]>
Commit | Line | Data |
---|---|---|
e855c69d | 1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
85ec4feb | 2 | Copyright (C) 2006-2018 Free Software Foundation, Inc. |
e855c69d AB |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8 | Software Foundation; either version 3, or (at your option) any later | |
9 | version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING3. If not see | |
18 | <http://www.gnu.org/licenses/>. */ | |
19 | ||
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
c7131fb2 | 23 | #include "backend.h" |
9fdcd34e | 24 | #include "cfghooks.h" |
c7131fb2 | 25 | #include "tree.h" |
e855c69d | 26 | #include "rtl.h" |
c7131fb2 | 27 | #include "df.h" |
4d0cdd0c | 28 | #include "memmodel.h" |
e855c69d | 29 | #include "tm_p.h" |
60393bbc AM |
30 | #include "cfgrtl.h" |
31 | #include "cfganal.h" | |
32 | #include "cfgbuild.h" | |
e855c69d AB |
33 | #include "insn-config.h" |
34 | #include "insn-attr.h" | |
e855c69d AB |
35 | #include "recog.h" |
36 | #include "params.h" | |
37 | #include "target.h" | |
e855c69d | 38 | #include "sched-int.h" |
5936d944 | 39 | #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */ |
e855c69d AB |
40 | |
41 | #ifdef INSN_SCHEDULING | |
c7131fb2 AM |
42 | #include "regset.h" |
43 | #include "cfgloop.h" | |
e855c69d AB |
44 | #include "sel-sched-ir.h" |
45 | /* We don't have to use it except for sel_print_insn. */ | |
46 | #include "sel-sched-dump.h" | |
47 | ||
48 | /* A vector holding bb info for whole scheduling pass. */ | |
7de76362 | 49 | vec<sel_global_bb_info_def> sel_global_bb_info; |
e855c69d AB |
50 | |
51 | /* A vector holding bb info. */ | |
7de76362 | 52 | vec<sel_region_bb_info_def> sel_region_bb_info; |
e855c69d AB |
53 | |
54 | /* A pool for allocating all lists. */ | |
fcb87c50 | 55 | object_allocator<_list_node> sched_lists_pool ("sel-sched-lists"); |
e855c69d AB |
56 | |
57 | /* This contains information about successors for compute_av_set. */ | |
58 | struct succs_info current_succs; | |
59 | ||
60 | /* Data structure to describe interaction with the generic scheduler utils. */ | |
61 | static struct common_sched_info_def sel_common_sched_info; | |
62 | ||
63 | /* The loop nest being pipelined. */ | |
64 | struct loop *current_loop_nest; | |
65 | ||
66 | /* LOOP_NESTS is a vector containing the corresponding loop nest for | |
67 | each region. */ | |
7de76362 | 68 | static vec<loop_p> loop_nests; |
e855c69d AB |
69 | |
70 | /* Saves blocks already in loop regions, indexed by bb->index. */ | |
71 | static sbitmap bbs_in_loop_rgns = NULL; | |
72 | ||
73 | /* CFG hooks that are saved before changing create_basic_block hook. */ | |
74 | static struct cfg_hooks orig_cfg_hooks; | |
75 | \f | |
76 | ||
77 | /* Array containing reverse topological index of function basic blocks, | |
78 | indexed by BB->INDEX. */ | |
79 | static int *rev_top_order_index = NULL; | |
80 | ||
81 | /* Length of the above array. */ | |
82 | static int rev_top_order_index_len = -1; | |
83 | ||
84 | /* A regset pool structure. */ | |
85 | static struct | |
86 | { | |
87 | /* The stack to which regsets are returned. */ | |
88 | regset *v; | |
89 | ||
90 | /* Its pointer. */ | |
91 | int n; | |
92 | ||
93 | /* Its size. */ | |
94 | int s; | |
95 | ||
96 | /* In VV we save all generated regsets so that, when destructing the | |
97 | pool, we can compare it with V and check that every regset was returned | |
98 | back to pool. */ | |
99 | regset *vv; | |
100 | ||
101 | /* The pointer of VV stack. */ | |
102 | int nn; | |
103 | ||
104 | /* Its size. */ | |
105 | int ss; | |
106 | ||
107 | /* The difference between allocated and returned regsets. */ | |
108 | int diff; | |
109 | } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 }; | |
110 | ||
111 | /* This represents the nop pool. */ | |
112 | static struct | |
113 | { | |
114 | /* The vector which holds previously emitted nops. */ | |
115 | insn_t *v; | |
116 | ||
117 | /* Its pointer. */ | |
118 | int n; | |
119 | ||
120 | /* Its size. */ | |
b8698a0f | 121 | int s; |
e855c69d AB |
122 | } nop_pool = { NULL, 0, 0 }; |
123 | ||
124 | /* The pool for basic block notes. */ | |
66e8df53 | 125 | static vec<rtx_note *> bb_note_pool; |
e855c69d AB |
126 | |
127 | /* A NOP pattern used to emit placeholder insns. */ | |
128 | rtx nop_pattern = NULL_RTX; | |
129 | /* A special instruction that resides in EXIT_BLOCK. | |
130 | EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */ | |
c5db5458 | 131 | rtx_insn *exit_insn = NULL; |
e855c69d | 132 | |
b8698a0f | 133 | /* TRUE if while scheduling current region, which is loop, its preheader |
e855c69d AB |
134 | was removed. */ |
135 | bool preheader_removed = false; | |
136 | \f | |
137 | ||
138 | /* Forward static declarations. */ | |
139 | static void fence_clear (fence_t); | |
140 | ||
141 | static void deps_init_id (idata_t, insn_t, bool); | |
142 | static void init_id_from_df (idata_t, insn_t, bool); | |
143 | static expr_t set_insn_init (expr_t, vinsn_t, int); | |
144 | ||
145 | static void cfg_preds (basic_block, insn_t **, int *); | |
146 | static void prepare_insn_expr (insn_t, int); | |
9771b263 | 147 | static void free_history_vect (vec<expr_history_def> &); |
e855c69d AB |
148 | |
149 | static void move_bb_info (basic_block, basic_block); | |
150 | static void remove_empty_bb (basic_block, bool); | |
262d8232 | 151 | static void sel_merge_blocks (basic_block, basic_block); |
e855c69d | 152 | static void sel_remove_loop_preheader (void); |
753de8cf | 153 | static bool bb_has_removable_jump_to_p (basic_block, basic_block); |
e855c69d AB |
154 | |
155 | static bool insn_is_the_only_one_in_bb_p (insn_t); | |
156 | static void create_initial_data_sets (basic_block); | |
157 | ||
b5b8b0ac | 158 | static void free_av_set (basic_block); |
e855c69d AB |
159 | static void invalidate_av_set (basic_block); |
160 | static void extend_insn_data (void); | |
92e265ac | 161 | static void sel_init_new_insn (insn_t, int, int = -1); |
e855c69d AB |
162 | static void finish_insns (void); |
163 | \f | |
164 | /* Various list functions. */ | |
165 | ||
166 | /* Copy an instruction list L. */ | |
167 | ilist_t | |
168 | ilist_copy (ilist_t l) | |
169 | { | |
170 | ilist_t head = NULL, *tailp = &head; | |
171 | ||
172 | while (l) | |
173 | { | |
174 | ilist_add (tailp, ILIST_INSN (l)); | |
175 | tailp = &ILIST_NEXT (*tailp); | |
176 | l = ILIST_NEXT (l); | |
177 | } | |
178 | ||
179 | return head; | |
180 | } | |
181 | ||
182 | /* Invert an instruction list L. */ | |
183 | ilist_t | |
184 | ilist_invert (ilist_t l) | |
185 | { | |
186 | ilist_t res = NULL; | |
187 | ||
188 | while (l) | |
189 | { | |
190 | ilist_add (&res, ILIST_INSN (l)); | |
191 | l = ILIST_NEXT (l); | |
192 | } | |
193 | ||
194 | return res; | |
195 | } | |
196 | ||
197 | /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */ | |
198 | void | |
199 | blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc) | |
200 | { | |
201 | bnd_t bnd; | |
202 | ||
203 | _list_add (lp); | |
204 | bnd = BLIST_BND (*lp); | |
205 | ||
6144a836 | 206 | BND_TO (bnd) = to; |
e855c69d AB |
207 | BND_PTR (bnd) = ptr; |
208 | BND_AV (bnd) = NULL; | |
209 | BND_AV1 (bnd) = NULL; | |
210 | BND_DC (bnd) = dc; | |
211 | } | |
212 | ||
213 | /* Remove the list note pointed to by LP. */ | |
214 | void | |
215 | blist_remove (blist_t *lp) | |
216 | { | |
217 | bnd_t b = BLIST_BND (*lp); | |
218 | ||
219 | av_set_clear (&BND_AV (b)); | |
220 | av_set_clear (&BND_AV1 (b)); | |
221 | ilist_clear (&BND_PTR (b)); | |
222 | ||
223 | _list_remove (lp); | |
224 | } | |
225 | ||
226 | /* Init a fence tail L. */ | |
227 | void | |
228 | flist_tail_init (flist_tail_t l) | |
229 | { | |
230 | FLIST_TAIL_HEAD (l) = NULL; | |
231 | FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l); | |
232 | } | |
233 | ||
234 | /* Try to find fence corresponding to INSN in L. */ | |
235 | fence_t | |
236 | flist_lookup (flist_t l, insn_t insn) | |
237 | { | |
238 | while (l) | |
239 | { | |
240 | if (FENCE_INSN (FLIST_FENCE (l)) == insn) | |
241 | return FLIST_FENCE (l); | |
242 | ||
243 | l = FLIST_NEXT (l); | |
244 | } | |
245 | ||
246 | return NULL; | |
247 | } | |
248 | ||
249 | /* Init the fields of F before running fill_insns. */ | |
250 | static void | |
251 | init_fence_for_scheduling (fence_t f) | |
252 | { | |
253 | FENCE_BNDS (f) = NULL; | |
254 | FENCE_PROCESSED_P (f) = false; | |
255 | FENCE_SCHEDULED_P (f) = false; | |
256 | } | |
257 | ||
258 | /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */ | |
259 | static void | |
b8698a0f | 260 | flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, |
6144a836 | 261 | insn_t last_scheduled_insn, vec<rtx_insn *, va_gc> *executing_insns, |
b8698a0f | 262 | int *ready_ticks, int ready_ticks_size, insn_t sched_next, |
136e01a3 | 263 | int cycle, int cycle_issued_insns, int issue_more, |
e855c69d AB |
264 | bool starts_cycle_p, bool after_stall_p) |
265 | { | |
266 | fence_t f; | |
267 | ||
268 | _list_add (lp); | |
269 | f = FLIST_FENCE (*lp); | |
270 | ||
271 | FENCE_INSN (f) = insn; | |
272 | ||
273 | gcc_assert (state != NULL); | |
274 | FENCE_STATE (f) = state; | |
275 | ||
276 | FENCE_CYCLE (f) = cycle; | |
277 | FENCE_ISSUED_INSNS (f) = cycle_issued_insns; | |
278 | FENCE_STARTS_CYCLE_P (f) = starts_cycle_p; | |
279 | FENCE_AFTER_STALL_P (f) = after_stall_p; | |
280 | ||
281 | gcc_assert (dc != NULL); | |
282 | FENCE_DC (f) = dc; | |
283 | ||
284 | gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL); | |
285 | FENCE_TC (f) = tc; | |
286 | ||
287 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
136e01a3 | 288 | FENCE_ISSUE_MORE (f) = issue_more; |
e855c69d AB |
289 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
290 | FENCE_READY_TICKS (f) = ready_ticks; | |
291 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
292 | FENCE_SCHED_NEXT (f) = sched_next; | |
293 | ||
294 | init_fence_for_scheduling (f); | |
295 | } | |
296 | ||
297 | /* Remove the head node of the list pointed to by LP. */ | |
298 | static void | |
299 | flist_remove (flist_t *lp) | |
300 | { | |
301 | if (FENCE_INSN (FLIST_FENCE (*lp))) | |
302 | fence_clear (FLIST_FENCE (*lp)); | |
303 | _list_remove (lp); | |
304 | } | |
305 | ||
306 | /* Clear the fence list pointed to by LP. */ | |
307 | void | |
308 | flist_clear (flist_t *lp) | |
309 | { | |
310 | while (*lp) | |
311 | flist_remove (lp); | |
312 | } | |
313 | ||
314 | /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */ | |
315 | void | |
316 | def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call) | |
317 | { | |
318 | def_t d; | |
b8698a0f | 319 | |
e855c69d AB |
320 | _list_add (dl); |
321 | d = DEF_LIST_DEF (*dl); | |
322 | ||
323 | d->orig_insn = original_insn; | |
324 | d->crosses_call = crosses_call; | |
325 | } | |
326 | \f | |
327 | ||
328 | /* Functions to work with target contexts. */ | |
329 | ||
b8698a0f | 330 | /* Bulk target context. It is convenient for debugging purposes to ensure |
e855c69d AB |
331 | that there are no uninitialized (null) target contexts. */ |
332 | static tc_t bulk_tc = (tc_t) 1; | |
333 | ||
b8698a0f | 334 | /* Target hooks wrappers. In the future we can provide some default |
e855c69d AB |
335 | implementations for them. */ |
336 | ||
337 | /* Allocate a store for the target context. */ | |
338 | static tc_t | |
339 | alloc_target_context (void) | |
340 | { | |
341 | return (targetm.sched.alloc_sched_context | |
342 | ? targetm.sched.alloc_sched_context () : bulk_tc); | |
343 | } | |
344 | ||
345 | /* Init target context TC. | |
346 | If CLEAN_P is true, then make TC as it is beginning of the scheduler. | |
347 | Overwise, copy current backend context to TC. */ | |
348 | static void | |
349 | init_target_context (tc_t tc, bool clean_p) | |
350 | { | |
351 | if (targetm.sched.init_sched_context) | |
352 | targetm.sched.init_sched_context (tc, clean_p); | |
353 | } | |
354 | ||
355 | /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as | |
356 | int init_target_context (). */ | |
357 | tc_t | |
358 | create_target_context (bool clean_p) | |
359 | { | |
360 | tc_t tc = alloc_target_context (); | |
361 | ||
362 | init_target_context (tc, clean_p); | |
363 | return tc; | |
364 | } | |
365 | ||
366 | /* Copy TC to the current backend context. */ | |
367 | void | |
368 | set_target_context (tc_t tc) | |
369 | { | |
370 | if (targetm.sched.set_sched_context) | |
371 | targetm.sched.set_sched_context (tc); | |
372 | } | |
373 | ||
374 | /* TC is about to be destroyed. Free any internal data. */ | |
375 | static void | |
376 | clear_target_context (tc_t tc) | |
377 | { | |
378 | if (targetm.sched.clear_sched_context) | |
379 | targetm.sched.clear_sched_context (tc); | |
380 | } | |
381 | ||
382 | /* Clear and free it. */ | |
383 | static void | |
384 | delete_target_context (tc_t tc) | |
385 | { | |
386 | clear_target_context (tc); | |
387 | ||
388 | if (targetm.sched.free_sched_context) | |
389 | targetm.sched.free_sched_context (tc); | |
390 | } | |
391 | ||
392 | /* Make a copy of FROM in TO. | |
393 | NB: May be this should be a hook. */ | |
394 | static void | |
395 | copy_target_context (tc_t to, tc_t from) | |
396 | { | |
397 | tc_t tmp = create_target_context (false); | |
398 | ||
399 | set_target_context (from); | |
400 | init_target_context (to, false); | |
401 | ||
402 | set_target_context (tmp); | |
403 | delete_target_context (tmp); | |
404 | } | |
405 | ||
406 | /* Create a copy of TC. */ | |
407 | static tc_t | |
408 | create_copy_of_target_context (tc_t tc) | |
409 | { | |
410 | tc_t copy = alloc_target_context (); | |
411 | ||
412 | copy_target_context (copy, tc); | |
413 | ||
414 | return copy; | |
415 | } | |
416 | ||
417 | /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P | |
418 | is the same as in init_target_context (). */ | |
419 | void | |
420 | reset_target_context (tc_t tc, bool clean_p) | |
421 | { | |
422 | clear_target_context (tc); | |
423 | init_target_context (tc, clean_p); | |
424 | } | |
425 | \f | |
b8698a0f | 426 | /* Functions to work with dependence contexts. |
88302d54 | 427 | Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence |
e855c69d AB |
428 | context. It accumulates information about processed insns to decide if |
429 | current insn is dependent on the processed ones. */ | |
430 | ||
431 | /* Make a copy of FROM in TO. */ | |
432 | static void | |
433 | copy_deps_context (deps_t to, deps_t from) | |
434 | { | |
bcf33775 | 435 | init_deps (to, false); |
e855c69d AB |
436 | deps_join (to, from); |
437 | } | |
438 | ||
439 | /* Allocate store for dep context. */ | |
440 | static deps_t | |
441 | alloc_deps_context (void) | |
442 | { | |
88302d54 | 443 | return XNEW (struct deps_desc); |
e855c69d AB |
444 | } |
445 | ||
446 | /* Allocate and initialize dep context. */ | |
447 | static deps_t | |
448 | create_deps_context (void) | |
449 | { | |
450 | deps_t dc = alloc_deps_context (); | |
451 | ||
bcf33775 | 452 | init_deps (dc, false); |
e855c69d AB |
453 | return dc; |
454 | } | |
455 | ||
456 | /* Create a copy of FROM. */ | |
457 | static deps_t | |
458 | create_copy_of_deps_context (deps_t from) | |
459 | { | |
460 | deps_t to = alloc_deps_context (); | |
461 | ||
462 | copy_deps_context (to, from); | |
463 | return to; | |
464 | } | |
465 | ||
466 | /* Clean up internal data of DC. */ | |
467 | static void | |
468 | clear_deps_context (deps_t dc) | |
469 | { | |
470 | free_deps (dc); | |
471 | } | |
472 | ||
473 | /* Clear and free DC. */ | |
474 | static void | |
475 | delete_deps_context (deps_t dc) | |
476 | { | |
477 | clear_deps_context (dc); | |
478 | free (dc); | |
479 | } | |
480 | ||
481 | /* Clear and init DC. */ | |
482 | static void | |
483 | reset_deps_context (deps_t dc) | |
484 | { | |
485 | clear_deps_context (dc); | |
bcf33775 | 486 | init_deps (dc, false); |
e855c69d AB |
487 | } |
488 | ||
b8698a0f | 489 | /* This structure describes the dependence analysis hooks for advancing |
e855c69d AB |
490 | dependence context. */ |
491 | static struct sched_deps_info_def advance_deps_context_sched_deps_info = | |
492 | { | |
493 | NULL, | |
494 | ||
495 | NULL, /* start_insn */ | |
496 | NULL, /* finish_insn */ | |
497 | NULL, /* start_lhs */ | |
498 | NULL, /* finish_lhs */ | |
499 | NULL, /* start_rhs */ | |
500 | NULL, /* finish_rhs */ | |
501 | haifa_note_reg_set, | |
502 | haifa_note_reg_clobber, | |
503 | haifa_note_reg_use, | |
504 | NULL, /* note_mem_dep */ | |
505 | NULL, /* note_dep */ | |
506 | ||
507 | 0, 0, 0 | |
508 | }; | |
509 | ||
510 | /* Process INSN and add its impact on DC. */ | |
511 | void | |
512 | advance_deps_context (deps_t dc, insn_t insn) | |
513 | { | |
514 | sched_deps_info = &advance_deps_context_sched_deps_info; | |
6144a836 | 515 | deps_analyze_insn (dc, insn); |
e855c69d AB |
516 | } |
517 | \f | |
518 | ||
519 | /* Functions to work with DFA states. */ | |
520 | ||
521 | /* Allocate store for a DFA state. */ | |
522 | static state_t | |
523 | state_alloc (void) | |
524 | { | |
525 | return xmalloc (dfa_state_size); | |
526 | } | |
527 | ||
528 | /* Allocate and initialize DFA state. */ | |
529 | static state_t | |
530 | state_create (void) | |
531 | { | |
532 | state_t state = state_alloc (); | |
533 | ||
534 | state_reset (state); | |
535 | advance_state (state); | |
536 | return state; | |
537 | } | |
538 | ||
539 | /* Free DFA state. */ | |
540 | static void | |
541 | state_free (state_t state) | |
542 | { | |
543 | free (state); | |
544 | } | |
545 | ||
546 | /* Make a copy of FROM in TO. */ | |
547 | static void | |
548 | state_copy (state_t to, state_t from) | |
549 | { | |
550 | memcpy (to, from, dfa_state_size); | |
551 | } | |
552 | ||
553 | /* Create a copy of FROM. */ | |
554 | static state_t | |
555 | state_create_copy (state_t from) | |
556 | { | |
557 | state_t to = state_alloc (); | |
558 | ||
559 | state_copy (to, from); | |
560 | return to; | |
561 | } | |
562 | \f | |
563 | ||
564 | /* Functions to work with fences. */ | |
565 | ||
566 | /* Clear the fence. */ | |
567 | static void | |
568 | fence_clear (fence_t f) | |
569 | { | |
570 | state_t s = FENCE_STATE (f); | |
571 | deps_t dc = FENCE_DC (f); | |
572 | void *tc = FENCE_TC (f); | |
573 | ||
574 | ilist_clear (&FENCE_BNDS (f)); | |
575 | ||
576 | gcc_assert ((s != NULL && dc != NULL && tc != NULL) | |
577 | || (s == NULL && dc == NULL && tc == NULL)); | |
578 | ||
04695783 | 579 | free (s); |
e855c69d AB |
580 | |
581 | if (dc != NULL) | |
582 | delete_deps_context (dc); | |
583 | ||
584 | if (tc != NULL) | |
585 | delete_target_context (tc); | |
9771b263 | 586 | vec_free (FENCE_EXECUTING_INSNS (f)); |
e855c69d AB |
587 | free (FENCE_READY_TICKS (f)); |
588 | FENCE_READY_TICKS (f) = NULL; | |
589 | } | |
590 | ||
591 | /* Init a list of fences with successors of OLD_FENCE. */ | |
592 | void | |
593 | init_fences (insn_t old_fence) | |
594 | { | |
595 | insn_t succ; | |
596 | succ_iterator si; | |
597 | bool first = true; | |
598 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f L |
599 | |
600 | FOR_EACH_SUCC_1 (succ, si, old_fence, | |
e855c69d AB |
601 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
602 | { | |
b8698a0f | 603 | |
e855c69d AB |
604 | if (first) |
605 | first = false; | |
606 | else | |
607 | gcc_assert (flag_sel_sched_pipelining_outer_loops); | |
608 | ||
609 | flist_add (&fences, succ, | |
610 | state_create (), | |
611 | create_deps_context () /* dc */, | |
612 | create_target_context (true) /* tc */, | |
6144a836 | 613 | NULL /* last_scheduled_insn */, |
e855c69d AB |
614 | NULL, /* executing_insns */ |
615 | XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ | |
616 | ready_ticks_size, | |
6144a836 | 617 | NULL /* sched_next */, |
b8698a0f | 618 | 1 /* cycle */, 0 /* cycle_issued_insns */, |
136e01a3 | 619 | issue_rate, /* issue_more */ |
b8698a0f | 620 | 1 /* starts_cycle_p */, 0 /* after_stall_p */); |
e855c69d AB |
621 | } |
622 | } | |
623 | ||
624 | /* Merges two fences (filling fields of fence F with resulting values) by | |
625 | following rules: 1) state, target context and last scheduled insn are | |
b8698a0f | 626 | propagated from fallthrough edge if it is available; |
e855c69d | 627 | 2) deps context and cycle is propagated from more probable edge; |
b8698a0f | 628 | 3) all other fields are set to corresponding constant values. |
e855c69d | 629 | |
b8698a0f | 630 | INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, |
136e01a3 AB |
631 | READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE |
632 | and AFTER_STALL_P are the corresponding fields of the second fence. */ | |
e855c69d AB |
633 | static void |
634 | merge_fences (fence_t f, insn_t insn, | |
b8698a0f | 635 | state_t state, deps_t dc, void *tc, |
6144a836 DM |
636 | rtx_insn *last_scheduled_insn, |
637 | vec<rtx_insn *, va_gc> *executing_insns, | |
e855c69d | 638 | int *ready_ticks, int ready_ticks_size, |
136e01a3 | 639 | rtx sched_next, int cycle, int issue_more, bool after_stall_p) |
e855c69d AB |
640 | { |
641 | insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f); | |
642 | ||
643 | gcc_assert (sel_bb_head_p (FENCE_INSN (f)) | |
644 | && !sched_next && !FENCE_SCHED_NEXT (f)); | |
645 | ||
b8698a0f | 646 | /* Check if we can decide which path fences came. |
e855c69d AB |
647 | If we can't (or don't want to) - reset all. */ |
648 | if (last_scheduled_insn == NULL | |
649 | || last_scheduled_insn_old == NULL | |
b8698a0f L |
650 | /* This is a case when INSN is reachable on several paths from |
651 | one insn (this can happen when pipelining of outer loops is on and | |
652 | there are two edges: one going around of inner loop and the other - | |
e855c69d AB |
653 | right through it; in such case just reset everything). */ |
654 | || last_scheduled_insn == last_scheduled_insn_old) | |
655 | { | |
656 | state_reset (FENCE_STATE (f)); | |
657 | state_free (state); | |
b8698a0f | 658 | |
e855c69d AB |
659 | reset_deps_context (FENCE_DC (f)); |
660 | delete_deps_context (dc); | |
b8698a0f | 661 | |
e855c69d AB |
662 | reset_target_context (FENCE_TC (f), true); |
663 | delete_target_context (tc); | |
664 | ||
665 | if (cycle > FENCE_CYCLE (f)) | |
666 | FENCE_CYCLE (f) = cycle; | |
667 | ||
668 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; | |
136e01a3 | 669 | FENCE_ISSUE_MORE (f) = issue_rate; |
9771b263 | 670 | vec_free (executing_insns); |
e855c69d AB |
671 | free (ready_ticks); |
672 | if (FENCE_EXECUTING_INSNS (f)) | |
9771b263 DN |
673 | FENCE_EXECUTING_INSNS (f)->block_remove (0, |
674 | FENCE_EXECUTING_INSNS (f)->length ()); | |
e855c69d AB |
675 | if (FENCE_READY_TICKS (f)) |
676 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
677 | } | |
678 | else | |
679 | { | |
680 | edge edge_old = NULL, edge_new = NULL; | |
681 | edge candidate; | |
682 | succ_iterator si; | |
683 | insn_t succ; | |
b8698a0f | 684 | |
e855c69d AB |
685 | /* Find fallthrough edge. */ |
686 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); | |
0fd4b31d | 687 | candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb); |
e855c69d AB |
688 | |
689 | if (!candidate | |
690 | || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn) | |
691 | && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old))) | |
692 | { | |
693 | /* No fallthrough edge leading to basic block of INSN. */ | |
694 | state_reset (FENCE_STATE (f)); | |
695 | state_free (state); | |
b8698a0f | 696 | |
e855c69d AB |
697 | reset_target_context (FENCE_TC (f), true); |
698 | delete_target_context (tc); | |
b8698a0f | 699 | |
e855c69d | 700 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; |
136e01a3 | 701 | FENCE_ISSUE_MORE (f) = issue_rate; |
e855c69d AB |
702 | } |
703 | else | |
704 | if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn)) | |
705 | { | |
b8698a0f | 706 | /* Would be weird if same insn is successor of several fallthrough |
e855c69d AB |
707 | edges. */ |
708 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
709 | != BLOCK_FOR_INSN (last_scheduled_insn_old)); | |
710 | ||
711 | state_free (FENCE_STATE (f)); | |
712 | FENCE_STATE (f) = state; | |
713 | ||
714 | delete_target_context (FENCE_TC (f)); | |
715 | FENCE_TC (f) = tc; | |
716 | ||
717 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
136e01a3 | 718 | FENCE_ISSUE_MORE (f) = issue_more; |
e855c69d AB |
719 | } |
720 | else | |
721 | { | |
722 | /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ | |
723 | state_free (state); | |
724 | delete_target_context (tc); | |
725 | ||
726 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
727 | != BLOCK_FOR_INSN (last_scheduled_insn)); | |
728 | } | |
729 | ||
730 | /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ | |
731 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, | |
732 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
733 | { | |
734 | if (succ == insn) | |
735 | { | |
736 | /* No same successor allowed from several edges. */ | |
737 | gcc_assert (!edge_old); | |
738 | edge_old = si.e1; | |
739 | } | |
740 | } | |
741 | /* Find edge of second predecessor (last_scheduled_insn->insn). */ | |
742 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, | |
743 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
744 | { | |
745 | if (succ == insn) | |
746 | { | |
747 | /* No same successor allowed from several edges. */ | |
748 | gcc_assert (!edge_new); | |
749 | edge_new = si.e1; | |
750 | } | |
751 | } | |
752 | ||
753 | /* Check if we can choose most probable predecessor. */ | |
754 | if (edge_old == NULL || edge_new == NULL) | |
755 | { | |
756 | reset_deps_context (FENCE_DC (f)); | |
757 | delete_deps_context (dc); | |
9771b263 | 758 | vec_free (executing_insns); |
e855c69d | 759 | free (ready_ticks); |
b8698a0f | 760 | |
e855c69d AB |
761 | FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); |
762 | if (FENCE_EXECUTING_INSNS (f)) | |
9771b263 DN |
763 | FENCE_EXECUTING_INSNS (f)->block_remove (0, |
764 | FENCE_EXECUTING_INSNS (f)->length ()); | |
e855c69d AB |
765 | if (FENCE_READY_TICKS (f)) |
766 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
767 | } | |
768 | else | |
769 | if (edge_new->probability > edge_old->probability) | |
770 | { | |
771 | delete_deps_context (FENCE_DC (f)); | |
772 | FENCE_DC (f) = dc; | |
9771b263 | 773 | vec_free (FENCE_EXECUTING_INSNS (f)); |
e855c69d AB |
774 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
775 | free (FENCE_READY_TICKS (f)); | |
776 | FENCE_READY_TICKS (f) = ready_ticks; | |
777 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
778 | FENCE_CYCLE (f) = cycle; | |
779 | } | |
780 | else | |
781 | { | |
782 | /* Leave DC and CYCLE untouched. */ | |
783 | delete_deps_context (dc); | |
9771b263 | 784 | vec_free (executing_insns); |
e855c69d AB |
785 | free (ready_ticks); |
786 | } | |
787 | } | |
788 | ||
789 | /* Fill remaining invariant fields. */ | |
790 | if (after_stall_p) | |
791 | FENCE_AFTER_STALL_P (f) = 1; | |
792 | ||
793 | FENCE_ISSUED_INSNS (f) = 0; | |
794 | FENCE_STARTS_CYCLE_P (f) = 1; | |
795 | FENCE_SCHED_NEXT (f) = NULL; | |
796 | } | |
797 | ||
b8698a0f | 798 | /* Add a new fence to NEW_FENCES list, initializing it from all |
e855c69d AB |
799 | other parameters. */ |
800 | static void | |
801 | add_to_fences (flist_tail_t new_fences, insn_t insn, | |
6144a836 DM |
802 | state_t state, deps_t dc, void *tc, |
803 | rtx_insn *last_scheduled_insn, | |
804 | vec<rtx_insn *, va_gc> *executing_insns, int *ready_ticks, | |
805 | int ready_ticks_size, rtx_insn *sched_next, int cycle, | |
136e01a3 AB |
806 | int cycle_issued_insns, int issue_rate, |
807 | bool starts_cycle_p, bool after_stall_p) | |
e855c69d AB |
808 | { |
809 | fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn); | |
810 | ||
811 | if (! f) | |
812 | { | |
813 | flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc, | |
b8698a0f | 814 | last_scheduled_insn, executing_insns, ready_ticks, |
e855c69d | 815 | ready_ticks_size, sched_next, cycle, cycle_issued_insns, |
136e01a3 | 816 | issue_rate, starts_cycle_p, after_stall_p); |
e855c69d AB |
817 | |
818 | FLIST_TAIL_TAILP (new_fences) | |
819 | = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences)); | |
820 | } | |
821 | else | |
822 | { | |
b8698a0f L |
823 | merge_fences (f, insn, state, dc, tc, last_scheduled_insn, |
824 | executing_insns, ready_ticks, ready_ticks_size, | |
136e01a3 | 825 | sched_next, cycle, issue_rate, after_stall_p); |
e855c69d AB |
826 | } |
827 | } | |
828 | ||
829 | /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */ | |
830 | void | |
831 | move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences) | |
832 | { | |
833 | fence_t f, old; | |
834 | flist_t *tailp = FLIST_TAIL_TAILP (new_fences); | |
835 | ||
836 | old = FLIST_FENCE (old_fences); | |
b8698a0f | 837 | f = flist_lookup (FLIST_TAIL_HEAD (new_fences), |
e855c69d AB |
838 | FENCE_INSN (FLIST_FENCE (old_fences))); |
839 | if (f) | |
840 | { | |
841 | merge_fences (f, old->insn, old->state, old->dc, old->tc, | |
842 | old->last_scheduled_insn, old->executing_insns, | |
843 | old->ready_ticks, old->ready_ticks_size, | |
136e01a3 | 844 | old->sched_next, old->cycle, old->issue_more, |
e855c69d AB |
845 | old->after_stall_p); |
846 | } | |
847 | else | |
848 | { | |
849 | _list_add (tailp); | |
850 | FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp); | |
851 | *FLIST_FENCE (*tailp) = *old; | |
852 | init_fence_for_scheduling (FLIST_FENCE (*tailp)); | |
853 | } | |
854 | FENCE_INSN (old) = NULL; | |
855 | } | |
856 | ||
b8698a0f | 857 | /* Add a new fence to NEW_FENCES list and initialize most of its data |
