]>
Commit | Line | Data |
---|---|---|
058e97ec | 1 | /* IRA allocation based on graph coloring. |
8d9254fc | 2 | Copyright (C) 2006-2020 Free Software Foundation, Inc. |
058e97ec VM |
3 | Contributed by Vladimir Makarov <vmakarov@redhat.com>. |
4 | ||
5 | This file is part of GCC. | |
6 | ||
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. | |
11 | ||
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. | |
16 | ||
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/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
c7131fb2 | 24 | #include "backend.h" |
957060b5 | 25 | #include "target.h" |
058e97ec | 26 | #include "rtl.h" |
957060b5 AM |
27 | #include "tree.h" |
28 | #include "predict.h" | |
c7131fb2 | 29 | #include "df.h" |
4d0cdd0c | 30 | #include "memmodel.h" |
058e97ec | 31 | #include "tm_p.h" |
957060b5 | 32 | #include "insn-config.h" |
058e97ec | 33 | #include "regs.h" |
957060b5 AM |
34 | #include "ira.h" |
35 | #include "ira-int.h" | |
058e97ec | 36 | #include "reload.h" |
c7131fb2 | 37 | #include "cfgloop.h" |
058e97ec | 38 | |
27508f5f | 39 | typedef struct allocno_hard_regs *allocno_hard_regs_t; |
1756cb66 VM |
40 | |
41 | /* The structure contains information about hard registers can be | |
27508f5f | 42 | assigned to allocnos. Usually it is allocno profitable hard |
1756cb66 VM |
43 | registers but in some cases this set can be a bit different. Major |
44 | reason of the difference is a requirement to use hard register sets | |
45 | that form a tree or a forest (set of trees), i.e. hard register set | |
46 | of a node should contain hard register sets of its subnodes. */ | |
27508f5f | 47 | struct allocno_hard_regs |
1756cb66 VM |
48 | { |
49 | /* Hard registers can be assigned to an allocno. */ | |
50 | HARD_REG_SET set; | |
51 | /* Overall (spilling) cost of all allocnos with given register | |
52 | set. */ | |
a9243bfc | 53 | int64_t cost; |
1756cb66 VM |
54 | }; |
55 | ||
27508f5f | 56 | typedef struct allocno_hard_regs_node *allocno_hard_regs_node_t; |
1756cb66 | 57 | |
27508f5f | 58 | /* A node representing allocno hard registers. Such nodes form a |
1756cb66 | 59 | forest (set of trees). Each subnode of given node in the forest |
27508f5f | 60 | refers for hard register set (usually allocno profitable hard |
1756cb66 VM |
61 | register set) which is a subset of one referred from given |
62 | node. */ | |
27508f5f | 63 | struct allocno_hard_regs_node |
1756cb66 VM |
64 | { |
65 | /* Set up number of the node in preorder traversing of the forest. */ | |
66 | int preorder_num; | |
67 | /* Used for different calculation like finding conflict size of an | |
68 | allocno. */ | |
69 | int check; | |
70 | /* Used for calculation of conflict size of an allocno. The | |
27508f5f | 71 | conflict size of the allocno is maximal number of given allocno |
1756cb66 VM |
72 | hard registers needed for allocation of the conflicting allocnos. |
73 | Given allocno is trivially colored if this number plus the number | |
74 | of hard registers needed for given allocno is not greater than | |
75 | the number of given allocno hard register set. */ | |
76 | int conflict_size; | |
77 | /* The number of hard registers given by member hard_regs. */ | |
78 | int hard_regs_num; | |
79 | /* The following member is used to form the final forest. */ | |
80 | bool used_p; | |
81 | /* Pointer to the corresponding profitable hard registers. */ | |
27508f5f | 82 | allocno_hard_regs_t hard_regs; |
1756cb66 VM |
83 | /* Parent, first subnode, previous and next node with the same |
84 | parent in the forest. */ | |
27508f5f | 85 | allocno_hard_regs_node_t parent, first, prev, next; |
1756cb66 VM |
86 | }; |
87 | ||
3b6d1699 VM |
88 | /* Info about changing hard reg costs of an allocno. */ |
89 | struct update_cost_record | |
90 | { | |
91 | /* Hard regno for which we changed the cost. */ | |
92 | int hard_regno; | |
93 | /* Divisor used when we changed the cost of HARD_REGNO. */ | |
94 | int divisor; | |
95 | /* Next record for given allocno. */ | |
96 | struct update_cost_record *next; | |
97 | }; | |
98 | ||
1756cb66 VM |
99 | /* To decrease footprint of ira_allocno structure we store all data |
100 | needed only for coloring in the following structure. */ | |
101 | struct allocno_color_data | |
102 | { | |
103 | /* TRUE value means that the allocno was not removed yet from the | |
df3e3493 | 104 | conflicting graph during coloring. */ |
1756cb66 VM |
105 | unsigned int in_graph_p : 1; |
106 | /* TRUE if it is put on the stack to make other allocnos | |
107 | colorable. */ | |
108 | unsigned int may_be_spilled_p : 1; | |
27508f5f | 109 | /* TRUE if the allocno is trivially colorable. */ |
1756cb66 VM |
110 | unsigned int colorable_p : 1; |
111 | /* Number of hard registers of the allocno class really | |
112 | available for the allocno allocation. It is number of the | |
113 | profitable hard regs. */ | |
114 | int available_regs_num; | |
8c679205 VM |
115 | /* Sum of frequencies of hard register preferences of all |
116 | conflicting allocnos which are not the coloring stack yet. */ | |
117 | int conflict_allocno_hard_prefs; | |
1756cb66 VM |
118 | /* Allocnos in a bucket (used in coloring) chained by the following |
119 | two members. */ | |
120 | ira_allocno_t next_bucket_allocno; | |
121 | ira_allocno_t prev_bucket_allocno; | |
122 | /* Used for temporary purposes. */ | |
123 | int temp; | |
27508f5f VM |
124 | /* Used to exclude repeated processing. */ |
125 | int last_process; | |
1756cb66 VM |
126 | /* Profitable hard regs available for this pseudo allocation. It |
127 | means that the set excludes unavailable hard regs and hard regs | |
128 | conflicting with given pseudo. They should be of the allocno | |
129 | class. */ | |
130 | HARD_REG_SET profitable_hard_regs; | |
27508f5f VM |
131 | /* The allocno hard registers node. */ |
132 | allocno_hard_regs_node_t hard_regs_node; | |
133 | /* Array of structures allocno_hard_regs_subnode representing | |
134 | given allocno hard registers node (the 1st element in the array) | |
135 | and all its subnodes in the tree (forest) of allocno hard | |
1756cb66 VM |
136 | register nodes (see comments above). */ |
137 | int hard_regs_subnodes_start; | |
2b9c63a2 | 138 | /* The length of the previous array. */ |
1756cb66 | 139 | int hard_regs_subnodes_num; |
3b6d1699 VM |
140 | /* Records about updating allocno hard reg costs from copies. If |
141 | the allocno did not get expected hard register, these records are | |
142 | used to restore original hard reg costs of allocnos connected to | |
143 | this allocno by copies. */ | |
144 | struct update_cost_record *update_cost_records; | |
bf08fb16 VM |
145 | /* Threads. We collect allocnos connected by copies into threads |
146 | and try to assign hard regs to allocnos by threads. */ | |
147 | /* Allocno representing all thread. */ | |
148 | ira_allocno_t first_thread_allocno; | |
149 | /* Allocnos in thread forms a cycle list through the following | |
150 | member. */ | |
151 | ira_allocno_t next_thread_allocno; | |
152 | /* All thread frequency. Defined only for first thread allocno. */ | |
153 | int thread_freq; | |
897a7308 VM |
154 | /* Sum of frequencies of hard register preferences of the allocno. */ |
155 | int hard_reg_prefs; | |
1756cb66 VM |
156 | }; |
157 | ||
158 | /* See above. */ | |
27508f5f | 159 | typedef struct allocno_color_data *allocno_color_data_t; |
1756cb66 | 160 | |
27508f5f VM |
161 | /* Container for storing allocno data concerning coloring. */ |
162 | static allocno_color_data_t allocno_color_data; | |
1756cb66 VM |
163 | |
164 | /* Macro to access the data concerning coloring. */ | |
27508f5f VM |
165 | #define ALLOCNO_COLOR_DATA(a) ((allocno_color_data_t) ALLOCNO_ADD_DATA (a)) |
166 | ||
167 | /* Used for finding allocno colorability to exclude repeated allocno | |
168 | processing and for updating preferencing to exclude repeated | |
169 | allocno processing during assignment. */ | |
170 | static int curr_allocno_process; | |
1756cb66 | 171 | |
058e97ec VM |
172 | /* This file contains code for regional graph coloring, spill/restore |
173 | code placement optimization, and code helping the reload pass to do | |
174 | a better job. */ | |
175 | ||
176 | /* Bitmap of allocnos which should be colored. */ | |
177 | static bitmap coloring_allocno_bitmap; | |
178 | ||
179 | /* Bitmap of allocnos which should be taken into account during | |
180 | coloring. In general case it contains allocnos from | |
181 | coloring_allocno_bitmap plus other already colored conflicting | |
182 | allocnos. */ | |
183 | static bitmap consideration_allocno_bitmap; | |
184 | ||
058e97ec VM |
185 | /* All allocnos sorted according their priorities. */ |
186 | static ira_allocno_t *sorted_allocnos; | |
187 | ||
188 | /* Vec representing the stack of allocnos used during coloring. */ | |
9771b263 | 189 | static vec<ira_allocno_t> allocno_stack_vec; |
058e97ec | 190 | |
71af27d2 OH |
191 | /* Helper for qsort comparison callbacks - return a positive integer if |
192 | X > Y, or a negative value otherwise. Use a conditional expression | |
193 | instead of a difference computation to insulate from possible overflow | |
194 | issues, e.g. X - Y < 0 for some X > 0 and Y < 0. */ | |
195 | #define SORTGT(x,y) (((x) > (y)) ? 1 : -1) | |
196 | ||
058e97ec VM |
197 | \f |
198 | ||
27508f5f | 199 | /* Definition of vector of allocno hard registers. */ |
fe82cdfb | 200 | |
27508f5f | 201 | /* Vector of unique allocno hard registers. */ |
9771b263 | 202 | static vec<allocno_hard_regs_t> allocno_hard_regs_vec; |
1756cb66 | 203 | |
8d67ee55 | 204 | struct allocno_hard_regs_hasher : nofree_ptr_hash <allocno_hard_regs> |
1756cb66 | 205 | { |
67f58944 TS |
206 | static inline hashval_t hash (const allocno_hard_regs *); |
207 | static inline bool equal (const allocno_hard_regs *, | |
208 | const allocno_hard_regs *); | |
4a8fb1a1 | 209 | }; |
1756cb66 | 210 | |
4a8fb1a1 LC |
211 | /* Returns hash value for allocno hard registers V. */ |
212 | inline hashval_t | |
67f58944 | 213 | allocno_hard_regs_hasher::hash (const allocno_hard_regs *hv) |
4a8fb1a1 | 214 | { |
1756cb66 VM |
215 | return iterative_hash (&hv->set, sizeof (HARD_REG_SET), 0); |
216 | } | |
217 | ||
27508f5f | 218 | /* Compares allocno hard registers V1 and V2. */ |
4a8fb1a1 | 219 | inline bool |
67f58944 TS |
220 | allocno_hard_regs_hasher::equal (const allocno_hard_regs *hv1, |
221 | const allocno_hard_regs *hv2) | |
1756cb66 | 222 | { |
a8579651 | 223 | return hv1->set == hv2->set; |
1756cb66 VM |
224 | } |
225 | ||
27508f5f | 226 | /* Hash table of unique allocno hard registers. */ |
c203e8a7 | 227 | static hash_table<allocno_hard_regs_hasher> *allocno_hard_regs_htab; |
1756cb66 | 228 | |
27508f5f VM |
229 | /* Return allocno hard registers in the hash table equal to HV. */ |
230 | static allocno_hard_regs_t | |
231 | find_hard_regs (allocno_hard_regs_t hv) | |
1756cb66 | 232 | { |
c203e8a7 | 233 | return allocno_hard_regs_htab->find (hv); |
1756cb66 VM |
234 | } |
235 | ||
236 | /* Insert allocno hard registers HV in the hash table (if it is not | |
237 | there yet) and return the value which in the table. */ | |
27508f5f VM |
238 | static allocno_hard_regs_t |
239 | insert_hard_regs (allocno_hard_regs_t hv) | |
1756cb66 | 240 | { |
c203e8a7 | 241 | allocno_hard_regs **slot = allocno_hard_regs_htab->find_slot (hv, INSERT); |
1756cb66 VM |
242 | |
243 | if (*slot == NULL) | |
244 | *slot = hv; | |
4a8fb1a1 | 245 | return *slot; |
1756cb66 VM |
246 | } |
247 | ||
27508f5f | 248 | /* Initialize data concerning allocno hard registers. */ |
1756cb66 | 249 | static void |
27508f5f | 250 | init_allocno_hard_regs (void) |
1756cb66 | 251 | { |
9771b263 | 252 | allocno_hard_regs_vec.create (200); |
c203e8a7 TS |
253 | allocno_hard_regs_htab |
254 | = new hash_table<allocno_hard_regs_hasher> (200); | |
1756cb66 VM |
255 | } |
256 | ||
27508f5f | 257 | /* Add (or update info about) allocno hard registers with SET and |
1756cb66 | 258 | COST. */ |
27508f5f | 259 | static allocno_hard_regs_t |
a9243bfc | 260 | add_allocno_hard_regs (HARD_REG_SET set, int64_t cost) |
1756cb66 | 261 | { |
27508f5f VM |
262 | struct allocno_hard_regs temp; |
263 | allocno_hard_regs_t hv; | |
1756cb66 VM |
264 | |
265 | gcc_assert (! hard_reg_set_empty_p (set)); | |
6576d245 | 266 | temp.set = set; |
1756cb66 VM |
267 | if ((hv = find_hard_regs (&temp)) != NULL) |
268 | hv->cost += cost; | |
269 | else | |
270 | { | |
27508f5f VM |
271 | hv = ((struct allocno_hard_regs *) |
272 | ira_allocate (sizeof (struct allocno_hard_regs))); | |
6576d245 | 273 | hv->set = set; |
1756cb66 | 274 | hv->cost = cost; |
9771b263 | 275 | allocno_hard_regs_vec.safe_push (hv); |
1756cb66 VM |
276 | insert_hard_regs (hv); |
277 | } | |
278 | return hv; | |
279 | } | |
280 | ||
281 | /* Finalize data concerning allocno hard registers. */ | |
282 | static void | |
27508f5f | 283 | finish_allocno_hard_regs (void) |
1756cb66 VM |
284 | { |
285 | int i; | |
27508f5f | 286 | allocno_hard_regs_t hv; |
1756cb66 VM |
287 | |
288 | for (i = 0; | |
9771b263 | 289 | allocno_hard_regs_vec.iterate (i, &hv); |
1756cb66 VM |
290 | i++) |
291 | ira_free (hv); | |
c203e8a7 TS |
292 | delete allocno_hard_regs_htab; |
293 | allocno_hard_regs_htab = NULL; | |
9771b263 | 294 | allocno_hard_regs_vec.release (); |
1756cb66 VM |
295 | } |
296 | ||
297 | /* Sort hard regs according to their frequency of usage. */ | |
298 | static int | |
27508f5f | 299 | allocno_hard_regs_compare (const void *v1p, const void *v2p) |
1756cb66 | 300 | { |
27508f5f VM |
301 | allocno_hard_regs_t hv1 = *(const allocno_hard_regs_t *) v1p; |
302 | allocno_hard_regs_t hv2 = *(const allocno_hard_regs_t *) v2p; | |
1756cb66 VM |
303 | |
304 | if (hv2->cost > hv1->cost) | |
305 | return 1; | |
306 | else if (hv2->cost < hv1->cost) | |
307 | return -1; | |
5804f627 | 308 | return SORTGT (allocno_hard_regs_hasher::hash(hv2), allocno_hard_regs_hasher::hash(hv1)); |
1756cb66 VM |
309 | } |
310 | ||
311 | \f | |
312 | ||
313 | /* Used for finding a common ancestor of two allocno hard registers | |
314 | nodes in the forest. We use the current value of | |
315 | 'node_check_tick' to mark all nodes from one node to the top and | |
316 | then walking up from another node until we find a marked node. | |
317 | ||
318 | It is also used to figure out allocno colorability as a mark that | |
319 | we already reset value of member 'conflict_size' for the forest | |
320 | node corresponding to the processed allocno. */ | |
321 | static int node_check_tick; | |
322 | ||
323 | /* Roots of the forest containing hard register sets can be assigned | |
27508f5f VM |
324 | to allocnos. */ |
325 | static allocno_hard_regs_node_t hard_regs_roots; | |
1756cb66 | 326 | |
27508f5f | 327 | /* Definition of vector of allocno hard register nodes. */ |
1756cb66 VM |
328 | |
329 | /* Vector used to create the forest. */ | |
9771b263 | 330 | static vec<allocno_hard_regs_node_t> hard_regs_node_vec; |
1756cb66 | 331 | |
27508f5f | 332 | /* Create and return allocno hard registers node containing allocno |
1756cb66 | 333 | hard registers HV. */ |
27508f5f VM |
334 | static allocno_hard_regs_node_t |
335 | create_new_allocno_hard_regs_node (allocno_hard_regs_t hv) | |
1756cb66 | 336 | { |
27508f5f | 337 | allocno_hard_regs_node_t new_node; |
1756cb66 | 338 | |
27508f5f VM |
339 | new_node = ((struct allocno_hard_regs_node *) |
340 | ira_allocate (sizeof (struct allocno_hard_regs_node))); | |
1756cb66 VM |
341 | new_node->check = 0; |
342 | new_node->hard_regs = hv; | |
343 | new_node->hard_regs_num = hard_reg_set_size (hv->set); | |
344 | new_node->first = NULL; | |
345 | new_node->used_p = false; | |
346 | return new_node; | |
347 | } | |
348 | ||
27508f5f | 349 | /* Add allocno hard registers node NEW_NODE to the forest on its level |
1756cb66 VM |
350 | given by ROOTS. */ |
351 | static void | |
27508f5f VM |
352 | add_new_allocno_hard_regs_node_to_forest (allocno_hard_regs_node_t *roots, |
353 | allocno_hard_regs_node_t new_node) | |
1756cb66 VM |
354 | { |
355 | new_node->next = *roots; | |
356 | if (new_node->next != NULL) | |
357 | new_node->next->prev = new_node; | |
358 | new_node->prev = NULL; | |
359 | *roots = new_node; | |
360 | } | |
361 | ||
27508f5f | 362 | /* Add allocno hard registers HV (or its best approximation if it is |
1756cb66 VM |
363 | not possible) to the forest on its level given by ROOTS. */ |
364 | static void | |
27508f5f VM |
365 | add_allocno_hard_regs_to_forest (allocno_hard_regs_node_t *roots, |
366 | allocno_hard_regs_t hv) | |
1756cb66 VM |
367 | { |
368 | unsigned int i, start; | |
27508f5f | 369 | allocno_hard_regs_node_t node, prev, new_node; |
1756cb66 | 370 | HARD_REG_SET temp_set; |
27508f5f | 371 | allocno_hard_regs_t hv2; |
1756cb66 | 372 | |
9771b263 | 373 | start = hard_regs_node_vec.length (); |
1756cb66 VM |
374 | for (node = *roots; node != NULL; node = node->next) |
375 | { | |
a8579651 | 376 | if (hv->set == node->hard_regs->set) |
1756cb66 VM |
377 | return; |
378 | if (hard_reg_set_subset_p (hv->set, node->hard_regs->set)) | |
379 | { | |
27508f5f | 380 | add_allocno_hard_regs_to_forest (&node->first, hv); |
1756cb66 VM |
381 | return; |
382 | } | |
383 | if (hard_reg_set_subset_p (node->hard_regs->set, hv->set)) | |
9771b263 | 384 | hard_regs_node_vec.safe_push (node); |
1756cb66 VM |
385 | else if (hard_reg_set_intersect_p (hv->set, node->hard_regs->set)) |
386 | { | |
dc333d8f | 387 | temp_set = hv->set & node->hard_regs->set; |
27508f5f VM |
388 | hv2 = add_allocno_hard_regs (temp_set, hv->cost); |
389 | add_allocno_hard_regs_to_forest (&node->first, hv2); | |
1756cb66 VM |
390 | } |
391 | } | |
9771b263 | 392 | if (hard_regs_node_vec.length () |
1756cb66 VM |
393 | > start + 1) |
394 | { | |
395 | /* Create a new node which contains nodes in hard_regs_node_vec. */ | |
396 | CLEAR_HARD_REG_SET (temp_set); | |
397 | for (i = start; | |
9771b263 | 398 | i < hard_regs_node_vec.length (); |
1756cb66 VM |
399 | i++) |
400 | { | |
9771b263 | 401 | node = hard_regs_node_vec[i]; |
44942965 | 402 | temp_set |= node->hard_regs->set; |
1756cb66 | 403 | } |
27508f5f VM |
404 | hv = add_allocno_hard_regs (temp_set, hv->cost); |
405 | new_node = create_new_allocno_hard_regs_node (hv); | |
1756cb66 VM |
406 | prev = NULL; |
407 | for (i = start; | |
9771b263 | 408 | i < hard_regs_node_vec.length (); |
1756cb66 VM |
409 | i++) |
410 | { | |
9771b263 | 411 | node = hard_regs_node_vec[i]; |
1756cb66 VM |
412 | if (node->prev == NULL) |
413 | *roots = node->next; | |
414 | else | |
415 | node->prev->next = node->next; | |
416 | if (node->next != NULL) | |
417 | node->next->prev = node->prev; | |
418 | if (prev == NULL) | |
419 | new_node->first = node; | |
420 | else | |
421 | prev->next = node; | |
422 | node->prev = prev; | |
423 | node->next = NULL; | |
424 | prev = node; | |
425 | } | |
27508f5f | 426 | add_new_allocno_hard_regs_node_to_forest (roots, new_node); |
1756cb66 | 427 | } |
9771b263 | 428 | hard_regs_node_vec.truncate (start); |
1756cb66 VM |
429 | } |
430 | ||
27508f5f | 431 | /* Add allocno hard registers nodes starting with the forest level |
1756cb66 VM |
432 | given by FIRST which contains biggest set inside SET. */ |
433 | static void | |
27508f5f | 434 | collect_allocno_hard_regs_cover (allocno_hard_regs_node_t first, |
1756cb66 VM |
435 | HARD_REG_SET set) |
436 | { | |
27508f5f | 437 | allocno_hard_regs_node_t node; |
1756cb66 VM |
438 | |
439 | ira_assert (first != NULL); | |
440 | for (node = first; node != NULL; node = node->next) | |
441 | if (hard_reg_set_subset_p (node->hard_regs->set, set)) | |
9771b263 | 442 | hard_regs_node_vec.safe_push (node); |
1756cb66 | 443 | else if (hard_reg_set_intersect_p (set, node->hard_regs->set)) |
27508f5f | 444 | collect_allocno_hard_regs_cover (node->first, set); |
1756cb66 VM |
445 | } |
446 | ||
27508f5f | 447 | /* Set up field parent as PARENT in all allocno hard registers nodes |
1756cb66 VM |
448 | in forest given by FIRST. */ |
449 | static void | |
27508f5f VM |
450 | setup_allocno_hard_regs_nodes_parent (allocno_hard_regs_node_t first, |
451 | allocno_hard_regs_node_t parent) | |
1756cb66 | 452 | { |
27508f5f | 453 | allocno_hard_regs_node_t node; |
1756cb66 VM |
454 | |
455 | for (node = first; node != NULL; node = node->next) | |
456 | { | |
457 | node->parent = parent; | |
27508f5f | 458 | setup_allocno_hard_regs_nodes_parent (node->first, node); |
1756cb66 VM |
459 | } |
460 | } | |
461 | ||
27508f5f | 462 | /* Return allocno hard registers node which is a first common ancestor |
1756cb66 | 463 | node of FIRST and SECOND in the forest. */ |
27508f5f VM |
464 | static allocno_hard_regs_node_t |
465 | first_common_ancestor_node (allocno_hard_regs_node_t first, | |
466 | allocno_hard_regs_node_t second) | |
1756cb66 | 467 | { |
27508f5f | 468 | allocno_hard_regs_node_t node; |
1756cb66 VM |
469 | |
470 | node_check_tick++; | |
471 | for (node = first; node != NULL; node = node->parent) | |
472 | node->check = node_check_tick; | |
473 | for (node = second; node != NULL; node = node->parent) | |
474 | if (node->check == node_check_tick) | |
475 | return node; | |
476 | return first_common_ancestor_node (second, first); | |
477 | } | |
478 | ||
479 | /* Print hard reg set SET to F. */ | |
480 | static void | |
481 | print_hard_reg_set (FILE *f, HARD_REG_SET set, bool new_line_p) | |
482 | { | |
a5e3dd5d | 483 | int i, start, end; |
1756cb66 | 484 | |
a5e3dd5d | 485 | for (start = end = -1, i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
1756cb66 | 486 | { |
a5e3dd5d HPN |
487 | bool reg_included = TEST_HARD_REG_BIT (set, i); |
488 | ||
489 | if (reg_included) | |
1756cb66 | 490 | { |
a5e3dd5d | 491 | if (start == -1) |
1756cb66 | 492 | start = i; |
a5e3dd5d | 493 | end = i; |
1756cb66 | 494 | } |
a5e3dd5d | 495 | if (start >= 0 && (!reg_included || i == FIRST_PSEUDO_REGISTER - 1)) |
1756cb66 | 496 | { |
a5e3dd5d | 497 | if (start == end) |
1756cb66 | 498 | fprintf (f, " %d", start); |
a5e3dd5d HPN |
499 | else if (start == end + 1) |
500 | fprintf (f, " %d %d", start, end); | |
1756cb66 | 501 | else |
a5e3dd5d | 502 | fprintf (f, " %d-%d", start, end); |
1756cb66 VM |
503 | start = -1; |
504 | } | |
505 | } | |
506 | if (new_line_p) | |
507 | fprintf (f, "\n"); | |
508 | } | |
509 | ||
27508f5f | 510 | /* Print allocno hard register subforest given by ROOTS and its LEVEL |
1756cb66 VM |
511 | to F. */ |
512 | static void | |
27508f5f | 513 | print_hard_regs_subforest (FILE *f, allocno_hard_regs_node_t roots, |
1756cb66 VM |
514 | int level) |
515 | { | |
516 | int i; | |
27508f5f | 517 | allocno_hard_regs_node_t node; |
1756cb66 VM |
518 | |
519 | for (node = roots; node != NULL; node = node->next) | |
520 | { | |
521 | fprintf (f, " "); | |
522 | for (i = 0; i < level * 2; i++) | |
523 | fprintf (f, " "); | |
524 | fprintf (f, "%d:(", node->preorder_num); | |
525 | print_hard_reg_set (f, node->hard_regs->set, false); | |
16998094 | 526 | fprintf (f, ")@%" PRId64"\n", node->hard_regs->cost); |
1756cb66 VM |
527 | print_hard_regs_subforest (f, node->first, level + 1); |
528 | } | |
529 | } | |
530 | ||
27508f5f | 531 | /* Print the allocno hard register forest to F. */ |
1756cb66 VM |
532 | static void |
533 | print_hard_regs_forest (FILE *f) | |
534 | { | |
535 | fprintf (f, " Hard reg set forest:\n"); | |
536 | print_hard_regs_subforest (f, hard_regs_roots, 1); | |
537 | } | |
538 | ||
27508f5f | 539 | /* Print the allocno hard register forest to stderr. */ |
1756cb66 VM |
540 | void |
541 | ira_debug_hard_regs_forest (void) | |
542 | { | |
543 | print_hard_regs_forest (stderr); | |
544 | } | |
545 | ||
27508f5f | 546 | /* Remove unused allocno hard registers nodes from forest given by its |
1756cb66 VM |
547 | *ROOTS. */ |
548 | static void | |
27508f5f | 549 | remove_unused_allocno_hard_regs_nodes (allocno_hard_regs_node_t *roots) |
1756cb66 | 550 | { |
27508f5f | 551 | allocno_hard_regs_node_t node, prev, next, last; |
1756cb66 VM |
552 | |
553 | for (prev = NULL, node = *roots; node != NULL; node = next) | |
554 | { | |
555 | next = node->next; | |
556 | if (node->used_p) | |
557 | { | |
27508f5f | 558 | remove_unused_allocno_hard_regs_nodes (&node->first); |
1756cb66 VM |
559 | prev = node; |
560 | } | |
561 | else | |
562 | { | |
563 | for (last = node->first; | |
564 | last != NULL && last->next != NULL; | |
565 | last = last->next) | |
566 | ; | |
567 | if (last != NULL) | |
568 | { | |
569 | if (prev == NULL) | |
570 | *roots = node->first; | |
571 | else | |
572 | prev->next = node->first; | |
573 | if (next != NULL) | |
574 | next->prev = last; | |
575 | last->next = next; | |
576 | next = node->first; | |
577 | } | |
578 | else | |
579 | { | |
580 | if (prev == NULL) | |
581 | *roots = next; | |
582 | else | |
583 | prev->next = next; | |
584 | if (next != NULL) | |
585 | next->prev = prev; | |
586 | } | |
587 | ira_free (node); | |
588 | } | |
589 | } | |
590 | } | |
591 | ||
27508f5f | 592 | /* Set up fields preorder_num starting with START_NUM in all allocno |
1756cb66 VM |
593 | hard registers nodes in forest given by FIRST. Return biggest set |
594 | PREORDER_NUM increased by 1. */ | |
595 | static int | |
27508f5f VM |
596 | enumerate_allocno_hard_regs_nodes (allocno_hard_regs_node_t first, |
597 | allocno_hard_regs_node_t parent, | |
598 | int start_num) | |
1756cb66 | 599 | { |
27508f5f | 600 | allocno_hard_regs_node_t node; |
1756cb66 VM |
601 | |
602 | for (node = first; node != NULL; node = node->next) | |
603 | { | |
604 | node->preorder_num = start_num++; | |
605 | node->parent = parent; | |
27508f5f VM |
606 | start_num = enumerate_allocno_hard_regs_nodes (node->first, node, |
607 | start_num); | |
1756cb66 VM |
608 | } |
609 | return start_num; | |
610 | } | |
611 | ||
27508f5f VM |
612 | /* Number of allocno hard registers nodes in the forest. */ |
613 | static int allocno_hard_regs_nodes_num; | |
1756cb66 | 614 | |
27508f5f VM |
615 | /* Table preorder number of allocno hard registers node in the forest |
616 | -> the allocno hard registers node. */ | |
617 | static allocno_hard_regs_node_t *allocno_hard_regs_nodes; | |
1756cb66 VM |
618 | |
619 | /* See below. */ | |
27508f5f | 620 | typedef struct allocno_hard_regs_subnode *allocno_hard_regs_subnode_t; |
1756cb66 VM |
621 | |
622 | /* The structure is used to describes all subnodes (not only immediate | |
27508f5f | 623 | ones) in the mentioned above tree for given allocno hard register |
1756cb66 VM |
624 | node. The usage of such data accelerates calculation of |
625 | colorability of given allocno. */ | |
27508f5f | 626 | struct allocno_hard_regs_subnode |
1756cb66 VM |
627 | { |
628 | /* The conflict size of conflicting allocnos whose hard register | |
629 | sets are equal sets (plus supersets if given node is given | |
27508f5f | 630 | allocno hard registers node) of one in the given node. */ |
1756cb66 VM |
631 | int left_conflict_size; |
632 | /* The summary conflict size of conflicting allocnos whose hard | |
633 | register sets are strict subsets of one in the given node. | |
634 | Overall conflict size is | |
635 | left_conflict_subnodes_size | |
636 | + MIN (max_node_impact - left_conflict_subnodes_size, | |
637 | left_conflict_size) | |
638 | */ | |
639 | short left_conflict_subnodes_size; | |
640 | short max_node_impact; | |
