]>
Commit | Line | Data |
---|---|---|
55a2c322 | 1 | /* Assign reload pseudos. |
5624e564 | 2 | Copyright (C) 2010-2015 Free Software Foundation, Inc. |
55a2c322 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 | ||
22 | /* This file's main objective is to assign hard registers to reload | |
23 | pseudos. It also tries to allocate hard registers to other | |
24 | pseudos, but at a lower priority than the reload pseudos. The pass | |
25 | does not transform the RTL. | |
26 | ||
27 | We must allocate a hard register to every reload pseudo. We try to | |
28 | increase the chances of finding a viable allocation by assigning | |
29 | the pseudos in order of fewest available hard registers first. If | |
30 | we still fail to find a hard register, we spill other (non-reload) | |
31 | pseudos in order to make room. | |
32 | ||
33 | find_hard_regno_for finds hard registers for allocation without | |
34 | spilling. spill_for does the same with spilling. Both functions | |
35 | use a cost model to determine the most profitable choice of hard | |
36 | and spill registers. | |
37 | ||
38 | Once we have finished allocating reload pseudos, we also try to | |
39 | assign registers to other (non-reload) pseudos. This is useful if | |
40 | hard registers were freed up by the spilling just described. | |
41 | ||
42 | We try to assign hard registers by collecting pseudos into threads. | |
43 | These threads contain reload and inheritance pseudos that are | |
44 | connected by copies (move insns). Doing this improves the chances | |
45 | of pseudos in the thread getting the same hard register and, as a | |
46 | result, of allowing some move insns to be deleted. | |
47 | ||
48 | When we assign a hard register to a pseudo, we decrease the cost of | |
49 | using the same hard register for pseudos that are connected by | |
50 | copies. | |
51 | ||
52 | If two hard registers have the same frequency-derived cost, we | |
53 | prefer hard registers with higher priorities. The mapping of | |
54 | registers to priorities is controlled by the register_priority | |
55 | target hook. For example, x86-64 has a few register priorities: | |
56 | hard registers with and without REX prefixes have different | |
57 | priorities. This permits us to generate smaller code as insns | |
58 | without REX prefixes are shorter. | |
59 | ||
60 | If a few hard registers are still equally good for the assignment, | |
61 | we choose the least used hard register. It is called leveling and | |
62 | may be profitable for some targets. | |
63 | ||
64 | Only insns with changed allocation pseudos are processed on the | |
65 | next constraint pass. | |
66 | ||
67 | The pseudo live-ranges are used to find conflicting pseudos. | |
68 | ||
69 | For understanding the code, it is important to keep in mind that | |
70 | inheritance, split, and reload pseudos created since last | |
71 | constraint pass have regno >= lra_constraint_new_regno_start. | |
72 | Inheritance and split pseudos created on any pass are in the | |
73 | corresponding bitmaps. Inheritance and split pseudos since the | |
74 | last constraint pass have also the corresponding non-negative | |
75 | restore_regno. */ | |
76 | ||
77 | #include "config.h" | |
78 | #include "system.h" | |
79 | #include "coretypes.h" | |
c7131fb2 | 80 | #include "backend.h" |
9fdcd34e | 81 | #include "predict.h" |
c7131fb2 | 82 | #include "tree.h" |
55a2c322 | 83 | #include "rtl.h" |
c7131fb2 | 84 | #include "df.h" |
ce940020 | 85 | #include "rtl-error.h" |
55a2c322 VM |
86 | #include "tm_p.h" |
87 | #include "target.h" | |
88 | #include "insn-config.h" | |
89 | #include "recog.h" | |
90 | #include "output.h" | |
91 | #include "regs.h" | |
36566b39 | 92 | #include "flags.h" |
36566b39 | 93 | #include "alias.h" |
36566b39 PK |
94 | #include "expmed.h" |
95 | #include "dojump.h" | |
96 | #include "explow.h" | |
97 | #include "calls.h" | |
98 | #include "emit-rtl.h" | |
99 | #include "varasm.h" | |
100 | #include "stmt.h" | |
55a2c322 | 101 | #include "expr.h" |
55a2c322 | 102 | #include "except.h" |
55a2c322 VM |
103 | #include "ira.h" |
104 | #include "sparseset.h" | |
88def637 | 105 | #include "params.h" |
c7131fb2 AM |
106 | #include "lra.h" |
107 | #include "insn-attr.h" | |
108 | #include "insn-codes.h" | |
55a2c322 VM |
109 | #include "lra-int.h" |
110 | ||
f54437d5 VM |
111 | /* Current iteration number of the pass and current iteration number |
112 | of the pass after the latest spill pass when any former reload | |
113 | pseudo was spilled. */ | |
114 | int lra_assignment_iter; | |
115 | int lra_assignment_iter_after_spill; | |
116 | ||
117 | /* Flag of spilling former reload pseudos on this pass. */ | |
118 | static bool former_reload_pseudo_spill_p; | |
119 | ||
55a2c322 VM |
120 | /* Array containing corresponding values of function |
121 | lra_get_allocno_class. It is used to speed up the code. */ | |
122 | static enum reg_class *regno_allocno_class_array; | |
123 | ||
124 | /* Information about the thread to which a pseudo belongs. Threads are | |
125 | a set of connected reload and inheritance pseudos with the same set of | |
126 | available hard registers. Lone registers belong to their own threads. */ | |
127 | struct regno_assign_info | |
128 | { | |
129 | /* First/next pseudo of the same thread. */ | |
130 | int first, next; | |
131 | /* Frequency of the thread (execution frequency of only reload | |
132 | pseudos in the thread when the thread contains a reload pseudo). | |
133 | Defined only for the first thread pseudo. */ | |
134 | int freq; | |
135 | }; | |
136 | ||
137 | /* Map regno to the corresponding regno assignment info. */ | |
138 | static struct regno_assign_info *regno_assign_info; | |
139 | ||
bc404e1b VM |
140 | /* All inherited, subreg or optional pseudos created before last spill |
141 | sub-pass. Such pseudos are permitted to get memory instead of hard | |
142 | regs. */ | |
143 | static bitmap_head non_reload_pseudos; | |
144 | ||
55a2c322 VM |
145 | /* Process a pseudo copy with execution frequency COPY_FREQ connecting |
146 | REGNO1 and REGNO2 to form threads. */ | |
147 | static void | |
148 | process_copy_to_form_thread (int regno1, int regno2, int copy_freq) | |
149 | { | |
150 | int last, regno1_first, regno2_first; | |
151 | ||
152 | lra_assert (regno1 >= lra_constraint_new_regno_start | |
153 | && regno2 >= lra_constraint_new_regno_start); | |
154 | regno1_first = regno_assign_info[regno1].first; | |
155 | regno2_first = regno_assign_info[regno2].first; | |
156 | if (regno1_first != regno2_first) | |
157 | { | |
158 | for (last = regno2_first; | |
159 | regno_assign_info[last].next >= 0; | |
160 | last = regno_assign_info[last].next) | |
161 | regno_assign_info[last].first = regno1_first; | |
162 | regno_assign_info[last].first = regno1_first; | |
163 | regno_assign_info[last].next = regno_assign_info[regno1_first].next; | |
164 | regno_assign_info[regno1_first].next = regno2_first; | |
165 | regno_assign_info[regno1_first].freq | |
166 | += regno_assign_info[regno2_first].freq; | |
167 | } | |
168 | regno_assign_info[regno1_first].freq -= 2 * copy_freq; | |
169 | lra_assert (regno_assign_info[regno1_first].freq >= 0); | |
170 | } | |
171 | ||
172 | /* Initialize REGNO_ASSIGN_INFO and form threads. */ | |
173 | static void | |
174 | init_regno_assign_info (void) | |
175 | { | |
176 | int i, regno1, regno2, max_regno = max_reg_num (); | |
177 | lra_copy_t cp; | |
f4eafc30 | 178 | |
55a2c322 VM |
179 | regno_assign_info = XNEWVEC (struct regno_assign_info, max_regno); |
180 | for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) | |
181 | { | |
182 | regno_assign_info[i].first = i; | |
183 | regno_assign_info[i].next = -1; | |
184 | regno_assign_info[i].freq = lra_reg_info[i].freq; | |
185 | } | |
186 | /* Form the threads. */ | |
187 | for (i = 0; (cp = lra_get_copy (i)) != NULL; i++) | |
188 | if ((regno1 = cp->regno1) >= lra_constraint_new_regno_start | |
189 | && (regno2 = cp->regno2) >= lra_constraint_new_regno_start | |
190 | && reg_renumber[regno1] < 0 && lra_reg_info[regno1].nrefs != 0 | |
191 | && reg_renumber[regno2] < 0 && lra_reg_info[regno2].nrefs != 0 | |
192 | && (ira_class_hard_regs_num[regno_allocno_class_array[regno1]] | |
193 | == ira_class_hard_regs_num[regno_allocno_class_array[regno2]])) | |
194 | process_copy_to_form_thread (regno1, regno2, cp->freq); | |
195 | } | |
196 | ||
197 | /* Free REGNO_ASSIGN_INFO. */ | |
198 | static void | |
199 | finish_regno_assign_info (void) | |
200 | { | |
201 | free (regno_assign_info); | |
202 | } | |
203 | ||
204 | /* The function is used to sort *reload* and *inheritance* pseudos to | |
205 | try to assign them hard registers. We put pseudos from the same | |
206 | thread always nearby. */ | |
207 | static int | |
208 | reload_pseudo_compare_func (const void *v1p, const void *v2p) | |
209 | { | |
210 | int r1 = *(const int *) v1p, r2 = *(const int *) v2p; | |
211 | enum reg_class cl1 = regno_allocno_class_array[r1]; | |
212 | enum reg_class cl2 = regno_allocno_class_array[r2]; | |
213 | int diff; | |
f4eafc30 | 214 | |
55a2c322 VM |
215 | lra_assert (r1 >= lra_constraint_new_regno_start |
216 | && r2 >= lra_constraint_new_regno_start); | |
f4eafc30 | 217 | |
55a2c322 VM |
