Commit | Line | Data |
---|---|---|
003757ed MD |
1 | /* Instruction scheduling pass. This file computes dependencies between |
2 | instructions. | |
3 | Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, | |
4 | 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc. | |
5 | Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by, | |
6 | and currently maintained by, Jim Wilson (wilson@cygnus.com) | |
7 | ||
8 | This file is part of GCC. | |
9 | ||
10 | GCC is free software; you can redistribute it and/or modify it under | |
11 | the terms of the GNU General Public License as published by the Free | |
12 | Software Foundation; either version 2, or (at your option) any later | |
13 | version. | |
14 | ||
15 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
16 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
17 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
18 | for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
21 | along with GCC; see the file COPYING. If not, write to the Free | |
22 | Software Foundation, 59 Temple Place - Suite 330, Boston, MA | |
23 | 02111-1307, USA. */ | |
24 | \f | |
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
28 | #include "tm.h" | |
29 | #include "toplev.h" | |
30 | #include "rtl.h" | |
31 | #include "tm_p.h" | |
32 | #include "hard-reg-set.h" | |
33 | #include "basic-block.h" | |
34 | #include "regs.h" | |
35 | #include "function.h" | |
36 | #include "flags.h" | |
37 | #include "insn-config.h" | |
38 | #include "insn-attr.h" | |
39 | #include "except.h" | |
40 | #include "toplev.h" | |
41 | #include "recog.h" | |
42 | #include "sched-int.h" | |
43 | #include "params.h" | |
44 | #include "cselib.h" | |
45 | #include "df.h" | |
46 | ||
47 | ||
48 | static regset_head reg_pending_sets_head; | |
49 | static regset_head reg_pending_clobbers_head; | |
50 | static regset_head reg_pending_uses_head; | |
51 | ||
52 | static regset reg_pending_sets; | |
53 | static regset reg_pending_clobbers; | |
54 | static regset reg_pending_uses; | |
55 | ||
56 | /* The following enumeration values tell us what dependencies we | |
57 | should use to implement the barrier. We use true-dependencies for | |
58 | TRUE_BARRIER and anti-dependencies for MOVE_BARRIER. */ | |
59 | enum reg_pending_barrier_mode | |
60 | { | |
61 | NOT_A_BARRIER = 0, | |
62 | MOVE_BARRIER, | |
63 | TRUE_BARRIER | |
64 | }; | |
65 | ||
66 | static enum reg_pending_barrier_mode reg_pending_barrier; | |
67 | ||
68 | /* To speed up the test for duplicate dependency links we keep a | |
69 | record of dependencies created by add_dependence when the average | |
70 | number of instructions in a basic block is very large. | |
71 | ||
72 | Studies have shown that there is typically around 5 instructions between | |
73 | branches for typical C code. So we can make a guess that the average | |
74 | basic block is approximately 5 instructions long; we will choose 100X | |
75 | the average size as a very large basic block. | |
76 | ||
77 | Each insn has associated bitmaps for its dependencies. Each bitmap | |
78 | has enough entries to represent a dependency on any other insn in | |
79 | the insn chain. All bitmap for true dependencies cache is | |
80 | allocated then the rest two ones are also allocated. */ | |
81 | static bitmap_head *true_dependency_cache; | |
82 | static bitmap_head *anti_dependency_cache; | |
83 | static bitmap_head *output_dependency_cache; | |
84 | int cache_size; | |
85 | ||
86 | /* To speed up checking consistency of formed forward insn | |
87 | dependencies we use the following cache. Another possible solution | |
88 | could be switching off checking duplication of insns in forward | |
89 | dependencies. */ | |
90 | #ifdef ENABLE_CHECKING | |
91 | static bitmap_head *forward_dependency_cache; | |
92 | #endif | |
93 | ||
94 | static int deps_may_trap_p (rtx); | |
95 | static void add_dependence_list (rtx, rtx, enum reg_note); | |
96 | static void add_dependence_list_and_free (rtx, rtx *, enum reg_note); | |
97 | static void set_sched_group_p (rtx); | |
98 | ||
99 | static void flush_pending_lists (struct deps *, rtx, int, int); | |
100 | static void sched_analyze_1 (struct deps *, rtx, rtx); | |
101 | static void sched_analyze_2 (struct deps *, rtx, rtx); | |
102 | static void sched_analyze_insn (struct deps *, rtx, rtx, rtx); | |
103 | ||
104 | static rtx get_condition (rtx); | |
105 | static int conditions_mutex_p (rtx, rtx); | |
106 | \f | |
107 | /* Return nonzero if a load of the memory reference MEM can cause a trap. */ | |
108 | ||
109 | static int | |
110 | deps_may_trap_p (rtx mem) | |
111 | { | |
112 | rtx addr = XEXP (mem, 0); | |
113 | ||
114 | if (REG_P (addr) && REGNO (addr) >= FIRST_PSEUDO_REGISTER) | |
115 | { | |
116 | rtx t = get_reg_known_value (REGNO (addr)); | |
117 | if (t) | |
118 | addr = t; | |
119 | } | |
120 | return rtx_addr_can_trap_p (addr); | |
121 | } | |
122 | \f | |
123 | /* Return the INSN_LIST containing INSN in LIST, or NULL | |
124 | if LIST does not contain INSN. */ | |
125 | ||
126 | rtx | |
127 | find_insn_list (rtx insn, rtx list) | |
128 | { | |
129 | while (list) | |
130 | { | |
131 | if (XEXP (list, 0) == insn) | |
132 | return list; | |
133 | list = XEXP (list, 1); | |
134 | } | |
135 | return 0; | |
136 | } | |
137 | \f | |
138 | /* Find the condition under which INSN is executed. */ | |
139 | ||
140 | static rtx | |
141 | get_condition (rtx insn) | |
142 | { | |
143 | rtx pat = PATTERN (insn); | |
144 | rtx cond; | |
145 | ||
146 | if (pat == 0) | |
147 | return 0; | |
148 | if (GET_CODE (pat) == COND_EXEC) | |
149 | return COND_EXEC_TEST (pat); | |
150 | if (GET_CODE (insn) != JUMP_INSN) | |
151 | return 0; | |
152 | if (GET_CODE (pat) != SET || SET_SRC (pat) != pc_rtx) | |
153 | return 0; | |
154 | if (GET_CODE (SET_DEST (pat)) != IF_THEN_ELSE) | |
155 | return 0; | |
156 | pat = SET_DEST (pat); | |
157 | cond = XEXP (pat, 0); | |
158 | if (GET_CODE (XEXP (cond, 1)) == LABEL_REF | |
159 | && XEXP (cond, 2) == pc_rtx) | |
160 | return cond; | |
161 | else if (GET_CODE (XEXP (cond, 2)) == LABEL_REF | |
162 | && XEXP (cond, 1) == pc_rtx) | |
163 | return gen_rtx_fmt_ee (reverse_condition (GET_CODE (cond)), GET_MODE (cond), | |
164 | XEXP (cond, 0), XEXP (cond, 1)); | |
165 | else | |
166 | return 0; | |
167 | } | |
168 | ||
169 | /* Return nonzero if conditions COND1 and COND2 can never be both true. */ | |
170 | ||
171 | static int | |
172 | conditions_mutex_p (rtx cond1, rtx cond2) | |
173 | { | |
174 | if (GET_RTX_CLASS (GET_CODE (cond1)) == '<' | |
175 | && GET_RTX_CLASS (GET_CODE (cond2)) == '<' | |
176 | && GET_CODE (cond1) == reverse_condition (GET_CODE (cond2)) | |
177 | && XEXP (cond1, 0) == XEXP (cond2, 0) | |
178 | && XEXP (cond1, 1) == XEXP (cond2, 1)) | |
179 | return 1; | |
180 | return 0; | |
181 | } | |
182 | \f | |
183 | /* Add ELEM wrapped in an INSN_LIST with reg note kind DEP_TYPE to the | |
184 | LOG_LINKS of INSN, if not already there. DEP_TYPE indicates the | |
185 | type of dependence that this link represents. The function returns | |
186 | nonzero if a new entry has been added to insn's LOG_LINK. */ | |
187 | ||
188 | int | |
189 | add_dependence (rtx insn, rtx elem, enum reg_note dep_type) | |
190 | { | |
191 | rtx link; | |
192 | int present_p; | |
193 | rtx cond1, cond2; | |
194 | ||
195 | /* Don't depend an insn on itself. */ | |
196 | if (insn == elem) | |
197 | return 0; | |