641 | }; | |
642 | ||
27508f5f VM |
643 | /* Container for hard regs subnodes of all allocnos. */ |
644 | static allocno_hard_regs_subnode_t allocno_hard_regs_subnodes; | |
1756cb66 | 645 | |
27508f5f VM |
646 | /* Table (preorder number of allocno hard registers node in the |
647 | forest, preorder number of allocno hard registers subnode) -> index | |
1756cb66 VM |
648 | of the subnode relative to the node. -1 if it is not a |
649 | subnode. */ | |
27508f5f | 650 | static int *allocno_hard_regs_subnode_index; |
1756cb66 | 651 | |
27508f5f VM |
652 | /* Setup arrays ALLOCNO_HARD_REGS_NODES and |
653 | ALLOCNO_HARD_REGS_SUBNODE_INDEX. */ | |
1756cb66 | 654 | static void |
27508f5f | 655 | setup_allocno_hard_regs_subnode_index (allocno_hard_regs_node_t first) |
1756cb66 | 656 | { |
27508f5f | 657 | allocno_hard_regs_node_t node, parent; |
1756cb66 VM |
658 | int index; |
659 | ||
660 | for (node = first; node != NULL; node = node->next) | |
661 | { | |
27508f5f | 662 | allocno_hard_regs_nodes[node->preorder_num] = node; |
1756cb66 VM |
663 | for (parent = node; parent != NULL; parent = parent->parent) |
664 | { | |
27508f5f VM |
665 | index = parent->preorder_num * allocno_hard_regs_nodes_num; |
666 | allocno_hard_regs_subnode_index[index + node->preorder_num] | |
1756cb66 VM |
667 | = node->preorder_num - parent->preorder_num; |
668 | } | |
27508f5f | 669 | setup_allocno_hard_regs_subnode_index (node->first); |
1756cb66 VM |
670 | } |
671 | } | |
672 | ||
27508f5f | 673 | /* Count all allocno hard registers nodes in tree ROOT. */ |
1756cb66 | 674 | static int |
27508f5f | 675 | get_allocno_hard_regs_subnodes_num (allocno_hard_regs_node_t root) |
1756cb66 VM |
676 | { |
677 | int len = 1; | |
678 | ||
679 | for (root = root->first; root != NULL; root = root->next) | |
27508f5f | 680 | len += get_allocno_hard_regs_subnodes_num (root); |
1756cb66 VM |
681 | return len; |
682 | } | |
683 | ||
27508f5f | 684 | /* Build the forest of allocno hard registers nodes and assign each |
1756cb66 VM |
685 | allocno a node from the forest. */ |
686 | static void | |
27508f5f | 687 | form_allocno_hard_regs_nodes_forest (void) |
1756cb66 VM |
688 | { |
689 | unsigned int i, j, size, len; | |
27508f5f | 690 | int start; |
1756cb66 | 691 | ira_allocno_t a; |
27508f5f | 692 | allocno_hard_regs_t hv; |
1756cb66 VM |
693 | bitmap_iterator bi; |
694 | HARD_REG_SET temp; | |
27508f5f VM |
695 | allocno_hard_regs_node_t node, allocno_hard_regs_node; |
696 | allocno_color_data_t allocno_data; | |
1756cb66 VM |
697 | |
698 | node_check_tick = 0; | |
27508f5f | 699 | init_allocno_hard_regs (); |
1756cb66 | 700 | hard_regs_roots = NULL; |
9771b263 | 701 | hard_regs_node_vec.create (100); |
1756cb66 VM |
702 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
703 | if (! TEST_HARD_REG_BIT (ira_no_alloc_regs, i)) | |
704 | { | |
705 | CLEAR_HARD_REG_SET (temp); | |
706 | SET_HARD_REG_BIT (temp, i); | |
27508f5f VM |
707 | hv = add_allocno_hard_regs (temp, 0); |
708 | node = create_new_allocno_hard_regs_node (hv); | |
709 | add_new_allocno_hard_regs_node_to_forest (&hard_regs_roots, node); | |
1756cb66 | 710 | } |
9771b263 | 711 | start = allocno_hard_regs_vec.length (); |
1756cb66 VM |
712 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) |
713 | { | |
714 | a = ira_allocnos[i]; | |
27508f5f VM |
715 | allocno_data = ALLOCNO_COLOR_DATA (a); |
716 | ||
717 | if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs)) | |
718 | continue; | |
719 | hv = (add_allocno_hard_regs | |
720 | (allocno_data->profitable_hard_regs, | |
721 | ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a))); | |
1756cb66 | 722 | } |
d15e5131 | 723 | temp = ~ira_no_alloc_regs; |
27508f5f | 724 | add_allocno_hard_regs (temp, 0); |
9771b263 DN |
725 | qsort (allocno_hard_regs_vec.address () + start, |
726 | allocno_hard_regs_vec.length () - start, | |
27508f5f | 727 | sizeof (allocno_hard_regs_t), allocno_hard_regs_compare); |
1756cb66 | 728 | for (i = start; |
9771b263 | 729 | allocno_hard_regs_vec.iterate (i, &hv); |
1756cb66 VM |
730 | i++) |
731 | { | |
27508f5f | 732 | add_allocno_hard_regs_to_forest (&hard_regs_roots, hv); |
9771b263 | 733 | ira_assert (hard_regs_node_vec.length () == 0); |
1756cb66 VM |
734 | } |
735 | /* We need to set up parent fields for right work of | |
736 | first_common_ancestor_node. */ | |
27508f5f | 737 | setup_allocno_hard_regs_nodes_parent (hard_regs_roots, NULL); |
1756cb66 VM |
738 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) |
739 | { | |
740 | a = ira_allocnos[i]; | |
27508f5f VM |
741 | allocno_data = ALLOCNO_COLOR_DATA (a); |
742 | if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs)) | |
743 | continue; | |
9771b263 | 744 | hard_regs_node_vec.truncate (0); |
27508f5f VM |
745 | collect_allocno_hard_regs_cover (hard_regs_roots, |
746 | allocno_data->profitable_hard_regs); | |
747 | allocno_hard_regs_node = NULL; | |
9771b263 | 748 | for (j = 0; hard_regs_node_vec.iterate (j, &node); j++) |
27508f5f VM |
749 | allocno_hard_regs_node |
750 | = (j == 0 | |
751 | ? node | |
752 | : first_common_ancestor_node (node, allocno_hard_regs_node)); | |
753 | /* That is a temporary storage. */ | |
754 | allocno_hard_regs_node->used_p = true; | |
755 | allocno_data->hard_regs_node = allocno_hard_regs_node; | |
1756cb66 VM |
756 | } |
757 | ira_assert (hard_regs_roots->next == NULL); | |
758 | hard_regs_roots->used_p = true; | |
27508f5f VM |
759 | remove_unused_allocno_hard_regs_nodes (&hard_regs_roots); |
760 | allocno_hard_regs_nodes_num | |
761 | = enumerate_allocno_hard_regs_nodes (hard_regs_roots, NULL, 0); | |
762 | allocno_hard_regs_nodes | |
763 | = ((allocno_hard_regs_node_t *) | |
764 | ira_allocate (allocno_hard_regs_nodes_num | |
765 | * sizeof (allocno_hard_regs_node_t))); | |
766 | size = allocno_hard_regs_nodes_num * allocno_hard_regs_nodes_num; | |
767 | allocno_hard_regs_subnode_index | |
1756cb66 VM |
768 | = (int *) ira_allocate (size * sizeof (int)); |
769 | for (i = 0; i < size; i++) | |
27508f5f VM |
770 | allocno_hard_regs_subnode_index[i] = -1; |
771 | setup_allocno_hard_regs_subnode_index (hard_regs_roots); | |
1756cb66 VM |
772 | start = 0; |
773 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) | |
774 | { | |
775 | a = ira_allocnos[i]; | |
27508f5f VM |
776 | allocno_data = ALLOCNO_COLOR_DATA (a); |
777 | if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs)) | |
778 | continue; | |
779 | len = get_allocno_hard_regs_subnodes_num (allocno_data->hard_regs_node); | |
780 | allocno_data->hard_regs_subnodes_start = start; | |
781 | allocno_data->hard_regs_subnodes_num = len; | |
782 | start += len; | |
1756cb66 | 783 | } |
27508f5f VM |
784 | allocno_hard_regs_subnodes |
785 | = ((allocno_hard_regs_subnode_t) | |
786 | ira_allocate (sizeof (struct allocno_hard_regs_subnode) * start)); | |
9771b263 | 787 | hard_regs_node_vec.release (); |
1756cb66 VM |
788 | } |
789 | ||
27508f5f | 790 | /* Free tree of allocno hard registers nodes given by its ROOT. */ |
1756cb66 | 791 | static void |
27508f5f | 792 | finish_allocno_hard_regs_nodes_tree (allocno_hard_regs_node_t root) |
1756cb66 | 793 | { |
27508f5f | 794 | allocno_hard_regs_node_t child, next; |
1756cb66 VM |
795 | |
796 | for (child = root->first; child != NULL; child = next) | |
797 | { | |
798 | next = child->next; | |
27508f5f | 799 | finish_allocno_hard_regs_nodes_tree (child); |
1756cb66 VM |
800 | } |
801 | ira_free (root); | |
802 | } | |
803 | ||
27508f5f | 804 | /* Finish work with the forest of allocno hard registers nodes. */ |
1756cb66 | 805 | static void |
27508f5f | 806 | finish_allocno_hard_regs_nodes_forest (void) |
1756cb66 | 807 | { |
27508f5f | 808 | allocno_hard_regs_node_t node, next; |
1756cb66 | 809 | |
27508f5f | 810 | ira_free (allocno_hard_regs_subnodes); |
1756cb66 VM |
811 | for (node = hard_regs_roots; node != NULL; node = next) |
812 | { | |
813 | next = node->next; | |
27508f5f | 814 | finish_allocno_hard_regs_nodes_tree (node); |
1756cb66 | 815 | } |
27508f5f VM |
816 | ira_free (allocno_hard_regs_nodes); |
817 | ira_free (allocno_hard_regs_subnode_index); | |
818 | finish_allocno_hard_regs (); | |
1756cb66 VM |
819 | } |
820 | ||
821 | /* Set up left conflict sizes and left conflict subnodes sizes of hard | |
822 | registers subnodes of allocno A. Return TRUE if allocno A is | |
823 | trivially colorable. */ | |
3553f0bb | 824 | static bool |
1756cb66 | 825 | setup_left_conflict_sizes_p (ira_allocno_t a) |
3553f0bb | 826 | { |
27508f5f VM |
827 | int i, k, nobj, start; |
828 | int conflict_size, left_conflict_subnodes_size, node_preorder_num; | |
1756cb66 | 829 | allocno_color_data_t data; |
27508f5f VM |
830 | HARD_REG_SET profitable_hard_regs; |
831 | allocno_hard_regs_subnode_t subnodes; | |
832 | allocno_hard_regs_node_t node; | |
833 | HARD_REG_SET node_set; | |
ac0ab4f7 | 834 | |
1756cb66 | 835 | nobj = ALLOCNO_NUM_OBJECTS (a); |
1756cb66 | 836 | data = ALLOCNO_COLOR_DATA (a); |
27508f5f | 837 | subnodes = allocno_hard_regs_subnodes + data->hard_regs_subnodes_start; |
6576d245 | 838 | profitable_hard_regs = data->profitable_hard_regs; |
27508f5f VM |
839 | node = data->hard_regs_node; |
840 | node_preorder_num = node->preorder_num; | |
6576d245 | 841 | node_set = node->hard_regs->set; |
27508f5f | 842 | node_check_tick++; |
1756cb66 VM |
843 | for (k = 0; k < nobj; k++) |
844 | { | |
1756cb66 VM |
845 | ira_object_t obj = ALLOCNO_OBJECT (a, k); |
846 | ira_object_t conflict_obj; | |
847 | ira_object_conflict_iterator oci; | |
1756cb66 | 848 | |
1756cb66 VM |
849 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) |
850 | { | |
851 | int size; | |
852 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); | |
27508f5f | 853 | allocno_hard_regs_node_t conflict_node, temp_node; |
1756cb66 | 854 | HARD_REG_SET conflict_node_set; |
27508f5f | 855 | allocno_color_data_t conflict_data; |
1756cb66 | 856 | |
27508f5f | 857 | conflict_data = ALLOCNO_COLOR_DATA (conflict_a); |
1756cb66 VM |
858 | if (! ALLOCNO_COLOR_DATA (conflict_a)->in_graph_p |
859 | || ! hard_reg_set_intersect_p (profitable_hard_regs, | |
27508f5f | 860 | conflict_data |
1756cb66 VM |
861 | ->profitable_hard_regs)) |
862 | continue; | |
27508f5f | 863 | conflict_node = conflict_data->hard_regs_node; |
6576d245 | 864 | conflict_node_set = conflict_node->hard_regs->set; |
1756cb66 VM |
865 | if (hard_reg_set_subset_p (node_set, conflict_node_set)) |
866 | temp_node = node; | |
867 | else | |
868 | { | |
869 | ira_assert (hard_reg_set_subset_p (conflict_node_set, node_set)); | |
870 | temp_node = conflict_node; | |
871 | } | |
872 | if (temp_node->check != node_check_tick) | |
873 | { | |
874 | temp_node->check = node_check_tick; | |
875 | temp_node->conflict_size = 0; | |
876 | } | |
877 | size = (ira_reg_class_max_nregs | |
878 | [ALLOCNO_CLASS (conflict_a)][ALLOCNO_MODE (conflict_a)]); | |
879 | if (ALLOCNO_NUM_OBJECTS (conflict_a) > 1) | |
880 | /* We will deal with the subwords individually. */ | |
881 | size = 1; | |
882 | temp_node->conflict_size += size; | |
883 | } | |
27508f5f VM |
884 | } |
885 | for (i = 0; i < data->hard_regs_subnodes_num; i++) | |
886 | { | |
887 | allocno_hard_regs_node_t temp_node; | |
888 | ||
889 | temp_node = allocno_hard_regs_nodes[i + node_preorder_num]; | |
890 | ira_assert (temp_node->preorder_num == i + node_preorder_num); | |
891 | subnodes[i].left_conflict_size = (temp_node->check != node_check_tick | |
892 | ? 0 : temp_node->conflict_size); | |
893 | if (hard_reg_set_subset_p (temp_node->hard_regs->set, | |
894 | profitable_hard_regs)) | |
895 | subnodes[i].max_node_impact = temp_node->hard_regs_num; | |
896 | else | |
1756cb66 | 897 | { |
27508f5f VM |
898 | HARD_REG_SET temp_set; |
899 | int j, n, hard_regno; | |
900 | enum reg_class aclass; | |
901 | ||
dc333d8f | 902 | temp_set = temp_node->hard_regs->set & profitable_hard_regs; |
27508f5f VM |
903 | aclass = ALLOCNO_CLASS (a); |
904 | for (n = 0, j = ira_class_hard_regs_num[aclass] - 1; j >= 0; j--) | |
1756cb66 | 905 | { |
27508f5f VM |
906 | hard_regno = ira_class_hard_regs[aclass][j]; |
907 | if (TEST_HARD_REG_BIT (temp_set, hard_regno)) | |
908 | n++; | |
1756cb66 | 909 | } |
27508f5f | 910 | subnodes[i].max_node_impact = n; |
1756cb66 | 911 | } |
27508f5f VM |
912 | subnodes[i].left_conflict_subnodes_size = 0; |
913 | } | |
914 | start = node_preorder_num * allocno_hard_regs_nodes_num; | |
6e3957da | 915 | for (i = data->hard_regs_subnodes_num - 1; i > 0; i--) |
27508f5f VM |
916 | { |
917 | int size, parent_i; | |
918 | allocno_hard_regs_node_t parent; | |
919 | ||
920 | size = (subnodes[i].left_conflict_subnodes_size | |
921 | + MIN (subnodes[i].max_node_impact | |
922 | - subnodes[i].left_conflict_subnodes_size, | |
923 | subnodes[i].left_conflict_size)); | |
924 | parent = allocno_hard_regs_nodes[i + node_preorder_num]->parent; | |
6e3957da | 925 | gcc_checking_assert(parent); |
27508f5f VM |
926 | parent_i |
927 | = allocno_hard_regs_subnode_index[start + parent->preorder_num]; | |
6e3957da | 928 | gcc_checking_assert(parent_i >= 0); |
27508f5f | 929 | subnodes[parent_i].left_conflict_subnodes_size += size; |
1756cb66 | 930 | } |
27508f5f VM |
931 | left_conflict_subnodes_size = subnodes[0].left_conflict_subnodes_size; |
932 | conflict_size | |
32721b2c ZZ |
933 | = (left_conflict_subnodes_size |
934 | + MIN (subnodes[0].max_node_impact - left_conflict_subnodes_size, | |
935 | subnodes[0].left_conflict_size)); | |
1756cb66 VM |
936 | conflict_size += ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]; |
937 | data->colorable_p = conflict_size <= data->available_regs_num; | |
938 | return data->colorable_p; | |
939 | } | |
ac0ab4f7 | 940 | |
1756cb66 | 941 | /* Update left conflict sizes of hard registers subnodes of allocno A |
27508f5f VM |
942 | after removing allocno REMOVED_A with SIZE from the conflict graph. |
943 | Return TRUE if A is trivially colorable. */ | |
1756cb66 VM |
944 | static bool |
945 | update_left_conflict_sizes_p (ira_allocno_t a, | |
27508f5f | 946 | ira_allocno_t removed_a, int size) |
1756cb66 | 947 | { |
27508f5f | 948 | int i, conflict_size, before_conflict_size, diff, start; |
1756cb66 | 949 | int node_preorder_num, parent_i; |
27508f5f VM |
950 | allocno_hard_regs_node_t node, removed_node, parent; |
951 | allocno_hard_regs_subnode_t subnodes; | |
1756cb66 | 952 | allocno_color_data_t data = ALLOCNO_COLOR_DATA (a); |
1756cb66 VM |
953 | |
954 | ira_assert (! data->colorable_p); | |
27508f5f VM |
955 | node = data->hard_regs_node; |
956 | node_preorder_num = node->preorder_num; | |
957 | removed_node = ALLOCNO_COLOR_DATA (removed_a)->hard_regs_node; | |
958 | ira_assert (hard_reg_set_subset_p (removed_node->hard_regs->set, | |
959 | node->hard_regs->set) | |
960 | || hard_reg_set_subset_p (node->hard_regs->set, | |
961 | removed_node->hard_regs->set)); | |
962 | start = node_preorder_num * allocno_hard_regs_nodes_num; | |
963 | i = allocno_hard_regs_subnode_index[start + removed_node->preorder_num]; | |
964 | if (i < 0) | |
965 | i = 0; | |
966 | subnodes = allocno_hard_regs_subnodes + data->hard_regs_subnodes_start; | |
967 | before_conflict_size | |
968 | = (subnodes[i].left_conflict_subnodes_size | |
969 | + MIN (subnodes[i].max_node_impact | |
970 | - subnodes[i].left_conflict_subnodes_size, | |
971 | subnodes[i].left_conflict_size)); | |
972 | subnodes[i].left_conflict_size -= size; | |
973 | for (;;) | |
ac0ab4f7 | 974 | { |
27508f5f VM |
975 | conflict_size |
976 | = (subnodes[i].left_conflict_subnodes_size | |
977 | + MIN (subnodes[i].max_node_impact | |
978 | - subnodes[i].left_conflict_subnodes_size, | |
979 | subnodes[i].left_conflict_size)); | |
980 | if ((diff = before_conflict_size - conflict_size) == 0) | |
981 | break; | |
982 | ira_assert (conflict_size < before_conflict_size); | |
983 | parent = allocno_hard_regs_nodes[i + node_preorder_num]->parent; | |
984 | if (parent == NULL) | |
985 | break; | |
986 | parent_i | |
987 | = allocno_hard_regs_subnode_index[start + parent->preorder_num]; | |
988 | if (parent_i < 0) | |
989 | break; | |
990 | i = parent_i; | |
1756cb66 VM |
991 | before_conflict_size |
992 | = (subnodes[i].left_conflict_subnodes_size | |
993 | + MIN (subnodes[i].max_node_impact | |
994 | - subnodes[i].left_conflict_subnodes_size, | |
995 | subnodes[i].left_conflict_size)); | |
27508f5f | 996 | subnodes[i].left_conflict_subnodes_size -= diff; |
ac0ab4f7 | 997 | } |
27508f5f VM |
998 | if (i != 0 |
999 | || (conflict_size | |
1000 | + ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)] | |
1001 | > data->available_regs_num)) | |
1002 | return false; | |
1003 | data->colorable_p = true; | |
1004 | return true; | |
3553f0bb VM |
1005 | } |
1006 | ||
27508f5f | 1007 | /* Return true if allocno A has empty profitable hard regs. */ |
3553f0bb | 1008 | static bool |
1756cb66 | 1009 | empty_profitable_hard_regs (ira_allocno_t a) |
3553f0bb | 1010 | { |
27508f5f | 1011 | allocno_color_data_t data = ALLOCNO_COLOR_DATA (a); |
1756cb66 | 1012 | |
27508f5f | 1013 | return hard_reg_set_empty_p (data->profitable_hard_regs); |
3553f0bb VM |
1014 | } |
1015 | ||
1756cb66 VM |
1016 | /* Set up profitable hard registers for each allocno being |
1017 | colored. */ | |
1018 | static void | |
1019 | setup_profitable_hard_regs (void) | |
1020 | { | |
1021 | unsigned int i; | |
1022 | int j, k, nobj, hard_regno, nregs, class_size; | |
1023 | ira_allocno_t a; | |
1024 | bitmap_iterator bi; | |
1025 | enum reg_class aclass; | |
ef4bddc2 | 1026 | machine_mode mode; |
27508f5f | 1027 | allocno_color_data_t data; |
1756cb66 | 1028 | |
8d189b3f VM |
1029 | /* Initial set up from allocno classes and explicitly conflicting |
1030 | hard regs. */ | |
1756cb66 VM |
1031 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) |
1032 | { | |
1033 | a = ira_allocnos[i]; | |
1034 | if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS) | |
1035 | continue; | |
27508f5f VM |
1036 | data = ALLOCNO_COLOR_DATA (a); |
1037 | if (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL | |
b81a2f0d VM |
1038 | && ALLOCNO_CLASS_COST (a) > ALLOCNO_MEMORY_COST (a) |
1039 | /* Do not empty profitable regs for static chain pointer | |
1040 | pseudo when non-local goto is used. */ | |
1041 | && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a))) | |
27508f5f VM |
1042 | CLEAR_HARD_REG_SET (data->profitable_hard_regs); |
1043 | else | |
1756cb66 | 1044 | { |
a2c19e93 | 1045 | mode = ALLOCNO_MODE (a); |
6576d245 RS |
1046 | data->profitable_hard_regs |
1047 | = ira_useful_class_mode_regs[aclass][mode]; | |
27508f5f VM |
1048 | nobj = ALLOCNO_NUM_OBJECTS (a); |
1049 | for (k = 0; k < nobj; k++) | |
1756cb66 | 1050 | { |
27508f5f VM |
1051 | ira_object_t obj = ALLOCNO_OBJECT (a, k); |
1052 | ||
d15e5131 RS |
1053 | data->profitable_hard_regs |
1054 | &= ~OBJECT_TOTAL_CONFLICT_HARD_REGS (obj); | |
1756cb66 VM |
1055 | } |
1056 | } | |
1057 | } | |
8d189b3f | 1058 | /* Exclude hard regs already assigned for conflicting objects. */ |
1756cb66 VM |
1059 | EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, i, bi) |
1060 | { | |
1061 | a = ira_allocnos[i]; | |
1062 | if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS | |
1063 | || ! ALLOCNO_ASSIGNED_P (a) | |
1064 | || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0) | |
1065 | continue; | |
1066 | mode = ALLOCNO_MODE (a); | |
ad474626 | 1067 | nregs = hard_regno_nregs (hard_regno, mode); |
1756cb66 VM |
1068 | nobj = ALLOCNO_NUM_OBJECTS (a); |
1069 | for (k = 0; k < nobj; k++) | |
1070 | { | |
1071 | ira_object_t obj = ALLOCNO_OBJECT (a, k); | |
1072 | ira_object_t conflict_obj; | |
1073 | ira_object_conflict_iterator oci; | |
1074 | ||
1075 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
1076 | { | |
27508f5f VM |
1077 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); |
1078 | ||
1079 | /* We can process the conflict allocno repeatedly with | |
1080 | the same result. */ | |
1756cb66 VM |
1081 | if (nregs == nobj && nregs > 1) |
1082 | { | |
1083 | int num = OBJECT_SUBWORD (conflict_obj); | |
1084 | ||
2805e6c0 | 1085 | if (REG_WORDS_BIG_ENDIAN) |
1756cb66 | 1086 | CLEAR_HARD_REG_BIT |
27508f5f | 1087 | (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs, |
1756cb66 VM |
1088 | hard_regno + nobj - num - 1); |
1089 | else | |
1090 | CLEAR_HARD_REG_BIT | |
27508f5f | 1091 | (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs, |
1756cb66 VM |
1092 | hard_regno + num); |
1093 | } | |
1094 | else | |
d15e5131 RS |
1095 | ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs |
1096 | &= ~ira_reg_mode_hard_regset[hard_regno][mode]; | |
1756cb66 VM |
1097 | } |
1098 | } | |
1099 | } | |
8d189b3f | 1100 | /* Exclude too costly hard regs. */ |
1756cb66 VM |
1101 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) |
1102 | { | |
1103 | int min_cost = INT_MAX; | |
1104 | int *costs; | |
1105 | ||
1106 | a = ira_allocnos[i]; | |
1107 | if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS | |
1108 | || empty_profitable_hard_regs (a)) | |
1109 | continue; | |
27508f5f | 1110 | data = ALLOCNO_COLOR_DATA (a); |
27508f5f VM |
1111 | if ((costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a)) != NULL |
1112 | || (costs = ALLOCNO_HARD_REG_COSTS (a)) != NULL) | |
1756cb66 | 1113 | { |
27508f5f VM |
1114 | class_size = ira_class_hard_regs_num[aclass]; |
1115 | for (j = 0; j < class_size; j++) | |
1756cb66 | 1116 | { |
27508f5f VM |
1117 | hard_regno = ira_class_hard_regs[aclass][j]; |
1118 | if (! TEST_HARD_REG_BIT (data->profitable_hard_regs, | |
1119 | hard_regno)) | |
1120 | continue; | |
b81a2f0d VM |
1121 | if (ALLOCNO_UPDATED_MEMORY_COST (a) < costs[j] |
1122 | /* Do not remove HARD_REGNO for static chain pointer | |
1123 | pseudo when non-local goto is used. */ | |
1124 | && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a))) | |
27508f5f VM |
1125 | CLEAR_HARD_REG_BIT (data->profitable_hard_regs, |
1126 | hard_regno); | |
1127 | else if (min_cost > costs[j]) | |
1128 | min_cost = costs[j]; | |
1756cb66 | 1129 | } |
1756cb66 | 1130 | } |
27508f5f | 1131 | else if (ALLOCNO_UPDATED_MEMORY_COST (a) |
b81a2f0d VM |
1132 | < ALLOCNO_UPDATED_CLASS_COST (a) |
1133 | /* Do not empty profitable regs for static chain | |
1134 | pointer pseudo when non-local goto is used. */ | |
1135 | && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a))) | |
27508f5f | 1136 | CLEAR_HARD_REG_SET (data->profitable_hard_regs); |
1756cb66 VM |
1137 | if (ALLOCNO_UPDATED_CLASS_COST (a) > min_cost) |
1138 | ALLOCNO_UPDATED_CLASS_COST (a) = min_cost; | |
1139 | } | |
1140 | } | |
3553f0bb VM |
1141 | |
1142 | \f | |
1143 | ||
058e97ec VM |
1144 | /* This page contains functions used to choose hard registers for |
1145 | allocnos. */ | |
1146 | ||
3b6d1699 | 1147 | /* Pool for update cost records. */ |
fb0b2914 | 1148 | static object_allocator<update_cost_record> update_cost_record_pool |
fcb87c50 | 1149 | ("update cost records"); |
3b6d1699 VM |
1150 | |
1151 | /* Return new update cost record with given params. */ | |
1152 | static struct update_cost_record * | |
1153 | get_update_cost_record (int hard_regno, int divisor, | |
1154 | struct update_cost_record *next) | |
1155 | { | |
1156 | struct update_cost_record *record; | |
1157 | ||
8b17d27f | 1158 | record = update_cost_record_pool.allocate (); |
3b6d1699 VM |
1159 | record->hard_regno = hard_regno; |
1160 | record->divisor = divisor; | |
1161 | record->next = next; | |
1162 | return record; | |
1163 | } | |
1164 | ||
1165 | /* Free memory for all records in LIST. */ | |
1166 | static void | |
1167 | free_update_cost_record_list (struct update_cost_record *list) | |
1168 | { | |
1169 | struct update_cost_record *next; | |
1170 | ||
1171 | while (list != NULL) | |
1172 | { | |
1173 | next = list->next; | |
8b17d27f | 1174 | update_cost_record_pool.remove (list); |
3b6d1699 VM |
1175 | list = next; |
1176 | } | |
1177 | } | |
1178 | ||
1179 | /* Free memory allocated for all update cost records. */ | |
1180 | static void | |
1181 | finish_update_cost_records (void) | |
1182 | { | |
8b17d27f | 1183 | update_cost_record_pool.release (); |
3b6d1699 VM |
1184 | } |
1185 | ||
058e97ec VM |
1186 | /* Array whose element value is TRUE if the corresponding hard |
1187 | register was already allocated for an allocno. */ | |
1188 | static bool allocated_hardreg_p[FIRST_PSEUDO_REGISTER]; | |
1189 | ||
f754734f | 1190 | /* Describes one element in a queue of allocnos whose costs need to be |
1756cb66 VM |
1191 | updated. Each allocno in the queue is known to have an allocno |
1192 | class. */ | |
f35bf7a9 RS |
1193 | struct update_cost_queue_elem |
1194 | { | |
f754734f RS |
1195 | /* This element is in the queue iff CHECK == update_cost_check. */ |
1196 | int check; | |
1197 | ||
1198 | /* COST_HOP_DIVISOR**N, where N is the length of the shortest path | |
1199 | connecting this allocno to the one being allocated. */ | |
1200 | int divisor; | |
1201 | ||
3133bed5 VM |
1202 | /* Allocno from which we started chaining costs of connected |
1203 | allocnos. */ | |
1204 | ira_allocno_t start; | |
1205 | ||
df3e3493 | 1206 | /* Allocno from which we are chaining costs of connected allocnos. |
3b6d1699 VM |
1207 | It is used not go back in graph of allocnos connected by |
1208 | copies. */ | |
1209 | ira_allocno_t from; | |
1210 | ||
f754734f RS |
1211 | /* The next allocno in the queue, or null if this is the last element. */ |
1212 | ira_allocno_t next; | |
1213 | }; | |
1214 | ||
1215 | /* The first element in a queue of allocnos whose copy costs need to be | |
1216 | updated. Null if the queue is empty. */ | |
1217 | static ira_allocno_t update_cost_queue; | |
1218 | ||
1219 | /* The last element in the queue described by update_cost_queue. | |
1220 | Not valid if update_cost_queue is null. */ | |
1221 | static struct update_cost_queue_elem *update_cost_queue_tail; | |
1222 | ||
1223 | /* A pool of elements in the queue described by update_cost_queue. | |
1224 | Elements are indexed by ALLOCNO_NUM. */ | |
1225 | static struct update_cost_queue_elem *update_cost_queue_elems; | |
058e97ec | 1226 | |
3b6d1699 | 1227 | /* The current value of update_costs_from_copies call count. */ |
058e97ec VM |
1228 | static int update_cost_check; |
1229 | ||
1230 | /* Allocate and initialize data necessary for function | |
c73ccc80 | 1231 | update_costs_from_copies. */ |
058e97ec VM |
1232 | static void |
1233 | initiate_cost_update (void) | |
1234 | { | |
f754734f RS |
1235 | size_t size; |
1236 | ||
1237 | size = ira_allocnos_num * sizeof (struct update_cost_queue_elem); | |
1238 | update_cost_queue_elems | |
1239 | = (struct update_cost_queue_elem *) ira_allocate (size); | |
1240 | memset (update_cost_queue_elems, 0, size); | |
058e97ec VM |
1241 | update_cost_check = 0; |
1242 | } | |
1243 | ||
3b6d1699 | 1244 | /* Deallocate data used by function update_costs_from_copies. */ |
058e97ec VM |
1245 | static void |
1246 | finish_cost_update (void) | |
1247 | { | |
0eeb2240 | 1248 | ira_free (update_cost_queue_elems); |
3b6d1699 | 1249 | finish_update_cost_records (); |
058e97ec VM |
1250 | } |
1251 | ||
a7f32992 VM |
1252 | /* When we traverse allocnos to update hard register costs, the cost |
1253 | divisor will be multiplied by the following macro value for each | |
1254 | hop from given allocno to directly connected allocnos. */ | |