218 | /* Prefer to assign reload registers with smaller classes first to |
219 | guarantee assignment to all reload registers. */ | |
220 | if ((diff = (ira_class_hard_regs_num[cl1] | |
221 | - ira_class_hard_regs_num[cl2])) != 0) | |
222 | return diff; | |
bc404e1b VM |
223 | if ((diff |
224 | = (ira_reg_class_max_nregs[cl2][lra_reg_info[r2].biggest_mode] | |
225 | - ira_reg_class_max_nregs[cl1][lra_reg_info[r1].biggest_mode])) != 0 | |
226 | /* The code below executes rarely as nregs == 1 in most cases. | |
227 | So we should not worry about using faster data structures to | |
228 | check reload pseudos. */ | |
229 | && ! bitmap_bit_p (&non_reload_pseudos, r1) | |
230 | && ! bitmap_bit_p (&non_reload_pseudos, r2)) | |
231 | return diff; | |
55a2c322 VM |
232 | if ((diff = (regno_assign_info[regno_assign_info[r2].first].freq |
233 | - regno_assign_info[regno_assign_info[r1].first].freq)) != 0) | |
234 | return diff; | |
47918951 VM |
235 | /* Allocate bigger pseudos first to avoid register file |
236 | fragmentation. */ | |
237 | if ((diff | |
238 | = (ira_reg_class_max_nregs[cl2][lra_reg_info[r2].biggest_mode] | |
239 | - ira_reg_class_max_nregs[cl1][lra_reg_info[r1].biggest_mode])) != 0) | |
240 | return diff; | |
55a2c322 VM |
241 | /* Put pseudos from the thread nearby. */ |
242 | if ((diff = regno_assign_info[r1].first - regno_assign_info[r2].first) != 0) | |
243 | return diff; | |
244 | /* If regs are equally good, sort by their numbers, so that the | |
245 | results of qsort leave nothing to chance. */ | |
246 | return r1 - r2; | |
247 | } | |
248 | ||
249 | /* The function is used to sort *non-reload* pseudos to try to assign | |
250 | them hard registers. The order calculation is simpler than in the | |
251 | previous function and based on the pseudo frequency usage. */ | |
252 | static int | |
253 | pseudo_compare_func (const void *v1p, const void *v2p) | |
254 | { | |
255 | int r1 = *(const int *) v1p, r2 = *(const int *) v2p; | |
256 | int diff; | |
257 | ||
b81a2f0d VM |
258 | /* Assign hard reg to static chain pointer first pseudo when |
259 | non-local goto is used. */ | |
260 | if (non_spilled_static_chain_regno_p (r1)) | |
261 | return -1; | |
262 | else if (non_spilled_static_chain_regno_p (r2)) | |
263 | return 1; | |
264 | ||
55a2c322 VM |
265 | /* Prefer to assign more frequently used registers first. */ |
266 | if ((diff = lra_reg_info[r2].freq - lra_reg_info[r1].freq) != 0) | |
267 | return diff; | |
f4eafc30 | 268 | |
55a2c322 VM |
269 | /* If regs are equally good, sort by their numbers, so that the |
270 | results of qsort leave nothing to chance. */ | |
271 | return r1 - r2; | |
272 | } | |
273 | ||
274 | /* Arrays of size LRA_LIVE_MAX_POINT mapping a program point to the | |
275 | pseudo live ranges with given start point. We insert only live | |
276 | ranges of pseudos interesting for assignment purposes. They are | |
277 | reload pseudos and pseudos assigned to hard registers. */ | |
278 | static lra_live_range_t *start_point_ranges; | |
279 | ||
280 | /* Used as a flag that a live range is not inserted in the start point | |
281 | chain. */ | |
282 | static struct lra_live_range not_in_chain_mark; | |
283 | ||
284 | /* Create and set up START_POINT_RANGES. */ | |
285 | static void | |
286 | create_live_range_start_chains (void) | |
287 | { | |
288 | int i, max_regno; | |
289 | lra_live_range_t r; | |
290 | ||
291 | start_point_ranges = XCNEWVEC (lra_live_range_t, lra_live_max_point); | |
292 | max_regno = max_reg_num (); | |
293 | for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) | |
294 | if (i >= lra_constraint_new_regno_start || reg_renumber[i] >= 0) | |
295 | { | |
296 | for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next) | |
297 | { | |
298 | r->start_next = start_point_ranges[r->start]; | |
299 | start_point_ranges[r->start] = r; | |
300 | } | |
301 | } | |
302 | else | |
303 | { | |
304 | for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next) | |
305 | r->start_next = ¬_in_chain_mark; | |
306 | } | |
307 | } | |
308 | ||
309 | /* Insert live ranges of pseudo REGNO into start chains if they are | |
310 | not there yet. */ | |
311 | static void | |
312 | insert_in_live_range_start_chain (int regno) | |
313 | { | |
314 | lra_live_range_t r = lra_reg_info[regno].live_ranges; | |
315 | ||
316 | if (r->start_next != ¬_in_chain_mark) | |
317 | return; | |
318 | for (; r != NULL; r = r->next) | |
319 | { | |
320 | r->start_next = start_point_ranges[r->start]; | |
321 | start_point_ranges[r->start] = r; | |
322 | } | |
323 | } | |
324 | ||
325 | /* Free START_POINT_RANGES. */ | |
326 | static void | |
327 | finish_live_range_start_chains (void) | |
328 | { | |
329 | gcc_assert (start_point_ranges != NULL); | |
330 | free (start_point_ranges); | |
331 | start_point_ranges = NULL; | |
332 | } | |
333 | ||
334 | /* Map: program point -> bitmap of all pseudos living at the point and | |
335 | assigned to hard registers. */ | |
336 | static bitmap_head *live_hard_reg_pseudos; | |
337 | static bitmap_obstack live_hard_reg_pseudos_bitmap_obstack; | |
338 | ||
339 | /* reg_renumber corresponding to pseudos marked in | |
340 | live_hard_reg_pseudos. reg_renumber might be not matched to | |
341 | live_hard_reg_pseudos but live_pseudos_reg_renumber always reflects | |
342 | live_hard_reg_pseudos. */ | |
343 | static int *live_pseudos_reg_renumber; | |
344 | ||
345 | /* Sparseset used to calculate living hard reg pseudos for some program | |
346 | point range. */ | |
347 | static sparseset live_range_hard_reg_pseudos; | |
348 | ||
349 | /* Sparseset used to calculate living reload/inheritance pseudos for | |
350 | some program point range. */ | |
351 | static sparseset live_range_reload_inheritance_pseudos; | |
352 | ||
353 | /* Allocate and initialize the data about living pseudos at program | |
354 | points. */ | |
355 | static void | |
356 | init_lives (void) | |
357 | { | |
358 | int i, max_regno = max_reg_num (); | |
359 | ||
360 | live_range_hard_reg_pseudos = sparseset_alloc (max_regno); | |
361 | live_range_reload_inheritance_pseudos = sparseset_alloc (max_regno); | |
362 | live_hard_reg_pseudos = XNEWVEC (bitmap_head, lra_live_max_point); | |
363 | bitmap_obstack_initialize (&live_hard_reg_pseudos_bitmap_obstack); | |
364 | for (i = 0; i < lra_live_max_point; i++) | |
365 | bitmap_initialize (&live_hard_reg_pseudos[i], | |
366 | &live_hard_reg_pseudos_bitmap_obstack); | |
367 | live_pseudos_reg_renumber = XNEWVEC (int, max_regno); | |
368 | for (i = 0; i < max_regno; i++) | |
369 | live_pseudos_reg_renumber[i] = -1; | |
370 | } | |
371 | ||
372 | /* Free the data about living pseudos at program points. */ | |
373 | static void | |
374 | finish_lives (void) | |
375 | { | |
376 | sparseset_free (live_range_hard_reg_pseudos); | |
377 | sparseset_free (live_range_reload_inheritance_pseudos); | |
378 | free (live_hard_reg_pseudos); | |
379 | bitmap_obstack_release (&live_hard_reg_pseudos_bitmap_obstack); | |
380 | free (live_pseudos_reg_renumber); | |
381 | } | |
382 | ||
383 | /* Update the LIVE_HARD_REG_PSEUDOS and LIVE_PSEUDOS_REG_RENUMBER | |
384 | entries for pseudo REGNO. Assume that the register has been | |
385 | spilled if FREE_P, otherwise assume that it has been assigned | |
386 | reg_renumber[REGNO] (if >= 0). We also insert the pseudo live | |
387 | ranges in the start chains when it is assumed to be assigned to a | |
388 | hard register because we use the chains of pseudos assigned to hard | |
389 | registers during allocation. */ | |
390 | static void | |
391 | update_lives (int regno, bool free_p) | |
392 | { | |
393 | int p; | |
394 | lra_live_range_t r; | |
395 | ||
396 | if (reg_renumber[regno] < 0) | |
397 | return; | |
398 | live_pseudos_reg_renumber[regno] = free_p ? -1 : reg_renumber[regno]; | |
399 | for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next) | |
400 | { | |
401 | for (p = r->start; p <= r->finish; p++) | |
402 | if (free_p) | |
403 | bitmap_clear_bit (&live_hard_reg_pseudos[p], regno); | |
404 | else | |
405 | { | |
406 | bitmap_set_bit (&live_hard_reg_pseudos[p], regno); | |
407 | insert_in_live_range_start_chain (regno); | |
408 | } | |
409 | } | |
410 | } | |
411 | ||
412 | /* Sparseset used to calculate reload pseudos conflicting with a given | |
413 | pseudo when we are trying to find a hard register for the given | |
414 | pseudo. */ | |
415 | static sparseset conflict_reload_and_inheritance_pseudos; | |
416 | ||
417 | /* Map: program point -> bitmap of all reload and inheritance pseudos | |
418 | living at the point. */ | |
419 | static bitmap_head *live_reload_and_inheritance_pseudos; | |
420 | static bitmap_obstack live_reload_and_inheritance_pseudos_bitmap_obstack; | |
421 | ||
422 | /* Allocate and initialize data about living reload pseudos at any | |
423 | given program point. */ | |
424 | static void | |
425 | init_live_reload_and_inheritance_pseudos (void) | |
426 | { | |
427 | int i, p, max_regno = max_reg_num (); | |
428 | lra_live_range_t r; | |
f4eafc30 | 429 | |
55a2c322 VM |
430 | conflict_reload_and_inheritance_pseudos = sparseset_alloc (max_regno); |
431 | live_reload_and_inheritance_pseudos = XNEWVEC (bitmap_head, lra_live_max_point); | |
