Suppress -fstack-protector warning on hppa.
[official-gcc.git] / gcc / ira-color.cc
blobf5f76e8089c9c229bbac77d9f24db5279800d763
1 /* IRA allocation based on graph coloring.
2 Copyright (C) 2006-2022 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "backend.h"
25 #include "target.h"
26 #include "rtl.h"
27 #include "tree.h"
28 #include "predict.h"
29 #include "df.h"
30 #include "memmodel.h"
31 #include "tm_p.h"
32 #include "insn-config.h"
33 #include "regs.h"
34 #include "ira.h"
35 #include "ira-int.h"
36 #include "reload.h"
37 #include "cfgloop.h"
39 /* To prevent soft conflict detection becoming quadratic in the
40 loop depth. Only for very pathological cases, so it hardly
41 seems worth a --param. */
42 const int max_soft_conflict_loop_depth = 64;
44 typedef struct allocno_hard_regs *allocno_hard_regs_t;
46 /* The structure contains information about hard registers can be
47 assigned to allocnos. Usually it is allocno profitable hard
48 registers but in some cases this set can be a bit different. Major
49 reason of the difference is a requirement to use hard register sets
50 that form a tree or a forest (set of trees), i.e. hard register set
51 of a node should contain hard register sets of its subnodes. */
52 struct allocno_hard_regs
54 /* Hard registers can be assigned to an allocno. */
55 HARD_REG_SET set;
56 /* Overall (spilling) cost of all allocnos with given register
57 set. */
58 int64_t cost;
61 typedef struct allocno_hard_regs_node *allocno_hard_regs_node_t;
63 /* A node representing allocno hard registers. Such nodes form a
64 forest (set of trees). Each subnode of given node in the forest
65 refers for hard register set (usually allocno profitable hard
66 register set) which is a subset of one referred from given
67 node. */
68 struct allocno_hard_regs_node
70 /* Set up number of the node in preorder traversing of the forest. */
71 int preorder_num;
72 /* Used for different calculation like finding conflict size of an
73 allocno. */
74 int check;
75 /* Used for calculation of conflict size of an allocno. The
76 conflict size of the allocno is maximal number of given allocno
77 hard registers needed for allocation of the conflicting allocnos.
78 Given allocno is trivially colored if this number plus the number
79 of hard registers needed for given allocno is not greater than
80 the number of given allocno hard register set. */
81 int conflict_size;
82 /* The number of hard registers given by member hard_regs. */
83 int hard_regs_num;
84 /* The following member is used to form the final forest. */
85 bool used_p;
86 /* Pointer to the corresponding profitable hard registers. */
87 allocno_hard_regs_t hard_regs;
88 /* Parent, first subnode, previous and next node with the same
89 parent in the forest. */
90 allocno_hard_regs_node_t parent, first, prev, next;
93 /* Info about changing hard reg costs of an allocno. */
94 struct update_cost_record
96 /* Hard regno for which we changed the cost. */
97 int hard_regno;
98 /* Divisor used when we changed the cost of HARD_REGNO. */
99 int divisor;
100 /* Next record for given allocno. */
101 struct update_cost_record *next;
104 /* To decrease footprint of ira_allocno structure we store all data
105 needed only for coloring in the following structure. */
106 struct allocno_color_data
108 /* TRUE value means that the allocno was not removed yet from the
109 conflicting graph during coloring. */
110 unsigned int in_graph_p : 1;
111 /* TRUE if it is put on the stack to make other allocnos
112 colorable. */
113 unsigned int may_be_spilled_p : 1;
114 /* TRUE if the allocno is trivially colorable. */
115 unsigned int colorable_p : 1;
116 /* Number of hard registers of the allocno class really
117 available for the allocno allocation. It is number of the
118 profitable hard regs. */
119 int available_regs_num;
120 /* Sum of frequencies of hard register preferences of all
121 conflicting allocnos which are not the coloring stack yet. */
122 int conflict_allocno_hard_prefs;
123 /* Allocnos in a bucket (used in coloring) chained by the following
124 two members. */
125 ira_allocno_t next_bucket_allocno;
126 ira_allocno_t prev_bucket_allocno;
127 /* Used for temporary purposes. */
128 int temp;
129 /* Used to exclude repeated processing. */
130 int last_process;
131 /* Profitable hard regs available for this pseudo allocation. It
132 means that the set excludes unavailable hard regs and hard regs
133 conflicting with given pseudo. They should be of the allocno
134 class. */
135 HARD_REG_SET profitable_hard_regs;
136 /* The allocno hard registers node. */
137 allocno_hard_regs_node_t hard_regs_node;
138 /* Array of structures allocno_hard_regs_subnode representing
139 given allocno hard registers node (the 1st element in the array)
140 and all its subnodes in the tree (forest) of allocno hard
141 register nodes (see comments above). */
142 int hard_regs_subnodes_start;
143 /* The length of the previous array. */
144 int hard_regs_subnodes_num;
145 /* Records about updating allocno hard reg costs from copies. If
146 the allocno did not get expected hard register, these records are
147 used to restore original hard reg costs of allocnos connected to
148 this allocno by copies. */
149 struct update_cost_record *update_cost_records;
150 /* Threads. We collect allocnos connected by copies into threads
151 and try to assign hard regs to allocnos by threads. */
152 /* Allocno representing all thread. */
153 ira_allocno_t first_thread_allocno;
154 /* Allocnos in thread forms a cycle list through the following
155 member. */
156 ira_allocno_t next_thread_allocno;
157 /* All thread frequency. Defined only for first thread allocno. */
158 int thread_freq;
159 /* Sum of frequencies of hard register preferences of the allocno. */
160 int hard_reg_prefs;
163 /* See above. */
164 typedef struct allocno_color_data *allocno_color_data_t;
166 /* Container for storing allocno data concerning coloring. */
167 static allocno_color_data_t allocno_color_data;
169 /* Macro to access the data concerning coloring. */
170 #define ALLOCNO_COLOR_DATA(a) ((allocno_color_data_t) ALLOCNO_ADD_DATA (a))
172 /* Used for finding allocno colorability to exclude repeated allocno
173 processing and for updating preferencing to exclude repeated
174 allocno processing during assignment. */
175 static int curr_allocno_process;
177 /* This file contains code for regional graph coloring, spill/restore
178 code placement optimization, and code helping the reload pass to do
179 a better job. */
181 /* Bitmap of allocnos which should be colored. */
182 static bitmap coloring_allocno_bitmap;
184 /* Bitmap of allocnos which should be taken into account during
185 coloring. In general case it contains allocnos from
186 coloring_allocno_bitmap plus other already colored conflicting
187 allocnos. */
188 static bitmap consideration_allocno_bitmap;
190 /* All allocnos sorted according their priorities. */
191 static ira_allocno_t *sorted_allocnos;
193 /* Vec representing the stack of allocnos used during coloring. */
194 static vec<ira_allocno_t> allocno_stack_vec;
196 /* Helper for qsort comparison callbacks - return a positive integer if
197 X > Y, or a negative value otherwise. Use a conditional expression
198 instead of a difference computation to insulate from possible overflow
199 issues, e.g. X - Y < 0 for some X > 0 and Y < 0. */
200 #define SORTGT(x,y) (((x) > (y)) ? 1 : -1)
204 /* Definition of vector of allocno hard registers. */
206 /* Vector of unique allocno hard registers. */
207 static vec<allocno_hard_regs_t> allocno_hard_regs_vec;
209 struct allocno_hard_regs_hasher : nofree_ptr_hash <allocno_hard_regs>
211 static inline hashval_t hash (const allocno_hard_regs *);
212 static inline bool equal (const allocno_hard_regs *,
213 const allocno_hard_regs *);
216 /* Returns hash value for allocno hard registers V. */
217 inline hashval_t
218 allocno_hard_regs_hasher::hash (const allocno_hard_regs *hv)
220 return iterative_hash (&hv->set, sizeof (HARD_REG_SET), 0);
223 /* Compares allocno hard registers V1 and V2. */
224 inline bool
225 allocno_hard_regs_hasher::equal (const allocno_hard_regs *hv1,
226 const allocno_hard_regs *hv2)
228 return hv1->set == hv2->set;
231 /* Hash table of unique allocno hard registers. */
232 static hash_table<allocno_hard_regs_hasher> *allocno_hard_regs_htab;
234 /* Return allocno hard registers in the hash table equal to HV. */
235 static allocno_hard_regs_t
236 find_hard_regs (allocno_hard_regs_t hv)
238 return allocno_hard_regs_htab->find (hv);
241 /* Insert allocno hard registers HV in the hash table (if it is not
242 there yet) and return the value which in the table. */
243 static allocno_hard_regs_t
244 insert_hard_regs (allocno_hard_regs_t hv)
246 allocno_hard_regs **slot = allocno_hard_regs_htab->find_slot (hv, INSERT);
248 if (*slot == NULL)
249 *slot = hv;
250 return *slot;
253 /* Initialize data concerning allocno hard registers. */
254 static void
255 init_allocno_hard_regs (void)
257 allocno_hard_regs_vec.create (200);
258 allocno_hard_regs_htab
259 = new hash_table<allocno_hard_regs_hasher> (200);
262 /* Add (or update info about) allocno hard registers with SET and
263 COST. */
264 static allocno_hard_regs_t
265 add_allocno_hard_regs (HARD_REG_SET set, int64_t cost)
267 struct allocno_hard_regs temp;
268 allocno_hard_regs_t hv;
270 gcc_assert (! hard_reg_set_empty_p (set));
271 temp.set = set;
272 if ((hv = find_hard_regs (&temp)) != NULL)
273 hv->cost += cost;
274 else
276 hv = ((struct allocno_hard_regs *)
277 ira_allocate (sizeof (struct allocno_hard_regs)));
278 hv->set = set;
279 hv->cost = cost;
280 allocno_hard_regs_vec.safe_push (hv);
281 insert_hard_regs (hv);
283 return hv;
286 /* Finalize data concerning allocno hard registers. */
287 static void
288 finish_allocno_hard_regs (void)
290 int i;
291 allocno_hard_regs_t hv;
293 for (i = 0;
294 allocno_hard_regs_vec.iterate (i, &hv);
295 i++)
296 ira_free (hv);
297 delete allocno_hard_regs_htab;
298 allocno_hard_regs_htab = NULL;
299 allocno_hard_regs_vec.release ();
302 /* Sort hard regs according to their frequency of usage. */
303 static int
304 allocno_hard_regs_compare (const void *v1p, const void *v2p)
306 allocno_hard_regs_t hv1 = *(const allocno_hard_regs_t *) v1p;
307 allocno_hard_regs_t hv2 = *(const allocno_hard_regs_t *) v2p;
309 if (hv2->cost > hv1->cost)
310 return 1;
311 else if (hv2->cost < hv1->cost)
312 return -1;
313 return SORTGT (allocno_hard_regs_hasher::hash(hv2), allocno_hard_regs_hasher::hash(hv1));
318 /* Used for finding a common ancestor of two allocno hard registers
319 nodes in the forest. We use the current value of
320 'node_check_tick' to mark all nodes from one node to the top and
321 then walking up from another node until we find a marked node.
323 It is also used to figure out allocno colorability as a mark that
324 we already reset value of member 'conflict_size' for the forest
325 node corresponding to the processed allocno. */
326 static int node_check_tick;
328 /* Roots of the forest containing hard register sets can be assigned
329 to allocnos. */
330 static allocno_hard_regs_node_t hard_regs_roots;
332 /* Definition of vector of allocno hard register nodes. */
334 /* Vector used to create the forest. */
335 static vec<allocno_hard_regs_node_t> hard_regs_node_vec;
337 /* Create and return allocno hard registers node containing allocno
338 hard registers HV. */
339 static allocno_hard_regs_node_t
340 create_new_allocno_hard_regs_node (allocno_hard_regs_t hv)
342 allocno_hard_regs_node_t new_node;
344 new_node = ((struct allocno_hard_regs_node *)
345 ira_allocate (sizeof (struct allocno_hard_regs_node)));
346 new_node->check = 0;
347 new_node->hard_regs = hv;
348 new_node->hard_regs_num = hard_reg_set_size (hv->set);
349 new_node->first = NULL;
350 new_node->used_p = false;
351 return new_node;
354 /* Add allocno hard registers node NEW_NODE to the forest on its level
355 given by ROOTS. */
356 static void
357 add_new_allocno_hard_regs_node_to_forest (allocno_hard_regs_node_t *roots,
358 allocno_hard_regs_node_t new_node)
360 new_node->next = *roots;
361 if (new_node->next != NULL)
362 new_node->next->prev = new_node;
363 new_node->prev = NULL;
364 *roots = new_node;
367 /* Add allocno hard registers HV (or its best approximation if it is
368 not possible) to the forest on its level given by ROOTS. */
369 static void
370 add_allocno_hard_regs_to_forest (allocno_hard_regs_node_t *roots,
371 allocno_hard_regs_t hv)
373 unsigned int i, start;
374 allocno_hard_regs_node_t node, prev, new_node;
375 HARD_REG_SET temp_set;
376 allocno_hard_regs_t hv2;
378 start = hard_regs_node_vec.length ();
379 for (node = *roots; node != NULL; node = node->next)
381 if (hv->set == node->hard_regs->set)
382 return;
383 if (hard_reg_set_subset_p (hv->set, node->hard_regs->set))
385 add_allocno_hard_regs_to_forest (&node->first, hv);
386 return;
388 if (hard_reg_set_subset_p (node->hard_regs->set, hv->set))
389 hard_regs_node_vec.safe_push (node);
390 else if (hard_reg_set_intersect_p (hv->set, node->hard_regs->set))
392 temp_set = hv->set & node->hard_regs->set;
393 hv2 = add_allocno_hard_regs (temp_set, hv->cost);
394 add_allocno_hard_regs_to_forest (&node->first, hv2);
397 if (hard_regs_node_vec.length ()
398 > start + 1)
400 /* Create a new node which contains nodes in hard_regs_node_vec. */
401 CLEAR_HARD_REG_SET (temp_set);
402 for (i = start;
403 i < hard_regs_node_vec.length ();
404 i++)
406 node = hard_regs_node_vec[i];
407 temp_set |= node->hard_regs->set;
409 hv = add_allocno_hard_regs (temp_set, hv->cost);
410 new_node = create_new_allocno_hard_regs_node (hv);
411 prev = NULL;
412 for (i = start;
413 i < hard_regs_node_vec.length ();
414 i++)
416 node = hard_regs_node_vec[i];
417 if (node->prev == NULL)
418 *roots = node->next;
419 else
420 node->prev->next = node->next;
421 if (node->next != NULL)
422 node->next->prev = node->prev;
423 if (prev == NULL)
424 new_node->first = node;
425 else
426 prev->next = node;
427 node->prev = prev;
428 node->next = NULL;
429 prev = node;
431 add_new_allocno_hard_regs_node_to_forest (roots, new_node);
433 hard_regs_node_vec.truncate (start);
436 /* Add allocno hard registers nodes starting with the forest level
437 given by FIRST which contains biggest set inside SET. */
438 static void
439 collect_allocno_hard_regs_cover (allocno_hard_regs_node_t first,
440 HARD_REG_SET set)
442 allocno_hard_regs_node_t node;
444 ira_assert (first != NULL);
445 for (node = first; node != NULL; node = node->next)
446 if (hard_reg_set_subset_p (node->hard_regs->set, set))
447 hard_regs_node_vec.safe_push (node);
448 else if (hard_reg_set_intersect_p (set, node->hard_regs->set))
449 collect_allocno_hard_regs_cover (node->first, set);
452 /* Set up field parent as PARENT in all allocno hard registers nodes
453 in forest given by FIRST. */
454 static void
455 setup_allocno_hard_regs_nodes_parent (allocno_hard_regs_node_t first,
456 allocno_hard_regs_node_t parent)
458 allocno_hard_regs_node_t node;
460 for (node = first; node != NULL; node = node->next)
462 node->parent = parent;
463 setup_allocno_hard_regs_nodes_parent (node->first, node);
467 /* Return allocno hard registers node which is a first common ancestor
468 node of FIRST and SECOND in the forest. */
469 static allocno_hard_regs_node_t
470 first_common_ancestor_node (allocno_hard_regs_node_t first,
471 allocno_hard_regs_node_t second)
473 allocno_hard_regs_node_t node;
475 node_check_tick++;
476 for (node = first; node != NULL; node = node->parent)
477 node->check = node_check_tick;
478 for (node = second; node != NULL; node = node->parent)
479 if (node->check == node_check_tick)
480 return node;
481 return first_common_ancestor_node (second, first);
484 /* Print hard reg set SET to F. */
485 static void
486 print_hard_reg_set (FILE *f, HARD_REG_SET set, bool new_line_p)
488 int i, start, end;
490 for (start = end = -1, i = 0; i < FIRST_PSEUDO_REGISTER; i++)
492 bool reg_included = TEST_HARD_REG_BIT (set, i);
494 if (reg_included)
496 if (start == -1)
497 start = i;
498 end = i;
500 if (start >= 0 && (!reg_included || i == FIRST_PSEUDO_REGISTER - 1))
502 if (start == end)
503 fprintf (f, " %d", start);
504 else if (start == end + 1)
505 fprintf (f, " %d %d", start, end);
506 else
507 fprintf (f, " %d-%d", start, end);
508 start = -1;
511 if (new_line_p)
512 fprintf (f, "\n");
515 /* Print allocno hard register subforest given by ROOTS and its LEVEL
516 to F. */
517 static void
518 print_hard_regs_subforest (FILE *f, allocno_hard_regs_node_t roots,
519 int level)
521 int i;
522 allocno_hard_regs_node_t node;
524 for (node = roots; node != NULL; node = node->next)
526 fprintf (f, " ");
527 for (i = 0; i < level * 2; i++)
528 fprintf (f, " ");
529 fprintf (f, "%d:(", node->preorder_num);
530 print_hard_reg_set (f, node->hard_regs->set, false);
531 fprintf (f, ")@%" PRId64"\n", node->hard_regs->cost);
532 print_hard_regs_subforest (f, node->first, level + 1);
536 /* Print the allocno hard register forest to F. */
537 static void
538 print_hard_regs_forest (FILE *f)
540 fprintf (f, " Hard reg set forest:\n");
541 print_hard_regs_subforest (f, hard_regs_roots, 1);
544 /* Print the allocno hard register forest to stderr. */
545 void
546 ira_debug_hard_regs_forest (void)
548 print_hard_regs_forest (stderr);
551 /* Remove unused allocno hard registers nodes from forest given by its
552 *ROOTS. */
553 static void
554 remove_unused_allocno_hard_regs_nodes (allocno_hard_regs_node_t *roots)
556 allocno_hard_regs_node_t node, prev, next, last;
558 for (prev = NULL, node = *roots; node != NULL; node = next)
560 next = node->next;
561 if (node->used_p)
563 remove_unused_allocno_hard_regs_nodes (&node->first);
564 prev = node;
566 else
568 for (last = node->first;
569 last != NULL && last->next != NULL;
570 last = last->next)
572 if (last != NULL)
574 if (prev == NULL)
575 *roots = node->first;
576 else
577 prev->next = node->first;
578 if (next != NULL)
579 next->prev = last;
580 last->next = next;
581 next = node->first;
583 else
585 if (prev == NULL)
586 *roots = next;
587 else
588 prev->next = next;
589 if (next != NULL)
590 next->prev = prev;
592 ira_free (node);
597 /* Set up fields preorder_num starting with START_NUM in all allocno
598 hard registers nodes in forest given by FIRST. Return biggest set
599 PREORDER_NUM increased by 1. */
600 static int
601 enumerate_allocno_hard_regs_nodes (allocno_hard_regs_node_t first,
602 allocno_hard_regs_node_t parent,
603 int start_num)
605 allocno_hard_regs_node_t node;
607 for (node = first; node != NULL; node = node->next)
609 node->preorder_num = start_num++;
610 node->parent = parent;
611 start_num = enumerate_allocno_hard_regs_nodes (node->first, node,
612 start_num);
614 return start_num;
617 /* Number of allocno hard registers nodes in the forest. */
618 static int allocno_hard_regs_nodes_num;
620 /* Table preorder number of allocno hard registers node in the forest
621 -> the allocno hard registers node. */
622 static allocno_hard_regs_node_t *allocno_hard_regs_nodes;
624 /* See below. */
625 typedef struct allocno_hard_regs_subnode *allocno_hard_regs_subnode_t;
627 /* The structure is used to describes all subnodes (not only immediate
628 ones) in the mentioned above tree for given allocno hard register
629 node. The usage of such data accelerates calculation of
630 colorability of given allocno. */
631 struct allocno_hard_regs_subnode
633 /* The conflict size of conflicting allocnos whose hard register
634 sets are equal sets (plus supersets if given node is given
635 allocno hard registers node) of one in the given node. */
636 int left_conflict_size;
637 /* The summary conflict size of conflicting allocnos whose hard
638 register sets are strict subsets of one in the given node.