1255 | #define COST_HOP_DIVISOR 4 | |
1256 | ||
f754734f | 1257 | /* Start a new cost-updating pass. */ |
058e97ec | 1258 | static void |
f754734f | 1259 | start_update_cost (void) |
058e97ec | 1260 | { |
f754734f RS |
1261 | update_cost_check++; |
1262 | update_cost_queue = NULL; | |
1263 | } | |
058e97ec | 1264 | |
3133bed5 | 1265 | /* Add (ALLOCNO, START, FROM, DIVISOR) to the end of update_cost_queue, unless |
1756cb66 | 1266 | ALLOCNO is already in the queue, or has NO_REGS class. */ |
f754734f | 1267 | static inline void |
3133bed5 VM |
1268 | queue_update_cost (ira_allocno_t allocno, ira_allocno_t start, |
1269 | ira_allocno_t from, int divisor) | |
f754734f RS |
1270 | { |
1271 | struct update_cost_queue_elem *elem; | |
1272 | ||
1273 | elem = &update_cost_queue_elems[ALLOCNO_NUM (allocno)]; | |
1274 | if (elem->check != update_cost_check | |
1756cb66 | 1275 | && ALLOCNO_CLASS (allocno) != NO_REGS) |
058e97ec | 1276 | { |
f754734f | 1277 | elem->check = update_cost_check; |
3133bed5 | 1278 | elem->start = start; |
3b6d1699 | 1279 | elem->from = from; |
f754734f RS |
1280 | elem->divisor = divisor; |
1281 | elem->next = NULL; | |
1282 | if (update_cost_queue == NULL) | |
1283 | update_cost_queue = allocno; | |
058e97ec | 1284 | else |
f754734f RS |
1285 | update_cost_queue_tail->next = allocno; |
1286 | update_cost_queue_tail = elem; | |
058e97ec VM |
1287 | } |
1288 | } | |
1289 | ||
3b6d1699 | 1290 | /* Try to remove the first element from update_cost_queue. Return |
3133bed5 VM |
1291 | false if the queue was empty, otherwise make (*ALLOCNO, *START, |
1292 | *FROM, *DIVISOR) describe the removed element. */ | |
f754734f | 1293 | static inline bool |
3133bed5 VM |
1294 | get_next_update_cost (ira_allocno_t *allocno, ira_allocno_t *start, |
1295 | ira_allocno_t *from, int *divisor) | |
058e97ec | 1296 | { |
f754734f RS |
1297 | struct update_cost_queue_elem *elem; |
1298 | ||
1299 | if (update_cost_queue == NULL) | |
1300 | return false; | |
1301 | ||
1302 | *allocno = update_cost_queue; | |
1303 | elem = &update_cost_queue_elems[ALLOCNO_NUM (*allocno)]; | |
3133bed5 | 1304 | *start = elem->start; |
3b6d1699 | 1305 | *from = elem->from; |
f754734f RS |
1306 | *divisor = elem->divisor; |
1307 | update_cost_queue = elem->next; | |
1308 | return true; | |
058e97ec VM |
1309 | } |
1310 | ||
86f0bef3 VM |
1311 | /* Increase costs of HARD_REGNO by UPDATE_COST and conflict cost by |
1312 | UPDATE_CONFLICT_COST for ALLOCNO. Return true if we really | |
1313 | modified the cost. */ | |
3b6d1699 | 1314 | static bool |
86f0bef3 VM |
1315 | update_allocno_cost (ira_allocno_t allocno, int hard_regno, |
1316 | int update_cost, int update_conflict_cost) | |
3b6d1699 VM |
1317 | { |
1318 | int i; | |
1319 | enum reg_class aclass = ALLOCNO_CLASS (allocno); | |
1320 | ||
1321 | i = ira_class_hard_reg_index[aclass][hard_regno]; | |
1322 | if (i < 0) | |
1323 | return false; | |
1324 | ira_allocate_and_set_or_copy_costs | |
1325 | (&ALLOCNO_UPDATED_HARD_REG_COSTS (allocno), aclass, | |
1326 | ALLOCNO_UPDATED_CLASS_COST (allocno), | |
1327 | ALLOCNO_HARD_REG_COSTS (allocno)); | |
1328 | ira_allocate_and_set_or_copy_costs | |
1329 | (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno), | |
1330 | aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (allocno)); | |
1331 | ALLOCNO_UPDATED_HARD_REG_COSTS (allocno)[i] += update_cost; | |
86f0bef3 | 1332 | ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno)[i] += update_conflict_cost; |
3b6d1699 VM |
1333 | return true; |
1334 | } | |
1335 | ||
3133bed5 VM |
1336 | /* Return TRUE if allocnos A1 and A2 conflicts. Here we are |
1337 | interesting only in conflicts of allocnos with intersected allocno | |
1338 | classes. */ | |
1339 | static bool | |
1340 | allocnos_conflict_p (ira_allocno_t a1, ira_allocno_t a2) | |
1341 | { | |
1342 | ira_object_t obj, conflict_obj; | |
1343 | ira_object_conflict_iterator oci; | |
1344 | int word, nwords = ALLOCNO_NUM_OBJECTS (a1); | |
1345 | ||
1346 | for (word = 0; word < nwords; word++) | |
1347 | { | |
1348 | obj = ALLOCNO_OBJECT (a1, word); | |
1349 | /* Take preferences of conflicting allocnos into account. */ | |
1350 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
1351 | if (OBJECT_ALLOCNO (conflict_obj) == a2) | |
1352 | return true; | |
1353 | } | |
1354 | return false; | |
1355 | } | |
1356 | ||
3b6d1699 VM |
1357 | /* Update (decrease if DECR_P) HARD_REGNO cost of allocnos connected |
1358 | by copies to ALLOCNO to increase chances to remove some copies as | |
74dc179a VM |
1359 | the result of subsequent assignment. Update conflict costs. |
1360 | Record cost updates if RECORD_P is true. */ | |
a7f32992 | 1361 | static void |
3b6d1699 | 1362 | update_costs_from_allocno (ira_allocno_t allocno, int hard_regno, |
74dc179a | 1363 | int divisor, bool decr_p, bool record_p) |
a7f32992 | 1364 | { |
86f0bef3 | 1365 | int cost, update_cost, update_conflict_cost; |
ef4bddc2 | 1366 | machine_mode mode; |
1756cb66 | 1367 | enum reg_class rclass, aclass; |
3133bed5 | 1368 | ira_allocno_t another_allocno, start = allocno, from = NULL; |
a7f32992 VM |
1369 | ira_copy_t cp, next_cp; |
1370 | ||
f754734f | 1371 | rclass = REGNO_REG_CLASS (hard_regno); |
f754734f | 1372 | do |
a7f32992 | 1373 | { |
f754734f | 1374 | mode = ALLOCNO_MODE (allocno); |
1756cb66 | 1375 | ira_init_register_move_cost_if_necessary (mode); |
f754734f | 1376 | for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp) |
a7f32992 | 1377 | { |
f754734f | 1378 | if (cp->first == allocno) |
a7f32992 | 1379 | { |
f754734f RS |
1380 | next_cp = cp->next_first_allocno_copy; |
1381 | another_allocno = cp->second; | |
1382 | } | |
1383 | else if (cp->second == allocno) | |
1384 | { | |
1385 | next_cp = cp->next_second_allocno_copy; | |
1386 | another_allocno = cp->first; | |
a7f32992 | 1387 | } |
f754734f RS |
1388 | else |
1389 | gcc_unreachable (); | |
1390 | ||
3133bed5 | 1391 | if (another_allocno == from |
74dc179a VM |
1392 | || (ALLOCNO_COLOR_DATA (another_allocno) != NULL |
1393 | && (ALLOCNO_COLOR_DATA (allocno)->first_thread_allocno | |
1394 | != ALLOCNO_COLOR_DATA (another_allocno)->first_thread_allocno))) | |
3b6d1699 VM |
1395 | continue; |
1396 | ||
1756cb66 VM |
1397 | aclass = ALLOCNO_CLASS (another_allocno); |
1398 | if (! TEST_HARD_REG_BIT (reg_class_contents[aclass], | |
6042d1dd | 1399 | hard_regno) |
f754734f RS |
1400 | || ALLOCNO_ASSIGNED_P (another_allocno)) |
1401 | continue; | |
1402 | ||
b3ad445f RS |
1403 | /* If we have different modes use the smallest one. It is |
1404 | a sub-register move. It is hard to predict what LRA | |
1405 | will reload (the pseudo or its sub-register) but LRA | |
1406 | will try to minimize the data movement. Also for some | |
1407 | register classes bigger modes might be invalid, | |
1408 | e.g. DImode for AREG on x86. For such cases the | |
1409 | register move cost will be maximal. */ | |
1410 | mode = narrower_subreg_mode (mode, ALLOCNO_MODE (cp->second)); | |
0046f8d7 | 1411 | ira_init_register_move_cost_if_necessary (mode); |
e2323a2b | 1412 | |
f754734f | 1413 | cost = (cp->second == allocno |
1756cb66 VM |
1414 | ? ira_register_move_cost[mode][rclass][aclass] |
1415 | : ira_register_move_cost[mode][aclass][rclass]); | |
f754734f RS |
1416 | if (decr_p) |
1417 | cost = -cost; | |
1418 | ||
3133bed5 | 1419 | update_cost = cp->freq * cost / divisor; |
74dc179a | 1420 | update_conflict_cost = update_cost; |
86f0bef3 | 1421 | |
74dc179a VM |
1422 | if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL) |
1423 | fprintf (ira_dump_file, | |
1424 | " a%dr%d (hr%d): update cost by %d, conflict cost by %d\n", | |
1425 | ALLOCNO_NUM (another_allocno), ALLOCNO_REGNO (another_allocno), | |
1426 | hard_regno, update_cost, update_conflict_cost); | |
f754734f RS |
1427 | if (update_cost == 0) |
1428 | continue; | |
1429 | ||
86f0bef3 VM |
1430 | if (! update_allocno_cost (another_allocno, hard_regno, |
1431 | update_cost, update_conflict_cost)) | |
1756cb66 | 1432 | continue; |
3133bed5 VM |
1433 | queue_update_cost (another_allocno, start, allocno, |
1434 | divisor * COST_HOP_DIVISOR); | |
3b6d1699 VM |
1435 | if (record_p && ALLOCNO_COLOR_DATA (another_allocno) != NULL) |
1436 | ALLOCNO_COLOR_DATA (another_allocno)->update_cost_records | |
1437 | = get_update_cost_record (hard_regno, divisor, | |
1438 | ALLOCNO_COLOR_DATA (another_allocno) | |
1439 | ->update_cost_records); | |
a7f32992 | 1440 | } |
a7f32992 | 1441 | } |
3133bed5 | 1442 | while (get_next_update_cost (&allocno, &start, &from, &divisor)); |
3b6d1699 VM |
1443 | } |
1444 | ||
1445 | /* Decrease preferred ALLOCNO hard register costs and costs of | |
1446 | allocnos connected to ALLOCNO through copy. */ | |
1447 | static void | |
1448 | update_costs_from_prefs (ira_allocno_t allocno) | |
1449 | { | |
1450 | ira_pref_t pref; | |
1451 | ||
1452 | start_update_cost (); | |
1453 | for (pref = ALLOCNO_PREFS (allocno); pref != NULL; pref = pref->next_pref) | |
74dc179a VM |
1454 | { |
1455 | if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL) | |
1456 | fprintf (ira_dump_file, " Start updating from pref of hr%d for a%dr%d:\n", | |
1457 | pref->hard_regno, ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno)); | |
1458 | update_costs_from_allocno (allocno, pref->hard_regno, | |
1459 | COST_HOP_DIVISOR, true, true); | |
1460 | } | |
3b6d1699 VM |
1461 | } |
1462 | ||
1463 | /* Update (decrease if DECR_P) the cost of allocnos connected to | |
1464 | ALLOCNO through copies to increase chances to remove some copies as | |
1465 | the result of subsequent assignment. ALLOCNO was just assigned to | |
c73ccc80 | 1466 | a hard register. Record cost updates if RECORD_P is true. */ |
3b6d1699 | 1467 | static void |
c73ccc80 | 1468 | update_costs_from_copies (ira_allocno_t allocno, bool decr_p, bool record_p) |
3b6d1699 VM |
1469 | { |
1470 | int hard_regno; | |
1471 | ||
1472 | hard_regno = ALLOCNO_HARD_REGNO (allocno); | |
1473 | ira_assert (hard_regno >= 0 && ALLOCNO_CLASS (allocno) != NO_REGS); | |
1474 | start_update_cost (); | |
74dc179a VM |
1475 | if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL) |
1476 | fprintf (ira_dump_file, " Start updating from a%dr%d by copies:\n", | |
1477 | ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno)); | |
1478 | update_costs_from_allocno (allocno, hard_regno, 1, decr_p, record_p); | |
3b6d1699 VM |
1479 | } |
1480 | ||
8c679205 VM |
1481 | /* Update conflict_allocno_hard_prefs of allocnos conflicting with |
1482 | ALLOCNO. */ | |
1483 | static void | |
1484 | update_conflict_allocno_hard_prefs (ira_allocno_t allocno) | |
1485 | { | |
1486 | int l, nr = ALLOCNO_NUM_OBJECTS (allocno); | |
1487 | ||
1488 | for (l = 0; l < nr; l++) | |
1489 | { | |
1490 | ira_object_t conflict_obj, obj = ALLOCNO_OBJECT (allocno, l); | |
1491 | ira_object_conflict_iterator oci; | |
1492 | ||
1493 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
1494 | { | |
1495 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); | |
1496 | allocno_color_data_t conflict_data = ALLOCNO_COLOR_DATA (conflict_a); | |
1497 | ira_pref_t pref; | |
1498 | ||
1499 | if (!(hard_reg_set_intersect_p | |
1500 | (ALLOCNO_COLOR_DATA (allocno)->profitable_hard_regs, | |
1501 | conflict_data->profitable_hard_regs))) | |
1502 | continue; | |
1503 | for (pref = ALLOCNO_PREFS (allocno); | |
1504 | pref != NULL; | |
1505 | pref = pref->next_pref) | |
1506 | conflict_data->conflict_allocno_hard_prefs += pref->freq; | |
1507 | } | |
1508 | } | |
1509 | } | |
1510 | ||
3b6d1699 VM |
1511 | /* Restore costs of allocnos connected to ALLOCNO by copies as it was |
1512 | before updating costs of these allocnos from given allocno. This | |
1513 | is a wise thing to do as if given allocno did not get an expected | |
1514 | hard reg, using smaller cost of the hard reg for allocnos connected | |
1515 | by copies to given allocno becomes actually misleading. Free all | |
1516 | update cost records for ALLOCNO as we don't need them anymore. */ | |
1517 | static void | |
1518 | restore_costs_from_copies (ira_allocno_t allocno) | |
1519 | { | |
1520 | struct update_cost_record *records, *curr; | |
1521 | ||
1522 | if (ALLOCNO_COLOR_DATA (allocno) == NULL) | |
1523 | return; | |
1524 | records = ALLOCNO_COLOR_DATA (allocno)->update_cost_records; | |
1525 | start_update_cost (); | |
74dc179a VM |
1526 | if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL) |
1527 | fprintf (ira_dump_file, " Start restoring from a%dr%d:\n", | |
1528 | ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno)); | |
3b6d1699 VM |
1529 | for (curr = records; curr != NULL; curr = curr->next) |
1530 | update_costs_from_allocno (allocno, curr->hard_regno, | |
74dc179a | 1531 | curr->divisor, true, false); |
3b6d1699 VM |
1532 | free_update_cost_record_list (records); |
1533 | ALLOCNO_COLOR_DATA (allocno)->update_cost_records = NULL; | |
f754734f RS |
1534 | } |
1535 | ||
7db7ed3c | 1536 | /* This function updates COSTS (decrease if DECR_P) for hard_registers |
1756cb66 | 1537 | of ACLASS by conflict costs of the unassigned allocnos |
7db7ed3c VM |
1538 | connected by copies with allocnos in update_cost_queue. This |
1539 | update increases chances to remove some copies. */ | |
f754734f | 1540 | static void |
1756cb66 | 1541 | update_conflict_hard_regno_costs (int *costs, enum reg_class aclass, |
7db7ed3c | 1542 | bool decr_p) |
f754734f RS |
1543 | { |
1544 | int i, cost, class_size, freq, mult, div, divisor; | |
7db7ed3c | 1545 | int index, hard_regno; |
f754734f RS |
1546 | int *conflict_costs; |
1547 | bool cont_p; | |
1756cb66 | 1548 | enum reg_class another_aclass; |
3133bed5 | 1549 | ira_allocno_t allocno, another_allocno, start, from; |
f754734f RS |
1550 | ira_copy_t cp, next_cp; |
1551 | ||
3133bed5 | 1552 | while (get_next_update_cost (&allocno, &start, &from, &divisor)) |
f754734f RS |
1553 | for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp) |
1554 | { | |
1555 | if (cp->first == allocno) | |
1556 | { | |
1557 | next_cp = cp->next_first_allocno_copy; | |
1558 | another_allocno = cp->second; | |
1559 | } | |
1560 | else if (cp->second == allocno) | |
1561 | { | |
1562 | next_cp = cp->next_second_allocno_copy; | |
1563 | another_allocno = cp->first; | |
1564 | } | |
1565 | else | |
1566 | gcc_unreachable (); | |
3b6d1699 | 1567 | |
3133bed5 VM |
1568 | if (another_allocno == from |
1569 | || allocnos_conflict_p (another_allocno, start)) | |
3b6d1699 VM |
1570 | continue; |
1571 | ||
1756cb66 VM |
1572 | another_aclass = ALLOCNO_CLASS (another_allocno); |
1573 | if (! ira_reg_classes_intersect_p[aclass][another_aclass] | |
f754734f | 1574 | || ALLOCNO_ASSIGNED_P (another_allocno) |
1756cb66 | 1575 | || ALLOCNO_COLOR_DATA (another_allocno)->may_be_spilled_p) |
f754734f | 1576 | continue; |
1756cb66 | 1577 | class_size = ira_class_hard_regs_num[another_aclass]; |
f754734f RS |
1578 | ira_allocate_and_copy_costs |
1579 | (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno), | |
1756cb66 | 1580 | another_aclass, ALLOCNO_CONFLICT_HARD_REG_COSTS (another_allocno)); |
f754734f RS |
1581 | conflict_costs |
1582 | = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno); | |
1583 | if (conflict_costs == NULL) | |
1584 | cont_p = true; | |
1585 | else | |
1586 | { | |
1587 | mult = cp->freq; | |
1588 | freq = ALLOCNO_FREQ (another_allocno); | |
1589 | if (freq == 0) | |
1590 | freq = 1; | |
1591 | div = freq * divisor; | |
1592 | cont_p = false; | |
1593 | for (i = class_size - 1; i >= 0; i--) | |
1594 | { | |
1756cb66 | 1595 | hard_regno = ira_class_hard_regs[another_aclass][i]; |
7db7ed3c | 1596 | ira_assert (hard_regno >= 0); |
1756cb66 | 1597 | index = ira_class_hard_reg_index[aclass][hard_regno]; |
7db7ed3c VM |
1598 | if (index < 0) |
1599 | continue; | |
7879aabe | 1600 | cost = (int) (((int64_t) conflict_costs [i] * mult) / div); |
f754734f RS |
1601 | if (cost == 0) |
1602 | continue; | |
1603 | cont_p = true; | |
1604 | if (decr_p) | |
1605 | cost = -cost; | |
7db7ed3c | 1606 | costs[index] += cost; |
f754734f RS |
1607 | } |
1608 | } | |
1609 | /* Probably 5 hops will be enough. */ | |
1610 | if (cont_p | |
1611 | && divisor <= (COST_HOP_DIVISOR | |
1612 | * COST_HOP_DIVISOR | |
1613 | * COST_HOP_DIVISOR | |
1614 | * COST_HOP_DIVISOR)) | |
3133bed5 | 1615 | queue_update_cost (another_allocno, start, from, divisor * COST_HOP_DIVISOR); |
f754734f | 1616 | } |
a7f32992 VM |
1617 | } |
1618 | ||
27508f5f VM |
1619 | /* Set up conflicting (through CONFLICT_REGS) for each object of |
1620 | allocno A and the start allocno profitable regs (through | |
1621 | START_PROFITABLE_REGS). Remember that the start profitable regs | |
67914693 | 1622 | exclude hard regs which cannot hold value of mode of allocno A. |
27508f5f VM |
1623 | This covers mostly cases when multi-register value should be |
1624 | aligned. */ | |
1756cb66 | 1625 | static inline void |
27508f5f VM |
1626 | get_conflict_and_start_profitable_regs (ira_allocno_t a, bool retry_p, |
1627 | HARD_REG_SET *conflict_regs, | |
1628 | HARD_REG_SET *start_profitable_regs) | |
1756cb66 VM |
1629 | { |
1630 | int i, nwords; | |
1631 | ira_object_t obj; | |
1632 | ||
1633 | nwords = ALLOCNO_NUM_OBJECTS (a); | |
1634 | for (i = 0; i < nwords; i++) | |
1635 | { | |
1636 | obj = ALLOCNO_OBJECT (a, i); | |
6576d245 | 1637 | conflict_regs[i] = OBJECT_TOTAL_CONFLICT_HARD_REGS (obj); |
1756cb66 | 1638 | } |
27508f5f | 1639 | if (retry_p) |
d15e5131 RS |
1640 | *start_profitable_regs |
1641 | = (reg_class_contents[ALLOCNO_CLASS (a)] | |
1642 | &~ (ira_prohibited_class_mode_regs | |
1643 | [ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)])); | |
27508f5f | 1644 | else |
6576d245 | 1645 | *start_profitable_regs = ALLOCNO_COLOR_DATA (a)->profitable_hard_regs; |
1756cb66 VM |
1646 | } |
1647 | ||
27508f5f VM |
1648 | /* Return true if HARD_REGNO is ok for assigning to allocno A with |
1649 | PROFITABLE_REGS and whose objects have CONFLICT_REGS. */ | |
1756cb66 VM |
1650 | static inline bool |
1651 | check_hard_reg_p (ira_allocno_t a, int hard_regno, | |
27508f5f | 1652 | HARD_REG_SET *conflict_regs, HARD_REG_SET profitable_regs) |
1756cb66 VM |
1653 | { |
1654 | int j, nwords, nregs; | |
8d189b3f | 1655 | enum reg_class aclass; |
ef4bddc2 | 1656 | machine_mode mode; |
1756cb66 | 1657 | |
8d189b3f VM |
1658 | aclass = ALLOCNO_CLASS (a); |
1659 | mode = ALLOCNO_MODE (a); | |
1660 | if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[aclass][mode], | |
1661 | hard_regno)) | |
1662 | return false; | |
27508f5f VM |
1663 | /* Checking only profitable hard regs. */ |
1664 | if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno)) | |
1665 | return false; | |
ad474626 | 1666 | nregs = hard_regno_nregs (hard_regno, mode); |
1756cb66 VM |
1667 | nwords = ALLOCNO_NUM_OBJECTS (a); |
1668 | for (j = 0; j < nregs; j++) | |
1669 | { | |
1670 | int k; | |
1671 | int set_to_test_start = 0, set_to_test_end = nwords; | |
1672 | ||
1673 | if (nregs == nwords) | |
1674 | { | |
2805e6c0 | 1675 | if (REG_WORDS_BIG_ENDIAN) |
1756cb66 VM |
1676 | set_to_test_start = nwords - j - 1; |
1677 | else | |
1678 | set_to_test_start = j; | |
1679 | set_to_test_end = set_to_test_start + 1; | |
1680 | } | |
1681 | for (k = set_to_test_start; k < set_to_test_end; k++) | |
27508f5f | 1682 | if (TEST_HARD_REG_BIT (conflict_regs[k], hard_regno + j)) |
1756cb66 VM |
1683 | break; |
1684 | if (k != set_to_test_end) | |
1685 | break; | |
1686 | } | |
1687 | return j == nregs; | |
1688 | } | |
9181a6e5 VM |
1689 | |
1690 | /* Return number of registers needed to be saved and restored at | |
1691 | function prologue/epilogue if we allocate HARD_REGNO to hold value | |
1692 | of MODE. */ | |
1693 | static int | |
ef4bddc2 | 1694 | calculate_saved_nregs (int hard_regno, machine_mode mode) |
9181a6e5 VM |
1695 | { |
1696 | int i; | |
1697 | int nregs = 0; | |
1698 | ||
1699 | ira_assert (hard_regno >= 0); | |
ad474626 | 1700 | for (i = hard_regno_nregs (hard_regno, mode) - 1; i >= 0; i--) |
9181a6e5 | 1701 | if (!allocated_hardreg_p[hard_regno + i] |
6c476222 | 1702 | && !crtl->abi->clobbers_full_reg_p (hard_regno + i) |
9181a6e5 VM |
1703 | && !LOCAL_REGNO (hard_regno + i)) |
1704 | nregs++; | |
1705 | return nregs; | |
1706 | } | |
1756cb66 | 1707 | |
22b0982c VM |
1708 | /* Choose a hard register for allocno A. If RETRY_P is TRUE, it means |
1709 | that the function called from function | |
1756cb66 VM |
1710 | `ira_reassign_conflict_allocnos' and `allocno_reload_assign'. In |
1711 | this case some allocno data are not defined or updated and we | |
1712 | should not touch these data. The function returns true if we | |
1713 | managed to assign a hard register to the allocno. | |
1714 | ||
1715 | To assign a hard register, first of all we calculate all conflict | |
1716 | hard registers which can come from conflicting allocnos with | |
1717 | already assigned hard registers. After that we find first free | |
1718 | hard register with the minimal cost. During hard register cost | |
1719 | calculation we take conflict hard register costs into account to | |
1720 | give a chance for conflicting allocnos to get a better hard | |
1721 | register in the future. | |
1722 | ||
1723 | If the best hard register cost is bigger than cost of memory usage | |
1724 | for the allocno, we don't assign a hard register to given allocno | |
1725 | at all. | |
1726 | ||
1727 | If we assign a hard register to the allocno, we update costs of the | |
1728 | hard register for allocnos connected by copies to improve a chance | |
1729 | to coalesce insns represented by the copies when we assign hard | |
1730 | registers to the allocnos connected by the copies. */ | |
058e97ec | 1731 | static bool |
22b0982c | 1732 | assign_hard_reg (ira_allocno_t a, bool retry_p) |
058e97ec | 1733 | { |
27508f5f | 1734 | HARD_REG_SET conflicting_regs[2], profitable_hard_regs; |
fbddb81d | 1735 | int i, j, hard_regno, best_hard_regno, class_size; |
22b0982c | 1736 | int cost, mem_cost, min_cost, full_cost, min_full_cost, nwords, word; |
058e97ec | 1737 | int *a_costs; |
1756cb66 | 1738 | enum reg_class aclass; |
ef4bddc2 | 1739 | machine_mode mode; |
058e97ec | 1740 | static int costs[FIRST_PSEUDO_REGISTER], full_costs[FIRST_PSEUDO_REGISTER]; |
fbddb81d | 1741 | int saved_nregs; |
a5c011cd MP |
1742 | enum reg_class rclass; |
1743 | int add_cost; | |
058e97ec VM |
1744 | #ifdef STACK_REGS |
1745 | bool no_stack_reg_p; | |
1746 | #endif | |
1747 | ||
22b0982c | 1748 | ira_assert (! ALLOCNO_ASSIGNED_P (a)); |
27508f5f VM |
1749 | get_conflict_and_start_profitable_regs (a, retry_p, |
1750 | conflicting_regs, | |
1751 | &profitable_hard_regs); | |
1756cb66 VM |
1752 | aclass = ALLOCNO_CLASS (a); |
1753 | class_size = ira_class_hard_regs_num[aclass]; | |
058e97ec VM |
1754 | best_hard_regno = -1; |
1755 | memset (full_costs, 0, sizeof (int) * class_size); | |
1756 | mem_cost = 0; | |
058e97ec VM |
1757 | memset (costs, 0, sizeof (int) * class_size); |
1758 | memset (full_costs, 0, sizeof (int) * class_size); | |
1759 | #ifdef STACK_REGS | |
1760 | no_stack_reg_p = false; | |
1761 | #endif | |
1756cb66 VM |
1762 | if (! retry_p) |
1763 | start_update_cost (); | |
22b0982c VM |
1764 | mem_cost += ALLOCNO_UPDATED_MEMORY_COST (a); |
1765 | ||
1766 | ira_allocate_and_copy_costs (&ALLOCNO_UPDATED_HARD_REG_COSTS (a), | |
1756cb66 | 1767 | aclass, ALLOCNO_HARD_REG_COSTS (a)); |
22b0982c | 1768 | a_costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a); |
058e97ec | 1769 | #ifdef STACK_REGS |
22b0982c | 1770 | no_stack_reg_p = no_stack_reg_p || ALLOCNO_TOTAL_NO_STACK_REG_P (a); |
058e97ec | 1771 | #endif |
1756cb66 | 1772 | cost = ALLOCNO_UPDATED_CLASS_COST (a); |
22b0982c VM |
1773 | for (i = 0; i < class_size; i++) |
1774 | if (a_costs != NULL) | |
1775 | { | |
1776 | costs[i] += a_costs[i]; | |
1777 | full_costs[i] += a_costs[i]; | |
1778 | } | |
1779 | else | |
1780 | { | |
1781 | costs[i] += cost; | |
1782 | full_costs[i] += cost; | |
1783 | } | |
1756cb66 | 1784 | nwords = ALLOCNO_NUM_OBJECTS (a); |
27508f5f | 1785 | curr_allocno_process++; |
22b0982c VM |
1786 | for (word = 0; word < nwords; word++) |
1787 | { | |
1788 | ira_object_t conflict_obj; | |
1789 | ira_object_t obj = ALLOCNO_OBJECT (a, word); | |
1790 | ira_object_conflict_iterator oci; | |
1791 | ||
22b0982c VM |
1792 | /* Take preferences of conflicting allocnos into account. */ |
1793 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
1756cb66 | 1794 | { |
22b0982c | 1795 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); |
1756cb66 | 1796 | enum reg_class conflict_aclass; |
4ef20c29 | 1797 | allocno_color_data_t data = ALLOCNO_COLOR_DATA (conflict_a); |
1756cb66 | 1798 | |
22b0982c VM |
1799 | /* Reload can give another class so we need to check all |
1800 | allocnos. */ | |
1756cb66 | 1801 | if (!retry_p |
06fbce66 ZZ |
1802 | && ((!ALLOCNO_ASSIGNED_P (conflict_a) |
1803 | || ALLOCNO_HARD_REGNO (conflict_a) < 0) | |
1804 | && !(hard_reg_set_intersect_p | |
1805 | (profitable_hard_regs, | |
1806 | ALLOCNO_COLOR_DATA | |
1807 | (conflict_a)->profitable_hard_regs)))) | |
1808 | { | |
1809 | /* All conflict allocnos are in consideration bitmap | |
1810 | when retry_p is false. It might change in future and | |
1811 | if it happens the assert will be broken. It means | |
1812 | the code should be modified for the new | |
1813 | assumptions. */ | |
1814 | ira_assert (bitmap_bit_p (consideration_allocno_bitmap, | |
1815 | ALLOCNO_NUM (conflict_a))); | |
1816 | continue; | |
1817 | } | |
1756cb66 | 1818 | conflict_aclass = ALLOCNO_CLASS (conflict_a); |
22b0982c | 1819 | ira_assert (ira_reg_classes_intersect_p |
1756cb66 | 1820 | [aclass][conflict_aclass]); |
22b0982c | 1821 | if (ALLOCNO_ASSIGNED_P (conflict_a)) |
fa86d337 | 1822 | { |
22b0982c VM |
1823 | hard_regno = ALLOCNO_HARD_REGNO (conflict_a); |
1824 | if (hard_regno >= 0 | |
b8faca75 VM |
1825 | && (ira_hard_reg_set_intersection_p |
1826 | (hard_regno, ALLOCNO_MODE (conflict_a), | |
1827 | reg_class_contents[aclass]))) | |
fa86d337 | 1828 | { |
22b0982c | 1829 | int n_objects = ALLOCNO_NUM_OBJECTS (conflict_a); |
4648deb4 | 1830 | int conflict_nregs; |
1756cb66 | 1831 | |
4648deb4 | 1832 | mode = ALLOCNO_MODE (conflict_a); |
ad474626 | 1833 | conflict_nregs = hard_regno_nregs (hard_regno, mode); |
22b0982c | 1834 | if (conflict_nregs == n_objects && conflict_nregs > 1) |
fa86d337 | 1835 | { |
22b0982c | 1836 | int num = OBJECT_SUBWORD (conflict_obj); |
ac0ab4f7 | 1837 | |
2805e6c0 | 1838 | if (REG_WORDS_BIG_ENDIAN) |
22b0982c VM |
1839 | SET_HARD_REG_BIT (conflicting_regs[word], |
1840 | hard_regno + n_objects - num - 1); | |
1841 | else | |
1842 | SET_HARD_REG_BIT (conflicting_regs[word], | |
1843 | hard_regno + num); | |
ac0ab4f7 | 1844 | } |
22b0982c | 1845 | else |
44942965 RS |
1846 | conflicting_regs[word] |
1847 | |= ira_reg_mode_hard_regset[hard_regno][mode]; | |
27508f5f | 1848 | if (hard_reg_set_subset_p (profitable_hard_regs, |
22b0982c VM |
1849 | conflicting_regs[word])) |
1850 | goto fail; | |
fa86d337 BS |
1851 | } |
1852 | } | |
1756cb66 | 1853 | else if (! retry_p |
27508f5f VM |
1854 | && ! ALLOCNO_COLOR_DATA (conflict_a)->may_be_spilled_p |
1855 | /* Don't process the conflict allocno twice. */ | |
1856 | && (ALLOCNO_COLOR_DATA (conflict_a)->last_process | |
1857 | != curr_allocno_process)) | |
22b0982c VM |
1858 | { |