432 | bitmap_obstack_initialize (&live_reload_and_inheritance_pseudos_bitmap_obstack); | |
433 | for (p = 0; p < lra_live_max_point; p++) | |
434 | bitmap_initialize (&live_reload_and_inheritance_pseudos[p], | |
435 | &live_reload_and_inheritance_pseudos_bitmap_obstack); | |
436 | for (i = lra_constraint_new_regno_start; i < max_regno; i++) | |
437 | { | |
438 | for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next) | |
439 | for (p = r->start; p <= r->finish; p++) | |
440 | bitmap_set_bit (&live_reload_and_inheritance_pseudos[p], i); | |
441 | } | |
442 | } | |
443 | ||
444 | /* Finalize data about living reload pseudos at any given program | |
445 | point. */ | |
446 | static void | |
447 | finish_live_reload_and_inheritance_pseudos (void) | |
448 | { | |
449 | sparseset_free (conflict_reload_and_inheritance_pseudos); | |
450 | free (live_reload_and_inheritance_pseudos); | |
451 | bitmap_obstack_release (&live_reload_and_inheritance_pseudos_bitmap_obstack); | |
452 | } | |
453 | ||
454 | /* The value used to check that cost of given hard reg is really | |
455 | defined currently. */ | |
456 | static int curr_hard_regno_costs_check = 0; | |
457 | /* Array used to check that cost of the corresponding hard reg (the | |
458 | array element index) is really defined currently. */ | |
459 | static int hard_regno_costs_check[FIRST_PSEUDO_REGISTER]; | |
460 | /* The current costs of allocation of hard regs. Defined only if the | |
461 | value of the corresponding element of the previous array is equal to | |
462 | CURR_HARD_REGNO_COSTS_CHECK. */ | |
463 | static int hard_regno_costs[FIRST_PSEUDO_REGISTER]; | |
464 | ||
465 | /* Adjust cost of HARD_REGNO by INCR. Reset the cost first if it is | |
466 | not defined yet. */ | |
467 | static inline void | |
468 | adjust_hard_regno_cost (int hard_regno, int incr) | |
469 | { | |
470 | if (hard_regno_costs_check[hard_regno] != curr_hard_regno_costs_check) | |
471 | hard_regno_costs[hard_regno] = 0; | |
472 | hard_regno_costs_check[hard_regno] = curr_hard_regno_costs_check; | |
473 | hard_regno_costs[hard_regno] += incr; | |
474 | } | |
475 | ||
476 | /* Try to find a free hard register for pseudo REGNO. Return the | |
477 | hard register on success and set *COST to the cost of using | |
478 | that register. (If several registers have equal cost, the one with | |
479 | the highest priority wins.) Return -1 on failure. | |
480 | ||
9e038952 VM |
481 | If FIRST_P, return the first available hard reg ignoring other |
482 | criteria, e.g. allocation cost. This approach results in less hard | |
483 | reg pool fragmentation and permit to allocate hard regs to reload | |
484 | pseudos in complicated situations where pseudo sizes are different. | |
485 | ||
55a2c322 | 486 | If TRY_ONLY_HARD_REGNO >= 0, consider only that hard register, |
34349d55 VM |
487 | otherwise consider all hard registers in REGNO's class. |
488 | ||
489 | If REGNO_SET is not empty, only hard registers from the set are | |
490 | considered. */ | |
55a2c322 | 491 | static int |
34349d55 VM |
492 | find_hard_regno_for_1 (int regno, int *cost, int try_only_hard_regno, |
493 | bool first_p, HARD_REG_SET regno_set) | |
55a2c322 VM |
494 | { |
495 | HARD_REG_SET conflict_set; | |
496 | int best_cost = INT_MAX, best_priority = INT_MIN, best_usage = INT_MAX; | |
497 | lra_live_range_t r; | |
498 | int p, i, j, rclass_size, best_hard_regno, priority, hard_regno; | |
499 | int hr, conflict_hr, nregs; | |
ef4bddc2 | 500 | machine_mode biggest_mode; |
55a2c322 | 501 | unsigned int k, conflict_regno; |
d70a81dd | 502 | int offset, val, biggest_nregs, nregs_diff; |
55a2c322 VM |
503 | enum reg_class rclass; |
504 | bitmap_iterator bi; | |
505 | bool *rclass_intersect_p; | |
a4971e68 | 506 | HARD_REG_SET impossible_start_hard_regs, available_regs; |
55a2c322 | 507 | |
34349d55 VM |
508 | if (hard_reg_set_empty_p (regno_set)) |
509 | COPY_HARD_REG_SET (conflict_set, lra_no_alloc_regs); | |
510 | else | |
511 | { | |
512 | COMPL_HARD_REG_SET (conflict_set, regno_set); | |
513 | IOR_HARD_REG_SET (conflict_set, lra_no_alloc_regs); | |
514 | } | |
55a2c322 VM |
515 | rclass = regno_allocno_class_array[regno]; |
516 | rclass_intersect_p = ira_reg_classes_intersect_p[rclass]; | |
517 | curr_hard_regno_costs_check++; | |
518 | sparseset_clear (conflict_reload_and_inheritance_pseudos); | |
519 | sparseset_clear (live_range_hard_reg_pseudos); | |
520 | IOR_HARD_REG_SET (conflict_set, lra_reg_info[regno].conflict_hard_regs); | |
521 | biggest_mode = lra_reg_info[regno].biggest_mode; | |
522 | for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next) | |
523 | { | |
524 | EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[r->start], 0, k, bi) | |
525 | if (rclass_intersect_p[regno_allocno_class_array[k]]) | |
526 | sparseset_set_bit (live_range_hard_reg_pseudos, k); | |
527 | EXECUTE_IF_SET_IN_BITMAP (&live_reload_and_inheritance_pseudos[r->start], | |
528 | 0, k, bi) | |
529 | if (lra_reg_info[k].preferred_hard_regno1 >= 0 | |
530 | && live_pseudos_reg_renumber[k] < 0 | |
531 | && rclass_intersect_p[regno_allocno_class_array[k]]) | |
532 | sparseset_set_bit (conflict_reload_and_inheritance_pseudos, k); | |
533 | for (p = r->start + 1; p <= r->finish; p++) | |
534 | { | |
535 | lra_live_range_t r2; | |
f4eafc30 | 536 | |
55a2c322 VM |
537 | for (r2 = start_point_ranges[p]; |
538 | r2 != NULL; | |
539 | r2 = r2->start_next) | |
540 | { | |
541 | if (r2->regno >= lra_constraint_new_regno_start | |
542 | && lra_reg_info[r2->regno].preferred_hard_regno1 >= 0 | |
543 | && live_pseudos_reg_renumber[r2->regno] < 0 | |
544 | && rclass_intersect_p[regno_allocno_class_array[r2->regno]]) | |
545 | sparseset_set_bit (conflict_reload_and_inheritance_pseudos, | |
546 | r2->regno); | |
547 | if (live_pseudos_reg_renumber[r2->regno] >= 0 | |
548 | && rclass_intersect_p[regno_allocno_class_array[r2->regno]]) | |
549 | sparseset_set_bit (live_range_hard_reg_pseudos, r2->regno); | |
550 | } | |
551 | } | |
552 | } | |
553 | if ((hard_regno = lra_reg_info[regno].preferred_hard_regno1) >= 0) | |
554 | { | |
555 | adjust_hard_regno_cost | |
556 | (hard_regno, -lra_reg_info[regno].preferred_hard_regno_profit1); | |
557 | if ((hard_regno = lra_reg_info[regno].preferred_hard_regno2) >= 0) | |
558 | adjust_hard_regno_cost | |
559 | (hard_regno, -lra_reg_info[regno].preferred_hard_regno_profit2); | |
560 | } | |
561 | #ifdef STACK_REGS | |
562 | if (lra_reg_info[regno].no_stack_p) | |
563 | for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++) | |
564 | SET_HARD_REG_BIT (conflict_set, i); | |
565 | #endif | |
566 | sparseset_clear_bit (conflict_reload_and_inheritance_pseudos, regno); | |
567 | val = lra_reg_info[regno].val; | |
d70a81dd | 568 | offset = lra_reg_info[regno].offset; |
55a2c322 VM |
569 | CLEAR_HARD_REG_SET (impossible_start_hard_regs); |
570 | EXECUTE_IF_SET_IN_SPARSESET (live_range_hard_reg_pseudos, conflict_regno) | |
d70a81dd | 571 | if (lra_reg_val_equal_p (conflict_regno, val, offset)) |
55a2c322 VM |
572 | { |
573 | conflict_hr = live_pseudos_reg_renumber[conflict_regno]; | |
574 | nregs = (hard_regno_nregs[conflict_hr] | |
575 | [lra_reg_info[conflict_regno].biggest_mode]); | |
576 | /* Remember about multi-register pseudos. For example, 2 hard | |
577 | register pseudos can start on the same hard register but can | |
f4eafc30 | 578 | not start on HR and HR+1/HR-1. */ |
55a2c322 VM |
579 | for (hr = conflict_hr + 1; |
580 | hr < FIRST_PSEUDO_REGISTER && hr < conflict_hr + nregs; | |
581 | hr++) | |
582 | SET_HARD_REG_BIT (impossible_start_hard_regs, hr); | |
583 | for (hr = conflict_hr - 1; | |
584 | hr >= 0 && hr + hard_regno_nregs[hr][biggest_mode] > conflict_hr; | |
585 | hr--) | |
586 | SET_HARD_REG_BIT (impossible_start_hard_regs, hr); | |
587 | } | |
588 | else | |
589 | { | |
590 | add_to_hard_reg_set (&conflict_set, | |
591 | lra_reg_info[conflict_regno].biggest_mode, | |
592 | live_pseudos_reg_renumber[conflict_regno]); | |
593 | if (hard_reg_set_subset_p (reg_class_contents[rclass], | |
594 | conflict_set)) | |
595 | return -1; | |
596 | } | |
597 | EXECUTE_IF_SET_IN_SPARSESET (conflict_reload_and_inheritance_pseudos, | |
598 | conflict_regno) | |
d70a81dd | 599 | if (!lra_reg_val_equal_p (conflict_regno, val, offset)) |
55a2c322 VM |
600 | { |
601 | lra_assert (live_pseudos_reg_renumber[conflict_regno] < 0); | |
602 | if ((hard_regno | |
603 | = lra_reg_info[conflict_regno].preferred_hard_regno1) >= 0) | |
604 | { | |
605 | adjust_hard_regno_cost | |
606 | (hard_regno, | |
607 | lra_reg_info[conflict_regno].preferred_hard_regno_profit1); | |
608 | if ((hard_regno | |
609 | = lra_reg_info[conflict_regno].preferred_hard_regno2) >= 0) | |
610 | adjust_hard_regno_cost | |
611 | (hard_regno, | |
612 | lra_reg_info[conflict_regno].preferred_hard_regno_profit2); | |
613 | } | |
614 | } | |
615 | /* Make sure that all registers in a multi-word pseudo belong to the | |
616 | required class. */ | |
617 | IOR_COMPL_HARD_REG_SET (conflict_set, reg_class_contents[rclass]); | |
618 | lra_assert (rclass != NO_REGS); | |
619 | rclass_size = ira_class_hard_regs_num[rclass]; | |
620 | best_hard_regno = -1; | |
621 | hard_regno = ira_class_hard_regs[rclass][0]; | |
622 | biggest_nregs = hard_regno_nregs[hard_regno][biggest_mode]; | |
