark some edges as forbidden */
1038 int arity = get_irn_arity(node);
1039 for (int i = 0; i < arity; ++i) {
1040 ir_node *op = get_irn_n(node, i);
1041 if (!arch_irn_consider_in_reg_alloc(cls, op))
1042 continue;
1043
1044 const arch_register_req_t *req = arch_get_irn_register_req_in(node, i);
1045 if (!(req->type & arch_register_req_type_limited))
1046 continue;
1047
1048 const unsigned *limited = req->limited;
1049 const arch_register_t *reg = arch_get_irn_register(op);
1050 unsigned current_reg = reg->index;
1051 for (unsigned r = 0; r < n_regs; ++r) {
1052 if (rbitset_is_set(limited, r))
1053 continue;
1054
1055 rbitset_set(forbidden_edges, current_reg*n_regs + r);
1056 }
1057 }
1058
1059 /* add all combinations, except for not allowed ones */
1060 for (unsigned l = 0; l < n_regs; ++l) {
1061 if (!rbitset_is_set(normal_regs, l)) {
1062 char name[15];
1063 snprintf(name, sizeof(name), "%u_to_%u", l, l);
1064 lpp_vars[l*n_regs+l] = lpp_add_var(lpp, name, lpp_binary, 1);
1065 continue;
1066 }
1067
1068 for (unsigned r = 0; r < n_regs; ++r) {
1069 if (!rbitset_is_set(normal_regs, r))
1070 continue;
1071 if (rbitset_is_set(forbidden_edges, l*n_regs + r))
1072 continue;
1073 /* livethrough values may not use constrained output registers */
1074 if (rbitset_is_set(live_through_regs, l)
1075 && rbitset_is_set(forbidden_regs, r))
1076 continue;
1077
1078 char name[15];
1079 snprintf(name, sizeof(name), "%u_to_%u", l, r);
1080
1081 double costs = l==r ? 9 : 8;
1082 lpp_vars[l*n_regs+r]
1083 = lpp_add_var(lpp, name, lpp_binary, costs);
1084 assert(lpp_vars[l*n_regs+r] > 0);
1085 }
1086 }
1087 /* add constraints */
1088 for (unsigned l = 0; l < n_regs; ++l) {
1089 /* only 1 destination per register */
1090 int constraint = -1;
1091 for (unsigned r = 0; r < n_regs; ++r) {
1092 int var = lpp_vars[l*n_regs+r];
1093 if (var == 0)
1094 continue;
1095 if (constraint < 0) {
1096 char name[64];
1097 snprintf(name, sizeof(name), "%u_to_dest", l);
1098 constraint = lpp_add_cst(lpp, name, lpp_equal, 1);
1099 }
1100 lpp_set_factor_fast(lpp, constraint, var, 1);
1101 }
1102 /* each destination used by at most 1 value */
1103 constraint = -1;
1104 for (unsigned r = 0; r < n_regs; ++r) {
1105 int var = lpp_vars[r*n_regs+l];
1106 if (var == 0)
1107 continue;
1108 if (constraint < 0) {
1109 char name[64];
1110 snprintf(name, sizeof(name), "one_to_%u", l);
1111 constraint = lpp_add_cst(lpp, name, lpp_less_equal, 1);
1112 }
1113 lpp_set_factor_fast(lpp, constraint, var, 1);
1114 }
1115 }
1116
1117 lpp_dump_plain(lpp, fopen("lppdump.txt", "w"));
1118
1119 /* solve lpp */
1120 lpp_solve(lpp, be_options.ilp_server, be_options.ilp_solver);
1121 if (!lpp_is_sol_valid(lpp))
1122 panic("ilp solution not valid!");
1123
1124 unsigned *assignment = ALLOCAN(unsigned, n_regs);
1125 for (unsigned l = 0; l < n_regs; ++l) {
1126 unsigned dest_reg = (unsigned)-1;
1127 for (unsigned r = 0; r < n_regs; ++r) {
1128 int var = lpp_vars[l*n_regs+r];
1129 if (var == 0)
1130 continue;
1131 double val = lpp_get_var_sol(lpp, var);
1132 if (val == 1) {
1133 assert(dest_reg == (unsigned)-1);
1134 dest_reg = r;
1135 }
1136 }
1137 assert(dest_reg != (unsigned)-1);
1138 assignment[dest_reg] = l;
1139 }
1140
1141 fprintf(stderr, "Assignment: ");
1142 for (unsigned l = 0; l < n_regs; ++l) {
1143 fprintf(stderr, "%u ", assignment[l]);
1144 }
1145 fprintf(stderr, "\n");
1146 fflush(stdout);
1147 permute_values(live_nodes, node, assignment);
1148 lpp_free(lpp);
1149 }
1150
is_aligned(unsigned num,unsigned alignment)1151 static bool is_aligned(unsigned num, unsigned alignment)
1152 {
1153 unsigned mask = alignment-1;
1154 assert(is_po2(alignment));
1155 return (num&mask) == 0;
1156 }
1157
1158 /**
1159 * Enforce constraints at a node by live range splits.