1859 | int k, *conflict_costs; | |
1860 | ||
27508f5f VM |
1861 | ALLOCNO_COLOR_DATA (conflict_a)->last_process |
1862 | = curr_allocno_process; | |
22b0982c VM |
1863 | ira_allocate_and_copy_costs |
1864 | (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a), | |
1756cb66 | 1865 | conflict_aclass, |
22b0982c VM |
1866 | ALLOCNO_CONFLICT_HARD_REG_COSTS (conflict_a)); |
1867 | conflict_costs | |
1868 | = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a); | |
1869 | if (conflict_costs != NULL) | |
1870 | for (j = class_size - 1; j >= 0; j--) | |
1871 | { | |
1756cb66 | 1872 | hard_regno = ira_class_hard_regs[aclass][j]; |
22b0982c | 1873 | ira_assert (hard_regno >= 0); |
1756cb66 | 1874 | k = ira_class_hard_reg_index[conflict_aclass][hard_regno]; |
4ef20c29 ZC |
1875 | if (k < 0 |
1876 | /* If HARD_REGNO is not available for CONFLICT_A, | |
1877 | the conflict would be ignored, since HARD_REGNO | |
1878 | will never be assigned to CONFLICT_A. */ | |
1879 | || !TEST_HARD_REG_BIT (data->profitable_hard_regs, | |
1880 | hard_regno)) | |
22b0982c VM |
1881 | continue; |
1882 | full_costs[j] -= conflict_costs[k]; | |
1883 | } | |
3133bed5 | 1884 | queue_update_cost (conflict_a, conflict_a, NULL, COST_HOP_DIVISOR); |
22b0982c | 1885 | } |
fa86d337 | 1886 | } |
058e97ec | 1887 | } |
1756cb66 VM |
1888 | if (! retry_p) |
1889 | /* Take into account preferences of allocnos connected by copies to | |
1890 | the conflict allocnos. */ | |
1891 | update_conflict_hard_regno_costs (full_costs, aclass, true); | |
f754734f | 1892 | |
a7f32992 VM |
1893 | /* Take preferences of allocnos connected by copies into |
1894 | account. */ | |
1756cb66 VM |
1895 | if (! retry_p) |
1896 | { | |
1897 | start_update_cost (); | |
3133bed5 | 1898 | queue_update_cost (a, a, NULL, COST_HOP_DIVISOR); |
1756cb66 VM |
1899 | update_conflict_hard_regno_costs (full_costs, aclass, false); |
1900 | } | |
058e97ec VM |
1901 | min_cost = min_full_cost = INT_MAX; |
1902 | /* We don't care about giving callee saved registers to allocnos no | |
1903 | living through calls because call clobbered registers are | |
1904 | allocated first (it is usual practice to put them first in | |
1905 | REG_ALLOC_ORDER). */ | |
1756cb66 | 1906 | mode = ALLOCNO_MODE (a); |
058e97ec VM |
1907 | for (i = 0; i < class_size; i++) |
1908 | { | |
1756cb66 | 1909 | hard_regno = ira_class_hard_regs[aclass][i]; |
058e97ec VM |
1910 | #ifdef STACK_REGS |
1911 | if (no_stack_reg_p | |
1912 | && FIRST_STACK_REG <= hard_regno && hard_regno <= LAST_STACK_REG) | |
1913 | continue; | |
1914 | #endif | |
1756cb66 VM |
1915 | if (! check_hard_reg_p (a, hard_regno, |
1916 | conflicting_regs, profitable_hard_regs)) | |
058e97ec VM |
1917 | continue; |
1918 | cost = costs[i]; | |
1919 | full_cost = full_costs[i]; | |
ed15c598 | 1920 | if (!HONOR_REG_ALLOC_ORDER) |
058e97ec | 1921 | { |
ed15c598 KC |
1922 | if ((saved_nregs = calculate_saved_nregs (hard_regno, mode)) != 0) |
1923 | /* We need to save/restore the hard register in | |
1924 | epilogue/prologue. Therefore we increase the cost. */ | |
1925 | { | |
1926 | rclass = REGNO_REG_CLASS (hard_regno); | |
1927 | add_cost = ((ira_memory_move_cost[mode][rclass][0] | |
1928 | + ira_memory_move_cost[mode][rclass][1]) | |
ad474626 RS |
1929 | * saved_nregs / hard_regno_nregs (hard_regno, |
1930 | mode) - 1); | |
ed15c598 KC |
1931 | cost += add_cost; |
1932 | full_cost += add_cost; | |
1933 | } | |
058e97ec VM |
1934 | } |
1935 | if (min_cost > cost) | |
1936 | min_cost = cost; | |
5dc1390b | 1937 | if (min_full_cost > full_cost) |
058e97ec VM |
1938 | { |
1939 | min_full_cost = full_cost; | |
1940 | best_hard_regno = hard_regno; | |
1941 | ira_assert (hard_regno >= 0); | |
1942 | } | |
3133bed5 VM |
1943 | if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL) |
1944 | fprintf (ira_dump_file, "(%d=%d,%d) ", hard_regno, cost, full_cost); | |
058e97ec | 1945 | } |
74dc179a VM |
1946 | if (internal_flag_ira_verbose > 5 && ira_dump_file != NULL) |
1947 | fprintf (ira_dump_file, "\n"); | |
b81a2f0d VM |
1948 | if (min_full_cost > mem_cost |
1949 | /* Do not spill static chain pointer pseudo when non-local goto | |
1950 | is used. */ | |
1951 | && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a))) | |
058e97ec VM |
1952 | { |
1953 | if (! retry_p && internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
1954 | fprintf (ira_dump_file, "(memory is more profitable %d vs %d) ", | |
1955 | mem_cost, min_full_cost); | |
1956 | best_hard_regno = -1; | |
1957 | } | |
1958 | fail: | |
058e97ec | 1959 | if (best_hard_regno >= 0) |
9181a6e5 | 1960 | { |
ad474626 | 1961 | for (i = hard_regno_nregs (best_hard_regno, mode) - 1; i >= 0; i--) |
34672f15 | 1962 | allocated_hardreg_p[best_hard_regno + i] = true; |
9181a6e5 | 1963 | } |
c73ccc80 VM |
1964 | if (! retry_p) |
1965 | restore_costs_from_copies (a); | |
22b0982c VM |
1966 | ALLOCNO_HARD_REGNO (a) = best_hard_regno; |
1967 | ALLOCNO_ASSIGNED_P (a) = true; | |
1968 | if (best_hard_regno >= 0) | |
c73ccc80 | 1969 | update_costs_from_copies (a, true, ! retry_p); |
1756cb66 | 1970 | ira_assert (ALLOCNO_CLASS (a) == aclass); |
2b9c63a2 | 1971 | /* We don't need updated costs anymore. */ |
22b0982c | 1972 | ira_free_allocno_updated_costs (a); |
058e97ec VM |
1973 | return best_hard_regno >= 0; |
1974 | } | |
1975 | ||
1976 | \f | |
1977 | ||
bf08fb16 VM |
1978 | /* An array used to sort copies. */ |
1979 | static ira_copy_t *sorted_copies; | |
1980 | ||
0550a77b VM |
1981 | /* If allocno A is a cap, return non-cap allocno from which A is |
1982 | created. Otherwise, return A. */ | |
1983 | static ira_allocno_t | |
1984 | get_cap_member (ira_allocno_t a) | |
1985 | { | |
1986 | ira_allocno_t member; | |
1987 | ||
1988 | while ((member = ALLOCNO_CAP_MEMBER (a)) != NULL) | |
1989 | a = member; | |
1990 | return a; | |
1991 | } | |
1992 | ||
bf08fb16 VM |
1993 | /* Return TRUE if live ranges of allocnos A1 and A2 intersect. It is |
1994 | used to find a conflict for new allocnos or allocnos with the | |
1995 | different allocno classes. */ | |
1996 | static bool | |
1997 | allocnos_conflict_by_live_ranges_p (ira_allocno_t a1, ira_allocno_t a2) | |
1998 | { | |
1999 | rtx reg1, reg2; | |
2000 | int i, j; | |
2001 | int n1 = ALLOCNO_NUM_OBJECTS (a1); | |
2002 | int n2 = ALLOCNO_NUM_OBJECTS (a2); | |
2003 | ||
2004 | if (a1 == a2) | |
2005 | return false; | |
2006 | reg1 = regno_reg_rtx[ALLOCNO_REGNO (a1)]; | |
2007 | reg2 = regno_reg_rtx[ALLOCNO_REGNO (a2)]; | |
2008 | if (reg1 != NULL && reg2 != NULL | |
2009 | && ORIGINAL_REGNO (reg1) == ORIGINAL_REGNO (reg2)) | |
2010 | return false; | |
2011 | ||
0550a77b VM |
2012 | /* We don't keep live ranges for caps because they can be quite big. |
2013 | Use ranges of non-cap allocno from which caps are created. */ | |
2014 | a1 = get_cap_member (a1); | |
2015 | a2 = get_cap_member (a2); | |
bf08fb16 VM |
2016 | for (i = 0; i < n1; i++) |
2017 | { | |
2018 | ira_object_t c1 = ALLOCNO_OBJECT (a1, i); | |
2019 | ||
2020 | for (j = 0; j < n2; j++) | |
2021 | { | |
2022 | ira_object_t c2 = ALLOCNO_OBJECT (a2, j); | |
2023 | ||
2024 | if (ira_live_ranges_intersect_p (OBJECT_LIVE_RANGES (c1), | |
2025 | OBJECT_LIVE_RANGES (c2))) | |
2026 | return true; | |
2027 | } | |
2028 | } | |
2029 | return false; | |
2030 | } | |
2031 | ||
2032 | /* The function is used to sort copies according to their execution | |
2033 | frequencies. */ | |
2034 | static int | |
2035 | copy_freq_compare_func (const void *v1p, const void *v2p) | |
2036 | { | |
2037 | ira_copy_t cp1 = *(const ira_copy_t *) v1p, cp2 = *(const ira_copy_t *) v2p; | |
2038 | int pri1, pri2; | |
2039 | ||
2040 | pri1 = cp1->freq; | |
2041 | pri2 = cp2->freq; | |
2042 | if (pri2 - pri1) | |
2043 | return pri2 - pri1; | |
2044 | ||
df3e3493 | 2045 | /* If frequencies are equal, sort by copies, so that the results of |
bf08fb16 VM |
2046 | qsort leave nothing to chance. */ |
2047 | return cp1->num - cp2->num; | |
2048 | } | |
2049 | ||
2050 | \f | |
2051 | ||
2052 | /* Return true if any allocno from thread of A1 conflicts with any | |
2053 | allocno from thread A2. */ | |
2054 | static bool | |
2055 | allocno_thread_conflict_p (ira_allocno_t a1, ira_allocno_t a2) | |
2056 | { | |
2057 | ira_allocno_t a, conflict_a; | |
2058 | ||
2059 | for (a = ALLOCNO_COLOR_DATA (a2)->next_thread_allocno;; | |
2060 | a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno) | |
2061 | { | |
2062 | for (conflict_a = ALLOCNO_COLOR_DATA (a1)->next_thread_allocno;; | |
2063 | conflict_a = ALLOCNO_COLOR_DATA (conflict_a)->next_thread_allocno) | |
2064 | { | |
2065 | if (allocnos_conflict_by_live_ranges_p (a, conflict_a)) | |
2066 | return true; | |
2067 | if (conflict_a == a1) | |
2068 | break; | |
2069 | } | |
2070 | if (a == a2) | |
2071 | break; | |
2072 | } | |
2073 | return false; | |
2074 | } | |
2075 | ||
2076 | /* Merge two threads given correspondingly by their first allocnos T1 | |
2077 | and T2 (more accurately merging T2 into T1). */ | |
2078 | static void | |
2079 | merge_threads (ira_allocno_t t1, ira_allocno_t t2) | |
2080 | { | |
2081 | ira_allocno_t a, next, last; | |
2082 | ||
2083 | gcc_assert (t1 != t2 | |
2084 | && ALLOCNO_COLOR_DATA (t1)->first_thread_allocno == t1 | |
2085 | && ALLOCNO_COLOR_DATA (t2)->first_thread_allocno == t2); | |
2086 | for (last = t2, a = ALLOCNO_COLOR_DATA (t2)->next_thread_allocno;; | |
2087 | a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno) | |
2088 | { | |
2089 | ALLOCNO_COLOR_DATA (a)->first_thread_allocno = t1; | |
2090 | if (a == t2) | |
2091 | break; | |
2092 | last = a; | |
2093 | } | |
2094 | next = ALLOCNO_COLOR_DATA (t1)->next_thread_allocno; | |
2095 | ALLOCNO_COLOR_DATA (t1)->next_thread_allocno = t2; | |
2096 | ALLOCNO_COLOR_DATA (last)->next_thread_allocno = next; | |
2097 | ALLOCNO_COLOR_DATA (t1)->thread_freq += ALLOCNO_COLOR_DATA (t2)->thread_freq; | |
2098 | } | |
2099 | ||
df3e3493 | 2100 | /* Create threads by processing CP_NUM copies from sorted copies. We |
bf08fb16 VM |
2101 | process the most expensive copies first. */ |
2102 | static void | |
2103 | form_threads_from_copies (int cp_num) | |
2104 | { | |
2105 | ira_allocno_t a, thread1, thread2; | |
2106 | ira_copy_t cp; | |
2107 | int i, n; | |
2108 | ||
2109 | qsort (sorted_copies, cp_num, sizeof (ira_copy_t), copy_freq_compare_func); | |
2110 | /* Form threads processing copies, most frequently executed | |
2111 | first. */ | |
2112 | for (; cp_num != 0;) | |
2113 | { | |
2114 | for (i = 0; i < cp_num; i++) | |
2115 | { | |
2116 | cp = sorted_copies[i]; | |
2117 | thread1 = ALLOCNO_COLOR_DATA (cp->first)->first_thread_allocno; | |
2118 | thread2 = ALLOCNO_COLOR_DATA (cp->second)->first_thread_allocno; | |
2119 | if (thread1 == thread2) | |
2120 | continue; | |
2121 | if (! allocno_thread_conflict_p (thread1, thread2)) | |
2122 | { | |
2123 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
2124 | fprintf | |
2125 | (ira_dump_file, | |
74dc179a | 2126 | " Forming thread by copy %d:a%dr%d-a%dr%d (freq=%d):\n", |
bf08fb16 VM |
2127 | cp->num, ALLOCNO_NUM (cp->first), ALLOCNO_REGNO (cp->first), |
2128 | ALLOCNO_NUM (cp->second), ALLOCNO_REGNO (cp->second), | |
2129 | cp->freq); | |
2130 | merge_threads (thread1, thread2); | |
2131 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
2132 | { | |
2133 | thread1 = ALLOCNO_COLOR_DATA (thread1)->first_thread_allocno; | |
74dc179a | 2134 | fprintf (ira_dump_file, " Result (freq=%d): a%dr%d(%d)", |
bf08fb16 VM |
2135 | ALLOCNO_COLOR_DATA (thread1)->thread_freq, |
2136 | ALLOCNO_NUM (thread1), ALLOCNO_REGNO (thread1), | |
2137 | ALLOCNO_FREQ (thread1)); | |
2138 | for (a = ALLOCNO_COLOR_DATA (thread1)->next_thread_allocno; | |
2139 | a != thread1; | |
2140 | a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno) | |
2141 | fprintf (ira_dump_file, " a%dr%d(%d)", | |
2142 | ALLOCNO_NUM (a), ALLOCNO_REGNO (a), | |
2143 | ALLOCNO_FREQ (a)); | |
2144 | fprintf (ira_dump_file, "\n"); | |
2145 | } | |
2146 | i++; | |
2147 | break; | |
2148 | } | |
2149 | } | |
2150 | /* Collect the rest of copies. */ | |
2151 | for (n = 0; i < cp_num; i++) | |
2152 | { | |
2153 | cp = sorted_copies[i]; | |
2154 | if (ALLOCNO_COLOR_DATA (cp->first)->first_thread_allocno | |
2155 | != ALLOCNO_COLOR_DATA (cp->second)->first_thread_allocno) | |
2156 | sorted_copies[n++] = cp; | |
2157 | } | |
2158 | cp_num = n; | |
2159 | } | |
2160 | } | |
2161 | ||
2162 | /* Create threads by processing copies of all alocnos from BUCKET. We | |
2163 | process the most expensive copies first. */ | |
2164 | static void | |
2165 | form_threads_from_bucket (ira_allocno_t bucket) | |
2166 | { | |
2167 | ira_allocno_t a; | |
2168 | ira_copy_t cp, next_cp; | |
2169 | int cp_num = 0; | |
2170 | ||
2171 | for (a = bucket; a != NULL; a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno) | |
2172 | { | |
2173 | for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp) | |
2174 | { | |
2175 | if (cp->first == a) | |
2176 | { | |
2177 | next_cp = cp->next_first_allocno_copy; | |
2178 | sorted_copies[cp_num++] = cp; | |
2179 | } | |
2180 | else if (cp->second == a) | |
2181 | next_cp = cp->next_second_allocno_copy; | |
2182 | else | |
2183 | gcc_unreachable (); | |
2184 | } | |
2185 | } | |
2186 | form_threads_from_copies (cp_num); | |
2187 | } | |
2188 | ||
2189 | /* Create threads by processing copies of colorable allocno A. We | |
2190 | process most expensive copies first. */ | |
2191 | static void | |
2192 | form_threads_from_colorable_allocno (ira_allocno_t a) | |
2193 | { | |
2194 | ira_allocno_t another_a; | |
2195 | ira_copy_t cp, next_cp; | |
2196 | int cp_num = 0; | |
2197 | ||
74dc179a VM |
2198 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
2199 | fprintf (ira_dump_file, " Forming thread from allocno a%dr%d:\n", | |
2200 | ALLOCNO_NUM (a), ALLOCNO_REGNO (a)); | |
bf08fb16 VM |
2201 | for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp) |
2202 | { | |
2203 | if (cp->first == a) | |
2204 | { | |
2205 | next_cp = cp->next_first_allocno_copy; | |
2206 | another_a = cp->second; | |
2207 | } | |
2208 | else if (cp->second == a) | |
2209 | { | |
2210 | next_cp = cp->next_second_allocno_copy; | |
2211 | another_a = cp->first; | |
2212 | } | |
2213 | else | |
2214 | gcc_unreachable (); | |
2215 | if ((! ALLOCNO_COLOR_DATA (another_a)->in_graph_p | |
2216 | && !ALLOCNO_COLOR_DATA (another_a)->may_be_spilled_p) | |
2217 | || ALLOCNO_COLOR_DATA (another_a)->colorable_p) | |
2218 | sorted_copies[cp_num++] = cp; | |
2219 | } | |
2220 | form_threads_from_copies (cp_num); | |
2221 | } | |
2222 | ||
2223 | /* Form initial threads which contain only one allocno. */ | |
2224 | static void | |
2225 | init_allocno_threads (void) | |
2226 | { | |
2227 | ira_allocno_t a; | |
2228 | unsigned int j; | |
2229 | bitmap_iterator bi; | |
897a7308 | 2230 | ira_pref_t pref; |
bf08fb16 VM |
2231 | |
2232 | EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi) | |
2233 | { | |
2234 | a = ira_allocnos[j]; | |
2235 | /* Set up initial thread data: */ | |
2236 | ALLOCNO_COLOR_DATA (a)->first_thread_allocno | |
2237 | = ALLOCNO_COLOR_DATA (a)->next_thread_allocno = a; | |
2238 | ALLOCNO_COLOR_DATA (a)->thread_freq = ALLOCNO_FREQ (a); | |
897a7308 VM |
2239 | ALLOCNO_COLOR_DATA (a)->hard_reg_prefs = 0; |
2240 | for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = pref->next_pref) | |
2241 | ALLOCNO_COLOR_DATA (a)->hard_reg_prefs += pref->freq; | |
bf08fb16 VM |
2242 | } |
2243 | } | |
2244 | ||
2245 | \f | |
2246 | ||
058e97ec VM |
2247 | /* This page contains the allocator based on the Chaitin-Briggs algorithm. */ |
2248 | ||
2249 | /* Bucket of allocnos that can colored currently without spilling. */ | |
2250 | static ira_allocno_t colorable_allocno_bucket; | |
2251 | ||
2252 | /* Bucket of allocnos that might be not colored currently without | |
2253 | spilling. */ | |
2254 | static ira_allocno_t uncolorable_allocno_bucket; | |
2255 | ||
1756cb66 VM |
2256 | /* The current number of allocnos in the uncolorable_bucket. */ |
2257 | static int uncolorable_allocnos_num; | |
058e97ec | 2258 | |
30ea859e VM |
2259 | /* Return the current spill priority of allocno A. The less the |
2260 | number, the more preferable the allocno for spilling. */ | |
1756cb66 | 2261 | static inline int |
30ea859e VM |
2262 | allocno_spill_priority (ira_allocno_t a) |
2263 | { | |
1756cb66 VM |
2264 | allocno_color_data_t data = ALLOCNO_COLOR_DATA (a); |
2265 | ||
2266 | return (data->temp | |
2267 | / (ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a) | |
2268 | * ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)] | |
30ea859e VM |
2269 | + 1)); |
2270 | } | |
2271 | ||
1756cb66 | 2272 | /* Add allocno A to bucket *BUCKET_PTR. A should be not in a bucket |
058e97ec VM |
2273 | before the call. */ |
2274 | static void | |
1756cb66 | 2275 | add_allocno_to_bucket (ira_allocno_t a, ira_allocno_t *bucket_ptr) |
058e97ec | 2276 | { |
1756cb66 VM |
2277 | ira_allocno_t first_a; |
2278 | allocno_color_data_t data; | |
058e97ec VM |
2279 | |
2280 | if (bucket_ptr == &uncolorable_allocno_bucket | |
1756cb66 | 2281 | && ALLOCNO_CLASS (a) != NO_REGS) |
058e97ec | 2282 | { |
1756cb66 VM |
2283 | uncolorable_allocnos_num++; |
2284 | ira_assert (uncolorable_allocnos_num > 0); | |
058e97ec | 2285 | } |
1756cb66 VM |
2286 | first_a = *bucket_ptr; |
2287 | data = ALLOCNO_COLOR_DATA (a); | |
2288 | data->next_bucket_allocno = first_a; | |
2289 | data->prev_bucket_allocno = NULL; | |
2290 | if (first_a != NULL) | |
2291 | ALLOCNO_COLOR_DATA (first_a)->prev_bucket_allocno = a; | |
2292 | *bucket_ptr = a; | |
058e97ec VM |
2293 | } |
2294 | ||
058e97ec VM |
2295 | /* Compare two allocnos to define which allocno should be pushed first |
2296 | into the coloring stack. If the return is a negative number, the | |
2297 | allocno given by the first parameter will be pushed first. In this | |
2298 | case such allocno has less priority than the second one and the | |
2299 | hard register will be assigned to it after assignment to the second | |
2300 | one. As the result of such assignment order, the second allocno | |
2301 | has a better chance to get the best hard register. */ | |
2302 | static int | |
2303 | bucket_allocno_compare_func (const void *v1p, const void *v2p) | |
2304 | { | |
2305 | ira_allocno_t a1 = *(const ira_allocno_t *) v1p; | |
2306 | ira_allocno_t a2 = *(const ira_allocno_t *) v2p; | |
8c679205 | 2307 | int diff, freq1, freq2, a1_num, a2_num, pref1, pref2; |
bf08fb16 VM |
2308 | ira_allocno_t t1 = ALLOCNO_COLOR_DATA (a1)->first_thread_allocno; |
2309 | ira_allocno_t t2 = ALLOCNO_COLOR_DATA (a2)->first_thread_allocno; | |
9c3b0346 VM |
2310 | int cl1 = ALLOCNO_CLASS (a1), cl2 = ALLOCNO_CLASS (a2); |
2311 | ||
bf08fb16 VM |
2312 | freq1 = ALLOCNO_COLOR_DATA (t1)->thread_freq; |
2313 | freq2 = ALLOCNO_COLOR_DATA (t2)->thread_freq; | |
2314 | if ((diff = freq1 - freq2) != 0) | |
2315 | return diff; | |
2316 | ||
2317 | if ((diff = ALLOCNO_NUM (t2) - ALLOCNO_NUM (t1)) != 0) | |
2318 | return diff; | |
2319 | ||
9c3b0346 VM |
2320 | /* Push pseudos requiring less hard registers first. It means that |
2321 | we will assign pseudos requiring more hard registers first | |
2322 | avoiding creation small holes in free hard register file into | |
67914693 | 2323 | which the pseudos requiring more hard registers cannot fit. */ |
9c3b0346 VM |
2324 | if ((diff = (ira_reg_class_max_nregs[cl1][ALLOCNO_MODE (a1)] |
2325 | - ira_reg_class_max_nregs[cl2][ALLOCNO_MODE (a2)])) != 0) | |
058e97ec | 2326 | return diff; |
bf08fb16 VM |
2327 | |
2328 | freq1 = ALLOCNO_FREQ (a1); | |
2329 | freq2 = ALLOCNO_FREQ (a2); | |
2330 | if ((diff = freq1 - freq2) != 0) | |
058e97ec | 2331 | return diff; |
bf08fb16 | 2332 | |
1756cb66 VM |
2333 | a1_num = ALLOCNO_COLOR_DATA (a1)->available_regs_num; |
2334 | a2_num = ALLOCNO_COLOR_DATA (a2)->available_regs_num; | |
2335 | if ((diff = a2_num - a1_num) != 0) | |
99710245 | 2336 | return diff; |
3133bed5 VM |
2337 | /* Push allocnos with minimal conflict_allocno_hard_prefs first. */ |
2338 | pref1 = ALLOCNO_COLOR_DATA (a1)->conflict_allocno_hard_prefs; | |
2339 | pref2 = ALLOCNO_COLOR_DATA (a2)->conflict_allocno_hard_prefs; | |
2340 | if ((diff = pref1 - pref2) != 0) | |
2341 | return diff; | |
058e97ec VM |
2342 | return ALLOCNO_NUM (a2) - ALLOCNO_NUM (a1); |
2343 | } | |
2344 | ||
2345 | /* Sort bucket *BUCKET_PTR and return the result through | |
2346 | BUCKET_PTR. */ | |
2347 | static void | |
1756cb66 VM |
2348 | sort_bucket (ira_allocno_t *bucket_ptr, |
2349 | int (*compare_func) (const void *, const void *)) | |
058e97ec VM |
2350 | { |
2351 | ira_allocno_t a, head; | |
2352 | int n; | |
2353 | ||
1756cb66 VM |
2354 | for (n = 0, a = *bucket_ptr; |
2355 | a != NULL; | |
2356 | a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno) | |
058e97ec VM |
2357 | sorted_allocnos[n++] = a; |
2358 | if (n <= 1) | |
2359 | return; | |
1756cb66 | 2360 | qsort (sorted_allocnos, n, sizeof (ira_allocno_t), compare_func); |
058e97ec VM |
2361 | head = NULL; |
2362 | for (n--; n >= 0; n--) | |
2363 | { | |
2364 | a = sorted_allocnos[n]; | |
1756cb66 VM |
2365 | ALLOCNO_COLOR_DATA (a)->next_bucket_allocno = head; |
2366 | ALLOCNO_COLOR_DATA (a)->prev_bucket_allocno = NULL; | |
058e97ec | 2367 | if (head != NULL) |
1756cb66 | 2368 | ALLOCNO_COLOR_DATA (head)->prev_bucket_allocno = a; |
058e97ec VM |
2369 | head = a; |
2370 | } | |
2371 | *bucket_ptr = head; | |
2372 | } | |
2373 | ||
bf08fb16 | 2374 | /* Add ALLOCNO to colorable bucket maintaining the order according |
058e97ec VM |
2375 | their priority. ALLOCNO should be not in a bucket before the |
2376 | call. */ | |
2377 | static void | |
bf08fb16 | 2378 | add_allocno_to_ordered_colorable_bucket (ira_allocno_t allocno) |
058e97ec VM |
2379 | { |
2380 | ira_allocno_t before, after; | |
058e97ec | 2381 | |
bf08fb16 VM |
2382 | form_threads_from_colorable_allocno (allocno); |
2383 | for (before = colorable_allocno_bucket, after = NULL; | |
058e97ec | 2384 | before != NULL; |
1756cb66 VM |
2385 | after = before, |
2386 | before = ALLOCNO_COLOR_DATA (before)->next_bucket_allocno) | |
058e97ec VM |
2387 | if (bucket_allocno_compare_func (&allocno, &before) < 0) |
2388 | break; | |
1756cb66 VM |
2389 | ALLOCNO_COLOR_DATA (allocno)->next_bucket_allocno = before; |
2390 | ALLOCNO_COLOR_DATA (allocno)->prev_bucket_allocno = after; | |
058e97ec | 2391 | if (after == NULL) |
bf08fb16 | 2392 | colorable_allocno_bucket = allocno; |
058e97ec | 2393 | else |
1756cb66 | 2394 | ALLOCNO_COLOR_DATA (after)->next_bucket_allocno = allocno; |
058e97ec | 2395 | if (before != NULL) |
1756cb66 | 2396 | ALLOCNO_COLOR_DATA (before)->prev_bucket_allocno = allocno; |
058e97ec VM |
2397 | } |
2398 | ||
2399 | /* Delete ALLOCNO from bucket *BUCKET_PTR. It should be there before | |
2400 | the call. */ | |
2401 | static void | |
2402 | delete_allocno_from_bucket (ira_allocno_t allocno, ira_allocno_t *bucket_ptr) | |
2403 | { | |
2404 | ira_allocno_t prev_allocno, next_allocno; | |
058e97ec VM |
2405 | |
2406 | if (bucket_ptr == &uncolorable_allocno_bucket | |
1756cb66 | 2407 | && ALLOCNO_CLASS (allocno) != NO_REGS) |
058e97ec | 2408 | { |
1756cb66 VM |
2409 | uncolorable_allocnos_num--; |
2410 | ira_assert (uncolorable_allocnos_num >= 0); | |
058e97ec | 2411 | } |
1756cb66 VM |
2412 | prev_allocno = ALLOCNO_COLOR_DATA (allocno)->prev_bucket_allocno; |
2413 | next_allocno = ALLOCNO_COLOR_DATA (allocno)->next_bucket_allocno; | |
058e97ec | 2414 | if (prev_allocno != NULL) |
1756cb66 | 2415 | ALLOCNO_COLOR_DATA (prev_allocno)->next_bucket_allocno = next_allocno; |
058e97ec VM |
2416 | else |
2417 | { | |
2418 | ira_assert (*bucket_ptr == allocno); | |
2419 | *bucket_ptr = next_allocno; | |
2420 | } | |
2421 | if (next_allocno != NULL) | |
1756cb66 | 2422 | ALLOCNO_COLOR_DATA (next_allocno)->prev_bucket_allocno = prev_allocno; |
058e97ec VM |
2423 | } |
2424 | ||
22b0982c | 2425 | /* Put allocno A onto the coloring stack without removing it from its |
058e97ec VM |
2426 | bucket. Pushing allocno to the coloring stack can result in moving |
2427 | conflicting allocnos from the uncolorable bucket to the colorable | |
8c679205 VM |
2428 | one. Update conflict_allocno_hard_prefs of the conflicting |
2429 | allocnos which are not on stack yet. */ | |
058e97ec | 2430 | static void |
22b0982c | 2431 | push_allocno_to_stack (ira_allocno_t a) |
058e97ec | 2432 | { |
1756cb66 VM |
2433 | enum reg_class aclass; |
2434 | allocno_color_data_t data, conflict_data; | |
2435 | int size, i, n = ALLOCNO_NUM_OBJECTS (a); | |
2436 | ||
2437 | data = ALLOCNO_COLOR_DATA (a); | |
2438 | data->in_graph_p = false; | |
9771b263 | 2439 | allocno_stack_vec.safe_push (a); |
1756cb66 VM |
2440 | aclass = ALLOCNO_CLASS (a); |
2441 | if (aclass == NO_REGS) | |
058e97ec | 2442 | return; |
1756cb66 VM |
2443 | size = ira_reg_class_max_nregs[aclass][ALLOCNO_MODE (a)]; |
2444 | if (n > 1) | |
ac0ab4f7 BS |
2445 | { |
2446 | /* We will deal with the subwords individually. */ | |
22b0982c | 2447 | gcc_assert (size == ALLOCNO_NUM_OBJECTS (a)); |
ac0ab4f7 BS |
2448 | size = 1; |
2449 | } | |
22b0982c | 2450 | for (i = 0; i < n; i++) |
058e97ec | 2451 | { |
22b0982c | 2452 | ira_object_t obj = ALLOCNO_OBJECT (a, i); |
22b0982c VM |
2453 | ira_object_t conflict_obj; |
2454 | ira_object_conflict_iterator oci; | |
2455 | ||
2456 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
548a6322 | 2457 | { |
22b0982c | 2458 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); |
8c679205 VM |
2459 | ira_pref_t pref; |
2460 | ||
1756cb66 | 2461 | conflict_data = ALLOCNO_COLOR_DATA (conflict_a); |
8c679205 | 2462 | if (! conflict_data->in_graph_p |
1756cb66 VM |
2463 | || ALLOCNO_ASSIGNED_P (conflict_a) |
2464 | || !(hard_reg_set_intersect_p | |
27508f5f VM |
2465 | (ALLOCNO_COLOR_DATA (a)->profitable_hard_regs, |
2466 | conflict_data->profitable_hard_regs))) | |
22b0982c | 2467 | continue; |
8c679205 VM |
2468 | for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = pref->next_pref) |
2469 | conflict_data->conflict_allocno_hard_prefs -= pref->freq; | |
2470 | if (conflict_data->colorable_p) | |
2471 | continue; | |
1756cb66 VM |
2472 | ira_assert (bitmap_bit_p (coloring_allocno_bitmap, |
2473 | ALLOCNO_NUM (conflict_a))); | |
27508f5f | 2474 | if (update_left_conflict_sizes_p (conflict_a, a, size)) |
22b0982c VM |
2475 | { |
2476 | delete_allocno_from_bucket | |
27508f5f | 2477 | (conflict_a, &uncolorable_allocno_bucket); |
bf08fb16 | 2478 | add_allocno_to_ordered_colorable_bucket (conflict_a); |
1756cb66 VM |
2479 | if (internal_flag_ira_verbose > 4 && ira_dump_file != NULL) |
2480 | { | |
2481 | fprintf (ira_dump_file, " Making"); | |
2482 | ira_print_expanded_allocno (conflict_a); | |
2483 | fprintf (ira_dump_file, " colorable\n"); | |
2484 | } | |
548a6322 | 2485 | } |
1756cb66 | 2486 | |
548a6322 | 2487 | } |
058e97ec VM |
2488 | } |
2489 | } | |
2490 | ||
2491 | /* Put ALLOCNO onto the coloring stack and remove it from its bucket. | |
2492 | The allocno is in the colorable bucket if COLORABLE_P is TRUE. */ | |
2493 | static void | |
2494 | remove_allocno_from_bucket_and_push (ira_allocno_t allocno, bool colorable_p) | |
2495 | { | |
058e97ec VM |
2496 | if (colorable_p) |
2497 | delete_allocno_from_bucket (allocno, &colorable_allocno_bucket); | |
2498 | else | |
