]>
Commit | Line | Data |
---|---|---|
e855c69d | 1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
32e8bb8e | 2 | Copyright (C) 2006, 2007, 2008, 2009 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" | |
23 | #include "tm.h" | |
24 | #include "toplev.h" | |
25 | #include "rtl.h" | |
26 | #include "tm_p.h" | |
27 | #include "hard-reg-set.h" | |
28 | #include "regs.h" | |
29 | #include "function.h" | |
30 | #include "flags.h" | |
31 | #include "insn-config.h" | |
32 | #include "insn-attr.h" | |
33 | #include "except.h" | |
34 | #include "toplev.h" | |
35 | #include "recog.h" | |
36 | #include "params.h" | |
37 | #include "target.h" | |
38 | #include "timevar.h" | |
39 | #include "tree-pass.h" | |
40 | #include "sched-int.h" | |
41 | #include "ggc.h" | |
42 | #include "tree.h" | |
43 | #include "vec.h" | |
44 | #include "langhooks.h" | |
45 | #include "rtlhooks-def.h" | |
46 | ||
47 | #ifdef INSN_SCHEDULING | |
48 | #include "sel-sched-ir.h" | |
49 | /* We don't have to use it except for sel_print_insn. */ | |
50 | #include "sel-sched-dump.h" | |
51 | ||
52 | /* A vector holding bb info for whole scheduling pass. */ | |
53 | VEC(sel_global_bb_info_def, heap) *sel_global_bb_info = NULL; | |
54 | ||
55 | /* A vector holding bb info. */ | |
56 | VEC(sel_region_bb_info_def, heap) *sel_region_bb_info = NULL; | |
57 | ||
58 | /* A pool for allocating all lists. */ | |
59 | alloc_pool sched_lists_pool; | |
60 | ||
61 | /* This contains information about successors for compute_av_set. */ | |
62 | struct succs_info current_succs; | |
63 | ||
64 | /* Data structure to describe interaction with the generic scheduler utils. */ | |
65 | static struct common_sched_info_def sel_common_sched_info; | |
66 | ||
67 | /* The loop nest being pipelined. */ | |
68 | struct loop *current_loop_nest; | |
69 | ||
70 | /* LOOP_NESTS is a vector containing the corresponding loop nest for | |
71 | each region. */ | |
72 | static VEC(loop_p, heap) *loop_nests = NULL; | |
73 | ||
74 | /* Saves blocks already in loop regions, indexed by bb->index. */ | |
75 | static sbitmap bbs_in_loop_rgns = NULL; | |
76 | ||
77 | /* CFG hooks that are saved before changing create_basic_block hook. */ | |
78 | static struct cfg_hooks orig_cfg_hooks; | |
79 | \f | |
80 | ||
81 | /* Array containing reverse topological index of function basic blocks, | |
82 | indexed by BB->INDEX. */ | |
83 | static int *rev_top_order_index = NULL; | |
84 | ||
85 | /* Length of the above array. */ | |
86 | static int rev_top_order_index_len = -1; | |
87 | ||
88 | /* A regset pool structure. */ | |
89 | static struct | |
90 | { | |
91 | /* The stack to which regsets are returned. */ | |
92 | regset *v; | |
93 | ||
94 | /* Its pointer. */ | |
95 | int n; | |
96 | ||
97 | /* Its size. */ | |
98 | int s; | |
99 | ||
100 | /* In VV we save all generated regsets so that, when destructing the | |
101 | pool, we can compare it with V and check that every regset was returned | |
102 | back to pool. */ | |
103 | regset *vv; | |
104 | ||
105 | /* The pointer of VV stack. */ | |
106 | int nn; | |
107 | ||
108 | /* Its size. */ | |
109 | int ss; | |
110 | ||
111 | /* The difference between allocated and returned regsets. */ | |
112 | int diff; | |
113 | } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 }; | |
114 | ||
115 | /* This represents the nop pool. */ | |
116 | static struct | |
117 | { | |
118 | /* The vector which holds previously emitted nops. */ | |
119 | insn_t *v; | |
120 | ||
121 | /* Its pointer. */ | |
122 | int n; | |
123 | ||
124 | /* Its size. */ | |
b8698a0f | 125 | int s; |
e855c69d AB |
126 | } nop_pool = { NULL, 0, 0 }; |
127 | ||
128 | /* The pool for basic block notes. */ | |
129 | static rtx_vec_t bb_note_pool; | |
130 | ||
131 | /* A NOP pattern used to emit placeholder insns. */ | |
132 | rtx nop_pattern = NULL_RTX; | |
133 | /* A special instruction that resides in EXIT_BLOCK. | |
134 | EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */ | |
135 | rtx exit_insn = NULL_RTX; | |
136 | ||
b8698a0f | 137 | /* TRUE if while scheduling current region, which is loop, its preheader |
e855c69d AB |
138 | was removed. */ |
139 | bool preheader_removed = false; | |
140 | \f | |
141 | ||
142 | /* Forward static declarations. */ | |
143 | static void fence_clear (fence_t); | |
144 | ||
145 | static void deps_init_id (idata_t, insn_t, bool); | |
146 | static void init_id_from_df (idata_t, insn_t, bool); | |
147 | static expr_t set_insn_init (expr_t, vinsn_t, int); | |
148 | ||
149 | static void cfg_preds (basic_block, insn_t **, int *); | |
150 | static void prepare_insn_expr (insn_t, int); | |
151 | static void free_history_vect (VEC (expr_history_def, heap) **); | |
152 | ||
153 | static void move_bb_info (basic_block, basic_block); | |
154 | static void remove_empty_bb (basic_block, bool); | |
155 | static void sel_remove_loop_preheader (void); | |
156 | ||
157 | static bool insn_is_the_only_one_in_bb_p (insn_t); | |
158 | static void create_initial_data_sets (basic_block); | |
159 | ||
b5b8b0ac | 160 | static void free_av_set (basic_block); |
e855c69d AB |
161 | static void invalidate_av_set (basic_block); |
162 | static void extend_insn_data (void); | |
163 | static void sel_init_new_insn (insn_t, int); | |
164 | static void finish_insns (void); | |
165 | \f | |
166 | /* Various list functions. */ | |
167 | ||
168 | /* Copy an instruction list L. */ | |
169 | ilist_t | |
170 | ilist_copy (ilist_t l) | |
171 | { | |
172 | ilist_t head = NULL, *tailp = &head; | |
173 | ||
174 | while (l) | |
175 | { | |
176 | ilist_add (tailp, ILIST_INSN (l)); | |
177 | tailp = &ILIST_NEXT (*tailp); | |
178 | l = ILIST_NEXT (l); | |
179 | } | |
180 | ||
181 | return head; | |
182 | } | |
183 | ||
184 | /* Invert an instruction list L. */ | |
185 | ilist_t | |
186 | ilist_invert (ilist_t l) | |
187 | { | |
188 | ilist_t res = NULL; | |
189 | ||
190 | while (l) | |
191 | { | |
192 | ilist_add (&res, ILIST_INSN (l)); | |
193 | l = ILIST_NEXT (l); | |
194 | } | |
195 | ||
196 | return res; | |
197 | } | |
198 | ||
199 | /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */ | |
200 | void | |
201 | blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc) | |
202 | { | |
203 | bnd_t bnd; | |
204 | ||
205 | _list_add (lp); | |
206 | bnd = BLIST_BND (*lp); | |
207 | ||
208 | BND_TO (bnd) = to; | |
209 | BND_PTR (bnd) = ptr; | |
210 | BND_AV (bnd) = NULL; | |