1160 *
1161 * @param live_nodes the set of live nodes, might be changed
1162 * @param node the current node
1163 */
enforce_constraints(ir_nodeset_t * live_nodes,ir_node * node,unsigned * forbidden_regs)1164 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1165 unsigned *forbidden_regs)
1166 {
1167 /* see if any use constraints are not met and whether double-width
1168 * values are involved */
1169 bool double_width = false;
1170 bool good = true;
1171 int arity = get_irn_arity(node);
1172 for (int i = 0; i < arity; ++i) {
1173 ir_node *op = get_irn_n(node, i);
1174 if (!arch_irn_consider_in_reg_alloc(cls, op))
1175 continue;
1176
1177 /* are there any limitations for the i'th operand? */
1178 const arch_register_req_t *req = arch_get_irn_register_req_in(node, i);
1179 if (req->width > 1)
1180 double_width = true;
1181 const arch_register_t *reg = arch_get_irn_register(op);
1182 unsigned reg_index = reg->index;
1183 if (req->type & arch_register_req_type_aligned) {
1184 if (!is_aligned(reg_index, req->width)) {
1185 good = false;
1186 continue;
1187 }
1188 }
1189 if (!(req->type & arch_register_req_type_limited))
1190 continue;
1191
1192 const unsigned *limited = req->limited;
1193 if (!rbitset_is_set(limited, reg_index)) {
1194 /* found an assignment outside the limited set */
1195 good = false;
1196 continue;
1197 }
1198 }
1199
1200 /* is any of the live-throughs using a constrained output register? */
1201 unsigned *live_through_regs = NULL;
1202 be_foreach_definition(node, cls, value,
1203 (void)value;
1204 if (req_->width > 1)
1205 double_width = true;
1206 if (! (req_->type & arch_register_req_type_limited))
1207 continue;
1208 if (live_through_regs == NULL) {
1209 rbitset_alloca(live_through_regs, n_regs);
1210 determine_live_through_regs(live_through_regs, node);
1211 }
1212 rbitset_or(forbidden_regs, req_->limited, n_regs);
1213 if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
1214 good = false;
1215 );
1216
1217 if (good)
1218 return;
1219
1220 /* create these arrays if we haven't yet */
1221 if (live_through_regs == NULL) {
1222 rbitset_alloca(live_through_regs, n_regs);
1223 }
1224
1225 if (double_width) {
1226 /* only the ILP variant can solve this yet */
1227 solve_lpp(live_nodes, node, forbidden_regs, live_through_regs);
1228 return;
1229 }
1230
1231 /* at this point we have to construct a bipartite matching problem to see
1232 * which values should go to which registers
1233 * Note: We're building the matrix in "reverse" - source registers are
1234 * right, destinations left because this will produce the solution
1235 * in the format required for permute_values.