2499 | delete_allocno_from_bucket (allocno, &uncolorable_allocno_bucket); | |
2500 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
2501 | { | |
2502 | fprintf (ira_dump_file, " Pushing"); | |
22b0982c | 2503 | ira_print_expanded_allocno (allocno); |
30ea859e | 2504 | if (colorable_p) |
1756cb66 VM |
2505 | fprintf (ira_dump_file, "(cost %d)\n", |
2506 | ALLOCNO_COLOR_DATA (allocno)->temp); | |
30ea859e VM |
2507 | else |
2508 | fprintf (ira_dump_file, "(potential spill: %spri=%d, cost=%d)\n", | |
2509 | ALLOCNO_BAD_SPILL_P (allocno) ? "bad spill, " : "", | |
1756cb66 VM |
2510 | allocno_spill_priority (allocno), |
2511 | ALLOCNO_COLOR_DATA (allocno)->temp); | |
2512 | } | |
058e97ec | 2513 | if (! colorable_p) |
1756cb66 | 2514 | ALLOCNO_COLOR_DATA (allocno)->may_be_spilled_p = true; |
548a6322 | 2515 | push_allocno_to_stack (allocno); |
058e97ec VM |
2516 | } |
2517 | ||
2518 | /* Put all allocnos from colorable bucket onto the coloring stack. */ | |
2519 | static void | |
2520 | push_only_colorable (void) | |
2521 | { | |
74dc179a VM |
2522 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
2523 | fprintf (ira_dump_file, " Forming thread from colorable bucket:\n"); | |
bf08fb16 | 2524 | form_threads_from_bucket (colorable_allocno_bucket); |
74dc179a VM |
2525 | for (ira_allocno_t a = colorable_allocno_bucket; |
2526 | a != NULL; | |
2527 | a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno) | |
2528 | update_costs_from_prefs (a); | |
1756cb66 | 2529 | sort_bucket (&colorable_allocno_bucket, bucket_allocno_compare_func); |
058e97ec VM |
2530 | for (;colorable_allocno_bucket != NULL;) |
2531 | remove_allocno_from_bucket_and_push (colorable_allocno_bucket, true); | |
2532 | } | |
2533 | ||
058e97ec | 2534 | /* Return the frequency of exit edges (if EXIT_P) or entry from/to the |
b8698a0f | 2535 | loop given by its LOOP_NODE. */ |
058e97ec VM |
2536 | int |
2537 | ira_loop_edge_freq (ira_loop_tree_node_t loop_node, int regno, bool exit_p) | |
2538 | { | |
2539 | int freq, i; | |
2540 | edge_iterator ei; | |
2541 | edge e; | |
9771b263 | 2542 | vec<edge> edges; |
058e97ec | 2543 | |
2608d841 | 2544 | ira_assert (current_loops != NULL && loop_node->loop != NULL |
058e97ec VM |
2545 | && (regno < 0 || regno >= FIRST_PSEUDO_REGISTER)); |
2546 | freq = 0; | |
2547 | if (! exit_p) | |
2548 | { | |
2549 | FOR_EACH_EDGE (e, ei, loop_node->loop->header->preds) | |
2550 | if (e->src != loop_node->loop->latch | |
2551 | && (regno < 0 | |
bf744527 SB |
2552 | || (bitmap_bit_p (df_get_live_out (e->src), regno) |
2553 | && bitmap_bit_p (df_get_live_in (e->dest), regno)))) | |
058e97ec VM |
2554 | freq += EDGE_FREQUENCY (e); |
2555 | } | |
2556 | else | |
2557 | { | |
2558 | edges = get_loop_exit_edges (loop_node->loop); | |
9771b263 | 2559 | FOR_EACH_VEC_ELT (edges, i, e) |
058e97ec | 2560 | if (regno < 0 |
bf744527 SB |
2561 | || (bitmap_bit_p (df_get_live_out (e->src), regno) |
2562 | && bitmap_bit_p (df_get_live_in (e->dest), regno))) | |
058e97ec | 2563 | freq += EDGE_FREQUENCY (e); |
9771b263 | 2564 | edges.release (); |
058e97ec VM |
2565 | } |
2566 | ||
2567 | return REG_FREQ_FROM_EDGE_FREQ (freq); | |
2568 | } | |
2569 | ||
2570 | /* Calculate and return the cost of putting allocno A into memory. */ | |
2571 | static int | |
2572 | calculate_allocno_spill_cost (ira_allocno_t a) | |
2573 | { | |
2574 | int regno, cost; | |
ef4bddc2 | 2575 | machine_mode mode; |
058e97ec VM |
2576 | enum reg_class rclass; |
2577 | ira_allocno_t parent_allocno; | |
2578 | ira_loop_tree_node_t parent_node, loop_node; | |
2579 | ||
2580 | regno = ALLOCNO_REGNO (a); | |
1756cb66 | 2581 | cost = ALLOCNO_UPDATED_MEMORY_COST (a) - ALLOCNO_UPDATED_CLASS_COST (a); |
058e97ec VM |
2582 | if (ALLOCNO_CAP (a) != NULL) |
2583 | return cost; | |
2584 | loop_node = ALLOCNO_LOOP_TREE_NODE (a); | |
2585 | if ((parent_node = loop_node->parent) == NULL) | |
2586 | return cost; | |
2587 | if ((parent_allocno = parent_node->regno_allocno_map[regno]) == NULL) | |
2588 | return cost; | |
2589 | mode = ALLOCNO_MODE (a); | |
1756cb66 | 2590 | rclass = ALLOCNO_CLASS (a); |
058e97ec VM |
2591 | if (ALLOCNO_HARD_REGNO (parent_allocno) < 0) |
2592 | cost -= (ira_memory_move_cost[mode][rclass][0] | |
2593 | * ira_loop_edge_freq (loop_node, regno, true) | |
2594 | + ira_memory_move_cost[mode][rclass][1] | |
2595 | * ira_loop_edge_freq (loop_node, regno, false)); | |
2596 | else | |
1756cb66 VM |
2597 | { |
2598 | ira_init_register_move_cost_if_necessary (mode); | |
2599 | cost += ((ira_memory_move_cost[mode][rclass][1] | |
2600 | * ira_loop_edge_freq (loop_node, regno, true) | |
2601 | + ira_memory_move_cost[mode][rclass][0] | |
2602 | * ira_loop_edge_freq (loop_node, regno, false)) | |
2603 | - (ira_register_move_cost[mode][rclass][rclass] | |
2604 | * (ira_loop_edge_freq (loop_node, regno, false) | |
2605 | + ira_loop_edge_freq (loop_node, regno, true)))); | |
2606 | } | |
058e97ec VM |
2607 | return cost; |
2608 | } | |
2609 | ||
1756cb66 VM |
2610 | /* Used for sorting allocnos for spilling. */ |
2611 | static inline int | |
2612 | allocno_spill_priority_compare (ira_allocno_t a1, ira_allocno_t a2) | |
058e97ec VM |
2613 | { |
2614 | int pri1, pri2, diff; | |
b8698a0f | 2615 | |
b81a2f0d VM |
2616 | /* Avoid spilling static chain pointer pseudo when non-local goto is |
2617 | used. */ | |
2618 | if (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a1))) | |
2619 | return 1; | |
2620 | else if (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a2))) | |
2621 | return -1; | |
1756cb66 VM |
2622 | if (ALLOCNO_BAD_SPILL_P (a1) && ! ALLOCNO_BAD_SPILL_P (a2)) |
2623 | return 1; | |
2624 | if (ALLOCNO_BAD_SPILL_P (a2) && ! ALLOCNO_BAD_SPILL_P (a1)) | |
2625 | return -1; | |
2626 | pri1 = allocno_spill_priority (a1); | |
2627 | pri2 = allocno_spill_priority (a2); | |
058e97ec VM |
2628 | if ((diff = pri1 - pri2) != 0) |
2629 | return diff; | |
1756cb66 VM |
2630 | if ((diff |
2631 | = ALLOCNO_COLOR_DATA (a1)->temp - ALLOCNO_COLOR_DATA (a2)->temp) != 0) | |
058e97ec VM |
2632 | return diff; |
2633 | return ALLOCNO_NUM (a1) - ALLOCNO_NUM (a2); | |
2634 | } | |
2635 | ||
1756cb66 VM |
2636 | /* Used for sorting allocnos for spilling. */ |
2637 | static int | |
2638 | allocno_spill_sort_compare (const void *v1p, const void *v2p) | |
99710245 | 2639 | { |
1756cb66 VM |
2640 | ira_allocno_t p1 = *(const ira_allocno_t *) v1p; |
2641 | ira_allocno_t p2 = *(const ira_allocno_t *) v2p; | |
99710245 | 2642 | |
1756cb66 | 2643 | return allocno_spill_priority_compare (p1, p2); |
058e97ec VM |
2644 | } |
2645 | ||
2646 | /* Push allocnos to the coloring stack. The order of allocnos in the | |
1756cb66 VM |
2647 | stack defines the order for the subsequent coloring. */ |
2648 | static void | |
2649 | push_allocnos_to_stack (void) | |
2650 | { | |
2651 | ira_allocno_t a; | |
2652 | int cost; | |
2653 | ||
2654 | /* Calculate uncolorable allocno spill costs. */ | |
2655 | for (a = uncolorable_allocno_bucket; | |
2656 | a != NULL; | |
2657 | a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno) | |
2658 | if (ALLOCNO_CLASS (a) != NO_REGS) | |
2659 | { | |
2660 | cost = calculate_allocno_spill_cost (a); | |
2661 | /* ??? Remove cost of copies between the coalesced | |
2662 | allocnos. */ | |
2663 | ALLOCNO_COLOR_DATA (a)->temp = cost; | |
2664 | } | |
2665 | sort_bucket (&uncolorable_allocno_bucket, allocno_spill_sort_compare); | |
2666 | for (;;) | |
2667 | { | |
2668 | push_only_colorable (); | |
2669 | a = uncolorable_allocno_bucket; | |
2670 | if (a == NULL) | |
2671 | break; | |
2672 | remove_allocno_from_bucket_and_push (a, false); | |
058e97ec VM |
2673 | } |
2674 | ira_assert (colorable_allocno_bucket == NULL | |
2675 | && uncolorable_allocno_bucket == NULL); | |
1756cb66 | 2676 | ira_assert (uncolorable_allocnos_num == 0); |
058e97ec VM |
2677 | } |
2678 | ||
2679 | /* Pop the coloring stack and assign hard registers to the popped | |
2680 | allocnos. */ | |
2681 | static void | |
2682 | pop_allocnos_from_stack (void) | |
2683 | { | |
2684 | ira_allocno_t allocno; | |
1756cb66 | 2685 | enum reg_class aclass; |
058e97ec | 2686 | |
9771b263 | 2687 | for (;allocno_stack_vec.length () != 0;) |
058e97ec | 2688 | { |
9771b263 | 2689 | allocno = allocno_stack_vec.pop (); |
1756cb66 | 2690 | aclass = ALLOCNO_CLASS (allocno); |
058e97ec VM |
2691 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
2692 | { | |
2693 | fprintf (ira_dump_file, " Popping"); | |
22b0982c | 2694 | ira_print_expanded_allocno (allocno); |
058e97ec VM |
2695 | fprintf (ira_dump_file, " -- "); |
2696 | } | |
1756cb66 | 2697 | if (aclass == NO_REGS) |
058e97ec VM |
2698 | { |
2699 | ALLOCNO_HARD_REGNO (allocno) = -1; | |
2700 | ALLOCNO_ASSIGNED_P (allocno) = true; | |
2701 | ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (allocno) == NULL); | |
2702 | ira_assert | |
2703 | (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno) == NULL); | |
2704 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
2705 | fprintf (ira_dump_file, "assign memory\n"); | |
2706 | } | |
2707 | else if (assign_hard_reg (allocno, false)) | |
2708 | { | |
2709 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
74dc179a | 2710 | fprintf (ira_dump_file, " assign reg %d\n", |
058e97ec VM |
2711 | ALLOCNO_HARD_REGNO (allocno)); |
2712 | } | |
2713 | else if (ALLOCNO_ASSIGNED_P (allocno)) | |
2714 | { | |
2715 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3b6d1699 VM |
2716 | fprintf (ira_dump_file, "spill%s\n", |
2717 | ALLOCNO_COLOR_DATA (allocno)->may_be_spilled_p | |
2718 | ? "" : "!"); | |
058e97ec | 2719 | } |
1756cb66 | 2720 | ALLOCNO_COLOR_DATA (allocno)->in_graph_p = true; |
ac0ab4f7 BS |
2721 | } |
2722 | } | |
2723 | ||
22b0982c | 2724 | /* Set up number of available hard registers for allocno A. */ |
058e97ec | 2725 | static void |
22b0982c | 2726 | setup_allocno_available_regs_num (ira_allocno_t a) |
058e97ec | 2727 | { |
27508f5f | 2728 | int i, n, hard_regno, hard_regs_num, nwords; |
1756cb66 | 2729 | enum reg_class aclass; |
1756cb66 | 2730 | allocno_color_data_t data; |
058e97ec | 2731 | |
1756cb66 VM |
2732 | aclass = ALLOCNO_CLASS (a); |
2733 | data = ALLOCNO_COLOR_DATA (a); | |
2734 | data->available_regs_num = 0; | |
2735 | if (aclass == NO_REGS) | |
058e97ec | 2736 | return; |
1756cb66 | 2737 | hard_regs_num = ira_class_hard_regs_num[aclass]; |
1756cb66 | 2738 | nwords = ALLOCNO_NUM_OBJECTS (a); |
058e97ec | 2739 | for (n = 0, i = hard_regs_num - 1; i >= 0; i--) |
478ab26d | 2740 | { |
1756cb66 | 2741 | hard_regno = ira_class_hard_regs[aclass][i]; |
27508f5f VM |
2742 | /* Checking only profitable hard regs. */ |
2743 | if (TEST_HARD_REG_BIT (data->profitable_hard_regs, hard_regno)) | |
478ab26d VM |
2744 | n++; |
2745 | } | |
1756cb66 VM |
2746 | data->available_regs_num = n; |
2747 | if (internal_flag_ira_verbose <= 2 || ira_dump_file == NULL) | |
2748 | return; | |
2749 | fprintf | |
2750 | (ira_dump_file, | |
27508f5f | 2751 | " Allocno a%dr%d of %s(%d) has %d avail. regs ", |
1756cb66 VM |
2752 | ALLOCNO_NUM (a), ALLOCNO_REGNO (a), |
2753 | reg_class_names[aclass], ira_class_hard_regs_num[aclass], n); | |
27508f5f VM |
2754 | print_hard_reg_set (ira_dump_file, data->profitable_hard_regs, false); |
2755 | fprintf (ira_dump_file, ", %snode: ", | |
a8579651 | 2756 | data->profitable_hard_regs == data->hard_regs_node->hard_regs->set |
27508f5f VM |
2757 | ? "" : "^"); |
2758 | print_hard_reg_set (ira_dump_file, | |
2759 | data->hard_regs_node->hard_regs->set, false); | |
1756cb66 | 2760 | for (i = 0; i < nwords; i++) |
22b0982c | 2761 | { |
1756cb66 | 2762 | ira_object_t obj = ALLOCNO_OBJECT (a, i); |
ac0ab4f7 | 2763 | |
1756cb66 | 2764 | if (nwords != 1) |
22b0982c | 2765 | { |
1756cb66 VM |
2766 | if (i != 0) |
2767 | fprintf (ira_dump_file, ", "); | |
2768 | fprintf (ira_dump_file, " obj %d", i); | |
22b0982c | 2769 | } |
1756cb66 VM |
2770 | fprintf (ira_dump_file, " (confl regs = "); |
2771 | print_hard_reg_set (ira_dump_file, OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), | |
2772 | false); | |
27508f5f | 2773 | fprintf (ira_dump_file, ")"); |
22b0982c | 2774 | } |
1756cb66 | 2775 | fprintf (ira_dump_file, "\n"); |
058e97ec VM |
2776 | } |
2777 | ||
2778 | /* Put ALLOCNO in a bucket corresponding to its number and size of its | |
2779 | conflicting allocnos and hard registers. */ | |
2780 | static void | |
2781 | put_allocno_into_bucket (ira_allocno_t allocno) | |
2782 | { | |
1756cb66 | 2783 | ALLOCNO_COLOR_DATA (allocno)->in_graph_p = true; |
058e97ec | 2784 | setup_allocno_available_regs_num (allocno); |
1756cb66 | 2785 | if (setup_left_conflict_sizes_p (allocno)) |
548a6322 | 2786 | add_allocno_to_bucket (allocno, &colorable_allocno_bucket); |
058e97ec | 2787 | else |
548a6322 | 2788 | add_allocno_to_bucket (allocno, &uncolorable_allocno_bucket); |
058e97ec VM |
2789 | } |
2790 | ||
22b0982c VM |
2791 | /* Map: allocno number -> allocno priority. */ |
2792 | static int *allocno_priorities; | |
058e97ec | 2793 | |
22b0982c VM |
2794 | /* Set up priorities for N allocnos in array |
2795 | CONSIDERATION_ALLOCNOS. */ | |
058e97ec | 2796 | static void |
22b0982c | 2797 | setup_allocno_priorities (ira_allocno_t *consideration_allocnos, int n) |
058e97ec | 2798 | { |
22b0982c VM |
2799 | int i, length, nrefs, priority, max_priority, mult; |
2800 | ira_allocno_t a; | |
058e97ec | 2801 | |
22b0982c VM |
2802 | max_priority = 0; |
2803 | for (i = 0; i < n; i++) | |
7db7ed3c VM |
2804 | { |
2805 | a = consideration_allocnos[i]; | |
2806 | nrefs = ALLOCNO_NREFS (a); | |
2807 | ira_assert (nrefs >= 0); | |
2808 | mult = floor_log2 (ALLOCNO_NREFS (a)) + 1; | |
2809 | ira_assert (mult >= 0); | |
2810 | allocno_priorities[ALLOCNO_NUM (a)] | |
2811 | = priority | |
2812 | = (mult | |
1756cb66 VM |
2813 | * (ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a)) |
2814 | * ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]); | |
7db7ed3c VM |
2815 | if (priority < 0) |
2816 | priority = -priority; | |
2817 | if (max_priority < priority) | |
2818 | max_priority = priority; | |
2819 | } | |
2820 | mult = max_priority == 0 ? 1 : INT_MAX / max_priority; | |
2821 | for (i = 0; i < n; i++) | |
2822 | { | |
2823 | a = consideration_allocnos[i]; | |
2824 | length = ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a); | |
ac0ab4f7 BS |
2825 | if (ALLOCNO_NUM_OBJECTS (a) > 1) |
2826 | length /= ALLOCNO_NUM_OBJECTS (a); | |
7db7ed3c VM |
2827 | if (length <= 0) |
2828 | length = 1; | |
2829 | allocno_priorities[ALLOCNO_NUM (a)] | |
2830 | = allocno_priorities[ALLOCNO_NUM (a)] * mult / length; | |
2831 | } | |
2832 | } | |
2833 | ||
1756cb66 VM |
2834 | /* Sort allocnos according to the profit of usage of a hard register |
2835 | instead of memory for them. */ | |
2836 | static int | |
2837 | allocno_cost_compare_func (const void *v1p, const void *v2p) | |
2838 | { | |
2839 | ira_allocno_t p1 = *(const ira_allocno_t *) v1p; | |
2840 | ira_allocno_t p2 = *(const ira_allocno_t *) v2p; | |
2841 | int c1, c2; | |
2842 | ||
2843 | c1 = ALLOCNO_UPDATED_MEMORY_COST (p1) - ALLOCNO_UPDATED_CLASS_COST (p1); | |
2844 | c2 = ALLOCNO_UPDATED_MEMORY_COST (p2) - ALLOCNO_UPDATED_CLASS_COST (p2); | |
2845 | if (c1 - c2) | |
2846 | return c1 - c2; | |
2847 | ||
2848 | /* If regs are equally good, sort by allocno numbers, so that the | |
2849 | results of qsort leave nothing to chance. */ | |
2850 | return ALLOCNO_NUM (p1) - ALLOCNO_NUM (p2); | |
2851 | } | |
2852 | ||
da178d56 VM |
2853 | /* Return savings on removed copies when ALLOCNO is assigned to |
2854 | HARD_REGNO. */ | |
2855 | static int | |
2856 | allocno_copy_cost_saving (ira_allocno_t allocno, int hard_regno) | |
2857 | { | |
2858 | int cost = 0; | |
b8506a8a | 2859 | machine_mode allocno_mode = ALLOCNO_MODE (allocno); |
da178d56 VM |
2860 | enum reg_class rclass; |
2861 | ira_copy_t cp, next_cp; | |
2862 | ||
2863 | rclass = REGNO_REG_CLASS (hard_regno); | |
c4b1942c VM |
2864 | if (ira_reg_class_max_nregs[rclass][allocno_mode] |
2865 | > ira_class_hard_regs_num[rclass]) | |
2866 | /* For the above condition the cost can be wrong. Use the allocno | |
2867 | class in this case. */ | |
2868 | rclass = ALLOCNO_CLASS (allocno); | |
da178d56 VM |
2869 | for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp) |
2870 | { | |
2871 | if (cp->first == allocno) | |
2872 | { | |
2873 | next_cp = cp->next_first_allocno_copy; | |
2874 | if (ALLOCNO_HARD_REGNO (cp->second) != hard_regno) | |
2875 | continue; | |
2876 | } | |
2877 | else if (cp->second == allocno) | |
2878 | { | |
2879 | next_cp = cp->next_second_allocno_copy; | |
2880 | if (ALLOCNO_HARD_REGNO (cp->first) != hard_regno) | |
2881 | continue; | |
2882 | } | |
2883 | else | |
2884 | gcc_unreachable (); | |
11f2ce1f | 2885 | ira_init_register_move_cost_if_necessary (allocno_mode); |
c4b1942c | 2886 | cost += cp->freq * ira_register_move_cost[allocno_mode][rclass][rclass]; |
da178d56 VM |
2887 | } |
2888 | return cost; | |
2889 | } | |
2890 | ||
1756cb66 VM |
2891 | /* We used Chaitin-Briggs coloring to assign as many pseudos as |
2892 | possible to hard registers. Let us try to improve allocation with | |
2893 | cost point of view. This function improves the allocation by | |
2894 | spilling some allocnos and assigning the freed hard registers to | |
2895 | other allocnos if it decreases the overall allocation cost. */ | |
2896 | static void | |
2897 | improve_allocation (void) | |
2898 | { | |
2899 | unsigned int i; | |
2900 | int j, k, n, hregno, conflict_hregno, base_cost, class_size, word, nwords; | |
2901 | int check, spill_cost, min_cost, nregs, conflict_nregs, r, best; | |
2902 | bool try_p; | |
2903 | enum reg_class aclass; | |
ef4bddc2 | 2904 | machine_mode mode; |
1756cb66 VM |
2905 | int *allocno_costs; |
2906 | int costs[FIRST_PSEUDO_REGISTER]; | |
27508f5f | 2907 | HARD_REG_SET conflicting_regs[2], profitable_hard_regs; |
1756cb66 VM |
2908 | ira_allocno_t a; |
2909 | bitmap_iterator bi; | |
2910 | ||
b81a2f0d VM |
2911 | /* Don't bother to optimize the code with static chain pointer and |
2912 | non-local goto in order not to spill the chain pointer | |
2913 | pseudo. */ | |
2914 | if (cfun->static_chain_decl && crtl->has_nonlocal_goto) | |
2915 | return; | |
1756cb66 VM |
2916 | /* Clear counts used to process conflicting allocnos only once for |
2917 | each allocno. */ | |
2918 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) | |
2919 | ALLOCNO_COLOR_DATA (ira_allocnos[i])->temp = 0; | |
2920 | check = n = 0; | |
2921 | /* Process each allocno and try to assign a hard register to it by | |
2922 | spilling some its conflicting allocnos. */ | |
2923 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) | |
2924 | { | |
2925 | a = ira_allocnos[i]; | |
2926 | ALLOCNO_COLOR_DATA (a)->temp = 0; | |
2927 | if (empty_profitable_hard_regs (a)) | |
2928 | continue; | |
2929 | check++; | |
2930 | aclass = ALLOCNO_CLASS (a); | |
da178d56 | 2931 | allocno_costs = ALLOCNO_HARD_REG_COSTS (a); |
1756cb66 VM |
2932 | if ((hregno = ALLOCNO_HARD_REGNO (a)) < 0) |
2933 | base_cost = ALLOCNO_UPDATED_MEMORY_COST (a); | |
2934 | else if (allocno_costs == NULL) | |
2935 | /* It means that assigning a hard register is not profitable | |
2936 | (we don't waste memory for hard register costs in this | |
2937 | case). */ | |
2938 | continue; | |
2939 | else | |
da178d56 VM |
2940 | base_cost = (allocno_costs[ira_class_hard_reg_index[aclass][hregno]] |
2941 | - allocno_copy_cost_saving (a, hregno)); | |
1756cb66 | 2942 | try_p = false; |
27508f5f VM |
2943 | get_conflict_and_start_profitable_regs (a, false, |
2944 | conflicting_regs, | |
2945 | &profitable_hard_regs); | |
1756cb66 VM |
2946 | class_size = ira_class_hard_regs_num[aclass]; |
2947 | /* Set up cost improvement for usage of each profitable hard | |
2948 | register for allocno A. */ | |
2949 | for (j = 0; j < class_size; j++) | |
2950 | { | |
2951 | hregno = ira_class_hard_regs[aclass][j]; | |
2952 | if (! check_hard_reg_p (a, hregno, | |
2953 | conflicting_regs, profitable_hard_regs)) | |
2954 | continue; | |
2955 | ira_assert (ira_class_hard_reg_index[aclass][hregno] == j); | |
2956 | k = allocno_costs == NULL ? 0 : j; | |
2957 | costs[hregno] = (allocno_costs == NULL | |
2958 | ? ALLOCNO_UPDATED_CLASS_COST (a) : allocno_costs[k]); | |
da178d56 | 2959 | costs[hregno] -= allocno_copy_cost_saving (a, hregno); |
1756cb66 VM |
2960 | costs[hregno] -= base_cost; |
2961 | if (costs[hregno] < 0) | |
2962 | try_p = true; | |
2963 | } | |
2964 | if (! try_p) | |
2965 | /* There is no chance to improve the allocation cost by | |
2966 | assigning hard register to allocno A even without spilling | |
2967 | conflicting allocnos. */ | |
2968 | continue; | |
2969 | mode = ALLOCNO_MODE (a); | |
2970 | nwords = ALLOCNO_NUM_OBJECTS (a); | |
2971 | /* Process each allocno conflicting with A and update the cost | |
2972 | improvement for profitable hard registers of A. To use a | |
2973 | hard register for A we need to spill some conflicting | |
2974 | allocnos and that creates penalty for the cost | |
2975 | improvement. */ | |
2976 | for (word = 0; word < nwords; word++) | |
2977 | { | |
2978 | ira_object_t conflict_obj; | |
2979 | ira_object_t obj = ALLOCNO_OBJECT (a, word); | |
2980 | ira_object_conflict_iterator oci; | |
2981 | ||
2982 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
2983 | { | |
2984 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); | |
2985 | ||
2986 | if (ALLOCNO_COLOR_DATA (conflict_a)->temp == check) | |
2987 | /* We already processed this conflicting allocno | |
2988 | because we processed earlier another object of the | |
2989 | conflicting allocno. */ | |
2990 | continue; | |
2991 | ALLOCNO_COLOR_DATA (conflict_a)->temp = check; | |
2992 | if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0) | |
2993 | continue; | |
2994 | spill_cost = ALLOCNO_UPDATED_MEMORY_COST (conflict_a); | |
2995 | k = (ira_class_hard_reg_index | |
2996 | [ALLOCNO_CLASS (conflict_a)][conflict_hregno]); | |
2997 | ira_assert (k >= 0); | |
da178d56 | 2998 | if ((allocno_costs = ALLOCNO_HARD_REG_COSTS (conflict_a)) |
1756cb66 VM |
2999 | != NULL) |
3000 | spill_cost -= allocno_costs[k]; | |
1756cb66 VM |
3001 | else |
3002 | spill_cost -= ALLOCNO_UPDATED_CLASS_COST (conflict_a); | |
da178d56 VM |
3003 | spill_cost |
3004 | += allocno_copy_cost_saving (conflict_a, conflict_hregno); | |
ad474626 RS |
3005 | conflict_nregs = hard_regno_nregs (conflict_hregno, |
3006 | ALLOCNO_MODE (conflict_a)); | |
1756cb66 | 3007 | for (r = conflict_hregno; |
4edd6298 | 3008 | r >= 0 && (int) end_hard_regno (mode, r) > conflict_hregno; |
1756cb66 VM |
3009 | r--) |
3010 | if (check_hard_reg_p (a, r, | |
3011 | conflicting_regs, profitable_hard_regs)) | |
3012 | costs[r] += spill_cost; | |
3013 | for (r = conflict_hregno + 1; | |
3014 | r < conflict_hregno + conflict_nregs; | |
3015 | r++) | |
3016 | if (check_hard_reg_p (a, r, | |
3017 | conflicting_regs, profitable_hard_regs)) | |
3018 | costs[r] += spill_cost; | |
3019 | } | |
3020 | } | |
3021 | min_cost = INT_MAX; | |
3022 | best = -1; | |
3023 | /* Now we choose hard register for A which results in highest | |
3024 | allocation cost improvement. */ | |
3025 | for (j = 0; j < class_size; j++) | |
3026 | { | |
3027 | hregno = ira_class_hard_regs[aclass][j]; | |
3028 | if (check_hard_reg_p (a, hregno, | |
3029 | conflicting_regs, profitable_hard_regs) | |
3030 | && min_cost > costs[hregno]) | |
3031 | { | |
3032 | best = hregno; | |
3033 | min_cost = costs[hregno]; | |
3034 | } | |
3035 | } | |
3036 | if (min_cost >= 0) | |
3037 | /* We are in a situation when assigning any hard register to A | |
3038 | by spilling some conflicting allocnos does not improve the | |
3039 | allocation cost. */ | |
3040 | continue; | |
ad474626 | 3041 | nregs = hard_regno_nregs (best, mode); |
1756cb66 VM |
3042 | /* Now spill conflicting allocnos which contain a hard register |
3043 | of A when we assign the best chosen hard register to it. */ | |
3044 | for (word = 0; word < nwords; word++) | |
3045 | { | |
3046 | ira_object_t conflict_obj; | |
3047 | ira_object_t obj = ALLOCNO_OBJECT (a, word); | |
3048 | ira_object_conflict_iterator oci; | |
3049 | ||
3050 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
3051 | { | |
3052 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); | |
3053 | ||
3054 | if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0) | |
3055 | continue; | |
ad474626 RS |
3056 | conflict_nregs = hard_regno_nregs (conflict_hregno, |
3057 | ALLOCNO_MODE (conflict_a)); | |
1756cb66 VM |
3058 | if (best + nregs <= conflict_hregno |
3059 | || conflict_hregno + conflict_nregs <= best) | |
3060 | /* No intersection. */ | |
3061 | continue; | |
3062 | ALLOCNO_HARD_REGNO (conflict_a) = -1; | |
3063 | sorted_allocnos[n++] = conflict_a; | |
3064 | if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL) | |
3065 | fprintf (ira_dump_file, "Spilling a%dr%d for a%dr%d\n", | |
3066 | ALLOCNO_NUM (conflict_a), ALLOCNO_REGNO (conflict_a), | |
3067 | ALLOCNO_NUM (a), ALLOCNO_REGNO (a)); | |
3068 | } | |
3069 | } | |
3070 | /* Assign the best chosen hard register to A. */ | |
3071 | ALLOCNO_HARD_REGNO (a) = best; | |
3072 | if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL) | |
3073 | fprintf (ira_dump_file, "Assigning %d to a%dr%d\n", | |
3074 | best, ALLOCNO_NUM (a), ALLOCNO_REGNO (a)); | |
3075 | } | |
3076 | if (n == 0) | |
3077 | return; | |
3078 | /* We spilled some allocnos to assign their hard registers to other | |
3079 | allocnos. The spilled allocnos are now in array | |
3080 | 'sorted_allocnos'. There is still a possibility that some of the | |
3081 | spilled allocnos can get hard registers. So let us try assign | |
3082 | them hard registers again (just a reminder -- function | |
3083 | 'assign_hard_reg' assigns hard registers only if it is possible | |
3084 | and profitable). We process the spilled allocnos with biggest | |
3085 | benefit to get hard register first -- see function | |
3086 | 'allocno_cost_compare_func'. */ | |
3087 | qsort (sorted_allocnos, n, sizeof (ira_allocno_t), | |
3088 | allocno_cost_compare_func); | |
3089 | for (j = 0; j < n; j++) | |
3090 | { | |
3091 | a = sorted_allocnos[j]; | |
3092 | ALLOCNO_ASSIGNED_P (a) = false; | |
3093 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3094 | { | |
3095 | fprintf (ira_dump_file, " "); | |
3096 | ira_print_expanded_allocno (a); | |
3097 | fprintf (ira_dump_file, " -- "); | |
3098 | } | |
3099 | if (assign_hard_reg (a, false)) | |
3100 | { | |
3101 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3102 | fprintf (ira_dump_file, "assign hard reg %d\n", | |
3103 | ALLOCNO_HARD_REGNO (a)); | |
3104 | } | |
3105 | else | |
3106 | { | |
3107 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3108 | fprintf (ira_dump_file, "assign memory\n"); | |
3109 | } | |
3110 | } | |
3111 | } | |
3112 | ||
aeb9f7cf | 3113 | /* Sort allocnos according to their priorities. */ |
7db7ed3c VM |
3114 | static int |
3115 | allocno_priority_compare_func (const void *v1p, const void *v2p) | |
3116 | { | |
3117 | ira_allocno_t a1 = *(const ira_allocno_t *) v1p; | |
3118 | ira_allocno_t a2 = *(const ira_allocno_t *) v2p; | |
158ec018 | 3119 | int pri1, pri2, diff; |
7db7ed3c | 3120 | |
b81a2f0d VM |
3121 | /* Assign hard reg to static chain pointer pseudo first when |