639 Overall conflict size is
640 left_conflict_subnodes_size
641 + MIN (max_node_impact - left_conflict_subnodes_size,
642 left_conflict_size)
644 short left_conflict_subnodes_size;
645 short max_node_impact;
648 /* Container for hard regs subnodes of all allocnos. */
649 static allocno_hard_regs_subnode_t allocno_hard_regs_subnodes;
651 /* Table (preorder number of allocno hard registers node in the
652 forest, preorder number of allocno hard registers subnode) -> index
653 of the subnode relative to the node. -1 if it is not a
654 subnode. */
655 static int *allocno_hard_regs_subnode_index;
657 /* Setup arrays ALLOCNO_HARD_REGS_NODES and
658 ALLOCNO_HARD_REGS_SUBNODE_INDEX. */
659 static void
660 setup_allocno_hard_regs_subnode_index (allocno_hard_regs_node_t first)
662 allocno_hard_regs_node_t node, parent;
663 int index;
665 for (node = first; node != NULL; node = node->next)
667 allocno_hard_regs_nodes[node->preorder_num] = node;
668 for (parent = node; parent != NULL; parent = parent->parent)
670 index = parent->preorder_num * allocno_hard_regs_nodes_num;
671 allocno_hard_regs_subnode_index[index + node->preorder_num]
672 = node->preorder_num - parent->preorder_num;
674 setup_allocno_hard_regs_subnode_index (node->first);
678 /* Count all allocno hard registers nodes in tree ROOT. */
679 static int
680 get_allocno_hard_regs_subnodes_num (allocno_hard_regs_node_t root)
682 int len = 1;
684 for (root = root->first; root != NULL; root = root->next)
685 len += get_allocno_hard_regs_subnodes_num (root);
686 return len;
689 /* Build the forest of allocno hard registers nodes and assign each
690 allocno a node from the forest. */
691 static void
692 form_allocno_hard_regs_nodes_forest (void)
694 unsigned int i, j, size, len;
695 int start;
696 ira_allocno_t a;
697 allocno_hard_regs_t hv;
698 bitmap_iterator bi;
699 HARD_REG_SET temp;
700 allocno_hard_regs_node_t node, allocno_hard_regs_node;
701 allocno_color_data_t allocno_data;
703 node_check_tick = 0;
704 init_allocno_hard_regs ();
705 hard_regs_roots = NULL;
706 hard_regs_node_vec.create (100);
707 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
708 if (! TEST_HARD_REG_BIT (ira_no_alloc_regs, i))
710 CLEAR_HARD_REG_SET (temp);
711 SET_HARD_REG_BIT (temp, i);
712 hv = add_allocno_hard_regs (temp, 0);
713 node = create_new_allocno_hard_regs_node (hv);
714 add_new_allocno_hard_regs_node_to_forest (&hard_regs_roots, node);
716 start = allocno_hard_regs_vec.length ();
717 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
719 a = ira_allocnos[i];
720 allocno_data = ALLOCNO_COLOR_DATA (a);
722 if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs))
723 continue;
724 hv = (add_allocno_hard_regs
725 (allocno_data->profitable_hard_regs,
726 ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a)));
728 temp = ~ira_no_alloc_regs;
729 add_allocno_hard_regs (temp, 0);
730 qsort (allocno_hard_regs_vec.address () + start,
731 allocno_hard_regs_vec.length () - start,
732 sizeof (allocno_hard_regs_t), allocno_hard_regs_compare);
733 for (i = start;
734 allocno_hard_regs_vec.iterate (i, &hv);
735 i++)
737 add_allocno_hard_regs_to_forest (&hard_regs_roots, hv);
738 ira_assert (hard_regs_node_vec.length () == 0);
740 /* We need to set up parent fields for right work of
741 first_common_ancestor_node. */
742 setup_allocno_hard_regs_nodes_parent (hard_regs_roots, NULL);
743 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
745 a = ira_allocnos[i];
746 allocno_data = ALLOCNO_COLOR_DATA (a);
747 if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs))
748 continue;
749 hard_regs_node_vec.truncate (0);
750 collect_allocno_hard_regs_cover (hard_regs_roots,
751 allocno_data->profitable_hard_regs);
752 allocno_hard_regs_node = NULL;
753 for (j = 0; hard_regs_node_vec.iterate (j, &node); j++)
754 allocno_hard_regs_node
755 = (j == 0
756 ? node
757 : first_common_ancestor_node (node, allocno_hard_regs_node));
758 /* That is a temporary storage. */
759 allocno_hard_regs_node->used_p = true;
760 allocno_data->hard_regs_node = allocno_hard_regs_node;
762 ira_assert (hard_regs_roots->next == NULL);
763 hard_regs_roots->used_p = true;
764 remove_unused_allocno_hard_regs_nodes (&hard_regs_roots);
765 allocno_hard_regs_nodes_num
766 = enumerate_allocno_hard_regs_nodes (hard_regs_roots, NULL, 0);
767 allocno_hard_regs_nodes
768 = ((allocno_hard_regs_node_t *)
769 ira_allocate (allocno_hard_regs_nodes_num
770 * sizeof (allocno_hard_regs_node_t)));
771 size = allocno_hard_regs_nodes_num * allocno_hard_regs_nodes_num;
772 allocno_hard_regs_subnode_index
773 = (int *) ira_allocate (size * sizeof (int));
774 for (i = 0; i < size; i++)
775 allocno_hard_regs_subnode_index[i] = -1;
776 setup_allocno_hard_regs_subnode_index (hard_regs_roots);
777 start = 0;
778 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
780 a = ira_allocnos[i];
781 allocno_data = ALLOCNO_COLOR_DATA (a);
782 if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs))
783 continue;
784 len = get_allocno_hard_regs_subnodes_num (allocno_data->hard_regs_node);
785 allocno_data->hard_regs_subnodes_start = start;
786 allocno_data->hard_regs_subnodes_num = len;
787 start += len;
789 allocno_hard_regs_subnodes
790 = ((allocno_hard_regs_subnode_t)
791 ira_allocate (sizeof (struct allocno_hard_regs_subnode) * start));
792 hard_regs_node_vec.release ();
795 /* Free tree of allocno hard registers nodes given by its ROOT. */
796 static void
797 finish_allocno_hard_regs_nodes_tree (allocno_hard_regs_node_t root)
799 allocno_hard_regs_node_t child, next;
801 for (child = root->first; child != NULL; child = next)
803 next = child->next;
804 finish_allocno_hard_regs_nodes_tree (child);
806 ira_free (root);
809 /* Finish work with the forest of allocno hard registers nodes. */
810 static void
811 finish_allocno_hard_regs_nodes_forest (void)
813 allocno_hard_regs_node_t node, next;
815 ira_free (allocno_hard_regs_subnodes);
816 for (node = hard_regs_roots; node != NULL; node = next)
818 next = node->next;
819 finish_allocno_hard_regs_nodes_tree (node);
821 ira_free (allocno_hard_regs_nodes);
822 ira_free (allocno_hard_regs_subnode_index);
823 finish_allocno_hard_regs ();
826 /* Set up left conflict sizes and left conflict subnodes sizes of hard
827 registers subnodes of allocno A. Return TRUE if allocno A is
828 trivially colorable. */
829 static bool
830 setup_left_conflict_sizes_p (ira_allocno_t a)
832 int i, k, nobj, start;
833 int conflict_size, left_conflict_subnodes_size, node_preorder_num;
834 allocno_color_data_t data;
835 HARD_REG_SET profitable_hard_regs;
836 allocno_hard_regs_subnode_t subnodes;
837 allocno_hard_regs_node_t node;
838 HARD_REG_SET node_set;
840 nobj = ALLOCNO_NUM_OBJECTS (a);
841 data = ALLOCNO_COLOR_DATA (a);
842 subnodes = allocno_hard_regs_subnodes + data->hard_regs_subnodes_start;
843 profitable_hard_regs = data->profitable_hard_regs;
844 node = data->hard_regs_node;
845 node_preorder_num = node->preorder_num;
846 node_set = node->hard_regs->set;
847 node_check_tick++;
848 for (k = 0; k < nobj; k++)
850 ira_object_t obj = ALLOCNO_OBJECT (a, k);
851 ira_object_t conflict_obj;
852 ira_object_conflict_iterator oci;
854 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
856 int size;
857 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
858 allocno_hard_regs_node_t conflict_node, temp_node;
859 HARD_REG_SET conflict_node_set;
860 allocno_color_data_t conflict_data;
862 conflict_data = ALLOCNO_COLOR_DATA (conflict_a);
863 if (! ALLOCNO_COLOR_DATA (conflict_a)->in_graph_p
864 || ! hard_reg_set_intersect_p (profitable_hard_regs,
865 conflict_data
866 ->profitable_hard_regs))
867 continue;
868 conflict_node = conflict_data->hard_regs_node;
869 conflict_node_set = conflict_node->hard_regs->set;
870 if (hard_reg_set_subset_p (node_set, conflict_node_set))
871 temp_node = node;
872 else
874 ira_assert (hard_reg_set_subset_p (conflict_node_set, node_set));
875 temp_node = conflict_node;
877 if (temp_node->check != node_check_tick)
879 temp_node->check = node_check_tick;
880 temp_node->conflict_size = 0;
882 size = (ira_reg_class_max_nregs
883 [ALLOCNO_CLASS (conflict_a)][ALLOCNO_MODE (conflict_a)]);
884 if (ALLOCNO_NUM_OBJECTS (conflict_a) > 1)
885 /* We will deal with the subwords individually. */
886 size = 1;
887 temp_node->conflict_size += size;
890 for (i = 0; i < data->hard_regs_subnodes_num; i++)
892 allocno_hard_regs_node_t temp_node;
894 temp_node = allocno_hard_regs_nodes[i + node_preorder_num];
895 ira_assert (temp_node->preorder_num == i + node_preorder_num);
896 subnodes[i].left_conflict_size = (temp_node->check != node_check_tick
897 ? 0 : temp_node->conflict_size);
898 if (hard_reg_set_subset_p (temp_node->hard_regs->set,
899 profitable_hard_regs))
900 subnodes[i].max_node_impact = temp_node->hard_regs_num;
901 else
903 HARD_REG_SET temp_set;
904 int j, n, hard_regno;
905 enum reg_class aclass;
907 temp_set = temp_node->hard_regs->set & profitable_hard_regs;
908 aclass = ALLOCNO_CLASS (a);
909 for (n = 0, j = ira_class_hard_regs_num[aclass] - 1; j >= 0; j--)
911 hard_regno = ira_class_hard_regs[aclass][j];
912 if (TEST_HARD_REG_BIT (temp_set, hard_regno))
913 n++;
915 subnodes[i].max_node_impact = n;
917 subnodes[i].left_conflict_subnodes_size = 0;
919 start = node_preorder_num * allocno_hard_regs_nodes_num;
920 for (i = data->hard_regs_subnodes_num - 1; i > 0; i--)
922 int size, parent_i;
923 allocno_hard_regs_node_t parent;
925 size = (subnodes[i].left_conflict_subnodes_size
926 + MIN (subnodes[i].max_node_impact
927 - subnodes[i].left_conflict_subnodes_size,
928 subnodes[i].left_conflict_size));
929 parent = allocno_hard_regs_nodes[i + node_preorder_num]->parent;
930 gcc_checking_assert(parent);
931 parent_i
932 = allocno_hard_regs_subnode_index[start + parent->preorder_num];
933 gcc_checking_assert(parent_i >= 0);
934 subnodes[parent_i].left_conflict_subnodes_size += size;
936 left_conflict_subnodes_size = subnodes[0].left_conflict_subnodes_size;
937 conflict_size
938 = (left_conflict_subnodes_size
939 + MIN (subnodes[0].max_node_impact - left_conflict_subnodes_size,
940 subnodes[0].left_conflict_size));
941 conflict_size += ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)];
942 data->colorable_p = conflict_size <= data->available_regs_num;
943 return data->colorable_p;
946 /* Update left conflict sizes of hard registers subnodes of allocno A
947 after removing allocno REMOVED_A with SIZE from the conflict graph.
948 Return TRUE if A is trivially colorable. */
949 static bool
950 update_left_conflict_sizes_p (ira_allocno_t a,
951 ira_allocno_t removed_a, int size)
953 int i, conflict_size, before_conflict_size, diff, start;
954 int node_preorder_num, parent_i;
955 allocno_hard_regs_node_t node, removed_node, parent;
956 allocno_hard_regs_subnode_t subnodes;
957 allocno_color_data_t data = ALLOCNO_COLOR_DATA (a);
959 ira_assert (! data->colorable_p);
960 node = data->hard_regs_node;
961 node_preorder_num = node->preorder_num;
962 removed_node = ALLOCNO_COLOR_DATA (removed_a)->hard_regs_node;
963 ira_assert (hard_reg_set_subset_p (removed_node->hard_regs->set,
964 node->hard_regs->set)
965 || hard_reg_set_subset_p (node->hard_regs->set,
966 removed_node->hard_regs->set));
967 start = node_preorder_num * allocno_hard_regs_nodes_num;
968 i = allocno_hard_regs_subnode_index[start + removed_node->preorder_num];
969 if (i < 0)
970 i = 0;
971 subnodes = allocno_hard_regs_subnodes + data->hard_regs_subnodes_start;
972 before_conflict_size
973 = (subnodes[i].left_conflict_subnodes_size
974 + MIN (subnodes[i].max_node_impact
975 - subnodes[i].left_conflict_subnodes_size,
976 subnodes[i].left_conflict_size));
977 subnodes[i].left_conflict_size -= size;
978 for (;;)
980 conflict_size
981 = (subnodes[i].left_conflict_subnodes_size
982 + MIN (subnodes[i].max_node_impact
983 - subnodes[i].left_conflict_subnodes_size,
984 subnodes[i].left_conflict_size));
985 if ((diff = before_conflict_size - conflict_size) == 0)
986 break;
987 ira_assert (conflict_size < before_conflict_size);
988 parent = allocno_hard_regs_nodes[i + node_preorder_num]->parent;
989 if (parent == NULL)
990 break;
991 parent_i
992 = allocno_hard_regs_subnode_index[start + parent->preorder_num];
993 if (parent_i < 0)
994 break;
995 i = parent_i;
996 before_conflict_size
997 = (subnodes[i].left_conflict_subnodes_size
998 + MIN (subnodes[i].max_node_impact
999 - subnodes[i].left_conflict_subnodes_size,
1000 subnodes[i].left_conflict_size));
1001 subnodes[i].left_conflict_subnodes_size -= diff;
1003 if (i != 0
1004 || (conflict_size
1005 + ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]
1006 > data->available_regs_num))
1007 return false;
1008 data->colorable_p = true;
1009 return true;
1012 /* Return true if allocno A has empty profitable hard regs. */
1013 static bool
1014 empty_profitable_hard_regs (ira_allocno_t a)
1016 allocno_color_data_t data = ALLOCNO_COLOR_DATA (a);
1018 return hard_reg_set_empty_p (data->profitable_hard_regs);
1021 /* Set up profitable hard registers for each allocno being
1022 colored. */
1023 static void
1024 setup_profitable_hard_regs (void)
1026 unsigned int i;
1027 int j, k, nobj, hard_regno, nregs, class_size;
1028 ira_allocno_t a;
1029 bitmap_iterator bi;
1030 enum reg_class aclass;
1031 machine_mode mode;
1032 allocno_color_data_t data;
1034 /* Initial set up from allocno classes and explicitly conflicting
1035 hard regs. */
1036 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
1038 a = ira_allocnos[i];
1039 if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS)
1040 continue;
1041 data = ALLOCNO_COLOR_DATA (a);
1042 if (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL
1043 && ALLOCNO_CLASS_COST (a) > ALLOCNO_MEMORY_COST (a)
1044 /* Do not empty profitable regs for static chain pointer
1045 pseudo when non-local goto is used. */
1046 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a)))
1047 CLEAR_HARD_REG_SET (data->profitable_hard_regs);
1048 else
1050 mode = ALLOCNO_MODE (a);
1051 data->profitable_hard_regs
1052 = ira_useful_class_mode_regs[aclass][mode];
1053 nobj = ALLOCNO_NUM_OBJECTS (a);
1054 for (k = 0; k < nobj; k++)
1056 ira_object_t obj = ALLOCNO_OBJECT (a, k);
1058 data->profitable_hard_regs
1059 &= ~OBJECT_TOTAL_CONFLICT_HARD_REGS (obj);
1063 /* Exclude hard regs already assigned for conflicting objects. */
1064 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, i, bi)
1066 a = ira_allocnos[i];
1067 if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS
1068 || ! ALLOCNO_ASSIGNED_P (a)
1069 || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0)
1070 continue;
1071 mode = ALLOCNO_MODE (a);
1072 nregs = hard_regno_nregs (hard_regno, mode);
1073 nobj = ALLOCNO_NUM_OBJECTS (a);
1074 for (k = 0; k < nobj; k++)
1076 ira_object_t obj = ALLOCNO_OBJECT (a, k);
1077 ira_object_t conflict_obj;
1078 ira_object_conflict_iterator oci;
1080 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
1082 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
1084 /* We can process the conflict allocno repeatedly with
1085 the same result. */
1086 if (nregs == nobj && nregs > 1)
1088 int num = OBJECT_SUBWORD (conflict_obj);
1090 if (REG_WORDS_BIG_ENDIAN)
1091 CLEAR_HARD_REG_BIT
1092 (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs,
1093 hard_regno + nobj - num - 1);
1094 else
1095 CLEAR_HARD_REG_BIT
1096 (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs,
1097 hard_regno + num);
1099 else
1100 ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs
1101 &= ~ira_reg_mode_hard_regset[hard_regno][mode];
1105 /* Exclude too costly hard regs. */
1106 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
1108 int min_cost = INT_MAX;
1109 int *costs;
1111 a = ira_allocnos[i];
1112 if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS
1113 || empty_profitable_hard_regs (a))
1114 continue;
1115 data = ALLOCNO_COLOR_DATA (a);
1116 if ((costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a)) != NULL
1117 || (costs = ALLOCNO_HARD_REG_COSTS (a)) != NULL)
1119 class_size = ira_class_hard_regs_num[aclass];
1120 for (j = 0; j < class_size; j++)
1122 hard_regno = ira_class_hard_regs[aclass][j];
1123 if (! TEST_HARD_REG_BIT (data->profitable_hard_regs,
1124 hard_regno))
1125 continue;
1126 if (ALLOCNO_UPDATED_MEMORY_COST (a) < costs[j]
1127 /* Do not remove HARD_REGNO for static chain pointer
1128 pseudo when non-local goto is used. */
1129 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a)))
1130 CLEAR_HARD_REG_BIT (data->profitable_hard_regs,
1131 hard_regno);
1132 else if (min_cost > costs[j])
1133 min_cost = costs[j];
1136 else if (ALLOCNO_UPDATED_MEMORY_COST (a)
1137 < ALLOCNO_UPDATED_CLASS_COST (a)
1138 /* Do not empty profitable regs for static chain
1139 pointer pseudo when non-local goto is used. */
1140 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a)))
1141 CLEAR_HARD_REG_SET (data->profitable_hard_regs);
1142 if (ALLOCNO_UPDATED_CLASS_COST (a) > min_cost)
1143 ALLOCNO_UPDATED_CLASS_COST (a) = min_cost;
1149 /* This page contains functions used to choose hard registers for
1150 allocnos. */
1152 /* Pool for update cost records. */
1153 static object_allocator<update_cost_record> update_cost_record_pool
1154 ("update cost records");
1156 /* Return new update cost record with given params. */
1157 static struct update_cost_record *
1158 get_update_cost_record (int hard_regno, int divisor,
1159 struct update_cost_record *next)
1161 struct update_cost_record *record;
1163 record = update_cost_record_pool.allocate ();
1164 record->hard_regno = hard_regno;
1165 record->divisor = divisor;
1166 record->next = next;
1167 return record;
1170 /* Free memory for all records in LIST. */
1171 static void
1172 free_update_cost_record_list (struct update_cost_record *list)
1174 struct update_cost_record *next;
1176 while (list != NULL)
1178 next = list->next;
1179 update_cost_record_pool.remove (list);
1180 list = next;
1184 /* Free memory allocated for all update cost records. */
1185 static void
1186 finish_update_cost_records (void)
1188 update_cost_record_pool.release ();
1191 /* Array whose element value is TRUE if the corresponding hard
1192 register was already allocated for an allocno. */
1193 static bool allocated_hardreg_p[FIRST_PSEUDO_REGISTER];
1195 /* Describes one element in a queue of allocnos whose costs need to be
1196 updated. Each allocno in the queue is known to have an allocno
1197 class. */
1198 struct update_cost_queue_elem
1200 /* This element is in the queue iff CHECK == update_cost_check. */
1201 int check;
1203 /* COST_HOP_DIVISOR**N, where N is the length of the shortest path
1204 connecting this allocno to the one being allocated. */
1205 int divisor;
1207 /* Allocno from which we started chaining costs of connected
1208 allocnos. */
1209 ira_allocno_t start;
1211 /* Allocno from which we are chaining costs of connected allocnos.
1212 It is used not go back in graph of allocnos connected by
1213 copies. */
1214 ira_allocno_t from;
1216 /* The next allocno in the queue, or null if this is the last element. */
1217 ira_allocno_t next;
1220 /* The first element in a queue of allocnos whose copy costs need to be
1221 updated. Null if the queue is empty. */
1222 static ira_allocno_t update_cost_queue;
1224 /* The last element in the queue described by update_cost_queue.
1225 Not valid if update_cost_queue is null. */
1226 static struct update_cost_queue_elem *update_cost_queue_tail;
1228 /* A pool of elements in the queue described by update_cost_queue.
1229 Elements are indexed by ALLOCNO_NUM. */
1230 static struct update_cost_queue_elem *update_cost_queue_elems;
1232 /* The current value of update_costs_from_copies call count. */
1233 static int update_cost_check;
1235 /* Allocate and initialize data necessary for function
1236 update_costs_from_copies. */
1237 static void
1238 initiate_cost_update (void)
1240 size_t size;
1242 size = ira_allocnos_num * sizeof (struct update_cost_queue_elem);
1243 update_cost_queue_elems
1244 = (struct update_cost_queue_elem *) ira_allocate (size);
1245 memset (update_cost_queue_elems, 0, size);
1246 update_cost_check = 0;
1249 /* Deallocate data used by function update_costs_from_copies. */
1250 static void
1251 finish_cost_update (void)
1253 ira_free (update_cost_queue_elems);
1254 finish_update_cost_records ();
1257 /* When we traverse allocnos to update hard register costs, the cost
1258 divisor will be multiplied by the following macro value for each
1259 hop from given allocno to directly connected allocnos. */
1260 #define COST_HOP_DIVISOR 4
1262 /* Start a new cost-updating pass. */
1263 static void
1264 start_update_cost (void)
1266 update_cost_check++;
1267 update_cost_queue = NULL;
1270 /* Add (ALLOCNO, START, FROM, DIVISOR) to the end of update_cost_queue, unless
1271 ALLOCNO is already in the queue, or has NO_REGS class. */
1272 static inline void
1273 queue_update_cost (ira_allocno_t allocno, ira_allocno_t start,
1274 ira_allocno_t from, int divisor)
1276 struct update_cost_queue_elem *elem;
1278 elem = &update_cost_queue_elems[ALLOCNO_NUM (allocno)];
1279 if (elem->check != update_cost_check
1280 && ALLOCNO_CLASS (allocno) != NO_REGS)
1282 elem->check = update_cost_check;
1283 elem->start = start;
1284 elem->from = from;
1285 elem->divisor = divisor;
1286 elem->next = NULL;
1287 if (update_cost_queue == NULL)
1288 update_cost_queue = allocno;
1289 else
1290 update_cost_queue_tail->next = allocno;
1291 update_cost_queue_tail = elem;
1295 /* Try to remove the first element from update_cost_queue. Return
1296 false if the queue was empty, otherwise make (*ALLOCNO, *START,
1297 *FROM, *DIVISOR) describe the removed element. */
1298 static inline bool
1299 get_next_update_cost (ira_allocno_t *allocno, ira_allocno_t *start,
1300 ira_allocno_t *from, int *divisor)
1302 struct update_cost_queue_elem *elem;
1304 if (update_cost_queue == NULL)
1305 return false;
1307 *allocno = update_cost_queue;
1308 elem = &update_cost_queue_elems[ALLOCNO_NUM (*allocno)];
1309 *start = elem->start;
1310 *from = elem->from;
1311 *divisor = elem->divisor;
1312 update_cost_queue = elem->next;
1313 return true;
1316 /* Increase costs of HARD_REGNO by UPDATE_COST and conflict cost by
1317 UPDATE_CONFLICT_COST for ALLOCNO. Return true if we really
1318 modified the cost. */
1319 static bool
1320 update_allocno_cost (ira_allocno_t allocno, int hard_regno,
1321 int update_cost, int update_conflict_cost)
1323 int i;
1324 enum reg_class aclass = ALLOCNO_CLASS (allocno);
1326 i = ira_class_hard_reg_index[aclass][hard_regno];
1327 if (i < 0)
1328 return false;
1329 ira_allocate_and_set_or_copy_costs
1330 (&ALLOCNO_UPDATED_HARD_REG_COSTS (allocno), aclass,
1331 ALLOCNO_UPDATED_CLASS_COST (allocno),
1332 ALLOCNO_HARD_REG_COSTS (allocno));
1333 ira_allocate_and_set_or_copy_costs
1334 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno),
1335 aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (allocno));
1336 ALLOCNO_UPDATED_HARD_REG_COSTS (allocno)[i] += update_cost;
1337 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno)[i] += update_conflict_cost;
1338 return true;
1341 /* Return TRUE if the object OBJ conflicts with the allocno A. */
1342 static bool
1343 object_conflicts_with_allocno_p (ira_object_t obj, ira_allocno_t a)
1345 if (!OBJECT_CONFLICT_VEC_P (obj))
1346 for (int word = 0; word < ALLOCNO_NUM_OBJECTS (a); word++)
1348 ira_object_t another_obj = ALLOCNO_OBJECT (a, word);
1349 if (OBJECT_CONFLICT_ID (another_obj) >= OBJECT_MIN (obj)
1350 && OBJECT_CONFLICT_ID (another_obj) <= OBJECT_MAX (obj)
1351 && TEST_MINMAX_SET_BIT (OBJECT_CONFLICT_BITVEC (obj),
1352 OBJECT_CONFLICT_ID (another_obj),
1353 OBJECT_MIN (obj), OBJECT_MAX (obj)))
1354 return true;
1356 else
1358 /* If this linear walk ever becomes a bottleneck we could add a
1359 conflict_vec_sorted_p flag and if not set, sort the conflicts after
1360 their ID so we can use a binary search. That would also require
1361 tracking the actual number of conflicts in the vector to not rely
1362 on the NULL termination. */
1363 ira_object_conflict_iterator oci;
1364 ira_object_t conflict_obj;
1365 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
1366 if (OBJECT_ALLOCNO (conflict_obj) == a)
1367 return true;
1369 return false;
1372 /* Return TRUE if allocnos A1 and A2 conflicts. Here we are
1373 interested only in conflicts of allocnos with intersecting allocno
1374 classes. */
1375 static bool
1376 allocnos_conflict_p (ira_allocno_t a1, ira_allocno_t a2)
1378 /* Compute the upper bound for the linear iteration when the object
1379 conflicts are represented as a sparse vector. In particular this
1380 will make sure we prefer O(1) bitvector testing. */
1381 int num_conflicts_in_vec1 = 0, num_conflicts_in_vec2 = 0;
1382 for (int word = 0; word < ALLOCNO_NUM_OBJECTS (a1); ++word)
1383 if (OBJECT_CONFLICT_VEC_P (ALLOCNO_OBJECT (a1, word)))
1384 num_conflicts_in_vec1 += OBJECT_NUM_CONFLICTS (ALLOCNO_OBJECT (a1, word));
1385 for (int word = 0; word < ALLOCNO_NUM_OBJECTS (a2); ++word)
1386 if (OBJECT_CONFLICT_VEC_P (ALLOCNO_OBJECT (a2, word)))
1387 num_conflicts_in_vec2 += OBJECT_NUM_CONFLICTS (ALLOCNO_OBJECT (a2, word));
1388 if (num_conflicts_in_vec2 < num_conflicts_in_vec1)
1389 std::swap (a1, a2);
1391 for (int word = 0; word < ALLOCNO_NUM_OBJECTS (a1); word++)
1393 ira_object_t obj = ALLOCNO_OBJECT (a1, word);
1394 /* Take preferences of conflicting allocnos into account. */
1395 if (object_conflicts_with_allocno_p (obj, a2))
1396 return true;
1398 return false;
1401 /* Update (decrease if DECR_P) HARD_REGNO cost of allocnos connected
1402 by copies to ALLOCNO to increase chances to remove some copies as
1403 the result of subsequent assignment. Update conflict costs.