1236 */
1237 hungarian_problem_t *bp
1238 = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1239
1240 /* add all combinations, then remove not allowed ones */
1241 for (unsigned l = 0; l < n_regs; ++l) {
1242 if (!rbitset_is_set(normal_regs, l)) {
1243 hungarian_add(bp, l, l, 1);
1244 continue;
1245 }
1246
1247 for (unsigned r = 0; r < n_regs; ++r) {
1248 if (!rbitset_is_set(normal_regs, r))
1249 continue;
1250 /* livethrough values may not use constrainted output registers */
1251 if (rbitset_is_set(live_through_regs, l)
1252 && rbitset_is_set(forbidden_regs, r))
1253 continue;
1254
1255 hungarian_add(bp, r, l, l == r ? 9 : 8);
1256 }
1257 }
1258
1259 for (int i = 0; i < arity; ++i) {
1260 ir_node *op = get_irn_n(node, i);
1261 if (!arch_irn_consider_in_reg_alloc(cls, op))
1262 continue;
1263
1264 const arch_register_req_t *req = arch_get_irn_register_req_in(node, i);
1265 if (!(req->type & arch_register_req_type_limited))
1266 continue;
1267
1268 const unsigned *limited = req->limited;
1269 const arch_register_t *reg = arch_get_irn_register(op);
1270 unsigned current_reg = reg->index;
1271 for (unsigned r = 0; r < n_regs; ++r) {
1272 if (rbitset_is_set(limited, r))
1273 continue;
1274 hungarian_remove(bp, r, current_reg);
1275 }
1276 }
1277
1278 //hungarian_print_cost_matrix(bp, 1);
1279 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1280
1281 unsigned *assignment = ALLOCAN(unsigned, n_regs);
1282 int res = hungarian_solve(bp, assignment, NULL, 0);
1283 assert(res == 0);
1284
1285 #if 0
1286 fprintf(stderr, "Swap result:");
1287 for (i = 0; i < (int) n_regs; ++i) {
1288 fprintf(stderr, " %d", assignment[i]);
1289 }
1290 fprintf(stderr, "\n");
1291 #endif
1292
1293 hungarian_free(bp);
1294
1295 permute_values(live_nodes, node, assignment);
1296 }
1297
1298 /** test whether a node @p n is a copy of the value of node @p of */
is_copy_of(ir_node * value,ir_node * test_value)1299 static bool is_copy_of(ir_node *value, ir_node *test_value)
1300 {
1301 if (value == test_value)
1302 return true;
1303
1304 allocation_info_t *info = get_allocation_info(value);
1305 allocation_info_t *test_info = get_allocation_info(test_value);
1306 return test_info->original_value == info->original_value;
1307 }
1308
1309 /**
1310 * find a value in the end-assignment of a basic block
1311 * @returns the index into the assignment array if found
1312 * -1 if not found
1313 */
find_value_in_block_info(block_info_t * info,ir_node * value)1314 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1315 {
1316 ir_node **end_assignments = info->assignments;
1317 for (unsigned r = 0; r < n_regs; ++r) {
1318 ir_node *a_value = end_assignments[r];
1319
1320 if (a_value == NULL)
1321 continue;
1322 if (is_copy_of(a_value, value))
1323 return (int) r;
1324 }
1325
1326 return -1;
1327 }
1328
1329 /**
1330 * Create the necessary permutations at the end of a basic block to fullfill
1331 * the register assignment for phi-nodes in the next block
1332 */
add_phi_permutations(ir_node * block,int p)1333 static void add_phi_permutations(ir_node *block, int p)
1334 {
1335 ir_node *pred = get_Block_cfgpred_block(block, p);
1336 block_info_t *pred_info = get_block_info(pred);
1337
1338 /* predecessor not processed yet? nothing to do */
1339 if (!pred_info->processed)
1340 return;
1341
1342 unsigned *permutation = ALLOCAN(unsigned, n_regs);
1343 for (unsigned r = 0; r < n_regs; ++r) {
1344 permutation[r] = r;
1345 }
1346
1347 /* check phi nodes */
1348 bool need_permutation = false;
1349 ir_node *phi = sched_first(block);
1350 for ( ; is_Phi(phi); phi = sched_next(phi)) {
1351 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1352 continue;
1353
1354 ir_node *phi_pred = get_Phi_pred(phi, p);
1355 int a = find_value_in_block_info(pred_info, phi_pred);
1356 assert(a >= 0);
1357
1358 const arch_register_t *reg = arch_get_irn_register(phi);
1359 int regn = reg->index;
1360 /* same register? nothing to do */
1361 if (regn == a)
1362 continue;
1363
1364 ir_node *op = pred_info->assignments[a];
1365 const arch_register_t *op_reg = arch_get_irn_register(op);
1366 /* virtual or joker registers are ok too */
1367 if ((op_reg->type & arch_register_type_joker)
1368 || (op_reg->type & arch_register_type_virtual))
1369 continue;
1370
1371 permutation[regn] = a;
1372 need_permutation = true;
1373 }
1374
1375 if (need_permutation) {
1376 /* permute values at end of predecessor */
1377 ir_node **old_assignments = assignments;
1378 assignments = pred_info->assignments;
1379 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1380 permutation);
1381 assignments = old_assignments;
1382 }
1383
1384 /* change phi nodes to use the copied values */
1385 phi = sched_first(block);
1386 for ( ; is_Phi(phi); phi = sched_next(phi)) {
1387 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1388 continue;
1389
1390 /* we have permuted all values into the correct registers so we can
1391 simply query which value occupies the phis register in the
1392 predecessor */
1393 int a = arch_get_irn_register(phi)->index;
1394 ir_node *op = pred_info->assignments[a];
1395 set_Phi_pred(phi, p, op);
1396 }
1397 }
1398
1399 /**
1400 * Set preferences for a phis register based on the registers used on the
1401 * phi inputs.