3122 | non-local goto is used. */ | |
158ec018 AM |
3123 | if ((diff = (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a2)) |
3124 | - non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a1)))) != 0) | |
3125 | return diff; | |
7db7ed3c VM |
3126 | pri1 = allocno_priorities[ALLOCNO_NUM (a1)]; |
3127 | pri2 = allocno_priorities[ALLOCNO_NUM (a2)]; | |
71af27d2 OH |
3128 | if (pri2 != pri1) |
3129 | return SORTGT (pri2, pri1); | |
7db7ed3c VM |
3130 | |
3131 | /* If regs are equally good, sort by allocnos, so that the results of | |
3132 | qsort leave nothing to chance. */ | |
3133 | return ALLOCNO_NUM (a1) - ALLOCNO_NUM (a2); | |
3134 | } | |
3135 | ||
058e97ec VM |
3136 | /* Chaitin-Briggs coloring for allocnos in COLORING_ALLOCNO_BITMAP |
3137 | taking into account allocnos in CONSIDERATION_ALLOCNO_BITMAP. */ | |
3138 | static void | |
3139 | color_allocnos (void) | |
3140 | { | |
7db7ed3c | 3141 | unsigned int i, n; |
058e97ec VM |
3142 | bitmap_iterator bi; |
3143 | ira_allocno_t a; | |
3144 | ||
76763a6d | 3145 | setup_profitable_hard_regs (); |
3b6d1699 VM |
3146 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) |
3147 | { | |
3b6d1699 VM |
3148 | allocno_color_data_t data; |
3149 | ira_pref_t pref, next_pref; | |
3150 | ||
3151 | a = ira_allocnos[i]; | |
3b6d1699 | 3152 | data = ALLOCNO_COLOR_DATA (a); |
8c679205 | 3153 | data->conflict_allocno_hard_prefs = 0; |
3b6d1699 VM |
3154 | for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = next_pref) |
3155 | { | |
3156 | next_pref = pref->next_pref; | |
3157 | if (! ira_hard_reg_in_set_p (pref->hard_regno, | |
3158 | ALLOCNO_MODE (a), | |
3159 | data->profitable_hard_regs)) | |
3160 | ira_remove_pref (pref); | |
3161 | } | |
3162 | } | |
8c679205 | 3163 | |
7db7ed3c | 3164 | if (flag_ira_algorithm == IRA_ALGORITHM_PRIORITY) |
058e97ec | 3165 | { |
7db7ed3c VM |
3166 | n = 0; |
3167 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) | |
058e97ec | 3168 | { |
7db7ed3c | 3169 | a = ira_allocnos[i]; |
1756cb66 | 3170 | if (ALLOCNO_CLASS (a) == NO_REGS) |
058e97ec | 3171 | { |
7db7ed3c VM |
3172 | ALLOCNO_HARD_REGNO (a) = -1; |
3173 | ALLOCNO_ASSIGNED_P (a) = true; | |
3174 | ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL); | |
3175 | ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a) == NULL); | |
3176 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3177 | { | |
3178 | fprintf (ira_dump_file, " Spill"); | |
22b0982c | 3179 | ira_print_expanded_allocno (a); |
7db7ed3c VM |
3180 | fprintf (ira_dump_file, "\n"); |
3181 | } | |
3182 | continue; | |
058e97ec | 3183 | } |
7db7ed3c VM |
3184 | sorted_allocnos[n++] = a; |
3185 | } | |
3186 | if (n != 0) | |
3187 | { | |
3188 | setup_allocno_priorities (sorted_allocnos, n); | |
3189 | qsort (sorted_allocnos, n, sizeof (ira_allocno_t), | |
3190 | allocno_priority_compare_func); | |
3191 | for (i = 0; i < n; i++) | |
3192 | { | |
3193 | a = sorted_allocnos[i]; | |
3194 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3195 | { | |
3196 | fprintf (ira_dump_file, " "); | |
22b0982c | 3197 | ira_print_expanded_allocno (a); |
7db7ed3c VM |
3198 | fprintf (ira_dump_file, " -- "); |
3199 | } | |
3200 | if (assign_hard_reg (a, false)) | |
3201 | { | |
3202 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3203 | fprintf (ira_dump_file, "assign hard reg %d\n", | |
3204 | ALLOCNO_HARD_REGNO (a)); | |
3205 | } | |
3206 | else | |
3207 | { | |
3208 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3209 | fprintf (ira_dump_file, "assign memory\n"); | |
3210 | } | |
3211 | } | |
3212 | } | |
3213 | } | |
3214 | else | |
3215 | { | |
27508f5f | 3216 | form_allocno_hard_regs_nodes_forest (); |
1756cb66 VM |
3217 | if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL) |
3218 | print_hard_regs_forest (ira_dump_file); | |
7db7ed3c VM |
3219 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) |
3220 | { | |
3221 | a = ira_allocnos[i]; | |
1756cb66 | 3222 | if (ALLOCNO_CLASS (a) != NO_REGS && ! empty_profitable_hard_regs (a)) |
3b6d1699 VM |
3223 | { |
3224 | ALLOCNO_COLOR_DATA (a)->in_graph_p = true; | |
8c679205 | 3225 | update_conflict_allocno_hard_prefs (a); |
3b6d1699 | 3226 | } |
1756cb66 | 3227 | else |
7db7ed3c VM |
3228 | { |
3229 | ALLOCNO_HARD_REGNO (a) = -1; | |
3230 | ALLOCNO_ASSIGNED_P (a) = true; | |
1756cb66 VM |
3231 | /* We don't need updated costs anymore. */ |
3232 | ira_free_allocno_updated_costs (a); | |
7db7ed3c VM |
3233 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) |
3234 | { | |
3235 | fprintf (ira_dump_file, " Spill"); | |
22b0982c | 3236 | ira_print_expanded_allocno (a); |
7db7ed3c VM |
3237 | fprintf (ira_dump_file, "\n"); |
3238 | } | |
7db7ed3c | 3239 | } |
1756cb66 VM |
3240 | } |
3241 | /* Put the allocnos into the corresponding buckets. */ | |
3242 | colorable_allocno_bucket = NULL; | |
3243 | uncolorable_allocno_bucket = NULL; | |
3244 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) | |
3245 | { | |
3246 | a = ira_allocnos[i]; | |
3247 | if (ALLOCNO_COLOR_DATA (a)->in_graph_p) | |
3248 | put_allocno_into_bucket (a); | |
058e97ec | 3249 | } |
7db7ed3c VM |
3250 | push_allocnos_to_stack (); |
3251 | pop_allocnos_from_stack (); | |
27508f5f | 3252 | finish_allocno_hard_regs_nodes_forest (); |
058e97ec | 3253 | } |
1756cb66 | 3254 | improve_allocation (); |
058e97ec VM |
3255 | } |
3256 | ||
3257 | \f | |
3258 | ||
2b9c63a2 | 3259 | /* Output information about the loop given by its LOOP_TREE_NODE. */ |
058e97ec VM |
3260 | static void |
3261 | print_loop_title (ira_loop_tree_node_t loop_tree_node) | |
3262 | { | |
3263 | unsigned int j; | |
3264 | bitmap_iterator bi; | |
ea1c67e6 VM |
3265 | ira_loop_tree_node_t subloop_node, dest_loop_node; |
3266 | edge e; | |
3267 | edge_iterator ei; | |
058e97ec | 3268 | |
2608d841 VM |
3269 | if (loop_tree_node->parent == NULL) |
3270 | fprintf (ira_dump_file, | |
3271 | "\n Loop 0 (parent -1, header bb%d, depth 0)\n bbs:", | |
3272 | NUM_FIXED_BLOCKS); | |
3273 | else | |
3274 | { | |
3275 | ira_assert (current_loops != NULL && loop_tree_node->loop != NULL); | |
3276 | fprintf (ira_dump_file, | |
3277 | "\n Loop %d (parent %d, header bb%d, depth %d)\n bbs:", | |
3278 | loop_tree_node->loop_num, loop_tree_node->parent->loop_num, | |
3279 | loop_tree_node->loop->header->index, | |
3280 | loop_depth (loop_tree_node->loop)); | |
3281 | } | |
ea1c67e6 VM |
3282 | for (subloop_node = loop_tree_node->children; |
3283 | subloop_node != NULL; | |
3284 | subloop_node = subloop_node->next) | |
3285 | if (subloop_node->bb != NULL) | |
3286 | { | |
3287 | fprintf (ira_dump_file, " %d", subloop_node->bb->index); | |
3288 | FOR_EACH_EDGE (e, ei, subloop_node->bb->succs) | |
fefa31b5 | 3289 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
ea1c67e6 VM |
3290 | && ((dest_loop_node = IRA_BB_NODE (e->dest)->parent) |
3291 | != loop_tree_node)) | |
3292 | fprintf (ira_dump_file, "(->%d:l%d)", | |
2608d841 | 3293 | e->dest->index, dest_loop_node->loop_num); |
ea1c67e6 VM |
3294 | } |
3295 | fprintf (ira_dump_file, "\n all:"); | |
49d988e7 | 3296 | EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->all_allocnos, 0, j, bi) |
058e97ec VM |
3297 | fprintf (ira_dump_file, " %dr%d", j, ALLOCNO_REGNO (ira_allocnos[j])); |
3298 | fprintf (ira_dump_file, "\n modified regnos:"); | |
3299 | EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->modified_regnos, 0, j, bi) | |
3300 | fprintf (ira_dump_file, " %d", j); | |
3301 | fprintf (ira_dump_file, "\n border:"); | |
3302 | EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->border_allocnos, 0, j, bi) | |
3303 | fprintf (ira_dump_file, " %dr%d", j, ALLOCNO_REGNO (ira_allocnos[j])); | |
3304 | fprintf (ira_dump_file, "\n Pressure:"); | |
1756cb66 | 3305 | for (j = 0; (int) j < ira_pressure_classes_num; j++) |
058e97ec | 3306 | { |
1756cb66 | 3307 | enum reg_class pclass; |
b8698a0f | 3308 | |
1756cb66 VM |
3309 | pclass = ira_pressure_classes[j]; |
3310 | if (loop_tree_node->reg_pressure[pclass] == 0) | |
058e97ec | 3311 | continue; |
1756cb66 VM |
3312 | fprintf (ira_dump_file, " %s=%d", reg_class_names[pclass], |
3313 | loop_tree_node->reg_pressure[pclass]); | |
058e97ec VM |
3314 | } |
3315 | fprintf (ira_dump_file, "\n"); | |
3316 | } | |
3317 | ||
3318 | /* Color the allocnos inside loop (in the extreme case it can be all | |
3319 | of the function) given the corresponding LOOP_TREE_NODE. The | |
3320 | function is called for each loop during top-down traverse of the | |
3321 | loop tree. */ | |
3322 | static void | |
3323 | color_pass (ira_loop_tree_node_t loop_tree_node) | |
3324 | { | |
27508f5f | 3325 | int regno, hard_regno, index = -1, n; |
058e97ec VM |
3326 | int cost, exit_freq, enter_freq; |
3327 | unsigned int j; | |
3328 | bitmap_iterator bi; | |
ef4bddc2 | 3329 | machine_mode mode; |
1756cb66 | 3330 | enum reg_class rclass, aclass, pclass; |
058e97ec VM |
3331 | ira_allocno_t a, subloop_allocno; |
3332 | ira_loop_tree_node_t subloop_node; | |
3333 | ||
3334 | ira_assert (loop_tree_node->bb == NULL); | |
3335 | if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL) | |
3336 | print_loop_title (loop_tree_node); | |
3337 | ||
49d988e7 | 3338 | bitmap_copy (coloring_allocno_bitmap, loop_tree_node->all_allocnos); |
058e97ec | 3339 | bitmap_copy (consideration_allocno_bitmap, coloring_allocno_bitmap); |
27508f5f | 3340 | n = 0; |
1756cb66 VM |
3341 | EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi) |
3342 | { | |
3343 | a = ira_allocnos[j]; | |
3344 | n++; | |
1756cb66 VM |
3345 | if (! ALLOCNO_ASSIGNED_P (a)) |
3346 | continue; | |
3347 | bitmap_clear_bit (coloring_allocno_bitmap, ALLOCNO_NUM (a)); | |
3348 | } | |
3349 | allocno_color_data | |
3350 | = (allocno_color_data_t) ira_allocate (sizeof (struct allocno_color_data) | |
3351 | * n); | |
3352 | memset (allocno_color_data, 0, sizeof (struct allocno_color_data) * n); | |
27508f5f VM |
3353 | curr_allocno_process = 0; |
3354 | n = 0; | |
058e97ec VM |
3355 | EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi) |
3356 | { | |
3357 | a = ira_allocnos[j]; | |
1756cb66 VM |
3358 | ALLOCNO_ADD_DATA (a) = allocno_color_data + n; |
3359 | n++; | |
058e97ec | 3360 | } |
bf08fb16 | 3361 | init_allocno_threads (); |
058e97ec VM |
3362 | /* Color all mentioned allocnos including transparent ones. */ |
3363 | color_allocnos (); | |
3364 | /* Process caps. They are processed just once. */ | |
7db7ed3c VM |
3365 | if (flag_ira_region == IRA_REGION_MIXED |
3366 | || flag_ira_region == IRA_REGION_ALL) | |
49d988e7 | 3367 | EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->all_allocnos, 0, j, bi) |
058e97ec VM |
3368 | { |
3369 | a = ira_allocnos[j]; | |
3370 | if (ALLOCNO_CAP_MEMBER (a) == NULL) | |
3371 | continue; | |
3372 | /* Remove from processing in the next loop. */ | |
3373 | bitmap_clear_bit (consideration_allocno_bitmap, j); | |
1756cb66 VM |
3374 | rclass = ALLOCNO_CLASS (a); |
3375 | pclass = ira_pressure_class_translate[rclass]; | |
7db7ed3c | 3376 | if (flag_ira_region == IRA_REGION_MIXED |
1756cb66 | 3377 | && (loop_tree_node->reg_pressure[pclass] |
f508f827 | 3378 | <= ira_class_hard_regs_num[pclass])) |
058e97ec VM |
3379 | { |
3380 | mode = ALLOCNO_MODE (a); | |
3381 | hard_regno = ALLOCNO_HARD_REGNO (a); | |
3382 | if (hard_regno >= 0) | |
3383 | { | |
3384 | index = ira_class_hard_reg_index[rclass][hard_regno]; | |
3385 | ira_assert (index >= 0); | |
3386 | } | |
3387 | regno = ALLOCNO_REGNO (a); | |
3388 | subloop_allocno = ALLOCNO_CAP_MEMBER (a); | |
3389 | subloop_node = ALLOCNO_LOOP_TREE_NODE (subloop_allocno); | |
3390 | ira_assert (!ALLOCNO_ASSIGNED_P (subloop_allocno)); | |
3391 | ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno; | |
3392 | ALLOCNO_ASSIGNED_P (subloop_allocno) = true; | |
3393 | if (hard_regno >= 0) | |
c73ccc80 | 3394 | update_costs_from_copies (subloop_allocno, true, true); |
2b9c63a2 | 3395 | /* We don't need updated costs anymore. */ |
058e97ec VM |
3396 | ira_free_allocno_updated_costs (subloop_allocno); |
3397 | } | |
3398 | } | |
3399 | /* Update costs of the corresponding allocnos (not caps) in the | |
3400 | subloops. */ | |
3401 | for (subloop_node = loop_tree_node->subloops; | |
3402 | subloop_node != NULL; | |
3403 | subloop_node = subloop_node->subloop_next) | |
3404 | { | |
3405 | ira_assert (subloop_node->bb == NULL); | |
3406 | EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi) | |
3407 | { | |
3408 | a = ira_allocnos[j]; | |
3409 | ira_assert (ALLOCNO_CAP_MEMBER (a) == NULL); | |
3410 | mode = ALLOCNO_MODE (a); | |
1756cb66 VM |
3411 | rclass = ALLOCNO_CLASS (a); |
3412 | pclass = ira_pressure_class_translate[rclass]; | |
058e97ec | 3413 | hard_regno = ALLOCNO_HARD_REGNO (a); |
7db7ed3c | 3414 | /* Use hard register class here. ??? */ |
058e97ec VM |
3415 | if (hard_regno >= 0) |
3416 | { | |
3417 | index = ira_class_hard_reg_index[rclass][hard_regno]; | |
3418 | ira_assert (index >= 0); | |
3419 | } | |
3420 | regno = ALLOCNO_REGNO (a); | |
3421 | /* ??? conflict costs */ | |
3422 | subloop_allocno = subloop_node->regno_allocno_map[regno]; | |
3423 | if (subloop_allocno == NULL | |
3424 | || ALLOCNO_CAP (subloop_allocno) != NULL) | |
3425 | continue; | |
1756cb66 | 3426 | ira_assert (ALLOCNO_CLASS (subloop_allocno) == rclass); |
49d988e7 VM |
3427 | ira_assert (bitmap_bit_p (subloop_node->all_allocnos, |
3428 | ALLOCNO_NUM (subloop_allocno))); | |
bcb21886 KY |
3429 | if ((flag_ira_region == IRA_REGION_MIXED |
3430 | && (loop_tree_node->reg_pressure[pclass] | |
3431 | <= ira_class_hard_regs_num[pclass])) | |
3432 | || (pic_offset_table_rtx != NULL | |
3c20c9bc VM |
3433 | && regno == (int) REGNO (pic_offset_table_rtx)) |
3434 | /* Avoid overlapped multi-registers. Moves between them | |
3435 | might result in wrong code generation. */ | |
3436 | || (hard_regno >= 0 | |
3437 | && ira_reg_class_max_nregs[pclass][mode] > 1)) | |
058e97ec VM |
3438 | { |
3439 | if (! ALLOCNO_ASSIGNED_P (subloop_allocno)) | |
3440 | { | |
3441 | ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno; | |
3442 | ALLOCNO_ASSIGNED_P (subloop_allocno) = true; | |
3443 | if (hard_regno >= 0) | |
c73ccc80 | 3444 | update_costs_from_copies (subloop_allocno, true, true); |
2b9c63a2 | 3445 | /* We don't need updated costs anymore. */ |
058e97ec VM |
3446 | ira_free_allocno_updated_costs (subloop_allocno); |
3447 | } | |
3448 | continue; | |
3449 | } | |
3450 | exit_freq = ira_loop_edge_freq (subloop_node, regno, true); | |
3451 | enter_freq = ira_loop_edge_freq (subloop_node, regno, false); | |
3452 | ira_assert (regno < ira_reg_equiv_len); | |
55a2c322 | 3453 | if (ira_equiv_no_lvalue_p (regno)) |
058e97ec VM |
3454 | { |
3455 | if (! ALLOCNO_ASSIGNED_P (subloop_allocno)) | |
3456 | { | |
3457 | ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno; | |
3458 | ALLOCNO_ASSIGNED_P (subloop_allocno) = true; | |
3459 | if (hard_regno >= 0) | |
c73ccc80 | 3460 | update_costs_from_copies (subloop_allocno, true, true); |
2b9c63a2 | 3461 | /* We don't need updated costs anymore. */ |
058e97ec VM |
3462 | ira_free_allocno_updated_costs (subloop_allocno); |
3463 | } | |
3464 | } | |
3465 | else if (hard_regno < 0) | |
3466 | { | |
3467 | ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno) | |
3468 | -= ((ira_memory_move_cost[mode][rclass][1] * enter_freq) | |
3469 | + (ira_memory_move_cost[mode][rclass][0] * exit_freq)); | |
3470 | } | |
3471 | else | |
3472 | { | |
1756cb66 VM |
3473 | aclass = ALLOCNO_CLASS (subloop_allocno); |
3474 | ira_init_register_move_cost_if_necessary (mode); | |
3475 | cost = (ira_register_move_cost[mode][rclass][rclass] | |
058e97ec | 3476 | * (exit_freq + enter_freq)); |
cb1ca6ac | 3477 | ira_allocate_and_set_or_copy_costs |
1756cb66 VM |
3478 | (&ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno), aclass, |
3479 | ALLOCNO_UPDATED_CLASS_COST (subloop_allocno), | |
cb1ca6ac VM |
3480 | ALLOCNO_HARD_REG_COSTS (subloop_allocno)); |
3481 | ira_allocate_and_set_or_copy_costs | |
3482 | (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno), | |
1756cb66 | 3483 | aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (subloop_allocno)); |
cb1ca6ac VM |
3484 | ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index] -= cost; |
3485 | ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno)[index] | |
058e97ec | 3486 | -= cost; |
1756cb66 | 3487 | if (ALLOCNO_UPDATED_CLASS_COST (subloop_allocno) |
cb1ca6ac | 3488 | > ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index]) |
1756cb66 | 3489 | ALLOCNO_UPDATED_CLASS_COST (subloop_allocno) |
cb1ca6ac | 3490 | = ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index]; |
058e97ec VM |
3491 | ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno) |
3492 | += (ira_memory_move_cost[mode][rclass][0] * enter_freq | |
3493 | + ira_memory_move_cost[mode][rclass][1] * exit_freq); | |
058e97ec VM |
3494 | } |
3495 | } | |
3496 | } | |
1756cb66 | 3497 | ira_free (allocno_color_data); |
bf08fb16 | 3498 | EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi) |
1756cb66 VM |
3499 | { |
3500 | a = ira_allocnos[j]; | |
3501 | ALLOCNO_ADD_DATA (a) = NULL; | |
1756cb66 | 3502 | } |
058e97ec VM |
3503 | } |
3504 | ||
3505 | /* Initialize the common data for coloring and calls functions to do | |
3506 | Chaitin-Briggs and regional coloring. */ | |
3507 | static void | |
3508 | do_coloring (void) | |
3509 | { | |
3510 | coloring_allocno_bitmap = ira_allocate_bitmap (); | |
058e97ec VM |
3511 | if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL) |
3512 | fprintf (ira_dump_file, "\n**** Allocnos coloring:\n\n"); | |
b8698a0f | 3513 | |
058e97ec VM |
3514 | ira_traverse_loop_tree (false, ira_loop_tree_root, color_pass, NULL); |
3515 | ||
3516 | if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL) | |
3517 | ira_print_disposition (ira_dump_file); | |
3518 | ||
058e97ec | 3519 | ira_free_bitmap (coloring_allocno_bitmap); |
058e97ec VM |
3520 | } |
3521 | ||
3522 | \f | |
3523 | ||
3524 | /* Move spill/restore code, which are to be generated in ira-emit.c, | |
3525 | to less frequent points (if it is profitable) by reassigning some | |
3526 | allocnos (in loop with subloops containing in another loop) to | |
3527 | memory which results in longer live-range where the corresponding | |
3528 | pseudo-registers will be in memory. */ | |
3529 | static void | |
3530 | move_spill_restore (void) | |
3531 | { | |
3532 | int cost, regno, hard_regno, hard_regno2, index; | |
3533 | bool changed_p; | |
3534 | int enter_freq, exit_freq; | |
ef4bddc2 | 3535 | machine_mode mode; |
058e97ec VM |
3536 | enum reg_class rclass; |
3537 | ira_allocno_t a, parent_allocno, subloop_allocno; | |
3538 | ira_loop_tree_node_t parent, loop_node, subloop_node; | |
3539 | ira_allocno_iterator ai; | |
3540 | ||
3541 | for (;;) | |
3542 | { | |
3543 | changed_p = false; | |
3544 | if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL) | |
3545 | fprintf (ira_dump_file, "New iteration of spill/restore move\n"); | |
3546 | FOR_EACH_ALLOCNO (a, ai) | |
3547 | { | |
3548 | regno = ALLOCNO_REGNO (a); | |
3549 | loop_node = ALLOCNO_LOOP_TREE_NODE (a); | |
3550 | if (ALLOCNO_CAP_MEMBER (a) != NULL | |
3551 | || ALLOCNO_CAP (a) != NULL | |
3552 | || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0 | |
3553 | || loop_node->children == NULL | |
3554 | /* don't do the optimization because it can create | |
3555 | copies and the reload pass can spill the allocno set | |
3556 | by copy although the allocno will not get memory | |
3557 | slot. */ | |
55a2c322 | 3558 | || ira_equiv_no_lvalue_p (regno) |
b81a2f0d VM |
3559 | || !bitmap_bit_p (loop_node->border_allocnos, ALLOCNO_NUM (a)) |
3560 | /* Do not spill static chain pointer pseudo when | |
3561 | non-local goto is used. */ | |
3562 | || non_spilled_static_chain_regno_p (regno)) | |
058e97ec VM |
3563 | continue; |
3564 | mode = ALLOCNO_MODE (a); | |
1756cb66 | 3565 | rclass = ALLOCNO_CLASS (a); |
058e97ec VM |
3566 | index = ira_class_hard_reg_index[rclass][hard_regno]; |
3567 | ira_assert (index >= 0); | |
3568 | cost = (ALLOCNO_MEMORY_COST (a) | |
3569 | - (ALLOCNO_HARD_REG_COSTS (a) == NULL | |
1756cb66 | 3570 | ? ALLOCNO_CLASS_COST (a) |
058e97ec | 3571 | : ALLOCNO_HARD_REG_COSTS (a)[index])); |
1756cb66 | 3572 | ira_init_register_move_cost_if_necessary (mode); |
058e97ec VM |
3573 | for (subloop_node = loop_node->subloops; |
3574 | subloop_node != NULL; | |
3575 | subloop_node = subloop_node->subloop_next) | |
3576 | { | |
3577 | ira_assert (subloop_node->bb == NULL); | |
3578 | subloop_allocno = subloop_node->regno_allocno_map[regno]; | |
3579 | if (subloop_allocno == NULL) | |
3580 | continue; | |
1756cb66 | 3581 | ira_assert (rclass == ALLOCNO_CLASS (subloop_allocno)); |
058e97ec VM |
3582 | /* We have accumulated cost. To get the real cost of |
3583 | allocno usage in the loop we should subtract costs of | |
3584 | the subloop allocnos. */ | |
3585 | cost -= (ALLOCNO_MEMORY_COST (subloop_allocno) | |
3586 | - (ALLOCNO_HARD_REG_COSTS (subloop_allocno) == NULL | |
1756cb66 | 3587 | ? ALLOCNO_CLASS_COST (subloop_allocno) |
058e97ec VM |
3588 | : ALLOCNO_HARD_REG_COSTS (subloop_allocno)[index])); |
3589 | exit_freq = ira_loop_edge_freq (subloop_node, regno, true); | |
3590 | enter_freq = ira_loop_edge_freq (subloop_node, regno, false); | |
3591 | if ((hard_regno2 = ALLOCNO_HARD_REGNO (subloop_allocno)) < 0) | |
3592 | cost -= (ira_memory_move_cost[mode][rclass][0] * exit_freq | |
3593 | + ira_memory_move_cost[mode][rclass][1] * enter_freq); | |
3594 | else | |
3595 | { | |
3596 | cost | |
3597 | += (ira_memory_move_cost[mode][rclass][0] * exit_freq | |
3598 | + ira_memory_move_cost[mode][rclass][1] * enter_freq); | |
3599 | if (hard_regno2 != hard_regno) | |
1756cb66 | 3600 | cost -= (ira_register_move_cost[mode][rclass][rclass] |
058e97ec VM |
3601 | * (exit_freq + enter_freq)); |
3602 | } | |
3603 | } | |
3604 | if ((parent = loop_node->parent) != NULL | |
3605 | && (parent_allocno = parent->regno_allocno_map[regno]) != NULL) | |
3606 | { | |
1756cb66 | 3607 | ira_assert (rclass == ALLOCNO_CLASS (parent_allocno)); |
058e97ec VM |
3608 | exit_freq = ira_loop_edge_freq (loop_node, regno, true); |
3609 | enter_freq = ira_loop_edge_freq (loop_node, regno, false); | |
3610 | if ((hard_regno2 = ALLOCNO_HARD_REGNO (parent_allocno)) < 0) | |
3611 | cost -= (ira_memory_move_cost[mode][rclass][0] * exit_freq | |
3612 | + ira_memory_move_cost[mode][rclass][1] * enter_freq); | |
3613 | else | |
3614 | { | |
3615 | cost | |
3616 | += (ira_memory_move_cost[mode][rclass][1] * exit_freq | |
3617 | + ira_memory_move_cost[mode][rclass][0] * enter_freq); | |
3618 | if (hard_regno2 != hard_regno) | |
1756cb66 | 3619 | cost -= (ira_register_move_cost[mode][rclass][rclass] |
058e97ec VM |
3620 | * (exit_freq + enter_freq)); |
3621 | } | |
3622 | } | |
3623 | if (cost < 0) | |
3624 | { | |
3625 | ALLOCNO_HARD_REGNO (a) = -1; | |
3626 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3627 | { | |
3628 | fprintf | |
3629 | (ira_dump_file, | |
3630 | " Moving spill/restore for a%dr%d up from loop %d", | |
2608d841 | 3631 | ALLOCNO_NUM (a), regno, loop_node->loop_num); |
058e97ec VM |
3632 | fprintf (ira_dump_file, " - profit %d\n", -cost); |
3633 | } | |
3634 | changed_p = true; | |
3635 | } | |
3636 | } | |
3637 | if (! changed_p) | |
3638 | break; | |
3639 | } | |
3640 | } | |
3641 | ||
3642 | \f | |
3643 | ||
3644 | /* Update current hard reg costs and current conflict hard reg costs | |
3645 | for allocno A. It is done by processing its copies containing | |
3646 | other allocnos already assigned. */ | |
3647 | static void | |
3648 | update_curr_costs (ira_allocno_t a) | |
3649 | { | |
3650 | int i, hard_regno, cost; | |
ef4bddc2 | 3651 | machine_mode mode; |
1756cb66 | 3652 | enum reg_class aclass, rclass; |
058e97ec VM |
3653 | ira_allocno_t another_a; |
3654 | ira_copy_t cp, next_cp; | |
3655 | ||
bdf0eb06 | 3656 | ira_free_allocno_updated_costs (a); |
058e97ec | 3657 | ira_assert (! ALLOCNO_ASSIGNED_P (a)); |
1756cb66 VM |
3658 | aclass = ALLOCNO_CLASS (a); |
3659 | if (aclass == NO_REGS) | |
058e97ec VM |
3660 | return; |
3661 | mode = ALLOCNO_MODE (a); | |
1756cb66 | 3662 | ira_init_register_move_cost_if_necessary (mode); |
058e97ec VM |
3663 | for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp) |
3664 | { | |
3665 | if (cp->first == a) | |
3666 | { | |
3667 | next_cp = cp->next_first_allocno_copy; | |
3668 | another_a = cp->second; | |
3669 | } | |
3670 | else if (cp->second == a) | |
3671 | { | |
3672 | next_cp = cp->next_second_allocno_copy; | |
3673 | another_a = cp->first; | |
3674 | } | |
3675 | else | |
3676 | gcc_unreachable (); | |
1756cb66 | 3677 | if (! ira_reg_classes_intersect_p[aclass][ALLOCNO_CLASS (another_a)] |
058e97ec VM |
3678 | || ! ALLOCNO_ASSIGNED_P (another_a) |
3679 | || (hard_regno = ALLOCNO_HARD_REGNO (another_a)) < 0) | |
3680 | continue; | |
3681 | rclass = REGNO_REG_CLASS (hard_regno); | |
1756cb66 | 3682 | i = ira_class_hard_reg_index[aclass][hard_regno]; |
7db7ed3c VM |
3683 | if (i < 0) |
3684 | continue; | |
058e97ec | 3685 | cost = (cp->first == a |
1756cb66 VM |
3686 | ? ira_register_move_cost[mode][rclass][aclass] |
3687 | : ira_register_move_cost[mode][aclass][rclass]); | |
058e97ec | 3688 | ira_allocate_and_set_or_copy_costs |
1756cb66 | 3689 | (&ALLOCNO_UPDATED_HARD_REG_COSTS (a), aclass, ALLOCNO_CLASS_COST (a), |
058e97ec VM |
3690 | ALLOCNO_HARD_REG_COSTS (a)); |
3691 | ira_allocate_and_set_or_copy_costs | |
3692 | (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a), | |
1756cb66 | 3693 | aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (a)); |
058e97ec VM |
3694 | ALLOCNO_UPDATED_HARD_REG_COSTS (a)[i] -= cp->freq * cost; |
3695 | ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a)[i] -= cp->freq * cost; | |
3696 | } | |
3697 | } | |
3698 | ||
058e97ec VM |
3699 | /* Try to assign hard registers to the unassigned allocnos and |
3700 | allocnos conflicting with them or conflicting with allocnos whose | |
3701 | regno >= START_REGNO. The function is called after ira_flattening, | |
3702 | so more allocnos (including ones created in ira-emit.c) will have a | |
3703 | chance to get a hard register. We use simple assignment algorithm | |
3704 | based on priorities. */ | |
3705 | void | |
3706 | ira_reassign_conflict_allocnos (int start_regno) | |
3707 | { | |
3708 | int i, allocnos_to_color_num; | |
fa86d337 | 3709 | ira_allocno_t a; |
1756cb66 | 3710 | enum reg_class aclass; |
058e97ec VM |
3711 | bitmap allocnos_to_color; |
3712 | ira_allocno_iterator ai; | |
3713 | ||
3714 | allocnos_to_color = ira_allocate_bitmap (); | |
3715 | allocnos_to_color_num = 0; | |
3716 | FOR_EACH_ALLOCNO (a, ai) | |
3717 | { | |
ac0ab4f7 | 3718 | int n = ALLOCNO_NUM_OBJECTS (a); |
fa86d337 | 3719 | |
058e97ec VM |
3720 | if (! ALLOCNO_ASSIGNED_P (a) |
3721 | && ! bitmap_bit_p (allocnos_to_color, ALLOCNO_NUM (a))) | |
3722 | { | |
1756cb66 | 3723 | if (ALLOCNO_CLASS (a) != NO_REGS) |
058e97ec VM |
3724 | sorted_allocnos[allocnos_to_color_num++] = a; |
3725 | else | |
3726 | { | |
3727 | ALLOCNO_ASSIGNED_P (a) = true; | |
3728 | ALLOCNO_HARD_REGNO (a) = -1; | |
3729 | ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL); | |
3730 | ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a) == NULL); | |
3731 | } | |
3732 | bitmap_set_bit (allocnos_to_color, ALLOCNO_NUM (a)); | |