e855c69d AB |
858 | as a clean one. */ |
859 | void | |
860 | add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
861 | { | |
862 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f | 863 | |
e855c69d AB |
864 | add_to_fences (new_fences, |
865 | succ, state_create (), create_deps_context (), | |
866 | create_target_context (true), | |
6144a836 | 867 | NULL, NULL, |
e855c69d | 868 | XCNEWVEC (int, ready_ticks_size), ready_ticks_size, |
6144a836 | 869 | NULL, FENCE_CYCLE (fence) + 1, |
136e01a3 | 870 | 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence)); |
e855c69d AB |
871 | } |
872 | ||
b8698a0f | 873 | /* Add a new fence to NEW_FENCES list and initialize all of its data |
e855c69d AB |
874 | from FENCE and SUCC. */ |
875 | void | |
876 | add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
877 | { | |
b8698a0f | 878 | int * new_ready_ticks |
e855c69d | 879 | = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); |
b8698a0f | 880 | |
e855c69d AB |
881 | memcpy (new_ready_ticks, FENCE_READY_TICKS (fence), |
882 | FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); | |
883 | add_to_fences (new_fences, | |
884 | succ, state_create_copy (FENCE_STATE (fence)), | |
885 | create_copy_of_deps_context (FENCE_DC (fence)), | |
886 | create_copy_of_target_context (FENCE_TC (fence)), | |
b8698a0f | 887 | FENCE_LAST_SCHEDULED_INSN (fence), |
9771b263 | 888 | vec_safe_copy (FENCE_EXECUTING_INSNS (fence)), |
e855c69d AB |
889 | new_ready_ticks, |
890 | FENCE_READY_TICKS_SIZE (fence), | |
891 | FENCE_SCHED_NEXT (fence), | |
892 | FENCE_CYCLE (fence), | |
893 | FENCE_ISSUED_INSNS (fence), | |
136e01a3 | 894 | FENCE_ISSUE_MORE (fence), |
e855c69d AB |
895 | FENCE_STARTS_CYCLE_P (fence), |
896 | FENCE_AFTER_STALL_P (fence)); | |
897 | } | |
898 | \f | |
899 | ||
900 | /* Functions to work with regset and nop pools. */ | |
901 | ||
902 | /* Returns the new regset from pool. It might have some of the bits set | |
903 | from the previous usage. */ | |
904 | regset | |
905 | get_regset_from_pool (void) | |
906 | { | |
907 | regset rs; | |
908 | ||
909 | if (regset_pool.n != 0) | |
910 | rs = regset_pool.v[--regset_pool.n]; | |
911 | else | |
912 | /* We need to create the regset. */ | |
913 | { | |
914 | rs = ALLOC_REG_SET (®_obstack); | |
915 | ||
916 | if (regset_pool.nn == regset_pool.ss) | |
917 | regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, | |
918 | (regset_pool.ss = 2 * regset_pool.ss + 1)); | |
919 | regset_pool.vv[regset_pool.nn++] = rs; | |
920 | } | |
921 | ||
922 | regset_pool.diff++; | |
923 | ||
924 | return rs; | |
925 | } | |
926 | ||
927 | /* Same as above, but returns the empty regset. */ | |
928 | regset | |
929 | get_clear_regset_from_pool (void) | |
930 | { | |
931 | regset rs = get_regset_from_pool (); | |
932 | ||
933 | CLEAR_REG_SET (rs); | |
934 | return rs; | |
935 | } | |
936 | ||
937 | /* Return regset RS to the pool for future use. */ | |
938 | void | |
939 | return_regset_to_pool (regset rs) | |
940 | { | |
9ef1bf71 | 941 | gcc_assert (rs); |
e855c69d AB |
942 | regset_pool.diff--; |
943 | ||
944 | if (regset_pool.n == regset_pool.s) | |
945 | regset_pool.v = XRESIZEVEC (regset, regset_pool.v, | |
946 | (regset_pool.s = 2 * regset_pool.s + 1)); | |
947 | regset_pool.v[regset_pool.n++] = rs; | |
948 | } | |
949 | ||
950 | /* This is used as a qsort callback for sorting regset pool stacks. | |
951 | X and XX are addresses of two regsets. They are never equal. */ | |
952 | static int | |
953 | cmp_v_in_regset_pool (const void *x, const void *xx) | |
954 | { | |
d38933a0 SB |
955 | uintptr_t r1 = (uintptr_t) *((const regset *) x); |
956 | uintptr_t r2 = (uintptr_t) *((const regset *) xx); | |
957 | if (r1 > r2) | |
958 | return 1; | |
959 | else if (r1 < r2) | |
960 | return -1; | |
961 | gcc_unreachable (); | |
e855c69d AB |
962 | } |
963 | ||
b2b29377 | 964 | /* Free the regset pool possibly checking for memory leaks. */ |
e855c69d AB |
965 | void |
966 | free_regset_pool (void) | |
967 | { | |
b2b29377 MM |
968 | if (flag_checking) |
969 | { | |
970 | regset *v = regset_pool.v; | |
971 | int i = 0; | |
972 | int n = regset_pool.n; | |
b8698a0f | 973 | |
b2b29377 MM |
974 | regset *vv = regset_pool.vv; |
975 | int ii = 0; | |
976 | int nn = regset_pool.nn; | |
b8698a0f | 977 | |
b2b29377 | 978 | int diff = 0; |
b8698a0f | 979 | |
b2b29377 | 980 | gcc_assert (n <= nn); |
b8698a0f | 981 | |
b2b29377 MM |
982 | /* Sort both vectors so it will be possible to compare them. */ |
983 | qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); | |
984 | qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); | |
b8698a0f | 985 | |
b2b29377 MM |
986 | while (ii < nn) |
987 | { | |
988 | if (v[i] == vv[ii]) | |
989 | i++; | |
990 | else | |
991 | /* VV[II] was lost. */ | |
992 | diff++; | |
b8698a0f | 993 | |
b2b29377 MM |
994 | ii++; |
995 | } | |
b8698a0f | 996 | |
b2b29377 MM |
997 | gcc_assert (diff == regset_pool.diff); |
998 | } | |
b8698a0f | 999 | |
e855c69d AB |
1000 | /* If not true - we have a memory leak. */ |
1001 | gcc_assert (regset_pool.diff == 0); | |
b8698a0f | 1002 | |
e855c69d AB |
1003 | while (regset_pool.n) |
1004 | { | |
1005 | --regset_pool.n; | |
1006 | FREE_REG_SET (regset_pool.v[regset_pool.n]); | |
1007 | } | |
1008 | ||
1009 | free (regset_pool.v); | |
1010 | regset_pool.v = NULL; | |
1011 | regset_pool.s = 0; | |
b8698a0f | 1012 | |
e855c69d AB |
1013 | free (regset_pool.vv); |
1014 | regset_pool.vv = NULL; | |
1015 | regset_pool.nn = 0; | |
1016 | regset_pool.ss = 0; | |
1017 | ||
1018 | regset_pool.diff = 0; | |
1019 | } | |
1020 | \f | |
1021 | ||
b8698a0f L |
1022 | /* Functions to work with nop pools. NOP insns are used as temporary |
1023 | placeholders of the insns being scheduled to allow correct update of | |
e855c69d AB |
1024 | the data sets. When update is finished, NOPs are deleted. */ |
1025 | ||
1026 | /* A vinsn that is used to represent a nop. This vinsn is shared among all | |
1027 | nops sel-sched generates. */ | |
1028 | static vinsn_t nop_vinsn = NULL; | |
1029 | ||
1030 | /* Emit a nop before INSN, taking it from pool. */ | |
1031 | insn_t | |
1032 | get_nop_from_pool (insn_t insn) | |
1033 | { | |
6144a836 | 1034 | rtx nop_pat; |
e855c69d AB |
1035 | insn_t nop; |
1036 | bool old_p = nop_pool.n != 0; | |
1037 | int flags; | |
1038 | ||
1039 | if (old_p) | |
6144a836 | 1040 | nop_pat = nop_pool.v[--nop_pool.n]; |
e855c69d | 1041 | else |
6144a836 | 1042 | nop_pat = nop_pattern; |
e855c69d | 1043 | |
6144a836 | 1044 | nop = emit_insn_before (nop_pat, insn); |
e855c69d AB |
1045 | |
1046 | if (old_p) | |
1047 | flags = INSN_INIT_TODO_SSID; | |
1048 | else | |
1049 | flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; | |
1050 | ||
1051 | set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); | |
1052 | sel_init_new_insn (nop, flags); | |
1053 | ||
1054 | return nop; | |
1055 | } | |
1056 | ||
1057 | /* Remove NOP from the instruction stream and return it to the pool. */ | |
1058 | void | |
b5b8b0ac | 1059 | return_nop_to_pool (insn_t nop, bool full_tidying) |
e855c69d AB |
1060 | { |
1061 | gcc_assert (INSN_IN_STREAM_P (nop)); | |
b5b8b0ac | 1062 | sel_remove_insn (nop, false, full_tidying); |
e855c69d | 1063 | |
1f397f45 | 1064 | /* We'll recycle this nop. */ |
4654c0cf | 1065 | nop->set_undeleted (); |
1f397f45 | 1066 | |
e855c69d | 1067 | if (nop_pool.n == nop_pool.s) |
6144a836 | 1068 | nop_pool.v = XRESIZEVEC (rtx_insn *, nop_pool.v, |
e855c69d AB |
1069 | (nop_pool.s = 2 * nop_pool.s + 1)); |
1070 | nop_pool.v[nop_pool.n++] = nop; | |
1071 | } | |
1072 | ||
1073 | /* Free the nop pool. */ | |
1074 | void | |
1075 | free_nop_pool (void) | |
1076 | { | |
1077 | nop_pool.n = 0; | |
1078 | nop_pool.s = 0; | |
1079 | free (nop_pool.v); | |
1080 | nop_pool.v = NULL; | |
1081 | } | |
1082 | \f | |
1083 | ||
b8698a0f | 1084 | /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb. |
e855c69d AB |
1085 | The callback is given two rtxes XX and YY and writes the new rtxes |
1086 | to NX and NY in case some needs to be skipped. */ | |
1087 | static int | |
1088 | skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) | |
1089 | { | |
1090 | const_rtx x = *xx; | |
1091 | const_rtx y = *yy; | |
b8698a0f | 1092 | |
e855c69d AB |
1093 | if (GET_CODE (x) == UNSPEC |
1094 | && (targetm.sched.skip_rtx_p == NULL | |
1095 | || targetm.sched.skip_rtx_p (x))) | |
1096 | { | |
1097 | *nx = XVECEXP (x, 0, 0); | |
1098 | *ny = CONST_CAST_RTX (y); | |
1099 | return 1; | |
1100 | } | |
b8698a0f | 1101 | |
e855c69d AB |
1102 | if (GET_CODE (y) == UNSPEC |
1103 | && (targetm.sched.skip_rtx_p == NULL | |
1104 | || targetm.sched.skip_rtx_p (y))) | |
1105 | { | |
1106 | *nx = CONST_CAST_RTX (x); | |
1107 | *ny = XVECEXP (y, 0, 0); | |
1108 | return 1; | |
1109 | } | |
b8698a0f | 1110 | |
e855c69d AB |
1111 | return 0; |
1112 | } | |
1113 | ||
b8698a0f | 1114 | /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way |
e855c69d AB |
1115 | to support ia64 speculation. When changes are needed, new rtx X and new mode |
1116 | NMODE are written, and the callback returns true. */ | |
1117 | static int | |
ef4bddc2 RS |
1118 | hash_with_unspec_callback (const_rtx x, machine_mode mode ATTRIBUTE_UNUSED, |
1119 | rtx *nx, machine_mode* nmode) | |
e855c69d | 1120 | { |
b8698a0f | 1121 | if (GET_CODE (x) == UNSPEC |
e855c69d AB |
1122 | && targetm.sched.skip_rtx_p |
1123 | && targetm.sched.skip_rtx_p (x)) | |
1124 | { | |
1125 | *nx = XVECEXP (x, 0 ,0); | |
32e8bb8e | 1126 | *nmode = VOIDmode; |
e855c69d AB |
1127 | return 1; |
1128 | } | |
b8698a0f | 1129 | |
e855c69d AB |
1130 | return 0; |
1131 | } | |
1132 | ||
1133 | /* Returns LHS and RHS are ok to be scheduled separately. */ | |
1134 | static bool | |
1135 | lhs_and_rhs_separable_p (rtx lhs, rtx rhs) | |
1136 | { | |
1137 | if (lhs == NULL || rhs == NULL) | |
1138 | return false; | |
1139 | ||
807e902e KZ |
1140 | /* Do not schedule constants as rhs: no point to use reg, if const |
1141 | can be used. Moreover, scheduling const as rhs may lead to mode | |
1142 | mismatch cause consts don't have modes but they could be merged | |
1143 | from branches where the same const used in different modes. */ | |
e855c69d AB |
1144 | if (CONSTANT_P (rhs)) |
1145 | return false; | |
1146 | ||
1147 | /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ | |
1148 | if (COMPARISON_P (rhs)) | |
1149 | return false; | |
1150 | ||
1151 | /* Do not allow single REG to be an rhs. */ | |
1152 | if (REG_P (rhs)) | |
1153 | return false; | |
1154 | ||
b8698a0f | 1155 | /* See comment at find_used_regs_1 (*1) for explanation of this |
e855c69d AB |
1156 | restriction. */ |
1157 | /* FIXME: remove this later. */ | |
1158 | if (MEM_P (lhs)) | |
1159 | return false; | |
1160 | ||
1161 | /* This will filter all tricky things like ZERO_EXTRACT etc. | |
1162 | For now we don't handle it. */ | |
1163 | if (!REG_P (lhs) && !MEM_P (lhs)) | |
1164 | return false; | |
1165 | ||
1166 | return true; | |
1167 | } | |
1168 | ||
b8698a0f L |
1169 | /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When |
1170 | FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is | |
e855c69d AB |
1171 | used e.g. for insns from recovery blocks. */ |
1172 | static void | |
1173 | vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) | |
1174 | { | |
1175 | hash_rtx_callback_function hrcf; | |
1176 | int insn_class; | |
1177 | ||
3d30f4e8 | 1178 | VINSN_INSN_RTX (vi) = insn; |
e855c69d AB |
1179 | VINSN_COUNT (vi) = 0; |
1180 | vi->cost = -1; | |
b8698a0f | 1181 | |
9ef1bf71 AM |
1182 | if (INSN_NOP_P (insn)) |
1183 | return; | |
1184 | ||
e855c69d AB |
1185 | if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) |
1186 | init_id_from_df (VINSN_ID (vi), insn, force_unique_p); | |
1187 | else | |
1188 | deps_init_id (VINSN_ID (vi), insn, force_unique_p); | |
b8698a0f | 1189 | |
e855c69d AB |
1190 | /* Hash vinsn depending on whether it is separable or not. */ |
1191 | hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; | |
1192 | if (VINSN_SEPARABLE_P (vi)) | |
1193 | { | |
1194 | rtx rhs = VINSN_RHS (vi); | |
1195 | ||
1196 | VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs), | |
1197 | NULL, NULL, false, hrcf); | |
1198 | VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi), | |
1199 | VOIDmode, NULL, NULL, | |
1200 | false, hrcf); | |
1201 | } | |
1202 | else | |
1203 | { | |
1204 | VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode, | |
1205 | NULL, NULL, false, hrcf); | |
1206 | VINSN_HASH_RTX (vi) = VINSN_HASH (vi); | |
1207 | } | |
b8698a0f | 1208 | |
e855c69d AB |
1209 | insn_class = haifa_classify_insn (insn); |
1210 | if (insn_class >= 2 | |
1211 | && (!targetm.sched.get_insn_spec_ds | |
1212 | || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) | |
1213 | == 0))) | |
1214 | VINSN_MAY_TRAP_P (vi) = true; | |
1215 | else | |
1216 | VINSN_MAY_TRAP_P (vi) = false; | |
1217 | } | |
1218 | ||
1219 | /* Indicate that VI has become the part of an rtx object. */ | |
1220 | void | |
1221 | vinsn_attach (vinsn_t vi) | |
1222 | { | |
1223 | /* Assert that VI is not pending for deletion. */ | |
1224 | gcc_assert (VINSN_INSN_RTX (vi)); | |
1225 | ||
1226 | VINSN_COUNT (vi)++; | |
1227 | } | |
1228 | ||
b8698a0f | 1229 | /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct |
e855c69d AB |
1230 | VINSN_TYPE (VI). */ |
1231 | static vinsn_t | |
1232 | vinsn_create (insn_t insn, bool force_unique_p) | |
1233 | { | |
1234 | vinsn_t vi = XCNEW (struct vinsn_def); | |
1235 | ||
1236 | vinsn_init (vi, insn, force_unique_p); | |
1237 | return vi; | |
1238 | } | |
1239 | ||
1240 | /* Return a copy of VI. When REATTACH_P is true, detach VI and attach | |
1241 | the copy. */ | |
b8698a0f | 1242 | vinsn_t |
e855c69d AB |
1243 | vinsn_copy (vinsn_t vi, bool reattach_p) |
1244 | { | |
9ee1fbb1 | 1245 | rtx_insn *copy; |
e855c69d AB |
1246 | bool unique = VINSN_UNIQUE_P (vi); |
1247 | vinsn_t new_vi; | |
b8698a0f | 1248 | |
e855c69d AB |
1249 | copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); |
1250 | new_vi = create_vinsn_from_insn_rtx (copy, unique); | |
1251 | if (reattach_p) | |
1252 | { | |
1253 | vinsn_detach (vi); | |
1254 | vinsn_attach (new_vi); | |
1255 | } | |
1256 | ||
1257 | return new_vi; | |
1258 | } | |
1259 | ||
1260 | /* Delete the VI vinsn and free its data. */ | |
1261 | static void | |
1262 | vinsn_delete (vinsn_t vi) | |
1263 | { | |
1264 | gcc_assert (VINSN_COUNT (vi) == 0); | |
1265 | ||
9ef1bf71 AM |
1266 | if (!INSN_NOP_P (VINSN_INSN_RTX (vi))) |
1267 | { | |
1268 | return_regset_to_pool (VINSN_REG_SETS (vi)); | |
1269 | return_regset_to_pool (VINSN_REG_USES (vi)); | |
1270 | return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); | |
1271 | } | |
e855c69d AB |
1272 | |
1273 | free (vi); | |
1274 | } | |
1275 | ||
b8698a0f | 1276 | /* Indicate that VI is no longer a part of some rtx object. |
e855c69d AB |
1277 | Remove VI if it is no longer needed. */ |
1278 | void | |
1279 | vinsn_detach (vinsn_t vi) | |
1280 | { | |
1281 | gcc_assert (VINSN_COUNT (vi) > 0); | |
1282 | ||
1283 | if (--VINSN_COUNT (vi) == 0) | |
1284 | vinsn_delete (vi); | |
1285 | } | |
1286 | ||
1287 | /* Returns TRUE if VI is a branch. */ | |
1288 | bool | |
1289 | vinsn_cond_branch_p (vinsn_t vi) | |
1290 | { | |
1291 | insn_t insn; | |
1292 | ||
1293 | if (!VINSN_UNIQUE_P (vi)) | |
1294 | return false; | |
1295 | ||
1296 | insn = VINSN_INSN_RTX (vi); | |
1297 | if (BB_END (BLOCK_FOR_INSN (insn)) != insn) | |
1298 | return false; | |
1299 | ||
1300 | return control_flow_insn_p (insn); | |
1301 | } | |
1302 | ||
1303 | /* Return latency of INSN. */ | |
1304 | static int | |
647d790d | 1305 | sel_insn_rtx_cost (rtx_insn *insn) |
e855c69d AB |
1306 | { |
1307 | int cost; | |
1308 | ||
1309 | /* A USE insn, or something else we don't need to | |
1310 | understand. We can't pass these directly to | |
1311 | result_ready_cost or insn_default_latency because it will | |
1312 | trigger a fatal error for unrecognizable insns. */ | |
1313 | if (recog_memoized (insn) < 0) | |
1314 | cost = 0; | |
1315 | else | |
1316 | { | |
1317 | cost = insn_default_latency (insn); | |
1318 | ||
1319 | if (cost < 0) | |
1320 | cost = 0; | |
1321 | } | |
1322 | ||
1323 | return cost; | |
1324 | } | |
1325 | ||
1326 | /* Return the cost of the VI. | |
ffc1ded5 | 1327 | !!! FIXME: Unify with haifa-sched.c: insn_sched_cost (). */ |
e855c69d AB |
1328 | int |
1329 | sel_vinsn_cost (vinsn_t vi) | |
1330 | { | |
1331 | int cost = vi->cost; | |
1332 | ||
1333 | if (cost < 0) | |
1334 | { | |
1335 | cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); | |
1336 | vi->cost = cost; | |
1337 | } | |
1338 | ||
1339 | return cost; | |
1340 | } | |
1341 | \f | |
1342 | ||
1343 | /* Functions for insn emitting. */ | |
1344 | ||
1345 | /* Emit new insn after AFTER based on PATTERN and initialize its data from | |
1346 | EXPR and SEQNO. */ | |
1347 | insn_t | |
1348 | sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) | |
1349 | { | |
1350 | insn_t new_insn; | |
1351 | ||
1352 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); | |
1353 | ||
1354 | new_insn = emit_insn_after (pattern, after); | |
1355 | set_insn_init (expr, NULL, seqno); | |
1356 | sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); | |
1357 | ||
1358 | return new_insn; | |
1359 | } | |
1360 | ||
1361 | /* Force newly generated vinsns to be unique. */ | |
1362 | static bool init_insn_force_unique_p = false; | |
1363 | ||
1364 | /* Emit new speculation recovery insn after AFTER based on PATTERN and | |
1365 | initialize its data from EXPR and SEQNO. */ | |
1366 | insn_t | |
1367 | sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, | |
1368 | insn_t after) | |
1369 | { | |
1370 | insn_t insn; | |
1371 | ||
1372 | gcc_assert (!init_insn_force_unique_p); | |
1373 | ||
1374 | init_insn_force_unique_p = true; | |
1375 | insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); | |
1376 | CANT_MOVE (insn) = 1; | |
1377 | init_insn_force_unique_p = false; | |
1378 | ||
1379 | return insn; | |
1380 | } | |
1381 | ||
1382 | /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, | |
b8698a0f L |
1383 | take it as a new vinsn instead of EXPR's vinsn. |
1384 | We simplify insns later, after scheduling region in | |
e855c69d AB |
1385 | simplify_changed_insns. */ |
1386 | insn_t | |
b8698a0f | 1387 | sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
e855c69d AB |
1388 | insn_t after) |
1389 | { | |
1390 | expr_t emit_expr; | |
1391 | insn_t insn; | |
1392 | int flags; | |
b8698a0f L |
1393 | |
1394 | emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), | |
e855c69d AB |
1395 | seqno); |
1396 | insn = EXPR_INSN_RTX (emit_expr); | |
fe08255d AB |
1397 | |
1398 | /* The insn may come from the transformation cache, which may hold already | |
1399 | deleted insns, so mark it as not deleted. */ | |
4654c0cf | 1400 | insn->set_undeleted (); |
fe08255d | 1401 | |
b8698a0f | 1402 | add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); |
e855c69d AB |
1403 | |
1404 | flags = INSN_INIT_TODO_SSID; | |
1405 | if (INSN_LUID (insn) == 0) | |
1406 | flags |= INSN_INIT_TODO_LUID; | |
1407 | sel_init_new_insn (insn, flags); | |
1408 | ||
1409 | return insn; | |
1410 | } | |
1411 | ||
1412 | /* Move insn from EXPR after AFTER. */ | |
1413 | insn_t | |
1414 | sel_move_insn (expr_t expr, int seqno, insn_t after) | |
1415 | { | |
1416 | insn_t insn = EXPR_INSN_RTX (expr); | |
1417 | basic_block bb = BLOCK_FOR_INSN (after); | |
1418 | insn_t next = NEXT_INSN (after); | |
1419 | ||
1420 | /* Assert that in move_op we disconnected this insn properly. */ | |
1421 | gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); | |
0f82e5c9 DM |
1422 | SET_PREV_INSN (insn) = after; |
1423 | SET_NEXT_INSN (insn) = next; | |
e855c69d | 1424 | |
0f82e5c9 DM |
1425 | SET_NEXT_INSN (after) = insn; |
1426 | SET_PREV_INSN (next) = insn; | |
e855c69d AB |
1427 | |
1428 | /* Update links from insn to bb and vice versa. */ | |
1429 | df_insn_change_bb (insn, bb); | |
1430 | if (BB_END (bb) == after) | |
1130d5e3 | 1431 | BB_END (bb) = insn; |
b8698a0f | 1432 | |
e855c69d AB |
1433 | prepare_insn_expr (insn, seqno); |
1434 | return insn; | |
1435 | } | |
1436 | ||
1437 | \f | |
1438 | /* Functions to work with right-hand sides. */ | |
1439 | ||
b8698a0f | 1440 | /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector |
e855c69d | 1441 | VECT and return true when found. Use NEW_VINSN for comparison only when |
b8698a0f L |
1442 | COMPARE_VINSNS is true. Write to INDP the index on which |
1443 | the search has stopped, such that inserting the new element at INDP will | |
e855c69d AB |
1444 | retain VECT's sort order. */ |
1445 | static bool | |
9771b263 | 1446 | find_in_history_vect_1 (vec<expr_history_def> vect, |
b8698a0f | 1447 | unsigned uid, vinsn_t new_vinsn, |
e855c69d AB |
1448 | bool compare_vinsns, int *indp) |
1449 | { | |
1450 | expr_history_def *arr; | |
9771b263 | 1451 | int i, j, len = vect.length (); |
e855c69d AB |
1452 | |
1453 | if (len == 0) | |
1454 | { | |
1455 | *indp = 0; | |
1456 | return false; | |
1457 | } | |
1458 | ||
9771b263 | 1459 | arr = vect.address (); |
e855c69d AB |
1460 | i = 0, j = len - 1; |
1461 | ||
1462 | while (i <= j) | |
1463 | { | |
1464 | unsigned auid = arr[i].uid; | |
b8698a0f | 1465 | vinsn_t avinsn = arr[i].new_expr_vinsn; |
e855c69d AB |
1466 | |
1467 | if (auid == uid | |
b8698a0f L |
1468 | /* When undoing transformation on a bookkeeping copy, the new vinsn |
1469 | may not be exactly equal to the one that is saved in the vector. | |
e855c69d AB |
1470 | This is because the insn whose copy we're checking was possibly |
1471 | substituted itself. */ | |
b8698a0f | 1472 | && (! compare_vinsns |
e855c69d AB |
1473 | || vinsn_equal_p (avinsn, new_vinsn))) |
1474 | { | |
1475 | *indp = i; | |
1476 | return true; | |
1477 | } | |
1478 | else if (auid > uid) | |
1479 | break; | |
1480 | i++; | |
1481 | } | |
1482 | ||
1483 | *indp = i; | |
1484 | return false; | |
1485 | } | |
1486 | ||
b8698a0f L |
1487 | /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return |
1488 | the position found or -1, if no such value is in vector. | |
e855c69d AB |
1489 | Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ |
1490 | int | |
9771b263 | 1491 | find_in_history_vect (vec<expr_history_def> vect, rtx insn, |
e855c69d AB |
1492 | vinsn_t new_vinsn, bool originators_p) |
1493 | { | |
1494 | int ind; | |
1495 | ||
b8698a0f | 1496 | if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn, |
e855c69d AB |
1497 | false, &ind)) |
1498 | return ind; | |
1499 | ||
1500 | if (INSN_ORIGINATORS (insn) && originators_p) | |
1501 | { | |
1502 | unsigned uid; | |
1503 | bitmap_iterator bi; | |
1504 | ||
1505 | EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) | |
1506 | if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind)) | |
1507 | return ind; | |
1508 | } | |
b8698a0f | 1509 | |
e855c69d AB |
1510 | return -1; |
1511 | } | |
1512 | ||
b8698a0f L |
1513 | /* Insert new element in a sorted history vector pointed to by PVECT, |
1514 | if it is not there already. The element is searched using | |
e855c69d AB |
1515 | UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save |
1516 | the history of a transformation. */ | |
1517 | void | |
9771b263 | 1518 | insert_in_history_vect (vec<expr_history_def> *pvect, |
e855c69d | 1519 | unsigned uid, enum local_trans_type type, |
b8698a0f | 1520 | vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, |
e855c69d AB |
1521 | ds_t spec_ds) |
1522 | { | |
9771b263 | 1523 | vec<expr_history_def> vect = *pvect; |
e855c69d AB |
1524 | expr_history_def temp; |
1525 | bool res; | |
1526 | int ind; | |
1527 | ||
1528 | res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind); | |
1529 | ||
1530 | if (res) | |
1531 | { | |
9771b263 | 1532 | expr_history_def *phist = &vect[ind]; |
e855c69d | 1533 | |
b8698a0f | 1534 | /* It is possible that speculation types of expressions that were |
e855c69d AB |
1535 | propagated through different paths will be different here. In this |
1536 | case, merge the status to get the correct check later. */ | |
1537 | if (phist->spec_ds != spec_ds) | |
1538 | phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); | |
1539 | return; | |
1540 | } | |
b8698a0f | 1541 | |
e855c69d AB |
1542 | temp.uid = uid; |
1543 | temp.old_expr_vinsn = old_expr_vinsn; | |
b8698a0f | 1544 | temp.new_expr_vinsn = new_expr_vinsn; |
e855c69d AB |
1545 | temp.spec_ds = spec_ds; |
1546 | temp.type = type; | |
1547 | ||
1548 | vinsn_attach (old_expr_vinsn); | |
1549 | vinsn_attach (new_expr_vinsn); | |
9771b263 | 1550 | vect.safe_insert (ind, temp); |
e855c69d AB |
1551 | *pvect = vect; |
1552 | } | |
1553 | ||
1554 | /* Free history vector PVECT. */ | |
1555 | static void | |
9771b263 | 1556 | free_history_vect (vec<expr_history_def> &pvect) |
e855c69d AB |
1557 | { |
1558 | unsigned i; | |
1559 | expr_history_def *phist; | |
1560 | ||
9771b263 | 1561 | if (! pvect.exists ()) |
e855c69d | 1562 | return; |
b8698a0f | 1563 | |
9771b263 | 1564 | for (i = 0; pvect.iterate (i, &phist); i++) |
e855c69d AB |
1565 | { |
1566 | vinsn_detach (phist->old_expr_vinsn); | |
1567 | vinsn_detach (phist->new_expr_vinsn); | |
1568 | } | |
b8698a0f | 1569 | |
9771b263 | 1570 | pvect.release (); |
e855c69d AB |
1571 | } |
1572 | ||
5d369d58 AB |
1573 | /* Merge vector FROM to PVECT. */ |
1574 | static void | |
9771b263 DN |
1575 | merge_history_vect (vec<expr_history_def> *pvect, |
1576 | vec<expr_history_def> from) | |
5d369d58 AB |
1577 | { |
1578 | expr_history_def *phist; | |
1579 | int i; | |
1580 | ||
1581 | /* We keep this vector sorted. */ | |
9771b263 | 1582 | for (i = 0; from.iterate (i, &phist); i++) |
5d369d58 AB |
1583 | insert_in_history_vect (pvect, phist->uid, phist->type, |
1584 | phist->old_expr_vinsn, phist->new_expr_vinsn, | |
1585 | phist->spec_ds); | |
1586 | } | |
e855c69d AB |
1587 | |
1588 | /* Compare two vinsns as rhses if possible and as vinsns otherwise. */ | |
1589 | bool | |
1590 | vinsn_equal_p (vinsn_t x, vinsn_t y) | |
1591 | { | |
1592 | rtx_equal_p_callback_function repcf; | |
1593 | ||
1594 | if (x == y) | |
1595 | return true; | |
1596 | ||
1597 | if (VINSN_TYPE (x) != VINSN_TYPE (y)) | |
1598 | return false; | |
1599 | ||
1600 | if (VINSN_HASH (x) != VINSN_HASH (y)) | |
1601 | return false; | |
1602 | ||
1603 | repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL; | |
b8698a0f | 1604 | if (VINSN_SEPARABLE_P (x)) |
e855c69d AB |
1605 | { |
1606 | /* Compare RHSes of VINSNs. */ | |
1607 | gcc_assert (VINSN_RHS (x)); | |
1608 | gcc_assert (VINSN_RHS (y)); | |
1609 | ||
1610 | return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf); | |
1611 | } | |
1612 | ||
1613 | return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); | |
1614 | } | |
1615 | \f | |
1616 | ||
1617 | /* Functions for working with expressions. */ | |
1618 | ||
1619 | /* Initialize EXPR. */ | |
1620 | static void | |
1621 | init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority, | |
1622 | int sched_times, int orig_bb_index, ds_t spec_done_ds, | |
1623 | ds_t spec_to_check_ds, int orig_sched_cycle, | |
9771b263 DN |
1624 | vec<expr_history_def> history, |
1625 | signed char target_available, | |
e855c69d AB |
1626 | bool was_substituted, bool was_renamed, bool needs_spec_check_p, |
1627 | bool cant_move) | |
1628 | { | |
1629 | vinsn_attach (vi); | |
1630 | ||
1631 | EXPR_VINSN (expr) = vi; | |
1632 | EXPR_SPEC (expr) = spec; | |
1633 | EXPR_USEFULNESS (expr) = use; | |
1634 | EXPR_PRIORITY (expr) = priority; | |
1635 | EXPR_PRIORITY_ADJ (expr) = 0; | |
1636 | EXPR_SCHED_TIMES (expr) = sched_times; | |
1637 | EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; | |
1638 | EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; | |
1639 | EXPR_SPEC_DONE_DS (expr) = spec_done_ds; | |
1640 | EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; | |
1641 | ||
9771b263 | 1642 | if (history.exists ()) |
e855c69d AB |
1643 | EXPR_HISTORY_OF_CHANGES (expr) = history; |
1644 | else | |
9771b263 | 1645 | EXPR_HISTORY_OF_CHANGES (expr).create (0); |
e855c69d AB |
1646 | |
1647 | EXPR_TARGET_AVAILABLE (expr) = target_available; | |
1648 | EXPR_WAS_SUBSTITUTED (expr) = was_substituted; | |
1649 | EXPR_WAS_RENAMED (expr) = was_renamed; | |
1650 | EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; | |
1651 | EXPR_CANT_MOVE (expr) = cant_move; | |
1652 | } | |
1653 | ||
1654 | /* Make a copy of the expr FROM into the expr TO. */ | |
1655 | void | |
1656 | copy_expr (expr_t to, expr_t from) | |
1657 | { | |
6e1aa848 | 1658 | vec<expr_history_def> temp = vNULL; |
e855c69d | 1659 | |
9771b263 | 1660 | if (EXPR_HISTORY_OF_CHANGES (from).exists ()) |
e855c69d AB |
1661 | { |
1662 | unsigned i; | |
1663 | expr_history_def *phist; | |
1664 | ||
9771b263 | 1665 | temp = EXPR_HISTORY_OF_CHANGES (from).copy (); |
b8698a0f | 1666 | for (i = 0; |
9771b263 | 1667 | temp.iterate (i, &phist); |
e855c69d AB |
1668 | i++) |
1669 | { | |
1670 | vinsn_attach (phist->old_expr_vinsn); | |
1671 | vinsn_attach (phist->new_expr_vinsn); | |
1672 | } | |
1673 | } | |
1674 | ||
b8698a0f | 1675 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), |
e855c69d AB |
1676 | EXPR_USEFULNESS (from), EXPR_PRIORITY (from), |
1677 | EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), | |
b8698a0f | 1678 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), |
e855c69d | 1679 | EXPR_ORIG_SCHED_CYCLE (from), temp, |
b8698a0f | 1680 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