623 | nregs_diff = (biggest_nregs | |
624 | - hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (regno)]); | |
a4971e68 VM |
625 | COPY_HARD_REG_SET (available_regs, reg_class_contents[rclass]); |
626 | AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs); | |
55a2c322 VM |
627 | for (i = 0; i < rclass_size; i++) |
628 | { | |
629 | if (try_only_hard_regno >= 0) | |
630 | hard_regno = try_only_hard_regno; | |
631 | else | |
632 | hard_regno = ira_class_hard_regs[rclass][i]; | |
633 | if (! overlaps_hard_reg_set_p (conflict_set, | |
634 | PSEUDO_REGNO_MODE (regno), hard_regno) | |
635 | /* We can not use prohibited_class_mode_regs because it is | |
636 | not defined for all classes. */ | |
637 | && HARD_REGNO_MODE_OK (hard_regno, PSEUDO_REGNO_MODE (regno)) | |
638 | && ! TEST_HARD_REG_BIT (impossible_start_hard_regs, hard_regno) | |
639 | && (nregs_diff == 0 | |
a1b46e46 JR |
640 | || (WORDS_BIG_ENDIAN |
641 | ? (hard_regno - nregs_diff >= 0 | |
a4971e68 | 642 | && TEST_HARD_REG_BIT (available_regs, |
a1b46e46 | 643 | hard_regno - nregs_diff)) |
a4971e68 | 644 | : TEST_HARD_REG_BIT (available_regs, |
a1b46e46 | 645 | hard_regno + nregs_diff)))) |
55a2c322 VM |
646 | { |
647 | if (hard_regno_costs_check[hard_regno] | |
648 | != curr_hard_regno_costs_check) | |
649 | { | |
650 | hard_regno_costs_check[hard_regno] = curr_hard_regno_costs_check; | |
651 | hard_regno_costs[hard_regno] = 0; | |
652 | } | |
653 | for (j = 0; | |
654 | j < hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (regno)]; | |
655 | j++) | |
656 | if (! TEST_HARD_REG_BIT (call_used_reg_set, hard_regno + j) | |
657 | && ! df_regs_ever_live_p (hard_regno + j)) | |
658 | /* It needs save restore. */ | |
659 | hard_regno_costs[hard_regno] | |
fef37404 VM |
660 | += (2 |
661 | * REG_FREQ_FROM_BB (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb) | |
662 | + 1); | |
55a2c322 VM |
663 | priority = targetm.register_priority (hard_regno); |
664 | if (best_hard_regno < 0 || hard_regno_costs[hard_regno] < best_cost | |
665 | || (hard_regno_costs[hard_regno] == best_cost | |
666 | && (priority > best_priority | |
3b9ceb4b | 667 | || (targetm.register_usage_leveling_p () |
55a2c322 VM |
668 | && priority == best_priority |
669 | && best_usage > lra_hard_reg_usage[hard_regno])))) | |
670 | { | |
671 | best_hard_regno = hard_regno; | |
672 | best_cost = hard_regno_costs[hard_regno]; | |
673 | best_priority = priority; | |
674 | best_usage = lra_hard_reg_usage[hard_regno]; | |
675 | } | |
676 | } | |
9e038952 | 677 | if (try_only_hard_regno >= 0 || (first_p && best_hard_regno >= 0)) |
55a2c322 VM |
678 | break; |
679 | } | |
680 | if (best_hard_regno >= 0) | |
681 | *cost = best_cost - lra_reg_info[regno].freq; | |
682 | return best_hard_regno; | |
683 | } | |
684 | ||
34349d55 VM |
685 | /* A wrapper for find_hard_regno_for_1 (see comments for that function |
686 | description). This function tries to find a hard register for | |
687 | preferred class first if it is worth. */ | |
688 | static int | |
689 | find_hard_regno_for (int regno, int *cost, int try_only_hard_regno, bool first_p) | |
690 | { | |
691 | int hard_regno; | |
692 | HARD_REG_SET regno_set; | |
693 | ||
694 | /* Only original pseudos can have a different preferred class. */ | |
695 | if (try_only_hard_regno < 0 && regno < lra_new_regno_start) | |
696 | { | |
697 | enum reg_class pref_class = reg_preferred_class (regno); | |
698 | ||
699 | if (regno_allocno_class_array[regno] != pref_class) | |
700 | { | |
701 | hard_regno = find_hard_regno_for_1 (regno, cost, -1, first_p, | |
702 | reg_class_contents[pref_class]); | |
703 | if (hard_regno >= 0) | |
704 | return hard_regno; | |
705 | } | |
706 | } | |
707 | CLEAR_HARD_REG_SET (regno_set); | |
708 | return find_hard_regno_for_1 (regno, cost, try_only_hard_regno, first_p, | |
709 | regno_set); | |
710 | } | |
711 | ||
55a2c322 VM |
712 | /* Current value used for checking elements in |
713 | update_hard_regno_preference_check. */ | |
714 | static int curr_update_hard_regno_preference_check; | |
715 | /* If an element value is equal to the above variable value, then the | |
716 | corresponding regno has been processed for preference | |
717 | propagation. */ | |
718 | static int *update_hard_regno_preference_check; | |
719 | ||
720 | /* Update the preference for using HARD_REGNO for pseudos that are | |
721 | connected directly or indirectly with REGNO. Apply divisor DIV | |
722 | to any preference adjustments. | |
723 | ||
724 | The more indirectly a pseudo is connected, the smaller its effect | |
725 | should be. We therefore increase DIV on each "hop". */ | |
726 | static void | |
727 | update_hard_regno_preference (int regno, int hard_regno, int div) | |
728 | { | |
729 | int another_regno, cost; | |
730 | lra_copy_t cp, next_cp; | |
731 | ||
732 | /* Search depth 5 seems to be enough. */ | |
733 | if (div > (1 << 5)) | |
734 | return; | |
735 | for (cp = lra_reg_info[regno].copies; cp != NULL; cp = next_cp) | |
736 | { | |
737 | if (cp->regno1 == regno) | |
738 | { | |
739 | next_cp = cp->regno1_next; | |
740 | another_regno = cp->regno2; | |
741 | } | |
742 | else if (cp->regno2 == regno) | |
743 | { | |
744 | next_cp = cp->regno2_next; | |
745 | another_regno = cp->regno1; | |
746 | } | |
747 | else | |
748 | gcc_unreachable (); | |
749 | if (reg_renumber[another_regno] < 0 | |
750 | && (update_hard_regno_preference_check[another_regno] | |
751 | != curr_update_hard_regno_preference_check)) | |
752 | { | |
753 | update_hard_regno_preference_check[another_regno] | |
754 | = curr_update_hard_regno_preference_check; | |
755 | cost = cp->freq < div ? 1 : cp->freq / div; | |
756 | lra_setup_reload_pseudo_preferenced_hard_reg | |
757 | (another_regno, hard_regno, cost); | |
758 | update_hard_regno_preference (another_regno, hard_regno, div * 2); | |
759 | } | |
760 | } | |
761 | } | |
762 | ||
2b778c9d VM |
763 | /* Return prefix title for pseudo REGNO. */ |
764 | static const char * | |
765 | pseudo_prefix_title (int regno) | |
766 | { | |
767 | return | |
768 | (regno < lra_constraint_new_regno_start ? "" | |
769 | : bitmap_bit_p (&lra_inheritance_pseudos, regno) ? "inheritance " | |
770 | : bitmap_bit_p (&lra_split_regs, regno) ? "split " | |
771 | : bitmap_bit_p (&lra_optional_reload_pseudos, regno) ? "optional reload " | |
772 | : bitmap_bit_p (&lra_subreg_reload_pseudos, regno) ? "subreg reload " | |
773 | : "reload "); | |
774 | } | |
775 | ||
55a2c322 VM |
776 | /* Update REG_RENUMBER and other pseudo preferences by assignment of |
777 | HARD_REGNO to pseudo REGNO and print about it if PRINT_P. */ | |
778 | void | |
779 | lra_setup_reg_renumber (int regno, int hard_regno, bool print_p) | |
780 | { | |
781 | int i, hr; | |
782 | ||
783 | /* We can not just reassign hard register. */ | |
784 | lra_assert (hard_regno < 0 || reg_renumber[regno] < 0); | |
785 | if ((hr = hard_regno) < 0) | |
786 | hr = reg_renumber[regno]; | |
787 | reg_renumber[regno] = hard_regno; | |
788 | lra_assert (hr >= 0); | |
789 | for (i = 0; i < hard_regno_nregs[hr][PSEUDO_REGNO_MODE (regno)]; i++) | |
790 | if (hard_regno < 0) | |
791 | lra_hard_reg_usage[hr + i] -= lra_reg_info[regno].freq; | |
792 | else | |
793 | lra_hard_reg_usage[hr + i] += lra_reg_info[regno].freq; | |
794 | if (print_p && lra_dump_file != NULL) | |
795 | fprintf (lra_dump_file, " Assign %d to %sr%d (freq=%d)\n", | |
2b778c9d | 796 | reg_renumber[regno], pseudo_prefix_title (regno), |
55a2c322 VM |
797 | regno, lra_reg_info[regno].freq); |
798 | if (hard_regno >= 0) | |
799 | { | |
800 | curr_update_hard_regno_preference_check++; | |
801 | update_hard_regno_preference (regno, hard_regno, 1); | |
802 | } | |
803 | } | |
804 | ||
805 | /* Pseudos which occur in insns containing a particular pseudo. */ | |
806 | static bitmap_head insn_conflict_pseudos; | |
807 | ||
808 | /* Bitmaps used to contain spill pseudos for given pseudo hard regno | |
809 | and best spill pseudos for given pseudo (and best hard regno). */ | |
810 | static bitmap_head spill_pseudos_bitmap, best_spill_pseudos_bitmap; | |
811 | ||
812 | /* Current pseudo check for validity of elements in | |
813 | TRY_HARD_REG_PSEUDOS. */ | |
814 | static int curr_pseudo_check; | |
815 | /* Array used for validity of elements in TRY_HARD_REG_PSEUDOS. */ | |
816 | static int try_hard_reg_pseudos_check[FIRST_PSEUDO_REGISTER]; | |
817 | /* Pseudos who hold given hard register at the considered points. */ | |
818 | static bitmap_head try_hard_reg_pseudos[FIRST_PSEUDO_REGISTER]; | |
819 | ||
820 | /* Set up try_hard_reg_pseudos for given program point P and class | |
821 | RCLASS. Those are pseudos living at P and assigned to a hard | |
822 | register of RCLASS. In other words, those are pseudos which can be | |
823 | spilled to assign a hard register of RCLASS to a pseudo living at | |
824 | P. */ | |
825 | static void | |
826 | setup_try_hard_regno_pseudos (int p, enum reg_class rclass) | |
827 | { | |
828 | int i, hard_regno; | |
ef4bddc2 | 829 | machine_mode mode; |
55a2c322 VM |
830 | unsigned int spill_regno; |
831 | bitmap_iterator bi; | |
832 | ||
833 | /* Find what pseudos could be spilled. */ | |
834 | EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[p], 0, spill_regno, bi) | |