1402 */
adapt_phi_prefs(ir_node * phi)1403 static void adapt_phi_prefs(ir_node *phi)
1404 {
1405 ir_node *block = get_nodes_block(phi);
1406 allocation_info_t *info = get_allocation_info(phi);
1407
1408 int arity = get_irn_arity(phi);
1409 for (int i = 0; i < arity; ++i) {
1410 ir_node *op = get_irn_n(phi, i);
1411 const arch_register_t *reg = arch_get_irn_register(op);
1412
1413 if (reg == NULL)
1414 continue;
1415 /* we only give the bonus if the predecessor already has registers
1416 * assigned, otherwise we only see a dummy value
1417 * and any conclusions about its register are useless */
1418 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1419 block_info_t *pred_block_info = get_block_info(pred_block);
1420 if (!pred_block_info->processed)
1421 continue;
1422
1423 /* give bonus for already assigned register */
1424 float weight = (float)get_block_execfreq(pred_block);
1425 info->prefs[reg->index] += weight * AFF_PHI;
1426 }
1427 }
1428
1429 /**
1430 * After a phi has been assigned a register propagate preference inputs
1431 * to the phi inputs.
1432 */
propagate_phi_register(ir_node * phi,unsigned assigned_r)1433 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1434 {
1435 ir_node *block = get_nodes_block(phi);
1436
1437 int arity = get_irn_arity(phi);
1438 for (int i = 0; i < arity; ++i) {
1439 ir_node *op = get_Phi_pred(phi, i);
1440 allocation_info_t *info = get_allocation_info(op);
1441 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1442 float weight
1443 = (float)get_block_execfreq(pred_block) * AFF_PHI;
1444
1445 if (info->prefs[assigned_r] >= weight)
1446 continue;
1447
1448 /* promote the prefered register */
1449 for (unsigned r = 0; r < n_regs; ++r) {
1450 if (info->prefs[r] > -weight) {
1451 info->prefs[r] = -weight;
1452 }
1453 }
1454 info->prefs[assigned_r] = weight;
1455
1456 if (is_Phi(op))
1457 propagate_phi_register(op, assigned_r);
1458 }
1459 }
1460
assign_phi_registers(ir_node * block)1461 static void assign_phi_registers(ir_node *block)
1462 {
1463 /* count phi nodes */
1464 int n_phis = 0;
1465 sched_foreach(block, node) {
1466 if (!is_Phi(node))
1467 break;
1468 if (!arch_irn_consider_in_reg_alloc(cls, node))
1469 continue;
1470 ++n_phis;
1471 }
1472
1473 if (n_phis == 0)
1474 return;
1475
1476 /* build a bipartite matching problem for all phi nodes */
1477 hungarian_problem_t *bp
1478 = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1479 int n = 0;
1480 sched_foreach(block, node) {
1481 if (!is_Phi(node))
1482 break;
1483 if (!arch_irn_consider_in_reg_alloc(cls, node))
1484 continue;
1485
1486 /* give boni for predecessor colorings */
1487 adapt_phi_prefs(node);
1488 /* add stuff to bipartite problem */
1489 allocation_info_t *info = get_allocation_info(node);
1490 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1491 for (unsigned r = 0; r < n_regs; ++r) {
1492 if (!rbitset_is_set(normal_regs, r))
1493 continue;
1494
1495 float costs = info->prefs[r];
1496 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1497 costs *= 100;
1498 costs += 10000;
1499 hungarian_add(bp, n, r, (int)costs);
1500 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1501 info->prefs[r]));
1502 }
1503 DB((dbg, LEVEL_3, "\n"));
1504 ++n;
1505 }
1506
1507 //hungarian_print_cost_matrix(bp, 7);
1508 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1509
1510 unsigned *assignment = ALLOCAN(unsigned, n_regs);
1511 int res = hungarian_solve(bp, assignment, NULL, 0);
1512 assert(res == 0);
1513
1514 /* apply results */
1515 n = 0;
1516 sched_foreach(block, node) {