e855c69d AB |
1681 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1682 | EXPR_CANT_MOVE (from)); | |
1683 | } | |
1684 | ||
b8698a0f | 1685 | /* Same, but the final expr will not ever be in av sets, so don't copy |
e855c69d AB |
1686 | "uninteresting" data such as bitmap cache. */ |
1687 | void | |
1688 | copy_expr_onside (expr_t to, expr_t from) | |
1689 | { | |
1690 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), | |
1691 | EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0, | |
9771b263 | 1692 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0, |
6e1aa848 | 1693 | vNULL, |
e855c69d AB |
1694 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
1695 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), | |
1696 | EXPR_CANT_MOVE (from)); | |
1697 | } | |
1698 | ||
1699 | /* Prepare the expr of INSN for scheduling. Used when moving insn and when | |
1700 | initializing new insns. */ | |
1701 | static void | |
1702 | prepare_insn_expr (insn_t insn, int seqno) | |
1703 | { | |
1704 | expr_t expr = INSN_EXPR (insn); | |
1705 | ds_t ds; | |
b8698a0f | 1706 | |
e855c69d AB |
1707 | INSN_SEQNO (insn) = seqno; |
1708 | EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); | |
1709 | EXPR_SPEC (expr) = 0; | |
1710 | EXPR_ORIG_SCHED_CYCLE (expr) = 0; | |
1711 | EXPR_WAS_SUBSTITUTED (expr) = 0; | |
1712 | EXPR_WAS_RENAMED (expr) = 0; | |
1713 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
1714 | INSN_LIVE_VALID_P (insn) = false; | |
1715 | ||
1716 | /* ??? If this expression is speculative, make its dependence | |
1717 | as weak as possible. We can filter this expression later | |
1718 | in process_spec_exprs, because we do not distinguish | |
1719 | between the status we got during compute_av_set and the | |
1720 | existing status. To be fixed. */ | |
1721 | ds = EXPR_SPEC_DONE_DS (expr); | |
1722 | if (ds) | |
1723 | EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); | |
1724 | ||
9771b263 | 1725 | free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); |
e855c69d AB |
1726 | } |
1727 | ||
1728 | /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1729 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1730 | a split point. */ |
1731 | static void | |
1732 | update_target_availability (expr_t to, expr_t from, insn_t split_point) | |
1733 | { | |
b8698a0f | 1734 | if (EXPR_TARGET_AVAILABLE (to) < 0 |
e855c69d AB |
1735 | || EXPR_TARGET_AVAILABLE (from) < 0) |
1736 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1737 | else | |
1738 | { | |
1739 | /* We try to detect the case when one of the expressions | |
1740 | can only be reached through another one. In this case, | |
1741 | we can do better. */ | |
1742 | if (split_point == NULL) | |
1743 | { | |
1744 | int toind, fromind; | |
1745 | ||
1746 | toind = EXPR_ORIG_BB_INDEX (to); | |
1747 | fromind = EXPR_ORIG_BB_INDEX (from); | |
b8698a0f | 1748 | |
e855c69d | 1749 | if (toind && toind == fromind) |
b8698a0f | 1750 | /* Do nothing -- everything is done in |
e855c69d AB |
1751 | merge_with_other_exprs. */ |
1752 | ; | |
1753 | else | |
1754 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1755 | } | |
854b5fd7 AM |
1756 | else if (EXPR_TARGET_AVAILABLE (from) == 0 |
1757 | && EXPR_LHS (from) | |
1758 | && REG_P (EXPR_LHS (from)) | |
1759 | && REGNO (EXPR_LHS (to)) != REGNO (EXPR_LHS (from))) | |
1760 | EXPR_TARGET_AVAILABLE (to) = -1; | |
e855c69d AB |
1761 | else |
1762 | EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); | |
1763 | } | |
1764 | } | |
1765 | ||
1766 | /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1767 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1768 | a split point. */ |
1769 | static void | |
1770 | update_speculative_bits (expr_t to, expr_t from, insn_t split_point) | |
1771 | { | |
1772 | ds_t old_to_ds, old_from_ds; | |
1773 | ||
1774 | old_to_ds = EXPR_SPEC_DONE_DS (to); | |
1775 | old_from_ds = EXPR_SPEC_DONE_DS (from); | |
b8698a0f | 1776 | |
e855c69d AB |
1777 | EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); |
1778 | EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); | |
1779 | EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); | |
1780 | ||
1781 | /* When merging e.g. control & data speculative exprs, or a control | |
b8698a0f | 1782 | speculative with a control&data speculative one, we really have |
e855c69d AB |
1783 | to change vinsn too. Also, when speculative status is changed, |
1784 | we also need to record this as a transformation in expr's history. */ | |
1785 | if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) | |
1786 | { | |
1787 | old_to_ds = ds_get_speculation_types (old_to_ds); | |
1788 | old_from_ds = ds_get_speculation_types (old_from_ds); | |
b8698a0f | 1789 | |
e855c69d AB |
1790 | if (old_to_ds != old_from_ds) |
1791 | { | |
1792 | ds_t record_ds; | |
b8698a0f L |
1793 | |
1794 | /* When both expressions are speculative, we need to change | |
e855c69d AB |
1795 | the vinsn first. */ |
1796 | if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) | |
1797 | { | |
1798 | int res; | |
b8698a0f | 1799 | |
e855c69d AB |
1800 | res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); |
1801 | gcc_assert (res >= 0); | |
1802 | } | |
1803 | ||
1804 | if (split_point != NULL) | |
1805 | { | |
1806 | /* Record the change with proper status. */ | |
1807 | record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; | |
1808 | record_ds &= ~(old_to_ds & SPECULATIVE); | |
1809 | record_ds &= ~(old_from_ds & SPECULATIVE); | |
b8698a0f L |
1810 | |
1811 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), | |
1812 | INSN_UID (split_point), TRANS_SPECULATION, | |
e855c69d AB |
1813 | EXPR_VINSN (from), EXPR_VINSN (to), |
1814 | record_ds); | |
1815 | } | |
1816 | } | |
1817 | } | |
1818 | } | |
1819 | ||
1820 | ||
1821 | /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, | |
1822 | this is done along different paths. */ | |
1823 | void | |
1824 | merge_expr_data (expr_t to, expr_t from, insn_t split_point) | |
1825 | { | |
bf3a40e9 DM |
1826 | /* Choose the maximum of the specs of merged exprs. This is required |
1827 | for correctness of bookkeeping. */ | |
1828 | if (EXPR_SPEC (to) < EXPR_SPEC (from)) | |
e855c69d AB |
1829 | EXPR_SPEC (to) = EXPR_SPEC (from); |
1830 | ||
1831 | if (split_point) | |
1832 | EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); | |
1833 | else | |
b8698a0f | 1834 | EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), |
e855c69d AB |
1835 | EXPR_USEFULNESS (from)); |
1836 | ||
1837 | if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) | |
1838 | EXPR_PRIORITY (to) = EXPR_PRIORITY (from); | |
1839 | ||
8e9a9b01 AB |
1840 | /* We merge sched-times half-way to the larger value to avoid the endless |
1841 | pipelining of unneeded insns. The average seems to be good compromise | |
1842 | between pipelining opportunities and avoiding extra work. */ | |
1843 | if (EXPR_SCHED_TIMES (to) != EXPR_SCHED_TIMES (from)) | |
1844 | EXPR_SCHED_TIMES (to) = ((EXPR_SCHED_TIMES (from) + EXPR_SCHED_TIMES (to) | |
1845 | + 1) / 2); | |
e855c69d AB |
1846 | |
1847 | if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) | |
1848 | EXPR_ORIG_BB_INDEX (to) = 0; | |
1849 | ||
b8698a0f | 1850 | EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), |
e855c69d AB |
1851 | EXPR_ORIG_SCHED_CYCLE (from)); |
1852 | ||
e855c69d AB |
1853 | EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); |
1854 | EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); | |
1855 | EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); | |
1856 | ||
5d369d58 AB |
1857 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to), |
1858 | EXPR_HISTORY_OF_CHANGES (from)); | |
e855c69d AB |
1859 | update_target_availability (to, from, split_point); |
1860 | update_speculative_bits (to, from, split_point); | |
1861 | } | |
1862 | ||
1863 | /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal | |
b8698a0f | 1864 | in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions |
e855c69d AB |
1865 | are merged from different successors at a split point. */ |
1866 | void | |
1867 | merge_expr (expr_t to, expr_t from, insn_t split_point) | |
1868 | { | |
1869 | vinsn_t to_vi = EXPR_VINSN (to); | |
1870 | vinsn_t from_vi = EXPR_VINSN (from); | |
1871 | ||
1872 | gcc_assert (vinsn_equal_p (to_vi, from_vi)); | |
1873 | ||
1874 | /* Make sure that speculative pattern is propagated into exprs that | |
1875 | have non-speculative one. This will provide us with consistent | |
1876 | speculative bits and speculative patterns inside expr. */ | |
11a6609c AB |
1877 | if (EXPR_SPEC_DONE_DS (to) == 0 |
1878 | && (EXPR_SPEC_DONE_DS (from) != 0 | |
1879 | /* Do likewise for volatile insns, so that we always retain | |
1880 | the may_trap_p bit on the resulting expression. However, | |
1881 | avoid propagating the trapping bit into the instructions | |
1882 | already speculated. This would result in replacing the | |
1883 | speculative pattern with the non-speculative one and breaking | |
1884 | the speculation support. */ | |
1885 | || (!VINSN_MAY_TRAP_P (EXPR_VINSN (to)) | |
1886 | && VINSN_MAY_TRAP_P (EXPR_VINSN (from))))) | |
e855c69d AB |
1887 | change_vinsn_in_expr (to, EXPR_VINSN (from)); |
1888 | ||
1889 | merge_expr_data (to, from, split_point); | |
1890 | gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); | |
1891 | } | |
1892 | ||
1893 | /* Clear the information of this EXPR. */ | |
1894 | void | |
1895 | clear_expr (expr_t expr) | |
1896 | { | |
b8698a0f | 1897 | |
e855c69d AB |
1898 | vinsn_detach (EXPR_VINSN (expr)); |
1899 | EXPR_VINSN (expr) = NULL; | |
1900 | ||
9771b263 | 1901 | free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); |
e855c69d AB |
1902 | } |
1903 | ||
1904 | /* For a given LV_SET, mark EXPR having unavailable target register. */ | |
1905 | static void | |
1906 | set_unavailable_target_for_expr (expr_t expr, regset lv_set) | |
1907 | { | |
1908 | if (EXPR_SEPARABLE_P (expr)) | |
1909 | { | |
1910 | if (REG_P (EXPR_LHS (expr)) | |
cf3d5824 | 1911 | && register_unavailable_p (lv_set, EXPR_LHS (expr))) |
e855c69d | 1912 | { |
b8698a0f L |
1913 | /* If it's an insn like r1 = use (r1, ...), and it exists in |
1914 | different forms in each of the av_sets being merged, we can't say | |
1915 | whether original destination register is available or not. | |
1916 | However, this still works if destination register is not used | |
e855c69d AB |
1917 | in the original expression: if the branch at which LV_SET we're |
1918 | looking here is not actually 'other branch' in sense that same | |
b8698a0f | 1919 | expression is available through it (but it can't be determined |
e855c69d | 1920 | at computation stage because of transformations on one of the |
b8698a0f L |
1921 | branches), it still won't affect the availability. |
1922 | Liveness of a register somewhere on a code motion path means | |
1923 | it's either read somewhere on a codemotion path, live on | |
e855c69d AB |
1924 | 'other' branch, live at the point immediately following |
1925 | the original operation, or is read by the original operation. | |
1926 | The latter case is filtered out in the condition below. | |
1927 | It still doesn't cover the case when register is defined and used | |
1928 | somewhere within the code motion path, and in this case we could | |
1929 | miss a unifying code motion along both branches using a renamed | |
1930 | register, but it won't affect a code correctness since upon | |
1931 | an actual code motion a bookkeeping code would be generated. */ | |
cf3d5824 SG |
1932 | if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
1933 | EXPR_LHS (expr))) | |
e855c69d AB |
1934 | EXPR_TARGET_AVAILABLE (expr) = -1; |
1935 | else | |
1936 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1937 | } | |
1938 | } | |
1939 | else | |
1940 | { | |
1941 | unsigned regno; | |
1942 | reg_set_iterator rsi; | |
b8698a0f L |
1943 | |
1944 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), | |
e855c69d AB |
1945 | 0, regno, rsi) |
1946 | if (bitmap_bit_p (lv_set, regno)) | |
1947 | { | |
1948 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1949 | break; | |
1950 | } | |
1951 | ||
1952 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), | |
1953 | 0, regno, rsi) | |
1954 | if (bitmap_bit_p (lv_set, regno)) | |
1955 | { | |
1956 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1957 | break; | |
1958 | } | |
1959 | } | |
1960 | } | |
1961 | ||
b8698a0f | 1962 | /* Try to make EXPR speculative. Return 1 when EXPR's pattern |
e855c69d AB |
1963 | or dependence status have changed, 2 when also the target register |
1964 | became unavailable, 0 if nothing had to be changed. */ | |
1965 | int | |
1966 | speculate_expr (expr_t expr, ds_t ds) | |
1967 | { | |
1968 | int res; | |
9ee1fbb1 | 1969 | rtx_insn *orig_insn_rtx; |
e855c69d AB |
1970 | rtx spec_pat; |
1971 | ds_t target_ds, current_ds; | |
1972 | ||
1973 | /* Obtain the status we need to put on EXPR. */ | |
1974 | target_ds = (ds & SPECULATIVE); | |
1975 | current_ds = EXPR_SPEC_DONE_DS (expr); | |
1976 | ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); | |
1977 | ||
1978 | orig_insn_rtx = EXPR_INSN_RTX (expr); | |
1979 | ||
1980 | res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); | |
1981 | ||
1982 | switch (res) | |
1983 | { | |
1984 | case 0: | |
1985 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1986 | return current_ds != ds ? 1 : 0; | |
b8698a0f | 1987 | |
e855c69d AB |
1988 | case 1: |
1989 | { | |
9ee1fbb1 DM |
1990 | rtx_insn *spec_insn_rtx = |
1991 | create_insn_rtx_from_pattern (spec_pat, NULL_RTX); | |
e855c69d AB |
1992 | vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); |
1993 | ||
1994 | change_vinsn_in_expr (expr, spec_vinsn); | |
1995 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1996 | EXPR_NEEDS_SPEC_CHECK_P (expr) = true; | |
1997 | ||
b8698a0f | 1998 | /* Do not allow clobbering the address register of speculative |
e855c69d | 1999 | insns. */ |
cf3d5824 SG |
2000 | if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
2001 | expr_dest_reg (expr))) | |
e855c69d AB |
2002 | { |
2003 | EXPR_TARGET_AVAILABLE (expr) = false; | |
2004 | return 2; | |
2005 | } | |
2006 | ||
2007 | return 1; | |
2008 | } | |
2009 | ||
2010 | case -1: | |
2011 | return -1; | |
2012 | ||
2013 | default: | |
2014 | gcc_unreachable (); | |
2015 | return -1; | |
2016 | } | |
2017 | } | |
2018 | ||
2019 | /* Return a destination register, if any, of EXPR. */ | |
2020 | rtx | |
2021 | expr_dest_reg (expr_t expr) | |
2022 | { | |
2023 | rtx dest = VINSN_LHS (EXPR_VINSN (expr)); | |
2024 | ||
2025 | if (dest != NULL_RTX && REG_P (dest)) | |
2026 | return dest; | |
2027 | ||
2028 | return NULL_RTX; | |
2029 | } | |
2030 | ||
2031 | /* Returns the REGNO of the R's destination. */ | |
2032 | unsigned | |
2033 | expr_dest_regno (expr_t expr) | |
2034 | { | |
2035 | rtx dest = expr_dest_reg (expr); | |
2036 | ||
2037 | gcc_assert (dest != NULL_RTX); | |
2038 | return REGNO (dest); | |
2039 | } | |
2040 | ||
b8698a0f | 2041 | /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in |
e855c69d AB |
2042 | AV_SET having unavailable target register. */ |
2043 | void | |
2044 | mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) | |
2045 | { | |
2046 | expr_t expr; | |
2047 | av_set_iterator avi; | |
2048 | ||
2049 | FOR_EACH_EXPR (expr, avi, join_set) | |
2050 | if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) | |
2051 | set_unavailable_target_for_expr (expr, lv_set); | |
2052 | } | |
2053 | \f | |
2054 | ||
cf3d5824 SG |
2055 | /* Returns true if REG (at least partially) is present in REGS. */ |
2056 | bool | |
2057 | register_unavailable_p (regset regs, rtx reg) | |
2058 | { | |
2059 | unsigned regno, end_regno; | |
2060 | ||
2061 | regno = REGNO (reg); | |
2062 | if (bitmap_bit_p (regs, regno)) | |
2063 | return true; | |
2064 | ||
2065 | end_regno = END_REGNO (reg); | |
2066 | ||
2067 | while (++regno < end_regno) | |
2068 | if (bitmap_bit_p (regs, regno)) | |
2069 | return true; | |
2070 | ||
2071 | return false; | |
2072 | } | |
2073 | ||
e855c69d AB |
2074 | /* Av set functions. */ |
2075 | ||
2076 | /* Add a new element to av set SETP. | |
2077 | Return the element added. */ | |
2078 | static av_set_t | |
2079 | av_set_add_element (av_set_t *setp) | |
2080 | { | |
2081 | /* Insert at the beginning of the list. */ | |
2082 | _list_add (setp); | |
2083 | return *setp; | |
2084 | } | |
2085 | ||
2086 | /* Add EXPR to SETP. */ | |
2087 | void | |
2088 | av_set_add (av_set_t *setp, expr_t expr) | |
2089 | { | |
2090 | av_set_t elem; | |
b8698a0f | 2091 | |
e855c69d AB |
2092 | gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); |
2093 | elem = av_set_add_element (setp); | |
2094 | copy_expr (_AV_SET_EXPR (elem), expr); | |
2095 | } | |
2096 | ||
2097 | /* Same, but do not copy EXPR. */ | |
2098 | static void | |
2099 | av_set_add_nocopy (av_set_t *setp, expr_t expr) | |
2100 | { | |
2101 | av_set_t elem; | |
2102 | ||
2103 | elem = av_set_add_element (setp); | |
2104 | *_AV_SET_EXPR (elem) = *expr; | |
2105 | } | |
2106 | ||
2107 | /* Remove expr pointed to by IP from the av_set. */ | |
2108 | void | |
2109 | av_set_iter_remove (av_set_iterator *ip) | |
2110 | { | |
2111 | clear_expr (_AV_SET_EXPR (*ip->lp)); | |
2112 | _list_iter_remove (ip); | |
2113 | } | |
2114 | ||
2115 | /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the | |
2116 | sense of vinsn_equal_p function. Return NULL if no such expr is | |
2117 | in SET was found. */ | |
2118 | expr_t | |
2119 | av_set_lookup (av_set_t set, vinsn_t sought_vinsn) | |
2120 | { | |
2121 | expr_t expr; | |
2122 | av_set_iterator i; | |
2123 | ||
2124 | FOR_EACH_EXPR (expr, i, set) | |
2125 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2126 | return expr; | |
2127 | return NULL; | |
2128 | } | |
2129 | ||
2130 | /* Same, but also remove the EXPR found. */ | |
2131 | static expr_t | |
2132 | av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) | |
2133 | { | |
2134 | expr_t expr; | |
2135 | av_set_iterator i; | |
2136 | ||
2137 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2138 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2139 | { | |
2140 | _list_iter_remove_nofree (&i); | |
2141 | return expr; | |
2142 | } | |
2143 | return NULL; | |
2144 | } | |
2145 | ||
2146 | /* Search for an expr in SET, such that it's equivalent to EXPR in the | |
2147 | sense of vinsn_equal_p function of their vinsns, but not EXPR itself. | |
2148 | Returns NULL if no such expr is in SET was found. */ | |
2149 | static expr_t | |
2150 | av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) | |
2151 | { | |
2152 | expr_t cur_expr; | |
2153 | av_set_iterator i; | |
2154 | ||
2155 | FOR_EACH_EXPR (cur_expr, i, set) | |
2156 | { | |
2157 | if (cur_expr == expr) | |
2158 | continue; | |
2159 | if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) | |
2160 | return cur_expr; | |
2161 | } | |
2162 | ||
2163 | return NULL; | |
2164 | } | |
2165 | ||
2166 | /* If other expression is already in AVP, remove one of them. */ | |
2167 | expr_t | |
2168 | merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) | |
2169 | { | |
2170 | expr_t expr2; | |
2171 | ||
2172 | expr2 = av_set_lookup_other_equiv_expr (*avp, expr); | |
2173 | if (expr2 != NULL) | |
2174 | { | |
2175 | /* Reset target availability on merge, since taking it only from one | |
2176 | of the exprs would be controversial for different code. */ | |
2177 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2178 | EXPR_USEFULNESS (expr2) = 0; | |
2179 | ||
2180 | merge_expr (expr2, expr, NULL); | |
b8698a0f | 2181 | |
e855c69d AB |
2182 | /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ |
2183 | EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; | |
b8698a0f | 2184 | |
e855c69d AB |
2185 | av_set_iter_remove (ip); |
2186 | return expr2; | |
2187 | } | |
2188 | ||
2189 | return expr; | |
2190 | } | |
2191 | ||
2192 | /* Return true if there is an expr that correlates to VI in SET. */ | |
2193 | bool | |
2194 | av_set_is_in_p (av_set_t set, vinsn_t vi) | |
2195 | { | |
2196 | return av_set_lookup (set, vi) != NULL; | |
2197 | } | |
2198 | ||
2199 | /* Return a copy of SET. */ | |
2200 | av_set_t | |
2201 | av_set_copy (av_set_t set) | |
2202 | { | |
2203 | expr_t expr; | |
2204 | av_set_iterator i; | |
2205 | av_set_t res = NULL; | |
2206 | ||
2207 | FOR_EACH_EXPR (expr, i, set) | |
2208 | av_set_add (&res, expr); | |
2209 | ||
2210 | return res; | |
2211 | } | |
2212 | ||
2213 | /* Join two av sets that do not have common elements by attaching second set | |
2214 | (pointed to by FROMP) to the end of first set (TO_TAILP must point to | |
2215 | _AV_SET_NEXT of first set's last element). */ | |
2216 | static void | |
2217 | join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) | |
2218 | { | |
2219 | gcc_assert (*to_tailp == NULL); | |
2220 | *to_tailp = *fromp; | |
2221 | *fromp = NULL; | |
2222 | } | |
2223 | ||
2224 | /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set | |
2225 | pointed to by FROMP afterwards. */ | |
2226 | void | |
2227 | av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) | |
2228 | { | |
2229 | expr_t expr1; | |
2230 | av_set_iterator i; | |
2231 | ||
2232 | /* Delete from TOP all exprs, that present in FROMP. */ | |
2233 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2234 | { | |
2235 | expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1)); | |
2236 | ||
2237 | if (expr2) | |
2238 | { | |
2239 | merge_expr (expr2, expr1, insn); | |
2240 | av_set_iter_remove (&i); | |
2241 | } | |
2242 | } | |
2243 | ||
2244 | join_distinct_sets (i.lp, fromp); | |
2245 | } | |
2246 | ||
b8698a0f | 2247 | /* Same as above, but also update availability of target register in |
e855c69d AB |
2248 | TOP judging by TO_LV_SET and FROM_LV_SET. */ |
2249 | void | |
2250 | av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, | |
2251 | regset from_lv_set, insn_t insn) | |
2252 | { | |
2253 | expr_t expr1; | |
2254 | av_set_iterator i; | |
2255 | av_set_t *to_tailp, in_both_set = NULL; | |
2256 | ||
2257 | /* Delete from TOP all expres, that present in FROMP. */ | |
2258 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2259 | { | |
2260 | expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1)); | |
2261 | ||
2262 | if (expr2) | |
2263 | { | |
b8698a0f | 2264 | /* It may be that the expressions have different destination |
e855c69d AB |
2265 | registers, in which case we need to check liveness here. */ |
2266 | if (EXPR_SEPARABLE_P (expr1)) | |
2267 | { | |
b8698a0f | 2268 | int regno1 = (REG_P (EXPR_LHS (expr1)) |
e855c69d | 2269 | ? (int) expr_dest_regno (expr1) : -1); |
b8698a0f | 2270 | int regno2 = (REG_P (EXPR_LHS (expr2)) |
e855c69d | 2271 | ? (int) expr_dest_regno (expr2) : -1); |
b8698a0f L |
2272 | |
2273 | /* ??? We don't have a way to check restrictions for | |
e855c69d AB |
2274 | *other* register on the current path, we did it only |
2275 | for the current target register. Give up. */ | |
2276 | if (regno1 != regno2) | |
2277 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2278 | } | |
2279 | else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) | |
2280 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2281 | ||
2282 | merge_expr (expr2, expr1, insn); | |
2283 | av_set_add_nocopy (&in_both_set, expr2); | |
2284 | av_set_iter_remove (&i); | |
2285 | } | |
2286 | else | |
b8698a0f | 2287 | /* EXPR1 is present in TOP, but not in FROMP. Check it on |
e855c69d AB |
2288 | FROM_LV_SET. */ |
2289 | set_unavailable_target_for_expr (expr1, from_lv_set); | |
2290 | } | |
2291 | to_tailp = i.lp; | |
2292 | ||
2293 | /* These expressions are not present in TOP. Check liveness | |
2294 | restrictions on TO_LV_SET. */ | |
2295 | FOR_EACH_EXPR (expr1, i, *fromp) | |
2296 | set_unavailable_target_for_expr (expr1, to_lv_set); | |
2297 | ||
2298 | join_distinct_sets (i.lp, &in_both_set); | |
2299 | join_distinct_sets (to_tailp, fromp); | |
2300 | } | |
2301 | ||
2302 | /* Clear av_set pointed to by SETP. */ | |
2303 | void | |
2304 | av_set_clear (av_set_t *setp) | |
2305 | { | |
2306 | expr_t expr; | |
2307 | av_set_iterator i; | |
2308 | ||
2309 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2310 | av_set_iter_remove (&i); | |
2311 | ||
2312 | gcc_assert (*setp == NULL); | |
2313 | } | |
2314 | ||
2315 | /* Leave only one non-speculative element in the SETP. */ | |
2316 | void | |
2317 | av_set_leave_one_nonspec (av_set_t *setp) | |
2318 | { | |
2319 | expr_t expr; | |
2320 | av_set_iterator i; | |
2321 | bool has_one_nonspec = false; | |
2322 | ||
b8698a0f | 2323 | /* Keep all speculative exprs, and leave one non-speculative |
e855c69d AB |
2324 | (the first one). */ |
2325 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2326 | { | |
2327 | if (!EXPR_SPEC_DONE_DS (expr)) | |
2328 | { | |
2329 | if (has_one_nonspec) | |
2330 | av_set_iter_remove (&i); | |
2331 | else | |
2332 | has_one_nonspec = true; | |
2333 | } | |
2334 | } | |
2335 | } | |
2336 | ||
2337 | /* Return the N'th element of the SET. */ | |
2338 | expr_t | |
2339 | av_set_element (av_set_t set, int n) | |
2340 | { | |
2341 | expr_t expr; | |
2342 | av_set_iterator i; | |
2343 | ||
2344 | FOR_EACH_EXPR (expr, i, set) | |
2345 | if (n-- == 0) | |
2346 | return expr; | |
2347 | ||
2348 | gcc_unreachable (); | |
2349 | return NULL; | |
2350 | } | |
2351 | ||
2352 | /* Deletes all expressions from AVP that are conditional branches (IFs). */ | |
2353 | void | |
2354 | av_set_substract_cond_branches (av_set_t *avp) | |
2355 | { | |
2356 | av_set_iterator i; | |
2357 | expr_t expr; | |
2358 | ||
2359 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2360 | if (vinsn_cond_branch_p (EXPR_VINSN (expr))) | |
2361 | av_set_iter_remove (&i); | |
2362 | } | |
2363 | ||
b8698a0f | 2364 | /* Multiplies usefulness attribute of each member of av-set *AVP by |
e855c69d AB |
2365 | value PROB / ALL_PROB. */ |
2366 | void | |
2367 | av_set_split_usefulness (av_set_t av, int prob, int all_prob) | |
2368 | { | |
2369 | av_set_iterator i; | |
2370 | expr_t expr; | |
2371 | ||
2372 | FOR_EACH_EXPR (expr, i, av) | |
b8698a0f | 2373 | EXPR_USEFULNESS (expr) = (all_prob |
e855c69d AB |
2374 | ? (EXPR_USEFULNESS (expr) * prob) / all_prob |
2375 | : 0); | |
2376 | } | |
2377 | ||
2378 | /* Leave in AVP only those expressions, which are present in AV, | |
5d369d58 | 2379 | and return it, merging history expressions. */ |
e855c69d | 2380 | void |
5d369d58 | 2381 | av_set_code_motion_filter (av_set_t *avp, av_set_t av) |
e855c69d AB |
2382 | { |
2383 | av_set_iterator i; | |
5d369d58 | 2384 | expr_t expr, expr2; |
e855c69d AB |
2385 | |
2386 | FOR_EACH_EXPR_1 (expr, i, avp) | |
5d369d58 | 2387 | if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL) |
e855c69d | 2388 | av_set_iter_remove (&i); |
5d369d58 AB |
2389 | else |
2390 | /* When updating av sets in bookkeeping blocks, we can add more insns | |
2391 | there which will be transformed but the upper av sets will not | |
2392 | reflect those transformations. We then fail to undo those | |
2393 | when searching for such insns. So merge the history saved | |
2394 | in the av set of the block we are processing. */ | |
2395 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2396 | EXPR_HISTORY_OF_CHANGES (expr2)); | |
e855c69d AB |
2397 | } |
2398 | ||
2399 | \f | |
2400 | ||
2401 | /* Dependence hooks to initialize insn data. */ | |
2402 | ||
2403 | /* This is used in hooks callable from dependence analysis when initializing | |
2404 | instruction's data. */ | |
2405 | static struct | |
2406 | { | |
2407 | /* Where the dependence was found (lhs/rhs). */ | |
2408 | deps_where_t where; | |
2409 | ||
2410 | /* The actual data object to initialize. */ | |
2411 | idata_t id; | |
2412 | ||
2413 | /* True when the insn should not be made clonable. */ | |
2414 | bool force_unique_p; | |
2415 | ||
2416 | /* True when insn should be treated as of type USE, i.e. never renamed. */ | |
2417 | bool force_use_p; | |
2418 | } deps_init_id_data; | |
2419 | ||
2420 | ||
b8698a0f | 2421 | /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be |
e855c69d AB |
2422 | clonable. */ |
2423 | static void | |
2424 | setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) | |
2425 | { | |
2426 | int type; | |
b8698a0f | 2427 | |
e855c69d AB |
2428 | /* Determine whether INSN could be cloned and return appropriate vinsn type. |
2429 | That clonable insns which can be separated into lhs and rhs have type SET. | |
2430 | Other clonable insns have type USE. */ | |
2431 | type = GET_CODE (insn); | |
2432 | ||
2433 | /* Only regular insns could be cloned. */ | |
2434 | if (type == INSN && !force_unique_p) | |
2435 | type = SET; | |
2436 | else if (type == JUMP_INSN && simplejump_p (insn)) | |
2437 | type = PC; | |
b5b8b0ac AO |
2438 | else if (type == DEBUG_INSN) |
2439 | type = !force_unique_p ? USE : INSN; | |
b8698a0f | 2440 | |
e855c69d AB |
2441 | IDATA_TYPE (id) = type; |
2442 | IDATA_REG_SETS (id) = get_clear_regset_from_pool (); | |
2443 | IDATA_REG_USES (id) = get_clear_regset_from_pool (); | |
2444 | IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); | |
2445 | } | |
2446 | ||
2447 | /* Start initializing insn data. */ | |
2448 | static void | |
2449 | deps_init_id_start_insn (insn_t insn) | |
2450 | { | |
2451 | gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); | |
2452 | ||
2453 | setup_id_for_insn (deps_init_id_data.id, insn, | |
2454 | deps_init_id_data.force_unique_p); | |