835 | { | |
836 | mode = PSEUDO_REGNO_MODE (spill_regno); | |
837 | hard_regno = live_pseudos_reg_renumber[spill_regno]; | |
838 | if (overlaps_hard_reg_set_p (reg_class_contents[rclass], | |
839 | mode, hard_regno)) | |
840 | { | |
841 | for (i = hard_regno_nregs[hard_regno][mode] - 1; i >= 0; i--) | |
842 | { | |
843 | if (try_hard_reg_pseudos_check[hard_regno + i] | |
844 | != curr_pseudo_check) | |
845 | { | |
846 | try_hard_reg_pseudos_check[hard_regno + i] | |
847 | = curr_pseudo_check; | |
848 | bitmap_clear (&try_hard_reg_pseudos[hard_regno + i]); | |
849 | } | |
850 | bitmap_set_bit (&try_hard_reg_pseudos[hard_regno + i], | |
851 | spill_regno); | |
852 | } | |
853 | } | |
854 | } | |
855 | } | |
856 | ||
857 | /* Assign temporarily HARD_REGNO to pseudo REGNO. Temporary | |
858 | assignment means that we might undo the data change. */ | |
859 | static void | |
860 | assign_temporarily (int regno, int hard_regno) | |
861 | { | |
862 | int p; | |
863 | lra_live_range_t r; | |
864 | ||
865 | for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next) | |
866 | { | |
867 | for (p = r->start; p <= r->finish; p++) | |
868 | if (hard_regno < 0) | |
869 | bitmap_clear_bit (&live_hard_reg_pseudos[p], regno); | |
870 | else | |
871 | { | |
872 | bitmap_set_bit (&live_hard_reg_pseudos[p], regno); | |
873 | insert_in_live_range_start_chain (regno); | |
874 | } | |
875 | } | |
876 | live_pseudos_reg_renumber[regno] = hard_regno; | |
877 | } | |
878 | ||
879 | /* Array used for sorting reload pseudos for subsequent allocation | |
880 | after spilling some pseudo. */ | |
881 | static int *sorted_reload_pseudos; | |
882 | ||
883 | /* Spill some pseudos for a reload pseudo REGNO and return hard | |
884 | register which should be used for pseudo after spilling. The | |
885 | function adds spilled pseudos to SPILLED_PSEUDO_BITMAP. When we | |
886 | choose hard register (and pseudos occupying the hard registers and | |
887 | to be spilled), we take into account not only how REGNO will | |
888 | benefit from the spills but also how other reload pseudos not yet | |
889 | assigned to hard registers benefit from the spills too. In very | |
9e038952 VM |
890 | rare cases, the function can fail and return -1. |
891 | ||
892 | If FIRST_P, return the first available hard reg ignoring other | |
893 | criteria, e.g. allocation cost and cost of spilling non-reload | |
894 | pseudos. This approach results in less hard reg pool fragmentation | |
895 | and permit to allocate hard regs to reload pseudos in complicated | |
896 | situations where pseudo sizes are different. */ | |
55a2c322 | 897 | static int |
9e038952 | 898 | spill_for (int regno, bitmap spilled_pseudo_bitmap, bool first_p) |
55a2c322 VM |
899 | { |
900 | int i, j, n, p, hard_regno, best_hard_regno, cost, best_cost, rclass_size; | |
901 | int reload_hard_regno, reload_cost; | |
b81a2f0d | 902 | bool static_p, best_static_p; |
ef4bddc2 | 903 | machine_mode mode; |
55a2c322 | 904 | enum reg_class rclass; |
55a2c322 VM |
905 | unsigned int spill_regno, reload_regno, uid; |
906 | int insn_pseudos_num, best_insn_pseudos_num; | |
8fd827b8 | 907 | int bad_spills_num, smallest_bad_spills_num; |
55a2c322 VM |
908 | lra_live_range_t r; |
909 | bitmap_iterator bi; | |
910 | ||
911 | rclass = regno_allocno_class_array[regno]; | |
912 | lra_assert (reg_renumber[regno] < 0 && rclass != NO_REGS); | |
913 | bitmap_clear (&insn_conflict_pseudos); | |
914 | bitmap_clear (&best_spill_pseudos_bitmap); | |
915 | EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi) | |
916 | { | |
917 | struct lra_insn_reg *ir; | |
f4eafc30 | 918 | |
55a2c322 VM |
919 | for (ir = lra_get_insn_regs (uid); ir != NULL; ir = ir->next) |
920 | if (ir->regno >= FIRST_PSEUDO_REGISTER) | |
921 | bitmap_set_bit (&insn_conflict_pseudos, ir->regno); | |
922 | } | |
923 | best_hard_regno = -1; | |
924 | best_cost = INT_MAX; | |
b81a2f0d | 925 | best_static_p = TRUE; |
55a2c322 | 926 | best_insn_pseudos_num = INT_MAX; |
8fd827b8 | 927 | smallest_bad_spills_num = INT_MAX; |
55a2c322 VM |
928 | rclass_size = ira_class_hard_regs_num[rclass]; |
929 | mode = PSEUDO_REGNO_MODE (regno); | |
930 | /* Invalidate try_hard_reg_pseudos elements. */ | |
931 | curr_pseudo_check++; | |
932 | for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next) | |
933 | for (p = r->start; p <= r->finish; p++) | |
934 | setup_try_hard_regno_pseudos (p, rclass); | |
935 | for (i = 0; i < rclass_size; i++) | |
936 | { | |
937 | hard_regno = ira_class_hard_regs[rclass][i]; | |
938 | bitmap_clear (&spill_pseudos_bitmap); | |
939 | for (j = hard_regno_nregs[hard_regno][mode] - 1; j >= 0; j--) | |
940 | { | |
941 | if (try_hard_reg_pseudos_check[hard_regno + j] != curr_pseudo_check) | |
942 | continue; | |
943 | lra_assert (!bitmap_empty_p (&try_hard_reg_pseudos[hard_regno + j])); | |
944 | bitmap_ior_into (&spill_pseudos_bitmap, | |
945 | &try_hard_reg_pseudos[hard_regno + j]); | |
946 | } | |
947 | /* Spill pseudos. */ | |
b81a2f0d | 948 | static_p = false; |
55a2c322 | 949 | EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi) |
bcb21886 KY |
950 | if ((pic_offset_table_rtx != NULL |
951 | && spill_regno == REGNO (pic_offset_table_rtx)) | |
952 | || ((int) spill_regno >= lra_constraint_new_regno_start | |
953 | && ! bitmap_bit_p (&lra_inheritance_pseudos, spill_regno) | |
954 | && ! bitmap_bit_p (&lra_split_regs, spill_regno) | |
955 | && ! bitmap_bit_p (&lra_subreg_reload_pseudos, spill_regno) | |
956 | && ! bitmap_bit_p (&lra_optional_reload_pseudos, spill_regno))) | |
55a2c322 | 957 | goto fail; |
b81a2f0d VM |
958 | else if (non_spilled_static_chain_regno_p (spill_regno)) |
959 | static_p = true; | |
55a2c322 | 960 | insn_pseudos_num = 0; |
8fd827b8 | 961 | bad_spills_num = 0; |
55a2c322 VM |
962 | if (lra_dump_file != NULL) |
963 | fprintf (lra_dump_file, " Trying %d:", hard_regno); | |
964 | sparseset_clear (live_range_reload_inheritance_pseudos); | |
965 | EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi) | |
966 | { | |
967 | if (bitmap_bit_p (&insn_conflict_pseudos, spill_regno)) | |
968 | insn_pseudos_num++; | |
8fd827b8 VM |
969 | if (spill_regno >= (unsigned int) lra_bad_spill_regno_start) |
970 | bad_spills_num++; | |
55a2c322 VM |
971 | for (r = lra_reg_info[spill_regno].live_ranges; |
972 | r != NULL; | |
973 | r = r->next) | |
974 | { | |
975 | for (p = r->start; p <= r->finish; p++) | |
976 | { | |
977 | lra_live_range_t r2; | |
f4eafc30 | 978 | |
55a2c322 VM |
979 | for (r2 = start_point_ranges[p]; |
980 | r2 != NULL; | |
981 | r2 = r2->start_next) | |
982 | if (r2->regno >= lra_constraint_new_regno_start) | |
983 | sparseset_set_bit (live_range_reload_inheritance_pseudos, | |
984 | r2->regno); | |
985 | } | |
986 | } | |
987 | } | |
295d875c | 988 | n = 0; |
88def637 | 989 | if (sparseset_cardinality (live_range_reload_inheritance_pseudos) |
bb750f4f | 990 | <= (unsigned)LRA_MAX_CONSIDERED_RELOAD_PSEUDOS) |
88def637 VM |
991 | EXECUTE_IF_SET_IN_SPARSESET (live_range_reload_inheritance_pseudos, |
992 | reload_regno) | |
993 | if ((int) reload_regno != regno | |
994 | && (ira_reg_classes_intersect_p | |
995 | [rclass][regno_allocno_class_array[reload_regno]]) | |
996 | && live_pseudos_reg_renumber[reload_regno] < 0 | |
9e038952 | 997 | && find_hard_regno_for (reload_regno, &cost, -1, first_p) < 0) |
88def637 | 998 | sorted_reload_pseudos[n++] = reload_regno; |
295d875c VM |
999 | EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi) |
1000 | { | |
1001 | update_lives (spill_regno, true); | |
1002 | if (lra_dump_file != NULL) | |
1003 | fprintf (lra_dump_file, " spill %d(freq=%d)", | |
1004 | spill_regno, lra_reg_info[spill_regno].freq); | |
1005 | } | |
9e038952 | 1006 | hard_regno = find_hard_regno_for (regno, &cost, -1, first_p); |
55a2c322 VM |
1007 | if (hard_regno >= 0) |
1008 | { | |
1009 | assign_temporarily (regno, hard_regno); | |
55a2c322 VM |
1010 | qsort (sorted_reload_pseudos, n, sizeof (int), |
1011 | reload_pseudo_compare_func); | |
1012 | for (j = 0; j < n; j++) | |
1013 | { | |
1014 | reload_regno = sorted_reload_pseudos[j]; | |
1015 | lra_assert (live_pseudos_reg_renumber[reload_regno] < 0); | |
1016 | if ((reload_hard_regno | |
1017 | = find_hard_regno_for (reload_regno, | |
9e038952 | 1018 | &reload_cost, -1, first_p)) >= 0) |
55a2c322 VM |
1019 | { |
1020 | if (lra_dump_file != NULL) | |
1021 | fprintf (lra_dump_file, " assign %d(cost=%d)", | |
1022 | reload_regno, reload_cost); | |
1023 | assign_temporarily (reload_regno, reload_hard_regno); | |
1024 | cost += reload_cost; | |
1025 | } | |
1026 | } | |
1027 | EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi) | |
1028 | { | |
0cc97fc5 | 1029 | rtx_insn_list *x; |
f4eafc30 | 1030 | |
55a2c322 VM |
1031 | cost += lra_reg_info[spill_regno].freq; |
1032 | if (ira_reg_equiv[spill_regno].memory != NULL | |
1033 | || ira_reg_equiv[spill_regno].constant != NULL) | |
1034 | for (x = ira_reg_equiv[spill_regno].init_insns; | |
1035 | x != NULL; | |
0cc97fc5 DM |
1036 | x = x->next ()) |
1037 | cost -= REG_FREQ_FROM_BB (BLOCK_FOR_INSN (x->insn ())); | |
55a2c322 | 1038 | } |
b81a2f0d VM |
1039 | /* Avoid spilling static chain pointer pseudo when non-local |