198 | ||
199 | /* We can get a dependency on deleted insns due to optimizations in | |
200 | the register allocation and reloading or due to splitting. Any | |
201 | such dependency is useless and can be ignored. */ | |
202 | if (GET_CODE (elem) == NOTE) | |
203 | return 0; | |
204 | ||
205 | /* flow.c doesn't handle conditional lifetimes entirely correctly; | |
206 | calls mess up the conditional lifetimes. */ | |
207 | /* ??? add_dependence is the wrong place to be eliding dependencies, | |
208 | as that forgets that the condition expressions themselves may | |
209 | be dependent. */ | |
210 | if (GET_CODE (insn) != CALL_INSN && GET_CODE (elem) != CALL_INSN) | |
211 | { | |
212 | cond1 = get_condition (insn); | |
213 | cond2 = get_condition (elem); | |
214 | if (cond1 && cond2 | |
215 | && conditions_mutex_p (cond1, cond2) | |
216 | /* Make sure first instruction doesn't affect condition of second | |
217 | instruction if switched. */ | |
218 | && !modified_in_p (cond1, elem) | |
219 | /* Make sure second instruction doesn't affect condition of first | |
220 | instruction if switched. */ | |
221 | && !modified_in_p (cond2, insn)) | |
222 | return 0; | |
223 | } | |
224 | ||
225 | present_p = 1; | |
226 | #ifdef INSN_SCHEDULING | |
227 | /* ??? No good way to tell from here whether we're doing interblock | |
228 | scheduling. Possibly add another callback. */ | |
229 | #if 0 | |
230 | /* (This code is guarded by INSN_SCHEDULING, otherwise INSN_BB is undefined.) | |
231 | No need for interblock dependences with calls, since | |
232 | calls are not moved between blocks. Note: the edge where | |
233 | elem is a CALL is still required. */ | |
234 | if (GET_CODE (insn) == CALL_INSN | |
235 | && (INSN_BB (elem) != INSN_BB (insn))) | |
236 | return 0; | |
237 | #endif | |
238 | ||
239 | /* If we already have a dependency for ELEM, then we do not need to | |
240 | do anything. Avoiding the list walk below can cut compile times | |
241 | dramatically for some code. */ | |
242 | if (true_dependency_cache != NULL) | |
243 | { | |
244 | enum reg_note present_dep_type = 0; | |
245 | ||
246 | if (anti_dependency_cache == NULL || output_dependency_cache == NULL) | |
247 | abort (); | |
248 | if (bitmap_bit_p (&true_dependency_cache[INSN_LUID (insn)], | |
249 | INSN_LUID (elem))) | |
250 | /* Do nothing (present_set_type is already 0). */ | |
251 | ; | |
252 | else if (bitmap_bit_p (&anti_dependency_cache[INSN_LUID (insn)], | |
253 | INSN_LUID (elem))) | |
254 | present_dep_type = REG_DEP_ANTI; | |
255 | else if (bitmap_bit_p (&output_dependency_cache[INSN_LUID (insn)], | |
256 | INSN_LUID (elem))) | |
257 | present_dep_type = REG_DEP_OUTPUT; | |
258 | else | |
259 | present_p = 0; | |
260 | if (present_p && (int) dep_type >= (int) present_dep_type) | |
261 | return 0; | |
262 | } | |
263 | #endif | |
264 | ||
265 | /* Check that we don't already have this dependence. */ | |
266 | if (present_p) | |
267 | for (link = LOG_LINKS (insn); link; link = XEXP (link, 1)) | |
268 | if (XEXP (link, 0) == elem) | |
269 | { | |
270 | #ifdef INSN_SCHEDULING | |
271 | /* Clear corresponding cache entry because type of the link | |
272 | may be changed. */ | |
273 | if (true_dependency_cache != NULL) | |
274 | { | |
275 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI) | |
276 | bitmap_clear_bit (&anti_dependency_cache[INSN_LUID (insn)], | |
277 | INSN_LUID (elem)); | |
278 | else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT | |
279 | && output_dependency_cache) | |
280 | bitmap_clear_bit (&output_dependency_cache[INSN_LUID (insn)], | |
281 | INSN_LUID (elem)); | |
282 | else | |
283 | abort (); | |
284 | } | |
285 | #endif | |
286 | ||
287 | /* If this is a more restrictive type of dependence than the existing | |
288 | one, then change the existing dependence to this type. */ | |
289 | if ((int) dep_type < (int) REG_NOTE_KIND (link)) | |
290 | PUT_REG_NOTE_KIND (link, dep_type); | |
291 | ||
292 | #ifdef INSN_SCHEDULING | |
293 | /* If we are adding a dependency to INSN's LOG_LINKs, then | |
294 | note that in the bitmap caches of dependency information. */ | |
295 | if (true_dependency_cache != NULL) | |
296 | { | |
297 | if ((int) REG_NOTE_KIND (link) == 0) | |
298 | bitmap_set_bit (&true_dependency_cache[INSN_LUID (insn)], | |
299 | INSN_LUID (elem)); | |
300 | else if (REG_NOTE_KIND (link) == REG_DEP_ANTI) | |
301 | bitmap_set_bit (&anti_dependency_cache[INSN_LUID (insn)], | |
302 | INSN_LUID (elem)); | |
303 | else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT) | |
304 | bitmap_set_bit (&output_dependency_cache[INSN_LUID (insn)], | |
305 | INSN_LUID (elem)); | |
306 | } | |
307 | #endif | |
308 | return 0; | |
309 | } | |
310 | /* Might want to check one level of transitivity to save conses. */ | |
311 | ||
312 | link = alloc_INSN_LIST (elem, LOG_LINKS (insn)); | |
313 | LOG_LINKS (insn) = link; | |
314 | ||
315 | /* Insn dependency, not data dependency. */ | |
316 | PUT_REG_NOTE_KIND (link, dep_type); | |
317 | ||
318 | #ifdef INSN_SCHEDULING | |
319 | /* If we are adding a dependency to INSN's LOG_LINKs, then note that | |
320 | in the bitmap caches of dependency information. */ | |
321 | if (true_dependency_cache != NULL) | |
322 | { | |
323 | if ((int) dep_type == 0) | |
324 | bitmap_set_bit (&true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); | |
325 | else if (dep_type == REG_DEP_ANTI) | |
326 | bitmap_set_bit (&anti_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); | |
327 | else if (dep_type == REG_DEP_OUTPUT) | |
328 | bitmap_set_bit (&output_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); | |
329 | } | |
330 | #endif | |
331 | return 1; | |
332 | } | |
333 | ||
334 | /* A convenience wrapper to operate on an entire list. */ | |
335 | ||
336 | static void | |
337 | add_dependence_list (rtx insn, rtx list, enum reg_note dep_type) | |
338 | { | |
339 | for (; list; list = XEXP (list, 1)) | |
340 | add_dependence (insn, XEXP (list, 0), dep_type); | |
341 | } | |
342 | ||
343 | /* Similar, but free *LISTP at the same time. */ | |
344 | ||
345 | static void | |
346 | add_dependence_list_and_free (rtx insn, rtx *listp, enum reg_note dep_type) | |
347 | { | |
348 | rtx list, next; | |
349 | for (list = *listp, *listp = NULL; list ; list = next) | |
350 | { | |
351 | next = XEXP (list, 1); | |
352 | add_dependence (insn, XEXP (list, 0), dep_type); | |
353 | free_INSN_LIST_node (list); | |
354 | } | |
355 | } | |
356 | ||
357 | /* Set SCHED_GROUP_P and care for the rest of the bookkeeping that | |
358 | goes along with that. */ | |
359 | ||
360 | static void | |
361 | set_sched_group_p (rtx insn) | |
362 | { | |
363 | rtx prev; | |
364 | ||
365 | SCHED_GROUP_P (insn) = 1; | |
366 | ||
367 | prev = prev_nonnote_insn (insn); | |
368 | add_dependence (insn, prev, REG_DEP_ANTI); | |
369 | } | |
370 | \f | |
371 | /* Process an insn's memory dependencies. There are four kinds of | |
372 | dependencies: | |
373 | ||
374 | (0) read dependence: read follows read | |
375 | (1) true dependence: read follows write | |
376 | (2) anti dependence: write follows read | |
377 | (3) output dependence: write follows write | |
378 | ||
379 | We are careful to build only dependencies which actually exist, and | |
380 | use transitivity to avoid building too many links. */ | |
381 | ||
382 | /* Add an INSN and MEM reference pair to a pending INSN_LIST and MEM_LIST. | |