1517 if (!is_Phi(node))
1518 break;
1519 if (!arch_irn_consider_in_reg_alloc(cls, node))
1520 continue;
1521 const arch_register_req_t *req
1522 = arch_get_irn_register_req(node);
1523
1524 unsigned r = assignment[n++];
1525 assert(rbitset_is_set(normal_regs, r));
1526 const arch_register_t *reg = arch_register_for_index(cls, r);
1527 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1528 use_reg(node, reg, req->width);
1529
1530 /* adapt preferences for phi inputs */
1531 propagate_phi_register(node, r);
1532 }
1533 }
1534
allocate_reg_req(ir_graph * irg)1535 static arch_register_req_t *allocate_reg_req(ir_graph *irg)
1536 {
1537 struct obstack *obst = be_get_be_obst(irg);
1538 arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
1539 return req;
1540 }
1541
1542 /**
1543 * Walker: assign registers to all nodes of a block that
1544 * need registers from the currently considered register class.
1545 */
allocate_coalesce_block(ir_node * block,void * data)1546 static void allocate_coalesce_block(ir_node *block, void *data)
1547 {
1548 (void) data;
1549 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1550
1551 /* clear assignments */
1552 block_info_t *block_info = get_block_info(block);
1553 assignments = block_info->assignments;
1554
1555 ir_nodeset_t live_nodes;
1556 ir_nodeset_init(&live_nodes);
1557
1558 /* gather regalloc infos of predecessor blocks */
1559 int n_preds = get_Block_n_cfgpreds(block);
1560 block_info_t **pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1561 for (int i = 0; i < n_preds; ++i) {
1562 ir_node *pred = get_Block_cfgpred_block(block, i);
1563 block_info_t *pred_info = get_block_info(pred);
1564 pred_block_infos[i] = pred_info;
1565 }
1566
1567 ir_node **phi_ins = ALLOCAN(ir_node*, n_preds);
1568
1569 /* collect live-in nodes and preassigned values */
1570 be_lv_foreach(lv, block, be_lv_state_in, node) {
1571 const arch_register_req_t *req = arch_get_irn_register_req(node);
1572 if (req->cls != cls)
1573 continue;
1574
1575 if (req->type & arch_register_req_type_ignore) {
1576 allocation_info_t *info = get_allocation_info(node);
1577 info->current_value = node;
1578
1579 const arch_register_t *reg = arch_get_irn_register(node);
1580 assert(reg != NULL); /* ignore values must be preassigned */
1581 use_reg(node, reg, req->width);
1582 continue;
1583 }
1584
1585 /* check all predecessors for this value, if it is not everywhere the
1586 same or unknown then we have to construct a phi
1587 (we collect the potential phi inputs here) */
1588 bool need_phi = false;
1589 for (int p = 0; p < n_preds; ++p) {
1590 block_info_t *pred_info = pred_block_infos[p];
1591
1592 if (!pred_info->processed) {
1593 /* use node for now, it will get fixed later */
1594 phi_ins[p] = node;
1595 need_phi = true;
1596 } else {
1597 int a = find_value_in_block_info(pred_info, node);
1598
1599 /* must live out of predecessor */
1600 assert(a >= 0);
1601 phi_ins[p] = pred_info->assignments[a];
1602 /* different value from last time? then we need a phi */
1603 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1604 need_phi = true;
1605 }
1606 }
1607 }
1608
1609 if (need_phi) {
1610 ir_mode *mode = get_irn_mode(node);
1611 const arch_register_req_t *phi_req = cls->class_req;
1612 if (req->width > 1) {
1613 arch_register_req_t *new_req = allocate_reg_req(irg);
1614 new_req->cls = cls;
1615 new_req->type = req->type & arch_register_req_type_aligned;
1616 new_req->width = req->width;
1617 phi_req = new_req;
1618 }
1619 ir_node *phi = be_new_Phi(block, n_preds, phi_ins, mode,
1620 phi_req);
1621