1040 | goto is used. */ | |
1041 | if ((! static_p && best_static_p) | |
1042 | || (static_p == best_static_p | |
1043 | && (best_insn_pseudos_num > insn_pseudos_num | |
1044 | || (best_insn_pseudos_num == insn_pseudos_num | |
1045 | && (bad_spills_num < smallest_bad_spills_num | |
1046 | || (bad_spills_num == smallest_bad_spills_num | |
1047 | && best_cost > cost)))))) | |
55a2c322 VM |
1048 | { |
1049 | best_insn_pseudos_num = insn_pseudos_num; | |
54e915b3 | 1050 | smallest_bad_spills_num = bad_spills_num; |
b81a2f0d | 1051 | best_static_p = static_p; |
55a2c322 VM |
1052 | best_cost = cost; |
1053 | best_hard_regno = hard_regno; | |
1054 | bitmap_copy (&best_spill_pseudos_bitmap, &spill_pseudos_bitmap); | |
1055 | if (lra_dump_file != NULL) | |
54e915b3 VM |
1056 | fprintf (lra_dump_file, |
1057 | " Now best %d(cost=%d, bad_spills=%d, insn_pseudos=%d)\n", | |
1058 | hard_regno, cost, bad_spills_num, insn_pseudos_num); | |
55a2c322 VM |
1059 | } |
1060 | assign_temporarily (regno, -1); | |
1061 | for (j = 0; j < n; j++) | |
1062 | { | |
1063 | reload_regno = sorted_reload_pseudos[j]; | |
1064 | if (live_pseudos_reg_renumber[reload_regno] >= 0) | |
1065 | assign_temporarily (reload_regno, -1); | |
1066 | } | |
1067 | } | |
1068 | if (lra_dump_file != NULL) | |
1069 | fprintf (lra_dump_file, "\n"); | |
1070 | /* Restore the live hard reg pseudo info for spilled pseudos. */ | |
1071 | EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi) | |
1072 | update_lives (spill_regno, false); | |
1073 | fail: | |
1074 | ; | |
1075 | } | |
1076 | /* Spill: */ | |
1077 | EXECUTE_IF_SET_IN_BITMAP (&best_spill_pseudos_bitmap, 0, spill_regno, bi) | |
1078 | { | |
f54437d5 VM |
1079 | if ((int) spill_regno >= lra_constraint_new_regno_start) |
1080 | former_reload_pseudo_spill_p = true; | |
55a2c322 VM |
1081 | if (lra_dump_file != NULL) |
1082 | fprintf (lra_dump_file, " Spill %sr%d(hr=%d, freq=%d) for r%d\n", | |
2b778c9d | 1083 | pseudo_prefix_title (spill_regno), |
55a2c322 VM |
1084 | spill_regno, reg_renumber[spill_regno], |
1085 | lra_reg_info[spill_regno].freq, regno); | |
1086 | update_lives (spill_regno, true); | |
1087 | lra_setup_reg_renumber (spill_regno, -1, false); | |
1088 | } | |
1089 | bitmap_ior_into (spilled_pseudo_bitmap, &best_spill_pseudos_bitmap); | |
1090 | return best_hard_regno; | |
1091 | } | |
1092 | ||
1093 | /* Assign HARD_REGNO to REGNO. */ | |
1094 | static void | |
1095 | assign_hard_regno (int hard_regno, int regno) | |
1096 | { | |
1097 | int i; | |
1098 | ||
1099 | lra_assert (hard_regno >= 0); | |
1100 | lra_setup_reg_renumber (regno, hard_regno, true); | |
1101 | update_lives (regno, false); | |
1102 | for (i = 0; | |
1103 | i < hard_regno_nregs[hard_regno][lra_reg_info[regno].biggest_mode]; | |
1104 | i++) | |
1105 | df_set_regs_ever_live (hard_regno + i, true); | |
1106 | } | |
1107 | ||
1108 | /* Array used for sorting different pseudos. */ | |
1109 | static int *sorted_pseudos; | |
1110 | ||
1111 | /* The constraints pass is allowed to create equivalences between | |
1112 | pseudos that make the current allocation "incorrect" (in the sense | |
1113 | that pseudos are assigned to hard registers from their own conflict | |
1114 | sets). The global variable lra_risky_transformations_p says | |
1115 | whether this might have happened. | |
1116 | ||
1117 | Process pseudos assigned to hard registers (less frequently used | |
1118 | first), spill if a conflict is found, and mark the spilled pseudos | |
1119 | in SPILLED_PSEUDO_BITMAP. Set up LIVE_HARD_REG_PSEUDOS from | |
1120 | pseudos, assigned to hard registers. */ | |
1121 | static void | |
1122 | setup_live_pseudos_and_spill_after_risky_transforms (bitmap | |
1123 | spilled_pseudo_bitmap) | |
1124 | { | |
1125 | int p, i, j, n, regno, hard_regno; | |
1126 | unsigned int k, conflict_regno; | |
d70a81dd | 1127 | int val, offset; |
55a2c322 | 1128 | HARD_REG_SET conflict_set; |
ef4bddc2 | 1129 | machine_mode mode; |
55a2c322 VM |
1130 | lra_live_range_t r; |
1131 | bitmap_iterator bi; | |
1132 | int max_regno = max_reg_num (); | |
1133 | ||
1134 | if (! lra_risky_transformations_p) | |
1135 | { | |
1136 | for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) | |
1137 | if (reg_renumber[i] >= 0 && lra_reg_info[i].nrefs > 0) | |
1138 | update_lives (i, false); | |
1139 | return; | |
1140 | } | |
1141 | for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) | |
bcb21886 KY |
1142 | if ((pic_offset_table_rtx == NULL_RTX |
1143 | || i != (int) REGNO (pic_offset_table_rtx)) | |
1144 | && reg_renumber[i] >= 0 && lra_reg_info[i].nrefs > 0) | |
55a2c322 VM |
1145 | sorted_pseudos[n++] = i; |
1146 | qsort (sorted_pseudos, n, sizeof (int), pseudo_compare_func); | |
bcb21886 KY |
1147 | if (pic_offset_table_rtx != NULL_RTX |
1148 | && (regno = REGNO (pic_offset_table_rtx)) >= FIRST_PSEUDO_REGISTER | |
1149 | && reg_renumber[regno] >= 0 && lra_reg_info[regno].nrefs > 0) | |
1150 | sorted_pseudos[n++] = regno; | |
55a2c322 VM |
1151 | for (i = n - 1; i >= 0; i--) |
1152 | { | |
1153 | regno = sorted_pseudos[i]; | |
1154 | hard_regno = reg_renumber[regno]; | |
1155 | lra_assert (hard_regno >= 0); | |
1156 | mode = lra_reg_info[regno].biggest_mode; | |
1157 | sparseset_clear (live_range_hard_reg_pseudos); | |
1158 | for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next) | |
1159 | { | |
1160 | EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[r->start], 0, k, bi) | |
1161 | sparseset_set_bit (live_range_hard_reg_pseudos, k); | |
1162 | for (p = r->start + 1; p <= r->finish; p++) | |
1163 | { | |
1164 | lra_live_range_t r2; | |
f4eafc30 | 1165 | |
55a2c322 VM |
1166 | for (r2 = start_point_ranges[p]; |
1167 | r2 != NULL; | |
1168 | r2 = r2->start_next) | |
1169 | if (live_pseudos_reg_renumber[r2->regno] >= 0) | |
1170 | sparseset_set_bit (live_range_hard_reg_pseudos, r2->regno); | |
1171 | } | |
1172 | } | |
1173 | COPY_HARD_REG_SET (conflict_set, lra_no_alloc_regs); | |
1174 | IOR_HARD_REG_SET (conflict_set, lra_reg_info[regno].conflict_hard_regs); | |
1175 | val = lra_reg_info[regno].val; | |
d70a81dd | 1176 | offset = lra_reg_info[regno].offset; |
55a2c322 | 1177 | EXECUTE_IF_SET_IN_SPARSESET (live_range_hard_reg_pseudos, conflict_regno) |
d70a81dd | 1178 | if (!lra_reg_val_equal_p (conflict_regno, val, offset) |
55a2c322 VM |
1179 | /* If it is multi-register pseudos they should start on |
1180 | the same hard register. */ | |
1181 | || hard_regno != reg_renumber[conflict_regno]) | |
1182 | add_to_hard_reg_set (&conflict_set, | |
1183 | lra_reg_info[conflict_regno].biggest_mode, | |
1184 | reg_renumber[conflict_regno]); | |
1185 | if (! overlaps_hard_reg_set_p (conflict_set, mode, hard_regno)) | |
1186 | { | |
1187 | update_lives (regno, false); | |
1188 | continue; | |
1189 | } | |
1190 | bitmap_set_bit (spilled_pseudo_bitmap, regno); | |
1191 | for (j = 0; | |
1192 | j < hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (regno)]; | |
1193 | j++) | |
1194 | lra_hard_reg_usage[hard_regno + j] -= lra_reg_info[regno].freq; | |
1195 | reg_renumber[regno] = -1; | |
f54437d5 VM |
1196 | if (regno >= lra_constraint_new_regno_start) |
1197 | former_reload_pseudo_spill_p = true; | |
55a2c322 VM |
1198 | if (lra_dump_file != NULL) |
1199 | fprintf (lra_dump_file, " Spill r%d after risky transformations\n", | |
1200 | regno); | |
1201 | } | |
1202 | } | |
1203 | ||
1204 | /* Improve allocation by assigning the same hard regno of inheritance | |
1205 | pseudos to the connected pseudos. We need this because inheritance | |
1206 | pseudos are allocated after reload pseudos in the thread and when | |
1207 | we assign a hard register to a reload pseudo we don't know yet that | |
1208 | the connected inheritance pseudos can get the same hard register. | |
1209 | Add pseudos with changed allocation to bitmap CHANGED_PSEUDOS. */ | |
1210 | static void | |
1211 | improve_inheritance (bitmap changed_pseudos) | |
1212 | { | |
1213 | unsigned int k; | |
1214 | int regno, another_regno, hard_regno, another_hard_regno, cost, i, n; | |
1215 | lra_copy_t cp, next_cp; | |
1216 | bitmap_iterator bi; | |
1217 | ||
8e3a4869 VM |
1218 | if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES) |
1219 | return; | |
55a2c322 VM |
1220 | n = 0; |
1221 | EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, k, bi) | |
1222 | if (reg_renumber[k] >= 0 && lra_reg_info[k].nrefs != 0) | |
1223 | sorted_pseudos[n++] = k; | |
1224 | qsort (sorted_pseudos, n, sizeof (int), pseudo_compare_func); | |
1225 | for (i = 0; i < n; i++) | |
1226 | { | |
1227 | regno = sorted_pseudos[i]; | |
1228 | hard_regno = reg_renumber[regno]; | |
1229 | lra_assert (hard_regno >= 0); | |
1230 | for (cp = lra_reg_info[regno].copies; cp != NULL; cp = next_cp) | |
1231 | { | |
1232 | if (cp->regno1 == regno) | |
1233 | { | |
1234 | next_cp = cp->regno1_next; | |
1235 | another_regno = cp->regno2; | |
1236 | } | |
1237 | else if (cp->regno2 == regno) | |
1238 | { | |
1239 | next_cp = cp->regno2_next; | |
1240 | another_regno = cp->regno1; | |
1241 | } | |
1242 | else | |
1243 | gcc_unreachable (); | |
1244 | /* Don't change reload pseudo allocation. It might have | |
1245 | this allocation for a purpose and changing it can result | |
1246 | in LRA cycling. */ | |