383 | The MEM is a memory reference contained within INSN, which we are saving | |
384 | so that we can do memory aliasing on it. */ | |
385 | ||
386 | void | |
387 | add_insn_mem_dependence (struct deps *deps, rtx *insn_list, rtx *mem_list, | |
388 | rtx insn, rtx mem) | |
389 | { | |
390 | rtx link; | |
391 | ||
392 | link = alloc_INSN_LIST (insn, *insn_list); | |
393 | *insn_list = link; | |
394 | ||
395 | if (current_sched_info->use_cselib) | |
396 | { | |
397 | mem = shallow_copy_rtx (mem); | |
398 | XEXP (mem, 0) = cselib_subst_to_values (XEXP (mem, 0)); | |
399 | } | |
400 | link = alloc_EXPR_LIST (VOIDmode, canon_rtx (mem), *mem_list); | |
401 | *mem_list = link; | |
402 | ||
403 | deps->pending_lists_length++; | |
404 | } | |
405 | ||
406 | /* Make a dependency between every memory reference on the pending lists | |
407 | and INSN, thus flushing the pending lists. FOR_READ is true if emitting | |
408 | dependencies for a read operation, similarly with FOR_WRITE. */ | |
409 | ||
410 | static void | |
411 | flush_pending_lists (struct deps *deps, rtx insn, int for_read, | |
412 | int for_write) | |
413 | { | |
414 | if (for_write) | |
415 | { | |
416 | add_dependence_list_and_free (insn, &deps->pending_read_insns, | |
417 | REG_DEP_ANTI); | |
418 | free_EXPR_LIST_list (&deps->pending_read_mems); | |
419 | } | |
420 | ||
421 | add_dependence_list_and_free (insn, &deps->pending_write_insns, | |
422 | for_read ? REG_DEP_ANTI : REG_DEP_OUTPUT); | |
423 | free_EXPR_LIST_list (&deps->pending_write_mems); | |
424 | deps->pending_lists_length = 0; | |
425 | ||
426 | add_dependence_list_and_free (insn, &deps->last_pending_memory_flush, | |
427 | for_read ? REG_DEP_ANTI : REG_DEP_OUTPUT); | |
428 | deps->last_pending_memory_flush = alloc_INSN_LIST (insn, NULL_RTX); | |
429 | deps->pending_flush_length = 1; | |
430 | } | |
431 | \f | |
432 | /* Analyze a single SET, CLOBBER, PRE_DEC, POST_DEC, PRE_INC or POST_INC | |
433 | rtx, X, creating all dependencies generated by the write to the | |
434 | destination of X, and reads of everything mentioned. */ | |
435 | ||
436 | static void | |
437 | sched_analyze_1 (struct deps *deps, rtx x, rtx insn) | |
438 | { | |
439 | int regno; | |
440 | rtx dest = XEXP (x, 0); | |
441 | enum rtx_code code = GET_CODE (x); | |
442 | ||
443 | if (dest == 0) | |
444 | return; | |
445 | ||
446 | if (GET_CODE (dest) == PARALLEL) | |
447 | { | |
448 | int i; | |
449 | ||
450 | for (i = XVECLEN (dest, 0) - 1; i >= 0; i--) | |
451 | if (XEXP (XVECEXP (dest, 0, i), 0) != 0) | |
452 | sched_analyze_1 (deps, | |
453 | gen_rtx_CLOBBER (VOIDmode, | |
454 | XEXP (XVECEXP (dest, 0, i), 0)), | |
455 | insn); | |
456 | ||
457 | if (GET_CODE (x) == SET) | |
458 | sched_analyze_2 (deps, SET_SRC (x), insn); | |
459 | return; | |
460 | } | |
461 | ||
462 | while (GET_CODE (dest) == STRICT_LOW_PART || GET_CODE (dest) == SUBREG | |
463 | || GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT) | |
464 | { | |
465 | if (GET_CODE (dest) == STRICT_LOW_PART | |
466 | || GET_CODE (dest) == ZERO_EXTRACT | |
467 | || GET_CODE (dest) == SIGN_EXTRACT | |
468 | || read_modify_subreg_p (dest)) | |
469 | { | |
470 | /* These both read and modify the result. We must handle | |
471 | them as writes to get proper dependencies for following | |
472 | instructions. We must handle them as reads to get proper | |
473 | dependencies from this to previous instructions. | |
474 | Thus we need to call sched_analyze_2. */ | |
475 | ||
476 | sched_analyze_2 (deps, XEXP (dest, 0), insn); | |
477 | } | |
478 | if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT) | |
479 | { | |
480 | /* The second and third arguments are values read by this insn. */ | |
481 | sched_analyze_2 (deps, XEXP (dest, 1), insn); | |
482 | sched_analyze_2 (deps, XEXP (dest, 2), insn); | |
483 | } | |
484 | dest = XEXP (dest, 0); | |
485 | } | |
486 | ||
487 | if (GET_CODE (dest) == REG) | |
488 | { | |
489 | regno = REGNO (dest); | |
490 | ||
491 | /* A hard reg in a wide mode may really be multiple registers. | |
492 | If so, mark all of them just like the first. */ | |
493 | if (regno < FIRST_PSEUDO_REGISTER) | |
494 | { | |
495 | int i = HARD_REGNO_NREGS (regno, GET_MODE (dest)); | |
496 | if (code == SET) | |
497 | { | |
498 | while (--i >= 0) | |
499 | SET_REGNO_REG_SET (reg_pending_sets, regno + i); | |
500 | } | |
501 | else | |
502 | { | |
503 | while (--i >= 0) | |
504 | SET_REGNO_REG_SET (reg_pending_clobbers, regno + i); | |
505 | } | |
506 | } | |
507 | /* ??? Reload sometimes emits USEs and CLOBBERs of pseudos that | |
508 | it does not reload. Ignore these as they have served their | |
509 | purpose already. */ | |
510 | else if (regno >= deps->max_reg) | |
511 | { | |
512 | if (GET_CODE (PATTERN (insn)) != USE | |
513 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
514 | abort (); | |
515 | } | |
516 | else | |
517 | { | |
518 | if (code == SET) | |
519 | SET_REGNO_REG_SET (reg_pending_sets, regno); | |
520 | else | |
521 | SET_REGNO_REG_SET (reg_pending_clobbers, regno); | |
522 | ||
523 | /* Pseudos that are REG_EQUIV to something may be replaced | |
524 | by that during reloading. We need only add dependencies for | |
525 | the address in the REG_EQUIV note. */ | |
526 | if (!reload_completed && get_reg_known_equiv_p (regno)) | |
527 | { | |
528 | rtx t = get_reg_known_value (regno); | |
529 | if (GET_CODE (t) == MEM) | |
530 | sched_analyze_2 (deps, XEXP (t, 0), insn); | |
531 | } | |
532 | ||
533 | /* Don't let it cross a call after scheduling if it doesn't | |
534 | already cross one. */ | |
535 | if (REG_N_CALLS_CROSSED (regno) == 0) | |
536 | add_dependence_list (insn, deps->last_function_call, REG_DEP_ANTI); | |
537 | } | |
538 | } | |
539 | else if (GET_CODE (dest) == MEM) | |
540 | { | |
541 | /* Writing memory. */ | |
542 | rtx t = dest; | |
543 | ||
544 | if (current_sched_info->use_cselib) | |
545 | { | |
546 | t = shallow_copy_rtx (dest); | |
547 | cselib_lookup (XEXP (t, 0), Pmode, 1); | |
548 | XEXP (t, 0) = cselib_subst_to_values (XEXP (t, 0)); | |
549 | } | |
550 | t = canon_rtx (t); | |
551 | ||
552 | if (deps->pending_lists_length > MAX_PENDING_LIST_LENGTH) | |
553 | { | |
554 | /* Flush all pending reads and writes to prevent the pending lists | |
555 | from getting any larger. Insn scheduling runs too slowly when | |
556 | these lists get long. When compiling GCC with itself, | |
557 | this flush occurs 8 times for sparc, and 10 times for m88k using | |
558 | the default value of 32. */ | |
559 | flush_pending_lists (deps, insn, false, true); | |
560 | } | |
561 | else | |
562 | { | |
563 | rtx pending, pending_mem; | |
564 | ||
565 | pending = deps->pending_read_insns; | |
566 | pending_mem = deps->pending_read_mems; | |
567 | while (pending) | |
568 | { | |
569 | if (anti_dependence (XEXP (pending_mem, 0), t)) | |
570 | add_dependence (insn, XEXP (pending, 0), REG_DEP_ANTI); | |
571 | ||
572 | pending = XEXP (pending, 1); | |
573 | pending_mem = XEXP (pending_mem, 1); | |
574 | } | |
575 | ||
576 | pending = deps->pending_write_insns; | |
577 | pending_mem = deps->pending_write_mems; | |
578 | while (pending) | |
579 | { | |
580 | if (output_dependence (XEXP (pending_mem, 0), t)) | |
581 | add_dependence (insn, XEXP (pending, 0), REG_DEP_OUTPUT); | |
582 | ||
583 | pending = XEXP (pending, 1); | |
584 | pending_mem = XEXP (pending_mem, 1); | |
585 | } | |