2455 | deps_init_id_data.where = DEPS_IN_INSN; | |
2456 | } | |
2457 | ||
2458 | /* Start initializing lhs data. */ | |
2459 | static void | |
2460 | deps_init_id_start_lhs (rtx lhs) | |
2461 | { | |
2462 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2463 | gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); | |
2464 | ||
2465 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2466 | { | |
2467 | IDATA_LHS (deps_init_id_data.id) = lhs; | |
2468 | deps_init_id_data.where = DEPS_IN_LHS; | |
2469 | } | |
2470 | } | |
2471 | ||
2472 | /* Finish initializing lhs data. */ | |
2473 | static void | |
2474 | deps_init_id_finish_lhs (void) | |
2475 | { | |
2476 | deps_init_id_data.where = DEPS_IN_INSN; | |
2477 | } | |
2478 | ||
2479 | /* Note a set of REGNO. */ | |
2480 | static void | |
2481 | deps_init_id_note_reg_set (int regno) | |
2482 | { | |
2483 | haifa_note_reg_set (regno); | |
2484 | ||
2485 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2486 | deps_init_id_data.force_use_p = true; | |
2487 | ||
2488 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2489 | SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); | |
2490 | ||
2491 | #ifdef STACK_REGS | |
b8698a0f | 2492 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2493 | renaming to avoid issues with find_used_regs. */ |
2494 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2495 | deps_init_id_data.force_use_p = true; | |
2496 | #endif | |
2497 | } | |
2498 | ||
2499 | /* Note a clobber of REGNO. */ | |
2500 | static void | |
2501 | deps_init_id_note_reg_clobber (int regno) | |
2502 | { | |
2503 | haifa_note_reg_clobber (regno); | |
2504 | ||
2505 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2506 | deps_init_id_data.force_use_p = true; | |
2507 | ||
2508 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2509 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); | |
2510 | } | |
2511 | ||
2512 | /* Note a use of REGNO. */ | |
2513 | static void | |
2514 | deps_init_id_note_reg_use (int regno) | |
2515 | { | |
2516 | haifa_note_reg_use (regno); | |
2517 | ||
2518 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2519 | SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); | |
2520 | } | |
2521 | ||
2522 | /* Start initializing rhs data. */ | |
2523 | static void | |
2524 | deps_init_id_start_rhs (rtx rhs) | |
2525 | { | |
2526 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2527 | ||
2528 | /* And there was no sel_deps_reset_to_insn (). */ | |
2529 | if (IDATA_LHS (deps_init_id_data.id) != NULL) | |
2530 | { | |
2531 | IDATA_RHS (deps_init_id_data.id) = rhs; | |
2532 | deps_init_id_data.where = DEPS_IN_RHS; | |
2533 | } | |
2534 | } | |
2535 | ||
2536 | /* Finish initializing rhs data. */ | |
2537 | static void | |
2538 | deps_init_id_finish_rhs (void) | |
2539 | { | |
2540 | gcc_assert (deps_init_id_data.where == DEPS_IN_RHS | |
2541 | || deps_init_id_data.where == DEPS_IN_INSN); | |
2542 | deps_init_id_data.where = DEPS_IN_INSN; | |
2543 | } | |
2544 | ||
2545 | /* Finish initializing insn data. */ | |
2546 | static void | |
2547 | deps_init_id_finish_insn (void) | |
2548 | { | |
2549 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2550 | ||
2551 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2552 | { | |
2553 | rtx lhs = IDATA_LHS (deps_init_id_data.id); | |
2554 | rtx rhs = IDATA_RHS (deps_init_id_data.id); | |
2555 | ||
2556 | if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) | |
2557 | || deps_init_id_data.force_use_p) | |
2558 | { | |
b8698a0f | 2559 | /* This should be a USE, as we don't want to schedule its RHS |
e855c69d | 2560 | separately. However, we still want to have them recorded |
b8698a0f | 2561 | for the purposes of substitution. That's why we don't |
e855c69d AB |
2562 | simply call downgrade_to_use () here. */ |
2563 | gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); | |
2564 | gcc_assert (!lhs == !rhs); | |
2565 | ||
2566 | IDATA_TYPE (deps_init_id_data.id) = USE; | |
2567 | } | |
2568 | } | |
2569 | ||
2570 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2571 | } | |
2572 | ||
2573 | /* This is dependence info used for initializing insn's data. */ | |
2574 | static struct sched_deps_info_def deps_init_id_sched_deps_info; | |
2575 | ||
2576 | /* This initializes most of the static part of the above structure. */ | |
2577 | static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = | |
2578 | { | |
2579 | NULL, | |
2580 | ||
2581 | deps_init_id_start_insn, | |
2582 | deps_init_id_finish_insn, | |
2583 | deps_init_id_start_lhs, | |
2584 | deps_init_id_finish_lhs, | |
2585 | deps_init_id_start_rhs, | |
2586 | deps_init_id_finish_rhs, | |
2587 | deps_init_id_note_reg_set, | |
2588 | deps_init_id_note_reg_clobber, | |
2589 | deps_init_id_note_reg_use, | |
2590 | NULL, /* note_mem_dep */ | |
2591 | NULL, /* note_dep */ | |
2592 | ||
2593 | 0, /* use_cselib */ | |
2594 | 0, /* use_deps_list */ | |
2595 | 0 /* generate_spec_deps */ | |
2596 | }; | |
2597 | ||
2598 | /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, | |
2599 | we don't actually need information about lhs and rhs. */ | |
2600 | static void | |
2601 | setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) | |
2602 | { | |
2603 | rtx pat = PATTERN (insn); | |
b8698a0f | 2604 | |
481683e1 | 2605 | if (NONJUMP_INSN_P (insn) |
b8698a0f | 2606 | && GET_CODE (pat) == SET |
e855c69d AB |
2607 | && !force_unique_p) |
2608 | { | |
2609 | IDATA_RHS (id) = SET_SRC (pat); | |
2610 | IDATA_LHS (id) = SET_DEST (pat); | |
2611 | } | |
2612 | else | |
2613 | IDATA_LHS (id) = IDATA_RHS (id) = NULL; | |
2614 | } | |
2615 | ||
2616 | /* Possibly downgrade INSN to USE. */ | |
2617 | static void | |
2618 | maybe_downgrade_id_to_use (idata_t id, insn_t insn) | |
2619 | { | |
2620 | bool must_be_use = false; | |
bfac633a | 2621 | df_ref def; |
e855c69d AB |
2622 | rtx lhs = IDATA_LHS (id); |
2623 | rtx rhs = IDATA_RHS (id); | |
b8698a0f | 2624 | |
e855c69d AB |
2625 | /* We downgrade only SETs. */ |
2626 | if (IDATA_TYPE (id) != SET) | |
2627 | return; | |
2628 | ||
2629 | if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) | |
2630 | { | |
2631 | IDATA_TYPE (id) = USE; | |
2632 | return; | |
2633 | } | |
b8698a0f | 2634 | |
bfac633a | 2635 | FOR_EACH_INSN_DEF (def, insn) |
e855c69d | 2636 | { |
e855c69d AB |
2637 | if (DF_REF_INSN (def) |
2638 | && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) | |
2639 | && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) | |
2640 | { | |
2641 | must_be_use = true; | |
2642 | break; | |
2643 | } | |
2644 | ||
2645 | #ifdef STACK_REGS | |
b8698a0f | 2646 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2647 | renaming to avoid issues with find_used_regs. */ |
2648 | if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) | |
2649 | { | |
2650 | must_be_use = true; | |
2651 | break; | |
2652 | } | |
2653 | #endif | |
b8698a0f L |
2654 | } |
2655 | ||
e855c69d AB |
2656 | if (must_be_use) |
2657 | IDATA_TYPE (id) = USE; | |
2658 | } | |
2659 | ||
74ea9ab6 AB |
2660 | /* Setup implicit register clobbers calculated by sched-deps for INSN |
2661 | before reload and save them in ID. */ | |
2662 | static void | |
2663 | setup_id_implicit_regs (idata_t id, insn_t insn) | |
2664 | { | |
2665 | if (reload_completed) | |
2666 | return; | |
2667 | ||
2668 | HARD_REG_SET temp; | |
2669 | unsigned regno; | |
2670 | hard_reg_set_iterator hrsi; | |
2671 | ||
2672 | get_implicit_reg_pending_clobbers (&temp, insn); | |
2673 | EXECUTE_IF_SET_IN_HARD_REG_SET (temp, 0, regno, hrsi) | |
2674 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); | |
2675 | } | |
2676 | ||
e855c69d AB |
2677 | /* Setup register sets describing INSN in ID. */ |
2678 | static void | |
2679 | setup_id_reg_sets (idata_t id, insn_t insn) | |
2680 | { | |
bfac633a RS |
2681 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
2682 | df_ref def, use; | |
e855c69d | 2683 | regset tmp = get_clear_regset_from_pool (); |
b8698a0f | 2684 | |
bfac633a | 2685 | FOR_EACH_INSN_INFO_DEF (def, insn_info) |
e855c69d | 2686 | { |
e855c69d | 2687 | unsigned int regno = DF_REF_REGNO (def); |
b8698a0f | 2688 | |
e855c69d AB |
2689 | /* Post modifies are treated like clobbers by sched-deps.c. */ |
2690 | if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER | |
2691 | | DF_REF_PRE_POST_MODIFY))) | |
2692 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); | |
2693 | else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) | |
2694 | { | |
2695 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); | |
2696 | ||
2697 | #ifdef STACK_REGS | |
b8698a0f | 2698 | /* For stack registers, treat writes to them as writes |
e855c69d AB |
2699 | to the first one to be consistent with sched-deps.c. */ |
2700 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2701 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); | |
2702 | #endif | |
2703 | } | |
2704 | /* Mark special refs that generate read/write def pair. */ | |
2705 | if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) | |
2706 | || regno == STACK_POINTER_REGNUM) | |
2707 | bitmap_set_bit (tmp, regno); | |
2708 | } | |
b8698a0f | 2709 | |
bfac633a | 2710 | FOR_EACH_INSN_INFO_USE (use, insn_info) |
e855c69d | 2711 | { |
e855c69d AB |
2712 | unsigned int regno = DF_REF_REGNO (use); |
2713 | ||
2714 | /* When these refs are met for the first time, skip them, as | |
2715 | these uses are just counterparts of some defs. */ | |
2716 | if (bitmap_bit_p (tmp, regno)) | |
2717 | bitmap_clear_bit (tmp, regno); | |
2718 | else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) | |
2719 | { | |
2720 | SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); | |
2721 | ||
2722 | #ifdef STACK_REGS | |
b8698a0f | 2723 | /* For stack registers, treat reads from them as reads from |
e855c69d AB |
2724 | the first one to be consistent with sched-deps.c. */ |
2725 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2726 | SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); | |
2727 | #endif | |
2728 | } | |
2729 | } | |
2730 | ||
74ea9ab6 AB |
2731 | /* Also get implicit reg clobbers from sched-deps. */ |
2732 | setup_id_implicit_regs (id, insn); | |
2733 | ||
e855c69d AB |
2734 | return_regset_to_pool (tmp); |
2735 | } | |
2736 | ||
2737 | /* Initialize instruction data for INSN in ID using DF's data. */ | |
2738 | static void | |
2739 | init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) | |
2740 | { | |
2741 | gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); | |
2742 | ||
2743 | setup_id_for_insn (id, insn, force_unique_p); | |
2744 | setup_id_lhs_rhs (id, insn, force_unique_p); | |
2745 | ||
2746 | if (INSN_NOP_P (insn)) | |
2747 | return; | |
2748 | ||
2749 | maybe_downgrade_id_to_use (id, insn); | |
2750 | setup_id_reg_sets (id, insn); | |
2751 | } | |
2752 | ||
2753 | /* Initialize instruction data for INSN in ID. */ | |
2754 | static void | |
2755 | deps_init_id (idata_t id, insn_t insn, bool force_unique_p) | |
2756 | { | |
88302d54 | 2757 | struct deps_desc _dc, *dc = &_dc; |
e855c69d AB |
2758 | |
2759 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2760 | deps_init_id_data.id = id; | |
2761 | deps_init_id_data.force_unique_p = force_unique_p; | |
2762 | deps_init_id_data.force_use_p = false; | |
2763 | ||
bcf33775 | 2764 | init_deps (dc, false); |
e855c69d AB |
2765 | memcpy (&deps_init_id_sched_deps_info, |
2766 | &const_deps_init_id_sched_deps_info, | |
2767 | sizeof (deps_init_id_sched_deps_info)); | |
e855c69d AB |
2768 | if (spec_info != NULL) |
2769 | deps_init_id_sched_deps_info.generate_spec_deps = 1; | |
e855c69d AB |
2770 | sched_deps_info = &deps_init_id_sched_deps_info; |
2771 | ||
6144a836 | 2772 | deps_analyze_insn (dc, insn); |
74ea9ab6 AB |
2773 | /* Implicit reg clobbers received from sched-deps separately. */ |
2774 | setup_id_implicit_regs (id, insn); | |
e855c69d AB |
2775 | |
2776 | free_deps (dc); | |
e855c69d AB |
2777 | deps_init_id_data.id = NULL; |
2778 | } | |
2779 | ||
2780 | \f | |
a95b23b4 BS |
2781 | struct sched_scan_info_def |
2782 | { | |
2783 | /* This hook notifies scheduler frontend to extend its internal per basic | |
2784 | block data structures. This hook should be called once before a series of | |
2785 | calls to bb_init (). */ | |
2786 | void (*extend_bb) (void); | |
2787 | ||
2788 | /* This hook makes scheduler frontend to initialize its internal data | |
2789 | structures for the passed basic block. */ | |
2790 | void (*init_bb) (basic_block); | |
2791 | ||
2792 | /* This hook notifies scheduler frontend to extend its internal per insn data | |
2793 | structures. This hook should be called once before a series of calls to | |
2794 | insn_init (). */ | |
2795 | void (*extend_insn) (void); | |
2796 | ||
2797 | /* This hook makes scheduler frontend to initialize its internal data | |
2798 | structures for the passed insn. */ | |
6144a836 | 2799 | void (*init_insn) (insn_t); |
a95b23b4 BS |
2800 | }; |
2801 | ||
2802 | /* A driver function to add a set of basic blocks (BBS) to the | |
2803 | scheduling region. */ | |
2804 | static void | |
2805 | sched_scan (const struct sched_scan_info_def *ssi, bb_vec_t bbs) | |
2806 | { | |
2807 | unsigned i; | |
2808 | basic_block bb; | |
2809 | ||
2810 | if (ssi->extend_bb) | |
2811 | ssi->extend_bb (); | |
2812 | ||
2813 | if (ssi->init_bb) | |
9771b263 | 2814 | FOR_EACH_VEC_ELT (bbs, i, bb) |
a95b23b4 BS |
2815 | ssi->init_bb (bb); |
2816 | ||
2817 | if (ssi->extend_insn) | |
2818 | ssi->extend_insn (); | |
2819 | ||
2820 | if (ssi->init_insn) | |
9771b263 | 2821 | FOR_EACH_VEC_ELT (bbs, i, bb) |
a95b23b4 | 2822 | { |
6144a836 | 2823 | rtx_insn *insn; |
a95b23b4 BS |
2824 | |
2825 | FOR_BB_INSNS (bb, insn) | |
2826 | ssi->init_insn (insn); | |
2827 | } | |
2828 | } | |
e855c69d AB |
2829 | |
2830 | /* Implement hooks for collecting fundamental insn properties like if insn is | |
2831 | an ASM or is within a SCHED_GROUP. */ | |
2832 | ||
2833 | /* True when a "one-time init" data for INSN was already inited. */ | |
2834 | static bool | |
2835 | first_time_insn_init (insn_t insn) | |
2836 | { | |
2837 | return INSN_LIVE (insn) == NULL; | |
2838 | } | |
2839 | ||
2840 | /* Hash an entry in a transformed_insns hashtable. */ | |
2841 | static hashval_t | |
2842 | hash_transformed_insns (const void *p) | |
2843 | { | |
2844 | return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2845 | } | |
2846 | ||
2847 | /* Compare the entries in a transformed_insns hashtable. */ | |
2848 | static int | |
2849 | eq_transformed_insns (const void *p, const void *q) | |
2850 | { | |
9ee1fbb1 DM |
2851 | rtx_insn *i1 = |
2852 | VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2853 | rtx_insn *i2 = | |
2854 | VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); | |
e855c69d AB |
2855 | |
2856 | if (INSN_UID (i1) == INSN_UID (i2)) | |
2857 | return 1; | |
2858 | return rtx_equal_p (PATTERN (i1), PATTERN (i2)); | |
2859 | } | |
2860 | ||
2861 | /* Free an entry in a transformed_insns hashtable. */ | |
2862 | static void | |
2863 | free_transformed_insns (void *p) | |
2864 | { | |
2865 | struct transformed_insns *pti = (struct transformed_insns *) p; | |
2866 | ||
2867 | vinsn_detach (pti->vinsn_old); | |
2868 | vinsn_detach (pti->vinsn_new); | |
2869 | free (pti); | |
2870 | } | |
2871 | ||
b8698a0f | 2872 | /* Init the s_i_d data for INSN which should be inited just once, when |
e855c69d AB |
2873 | we first see the insn. */ |
2874 | static void | |
2875 | init_first_time_insn_data (insn_t insn) | |
2876 | { | |
2877 | /* This should not be set if this is the first time we init data for | |
2878 | insn. */ | |
2879 | gcc_assert (first_time_insn_init (insn)); | |
b8698a0f | 2880 | |
e855c69d AB |
2881 | /* These are needed for nops too. */ |
2882 | INSN_LIVE (insn) = get_regset_from_pool (); | |
2883 | INSN_LIVE_VALID_P (insn) = false; | |
bcf33775 | 2884 | |
e855c69d AB |
2885 | if (!INSN_NOP_P (insn)) |
2886 | { | |
2887 | INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); | |
2888 | INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); | |
b8698a0f | 2889 | INSN_TRANSFORMED_INSNS (insn) |
e855c69d AB |
2890 | = htab_create (16, hash_transformed_insns, |
2891 | eq_transformed_insns, free_transformed_insns); | |
bcf33775 | 2892 | init_deps (&INSN_DEPS_CONTEXT (insn), true); |
e855c69d AB |
2893 | } |
2894 | } | |
2895 | ||
b8698a0f | 2896 | /* Free almost all above data for INSN that is scheduled already. |
bcf33775 AB |
2897 | Used for extra-large basic blocks. */ |
2898 | void | |
2899 | free_data_for_scheduled_insn (insn_t insn) | |
e855c69d AB |
2900 | { |
2901 | gcc_assert (! first_time_insn_init (insn)); | |
b8698a0f | 2902 | |
bcf33775 AB |
2903 | if (! INSN_ANALYZED_DEPS (insn)) |
2904 | return; | |
b8698a0f | 2905 | |
e855c69d AB |
2906 | BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); |
2907 | BITMAP_FREE (INSN_FOUND_DEPS (insn)); | |
2908 | htab_delete (INSN_TRANSFORMED_INSNS (insn)); | |
b8698a0f | 2909 | |
e855c69d AB |
2910 | /* This is allocated only for bookkeeping insns. */ |
2911 | if (INSN_ORIGINATORS (insn)) | |
2912 | BITMAP_FREE (INSN_ORIGINATORS (insn)); | |
2913 | free_deps (&INSN_DEPS_CONTEXT (insn)); | |
bcf33775 AB |
2914 | |
2915 | INSN_ANALYZED_DEPS (insn) = NULL; | |
2916 | ||
b8698a0f | 2917 | /* Clear the readonly flag so we would ICE when trying to recalculate |
bcf33775 AB |
2918 | the deps context (as we believe that it should not happen). */ |
2919 | (&INSN_DEPS_CONTEXT (insn))->readonly = 0; | |
2920 | } | |
2921 | ||
2922 | /* Free the same data as above for INSN. */ | |
2923 | static void | |
2924 | free_first_time_insn_data (insn_t insn) | |
2925 | { | |
2926 | gcc_assert (! first_time_insn_init (insn)); | |
2927 | ||
2928 | free_data_for_scheduled_insn (insn); | |
2929 | return_regset_to_pool (INSN_LIVE (insn)); | |
2930 | INSN_LIVE (insn) = NULL; | |
2931 | INSN_LIVE_VALID_P (insn) = false; | |
e855c69d AB |
2932 | } |
2933 | ||
2934 | /* Initialize region-scope data structures for basic blocks. */ | |
2935 | static void | |
2936 | init_global_and_expr_for_bb (basic_block bb) | |
2937 | { | |
2938 | if (sel_bb_empty_p (bb)) | |
2939 | return; | |
2940 | ||
2941 | invalidate_av_set (bb); | |
2942 | } | |
2943 | ||
2944 | /* Data for global dependency analysis (to initialize CANT_MOVE and | |
2945 | SCHED_GROUP_P). */ | |
2946 | static struct | |
2947 | { | |
2948 | /* Previous insn. */ | |
2949 | insn_t prev_insn; | |
2950 | } init_global_data; | |
2951 | ||
2952 | /* Determine if INSN is in the sched_group, is an asm or should not be | |
2953 | cloned. After that initialize its expr. */ | |
2954 | static void | |
2955 | init_global_and_expr_for_insn (insn_t insn) | |
2956 | { | |
2957 | if (LABEL_P (insn)) | |
2958 | return; | |
2959 | ||
2960 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2961 | { | |
6144a836 | 2962 | init_global_data.prev_insn = NULL; |
e855c69d AB |
2963 | return; |
2964 | } | |
2965 | ||
2966 | gcc_assert (INSN_P (insn)); | |
2967 | ||
2968 | if (SCHED_GROUP_P (insn)) | |
2969 | /* Setup a sched_group. */ | |
2970 | { | |
2971 | insn_t prev_insn = init_global_data.prev_insn; | |
2972 | ||
2973 | if (prev_insn) | |
2974 | INSN_SCHED_NEXT (prev_insn) = insn; | |
2975 | ||
2976 | init_global_data.prev_insn = insn; | |
2977 | } | |
2978 | else | |
6144a836 | 2979 | init_global_data.prev_insn = NULL; |
e855c69d AB |
2980 | |
2981 | if (GET_CODE (PATTERN (insn)) == ASM_INPUT | |
2982 | || asm_noperands (PATTERN (insn)) >= 0) | |
2983 | /* Mark INSN as an asm. */ | |
2984 | INSN_ASM_P (insn) = true; | |
2985 | ||
2986 | { | |
2987 | bool force_unique_p; | |
2988 | ds_t spec_done_ds; | |
2989 | ||
cfeb0fa8 AB |
2990 | /* Certain instructions cannot be cloned, and frame related insns and |
2991 | the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of | |
2992 | their block. */ | |
2993 | if (prologue_epilogue_contains (insn)) | |
2994 | { | |
2995 | if (RTX_FRAME_RELATED_P (insn)) | |
2996 | CANT_MOVE (insn) = 1; | |
2997 | else | |
2998 | { | |
2999 | rtx note; | |
3000 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
3001 | if (REG_NOTE_KIND (note) == REG_SAVE_NOTE | |
3002 | && ((enum insn_note) INTVAL (XEXP (note, 0)) | |
3003 | == NOTE_INSN_EPILOGUE_BEG)) | |
3004 | { | |
3005 | CANT_MOVE (insn) = 1; | |
3006 | break; | |
3007 | } | |
3008 | } | |
3009 | force_unique_p = true; | |
3010 | } | |
e855c69d | 3011 | else |
cfeb0fa8 AB |
3012 | if (CANT_MOVE (insn) |
3013 | || INSN_ASM_P (insn) | |
3014 | || SCHED_GROUP_P (insn) | |
07643d76 | 3015 | || CALL_P (insn) |
cfeb0fa8 AB |
3016 | /* Exception handling insns are always unique. */ |
3017 | || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) | |
3018 | /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ | |
6bf2d156 DM |
3019 | || control_flow_insn_p (insn) |
3020 | || volatile_insn_p (PATTERN (insn)) | |
3021 | || (targetm.cannot_copy_insn_p | |
3022 | && targetm.cannot_copy_insn_p (insn))) | |
cfeb0fa8 AB |
3023 | force_unique_p = true; |
3024 | else | |
3025 | force_unique_p = false; | |
e855c69d AB |
3026 | |
3027 | if (targetm.sched.get_insn_spec_ds) | |
3028 | { | |
3029 | spec_done_ds = targetm.sched.get_insn_spec_ds (insn); | |
3030 | spec_done_ds = ds_get_max_dep_weak (spec_done_ds); | |
3031 | } | |
3032 | else | |
3033 | spec_done_ds = 0; | |
3034 | ||
3035 | /* Initialize INSN's expr. */ | |
3036 | init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0, | |
3037 | REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn), | |
6e1aa848 | 3038 | spec_done_ds, 0, 0, vNULL, true, |
9771b263 | 3039 | false, false, false, CANT_MOVE (insn)); |
e855c69d AB |
3040 | } |
3041 | ||
3042 | init_first_time_insn_data (insn); | |
3043 | } | |
3044 | ||
3045 | /* Scan the region and initialize instruction data for basic blocks BBS. */ | |
3046 | void | |
3047 | sel_init_global_and_expr (bb_vec_t bbs) | |
3048 | { | |
3049 | /* ??? It would be nice to implement push / pop scheme for sched_infos. */ | |
3050 | const struct sched_scan_info_def ssi = | |
3051 | { | |
3052 | NULL, /* extend_bb */ | |
3053 | init_global_and_expr_for_bb, /* init_bb */ | |
3054 | extend_insn_data, /* extend_insn */ | |
3055 | init_global_and_expr_for_insn /* init_insn */ | |
3056 | }; | |
b8698a0f | 3057 | |
a95b23b4 | 3058 | sched_scan (&ssi, bbs); |
e855c69d AB |
3059 | } |
3060 | ||
3061 | /* Finalize region-scope data structures for basic blocks. */ | |
3062 | static void | |
3063 | finish_global_and_expr_for_bb (basic_block bb) | |
3064 | { | |
3065 | av_set_clear (&BB_AV_SET (bb)); | |
3066 | BB_AV_LEVEL (bb) = 0; | |
3067 | } | |
3068 | ||
3069 | /* Finalize INSN's data. */ | |
3070 | static void | |
3071 | finish_global_and_expr_insn (insn_t insn) | |
3072 | { | |
3073 | if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) | |
3074 | return; | |
3075 | ||
3076 | gcc_assert (INSN_P (insn)); | |
3077 | ||
3078 | if (INSN_LUID (insn) > 0) | |
3079 | { | |
3080 | free_first_time_insn_data (insn); | |
3081 | INSN_WS_LEVEL (insn) = 0; | |
3082 | CANT_MOVE (insn) = 0; | |
b8698a0f L |
3083 | |
3084 | /* We can no longer assert this, as vinsns of this insn could be | |
3085 | easily live in other insn's caches. This should be changed to | |
e855c69d AB |
3086 | a counter-like approach among all vinsns. */ |
3087 | gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); | |
3088 | clear_expr (INSN_EXPR (insn)); | |
3089 | } | |
3090 | } | |
3091 | ||
3092 | /* Finalize per instruction data for the whole region. */ | |
3093 | void | |
3094 | sel_finish_global_and_expr (void) | |
3095 | { | |
3096 | { | |
3097 | bb_vec_t bbs; | |
3098 | int i; | |
3099 | ||
9771b263 | 3100 | bbs.create (current_nr_blocks); |
e855c69d AB |
3101 | |
3102 | for (i = 0; i < current_nr_blocks; i++) | |
06e28de2 | 3103 | bbs.quick_push (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))); |
e855c69d AB |
3104 | |
3105 | /* Clear AV_SETs and INSN_EXPRs. */ | |
3106 | { | |
3107 | const struct sched_scan_info_def ssi = | |
3108 | { | |
3109 | NULL, /* extend_bb */ | |
3110 | finish_global_and_expr_for_bb, /* init_bb */ | |
3111 | NULL, /* extend_insn */ | |
3112 | finish_global_and_expr_insn /* init_insn */ | |
3113 | }; | |
3114 | ||
a95b23b4 | 3115 | sched_scan (&ssi, bbs); |
e855c69d AB |
3116 | } |
3117 | ||
9771b263 | 3118 | bbs.release (); |
e855c69d AB |
3119 | } |
3120 | ||
3121 | finish_insns (); | |
3122 | } | |
3123 | \f | |
3124 | ||
b8698a0f L |
3125 | /* In the below hooks, we merely calculate whether or not a dependence |
3126 | exists, and in what part of insn. However, we will need more data | |
e855c69d AB |
3127 | when we'll start caching dependence requests. */ |
3128 | ||
3129 | /* Container to hold information for dependency analysis. */ | |
3130 | static struct | |
3131 | { | |
3132 | deps_t dc; | |
3133 | ||
3134 | /* A variable to track which part of rtx we are scanning in | |
3135 | sched-deps.c: sched_analyze_insn (). */ | |
3136 | deps_where_t where; | |
3137 | ||
3138 | /* Current producer. */ | |
3139 | insn_t pro; | |
3140 | ||
3141 | /* Current consumer. */ | |
3142 | vinsn_t con; | |
3143 | ||
3144 | /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. | |
3145 | X is from { INSN, LHS, RHS }. */ | |
3146 | ds_t has_dep_p[DEPS_IN_NOWHERE]; | |
3147 | } has_dependence_data; | |
3148 | ||
3149 | /* Start analyzing dependencies of INSN. */ | |
3150 | static void | |
3151 | has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) | |
3152 | { | |
3153 | gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); | |
3154 | ||
3155 | has_dependence_data.where = DEPS_IN_INSN; | |
3156 | } | |
3157 | ||
3158 | /* Finish analyzing dependencies of an insn. */ | |
3159 | static void | |
3160 | has_dependence_finish_insn (void) | |
3161 | { | |
3162 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3163 | ||
3164 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3165 | } | |
3166 | ||
3167 | /* Start analyzing dependencies of LHS. */ | |
3168 | static void | |
3169 | has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) | |
3170 | { | |
3171 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3172 | ||
3173 | if (VINSN_LHS (has_dependence_data.con) != NULL) | |
3174 | has_dependence_data.where = DEPS_IN_LHS; | |
3175 | } | |
3176 | ||
3177 | /* Finish analyzing dependencies of an lhs. */ | |
3178 | static void | |
3179 | has_dependence_finish_lhs (void) | |
3180 | { | |
3181 | has_dependence_data.where = DEPS_IN_INSN; | |
3182 | } | |
3183 | ||
3184 | /* Start analyzing dependencies of RHS. */ | |
3185 | static void | |
3186 | has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) | |
3187 | { | |
3188 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3189 | ||
3190 | if (VINSN_RHS (has_dependence_data.con) != NULL) | |
3191 | has_dependence_data.where = DEPS_IN_RHS; | |
3192 | } | |
3193 | ||
3194 | /* Start analyzing dependencies of an rhs. */ | |
3195 | static void | |
3196 | has_dependence_finish_rhs (void) | |
3197 | { | |
3198 | gcc_assert (has_dependence_data.where == DEPS_IN_RHS | |
3199 | || has_dependence_data.where == DEPS_IN_INSN); | |
3200 | ||
3201 | has_dependence_data.where = DEPS_IN_INSN; | |
3202 | } | |
3203 | ||
3204 | /* Note a set of REGNO. */ | |
3205 | static void | |
3206 | has_dependence_note_reg_set (int regno) | |
3207 | { | |
3208 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3209 | ||
3210 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3211 | VINSN_INSN_RTX | |
3212 | (has_dependence_data.con))) | |
3213 | { | |
3214 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3215 | ||
3216 | if (reg_last->sets != NULL | |
3217 | || reg_last->clobbers != NULL) | |
3218 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
3219 | ||
50919d13 | 3220 | if (reg_last->uses || reg_last->implicit_sets) |
e855c69d AB |
3221 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3222 | } | |
3223 | } | |
3224 | ||
3225 | /* Note a clobber of REGNO. */ | |
3226 | static void | |
3227 | has_dependence_note_reg_clobber (int regno) | |
3228 | { | |
3229 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3230 | ||
3231 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3232 | VINSN_INSN_RTX | |
3233 | (has_dependence_data.con))) | |
3234 | { | |
3235 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3236 | ||
3237 | if (reg_last->sets) | |
3238 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
b8698a0f | 3239 | |
50919d13 | 3240 | if (reg_last->uses || reg_last->implicit_sets) |
e855c69d AB |
3241 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3242 | } | |
3243 | } | |
3244 | ||
3245 | /* Note a use of REGNO. */ | |
3246 | static void | |
3247 | has_dependence_note_reg_use (int regno) | |
3248 | { | |
3249 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3250 | ||
3251 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3252 | VINSN_INSN_RTX | |
3253 | (has_dependence_data.con))) | |
3254 | { | |
3255 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3256 | ||
3257 | if (reg_last->sets) | |
3258 | *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; | |
3259 | ||
50919d13 | 3260 | if (reg_last->clobbers || reg_last->implicit_sets) |
e855c69d AB |
3261 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3262 | ||
0d4acd90 AB |
3263 | /* Merge BE_IN_SPEC bits into *DSP when the dependency producer |
3264 | is actually a check insn. We need to do this for any register | |
3265 | read-read dependency with the check unless we track properly | |
3266 | all registers written by BE_IN_SPEC-speculated insns, as | |