1622 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1623 #ifdef DEBUG_libfirm
1624 for (int pi = 0; pi < n_preds; ++pi) {
1625 DB((dbg, LEVEL_3, " %+F", phi_ins[pi]));
1626 }
1627 DB((dbg, LEVEL_3, "\n"));
1628 #endif
1629 mark_as_copy_of(phi, node);
1630 sched_add_after(block, phi);
1631
1632 node = phi;
1633 } else {
1634 allocation_info_t *info = get_allocation_info(node);
1635 info->current_value = phi_ins[0];
1636
1637 /* Grab 1 of the inputs we constructed (might not be the same as
1638 * "node" as we could see the same copy of the value in all
1639 * predecessors */
1640 node = phi_ins[0];
1641 }
1642
1643 /* if the node already has a register assigned use it */
1644 const arch_register_t *reg = arch_get_irn_register(node);
1645 if (reg != NULL) {
1646 use_reg(node, reg, req->width);
1647 }
1648
1649 /* remember that this node is live at the beginning of the block */
1650 ir_nodeset_insert(&live_nodes, node);
1651 }
1652
1653 unsigned *forbidden_regs; /**< collects registers which must
1654 not be used for optimistic splits */
1655 rbitset_alloca(forbidden_regs, n_regs);
1656
1657 /* handle phis... */
1658 assign_phi_registers(block);
1659
1660 /* all live-ins must have a register */
1661 #ifndef NDEBUG
1662 foreach_ir_nodeset(&live_nodes, node, iter) {
1663 const arch_register_t *reg = arch_get_irn_register(node);
1664 assert(reg != NULL);
1665 }
1666 #endif
1667
1668 /* assign instructions in the block */
1669 sched_foreach(block, node) {
1670 /* phis are already assigned */
1671 if (is_Phi(node))
1672 continue;
1673
1674 rewire_inputs(node);
1675
1676 /* enforce use constraints */
1677 rbitset_clear_all(forbidden_regs, n_regs);
1678 enforce_constraints(&live_nodes, node, forbidden_regs);
1679
1680 rewire_inputs(node);
1681
1682 /* we may not use registers used for inputs for optimistic splits */
1683 int arity = get_irn_arity(node);
1684 for (int i = 0; i < arity; ++i) {
1685 ir_node *op = get_irn_n(node, i);
1686 if (!arch_irn_consider_in_reg_alloc(cls, op))
1687 continue;
1688
1689 const arch_register_t *reg = arch_get_irn_register(op);
1690 rbitset_set(forbidden_regs, reg->index);
1691 }
1692
1693 /* free registers of values last used at this instruction */
1694 free_last_uses(&live_nodes, node);
1695
1696 /* assign output registers */
1697 be_foreach_definition_(node, cls, value,
1698 assign_reg(block, value, forbidden_regs);
1699 );
1700 }
1701
1702 ir_nodeset_destroy(&live_nodes);
1703 assignments = NULL;
1704
1705 block_info->processed = true;
1706
1707 /* permute values at end of predecessor blocks in case of phi-nodes */
1708 if (n_preds > 1) {
1709 for (int p = 0; p < n_preds; ++p) {
1710 add_phi_permutations(block, p);
1711 }
1712 }
1713
1714 /* if we have exactly 1 successor then we might be able to produce phi
1715 copies now */
1716 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1717 const ir_edge_t *edge
1718 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1719 ir_node *succ = get_edge_src_irn(edge);
1720 int p = get_edge_src_pos(edge);
1721 block_info_t *succ_info = get_block_info(succ);
1722
1723 if (succ_info->processed) {
1724 add_phi_permutations(succ, p);
1725 }
1726 }
1727 }
1728
1729 typedef struct block_costs_t block_costs_t;
1730 struct block_costs_t {
1731 float costs; /**< costs of the block */
1732 int dfs_num; /**< depth first search number (to detect backedges) */
1733 };
1734
cmp_block_costs(const void * d1,const void * d2)1735 static int cmp_block_costs(const void *d1, const void *d2)
1736 {
1737 const ir_node * const *block1 = (const ir_node**)d1;