586 | ||
587 | add_dependence_list (insn, deps->last_pending_memory_flush, | |
588 | REG_DEP_ANTI); | |
589 | ||
590 | add_insn_mem_dependence (deps, &deps->pending_write_insns, | |
591 | &deps->pending_write_mems, insn, dest); | |
592 | } | |
593 | sched_analyze_2 (deps, XEXP (dest, 0), insn); | |
594 | } | |
595 | ||
596 | /* Analyze reads. */ | |
597 | if (GET_CODE (x) == SET) | |
598 | sched_analyze_2 (deps, SET_SRC (x), insn); | |
599 | } | |
600 | ||
601 | /* Analyze the uses of memory and registers in rtx X in INSN. */ | |
602 | ||
603 | static void | |
604 | sched_analyze_2 (struct deps *deps, rtx x, rtx insn) | |
605 | { | |
606 | int i; | |
607 | int j; | |
608 | enum rtx_code code; | |
609 | const char *fmt; | |
610 | ||
611 | if (x == 0) | |
612 | return; | |
613 | ||
614 | code = GET_CODE (x); | |
615 | ||
616 | switch (code) | |
617 | { | |
618 | case CONST_INT: | |
619 | case CONST_DOUBLE: | |
620 | case CONST_VECTOR: | |
621 | case SYMBOL_REF: | |
622 | case CONST: | |
623 | case LABEL_REF: | |
624 | /* Ignore constants. Note that we must handle CONST_DOUBLE here | |
625 | because it may have a cc0_rtx in its CONST_DOUBLE_CHAIN field, but | |
626 | this does not mean that this insn is using cc0. */ | |
627 | return; | |
628 | ||
629 | #ifdef HAVE_cc0 | |
630 | case CC0: | |
631 | /* User of CC0 depends on immediately preceding insn. */ | |
632 | set_sched_group_p (insn); | |
633 | /* Don't move CC0 setter to another block (it can set up the | |
634 | same flag for previous CC0 users which is safe). */ | |
635 | CANT_MOVE (prev_nonnote_insn (insn)) = 1; | |
636 | return; | |
637 | #endif | |
638 | ||
639 | case REG: | |
640 | { | |
641 | int regno = REGNO (x); | |
642 | if (regno < FIRST_PSEUDO_REGISTER) | |
643 | { | |
644 | int i = HARD_REGNO_NREGS (regno, GET_MODE (x)); | |
645 | while (--i >= 0) | |
646 | SET_REGNO_REG_SET (reg_pending_uses, regno + i); | |
647 | } | |
648 | /* ??? Reload sometimes emits USEs and CLOBBERs of pseudos that | |
649 | it does not reload. Ignore these as they have served their | |
650 | purpose already. */ | |
651 | else if (regno >= deps->max_reg) | |
652 | { | |
653 | if (GET_CODE (PATTERN (insn)) != USE | |
654 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
655 | abort (); | |
656 | } | |
657 | else | |
658 | { | |
659 | SET_REGNO_REG_SET (reg_pending_uses, regno); | |
660 | ||
661 | /* Pseudos that are REG_EQUIV to something may be replaced | |
662 | by that during reloading. We need only add dependencies for | |
663 | the address in the REG_EQUIV note. */ | |
664 | if (!reload_completed && get_reg_known_equiv_p (regno)) | |
665 | { | |
666 | rtx t = get_reg_known_value (regno); | |
667 | if (GET_CODE (t) == MEM) | |
668 | sched_analyze_2 (deps, XEXP (t, 0), insn); | |
669 | } | |
670 | ||
671 | /* If the register does not already cross any calls, then add this | |
672 | insn to the sched_before_next_call list so that it will still | |
673 | not cross calls after scheduling. */ | |
674 | if (REG_N_CALLS_CROSSED (regno) == 0) | |
675 | deps->sched_before_next_call | |
676 | = alloc_INSN_LIST (insn, deps->sched_before_next_call); | |
677 | } | |
678 | return; | |
679 | } | |
680 | ||
681 | case MEM: | |
682 | { | |
683 | /* Reading memory. */ | |
684 | rtx u; | |
685 | rtx pending, pending_mem; | |
686 | rtx t = x; | |
687 | ||
688 | if (current_sched_info->use_cselib) | |
689 | { | |
690 | t = shallow_copy_rtx (t); | |
691 | cselib_lookup (XEXP (t, 0), Pmode, 1); | |
692 | XEXP (t, 0) = cselib_subst_to_values (XEXP (t, 0)); | |
693 | } | |
694 | t = canon_rtx (t); | |
695 | pending = deps->pending_read_insns; | |
696 | pending_mem = deps->pending_read_mems; | |
697 | while (pending) | |
698 | { | |
699 | if (read_dependence (XEXP (pending_mem, 0), t)) | |
700 | add_dependence (insn, XEXP (pending, 0), REG_DEP_ANTI); | |
701 | ||
702 | pending = XEXP (pending, 1); | |
703 | pending_mem = XEXP (pending_mem, 1); | |
704 | } | |
705 | ||
706 | pending = deps->pending_write_insns; | |
707 | pending_mem = deps->pending_write_mems; | |
708 | while (pending) | |
709 | { | |
710 | if (true_dependence (XEXP (pending_mem, 0), VOIDmode, | |
711 | t, rtx_varies_p)) | |
712 | add_dependence (insn, XEXP (pending, 0), 0); | |
713 | ||
714 | pending = XEXP (pending, 1); | |
715 | pending_mem = XEXP (pending_mem, 1); | |
716 | } | |
717 | ||
718 | for (u = deps->last_pending_memory_flush; u; u = XEXP (u, 1)) | |
719 | if (GET_CODE (XEXP (u, 0)) != JUMP_INSN | |
720 | || deps_may_trap_p (x)) | |
721 | add_dependence (insn, XEXP (u, 0), REG_DEP_ANTI); | |
722 | ||
723 | /* Always add these dependencies to pending_reads, since | |
724 | this insn may be followed by a write. */ | |
725 | add_insn_mem_dependence (deps, &deps->pending_read_insns, | |
726 | &deps->pending_read_mems, insn, x); | |
727 | ||
728 | /* Take advantage of tail recursion here. */ | |
729 | sched_analyze_2 (deps, XEXP (x, 0), insn); | |
730 | return; | |
731 | } | |
732 | ||
733 | /* Force pending stores to memory in case a trap handler needs them. */ | |
734 | case TRAP_IF: | |
735 | flush_pending_lists (deps, insn, true, false); | |
736 | break; | |
737 | ||
738 | case ASM_OPERANDS: | |
739 | case ASM_INPUT: | |
740 | case UNSPEC_VOLATILE: | |
741 | { | |
742 | /* Traditional and volatile asm instructions must be considered to use | |
743 | and clobber all hard registers, all pseudo-registers and all of | |
744 | memory. So must TRAP_IF and UNSPEC_VOLATILE operations. | |
745 | ||
746 | Consider for instance a volatile asm that changes the fpu rounding | |
747 | mode. An insn should not be moved across this even if it only uses | |
748 | pseudo-regs because it might give an incorrectly rounded result. */ | |
749 | if (code != ASM_OPERANDS || MEM_VOLATILE_P (x)) | |
750 | reg_pending_barrier = TRUE_BARRIER; | |
751 | ||
752 | /* For all ASM_OPERANDS, we must traverse the vector of input operands. | |
753 | We can not just fall through here since then we would be confused | |
754 | by the ASM_INPUT rtx inside ASM_OPERANDS, which do not indicate | |
755 | traditional asms unlike their normal usage. */ | |
756 | ||
757 | if (code == ASM_OPERANDS) | |
758 | { | |
759 | for (j = 0; j < ASM_OPERANDS_INPUT_LENGTH (x); j++) | |
760 | sched_analyze_2 (deps, ASM_OPERANDS_INPUT (x, j), insn); | |
761 | return; | |
762 | } | |
763 | break; | |
764 | } | |
765 | ||
766 | case PRE_DEC: | |
767 | case POST_DEC: | |
768 | case PRE_INC: | |
769 | case POST_INC: | |
770 | /* These both read and modify the result. We must handle them as writes | |
771 | to get proper dependencies for following instructions. We must handle | |
772 | them as reads to get proper dependencies from this to previous | |
773 | instructions. Thus we need to pass them to both sched_analyze_1 | |
774 | and sched_analyze_2. We must call sched_analyze_2 first in order | |
775 | to get the proper antecedent for the read. */ | |
776 | sched_analyze_2 (deps, XEXP (x, 0), insn); | |
777 | sched_analyze_1 (deps, x, insn); | |
778 | return; | |
779 | ||
780 | case POST_MODIFY: | |
781 | case PRE_MODIFY: | |
782 | /* op0 = op0 + op1 */ | |
783 | sched_analyze_2 (deps, XEXP (x, 0), insn); | |
784 | sched_analyze_2 (deps, XEXP (x, 1), insn); | |
785 | sched_analyze_1 (deps, x, insn); | |
786 | return; | |
787 | ||
788 | default: | |
789 | break; | |
790 | } | |
791 | ||
792 | /* Other cases: walk the insn. */ | |