3267 | we don't have explicit dependence lists. See PR 53975. */ | |
e855c69d AB |
3268 | if (reg_last->uses) |
3269 | { | |
3270 | ds_t pro_spec_checked_ds; | |
3271 | ||
3272 | pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); | |
3273 | pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); | |
3274 | ||
0d4acd90 | 3275 | if (pro_spec_checked_ds != 0) |
e855c69d AB |
3276 | *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, |
3277 | NULL_RTX, NULL_RTX); | |
3278 | } | |
3279 | } | |
3280 | } | |
3281 | ||
3282 | /* Note a memory dependence. */ | |
3283 | static void | |
3284 | has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, | |
3285 | rtx pending_mem ATTRIBUTE_UNUSED, | |
3286 | insn_t pending_insn ATTRIBUTE_UNUSED, | |
3287 | ds_t ds ATTRIBUTE_UNUSED) | |
3288 | { | |
3289 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3290 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3291 | { | |
3292 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3293 | ||
3294 | *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); | |
3295 | } | |
3296 | } | |
3297 | ||
3298 | /* Note a dependence. */ | |
3299 | static void | |
3300 | has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED, | |
3301 | ds_t ds ATTRIBUTE_UNUSED) | |
3302 | { | |
3303 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3304 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3305 | { | |
3306 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3307 | ||
3308 | *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); | |
3309 | } | |
3310 | } | |
3311 | ||
3312 | /* Mark the insn as having a hard dependence that prevents speculation. */ | |
3313 | void | |
3314 | sel_mark_hard_insn (rtx insn) | |
3315 | { | |
3316 | int i; | |
3317 | ||
3318 | /* Only work when we're in has_dependence_p mode. | |
3319 | ??? This is a hack, this should actually be a hook. */ | |
3320 | if (!has_dependence_data.dc || !has_dependence_data.pro) | |
3321 | return; | |
3322 | ||
3323 | gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); | |
3324 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3325 | ||
3326 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3327 | has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; | |
3328 | } | |
3329 | ||
3330 | /* This structure holds the hooks for the dependency analysis used when | |
3331 | actually processing dependencies in the scheduler. */ | |
3332 | static struct sched_deps_info_def has_dependence_sched_deps_info; | |
3333 | ||
3334 | /* This initializes most of the fields of the above structure. */ | |
3335 | static const struct sched_deps_info_def const_has_dependence_sched_deps_info = | |
3336 | { | |
3337 | NULL, | |
3338 | ||
3339 | has_dependence_start_insn, | |
3340 | has_dependence_finish_insn, | |
3341 | has_dependence_start_lhs, | |
3342 | has_dependence_finish_lhs, | |
3343 | has_dependence_start_rhs, | |
3344 | has_dependence_finish_rhs, | |
3345 | has_dependence_note_reg_set, | |
3346 | has_dependence_note_reg_clobber, | |
3347 | has_dependence_note_reg_use, | |
3348 | has_dependence_note_mem_dep, | |
3349 | has_dependence_note_dep, | |
3350 | ||
3351 | 0, /* use_cselib */ | |
3352 | 0, /* use_deps_list */ | |
3353 | 0 /* generate_spec_deps */ | |
3354 | }; | |
3355 | ||
3356 | /* Initialize has_dependence_sched_deps_info with extra spec field. */ | |
3357 | static void | |
3358 | setup_has_dependence_sched_deps_info (void) | |
3359 | { | |
3360 | memcpy (&has_dependence_sched_deps_info, | |
3361 | &const_has_dependence_sched_deps_info, | |
3362 | sizeof (has_dependence_sched_deps_info)); | |
3363 | ||
3364 | if (spec_info != NULL) | |
3365 | has_dependence_sched_deps_info.generate_spec_deps = 1; | |
3366 | ||
3367 | sched_deps_info = &has_dependence_sched_deps_info; | |
3368 | } | |
3369 | ||
3370 | /* Remove all dependences found and recorded in has_dependence_data array. */ | |
3371 | void | |
3372 | sel_clear_has_dependence (void) | |
3373 | { | |
3374 | int i; | |
3375 | ||
3376 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3377 | has_dependence_data.has_dep_p[i] = 0; | |
3378 | } | |
3379 | ||
3380 | /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer | |
3381 | to the dependence information array in HAS_DEP_PP. */ | |
3382 | ds_t | |
3383 | has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) | |
3384 | { | |
3385 | int i; | |
3386 | ds_t ds; | |
88302d54 | 3387 | struct deps_desc *dc; |
e855c69d AB |
3388 | |
3389 | if (INSN_SIMPLEJUMP_P (pred)) | |
3390 | /* Unconditional jump is just a transfer of control flow. | |
3391 | Ignore it. */ | |
3392 | return false; | |
3393 | ||
3394 | dc = &INSN_DEPS_CONTEXT (pred); | |
bcf33775 AB |
3395 | |
3396 | /* We init this field lazily. */ | |
3397 | if (dc->reg_last == NULL) | |
3398 | init_deps_reg_last (dc); | |
b8698a0f | 3399 | |
e855c69d AB |
3400 | if (!dc->readonly) |
3401 | { | |
3402 | has_dependence_data.pro = NULL; | |
3403 | /* Initialize empty dep context with information about PRED. */ | |
3404 | advance_deps_context (dc, pred); | |
3405 | dc->readonly = 1; | |
3406 | } | |
3407 | ||
3408 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3409 | has_dependence_data.pro = pred; | |
3410 | has_dependence_data.con = EXPR_VINSN (expr); | |
3411 | has_dependence_data.dc = dc; | |
3412 | ||
3413 | sel_clear_has_dependence (); | |
3414 | ||
3415 | /* Now catch all dependencies that would be generated between PRED and | |
3416 | INSN. */ | |
3417 | setup_has_dependence_sched_deps_info (); | |
3418 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3419 | has_dependence_data.dc = NULL; | |
3420 | ||
3421 | /* When a barrier was found, set DEPS_IN_INSN bits. */ | |
3422 | if (dc->last_reg_pending_barrier == TRUE_BARRIER) | |
3423 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; | |
3424 | else if (dc->last_reg_pending_barrier == MOVE_BARRIER) | |
3425 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
3426 | ||
3427 | /* Do not allow stores to memory to move through checks. Currently | |
3428 | we don't move this to sched-deps.c as the check doesn't have | |
b8698a0f | 3429 | obvious places to which this dependence can be attached. |
e855c69d AB |
3430 | FIMXE: this should go to a hook. */ |
3431 | if (EXPR_LHS (expr) | |
3432 | && MEM_P (EXPR_LHS (expr)) | |
3433 | && sel_insn_is_speculation_check (pred)) | |
3434 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
b8698a0f | 3435 | |
e855c69d AB |
3436 | *has_dep_pp = has_dependence_data.has_dep_p; |
3437 | ds = 0; | |
3438 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3439 | ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], | |
3440 | NULL_RTX, NULL_RTX); | |
3441 | ||
3442 | return ds; | |
3443 | } | |
3444 | \f | |
3445 | ||
b8698a0f L |
3446 | /* Dependence hooks implementation that checks dependence latency constraints |
3447 | on the insns being scheduled. The entry point for these routines is | |
3448 | tick_check_p predicate. */ | |
e855c69d AB |
3449 | |
3450 | static struct | |
3451 | { | |
3452 | /* An expr we are currently checking. */ | |
3453 | expr_t expr; | |
3454 | ||
3455 | /* A minimal cycle for its scheduling. */ | |
3456 | int cycle; | |
3457 | ||
3458 | /* Whether we have seen a true dependence while checking. */ | |
3459 | bool seen_true_dep_p; | |
3460 | } tick_check_data; | |
3461 | ||
3462 | /* Update minimal scheduling cycle for tick_check_insn given that it depends | |
3463 | on PRO with status DS and weight DW. */ | |
3464 | static void | |
3465 | tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) | |
3466 | { | |
3467 | expr_t con_expr = tick_check_data.expr; | |
3468 | insn_t con_insn = EXPR_INSN_RTX (con_expr); | |
3469 | ||
3470 | if (con_insn != pro_insn) | |
3471 | { | |
3472 | enum reg_note dt; | |
3473 | int tick; | |
3474 | ||
3475 | if (/* PROducer was removed from above due to pipelining. */ | |
3476 | !INSN_IN_STREAM_P (pro_insn) | |
3477 | /* Or PROducer was originally on the next iteration regarding the | |
3478 | CONsumer. */ | |
3479 | || (INSN_SCHED_TIMES (pro_insn) | |
3480 | - EXPR_SCHED_TIMES (con_expr)) > 1) | |
3481 | /* Don't count this dependence. */ | |
3482 | return; | |
3483 | ||
3484 | dt = ds_to_dt (ds); | |
3485 | if (dt == REG_DEP_TRUE) | |
3486 | tick_check_data.seen_true_dep_p = true; | |
3487 | ||
3488 | gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); | |
3489 | ||
3490 | { | |
3491 | dep_def _dep, *dep = &_dep; | |
3492 | ||
3493 | init_dep (dep, pro_insn, con_insn, dt); | |
3494 | ||
3495 | tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); | |
3496 | } | |
3497 | ||
3498 | /* When there are several kinds of dependencies between pro and con, | |
3499 | only REG_DEP_TRUE should be taken into account. */ | |
3500 | if (tick > tick_check_data.cycle | |
3501 | && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) | |
3502 | tick_check_data.cycle = tick; | |
3503 | } | |
3504 | } | |
3505 | ||
3506 | /* An implementation of note_dep hook. */ | |
3507 | static void | |
3508 | tick_check_note_dep (insn_t pro, ds_t ds) | |
3509 | { | |
3510 | tick_check_dep_with_dw (pro, ds, 0); | |
3511 | } | |
3512 | ||
3513 | /* An implementation of note_mem_dep hook. */ | |
3514 | static void | |
3515 | tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) | |
3516 | { | |
3517 | dw_t dw; | |
3518 | ||
3519 | dw = (ds_to_dt (ds) == REG_DEP_TRUE | |
3520 | ? estimate_dep_weak (mem1, mem2) | |
3521 | : 0); | |
3522 | ||
3523 | tick_check_dep_with_dw (pro, ds, dw); | |
3524 | } | |
3525 | ||
3526 | /* This structure contains hooks for dependence analysis used when determining | |
3527 | whether an insn is ready for scheduling. */ | |
3528 | static struct sched_deps_info_def tick_check_sched_deps_info = | |
3529 | { | |
3530 | NULL, | |
3531 | ||
3532 | NULL, | |
3533 | NULL, | |
3534 | NULL, | |
3535 | NULL, | |
3536 | NULL, | |
3537 | NULL, | |
3538 | haifa_note_reg_set, | |
3539 | haifa_note_reg_clobber, | |
3540 | haifa_note_reg_use, | |
3541 | tick_check_note_mem_dep, | |
3542 | tick_check_note_dep, | |
3543 | ||
3544 | 0, 0, 0 | |
3545 | }; | |
3546 | ||
3547 | /* Estimate number of cycles from the current cycle of FENCE until EXPR can be | |
3548 | scheduled. Return 0 if all data from producers in DC is ready. */ | |
3549 | int | |
3550 | tick_check_p (expr_t expr, deps_t dc, fence_t fence) | |
3551 | { | |
3552 | int cycles_left; | |
3553 | /* Initialize variables. */ | |
3554 | tick_check_data.expr = expr; | |
3555 | tick_check_data.cycle = 0; | |
3556 | tick_check_data.seen_true_dep_p = false; | |
3557 | sched_deps_info = &tick_check_sched_deps_info; | |
b8698a0f | 3558 | |
e855c69d AB |
3559 | gcc_assert (!dc->readonly); |
3560 | dc->readonly = 1; | |
3561 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3562 | dc->readonly = 0; | |
3563 | ||
3564 | cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); | |
3565 | ||
3566 | return cycles_left >= 0 ? cycles_left : 0; | |
3567 | } | |
3568 | \f | |
3569 | ||
3570 | /* Functions to work with insns. */ | |
3571 | ||
3572 | /* Returns true if LHS of INSN is the same as DEST of an insn | |
3573 | being moved. */ | |
3574 | bool | |
3575 | lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) | |
3576 | { | |
3577 | rtx lhs = INSN_LHS (insn); | |
3578 | ||
3579 | if (lhs == NULL || dest == NULL) | |
3580 | return false; | |
b8698a0f | 3581 | |
e855c69d AB |
3582 | return rtx_equal_p (lhs, dest); |
3583 | } | |
3584 | ||
3585 | /* Return s_i_d entry of INSN. Callable from debugger. */ | |
3586 | sel_insn_data_def | |
3587 | insn_sid (insn_t insn) | |
3588 | { | |
3589 | return *SID (insn); | |
3590 | } | |
3591 | ||
3592 | /* True when INSN is a speculative check. We can tell this by looking | |
3593 | at the data structures of the selective scheduler, not by examining | |
3594 | the pattern. */ | |
3595 | bool | |
3596 | sel_insn_is_speculation_check (rtx insn) | |
3597 | { | |
9771b263 | 3598 | return s_i_d.exists () && !! INSN_SPEC_CHECKED_DS (insn); |
e855c69d AB |
3599 | } |
3600 | ||
b8698a0f | 3601 | /* Extracts machine mode MODE and destination location DST_LOC |
e855c69d AB |
3602 | for given INSN. */ |
3603 | void | |
ef4bddc2 | 3604 | get_dest_and_mode (rtx insn, rtx *dst_loc, machine_mode *mode) |
e855c69d AB |
3605 | { |
3606 | rtx pat = PATTERN (insn); | |
3607 | ||
3608 | gcc_assert (dst_loc); | |
3609 | gcc_assert (GET_CODE (pat) == SET); | |
3610 | ||
3611 | *dst_loc = SET_DEST (pat); | |
3612 | ||
3613 | gcc_assert (*dst_loc); | |
3614 | gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); | |
3615 | ||
3616 | if (mode) | |
3617 | *mode = GET_MODE (*dst_loc); | |
3618 | } | |
3619 | ||
b8698a0f | 3620 | /* Returns true when moving through JUMP will result in bookkeeping |
e855c69d AB |
3621 | creation. */ |
3622 | bool | |
3623 | bookkeeping_can_be_created_if_moved_through_p (insn_t jump) | |
3624 | { | |
3625 | insn_t succ; | |
3626 | succ_iterator si; | |
3627 | ||
3628 | FOR_EACH_SUCC (succ, si, jump) | |
3629 | if (sel_num_cfg_preds_gt_1 (succ)) | |
3630 | return true; | |
3631 | ||
3632 | return false; | |
3633 | } | |
3634 | ||
3635 | /* Return 'true' if INSN is the only one in its basic block. */ | |
3636 | static bool | |
3637 | insn_is_the_only_one_in_bb_p (insn_t insn) | |
3638 | { | |
3639 | return sel_bb_head_p (insn) && sel_bb_end_p (insn); | |
3640 | } | |
3641 | ||
b8698a0f | 3642 | /* Check that the region we're scheduling still has at most one |
e855c69d AB |
3643 | backedge. */ |
3644 | static void | |
3645 | verify_backedges (void) | |
3646 | { | |
3647 | if (pipelining_p) | |
3648 | { | |
3649 | int i, n = 0; | |
3650 | edge e; | |
3651 | edge_iterator ei; | |
b8698a0f | 3652 | |
e855c69d | 3653 | for (i = 0; i < current_nr_blocks; i++) |
06e28de2 | 3654 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))->succs) |
e855c69d AB |
3655 | if (in_current_region_p (e->dest) |
3656 | && BLOCK_TO_BB (e->dest->index) < i) | |
3657 | n++; | |
b8698a0f | 3658 | |
e855c69d AB |
3659 | gcc_assert (n <= 1); |
3660 | } | |
3661 | } | |
e855c69d AB |
3662 | \f |
3663 | ||
3664 | /* Functions to work with control flow. */ | |
3665 | ||
b59ab570 AM |
3666 | /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks |
3667 | are sorted in topological order (it might have been invalidated by | |
3668 | redirecting an edge). */ | |
3669 | static void | |
3670 | sel_recompute_toporder (void) | |
3671 | { | |
3672 | int i, n, rgn; | |
3673 | int *postorder, n_blocks; | |
3674 | ||
0cae8d31 | 3675 | postorder = XALLOCAVEC (int, n_basic_blocks_for_fn (cfun)); |
b59ab570 AM |
3676 | n_blocks = post_order_compute (postorder, false, false); |
3677 | ||
3678 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
3679 | for (n = 0, i = n_blocks - 1; i >= 0; i--) | |
3680 | if (CONTAINING_RGN (postorder[i]) == rgn) | |
3681 | { | |
3682 | BLOCK_TO_BB (postorder[i]) = n; | |
3683 | BB_TO_BLOCK (n) = postorder[i]; | |
3684 | n++; | |
3685 | } | |
3686 | ||
3687 | /* Assert that we updated info for all blocks. We may miss some blocks if | |
3688 | this function is called when redirecting an edge made a block | |
3689 | unreachable, but that block is not deleted yet. */ | |
3690 | gcc_assert (n == RGN_NR_BLOCKS (rgn)); | |
3691 | } | |
3692 | ||
e855c69d | 3693 | /* Tidy the possibly empty block BB. */ |
65592aad | 3694 | static bool |
5f33b972 | 3695 | maybe_tidy_empty_bb (basic_block bb) |
e855c69d | 3696 | { |
b7b5540a | 3697 | basic_block succ_bb, pred_bb, note_bb; |
9771b263 | 3698 | vec<basic_block> dom_bbs; |
f2c45f08 AM |
3699 | edge e; |
3700 | edge_iterator ei; | |
e855c69d AB |
3701 | bool rescan_p; |
3702 | ||
3703 | /* Keep empty bb only if this block immediately precedes EXIT and | |
762bffba AB |
3704 | has incoming non-fallthrough edge, or it has no predecessors or |
3705 | successors. Otherwise remove it. */ | |
b5b8b0ac | 3706 | if (!sel_bb_empty_p (bb) |
b8698a0f | 3707 | || (single_succ_p (bb) |
fefa31b5 | 3708 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
b8698a0f | 3709 | && (!single_pred_p (bb) |
762bffba AB |
3710 | || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU))) |
3711 | || EDGE_COUNT (bb->preds) == 0 | |
3712 | || EDGE_COUNT (bb->succs) == 0) | |
e855c69d AB |
3713 | return false; |
3714 | ||
f2c45f08 AM |
3715 | /* Do not attempt to redirect complex edges. */ |
3716 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3717 | if (e->flags & EDGE_COMPLEX) | |
3718 | return false; | |
5ef5a3b7 JJ |
3719 | else if (e->flags & EDGE_FALLTHRU) |
3720 | { | |
3721 | rtx note; | |
3722 | /* If prev bb ends with asm goto, see if any of the | |
3723 | ASM_OPERANDS_LABELs don't point to the fallthru | |
3724 | label. Do not attempt to redirect it in that case. */ | |
3725 | if (JUMP_P (BB_END (e->src)) | |
3726 | && (note = extract_asm_operands (PATTERN (BB_END (e->src))))) | |
3727 | { | |
3728 | int i, n = ASM_OPERANDS_LABEL_LENGTH (note); | |
3729 | ||
3730 | for (i = 0; i < n; ++i) | |
3731 | if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (bb)) | |
3732 | return false; | |
3733 | } | |
3734 | } | |
f2c45f08 | 3735 | |
e855c69d AB |
3736 | free_data_sets (bb); |
3737 | ||
3738 | /* Do not delete BB if it has more than one successor. | |
3739 | That can occur when we moving a jump. */ | |
3740 | if (!single_succ_p (bb)) | |
3741 | { | |
3742 | gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); | |
3743 | sel_merge_blocks (bb->prev_bb, bb); | |
3744 | return true; | |
3745 | } | |
3746 | ||
3747 | succ_bb = single_succ (bb); | |
3748 | rescan_p = true; | |
3749 | pred_bb = NULL; | |
9771b263 | 3750 | dom_bbs.create (0); |
e855c69d | 3751 | |
b7b5540a AB |
3752 | /* Save a pred/succ from the current region to attach the notes to. */ |
3753 | note_bb = NULL; | |
3754 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3755 | if (in_current_region_p (e->src)) | |
3756 | { | |
3757 | note_bb = e->src; | |
3758 | break; | |
3759 | } | |
3760 | if (note_bb == NULL) | |
3761 | note_bb = succ_bb; | |
3762 | ||
e855c69d AB |
3763 | /* Redirect all non-fallthru edges to the next bb. */ |
3764 | while (rescan_p) | |
3765 | { | |
e855c69d AB |
3766 | rescan_p = false; |
3767 | ||
3768 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3769 | { | |
3770 | pred_bb = e->src; | |
3771 | ||
3772 | if (!(e->flags & EDGE_FALLTHRU)) | |
3773 | { | |
5f33b972 | 3774 | /* We can not invalidate computed topological order by moving |
00c4e97c AB |
3775 | the edge destination block (E->SUCC) along a fallthru edge. |
3776 | ||
3777 | We will update dominators here only when we'll get | |
3778 | an unreachable block when redirecting, otherwise | |
3779 | sel_redirect_edge_and_branch will take care of it. */ | |
3780 | if (e->dest != bb | |
3781 | && single_pred_p (e->dest)) | |
9771b263 | 3782 | dom_bbs.safe_push (e->dest); |
5f33b972 | 3783 | sel_redirect_edge_and_branch (e, succ_bb); |
e855c69d AB |
3784 | rescan_p = true; |
3785 | break; | |
3786 | } | |
5f33b972 AM |
3787 | /* If the edge is fallthru, but PRED_BB ends in a conditional jump |
3788 | to BB (so there is no non-fallthru edge from PRED_BB to BB), we | |
3789 | still have to adjust it. */ | |
3790 | else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb))) | |
3791 | { | |
3792 | /* If possible, try to remove the unneeded conditional jump. */ | |
3793 | if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0 | |
3794 | && !IN_CURRENT_FENCE_P (BB_END (pred_bb))) | |
3795 | { | |
3796 | if (!sel_remove_insn (BB_END (pred_bb), false, false)) | |
3797 | tidy_fallthru_edge (e); | |
3798 | } | |
3799 | else | |
3800 | sel_redirect_edge_and_branch (e, succ_bb); | |
3801 | rescan_p = true; | |
3802 | break; | |
3803 | } | |
e855c69d AB |
3804 | } |
3805 | } | |
3806 | ||
e855c69d AB |
3807 | if (can_merge_blocks_p (bb->prev_bb, bb)) |
3808 | sel_merge_blocks (bb->prev_bb, bb); | |
3809 | else | |
e855c69d | 3810 | { |
262d8232 | 3811 | /* This is a block without fallthru predecessor. Just delete it. */ |
b7b5540a AB |
3812 | gcc_assert (note_bb); |
3813 | move_bb_info (note_bb, bb); | |
e855c69d AB |
3814 | remove_empty_bb (bb, true); |
3815 | } | |
3816 | ||
9771b263 | 3817 | if (!dom_bbs.is_empty ()) |
00c4e97c | 3818 | { |
9771b263 | 3819 | dom_bbs.safe_push (succ_bb); |
00c4e97c | 3820 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); |
9771b263 | 3821 | dom_bbs.release (); |
00c4e97c AB |
3822 | } |
3823 | ||
e855c69d AB |
3824 | return true; |
3825 | } | |
3826 | ||
b8698a0f | 3827 | /* Tidy the control flow after we have removed original insn from |
e855c69d AB |
3828 | XBB. Return true if we have removed some blocks. When FULL_TIDYING |
3829 | is true, also try to optimize control flow on non-empty blocks. */ | |
3830 | bool | |
3831 | tidy_control_flow (basic_block xbb, bool full_tidying) | |
3832 | { | |
3833 | bool changed = true; | |
b5b8b0ac | 3834 | insn_t first, last; |
b8698a0f | 3835 | |
e855c69d | 3836 | /* First check whether XBB is empty. */ |
5f33b972 | 3837 | changed = maybe_tidy_empty_bb (xbb); |
e855c69d AB |
3838 | if (changed || !full_tidying) |
3839 | return changed; | |
b8698a0f | 3840 | |
e855c69d | 3841 | /* Check if there is a unnecessary jump after insn left. */ |
753de8cf | 3842 | if (bb_has_removable_jump_to_p (xbb, xbb->next_bb) |
e855c69d AB |
3843 | && INSN_SCHED_TIMES (BB_END (xbb)) == 0 |
3844 | && !IN_CURRENT_FENCE_P (BB_END (xbb))) | |
3845 | { | |
ab6dceab AB |
3846 | /* We used to call sel_remove_insn here that can trigger tidy_control_flow |
3847 | before we fix up the fallthru edge. Correct that ordering by | |
3848 | explicitly doing the latter before the former. */ | |
3849 | clear_expr (INSN_EXPR (BB_END (xbb))); | |
e855c69d | 3850 | tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); |
ab6dceab AB |
3851 | if (tidy_control_flow (xbb, false)) |
3852 | return true; | |
e855c69d AB |
3853 | } |
3854 | ||
b5b8b0ac AO |
3855 | first = sel_bb_head (xbb); |
3856 | last = sel_bb_end (xbb); | |
3857 | if (MAY_HAVE_DEBUG_INSNS) | |
3858 | { | |
3859 | if (first != last && DEBUG_INSN_P (first)) | |
3860 | do | |
3861 | first = NEXT_INSN (first); | |
3862 | while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); | |
3863 | ||
3864 | if (first != last && DEBUG_INSN_P (last)) | |
3865 | do | |
3866 | last = PREV_INSN (last); | |
3867 | while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); | |
3868 | } | |
e855c69d | 3869 | /* Check if there is an unnecessary jump in previous basic block leading |
b8698a0f L |
3870 | to next basic block left after removing INSN from stream. |
3871 | If it is so, remove that jump and redirect edge to current | |
3872 | basic block (where there was INSN before deletion). This way | |
3873 | when NOP will be deleted several instructions later with its | |
3874 | basic block we will not get a jump to next instruction, which | |
e855c69d | 3875 | can be harmful. */ |
b5b8b0ac | 3876 | if (first == last |
e855c69d | 3877 | && !sel_bb_empty_p (xbb) |
b5b8b0ac | 3878 | && INSN_NOP_P (last) |
e855c69d AB |
3879 | /* Flow goes fallthru from current block to the next. */ |
3880 | && EDGE_COUNT (xbb->succs) == 1 | |
3881 | && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) | |
3882 | /* When successor is an EXIT block, it may not be the next block. */ | |
fefa31b5 | 3883 | && single_succ (xbb) != EXIT_BLOCK_PTR_FOR_FN (cfun) |
e855c69d AB |
3884 | /* And unconditional jump in previous basic block leads to |
3885 | next basic block of XBB and this jump can be safely removed. */ | |
3886 | && in_current_region_p (xbb->prev_bb) | |
753de8cf | 3887 | && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb) |
e855c69d AB |
3888 | && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 |
3889 | /* Also this jump is not at the scheduling boundary. */ | |
3890 | && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) | |
3891 | { | |
b59ab570 | 3892 | bool recompute_toporder_p; |
e855c69d AB |
3893 | /* Clear data structures of jump - jump itself will be removed |
3894 | by sel_redirect_edge_and_branch. */ | |
3895 | clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); | |
b59ab570 AM |
3896 | recompute_toporder_p |
3897 | = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); | |
3898 | ||
e855c69d AB |
3899 | gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); |
3900 | ||
3901 | /* It can turn out that after removing unused jump, basic block | |
3902 | that contained that jump, becomes empty too. In such case | |
3903 | remove it too. */ | |
3904 | if (sel_bb_empty_p (xbb->prev_bb)) | |
5f33b972 AM |
3905 | changed = maybe_tidy_empty_bb (xbb->prev_bb); |
3906 | if (recompute_toporder_p) | |
b59ab570 | 3907 | sel_recompute_toporder (); |
e855c69d | 3908 | } |
d787f788 | 3909 | |
b2b29377 MM |
3910 | /* TODO: use separate flag for CFG checking. */ |
3911 | if (flag_checking) | |
3912 | { | |
3913 | verify_backedges (); | |
3914 | verify_dominators (CDI_DOMINATORS); | |
3915 | } | |
d787f788 | 3916 | |
e855c69d AB |
3917 | return changed; |
3918 | } | |
3919 | ||
b59ab570 AM |
3920 | /* Purge meaningless empty blocks in the middle of a region. */ |
3921 | void | |
3922 | purge_empty_blocks (void) | |
3923 | { | |
9d0dcda1 | 3924 | int i; |
b59ab570 | 3925 | |
9d0dcda1 AM |
3926 | /* Do not attempt to delete the first basic block in the region. */ |
3927 | for (i = 1; i < current_nr_blocks; ) | |
b59ab570 | 3928 | { |
06e28de2 | 3929 | basic_block b = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
b59ab570 | 3930 | |
5f33b972 | 3931 | if (maybe_tidy_empty_bb (b)) |
b59ab570 AM |
3932 | continue; |
3933 | ||
3934 | i++; | |
3935 | } | |
3936 | } | |
3937 | ||
b8698a0f L |
3938 | /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, |
3939 | do not delete insn's data, because it will be later re-emitted. | |
e855c69d AB |
3940 | Return true if we have removed some blocks afterwards. */ |
3941 | bool | |
3942 | sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) | |
3943 | { | |
3944 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3945 | ||
3946 | gcc_assert (INSN_IN_STREAM_P (insn)); | |
3947 | ||
b5b8b0ac AO |
3948 | if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) |
3949 | { | |
3950 | expr_t expr; | |
3951 | av_set_iterator i; | |
3952 | ||
3953 | /* When we remove a debug insn that is head of a BB, it remains | |
3954 | in the AV_SET of the block, but it shouldn't. */ | |
3955 | FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) | |
3956 | if (EXPR_INSN_RTX (expr) == insn) | |
3957 | { | |
3958 | av_set_iter_remove (&i); | |
3959 | break; | |
3960 | } | |
3961 | } | |
3962 | ||
e855c69d | 3963 | if (only_disconnect) |
1f397f45 | 3964 | remove_insn (insn); |
e855c69d AB |
3965 | else |
3966 | { | |
1f397f45 | 3967 | delete_insn (insn); |
e855c69d AB |
3968 | clear_expr (INSN_EXPR (insn)); |
3969 | } | |
3970 | ||
1f397f45 SB |
3971 | /* It is necessary to NULL these fields in case we are going to re-insert |
3972 | INSN into the insns stream, as will usually happen in the ONLY_DISCONNECT | |
3973 | case, but also for NOPs that we will return to the nop pool. */ | |
0f82e5c9 DM |
3974 | SET_PREV_INSN (insn) = NULL_RTX; |
3975 | SET_NEXT_INSN (insn) = NULL_RTX; | |
1f397f45 | 3976 | set_block_for_insn (insn, NULL); |
e855c69d AB |
3977 | |
3978 | return tidy_control_flow (bb, full_tidying); | |
3979 | } | |
3980 | ||
3981 | /* Estimate number of the insns in BB. */ | |
3982 | static int | |
3983 | sel_estimate_number_of_insns (basic_block bb) | |
3984 | { | |
3985 | int res = 0; | |
3986 | insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); | |
3987 | ||
3988 | for (; insn != next_tail; insn = NEXT_INSN (insn)) | |
b5b8b0ac | 3989 | if (NONDEBUG_INSN_P (insn)) |
e855c69d AB |
3990 | res++; |
3991 | ||
3992 | return res; | |
3993 | } | |
3994 | ||
3995 | /* We don't need separate luids for notes or labels. */ | |
3996 | static int | |
3997 | sel_luid_for_non_insn (rtx x) | |
3998 | { | |
3999 | gcc_assert (NOTE_P (x) || LABEL_P (x)); | |
4000 | ||
4001 | return -1; | |
4002 | } | |
4003 | ||
0d9439b0 SG |
4004 | /* Find the proper seqno for inserting at INSN by successors. |
4005 | Return -1 if no successors with positive seqno exist. */ | |
e855c69d | 4006 | static int |
6144a836 | 4007 | get_seqno_by_succs (rtx_insn *insn) |
0d9439b0 SG |
4008 | { |
4009 | basic_block bb = BLOCK_FOR_INSN (insn); | |
6144a836 | 4010 | rtx_insn *tmp = insn, *end = BB_END (bb); |
0d9439b0 SG |
4011 | int seqno; |
4012 | insn_t succ = NULL; | |
4013 | succ_iterator si; | |
4014 | ||
4015 | while (tmp != end) | |
4016 | { | |
4017 | tmp = NEXT_INSN (tmp); | |
4018 | if (INSN_P (tmp)) | |
4019 | return INSN_SEQNO (tmp); | |
4020 | } | |
4021 | ||
4022 | seqno = INT_MAX; | |
4023 | ||
4024 | FOR_EACH_SUCC_1 (succ, si, end, SUCCS_NORMAL) | |
4025 | if (INSN_SEQNO (succ) > 0) | |
4026 | seqno = MIN (seqno, INSN_SEQNO (succ)); | |
4027 | ||
4028 | if (seqno == INT_MAX) | |
4029 | return -1; | |
4030 | ||
4031 | return seqno; | |
4032 | } | |
4033 | ||
92e265ac AB |
4034 | /* Compute seqno for INSN by its preds or succs. Use OLD_SEQNO to compute |
4035 | seqno in corner cases. */ | |
0d9439b0 | 4036 | static int |
92e265ac | 4037 | get_seqno_for_a_jump (insn_t insn, int old_seqno) |
e855c69d AB |
4038 | { |
4039 | int seqno; | |
4040 | ||
4041 | gcc_assert (INSN_SIMPLEJUMP_P (insn)); | |
4042 | ||
4043 | if (!sel_bb_head_p (insn)) | |
4044 | seqno = INSN_SEQNO (PREV_INSN (insn)); | |
4045 | else | |
4046 | { | |
4047 | basic_block bb = BLOCK_FOR_INSN (insn); | |
4048 | ||
4049 | if (single_pred_p (bb) | |