1738 const ir_node * const *block2 = (const ir_node**)d2;
1739 const block_costs_t *info1 = (const block_costs_t*)get_irn_link(*block1);
1740 const block_costs_t *info2 = (const block_costs_t*)get_irn_link(*block2);
1741 return QSORT_CMP(info2->costs, info1->costs);
1742 }
1743
determine_block_order(void)1744 static void determine_block_order(void)
1745 {
1746 ir_node **blocklist = be_get_cfgpostorder(irg);
1747 size_t n_blocks = ARR_LEN(blocklist);
1748 int dfs_num = 0;
1749 pdeq *worklist = new_pdeq();
1750 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1751 size_t order_p = 0;
1752
1753 /* clear block links... */
1754 for (size_t p = 0; p < n_blocks; ++p) {
1755 ir_node *block = blocklist[p];
1756 set_irn_link(block, NULL);
1757 }
1758
1759 /* walk blocks in reverse postorder, the costs for each block are the
1760 * sum of the costs of its predecessors (excluding the costs on backedges
1761 * which we can't determine) */
1762 for (size_t p = n_blocks; p > 0;) {
1763 block_costs_t *cost_info;
1764 ir_node *block = blocklist[--p];
1765
1766 float execfreq = (float)get_block_execfreq(block);
1767 float costs = execfreq;
1768 int n_cfgpreds = get_Block_n_cfgpreds(block);
1769 for (int p2 = 0; p2 < n_cfgpreds; ++p2) {
1770 ir_node *pred_block = get_Block_cfgpred_block(block, p2);
1771 block_costs_t *pred_costs = (block_costs_t*)get_irn_link(pred_block);
1772 /* we don't have any info for backedges */
1773 if (pred_costs == NULL)
1774 continue;
1775 costs += pred_costs->costs;
1776 }
1777
1778 cost_info = OALLOCZ(&obst, block_costs_t);
1779 cost_info->costs = costs;
1780 cost_info->dfs_num = dfs_num++;
1781 set_irn_link(block, cost_info);
1782 }
1783
1784 /* sort array by block costs */
1785 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1786
1787 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1788 inc_irg_block_visited(irg);
1789
1790 for (size_t p = 0; p < n_blocks; ++p) {
1791 ir_node *block = blocklist[p];
1792 if (Block_block_visited(block))
1793 continue;
1794
1795 /* continually add predecessors with highest costs to worklist
1796 * (without using backedges) */
1797 do {
1798 block_costs_t *info = (block_costs_t*)get_irn_link(block);
1799 ir_node *best_pred = NULL;
1800 float best_costs = -1;
1801 int n_cfgpred = get_Block_n_cfgpreds(block);
1802
1803 pdeq_putr(worklist, block);
1804 mark_Block_block_visited(block);
1805 for (int i = 0; i < n_cfgpred; ++i) {
1806 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1807 block_costs_t *pred_info = (block_costs_t*)get_irn_link(pred_block);
1808
1809 /* ignore backedges */
1810 if (pred_info->dfs_num > info->dfs_num)
1811 continue;
1812
1813 if (info->costs > best_costs) {
1814 best_costs = info->costs;
1815 best_pred = pred_block;
1816 }
1817 }
1818 block = best_pred;
1819 } while (block != NULL && !Block_block_visited(block));
1820
1821 /* now put all nodes in the worklist in our final order */
1822 while (!pdeq_empty(worklist)) {
1823 ir_node *pblock = (ir_node*)pdeq_getr(worklist);
1824 assert(order_p < n_blocks);
1825 order[order_p++] = pblock;
1826 }
1827 }
1828 assert(order_p == n_blocks);
1829 del_pdeq(worklist);
1830
1831 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1832
1833 DEL_ARR_F(blocklist);
1834
1835 obstack_free(&obst, NULL);
1836 obstack_init(&obst);
1837
1838 block_order = order;
1839 n_block_order = n_blocks;
1840 }
1841
free_block_order(void)1842 static void free_block_order(void)
1843 {
1844 xfree(block_order);
1845 }
1846
1847 /**
1848 * Run the register allocator for the current register class.