211 | BND_AV1 (bnd) = NULL; | |
212 | BND_DC (bnd) = dc; | |
213 | } | |
214 | ||
215 | /* Remove the list note pointed to by LP. */ | |
216 | void | |
217 | blist_remove (blist_t *lp) | |
218 | { | |
219 | bnd_t b = BLIST_BND (*lp); | |
220 | ||
221 | av_set_clear (&BND_AV (b)); | |
222 | av_set_clear (&BND_AV1 (b)); | |
223 | ilist_clear (&BND_PTR (b)); | |
224 | ||
225 | _list_remove (lp); | |
226 | } | |
227 | ||
228 | /* Init a fence tail L. */ | |
229 | void | |
230 | flist_tail_init (flist_tail_t l) | |
231 | { | |
232 | FLIST_TAIL_HEAD (l) = NULL; | |
233 | FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l); | |
234 | } | |
235 | ||
236 | /* Try to find fence corresponding to INSN in L. */ | |
237 | fence_t | |
238 | flist_lookup (flist_t l, insn_t insn) | |
239 | { | |
240 | while (l) | |
241 | { | |
242 | if (FENCE_INSN (FLIST_FENCE (l)) == insn) | |
243 | return FLIST_FENCE (l); | |
244 | ||
245 | l = FLIST_NEXT (l); | |
246 | } | |
247 | ||
248 | return NULL; | |
249 | } | |
250 | ||
251 | /* Init the fields of F before running fill_insns. */ | |
252 | static void | |
253 | init_fence_for_scheduling (fence_t f) | |
254 | { | |
255 | FENCE_BNDS (f) = NULL; | |
256 | FENCE_PROCESSED_P (f) = false; | |
257 | FENCE_SCHEDULED_P (f) = false; | |
258 | } | |
259 | ||
260 | /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */ | |
261 | static void | |
b8698a0f L |
262 | flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, |
263 | insn_t last_scheduled_insn, VEC(rtx,gc) *executing_insns, | |
264 | int *ready_ticks, int ready_ticks_size, insn_t sched_next, | |
265 | int cycle, int cycle_issued_insns, | |
e855c69d AB |
266 | bool starts_cycle_p, bool after_stall_p) |
267 | { | |
268 | fence_t f; | |
269 | ||
270 | _list_add (lp); | |
271 | f = FLIST_FENCE (*lp); | |
272 | ||
273 | FENCE_INSN (f) = insn; | |
274 | ||
275 | gcc_assert (state != NULL); | |
276 | FENCE_STATE (f) = state; | |
277 | ||
278 | FENCE_CYCLE (f) = cycle; | |
279 | FENCE_ISSUED_INSNS (f) = cycle_issued_insns; | |
280 | FENCE_STARTS_CYCLE_P (f) = starts_cycle_p; | |
281 | FENCE_AFTER_STALL_P (f) = after_stall_p; | |
282 | ||
283 | gcc_assert (dc != NULL); | |
284 | FENCE_DC (f) = dc; | |
285 | ||
286 | gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL); | |
287 | FENCE_TC (f) = tc; | |
288 | ||
289 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
290 | FENCE_EXECUTING_INSNS (f) = executing_insns; | |
291 | FENCE_READY_TICKS (f) = ready_ticks; | |
292 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
293 | FENCE_SCHED_NEXT (f) = sched_next; | |
294 | ||
295 | init_fence_for_scheduling (f); | |
296 | } | |
297 | ||
298 | /* Remove the head node of the list pointed to by LP. */ | |
299 | static void | |
300 | flist_remove (flist_t *lp) | |
301 | { | |
302 | if (FENCE_INSN (FLIST_FENCE (*lp))) | |
303 | fence_clear (FLIST_FENCE (*lp)); | |
304 | _list_remove (lp); | |
305 | } | |
306 | ||
307 | /* Clear the fence list pointed to by LP. */ | |
308 | void | |
309 | flist_clear (flist_t *lp) | |
310 | { | |
311 | while (*lp) | |
312 | flist_remove (lp); | |
313 | } | |
314 | ||
315 | /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */ | |
316 | void | |
317 | def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call) | |
318 | { | |
319 | def_t d; | |
b8698a0f | 320 | |
e855c69d AB |
321 | _list_add (dl); |
322 | d = DEF_LIST_DEF (*dl); | |
323 | ||
324 | d->orig_insn = original_insn; | |
325 | d->crosses_call = crosses_call; | |
326 | } | |
327 | \f | |
328 | ||
329 | /* Functions to work with target contexts. */ | |
330 | ||
b8698a0f | 331 | /* Bulk target context. It is convenient for debugging purposes to ensure |
e855c69d AB |
332 | that there are no uninitialized (null) target contexts. */ |
333 | static tc_t bulk_tc = (tc_t) 1; | |
334 | ||
b8698a0f | 335 | /* Target hooks wrappers. In the future we can provide some default |
e855c69d AB |
336 | implementations for them. */ |
337 | ||
338 | /* Allocate a store for the target context. */ | |
339 | static tc_t | |
340 | alloc_target_context (void) | |
341 | { | |
342 | return (targetm.sched.alloc_sched_context | |
343 | ? targetm.sched.alloc_sched_context () : bulk_tc); | |
344 | } | |
345 | ||
346 | /* Init target context TC. | |
347 | If CLEAN_P is true, then make TC as it is beginning of the scheduler. | |
348 | Overwise, copy current backend context to TC. */ | |
349 | static void | |
350 | init_target_context (tc_t tc, bool clean_p) | |
351 | { | |
352 | if (targetm.sched.init_sched_context) | |
353 | targetm.sched.init_sched_context (tc, clean_p); | |
354 | } | |
355 | ||
356 | /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as | |
357 | int init_target_context (). */ | |
358 | tc_t | |
359 | create_target_context (bool clean_p) | |
360 | { | |
361 | tc_t tc = alloc_target_context (); | |
362 | ||
363 | init_target_context (tc, clean_p); | |
364 | return tc; | |
365 | } | |
366 | ||
367 | /* Copy TC to the current backend context. */ | |
368 | void | |
369 | set_target_context (tc_t tc) | |
370 | { | |
371 | if (targetm.sched.set_sched_context) | |
372 | targetm.sched.set_sched_context (tc); | |
373 | } | |
374 | ||
375 | /* TC is about to be destroyed. Free any internal data. */ | |
376 | static void | |
377 | clear_target_context (tc_t tc) | |
378 | { | |
379 | if (targetm.sched.clear_sched_context) | |
380 | targetm.sched.clear_sched_context (tc); | |
381 | } | |
382 | ||
383 | /* Clear and free it. */ | |
384 | static void | |
385 | delete_target_context (tc_t tc) | |
386 | { | |
387 | clear_target_context (tc); | |
388 | ||
389 | if (targetm.sched.free_sched_context) | |
390 | targetm.sched.free_sched_context (tc); | |
391 | } | |
392 | ||
393 | /* Make a copy of FROM in TO. | |
394 | NB: May be this should be a hook. */ | |
395 | static void | |
396 | copy_target_context (tc_t to, tc_t from) | |
397 | { | |
398 | tc_t tmp = create_target_context (false); | |
399 | ||
400 | set_target_context (from); | |
401 | init_target_context (to, false); | |
402 | ||
403 | set_target_context (tmp); | |
404 | delete_target_context (tmp); | |
405 | } | |
406 | ||
407 | /* Create a copy of TC. */ | |
408 | static tc_t | |
409 | create_copy_of_target_context (tc_t tc) | |
410 | { | |
411 | tc_t copy = alloc_target_context (); | |
412 | ||
413 | copy_target_context (copy, tc); | |
414 | ||
415 | return copy; | |
416 | } | |