4050 | && !in_current_region_p (single_pred (bb))) | |
4051 | { | |
4052 | /* We can have preds outside a region when splitting edges | |
b8698a0f | 4053 | for pipelining of an outer loop. Use succ instead. |
e855c69d AB |
4054 | There should be only one of them. */ |
4055 | insn_t succ = NULL; | |
4056 | succ_iterator si; | |
4057 | bool first = true; | |
b8698a0f | 4058 | |
e855c69d AB |
4059 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
4060 | && current_loop_nest); | |
b8698a0f | 4061 | FOR_EACH_SUCC_1 (succ, si, insn, |
e855c69d AB |
4062 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
4063 | { | |
4064 | gcc_assert (first); | |
4065 | first = false; | |
4066 | } | |
4067 | ||
4068 | gcc_assert (succ != NULL); | |
4069 | seqno = INSN_SEQNO (succ); | |
4070 | } | |
4071 | else | |
4072 | { | |
4073 | insn_t *preds; | |
4074 | int n; | |
4075 | ||
4076 | cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); | |
e855c69d | 4077 | |
0d9439b0 SG |
4078 | gcc_assert (n > 0); |
4079 | /* For one predecessor, use simple method. */ | |
4080 | if (n == 1) | |
4081 | seqno = INSN_SEQNO (preds[0]); | |
4082 | else | |
4083 | seqno = get_seqno_by_preds (insn); | |
b8698a0f | 4084 | |
e855c69d AB |
4085 | free (preds); |
4086 | } | |
4087 | } | |
4088 | ||
0d9439b0 SG |
4089 | /* We were unable to find a good seqno among preds. */ |
4090 | if (seqno < 0) | |
4091 | seqno = get_seqno_by_succs (insn); | |
4092 | ||
92e265ac AB |
4093 | if (seqno < 0) |
4094 | { | |
4095 | /* The only case where this could be here legally is that the only | |
4096 | unscheduled insn was a conditional jump that got removed and turned | |
4097 | into this unconditional one. Initialize from the old seqno | |
4098 | of that jump passed down to here. */ | |
4099 | seqno = old_seqno; | |
4100 | } | |
0d9439b0 | 4101 | |
92e265ac | 4102 | gcc_assert (seqno >= 0); |
e855c69d AB |
4103 | return seqno; |
4104 | } | |
4105 | ||
da7ba240 AB |
4106 | /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors |
4107 | with positive seqno exist. */ | |
e855c69d | 4108 | int |
b32d5189 | 4109 | get_seqno_by_preds (rtx_insn *insn) |
e855c69d AB |
4110 | { |
4111 | basic_block bb = BLOCK_FOR_INSN (insn); | |
b32d5189 | 4112 | rtx_insn *tmp = insn, *head = BB_HEAD (bb); |
e855c69d AB |
4113 | insn_t *preds; |
4114 | int n, i, seqno; | |
4115 | ||
6598bb55 AB |
4116 | /* Loop backwards from INSN to HEAD including both. */ |
4117 | while (1) | |
0d9439b0 | 4118 | { |
0d9439b0 | 4119 | if (INSN_P (tmp)) |
6598bb55 AB |
4120 | return INSN_SEQNO (tmp); |
4121 | if (tmp == head) | |
4122 | break; | |
4123 | tmp = PREV_INSN (tmp); | |
0d9439b0 | 4124 | } |
b8698a0f | 4125 | |
e855c69d AB |
4126 | cfg_preds (bb, &preds, &n); |
4127 | for (i = 0, seqno = -1; i < n; i++) | |
4128 | seqno = MAX (seqno, INSN_SEQNO (preds[i])); | |
4129 | ||
e855c69d AB |
4130 | return seqno; |
4131 | } | |
4132 | ||
4133 | \f | |
4134 | ||
4135 | /* Extend pass-scope data structures for basic blocks. */ | |
4136 | void | |
4137 | sel_extend_global_bb_info (void) | |
4138 | { | |
8b1c6fd7 | 4139 | sel_global_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun)); |
e855c69d AB |
4140 | } |
4141 | ||
4142 | /* Extend region-scope data structures for basic blocks. */ | |
4143 | static void | |
4144 | extend_region_bb_info (void) | |
4145 | { | |
8b1c6fd7 | 4146 | sel_region_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun)); |
e855c69d AB |
4147 | } |
4148 | ||
4149 | /* Extend all data structures to fit for all basic blocks. */ | |
4150 | static void | |
4151 | extend_bb_info (void) | |
4152 | { | |
4153 | sel_extend_global_bb_info (); | |
4154 | extend_region_bb_info (); | |
4155 | } | |
4156 | ||
4157 | /* Finalize pass-scope data structures for basic blocks. */ | |
4158 | void | |
4159 | sel_finish_global_bb_info (void) | |
4160 | { | |
9771b263 | 4161 | sel_global_bb_info.release (); |
e855c69d AB |
4162 | } |
4163 | ||
4164 | /* Finalize region-scope data structures for basic blocks. */ | |
4165 | static void | |
4166 | finish_region_bb_info (void) | |
4167 | { | |
9771b263 | 4168 | sel_region_bb_info.release (); |
e855c69d AB |
4169 | } |
4170 | \f | |
4171 | ||
4172 | /* Data for each insn in current region. */ | |
7de76362 | 4173 | vec<sel_insn_data_def> s_i_d; |
e855c69d | 4174 | |
e855c69d AB |
4175 | /* Extend data structures for insns from current region. */ |
4176 | static void | |
4177 | extend_insn_data (void) | |
4178 | { | |
4179 | int reserve; | |
b8698a0f | 4180 | |
e855c69d AB |
4181 | sched_extend_target (); |
4182 | sched_deps_init (false); | |
4183 | ||
4184 | /* Extend data structures for insns from current region. */ | |
9771b263 DN |
4185 | reserve = (sched_max_luid + 1 - s_i_d.length ()); |
4186 | if (reserve > 0 && ! s_i_d.space (reserve)) | |
bcf33775 AB |
4187 | { |
4188 | int size; | |
4189 | ||
4190 | if (sched_max_luid / 2 > 1024) | |
4191 | size = sched_max_luid + 1024; | |
4192 | else | |
4193 | size = 3 * sched_max_luid / 2; | |
b8698a0f | 4194 | |
bcf33775 | 4195 | |
9771b263 | 4196 | s_i_d.safe_grow_cleared (size); |
bcf33775 | 4197 | } |
e855c69d AB |
4198 | } |
4199 | ||
4200 | /* Finalize data structures for insns from current region. */ | |
4201 | static void | |
4202 | finish_insns (void) | |
4203 | { | |
4204 | unsigned i; | |
4205 | ||
4206 | /* Clear here all dependence contexts that may have left from insns that were | |
4207 | removed during the scheduling. */ | |
9771b263 | 4208 | for (i = 0; i < s_i_d.length (); i++) |
e855c69d | 4209 | { |
9771b263 | 4210 | sel_insn_data_def *sid_entry = &s_i_d[i]; |
b8698a0f | 4211 | |
e855c69d AB |
4212 | if (sid_entry->live) |
4213 | return_regset_to_pool (sid_entry->live); | |
4214 | if (sid_entry->analyzed_deps) | |
4215 | { | |
4216 | BITMAP_FREE (sid_entry->analyzed_deps); | |
4217 | BITMAP_FREE (sid_entry->found_deps); | |
4218 | htab_delete (sid_entry->transformed_insns); | |
4219 | free_deps (&sid_entry->deps_context); | |
4220 | } | |
4221 | if (EXPR_VINSN (&sid_entry->expr)) | |
4222 | { | |
4223 | clear_expr (&sid_entry->expr); | |
b8698a0f | 4224 | |
e855c69d AB |
4225 | /* Also, clear CANT_MOVE bit here, because we really don't want it |
4226 | to be passed to the next region. */ | |
4227 | CANT_MOVE_BY_LUID (i) = 0; | |
4228 | } | |
4229 | } | |
b8698a0f | 4230 | |
9771b263 | 4231 | s_i_d.release (); |
e855c69d AB |
4232 | } |
4233 | ||
4234 | /* A proxy to pass initialization data to init_insn (). */ | |
4235 | static sel_insn_data_def _insn_init_ssid; | |
4236 | static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; | |
4237 | ||
4238 | /* If true create a new vinsn. Otherwise use the one from EXPR. */ | |
4239 | static bool insn_init_create_new_vinsn_p; | |
4240 | ||
4241 | /* Set all necessary data for initialization of the new insn[s]. */ | |
4242 | static expr_t | |
4243 | set_insn_init (expr_t expr, vinsn_t vi, int seqno) | |
4244 | { | |
4245 | expr_t x = &insn_init_ssid->expr; | |
4246 | ||
4247 | copy_expr_onside (x, expr); | |
4248 | if (vi != NULL) | |
4249 | { | |
4250 | insn_init_create_new_vinsn_p = false; | |
4251 | change_vinsn_in_expr (x, vi); | |
4252 | } | |
4253 | else | |
4254 | insn_init_create_new_vinsn_p = true; | |
4255 | ||
4256 | insn_init_ssid->seqno = seqno; | |
4257 | return x; | |
4258 | } | |
4259 | ||
4260 | /* Init data for INSN. */ | |
4261 | static void | |
4262 | init_insn_data (insn_t insn) | |
4263 | { | |
4264 | expr_t expr; | |
4265 | sel_insn_data_t ssid = insn_init_ssid; | |
4266 | ||
4267 | /* The fields mentioned below are special and hence are not being | |
4268 | propagated to the new insns. */ | |
4269 | gcc_assert (!ssid->asm_p && ssid->sched_next == NULL | |
4270 | && !ssid->after_stall_p && ssid->sched_cycle == 0); | |
4271 | gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); | |
4272 | ||
4273 | expr = INSN_EXPR (insn); | |
4274 | copy_expr (expr, &ssid->expr); | |
4275 | prepare_insn_expr (insn, ssid->seqno); | |
4276 | ||
4277 | if (insn_init_create_new_vinsn_p) | |
4278 | change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p)); | |
b8698a0f | 4279 | |
e855c69d AB |
4280 | if (first_time_insn_init (insn)) |
4281 | init_first_time_insn_data (insn); | |
4282 | } | |
4283 | ||
4284 | /* This is used to initialize spurious jumps generated by | |
92e265ac AB |
4285 | sel_redirect_edge (). OLD_SEQNO is used for initializing seqnos |
4286 | in corner cases within get_seqno_for_a_jump. */ | |
e855c69d | 4287 | static void |
92e265ac | 4288 | init_simplejump_data (insn_t insn, int old_seqno) |
e855c69d AB |
4289 | { |
4290 | init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0, | |
9771b263 | 4291 | REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0, |
6e1aa848 | 4292 | vNULL, true, false, false, |
e855c69d | 4293 | false, true); |
92e265ac | 4294 | INSN_SEQNO (insn) = get_seqno_for_a_jump (insn, old_seqno); |
e855c69d AB |
4295 | init_first_time_insn_data (insn); |
4296 | } | |
4297 | ||
b8698a0f | 4298 | /* Perform deferred initialization of insns. This is used to process |
92e265ac AB |
4299 | a new jump that may be created by redirect_edge. OLD_SEQNO is used |
4300 | for initializing simplejumps in init_simplejump_data. */ | |
4301 | static void | |
4302 | sel_init_new_insn (insn_t insn, int flags, int old_seqno) | |
e855c69d AB |
4303 | { |
4304 | /* We create data structures for bb when the first insn is emitted in it. */ | |
4305 | if (INSN_P (insn) | |
4306 | && INSN_IN_STREAM_P (insn) | |
4307 | && insn_is_the_only_one_in_bb_p (insn)) | |
4308 | { | |
4309 | extend_bb_info (); | |
4310 | create_initial_data_sets (BLOCK_FOR_INSN (insn)); | |
4311 | } | |
b8698a0f | 4312 | |
e855c69d | 4313 | if (flags & INSN_INIT_TODO_LUID) |
a95b23b4 BS |
4314 | { |
4315 | sched_extend_luids (); | |
4316 | sched_init_insn_luid (insn); | |
4317 | } | |
e855c69d AB |
4318 | |
4319 | if (flags & INSN_INIT_TODO_SSID) | |
4320 | { | |
4321 | extend_insn_data (); | |
4322 | init_insn_data (insn); | |
4323 | clear_expr (&insn_init_ssid->expr); | |
4324 | } | |
4325 | ||
4326 | if (flags & INSN_INIT_TODO_SIMPLEJUMP) | |
4327 | { | |
4328 | extend_insn_data (); | |
92e265ac | 4329 | init_simplejump_data (insn, old_seqno); |
e855c69d | 4330 | } |
b8698a0f | 4331 | |
e855c69d AB |
4332 | gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) |
4333 | == CONTAINING_RGN (BB_TO_BLOCK (0))); | |
4334 | } | |
4335 | \f | |
4336 | ||
4337 | /* Functions to init/finish work with lv sets. */ | |
4338 | ||
4339 | /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ | |
4340 | static void | |
4341 | init_lv_set (basic_block bb) | |
4342 | { | |
4343 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
4344 | ||
4345 | BB_LV_SET (bb) = get_regset_from_pool (); | |
b8698a0f | 4346 | COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); |
e855c69d AB |
4347 | BB_LV_SET_VALID_P (bb) = true; |
4348 | } | |
4349 | ||
4350 | /* Copy liveness information to BB from FROM_BB. */ | |
4351 | static void | |
4352 | copy_lv_set_from (basic_block bb, basic_block from_bb) | |
4353 | { | |
4354 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
b8698a0f | 4355 | |
e855c69d AB |
4356 | COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); |
4357 | BB_LV_SET_VALID_P (bb) = true; | |
b8698a0f | 4358 | } |
e855c69d AB |
4359 | |
4360 | /* Initialize lv set of all bb headers. */ | |
4361 | void | |
4362 | init_lv_sets (void) | |
4363 | { | |
4364 | basic_block bb; | |
4365 | ||
4366 | /* Initialize of LV sets. */ | |
11cd3bed | 4367 | FOR_EACH_BB_FN (bb, cfun) |
e855c69d AB |
4368 | init_lv_set (bb); |
4369 | ||
4370 | /* Don't forget EXIT_BLOCK. */ | |
fefa31b5 | 4371 | init_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e855c69d AB |
4372 | } |
4373 | ||
4374 | /* Release lv set of HEAD. */ | |
4375 | static void | |
4376 | free_lv_set (basic_block bb) | |
4377 | { | |
4378 | gcc_assert (BB_LV_SET (bb) != NULL); | |
4379 | ||
4380 | return_regset_to_pool (BB_LV_SET (bb)); | |
4381 | BB_LV_SET (bb) = NULL; | |
4382 | BB_LV_SET_VALID_P (bb) = false; | |
4383 | } | |
4384 | ||
4385 | /* Finalize lv sets of all bb headers. */ | |
4386 | void | |
4387 | free_lv_sets (void) | |
4388 | { | |
4389 | basic_block bb; | |
4390 | ||
4391 | /* Don't forget EXIT_BLOCK. */ | |
fefa31b5 | 4392 | free_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e855c69d AB |
4393 | |
4394 | /* Free LV sets. */ | |
11cd3bed | 4395 | FOR_EACH_BB_FN (bb, cfun) |
e855c69d AB |
4396 | if (BB_LV_SET (bb)) |
4397 | free_lv_set (bb); | |
4398 | } | |
4399 | ||
5c416724 DM |
4400 | /* Mark AV_SET for BB as invalid, so this set will be updated the next time |
4401 | compute_av() processes BB. This function is called when creating new basic | |
4402 | blocks, as well as for blocks (either new or existing) where new jumps are | |
4403 | created when the control flow is being updated. */ | |
e855c69d AB |
4404 | static void |
4405 | invalidate_av_set (basic_block bb) | |
4406 | { | |
e855c69d AB |
4407 | BB_AV_LEVEL (bb) = -1; |
4408 | } | |
4409 | ||
4410 | /* Create initial data sets for BB (they will be invalid). */ | |
4411 | static void | |
4412 | create_initial_data_sets (basic_block bb) | |
4413 | { | |
4414 | if (BB_LV_SET (bb)) | |
4415 | BB_LV_SET_VALID_P (bb) = false; | |
4416 | else | |
4417 | BB_LV_SET (bb) = get_regset_from_pool (); | |
4418 | invalidate_av_set (bb); | |
4419 | } | |
4420 | ||
4421 | /* Free av set of BB. */ | |
4422 | static void | |
4423 | free_av_set (basic_block bb) | |
4424 | { | |
4425 | av_set_clear (&BB_AV_SET (bb)); | |
4426 | BB_AV_LEVEL (bb) = 0; | |
4427 | } | |
4428 | ||
4429 | /* Free data sets of BB. */ | |
4430 | void | |
4431 | free_data_sets (basic_block bb) | |
4432 | { | |
4433 | free_lv_set (bb); | |
4434 | free_av_set (bb); | |
4435 | } | |
4436 | ||
e855c69d AB |
4437 | /* Exchange data sets of TO and FROM. */ |
4438 | void | |
4439 | exchange_data_sets (basic_block to, basic_block from) | |
4440 | { | |
6b4db501 MM |
4441 | /* Exchange lv sets of TO and FROM. */ |
4442 | std::swap (BB_LV_SET (from), BB_LV_SET (to)); | |
4443 | std::swap (BB_LV_SET_VALID_P (from), BB_LV_SET_VALID_P (to)); | |
4444 | ||
4445 | /* Exchange av sets of TO and FROM. */ | |
4446 | std::swap (BB_AV_SET (from), BB_AV_SET (to)); | |
4447 | std::swap (BB_AV_LEVEL (from), BB_AV_LEVEL (to)); | |
e855c69d AB |
4448 | } |
4449 | ||
4450 | /* Copy data sets of FROM to TO. */ | |
4451 | void | |
4452 | copy_data_sets (basic_block to, basic_block from) | |
4453 | { | |
4454 | gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); | |
4455 | gcc_assert (BB_AV_SET (to) == NULL); | |
4456 | ||
4457 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4458 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4459 | ||
4460 | if (BB_AV_SET_VALID_P (from)) | |
4461 | { | |
4462 | BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); | |
4463 | } | |
4464 | if (BB_LV_SET_VALID_P (from)) | |
4465 | { | |
4466 | gcc_assert (BB_LV_SET (to) != NULL); | |
4467 | COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); | |
4468 | } | |
4469 | } | |
4470 | ||
4471 | /* Return an av set for INSN, if any. */ | |
4472 | av_set_t | |
4473 | get_av_set (insn_t insn) | |
4474 | { | |
4475 | av_set_t av_set; | |
4476 | ||
4477 | gcc_assert (AV_SET_VALID_P (insn)); | |
4478 | ||
4479 | if (sel_bb_head_p (insn)) | |
4480 | av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); | |
4481 | else | |
4482 | av_set = NULL; | |
4483 | ||
4484 | return av_set; | |
4485 | } | |
4486 | ||
4487 | /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ | |
4488 | int | |
4489 | get_av_level (insn_t insn) | |
4490 | { | |
4491 | int av_level; | |
4492 | ||
4493 | gcc_assert (INSN_P (insn)); | |
4494 | ||
4495 | if (sel_bb_head_p (insn)) | |
4496 | av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); | |
4497 | else | |
4498 | av_level = INSN_WS_LEVEL (insn); | |
4499 | ||
4500 | return av_level; | |
4501 | } | |
4502 | ||
4503 | \f | |
4504 | ||
4505 | /* Variables to work with control-flow graph. */ | |
4506 | ||
4507 | /* The basic block that already has been processed by the sched_data_update (), | |
4508 | but hasn't been in sel_add_bb () yet. */ | |
7de76362 | 4509 | static vec<basic_block> last_added_blocks; |
e855c69d AB |
4510 | |
4511 | /* A pool for allocating successor infos. */ | |
4512 | static struct | |
4513 | { | |
4514 | /* A stack for saving succs_info structures. */ | |
4515 | struct succs_info *stack; | |
4516 | ||
4517 | /* Its size. */ | |
4518 | int size; | |
4519 | ||
4520 | /* Top of the stack. */ | |
4521 | int top; | |
4522 | ||
4523 | /* Maximal value of the top. */ | |
4524 | int max_top; | |
4525 | } succs_info_pool; | |
4526 | ||
4527 | /* Functions to work with control-flow graph. */ | |
4528 | ||
4529 | /* Return basic block note of BB. */ | |
c5db5458 | 4530 | rtx_insn * |
e855c69d AB |
4531 | sel_bb_head (basic_block bb) |
4532 | { | |
c5db5458 | 4533 | rtx_insn *head; |
e855c69d | 4534 | |
fefa31b5 | 4535 | if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
e855c69d AB |
4536 | { |
4537 | gcc_assert (exit_insn != NULL_RTX); | |
4538 | head = exit_insn; | |
4539 | } | |
4540 | else | |
4541 | { | |
e67d1102 | 4542 | rtx_note *note = bb_note (bb); |
e855c69d AB |
4543 | head = next_nonnote_insn (note); |
4544 | ||
89ad0f25 | 4545 | if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb)) |
c5db5458 | 4546 | head = NULL; |
e855c69d AB |
4547 | } |
4548 | ||
4549 | return head; | |
4550 | } | |
4551 | ||
4552 | /* Return true if INSN is a basic block header. */ | |
4553 | bool | |
4554 | sel_bb_head_p (insn_t insn) | |
4555 | { | |
4556 | return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn; | |
4557 | } | |
4558 | ||
4559 | /* Return last insn of BB. */ | |
c5db5458 | 4560 | rtx_insn * |
e855c69d AB |
4561 | sel_bb_end (basic_block bb) |
4562 | { | |
4563 | if (sel_bb_empty_p (bb)) | |
c5db5458 | 4564 | return NULL; |
e855c69d | 4565 | |
fefa31b5 | 4566 | gcc_assert (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e855c69d AB |
4567 | |
4568 | return BB_END (bb); | |
4569 | } | |
4570 | ||
4571 | /* Return true if INSN is the last insn in its basic block. */ | |
4572 | bool | |
4573 | sel_bb_end_p (insn_t insn) | |
4574 | { | |
4575 | return insn == sel_bb_end (BLOCK_FOR_INSN (insn)); | |
4576 | } | |
4577 | ||
4578 | /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ | |
4579 | bool | |
4580 | sel_bb_empty_p (basic_block bb) | |
4581 | { | |
4582 | return sel_bb_head (bb) == NULL; | |
4583 | } | |
4584 | ||
4585 | /* True when BB belongs to the current scheduling region. */ | |
4586 | bool | |
4587 | in_current_region_p (basic_block bb) | |
4588 | { | |
4589 | if (bb->index < NUM_FIXED_BLOCKS) | |
4590 | return false; | |
4591 | ||
4592 | return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4593 | } | |
4594 | ||
4595 | /* Return the block which is a fallthru bb of a conditional jump JUMP. */ | |
4596 | basic_block | |
68a1a6c0 | 4597 | fallthru_bb_of_jump (const rtx_insn *jump) |
e855c69d AB |
4598 | { |
4599 | if (!JUMP_P (jump)) | |
4600 | return NULL; | |
4601 | ||
e855c69d AB |
4602 | if (!any_condjump_p (jump)) |
4603 | return NULL; | |
4604 | ||
268bab85 AB |
4605 | /* A basic block that ends with a conditional jump may still have one successor |
4606 | (and be followed by a barrier), we are not interested. */ | |
4607 | if (single_succ_p (BLOCK_FOR_INSN (jump))) | |
4608 | return NULL; | |
4609 | ||
e855c69d AB |
4610 | return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; |
4611 | } | |
4612 | ||
4613 | /* Remove all notes from BB. */ | |
4614 | static void | |
4615 | init_bb (basic_block bb) | |
4616 | { | |
4617 | remove_notes (bb_note (bb), BB_END (bb)); | |
b311fd0f | 4618 | BB_NOTE_LIST (bb) = note_list; |
e855c69d AB |
4619 | } |
4620 | ||
4621 | void | |
a95b23b4 | 4622 | sel_init_bbs (bb_vec_t bbs) |
e855c69d AB |
4623 | { |
4624 | const struct sched_scan_info_def ssi = | |
4625 | { | |
4626 | extend_bb_info, /* extend_bb */ | |
4627 | init_bb, /* init_bb */ | |
4628 | NULL, /* extend_insn */ | |
4629 | NULL /* init_insn */ | |
4630 | }; | |
4631 | ||
a95b23b4 | 4632 | sched_scan (&ssi, bbs); |
e855c69d AB |
4633 | } |
4634 | ||
7898b93b | 4635 | /* Restore notes for the whole region. */ |
e855c69d | 4636 | static void |
7898b93b | 4637 | sel_restore_notes (void) |
e855c69d AB |
4638 | { |
4639 | int bb; | |
7898b93b | 4640 | insn_t insn; |
e855c69d AB |
4641 | |
4642 | for (bb = 0; bb < current_nr_blocks; bb++) | |
4643 | { | |
4644 | basic_block first, last; | |
4645 | ||
4646 | first = EBB_FIRST_BB (bb); | |
4647 | last = EBB_LAST_BB (bb)->next_bb; | |
4648 | ||
4649 | do | |
4650 | { | |
4651 | note_list = BB_NOTE_LIST (first); | |
4652 | restore_other_notes (NULL, first); | |
b311fd0f | 4653 | BB_NOTE_LIST (first) = NULL; |
e855c69d | 4654 | |
7898b93b AM |
4655 | FOR_BB_INSNS (first, insn) |
4656 | if (NONDEBUG_INSN_P (insn)) | |
4657 | reemit_notes (insn); | |
4658 | ||
e855c69d AB |
4659 | first = first->next_bb; |
4660 | } | |
4661 | while (first != last); | |
4662 | } | |
4663 | } | |
4664 | ||
4665 | /* Free per-bb data structures. */ | |
4666 | void | |
4667 | sel_finish_bbs (void) | |
4668 | { | |
7898b93b | 4669 | sel_restore_notes (); |
e855c69d AB |
4670 | |
4671 | /* Remove current loop preheader from this loop. */ | |
4672 | if (current_loop_nest) | |
4673 | sel_remove_loop_preheader (); | |
4674 | ||
4675 | finish_region_bb_info (); | |
4676 | } | |
4677 | ||
4678 | /* Return true if INSN has a single successor of type FLAGS. */ | |
4679 | bool | |
4680 | sel_insn_has_single_succ_p (insn_t insn, int flags) | |
4681 | { | |
4682 | insn_t succ; | |
4683 | succ_iterator si; | |
4684 | bool first_p = true; | |
4685 | ||
4686 | FOR_EACH_SUCC_1 (succ, si, insn, flags) | |
4687 | { | |
4688 | if (first_p) | |
4689 | first_p = false; | |
4690 | else | |
4691 | return false; | |
4692 | } | |
4693 | ||
4694 | return true; | |
4695 | } | |
4696 | ||
4697 | /* Allocate successor's info. */ | |
4698 | static struct succs_info * | |
4699 | alloc_succs_info (void) | |
4700 | { | |
4701 | if (succs_info_pool.top == succs_info_pool.max_top) | |
4702 | { | |
4703 | int i; | |
b8698a0f | 4704 | |
e855c69d AB |
4705 | if (++succs_info_pool.max_top >= succs_info_pool.size) |
4706 | gcc_unreachable (); | |
4707 | ||
4708 | i = ++succs_info_pool.top; | |
9771b263 DN |
4709 | succs_info_pool.stack[i].succs_ok.create (10); |
4710 | succs_info_pool.stack[i].succs_other.create (10); | |
4711 | succs_info_pool.stack[i].probs_ok.create (10); | |
e855c69d AB |
4712 | } |
4713 | else | |
4714 | succs_info_pool.top++; | |
4715 | ||
4716 | return &succs_info_pool.stack[succs_info_pool.top]; | |
4717 | } | |
4718 | ||
4719 | /* Free successor's info. */ | |
4720 | void | |
4721 | free_succs_info (struct succs_info * sinfo) | |
4722 | { | |
b8698a0f | 4723 | gcc_assert (succs_info_pool.top >= 0 |
e855c69d AB |
4724 | && &succs_info_pool.stack[succs_info_pool.top] == sinfo); |
4725 | succs_info_pool.top--; | |
4726 | ||
4727 | /* Clear stale info. */ | |
9771b263 DN |
4728 | sinfo->succs_ok.block_remove (0, sinfo->succs_ok.length ()); |
4729 | sinfo->succs_other.block_remove (0, sinfo->succs_other.length ()); | |
4730 | sinfo->probs_ok.block_remove (0, sinfo->probs_ok.length ()); | |
e855c69d AB |
4731 | sinfo->all_prob = 0; |
4732 | sinfo->succs_ok_n = 0; | |
4733 | sinfo->all_succs_n = 0; | |
4734 | } | |
4735 | ||
b8698a0f | 4736 | /* Compute successor info for INSN. FLAGS are the flags passed |
e855c69d AB |
4737 | to the FOR_EACH_SUCC_1 iterator. */ |
4738 | struct succs_info * | |
4739 | compute_succs_info (insn_t insn, short flags) | |
4740 | { | |
4741 | succ_iterator si; | |
4742 | insn_t succ; | |
4743 | struct succs_info *sinfo = alloc_succs_info (); | |
4744 | ||
4745 | /* Traverse *all* successors and decide what to do with each. */ | |
4746 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
4747 | { | |
4748 | /* FIXME: this doesn't work for skipping to loop exits, as we don't | |
4749 | perform code motion through inner loops. */ | |
4750 | short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; | |
4751 | ||
4752 | if (current_flags & flags) | |
4753 | { | |
9771b263 DN |
4754 | sinfo->succs_ok.safe_push (succ); |
4755 | sinfo->probs_ok.safe_push ( | |
4756 | /* FIXME: Improve calculation when skipping | |
4757 | inner loop to exits. */ | |
9aad8d93 JH |
4758 | si.bb_end |
4759 | ? (si.e1->probability.initialized_p () | |
4760 | ? si.e1->probability.to_reg_br_prob_base () | |
4761 | : 0) | |
4762 | : REG_BR_PROB_BASE); | |
e855c69d AB |
4763 | sinfo->succs_ok_n++; |
4764 | } | |
4765 | else | |
9771b263 | 4766 | sinfo->succs_other.safe_push (succ); |
e855c69d AB |
4767 | |
4768 | /* Compute all_prob. */ | |
4769 | if (!si.bb_end) | |
4770 | sinfo->all_prob = REG_BR_PROB_BASE; | |
357067f2 JH |
4771 | else if (si.e1->probability.initialized_p ()) |
4772 | sinfo->all_prob += si.e1->probability.to_reg_br_prob_base (); | |
e855c69d AB |
4773 | |
4774 | sinfo->all_succs_n++; | |
4775 | } | |
4776 | ||
4777 | return sinfo; | |
4778 | } | |
4779 | ||
b8698a0f | 4780 | /* Return the predecessors of BB in PREDS and their number in N. |
e855c69d AB |
4781 | Empty blocks are skipped. SIZE is used to allocate PREDS. */ |
4782 | static void | |
4783 | cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) | |
4784 | { | |
4785 | edge e; | |
4786 | edge_iterator ei; | |
4787 | ||
4788 | gcc_assert (BLOCK_TO_BB (bb->index) != 0); | |
4789 | ||
4790 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4791 | { | |
4792 | basic_block pred_bb = e->src; | |
4793 | insn_t bb_end = BB_END (pred_bb); | |
4794 | ||
3e6a3f6f AB |
4795 | if (!in_current_region_p (pred_bb)) |
4796 | { | |
4797 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
4798 | && current_loop_nest); | |
4799 | continue; | |
4800 | } | |
e855c69d AB |
4801 | |
4802 | if (sel_bb_empty_p (pred_bb)) | |
4803 | cfg_preds_1 (pred_bb, preds, n, size); | |
4804 | else | |
4805 | { | |
4806 | if (*n == *size) | |
b8698a0f | 4807 | *preds = XRESIZEVEC (insn_t, *preds, |
e855c69d AB |
4808 | (*size = 2 * *size + 1)); |
4809 | (*preds)[(*n)++] = bb_end; | |
4810 | } | |
4811 | } | |
4812 | ||
3e6a3f6f AB |
4813 | gcc_assert (*n != 0 |
4814 | || (flag_sel_sched_pipelining_outer_loops | |
4815 | && current_loop_nest)); | |
e855c69d AB |
4816 | } |
4817 | ||
b8698a0f L |
4818 | /* Find all predecessors of BB and record them in PREDS and their number |
4819 | in N. Empty blocks are skipped, and only normal (forward in-region) | |
e855c69d AB |
4820 | edges are processed. */ |
4821 | static void | |
4822 | cfg_preds (basic_block bb, insn_t **preds, int *n) | |
4823 | { | |
4824 | int size = 0; | |
4825 | ||
4826 | *preds = NULL; | |
4827 | *n = 0; | |
4828 | cfg_preds_1 (bb, preds, n, &size); | |
4829 | } | |
4830 | ||
4831 | /* Returns true if we are moving INSN through join point. */ | |
4832 | bool | |
4833 | sel_num_cfg_preds_gt_1 (insn_t insn) | |
4834 | { | |
4835 | basic_block bb; | |
4836 | ||
4837 | if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) | |
4838 | return false; | |
4839 | ||
4840 | bb = BLOCK_FOR_INSN (insn); | |
4841 | ||
4842 | while (1) | |
4843 | { | |
4844 | if (EDGE_COUNT (bb->preds) > 1) | |
4845 | return true; | |
4846 | ||
4847 | gcc_assert (EDGE_PRED (bb, 0)->dest == bb); | |
4848 | bb = EDGE_PRED (bb, 0)->src; | |
4849 | ||
4850 | if (!sel_bb_empty_p (bb)) | |
4851 | break; | |
4852 | } | |
4853 | ||
4854 | return false; | |
4855 | } | |
4856 | ||
b8698a0f | 4857 | /* Returns true when BB should be the end of an ebb. Adapted from the |
e855c69d AB |
4858 | code in sched-ebb.c. */ |
4859 | bool | |
4860 | bb_ends_ebb_p (basic_block bb) | |
4861 | { | |
4862 | basic_block next_bb = bb_next_bb (bb); | |
4863 | edge e; | |
b8698a0f | 4864 | |
fefa31b5 | 4865 | if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e855c69d AB |
4866 | || bitmap_bit_p (forced_ebb_heads, next_bb->index) |
4867 | || (LABEL_P (BB_HEAD (next_bb)) | |
4868 | /* NB: LABEL_NUSES () is not maintained outside of jump.c. | |
4869 | Work around that. */ | |
4870 | && !single_pred_p (next_bb))) | |
4871 | return true; | |
4872 | ||
4873 | if (!in_current_region_p (next_bb)) | |
4874 | return true; | |
4875 | ||
0fd4b31d NF |
4876 | e = find_fallthru_edge (bb->succs); |
4877 | if (e) | |
4878 | { | |
4879 | gcc_assert (e->dest == next_bb); | |
4880 | ||
4881 | return false; | |
4882 | } | |
e855c69d AB |
4883 | |
4884 | return true; | |
4885 | } | |
4886 | ||
4887 | /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a | |
4888 | successor of INSN. */ | |
4889 | bool | |
4890 | in_same_ebb_p (insn_t insn, insn_t succ) | |
4891 | { | |
4892 | basic_block ptr = BLOCK_FOR_INSN (insn); | |
4893 | ||
c3284718 | 4894 | for (;;) |
e855c69d AB |
4895 | { |
4896 | if (ptr == BLOCK_FOR_INSN (succ)) | |
4897 | return true; | |
b8698a0f | 4898 | |
e855c69d AB |
4899 | if (bb_ends_ebb_p (ptr)) |
4900 | return false; | |
4901 | ||
4902 | ptr = bb_next_bb (ptr); | |
4903 | } | |
4904 | ||
4905 | gcc_unreachable (); | |
4906 | return false; | |
4907 | } | |
4908 | ||
4909 | /* Recomputes the reverse topological order for the function and | |
4910 | saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also | |
4911 | modified appropriately. */ | |
4912 | static void | |
4913 | recompute_rev_top_order (void) | |
4914 | { | |
4915 | int *postorder; | |
4916 | int n_blocks, i; | |
4917 | ||
8b1c6fd7 DM |
4918 | if (!rev_top_order_index |