1404 Record cost updates if RECORD_P is true. */
1405 static void
1406 update_costs_from_allocno (ira_allocno_t allocno, int hard_regno,
1407 int divisor, bool decr_p, bool record_p)
1409 int cost, update_cost, update_conflict_cost;
1410 machine_mode mode;
1411 enum reg_class rclass, aclass;
1412 ira_allocno_t another_allocno, start = allocno, from = NULL;
1413 ira_copy_t cp, next_cp;
1415 rclass = REGNO_REG_CLASS (hard_regno);
1418 mode = ALLOCNO_MODE (allocno);
1419 ira_init_register_move_cost_if_necessary (mode);
1420 for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp)
1422 if (cp->first == allocno)
1424 next_cp = cp->next_first_allocno_copy;
1425 another_allocno = cp->second;
1427 else if (cp->second == allocno)
1429 next_cp = cp->next_second_allocno_copy;
1430 another_allocno = cp->first;
1432 else
1433 gcc_unreachable ();
1435 if (another_allocno == from
1436 || (ALLOCNO_COLOR_DATA (another_allocno) != NULL
1437 && (ALLOCNO_COLOR_DATA (allocno)->first_thread_allocno
1438 != ALLOCNO_COLOR_DATA (another_allocno)->first_thread_allocno)))
1439 continue;
1441 aclass = ALLOCNO_CLASS (another_allocno);
1442 if (! TEST_HARD_REG_BIT (reg_class_contents[aclass],
1443 hard_regno)
1444 || ALLOCNO_ASSIGNED_P (another_allocno))
1445 continue;
1447 /* If we have different modes use the smallest one. It is
1448 a sub-register move. It is hard to predict what LRA
1449 will reload (the pseudo or its sub-register) but LRA
1450 will try to minimize the data movement. Also for some
1451 register classes bigger modes might be invalid,
1452 e.g. DImode for AREG on x86. For such cases the
1453 register move cost will be maximal. */
1454 mode = narrower_subreg_mode (ALLOCNO_MODE (cp->first),
1455 ALLOCNO_MODE (cp->second));
1457 ira_init_register_move_cost_if_necessary (mode);
1459 cost = (cp->second == allocno
1460 ? ira_register_move_cost[mode][rclass][aclass]
1461 : ira_register_move_cost[mode][aclass][rclass]);
1462 if (decr_p)
1463 cost = -cost;
1465 update_cost = cp->freq * cost / divisor;
1466 update_conflict_cost = update_cost;
1468 if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
1469 fprintf (ira_dump_file,
1470 " a%dr%d (hr%d): update cost by %d, conflict cost by %d\n",
1471 ALLOCNO_NUM (another_allocno), ALLOCNO_REGNO (another_allocno),
1472 hard_regno, update_cost, update_conflict_cost);
1473 if (update_cost == 0)
1474 continue;
1476 if (! update_allocno_cost (another_allocno, hard_regno,
1477 update_cost, update_conflict_cost))
1478 continue;
1479 queue_update_cost (another_allocno, start, allocno,
1480 divisor * COST_HOP_DIVISOR);
1481 if (record_p && ALLOCNO_COLOR_DATA (another_allocno) != NULL)
1482 ALLOCNO_COLOR_DATA (another_allocno)->update_cost_records
1483 = get_update_cost_record (hard_regno, divisor,
1484 ALLOCNO_COLOR_DATA (another_allocno)
1485 ->update_cost_records);
1488 while (get_next_update_cost (&allocno, &start, &from, &divisor));
1491 /* Decrease preferred ALLOCNO hard register costs and costs of
1492 allocnos connected to ALLOCNO through copy. */
1493 static void
1494 update_costs_from_prefs (ira_allocno_t allocno)
1496 ira_pref_t pref;
1498 start_update_cost ();
1499 for (pref = ALLOCNO_PREFS (allocno); pref != NULL; pref = pref->next_pref)
1501 if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
1502 fprintf (ira_dump_file, " Start updating from pref of hr%d for a%dr%d:\n",
1503 pref->hard_regno, ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno));
1504 update_costs_from_allocno (allocno, pref->hard_regno,
1505 COST_HOP_DIVISOR, true, true);
1509 /* Update (decrease if DECR_P) the cost of allocnos connected to
1510 ALLOCNO through copies to increase chances to remove some copies as
1511 the result of subsequent assignment. ALLOCNO was just assigned to
1512 a hard register. Record cost updates if RECORD_P is true. */
1513 static void
1514 update_costs_from_copies (ira_allocno_t allocno, bool decr_p, bool record_p)
1516 int hard_regno;
1518 hard_regno = ALLOCNO_HARD_REGNO (allocno);
1519 ira_assert (hard_regno >= 0 && ALLOCNO_CLASS (allocno) != NO_REGS);
1520 start_update_cost ();
1521 if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
1522 fprintf (ira_dump_file, " Start updating from a%dr%d by copies:\n",
1523 ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno));
1524 update_costs_from_allocno (allocno, hard_regno, 1, decr_p, record_p);
1527 /* Update conflict_allocno_hard_prefs of allocnos conflicting with
1528 ALLOCNO. */
1529 static void
1530 update_conflict_allocno_hard_prefs (ira_allocno_t allocno)
1532 int l, nr = ALLOCNO_NUM_OBJECTS (allocno);
1534 for (l = 0; l < nr; l++)
1536 ira_object_t conflict_obj, obj = ALLOCNO_OBJECT (allocno, l);
1537 ira_object_conflict_iterator oci;
1539 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
1541 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
1542 allocno_color_data_t conflict_data = ALLOCNO_COLOR_DATA (conflict_a);
1543 ira_pref_t pref;
1545 if (!(hard_reg_set_intersect_p
1546 (ALLOCNO_COLOR_DATA (allocno)->profitable_hard_regs,
1547 conflict_data->profitable_hard_regs)))
1548 continue;
1549 for (pref = ALLOCNO_PREFS (allocno);
1550 pref != NULL;
1551 pref = pref->next_pref)
1552 conflict_data->conflict_allocno_hard_prefs += pref->freq;
1557 /* Restore costs of allocnos connected to ALLOCNO by copies as it was
1558 before updating costs of these allocnos from given allocno. This
1559 is a wise thing to do as if given allocno did not get an expected
1560 hard reg, using smaller cost of the hard reg for allocnos connected
1561 by copies to given allocno becomes actually misleading. Free all
1562 update cost records for ALLOCNO as we don't need them anymore. */
1563 static void
1564 restore_costs_from_copies (ira_allocno_t allocno)
1566 struct update_cost_record *records, *curr;
1568 if (ALLOCNO_COLOR_DATA (allocno) == NULL)
1569 return;
1570 records = ALLOCNO_COLOR_DATA (allocno)->update_cost_records;
1571 start_update_cost ();
1572 if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
1573 fprintf (ira_dump_file, " Start restoring from a%dr%d:\n",
1574 ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno));
1575 for (curr = records; curr != NULL; curr = curr->next)
1576 update_costs_from_allocno (allocno, curr->hard_regno,
1577 curr->divisor, true, false);
1578 free_update_cost_record_list (records);
1579 ALLOCNO_COLOR_DATA (allocno)->update_cost_records = NULL;
1582 /* This function updates COSTS (decrease if DECR_P) for hard_registers
1583 of ACLASS by conflict costs of the unassigned allocnos
1584 connected by copies with allocnos in update_cost_queue. This
1585 update increases chances to remove some copies. */
1586 static void
1587 update_conflict_hard_regno_costs (int *costs, enum reg_class aclass,
1588 bool decr_p)
1590 int i, cost, class_size, freq, mult, div, divisor;
1591 int index, hard_regno;
1592 int *conflict_costs;
1593 bool cont_p;
1594 enum reg_class another_aclass;
1595 ira_allocno_t allocno, another_allocno, start, from;
1596 ira_copy_t cp, next_cp;
1598 while (get_next_update_cost (&allocno, &start, &from, &divisor))
1599 for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp)
1601 if (cp->first == allocno)
1603 next_cp = cp->next_first_allocno_copy;
1604 another_allocno = cp->second;
1606 else if (cp->second == allocno)
1608 next_cp = cp->next_second_allocno_copy;
1609 another_allocno = cp->first;
1611 else
1612 gcc_unreachable ();
1614 another_aclass = ALLOCNO_CLASS (another_allocno);
1615 if (another_allocno == from
1616 || ALLOCNO_ASSIGNED_P (another_allocno)
1617 || ALLOCNO_COLOR_DATA (another_allocno)->may_be_spilled_p
1618 || ! ira_reg_classes_intersect_p[aclass][another_aclass])
1619 continue;
1620 if (allocnos_conflict_p (another_allocno, start))
1621 continue;
1623 class_size = ira_class_hard_regs_num[another_aclass];
1624 ira_allocate_and_copy_costs
1625 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno),
1626 another_aclass, ALLOCNO_CONFLICT_HARD_REG_COSTS (another_allocno));
1627 conflict_costs
1628 = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno);
1629 if (conflict_costs == NULL)
1630 cont_p = true;
1631 else
1633 mult = cp->freq;
1634 freq = ALLOCNO_FREQ (another_allocno);
1635 if (freq == 0)
1636 freq = 1;
1637 div = freq * divisor;
1638 cont_p = false;
1639 for (i = class_size - 1; i >= 0; i--)
1641 hard_regno = ira_class_hard_regs[another_aclass][i];
1642 ira_assert (hard_regno >= 0);
1643 index = ira_class_hard_reg_index[aclass][hard_regno];
1644 if (index < 0)
1645 continue;
1646 cost = (int) (((int64_t) conflict_costs [i] * mult) / div);
1647 if (cost == 0)
1648 continue;
1649 cont_p = true;
1650 if (decr_p)
1651 cost = -cost;
1652 costs[index] += cost;
1655 /* Probably 5 hops will be enough. */
1656 if (cont_p
1657 && divisor <= (COST_HOP_DIVISOR
1658 * COST_HOP_DIVISOR
1659 * COST_HOP_DIVISOR
1660 * COST_HOP_DIVISOR))
1661 queue_update_cost (another_allocno, start, from, divisor * COST_HOP_DIVISOR);
1665 /* Set up conflicting (through CONFLICT_REGS) for each object of
1666 allocno A and the start allocno profitable regs (through
1667 START_PROFITABLE_REGS). Remember that the start profitable regs
1668 exclude hard regs which cannot hold value of mode of allocno A.
1669 This covers mostly cases when multi-register value should be
1670 aligned. */
1671 static inline void
1672 get_conflict_and_start_profitable_regs (ira_allocno_t a, bool retry_p,
1673 HARD_REG_SET *conflict_regs,
1674 HARD_REG_SET *start_profitable_regs)
1676 int i, nwords;
1677 ira_object_t obj;
1679 nwords = ALLOCNO_NUM_OBJECTS (a);
1680 for (i = 0; i < nwords; i++)
1682 obj = ALLOCNO_OBJECT (a, i);
1683 conflict_regs[i] = OBJECT_TOTAL_CONFLICT_HARD_REGS (obj);
1685 if (retry_p)
1686 *start_profitable_regs
1687 = (reg_class_contents[ALLOCNO_CLASS (a)]
1688 &~ (ira_prohibited_class_mode_regs
1689 [ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]));
1690 else
1691 *start_profitable_regs = ALLOCNO_COLOR_DATA (a)->profitable_hard_regs;
1694 /* Return true if HARD_REGNO is ok for assigning to allocno A with
1695 PROFITABLE_REGS and whose objects have CONFLICT_REGS. */
1696 static inline bool
1697 check_hard_reg_p (ira_allocno_t a, int hard_regno,
1698 HARD_REG_SET *conflict_regs, HARD_REG_SET profitable_regs)
1700 int j, nwords, nregs;
1701 enum reg_class aclass;
1702 machine_mode mode;
1704 aclass = ALLOCNO_CLASS (a);
1705 mode = ALLOCNO_MODE (a);
1706 if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[aclass][mode],
1707 hard_regno))
1708 return false;
1709 /* Checking only profitable hard regs. */
1710 if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno))
1711 return false;
1712 nregs = hard_regno_nregs (hard_regno, mode);
1713 nwords = ALLOCNO_NUM_OBJECTS (a);
1714 for (j = 0; j < nregs; j++)
1716 int k;
1717 int set_to_test_start = 0, set_to_test_end = nwords;
1719 if (nregs == nwords)
1721 if (REG_WORDS_BIG_ENDIAN)
1722 set_to_test_start = nwords - j - 1;
1723 else
1724 set_to_test_start = j;
1725 set_to_test_end = set_to_test_start + 1;
1727 for (k = set_to_test_start; k < set_to_test_end; k++)
1728 if (TEST_HARD_REG_BIT (conflict_regs[k], hard_regno + j))
1729 break;
1730 if (k != set_to_test_end)
1731 break;
1733 return j == nregs;
1736 /* Return number of registers needed to be saved and restored at
1737 function prologue/epilogue if we allocate HARD_REGNO to hold value
1738 of MODE. */
1739 static int
1740 calculate_saved_nregs (int hard_regno, machine_mode mode)
1742 int i;
1743 int nregs = 0;
1745 ira_assert (hard_regno >= 0);
1746 for (i = hard_regno_nregs (hard_regno, mode) - 1; i >= 0; i--)
1747 if (!allocated_hardreg_p[hard_regno + i]
1748 && !crtl->abi->clobbers_full_reg_p (hard_regno + i)
1749 && !LOCAL_REGNO (hard_regno + i))
1750 nregs++;
1751 return nregs;
1754 /* Allocnos A1 and A2 are known to conflict. Check whether, in some loop L
1755 that is either the current loop or a nested subloop, the conflict is of
1756 the following form:
1758 - One allocno (X) is a cap allocno for some non-cap allocno X2.
1760 - X2 belongs to some loop L2.
1762 - The other allocno (Y) is a non-cap allocno.
1764 - Y is an ancestor of some allocno Y2 in L2. (Note that such a Y2
1765 must exist, given that X and Y conflict.)
1767 - Y2 is not referenced in L2 (that is, ALLOCNO_NREFS (Y2) == 0).
1769 - Y can use a different allocation from Y2.
1771 In this case, Y's register is live across L2 but is not used within it,
1772 whereas X's register is used only within L2. The conflict is therefore
1773 only "soft", in that it can easily be avoided by spilling Y2 inside L2
1774 without affecting any insn references.
1776 If the conflict does have this form, return the Y2 that would need to be
1777 spilled in order to allow X and Y (and thus A1 and A2) to use the same
1778 register. Return null otherwise. Returning null is conservatively correct;
1779 any nonnnull return value is an optimization. */
1780 ira_allocno_t
1781 ira_soft_conflict (ira_allocno_t a1, ira_allocno_t a2)
1783 /* Search for the loop L and its associated allocnos X and Y. */
1784 int search_depth = 0;
1785 while (ALLOCNO_CAP_MEMBER (a1) && ALLOCNO_CAP_MEMBER (a2))
1787 a1 = ALLOCNO_CAP_MEMBER (a1);
1788 a2 = ALLOCNO_CAP_MEMBER (a2);
1789 if (search_depth++ > max_soft_conflict_loop_depth)
1790 return nullptr;
1792 /* This must be true if A1 and A2 conflict. */
1793 ira_assert (ALLOCNO_LOOP_TREE_NODE (a1) == ALLOCNO_LOOP_TREE_NODE (a2));
1795 /* Make A1 the cap allocno (X in the comment above) and A2 the
1796 non-cap allocno (Y in the comment above). */
1797 if (ALLOCNO_CAP_MEMBER (a2))
1798 std::swap (a1, a2);
1799 if (!ALLOCNO_CAP_MEMBER (a1))
1800 return nullptr;
1802 /* Search for the real allocno that A1 caps (X2 in the comment above). */
1805 a1 = ALLOCNO_CAP_MEMBER (a1);
1806 if (search_depth++ > max_soft_conflict_loop_depth)
1807 return nullptr;
1809 while (ALLOCNO_CAP_MEMBER (a1));
1811 /* Find the associated allocno for A2 (Y2 in the comment above). */
1812 auto node = ALLOCNO_LOOP_TREE_NODE (a1);
1813 auto local_a2 = node->regno_allocno_map[ALLOCNO_REGNO (a2)];
1815 /* Find the parent of LOCAL_A2/Y2. LOCAL_A2 must be a descendant of A2
1816 for the conflict query to make sense, so this parent lookup must succeed.
1818 If the parent allocno has no references, it is usually cheaper to
1819 spill at that loop level instead. Keep searching until we find
1820 a parent allocno that does have references (but don't look past
1821 the starting allocno). */
1822 ira_allocno_t local_parent_a2;
1823 for (;;)
1825 local_parent_a2 = ira_parent_allocno (local_a2);
1826 if (local_parent_a2 == a2 || ALLOCNO_NREFS (local_parent_a2) != 0)
1827 break;
1828 local_a2 = local_parent_a2;
1830 if (CHECKING_P)
1832 /* Sanity check to make sure that the conflict we've been given
1833 makes sense. */
1834 auto test_a2 = local_parent_a2;
1835 while (test_a2 != a2)
1837 test_a2 = ira_parent_allocno (test_a2);
1838 ira_assert (test_a2);
1841 if (local_a2
1842 && ALLOCNO_NREFS (local_a2) == 0
1843 && ira_subloop_allocnos_can_differ_p (local_parent_a2))
1844 return local_a2;
1845 return nullptr;
1848 /* The caller has decided to allocate HREGNO to A and has proved that
1849 this is safe. However, the allocation might require the kind of
1850 spilling described in the comment above ira_soft_conflict.
1851 The caller has recorded that:
1853 - The allocnos in ALLOCNOS_TO_SPILL are the ones that would need
1854 to be spilled to satisfy soft conflicts for at least one allocation
1855 (not necessarily HREGNO).
1857 - The soft conflicts apply only to A allocations that overlap
1858 SOFT_CONFLICT_REGS.
1860 If allocating HREGNO is subject to any soft conflicts, record the
1861 subloop allocnos that need to be spilled. */
1862 static void
1863 spill_soft_conflicts (ira_allocno_t a, bitmap allocnos_to_spill,
1864 HARD_REG_SET soft_conflict_regs, int hregno)
1866 auto nregs = hard_regno_nregs (hregno, ALLOCNO_MODE (a));
1867 bitmap_iterator bi;
1868 unsigned int i;
1869 EXECUTE_IF_SET_IN_BITMAP (allocnos_to_spill, 0, i, bi)
1871 /* SPILL_A needs to be spilled for at least one allocation
1872 (not necessarily this one). */
1873 auto spill_a = ira_allocnos[i];
1875 /* Find the corresponding allocno for this loop. */
1876 auto conflict_a = spill_a;
1879 conflict_a = ira_parent_or_cap_allocno (conflict_a);
1880 ira_assert (conflict_a);
1882 while (ALLOCNO_LOOP_TREE_NODE (conflict_a)->level
1883 > ALLOCNO_LOOP_TREE_NODE (a)->level);
1885 ira_assert (ALLOCNO_LOOP_TREE_NODE (conflict_a)
1886 == ALLOCNO_LOOP_TREE_NODE (a));
1888 if (conflict_a == a)
1890 /* SPILL_A is a descendant of A. We don't know (and don't need
1891 to know) which cap allocnos have a soft conflict with A.
1892 All we need to do is test whether the soft conflict applies
1893 to the chosen allocation. */
1894 if (ira_hard_reg_set_intersection_p (hregno, ALLOCNO_MODE (a),
1895 soft_conflict_regs))
1896 ALLOCNO_MIGHT_CONFLICT_WITH_PARENT_P (spill_a) = true;
1898 else
1900 /* SPILL_A is a descendant of CONFLICT_A, which has a soft conflict
1901 with A. Test whether the soft conflict applies to the current
1902 allocation. */
1903 ira_assert (ira_soft_conflict (a, conflict_a) == spill_a);
1904 auto conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a);
1905 ira_assert (conflict_hregno >= 0);
1906 auto conflict_nregs = hard_regno_nregs (conflict_hregno,
1907 ALLOCNO_MODE (conflict_a));
1908 if (hregno + nregs > conflict_hregno
1909 && conflict_hregno + conflict_nregs > hregno)
1910 ALLOCNO_MIGHT_CONFLICT_WITH_PARENT_P (spill_a) = true;
1915 /* Choose a hard register for allocno A. If RETRY_P is TRUE, it means
1916 that the function called from function
1917 `ira_reassign_conflict_allocnos' and `allocno_reload_assign'. In
1918 this case some allocno data are not defined or updated and we
1919 should not touch these data. The function returns true if we
1920 managed to assign a hard register to the allocno.
1922 To assign a hard register, first of all we calculate all conflict
1923 hard registers which can come from conflicting allocnos with
1924 already assigned hard registers. After that we find first free
1925 hard register with the minimal cost. During hard register cost
1926 calculation we take conflict hard register costs into account to
1927 give a chance for conflicting allocnos to get a better hard
1928 register in the future.
1930 If the best hard register cost is bigger than cost of memory usage
1931 for the allocno, we don't assign a hard register to given allocno
1932 at all.