417 | ||
418 | /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P | |
419 | is the same as in init_target_context (). */ | |
420 | void | |
421 | reset_target_context (tc_t tc, bool clean_p) | |
422 | { | |
423 | clear_target_context (tc); | |
424 | init_target_context (tc, clean_p); | |
425 | } | |
426 | \f | |
b8698a0f | 427 | /* Functions to work with dependence contexts. |
e855c69d AB |
428 | Dc (aka deps context, aka deps_t, aka struct deps *) is short for dependence |
429 | context. It accumulates information about processed insns to decide if | |
430 | current insn is dependent on the processed ones. */ | |
431 | ||
432 | /* Make a copy of FROM in TO. */ | |
433 | static void | |
434 | copy_deps_context (deps_t to, deps_t from) | |
435 | { | |
bcf33775 | 436 | init_deps (to, false); |
e855c69d AB |
437 | deps_join (to, from); |
438 | } | |
439 | ||
440 | /* Allocate store for dep context. */ | |
441 | static deps_t | |
442 | alloc_deps_context (void) | |
443 | { | |
444 | return XNEW (struct deps); | |
445 | } | |
446 | ||
447 | /* Allocate and initialize dep context. */ | |
448 | static deps_t | |
449 | create_deps_context (void) | |
450 | { | |
451 | deps_t dc = alloc_deps_context (); | |
452 | ||
bcf33775 | 453 | init_deps (dc, false); |
e855c69d AB |
454 | return dc; |
455 | } | |
456 | ||
457 | /* Create a copy of FROM. */ | |
458 | static deps_t | |
459 | create_copy_of_deps_context (deps_t from) | |
460 | { | |
461 | deps_t to = alloc_deps_context (); | |
462 | ||
463 | copy_deps_context (to, from); | |
464 | return to; | |
465 | } | |
466 | ||
467 | /* Clean up internal data of DC. */ | |
468 | static void | |
469 | clear_deps_context (deps_t dc) | |
470 | { | |
471 | free_deps (dc); | |
472 | } | |
473 | ||
474 | /* Clear and free DC. */ | |
475 | static void | |
476 | delete_deps_context (deps_t dc) | |
477 | { | |
478 | clear_deps_context (dc); | |
479 | free (dc); | |
480 | } | |
481 | ||
482 | /* Clear and init DC. */ | |
483 | static void | |
484 | reset_deps_context (deps_t dc) | |
485 | { | |
486 | clear_deps_context (dc); | |
bcf33775 | 487 | init_deps (dc, false); |
e855c69d AB |
488 | } |
489 | ||
b8698a0f | 490 | /* This structure describes the dependence analysis hooks for advancing |
e855c69d AB |
491 | dependence context. */ |
492 | static struct sched_deps_info_def advance_deps_context_sched_deps_info = | |
493 | { | |
494 | NULL, | |
495 | ||
496 | NULL, /* start_insn */ | |
497 | NULL, /* finish_insn */ | |
498 | NULL, /* start_lhs */ | |
499 | NULL, /* finish_lhs */ | |
500 | NULL, /* start_rhs */ | |
501 | NULL, /* finish_rhs */ | |
502 | haifa_note_reg_set, | |
503 | haifa_note_reg_clobber, | |
504 | haifa_note_reg_use, | |
505 | NULL, /* note_mem_dep */ | |
506 | NULL, /* note_dep */ | |
507 | ||
508 | 0, 0, 0 | |
509 | }; | |
510 | ||
511 | /* Process INSN and add its impact on DC. */ | |
512 | void | |
513 | advance_deps_context (deps_t dc, insn_t insn) | |
514 | { | |
515 | sched_deps_info = &advance_deps_context_sched_deps_info; | |
516 | deps_analyze_insn (dc, insn); | |
517 | } | |
518 | \f | |
519 | ||
520 | /* Functions to work with DFA states. */ | |
521 | ||
522 | /* Allocate store for a DFA state. */ | |
523 | static state_t | |
524 | state_alloc (void) | |
525 | { | |
526 | return xmalloc (dfa_state_size); | |
527 | } | |
528 | ||
529 | /* Allocate and initialize DFA state. */ | |
530 | static state_t | |
531 | state_create (void) | |
532 | { | |
533 | state_t state = state_alloc (); | |
534 | ||
535 | state_reset (state); | |
536 | advance_state (state); | |
537 | return state; | |
538 | } | |
539 | ||
540 | /* Free DFA state. */ | |
541 | static void | |
542 | state_free (state_t state) | |
543 | { | |
544 | free (state); | |
545 | } | |
546 | ||
547 | /* Make a copy of FROM in TO. */ | |
548 | static void | |
549 | state_copy (state_t to, state_t from) | |
550 | { | |
551 | memcpy (to, from, dfa_state_size); | |
552 | } | |
553 | ||
554 | /* Create a copy of FROM. */ | |
555 | static state_t | |
556 | state_create_copy (state_t from) | |
557 | { | |
558 | state_t to = state_alloc (); | |
559 | ||
560 | state_copy (to, from); | |
561 | return to; | |
562 | } | |
563 | \f | |
564 | ||
565 | /* Functions to work with fences. */ | |
566 | ||
567 | /* Clear the fence. */ | |
568 | static void | |
569 | fence_clear (fence_t f) | |
570 | { | |
571 | state_t s = FENCE_STATE (f); | |
572 | deps_t dc = FENCE_DC (f); | |
573 | void *tc = FENCE_TC (f); | |
574 | ||
575 | ilist_clear (&FENCE_BNDS (f)); | |
576 | ||
577 | gcc_assert ((s != NULL && dc != NULL && tc != NULL) | |
578 | || (s == NULL && dc == NULL && tc == NULL)); | |
579 | ||
580 | if (s != NULL) | |
581 | free (s); | |
582 | ||
583 | if (dc != NULL) | |
584 | delete_deps_context (dc); | |
585 | ||
586 | if (tc != NULL) | |
587 | delete_target_context (tc); | |
588 | VEC_free (rtx, gc, FENCE_EXECUTING_INSNS (f)); | |
589 | free (FENCE_READY_TICKS (f)); | |
590 | FENCE_READY_TICKS (f) = NULL; | |
591 | } | |
592 | ||
593 | /* Init a list of fences with successors of OLD_FENCE. */ | |
594 | void | |
595 | init_fences (insn_t old_fence) | |
596 | { | |
597 | insn_t succ; | |
598 | succ_iterator si; | |
599 | bool first = true; | |
600 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f L |
601 | |
602 | FOR_EACH_SUCC_1 (succ, si, old_fence, | |
e855c69d AB |
603 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
604 | { | |
b8698a0f | 605 | |
e855c69d AB |
606 | if (first) |
607 | first = false; | |
608 | else | |
609 | gcc_assert (flag_sel_sched_pipelining_outer_loops); | |
610 | ||
611 | flist_add (&fences, succ, | |
612 | state_create (), | |
613 | create_deps_context () /* dc */, | |
614 | create_target_context (true) /* tc */, | |
b8698a0f | 615 | NULL_RTX /* last_scheduled_insn */, |
e855c69d AB |
616 | NULL, /* executing_insns */ |
617 | XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ | |
618 | ready_ticks_size, | |
619 | NULL_RTX /* sched_next */, | |
b8698a0f L |
620 | 1 /* cycle */, 0 /* cycle_issued_insns */, |
621 | 1 /* starts_cycle_p */, 0 /* after_stall_p */); | |
e855c69d AB |
622 | } |
623 | } | |
624 | ||
625 | /* Merges two fences (filling fields of fence F with resulting values) by | |
626 | following rules: 1) state, target context and last scheduled insn are | |
b8698a0f | 627 | propagated from fallthrough edge if it is available; |