4919 | || rev_top_order_index_len < last_basic_block_for_fn (cfun)) | |
e855c69d | 4920 | { |
8b1c6fd7 | 4921 | rev_top_order_index_len = last_basic_block_for_fn (cfun); |
e855c69d AB |
4922 | rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, |
4923 | rev_top_order_index_len); | |
4924 | } | |
4925 | ||
0cae8d31 | 4926 | postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
e855c69d AB |
4927 | |
4928 | n_blocks = post_order_compute (postorder, true, false); | |
0cae8d31 | 4929 | gcc_assert (n_basic_blocks_for_fn (cfun) == n_blocks); |
e855c69d AB |
4930 | |
4931 | /* Build reverse function: for each basic block with BB->INDEX == K | |
4932 | rev_top_order_index[K] is it's reverse topological sort number. */ | |
4933 | for (i = 0; i < n_blocks; i++) | |
4934 | { | |
4935 | gcc_assert (postorder[i] < rev_top_order_index_len); | |
4936 | rev_top_order_index[postorder[i]] = i; | |
4937 | } | |
4938 | ||
4939 | free (postorder); | |
4940 | } | |
4941 | ||
4942 | /* Clear all flags from insns in BB that could spoil its rescheduling. */ | |
4943 | void | |
4944 | clear_outdated_rtx_info (basic_block bb) | |
4945 | { | |
b32d5189 | 4946 | rtx_insn *insn; |
e855c69d AB |
4947 | |
4948 | FOR_BB_INSNS (bb, insn) | |
4949 | if (INSN_P (insn)) | |
4950 | { | |
4951 | SCHED_GROUP_P (insn) = 0; | |
4952 | INSN_AFTER_STALL_P (insn) = 0; | |
4953 | INSN_SCHED_TIMES (insn) = 0; | |
4954 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; | |
4955 | ||
4956 | /* We cannot use the changed caches, as previously we could ignore | |
b8698a0f | 4957 | the LHS dependence due to enabled renaming and transform |
e855c69d AB |
4958 | the expression, and currently we'll be unable to do this. */ |
4959 | htab_empty (INSN_TRANSFORMED_INSNS (insn)); | |
4960 | } | |
4961 | } | |
4962 | ||
4963 | /* Add BB_NOTE to the pool of available basic block notes. */ | |
4964 | static void | |
4965 | return_bb_to_pool (basic_block bb) | |
4966 | { | |
e67d1102 | 4967 | rtx_note *note = bb_note (bb); |
e855c69d AB |
4968 | |
4969 | gcc_assert (NOTE_BASIC_BLOCK (note) == bb | |
4970 | && bb->aux == NULL); | |
4971 | ||
4972 | /* It turns out that current cfg infrastructure does not support | |
4973 | reuse of basic blocks. Don't bother for now. */ | |
9771b263 | 4974 | /*bb_note_pool.safe_push (note);*/ |
e855c69d AB |
4975 | } |
4976 | ||
4977 | /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ | |
66e8df53 | 4978 | static rtx_note * |
e855c69d AB |
4979 | get_bb_note_from_pool (void) |
4980 | { | |
9771b263 | 4981 | if (bb_note_pool.is_empty ()) |
66e8df53 | 4982 | return NULL; |
e855c69d AB |
4983 | else |
4984 | { | |
66e8df53 | 4985 | rtx_note *note = bb_note_pool.pop (); |
e855c69d | 4986 | |
0f82e5c9 DM |
4987 | SET_PREV_INSN (note) = NULL_RTX; |
4988 | SET_NEXT_INSN (note) = NULL_RTX; | |
e855c69d AB |
4989 | |
4990 | return note; | |
4991 | } | |
4992 | } | |
4993 | ||
4994 | /* Free bb_note_pool. */ | |
4995 | void | |
4996 | free_bb_note_pool (void) | |
4997 | { | |
9771b263 | 4998 | bb_note_pool.release (); |
e855c69d AB |
4999 | } |
5000 | ||
5001 | /* Setup scheduler pool and successor structure. */ | |
5002 | void | |
5003 | alloc_sched_pools (void) | |
5004 | { | |
5005 | int succs_size; | |
5006 | ||
5007 | succs_size = MAX_WS + 1; | |
b8698a0f | 5008 | succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); |
e855c69d AB |
5009 | succs_info_pool.size = succs_size; |
5010 | succs_info_pool.top = -1; | |
5011 | succs_info_pool.max_top = -1; | |
e855c69d AB |
5012 | } |
5013 | ||
5014 | /* Free the pools. */ | |
5015 | void | |
5016 | free_sched_pools (void) | |
5017 | { | |
5018 | int i; | |
b8698a0f | 5019 | |
8bb6373a | 5020 | sched_lists_pool.release (); |
e855c69d | 5021 | gcc_assert (succs_info_pool.top == -1); |
85f5dbea | 5022 | for (i = 0; i <= succs_info_pool.max_top; i++) |
e855c69d | 5023 | { |
9771b263 DN |
5024 | succs_info_pool.stack[i].succs_ok.release (); |
5025 | succs_info_pool.stack[i].succs_other.release (); | |
5026 | succs_info_pool.stack[i].probs_ok.release (); | |
e855c69d AB |
5027 | } |
5028 | free (succs_info_pool.stack); | |
5029 | } | |
5030 | \f | |
5031 | ||
b8698a0f | 5032 | /* Returns a position in RGN where BB can be inserted retaining |
e855c69d AB |
5033 | topological order. */ |
5034 | static int | |
5035 | find_place_to_insert_bb (basic_block bb, int rgn) | |
5036 | { | |
5037 | bool has_preds_outside_rgn = false; | |
5038 | edge e; | |
5039 | edge_iterator ei; | |
b8698a0f | 5040 | |
e855c69d AB |
5041 | /* Find whether we have preds outside the region. */ |
5042 | FOR_EACH_EDGE (e, ei, bb->preds) | |
5043 | if (!in_current_region_p (e->src)) | |
5044 | { | |
5045 | has_preds_outside_rgn = true; | |
5046 | break; | |
5047 | } | |
b8698a0f | 5048 | |
e855c69d AB |
5049 | /* Recompute the top order -- needed when we have > 1 pred |
5050 | and in case we don't have preds outside. */ | |
5051 | if (flag_sel_sched_pipelining_outer_loops | |
5052 | && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) | |
5053 | { | |
5054 | int i, bbi = bb->index, cur_bbi; | |
5055 | ||
5056 | recompute_rev_top_order (); | |
5057 | for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) | |
5058 | { | |
5059 | cur_bbi = BB_TO_BLOCK (i); | |
b8698a0f | 5060 | if (rev_top_order_index[bbi] |
e855c69d AB |
5061 | < rev_top_order_index[cur_bbi]) |
5062 | break; | |
5063 | } | |
b8698a0f | 5064 | |
073a8998 | 5065 | /* We skipped the right block, so we increase i. We accommodate |
e855c69d AB |
5066 | it for increasing by step later, so we decrease i. */ |
5067 | return (i + 1) - 1; | |
5068 | } | |
5069 | else if (has_preds_outside_rgn) | |
5070 | { | |
5071 | /* This is the case when we generate an extra empty block | |
5072 | to serve as region head during pipelining. */ | |
5073 | e = EDGE_SUCC (bb, 0); | |
5074 | gcc_assert (EDGE_COUNT (bb->succs) == 1 | |
5075 | && in_current_region_p (EDGE_SUCC (bb, 0)->dest) | |
5076 | && (BLOCK_TO_BB (e->dest->index) == 0)); | |
5077 | return -1; | |
5078 | } | |
5079 | ||
5080 | /* We don't have preds outside the region. We should have | |
5081 | the only pred, because the multiple preds case comes from | |
5082 | the pipelining of outer loops, and that is handled above. | |
5083 | Just take the bbi of this single pred. */ | |
5084 | if (EDGE_COUNT (bb->succs) > 0) | |
5085 | { | |
5086 | int pred_bbi; | |
b8698a0f | 5087 | |
e855c69d | 5088 | gcc_assert (EDGE_COUNT (bb->preds) == 1); |
b8698a0f | 5089 | |
e855c69d AB |
5090 | pred_bbi = EDGE_PRED (bb, 0)->src->index; |
5091 | return BLOCK_TO_BB (pred_bbi); | |
5092 | } | |
5093 | else | |
5094 | /* BB has no successors. It is safe to put it in the end. */ | |
5095 | return current_nr_blocks - 1; | |
5096 | } | |
5097 | ||
5098 | /* Deletes an empty basic block freeing its data. */ | |
5099 | static void | |
5100 | delete_and_free_basic_block (basic_block bb) | |
5101 | { | |
5102 | gcc_assert (sel_bb_empty_p (bb)); | |
5103 | ||
5104 | if (BB_LV_SET (bb)) | |
5105 | free_lv_set (bb); | |
5106 | ||
5107 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
5108 | ||
b8698a0f L |
5109 | /* Can't assert av_set properties because we use sel_aremove_bb |
5110 | when removing loop preheader from the region. At the point of | |
e855c69d AB |
5111 | removing the preheader we already have deallocated sel_region_bb_info. */ |
5112 | gcc_assert (BB_LV_SET (bb) == NULL | |
5113 | && !BB_LV_SET_VALID_P (bb) | |
5114 | && BB_AV_LEVEL (bb) == 0 | |
5115 | && BB_AV_SET (bb) == NULL); | |
b8698a0f | 5116 | |
e855c69d AB |
5117 | delete_basic_block (bb); |
5118 | } | |
5119 | ||
5120 | /* Add BB to the current region and update the region data. */ | |
5121 | static void | |
5122 | add_block_to_current_region (basic_block bb) | |
5123 | { | |
5124 | int i, pos, bbi = -2, rgn; | |
5125 | ||
5126 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
5127 | bbi = find_place_to_insert_bb (bb, rgn); | |
5128 | bbi += 1; | |
5129 | pos = RGN_BLOCKS (rgn) + bbi; | |
5130 | ||
5131 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
5132 | && ebb_head[bbi] == pos); | |
b8698a0f | 5133 | |
e855c69d AB |
5134 | /* Make a place for the new block. */ |
5135 | extend_regions (); | |
5136 | ||
5137 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
5138 | BLOCK_TO_BB (rgn_bb_table[i])++; | |
b8698a0f | 5139 | |
e855c69d AB |
5140 | memmove (rgn_bb_table + pos + 1, |
5141 | rgn_bb_table + pos, | |
5142 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5143 | ||
5144 | /* Initialize data for BB. */ | |
5145 | rgn_bb_table[pos] = bb->index; | |
5146 | BLOCK_TO_BB (bb->index) = bbi; | |
5147 | CONTAINING_RGN (bb->index) = rgn; | |
5148 | ||
5149 | RGN_NR_BLOCKS (rgn)++; | |
b8698a0f | 5150 | |
e855c69d AB |
5151 | for (i = rgn + 1; i <= nr_regions; i++) |
5152 | RGN_BLOCKS (i)++; | |
5153 | } | |
5154 | ||
5155 | /* Remove BB from the current region and update the region data. */ | |
5156 | static void | |
5157 | remove_bb_from_region (basic_block bb) | |
5158 | { | |
5159 | int i, pos, bbi = -2, rgn; | |
5160 | ||
5161 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
5162 | bbi = BLOCK_TO_BB (bb->index); | |
5163 | pos = RGN_BLOCKS (rgn) + bbi; | |
5164 | ||
5165 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
5166 | && ebb_head[bbi] == pos); | |
5167 | ||
5168 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
5169 | BLOCK_TO_BB (rgn_bb_table[i])--; | |
5170 | ||
5171 | memmove (rgn_bb_table + pos, | |
5172 | rgn_bb_table + pos + 1, | |
5173 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5174 | ||
5175 | RGN_NR_BLOCKS (rgn)--; | |
5176 | for (i = rgn + 1; i <= nr_regions; i++) | |
5177 | RGN_BLOCKS (i)--; | |
5178 | } | |
5179 | ||
b8698a0f | 5180 | /* Add BB to the current region and update all data. If BB is NULL, add all |
e855c69d AB |
5181 | blocks from last_added_blocks vector. */ |
5182 | static void | |
5183 | sel_add_bb (basic_block bb) | |
5184 | { | |
5185 | /* Extend luids so that new notes will receive zero luids. */ | |
a95b23b4 | 5186 | sched_extend_luids (); |
e855c69d | 5187 | sched_init_bbs (); |
a95b23b4 | 5188 | sel_init_bbs (last_added_blocks); |
e855c69d | 5189 | |
b8698a0f | 5190 | /* When bb is passed explicitly, the vector should contain |
e855c69d AB |
5191 | the only element that equals to bb; otherwise, the vector |
5192 | should not be NULL. */ | |
9771b263 | 5193 | gcc_assert (last_added_blocks.exists ()); |
b8698a0f | 5194 | |
e855c69d AB |
5195 | if (bb != NULL) |
5196 | { | |
9771b263 DN |
5197 | gcc_assert (last_added_blocks.length () == 1 |
5198 | && last_added_blocks[0] == bb); | |
e855c69d AB |
5199 | add_block_to_current_region (bb); |
5200 | ||
5201 | /* We associate creating/deleting data sets with the first insn | |
5202 | appearing / disappearing in the bb. */ | |
5203 | if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) | |
5204 | create_initial_data_sets (bb); | |
b8698a0f | 5205 | |
9771b263 | 5206 | last_added_blocks.release (); |
e855c69d AB |
5207 | } |
5208 | else | |
5209 | /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ | |
5210 | { | |
5211 | int i; | |
5212 | basic_block temp_bb = NULL; | |
5213 | ||
b8698a0f | 5214 | for (i = 0; |
9771b263 | 5215 | last_added_blocks.iterate (i, &bb); i++) |
e855c69d AB |
5216 | { |
5217 | add_block_to_current_region (bb); | |
5218 | temp_bb = bb; | |
5219 | } | |
5220 | ||
b8698a0f | 5221 | /* We need to fetch at least one bb so we know the region |
e855c69d AB |
5222 | to update. */ |
5223 | gcc_assert (temp_bb != NULL); | |
5224 | bb = temp_bb; | |
5225 | ||
9771b263 | 5226 | last_added_blocks.release (); |
e855c69d AB |
5227 | } |
5228 | ||
5229 | rgn_setup_region (CONTAINING_RGN (bb->index)); | |
5230 | } | |
5231 | ||
b8698a0f | 5232 | /* Remove BB from the current region and update all data. |
e855c69d AB |
5233 | If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ |
5234 | static void | |
5235 | sel_remove_bb (basic_block bb, bool remove_from_cfg_p) | |
5236 | { | |
262d8232 AB |
5237 | unsigned idx = bb->index; |
5238 | ||
e855c69d | 5239 | gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); |
b8698a0f | 5240 | |
e855c69d AB |
5241 | remove_bb_from_region (bb); |
5242 | return_bb_to_pool (bb); | |
262d8232 | 5243 | bitmap_clear_bit (blocks_to_reschedule, idx); |
b8698a0f | 5244 | |
e855c69d | 5245 | if (remove_from_cfg_p) |
00c4e97c AB |
5246 | { |
5247 | basic_block succ = single_succ (bb); | |
5248 | delete_and_free_basic_block (bb); | |
5249 | set_immediate_dominator (CDI_DOMINATORS, succ, | |
5250 | recompute_dominator (CDI_DOMINATORS, succ)); | |
5251 | } | |
e855c69d | 5252 | |
262d8232 | 5253 | rgn_setup_region (CONTAINING_RGN (idx)); |
e855c69d AB |
5254 | } |
5255 | ||
5256 | /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ | |
5257 | static void | |
5258 | move_bb_info (basic_block merge_bb, basic_block empty_bb) | |
5259 | { | |
b7b5540a AB |
5260 | if (in_current_region_p (merge_bb)) |
5261 | concat_note_lists (BB_NOTE_LIST (empty_bb), | |
b311fd0f DM |
5262 | &BB_NOTE_LIST (merge_bb)); |
5263 | BB_NOTE_LIST (empty_bb) = NULL; | |
e855c69d AB |
5264 | |
5265 | } | |
5266 | ||
e855c69d AB |
5267 | /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from |
5268 | region, but keep it in CFG. */ | |
5269 | static void | |
5270 | remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) | |
5271 | { | |
5272 | /* The block should contain just a note or a label. | |
5273 | We try to check whether it is unused below. */ | |
5274 | gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) | |
5275 | || LABEL_P (BB_HEAD (empty_bb))); | |
5276 | ||
5277 | /* If basic block has predecessors or successors, redirect them. */ | |
5278 | if (remove_from_cfg_p | |
5279 | && (EDGE_COUNT (empty_bb->preds) > 0 | |
5280 | || EDGE_COUNT (empty_bb->succs) > 0)) | |
5281 | { | |
5282 | basic_block pred; | |
5283 | basic_block succ; | |
5284 | ||
5285 | /* We need to init PRED and SUCC before redirecting edges. */ | |
5286 | if (EDGE_COUNT (empty_bb->preds) > 0) | |
5287 | { | |
5288 | edge e; | |
5289 | ||
5290 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); | |
5291 | ||
5292 | e = EDGE_PRED (empty_bb, 0); | |
5293 | gcc_assert (e->src == empty_bb->prev_bb | |
5294 | && (e->flags & EDGE_FALLTHRU)); | |
5295 | ||
5296 | pred = empty_bb->prev_bb; | |
5297 | } | |
5298 | else | |
5299 | pred = NULL; | |
5300 | ||
5301 | if (EDGE_COUNT (empty_bb->succs) > 0) | |
5302 | { | |
5303 | /* We do not check fallthruness here as above, because | |
5304 | after removing a jump the edge may actually be not fallthru. */ | |
5305 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); | |
5306 | succ = EDGE_SUCC (empty_bb, 0)->dest; | |
5307 | } | |
5308 | else | |
5309 | succ = NULL; | |
5310 | ||
5311 | if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) | |
5312 | { | |
5313 | edge e = EDGE_PRED (empty_bb, 0); | |
5314 | ||
5315 | if (e->flags & EDGE_FALLTHRU) | |
5316 | redirect_edge_succ_nodup (e, succ); | |
5317 | else | |
5318 | sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); | |
5319 | } | |
5320 | ||
5321 | if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) | |
5322 | { | |
5323 | edge e = EDGE_SUCC (empty_bb, 0); | |
5324 | ||
5325 | if (find_edge (pred, e->dest) == NULL) | |
5326 | redirect_edge_pred (e, pred); | |
5327 | } | |
5328 | } | |
5329 | ||
5330 | /* Finish removing. */ | |
5331 | sel_remove_bb (empty_bb, remove_from_cfg_p); | |
5332 | } | |
5333 | ||
b8698a0f | 5334 | /* An implementation of create_basic_block hook, which additionally updates |
e855c69d AB |
5335 | per-bb data structures. */ |
5336 | static basic_block | |
5337 | sel_create_basic_block (void *headp, void *endp, basic_block after) | |
5338 | { | |
5339 | basic_block new_bb; | |
66e8df53 | 5340 | rtx_note *new_bb_note; |
b8698a0f L |
5341 | |
5342 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
9771b263 | 5343 | || !last_added_blocks.exists ()); |
e855c69d AB |
5344 | |
5345 | new_bb_note = get_bb_note_from_pool (); | |
5346 | ||
5347 | if (new_bb_note == NULL_RTX) | |
5348 | new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); | |
5349 | else | |
5350 | { | |
e00022e9 DM |
5351 | new_bb = create_basic_block_structure ((rtx_insn *) headp, |
5352 | (rtx_insn *) endp, | |
e855c69d AB |
5353 | new_bb_note, after); |
5354 | new_bb->aux = NULL; | |
5355 | } | |
5356 | ||
9771b263 | 5357 | last_added_blocks.safe_push (new_bb); |
e855c69d AB |
5358 | |
5359 | return new_bb; | |
5360 | } | |
5361 | ||
5362 | /* Implement sched_init_only_bb (). */ | |
5363 | static void | |
5364 | sel_init_only_bb (basic_block bb, basic_block after) | |
5365 | { | |
5366 | gcc_assert (after == NULL); | |
5367 | ||
5368 | extend_regions (); | |
5369 | rgn_make_new_region_out_of_new_block (bb); | |
5370 | } | |
5371 | ||
5372 | /* Update the latch when we've splitted or merged it from FROM block to TO. | |
5373 | This should be checked for all outer loops, too. */ | |
5374 | static void | |
5375 | change_loops_latches (basic_block from, basic_block to) | |
5376 | { | |
5377 | gcc_assert (from != to); | |
5378 | ||
5379 | if (current_loop_nest) | |
5380 | { | |
5381 | struct loop *loop; | |
5382 | ||
5383 | for (loop = current_loop_nest; loop; loop = loop_outer (loop)) | |
5384 | if (considered_for_pipelining_p (loop) && loop->latch == from) | |
5385 | { | |
5386 | gcc_assert (loop == current_loop_nest); | |
5387 | loop->latch = to; | |
5388 | gcc_assert (loop_latch_edge (loop)); | |
5389 | } | |
5390 | } | |
5391 | } | |
5392 | ||
b8698a0f | 5393 | /* Splits BB on two basic blocks, adding it to the region and extending |
e855c69d AB |
5394 | per-bb data structures. Returns the newly created bb. */ |
5395 | static basic_block | |
5396 | sel_split_block (basic_block bb, rtx after) | |
5397 | { | |
5398 | basic_block new_bb; | |
5399 | insn_t insn; | |
5400 | ||
5401 | new_bb = sched_split_block_1 (bb, after); | |
5402 | sel_add_bb (new_bb); | |
5403 | ||
5404 | /* This should be called after sel_add_bb, because this uses | |
b8698a0f | 5405 | CONTAINING_RGN for the new block, which is not yet initialized. |
e855c69d AB |
5406 | FIXME: this function may be a no-op now. */ |
5407 | change_loops_latches (bb, new_bb); | |
5408 | ||
5409 | /* Update ORIG_BB_INDEX for insns moved into the new block. */ | |
5410 | FOR_BB_INSNS (new_bb, insn) | |
5411 | if (INSN_P (insn)) | |
5412 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; | |
5413 | ||
5414 | if (sel_bb_empty_p (bb)) | |
5415 | { | |
5416 | gcc_assert (!sel_bb_empty_p (new_bb)); | |
5417 | ||
5418 | /* NEW_BB has data sets that need to be updated and BB holds | |
5419 | data sets that should be removed. Exchange these data sets | |
5420 | so that we won't lose BB's valid data sets. */ | |
5421 | exchange_data_sets (new_bb, bb); | |
5422 | free_data_sets (bb); | |
5423 | } | |
5424 | ||
5425 | if (!sel_bb_empty_p (new_bb) | |
5426 | && bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
5427 | bitmap_set_bit (blocks_to_reschedule, new_bb->index); | |
5428 | ||
5429 | return new_bb; | |
5430 | } | |
5431 | ||
5432 | /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. | |
5433 | Otherwise returns NULL. */ | |
9ee1fbb1 | 5434 | static rtx_insn * |
e855c69d AB |
5435 | check_for_new_jump (basic_block bb, int prev_max_uid) |
5436 | { | |
9ee1fbb1 | 5437 | rtx_insn *end; |
e855c69d AB |
5438 | |
5439 | end = sel_bb_end (bb); | |
5440 | if (end && INSN_UID (end) >= prev_max_uid) | |
5441 | return end; | |
5442 | return NULL; | |
5443 | } | |
5444 | ||
b8698a0f | 5445 | /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. |
e855c69d | 5446 | New means having UID at least equal to PREV_MAX_UID. */ |
9ee1fbb1 | 5447 | static rtx_insn * |
e855c69d AB |
5448 | find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) |
5449 | { | |
9ee1fbb1 | 5450 | rtx_insn *jump; |
e855c69d AB |
5451 | |
5452 | /* Return immediately if no new insns were emitted. */ | |
5453 | if (get_max_uid () == prev_max_uid) | |
5454 | return NULL; | |
b8698a0f | 5455 | |
e855c69d AB |
5456 | /* Now check both blocks for new jumps. It will ever be only one. */ |
5457 | if ((jump = check_for_new_jump (from, prev_max_uid))) | |
5458 | return jump; | |
5459 | ||
5460 | if (jump_bb != NULL | |
5461 | && (jump = check_for_new_jump (jump_bb, prev_max_uid))) | |
5462 | return jump; | |
5463 | return NULL; | |
5464 | } | |
5465 | ||
5466 | /* Splits E and adds the newly created basic block to the current region. | |
5467 | Returns this basic block. */ | |
5468 | basic_block | |
5469 | sel_split_edge (edge e) | |
5470 | { | |
5471 | basic_block new_bb, src, other_bb = NULL; | |
5472 | int prev_max_uid; | |
9ee1fbb1 | 5473 | rtx_insn *jump; |
e855c69d AB |
5474 | |
5475 | src = e->src; | |
5476 | prev_max_uid = get_max_uid (); | |
5477 | new_bb = split_edge (e); | |
5478 | ||
b8698a0f | 5479 | if (flag_sel_sched_pipelining_outer_loops |
e855c69d AB |
5480 | && current_loop_nest) |
5481 | { | |
5482 | int i; | |
5483 | basic_block bb; | |
5484 | ||
b8698a0f | 5485 | /* Some of the basic blocks might not have been added to the loop. |
e855c69d | 5486 | Add them here, until this is fixed in force_fallthru. */ |
b8698a0f | 5487 | for (i = 0; |
9771b263 | 5488 | last_added_blocks.iterate (i, &bb); i++) |
e855c69d AB |
5489 | if (!bb->loop_father) |
5490 | { | |
5491 | add_bb_to_loop (bb, e->dest->loop_father); | |
5492 | ||
5493 | gcc_assert (!other_bb && (new_bb->index != bb->index)); | |
5494 | other_bb = bb; | |
5495 | } | |
5496 | } | |
5497 | ||
5498 | /* Add all last_added_blocks to the region. */ | |
5499 | sel_add_bb (NULL); | |
5500 | ||
5501 | jump = find_new_jump (src, new_bb, prev_max_uid); | |
5502 | if (jump) | |
5503 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5504 | ||
5505 | /* Put the correct lv set on this block. */ | |
5506 | if (other_bb && !sel_bb_empty_p (other_bb)) | |
5507 | compute_live (sel_bb_head (other_bb)); | |
5508 | ||
5509 | return new_bb; | |
5510 | } | |
5511 | ||
5512 | /* Implement sched_create_empty_bb (). */ | |
5513 | static basic_block | |
5514 | sel_create_empty_bb (basic_block after) | |
5515 | { | |
5516 | basic_block new_bb; | |
5517 | ||
5518 | new_bb = sched_create_empty_bb_1 (after); | |
5519 | ||
5520 | /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit | |
5521 | later. */ | |
9771b263 DN |
5522 | gcc_assert (last_added_blocks.length () == 1 |
5523 | && last_added_blocks[0] == new_bb); | |
e855c69d | 5524 | |
9771b263 | 5525 | last_added_blocks.release (); |
e855c69d AB |
5526 | return new_bb; |
5527 | } | |
5528 | ||
5529 | /* Implement sched_create_recovery_block. ORIG_INSN is where block | |
5530 | will be splitted to insert a check. */ | |
5531 | basic_block | |
5532 | sel_create_recovery_block (insn_t orig_insn) | |
5533 | { | |
5534 | basic_block first_bb, second_bb, recovery_block; | |
5535 | basic_block before_recovery = NULL; | |
9ee1fbb1 | 5536 | rtx_insn *jump; |
e855c69d AB |
5537 | |
5538 | first_bb = BLOCK_FOR_INSN (orig_insn); | |
5539 | if (sel_bb_end_p (orig_insn)) | |
5540 | { | |
5541 | /* Avoid introducing an empty block while splitting. */ | |
5542 | gcc_assert (single_succ_p (first_bb)); | |
5543 | second_bb = single_succ (first_bb); | |
5544 | } | |
5545 | else | |
5546 | second_bb = sched_split_block (first_bb, orig_insn); | |
5547 | ||
5548 | recovery_block = sched_create_recovery_block (&before_recovery); | |
5549 | if (before_recovery) | |
fefa31b5 | 5550 | copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e855c69d AB |
5551 | |
5552 | gcc_assert (sel_bb_empty_p (recovery_block)); | |
5553 | sched_create_recovery_edges (first_bb, recovery_block, second_bb); | |
5554 | if (current_loops != NULL) | |
5555 | add_bb_to_loop (recovery_block, first_bb->loop_father); | |
b8698a0f | 5556 | |
e855c69d | 5557 | sel_add_bb (recovery_block); |
b8698a0f | 5558 | |
e855c69d AB |
5559 | jump = BB_END (recovery_block); |
5560 | gcc_assert (sel_bb_head (recovery_block) == jump); | |
5561 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5562 | ||
5563 | return recovery_block; | |
5564 | } | |
5565 | ||
5566 | /* Merge basic block B into basic block A. */ | |
262d8232 | 5567 | static void |
e855c69d AB |
5568 | sel_merge_blocks (basic_block a, basic_block b) |
5569 | { | |
262d8232 AB |
5570 | gcc_assert (sel_bb_empty_p (b) |
5571 | && EDGE_COUNT (b->preds) == 1 | |
5572 | && EDGE_PRED (b, 0)->src == b->prev_bb); | |
e855c69d | 5573 | |
262d8232 AB |
5574 | move_bb_info (b->prev_bb, b); |
5575 | remove_empty_bb (b, false); | |
5576 | merge_blocks (a, b); | |
e855c69d AB |
5577 | change_loops_latches (b, a); |
5578 | } | |
5579 | ||
5580 | /* A wrapper for redirect_edge_and_branch_force, which also initializes | |
92e265ac | 5581 | data structures for possibly created bb and insns. */ |
e855c69d AB |
5582 | void |
5583 | sel_redirect_edge_and_branch_force (edge e, basic_block to) | |
5584 | { | |
00c4e97c | 5585 | basic_block jump_bb, src, orig_dest = e->dest; |
e855c69d | 5586 | int prev_max_uid; |
9ee1fbb1 | 5587 | rtx_insn *jump; |
92e265ac | 5588 | int old_seqno = -1; |
b8698a0f | 5589 | |
00c4e97c AB |
5590 | /* This function is now used only for bookkeeping code creation, where |
5591 | we'll never get the single pred of orig_dest block and thus will not | |
5592 | hit unreachable blocks when updating dominator info. */ | |
5593 | gcc_assert (!sel_bb_empty_p (e->src) | |
5594 | && !single_pred_p (orig_dest)); | |
e855c69d AB |
5595 | src = e->src; |
5596 | prev_max_uid = get_max_uid (); | |
92e265ac AB |
5597 | /* Compute and pass old_seqno down to sel_init_new_insn only for the case |
5598 | when the conditional jump being redirected may become unconditional. */ | |
5599 | if (any_condjump_p (BB_END (src)) | |
5600 | && INSN_SEQNO (BB_END (src)) >= 0) | |
5601 | old_seqno = INSN_SEQNO (BB_END (src)); | |
e855c69d | 5602 | |
92e265ac | 5603 | jump_bb = redirect_edge_and_branch_force (e, to); |
e855c69d AB |
5604 | if (jump_bb != NULL) |
5605 | sel_add_bb (jump_bb); | |
5606 | ||
5607 | /* This function could not be used to spoil the loop structure by now, | |
5608 | thus we don't care to update anything. But check it to be sure. */ | |
5609 | if (current_loop_nest | |
5610 | && pipelining_p) | |
5611 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
b8698a0f | 5612 | |
e855c69d AB |
5613 | jump = find_new_jump (src, jump_bb, prev_max_uid); |
5614 | if (jump) | |
92e265ac AB |
5615 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, |
5616 | old_seqno); | |
00c4e97c AB |
5617 | set_immediate_dominator (CDI_DOMINATORS, to, |
5618 | recompute_dominator (CDI_DOMINATORS, to)); | |
5619 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5620 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
e855c69d AB |
5621 | } |
5622 | ||
b59ab570 AM |
5623 | /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by |
5624 | redirected edge are in reverse topological order. */ | |
5625 | bool | |
e855c69d AB |
5626 | sel_redirect_edge_and_branch (edge e, basic_block to) |
5627 | { | |
5628 | bool latch_edge_p; | |
00c4e97c | 5629 | basic_block src, orig_dest = e->dest; |
e855c69d | 5630 | int prev_max_uid; |
9ee1fbb1 | 5631 | rtx_insn *jump; |
f2c45f08 | 5632 | edge redirected; |
b59ab570 | 5633 | bool recompute_toporder_p = false; |
00c4e97c | 5634 | bool maybe_unreachable = single_pred_p (orig_dest); |
92e265ac | 5635 | int old_seqno = -1; |
e855c69d AB |
5636 | |
5637 | latch_edge_p = (pipelining_p | |
5638 | && current_loop_nest | |
5639 | && e == loop_latch_edge (current_loop_nest)); | |
5640 | ||
5641 | src = e->src; | |
5642 | prev_max_uid = get_max_uid (); | |
f2c45f08 | 5643 | |
92e265ac AB |
5644 | /* Compute and pass old_seqno down to sel_init_new_insn only for the case |
5645 | when the conditional jump being redirected may become unconditional. */ | |
5646 | if (any_condjump_p (BB_END (src)) | |
5647 | && INSN_SEQNO (BB_END (src)) >= 0) | |
5648 | old_seqno = INSN_SEQNO (BB_END (src)); | |
5649 | ||
f2c45f08 AM |
5650 | redirected = redirect_edge_and_branch (e, to); |
5651 | ||
9771b263 | 5652 | gcc_assert (redirected && !last_added_blocks.exists ()); |
e855c69d AB |
5653 | |
5654 | /* When we've redirected a latch edge, update the header. */ | |
5655 | if (latch_edge_p) | |
5656 | { | |
5657 | current_loop_nest->header = to; | |
5658 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
5659 | } | |
5660 | ||
b59ab570 AM |
5661 | /* In rare situations, the topological relation between the blocks connected |
5662 | by the redirected edge can change (see PR42245 for an example). Update | |
5663 | block_to_bb/bb_to_block. */ | |
5664 | if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) | |
5665 | && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) | |
5666 | recompute_toporder_p = true; | |
5667 | ||
e855c69d AB |
5668 | jump = find_new_jump (src, NULL, prev_max_uid); |
5669 | if (jump) | |
92e265ac | 5670 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, old_seqno); |
b59ab570 | 5671 | |
00c4e97c AB |
5672 | /* Only update dominator info when we don't have unreachable blocks. |
5673 | Otherwise we'll update in maybe_tidy_empty_bb. */ | |
5674 | if (!maybe_unreachable) | |
5675 | { | |
5676 | set_immediate_dominator (CDI_DOMINATORS, to, | |
5677 | recompute_dominator (CDI_DOMINATORS, to)); | |
5678 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5679 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
5680 | } | |
b59ab570 | 5681 | return recompute_toporder_p; |
e855c69d AB |
5682 | } |
5683 | ||