793 | fmt = GET_RTX_FORMAT (code); | |
794 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
795 | { | |
796 | if (fmt[i] == 'e') | |
797 | sched_analyze_2 (deps, XEXP (x, i), insn); | |
798 | else if (fmt[i] == 'E') | |
799 | for (j = 0; j < XVECLEN (x, i); j++) | |
800 | sched_analyze_2 (deps, XVECEXP (x, i, j), insn); | |
801 | } | |
802 | } | |
803 | ||
804 | /* Analyze an INSN with pattern X to find all dependencies. */ | |
805 | ||
806 | static void | |
807 | sched_analyze_insn (struct deps *deps, rtx x, rtx insn, rtx loop_notes) | |
808 | { | |
809 | RTX_CODE code = GET_CODE (x); | |
810 | rtx link; | |
811 | int i; | |
812 | ||
813 | if (code == COND_EXEC) | |
814 | { | |
815 | sched_analyze_2 (deps, COND_EXEC_TEST (x), insn); | |
816 | ||
817 | /* ??? Should be recording conditions so we reduce the number of | |
818 | false dependencies. */ | |
819 | x = COND_EXEC_CODE (x); | |
820 | code = GET_CODE (x); | |
821 | } | |
822 | if (code == SET || code == CLOBBER) | |
823 | { | |
824 | sched_analyze_1 (deps, x, insn); | |
825 | ||
826 | /* Bare clobber insns are used for letting life analysis, reg-stack | |
827 | and others know that a value is dead. Depend on the last call | |
828 | instruction so that reg-stack won't get confused. */ | |
829 | if (code == CLOBBER) | |
830 | add_dependence_list (insn, deps->last_function_call, REG_DEP_OUTPUT); | |
831 | } | |
832 | else if (code == PARALLEL) | |
833 | { | |
834 | int i; | |
835 | for (i = XVECLEN (x, 0) - 1; i >= 0; i--) | |
836 | { | |
837 | rtx sub = XVECEXP (x, 0, i); | |
838 | code = GET_CODE (sub); | |
839 | ||
840 | if (code == COND_EXEC) | |
841 | { | |
842 | sched_analyze_2 (deps, COND_EXEC_TEST (sub), insn); | |
843 | sub = COND_EXEC_CODE (sub); | |
844 | code = GET_CODE (sub); | |
845 | } | |
846 | if (code == SET || code == CLOBBER) | |
847 | sched_analyze_1 (deps, sub, insn); | |
848 | else | |
849 | sched_analyze_2 (deps, sub, insn); | |
850 | } | |
851 | } | |
852 | else | |
853 | sched_analyze_2 (deps, x, insn); | |
854 | ||
855 | /* Mark registers CLOBBERED or used by called function. */ | |
856 | if (GET_CODE (insn) == CALL_INSN) | |
857 | { | |
858 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1)) | |
859 | { | |
860 | if (GET_CODE (XEXP (link, 0)) == CLOBBER) | |
861 | sched_analyze_1 (deps, XEXP (link, 0), insn); | |
862 | else | |
863 | sched_analyze_2 (deps, XEXP (link, 0), insn); | |
864 | } | |
865 | if (find_reg_note (insn, REG_SETJMP, NULL)) | |
866 | reg_pending_barrier = MOVE_BARRIER; | |
867 | } | |
868 | ||
869 | if (GET_CODE (insn) == JUMP_INSN) | |
870 | { | |
871 | rtx next; | |
872 | next = next_nonnote_insn (insn); | |
873 | if (next && GET_CODE (next) == BARRIER) | |
874 | reg_pending_barrier = TRUE_BARRIER; | |
875 | else | |
876 | { | |
877 | rtx pending, pending_mem; | |
878 | regset_head tmp_uses, tmp_sets; | |
879 | INIT_REG_SET (&tmp_uses); | |
880 | INIT_REG_SET (&tmp_sets); | |
881 | ||
882 | (*current_sched_info->compute_jump_reg_dependencies) | |
883 | (insn, &deps->reg_conditional_sets, &tmp_uses, &tmp_sets); | |
884 | /* Make latency of jump equal to 0 by using anti-dependence. */ | |
885 | EXECUTE_IF_SET_IN_REG_SET (&tmp_uses, 0, i, | |
886 | { | |
887 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
888 | add_dependence_list (insn, reg_last->sets, REG_DEP_ANTI); | |
889 | add_dependence_list (insn, reg_last->clobbers, REG_DEP_ANTI); | |
890 | reg_last->uses_length++; | |
891 | reg_last->uses = alloc_INSN_LIST (insn, reg_last->uses); | |
892 | }); | |
893 | IOR_REG_SET (reg_pending_sets, &tmp_sets); | |
894 | ||
895 | CLEAR_REG_SET (&tmp_uses); | |
896 | CLEAR_REG_SET (&tmp_sets); | |
897 | ||
898 | /* All memory writes and volatile reads must happen before the | |
899 | jump. Non-volatile reads must happen before the jump iff | |
900 | the result is needed by the above register used mask. */ | |
901 | ||
902 | pending = deps->pending_write_insns; | |
903 | pending_mem = deps->pending_write_mems; | |
904 | while (pending) | |
905 | { | |
906 | add_dependence (insn, XEXP (pending, 0), REG_DEP_OUTPUT); | |
907 | pending = XEXP (pending, 1); | |
908 | pending_mem = XEXP (pending_mem, 1); | |
909 | } | |
910 | ||
911 | pending = deps->pending_read_insns; | |
912 | pending_mem = deps->pending_read_mems; | |
913 | while (pending) | |
914 | { | |
915 | if (MEM_VOLATILE_P (XEXP (pending_mem, 0))) | |
916 | add_dependence (insn, XEXP (pending, 0), REG_DEP_OUTPUT); | |
917 | pending = XEXP (pending, 1); | |
918 | pending_mem = XEXP (pending_mem, 1); | |
919 | } | |
920 | ||
921 | add_dependence_list (insn, deps->last_pending_memory_flush, | |
922 | REG_DEP_ANTI); | |
923 | } | |
924 | } | |
925 | ||
926 | /* If there is a {LOOP,EHREGION}_{BEG,END} note in the middle of a basic | |
927 | block, then we must be sure that no instructions are scheduled across it. | |
928 | Otherwise, the reg_n_refs info (which depends on loop_depth) would | |
929 | become incorrect. */ | |
930 | if (loop_notes) | |
931 | { | |
932 | rtx link; | |
933 | ||
934 | /* Update loop_notes with any notes from this insn. Also determine | |
935 | if any of the notes on the list correspond to instruction scheduling | |
936 | barriers (loop, eh & setjmp notes, but not range notes). */ | |
937 | link = loop_notes; | |
938 | while (XEXP (link, 1)) | |
939 | { | |
940 | if (INTVAL (XEXP (link, 0)) == NOTE_INSN_LOOP_BEG | |
941 | || INTVAL (XEXP (link, 0)) == NOTE_INSN_LOOP_END | |
942 | || INTVAL (XEXP (link, 0)) == NOTE_INSN_EH_REGION_BEG | |
943 | || INTVAL (XEXP (link, 0)) == NOTE_INSN_EH_REGION_END) | |
944 | reg_pending_barrier = MOVE_BARRIER; | |
945 | ||
946 | link = XEXP (link, 1); | |
947 | } | |
948 | XEXP (link, 1) = REG_NOTES (insn); | |
949 | REG_NOTES (insn) = loop_notes; | |
950 | } | |
951 | ||
952 | /* If this instruction can throw an exception, then moving it changes | |
953 | where block boundaries fall. This is mighty confusing elsewhere. | |
954 | Therefore, prevent such an instruction from being moved. */ | |
955 | if (can_throw_internal (insn)) | |
956 | reg_pending_barrier = MOVE_BARRIER; | |
957 | ||
958 | /* Add dependencies if a scheduling barrier was found. */ | |
959 | if (reg_pending_barrier) | |
960 | { | |
961 | /* In the case of barrier the most added dependencies are not | |
962 | real, so we use anti-dependence here. */ | |
963 | if (GET_CODE (PATTERN (insn)) == COND_EXEC) | |
964 | { | |
965 | EXECUTE_IF_SET_IN_REG_SET (&deps->reg_last_in_use, 0, i, | |
966 | { | |
967 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
968 | add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); | |
969 | add_dependence_list | |
970 | (insn, reg_last->sets, | |
971 | reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); | |
972 | add_dependence_list | |
973 | (insn, reg_last->clobbers, | |
974 | reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); | |
975 | }); | |
976 | } | |
977 | else | |
978 | { | |
979 | EXECUTE_IF_SET_IN_REG_SET (&deps->reg_last_in_use, 0, i, | |
980 | { | |
981 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
982 | add_dependence_list_and_free (insn, ®_last->uses, | |
983 | REG_DEP_ANTI); | |
984 | add_dependence_list_and_free | |
985 | (insn, ®_last->sets, | |
986 | reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); | |
987 | add_dependence_list_and_free | |
988 | (insn, ®_last->clobbers, | |