e855c69d | 628 | 2) deps context and cycle is propagated from more probable edge; |
b8698a0f | 629 | 3) all other fields are set to corresponding constant values. |
e855c69d | 630 | |
b8698a0f | 631 | INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, |
e855c69d AB |
632 | READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE and AFTER_STALL_P |
633 | are the corresponding fields of the second fence. */ | |
634 | static void | |
635 | merge_fences (fence_t f, insn_t insn, | |
b8698a0f | 636 | state_t state, deps_t dc, void *tc, |
e855c69d AB |
637 | rtx last_scheduled_insn, VEC(rtx, gc) *executing_insns, |
638 | int *ready_ticks, int ready_ticks_size, | |
639 | rtx sched_next, int cycle, bool after_stall_p) | |
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; | |
669 | VEC_free (rtx, gc, executing_insns); | |
670 | free (ready_ticks); | |
671 | if (FENCE_EXECUTING_INSNS (f)) | |
b8698a0f | 672 | VEC_block_remove (rtx, FENCE_EXECUTING_INSNS (f), 0, |
e855c69d AB |
673 | VEC_length (rtx, FENCE_EXECUTING_INSNS (f))); |
674 | if (FENCE_READY_TICKS (f)) | |
675 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
676 | } | |
677 | else | |
678 | { | |
679 | edge edge_old = NULL, edge_new = NULL; | |
680 | edge candidate; | |
681 | succ_iterator si; | |
682 | insn_t succ; | |
b8698a0f | 683 | |
e855c69d AB |
684 | /* Find fallthrough edge. */ |
685 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); | |
686 | candidate = find_fallthru_edge (BLOCK_FOR_INSN (insn)->prev_bb); | |
687 | ||
688 | if (!candidate | |
689 | || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn) | |
690 | && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old))) | |
691 | { | |
692 | /* No fallthrough edge leading to basic block of INSN. */ | |
693 | state_reset (FENCE_STATE (f)); | |
694 | state_free (state); | |
b8698a0f | 695 | |
e855c69d AB |
696 | reset_target_context (FENCE_TC (f), true); |
697 | delete_target_context (tc); | |
b8698a0f | 698 | |
e855c69d AB |
699 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; |
700 | } | |
701 | else | |
702 | if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn)) | |
703 | { | |
b8698a0f | 704 | /* Would be weird if same insn is successor of several fallthrough |
e855c69d AB |
705 | edges. */ |
706 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
707 | != BLOCK_FOR_INSN (last_scheduled_insn_old)); | |
708 | ||
709 | state_free (FENCE_STATE (f)); | |
710 | FENCE_STATE (f) = state; | |
711 | ||
712 | delete_target_context (FENCE_TC (f)); | |
713 | FENCE_TC (f) = tc; | |
714 | ||
715 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
716 | } | |
717 | else | |
718 | { | |
719 | /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ | |
720 | state_free (state); | |
721 | delete_target_context (tc); | |
722 | ||
723 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
724 | != BLOCK_FOR_INSN (last_scheduled_insn)); | |
725 | } | |
726 | ||
727 | /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ | |
728 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, | |
729 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
730 | { | |
731 | if (succ == insn) | |
732 | { | |
733 | /* No same successor allowed from several edges. */ | |
734 | gcc_assert (!edge_old); | |
735 | edge_old = si.e1; | |
736 | } | |
737 | } | |
738 | /* Find edge of second predecessor (last_scheduled_insn->insn). */ | |
739 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, | |
740 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
741 | { | |
742 | if (succ == insn) | |
743 | { | |
744 | /* No same successor allowed from several edges. */ | |
745 | gcc_assert (!edge_new); | |
746 | edge_new = si.e1; | |
747 | } | |
748 | } | |
749 | ||
750 | /* Check if we can choose most probable predecessor. */ | |
751 | if (edge_old == NULL || edge_new == NULL) | |
752 | { | |
753 | reset_deps_context (FENCE_DC (f)); | |
754 | delete_deps_context (dc); | |
755 | VEC_free (rtx, gc, executing_insns); | |
756 | free (ready_ticks); | |
b8698a0f | 757 | |
e855c69d AB |
758 | FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); |
759 | if (FENCE_EXECUTING_INSNS (f)) | |
b8698a0f | 760 | VEC_block_remove (rtx, FENCE_EXECUTING_INSNS (f), 0, |
e855c69d AB |
761 | VEC_length (rtx, FENCE_EXECUTING_INSNS (f))); |
762 | if (FENCE_READY_TICKS (f)) | |
763 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
764 | } | |
765 | else | |
766 | if (edge_new->probability > edge_old->probability) | |
767 | { | |
768 | delete_deps_context (FENCE_DC (f)); | |
769 | FENCE_DC (f) = dc; | |
770 | VEC_free (rtx, gc, FENCE_EXECUTING_INSNS (f)); | |
771 | FENCE_EXECUTING_INSNS (f) = executing_insns; | |
772 | free (FENCE_READY_TICKS (f)); | |
773 | FENCE_READY_TICKS (f) = ready_ticks; | |
774 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
775 | FENCE_CYCLE (f) = cycle; | |
776 | } | |
777 | else | |
778 | { | |
779 | /* Leave DC and CYCLE untouched. */ | |
780 | delete_deps_context (dc); | |
781 | VEC_free (rtx, gc, executing_insns); | |
782 | free (ready_ticks); | |
783 | } | |
784 | } | |
785 | ||
786 | /* Fill remaining invariant fields. */ | |
787 | if (after_stall_p) | |
788 | FENCE_AFTER_STALL_P (f) = 1; | |
789 | ||
790 | FENCE_ISSUED_INSNS (f) = 0; | |
791 | FENCE_STARTS_CYCLE_P (f) = 1; | |
792 | FENCE_SCHED_NEXT (f) = NULL; | |
793 | } | |
794 | ||
b8698a0f | 795 | /* Add a new fence to NEW_FENCES list, initializing it from all |
e855c69d AB |
796 | other parameters. */ |
797 | static void | |
798 | add_to_fences (flist_tail_t new_fences, insn_t insn, | |
b8698a0f L |
799 | state_t state, deps_t dc, void *tc, rtx last_scheduled_insn, |
800 | VEC(rtx, gc) *executing_insns, int *ready_ticks, | |
801 | int ready_ticks_size, rtx sched_next, int cycle, | |
e855c69d AB |
802 | int cycle_issued_insns, bool starts_cycle_p, bool after_stall_p) |
803 | { | |
804 | fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn); | |
805 | ||
806 | if (! f) | |
807 | { | |
808 | flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc, | |
b8698a0f | 809 | last_scheduled_insn, executing_insns, ready_ticks, |
e855c69d AB |
810 | ready_ticks_size, sched_next, cycle, cycle_issued_insns, |
811 | starts_cycle_p, after_stall_p); | |
812 | ||
813 | FLIST_TAIL_TAILP (new_fences) | |
814 | = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences)); | |
815 | } | |
816 | else | |
817 | { | |
b8698a0f L |