1934 If we assign a hard register to the allocno, we update costs of the
1935 hard register for allocnos connected by copies to improve a chance
1936 to coalesce insns represented by the copies when we assign hard
1937 registers to the allocnos connected by the copies. */
1938 static bool
1939 assign_hard_reg (ira_allocno_t a, bool retry_p)
1941 HARD_REG_SET conflicting_regs[2], profitable_hard_regs;
1942 int i, j, hard_regno, best_hard_regno, class_size;
1943 int cost, mem_cost, min_cost, full_cost, min_full_cost, nwords, word;
1944 int *a_costs;
1945 enum reg_class aclass;
1946 machine_mode mode;
1947 static int costs[FIRST_PSEUDO_REGISTER], full_costs[FIRST_PSEUDO_REGISTER];
1948 int saved_nregs;
1949 enum reg_class rclass;
1950 int add_cost;
1951 #ifdef STACK_REGS
1952 bool no_stack_reg_p;
1953 #endif
1954 auto_bitmap allocnos_to_spill;
1955 HARD_REG_SET soft_conflict_regs = {};
1957 ira_assert (! ALLOCNO_ASSIGNED_P (a));
1958 get_conflict_and_start_profitable_regs (a, retry_p,
1959 conflicting_regs,
1960 &profitable_hard_regs);
1961 aclass = ALLOCNO_CLASS (a);
1962 class_size = ira_class_hard_regs_num[aclass];
1963 best_hard_regno = -1;
1964 mem_cost = 0;
1965 memset (costs, 0, sizeof (int) * class_size);
1966 memset (full_costs, 0, sizeof (int) * class_size);
1967 #ifdef STACK_REGS
1968 no_stack_reg_p = false;
1969 #endif
1970 if (! retry_p)
1971 start_update_cost ();
1972 mem_cost += ALLOCNO_UPDATED_MEMORY_COST (a);
1974 ira_allocate_and_copy_costs (&ALLOCNO_UPDATED_HARD_REG_COSTS (a),
1975 aclass, ALLOCNO_HARD_REG_COSTS (a));
1976 a_costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a);
1977 #ifdef STACK_REGS
1978 no_stack_reg_p = no_stack_reg_p || ALLOCNO_TOTAL_NO_STACK_REG_P (a);
1979 #endif
1980 cost = ALLOCNO_UPDATED_CLASS_COST (a);
1981 for (i = 0; i < class_size; i++)
1982 if (a_costs != NULL)
1984 costs[i] += a_costs[i];
1985 full_costs[i] += a_costs[i];
1987 else
1989 costs[i] += cost;
1990 full_costs[i] += cost;
1992 nwords = ALLOCNO_NUM_OBJECTS (a);
1993 curr_allocno_process++;
1994 for (word = 0; word < nwords; word++)
1996 ira_object_t conflict_obj;
1997 ira_object_t obj = ALLOCNO_OBJECT (a, word);
1998 ira_object_conflict_iterator oci;
2000 /* Take preferences of conflicting allocnos into account. */
2001 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
2003 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
2004 enum reg_class conflict_aclass;
2005 allocno_color_data_t data = ALLOCNO_COLOR_DATA (conflict_a);
2007 /* Reload can give another class so we need to check all
2008 allocnos. */
2009 if (!retry_p
2010 && ((!ALLOCNO_ASSIGNED_P (conflict_a)
2011 || ALLOCNO_HARD_REGNO (conflict_a) < 0)
2012 && !(hard_reg_set_intersect_p
2013 (profitable_hard_regs,
2014 ALLOCNO_COLOR_DATA
2015 (conflict_a)->profitable_hard_regs))))
2017 /* All conflict allocnos are in consideration bitmap
2018 when retry_p is false. It might change in future and
2019 if it happens the assert will be broken. It means
2020 the code should be modified for the new
2021 assumptions. */
2022 ira_assert (bitmap_bit_p (consideration_allocno_bitmap,
2023 ALLOCNO_NUM (conflict_a)));
2024 continue;
2026 conflict_aclass = ALLOCNO_CLASS (conflict_a);
2027 ira_assert (ira_reg_classes_intersect_p
2028 [aclass][conflict_aclass]);
2029 if (ALLOCNO_ASSIGNED_P (conflict_a))
2031 hard_regno = ALLOCNO_HARD_REGNO (conflict_a);
2032 if (hard_regno >= 0
2033 && (ira_hard_reg_set_intersection_p
2034 (hard_regno, ALLOCNO_MODE (conflict_a),
2035 reg_class_contents[aclass])))
2037 int n_objects = ALLOCNO_NUM_OBJECTS (conflict_a);
2038 int conflict_nregs;
2040 mode = ALLOCNO_MODE (conflict_a);
2041 conflict_nregs = hard_regno_nregs (hard_regno, mode);
2042 auto spill_a = (retry_p
2043 ? nullptr
2044 : ira_soft_conflict (a, conflict_a));
2045 if (spill_a)
2047 if (bitmap_set_bit (allocnos_to_spill,
2048 ALLOCNO_NUM (spill_a)))
2050 ira_loop_border_costs border_costs (spill_a);
2051 auto cost = border_costs.spill_inside_loop_cost ();
2052 auto note_conflict = [&](int r)
2054 SET_HARD_REG_BIT (soft_conflict_regs, r);
2055 auto hri = ira_class_hard_reg_index[aclass][r];
2056 if (hri >= 0)
2058 costs[hri] += cost;
2059 full_costs[hri] += cost;
2062 for (int r = hard_regno;
2063 r >= 0 && (int) end_hard_regno (mode, r) > hard_regno;
2064 r--)
2065 note_conflict (r);
2066 for (int r = hard_regno + 1;
2067 r < hard_regno + conflict_nregs;
2068 r++)
2069 note_conflict (r);
2072 else
2074 if (conflict_nregs == n_objects && conflict_nregs > 1)
2076 int num = OBJECT_SUBWORD (conflict_obj);
2078 if (REG_WORDS_BIG_ENDIAN)
2079 SET_HARD_REG_BIT (conflicting_regs[word],
2080 hard_regno + n_objects - num - 1);
2081 else
2082 SET_HARD_REG_BIT (conflicting_regs[word],
2083 hard_regno + num);
2085 else
2086 conflicting_regs[word]
2087 |= ira_reg_mode_hard_regset[hard_regno][mode];
2088 if (hard_reg_set_subset_p (profitable_hard_regs,
2089 conflicting_regs[word]))
2090 goto fail;
2094 else if (! retry_p
2095 && ! ALLOCNO_COLOR_DATA (conflict_a)->may_be_spilled_p
2096 /* Don't process the conflict allocno twice. */
2097 && (ALLOCNO_COLOR_DATA (conflict_a)->last_process
2098 != curr_allocno_process))
2100 int k, *conflict_costs;
2102 ALLOCNO_COLOR_DATA (conflict_a)->last_process
2103 = curr_allocno_process;
2104 ira_allocate_and_copy_costs
2105 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a),
2106 conflict_aclass,
2107 ALLOCNO_CONFLICT_HARD_REG_COSTS (conflict_a));
2108 conflict_costs
2109 = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a);
2110 if (conflict_costs != NULL)
2111 for (j = class_size - 1; j >= 0; j--)
2113 hard_regno = ira_class_hard_regs[aclass][j];
2114 ira_assert (hard_regno >= 0);
2115 k = ira_class_hard_reg_index[conflict_aclass][hard_regno];
2116 if (k < 0
2117 /* If HARD_REGNO is not available for CONFLICT_A,
2118 the conflict would be ignored, since HARD_REGNO
2119 will never be assigned to CONFLICT_A. */
2120 || !TEST_HARD_REG_BIT (data->profitable_hard_regs,
2121 hard_regno))
2122 continue;
2123 full_costs[j] -= conflict_costs[k];
2125 queue_update_cost (conflict_a, conflict_a, NULL, COST_HOP_DIVISOR);
2129 if (! retry_p)
2130 /* Take into account preferences of allocnos connected by copies to
2131 the conflict allocnos. */
2132 update_conflict_hard_regno_costs (full_costs, aclass, true);
2134 /* Take preferences of allocnos connected by copies into
2135 account. */
2136 if (! retry_p)
2138 start_update_cost ();
2139 queue_update_cost (a, a, NULL, COST_HOP_DIVISOR);
2140 update_conflict_hard_regno_costs (full_costs, aclass, false);
2142 min_cost = min_full_cost = INT_MAX;
2143 /* We don't care about giving callee saved registers to allocnos no
2144 living through calls because call clobbered registers are
2145 allocated first (it is usual practice to put them first in
2146 REG_ALLOC_ORDER). */
2147 mode = ALLOCNO_MODE (a);
2148 for (i = 0; i < class_size; i++)
2150 hard_regno = ira_class_hard_regs[aclass][i];
2151 #ifdef STACK_REGS
2152 if (no_stack_reg_p
2153 && FIRST_STACK_REG <= hard_regno && hard_regno <= LAST_STACK_REG)
2154 continue;
2155 #endif
2156 if (! check_hard_reg_p (a, hard_regno,
2157 conflicting_regs, profitable_hard_regs))
2158 continue;
2159 cost = costs[i];
2160 full_cost = full_costs[i];
2161 if (!HONOR_REG_ALLOC_ORDER)
2163 if ((saved_nregs = calculate_saved_nregs (hard_regno, mode)) != 0)
2164 /* We need to save/restore the hard register in
2165 epilogue/prologue. Therefore we increase the cost. */
2167 rclass = REGNO_REG_CLASS (hard_regno);
2168 add_cost = ((ira_memory_move_cost[mode][rclass][0]
2169 + ira_memory_move_cost[mode][rclass][1])
2170 * saved_nregs / hard_regno_nregs (hard_regno,
2171 mode) - 1);
2172 cost += add_cost;
2173 full_cost += add_cost;
2176 if (min_cost > cost)
2177 min_cost = cost;
2178 if (min_full_cost > full_cost)
2180 min_full_cost = full_cost;
2181 best_hard_regno = hard_regno;
2182 ira_assert (hard_regno >= 0);
2184 if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
2185 fprintf (ira_dump_file, "(%d=%d,%d) ", hard_regno, cost, full_cost);
2187 if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL)
2188 fprintf (ira_dump_file, "\n");
2189 if (min_full_cost > mem_cost
2190 /* Do not spill static chain pointer pseudo when non-local goto
2191 is used. */
2192 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a)))
2194 if (! retry_p && internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2195 fprintf (ira_dump_file, "(memory is more profitable %d vs %d) ",
2196 mem_cost, min_full_cost);
2197 best_hard_regno = -1;
2199 fail:
2200 if (best_hard_regno >= 0)
2202 for (i = hard_regno_nregs (best_hard_regno, mode) - 1; i >= 0; i--)
2203 allocated_hardreg_p[best_hard_regno + i] = true;
2204 spill_soft_conflicts (a, allocnos_to_spill, soft_conflict_regs,
2205 best_hard_regno);
2207 if (! retry_p)
2208 restore_costs_from_copies (a);
2209 ALLOCNO_HARD_REGNO (a) = best_hard_regno;
2210 ALLOCNO_ASSIGNED_P (a) = true;
2211 if (best_hard_regno >= 0 && !retry_p)
2212 update_costs_from_copies (a, true, true);
2213 ira_assert (ALLOCNO_CLASS (a) == aclass);
2214 /* We don't need updated costs anymore. */
2215 ira_free_allocno_updated_costs (a);
2216 return best_hard_regno >= 0;
2221 /* An array used to sort copies. */
2222 static ira_copy_t *sorted_copies;
2224 /* If allocno A is a cap, return non-cap allocno from which A is
2225 created. Otherwise, return A. */
2226 static ira_allocno_t
2227 get_cap_member (ira_allocno_t a)
2229 ira_allocno_t member;
2231 while ((member = ALLOCNO_CAP_MEMBER (a)) != NULL)
2232 a = member;
2233 return a;
2236 /* Return TRUE if live ranges of allocnos A1 and A2 intersect. It is
2237 used to find a conflict for new allocnos or allocnos with the
2238 different allocno classes. */
2239 static bool
2240 allocnos_conflict_by_live_ranges_p (ira_allocno_t a1, ira_allocno_t a2)
2242 rtx reg1, reg2;
2243 int i, j;
2244 int n1 = ALLOCNO_NUM_OBJECTS (a1);
2245 int n2 = ALLOCNO_NUM_OBJECTS (a2);
2247 if (a1 == a2)
2248 return false;
2249 reg1 = regno_reg_rtx[ALLOCNO_REGNO (a1)];
2250 reg2 = regno_reg_rtx[ALLOCNO_REGNO (a2)];
2251 if (reg1 != NULL && reg2 != NULL
2252 && ORIGINAL_REGNO (reg1) == ORIGINAL_REGNO (reg2))
2253 return false;
2255 /* We don't keep live ranges for caps because they can be quite big.
2256 Use ranges of non-cap allocno from which caps are created. */
2257 a1 = get_cap_member (a1);
2258 a2 = get_cap_member (a2);
2259 for (i = 0; i < n1; i++)
2261 ira_object_t c1 = ALLOCNO_OBJECT (a1, i);
2263 for (j = 0; j < n2; j++)
2265 ira_object_t c2 = ALLOCNO_OBJECT (a2, j);
2267 if (ira_live_ranges_intersect_p (OBJECT_LIVE_RANGES (c1),
2268 OBJECT_LIVE_RANGES (c2)))
2269 return true;
2272 return false;
2275 /* The function is used to sort copies according to their execution
2276 frequencies. */
2277 static int
2278 copy_freq_compare_func (const void *v1p, const void *v2p)
2280 ira_copy_t cp1 = *(const ira_copy_t *) v1p, cp2 = *(const ira_copy_t *) v2p;
2281 int pri1, pri2;
2283 pri1 = cp1->freq;
2284 pri2 = cp2->freq;
2285 if (pri2 - pri1)
2286 return pri2 - pri1;
2288 /* If frequencies are equal, sort by copies, so that the results of
2289 qsort leave nothing to chance. */
2290 return cp1->num - cp2->num;
2295 /* Return true if any allocno from thread of A1 conflicts with any
2296 allocno from thread A2. */
2297 static bool
2298 allocno_thread_conflict_p (ira_allocno_t a1, ira_allocno_t a2)
2300 ira_allocno_t a, conflict_a;
2302 for (a = ALLOCNO_COLOR_DATA (a2)->next_thread_allocno;;
2303 a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno)
2305 for (conflict_a = ALLOCNO_COLOR_DATA (a1)->next_thread_allocno;;
2306 conflict_a = ALLOCNO_COLOR_DATA (conflict_a)->next_thread_allocno)
2308 if (allocnos_conflict_by_live_ranges_p (a, conflict_a))
2309 return true;
2310 if (conflict_a == a1)
2311 break;
2313 if (a == a2)
2314 break;
2316 return false;
2319 /* Merge two threads given correspondingly by their first allocnos T1
2320 and T2 (more accurately merging T2 into T1). */
2321 static void
2322 merge_threads (ira_allocno_t t1, ira_allocno_t t2)
2324 ira_allocno_t a, next, last;
2326 gcc_assert (t1 != t2
2327 && ALLOCNO_COLOR_DATA (t1)->first_thread_allocno == t1
2328 && ALLOCNO_COLOR_DATA (t2)->first_thread_allocno == t2);
2329 for (last = t2, a = ALLOCNO_COLOR_DATA (t2)->next_thread_allocno;;
2330 a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno)
2332 ALLOCNO_COLOR_DATA (a)->first_thread_allocno = t1;
2333 if (a == t2)
2334 break;
2335 last = a;
2337 next = ALLOCNO_COLOR_DATA (t1)->next_thread_allocno;
2338 ALLOCNO_COLOR_DATA (t1)->next_thread_allocno = t2;
2339 ALLOCNO_COLOR_DATA (last)->next_thread_allocno = next;
2340 ALLOCNO_COLOR_DATA (t1)->thread_freq += ALLOCNO_COLOR_DATA (t2)->thread_freq;
2343 /* Create threads by processing CP_NUM copies from sorted copies. We
2344 process the most expensive copies first. */
2345 static void
2346 form_threads_from_copies (int cp_num)
2348 ira_allocno_t a, thread1, thread2;
2349 ira_copy_t cp;
2351 qsort (sorted_copies, cp_num, sizeof (ira_copy_t), copy_freq_compare_func);
2352 /* Form threads processing copies, most frequently executed
2353 first. */
2354 for (int i = 0; i < cp_num; i++)
2356 cp = sorted_copies[i];
2357 thread1 = ALLOCNO_COLOR_DATA (cp->first)->first_thread_allocno;
2358 thread2 = ALLOCNO_COLOR_DATA (cp->second)->first_thread_allocno;
2359 if (thread1 == thread2)
2360 continue;
2361 if (! allocno_thread_conflict_p (thread1, thread2))
2363 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2364 fprintf
2365 (ira_dump_file,
2366 " Forming thread by copy %d:a%dr%d-a%dr%d (freq=%d):\n",
2367 cp->num, ALLOCNO_NUM (cp->first), ALLOCNO_REGNO (cp->first),
2368 ALLOCNO_NUM (cp->second), ALLOCNO_REGNO (cp->second),
2369 cp->freq);
2370 merge_threads (thread1, thread2);
2371 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2373 thread1 = ALLOCNO_COLOR_DATA (thread1)->first_thread_allocno;
2374 fprintf (ira_dump_file, " Result (freq=%d): a%dr%d(%d)",
2375 ALLOCNO_COLOR_DATA (thread1)->thread_freq,
2376 ALLOCNO_NUM (thread1), ALLOCNO_REGNO (thread1),
2377 ALLOCNO_FREQ (thread1));
2378 for (a = ALLOCNO_COLOR_DATA (thread1)->next_thread_allocno;
2379 a != thread1;
2380 a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno)
2381 fprintf (ira_dump_file, " a%dr%d(%d)",
2382 ALLOCNO_NUM (a), ALLOCNO_REGNO (a),
2383 ALLOCNO_FREQ (a));
2384 fprintf (ira_dump_file, "\n");
2390 /* Create threads by processing copies of all alocnos from BUCKET. We
2391 process the most expensive copies first. */
2392 static void
2393 form_threads_from_bucket (ira_allocno_t bucket)
2395 ira_allocno_t a;
2396 ira_copy_t cp, next_cp;
2397 int cp_num = 0;
2399 for (a = bucket; a != NULL; a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2401 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
2403 if (cp->first == a)
2405 next_cp = cp->next_first_allocno_copy;
2406 sorted_copies[cp_num++] = cp;
2408 else if (cp->second == a)
2409 next_cp = cp->next_second_allocno_copy;
2410 else
2411 gcc_unreachable ();
2414 form_threads_from_copies (cp_num);
2417 /* Create threads by processing copies of colorable allocno A. We
2418 process most expensive copies first. */
2419 static void
2420 form_threads_from_colorable_allocno (ira_allocno_t a)
2422 ira_allocno_t another_a;
2423 ira_copy_t cp, next_cp;
2424 int cp_num = 0;
2426 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2427 fprintf (ira_dump_file, " Forming thread from allocno a%dr%d:\n",
2428 ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
2429 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
2431 if (cp->first == a)
2433 next_cp = cp->next_first_allocno_copy;
2434 another_a = cp->second;
2436 else if (cp->second == a)
2438 next_cp = cp->next_second_allocno_copy;
2439 another_a = cp->first;
2441 else
2442 gcc_unreachable ();
2443 if ((! ALLOCNO_COLOR_DATA (another_a)->in_graph_p
2444 && !ALLOCNO_COLOR_DATA (another_a)->may_be_spilled_p)
2445 || ALLOCNO_COLOR_DATA (another_a)->colorable_p)
2446 sorted_copies[cp_num++] = cp;
2448 form_threads_from_copies (cp_num);
2451 /* Form initial threads which contain only one allocno. */
2452 static void
2453 init_allocno_threads (void)
2455 ira_allocno_t a;
2456 unsigned int j;
2457 bitmap_iterator bi;
2458 ira_pref_t pref;
2460 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
2462 a = ira_allocnos[j];
2463 /* Set up initial thread data: */
2464 ALLOCNO_COLOR_DATA (a)->first_thread_allocno
2465 = ALLOCNO_COLOR_DATA (a)->next_thread_allocno = a;
2466 ALLOCNO_COLOR_DATA (a)->thread_freq = ALLOCNO_FREQ (a);
2467 ALLOCNO_COLOR_DATA (a)->hard_reg_prefs = 0;
2468 for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = pref->next_pref)
2469 ALLOCNO_COLOR_DATA (a)->hard_reg_prefs += pref->freq;
2475 /* This page contains the allocator based on the Chaitin-Briggs algorithm. */
2477 /* Bucket of allocnos that can colored currently without spilling. */
2478 static ira_allocno_t colorable_allocno_bucket;
2480 /* Bucket of allocnos that might be not colored currently without
2481 spilling. */
2482 static ira_allocno_t uncolorable_allocno_bucket;
2484 /* The current number of allocnos in the uncolorable_bucket. */
2485 static int uncolorable_allocnos_num;
2487 /* Return the current spill priority of allocno A. The less the
2488 number, the more preferable the allocno for spilling. */
2489 static inline int
2490 allocno_spill_priority (ira_allocno_t a)
2492 allocno_color_data_t data = ALLOCNO_COLOR_DATA (a);
2494 return (data->temp
2495 / (ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a)
2496 * ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]
2497 + 1));
2500 /* Add allocno A to bucket *BUCKET_PTR. A should be not in a bucket
2501 before the call. */
2502 static void
2503 add_allocno_to_bucket (ira_allocno_t a, ira_allocno_t *bucket_ptr)
2505 ira_allocno_t first_a;
2506 allocno_color_data_t data;
2508 if (bucket_ptr == &uncolorable_allocno_bucket
2509 && ALLOCNO_CLASS (a) != NO_REGS)
2511 uncolorable_allocnos_num++;
2512 ira_assert (uncolorable_allocnos_num > 0);
2514 first_a = *bucket_ptr;
2515 data = ALLOCNO_COLOR_DATA (a);
2516 data->next_bucket_allocno = first_a;
2517 data->prev_bucket_allocno = NULL;
2518 if (first_a != NULL)
2519 ALLOCNO_COLOR_DATA (first_a)->prev_bucket_allocno = a;
2520 *bucket_ptr = a;
2523 /* Compare two allocnos to define which allocno should be pushed first
2524 into the coloring stack. If the return is a negative number, the
2525 allocno given by the first parameter will be pushed first. In this
2526 case such allocno has less priority than the second one and the
2527 hard register will be assigned to it after assignment to the second
2528 one. As the result of such assignment order, the second allocno
2529 has a better chance to get the best hard register. */
2530 static int
2531 bucket_allocno_compare_func (const void *v1p, const void *v2p)
2533 ira_allocno_t a1 = *(const ira_allocno_t *) v1p;
2534 ira_allocno_t a2 = *(const ira_allocno_t *) v2p;
2535 int diff, freq1, freq2, a1_num, a2_num, pref1, pref2;
2536 ira_allocno_t t1 = ALLOCNO_COLOR_DATA (a1)->first_thread_allocno;
2537 ira_allocno_t t2 = ALLOCNO_COLOR_DATA (a2)->first_thread_allocno;
2538 int cl1 = ALLOCNO_CLASS (a1), cl2 = ALLOCNO_CLASS (a2);
2540 freq1 = ALLOCNO_COLOR_DATA (t1)->thread_freq;
2541 freq2 = ALLOCNO_COLOR_DATA (t2)->thread_freq;
2542 if ((diff = freq1 - freq2) != 0)
2543 return diff;
2545 if ((diff = ALLOCNO_NUM (t2) - ALLOCNO_NUM (t1)) != 0)
2546 return diff;
2548 /* Push pseudos requiring less hard registers first. It means that
2549 we will assign pseudos requiring more hard registers first
2550 avoiding creation small holes in free hard register file into
2551 which the pseudos requiring more hard registers cannot fit. */
2552 if ((diff = (ira_reg_class_max_nregs[cl1][ALLOCNO_MODE (a1)]
2553 - ira_reg_class_max_nregs[cl2][ALLOCNO_MODE (a2)])) != 0)
2554 return diff;
2556 freq1 = ALLOCNO_FREQ (a1);
2557 freq2 = ALLOCNO_FREQ (a2);
2558 if ((diff = freq1 - freq2) != 0)
2559 return diff;
2561 a1_num = ALLOCNO_COLOR_DATA (a1)->available_regs_num;
2562 a2_num = ALLOCNO_COLOR_DATA (a2)->available_regs_num;
2563 if ((diff = a2_num - a1_num) != 0)
2564 return diff;
2565 /* Push allocnos with minimal conflict_allocno_hard_prefs first. */
2566 pref1 = ALLOCNO_COLOR_DATA (a1)->conflict_allocno_hard_prefs;
2567 pref2 = ALLOCNO_COLOR_DATA (a2)->conflict_allocno_hard_prefs;
2568 if ((diff = pref1 - pref2) != 0)
2569 return diff;
2570 return ALLOCNO_NUM (a2) - ALLOCNO_NUM (a1);
2573 /* Sort bucket *BUCKET_PTR and return the result through
2574 BUCKET_PTR. */
2575 static void
2576 sort_bucket (ira_allocno_t *bucket_ptr,
2577 int (*compare_func) (const void *, const void *))
2579 ira_allocno_t a, head;
2580 int n;
2582 for (n = 0, a = *bucket_ptr;
2583 a != NULL;
2584 a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2585 sorted_allocnos[n++] = a;
2586 if (n <= 1)
2587 return;
2588 qsort (sorted_allocnos, n, sizeof (ira_allocno_t), compare_func);
2589 head = NULL;
2590 for (n--; n >= 0; n--)
2592 a = sorted_allocnos[n];
2593 ALLOCNO_COLOR_DATA (a)->next_bucket_allocno = head;
2594 ALLOCNO_COLOR_DATA (a)->prev_bucket_allocno = NULL;
2595 if (head != NULL)
2596 ALLOCNO_COLOR_DATA (head)->prev_bucket_allocno = a;
2597 head = a;
2599 *bucket_ptr = head;
2602 /* Add ALLOCNO to colorable bucket maintaining the order according
2603 their priority. ALLOCNO should be not in a bucket before the
2604 call. */
2605 static void
2606 add_allocno_to_ordered_colorable_bucket (ira_allocno_t allocno)
2608 ira_allocno_t before, after;
2610 form_threads_from_colorable_allocno (allocno);
2611 for (before = colorable_allocno_bucket, after = NULL;
2612 before != NULL;
2613 after = before,
2614 before = ALLOCNO_COLOR_DATA (before)->next_bucket_allocno)
2615 if (bucket_allocno_compare_func (&allocno, &before) < 0)
2616 break;
2617 ALLOCNO_COLOR_DATA (allocno)->next_bucket_allocno = before;
2618 ALLOCNO_COLOR_DATA (allocno)->prev_bucket_allocno = after;
2619 if (after == NULL)
2620 colorable_allocno_bucket = allocno;
2621 else
2622 ALLOCNO_COLOR_DATA (after)->next_bucket_allocno = allocno;
2623 if (before != NULL)
2624 ALLOCNO_COLOR_DATA (before)->prev_bucket_allocno = allocno;
2627 /* Delete ALLOCNO from bucket *BUCKET_PTR. It should be there before
2628 the call. */
2629 static void
2630 delete_allocno_from_bucket (ira_allocno_t allocno, ira_allocno_t *bucket_ptr)
2632 ira_allocno_t prev_allocno, next_allocno;
2634 if (bucket_ptr == &uncolorable_allocno_bucket
2635 && ALLOCNO_CLASS (allocno) != NO_REGS)
2637 uncolorable_allocnos_num--;
2638 ira_assert (uncolorable_allocnos_num >= 0);
2640 prev_allocno = ALLOCNO_COLOR_DATA (allocno)->prev_bucket_allocno;
2641 next_allocno = ALLOCNO_COLOR_DATA (allocno)->next_bucket_allocno;
2642 if (prev_allocno != NULL)
2643 ALLOCNO_COLOR_DATA (prev_allocno)->next_bucket_allocno = next_allocno;
2644 else
2646 ira_assert (*bucket_ptr == allocno);
2647 *bucket_ptr = next_allocno;
2649 if (next_allocno != NULL)
2650 ALLOCNO_COLOR_DATA (next_allocno)->prev_bucket_allocno = prev_allocno;
2653 /* Put allocno A onto the coloring stack without removing it from its
2654 bucket. Pushing allocno to the coloring stack can result in moving
2655 conflicting allocnos from the uncolorable bucket to the colorable
2656 one. Update conflict_allocno_hard_prefs of the conflicting
2657 allocnos which are not on stack yet. */
2658 static void
2659 push_allocno_to_stack (ira_allocno_t a)
2661 enum reg_class aclass;
2662 allocno_color_data_t data, conflict_data;
2663 int size, i, n = ALLOCNO_NUM_OBJECTS (a);
2665 data = ALLOCNO_COLOR_DATA (a);
2666 data->in_graph_p = false;
2667 allocno_stack_vec.safe_push (a);
2668 aclass = ALLOCNO_CLASS (a);
2669 if (aclass == NO_REGS)
2670 return;
2671 size = ira_reg_class_max_nregs[aclass][ALLOCNO_MODE (a)];
2672 if (n > 1)
2674 /* We will deal with the subwords individually. */
2675 gcc_assert (size == ALLOCNO_NUM_OBJECTS (a));
2676 size = 1;
2678 for (i = 0; i < n; i++)
2680 ira_object_t obj = ALLOCNO_OBJECT (a, i);
2681 ira_object_t conflict_obj;
2682 ira_object_conflict_iterator oci;
2684 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
2686 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
2687 ira_pref_t pref;
2689 conflict_data = ALLOCNO_COLOR_DATA (conflict_a);
2690 if (! conflict_data->in_graph_p
2691 || ALLOCNO_ASSIGNED_P (conflict_a)
2692 || !(hard_reg_set_intersect_p
2693 (ALLOCNO_COLOR_DATA (a)->profitable_hard_regs,
2694 conflict_data->profitable_hard_regs)))
2695 continue;
2696 for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = pref->next_pref)
2697 conflict_data->conflict_allocno_hard_prefs -= pref->freq;
2698 if (conflict_data->colorable_p)
2699 continue;
2700 ira_assert (bitmap_bit_p (coloring_allocno_bitmap,
2701 ALLOCNO_NUM (conflict_a)));
2702 if (update_left_conflict_sizes_p (conflict_a, a, size))
2704 delete_allocno_from_bucket
2705 (conflict_a, &uncolorable_allocno_bucket);
2706 add_allocno_to_ordered_colorable_bucket (conflict_a);
2707 if (internal_flag_ira_verbose > 4 && ira_dump_file != NULL)
2709 fprintf (ira_dump_file, " Making");
2710 ira_print_expanded_allocno (conflict_a);
2711 fprintf (ira_dump_file, " colorable\n");
2719 /* Put ALLOCNO onto the coloring stack and remove it from its bucket.