1849 */
be_pref_alloc_cls(void)1850 static void be_pref_alloc_cls(void)
1851 {
1852 be_assure_live_sets(irg);
1853 lv = be_get_irg_liveness(irg);
1854
1855 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1856
1857 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1858
1859 be_clear_links(irg);
1860
1861 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1862 combine_congruence_classes();
1863
1864 for (size_t i = 0; i < n_block_order; ++i) {
1865 ir_node *block = block_order[i];
1866 allocate_coalesce_block(block, NULL);
1867 }
1868
1869 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1870 }
1871
dump(int mask,ir_graph * irg,const char * suffix)1872 static void dump(int mask, ir_graph *irg, const char *suffix)
1873 {
1874 if (be_options.dump_flags & mask)
1875 dump_ir_graph(irg, suffix);
1876 }
1877
1878 /**
1879 * Run the spiller on the current graph.
1880 */
spill(void)1881 static void spill(void)
1882 {
1883 /* make sure all nodes show their real register pressure */
1884 be_timer_push(T_RA_CONSTR);
1885 be_pre_spill_prepare_constr(irg, cls);
1886 be_timer_pop(T_RA_CONSTR);
1887
1888 dump(DUMP_RA, irg, "spillprepare");
1889
1890 /* spill */
1891 be_timer_push(T_RA_SPILL);
1892 be_do_spill(irg, cls);
1893 be_timer_pop(T_RA_SPILL);
1894
1895 be_timer_push(T_RA_SPILL_APPLY);
1896 check_for_memory_operands(irg);
1897 be_timer_pop(T_RA_SPILL_APPLY);
1898
1899 dump(DUMP_RA, irg, "spill");
1900 }
1901
1902 /**
1903 * The pref register allocator for a whole procedure.
1904 */
be_pref_alloc(ir_graph * new_irg)1905 static void be_pref_alloc(ir_graph *new_irg)
1906 {
1907 obstack_init(&obst);
1908
1909 irg = new_irg;
1910
1911 /* determine a good coloring order */
1912 determine_block_order();
1913
1914 const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
1915 int n_cls = arch_env->n_register_classes;
1916 for (int c = 0; c < n_cls; ++c) {
1917 cls = &arch_env->register_classes[c];
1918 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1919 continue;
1920
1921 stat_ev_ctx_push_str("regcls", cls->name);
1922
1923 n_regs = arch_register_class_n_regs(cls);
1924 normal_regs = rbitset_malloc(n_regs);
1925 be_set_allocatable_regs(irg, cls, normal_regs);
1926
1927 spill();
1928
1929 /* verify schedule and register pressure */
1930 be_timer_push(T_VERIFY);
1931 if (be_options.verify_option == BE_VERIFY_WARN) {
1932 be_verify_schedule(irg);
1933 be_verify_register_pressure(irg, cls);
1934 } else if (be_options.verify_option == BE_VERIFY_ASSERT) {
1935 assert(be_verify_schedule(irg) && "Schedule verification failed");
1936 assert(be_verify_register_pressure(irg, cls)
1937 && "Register pressure verification failed");
1938 }
1939 be_timer_pop(T_VERIFY);
1940
1941 be_timer_push(T_RA_COLOR);
1942 be_pref_alloc_cls();
1943 be_timer_pop(T_RA_COLOR);
1944
1945 /* we most probably constructed new Phis so liveness info is invalid
1946 * now */
1947 be_invalidate_live_sets(irg);
1948 free(normal_regs);
1949
1950 stat_ev_ctx_pop("regcls");
1951 }
1952
1953 free_block_order();
1954
1955 be_timer_push(T_RA_SPILL_APPLY);
1956 be_abi_fix_stack_nodes(irg);
1957 be_timer_pop(T_RA_SPILL_APPLY);
1958
1959 be_timer_push(T_VERIFY);
1960 if (be_options.verify_option == BE_VERIFY_WARN) {
1961 be_verify_register_allocation(irg);
1962 } else if (be_options.verify_option == BE_VERIFY_ASSERT) {
1963 assert(be_verify_register_allocation(irg)
1964 && "Register allocation invalid");
1965 }
1966 be_timer_pop(T_VERIFY);
1967
1968 obstack_free(&obst, NULL);
1969 }
1970
BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc)1971 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc)
1972 void be_init_pref_alloc(void)
1973 {
1974 static be_ra_t be_ra_pref = { be_pref_alloc };
1975 be_register_allocator("pref", &be_ra_pref);
1976 FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");
1977 }
1978