818 | merge_fences (f, insn, state, dc, tc, last_scheduled_insn, |
819 | executing_insns, ready_ticks, ready_ticks_size, | |
e855c69d AB |
820 | sched_next, cycle, after_stall_p); |
821 | } | |
822 | } | |
823 | ||
824 | /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */ | |
825 | void | |
826 | move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences) | |
827 | { | |
828 | fence_t f, old; | |
829 | flist_t *tailp = FLIST_TAIL_TAILP (new_fences); | |
830 | ||
831 | old = FLIST_FENCE (old_fences); | |
b8698a0f | 832 | f = flist_lookup (FLIST_TAIL_HEAD (new_fences), |
e855c69d AB |
833 | FENCE_INSN (FLIST_FENCE (old_fences))); |
834 | if (f) | |
835 | { | |
836 | merge_fences (f, old->insn, old->state, old->dc, old->tc, | |
837 | old->last_scheduled_insn, old->executing_insns, | |
838 | old->ready_ticks, old->ready_ticks_size, | |
b8698a0f | 839 | old->sched_next, old->cycle, |
e855c69d AB |
840 | old->after_stall_p); |
841 | } | |
842 | else | |
843 | { | |
844 | _list_add (tailp); | |
845 | FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp); | |
846 | *FLIST_FENCE (*tailp) = *old; | |
847 | init_fence_for_scheduling (FLIST_FENCE (*tailp)); | |
848 | } | |
849 | FENCE_INSN (old) = NULL; | |
850 | } | |
851 | ||
b8698a0f | 852 | /* Add a new fence to NEW_FENCES list and initialize most of its data |
e855c69d AB |
853 | as a clean one. */ |
854 | void | |
855 | add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
856 | { | |
857 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f | 858 | |
e855c69d AB |
859 | add_to_fences (new_fences, |
860 | succ, state_create (), create_deps_context (), | |
861 | create_target_context (true), | |
b8698a0f | 862 | NULL_RTX, NULL, |
e855c69d AB |
863 | XCNEWVEC (int, ready_ticks_size), ready_ticks_size, |
864 | NULL_RTX, FENCE_CYCLE (fence) + 1, | |
865 | 0, 1, FENCE_AFTER_STALL_P (fence)); | |
866 | } | |
867 | ||
b8698a0f | 868 | /* Add a new fence to NEW_FENCES list and initialize all of its data |
e855c69d AB |
869 | from FENCE and SUCC. */ |
870 | void | |
871 | add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
872 | { | |
b8698a0f | 873 | int * new_ready_ticks |
e855c69d | 874 | = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); |
b8698a0f | 875 | |
e855c69d AB |
876 | memcpy (new_ready_ticks, FENCE_READY_TICKS (fence), |
877 | FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); | |
878 | add_to_fences (new_fences, | |
879 | succ, state_create_copy (FENCE_STATE (fence)), | |
880 | create_copy_of_deps_context (FENCE_DC (fence)), | |
881 | create_copy_of_target_context (FENCE_TC (fence)), | |
b8698a0f | 882 | FENCE_LAST_SCHEDULED_INSN (fence), |
e855c69d AB |
883 | VEC_copy (rtx, gc, FENCE_EXECUTING_INSNS (fence)), |
884 | new_ready_ticks, | |
885 | FENCE_READY_TICKS_SIZE (fence), | |
886 | FENCE_SCHED_NEXT (fence), | |
887 | FENCE_CYCLE (fence), | |
888 | FENCE_ISSUED_INSNS (fence), | |
889 | FENCE_STARTS_CYCLE_P (fence), | |
890 | FENCE_AFTER_STALL_P (fence)); | |
891 | } | |
892 | \f | |
893 | ||
894 | /* Functions to work with regset and nop pools. */ | |
895 | ||
896 | /* Returns the new regset from pool. It might have some of the bits set | |
897 | from the previous usage. */ | |
898 | regset | |
899 | get_regset_from_pool (void) | |
900 | { | |
901 | regset rs; | |
902 | ||
903 | if (regset_pool.n != 0) | |
904 | rs = regset_pool.v[--regset_pool.n]; | |
905 | else | |
906 | /* We need to create the regset. */ | |
907 | { | |
908 | rs = ALLOC_REG_SET (®_obstack); | |
909 | ||
910 | if (regset_pool.nn == regset_pool.ss) | |
911 | regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, | |
912 | (regset_pool.ss = 2 * regset_pool.ss + 1)); | |
913 | regset_pool.vv[regset_pool.nn++] = rs; | |
914 | } | |
915 | ||
916 | regset_pool.diff++; | |
917 | ||
918 | return rs; | |
919 | } | |
920 | ||
921 | /* Same as above, but returns the empty regset. */ | |
922 | regset | |
923 | get_clear_regset_from_pool (void) | |
924 | { | |
925 | regset rs = get_regset_from_pool (); | |
926 | ||
927 | CLEAR_REG_SET (rs); | |
928 | return rs; | |
929 | } | |
930 | ||
931 | /* Return regset RS to the pool for future use. */ | |
932 | void | |
933 | return_regset_to_pool (regset rs) | |
934 | { | |
935 | regset_pool.diff--; | |
936 | ||
937 | if (regset_pool.n == regset_pool.s) | |
938 | regset_pool.v = XRESIZEVEC (regset, regset_pool.v, | |
939 | (regset_pool.s = 2 * regset_pool.s + 1)); | |
940 | regset_pool.v[regset_pool.n++] = rs; | |
941 | } | |
942 | ||
68ad446f | 943 | #ifdef ENABLE_CHECKING |
e855c69d AB |
944 | /* This is used as a qsort callback for sorting regset pool stacks. |
945 | X and XX are addresses of two regsets. They are never equal. */ | |
946 | static int | |
947 | cmp_v_in_regset_pool (const void *x, const void *xx) | |
948 | { | |
949 | return *((const regset *) x) - *((const regset *) xx); | |
950 | } | |
68ad446f | 951 | #endif |
e855c69d AB |
952 | |
953 | /* Free the regset pool possibly checking for memory leaks. */ | |
954 | void | |
955 | free_regset_pool (void) | |
956 | { | |
957 | #ifdef ENABLE_CHECKING | |
958 | { | |
959 | regset *v = regset_pool.v; | |
960 | int i = 0; | |
961 | int n = regset_pool.n; | |
b8698a0f | 962 | |
e855c69d AB |
963 | regset *vv = regset_pool.vv; |
964 | int ii = 0; | |
965 | int nn = regset_pool.nn; | |
b8698a0f | 966 | |
e855c69d | 967 | int diff = 0; |
b8698a0f | 968 | |
e855c69d | 969 | gcc_assert (n <= nn); |
b8698a0f | 970 | |
e855c69d AB |
971 | /* Sort both vectors so it will be possible to compare them. */ |
972 | qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); | |
973 | qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); | |
b8698a0f | 974 | |
e855c69d AB |
975 | while (ii < nn) |
976 | { | |
977 | if (v[i] == vv[ii]) | |
978 | i++; | |
979 | else | |
980 | /* VV[II] was lost. */ | |
981 | diff++; | |
b8698a0f | 982 | |
e855c69d AB |
983 | ii++; |
984 | } | |
b8698a0f | 985 | |
e855c69d AB |
986 | gcc_assert (diff == regset_pool.diff); |
987 | } | |
988 | #endif | |
b8698a0f | 989 | |
e855c69d AB |
990 | /* If not true - we have a memory leak. */ |
991 | gcc_assert (regset_pool.diff == 0); | |
b8698a0f | 992 | |
e855c69d AB |
993 | while (regset_pool.n) |
994 | { | |
995 | --regset_pool.n; | |
996 | FREE_REG_SET (regset_pool.v[regset_pool.n]); | |
997 | } | |
998 | ||
999 | free (regset_pool.v); | |
1000 | regset_pool.v = NULL; | |