3733 | } | |
3734 | if (ALLOCNO_REGNO (a) < start_regno | |
1756cb66 | 3735 | || (aclass = ALLOCNO_CLASS (a)) == NO_REGS) |
058e97ec | 3736 | continue; |
ac0ab4f7 | 3737 | for (i = 0; i < n; i++) |
058e97ec | 3738 | { |
ac0ab4f7 BS |
3739 | ira_object_t obj = ALLOCNO_OBJECT (a, i); |
3740 | ira_object_t conflict_obj; | |
3741 | ira_object_conflict_iterator oci; | |
1756cb66 | 3742 | |
ac0ab4f7 BS |
3743 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) |
3744 | { | |
3745 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); | |
1756cb66 | 3746 | |
ac0ab4f7 | 3747 | ira_assert (ira_reg_classes_intersect_p |
1756cb66 | 3748 | [aclass][ALLOCNO_CLASS (conflict_a)]); |
fcaa4ca4 | 3749 | if (!bitmap_set_bit (allocnos_to_color, ALLOCNO_NUM (conflict_a))) |
ac0ab4f7 | 3750 | continue; |
ac0ab4f7 BS |
3751 | sorted_allocnos[allocnos_to_color_num++] = conflict_a; |
3752 | } | |
058e97ec VM |
3753 | } |
3754 | } | |
3755 | ira_free_bitmap (allocnos_to_color); | |
3756 | if (allocnos_to_color_num > 1) | |
3757 | { | |
1ae64b0f | 3758 | setup_allocno_priorities (sorted_allocnos, allocnos_to_color_num); |
058e97ec VM |
3759 | qsort (sorted_allocnos, allocnos_to_color_num, sizeof (ira_allocno_t), |
3760 | allocno_priority_compare_func); | |
3761 | } | |
3762 | for (i = 0; i < allocnos_to_color_num; i++) | |
3763 | { | |
3764 | a = sorted_allocnos[i]; | |
3765 | ALLOCNO_ASSIGNED_P (a) = false; | |
058e97ec VM |
3766 | update_curr_costs (a); |
3767 | } | |
3768 | for (i = 0; i < allocnos_to_color_num; i++) | |
3769 | { | |
3770 | a = sorted_allocnos[i]; | |
3771 | if (assign_hard_reg (a, true)) | |
3772 | { | |
3773 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3774 | fprintf | |
3775 | (ira_dump_file, | |
3776 | " Secondary allocation: assign hard reg %d to reg %d\n", | |
3777 | ALLOCNO_HARD_REGNO (a), ALLOCNO_REGNO (a)); | |
3778 | } | |
3779 | } | |
3780 | } | |
3781 | ||
3782 | \f | |
3783 | ||
1756cb66 VM |
3784 | /* This page contains functions used to find conflicts using allocno |
3785 | live ranges. */ | |
3786 | ||
1756cb66 VM |
3787 | #ifdef ENABLE_IRA_CHECKING |
3788 | ||
3789 | /* Return TRUE if live ranges of pseudo-registers REGNO1 and REGNO2 | |
3790 | intersect. This should be used when there is only one region. | |
3791 | Currently this is used during reload. */ | |
3792 | static bool | |
3793 | conflict_by_live_ranges_p (int regno1, int regno2) | |
3794 | { | |
3795 | ira_allocno_t a1, a2; | |
3796 | ||
3797 | ira_assert (regno1 >= FIRST_PSEUDO_REGISTER | |
3798 | && regno2 >= FIRST_PSEUDO_REGISTER); | |
df3e3493 | 3799 | /* Reg info calculated by dataflow infrastructure can be different |
1756cb66 VM |
3800 | from one calculated by regclass. */ |
3801 | if ((a1 = ira_loop_tree_root->regno_allocno_map[regno1]) == NULL | |
3802 | || (a2 = ira_loop_tree_root->regno_allocno_map[regno2]) == NULL) | |
3803 | return false; | |
3804 | return allocnos_conflict_by_live_ranges_p (a1, a2); | |
3805 | } | |
3806 | ||
3807 | #endif | |
3808 | ||
3809 | \f | |
3810 | ||
058e97ec VM |
3811 | /* This page contains code to coalesce memory stack slots used by |
3812 | spilled allocnos. This results in smaller stack frame, better data | |
3813 | locality, and in smaller code for some architectures like | |
3814 | x86/x86_64 where insn size depends on address displacement value. | |
3815 | On the other hand, it can worsen insn scheduling after the RA but | |
3816 | in practice it is less important than smaller stack frames. */ | |
3817 | ||
22b0982c VM |
3818 | /* TRUE if we coalesced some allocnos. In other words, if we got |
3819 | loops formed by members first_coalesced_allocno and | |
3820 | next_coalesced_allocno containing more one allocno. */ | |
3821 | static bool allocno_coalesced_p; | |
3822 | ||
3823 | /* Bitmap used to prevent a repeated allocno processing because of | |
3824 | coalescing. */ | |
3825 | static bitmap processed_coalesced_allocno_bitmap; | |
3826 | ||
1756cb66 VM |
3827 | /* See below. */ |
3828 | typedef struct coalesce_data *coalesce_data_t; | |
3829 | ||
3830 | /* To decrease footprint of ira_allocno structure we store all data | |
3831 | needed only for coalescing in the following structure. */ | |
3832 | struct coalesce_data | |
3833 | { | |
3834 | /* Coalesced allocnos form a cyclic list. One allocno given by | |
3835 | FIRST represents all coalesced allocnos. The | |
3836 | list is chained by NEXT. */ | |
3837 | ira_allocno_t first; | |
3838 | ira_allocno_t next; | |
3839 | int temp; | |
3840 | }; | |
3841 | ||
3842 | /* Container for storing allocno data concerning coalescing. */ | |
3843 | static coalesce_data_t allocno_coalesce_data; | |
3844 | ||
3845 | /* Macro to access the data concerning coalescing. */ | |
3846 | #define ALLOCNO_COALESCE_DATA(a) ((coalesce_data_t) ALLOCNO_ADD_DATA (a)) | |
3847 | ||
22b0982c VM |
3848 | /* Merge two sets of coalesced allocnos given correspondingly by |
3849 | allocnos A1 and A2 (more accurately merging A2 set into A1 | |
3850 | set). */ | |
3851 | static void | |
3852 | merge_allocnos (ira_allocno_t a1, ira_allocno_t a2) | |
3853 | { | |
3854 | ira_allocno_t a, first, last, next; | |
3855 | ||
1756cb66 VM |
3856 | first = ALLOCNO_COALESCE_DATA (a1)->first; |
3857 | a = ALLOCNO_COALESCE_DATA (a2)->first; | |
3858 | if (first == a) | |
22b0982c | 3859 | return; |
1756cb66 VM |
3860 | for (last = a2, a = ALLOCNO_COALESCE_DATA (a2)->next;; |
3861 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
22b0982c | 3862 | { |
1756cb66 | 3863 | ALLOCNO_COALESCE_DATA (a)->first = first; |
22b0982c VM |
3864 | if (a == a2) |
3865 | break; | |
3866 | last = a; | |
3867 | } | |
1756cb66 VM |
3868 | next = allocno_coalesce_data[ALLOCNO_NUM (first)].next; |
3869 | allocno_coalesce_data[ALLOCNO_NUM (first)].next = a2; | |
3870 | allocno_coalesce_data[ALLOCNO_NUM (last)].next = next; | |
22b0982c VM |
3871 | } |
3872 | ||
1756cb66 VM |
3873 | /* Return TRUE if there are conflicting allocnos from two sets of |
3874 | coalesced allocnos given correspondingly by allocnos A1 and A2. We | |
3875 | use live ranges to find conflicts because conflicts are represented | |
3876 | only for allocnos of the same allocno class and during the reload | |
3877 | pass we coalesce allocnos for sharing stack memory slots. */ | |
22b0982c VM |
3878 | static bool |
3879 | coalesced_allocno_conflict_p (ira_allocno_t a1, ira_allocno_t a2) | |
3880 | { | |
1756cb66 | 3881 | ira_allocno_t a, conflict_a; |
22b0982c | 3882 | |
22b0982c VM |
3883 | if (allocno_coalesced_p) |
3884 | { | |
1756cb66 VM |
3885 | bitmap_clear (processed_coalesced_allocno_bitmap); |
3886 | for (a = ALLOCNO_COALESCE_DATA (a1)->next;; | |
3887 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
22b0982c | 3888 | { |
1756cb66 | 3889 | bitmap_set_bit (processed_coalesced_allocno_bitmap, ALLOCNO_NUM (a)); |
22b0982c VM |
3890 | if (a == a1) |
3891 | break; | |
3892 | } | |
3893 | } | |
1756cb66 VM |
3894 | for (a = ALLOCNO_COALESCE_DATA (a2)->next;; |
3895 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
22b0982c | 3896 | { |
1756cb66 VM |
3897 | for (conflict_a = ALLOCNO_COALESCE_DATA (a1)->next;; |
3898 | conflict_a = ALLOCNO_COALESCE_DATA (conflict_a)->next) | |
22b0982c | 3899 | { |
1756cb66 | 3900 | if (allocnos_conflict_by_live_ranges_p (a, conflict_a)) |
22b0982c | 3901 | return true; |
1756cb66 | 3902 | if (conflict_a == a1) |
22b0982c VM |
3903 | break; |
3904 | } | |
22b0982c VM |
3905 | if (a == a2) |
3906 | break; | |
3907 | } | |
3908 | return false; | |
3909 | } | |
3910 | ||
3911 | /* The major function for aggressive allocno coalescing. We coalesce | |
3912 | only spilled allocnos. If some allocnos have been coalesced, we | |
3913 | set up flag allocno_coalesced_p. */ | |
3914 | static void | |
3915 | coalesce_allocnos (void) | |
3916 | { | |
3917 | ira_allocno_t a; | |
bf08fb16 | 3918 | ira_copy_t cp, next_cp; |
22b0982c VM |
3919 | unsigned int j; |
3920 | int i, n, cp_num, regno; | |
3921 | bitmap_iterator bi; | |
3922 | ||
22b0982c VM |
3923 | cp_num = 0; |
3924 | /* Collect copies. */ | |
3925 | EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, j, bi) | |
3926 | { | |
3927 | a = ira_allocnos[j]; | |
3928 | regno = ALLOCNO_REGNO (a); | |
3929 | if (! ALLOCNO_ASSIGNED_P (a) || ALLOCNO_HARD_REGNO (a) >= 0 | |
55a2c322 | 3930 | || ira_equiv_no_lvalue_p (regno)) |
22b0982c VM |
3931 | continue; |
3932 | for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp) | |
3933 | { | |
3934 | if (cp->first == a) | |
3935 | { | |
3936 | next_cp = cp->next_first_allocno_copy; | |
3937 | regno = ALLOCNO_REGNO (cp->second); | |
3938 | /* For priority coloring we coalesce allocnos only with | |
1756cb66 | 3939 | the same allocno class not with intersected allocno |
22b0982c VM |
3940 | classes as it were possible. It is done for |
3941 | simplicity. */ | |
3942 | if ((cp->insn != NULL || cp->constraint_p) | |
3943 | && ALLOCNO_ASSIGNED_P (cp->second) | |
3944 | && ALLOCNO_HARD_REGNO (cp->second) < 0 | |
55a2c322 | 3945 | && ! ira_equiv_no_lvalue_p (regno)) |
22b0982c VM |
3946 | sorted_copies[cp_num++] = cp; |
3947 | } | |
3948 | else if (cp->second == a) | |
3949 | next_cp = cp->next_second_allocno_copy; | |
3950 | else | |
3951 | gcc_unreachable (); | |
3952 | } | |
3953 | } | |
3954 | qsort (sorted_copies, cp_num, sizeof (ira_copy_t), copy_freq_compare_func); | |
3955 | /* Coalesced copies, most frequently executed first. */ | |
3956 | for (; cp_num != 0;) | |
3957 | { | |
3958 | for (i = 0; i < cp_num; i++) | |
3959 | { | |
3960 | cp = sorted_copies[i]; | |
3961 | if (! coalesced_allocno_conflict_p (cp->first, cp->second)) | |
3962 | { | |
3963 | allocno_coalesced_p = true; | |
3964 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
3965 | fprintf | |
3966 | (ira_dump_file, | |
3967 | " Coalescing copy %d:a%dr%d-a%dr%d (freq=%d)\n", | |
3968 | cp->num, ALLOCNO_NUM (cp->first), ALLOCNO_REGNO (cp->first), | |
3969 | ALLOCNO_NUM (cp->second), ALLOCNO_REGNO (cp->second), | |
3970 | cp->freq); | |
3971 | merge_allocnos (cp->first, cp->second); | |
3972 | i++; | |
3973 | break; | |
3974 | } | |
3975 | } | |
3976 | /* Collect the rest of copies. */ | |
3977 | for (n = 0; i < cp_num; i++) | |
3978 | { | |
3979 | cp = sorted_copies[i]; | |
1756cb66 VM |
3980 | if (allocno_coalesce_data[ALLOCNO_NUM (cp->first)].first |
3981 | != allocno_coalesce_data[ALLOCNO_NUM (cp->second)].first) | |
22b0982c VM |
3982 | sorted_copies[n++] = cp; |
3983 | } | |
3984 | cp_num = n; | |
3985 | } | |
22b0982c VM |
3986 | } |
3987 | ||
058e97ec VM |
3988 | /* Usage cost and order number of coalesced allocno set to which |
3989 | given pseudo register belongs to. */ | |
3990 | static int *regno_coalesced_allocno_cost; | |
3991 | static int *regno_coalesced_allocno_num; | |
3992 | ||
3993 | /* Sort pseudos according frequencies of coalesced allocno sets they | |
3994 | belong to (putting most frequently ones first), and according to | |
3995 | coalesced allocno set order numbers. */ | |
3996 | static int | |
3997 | coalesced_pseudo_reg_freq_compare (const void *v1p, const void *v2p) | |
3998 | { | |
3999 | const int regno1 = *(const int *) v1p; | |
4000 | const int regno2 = *(const int *) v2p; | |
4001 | int diff; | |
4002 | ||
4003 | if ((diff = (regno_coalesced_allocno_cost[regno2] | |
4004 | - regno_coalesced_allocno_cost[regno1])) != 0) | |
4005 | return diff; | |
4006 | if ((diff = (regno_coalesced_allocno_num[regno1] | |
4007 | - regno_coalesced_allocno_num[regno2])) != 0) | |
4008 | return diff; | |
4009 | return regno1 - regno2; | |
4010 | } | |
4011 | ||
4012 | /* Widest width in which each pseudo reg is referred to (via subreg). | |
4013 | It is used for sorting pseudo registers. */ | |
bd5a2c67 | 4014 | static machine_mode *regno_max_ref_mode; |
058e97ec | 4015 | |
058e97ec VM |
4016 | /* Sort pseudos according their slot numbers (putting ones with |
4017 | smaller numbers first, or last when the frame pointer is not | |
4018 | needed). */ | |
4019 | static int | |
4020 | coalesced_pseudo_reg_slot_compare (const void *v1p, const void *v2p) | |
4021 | { | |
4022 | const int regno1 = *(const int *) v1p; | |
4023 | const int regno2 = *(const int *) v2p; | |
4024 | ira_allocno_t a1 = ira_regno_allocno_map[regno1]; | |
4025 | ira_allocno_t a2 = ira_regno_allocno_map[regno2]; | |
4026 | int diff, slot_num1, slot_num2; | |
bd5a2c67 | 4027 | machine_mode mode1, mode2; |
058e97ec VM |
4028 | |
4029 | if (a1 == NULL || ALLOCNO_HARD_REGNO (a1) >= 0) | |
4030 | { | |
4031 | if (a2 == NULL || ALLOCNO_HARD_REGNO (a2) >= 0) | |
004a6ce8 | 4032 | return regno1 - regno2; |
058e97ec VM |
4033 | return 1; |
4034 | } | |
4035 | else if (a2 == NULL || ALLOCNO_HARD_REGNO (a2) >= 0) | |
4036 | return -1; | |
4037 | slot_num1 = -ALLOCNO_HARD_REGNO (a1); | |
4038 | slot_num2 = -ALLOCNO_HARD_REGNO (a2); | |
4039 | if ((diff = slot_num1 - slot_num2) != 0) | |
4040 | return (frame_pointer_needed | |
e0bf0dc2 | 4041 | || (!FRAME_GROWS_DOWNWARD) == STACK_GROWS_DOWNWARD ? diff : -diff); |
bd5a2c67 RS |
4042 | mode1 = wider_subreg_mode (PSEUDO_REGNO_MODE (regno1), |
4043 | regno_max_ref_mode[regno1]); | |
4044 | mode2 = wider_subreg_mode (PSEUDO_REGNO_MODE (regno2), | |
4045 | regno_max_ref_mode[regno2]); | |
cf098191 RS |
4046 | if ((diff = compare_sizes_for_sort (GET_MODE_SIZE (mode2), |
4047 | GET_MODE_SIZE (mode1))) != 0) | |
058e97ec | 4048 | return diff; |
004a6ce8 | 4049 | return regno1 - regno2; |
058e97ec VM |
4050 | } |
4051 | ||
4052 | /* Setup REGNO_COALESCED_ALLOCNO_COST and REGNO_COALESCED_ALLOCNO_NUM | |
4053 | for coalesced allocno sets containing allocnos with their regnos | |
4054 | given in array PSEUDO_REGNOS of length N. */ | |
4055 | static void | |
4056 | setup_coalesced_allocno_costs_and_nums (int *pseudo_regnos, int n) | |
4057 | { | |
4058 | int i, num, regno, cost; | |
4059 | ira_allocno_t allocno, a; | |
4060 | ||
4061 | for (num = i = 0; i < n; i++) | |
4062 | { | |
4063 | regno = pseudo_regnos[i]; | |
4064 | allocno = ira_regno_allocno_map[regno]; | |
4065 | if (allocno == NULL) | |
4066 | { | |
4067 | regno_coalesced_allocno_cost[regno] = 0; | |
4068 | regno_coalesced_allocno_num[regno] = ++num; | |
4069 | continue; | |
4070 | } | |
1756cb66 | 4071 | if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno) |
058e97ec VM |
4072 | continue; |
4073 | num++; | |
1756cb66 VM |
4074 | for (cost = 0, a = ALLOCNO_COALESCE_DATA (allocno)->next;; |
4075 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
058e97ec VM |
4076 | { |
4077 | cost += ALLOCNO_FREQ (a); | |
4078 | if (a == allocno) | |
4079 | break; | |
4080 | } | |
1756cb66 VM |
4081 | for (a = ALLOCNO_COALESCE_DATA (allocno)->next;; |
4082 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
058e97ec VM |
4083 | { |
4084 | regno_coalesced_allocno_num[ALLOCNO_REGNO (a)] = num; | |
4085 | regno_coalesced_allocno_cost[ALLOCNO_REGNO (a)] = cost; | |
4086 | if (a == allocno) | |
4087 | break; | |
4088 | } | |
4089 | } | |
4090 | } | |
4091 | ||
4092 | /* Collect spilled allocnos representing coalesced allocno sets (the | |
4093 | first coalesced allocno). The collected allocnos are returned | |
4094 | through array SPILLED_COALESCED_ALLOCNOS. The function returns the | |
4095 | number of the collected allocnos. The allocnos are given by their | |
4096 | regnos in array PSEUDO_REGNOS of length N. */ | |
4097 | static int | |
4098 | collect_spilled_coalesced_allocnos (int *pseudo_regnos, int n, | |
4099 | ira_allocno_t *spilled_coalesced_allocnos) | |
4100 | { | |
4101 | int i, num, regno; | |
4102 | ira_allocno_t allocno; | |
4103 | ||
4104 | for (num = i = 0; i < n; i++) | |
4105 | { | |
4106 | regno = pseudo_regnos[i]; | |
4107 | allocno = ira_regno_allocno_map[regno]; | |
4108 | if (allocno == NULL || ALLOCNO_HARD_REGNO (allocno) >= 0 | |
1756cb66 | 4109 | || ALLOCNO_COALESCE_DATA (allocno)->first != allocno) |
058e97ec VM |
4110 | continue; |
4111 | spilled_coalesced_allocnos[num++] = allocno; | |
4112 | } | |
4113 | return num; | |
4114 | } | |
4115 | ||
3553f0bb VM |
4116 | /* Array of live ranges of size IRA_ALLOCNOS_NUM. Live range for |
4117 | given slot contains live ranges of coalesced allocnos assigned to | |
4118 | given slot. */ | |
b14151b5 | 4119 | static live_range_t *slot_coalesced_allocnos_live_ranges; |
b15a7ae6 | 4120 | |
3553f0bb VM |
4121 | /* Return TRUE if coalesced allocnos represented by ALLOCNO has live |
4122 | ranges intersected with live ranges of coalesced allocnos assigned | |
4123 | to slot with number N. */ | |
b15a7ae6 | 4124 | static bool |
3553f0bb | 4125 | slot_coalesced_allocno_live_ranges_intersect_p (ira_allocno_t allocno, int n) |
b15a7ae6 | 4126 | { |
b15a7ae6 | 4127 | ira_allocno_t a; |
b15a7ae6 | 4128 | |
1756cb66 VM |
4129 | for (a = ALLOCNO_COALESCE_DATA (allocno)->next;; |
4130 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
b15a7ae6 | 4131 | { |
ac0ab4f7 BS |
4132 | int i; |
4133 | int nr = ALLOCNO_NUM_OBJECTS (a); | |
0550a77b | 4134 | gcc_assert (ALLOCNO_CAP_MEMBER (a) == NULL); |
ac0ab4f7 BS |
4135 | for (i = 0; i < nr; i++) |
4136 | { | |
4137 | ira_object_t obj = ALLOCNO_OBJECT (a, i); | |
1756cb66 VM |
4138 | |
4139 | if (ira_live_ranges_intersect_p | |
4140 | (slot_coalesced_allocnos_live_ranges[n], | |
4141 | OBJECT_LIVE_RANGES (obj))) | |
ac0ab4f7 BS |
4142 | return true; |
4143 | } | |
b15a7ae6 VM |
4144 | if (a == allocno) |
4145 | break; | |
4146 | } | |
4147 | return false; | |
4148 | } | |
4149 | ||
3553f0bb VM |
4150 | /* Update live ranges of slot to which coalesced allocnos represented |
4151 | by ALLOCNO were assigned. */ | |
b15a7ae6 | 4152 | static void |
3553f0bb | 4153 | setup_slot_coalesced_allocno_live_ranges (ira_allocno_t allocno) |
b15a7ae6 | 4154 | { |
ac0ab4f7 | 4155 | int i, n; |
b15a7ae6 | 4156 | ira_allocno_t a; |
b14151b5 | 4157 | live_range_t r; |
b15a7ae6 | 4158 | |
1756cb66 VM |
4159 | n = ALLOCNO_COALESCE_DATA (allocno)->temp; |
4160 | for (a = ALLOCNO_COALESCE_DATA (allocno)->next;; | |
4161 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
b15a7ae6 | 4162 | { |
ac0ab4f7 | 4163 | int nr = ALLOCNO_NUM_OBJECTS (a); |
0550a77b | 4164 | gcc_assert (ALLOCNO_CAP_MEMBER (a) == NULL); |
ac0ab4f7 BS |
4165 | for (i = 0; i < nr; i++) |
4166 | { | |
4167 | ira_object_t obj = ALLOCNO_OBJECT (a, i); | |
1756cb66 | 4168 | |
ac0ab4f7 BS |
4169 | r = ira_copy_live_range_list (OBJECT_LIVE_RANGES (obj)); |
4170 | slot_coalesced_allocnos_live_ranges[n] | |
4171 | = ira_merge_live_ranges | |
1756cb66 | 4172 | (slot_coalesced_allocnos_live_ranges[n], r); |
ac0ab4f7 | 4173 | } |
b15a7ae6 VM |
4174 | if (a == allocno) |
4175 | break; | |
4176 | } | |
4177 | } | |
4178 | ||
058e97ec VM |
4179 | /* We have coalesced allocnos involving in copies. Coalesce allocnos |
4180 | further in order to share the same memory stack slot. Allocnos | |
4181 | representing sets of allocnos coalesced before the call are given | |
4182 | in array SPILLED_COALESCED_ALLOCNOS of length NUM. Return TRUE if | |
4183 | some allocnos were coalesced in the function. */ | |
4184 | static bool | |
4185 | coalesce_spill_slots (ira_allocno_t *spilled_coalesced_allocnos, int num) | |
4186 | { | |
3553f0bb | 4187 | int i, j, n, last_coalesced_allocno_num; |
058e97ec VM |
4188 | ira_allocno_t allocno, a; |
4189 | bool merged_p = false; | |
1240d76e | 4190 | bitmap set_jump_crosses = regstat_get_setjmp_crosses (); |
058e97ec | 4191 | |
3553f0bb | 4192 | slot_coalesced_allocnos_live_ranges |
b14151b5 | 4193 | = (live_range_t *) ira_allocate (sizeof (live_range_t) * ira_allocnos_num); |
3553f0bb | 4194 | memset (slot_coalesced_allocnos_live_ranges, 0, |
b14151b5 | 4195 | sizeof (live_range_t) * ira_allocnos_num); |
b15a7ae6 | 4196 | last_coalesced_allocno_num = 0; |
058e97ec VM |
4197 | /* Coalesce non-conflicting spilled allocnos preferring most |
4198 | frequently used. */ | |
4199 | for (i = 0; i < num; i++) | |
4200 | { | |
4201 | allocno = spilled_coalesced_allocnos[i]; | |
1756cb66 | 4202 | if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno |
1240d76e | 4203 | || bitmap_bit_p (set_jump_crosses, ALLOCNO_REGNO (allocno)) |
55a2c322 | 4204 | || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno))) |
058e97ec VM |
4205 | continue; |
4206 | for (j = 0; j < i; j++) | |
4207 | { | |
4208 | a = spilled_coalesced_allocnos[j]; | |
1756cb66 VM |
4209 | n = ALLOCNO_COALESCE_DATA (a)->temp; |
4210 | if (ALLOCNO_COALESCE_DATA (a)->first == a | |
1240d76e | 4211 | && ! bitmap_bit_p (set_jump_crosses, ALLOCNO_REGNO (a)) |
55a2c322 | 4212 | && ! ira_equiv_no_lvalue_p (ALLOCNO_REGNO (a)) |
3553f0bb | 4213 | && ! slot_coalesced_allocno_live_ranges_intersect_p (allocno, n)) |
b15a7ae6 VM |
4214 | break; |
4215 | } | |
4216 | if (j >= i) | |
4217 | { | |
4218 | /* No coalescing: set up number for coalesced allocnos | |
4219 | represented by ALLOCNO. */ | |
1756cb66 | 4220 | ALLOCNO_COALESCE_DATA (allocno)->temp = last_coalesced_allocno_num++; |
3553f0bb | 4221 | setup_slot_coalesced_allocno_live_ranges (allocno); |
b15a7ae6 VM |
4222 | } |
4223 | else | |
4224 | { | |
058e97ec VM |
4225 | allocno_coalesced_p = true; |
4226 | merged_p = true; | |
4227 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
4228 | fprintf (ira_dump_file, | |
4229 | " Coalescing spilled allocnos a%dr%d->a%dr%d\n", | |
4230 | ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno), | |
4231 | ALLOCNO_NUM (a), ALLOCNO_REGNO (a)); | |
1756cb66 VM |
4232 | ALLOCNO_COALESCE_DATA (allocno)->temp |
4233 | = ALLOCNO_COALESCE_DATA (a)->temp; | |
3553f0bb | 4234 | setup_slot_coalesced_allocno_live_ranges (allocno); |
058e97ec | 4235 | merge_allocnos (a, allocno); |
1756cb66 | 4236 | ira_assert (ALLOCNO_COALESCE_DATA (a)->first == a); |
058e97ec VM |
4237 | } |
4238 | } | |
3553f0bb | 4239 | for (i = 0; i < ira_allocnos_num; i++) |
9140d27b | 4240 | ira_finish_live_range_list (slot_coalesced_allocnos_live_ranges[i]); |
3553f0bb | 4241 | ira_free (slot_coalesced_allocnos_live_ranges); |
058e97ec VM |
4242 | return merged_p; |
4243 | } | |
4244 | ||
4245 | /* Sort pseudo-register numbers in array PSEUDO_REGNOS of length N for | |
4246 | subsequent assigning stack slots to them in the reload pass. To do | |
4247 | this we coalesce spilled allocnos first to decrease the number of | |
4248 | memory-memory move insns. This function is called by the | |
4249 | reload. */ | |
4250 | void | |
4251 | ira_sort_regnos_for_alter_reg (int *pseudo_regnos, int n, | |
bd5a2c67 | 4252 | machine_mode *reg_max_ref_mode) |
058e97ec VM |
4253 | { |
4254 | int max_regno = max_reg_num (); | |
4255 | int i, regno, num, slot_num; | |
4256 | ira_allocno_t allocno, a; | |
4257 | ira_allocno_iterator ai; | |
4258 | ira_allocno_t *spilled_coalesced_allocnos; | |
4259 | ||
9994ad20 KC |
4260 | ira_assert (! ira_use_lra_p); |
4261 | ||
058e97ec VM |
4262 | /* Set up allocnos can be coalesced. */ |
4263 | coloring_allocno_bitmap = ira_allocate_bitmap (); | |
4264 | for (i = 0; i < n; i++) | |
4265 | { | |
4266 | regno = pseudo_regnos[i]; | |
4267 | allocno = ira_regno_allocno_map[regno]; | |
4268 | if (allocno != NULL) | |
1756cb66 | 4269 | bitmap_set_bit (coloring_allocno_bitmap, ALLOCNO_NUM (allocno)); |
058e97ec VM |
4270 | } |
4271 | allocno_coalesced_p = false; | |
22b0982c | 4272 | processed_coalesced_allocno_bitmap = ira_allocate_bitmap (); |
1756cb66 VM |
4273 | allocno_coalesce_data |
4274 | = (coalesce_data_t) ira_allocate (sizeof (struct coalesce_data) | |
4275 | * ira_allocnos_num); | |
4276 | /* Initialize coalesce data for allocnos. */ | |
4277 | FOR_EACH_ALLOCNO (a, ai) | |
4278 | { | |
4279 | ALLOCNO_ADD_DATA (a) = allocno_coalesce_data + ALLOCNO_NUM (a); | |
4280 | ALLOCNO_COALESCE_DATA (a)->first = a; | |
4281 | ALLOCNO_COALESCE_DATA (a)->next = a; | |
4282 | } | |
22b0982c | 4283 | coalesce_allocnos (); |
058e97ec VM |
4284 | ira_free_bitmap (coloring_allocno_bitmap); |
4285 | regno_coalesced_allocno_cost | |
4286 | = (int *) ira_allocate (max_regno * sizeof (int)); | |
4287 | regno_coalesced_allocno_num | |
4288 | = (int *) ira_allocate (max_regno * sizeof (int)); | |
4289 | memset (regno_coalesced_allocno_num, 0, max_regno * sizeof (int)); | |
4290 | setup_coalesced_allocno_costs_and_nums (pseudo_regnos, n); | |
4291 | /* Sort regnos according frequencies of the corresponding coalesced | |
4292 | allocno sets. */ | |
4293 | qsort (pseudo_regnos, n, sizeof (int), coalesced_pseudo_reg_freq_compare); | |
4294 | spilled_coalesced_allocnos | |
4295 | = (ira_allocno_t *) ira_allocate (ira_allocnos_num | |
4296 | * sizeof (ira_allocno_t)); | |
4297 | /* Collect allocnos representing the spilled coalesced allocno | |
4298 | sets. */ | |
4299 | num = collect_spilled_coalesced_allocnos (pseudo_regnos, n, | |
4300 | spilled_coalesced_allocnos); | |
4301 | if (flag_ira_share_spill_slots | |
4302 | && coalesce_spill_slots (spilled_coalesced_allocnos, num)) | |
4303 | { | |
4304 | setup_coalesced_allocno_costs_and_nums (pseudo_regnos, n); | |
4305 | qsort (pseudo_regnos, n, sizeof (int), | |
4306 | coalesced_pseudo_reg_freq_compare); | |
4307 | num = collect_spilled_coalesced_allocnos (pseudo_regnos, n, | |
4308 | spilled_coalesced_allocnos); | |
4309 | } | |
4310 | ira_free_bitmap (processed_coalesced_allocno_bitmap); | |
4311 | allocno_coalesced_p = false; | |
4312 | /* Assign stack slot numbers to spilled allocno sets, use smaller | |
4313 | numbers for most frequently used coalesced allocnos. -1 is | |
4314 | reserved for dynamic search of stack slots for pseudos spilled by | |
4315 | the reload. */ | |
4316 | slot_num = 1; | |
4317 | for (i = 0; i < num; i++) | |
4318 | { | |
4319 | allocno = spilled_coalesced_allocnos[i]; | |
1756cb66 | 4320 | if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno |
058e97ec | 4321 | || ALLOCNO_HARD_REGNO (allocno) >= 0 |
55a2c322 | 4322 | || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno))) |
058e97ec VM |
4323 | continue; |
4324 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
4325 | fprintf (ira_dump_file, " Slot %d (freq,size):", slot_num); | |
4326 | slot_num++; | |
1756cb66 VM |
4327 | for (a = ALLOCNO_COALESCE_DATA (allocno)->next;; |
4328 | a = ALLOCNO_COALESCE_DATA (a)->next) | |
058e97ec VM |
4329 | { |
4330 | ira_assert (ALLOCNO_HARD_REGNO (a) < 0); | |
4331 | ALLOCNO_HARD_REGNO (a) = -slot_num; | |
4332 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
bd5a2c67 RS |
4333 | { |
4334 | machine_mode mode = wider_subreg_mode | |
4335 | (PSEUDO_REGNO_MODE (ALLOCNO_REGNO (a)), | |
4336 | reg_max_ref_mode[ALLOCNO_REGNO (a)]); | |
cf098191 RS |
4337 | fprintf (ira_dump_file, " a%dr%d(%d,", |
4338 | ALLOCNO_NUM (a), ALLOCNO_REGNO (a), ALLOCNO_FREQ (a)); | |
4339 | print_dec (GET_MODE_SIZE (mode), ira_dump_file, SIGNED); | |
4340 | fprintf (ira_dump_file, ")\n"); | |
bd5a2c67 | 4341 | } |
b8698a0f | 4342 | |
058e97ec VM |
4343 | if (a == allocno) |
4344 | break; | |
4345 | } | |
4346 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
4347 | fprintf (ira_dump_file, "\n"); | |
4348 | } | |
4349 | ira_spilled_reg_stack_slots_num = slot_num - 1; | |
4350 | ira_free (spilled_coalesced_allocnos); | |
4351 | /* Sort regnos according the slot numbers. */ | |
bd5a2c67 | 4352 | regno_max_ref_mode = reg_max_ref_mode; |
058e97ec | 4353 | qsort (pseudo_regnos, n, sizeof (int), coalesced_pseudo_reg_slot_compare); |
058e97ec | 4354 | FOR_EACH_ALLOCNO (a, ai) |