1247 | if ((another_regno < lra_constraint_new_regno_start | |
1248 | || bitmap_bit_p (&lra_inheritance_pseudos, another_regno)) | |
1249 | && (another_hard_regno = reg_renumber[another_regno]) >= 0 | |
1250 | && another_hard_regno != hard_regno) | |
1251 | { | |
1252 | if (lra_dump_file != NULL) | |
1253 | fprintf | |
1254 | (lra_dump_file, | |
1255 | " Improving inheritance for %d(%d) and %d(%d)...\n", | |
1256 | regno, hard_regno, another_regno, another_hard_regno); | |
1257 | update_lives (another_regno, true); | |
1258 | lra_setup_reg_renumber (another_regno, -1, false); | |
9e038952 VM |
1259 | if (hard_regno == find_hard_regno_for (another_regno, &cost, |
1260 | hard_regno, false)) | |
55a2c322 VM |
1261 | assign_hard_regno (hard_regno, another_regno); |
1262 | else | |
1263 | assign_hard_regno (another_hard_regno, another_regno); | |
1264 | bitmap_set_bit (changed_pseudos, another_regno); | |
1265 | } | |
1266 | } | |
1267 | } | |
1268 | } | |
1269 | ||
1270 | ||
1271 | /* Bitmap finally containing all pseudos spilled on this assignment | |
1272 | pass. */ | |
1273 | static bitmap_head all_spilled_pseudos; | |
1274 | /* All pseudos whose allocation was changed. */ | |
1275 | static bitmap_head changed_pseudo_bitmap; | |
1276 | ||
9e038952 VM |
1277 | |
1278 | /* Add to LIVE_RANGE_HARD_REG_PSEUDOS all pseudos conflicting with | |
1279 | REGNO and whose hard regs can be assigned to REGNO. */ | |
1280 | static void | |
1281 | find_all_spills_for (int regno) | |
1282 | { | |
1283 | int p; | |
1284 | lra_live_range_t r; | |
1285 | unsigned int k; | |
1286 | bitmap_iterator bi; | |
1287 | enum reg_class rclass; | |
1288 | bool *rclass_intersect_p; | |
1289 | ||
1290 | rclass = regno_allocno_class_array[regno]; | |
1291 | rclass_intersect_p = ira_reg_classes_intersect_p[rclass]; | |
1292 | for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next) | |
1293 | { | |
1294 | EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[r->start], 0, k, bi) | |
1295 | if (rclass_intersect_p[regno_allocno_class_array[k]]) | |
1296 | sparseset_set_bit (live_range_hard_reg_pseudos, k); | |
1297 | for (p = r->start + 1; p <= r->finish; p++) | |
1298 | { | |
1299 | lra_live_range_t r2; | |
1300 | ||
1301 | for (r2 = start_point_ranges[p]; | |
1302 | r2 != NULL; | |
1303 | r2 = r2->start_next) | |
1304 | { | |
1305 | if (live_pseudos_reg_renumber[r2->regno] >= 0 | |
1306 | && rclass_intersect_p[regno_allocno_class_array[r2->regno]]) | |
1307 | sparseset_set_bit (live_range_hard_reg_pseudos, r2->regno); | |
1308 | } | |
1309 | } | |
1310 | } | |
1311 | } | |
1312 | ||
55a2c322 VM |
1313 | /* Assign hard registers to reload pseudos and other pseudos. */ |
1314 | static void | |
1315 | assign_by_spills (void) | |
1316 | { | |
1317 | int i, n, nfails, iter, regno, hard_regno, cost, restore_regno; | |
cfa434f6 | 1318 | rtx_insn *insn; |
55a2c322 | 1319 | bitmap_head changed_insns, do_not_assign_nonreload_pseudos; |
9e038952 | 1320 | unsigned int u, conflict_regno; |
55a2c322 | 1321 | bitmap_iterator bi; |
992ca0f0 | 1322 | bool reload_p; |
55a2c322 VM |
1323 | int max_regno = max_reg_num (); |
1324 | ||
1325 | for (n = 0, i = lra_constraint_new_regno_start; i < max_regno; i++) | |
1326 | if (reg_renumber[i] < 0 && lra_reg_info[i].nrefs != 0 | |
1327 | && regno_allocno_class_array[i] != NO_REGS) | |
1328 | sorted_pseudos[n++] = i; | |
1329 | bitmap_initialize (&insn_conflict_pseudos, ®_obstack); | |
1330 | bitmap_initialize (&spill_pseudos_bitmap, ®_obstack); | |
1331 | bitmap_initialize (&best_spill_pseudos_bitmap, ®_obstack); | |
1332 | update_hard_regno_preference_check = XCNEWVEC (int, max_regno); | |
1333 | curr_update_hard_regno_preference_check = 0; | |
1334 | memset (try_hard_reg_pseudos_check, 0, sizeof (try_hard_reg_pseudos_check)); | |
1335 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
1336 | bitmap_initialize (&try_hard_reg_pseudos[i], ®_obstack); | |
1337 | curr_pseudo_check = 0; | |
1338 | bitmap_initialize (&changed_insns, ®_obstack); | |
1339 | bitmap_initialize (&non_reload_pseudos, ®_obstack); | |
1340 | bitmap_ior (&non_reload_pseudos, &lra_inheritance_pseudos, &lra_split_regs); | |
2b778c9d | 1341 | bitmap_ior_into (&non_reload_pseudos, &lra_subreg_reload_pseudos); |
55a2c322 VM |
1342 | bitmap_ior_into (&non_reload_pseudos, &lra_optional_reload_pseudos); |
1343 | for (iter = 0; iter <= 1; iter++) | |
1344 | { | |
1345 | qsort (sorted_pseudos, n, sizeof (int), reload_pseudo_compare_func); | |
1346 | nfails = 0; | |
1347 | for (i = 0; i < n; i++) | |
1348 | { | |
1349 | regno = sorted_pseudos[i]; | |
1350 | if (lra_dump_file != NULL) | |
1351 | fprintf (lra_dump_file, " Assigning to %d " | |
1352 | "(cl=%s, orig=%d, freq=%d, tfirst=%d, tfreq=%d)...\n", | |
1353 | regno, reg_class_names[regno_allocno_class_array[regno]], | |
1354 | ORIGINAL_REGNO (regno_reg_rtx[regno]), | |
1355 | lra_reg_info[regno].freq, regno_assign_info[regno].first, | |
1356 | regno_assign_info[regno_assign_info[regno].first].freq); | |
9e038952 | 1357 | hard_regno = find_hard_regno_for (regno, &cost, -1, iter == 1); |
992ca0f0 VM |
1358 | reload_p = ! bitmap_bit_p (&non_reload_pseudos, regno); |
1359 | if (hard_regno < 0 && reload_p) | |
9e038952 | 1360 | hard_regno = spill_for (regno, &all_spilled_pseudos, iter == 1); |
55a2c322 VM |
1361 | if (hard_regno < 0) |
1362 | { | |
992ca0f0 | 1363 | if (reload_p) |
55a2c322 VM |
1364 | sorted_pseudos[nfails++] = regno; |
1365 | } | |
1366 | else | |
1367 | { | |
1368 | /* This register might have been spilled by the previous | |
1369 | pass. Indicate that it is no longer spilled. */ | |
1370 | bitmap_clear_bit (&all_spilled_pseudos, regno); | |
1371 | assign_hard_regno (hard_regno, regno); | |
992ca0f0 VM |
1372 | if (! reload_p) |
1373 | /* As non-reload pseudo assignment is changed we | |
1374 | should reconsider insns referring for the | |
1375 | pseudo. */ | |
1376 | bitmap_set_bit (&changed_pseudo_bitmap, regno); | |
55a2c322 VM |
1377 | } |
1378 | } | |
1379 | if (nfails == 0) | |
1380 | break; | |
ce940020 VM |
1381 | if (iter > 0) |
1382 | { | |
bdf13188 EB |
1383 | /* We did not assign hard regs to reload pseudos after two iterations. |
1384 | Either it's an asm and something is wrong with the constraints, or | |
1385 | we have run out of spill registers; error out in either case. */ | |
327b20f5 | 1386 | bool asm_p = false; |
ce940020 VM |
1387 | bitmap_head failed_reload_insns; |
1388 | ||
1389 | bitmap_initialize (&failed_reload_insns, ®_obstack); | |
1390 | for (i = 0; i < nfails; i++) | |
c656b86b VM |
1391 | { |
1392 | regno = sorted_pseudos[i]; | |
1393 | bitmap_ior_into (&failed_reload_insns, | |
1394 | &lra_reg_info[regno].insn_bitmap); | |
1395 | /* Assign an arbitrary hard register of regno class to | |
bdf13188 | 1396 | avoid further trouble with this insn. */ |
c656b86b VM |
1397 | bitmap_clear_bit (&all_spilled_pseudos, regno); |
1398 | assign_hard_regno | |
1399 | (ira_class_hard_regs[regno_allocno_class_array[regno]][0], | |
1400 | regno); | |
1401 | } | |
ce940020 VM |
1402 | EXECUTE_IF_SET_IN_BITMAP (&failed_reload_insns, 0, u, bi) |
1403 | { | |
1404 | insn = lra_insn_recog_data[u]->insn; | |
1405 | if (asm_noperands (PATTERN (insn)) >= 0) | |
1406 | { | |
327b20f5 | 1407 | asm_p = true; |
ce940020 VM |
1408 | error_for_asm (insn, |
1409 | "%<asm%> operand has impossible constraints"); | |
e86c0101 SB |
1410 | /* Avoid further trouble with this insn. |
1411 | For asm goto, instead of fixing up all the edges | |
1412 | just clear the template and clear input operands | |
1413 | (asm goto doesn't have any output operands). */ | |
1414 | if (JUMP_P (insn)) | |
1415 | { | |
1416 | rtx asm_op = extract_asm_operands (PATTERN (insn)); | |
1417 | ASM_OPERANDS_TEMPLATE (asm_op) = ggc_strdup (""); | |
1418 | ASM_OPERANDS_INPUT_VEC (asm_op) = rtvec_alloc (0); | |
1419 | ASM_OPERANDS_INPUT_CONSTRAINT_VEC (asm_op) = rtvec_alloc (0); | |
1420 | lra_update_insn_regno_info (insn); | |
1421 | } | |
1422 | else | |
1423 | { | |
1424 | PATTERN (insn) = gen_rtx_USE (VOIDmode, const0_rtx); | |
1425 | lra_set_insn_deleted (insn); | |
1426 | } | |
ce940020 | 1427 | } |
327b20f5 | 1428 | else if (!asm_p) |
bdf13188 EB |
1429 | { |
1430 | error ("unable to find a register to spill"); | |
1431 | fatal_insn ("this is the insn:", insn); | |
1432 | } | |
ce940020 | 1433 | } |
ce940020 VM |
1434 | break; |
1435 | } | |
9e038952 VM |
1436 | /* This is a very rare event. We can not assign a hard register |
1437 | to reload pseudo because the hard register was assigned to | |
1438 | another reload pseudo on a previous assignment pass. For x86 | |
1439 | example, on the 1st pass we assigned CX (although another | |
1440 | hard register could be used for this) to reload pseudo in an | |
1441 | insn, on the 2nd pass we need CX (and only this) hard | |
1442 | register for a new reload pseudo in the same insn. Another | |
1443 | possible situation may occur in assigning to multi-regs | |
1444 | reload pseudos when hard regs pool is too fragmented even | |
1445 | after spilling non-reload pseudos. | |
1446 | ||
1447 | We should do something radical here to succeed. Here we | |
1448 | spill *all* conflicting pseudos and reassign them. */ | |
55a2c322 VM |