1001 | regset_pool.s = 0; | |
b8698a0f | 1002 | |
e855c69d AB |
1003 | free (regset_pool.vv); |
1004 | regset_pool.vv = NULL; | |
1005 | regset_pool.nn = 0; | |
1006 | regset_pool.ss = 0; | |
1007 | ||
1008 | regset_pool.diff = 0; | |
1009 | } | |
1010 | \f | |
1011 | ||
b8698a0f L |
1012 | /* Functions to work with nop pools. NOP insns are used as temporary |
1013 | placeholders of the insns being scheduled to allow correct update of | |
e855c69d AB |
1014 | the data sets. When update is finished, NOPs are deleted. */ |
1015 | ||
1016 | /* A vinsn that is used to represent a nop. This vinsn is shared among all | |
1017 | nops sel-sched generates. */ | |
1018 | static vinsn_t nop_vinsn = NULL; | |
1019 | ||
1020 | /* Emit a nop before INSN, taking it from pool. */ | |
1021 | insn_t | |
1022 | get_nop_from_pool (insn_t insn) | |
1023 | { | |
1024 | insn_t nop; | |
1025 | bool old_p = nop_pool.n != 0; | |
1026 | int flags; | |
1027 | ||
1028 | if (old_p) | |
1029 | nop = nop_pool.v[--nop_pool.n]; | |
1030 | else | |
1031 | nop = nop_pattern; | |
1032 | ||
1033 | nop = emit_insn_before (nop, insn); | |
1034 | ||
1035 | if (old_p) | |
1036 | flags = INSN_INIT_TODO_SSID; | |
1037 | else | |
1038 | flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; | |
1039 | ||
1040 | set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); | |
1041 | sel_init_new_insn (nop, flags); | |
1042 | ||
1043 | return nop; | |
1044 | } | |
1045 | ||
1046 | /* Remove NOP from the instruction stream and return it to the pool. */ | |
1047 | void | |
b5b8b0ac | 1048 | return_nop_to_pool (insn_t nop, bool full_tidying) |
e855c69d AB |
1049 | { |
1050 | gcc_assert (INSN_IN_STREAM_P (nop)); | |
b5b8b0ac | 1051 | sel_remove_insn (nop, false, full_tidying); |
e855c69d AB |
1052 | |
1053 | if (nop_pool.n == nop_pool.s) | |
b8698a0f | 1054 | nop_pool.v = XRESIZEVEC (rtx, nop_pool.v, |
e855c69d AB |
1055 | (nop_pool.s = 2 * nop_pool.s + 1)); |
1056 | nop_pool.v[nop_pool.n++] = nop; | |
1057 | } | |
1058 | ||
1059 | /* Free the nop pool. */ | |
1060 | void | |
1061 | free_nop_pool (void) | |
1062 | { | |
1063 | nop_pool.n = 0; | |
1064 | nop_pool.s = 0; | |
1065 | free (nop_pool.v); | |
1066 | nop_pool.v = NULL; | |
1067 | } | |
1068 | \f | |
1069 | ||
b8698a0f | 1070 | /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb. |
e855c69d AB |
1071 | The callback is given two rtxes XX and YY and writes the new rtxes |
1072 | to NX and NY in case some needs to be skipped. */ | |
1073 | static int | |
1074 | skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) | |
1075 | { | |
1076 | const_rtx x = *xx; | |
1077 | const_rtx y = *yy; | |
b8698a0f | 1078 | |
e855c69d AB |
1079 | if (GET_CODE (x) == UNSPEC |
1080 | && (targetm.sched.skip_rtx_p == NULL | |
1081 | || targetm.sched.skip_rtx_p (x))) | |
1082 | { | |
1083 | *nx = XVECEXP (x, 0, 0); | |
1084 | *ny = CONST_CAST_RTX (y); | |
1085 | return 1; | |
1086 | } | |
b8698a0f | 1087 | |
e855c69d AB |
1088 | if (GET_CODE (y) == UNSPEC |
1089 | && (targetm.sched.skip_rtx_p == NULL | |
1090 | || targetm.sched.skip_rtx_p (y))) | |
1091 | { | |
1092 | *nx = CONST_CAST_RTX (x); | |
1093 | *ny = XVECEXP (y, 0, 0); | |
1094 | return 1; | |
1095 | } | |
b8698a0f | 1096 | |
e855c69d AB |
1097 | return 0; |
1098 | } | |
1099 | ||
b8698a0f | 1100 | /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way |
e855c69d AB |
1101 | to support ia64 speculation. When changes are needed, new rtx X and new mode |
1102 | NMODE are written, and the callback returns true. */ | |
1103 | static int | |
1104 | hash_with_unspec_callback (const_rtx x, enum machine_mode mode ATTRIBUTE_UNUSED, | |
1105 | rtx *nx, enum machine_mode* nmode) | |
1106 | { | |
b8698a0f | 1107 | if (GET_CODE (x) == UNSPEC |
e855c69d AB |
1108 | && targetm.sched.skip_rtx_p |
1109 | && targetm.sched.skip_rtx_p (x)) | |
1110 | { | |
1111 | *nx = XVECEXP (x, 0 ,0); | |
32e8bb8e | 1112 | *nmode = VOIDmode; |
e855c69d AB |
1113 | return 1; |
1114 | } | |
b8698a0f | 1115 | |
e855c69d AB |
1116 | return 0; |
1117 | } | |
1118 | ||
1119 | /* Returns LHS and RHS are ok to be scheduled separately. */ | |
1120 | static bool | |
1121 | lhs_and_rhs_separable_p (rtx lhs, rtx rhs) | |
1122 | { | |
1123 | if (lhs == NULL || rhs == NULL) | |
1124 | return false; | |
1125 | ||
b8698a0f L |
1126 | /* Do not schedule CONST, CONST_INT and CONST_DOUBLE etc as rhs: no point |
1127 | to use reg, if const can be used. Moreover, scheduling const as rhs may | |
1128 | lead to mode mismatch cause consts don't have modes but they could be | |
e855c69d AB |
1129 | merged from branches where the same const used in different modes. */ |
1130 | if (CONSTANT_P (rhs)) | |
1131 | return false; | |
1132 | ||
1133 | /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ | |
1134 | if (COMPARISON_P (rhs)) | |
1135 | return false; | |
1136 | ||
1137 | /* Do not allow single REG to be an rhs. */ | |
1138 | if (REG_P (rhs)) | |
1139 | return false; | |
1140 | ||
b8698a0f | 1141 | /* See comment at find_used_regs_1 (*1) for explanation of this |
e855c69d AB |
1142 | restriction. */ |
1143 | /* FIXME: remove this later. */ | |
1144 | if (MEM_P (lhs)) | |
1145 | return false; | |
1146 | ||
1147 | /* This will filter all tricky things like ZERO_EXTRACT etc. | |
1148 | For now we don't handle it. */ | |
1149 | if (!REG_P (lhs) && !MEM_P (lhs)) | |
1150 | return false; | |
1151 | ||
1152 | return true; | |
1153 | } | |
1154 | ||
b8698a0f L |
1155 | /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When |
1156 | FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is | |
e855c69d AB |
1157 | used e.g. for insns from recovery blocks. */ |
1158 | static void | |
1159 | vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) | |
1160 | { | |
1161 | hash_rtx_callback_function hrcf; | |
1162 | int insn_class; | |
1163 | ||
1164 | VINSN_INSN_RTX (vi) = insn; | |
1165 | VINSN_COUNT (vi) = 0; | |
1166 | vi->cost = -1; | |
b8698a0f | 1167 | |
e855c69d AB |
1168 | if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) |
1169 | init_id_from_df (VINSN_ID (vi), insn, force_unique_p); | |
1170 | else | |
1171 | deps_init_id (VINSN_ID (vi), insn, force_unique_p); | |
b8698a0f | 1172 | |
e855c69d AB |
1173 | /* Hash vinsn depending on whether it is separable or not. */ |
1174 | hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; | |
1175 | if (VINSN_SEPARABLE_P (vi)) | |
1176 | { | |