1756cb66 VM |
4355 | ALLOCNO_ADD_DATA (a) = NULL; |
4356 | ira_free (allocno_coalesce_data); | |
058e97ec VM |
4357 | ira_free (regno_coalesced_allocno_num); |
4358 | ira_free (regno_coalesced_allocno_cost); | |
4359 | } | |
4360 | ||
4361 | \f | |
4362 | ||
4363 | /* This page contains code used by the reload pass to improve the | |
4364 | final code. */ | |
4365 | ||
4366 | /* The function is called from reload to mark changes in the | |
4367 | allocation of REGNO made by the reload. Remember that reg_renumber | |
4368 | reflects the change result. */ | |
4369 | void | |
4370 | ira_mark_allocation_change (int regno) | |
4371 | { | |
4372 | ira_allocno_t a = ira_regno_allocno_map[regno]; | |
4373 | int old_hard_regno, hard_regno, cost; | |
1756cb66 | 4374 | enum reg_class aclass = ALLOCNO_CLASS (a); |
058e97ec VM |
4375 | |
4376 | ira_assert (a != NULL); | |
4377 | hard_regno = reg_renumber[regno]; | |
4378 | if ((old_hard_regno = ALLOCNO_HARD_REGNO (a)) == hard_regno) | |
4379 | return; | |
4380 | if (old_hard_regno < 0) | |
4381 | cost = -ALLOCNO_MEMORY_COST (a); | |
4382 | else | |
4383 | { | |
1756cb66 | 4384 | ira_assert (ira_class_hard_reg_index[aclass][old_hard_regno] >= 0); |
058e97ec | 4385 | cost = -(ALLOCNO_HARD_REG_COSTS (a) == NULL |
1756cb66 | 4386 | ? ALLOCNO_CLASS_COST (a) |
058e97ec | 4387 | : ALLOCNO_HARD_REG_COSTS (a) |
1756cb66 | 4388 | [ira_class_hard_reg_index[aclass][old_hard_regno]]); |
c73ccc80 | 4389 | update_costs_from_copies (a, false, false); |
058e97ec VM |
4390 | } |
4391 | ira_overall_cost -= cost; | |
4392 | ALLOCNO_HARD_REGNO (a) = hard_regno; | |
4393 | if (hard_regno < 0) | |
4394 | { | |
4395 | ALLOCNO_HARD_REGNO (a) = -1; | |
4396 | cost += ALLOCNO_MEMORY_COST (a); | |
4397 | } | |
1756cb66 | 4398 | else if (ira_class_hard_reg_index[aclass][hard_regno] >= 0) |
058e97ec VM |
4399 | { |
4400 | cost += (ALLOCNO_HARD_REG_COSTS (a) == NULL | |
1756cb66 | 4401 | ? ALLOCNO_CLASS_COST (a) |
058e97ec | 4402 | : ALLOCNO_HARD_REG_COSTS (a) |
1756cb66 | 4403 | [ira_class_hard_reg_index[aclass][hard_regno]]); |
c73ccc80 | 4404 | update_costs_from_copies (a, true, false); |
058e97ec VM |
4405 | } |
4406 | else | |
4407 | /* Reload changed class of the allocno. */ | |
4408 | cost = 0; | |
4409 | ira_overall_cost += cost; | |
4410 | } | |
4411 | ||
4412 | /* This function is called when reload deletes memory-memory move. In | |
4413 | this case we marks that the allocation of the corresponding | |
4414 | allocnos should be not changed in future. Otherwise we risk to get | |
4415 | a wrong code. */ | |
4416 | void | |
4417 | ira_mark_memory_move_deletion (int dst_regno, int src_regno) | |
4418 | { | |
4419 | ira_allocno_t dst = ira_regno_allocno_map[dst_regno]; | |
4420 | ira_allocno_t src = ira_regno_allocno_map[src_regno]; | |
4421 | ||
4422 | ira_assert (dst != NULL && src != NULL | |
4423 | && ALLOCNO_HARD_REGNO (dst) < 0 | |
4424 | && ALLOCNO_HARD_REGNO (src) < 0); | |
4425 | ALLOCNO_DONT_REASSIGN_P (dst) = true; | |
4426 | ALLOCNO_DONT_REASSIGN_P (src) = true; | |
4427 | } | |
4428 | ||
4429 | /* Try to assign a hard register (except for FORBIDDEN_REGS) to | |
3631be48 | 4430 | allocno A and return TRUE in the case of success. */ |
058e97ec VM |
4431 | static bool |
4432 | allocno_reload_assign (ira_allocno_t a, HARD_REG_SET forbidden_regs) | |
4433 | { | |
4434 | int hard_regno; | |
1756cb66 | 4435 | enum reg_class aclass; |
058e97ec | 4436 | int regno = ALLOCNO_REGNO (a); |
ac0ab4f7 BS |
4437 | HARD_REG_SET saved[2]; |
4438 | int i, n; | |
058e97ec | 4439 | |
ac0ab4f7 BS |
4440 | n = ALLOCNO_NUM_OBJECTS (a); |
4441 | for (i = 0; i < n; i++) | |
4442 | { | |
4443 | ira_object_t obj = ALLOCNO_OBJECT (a, i); | |
6576d245 | 4444 | saved[i] = OBJECT_TOTAL_CONFLICT_HARD_REGS (obj); |
44942965 | 4445 | OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) |= forbidden_regs; |
ac0ab4f7 | 4446 | if (! flag_caller_saves && ALLOCNO_CALLS_CROSSED_NUM (a) != 0) |
6c476222 | 4447 | OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) |= ira_need_caller_save_regs (a); |
ac0ab4f7 | 4448 | } |
058e97ec | 4449 | ALLOCNO_ASSIGNED_P (a) = false; |
1756cb66 | 4450 | aclass = ALLOCNO_CLASS (a); |
058e97ec VM |
4451 | update_curr_costs (a); |
4452 | assign_hard_reg (a, true); | |
4453 | hard_regno = ALLOCNO_HARD_REGNO (a); | |
4454 | reg_renumber[regno] = hard_regno; | |
4455 | if (hard_regno < 0) | |
4456 | ALLOCNO_HARD_REGNO (a) = -1; | |
4457 | else | |
4458 | { | |
1756cb66 VM |
4459 | ira_assert (ira_class_hard_reg_index[aclass][hard_regno] >= 0); |
4460 | ira_overall_cost | |
4461 | -= (ALLOCNO_MEMORY_COST (a) | |
4462 | - (ALLOCNO_HARD_REG_COSTS (a) == NULL | |
4463 | ? ALLOCNO_CLASS_COST (a) | |
4464 | : ALLOCNO_HARD_REG_COSTS (a)[ira_class_hard_reg_index | |
4465 | [aclass][hard_regno]])); | |
3366b378 | 4466 | if (ira_need_caller_save_p (a, hard_regno)) |
058e97ec VM |
4467 | { |
4468 | ira_assert (flag_caller_saves); | |
4469 | caller_save_needed = 1; | |
4470 | } | |
4471 | } | |
4472 | ||
4473 | /* If we found a hard register, modify the RTL for the pseudo | |
4474 | register to show the hard register, and mark the pseudo register | |
4475 | live. */ | |
4476 | if (reg_renumber[regno] >= 0) | |
4477 | { | |
4478 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
4479 | fprintf (ira_dump_file, ": reassign to %d\n", reg_renumber[regno]); | |
4480 | SET_REGNO (regno_reg_rtx[regno], reg_renumber[regno]); | |
4481 | mark_home_live (regno); | |
4482 | } | |
4483 | else if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
4484 | fprintf (ira_dump_file, "\n"); | |
ac0ab4f7 BS |
4485 | for (i = 0; i < n; i++) |
4486 | { | |
4487 | ira_object_t obj = ALLOCNO_OBJECT (a, i); | |
6576d245 | 4488 | OBJECT_TOTAL_CONFLICT_HARD_REGS (obj) = saved[i]; |
ac0ab4f7 | 4489 | } |
058e97ec VM |
4490 | return reg_renumber[regno] >= 0; |
4491 | } | |
4492 | ||
4493 | /* Sort pseudos according their usage frequencies (putting most | |
4494 | frequently ones first). */ | |
4495 | static int | |
4496 | pseudo_reg_compare (const void *v1p, const void *v2p) | |
4497 | { | |
4498 | int regno1 = *(const int *) v1p; | |
4499 | int regno2 = *(const int *) v2p; | |
4500 | int diff; | |
4501 | ||
4502 | if ((diff = REG_FREQ (regno2) - REG_FREQ (regno1)) != 0) | |
4503 | return diff; | |
4504 | return regno1 - regno2; | |
4505 | } | |
4506 | ||
4507 | /* Try to allocate hard registers to SPILLED_PSEUDO_REGS (there are | |
4508 | NUM of them) or spilled pseudos conflicting with pseudos in | |
4509 | SPILLED_PSEUDO_REGS. Return TRUE and update SPILLED, if the | |
4510 | allocation has been changed. The function doesn't use | |
4511 | BAD_SPILL_REGS and hard registers in PSEUDO_FORBIDDEN_REGS and | |
4512 | PSEUDO_PREVIOUS_REGS for the corresponding pseudos. The function | |
4513 | is called by the reload pass at the end of each reload | |
4514 | iteration. */ | |
4515 | bool | |
4516 | ira_reassign_pseudos (int *spilled_pseudo_regs, int num, | |
4517 | HARD_REG_SET bad_spill_regs, | |
4518 | HARD_REG_SET *pseudo_forbidden_regs, | |
6190446b JL |
4519 | HARD_REG_SET *pseudo_previous_regs, |
4520 | bitmap spilled) | |
058e97ec | 4521 | { |
016f9d9d | 4522 | int i, n, regno; |
058e97ec | 4523 | bool changed_p; |
fa86d337 | 4524 | ira_allocno_t a; |
058e97ec | 4525 | HARD_REG_SET forbidden_regs; |
6190446b JL |
4526 | bitmap temp = BITMAP_ALLOC (NULL); |
4527 | ||
4528 | /* Add pseudos which conflict with pseudos already in | |
4529 | SPILLED_PSEUDO_REGS to SPILLED_PSEUDO_REGS. This is preferable | |
4530 | to allocating in two steps as some of the conflicts might have | |
4531 | a higher priority than the pseudos passed in SPILLED_PSEUDO_REGS. */ | |
4532 | for (i = 0; i < num; i++) | |
4533 | bitmap_set_bit (temp, spilled_pseudo_regs[i]); | |
4534 | ||
4535 | for (i = 0, n = num; i < n; i++) | |
4536 | { | |
ac0ab4f7 | 4537 | int nr, j; |
6190446b JL |
4538 | int regno = spilled_pseudo_regs[i]; |
4539 | bitmap_set_bit (temp, regno); | |
4540 | ||
4541 | a = ira_regno_allocno_map[regno]; | |
ac0ab4f7 BS |
4542 | nr = ALLOCNO_NUM_OBJECTS (a); |
4543 | for (j = 0; j < nr; j++) | |
fa86d337 | 4544 | { |
ac0ab4f7 BS |
4545 | ira_object_t conflict_obj; |
4546 | ira_object_t obj = ALLOCNO_OBJECT (a, j); | |
4547 | ira_object_conflict_iterator oci; | |
4548 | ||
4549 | FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci) | |
fa86d337 | 4550 | { |
ac0ab4f7 BS |
4551 | ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj); |
4552 | if (ALLOCNO_HARD_REGNO (conflict_a) < 0 | |
4553 | && ! ALLOCNO_DONT_REASSIGN_P (conflict_a) | |
fcaa4ca4 | 4554 | && bitmap_set_bit (temp, ALLOCNO_REGNO (conflict_a))) |
ac0ab4f7 BS |
4555 | { |
4556 | spilled_pseudo_regs[num++] = ALLOCNO_REGNO (conflict_a); | |
ac0ab4f7 BS |
4557 | /* ?!? This seems wrong. */ |
4558 | bitmap_set_bit (consideration_allocno_bitmap, | |
4559 | ALLOCNO_NUM (conflict_a)); | |
4560 | } | |
fa86d337 BS |
4561 | } |
4562 | } | |
6190446b | 4563 | } |
058e97ec VM |
4564 | |
4565 | if (num > 1) | |
4566 | qsort (spilled_pseudo_regs, num, sizeof (int), pseudo_reg_compare); | |
4567 | changed_p = false; | |
4568 | /* Try to assign hard registers to pseudos from | |
4569 | SPILLED_PSEUDO_REGS. */ | |
016f9d9d | 4570 | for (i = 0; i < num; i++) |
058e97ec VM |
4571 | { |
4572 | regno = spilled_pseudo_regs[i]; | |
44942965 RS |
4573 | forbidden_regs = (bad_spill_regs |
4574 | | pseudo_forbidden_regs[regno] | |
4575 | | pseudo_previous_regs[regno]); | |
058e97ec VM |
4576 | gcc_assert (reg_renumber[regno] < 0); |
4577 | a = ira_regno_allocno_map[regno]; | |
4578 | ira_mark_allocation_change (regno); | |
4579 | ira_assert (reg_renumber[regno] < 0); | |
4580 | if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL) | |
4581 | fprintf (ira_dump_file, | |
6190446b | 4582 | " Try Assign %d(a%d), cost=%d", regno, ALLOCNO_NUM (a), |
058e97ec | 4583 | ALLOCNO_MEMORY_COST (a) |
1756cb66 | 4584 | - ALLOCNO_CLASS_COST (a)); |
058e97ec VM |
4585 | allocno_reload_assign (a, forbidden_regs); |
4586 | if (reg_renumber[regno] >= 0) | |
4587 | { | |
4588 | CLEAR_REGNO_REG_SET (spilled, regno); | |
4589 | changed_p = true; | |
4590 | } | |
058e97ec | 4591 | } |
6190446b | 4592 | BITMAP_FREE (temp); |
058e97ec VM |
4593 | return changed_p; |
4594 | } | |
4595 | ||
4596 | /* The function is called by reload and returns already allocated | |
4597 | stack slot (if any) for REGNO with given INHERENT_SIZE and | |
4598 | TOTAL_SIZE. In the case of failure to find a slot which can be | |
4599 | used for REGNO, the function returns NULL. */ | |
4600 | rtx | |
80ce7eb4 RS |
4601 | ira_reuse_stack_slot (int regno, poly_uint64 inherent_size, |
4602 | poly_uint64 total_size) | |
058e97ec VM |
4603 | { |
4604 | unsigned int i; | |
4605 | int slot_num, best_slot_num; | |
4606 | int cost, best_cost; | |
4607 | ira_copy_t cp, next_cp; | |
4608 | ira_allocno_t another_allocno, allocno = ira_regno_allocno_map[regno]; | |
4609 | rtx x; | |
4610 | bitmap_iterator bi; | |
99b1c316 | 4611 | class ira_spilled_reg_stack_slot *slot = NULL; |
058e97ec | 4612 | |
9994ad20 KC |
4613 | ira_assert (! ira_use_lra_p); |
4614 | ||
80ce7eb4 RS |
4615 | ira_assert (known_eq (inherent_size, PSEUDO_REGNO_BYTES (regno)) |
4616 | && known_le (inherent_size, total_size) | |
058e97ec VM |
4617 | && ALLOCNO_HARD_REGNO (allocno) < 0); |
4618 | if (! flag_ira_share_spill_slots) | |
4619 | return NULL_RTX; | |
4620 | slot_num = -ALLOCNO_HARD_REGNO (allocno) - 2; | |
4621 | if (slot_num != -1) | |
4622 | { | |
4623 | slot = &ira_spilled_reg_stack_slots[slot_num]; | |
4624 | x = slot->mem; | |
4625 | } | |
4626 | else | |
4627 | { | |
4628 | best_cost = best_slot_num = -1; | |
4629 | x = NULL_RTX; | |
4630 | /* It means that the pseudo was spilled in the reload pass, try | |
4631 | to reuse a slot. */ | |
4632 | for (slot_num = 0; | |
4633 | slot_num < ira_spilled_reg_stack_slots_num; | |
4634 | slot_num++) | |
4635 | { | |
4636 | slot = &ira_spilled_reg_stack_slots[slot_num]; | |
4637 | if (slot->mem == NULL_RTX) | |
4638 | continue; | |
80ce7eb4 RS |
4639 | if (maybe_lt (slot->width, total_size) |
4640 | || maybe_lt (GET_MODE_SIZE (GET_MODE (slot->mem)), inherent_size)) | |
058e97ec | 4641 | continue; |
b8698a0f | 4642 | |
058e97ec VM |
4643 | EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs, |
4644 | FIRST_PSEUDO_REGISTER, i, bi) | |
4645 | { | |
4646 | another_allocno = ira_regno_allocno_map[i]; | |
1756cb66 VM |
4647 | if (allocnos_conflict_by_live_ranges_p (allocno, |
4648 | another_allocno)) | |
058e97ec VM |
4649 | goto cont; |
4650 | } | |
4651 | for (cost = 0, cp = ALLOCNO_COPIES (allocno); | |
4652 | cp != NULL; | |
4653 | cp = next_cp) | |
4654 | { | |
4655 | if (cp->first == allocno) | |
4656 | { | |
4657 | next_cp = cp->next_first_allocno_copy; | |
4658 | another_allocno = cp->second; | |
4659 | } | |
4660 | else if (cp->second == allocno) | |
4661 | { | |
4662 | next_cp = cp->next_second_allocno_copy; | |
4663 | another_allocno = cp->first; | |
4664 | } | |
4665 | else | |
4666 | gcc_unreachable (); | |
4667 | if (cp->insn == NULL_RTX) | |
4668 | continue; | |
4669 | if (bitmap_bit_p (&slot->spilled_regs, | |
4670 | ALLOCNO_REGNO (another_allocno))) | |
4671 | cost += cp->freq; | |
4672 | } | |
4673 | if (cost > best_cost) | |
4674 | { | |
4675 | best_cost = cost; | |
4676 | best_slot_num = slot_num; | |
4677 | } | |
4678 | cont: | |
4679 | ; | |
4680 | } | |
4681 | if (best_cost >= 0) | |
4682 | { | |
99b96649 EB |
4683 | slot_num = best_slot_num; |
4684 | slot = &ira_spilled_reg_stack_slots[slot_num]; | |
058e97ec VM |
4685 | SET_REGNO_REG_SET (&slot->spilled_regs, regno); |
4686 | x = slot->mem; | |
99b96649 | 4687 | ALLOCNO_HARD_REGNO (allocno) = -slot_num - 2; |
058e97ec VM |
4688 | } |
4689 | } | |
4690 | if (x != NULL_RTX) | |
4691 | { | |
80ce7eb4 | 4692 | ira_assert (known_ge (slot->width, total_size)); |
f7556aae | 4693 | #ifdef ENABLE_IRA_CHECKING |
058e97ec VM |
4694 | EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs, |
4695 | FIRST_PSEUDO_REGISTER, i, bi) | |
4696 | { | |
1756cb66 | 4697 | ira_assert (! conflict_by_live_ranges_p (regno, i)); |
058e97ec | 4698 | } |
f7556aae | 4699 | #endif |
058e97ec VM |
4700 | SET_REGNO_REG_SET (&slot->spilled_regs, regno); |
4701 | if (internal_flag_ira_verbose > 3 && ira_dump_file) | |
4702 | { | |
4703 | fprintf (ira_dump_file, " Assigning %d(freq=%d) slot %d of", | |
4704 | regno, REG_FREQ (regno), slot_num); | |
4705 | EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs, | |
4706 | FIRST_PSEUDO_REGISTER, i, bi) | |
4707 | { | |
4708 | if ((unsigned) regno != i) | |
4709 | fprintf (ira_dump_file, " %d", i); | |
4710 | } | |
4711 | fprintf (ira_dump_file, "\n"); | |
4712 | } | |
4713 | } | |
4714 | return x; | |
4715 | } | |
4716 | ||
4717 | /* This is called by reload every time a new stack slot X with | |
4718 | TOTAL_SIZE was allocated for REGNO. We store this info for | |
4719 | subsequent ira_reuse_stack_slot calls. */ | |
4720 | void | |
80ce7eb4 | 4721 | ira_mark_new_stack_slot (rtx x, int regno, poly_uint64 total_size) |
058e97ec | 4722 | { |
99b1c316 | 4723 | class ira_spilled_reg_stack_slot *slot; |
058e97ec VM |
4724 | int slot_num; |
4725 | ira_allocno_t allocno; | |
4726 | ||
9994ad20 KC |
4727 | ira_assert (! ira_use_lra_p); |
4728 | ||
80ce7eb4 | 4729 | ira_assert (known_le (PSEUDO_REGNO_BYTES (regno), total_size)); |
058e97ec VM |
4730 | allocno = ira_regno_allocno_map[regno]; |
4731 | slot_num = -ALLOCNO_HARD_REGNO (allocno) - 2; | |
4732 | if (slot_num == -1) | |
4733 | { | |
4734 | slot_num = ira_spilled_reg_stack_slots_num++; | |
4735 | ALLOCNO_HARD_REGNO (allocno) = -slot_num - 2; | |
4736 | } | |
4737 | slot = &ira_spilled_reg_stack_slots[slot_num]; | |
4738 | INIT_REG_SET (&slot->spilled_regs); | |
4739 | SET_REGNO_REG_SET (&slot->spilled_regs, regno); | |
4740 | slot->mem = x; | |
4741 | slot->width = total_size; | |
4742 | if (internal_flag_ira_verbose > 3 && ira_dump_file) | |
4743 | fprintf (ira_dump_file, " Assigning %d(freq=%d) a new slot %d\n", | |
4744 | regno, REG_FREQ (regno), slot_num); | |
4745 | } | |
4746 | ||
4747 | ||
4748 | /* Return spill cost for pseudo-registers whose numbers are in array | |
4749 | REGNOS (with a negative number as an end marker) for reload with | |
4750 | given IN and OUT for INSN. Return also number points (through | |
4751 | EXCESS_PRESSURE_LIVE_LENGTH) where the pseudo-register lives and | |
4752 | the register pressure is high, number of references of the | |
6c476222 RS |
4753 | pseudo-registers (through NREFS), the number of psuedo registers |
4754 | whose allocated register wouldn't need saving in the prologue | |
4755 | (through CALL_USED_COUNT), and the first hard regno occupied by the | |
058e97ec VM |
4756 | pseudo-registers (through FIRST_HARD_REGNO). */ |
4757 | static int | |
8c797f81 | 4758 | calculate_spill_cost (int *regnos, rtx in, rtx out, rtx_insn *insn, |
058e97ec VM |
4759 | int *excess_pressure_live_length, |
4760 | int *nrefs, int *call_used_count, int *first_hard_regno) | |
4761 | { | |
6c476222 | 4762 | int i, cost, regno, hard_regno, count, saved_cost; |
058e97ec VM |
4763 | bool in_p, out_p; |
4764 | int length; | |
4765 | ira_allocno_t a; | |
4766 | ||
4767 | *nrefs = 0; | |
4768 | for (length = count = cost = i = 0;; i++) | |
4769 | { | |
4770 | regno = regnos[i]; | |
4771 | if (regno < 0) | |
4772 | break; | |
4773 | *nrefs += REG_N_REFS (regno); | |
4774 | hard_regno = reg_renumber[regno]; | |
4775 | ira_assert (hard_regno >= 0); | |
4776 | a = ira_regno_allocno_map[regno]; | |
ac0ab4f7 | 4777 | length += ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a) / ALLOCNO_NUM_OBJECTS (a); |
1756cb66 | 4778 | cost += ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a); |
6c476222 RS |
4779 | if (in_hard_reg_set_p (crtl->abi->full_reg_clobbers (), |
4780 | ALLOCNO_MODE (a), hard_regno)) | |
058e97ec VM |
4781 | count++; |
4782 | in_p = in && REG_P (in) && (int) REGNO (in) == hard_regno; | |
4783 | out_p = out && REG_P (out) && (int) REGNO (out) == hard_regno; | |
4784 | if ((in_p || out_p) | |
4785 | && find_regno_note (insn, REG_DEAD, hard_regno) != NULL_RTX) | |
4786 | { | |
4787 | saved_cost = 0; | |
4788 | if (in_p) | |
4789 | saved_cost += ira_memory_move_cost | |
1756cb66 | 4790 | [ALLOCNO_MODE (a)][ALLOCNO_CLASS (a)][1]; |
058e97ec VM |
4791 | if (out_p) |
4792 | saved_cost | |
4793 | += ira_memory_move_cost | |
1756cb66 | 4794 | [ALLOCNO_MODE (a)][ALLOCNO_CLASS (a)][0]; |
058e97ec VM |
4795 | cost -= REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn)) * saved_cost; |
4796 | } | |
4797 | } | |
4798 | *excess_pressure_live_length = length; | |
4799 | *call_used_count = count; | |
4800 | hard_regno = -1; | |
4801 | if (regnos[0] >= 0) | |
4802 | { | |
4803 | hard_regno = reg_renumber[regnos[0]]; | |
4804 | } | |
4805 | *first_hard_regno = hard_regno; | |
4806 | return cost; | |
4807 | } | |
4808 | ||
4809 | /* Return TRUE if spilling pseudo-registers whose numbers are in array | |
4810 | REGNOS is better than spilling pseudo-registers with numbers in | |
4811 | OTHER_REGNOS for reload with given IN and OUT for INSN. The | |
4812 | function used by the reload pass to make better register spilling | |
4813 | decisions. */ | |
4814 | bool | |
4815 | ira_better_spill_reload_regno_p (int *regnos, int *other_regnos, | |
8c797f81 | 4816 | rtx in, rtx out, rtx_insn *insn) |
058e97ec VM |
4817 | { |
4818 | int cost, other_cost; | |
4819 | int length, other_length; | |
4820 | int nrefs, other_nrefs; | |
4821 | int call_used_count, other_call_used_count; | |
4822 | int hard_regno, other_hard_regno; | |
4823 | ||
b8698a0f | 4824 | cost = calculate_spill_cost (regnos, in, out, insn, |
058e97ec VM |
4825 | &length, &nrefs, &call_used_count, &hard_regno); |
4826 | other_cost = calculate_spill_cost (other_regnos, in, out, insn, | |
4827 | &other_length, &other_nrefs, | |
4828 | &other_call_used_count, | |
4829 | &other_hard_regno); | |
4830 | if (nrefs == 0 && other_nrefs != 0) | |
4831 | return true; | |
4832 | if (nrefs != 0 && other_nrefs == 0) | |
4833 | return false; | |
4834 | if (cost != other_cost) | |
4835 | return cost < other_cost; | |
4836 | if (length != other_length) | |
4837 | return length > other_length; | |
4838 | #ifdef REG_ALLOC_ORDER | |
4839 | if (hard_regno >= 0 && other_hard_regno >= 0) | |
4840 | return (inv_reg_alloc_order[hard_regno] | |
4841 | < inv_reg_alloc_order[other_hard_regno]); | |
4842 | #else | |
4843 | if (call_used_count != other_call_used_count) | |
4844 | return call_used_count > other_call_used_count; | |
4845 | #endif | |
4846 | return false; | |
4847 | } | |
4848 | ||
4849 | \f | |
4850 | ||
4851 | /* Allocate and initialize data necessary for assign_hard_reg. */ | |
4852 | void | |
4853 | ira_initiate_assign (void) | |
4854 | { | |
4855 | sorted_allocnos | |
4856 | = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t) | |
4857 | * ira_allocnos_num); | |
4858 | consideration_allocno_bitmap = ira_allocate_bitmap (); | |
4859 | initiate_cost_update (); | |
4860 | allocno_priorities = (int *) ira_allocate (sizeof (int) * ira_allocnos_num); | |
bf08fb16 VM |
4861 | sorted_copies = (ira_copy_t *) ira_allocate (ira_copies_num |
4862 | * sizeof (ira_copy_t)); | |
058e97ec VM |
4863 | } |
4864 | ||
4865 | /* Deallocate data used by assign_hard_reg. */ | |
4866 | void | |
4867 | ira_finish_assign (void) | |
4868 | { | |
4869 | ira_free (sorted_allocnos); | |
4870 | ira_free_bitmap (consideration_allocno_bitmap); | |
4871 | finish_cost_update (); | |
4872 | ira_free (allocno_priorities); | |
bf08fb16 | 4873 | ira_free (sorted_copies); |
058e97ec VM |
4874 | } |
4875 | ||
4876 | \f | |
4877 | ||
4878 | /* Entry function doing color-based register allocation. */ | |
cb1ca6ac VM |
4879 | static void |
4880 | color (void) | |
058e97ec | 4881 | { |
9771b263 | 4882 | allocno_stack_vec.create (ira_allocnos_num); |
058e97ec VM |
4883 | memset (allocated_hardreg_p, 0, sizeof (allocated_hardreg_p)); |
4884 | ira_initiate_assign (); | |
4885 | do_coloring (); | |
4886 | ira_finish_assign (); | |
9771b263 | 4887 | allocno_stack_vec.release (); |
058e97ec VM |
4888 | move_spill_restore (); |
4889 | } | |
4890 | ||
4891 | \f | |
4892 | ||
4893 | /* This page contains a simple register allocator without usage of | |
4894 | allocno conflicts. This is used for fast allocation for -O0. */ | |
4895 | ||
4896 | /* Do register allocation by not using allocno conflicts. It uses | |
4897 | only allocno live ranges. The algorithm is close to Chow's | |
4898 | priority coloring. */ | |
cb1ca6ac VM |
4899 | static void |
4900 | fast_allocation (void) | |
058e97ec | 4901 | { |
159fdc39 VM |
4902 | int i, j, k, num, class_size, hard_regno, best_hard_regno, cost, min_cost; |
4903 | int *costs; | |
058e97ec VM |
4904 | #ifdef STACK_REGS |
4905 | bool no_stack_reg_p; | |
4906 | #endif | |
1756cb66 | 4907 | enum reg_class aclass; |
ef4bddc2 | 4908 | machine_mode mode; |
058e97ec VM |
4909 | ira_allocno_t a; |
4910 | ira_allocno_iterator ai; | |
b14151b5 | 4911 | live_range_t r; |
058e97ec VM |
4912 | HARD_REG_SET conflict_hard_regs, *used_hard_regs; |
4913 | ||
058e97ec VM |
4914 | sorted_allocnos = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t) |
4915 | * ira_allocnos_num); | |
4916 | num = 0; | |
4917 | FOR_EACH_ALLOCNO (a, ai) | |
4918 | sorted_allocnos[num++] = a; | |
1ae64b0f VM |
4919 | allocno_priorities = (int *) ira_allocate (sizeof (int) * ira_allocnos_num); |
4920 | setup_allocno_priorities (sorted_allocnos, num); | |
4921 | used_hard_regs = (HARD_REG_SET *) ira_allocate (sizeof (HARD_REG_SET) | |
4922 | * ira_max_point); | |
4923 | for (i = 0; i < ira_max_point; i++) | |
4924 | CLEAR_HARD_REG_SET (used_hard_regs[i]); | |
311aab06 | 4925 | qsort (sorted_allocnos, num, sizeof (ira_allocno_t), |
058e97ec VM |
4926 | allocno_priority_compare_func); |
4927 | for (i = 0; i < num; i++) | |
4928 | { | |
ac0ab4f7 BS |
4929 | int nr, l; |
4930 | ||
058e97ec | 4931 | a = sorted_allocnos[i]; |
ac0ab4f7 BS |
4932 | nr = ALLOCNO_NUM_OBJECTS (a); |
4933 | CLEAR_HARD_REG_SET (conflict_hard_regs); | |
4934 | for (l = 0; l < nr; l++) | |
4935 | { | |
4936 | ira_object_t obj = ALLOCNO_OBJECT (a, l); | |
44942965 | 4937 | conflict_hard_regs |= OBJECT_CONFLICT_HARD_REGS (obj); |
ac0ab4f7 BS |
4938 | for (r = OBJECT_LIVE_RANGES (obj); r != NULL; r = r->next) |
4939 | for (j = r->start; j <= r->finish; j++) | |
44942965 | 4940 | conflict_hard_regs |= used_hard_regs[j]; |
ac0ab4f7 | 4941 | } |
1756cb66 | 4942 | aclass = ALLOCNO_CLASS (a); |
6b8d9676 VM |
4943 | ALLOCNO_ASSIGNED_P (a) = true; |
4944 | ALLOCNO_HARD_REGNO (a) = -1; | |
1756cb66 | 4945 | if (hard_reg_set_subset_p (reg_class_contents[aclass], |
058e97ec VM |
4946 | conflict_hard_regs)) |
4947 | continue; | |
4948 | mode = ALLOCNO_MODE (a); | |
4949 | #ifdef STACK_REGS | |
4950 | no_stack_reg_p = ALLOCNO_NO_STACK_REG_P (a); | |
4951 | #endif | |
1756cb66 | 4952 | class_size = ira_class_hard_regs_num[aclass]; |
159fdc39 VM |
4953 | costs = ALLOCNO_HARD_REG_COSTS (a); |
4954 | min_cost = INT_MAX; | |
4955 | best_hard_regno = -1; | |
058e97ec VM |
4956 | for (j = 0; j < class_size; j++) |
4957 | { | |
1756cb66 | 4958 | hard_regno = ira_class_hard_regs[aclass][j]; |
058e97ec VM |
4959 | #ifdef STACK_REGS |
4960 | if (no_stack_reg_p && FIRST_STACK_REG <= hard_regno | |
4961 | && hard_regno <= LAST_STACK_REG) | |
4962 | continue; | |
4963 | #endif | |
9181a6e5 | 4964 | if (ira_hard_reg_set_intersection_p (hard_regno, mode, conflict_hard_regs) |
058e97ec | 4965 | || (TEST_HARD_REG_BIT |
1756cb66 | 4966 | (ira_prohibited_class_mode_regs[aclass][mode], hard_regno))) |
058e97ec | 4967 | continue; |
159fdc39 VM |
4968 | if (costs == NULL) |
4969 | { | |
4970 | best_hard_regno = hard_regno; | |
4971 | break; | |
4972 | } | |
4973 | cost = costs[j]; | |
4974 | if (min_cost > cost) | |
ac0ab4f7 | 4975 | { |
159fdc39 VM |
4976 | min_cost = cost; |
4977 | best_hard_regno = hard_regno; | |
ac0ab4f7 | 4978 | } |
159fdc39 VM |
4979 | } |
4980 | if (best_hard_regno < 0) | |
4981 | continue; | |
4982 | ALLOCNO_HARD_REGNO (a) = hard_regno = best_hard_regno; | |
4983 | for (l = 0; l < nr; l++) | |
4984 | { | |
4985 | ira_object_t obj = ALLOCNO_OBJECT (a, l); | |
4986 | for (r = OBJECT_LIVE_RANGES (obj); r != NULL; r = r->next) | |
4987 | for (k = r->start; k <= r->finish; k++) | |
44942965 | 4988 | used_hard_regs[k] |= ira_reg_mode_hard_regset[hard_regno][mode]; |
058e97ec VM |
4989 | } |
4990 | } | |
4991 | ira_free (sorted_allocnos); | |
4992 | ira_free (used_hard_regs); | |
4993 | ira_free (allocno_priorities); | |
4994 | if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL) | |
4995 | ira_print_disposition (ira_dump_file); | |
4996 | } | |
cb1ca6ac VM |
4997 | |
4998 | \f | |
4999 | ||
5000 | /* Entry function doing coloring. */ | |
5001 | void | |
5002 | ira_color (void) | |
5003 | { | |
5004 | ira_allocno_t a; | |
5005 | ira_allocno_iterator ai; | |
5006 | ||
5007 | /* Setup updated costs. */ | |
5008 | FOR_EACH_ALLOCNO (a, ai) | |
5009 | { | |
5010 | ALLOCNO_UPDATED_MEMORY_COST (a) = ALLOCNO_MEMORY_COST (a); | |
1756cb66 | 5011 | ALLOCNO_UPDATED_CLASS_COST (a) = ALLOCNO_CLASS_COST (a); |
cb1ca6ac | 5012 | } |
311aab06 | 5013 | if (ira_conflicts_p) |
cb1ca6ac VM |
5014 | color (); |
5015 | else | |
5016 | fast_allocation (); | |
5017 | } |