2720 The allocno is in the colorable bucket if COLORABLE_P is TRUE. */
2721 static void
2722 remove_allocno_from_bucket_and_push (ira_allocno_t allocno, bool colorable_p)
2724 if (colorable_p)
2725 delete_allocno_from_bucket (allocno, &colorable_allocno_bucket);
2726 else
2727 delete_allocno_from_bucket (allocno, &uncolorable_allocno_bucket);
2728 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2730 fprintf (ira_dump_file, " Pushing");
2731 ira_print_expanded_allocno (allocno);
2732 if (colorable_p)
2733 fprintf (ira_dump_file, "(cost %d)\n",
2734 ALLOCNO_COLOR_DATA (allocno)->temp);
2735 else
2736 fprintf (ira_dump_file, "(potential spill: %spri=%d, cost=%d)\n",
2737 ALLOCNO_BAD_SPILL_P (allocno) ? "bad spill, " : "",
2738 allocno_spill_priority (allocno),
2739 ALLOCNO_COLOR_DATA (allocno)->temp);
2741 if (! colorable_p)
2742 ALLOCNO_COLOR_DATA (allocno)->may_be_spilled_p = true;
2743 push_allocno_to_stack (allocno);
2746 /* Put all allocnos from colorable bucket onto the coloring stack. */
2747 static void
2748 push_only_colorable (void)
2750 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2751 fprintf (ira_dump_file, " Forming thread from colorable bucket:\n");
2752 form_threads_from_bucket (colorable_allocno_bucket);
2753 for (ira_allocno_t a = colorable_allocno_bucket;
2754 a != NULL;
2755 a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2756 update_costs_from_prefs (a);
2757 sort_bucket (&colorable_allocno_bucket, bucket_allocno_compare_func);
2758 for (;colorable_allocno_bucket != NULL;)
2759 remove_allocno_from_bucket_and_push (colorable_allocno_bucket, true);
2762 /* Return the frequency of exit edges (if EXIT_P) or entry from/to the
2763 loop given by its LOOP_NODE. */
2765 ira_loop_edge_freq (ira_loop_tree_node_t loop_node, int regno, bool exit_p)
2767 int freq, i;
2768 edge_iterator ei;
2769 edge e;
2771 ira_assert (current_loops != NULL && loop_node->loop != NULL
2772 && (regno < 0 || regno >= FIRST_PSEUDO_REGISTER));
2773 freq = 0;
2774 if (! exit_p)
2776 FOR_EACH_EDGE (e, ei, loop_node->loop->header->preds)
2777 if (e->src != loop_node->loop->latch
2778 && (regno < 0
2779 || (bitmap_bit_p (df_get_live_out (e->src), regno)
2780 && bitmap_bit_p (df_get_live_in (e->dest), regno))))
2781 freq += EDGE_FREQUENCY (e);
2783 else
2785 auto_vec<edge> edges = get_loop_exit_edges (loop_node->loop);
2786 FOR_EACH_VEC_ELT (edges, i, e)
2787 if (regno < 0
2788 || (bitmap_bit_p (df_get_live_out (e->src), regno)
2789 && bitmap_bit_p (df_get_live_in (e->dest), regno)))
2790 freq += EDGE_FREQUENCY (e);
2793 return REG_FREQ_FROM_EDGE_FREQ (freq);
2796 /* Construct an object that describes the boundary between A and its
2797 parent allocno. */
2798 ira_loop_border_costs::ira_loop_border_costs (ira_allocno_t a)
2799 : m_mode (ALLOCNO_MODE (a)),
2800 m_class (ALLOCNO_CLASS (a)),
2801 m_entry_freq (ira_loop_edge_freq (ALLOCNO_LOOP_TREE_NODE (a),
2802 ALLOCNO_REGNO (a), false)),
2803 m_exit_freq (ira_loop_edge_freq (ALLOCNO_LOOP_TREE_NODE (a),
2804 ALLOCNO_REGNO (a), true))
2808 /* Calculate and return the cost of putting allocno A into memory. */
2809 static int
2810 calculate_allocno_spill_cost (ira_allocno_t a)
2812 int regno, cost;
2813 ira_allocno_t parent_allocno;
2814 ira_loop_tree_node_t parent_node, loop_node;
2816 regno = ALLOCNO_REGNO (a);
2817 cost = ALLOCNO_UPDATED_MEMORY_COST (a) - ALLOCNO_UPDATED_CLASS_COST (a);
2818 if (ALLOCNO_CAP (a) != NULL)
2819 return cost;
2820 loop_node = ALLOCNO_LOOP_TREE_NODE (a);
2821 if ((parent_node = loop_node->parent) == NULL)
2822 return cost;
2823 if ((parent_allocno = parent_node->regno_allocno_map[regno]) == NULL)
2824 return cost;
2825 ira_loop_border_costs border_costs (a);
2826 if (ALLOCNO_HARD_REGNO (parent_allocno) < 0)
2827 cost -= border_costs.spill_outside_loop_cost ();
2828 else
2829 cost += (border_costs.spill_inside_loop_cost ()
2830 - border_costs.move_between_loops_cost ());
2831 return cost;
2834 /* Used for sorting allocnos for spilling. */
2835 static inline int
2836 allocno_spill_priority_compare (ira_allocno_t a1, ira_allocno_t a2)
2838 int pri1, pri2, diff;
2840 /* Avoid spilling static chain pointer pseudo when non-local goto is
2841 used. */
2842 if (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a1)))
2843 return 1;
2844 else if (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a2)))
2845 return -1;
2846 if (ALLOCNO_BAD_SPILL_P (a1) && ! ALLOCNO_BAD_SPILL_P (a2))
2847 return 1;
2848 if (ALLOCNO_BAD_SPILL_P (a2) && ! ALLOCNO_BAD_SPILL_P (a1))
2849 return -1;
2850 pri1 = allocno_spill_priority (a1);
2851 pri2 = allocno_spill_priority (a2);
2852 if ((diff = pri1 - pri2) != 0)
2853 return diff;
2854 if ((diff
2855 = ALLOCNO_COLOR_DATA (a1)->temp - ALLOCNO_COLOR_DATA (a2)->temp) != 0)
2856 return diff;
2857 return ALLOCNO_NUM (a1) - ALLOCNO_NUM (a2);
2860 /* Used for sorting allocnos for spilling. */
2861 static int
2862 allocno_spill_sort_compare (const void *v1p, const void *v2p)
2864 ira_allocno_t p1 = *(const ira_allocno_t *) v1p;
2865 ira_allocno_t p2 = *(const ira_allocno_t *) v2p;
2867 return allocno_spill_priority_compare (p1, p2);
2870 /* Push allocnos to the coloring stack. The order of allocnos in the
2871 stack defines the order for the subsequent coloring. */
2872 static void
2873 push_allocnos_to_stack (void)
2875 ira_allocno_t a;
2876 int cost;
2878 /* Calculate uncolorable allocno spill costs. */
2879 for (a = uncolorable_allocno_bucket;
2880 a != NULL;
2881 a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2882 if (ALLOCNO_CLASS (a) != NO_REGS)
2884 cost = calculate_allocno_spill_cost (a);
2885 /* ??? Remove cost of copies between the coalesced
2886 allocnos. */
2887 ALLOCNO_COLOR_DATA (a)->temp = cost;
2889 sort_bucket (&uncolorable_allocno_bucket, allocno_spill_sort_compare);
2890 for (;;)
2892 push_only_colorable ();
2893 a = uncolorable_allocno_bucket;
2894 if (a == NULL)
2895 break;
2896 remove_allocno_from_bucket_and_push (a, false);
2898 ira_assert (colorable_allocno_bucket == NULL
2899 && uncolorable_allocno_bucket == NULL);
2900 ira_assert (uncolorable_allocnos_num == 0);
2903 /* Pop the coloring stack and assign hard registers to the popped
2904 allocnos. */
2905 static void
2906 pop_allocnos_from_stack (void)
2908 ira_allocno_t allocno;
2909 enum reg_class aclass;
2911 for (;allocno_stack_vec.length () != 0;)
2913 allocno = allocno_stack_vec.pop ();
2914 aclass = ALLOCNO_CLASS (allocno);
2915 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2917 fprintf (ira_dump_file, " Popping");
2918 ira_print_expanded_allocno (allocno);
2919 fprintf (ira_dump_file, " -- ");
2921 if (aclass == NO_REGS)
2923 ALLOCNO_HARD_REGNO (allocno) = -1;
2924 ALLOCNO_ASSIGNED_P (allocno) = true;
2925 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (allocno) == NULL);
2926 ira_assert
2927 (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno) == NULL);
2928 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2929 fprintf (ira_dump_file, "assign memory\n");
2931 else if (assign_hard_reg (allocno, false))
2933 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2934 fprintf (ira_dump_file, " assign reg %d\n",
2935 ALLOCNO_HARD_REGNO (allocno));
2937 else if (ALLOCNO_ASSIGNED_P (allocno))
2939 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2940 fprintf (ira_dump_file, "spill%s\n",
2941 ALLOCNO_COLOR_DATA (allocno)->may_be_spilled_p
2942 ? "" : "!");
2944 ALLOCNO_COLOR_DATA (allocno)->in_graph_p = true;
2948 /* Set up number of available hard registers for allocno A. */
2949 static void
2950 setup_allocno_available_regs_num (ira_allocno_t a)
2952 int i, n, hard_regno, hard_regs_num, nwords;
2953 enum reg_class aclass;
2954 allocno_color_data_t data;
2956 aclass = ALLOCNO_CLASS (a);
2957 data = ALLOCNO_COLOR_DATA (a);
2958 data->available_regs_num = 0;
2959 if (aclass == NO_REGS)
2960 return;
2961 hard_regs_num = ira_class_hard_regs_num[aclass];
2962 nwords = ALLOCNO_NUM_OBJECTS (a);
2963 for (n = 0, i = hard_regs_num - 1; i >= 0; i--)
2965 hard_regno = ira_class_hard_regs[aclass][i];
2966 /* Checking only profitable hard regs. */
2967 if (TEST_HARD_REG_BIT (data->profitable_hard_regs, hard_regno))
2968 n++;
2970 data->available_regs_num = n;
2971 if (internal_flag_ira_verbose <= 2 || ira_dump_file == NULL)
2972 return;
2973 fprintf
2974 (ira_dump_file,
2975 " Allocno a%dr%d of %s(%d) has %d avail. regs ",
2976 ALLOCNO_NUM (a), ALLOCNO_REGNO (a),
2977 reg_class_names[aclass], ira_class_hard_regs_num[aclass], n);
2978 print_hard_reg_set (ira_dump_file, data->profitable_hard_regs, false);
2979 fprintf (ira_dump_file, ", %snode: ",
2980 data->profitable_hard_regs == data->hard_regs_node->hard_regs->set
2981 ? "" : "^");
2982 print_hard_reg_set (ira_dump_file,
2983 data->hard_regs_node->hard_regs->set, false);
2984 for (i = 0; i < nwords; i++)
2986 ira_object_t obj = ALLOCNO_OBJECT (a, i);
2988 if (nwords != 1)
2990 if (i != 0)
2991 fprintf (ira_dump_file, ", ");
2992 fprintf (ira_dump_file, " obj %d", i);
2994 fprintf (ira_dump_file, " (confl regs = ");
2995 print_hard_reg_set (ira_dump_file, OBJECT_TOTAL_CONFLICT_HARD_REGS (obj),
2996 false);
2997 fprintf (ira_dump_file, ")");
2999 fprintf (ira_dump_file, "\n");
3002 /* Put ALLOCNO in a bucket corresponding to its number and size of its
3003 conflicting allocnos and hard registers. */
3004 static void
3005 put_allocno_into_bucket (ira_allocno_t allocno)
3007 ALLOCNO_COLOR_DATA (allocno)->in_graph_p = true;
3008 setup_allocno_available_regs_num (allocno);
3009 if (setup_left_conflict_sizes_p (allocno))
3010 add_allocno_to_bucket (allocno, &colorable_allocno_bucket);
3011 else
3012 add_allocno_to_bucket (allocno, &uncolorable_allocno_bucket);
3015 /* Map: allocno number -> allocno priority. */
3016 static int *allocno_priorities;
3018 /* Set up priorities for N allocnos in array
3019 CONSIDERATION_ALLOCNOS. */
3020 static void
3021 setup_allocno_priorities (ira_allocno_t *consideration_allocnos, int n)
3023 int i, length, nrefs, priority, max_priority, mult, diff;
3024 ira_allocno_t a;
3026 max_priority = 0;
3027 for (i = 0; i < n; i++)
3029 a = consideration_allocnos[i];
3030 nrefs = ALLOCNO_NREFS (a);
3031 ira_assert (nrefs >= 0);
3032 mult = floor_log2 (ALLOCNO_NREFS (a)) + 1;
3033 ira_assert (mult >= 0);
3034 mult *= ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)];
3035 diff = ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a);
3036 #ifdef __has_builtin
3037 #if __has_builtin(__builtin_smul_overflow)
3038 #define HAS_SMUL_OVERFLOW
3039 #endif
3040 #endif
3041 /* Multiplication can overflow for very large functions.
3042 Check the overflow and constrain the result if necessary: */
3043 #ifdef HAS_SMUL_OVERFLOW
3044 if (__builtin_smul_overflow (mult, diff, &priority)
3045 || priority < -INT_MAX)
3046 priority = diff >= 0 ? INT_MAX : -INT_MAX;
3047 #else
3048 static_assert
3049 (sizeof (long long) >= 2 * sizeof (int),
3050 "overflow code does not work for such int and long long sizes");
3051 long long priorityll = (long long) mult * diff;
3052 if (priorityll < -INT_MAX || priorityll > INT_MAX)
3053 priority = diff >= 0 ? INT_MAX : -INT_MAX;
3054 else
3055 priority = priorityll;
3056 #endif
3057 allocno_priorities[ALLOCNO_NUM (a)] = priority;
3058 if (priority < 0)
3059 priority = -priority;
3060 if (max_priority < priority)
3061 max_priority = priority;
3063 mult = max_priority == 0 ? 1 : INT_MAX / max_priority;
3064 for (i = 0; i < n; i++)
3066 a = consideration_allocnos[i];
3067 length = ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a);
3068 if (ALLOCNO_NUM_OBJECTS (a) > 1)
3069 length /= ALLOCNO_NUM_OBJECTS (a);
3070 if (length <= 0)
3071 length = 1;
3072 allocno_priorities[ALLOCNO_NUM (a)]
3073 = allocno_priorities[ALLOCNO_NUM (a)] * mult / length;
3077 /* Sort allocnos according to the profit of usage of a hard register
3078 instead of memory for them. */
3079 static int
3080 allocno_cost_compare_func (const void *v1p, const void *v2p)
3082 ira_allocno_t p1 = *(const ira_allocno_t *) v1p;
3083 ira_allocno_t p2 = *(const ira_allocno_t *) v2p;
3084 int c1, c2;
3086 c1 = ALLOCNO_UPDATED_MEMORY_COST (p1) - ALLOCNO_UPDATED_CLASS_COST (p1);
3087 c2 = ALLOCNO_UPDATED_MEMORY_COST (p2) - ALLOCNO_UPDATED_CLASS_COST (p2);
3088 if (c1 - c2)
3089 return c1 - c2;
3091 /* If regs are equally good, sort by allocno numbers, so that the
3092 results of qsort leave nothing to chance. */
3093 return ALLOCNO_NUM (p1) - ALLOCNO_NUM (p2);
3096 /* Return savings on removed copies when ALLOCNO is assigned to
3097 HARD_REGNO. */
3098 static int
3099 allocno_copy_cost_saving (ira_allocno_t allocno, int hard_regno)
3101 int cost = 0;
3102 machine_mode allocno_mode = ALLOCNO_MODE (allocno);
3103 enum reg_class rclass;
3104 ira_copy_t cp, next_cp;
3106 rclass = REGNO_REG_CLASS (hard_regno);
3107 if (ira_reg_class_max_nregs[rclass][allocno_mode]
3108 > ira_class_hard_regs_num[rclass])
3109 /* For the above condition the cost can be wrong. Use the allocno
3110 class in this case. */
3111 rclass = ALLOCNO_CLASS (allocno);
3112 for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp)
3114 if (cp->first == allocno)
3116 next_cp = cp->next_first_allocno_copy;
3117 if (ALLOCNO_HARD_REGNO (cp->second) != hard_regno)
3118 continue;
3120 else if (cp->second == allocno)
3122 next_cp = cp->next_second_allocno_copy;
3123 if (ALLOCNO_HARD_REGNO (cp->first) != hard_regno)
3124 continue;
3126 else
3127 gcc_unreachable ();
3128 ira_init_register_move_cost_if_necessary (allocno_mode);
3129 cost += cp->freq * ira_register_move_cost[allocno_mode][rclass][rclass];
3131 return cost;
3134 /* We used Chaitin-Briggs coloring to assign as many pseudos as
3135 possible to hard registers. Let us try to improve allocation with
3136 cost point of view. This function improves the allocation by
3137 spilling some allocnos and assigning the freed hard registers to
3138 other allocnos if it decreases the overall allocation cost. */
3139 static void
3140 improve_allocation (void)
3142 unsigned int i;
3143 int j, k, n, hregno, conflict_hregno, base_cost, class_size, word, nwords;
3144 int check, spill_cost, min_cost, nregs, conflict_nregs, r, best;
3145 bool try_p;
3146 enum reg_class aclass;
3147 machine_mode mode;
3148 int *allocno_costs;
3149 int costs[FIRST_PSEUDO_REGISTER];
3150 HARD_REG_SET conflicting_regs[2], profitable_hard_regs;
3151 ira_allocno_t a;
3152 bitmap_iterator bi;
3154 /* Don't bother to optimize the code with static chain pointer and
3155 non-local goto in order not to spill the chain pointer
3156 pseudo. */
3157 if (cfun->static_chain_decl && crtl->has_nonlocal_goto)
3158 return;
3159 /* Clear counts used to process conflicting allocnos only once for
3160 each allocno. */
3161 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3162 ALLOCNO_COLOR_DATA (ira_allocnos[i])->temp = 0;
3163 check = n = 0;
3164 /* Process each allocno and try to assign a hard register to it by
3165 spilling some its conflicting allocnos. */
3166 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3168 a = ira_allocnos[i];
3169 ALLOCNO_COLOR_DATA (a)->temp = 0;
3170 if (empty_profitable_hard_regs (a))
3171 continue;
3172 check++;
3173 aclass = ALLOCNO_CLASS (a);
3174 allocno_costs = ALLOCNO_HARD_REG_COSTS (a);
3175 if ((hregno = ALLOCNO_HARD_REGNO (a)) < 0)
3176 base_cost = ALLOCNO_UPDATED_MEMORY_COST (a);
3177 else if (allocno_costs == NULL)
3178 /* It means that assigning a hard register is not profitable
3179 (we don't waste memory for hard register costs in this
3180 case). */
3181 continue;
3182 else
3183 base_cost = (allocno_costs[ira_class_hard_reg_index[aclass][hregno]]
3184 - allocno_copy_cost_saving (a, hregno));
3185 try_p = false;
3186 get_conflict_and_start_profitable_regs (a, false,
3187 conflicting_regs,
3188 &profitable_hard_regs);
3189 class_size = ira_class_hard_regs_num[aclass];
3190 /* Set up cost improvement for usage of each profitable hard
3191 register for allocno A. */
3192 for (j = 0; j < class_size; j++)
3194 hregno = ira_class_hard_regs[aclass][j];
3195 if (! check_hard_reg_p (a, hregno,
3196 conflicting_regs, profitable_hard_regs))
3197 continue;
3198 ira_assert (ira_class_hard_reg_index[aclass][hregno] == j);
3199 k = allocno_costs == NULL ? 0 : j;
3200 costs[hregno] = (allocno_costs == NULL
3201 ? ALLOCNO_UPDATED_CLASS_COST (a) : allocno_costs[k]);
3202 costs[hregno] -= allocno_copy_cost_saving (a, hregno);
3203 costs[hregno] -= base_cost;
3204 if (costs[hregno] < 0)
3205 try_p = true;
3207 if (! try_p)
3208 /* There is no chance to improve the allocation cost by
3209 assigning hard register to allocno A even without spilling
3210 conflicting allocnos. */
3211 continue;
3212 auto_bitmap allocnos_to_spill;
3213 HARD_REG_SET soft_conflict_regs = {};
3214 mode = ALLOCNO_MODE (a);
3215 nwords = ALLOCNO_NUM_OBJECTS (a);
3216 /* Process each allocno conflicting with A and update the cost
3217 improvement for profitable hard registers of A. To use a
3218 hard register for A we need to spill some conflicting
3219 allocnos and that creates penalty for the cost
3220 improvement. */
3221 for (word = 0; word < nwords; word++)
3223 ira_object_t conflict_obj;
3224 ira_object_t obj = ALLOCNO_OBJECT (a, word);
3225 ira_object_conflict_iterator oci;
3227 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
3229 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
3231 if (ALLOCNO_COLOR_DATA (conflict_a)->temp == check)
3232 /* We already processed this conflicting allocno
3233 because we processed earlier another object of the
3234 conflicting allocno. */
3235 continue;
3236 ALLOCNO_COLOR_DATA (conflict_a)->temp = check;
3237 if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0)
3238 continue;
3239 auto spill_a = ira_soft_conflict (a, conflict_a);
3240 if (spill_a)
3242 if (!bitmap_set_bit (allocnos_to_spill,
3243 ALLOCNO_NUM (spill_a)))
3244 continue;
3245 ira_loop_border_costs border_costs (spill_a);
3246 spill_cost = border_costs.spill_inside_loop_cost ();
3248 else
3250 spill_cost = ALLOCNO_UPDATED_MEMORY_COST (conflict_a);
3251 k = (ira_class_hard_reg_index
3252 [ALLOCNO_CLASS (conflict_a)][conflict_hregno]);
3253 ira_assert (k >= 0);
3254 if ((allocno_costs = ALLOCNO_HARD_REG_COSTS (conflict_a))
3255 != NULL)
3256 spill_cost -= allocno_costs[k];
3257 else
3258 spill_cost -= ALLOCNO_UPDATED_CLASS_COST (conflict_a);
3259 spill_cost
3260 += allocno_copy_cost_saving (conflict_a, conflict_hregno);
3262 conflict_nregs = hard_regno_nregs (conflict_hregno,
3263 ALLOCNO_MODE (conflict_a));
3264 auto note_conflict = [&](int r)
3266 if (check_hard_reg_p (a, r,
3267 conflicting_regs, profitable_hard_regs))
3269 if (spill_a)
3270 SET_HARD_REG_BIT (soft_conflict_regs, r);
3271 costs[r] += spill_cost;
3274 for (r = conflict_hregno;
3275 r >= 0 && (int) end_hard_regno (mode, r) > conflict_hregno;
3276 r--)
3277 note_conflict (r);
3278 for (r = conflict_hregno + 1;
3279 r < conflict_hregno + conflict_nregs;
3280 r++)
3281 note_conflict (r);
3284 min_cost = INT_MAX;
3285 best = -1;
3286 /* Now we choose hard register for A which results in highest
3287 allocation cost improvement. */
3288 for (j = 0; j < class_size; j++)
3290 hregno = ira_class_hard_regs[aclass][j];
3291 if (check_hard_reg_p (a, hregno,
3292 conflicting_regs, profitable_hard_regs)
3293 && min_cost > costs[hregno])
3295 best = hregno;
3296 min_cost = costs[hregno];
3299 if (min_cost >= 0)
3300 /* We are in a situation when assigning any hard register to A
3301 by spilling some conflicting allocnos does not improve the
3302 allocation cost. */
3303 continue;
3304 spill_soft_conflicts (a, allocnos_to_spill, soft_conflict_regs, best);
3305 nregs = hard_regno_nregs (best, mode);
3306 /* Now spill conflicting allocnos which contain a hard register
3307 of A when we assign the best chosen hard register to it. */
3308 for (word = 0; word < nwords; word++)
3310 ira_object_t conflict_obj;
3311 ira_object_t obj = ALLOCNO_OBJECT (a, word);
3312 ira_object_conflict_iterator oci;
3314 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
3316 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
3318 if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0)
3319 continue;
3320 conflict_nregs = hard_regno_nregs (conflict_hregno,
3321 ALLOCNO_MODE (conflict_a));
3322 if (best + nregs <= conflict_hregno
3323 || conflict_hregno + conflict_nregs <= best)
3324 /* No intersection. */
3325 continue;
3326 ALLOCNO_HARD_REGNO (conflict_a) = -1;
3327 sorted_allocnos[n++] = conflict_a;
3328 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
3329 fprintf (ira_dump_file, "Spilling a%dr%d for a%dr%d\n",
3330 ALLOCNO_NUM (conflict_a), ALLOCNO_REGNO (conflict_a),
3331 ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
3334 /* Assign the best chosen hard register to A. */
3335 ALLOCNO_HARD_REGNO (a) = best;
3336 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
3337 fprintf (ira_dump_file, "Assigning %d to a%dr%d\n",
3338 best, ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
3340 if (n == 0)
3341 return;
3342 /* We spilled some allocnos to assign their hard registers to other
3343 allocnos. The spilled allocnos are now in array
3344 'sorted_allocnos'. There is still a possibility that some of the
3345 spilled allocnos can get hard registers. So let us try assign
3346 them hard registers again (just a reminder -- function
3347 'assign_hard_reg' assigns hard registers only if it is possible
3348 and profitable). We process the spilled allocnos with biggest
3349 benefit to get hard register first -- see function
3350 'allocno_cost_compare_func'. */
3351 qsort (sorted_allocnos, n, sizeof (ira_allocno_t),
3352 allocno_cost_compare_func);
3353 for (j = 0; j < n; j++)
3355 a = sorted_allocnos[j];
3356 ALLOCNO_ASSIGNED_P (a) = false;
3357 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3359 fprintf (ira_dump_file, " ");
3360 ira_print_expanded_allocno (a);
3361 fprintf (ira_dump_file, " -- ");
3363 if (assign_hard_reg (a, false))
3365 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3366 fprintf (ira_dump_file, "assign hard reg %d\n",
3367 ALLOCNO_HARD_REGNO (a));
3369 else
3371 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3372 fprintf (ira_dump_file, "assign memory\n");
3377 /* Sort allocnos according to their priorities. */
3378 static int
3379 allocno_priority_compare_func (const void *v1p, const void *v2p)
3381 ira_allocno_t a1 = *(const ira_allocno_t *) v1p;
3382 ira_allocno_t a2 = *(const ira_allocno_t *) v2p;
3383 int pri1, pri2, diff;
3385 /* Assign hard reg to static chain pointer pseudo first when
3386 non-local goto is used. */
3387 if ((diff = (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a2))
3388 - non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a1)))) != 0)
3389 return diff;
3390 pri1 = allocno_priorities[ALLOCNO_NUM (a1)];
3391 pri2 = allocno_priorities[ALLOCNO_NUM (a2)];
3392 if (pri2 != pri1)
3393 return SORTGT (pri2, pri1);