5684 | /* This variable holds the cfg hooks used by the selective scheduler. */ | |
5685 | static struct cfg_hooks sel_cfg_hooks; | |
5686 | ||
5687 | /* Register sel-sched cfg hooks. */ | |
5688 | void | |
5689 | sel_register_cfg_hooks (void) | |
5690 | { | |
5691 | sched_split_block = sel_split_block; | |
5692 | ||
5693 | orig_cfg_hooks = get_cfg_hooks (); | |
5694 | sel_cfg_hooks = orig_cfg_hooks; | |
5695 | ||
5696 | sel_cfg_hooks.create_basic_block = sel_create_basic_block; | |
5697 | ||
5698 | set_cfg_hooks (sel_cfg_hooks); | |
5699 | ||
5700 | sched_init_only_bb = sel_init_only_bb; | |
5701 | sched_split_block = sel_split_block; | |
5702 | sched_create_empty_bb = sel_create_empty_bb; | |
5703 | } | |
5704 | ||
5705 | /* Unregister sel-sched cfg hooks. */ | |
5706 | void | |
5707 | sel_unregister_cfg_hooks (void) | |
5708 | { | |
5709 | sched_create_empty_bb = NULL; | |
5710 | sched_split_block = NULL; | |
5711 | sched_init_only_bb = NULL; | |
5712 | ||
5713 | set_cfg_hooks (orig_cfg_hooks); | |
5714 | } | |
5715 | \f | |
5716 | ||
5717 | /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, | |
5718 | LABEL is where this jump should be directed. */ | |
9c068b73 | 5719 | rtx_insn * |
e855c69d AB |
5720 | create_insn_rtx_from_pattern (rtx pattern, rtx label) |
5721 | { | |
9c068b73 | 5722 | rtx_insn *insn_rtx; |
e855c69d AB |
5723 | |
5724 | gcc_assert (!INSN_P (pattern)); | |
5725 | ||
5726 | start_sequence (); | |
5727 | ||
5728 | if (label == NULL_RTX) | |
5729 | insn_rtx = emit_insn (pattern); | |
b5b8b0ac AO |
5730 | else if (DEBUG_INSN_P (label)) |
5731 | insn_rtx = emit_debug_insn (pattern); | |
e855c69d AB |
5732 | else |
5733 | { | |
5734 | insn_rtx = emit_jump_insn (pattern); | |
5735 | JUMP_LABEL (insn_rtx) = label; | |
5736 | ++LABEL_NUSES (label); | |
5737 | } | |
5738 | ||
5739 | end_sequence (); | |
5740 | ||
a95b23b4 | 5741 | sched_extend_luids (); |
e855c69d AB |
5742 | sched_extend_target (); |
5743 | sched_deps_init (false); | |
5744 | ||
5745 | /* Initialize INSN_CODE now. */ | |
5746 | recog_memoized (insn_rtx); | |
5747 | return insn_rtx; | |
5748 | } | |
5749 | ||
5750 | /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn | |
5751 | must not be clonable. */ | |
5752 | vinsn_t | |
6144a836 | 5753 | create_vinsn_from_insn_rtx (rtx_insn *insn_rtx, bool force_unique_p) |
e855c69d AB |
5754 | { |
5755 | gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); | |
5756 | ||
5757 | /* If VINSN_TYPE is not USE, retain its uniqueness. */ | |
5758 | return vinsn_create (insn_rtx, force_unique_p); | |
5759 | } | |
5760 | ||
5761 | /* Create a copy of INSN_RTX. */ | |
9c068b73 | 5762 | rtx_insn * |
e855c69d AB |
5763 | create_copy_of_insn_rtx (rtx insn_rtx) |
5764 | { | |
9c068b73 DM |
5765 | rtx_insn *res; |
5766 | rtx link; | |
e855c69d | 5767 | |
b5b8b0ac AO |
5768 | if (DEBUG_INSN_P (insn_rtx)) |
5769 | return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5770 | insn_rtx); | |
5771 | ||
e855c69d AB |
5772 | gcc_assert (NONJUMP_INSN_P (insn_rtx)); |
5773 | ||
5774 | res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5775 | NULL_RTX); | |
d734e6c4 | 5776 | |
1581a12c BS |
5777 | /* Locate the end of existing REG_NOTES in NEW_RTX. */ |
5778 | rtx *ptail = ®_NOTES (res); | |
5779 | while (*ptail != NULL_RTX) | |
5780 | ptail = &XEXP (*ptail, 1); | |
5781 | ||
d734e6c4 JJ |
5782 | /* Copy all REG_NOTES except REG_EQUAL/REG_EQUIV and REG_LABEL_OPERAND |
5783 | since mark_jump_label will make them. REG_LABEL_TARGETs are created | |
5784 | there too, but are supposed to be sticky, so we copy them. */ | |
5785 | for (link = REG_NOTES (insn_rtx); link; link = XEXP (link, 1)) | |
5786 | if (REG_NOTE_KIND (link) != REG_LABEL_OPERAND | |
5787 | && REG_NOTE_KIND (link) != REG_EQUAL | |
5788 | && REG_NOTE_KIND (link) != REG_EQUIV) | |
5789 | { | |
1581a12c BS |
5790 | *ptail = duplicate_reg_note (link); |
5791 | ptail = &XEXP (*ptail, 1); | |
d734e6c4 JJ |
5792 | } |
5793 | ||
e855c69d AB |
5794 | return res; |
5795 | } | |
5796 | ||
5797 | /* Change vinsn field of EXPR to hold NEW_VINSN. */ | |
5798 | void | |
5799 | change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) | |
5800 | { | |
5801 | vinsn_detach (EXPR_VINSN (expr)); | |
5802 | ||
5803 | EXPR_VINSN (expr) = new_vinsn; | |
5804 | vinsn_attach (new_vinsn); | |
5805 | } | |
5806 | ||
5807 | /* Helpers for global init. */ | |
5808 | /* This structure is used to be able to call existing bundling mechanism | |
5809 | and calculate insn priorities. */ | |
b8698a0f | 5810 | static struct haifa_sched_info sched_sel_haifa_sched_info = |
e855c69d AB |
5811 | { |
5812 | NULL, /* init_ready_list */ | |
5813 | NULL, /* can_schedule_ready_p */ | |
5814 | NULL, /* schedule_more_p */ | |
5815 | NULL, /* new_ready */ | |
5816 | NULL, /* rgn_rank */ | |
5817 | sel_print_insn, /* rgn_print_insn */ | |
5818 | contributes_to_priority, | |
356c23b3 | 5819 | NULL, /* insn_finishes_block_p */ |
e855c69d AB |
5820 | |
5821 | NULL, NULL, | |
5822 | NULL, NULL, | |
5823 | 0, 0, | |
5824 | ||
5825 | NULL, /* add_remove_insn */ | |
5826 | NULL, /* begin_schedule_ready */ | |
86014d07 | 5827 | NULL, /* begin_move_insn */ |
e855c69d | 5828 | NULL, /* advance_target_bb */ |
26965010 BS |
5829 | |
5830 | NULL, | |
5831 | NULL, | |
5832 | ||
e855c69d AB |
5833 | SEL_SCHED | NEW_BBS |
5834 | }; | |
5835 | ||
5836 | /* Setup special insns used in the scheduler. */ | |
b8698a0f | 5837 | void |
e855c69d AB |
5838 | setup_nop_and_exit_insns (void) |
5839 | { | |
5840 | gcc_assert (nop_pattern == NULL_RTX | |
5841 | && exit_insn == NULL_RTX); | |
5842 | ||
9ef1bf71 | 5843 | nop_pattern = constm1_rtx; |
e855c69d AB |
5844 | |
5845 | start_sequence (); | |
5846 | emit_insn (nop_pattern); | |
5847 | exit_insn = get_insns (); | |
5848 | end_sequence (); | |
fefa31b5 | 5849 | set_block_for_insn (exit_insn, EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e855c69d AB |
5850 | } |
5851 | ||
5852 | /* Free special insns used in the scheduler. */ | |
5853 | void | |
5854 | free_nop_and_exit_insns (void) | |
5855 | { | |
c5db5458 | 5856 | exit_insn = NULL; |
e855c69d AB |
5857 | nop_pattern = NULL_RTX; |
5858 | } | |
5859 | ||
5860 | /* Setup a special vinsn used in new insns initialization. */ | |
5861 | void | |
5862 | setup_nop_vinsn (void) | |
5863 | { | |
5864 | nop_vinsn = vinsn_create (exit_insn, false); | |
5865 | vinsn_attach (nop_vinsn); | |
5866 | } | |
5867 | ||
5868 | /* Free a special vinsn used in new insns initialization. */ | |
5869 | void | |
5870 | free_nop_vinsn (void) | |
5871 | { | |
5872 | gcc_assert (VINSN_COUNT (nop_vinsn) == 1); | |
5873 | vinsn_detach (nop_vinsn); | |
5874 | nop_vinsn = NULL; | |
5875 | } | |
5876 | ||
5877 | /* Call a set_sched_flags hook. */ | |
5878 | void | |
5879 | sel_set_sched_flags (void) | |
5880 | { | |
b8698a0f | 5881 | /* ??? This means that set_sched_flags were called, and we decided to |
e855c69d | 5882 | support speculation. However, set_sched_flags also modifies flags |
b8698a0f | 5883 | on current_sched_info, doing this only at global init. And we |
e855c69d AB |
5884 | sometimes change c_s_i later. So put the correct flags again. */ |
5885 | if (spec_info && targetm.sched.set_sched_flags) | |
5886 | targetm.sched.set_sched_flags (spec_info); | |
5887 | } | |
5888 | ||
5889 | /* Setup pointers to global sched info structures. */ | |
5890 | void | |
5891 | sel_setup_sched_infos (void) | |
5892 | { | |
5893 | rgn_setup_common_sched_info (); | |
5894 | ||
5895 | memcpy (&sel_common_sched_info, common_sched_info, | |
5896 | sizeof (sel_common_sched_info)); | |
5897 | ||
5898 | sel_common_sched_info.fix_recovery_cfg = NULL; | |
5899 | sel_common_sched_info.add_block = NULL; | |
5900 | sel_common_sched_info.estimate_number_of_insns | |
5901 | = sel_estimate_number_of_insns; | |
5902 | sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; | |
5903 | sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; | |
5904 | ||
5905 | common_sched_info = &sel_common_sched_info; | |
5906 | ||
5907 | current_sched_info = &sched_sel_haifa_sched_info; | |
b8698a0f | 5908 | current_sched_info->sched_max_insns_priority = |
e855c69d | 5909 | get_rgn_sched_max_insns_priority (); |
b8698a0f | 5910 | |
e855c69d AB |
5911 | sel_set_sched_flags (); |
5912 | } | |
5913 | \f | |
5914 | ||
5915 | /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, | |
5916 | *BB_ORD_INDEX after that is increased. */ | |
5917 | static void | |
5918 | sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) | |
5919 | { | |
5920 | RGN_NR_BLOCKS (rgn) += 1; | |
5921 | RGN_DONT_CALC_DEPS (rgn) = 0; | |
5922 | RGN_HAS_REAL_EBB (rgn) = 0; | |
5923 | CONTAINING_RGN (bb->index) = rgn; | |
5924 | BLOCK_TO_BB (bb->index) = *bb_ord_index; | |
5925 | rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; | |
5926 | (*bb_ord_index)++; | |
5927 | ||
5928 | /* FIXME: it is true only when not scheduling ebbs. */ | |
5929 | RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); | |
5930 | } | |
5931 | ||
5932 | /* Functions to support pipelining of outer loops. */ | |
5933 | ||
5934 | /* Creates a new empty region and returns it's number. */ | |
5935 | static int | |
5936 | sel_create_new_region (void) | |
5937 | { | |
5938 | int new_rgn_number = nr_regions; | |
5939 | ||
5940 | RGN_NR_BLOCKS (new_rgn_number) = 0; | |
5941 | ||
5942 | /* FIXME: This will work only when EBBs are not created. */ | |
5943 | if (new_rgn_number != 0) | |
b8698a0f | 5944 | RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + |
e855c69d AB |
5945 | RGN_NR_BLOCKS (new_rgn_number - 1); |
5946 | else | |
5947 | RGN_BLOCKS (new_rgn_number) = 0; | |
5948 | ||
5949 | /* Set the blocks of the next region so the other functions may | |
5950 | calculate the number of blocks in the region. */ | |
b8698a0f | 5951 | RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + |
e855c69d AB |
5952 | RGN_NR_BLOCKS (new_rgn_number); |
5953 | ||
5954 | nr_regions++; | |
5955 | ||
5956 | return new_rgn_number; | |
5957 | } | |
5958 | ||
5959 | /* If X has a smaller topological sort number than Y, returns -1; | |
5960 | if greater, returns 1. */ | |
5961 | static int | |
5962 | bb_top_order_comparator (const void *x, const void *y) | |
5963 | { | |
5964 | basic_block bb1 = *(const basic_block *) x; | |
5965 | basic_block bb2 = *(const basic_block *) y; | |
5966 | ||
b8698a0f L |
5967 | gcc_assert (bb1 == bb2 |
5968 | || rev_top_order_index[bb1->index] | |
e855c69d AB |
5969 | != rev_top_order_index[bb2->index]); |
5970 | ||
5971 | /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so | |
5972 | bbs with greater number should go earlier. */ | |
5973 | if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) | |
5974 | return -1; | |
5975 | else | |
5976 | return 1; | |
5977 | } | |
5978 | ||
b8698a0f | 5979 | /* Create a region for LOOP and return its number. If we don't want |
e855c69d AB |
5980 | to pipeline LOOP, return -1. */ |
5981 | static int | |
5982 | make_region_from_loop (struct loop *loop) | |
5983 | { | |
5984 | unsigned int i; | |
5985 | int new_rgn_number = -1; | |
5986 | struct loop *inner; | |
5987 | ||
5988 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5989 | int bb_ord_index = 0; | |
5990 | basic_block *loop_blocks; | |
5991 | basic_block preheader_block; | |
5992 | ||
b8698a0f | 5993 | if (loop->num_nodes |
e855c69d AB |
5994 | > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS)) |
5995 | return -1; | |
b8698a0f | 5996 | |
e855c69d AB |
5997 | /* Don't pipeline loops whose latch belongs to some of its inner loops. */ |
5998 | for (inner = loop->inner; inner; inner = inner->inner) | |
5999 | if (flow_bb_inside_loop_p (inner, loop->latch)) | |
6000 | return -1; | |
6001 | ||
6002 | loop->ninsns = num_loop_insns (loop); | |
6003 | if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS)) | |
6004 | return -1; | |
6005 | ||
6006 | loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); | |
6007 | ||
6008 | for (i = 0; i < loop->num_nodes; i++) | |
6009 | if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) | |
6010 | { | |
6011 | free (loop_blocks); | |
6012 | return -1; | |
6013 | } | |
6014 | ||
6015 | preheader_block = loop_preheader_edge (loop)->src; | |
6016 | gcc_assert (preheader_block); | |
6017 | gcc_assert (loop_blocks[0] == loop->header); | |
6018 | ||
6019 | new_rgn_number = sel_create_new_region (); | |
6020 | ||
6021 | sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number); | |
d7c028c0 | 6022 | bitmap_set_bit (bbs_in_loop_rgns, preheader_block->index); |
e855c69d AB |
6023 | |
6024 | for (i = 0; i < loop->num_nodes; i++) | |
6025 | { | |
6026 | /* Add only those blocks that haven't been scheduled in the inner loop. | |
6027 | The exception is the basic blocks with bookkeeping code - they should | |
b8698a0f | 6028 | be added to the region (and they actually don't belong to the loop |
e855c69d AB |
6029 | body, but to the region containing that loop body). */ |
6030 | ||
6031 | gcc_assert (new_rgn_number >= 0); | |
6032 | ||
d7c028c0 | 6033 | if (! bitmap_bit_p (bbs_in_loop_rgns, loop_blocks[i]->index)) |
e855c69d | 6034 | { |
b8698a0f | 6035 | sel_add_block_to_region (loop_blocks[i], &bb_ord_index, |
e855c69d | 6036 | new_rgn_number); |
d7c028c0 | 6037 | bitmap_set_bit (bbs_in_loop_rgns, loop_blocks[i]->index); |
e855c69d AB |
6038 | } |
6039 | } | |
6040 | ||
6041 | free (loop_blocks); | |
6042 | MARK_LOOP_FOR_PIPELINING (loop); | |
6043 | ||
6044 | return new_rgn_number; | |
6045 | } | |
6046 | ||
6047 | /* Create a new region from preheader blocks LOOP_BLOCKS. */ | |
6048 | void | |
9771b263 | 6049 | make_region_from_loop_preheader (vec<basic_block> *&loop_blocks) |
e855c69d AB |
6050 | { |
6051 | unsigned int i; | |
6052 | int new_rgn_number = -1; | |
6053 | basic_block bb; | |
6054 | ||
6055 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
6056 | int bb_ord_index = 0; | |
6057 | ||
6058 | new_rgn_number = sel_create_new_region (); | |
6059 | ||
9771b263 | 6060 | FOR_EACH_VEC_ELT (*loop_blocks, i, bb) |
e855c69d AB |
6061 | { |
6062 | gcc_assert (new_rgn_number >= 0); | |
6063 | ||
6064 | sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number); | |
6065 | } | |
6066 | ||
9771b263 | 6067 | vec_free (loop_blocks); |
e855c69d AB |
6068 | } |
6069 | ||
6070 | ||
6071 | /* Create region(s) from loop nest LOOP, such that inner loops will be | |
b8698a0f | 6072 | pipelined before outer loops. Returns true when a region for LOOP |
e855c69d AB |
6073 | is created. */ |
6074 | static bool | |
6075 | make_regions_from_loop_nest (struct loop *loop) | |
b8698a0f | 6076 | { |
e855c69d AB |
6077 | struct loop *cur_loop; |
6078 | int rgn_number; | |
6079 | ||
6080 | /* Traverse all inner nodes of the loop. */ | |
6081 | for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) | |
d7c028c0 | 6082 | if (! bitmap_bit_p (bbs_in_loop_rgns, cur_loop->header->index)) |
e855c69d AB |
6083 | return false; |
6084 | ||
6085 | /* At this moment all regular inner loops should have been pipelined. | |
6086 | Try to create a region from this loop. */ | |
6087 | rgn_number = make_region_from_loop (loop); | |
6088 | ||
6089 | if (rgn_number < 0) | |
6090 | return false; | |
6091 | ||
9771b263 | 6092 | loop_nests.safe_push (loop); |
e855c69d AB |
6093 | return true; |
6094 | } | |
6095 | ||
6096 | /* Initalize data structures needed. */ | |
6097 | void | |
6098 | sel_init_pipelining (void) | |
6099 | { | |
6100 | /* Collect loop information to be used in outer loops pipelining. */ | |
6101 | loop_optimizer_init (LOOPS_HAVE_PREHEADERS | |
6102 | | LOOPS_HAVE_FALLTHRU_PREHEADERS | |
6103 | | LOOPS_HAVE_RECORDED_EXITS | |
6104 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); | |
6105 | current_loop_nest = NULL; | |
6106 | ||
8b1c6fd7 | 6107 | bbs_in_loop_rgns = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
f61e445a | 6108 | bitmap_clear (bbs_in_loop_rgns); |
e855c69d AB |
6109 | |
6110 | recompute_rev_top_order (); | |
6111 | } | |
6112 | ||
6113 | /* Returns a struct loop for region RGN. */ | |
6114 | loop_p | |
6115 | get_loop_nest_for_rgn (unsigned int rgn) | |
6116 | { | |
6117 | /* Regions created with extend_rgns don't have corresponding loop nests, | |
6118 | because they don't represent loops. */ | |
9771b263 DN |
6119 | if (rgn < loop_nests.length ()) |
6120 | return loop_nests[rgn]; | |
e855c69d AB |
6121 | else |
6122 | return NULL; | |
6123 | } | |
6124 | ||
6125 | /* True when LOOP was included into pipelining regions. */ | |
6126 | bool | |
6127 | considered_for_pipelining_p (struct loop *loop) | |
6128 | { | |
6129 | if (loop_depth (loop) == 0) | |
6130 | return false; | |
6131 | ||
b8698a0f L |
6132 | /* Now, the loop could be too large or irreducible. Check whether its |
6133 | region is in LOOP_NESTS. | |
6134 | We determine the region number of LOOP as the region number of its | |
6135 | latch. We can't use header here, because this header could be | |
e855c69d AB |
6136 | just removed preheader and it will give us the wrong region number. |
6137 | Latch can't be used because it could be in the inner loop too. */ | |
8ec4d0ad | 6138 | if (LOOP_MARKED_FOR_PIPELINING_P (loop)) |
e855c69d AB |
6139 | { |
6140 | int rgn = CONTAINING_RGN (loop->latch->index); | |
6141 | ||
9771b263 | 6142 | gcc_assert ((unsigned) rgn < loop_nests.length ()); |
e855c69d AB |
6143 | return true; |
6144 | } | |
b8698a0f | 6145 | |
e855c69d AB |
6146 | return false; |
6147 | } | |
6148 | ||
b8698a0f | 6149 | /* Makes regions from the rest of the blocks, after loops are chosen |
e855c69d AB |
6150 | for pipelining. */ |
6151 | static void | |
6152 | make_regions_from_the_rest (void) | |
6153 | { | |
6154 | int cur_rgn_blocks; | |
6155 | int *loop_hdr; | |
6156 | int i; | |
6157 | ||
6158 | basic_block bb; | |
6159 | edge e; | |
6160 | edge_iterator ei; | |
6161 | int *degree; | |
e855c69d AB |
6162 | |
6163 | /* Index in rgn_bb_table where to start allocating new regions. */ | |
6164 | cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; | |
e855c69d | 6165 | |
b8698a0f | 6166 | /* Make regions from all the rest basic blocks - those that don't belong to |
e855c69d AB |
6167 | any loop or belong to irreducible loops. Prepare the data structures |
6168 | for extend_rgns. */ | |
6169 | ||
6170 | /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, | |
6171 | LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same | |
6172 | loop. */ | |
8b1c6fd7 DM |
6173 | loop_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
6174 | degree = XCNEWVEC (int, last_basic_block_for_fn (cfun)); | |
e855c69d AB |
6175 | |
6176 | ||
6177 | /* For each basic block that belongs to some loop assign the number | |
6178 | of innermost loop it belongs to. */ | |
8b1c6fd7 | 6179 | for (i = 0; i < last_basic_block_for_fn (cfun); i++) |
e855c69d AB |
6180 | loop_hdr[i] = -1; |
6181 | ||
11cd3bed | 6182 | FOR_EACH_BB_FN (bb, cfun) |
e855c69d | 6183 | { |
13dfd554 | 6184 | if (bb->loop_father && bb->loop_father->num != 0 |
e855c69d AB |
6185 | && !(bb->flags & BB_IRREDUCIBLE_LOOP)) |
6186 | loop_hdr[bb->index] = bb->loop_father->num; | |
6187 | } | |
6188 | ||
b8698a0f | 6189 | /* For each basic block degree is calculated as the number of incoming |
e855c69d AB |
6190 | edges, that are going out of bbs that are not yet scheduled. |
6191 | The basic blocks that are scheduled have degree value of zero. */ | |
11cd3bed | 6192 | FOR_EACH_BB_FN (bb, cfun) |
e855c69d AB |
6193 | { |
6194 | degree[bb->index] = 0; | |
6195 | ||
d7c028c0 | 6196 | if (!bitmap_bit_p (bbs_in_loop_rgns, bb->index)) |
e855c69d AB |
6197 | { |
6198 | FOR_EACH_EDGE (e, ei, bb->preds) | |
d7c028c0 | 6199 | if (!bitmap_bit_p (bbs_in_loop_rgns, e->src->index)) |
e855c69d AB |
6200 | degree[bb->index]++; |
6201 | } | |
6202 | else | |
6203 | degree[bb->index] = -1; | |
6204 | } | |
6205 | ||
6206 | extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); | |
6207 | ||
6208 | /* Any block that did not end up in a region is placed into a region | |
6209 | by itself. */ | |
11cd3bed | 6210 | FOR_EACH_BB_FN (bb, cfun) |
e855c69d AB |
6211 | if (degree[bb->index] >= 0) |
6212 | { | |
6213 | rgn_bb_table[cur_rgn_blocks] = bb->index; | |
6214 | RGN_NR_BLOCKS (nr_regions) = 1; | |
6215 | RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; | |
6216 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
6217 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
6218 | CONTAINING_RGN (bb->index) = nr_regions++; | |
6219 | BLOCK_TO_BB (bb->index) = 0; | |
6220 | } | |
6221 | ||
6222 | free (degree); | |
6223 | free (loop_hdr); | |
6224 | } | |
6225 | ||
6226 | /* Free data structures used in pipelining of loops. */ | |
6227 | void sel_finish_pipelining (void) | |
6228 | { | |
e855c69d AB |
6229 | struct loop *loop; |
6230 | ||
6231 | /* Release aux fields so we don't free them later by mistake. */ | |
f0bd40b1 | 6232 | FOR_EACH_LOOP (loop, 0) |
e855c69d AB |
6233 | loop->aux = NULL; |
6234 | ||
6235 | loop_optimizer_finalize (); | |
6236 | ||
9771b263 | 6237 | loop_nests.release (); |
e855c69d AB |
6238 | |
6239 | free (rev_top_order_index); | |
6240 | rev_top_order_index = NULL; | |
6241 | } | |
6242 | ||
b8698a0f | 6243 | /* This function replaces the find_rgns when |
e855c69d | 6244 | FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ |
b8698a0f | 6245 | void |
e855c69d AB |
6246 | sel_find_rgns (void) |
6247 | { | |
6248 | sel_init_pipelining (); | |
6249 | extend_regions (); | |
6250 | ||
6251 | if (current_loops) | |
6252 | { | |
6253 | loop_p loop; | |
e855c69d | 6254 | |
f0bd40b1 RB |
6255 | FOR_EACH_LOOP (loop, (flag_sel_sched_pipelining_outer_loops |
6256 | ? LI_FROM_INNERMOST | |
6257 | : LI_ONLY_INNERMOST)) | |
e855c69d AB |
6258 | make_regions_from_loop_nest (loop); |
6259 | } | |
6260 | ||
6261 | /* Make regions from all the rest basic blocks and schedule them. | |
b8698a0f | 6262 | These blocks include blocks that don't belong to any loop or belong |
e855c69d AB |
6263 | to irreducible loops. */ |
6264 | make_regions_from_the_rest (); | |
6265 | ||
6266 | /* We don't need bbs_in_loop_rgns anymore. */ | |
6267 | sbitmap_free (bbs_in_loop_rgns); | |
6268 | bbs_in_loop_rgns = NULL; | |
6269 | } | |
6270 | ||
ea4d630f AM |
6271 | /* Add the preheader blocks from previous loop to current region taking |
6272 | it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS. | |
e855c69d AB |
6273 | This function is only used with -fsel-sched-pipelining-outer-loops. */ |
6274 | void | |
ea4d630f | 6275 | sel_add_loop_preheaders (bb_vec_t *bbs) |
e855c69d AB |
6276 | { |
6277 | int i; | |
6278 | basic_block bb; | |
9771b263 | 6279 | vec<basic_block> *preheader_blocks |
e855c69d AB |
6280 | = LOOP_PREHEADER_BLOCKS (current_loop_nest); |
6281 | ||
9771b263 DN |
6282 | if (!preheader_blocks) |
6283 | return; | |
6284 | ||
6285 | for (i = 0; preheader_blocks->iterate (i, &bb); i++) | |
8ec4d0ad | 6286 | { |
9771b263 DN |
6287 | bbs->safe_push (bb); |
6288 | last_added_blocks.safe_push (bb); | |
e855c69d | 6289 | sel_add_bb (bb); |
8ec4d0ad | 6290 | } |
e855c69d | 6291 | |
9771b263 | 6292 | vec_free (preheader_blocks); |
e855c69d AB |
6293 | } |
6294 | ||
b8698a0f L |
6295 | /* While pipelining outer loops, returns TRUE if BB is a loop preheader. |
6296 | Please note that the function should also work when pipelining_p is | |
6297 | false, because it is used when deciding whether we should or should | |
e855c69d AB |
6298 | not reschedule pipelined code. */ |
6299 | bool | |
6300 | sel_is_loop_preheader_p (basic_block bb) | |
6301 | { | |
6302 | if (current_loop_nest) | |
6303 | { | |
6304 | struct loop *outer; | |
6305 | ||
6306 | if (preheader_removed) | |
6307 | return false; | |
6308 | ||
6309 | /* Preheader is the first block in the region. */ | |
6310 | if (BLOCK_TO_BB (bb->index) == 0) | |
6311 | return true; | |
6312 | ||
6313 | /* We used to find a preheader with the topological information. | |
6314 | Check that the above code is equivalent to what we did before. */ | |
6315 | ||
6316 | if (in_current_region_p (current_loop_nest->header)) | |
b8698a0f | 6317 | gcc_assert (!(BLOCK_TO_BB (bb->index) |
e855c69d AB |
6318 | < BLOCK_TO_BB (current_loop_nest->header->index))); |
6319 | ||
6320 | /* Support the situation when the latch block of outer loop | |
6321 | could be from here. */ | |
6322 | for (outer = loop_outer (current_loop_nest); | |
6323 | outer; | |
6324 | outer = loop_outer (outer)) | |
6325 | if (considered_for_pipelining_p (outer) && outer->latch == bb) | |
6326 | gcc_unreachable (); | |
6327 | } | |
6328 | ||
6329 | return false; | |
6330 | } | |
6331 | ||
753de8cf AM |
6332 | /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and |
6333 | can be removed, making the corresponding edge fallthrough (assuming that | |
6334 | all basic blocks between JUMP_BB and DEST_BB are empty). */ | |
6335 | static bool | |
6336 | bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb) | |
e855c69d | 6337 | { |
b4550bf7 AM |
6338 | if (!onlyjump_p (BB_END (jump_bb)) |
6339 | || tablejump_p (BB_END (jump_bb), NULL, NULL)) | |
e855c69d AB |
6340 | return false; |
6341 | ||
b8698a0f | 6342 | /* Several outgoing edges, abnormal edge or destination of jump is |
e855c69d AB |
6343 | not DEST_BB. */ |
6344 | if (EDGE_COUNT (jump_bb->succs) != 1 | |
753de8cf | 6345 | || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING) |
e855c69d AB |
6346 | || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) |
6347 | return false; | |
6348 | ||
6349 | /* If not anything of the upper. */ | |
6350 | return true; | |
6351 | } | |
6352 | ||
6353 | /* Removes the loop preheader from the current region and saves it in | |
b8698a0f | 6354 | PREHEADER_BLOCKS of the father loop, so they will be added later to |
e855c69d AB |
6355 | region that represents an outer loop. */ |
6356 | static void | |
6357 | sel_remove_loop_preheader (void) | |
6358 | { | |
6359 | int i, old_len; | |
6360 | int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
6361 | basic_block bb; | |
6362 | bool all_empty_p = true; | |
9771b263 | 6363 | vec<basic_block> *preheader_blocks |
e855c69d AB |
6364 | = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); |
6365 | ||
9771b263 DN |
6366 | vec_check_alloc (preheader_blocks, 0); |
6367 | ||
e855c69d | 6368 | gcc_assert (current_loop_nest); |
9771b263 | 6369 | old_len = preheader_blocks->length (); |
e855c69d AB |
6370 | |
6371 | /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ | |
6372 | for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) | |
6373 | { | |
06e28de2 | 6374 | bb = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
e855c69d | 6375 | |
b8698a0f | 6376 | /* If the basic block belongs to region, but doesn't belong to |
e855c69d AB |
6377 | corresponding loop, then it should be a preheader. */ |
6378 | if (sel_is_loop_preheader_p (bb)) | |
6379 | { | |
9771b263 | 6380 | preheader_blocks->safe_push (bb); |
e855c69d AB |
6381 | if (BB_END (bb) != bb_note (bb)) |
6382 | all_empty_p = false; | |
6383 | } | |
6384 | } | |
b8698a0f | 6385 | |
e855c69d | 6386 | /* Remove these blocks only after iterating over the whole region. */ |
9771b263 | 6387 | for (i = preheader_blocks->length () - 1; i >= old_len; i--) |
e855c69d | 6388 | { |
9771b263 | 6389 | bb = (*preheader_blocks)[i]; |
e855c69d AB |
6390 | sel_remove_bb (bb, false); |
6391 | } | |
6392 | ||
6393 | if (!considered_for_pipelining_p (loop_outer (current_loop_nest))) | |
6394 | { | |
6395 | if (!all_empty_p) | |
6396 | /* Immediately create new region from preheader. */ | |
9771b263 | 6397 | make_region_from_loop_preheader (preheader_blocks); |
e855c69d AB |
6398 | else |
6399 | { | |
6400 | /* If all preheader blocks are empty - dont create new empty region. | |
6401 | Instead, remove them completely. */ | |
9771b263 | 6402 | FOR_EACH_VEC_ELT (*preheader_blocks, i, bb) |
e855c69d AB |
6403 | { |
6404 | edge e; | |
6405 | edge_iterator ei; | |
6406 | basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; | |
6407 | ||
6408 | /* Redirect all incoming edges to next basic block. */ | |
6409 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) | |
6410 | { | |
6411 | if (! (e->flags & EDGE_FALLTHRU)) | |
6412 | redirect_edge_and_branch (e, bb->next_bb); | |
6413 | else | |
6414 | redirect_edge_succ (e, bb->next_bb); | |
6415 | } | |
6416 | gcc_assert (BB_NOTE_LIST (bb) == NULL); | |
6417 | delete_and_free_basic_block (bb); | |
6418 | ||
b8698a0f L |
6419 | /* Check if after deleting preheader there is a nonconditional |
6420 | jump in PREV_BB that leads to the next basic block NEXT_BB. | |
6421 | If it is so - delete this jump and clear data sets of its | |
e855c69d AB |
6422 | basic block if it becomes empty. */ |
6423 | if (next_bb->prev_bb == prev_bb | |
fefa31b5 | 6424 | && prev_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
753de8cf | 6425 | && bb_has_removable_jump_to_p (prev_bb, next_bb)) |
e855c69d AB |
6426 | { |
6427 | redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); | |
6428 | if (BB_END (prev_bb) == bb_note (prev_bb)) | |
6429 | free_data_sets (prev_bb); | |
6430 | } | |
00c4e97c AB |
6431 | |
6432 | set_immediate_dominator (CDI_DOMINATORS, next_bb, | |
6433 | recompute_dominator (CDI_DOMINATORS, | |
6434 | next_bb)); | |
e855c69d AB |
6435 | } |
6436 | } | |
9771b263 | 6437 | vec_free (preheader_blocks); |
e855c69d AB |
6438 | } |
6439 | else | |
6440 | /* Store preheader within the father's loop structure. */ | |
6441 | SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), | |
6442 | preheader_blocks); | |
6443 | } | |
68975683 | 6444 | |
e855c69d | 6445 | #endif |