989 | reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); | |
990 | reg_last->uses_length = 0; | |
991 | reg_last->clobbers_length = 0; | |
992 | }); | |
993 | } | |
994 | ||
995 | for (i = 0; i < deps->max_reg; i++) | |
996 | { | |
997 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
998 | reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); | |
999 | SET_REGNO_REG_SET (&deps->reg_last_in_use, i); | |
1000 | } | |
1001 | ||
1002 | flush_pending_lists (deps, insn, true, true); | |
1003 | CLEAR_REG_SET (&deps->reg_conditional_sets); | |
1004 | reg_pending_barrier = NOT_A_BARRIER; | |
1005 | } | |
1006 | else | |
1007 | { | |
1008 | /* If the current insn is conditional, we can't free any | |
1009 | of the lists. */ | |
1010 | if (GET_CODE (PATTERN (insn)) == COND_EXEC) | |
1011 | { | |
1012 | EXECUTE_IF_SET_IN_REG_SET (reg_pending_uses, 0, i, | |
1013 | { | |
1014 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1015 | add_dependence_list (insn, reg_last->sets, 0); | |
1016 | add_dependence_list (insn, reg_last->clobbers, 0); | |
1017 | reg_last->uses = alloc_INSN_LIST (insn, reg_last->uses); | |
1018 | reg_last->uses_length++; | |
1019 | }); | |
1020 | EXECUTE_IF_SET_IN_REG_SET (reg_pending_clobbers, 0, i, | |
1021 | { | |
1022 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1023 | add_dependence_list (insn, reg_last->sets, REG_DEP_OUTPUT); | |
1024 | add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); | |
1025 | reg_last->clobbers = alloc_INSN_LIST (insn, reg_last->clobbers); | |
1026 | reg_last->clobbers_length++; | |
1027 | }); | |
1028 | EXECUTE_IF_SET_IN_REG_SET (reg_pending_sets, 0, i, | |
1029 | { | |
1030 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1031 | add_dependence_list (insn, reg_last->sets, REG_DEP_OUTPUT); | |
1032 | add_dependence_list (insn, reg_last->clobbers, REG_DEP_OUTPUT); | |
1033 | add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); | |
1034 | reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); | |
1035 | SET_REGNO_REG_SET (&deps->reg_conditional_sets, i); | |
1036 | }); | |
1037 | } | |
1038 | else | |
1039 | { | |
1040 | EXECUTE_IF_SET_IN_REG_SET (reg_pending_uses, 0, i, | |
1041 | { | |
1042 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1043 | add_dependence_list (insn, reg_last->sets, 0); | |
1044 | add_dependence_list (insn, reg_last->clobbers, 0); | |
1045 | reg_last->uses_length++; | |
1046 | reg_last->uses = alloc_INSN_LIST (insn, reg_last->uses); | |
1047 | }); | |
1048 | EXECUTE_IF_SET_IN_REG_SET (reg_pending_clobbers, 0, i, | |
1049 | { | |
1050 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1051 | if (reg_last->uses_length > MAX_PENDING_LIST_LENGTH | |
1052 | || reg_last->clobbers_length > MAX_PENDING_LIST_LENGTH) | |
1053 | { | |
1054 | add_dependence_list_and_free (insn, ®_last->sets, | |
1055 | REG_DEP_OUTPUT); | |
1056 | add_dependence_list_and_free (insn, ®_last->uses, | |
1057 | REG_DEP_ANTI); | |
1058 | add_dependence_list_and_free (insn, ®_last->clobbers, | |
1059 | REG_DEP_OUTPUT); | |
1060 | reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); | |
1061 | reg_last->clobbers_length = 0; | |
1062 | reg_last->uses_length = 0; | |
1063 | } | |
1064 | else | |
1065 | { | |
1066 | add_dependence_list (insn, reg_last->sets, REG_DEP_OUTPUT); | |
1067 | add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); | |
1068 | } | |
1069 | reg_last->clobbers_length++; | |
1070 | reg_last->clobbers = alloc_INSN_LIST (insn, reg_last->clobbers); | |
1071 | }); | |
1072 | EXECUTE_IF_SET_IN_REG_SET (reg_pending_sets, 0, i, | |
1073 | { | |
1074 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1075 | add_dependence_list_and_free (insn, ®_last->sets, | |
1076 | REG_DEP_OUTPUT); | |
1077 | add_dependence_list_and_free (insn, ®_last->clobbers, | |
1078 | REG_DEP_OUTPUT); | |
1079 | add_dependence_list_and_free (insn, ®_last->uses, | |
1080 | REG_DEP_ANTI); | |
1081 | reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); | |
1082 | reg_last->uses_length = 0; | |
1083 | reg_last->clobbers_length = 0; | |
1084 | CLEAR_REGNO_REG_SET (&deps->reg_conditional_sets, i); | |
1085 | }); | |
1086 | } | |
1087 | ||
1088 | IOR_REG_SET (&deps->reg_last_in_use, reg_pending_uses); | |
1089 | IOR_REG_SET (&deps->reg_last_in_use, reg_pending_clobbers); | |
1090 | IOR_REG_SET (&deps->reg_last_in_use, reg_pending_sets); | |
1091 | } | |
1092 | CLEAR_REG_SET (reg_pending_uses); | |
1093 | CLEAR_REG_SET (reg_pending_clobbers); | |
1094 | CLEAR_REG_SET (reg_pending_sets); | |
1095 | ||
1096 | /* If we are currently in a libcall scheduling group, then mark the | |
1097 | current insn as being in a scheduling group and that it can not | |
1098 | be moved into a different basic block. */ | |
1099 | ||
1100 | if (deps->libcall_block_tail_insn) | |
1101 | { | |
1102 | set_sched_group_p (insn); | |
1103 | CANT_MOVE (insn) = 1; | |
1104 | } | |
1105 | ||
1106 | /* If a post-call group is still open, see if it should remain so. | |
1107 | This insn must be a simple move of a hard reg to a pseudo or | |
1108 | vice-versa. | |
1109 | ||
1110 | We must avoid moving these insns for correctness on | |
1111 | SMALL_REGISTER_CLASS machines, and for special registers like | |
1112 | PIC_OFFSET_TABLE_REGNUM. For simplicity, extend this to all | |
1113 | hard regs for all targets. */ | |
1114 | ||
1115 | if (deps->in_post_call_group_p) | |
1116 | { | |
1117 | rtx tmp, set = single_set (insn); | |
1118 | int src_regno, dest_regno; | |
1119 | ||
1120 | if (set == NULL) | |
1121 | goto end_call_group; | |
1122 | ||
1123 | tmp = SET_DEST (set); | |
1124 | if (GET_CODE (tmp) == SUBREG) | |
1125 | tmp = SUBREG_REG (tmp); | |
1126 | if (GET_CODE (tmp) == REG) | |
1127 | dest_regno = REGNO (tmp); | |
1128 | else | |
1129 | goto end_call_group; | |
1130 | ||
1131 | tmp = SET_SRC (set); | |
1132 | if (GET_CODE (tmp) == SUBREG) | |
1133 | tmp = SUBREG_REG (tmp); | |
1134 | if (GET_CODE (tmp) == REG) | |
1135 | src_regno = REGNO (tmp); | |
1136 | else | |
1137 | goto end_call_group; | |
1138 | ||
1139 | if (src_regno < FIRST_PSEUDO_REGISTER | |
1140 | || dest_regno < FIRST_PSEUDO_REGISTER) | |
1141 | { | |
1142 | set_sched_group_p (insn); | |
1143 | CANT_MOVE (insn) = 1; | |
1144 | } | |
1145 | else | |
1146 | { | |
1147 | end_call_group: | |
1148 | deps->in_post_call_group_p = false; | |
1149 | } | |
1150 | } | |
1151 | } | |
1152 | ||
1153 | /* Analyze every insn between HEAD and TAIL inclusive, creating LOG_LINKS | |
1154 | for every dependency. */ | |
1155 | ||
1156 | void | |
1157 | sched_analyze (struct deps *deps, rtx head, rtx tail) | |
1158 | { | |
1159 | rtx insn; | |
1160 | rtx loop_notes = 0; | |
1161 | ||
1162 | if (current_sched_info->use_cselib) | |
1163 | cselib_init (); | |
1164 | ||
1165 | for (insn = head;; insn = NEXT_INSN (insn)) | |
1166 | { | |
1167 | rtx link, end_seq, r0, set; | |
1168 | ||
1169 | if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN) | |
1170 | { | |
1171 | /* Clear out the stale LOG_LINKS from flow. */ | |
1172 | free_INSN_LIST_list (&LOG_LINKS (insn)); | |
1173 | ||
1174 | /* Make each JUMP_INSN a scheduling barrier for memory | |
1175 | references. */ | |
1176 | if (GET_CODE (insn) == JUMP_INSN) | |
1177 | { | |
1178 | /* Keep the list a reasonable size. */ | |
1179 | if (deps->pending_flush_length++ > MAX_PENDING_LIST_LENGTH) | |
1180 | flush_pending_lists (deps, insn, true, true); | |
1181 | else | |