3395 /* If regs are equally good, sort by allocnos, so that the results of
3396 qsort leave nothing to chance. */
3397 return ALLOCNO_NUM (a1) - ALLOCNO_NUM (a2);
3400 /* Chaitin-Briggs coloring for allocnos in COLORING_ALLOCNO_BITMAP
3401 taking into account allocnos in CONSIDERATION_ALLOCNO_BITMAP. */
3402 static void
3403 color_allocnos (void)
3405 unsigned int i, n;
3406 bitmap_iterator bi;
3407 ira_allocno_t a;
3409 setup_profitable_hard_regs ();
3410 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3412 allocno_color_data_t data;
3413 ira_pref_t pref, next_pref;
3415 a = ira_allocnos[i];
3416 data = ALLOCNO_COLOR_DATA (a);
3417 data->conflict_allocno_hard_prefs = 0;
3418 for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = next_pref)
3420 next_pref = pref->next_pref;
3421 if (! ira_hard_reg_in_set_p (pref->hard_regno,
3422 ALLOCNO_MODE (a),
3423 data->profitable_hard_regs))
3424 ira_remove_pref (pref);
3428 if (flag_ira_algorithm == IRA_ALGORITHM_PRIORITY)
3430 n = 0;
3431 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3433 a = ira_allocnos[i];
3434 if (ALLOCNO_CLASS (a) == NO_REGS)
3436 ALLOCNO_HARD_REGNO (a) = -1;
3437 ALLOCNO_ASSIGNED_P (a) = true;
3438 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL);
3439 ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a) == NULL);
3440 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3442 fprintf (ira_dump_file, " Spill");
3443 ira_print_expanded_allocno (a);
3444 fprintf (ira_dump_file, "\n");
3446 continue;
3448 sorted_allocnos[n++] = a;
3450 if (n != 0)
3452 setup_allocno_priorities (sorted_allocnos, n);
3453 qsort (sorted_allocnos, n, sizeof (ira_allocno_t),
3454 allocno_priority_compare_func);
3455 for (i = 0; i < n; i++)
3457 a = sorted_allocnos[i];
3458 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3460 fprintf (ira_dump_file, " ");
3461 ira_print_expanded_allocno (a);
3462 fprintf (ira_dump_file, " -- ");
3464 if (assign_hard_reg (a, false))
3466 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3467 fprintf (ira_dump_file, "assign hard reg %d\n",
3468 ALLOCNO_HARD_REGNO (a));
3470 else
3472 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3473 fprintf (ira_dump_file, "assign memory\n");
3478 else
3480 form_allocno_hard_regs_nodes_forest ();
3481 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
3482 print_hard_regs_forest (ira_dump_file);
3483 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3485 a = ira_allocnos[i];
3486 if (ALLOCNO_CLASS (a) != NO_REGS && ! empty_profitable_hard_regs (a))
3488 ALLOCNO_COLOR_DATA (a)->in_graph_p = true;
3489 update_conflict_allocno_hard_prefs (a);
3491 else
3493 ALLOCNO_HARD_REGNO (a) = -1;
3494 ALLOCNO_ASSIGNED_P (a) = true;
3495 /* We don't need updated costs anymore. */
3496 ira_free_allocno_updated_costs (a);
3497 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3499 fprintf (ira_dump_file, " Spill");
3500 ira_print_expanded_allocno (a);
3501 fprintf (ira_dump_file, "\n");
3505 /* Put the allocnos into the corresponding buckets. */
3506 colorable_allocno_bucket = NULL;
3507 uncolorable_allocno_bucket = NULL;
3508 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3510 a = ira_allocnos[i];
3511 if (ALLOCNO_COLOR_DATA (a)->in_graph_p)
3512 put_allocno_into_bucket (a);
3514 push_allocnos_to_stack ();
3515 pop_allocnos_from_stack ();
3516 finish_allocno_hard_regs_nodes_forest ();
3518 improve_allocation ();
3523 /* Output information about the loop given by its LOOP_TREE_NODE. */
3524 static void
3525 print_loop_title (ira_loop_tree_node_t loop_tree_node)
3527 unsigned int j;
3528 bitmap_iterator bi;
3529 ira_loop_tree_node_t subloop_node, dest_loop_node;
3530 edge e;
3531 edge_iterator ei;
3533 if (loop_tree_node->parent == NULL)
3534 fprintf (ira_dump_file,
3535 "\n Loop 0 (parent -1, header bb%d, depth 0)\n bbs:",
3536 NUM_FIXED_BLOCKS);
3537 else
3539 ira_assert (current_loops != NULL && loop_tree_node->loop != NULL);
3540 fprintf (ira_dump_file,
3541 "\n Loop %d (parent %d, header bb%d, depth %d)\n bbs:",
3542 loop_tree_node->loop_num, loop_tree_node->parent->loop_num,
3543 loop_tree_node->loop->header->index,
3544 loop_depth (loop_tree_node->loop));
3546 for (subloop_node = loop_tree_node->children;
3547 subloop_node != NULL;
3548 subloop_node = subloop_node->next)
3549 if (subloop_node->bb != NULL)
3551 fprintf (ira_dump_file, " %d", subloop_node->bb->index);
3552 FOR_EACH_EDGE (e, ei, subloop_node->bb->succs)
3553 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
3554 && ((dest_loop_node = IRA_BB_NODE (e->dest)->parent)
3555 != loop_tree_node))
3556 fprintf (ira_dump_file, "(->%d:l%d)",
3557 e->dest->index, dest_loop_node->loop_num);
3559 fprintf (ira_dump_file, "\n all:");
3560 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->all_allocnos, 0, j, bi)
3561 fprintf (ira_dump_file, " %dr%d", j, ALLOCNO_REGNO (ira_allocnos[j]));
3562 fprintf (ira_dump_file, "\n modified regnos:");
3563 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->modified_regnos, 0, j, bi)
3564 fprintf (ira_dump_file, " %d", j);
3565 fprintf (ira_dump_file, "\n border:");
3566 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->border_allocnos, 0, j, bi)
3567 fprintf (ira_dump_file, " %dr%d", j, ALLOCNO_REGNO (ira_allocnos[j]));
3568 fprintf (ira_dump_file, "\n Pressure:");
3569 for (j = 0; (int) j < ira_pressure_classes_num; j++)
3571 enum reg_class pclass;
3573 pclass = ira_pressure_classes[j];
3574 if (loop_tree_node->reg_pressure[pclass] == 0)
3575 continue;
3576 fprintf (ira_dump_file, " %s=%d", reg_class_names[pclass],
3577 loop_tree_node->reg_pressure[pclass]);
3579 fprintf (ira_dump_file, "\n");
3582 /* Color the allocnos inside loop (in the extreme case it can be all
3583 of the function) given the corresponding LOOP_TREE_NODE. The
3584 function is called for each loop during top-down traverse of the
3585 loop tree. */
3586 static void
3587 color_pass (ira_loop_tree_node_t loop_tree_node)
3589 int regno, hard_regno, index = -1, n;
3590 int cost;
3591 unsigned int j;
3592 bitmap_iterator bi;
3593 machine_mode mode;
3594 enum reg_class rclass, aclass;
3595 ira_allocno_t a, subloop_allocno;
3596 ira_loop_tree_node_t subloop_node;
3598 ira_assert (loop_tree_node->bb == NULL);
3599 if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL)
3600 print_loop_title (loop_tree_node);
3602 bitmap_copy (coloring_allocno_bitmap, loop_tree_node->all_allocnos);
3603 bitmap_copy (consideration_allocno_bitmap, coloring_allocno_bitmap);
3604 n = 0;
3605 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3607 a = ira_allocnos[j];
3608 n++;
3609 if (! ALLOCNO_ASSIGNED_P (a))
3610 continue;
3611 bitmap_clear_bit (coloring_allocno_bitmap, ALLOCNO_NUM (a));
3613 allocno_color_data
3614 = (allocno_color_data_t) ira_allocate (sizeof (struct allocno_color_data)
3615 * n);
3616 memset (allocno_color_data, 0, sizeof (struct allocno_color_data) * n);
3617 curr_allocno_process = 0;
3618 n = 0;
3619 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3621 a = ira_allocnos[j];
3622 ALLOCNO_ADD_DATA (a) = allocno_color_data + n;
3623 n++;
3625 init_allocno_threads ();
3626 /* Color all mentioned allocnos including transparent ones. */
3627 color_allocnos ();
3628 /* Process caps. They are processed just once. */
3629 if (flag_ira_region == IRA_REGION_MIXED
3630 || flag_ira_region == IRA_REGION_ALL)
3631 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->all_allocnos, 0, j, bi)
3633 a = ira_allocnos[j];
3634 if (ALLOCNO_CAP_MEMBER (a) == NULL)
3635 continue;
3636 /* Remove from processing in the next loop. */
3637 bitmap_clear_bit (consideration_allocno_bitmap, j);
3638 rclass = ALLOCNO_CLASS (a);
3639 subloop_allocno = ALLOCNO_CAP_MEMBER (a);
3640 subloop_node = ALLOCNO_LOOP_TREE_NODE (subloop_allocno);
3641 if (ira_single_region_allocno_p (a, subloop_allocno))
3643 mode = ALLOCNO_MODE (a);
3644 hard_regno = ALLOCNO_HARD_REGNO (a);
3645 if (hard_regno >= 0)
3647 index = ira_class_hard_reg_index[rclass][hard_regno];
3648 ira_assert (index >= 0);
3650 regno = ALLOCNO_REGNO (a);
3651 ira_assert (!ALLOCNO_ASSIGNED_P (subloop_allocno));
3652 ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno;
3653 ALLOCNO_ASSIGNED_P (subloop_allocno) = true;
3654 if (hard_regno >= 0)
3655 update_costs_from_copies (subloop_allocno, true, true);
3656 /* We don't need updated costs anymore. */
3657 ira_free_allocno_updated_costs (subloop_allocno);
3660 /* Update costs of the corresponding allocnos (not caps) in the
3661 subloops. */
3662 for (subloop_node = loop_tree_node->subloops;
3663 subloop_node != NULL;
3664 subloop_node = subloop_node->subloop_next)
3666 ira_assert (subloop_node->bb == NULL);
3667 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3669 a = ira_allocnos[j];
3670 ira_assert (ALLOCNO_CAP_MEMBER (a) == NULL);
3671 mode = ALLOCNO_MODE (a);
3672 rclass = ALLOCNO_CLASS (a);
3673 hard_regno = ALLOCNO_HARD_REGNO (a);
3674 /* Use hard register class here. ??? */
3675 if (hard_regno >= 0)
3677 index = ira_class_hard_reg_index[rclass][hard_regno];
3678 ira_assert (index >= 0);
3680 regno = ALLOCNO_REGNO (a);
3681 /* ??? conflict costs */
3682 subloop_allocno = subloop_node->regno_allocno_map[regno];
3683 if (subloop_allocno == NULL
3684 || ALLOCNO_CAP (subloop_allocno) != NULL)
3685 continue;
3686 ira_assert (ALLOCNO_CLASS (subloop_allocno) == rclass);
3687 ira_assert (bitmap_bit_p (subloop_node->all_allocnos,
3688 ALLOCNO_NUM (subloop_allocno)));
3689 if (ira_single_region_allocno_p (a, subloop_allocno)
3690 || !ira_subloop_allocnos_can_differ_p (a, hard_regno >= 0,
3691 false))
3693 gcc_assert (!ALLOCNO_MIGHT_CONFLICT_WITH_PARENT_P
3694 (subloop_allocno));
3695 if (! ALLOCNO_ASSIGNED_P (subloop_allocno))
3697 ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno;
3698 ALLOCNO_ASSIGNED_P (subloop_allocno) = true;
3699 if (hard_regno >= 0)
3700 update_costs_from_copies (subloop_allocno, true, true);
3701 /* We don't need updated costs anymore. */
3702 ira_free_allocno_updated_costs (subloop_allocno);
3705 else if (hard_regno < 0)
3707 /* If we allocate a register to SUBLOOP_ALLOCNO, we'll need
3708 to load the register on entry to the subloop and store
3709 the register back on exit from the subloop. This incurs
3710 a fixed cost for all registers. Since UPDATED_MEMORY_COST
3711 is (and should only be) used relative to the register costs
3712 for the same allocno, we can subtract this shared register
3713 cost from the memory cost. */
3714 ira_loop_border_costs border_costs (subloop_allocno);
3715 ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno)
3716 -= border_costs.spill_outside_loop_cost ();
3718 else
3720 ira_loop_border_costs border_costs (subloop_allocno);
3721 aclass = ALLOCNO_CLASS (subloop_allocno);
3722 ira_init_register_move_cost_if_necessary (mode);
3723 cost = border_costs.move_between_loops_cost ();
3724 ira_allocate_and_set_or_copy_costs
3725 (&ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno), aclass,
3726 ALLOCNO_UPDATED_CLASS_COST (subloop_allocno),
3727 ALLOCNO_HARD_REG_COSTS (subloop_allocno));
3728 ira_allocate_and_set_or_copy_costs
3729 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno),
3730 aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (subloop_allocno));
3731 ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index] -= cost;
3732 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno)[index]
3733 -= cost;
3734 if (ALLOCNO_UPDATED_CLASS_COST (subloop_allocno)
3735 > ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index])
3736 ALLOCNO_UPDATED_CLASS_COST (subloop_allocno)
3737 = ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index];
3738 /* If we spill SUBLOOP_ALLOCNO, we'll need to store HARD_REGNO
3739 on entry to the subloop and restore HARD_REGNO on exit from
3740 the subloop. */
3741 ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno)
3742 += border_costs.spill_inside_loop_cost ();
3746 ira_free (allocno_color_data);
3747 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3749 a = ira_allocnos[j];
3750 ALLOCNO_ADD_DATA (a) = NULL;
3754 /* Initialize the common data for coloring and calls functions to do
3755 Chaitin-Briggs and regional coloring. */
3756 static void
3757 do_coloring (void)
3759 coloring_allocno_bitmap = ira_allocate_bitmap ();
3760 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
3761 fprintf (ira_dump_file, "\n**** Allocnos coloring:\n\n");
3763 ira_traverse_loop_tree (false, ira_loop_tree_root, color_pass, NULL);
3765 if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL)
3766 ira_print_disposition (ira_dump_file);
3768 ira_free_bitmap (coloring_allocno_bitmap);
3773 /* Move spill/restore code, which are to be generated in ira-emit.cc,
3774 to less frequent points (if it is profitable) by reassigning some
3775 allocnos (in loop with subloops containing in another loop) to
3776 memory which results in longer live-range where the corresponding
3777 pseudo-registers will be in memory. */
3778 static void
3779 move_spill_restore (void)
3781 int cost, regno, hard_regno, hard_regno2, index;
3782 bool changed_p;
3783 machine_mode mode;
3784 enum reg_class rclass;
3785 ira_allocno_t a, parent_allocno, subloop_allocno;
3786 ira_loop_tree_node_t parent, loop_node, subloop_node;
3787 ira_allocno_iterator ai;
3789 for (;;)
3791 changed_p = false;
3792 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
3793 fprintf (ira_dump_file, "New iteration of spill/restore move\n");
3794 FOR_EACH_ALLOCNO (a, ai)
3796 regno = ALLOCNO_REGNO (a);
3797 loop_node = ALLOCNO_LOOP_TREE_NODE (a);
3798 if (ALLOCNO_CAP_MEMBER (a) != NULL
3799 || ALLOCNO_CAP (a) != NULL
3800 || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0
3801 || loop_node->children == NULL
3802 /* don't do the optimization because it can create
3803 copies and the reload pass can spill the allocno set
3804 by copy although the allocno will not get memory
3805 slot. */
3806 || ira_equiv_no_lvalue_p (regno)
3807 || !bitmap_bit_p (loop_node->border_allocnos, ALLOCNO_NUM (a))
3808 /* Do not spill static chain pointer pseudo when
3809 non-local goto is used. */
3810 || non_spilled_static_chain_regno_p (regno))
3811 continue;
3812 mode = ALLOCNO_MODE (a);
3813 rclass = ALLOCNO_CLASS (a);
3814 index = ira_class_hard_reg_index[rclass][hard_regno];
3815 ira_assert (index >= 0);
3816 cost = (ALLOCNO_MEMORY_COST (a)
3817 - (ALLOCNO_HARD_REG_COSTS (a) == NULL
3818 ? ALLOCNO_CLASS_COST (a)
3819 : ALLOCNO_HARD_REG_COSTS (a)[index]));
3820 ira_init_register_move_cost_if_necessary (mode);
3821 for (subloop_node = loop_node->subloops;
3822 subloop_node != NULL;
3823 subloop_node = subloop_node->subloop_next)
3825 ira_assert (subloop_node->bb == NULL);
3826 subloop_allocno = subloop_node->regno_allocno_map[regno];
3827 if (subloop_allocno == NULL)
3828 continue;
3829 ira_assert (rclass == ALLOCNO_CLASS (subloop_allocno));
3830 ira_loop_border_costs border_costs (subloop_allocno);
3832 /* We have accumulated cost. To get the real cost of
3833 allocno usage in the loop we should subtract the costs
3834 added by propagate_allocno_info for the subloop allocnos. */
3835 int reg_cost
3836 = (ALLOCNO_HARD_REG_COSTS (subloop_allocno) == NULL
3837 ? ALLOCNO_CLASS_COST (subloop_allocno)
3838 : ALLOCNO_HARD_REG_COSTS (subloop_allocno)[index]);
3840 int spill_cost
3841 = (border_costs.spill_inside_loop_cost ()
3842 + ALLOCNO_MEMORY_COST (subloop_allocno));
3844 /* If HARD_REGNO conflicts with SUBLOOP_A then
3845 propagate_allocno_info will have propagated
3846 the cost of spilling HARD_REGNO in SUBLOOP_NODE.
3847 (ira_subloop_allocnos_can_differ_p must be true
3848 in that case.) If HARD_REGNO is a caller-saved
3849 register, we might have modelled it in the same way.
3851 Otherwise, SPILL_COST acted as a cap on the propagated
3852 register cost, in cases where the allocations can differ. */
3853 auto conflicts = ira_total_conflict_hard_regs (subloop_allocno);
3854 if (TEST_HARD_REG_BIT (conflicts, hard_regno)
3855 || (ira_need_caller_save_p (subloop_allocno, hard_regno)
3856 && ira_caller_save_loop_spill_p (a, subloop_allocno,
3857 spill_cost)))
3858 reg_cost = spill_cost;
3859 else if (ira_subloop_allocnos_can_differ_p (a))
3860 reg_cost = MIN (reg_cost, spill_cost);
3862 cost -= ALLOCNO_MEMORY_COST (subloop_allocno) - reg_cost;
3864 if ((hard_regno2 = ALLOCNO_HARD_REGNO (subloop_allocno)) < 0)
3865 /* The register was spilled in the subloop. If we spill
3866 it in the outer loop too then we'll no longer need to
3867 save the register on entry to the subloop and restore
3868 the register on exit from the subloop. */
3869 cost -= border_costs.spill_inside_loop_cost ();
3870 else
3872 /* The register was also allocated in the subloop. If we
3873 spill it in the outer loop then we'll need to load the
3874 register on entry to the subloop and store the register
3875 back on exit from the subloop. */
3876 cost += border_costs.spill_outside_loop_cost ();
3877 if (hard_regno2 != hard_regno)
3878 cost -= border_costs.move_between_loops_cost ();
3881 if ((parent = loop_node->parent) != NULL
3882 && (parent_allocno = parent->regno_allocno_map[regno]) != NULL)
3884 ira_assert (rclass == ALLOCNO_CLASS (parent_allocno));
3885 ira_loop_border_costs border_costs (a);
3886 if ((hard_regno2 = ALLOCNO_HARD_REGNO (parent_allocno)) < 0)
3887 /* The register was spilled in the parent loop. If we spill
3888 it in this loop too then we'll no longer need to load the
3889 register on entry to this loop and save the register back
3890 on exit from this loop. */
3891 cost -= border_costs.spill_outside_loop_cost ();
3892 else
3894 /* The register was also allocated in the parent loop.
3895 If we spill it in this loop then we'll need to save
3896 the register on entry to this loop and restore the
3897 register on exit from this loop. */
3898 cost += border_costs.spill_inside_loop_cost ();
3899 if (hard_regno2 != hard_regno)
3900 cost -= border_costs.move_between_loops_cost ();
3903 if (cost < 0)
3905 ALLOCNO_HARD_REGNO (a) = -1;
3906 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3908 fprintf
3909 (ira_dump_file,
3910 " Moving spill/restore for a%dr%d up from loop %d",
3911 ALLOCNO_NUM (a), regno, loop_node->loop_num);
3912 fprintf (ira_dump_file, " - profit %d\n", -cost);
3914 changed_p = true;
3917 if (! changed_p)
3918 break;
3924 /* Update current hard reg costs and current conflict hard reg costs
3925 for allocno A. It is done by processing its copies containing
3926 other allocnos already assigned. */
3927 static void
3928 update_curr_costs (ira_allocno_t a)
3930 int i, hard_regno, cost;
3931 machine_mode mode;
3932 enum reg_class aclass, rclass;
3933 ira_allocno_t another_a;
3934 ira_copy_t cp, next_cp;
3936 ira_free_allocno_updated_costs (a);
3937 ira_assert (! ALLOCNO_ASSIGNED_P (a));
3938 aclass = ALLOCNO_CLASS (a);
3939 if (aclass == NO_REGS)
3940 return;
3941 mode = ALLOCNO_MODE (a);
3942 ira_init_register_move_cost_if_necessary (mode);
3943 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
3945 if (cp->first == a)
3947 next_cp = cp->next_first_allocno_copy;
3948 another_a = cp->second;
3950 else if (cp->second == a)
3952 next_cp = cp->next_second_allocno_copy;
3953 another_a = cp->first;
3955 else
3956 gcc_unreachable ();
3957 if (! ira_reg_classes_intersect_p[aclass][ALLOCNO_CLASS (another_a)]
3958 || ! ALLOCNO_ASSIGNED_P (another_a)
3959 || (hard_regno = ALLOCNO_HARD_REGNO (another_a)) < 0)
3960 continue;
3961 rclass = REGNO_REG_CLASS (hard_regno);
3962 i = ira_class_hard_reg_index[aclass][hard_regno];
3963 if (i < 0)
3964 continue;
3965 cost = (cp->first == a
3966 ? ira_register_move_cost[mode][rclass][aclass]
3967 : ira_register_move_cost[mode][aclass][rclass]);
3968 ira_allocate_and_set_or_copy_costs
3969 (&ALLOCNO_UPDATED_HARD_REG_COSTS (a), aclass, ALLOCNO_CLASS_COST (a),
3970 ALLOCNO_HARD_REG_COSTS (a));
3971 ira_allocate_and_set_or_copy_costs
3972 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a),
3973 aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (a));
3974 ALLOCNO_UPDATED_HARD_REG_COSTS (a)[i] -= cp->freq * cost;
3975 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a)[i] -= cp->freq * cost;
3979 /* Try to assign hard registers to the unassigned allocnos and
3980 allocnos conflicting with them or conflicting with allocnos whose
3981 regno >= START_REGNO. The function is called after ira_flattening,
3982 so more allocnos (including ones created in ira-emit.cc) will have a
3983 chance to get a hard register. We use simple assignment algorithm
3984 based on priorities. */
3985 void
3986 ira_reassign_conflict_allocnos (int start_regno)
3988 int i, allocnos_to_color_num;
3989 ira_allocno_t a;
3990 enum reg_class aclass;
3991 bitmap allocnos_to_color;
3992 ira_allocno_iterator ai;
3994 allocnos_to_color = ira_allocate_bitmap ();
3995 allocnos_to_color_num = 0;
3996 FOR_EACH_ALLOCNO (a, ai)
3998 int n = ALLOCNO_NUM_OBJECTS (a);
4000 if (! ALLOCNO_ASSIGNED_P (a)
4001 && ! bitmap_bit_p (allocnos_to_color, ALLOCNO_NUM (a)))
4003 if (ALLOCNO_CLASS (a) != NO_REGS)
4004 sorted_allocnos[allocnos_to_color_num++] = a;
4005 else
4007 ALLOCNO_ASSIGNED_P (a) = true;
4008 ALLOCNO_HARD_REGNO (a) = -1;
4009 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL);
4010 ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a) == NULL);
4012 bitmap_set_bit (allocnos_to_color, ALLOCNO_NUM (a));
4014 if (ALLOCNO_REGNO (a) < start_regno
4015 || (aclass = ALLOCNO_CLASS (a)) == NO_REGS)
4016 continue;
4017 for (i = 0; i < n; i++)
4019 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4020 ira_object_t conflict_obj;
4021 ira_object_conflict_iterator oci;
4023 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
4025 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
4027 ira_assert (ira_reg_classes_intersect_p
4028 [aclass][ALLOCNO_CLASS (conflict_a)]);
4029 if (!bitmap_set_bit (allocnos_to_color, ALLOCNO_NUM (conflict_a)))
4030 continue;
4031 sorted_allocnos[allocnos_to_color_num++] = conflict_a;
4035 ira_free_bitmap (allocnos_to_color);
4036 if (allocnos_to_color_num > 1)
4038 setup_allocno_priorities (sorted_allocnos, allocnos_to_color_num);
4039 qsort (sorted_allocnos, allocnos_to_color_num, sizeof (ira_allocno_t),
4040 allocno_priority_compare_func);
4042 for (i = 0; i < allocnos_to_color_num; i++)
4044 a = sorted_allocnos[i];
4045 ALLOCNO_ASSIGNED_P (a) = false;
4046 update_curr_costs (a);
4048 for (i = 0; i < allocnos_to_color_num; i++)
4050 a = sorted_allocnos[i];
4051 if (assign_hard_reg (a, true))
4053 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4054 fprintf
4055 (ira_dump_file,
4056 " Secondary allocation: assign hard reg %d to reg %d\n",
4057 ALLOCNO_HARD_REGNO (a), ALLOCNO_REGNO (a));
4064 /* This page contains functions used to find conflicts using allocno
4065 live ranges. */
4067 #ifdef ENABLE_IRA_CHECKING
4069 /* Return TRUE if live ranges of pseudo-registers REGNO1 and REGNO2
4070 intersect. This should be used when there is only one region.