1177 | rtx rhs = VINSN_RHS (vi); | |
1178 | ||
1179 | VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs), | |
1180 | NULL, NULL, false, hrcf); | |
1181 | VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi), | |
1182 | VOIDmode, NULL, NULL, | |
1183 | false, hrcf); | |
1184 | } | |
1185 | else | |
1186 | { | |
1187 | VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode, | |
1188 | NULL, NULL, false, hrcf); | |
1189 | VINSN_HASH_RTX (vi) = VINSN_HASH (vi); | |
1190 | } | |
b8698a0f | 1191 | |
e855c69d AB |
1192 | insn_class = haifa_classify_insn (insn); |
1193 | if (insn_class >= 2 | |
1194 | && (!targetm.sched.get_insn_spec_ds | |
1195 | || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) | |
1196 | == 0))) | |
1197 | VINSN_MAY_TRAP_P (vi) = true; | |
1198 | else | |
1199 | VINSN_MAY_TRAP_P (vi) = false; | |
1200 | } | |
1201 | ||
1202 | /* Indicate that VI has become the part of an rtx object. */ | |
1203 | void | |
1204 | vinsn_attach (vinsn_t vi) | |
1205 | { | |
1206 | /* Assert that VI is not pending for deletion. */ | |
1207 | gcc_assert (VINSN_INSN_RTX (vi)); | |
1208 | ||
1209 | VINSN_COUNT (vi)++; | |
1210 | } | |
1211 | ||
b8698a0f | 1212 | /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct |
e855c69d AB |
1213 | VINSN_TYPE (VI). */ |
1214 | static vinsn_t | |
1215 | vinsn_create (insn_t insn, bool force_unique_p) | |
1216 | { | |
1217 | vinsn_t vi = XCNEW (struct vinsn_def); | |
1218 | ||
1219 | vinsn_init (vi, insn, force_unique_p); | |
1220 | return vi; | |
1221 | } | |
1222 | ||
1223 | /* Return a copy of VI. When REATTACH_P is true, detach VI and attach | |
1224 | the copy. */ | |
b8698a0f | 1225 | vinsn_t |
e855c69d AB |
1226 | vinsn_copy (vinsn_t vi, bool reattach_p) |
1227 | { | |
1228 | rtx copy; | |
1229 | bool unique = VINSN_UNIQUE_P (vi); | |
1230 | vinsn_t new_vi; | |
b8698a0f | 1231 | |
e855c69d AB |
1232 | copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); |
1233 | new_vi = create_vinsn_from_insn_rtx (copy, unique); | |
1234 | if (reattach_p) | |
1235 | { | |
1236 | vinsn_detach (vi); | |
1237 | vinsn_attach (new_vi); | |
1238 | } | |
1239 | ||
1240 | return new_vi; | |
1241 | } | |
1242 | ||
1243 | /* Delete the VI vinsn and free its data. */ | |
1244 | static void | |
1245 | vinsn_delete (vinsn_t vi) | |
1246 | { | |
1247 | gcc_assert (VINSN_COUNT (vi) == 0); | |
1248 | ||
1249 | return_regset_to_pool (VINSN_REG_SETS (vi)); | |
1250 | return_regset_to_pool (VINSN_REG_USES (vi)); | |
1251 | return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); | |
1252 | ||
1253 | free (vi); | |
1254 | } | |
1255 | ||
b8698a0f | 1256 | /* Indicate that VI is no longer a part of some rtx object. |
e855c69d AB |
1257 | Remove VI if it is no longer needed. */ |
1258 | void | |
1259 | vinsn_detach (vinsn_t vi) | |
1260 | { | |
1261 | gcc_assert (VINSN_COUNT (vi) > 0); | |
1262 | ||
1263 | if (--VINSN_COUNT (vi) == 0) | |
1264 | vinsn_delete (vi); | |
1265 | } | |
1266 | ||
1267 | /* Returns TRUE if VI is a branch. */ | |
1268 | bool | |
1269 | vinsn_cond_branch_p (vinsn_t vi) | |
1270 | { | |
1271 | insn_t insn; | |
1272 | ||
1273 | if (!VINSN_UNIQUE_P (vi)) | |
1274 | return false; | |
1275 | ||
1276 | insn = VINSN_INSN_RTX (vi); | |
1277 | if (BB_END (BLOCK_FOR_INSN (insn)) != insn) | |
1278 | return false; | |
1279 | ||
1280 | return control_flow_insn_p (insn); | |
1281 | } | |
1282 | ||
1283 | /* Return latency of INSN. */ | |
1284 | static int | |
1285 | sel_insn_rtx_cost (rtx insn) | |
1286 | { | |
1287 | int cost; | |
1288 | ||
1289 | /* A USE insn, or something else we don't need to | |
1290 | understand. We can't pass these directly to | |
1291 | result_ready_cost or insn_default_latency because it will | |
1292 | trigger a fatal error for unrecognizable insns. */ | |
1293 | if (recog_memoized (insn) < 0) | |
1294 | cost = 0; | |
1295 | else | |
1296 | { | |
1297 | cost = insn_default_latency (insn); | |
1298 | ||
1299 | if (cost < 0) | |
1300 | cost = 0; | |
1301 | } | |
1302 | ||
1303 | return cost; | |
1304 | } | |
1305 | ||
1306 | /* Return the cost of the VI. | |
1307 | !!! FIXME: Unify with haifa-sched.c: insn_cost (). */ | |
1308 | int | |
1309 | sel_vinsn_cost (vinsn_t vi) | |
1310 | { | |
1311 | int cost = vi->cost; | |
1312 | ||
1313 | if (cost < 0) | |
1314 | { | |
1315 | cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); | |
1316 | vi->cost = cost; | |
1317 | } | |
1318 | ||
1319 | return cost; | |
1320 | } | |
1321 | \f | |
1322 | ||
1323 | /* Functions for insn emitting. */ | |
1324 | ||
1325 | /* Emit new insn after AFTER based on PATTERN and initialize its data from | |
1326 | EXPR and SEQNO. */ | |
1327 | insn_t | |
1328 | sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) | |
1329 | { | |
1330 | insn_t new_insn; | |
1331 | ||
1332 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); | |
1333 | ||
1334 | new_insn = emit_insn_after (pattern, after); | |
1335 | set_insn_init (expr, NULL, seqno); | |
1336 | sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); | |
1337 | ||
1338 | return new_insn; | |
1339 | } | |
1340 | ||
1341 | /* Force newly generated vinsns to be unique. */ | |
1342 | static bool init_insn_force_unique_p = false; | |
1343 | ||
1344 | /* Emit new speculation recovery insn after AFTER based on PATTERN and | |
1345 | initialize its data from EXPR and SEQNO. */ | |
1346 | insn_t | |
1347 | sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, | |
1348 | insn_t after) | |
1349 | { | |
1350 | insn_t insn; | |
1351 | ||
1352 | gcc_assert (!init_insn_force_unique_p); | |
1353 | ||
1354 | init_insn_force_unique_p = true; | |
1355 | insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); | |
1356 | CANT_MOVE (insn) = 1; | |
1357 | init_insn_force_unique_p = false; | |
1358 | ||
1359 | return insn; | |
1360 | } | |
1361 | ||
1362 | /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, | |
b8698a0f L |
1363 | take it as a new vinsn instead of EXPR's vinsn. |
1364 | We simplify insns later, after scheduling region in | |
e855c69d AB |
1365 | simplify_changed_insns. */ |
1366 | insn_t | |
b8698a0f | 1367 | sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
e855c69d AB |
1368 | insn_t after) |
1369 | { | |
1370 | expr_t emit_expr; | |
1371 | insn_t insn; | |
1372 | int flags; | |
b8698a0f L |
1373 | |
1374 | emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), | |
e855c69d AB |
1375 | seqno); |