1182 | deps->last_pending_memory_flush | |
1183 | = alloc_INSN_LIST (insn, deps->last_pending_memory_flush); | |
1184 | } | |
1185 | sched_analyze_insn (deps, PATTERN (insn), insn, loop_notes); | |
1186 | loop_notes = 0; | |
1187 | } | |
1188 | else if (GET_CODE (insn) == CALL_INSN) | |
1189 | { | |
1190 | int i; | |
1191 | ||
1192 | CANT_MOVE (insn) = 1; | |
1193 | ||
1194 | /* Clear out the stale LOG_LINKS from flow. */ | |
1195 | free_INSN_LIST_list (&LOG_LINKS (insn)); | |
1196 | ||
1197 | if (find_reg_note (insn, REG_SETJMP, NULL)) | |
1198 | { | |
1199 | /* This is setjmp. Assume that all registers, not just | |
1200 | hard registers, may be clobbered by this call. */ | |
1201 | reg_pending_barrier = MOVE_BARRIER; | |
1202 | } | |
1203 | else | |
1204 | { | |
1205 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
1206 | /* A call may read and modify global register variables. */ | |
1207 | if (global_regs[i]) | |
1208 | { | |
1209 | SET_REGNO_REG_SET (reg_pending_sets, i); | |
1210 | SET_REGNO_REG_SET (reg_pending_uses, i); | |
1211 | } | |
1212 | /* Other call-clobbered hard regs may be clobbered. | |
1213 | Since we only have a choice between 'might be clobbered' | |
1214 | and 'definitely not clobbered', we must include all | |
1215 | partly call-clobbered registers here. */ | |
1216 | else if (HARD_REGNO_CALL_PART_CLOBBERED (i, reg_raw_mode[i]) | |
1217 | || TEST_HARD_REG_BIT (regs_invalidated_by_call, i)) | |
1218 | SET_REGNO_REG_SET (reg_pending_clobbers, i); | |
1219 | /* We don't know what set of fixed registers might be used | |
1220 | by the function, but it is certain that the stack pointer | |
1221 | is among them, but be conservative. */ | |
1222 | else if (fixed_regs[i]) | |
1223 | SET_REGNO_REG_SET (reg_pending_uses, i); | |
1224 | /* The frame pointer is normally not used by the function | |
1225 | itself, but by the debugger. */ | |
1226 | /* ??? MIPS o32 is an exception. It uses the frame pointer | |
1227 | in the macro expansion of jal but does not represent this | |
1228 | fact in the call_insn rtl. */ | |
1229 | else if (i == FRAME_POINTER_REGNUM | |
1230 | || (i == HARD_FRAME_POINTER_REGNUM | |
1231 | && (! reload_completed || frame_pointer_needed))) | |
1232 | SET_REGNO_REG_SET (reg_pending_uses, i); | |
1233 | } | |
1234 | ||
1235 | /* For each insn which shouldn't cross a call, add a dependence | |
1236 | between that insn and this call insn. */ | |
1237 | add_dependence_list_and_free (insn, &deps->sched_before_next_call, | |
1238 | REG_DEP_ANTI); | |
1239 | ||
1240 | sched_analyze_insn (deps, PATTERN (insn), insn, loop_notes); | |
1241 | loop_notes = 0; | |
1242 | ||
1243 | /* In the absence of interprocedural alias analysis, we must flush | |
1244 | all pending reads and writes, and start new dependencies starting | |
1245 | from here. But only flush writes for constant calls (which may | |
1246 | be passed a pointer to something we haven't written yet). */ | |
1247 | flush_pending_lists (deps, insn, true, !CONST_OR_PURE_CALL_P (insn)); | |
1248 | ||
1249 | /* Remember the last function call for limiting lifetimes. */ | |
1250 | free_INSN_LIST_list (&deps->last_function_call); | |
1251 | deps->last_function_call = alloc_INSN_LIST (insn, NULL_RTX); | |
1252 | ||
1253 | /* Before reload, begin a post-call group, so as to keep the | |
1254 | lifetimes of hard registers correct. */ | |
1255 | if (! reload_completed) | |
1256 | deps->in_post_call_group_p = true; | |
1257 | } | |
1258 | ||
1259 | /* See comments on reemit_notes as to why we do this. | |
1260 | ??? Actually, the reemit_notes just say what is done, not why. */ | |
1261 | ||
1262 | if (GET_CODE (insn) == NOTE | |
1263 | && (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG | |
1264 | || NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END | |
1265 | || NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_BEG | |
1266 | || NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_END)) | |
1267 | { | |
1268 | rtx rtx_region; | |
1269 | ||
1270 | if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_BEG | |
1271 | || NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_END) | |
1272 | rtx_region = GEN_INT (NOTE_EH_HANDLER (insn)); | |
1273 | else | |
1274 | rtx_region = GEN_INT (0); | |
1275 | ||
1276 | loop_notes = alloc_EXPR_LIST (REG_SAVE_NOTE, | |
1277 | rtx_region, | |
1278 | loop_notes); | |
1279 | loop_notes = alloc_EXPR_LIST (REG_SAVE_NOTE, | |
1280 | GEN_INT (NOTE_LINE_NUMBER (insn)), | |
1281 | loop_notes); | |
1282 | CONST_OR_PURE_CALL_P (loop_notes) = CONST_OR_PURE_CALL_P (insn); | |
1283 | } | |
1284 | ||
1285 | if (current_sched_info->use_cselib) | |
1286 | cselib_process_insn (insn); | |
1287 | ||
1288 | /* Now that we have completed handling INSN, check and see if it is | |
1289 | a CLOBBER beginning a libcall block. If it is, record the | |
1290 | end of the libcall sequence. | |
1291 | ||
1292 | We want to schedule libcall blocks as a unit before reload. While | |
1293 | this restricts scheduling, it preserves the meaning of a libcall | |
1294 | block. | |
1295 | ||
1296 | As a side effect, we may get better code due to decreased register | |
1297 | pressure as well as less chance of a foreign insn appearing in | |
1298 | a libcall block. */ | |
1299 | if (!reload_completed | |
1300 | /* Note we may have nested libcall sequences. We only care about | |
1301 | the outermost libcall sequence. */ | |
1302 | && deps->libcall_block_tail_insn == 0 | |
1303 | /* The sequence must start with a clobber of a register. */ | |
1304 | && GET_CODE (insn) == INSN | |
1305 | && GET_CODE (PATTERN (insn)) == CLOBBER | |
1306 | && (r0 = XEXP (PATTERN (insn), 0), GET_CODE (r0) == REG) | |
1307 | && GET_CODE (XEXP (PATTERN (insn), 0)) == REG | |
1308 | /* The CLOBBER must also have a REG_LIBCALL note attached. */ | |
1309 | && (link = find_reg_note (insn, REG_LIBCALL, NULL_RTX)) != 0 | |
1310 | && (end_seq = XEXP (link, 0)) != 0 | |
1311 | /* The insn referenced by the REG_LIBCALL note must be a | |
1312 | simple nop copy with the same destination as the register | |
1313 | mentioned in the clobber. */ | |
1314 | && (set = single_set (end_seq)) != 0 | |
1315 | && SET_DEST (set) == r0 && SET_SRC (set) == r0 | |
1316 | /* And finally the insn referenced by the REG_LIBCALL must | |
1317 | also contain a REG_EQUAL note and a REG_RETVAL note. */ | |
1318 | && find_reg_note (end_seq, REG_EQUAL, NULL_RTX) != 0 | |
1319 | && find_reg_note (end_seq, REG_RETVAL, NULL_RTX) != 0) | |
1320 | deps->libcall_block_tail_insn = XEXP (link, 0); | |
1321 | ||
1322 | /* If we have reached the end of a libcall block, then close the | |
1323 | block. */ | |
1324 | if (deps->libcall_block_tail_insn == insn) | |
1325 | deps->libcall_block_tail_insn = 0; | |
1326 | ||
1327 | if (insn == tail) | |
1328 | { | |
1329 | if (current_sched_info->use_cselib) | |
1330 | cselib_finish (); | |
1331 | return; | |
1332 | } | |
1333 | } | |
1334 | abort (); | |
1335 | } | |
1336 | \f | |
1337 | ||
1338 | /* The following function adds forward dependence (FROM, TO) with | |
1339 | given DEP_TYPE. The forward dependence should be not exist before. */ | |
1340 | ||
1341 | void | |
1342 | add_forward_dependence (rtx from, rtx to, enum reg_note dep_type) | |
1343 | { | |
1344 | rtx new_link; | |
1345 | ||
1346 | #ifdef ENABLE_CHECKING | |
1347 | /* If add_dependence is working properly there should never | |