4071 Currently this is used during reload. */
4072 static bool
4073 conflict_by_live_ranges_p (int regno1, int regno2)
4075 ira_allocno_t a1, a2;
4077 ira_assert (regno1 >= FIRST_PSEUDO_REGISTER
4078 && regno2 >= FIRST_PSEUDO_REGISTER);
4079 /* Reg info calculated by dataflow infrastructure can be different
4080 from one calculated by regclass. */
4081 if ((a1 = ira_loop_tree_root->regno_allocno_map[regno1]) == NULL
4082 || (a2 = ira_loop_tree_root->regno_allocno_map[regno2]) == NULL)
4083 return false;
4084 return allocnos_conflict_by_live_ranges_p (a1, a2);
4087 #endif
4091 /* This page contains code to coalesce memory stack slots used by
4092 spilled allocnos. This results in smaller stack frame, better data
4093 locality, and in smaller code for some architectures like
4094 x86/x86_64 where insn size depends on address displacement value.
4095 On the other hand, it can worsen insn scheduling after the RA but
4096 in practice it is less important than smaller stack frames. */
4098 /* TRUE if we coalesced some allocnos. In other words, if we got
4099 loops formed by members first_coalesced_allocno and
4100 next_coalesced_allocno containing more one allocno. */
4101 static bool allocno_coalesced_p;
4103 /* Bitmap used to prevent a repeated allocno processing because of
4104 coalescing. */
4105 static bitmap processed_coalesced_allocno_bitmap;
4107 /* See below. */
4108 typedef struct coalesce_data *coalesce_data_t;
4110 /* To decrease footprint of ira_allocno structure we store all data
4111 needed only for coalescing in the following structure. */
4112 struct coalesce_data
4114 /* Coalesced allocnos form a cyclic list. One allocno given by
4115 FIRST represents all coalesced allocnos. The
4116 list is chained by NEXT. */
4117 ira_allocno_t first;
4118 ira_allocno_t next;
4119 int temp;
4122 /* Container for storing allocno data concerning coalescing. */
4123 static coalesce_data_t allocno_coalesce_data;
4125 /* Macro to access the data concerning coalescing. */
4126 #define ALLOCNO_COALESCE_DATA(a) ((coalesce_data_t) ALLOCNO_ADD_DATA (a))
4128 /* Merge two sets of coalesced allocnos given correspondingly by
4129 allocnos A1 and A2 (more accurately merging A2 set into A1
4130 set). */
4131 static void
4132 merge_allocnos (ira_allocno_t a1, ira_allocno_t a2)
4134 ira_allocno_t a, first, last, next;
4136 first = ALLOCNO_COALESCE_DATA (a1)->first;
4137 a = ALLOCNO_COALESCE_DATA (a2)->first;
4138 if (first == a)
4139 return;
4140 for (last = a2, a = ALLOCNO_COALESCE_DATA (a2)->next;;
4141 a = ALLOCNO_COALESCE_DATA (a)->next)
4143 ALLOCNO_COALESCE_DATA (a)->first = first;
4144 if (a == a2)
4145 break;
4146 last = a;
4148 next = allocno_coalesce_data[ALLOCNO_NUM (first)].next;
4149 allocno_coalesce_data[ALLOCNO_NUM (first)].next = a2;
4150 allocno_coalesce_data[ALLOCNO_NUM (last)].next = next;
4153 /* Return TRUE if there are conflicting allocnos from two sets of
4154 coalesced allocnos given correspondingly by allocnos A1 and A2. We
4155 use live ranges to find conflicts because conflicts are represented
4156 only for allocnos of the same allocno class and during the reload
4157 pass we coalesce allocnos for sharing stack memory slots. */
4158 static bool
4159 coalesced_allocno_conflict_p (ira_allocno_t a1, ira_allocno_t a2)
4161 ira_allocno_t a, conflict_a;
4163 if (allocno_coalesced_p)
4165 bitmap_clear (processed_coalesced_allocno_bitmap);
4166 for (a = ALLOCNO_COALESCE_DATA (a1)->next;;
4167 a = ALLOCNO_COALESCE_DATA (a)->next)
4169 bitmap_set_bit (processed_coalesced_allocno_bitmap, ALLOCNO_NUM (a));
4170 if (a == a1)
4171 break;
4174 for (a = ALLOCNO_COALESCE_DATA (a2)->next;;
4175 a = ALLOCNO_COALESCE_DATA (a)->next)
4177 for (conflict_a = ALLOCNO_COALESCE_DATA (a1)->next;;
4178 conflict_a = ALLOCNO_COALESCE_DATA (conflict_a)->next)
4180 if (allocnos_conflict_by_live_ranges_p (a, conflict_a))
4181 return true;
4182 if (conflict_a == a1)
4183 break;
4185 if (a == a2)
4186 break;
4188 return false;
4191 /* The major function for aggressive allocno coalescing. We coalesce
4192 only spilled allocnos. If some allocnos have been coalesced, we
4193 set up flag allocno_coalesced_p. */
4194 static void
4195 coalesce_allocnos (void)
4197 ira_allocno_t a;
4198 ira_copy_t cp, next_cp;
4199 unsigned int j;
4200 int i, n, cp_num, regno;
4201 bitmap_iterator bi;
4203 cp_num = 0;
4204 /* Collect copies. */
4205 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, j, bi)
4207 a = ira_allocnos[j];
4208 regno = ALLOCNO_REGNO (a);
4209 if (! ALLOCNO_ASSIGNED_P (a) || ALLOCNO_HARD_REGNO (a) >= 0
4210 || ira_equiv_no_lvalue_p (regno))
4211 continue;
4212 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
4214 if (cp->first == a)
4216 next_cp = cp->next_first_allocno_copy;
4217 regno = ALLOCNO_REGNO (cp->second);
4218 /* For priority coloring we coalesce allocnos only with
4219 the same allocno class not with intersected allocno
4220 classes as it were possible. It is done for
4221 simplicity. */
4222 if ((cp->insn != NULL || cp->constraint_p)
4223 && ALLOCNO_ASSIGNED_P (cp->second)
4224 && ALLOCNO_HARD_REGNO (cp->second) < 0
4225 && ! ira_equiv_no_lvalue_p (regno))
4226 sorted_copies[cp_num++] = cp;
4228 else if (cp->second == a)
4229 next_cp = cp->next_second_allocno_copy;
4230 else
4231 gcc_unreachable ();
4234 qsort (sorted_copies, cp_num, sizeof (ira_copy_t), copy_freq_compare_func);
4235 /* Coalesced copies, most frequently executed first. */
4236 for (; cp_num != 0;)
4238 for (i = 0; i < cp_num; i++)
4240 cp = sorted_copies[i];
4241 if (! coalesced_allocno_conflict_p (cp->first, cp->second))
4243 allocno_coalesced_p = true;
4244 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4245 fprintf
4246 (ira_dump_file,
4247 " Coalescing copy %d:a%dr%d-a%dr%d (freq=%d)\n",
4248 cp->num, ALLOCNO_NUM (cp->first), ALLOCNO_REGNO (cp->first),
4249 ALLOCNO_NUM (cp->second), ALLOCNO_REGNO (cp->second),
4250 cp->freq);
4251 merge_allocnos (cp->first, cp->second);
4252 i++;
4253 break;
4256 /* Collect the rest of copies. */
4257 for (n = 0; i < cp_num; i++)
4259 cp = sorted_copies[i];
4260 if (allocno_coalesce_data[ALLOCNO_NUM (cp->first)].first
4261 != allocno_coalesce_data[ALLOCNO_NUM (cp->second)].first)
4262 sorted_copies[n++] = cp;
4264 cp_num = n;
4268 /* Usage cost and order number of coalesced allocno set to which
4269 given pseudo register belongs to. */
4270 static int *regno_coalesced_allocno_cost;
4271 static int *regno_coalesced_allocno_num;
4273 /* Sort pseudos according frequencies of coalesced allocno sets they
4274 belong to (putting most frequently ones first), and according to
4275 coalesced allocno set order numbers. */
4276 static int
4277 coalesced_pseudo_reg_freq_compare (const void *v1p, const void *v2p)
4279 const int regno1 = *(const int *) v1p;
4280 const int regno2 = *(const int *) v2p;
4281 int diff;
4283 if ((diff = (regno_coalesced_allocno_cost[regno2]
4284 - regno_coalesced_allocno_cost[regno1])) != 0)
4285 return diff;
4286 if ((diff = (regno_coalesced_allocno_num[regno1]
4287 - regno_coalesced_allocno_num[regno2])) != 0)
4288 return diff;
4289 return regno1 - regno2;
4292 /* Widest width in which each pseudo reg is referred to (via subreg).
4293 It is used for sorting pseudo registers. */
4294 static machine_mode *regno_max_ref_mode;
4296 /* Sort pseudos according their slot numbers (putting ones with
4297 smaller numbers first, or last when the frame pointer is not
4298 needed). */
4299 static int
4300 coalesced_pseudo_reg_slot_compare (const void *v1p, const void *v2p)
4302 const int regno1 = *(const int *) v1p;
4303 const int regno2 = *(const int *) v2p;
4304 ira_allocno_t a1 = ira_regno_allocno_map[regno1];
4305 ira_allocno_t a2 = ira_regno_allocno_map[regno2];
4306 int diff, slot_num1, slot_num2;
4307 machine_mode mode1, mode2;
4309 if (a1 == NULL || ALLOCNO_HARD_REGNO (a1) >= 0)
4311 if (a2 == NULL || ALLOCNO_HARD_REGNO (a2) >= 0)
4312 return regno1 - regno2;
4313 return 1;
4315 else if (a2 == NULL || ALLOCNO_HARD_REGNO (a2) >= 0)
4316 return -1;
4317 slot_num1 = -ALLOCNO_HARD_REGNO (a1);
4318 slot_num2 = -ALLOCNO_HARD_REGNO (a2);
4319 if ((diff = slot_num1 - slot_num2) != 0)
4320 return (frame_pointer_needed
4321 || (!FRAME_GROWS_DOWNWARD) == STACK_GROWS_DOWNWARD ? diff : -diff);
4322 mode1 = wider_subreg_mode (PSEUDO_REGNO_MODE (regno1),
4323 regno_max_ref_mode[regno1]);
4324 mode2 = wider_subreg_mode (PSEUDO_REGNO_MODE (regno2),
4325 regno_max_ref_mode[regno2]);
4326 if ((diff = compare_sizes_for_sort (GET_MODE_SIZE (mode2),
4327 GET_MODE_SIZE (mode1))) != 0)
4328 return diff;
4329 return regno1 - regno2;
4332 /* Setup REGNO_COALESCED_ALLOCNO_COST and REGNO_COALESCED_ALLOCNO_NUM
4333 for coalesced allocno sets containing allocnos with their regnos
4334 given in array PSEUDO_REGNOS of length N. */
4335 static void
4336 setup_coalesced_allocno_costs_and_nums (int *pseudo_regnos, int n)
4338 int i, num, regno, cost;
4339 ira_allocno_t allocno, a;
4341 for (num = i = 0; i < n; i++)
4343 regno = pseudo_regnos[i];
4344 allocno = ira_regno_allocno_map[regno];
4345 if (allocno == NULL)
4347 regno_coalesced_allocno_cost[regno] = 0;
4348 regno_coalesced_allocno_num[regno] = ++num;
4349 continue;
4351 if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno)
4352 continue;
4353 num++;
4354 for (cost = 0, a = ALLOCNO_COALESCE_DATA (allocno)->next;;
4355 a = ALLOCNO_COALESCE_DATA (a)->next)
4357 cost += ALLOCNO_FREQ (a);
4358 if (a == allocno)
4359 break;
4361 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
4362 a = ALLOCNO_COALESCE_DATA (a)->next)
4364 regno_coalesced_allocno_num[ALLOCNO_REGNO (a)] = num;
4365 regno_coalesced_allocno_cost[ALLOCNO_REGNO (a)] = cost;
4366 if (a == allocno)
4367 break;
4372 /* Collect spilled allocnos representing coalesced allocno sets (the
4373 first coalesced allocno). The collected allocnos are returned
4374 through array SPILLED_COALESCED_ALLOCNOS. The function returns the
4375 number of the collected allocnos. The allocnos are given by their
4376 regnos in array PSEUDO_REGNOS of length N. */
4377 static int
4378 collect_spilled_coalesced_allocnos (int *pseudo_regnos, int n,
4379 ira_allocno_t *spilled_coalesced_allocnos)
4381 int i, num, regno;
4382 ira_allocno_t allocno;
4384 for (num = i = 0; i < n; i++)
4386 regno = pseudo_regnos[i];
4387 allocno = ira_regno_allocno_map[regno];
4388 if (allocno == NULL || ALLOCNO_HARD_REGNO (allocno) >= 0
4389 || ALLOCNO_COALESCE_DATA (allocno)->first != allocno)
4390 continue;
4391 spilled_coalesced_allocnos[num++] = allocno;
4393 return num;
4396 /* Array of live ranges of size IRA_ALLOCNOS_NUM. Live range for
4397 given slot contains live ranges of coalesced allocnos assigned to
4398 given slot. */
4399 static live_range_t *slot_coalesced_allocnos_live_ranges;
4401 /* Return TRUE if coalesced allocnos represented by ALLOCNO has live
4402 ranges intersected with live ranges of coalesced allocnos assigned
4403 to slot with number N. */
4404 static bool
4405 slot_coalesced_allocno_live_ranges_intersect_p (ira_allocno_t allocno, int n)
4407 ira_allocno_t a;
4409 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
4410 a = ALLOCNO_COALESCE_DATA (a)->next)
4412 int i;
4413 int nr = ALLOCNO_NUM_OBJECTS (a);
4414 gcc_assert (ALLOCNO_CAP_MEMBER (a) == NULL);
4415 for (i = 0; i < nr; i++)
4417 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4419 if (ira_live_ranges_intersect_p
4420 (slot_coalesced_allocnos_live_ranges[n],
4421 OBJECT_LIVE_RANGES (obj)))
4422 return true;
4424 if (a == allocno)
4425 break;
4427 return false;
4430 /* Update live ranges of slot to which coalesced allocnos represented
4431 by ALLOCNO were assigned. */
4432 static void
4433 setup_slot_coalesced_allocno_live_ranges (ira_allocno_t allocno)
4435 int i, n;
4436 ira_allocno_t a;
4437 live_range_t r;
4439 n = ALLOCNO_COALESCE_DATA (allocno)->temp;
4440 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
4441 a = ALLOCNO_COALESCE_DATA (a)->next)
4443 int nr = ALLOCNO_NUM_OBJECTS (a);
4444 gcc_assert (ALLOCNO_CAP_MEMBER (a) == NULL);
4445 for (i = 0; i < nr; i++)
4447 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4449 r = ira_copy_live_range_list (OBJECT_LIVE_RANGES (obj));
4450 slot_coalesced_allocnos_live_ranges[n]
4451 = ira_merge_live_ranges
4452 (slot_coalesced_allocnos_live_ranges[n], r);
4454 if (a == allocno)
4455 break;
4459 /* We have coalesced allocnos involving in copies. Coalesce allocnos
4460 further in order to share the same memory stack slot. Allocnos
4461 representing sets of allocnos coalesced before the call are given
4462 in array SPILLED_COALESCED_ALLOCNOS of length NUM. Return TRUE if
4463 some allocnos were coalesced in the function. */
4464 static bool
4465 coalesce_spill_slots (ira_allocno_t *spilled_coalesced_allocnos, int num)
4467 int i, j, n, last_coalesced_allocno_num;
4468 ira_allocno_t allocno, a;
4469 bool merged_p = false;
4470 bitmap set_jump_crosses = regstat_get_setjmp_crosses ();
4472 slot_coalesced_allocnos_live_ranges
4473 = (live_range_t *) ira_allocate (sizeof (live_range_t) * ira_allocnos_num);
4474 memset (slot_coalesced_allocnos_live_ranges, 0,
4475 sizeof (live_range_t) * ira_allocnos_num);
4476 last_coalesced_allocno_num = 0;
4477 /* Coalesce non-conflicting spilled allocnos preferring most
4478 frequently used. */
4479 for (i = 0; i < num; i++)
4481 allocno = spilled_coalesced_allocnos[i];
4482 if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno
4483 || bitmap_bit_p (set_jump_crosses, ALLOCNO_REGNO (allocno))
4484 || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno)))
4485 continue;
4486 for (j = 0; j < i; j++)
4488 a = spilled_coalesced_allocnos[j];
4489 n = ALLOCNO_COALESCE_DATA (a)->temp;
4490 if (ALLOCNO_COALESCE_DATA (a)->first == a
4491 && ! bitmap_bit_p (set_jump_crosses, ALLOCNO_REGNO (a))
4492 && ! ira_equiv_no_lvalue_p (ALLOCNO_REGNO (a))
4493 && ! slot_coalesced_allocno_live_ranges_intersect_p (allocno, n))
4494 break;
4496 if (j >= i)
4498 /* No coalescing: set up number for coalesced allocnos
4499 represented by ALLOCNO. */
4500 ALLOCNO_COALESCE_DATA (allocno)->temp = last_coalesced_allocno_num++;
4501 setup_slot_coalesced_allocno_live_ranges (allocno);
4503 else
4505 allocno_coalesced_p = true;
4506 merged_p = true;
4507 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4508 fprintf (ira_dump_file,
4509 " Coalescing spilled allocnos a%dr%d->a%dr%d\n",
4510 ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno),
4511 ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
4512 ALLOCNO_COALESCE_DATA (allocno)->temp
4513 = ALLOCNO_COALESCE_DATA (a)->temp;
4514 setup_slot_coalesced_allocno_live_ranges (allocno);
4515 merge_allocnos (a, allocno);
4516 ira_assert (ALLOCNO_COALESCE_DATA (a)->first == a);
4519 for (i = 0; i < ira_allocnos_num; i++)
4520 ira_finish_live_range_list (slot_coalesced_allocnos_live_ranges[i]);
4521 ira_free (slot_coalesced_allocnos_live_ranges);
4522 return merged_p;
4525 /* Sort pseudo-register numbers in array PSEUDO_REGNOS of length N for
4526 subsequent assigning stack slots to them in the reload pass. To do
4527 this we coalesce spilled allocnos first to decrease the number of
4528 memory-memory move insns. This function is called by the
4529 reload. */
4530 void
4531 ira_sort_regnos_for_alter_reg (int *pseudo_regnos, int n,
4532 machine_mode *reg_max_ref_mode)
4534 int max_regno = max_reg_num ();
4535 int i, regno, num, slot_num;
4536 ira_allocno_t allocno, a;
4537 ira_allocno_iterator ai;
4538 ira_allocno_t *spilled_coalesced_allocnos;
4540 ira_assert (! ira_use_lra_p);
4542 /* Set up allocnos can be coalesced. */
4543 coloring_allocno_bitmap = ira_allocate_bitmap ();
4544 for (i = 0; i < n; i++)
4546 regno = pseudo_regnos[i];
4547 allocno = ira_regno_allocno_map[regno];
4548 if (allocno != NULL)
4549 bitmap_set_bit (coloring_allocno_bitmap, ALLOCNO_NUM (allocno));
4551 allocno_coalesced_p = false;
4552 processed_coalesced_allocno_bitmap = ira_allocate_bitmap ();
4553 allocno_coalesce_data
4554 = (coalesce_data_t) ira_allocate (sizeof (struct coalesce_data)
4555 * ira_allocnos_num);
4556 /* Initialize coalesce data for allocnos. */
4557 FOR_EACH_ALLOCNO (a, ai)
4559 ALLOCNO_ADD_DATA (a) = allocno_coalesce_data + ALLOCNO_NUM (a);
4560 ALLOCNO_COALESCE_DATA (a)->first = a;
4561 ALLOCNO_COALESCE_DATA (a)->next = a;
4563 coalesce_allocnos ();
4564 ira_free_bitmap (coloring_allocno_bitmap);
4565 regno_coalesced_allocno_cost
4566 = (int *) ira_allocate (max_regno * sizeof (int));
4567 regno_coalesced_allocno_num
4568 = (int *) ira_allocate (max_regno * sizeof (int));
4569 memset (regno_coalesced_allocno_num, 0, max_regno * sizeof (int));
4570 setup_coalesced_allocno_costs_and_nums (pseudo_regnos, n);
4571 /* Sort regnos according frequencies of the corresponding coalesced
4572 allocno sets. */
4573 qsort (pseudo_regnos, n, sizeof (int), coalesced_pseudo_reg_freq_compare);
4574 spilled_coalesced_allocnos
4575 = (ira_allocno_t *) ira_allocate (ira_allocnos_num
4576 * sizeof (ira_allocno_t));
4577 /* Collect allocnos representing the spilled coalesced allocno
4578 sets. */
4579 num = collect_spilled_coalesced_allocnos (pseudo_regnos, n,
4580 spilled_coalesced_allocnos);
4581 if (flag_ira_share_spill_slots
4582 && coalesce_spill_slots (spilled_coalesced_allocnos, num))
4584 setup_coalesced_allocno_costs_and_nums (pseudo_regnos, n);
4585 qsort (pseudo_regnos, n, sizeof (int),
4586 coalesced_pseudo_reg_freq_compare);
4587 num = collect_spilled_coalesced_allocnos (pseudo_regnos, n,
4588 spilled_coalesced_allocnos);
4590 ira_free_bitmap (processed_coalesced_allocno_bitmap);
4591 allocno_coalesced_p = false;
4592 /* Assign stack slot numbers to spilled allocno sets, use smaller
4593 numbers for most frequently used coalesced allocnos. -1 is
4594 reserved for dynamic search of stack slots for pseudos spilled by
4595 the reload. */
4596 slot_num = 1;
4597 for (i = 0; i < num; i++)
4599 allocno = spilled_coalesced_allocnos[i];
4600 if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno
4601 || ALLOCNO_HARD_REGNO (allocno) >= 0
4602 || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno)))
4603 continue;
4604 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4605 fprintf (ira_dump_file, " Slot %d (freq,size):", slot_num);
4606 slot_num++;
4607 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
4608 a = ALLOCNO_COALESCE_DATA (a)->next)
4610 ira_assert (ALLOCNO_HARD_REGNO (a) < 0);
4611 ALLOCNO_HARD_REGNO (a) = -slot_num;
4612 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4614 machine_mode mode = wider_subreg_mode
4615 (PSEUDO_REGNO_MODE (ALLOCNO_REGNO (a)),
4616 reg_max_ref_mode[ALLOCNO_REGNO (a)]);
4617 fprintf (ira_dump_file, " a%dr%d(%d,",
4618 ALLOCNO_NUM (a), ALLOCNO_REGNO (a), ALLOCNO_FREQ (a));
4619 print_dec (GET_MODE_SIZE (mode), ira_dump_file, SIGNED);
4620 fprintf (ira_dump_file, ")\n");
4623 if (a == allocno)
4624 break;
4626 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4627 fprintf (ira_dump_file, "\n");
4629 ira_spilled_reg_stack_slots_num = slot_num - 1;
4630 ira_free (spilled_coalesced_allocnos);
4631 /* Sort regnos according the slot numbers. */
4632 regno_max_ref_mode = reg_max_ref_mode;
4633 qsort (pseudo_regnos, n, sizeof (int), coalesced_pseudo_reg_slot_compare);
4634 FOR_EACH_ALLOCNO (a, ai)
4635 ALLOCNO_ADD_DATA (a) = NULL;
4636 ira_free (allocno_coalesce_data);
4637 ira_free (regno_coalesced_allocno_num);
4638 ira_free (regno_coalesced_allocno_cost);
4643 /* This page contains code used by the reload pass to improve the
4644 final code. */
4646 /* The function is called from reload to mark changes in the
4647 allocation of REGNO made by the reload. Remember that reg_renumber
4648 reflects the change result. */
4649 void
4650 ira_mark_allocation_change (int regno)
4652 ira_allocno_t a = ira_regno_allocno_map[regno];
4653 int old_hard_regno, hard_regno, cost;
4654 enum reg_class aclass = ALLOCNO_CLASS (a);
4656 ira_assert (a != NULL);
4657 hard_regno = reg_renumber[regno];