1449 | if (lra_dump_file != NULL) |
1450 | fprintf (lra_dump_file, " 2nd iter for reload pseudo assignments:\n"); | |
9e038952 | 1451 | sparseset_clear (live_range_hard_reg_pseudos); |
55a2c322 VM |
1452 | for (i = 0; i < nfails; i++) |
1453 | { | |
1454 | if (lra_dump_file != NULL) | |
1455 | fprintf (lra_dump_file, " Reload r%d assignment failure\n", | |
1456 | sorted_pseudos[i]); | |
9e038952 VM |
1457 | find_all_spills_for (sorted_pseudos[i]); |
1458 | } | |
1459 | EXECUTE_IF_SET_IN_SPARSESET (live_range_hard_reg_pseudos, conflict_regno) | |
1460 | { | |
1461 | if ((int) conflict_regno >= lra_constraint_new_regno_start) | |
f54437d5 VM |
1462 | { |
1463 | sorted_pseudos[nfails++] = conflict_regno; | |
1464 | former_reload_pseudo_spill_p = true; | |
1465 | } | |
9e038952 VM |
1466 | if (lra_dump_file != NULL) |
1467 | fprintf (lra_dump_file, " Spill %s r%d(hr=%d, freq=%d)\n", | |
1468 | pseudo_prefix_title (conflict_regno), conflict_regno, | |
1469 | reg_renumber[conflict_regno], | |
1470 | lra_reg_info[conflict_regno].freq); | |
1471 | update_lives (conflict_regno, true); | |
1472 | lra_setup_reg_renumber (conflict_regno, -1, false); | |
55a2c322 | 1473 | } |
55a2c322 VM |
1474 | n = nfails; |
1475 | } | |
1476 | improve_inheritance (&changed_pseudo_bitmap); | |
1477 | bitmap_clear (&non_reload_pseudos); | |
1478 | bitmap_clear (&changed_insns); | |
1479 | if (! lra_simple_p) | |
1480 | { | |
1481 | /* We should not assign to original pseudos of inheritance | |
1482 | pseudos or split pseudos if any its inheritance pseudo did | |
1483 | not get hard register or any its split pseudo was not split | |
1484 | because undo inheritance/split pass will extend live range of | |
1485 | such inheritance or split pseudos. */ | |
1486 | bitmap_initialize (&do_not_assign_nonreload_pseudos, ®_obstack); | |
1487 | EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, u, bi) | |
1488 | if ((restore_regno = lra_reg_info[u].restore_regno) >= 0 | |
1489 | && reg_renumber[u] < 0 | |
1490 | && bitmap_bit_p (&lra_inheritance_pseudos, u)) | |
1491 | bitmap_set_bit (&do_not_assign_nonreload_pseudos, restore_regno); | |
1492 | EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, u, bi) | |
1493 | if ((restore_regno = lra_reg_info[u].restore_regno) >= 0 | |
1494 | && reg_renumber[u] >= 0) | |
1495 | bitmap_set_bit (&do_not_assign_nonreload_pseudos, restore_regno); | |
1496 | for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) | |
1497 | if (((i < lra_constraint_new_regno_start | |
1498 | && ! bitmap_bit_p (&do_not_assign_nonreload_pseudos, i)) | |
1499 | || (bitmap_bit_p (&lra_inheritance_pseudos, i) | |
1500 | && lra_reg_info[i].restore_regno >= 0) | |
1501 | || (bitmap_bit_p (&lra_split_regs, i) | |
1502 | && lra_reg_info[i].restore_regno >= 0) | |
2b778c9d | 1503 | || bitmap_bit_p (&lra_subreg_reload_pseudos, i) |
55a2c322 VM |
1504 | || bitmap_bit_p (&lra_optional_reload_pseudos, i)) |
1505 | && reg_renumber[i] < 0 && lra_reg_info[i].nrefs != 0 | |
1506 | && regno_allocno_class_array[i] != NO_REGS) | |
1507 | sorted_pseudos[n++] = i; | |
1508 | bitmap_clear (&do_not_assign_nonreload_pseudos); | |
1509 | if (n != 0 && lra_dump_file != NULL) | |
1510 | fprintf (lra_dump_file, " Reassigning non-reload pseudos\n"); | |
1511 | qsort (sorted_pseudos, n, sizeof (int), pseudo_compare_func); | |
1512 | for (i = 0; i < n; i++) | |
1513 | { | |
1514 | regno = sorted_pseudos[i]; | |
9e038952 | 1515 | hard_regno = find_hard_regno_for (regno, &cost, -1, false); |
55a2c322 VM |
1516 | if (hard_regno >= 0) |
1517 | { | |
1518 | assign_hard_regno (hard_regno, regno); | |
992ca0f0 VM |
1519 | /* We change allocation for non-reload pseudo on this |
1520 | iteration -- mark the pseudo for invalidation of used | |
1521 | alternatives of insns containing the pseudo. */ | |
55a2c322 VM |
1522 | bitmap_set_bit (&changed_pseudo_bitmap, regno); |
1523 | } | |
9afb455c VM |
1524 | else |
1525 | { | |
1526 | enum reg_class rclass = lra_get_allocno_class (regno); | |
1527 | enum reg_class spill_class; | |
1528 | ||
1df2287f VM |
1529 | if (targetm.spill_class == NULL |
1530 | || lra_reg_info[regno].restore_regno < 0 | |
9afb455c VM |
1531 | || ! bitmap_bit_p (&lra_inheritance_pseudos, regno) |
1532 | || (spill_class | |
1533 | = ((enum reg_class) | |
1534 | targetm.spill_class | |
1535 | ((reg_class_t) rclass, | |
1536 | PSEUDO_REGNO_MODE (regno)))) == NO_REGS) | |
1537 | continue; | |
1538 | regno_allocno_class_array[regno] = spill_class; | |
1539 | hard_regno = find_hard_regno_for (regno, &cost, -1, false); | |
1540 | if (hard_regno < 0) | |
1541 | regno_allocno_class_array[regno] = rclass; | |
1542 | else | |
1543 | { | |
1544 | setup_reg_classes | |
1545 | (regno, spill_class, spill_class, spill_class); | |
1546 | assign_hard_regno (hard_regno, regno); | |
1547 | bitmap_set_bit (&changed_pseudo_bitmap, regno); | |
1548 | } | |
1549 | } | |
55a2c322 VM |
1550 | } |
1551 | } | |
1552 | free (update_hard_regno_preference_check); | |
1553 | bitmap_clear (&best_spill_pseudos_bitmap); | |
1554 | bitmap_clear (&spill_pseudos_bitmap); | |
1555 | bitmap_clear (&insn_conflict_pseudos); | |
1556 | } | |
1557 | ||
1558 | ||
1559 | /* Entry function to assign hard registers to new reload pseudos | |
1560 | starting with LRA_CONSTRAINT_NEW_REGNO_START (by possible spilling | |
1561 | of old pseudos) and possibly to the old pseudos. The function adds | |
1562 | what insns to process for the next constraint pass. Those are all | |
1563 | insns who contains non-reload and non-inheritance pseudos with | |
1564 | changed allocation. | |
1565 | ||
1566 | Return true if we did not spill any non-reload and non-inheritance | |
1567 | pseudos. */ | |
1568 | bool | |
1569 | lra_assign (void) | |
1570 | { | |
1571 | int i; | |
1572 | unsigned int u; | |
1573 | bitmap_iterator bi; | |
1574 | bitmap_head insns_to_process; | |
1575 | bool no_spills_p; | |
1576 | int max_regno = max_reg_num (); | |
1577 | ||
1578 | timevar_push (TV_LRA_ASSIGN); | |
f54437d5 VM |
1579 | lra_assignment_iter++; |
1580 | if (lra_dump_file != NULL) | |
1581 | fprintf (lra_dump_file, "\n********** Assignment #%d: **********\n\n", | |
1582 | lra_assignment_iter); | |
55a2c322 VM |
1583 | init_lives (); |
1584 | sorted_pseudos = XNEWVEC (int, max_regno); | |
1585 | sorted_reload_pseudos = XNEWVEC (int, max_regno); | |
1586 | regno_allocno_class_array = XNEWVEC (enum reg_class, max_regno); | |
1587 | for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) | |
1588 | regno_allocno_class_array[i] = lra_get_allocno_class (i); | |
f54437d5 | 1589 | former_reload_pseudo_spill_p = false; |
55a2c322 VM |
1590 | init_regno_assign_info (); |
1591 | bitmap_initialize (&all_spilled_pseudos, ®_obstack); | |
1592 | create_live_range_start_chains (); | |
1593 | setup_live_pseudos_and_spill_after_risky_transforms (&all_spilled_pseudos); | |
b2b29377 | 1594 | if (flag_checking && !flag_ipa_ra) |
10e1bdb2 TV |
1595 | for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) |
1596 | if (lra_reg_info[i].nrefs != 0 && reg_renumber[i] >= 0 | |
1597 | && lra_reg_info[i].call_p | |
1598 | && overlaps_hard_reg_set_p (call_used_reg_set, | |
1599 | PSEUDO_REGNO_MODE (i), reg_renumber[i])) | |
1600 | gcc_unreachable (); | |
55a2c322 VM |
1601 | /* Setup insns to process on the next constraint pass. */ |
1602 | bitmap_initialize (&changed_pseudo_bitmap, ®_obstack); | |
1603 | init_live_reload_and_inheritance_pseudos (); | |
1604 | assign_by_spills (); | |
1605 | finish_live_reload_and_inheritance_pseudos (); | |
1606 | bitmap_ior_into (&changed_pseudo_bitmap, &all_spilled_pseudos); | |
1607 | no_spills_p = true; | |
1608 | EXECUTE_IF_SET_IN_BITMAP (&all_spilled_pseudos, 0, u, bi) | |
1609 | /* We ignore spilled pseudos created on last inheritance pass | |
1610 | because they will be removed. */ | |
1611 | if (lra_reg_info[u].restore_regno < 0) | |
1612 | { | |
1613 | no_spills_p = false; | |
1614 | break; | |
1615 | } | |
1616 | finish_live_range_start_chains (); | |
1617 | bitmap_clear (&all_spilled_pseudos); | |
1618 | bitmap_initialize (&insns_to_process, ®_obstack); | |
1619 | EXECUTE_IF_SET_IN_BITMAP (&changed_pseudo_bitmap, 0, u, bi) | |
1620 | bitmap_ior_into (&insns_to_process, &lra_reg_info[u].insn_bitmap); | |
1621 | bitmap_clear (&changed_pseudo_bitmap); | |
1622 | EXECUTE_IF_SET_IN_BITMAP (&insns_to_process, 0, u, bi) | |
1623 | { | |
1624 | lra_push_insn_by_uid (u); | |
1625 | /* Invalidate alternatives for insn should be processed. */ | |
1626 | lra_set_used_insn_alternative_by_uid (u, -1); | |
1627 | } | |
1628 | bitmap_clear (&insns_to_process); | |
1629 | finish_regno_assign_info (); | |
1630 | free (regno_allocno_class_array); | |
1631 | free (sorted_pseudos); | |
1632 | free (sorted_reload_pseudos); | |
1633 | finish_lives (); | |
1634 | timevar_pop (TV_LRA_ASSIGN); | |
f54437d5 VM |
1635 | if (former_reload_pseudo_spill_p) |
1636 | lra_assignment_iter_after_spill++; | |
1637 | if (lra_assignment_iter_after_spill > LRA_MAX_ASSIGNMENT_ITERATION_NUMBER) | |
1638 | internal_error | |
1639 | ("Maximum number of LRA assignment passes is achieved (%d)\n", | |
1640 | LRA_MAX_ASSIGNMENT_ITERATION_NUMBER); | |
55a2c322 VM |
1641 | return no_spills_p; |
1642 | } |