1348 | be notes, deleted insns or duplicates in the backward | |
1349 | links. Thus we need not check for them here. | |
1350 | ||
1351 | However, if we have enabled checking we might as well go | |
1352 | ahead and verify that add_dependence worked properly. */ | |
1353 | if (GET_CODE (from) == NOTE | |
1354 | || INSN_DELETED_P (from) | |
1355 | || (forward_dependency_cache != NULL | |
1356 | && bitmap_bit_p (&forward_dependency_cache[INSN_LUID (from)], | |
1357 | INSN_LUID (to))) | |
1358 | || (forward_dependency_cache == NULL | |
1359 | && find_insn_list (to, INSN_DEPEND (from)))) | |
1360 | abort (); | |
1361 | if (forward_dependency_cache != NULL) | |
1362 | bitmap_bit_p (&forward_dependency_cache[INSN_LUID (from)], | |
1363 | INSN_LUID (to)); | |
1364 | #endif | |
1365 | ||
1366 | new_link = alloc_INSN_LIST (to, INSN_DEPEND (from)); | |
1367 | ||
1368 | PUT_REG_NOTE_KIND (new_link, dep_type); | |
1369 | ||
1370 | INSN_DEPEND (from) = new_link; | |
1371 | INSN_DEP_COUNT (to) += 1; | |
1372 | } | |
1373 | ||
1374 | /* Examine insns in the range [ HEAD, TAIL ] and Use the backward | |
1375 | dependences from LOG_LINKS to build forward dependences in | |
1376 | INSN_DEPEND. */ | |
1377 | ||
1378 | void | |
1379 | compute_forward_dependences (rtx head, rtx tail) | |
1380 | { | |
1381 | rtx insn, link; | |
1382 | rtx next_tail; | |
1383 | ||
1384 | next_tail = NEXT_INSN (tail); | |
1385 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) | |
1386 | { | |
1387 | if (! INSN_P (insn)) | |
1388 | continue; | |
1389 | ||
1390 | for (link = LOG_LINKS (insn); link; link = XEXP (link, 1)) | |
1391 | add_forward_dependence (XEXP (link, 0), insn, REG_NOTE_KIND (link)); | |
1392 | } | |
1393 | } | |
1394 | \f | |
1395 | /* Initialize variables for region data dependence analysis. | |
1396 | n_bbs is the number of region blocks. */ | |
1397 | ||
1398 | void | |
1399 | init_deps (struct deps *deps) | |
1400 | { | |
1401 | int max_reg = (reload_completed ? FIRST_PSEUDO_REGISTER : max_reg_num ()); | |
1402 | ||
1403 | deps->max_reg = max_reg; | |
1404 | deps->reg_last = xcalloc (max_reg, sizeof (struct deps_reg)); | |
1405 | INIT_REG_SET (&deps->reg_last_in_use); | |
1406 | INIT_REG_SET (&deps->reg_conditional_sets); | |
1407 | ||
1408 | deps->pending_read_insns = 0; | |
1409 | deps->pending_read_mems = 0; | |
1410 | deps->pending_write_insns = 0; | |
1411 | deps->pending_write_mems = 0; | |
1412 | deps->pending_lists_length = 0; | |
1413 | deps->pending_flush_length = 0; | |
1414 | deps->last_pending_memory_flush = 0; | |
1415 | deps->last_function_call = 0; | |
1416 | deps->sched_before_next_call = 0; | |
1417 | deps->in_post_call_group_p = false; | |
1418 | deps->libcall_block_tail_insn = 0; | |
1419 | } | |
1420 | ||
1421 | /* Free insn lists found in DEPS. */ | |
1422 | ||
1423 | void | |
1424 | free_deps (struct deps *deps) | |
1425 | { | |
1426 | int i; | |
1427 | ||
1428 | free_INSN_LIST_list (&deps->pending_read_insns); | |
1429 | free_EXPR_LIST_list (&deps->pending_read_mems); | |
1430 | free_INSN_LIST_list (&deps->pending_write_insns); | |
1431 | free_EXPR_LIST_list (&deps->pending_write_mems); | |
1432 | free_INSN_LIST_list (&deps->last_pending_memory_flush); | |
1433 | ||
1434 | /* Without the EXECUTE_IF_SET, this loop is executed max_reg * nr_regions | |
1435 | times. For a testcase with 42000 regs and 8000 small basic blocks, | |
1436 | this loop accounted for nearly 60% (84 sec) of the total -O2 runtime. */ | |
1437 | EXECUTE_IF_SET_IN_REG_SET (&deps->reg_last_in_use, 0, i, | |
1438 | { | |
1439 | struct deps_reg *reg_last = &deps->reg_last[i]; | |
1440 | if (reg_last->uses) | |
1441 | free_INSN_LIST_list (®_last->uses); | |
1442 | if (reg_last->sets) | |
1443 | free_INSN_LIST_list (®_last->sets); | |
1444 | if (reg_last->clobbers) | |
1445 | free_INSN_LIST_list (®_last->clobbers); | |
1446 | }); | |
1447 | CLEAR_REG_SET (&deps->reg_last_in_use); | |
1448 | CLEAR_REG_SET (&deps->reg_conditional_sets); | |
1449 | ||
1450 | free (deps->reg_last); | |
1451 | } | |
1452 | ||
1453 | /* If it is profitable to use them, initialize caches for tracking | |
1454 | dependency information. LUID is the number of insns to be scheduled, | |
1455 | it is used in the estimate of profitability. */ | |
1456 | ||
1457 | void | |
1458 | init_dependency_caches (int luid) | |
1459 | { | |
1460 | /* ?!? We could save some memory by computing a per-region luid mapping | |
1461 | which could reduce both the number of vectors in the cache and the size | |
1462 | of each vector. Instead we just avoid the cache entirely unless the | |
1463 | average number of instructions in a basic block is very high. See | |
1464 | the comment before the declaration of true_dependency_cache for | |
1465 | what we consider "very high". */ | |
1466 | if (luid / n_basic_blocks > 100 * 5) | |
1467 | { | |
1468 | int i; | |
1469 | true_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); | |
1470 | anti_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); | |
1471 | output_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); | |
1472 | #ifdef ENABLE_CHECKING | |
1473 | forward_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); | |
1474 | #endif | |
1475 | for (i = 0; i < luid; i++) | |
1476 | { | |
1477 | bitmap_initialize (&true_dependency_cache[i], 0); | |
1478 | bitmap_initialize (&anti_dependency_cache[i], 0); | |
1479 | bitmap_initialize (&output_dependency_cache[i], 0); | |
1480 | #ifdef ENABLE_CHECKING | |
1481 | bitmap_initialize (&forward_dependency_cache[i], 0); | |
1482 | #endif | |
1483 | } | |
1484 | cache_size = luid; | |
1485 | } | |
1486 | } | |
1487 | ||
1488 | /* Free the caches allocated in init_dependency_caches. */ | |
1489 | ||
1490 | void | |
1491 | free_dependency_caches (void) | |
1492 | { | |
1493 | if (true_dependency_cache) | |
1494 | { | |
1495 | int i; | |
1496 | ||
1497 | for (i = 0; i < cache_size; i++) | |
1498 | { | |
1499 | bitmap_clear (&true_dependency_cache[i]); | |
1500 | bitmap_clear (&anti_dependency_cache[i]); | |
1501 | bitmap_clear (&output_dependency_cache[i]); | |
1502 | #ifdef ENABLE_CHECKING | |
1503 | bitmap_clear (&forward_dependency_cache[i]); | |
1504 | #endif | |
1505 | } | |
1506 | free (true_dependency_cache); | |
1507 | true_dependency_cache = NULL; | |
1508 | free (anti_dependency_cache); | |
1509 | anti_dependency_cache = NULL; | |
1510 | free (output_dependency_cache); | |
1511 | output_dependency_cache = NULL; | |
1512 | #ifdef ENABLE_CHECKING | |
1513 | free (forward_dependency_cache); | |
1514 | forward_dependency_cache = NULL; | |
1515 | #endif | |
1516 | } | |
1517 | } | |
1518 | ||
1519 | /* Initialize some global variables needed by the dependency analysis | |
1520 | code. */ | |
1521 | ||
1522 | void | |
1523 | init_deps_global (void) | |
1524 | { | |
1525 | reg_pending_sets = INITIALIZE_REG_SET (reg_pending_sets_head); | |
1526 | reg_pending_clobbers = INITIALIZE_REG_SET (reg_pending_clobbers_head); | |
1527 | reg_pending_uses = INITIALIZE_REG_SET (reg_pending_uses_head); | |
1528 | reg_pending_barrier = NOT_A_BARRIER; | |
1529 | } | |
1530 | ||
1531 | /* Free everything used by the dependency analysis code. */ | |
1532 | ||
1533 | void | |
1534 | finish_deps_global (void) | |
1535 | { | |
1536 | FREE_REG_SET (reg_pending_sets); | |
1537 | FREE_REG_SET (reg_pending_clobbers); | |
1538 | FREE_REG_SET (reg_pending_uses); | |
1539 | } |