4658 if ((old_hard_regno = ALLOCNO_HARD_REGNO (a)) == hard_regno)
4659 return;
4660 if (old_hard_regno < 0)
4661 cost = -ALLOCNO_MEMORY_COST (a);
4662 else
4664 ira_assert (ira_class_hard_reg_index[aclass][old_hard_regno] >= 0);
4665 cost = -(ALLOCNO_HARD_REG_COSTS (a) == NULL
4666 ? ALLOCNO_CLASS_COST (a)
4667 : ALLOCNO_HARD_REG_COSTS (a)
4668 [ira_class_hard_reg_index[aclass][old_hard_regno]]);
4669 update_costs_from_copies (a, false, false);
4671 ira_overall_cost -= cost;
4672 ALLOCNO_HARD_REGNO (a) = hard_regno;
4673 if (hard_regno < 0)
4675 ALLOCNO_HARD_REGNO (a) = -1;
4676 cost += ALLOCNO_MEMORY_COST (a);
4678 else if (ira_class_hard_reg_index[aclass][hard_regno] >= 0)
4680 cost += (ALLOCNO_HARD_REG_COSTS (a) == NULL
4681 ? ALLOCNO_CLASS_COST (a)
4682 : ALLOCNO_HARD_REG_COSTS (a)
4683 [ira_class_hard_reg_index[aclass][hard_regno]]);
4684 update_costs_from_copies (a, true, false);
4686 else
4687 /* Reload changed class of the allocno. */
4688 cost = 0;
4689 ira_overall_cost += cost;
4692 /* This function is called when reload deletes memory-memory move. In
4693 this case we marks that the allocation of the corresponding
4694 allocnos should be not changed in future. Otherwise we risk to get
4695 a wrong code. */
4696 void
4697 ira_mark_memory_move_deletion (int dst_regno, int src_regno)
4699 ira_allocno_t dst = ira_regno_allocno_map[dst_regno];
4700 ira_allocno_t src = ira_regno_allocno_map[src_regno];
4702 ira_assert (dst != NULL && src != NULL
4703 && ALLOCNO_HARD_REGNO (dst) < 0
4704 && ALLOCNO_HARD_REGNO (src) < 0);
4705 ALLOCNO_DONT_REASSIGN_P (dst) = true;
4706 ALLOCNO_DONT_REASSIGN_P (src) = true;
4709 /* Try to assign a hard register (except for FORBIDDEN_REGS) to
4710 allocno A and return TRUE in the case of success. */
4711 static bool
4712 allocno_reload_assign (ira_allocno_t a, HARD_REG_SET forbidden_regs)
4714 int hard_regno;
4715 enum reg_class aclass;
4716 int regno = ALLOCNO_REGNO (a);
4717 HARD_REG_SET saved[2];
4718 int i, n;
4720 n = ALLOCNO_NUM_OBJECTS (a);
4721 for (i = 0; i < n; i++)
4723 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4724 saved[i] = OBJECT_TOTAL_CONFLICT_HARD_REGS (obj);
4725 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) |= forbidden_regs;
4726 if (! flag_caller_saves && ALLOCNO_CALLS_CROSSED_NUM (a) != 0)
4727 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) |= ira_need_caller_save_regs (a);
4729 ALLOCNO_ASSIGNED_P (a) = false;
4730 aclass = ALLOCNO_CLASS (a);
4731 update_curr_costs (a);
4732 assign_hard_reg (a, true);
4733 hard_regno = ALLOCNO_HARD_REGNO (a);
4734 reg_renumber[regno] = hard_regno;
4735 if (hard_regno < 0)
4736 ALLOCNO_HARD_REGNO (a) = -1;
4737 else
4739 ira_assert (ira_class_hard_reg_index[aclass][hard_regno] >= 0);
4740 ira_overall_cost
4741 -= (ALLOCNO_MEMORY_COST (a)
4742 - (ALLOCNO_HARD_REG_COSTS (a) == NULL
4743 ? ALLOCNO_CLASS_COST (a)
4744 : ALLOCNO_HARD_REG_COSTS (a)[ira_class_hard_reg_index
4745 [aclass][hard_regno]]));
4746 if (ira_need_caller_save_p (a, hard_regno))
4748 ira_assert (flag_caller_saves);
4749 caller_save_needed = 1;
4753 /* If we found a hard register, modify the RTL for the pseudo
4754 register to show the hard register, and mark the pseudo register
4755 live. */
4756 if (reg_renumber[regno] >= 0)
4758 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4759 fprintf (ira_dump_file, ": reassign to %d\n", reg_renumber[regno]);
4760 SET_REGNO (regno_reg_rtx[regno], reg_renumber[regno]);
4761 mark_home_live (regno);
4763 else if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4764 fprintf (ira_dump_file, "\n");
4765 for (i = 0; i < n; i++)
4767 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4768 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) = saved[i];
4770 return reg_renumber[regno] >= 0;
4773 /* Sort pseudos according their usage frequencies (putting most
4774 frequently ones first). */
4775 static int
4776 pseudo_reg_compare (const void *v1p, const void *v2p)
4778 int regno1 = *(const int *) v1p;
4779 int regno2 = *(const int *) v2p;
4780 int diff;
4782 if ((diff = REG_FREQ (regno2) - REG_FREQ (regno1)) != 0)
4783 return diff;
4784 return regno1 - regno2;
4787 /* Try to allocate hard registers to SPILLED_PSEUDO_REGS (there are
4788 NUM of them) or spilled pseudos conflicting with pseudos in
4789 SPILLED_PSEUDO_REGS. Return TRUE and update SPILLED, if the
4790 allocation has been changed. The function doesn't use
4791 BAD_SPILL_REGS and hard registers in PSEUDO_FORBIDDEN_REGS and
4792 PSEUDO_PREVIOUS_REGS for the corresponding pseudos. The function
4793 is called by the reload pass at the end of each reload
4794 iteration. */
4795 bool
4796 ira_reassign_pseudos (int *spilled_pseudo_regs, int num,
4797 HARD_REG_SET bad_spill_regs,
4798 HARD_REG_SET *pseudo_forbidden_regs,
4799 HARD_REG_SET *pseudo_previous_regs,
4800 bitmap spilled)
4802 int i, n, regno;
4803 bool changed_p;
4804 ira_allocno_t a;
4805 HARD_REG_SET forbidden_regs;
4806 bitmap temp = BITMAP_ALLOC (NULL);
4808 /* Add pseudos which conflict with pseudos already in
4809 SPILLED_PSEUDO_REGS to SPILLED_PSEUDO_REGS. This is preferable
4810 to allocating in two steps as some of the conflicts might have
4811 a higher priority than the pseudos passed in SPILLED_PSEUDO_REGS. */
4812 for (i = 0; i < num; i++)
4813 bitmap_set_bit (temp, spilled_pseudo_regs[i]);
4815 for (i = 0, n = num; i < n; i++)
4817 int nr, j;
4818 int regno = spilled_pseudo_regs[i];
4819 bitmap_set_bit (temp, regno);
4821 a = ira_regno_allocno_map[regno];
4822 nr = ALLOCNO_NUM_OBJECTS (a);
4823 for (j = 0; j < nr; j++)
4825 ira_object_t conflict_obj;
4826 ira_object_t obj = ALLOCNO_OBJECT (a, j);
4827 ira_object_conflict_iterator oci;
4829 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
4831 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
4832 if (ALLOCNO_HARD_REGNO (conflict_a) < 0
4833 && ! ALLOCNO_DONT_REASSIGN_P (conflict_a)
4834 && bitmap_set_bit (temp, ALLOCNO_REGNO (conflict_a)))
4836 spilled_pseudo_regs[num++] = ALLOCNO_REGNO (conflict_a);
4837 /* ?!? This seems wrong. */
4838 bitmap_set_bit (consideration_allocno_bitmap,
4839 ALLOCNO_NUM (conflict_a));
4845 if (num > 1)
4846 qsort (spilled_pseudo_regs, num, sizeof (int), pseudo_reg_compare);
4847 changed_p = false;
4848 /* Try to assign hard registers to pseudos from
4849 SPILLED_PSEUDO_REGS. */
4850 for (i = 0; i < num; i++)
4852 regno = spilled_pseudo_regs[i];
4853 forbidden_regs = (bad_spill_regs
4854 | pseudo_forbidden_regs[regno]
4855 | pseudo_previous_regs[regno]);
4856 gcc_assert (reg_renumber[regno] < 0);
4857 a = ira_regno_allocno_map[regno];
4858 ira_mark_allocation_change (regno);
4859 ira_assert (reg_renumber[regno] < 0);
4860 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4861 fprintf (ira_dump_file,
4862 " Try Assign %d(a%d), cost=%d", regno, ALLOCNO_NUM (a),
4863 ALLOCNO_MEMORY_COST (a)
4864 - ALLOCNO_CLASS_COST (a));
4865 allocno_reload_assign (a, forbidden_regs);
4866 if (reg_renumber[regno] >= 0)
4868 CLEAR_REGNO_REG_SET (spilled, regno);
4869 changed_p = true;
4872 BITMAP_FREE (temp);
4873 return changed_p;
4876 /* The function is called by reload and returns already allocated
4877 stack slot (if any) for REGNO with given INHERENT_SIZE and
4878 TOTAL_SIZE. In the case of failure to find a slot which can be
4879 used for REGNO, the function returns NULL. */
4881 ira_reuse_stack_slot (int regno, poly_uint64 inherent_size,
4882 poly_uint64 total_size)
4884 unsigned int i;
4885 int slot_num, best_slot_num;
4886 int cost, best_cost;
4887 ira_copy_t cp, next_cp;
4888 ira_allocno_t another_allocno, allocno = ira_regno_allocno_map[regno];
4889 rtx x;
4890 bitmap_iterator bi;
4891 class ira_spilled_reg_stack_slot *slot = NULL;
4893 ira_assert (! ira_use_lra_p);
4895 ira_assert (known_eq (inherent_size, PSEUDO_REGNO_BYTES (regno))
4896 && known_le (inherent_size, total_size)
4897 && ALLOCNO_HARD_REGNO (allocno) < 0);
4898 if (! flag_ira_share_spill_slots)
4899 return NULL_RTX;
4900 slot_num = -ALLOCNO_HARD_REGNO (allocno) - 2;
4901 if (slot_num != -1)
4903 slot = &ira_spilled_reg_stack_slots[slot_num];
4904 x = slot->mem;
4906 else
4908 best_cost = best_slot_num = -1;
4909 x = NULL_RTX;
4910 /* It means that the pseudo was spilled in the reload pass, try
4911 to reuse a slot. */
4912 for (slot_num = 0;
4913 slot_num < ira_spilled_reg_stack_slots_num;
4914 slot_num++)
4916 slot = &ira_spilled_reg_stack_slots[slot_num];
4917 if (slot->mem == NULL_RTX)
4918 continue;
4919 if (maybe_lt (slot->width, total_size)
4920 || maybe_lt (GET_MODE_SIZE (GET_MODE (slot->mem)), inherent_size))
4921 continue;
4923 EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs,
4924 FIRST_PSEUDO_REGISTER, i, bi)
4926 another_allocno = ira_regno_allocno_map[i];
4927 if (allocnos_conflict_by_live_ranges_p (allocno,
4928 another_allocno))
4929 goto cont;
4931 for (cost = 0, cp = ALLOCNO_COPIES (allocno);
4932 cp != NULL;
4933 cp = next_cp)
4935 if (cp->first == allocno)
4937 next_cp = cp->next_first_allocno_copy;
4938 another_allocno = cp->second;
4940 else if (cp->second == allocno)
4942 next_cp = cp->next_second_allocno_copy;
4943 another_allocno = cp->first;
4945 else
4946 gcc_unreachable ();
4947 if (cp->insn == NULL_RTX)
4948 continue;
4949 if (bitmap_bit_p (&slot->spilled_regs,
4950 ALLOCNO_REGNO (another_allocno)))
4951 cost += cp->freq;
4953 if (cost > best_cost)
4955 best_cost = cost;
4956 best_slot_num = slot_num;
4958 cont:
4961 if (best_cost >= 0)
4963 slot_num = best_slot_num;
4964 slot = &ira_spilled_reg_stack_slots[slot_num];
4965 SET_REGNO_REG_SET (&slot->spilled_regs, regno);
4966 x = slot->mem;
4967 ALLOCNO_HARD_REGNO (allocno) = -slot_num - 2;
4970 if (x != NULL_RTX)
4972 ira_assert (known_ge (slot->width, total_size));
4973 #ifdef ENABLE_IRA_CHECKING
4974 EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs,
4975 FIRST_PSEUDO_REGISTER, i, bi)
4977 ira_assert (! conflict_by_live_ranges_p (regno, i));
4979 #endif
4980 SET_REGNO_REG_SET (&slot->spilled_regs, regno);
4981 if (internal_flag_ira_verbose > 3 && ira_dump_file)
4983 fprintf (ira_dump_file, " Assigning %d(freq=%d) slot %d of",
4984 regno, REG_FREQ (regno), slot_num);
4985 EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs,
4986 FIRST_PSEUDO_REGISTER, i, bi)
4988 if ((unsigned) regno != i)
4989 fprintf (ira_dump_file, " %d", i);
4991 fprintf (ira_dump_file, "\n");
4994 return x;
4997 /* This is called by reload every time a new stack slot X with
4998 TOTAL_SIZE was allocated for REGNO. We store this info for
4999 subsequent ira_reuse_stack_slot calls. */
5000 void
5001 ira_mark_new_stack_slot (rtx x, int regno, poly_uint64 total_size)
5003 class ira_spilled_reg_stack_slot *slot;
5004 int slot_num;
5005 ira_allocno_t allocno;
5007 ira_assert (! ira_use_lra_p);
5009 ira_assert (known_le (PSEUDO_REGNO_BYTES (regno), total_size));
5010 allocno = ira_regno_allocno_map[regno];
5011 slot_num = -ALLOCNO_HARD_REGNO (allocno) - 2;
5012 if (slot_num == -1)
5014 slot_num = ira_spilled_reg_stack_slots_num++;
5015 ALLOCNO_HARD_REGNO (allocno) = -slot_num - 2;
5017 slot = &ira_spilled_reg_stack_slots[slot_num];
5018 INIT_REG_SET (&slot->spilled_regs);
5019 SET_REGNO_REG_SET (&slot->spilled_regs, regno);
5020 slot->mem = x;
5021 slot->width = total_size;
5022 if (internal_flag_ira_verbose > 3 && ira_dump_file)
5023 fprintf (ira_dump_file, " Assigning %d(freq=%d) a new slot %d\n",
5024 regno, REG_FREQ (regno), slot_num);
5028 /* Return spill cost for pseudo-registers whose numbers are in array
5029 REGNOS (with a negative number as an end marker) for reload with
5030 given IN and OUT for INSN. Return also number points (through
5031 EXCESS_PRESSURE_LIVE_LENGTH) where the pseudo-register lives and
5032 the register pressure is high, number of references of the
5033 pseudo-registers (through NREFS), the number of psuedo registers
5034 whose allocated register wouldn't need saving in the prologue
5035 (through CALL_USED_COUNT), and the first hard regno occupied by the
5036 pseudo-registers (through FIRST_HARD_REGNO). */
5037 static int
5038 calculate_spill_cost (int *regnos, rtx in, rtx out, rtx_insn *insn,
5039 int *excess_pressure_live_length,
5040 int *nrefs, int *call_used_count, int *first_hard_regno)
5042 int i, cost, regno, hard_regno, count, saved_cost;
5043 bool in_p, out_p;
5044 int length;
5045 ira_allocno_t a;
5047 *nrefs = 0;
5048 for (length = count = cost = i = 0;; i++)
5050 regno = regnos[i];
5051 if (regno < 0)
5052 break;
5053 *nrefs += REG_N_REFS (regno);
5054 hard_regno = reg_renumber[regno];
5055 ira_assert (hard_regno >= 0);
5056 a = ira_regno_allocno_map[regno];
5057 length += ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a) / ALLOCNO_NUM_OBJECTS (a);
5058 cost += ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a);
5059 if (in_hard_reg_set_p (crtl->abi->full_reg_clobbers (),
5060 ALLOCNO_MODE (a), hard_regno))
5061 count++;
5062 in_p = in && REG_P (in) && (int) REGNO (in) == hard_regno;
5063 out_p = out && REG_P (out) && (int) REGNO (out) == hard_regno;
5064 if ((in_p || out_p)
5065 && find_regno_note (insn, REG_DEAD, hard_regno) != NULL_RTX)
5067 saved_cost = 0;
5068 if (in_p)
5069 saved_cost += ira_memory_move_cost
5070 [ALLOCNO_MODE (a)][ALLOCNO_CLASS (a)][1];
5071 if (out_p)
5072 saved_cost
5073 += ira_memory_move_cost
5074 [ALLOCNO_MODE (a)][ALLOCNO_CLASS (a)][0];
5075 cost -= REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn)) * saved_cost;
5078 *excess_pressure_live_length = length;
5079 *call_used_count = count;
5080 hard_regno = -1;
5081 if (regnos[0] >= 0)
5083 hard_regno = reg_renumber[regnos[0]];
5085 *first_hard_regno = hard_regno;
5086 return cost;
5089 /* Return TRUE if spilling pseudo-registers whose numbers are in array
5090 REGNOS is better than spilling pseudo-registers with numbers in
5091 OTHER_REGNOS for reload with given IN and OUT for INSN. The
5092 function used by the reload pass to make better register spilling
5093 decisions. */
5094 bool
5095 ira_better_spill_reload_regno_p (int *regnos, int *other_regnos,
5096 rtx in, rtx out, rtx_insn *insn)
5098 int cost, other_cost;
5099 int length, other_length;
5100 int nrefs, other_nrefs;
5101 int call_used_count, other_call_used_count;
5102 int hard_regno, other_hard_regno;
5104 cost = calculate_spill_cost (regnos, in, out, insn,
5105 &length, &nrefs, &call_used_count, &hard_regno);
5106 other_cost = calculate_spill_cost (other_regnos, in, out, insn,
5107 &other_length, &other_nrefs,
5108 &other_call_used_count,
5109 &other_hard_regno);
5110 if (nrefs == 0 && other_nrefs != 0)
5111 return true;
5112 if (nrefs != 0 && other_nrefs == 0)
5113 return false;
5114 if (cost != other_cost)
5115 return cost < other_cost;
5116 if (length != other_length)
5117 return length > other_length;
5118 #ifdef REG_ALLOC_ORDER
5119 if (hard_regno >= 0 && other_hard_regno >= 0)
5120 return (inv_reg_alloc_order[hard_regno]
5121 < inv_reg_alloc_order[other_hard_regno]);
5122 #else
5123 if (call_used_count != other_call_used_count)
5124 return call_used_count > other_call_used_count;
5125 #endif
5126 return false;
5131 /* Allocate and initialize data necessary for assign_hard_reg. */
5132 void
5133 ira_initiate_assign (void)
5135 sorted_allocnos
5136 = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t)
5137 * ira_allocnos_num);
5138 consideration_allocno_bitmap = ira_allocate_bitmap ();
5139 initiate_cost_update ();
5140 allocno_priorities = (int *) ira_allocate (sizeof (int) * ira_allocnos_num);
5141 sorted_copies = (ira_copy_t *) ira_allocate (ira_copies_num
5142 * sizeof (ira_copy_t));
5145 /* Deallocate data used by assign_hard_reg. */
5146 void
5147 ira_finish_assign (void)
5149 ira_free (sorted_allocnos);
5150 ira_free_bitmap (consideration_allocno_bitmap);
5151 finish_cost_update ();
5152 ira_free (allocno_priorities);
5153 ira_free (sorted_copies);
5158 /* Entry function doing color-based register allocation. */
5159 static void
5160 color (void)
5162 allocno_stack_vec.create (ira_allocnos_num);
5163 memset (allocated_hardreg_p, 0, sizeof (allocated_hardreg_p));
5164 ira_initiate_assign ();
5165 do_coloring ();
5166 ira_finish_assign ();
5167 allocno_stack_vec.release ();
5168 move_spill_restore ();
5173 /* This page contains a simple register allocator without usage of
5174 allocno conflicts. This is used for fast allocation for -O0. */
5176 /* Do register allocation by not using allocno conflicts. It uses
5177 only allocno live ranges. The algorithm is close to Chow's
5178 priority coloring. */
5179 static void
5180 fast_allocation (void)
5182 int i, j, k, num, class_size, hard_regno, best_hard_regno, cost, min_cost;
5183 int *costs;
5184 #ifdef STACK_REGS
5185 bool no_stack_reg_p;
5186 #endif
5187 enum reg_class aclass;
5188 machine_mode mode;
5189 ira_allocno_t a;
5190 ira_allocno_iterator ai;
5191 live_range_t r;
5192 HARD_REG_SET conflict_hard_regs, *used_hard_regs;
5194 sorted_allocnos = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t)
5195 * ira_allocnos_num);
5196 num = 0;
5197 FOR_EACH_ALLOCNO (a, ai)
5198 sorted_allocnos[num++] = a;
5199 allocno_priorities = (int *) ira_allocate (sizeof (int) * ira_allocnos_num);
5200 setup_allocno_priorities (sorted_allocnos, num);
5201 used_hard_regs = (HARD_REG_SET *) ira_allocate (sizeof (HARD_REG_SET)
5202 * ira_max_point);
5203 for (i = 0; i < ira_max_point; i++)
5204 CLEAR_HARD_REG_SET (used_hard_regs[i]);
5205 qsort (sorted_allocnos, num, sizeof (ira_allocno_t),
5206 allocno_priority_compare_func);
5207 for (i = 0; i < num; i++)
5209 int nr, l;
5211 a = sorted_allocnos[i];
5212 nr = ALLOCNO_NUM_OBJECTS (a);
5213 CLEAR_HARD_REG_SET (conflict_hard_regs);
5214 for (l = 0; l < nr; l++)
5216 ira_object_t obj = ALLOCNO_OBJECT (a, l);
5217 conflict_hard_regs |= OBJECT_CONFLICT_HARD_REGS (obj);
5218 for (r = OBJECT_LIVE_RANGES (obj); r != NULL; r = r->next)
5219 for (j = r->start; j <= r->finish; j++)
5220 conflict_hard_regs |= used_hard_regs[j];
5222 aclass = ALLOCNO_CLASS (a);
5223 ALLOCNO_ASSIGNED_P (a) = true;
5224 ALLOCNO_HARD_REGNO (a) = -1;
5225 if (hard_reg_set_subset_p (reg_class_contents[aclass],
5226 conflict_hard_regs))
5227 continue;
5228 mode = ALLOCNO_MODE (a);
5229 #ifdef STACK_REGS
5230 no_stack_reg_p = ALLOCNO_NO_STACK_REG_P (a);
5231 #endif
5232 class_size = ira_class_hard_regs_num[aclass];
5233 costs = ALLOCNO_HARD_REG_COSTS (a);
5234 min_cost = INT_MAX;
5235 best_hard_regno = -1;
5236 for (j = 0; j < class_size; j++)
5238 hard_regno = ira_class_hard_regs[aclass][j];
5239 #ifdef STACK_REGS
5240 if (no_stack_reg_p && FIRST_STACK_REG <= hard_regno
5241 && hard_regno <= LAST_STACK_REG)
5242 continue;
5243 #endif
5244 if (ira_hard_reg_set_intersection_p (hard_regno, mode, conflict_hard_regs)
5245 || (TEST_HARD_REG_BIT
5246 (ira_prohibited_class_mode_regs[aclass][mode], hard_regno)))
5247 continue;
5248 if (costs == NULL)
5250 best_hard_regno = hard_regno;
5251 break;
5253 cost = costs[j];
5254 if (min_cost > cost)
5256 min_cost = cost;
5257 best_hard_regno = hard_regno;
5260 if (best_hard_regno < 0)
5261 continue;
5262 ALLOCNO_HARD_REGNO (a) = hard_regno = best_hard_regno;
5263 for (l = 0; l < nr; l++)
5265 ira_object_t obj = ALLOCNO_OBJECT (a, l);
5266 for (r = OBJECT_LIVE_RANGES (obj); r != NULL; r = r->next)
5267 for (k = r->start; k <= r->finish; k++)
5268 used_hard_regs[k] |= ira_reg_mode_hard_regset[hard_regno][mode];
5271 ira_free (sorted_allocnos);
5272 ira_free (used_hard_regs);
5273 ira_free (allocno_priorities);
5274 if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL)
5275 ira_print_disposition (ira_dump_file);
5280 /* Entry function doing coloring. */
5281 void
5282 ira_color (void)
5284 ira_allocno_t a;
5285 ira_allocno_iterator ai;
5287 /* Setup updated costs. */
5288 FOR_EACH_ALLOCNO (a, ai)
5290 ALLOCNO_UPDATED_MEMORY_COST (a) = ALLOCNO_MEMORY_COST (a);
5291 ALLOCNO_UPDATED_CLASS_COST (a) = ALLOCNO_CLASS_COST (a);
5293 if (ira_conflicts_p)
5294 color ();
5295 else
5296 fast_allocation ();