1376 | insn = EXPR_INSN_RTX (emit_expr); | |
b8698a0f | 1377 | add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); |
e855c69d AB |
1378 | |
1379 | flags = INSN_INIT_TODO_SSID; | |
1380 | if (INSN_LUID (insn) == 0) | |
1381 | flags |= INSN_INIT_TODO_LUID; | |
1382 | sel_init_new_insn (insn, flags); | |
1383 | ||
1384 | return insn; | |
1385 | } | |
1386 | ||
1387 | /* Move insn from EXPR after AFTER. */ | |
1388 | insn_t | |
1389 | sel_move_insn (expr_t expr, int seqno, insn_t after) | |
1390 | { | |
1391 | insn_t insn = EXPR_INSN_RTX (expr); | |
1392 | basic_block bb = BLOCK_FOR_INSN (after); | |
1393 | insn_t next = NEXT_INSN (after); | |
1394 | ||
1395 | /* Assert that in move_op we disconnected this insn properly. */ | |
1396 | gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); | |
1397 | PREV_INSN (insn) = after; | |
1398 | NEXT_INSN (insn) = next; | |
1399 | ||
1400 | NEXT_INSN (after) = insn; | |
1401 | PREV_INSN (next) = insn; | |
1402 | ||
1403 | /* Update links from insn to bb and vice versa. */ | |
1404 | df_insn_change_bb (insn, bb); | |
1405 | if (BB_END (bb) == after) | |
1406 | BB_END (bb) = insn; | |
b8698a0f | 1407 | |
e855c69d AB |
1408 | prepare_insn_expr (insn, seqno); |
1409 | return insn; | |
1410 | } | |
1411 | ||
1412 | \f | |
1413 | /* Functions to work with right-hand sides. */ | |
1414 | ||
b8698a0f | 1415 | /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector |
e855c69d | 1416 | VECT and return true when found. Use NEW_VINSN for comparison only when |
b8698a0f L |
1417 | COMPARE_VINSNS is true. Write to INDP the index on which |
1418 | the search has stopped, such that inserting the new element at INDP will | |
e855c69d AB |
1419 | retain VECT's sort order. */ |
1420 | static bool | |
b8698a0f L |
1421 | find_in_history_vect_1 (VEC(expr_history_def, heap) *vect, |
1422 | unsigned uid, vinsn_t new_vinsn, | |
e855c69d AB |
1423 | bool compare_vinsns, int *indp) |
1424 | { | |
1425 | expr_history_def *arr; | |
1426 | int i, j, len = VEC_length (expr_history_def, vect); | |
1427 | ||
1428 | if (len == 0) | |
1429 | { | |
1430 | *indp = 0; | |
1431 | return false; | |
1432 | } | |
1433 | ||
1434 | arr = VEC_address (expr_history_def, vect); | |
1435 | i = 0, j = len - 1; | |
1436 | ||
1437 | while (i <= j) | |
1438 | { | |
1439 | unsigned auid = arr[i].uid; | |
b8698a0f | 1440 | vinsn_t avinsn = arr[i].new_expr_vinsn; |
e855c69d AB |
1441 | |
1442 | if (auid == uid | |
b8698a0f L |
1443 | /* When undoing transformation on a bookkeeping copy, the new vinsn |
1444 | may not be exactly equal to the one that is saved in the vector. | |
e855c69d AB |
1445 | This is because the insn whose copy we're checking was possibly |
1446 | substituted itself. */ | |
b8698a0f | 1447 | && (! compare_vinsns |
e855c69d AB |
1448 | || vinsn_equal_p (avinsn, new_vinsn))) |
1449 | { | |
1450 | *indp = i; | |
1451 | return true; | |
1452 | } | |
1453 | else if (auid > uid) | |
1454 | break; | |
1455 | i++; | |
1456 | } | |
1457 | ||
1458 | *indp = i; | |
1459 | return false; | |
1460 | } | |
1461 | ||
b8698a0f L |
1462 | /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return |
1463 | the position found or -1, if no such value is in vector. | |
e855c69d AB |
1464 | Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ |
1465 | int | |
b8698a0f | 1466 | find_in_history_vect (VEC(expr_history_def, heap) *vect, rtx insn, |
e855c69d AB |
1467 | vinsn_t new_vinsn, bool originators_p) |
1468 | { | |
1469 | int ind; | |
1470 | ||
b8698a0f | 1471 | if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn, |
e855c69d AB |
1472 | false, &ind)) |
1473 | return ind; | |
1474 | ||
1475 | if (INSN_ORIGINATORS (insn) && originators_p) | |
1476 | { | |
1477 | unsigned uid; | |
1478 | bitmap_iterator bi; | |
1479 | ||
1480 | EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) | |
1481 | if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind)) | |
1482 | return ind; | |
1483 | } | |
b8698a0f | 1484 | |
e855c69d AB |
1485 | return -1; |
1486 | } | |
1487 | ||
b8698a0f L |
1488 | /* Insert new element in a sorted history vector pointed to by PVECT, |
1489 | if it is not there already. The element is searched using | |
e855c69d AB |
1490 | UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save |
1491 | the history of a transformation. */ | |
1492 | void | |
1493 | insert_in_history_vect (VEC (expr_history_def, heap) **pvect, | |
1494 | unsigned uid, enum local_trans_type type, | |
b8698a0f | 1495 | vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, |
e855c69d AB |
1496 | ds_t spec_ds) |
1497 | { | |
1498 | VEC(expr_history_def, heap) *vect = *pvect; | |
1499 | expr_history_def temp; | |
1500 | bool res; | |
1501 | int ind; | |
1502 | ||
1503 | res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind); | |
1504 | ||
1505 | if (res) | |
1506 | { | |
1507 | expr_history_def *phist = VEC_index (expr_history_def, vect, ind); | |
1508 | ||
b8698a0f | 1509 | /* It is possible that speculation types of expressions that were |
e855c69d AB |
1510 | propagated through different paths will be different here. In this |
1511 | case, merge the status to get the correct check later. */ | |
1512 | if (phist->spec_ds != spec_ds) | |
1513 | phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); | |
1514 | return; | |
1515 | } | |
b8698a0f | 1516 | |
e855c69d AB |
1517 | temp.uid = uid; |
1518 | temp.old_expr_vinsn = old_expr_vinsn; | |
b8698a0f | 1519 | temp.new_expr_vinsn = new_expr_vinsn; |
e855c69d AB |
1520 | temp.spec_ds = spec_ds; |
1521 | temp.type = type; | |
1522 | ||
1523 | vinsn_attach (old_expr_vinsn); | |
1524 | vinsn_attach (new_expr_vinsn); | |
1525 | VEC_safe_insert (expr_history_def, heap, vect, ind, &temp); | |
1526 | *pvect = vect; | |
1527 | } | |
1528 | ||
1529 | /* Free history vector PVECT. */ | |
1530 | static void | |
1531 | free_history_vect (VEC (expr_history_def, heap) **pvect) | |
1532 | { | |
1533 | unsigned i; | |
1534 | expr_history_def *phist; | |
1535 | ||
1536 | if (! *pvect) | |
1537 | return; | |
b8698a0f L |
1538 | |
1539 | for (i = 0; | |
e855c69d AB |
1540 | VEC_iterate (expr_history_def, *pvect, i, phist); |
1541 | i++) | |
1542 | { | |
1543 | vinsn_detach (phist->old_expr_vinsn); | |
1544 | vinsn_detach (phist->new_expr_vinsn); | |
1545 | } | |
b8698a0f | 1546 | |
e855c69d AB |
1547 | VEC_free (expr_history_def, heap, *pvect); |
1548 | *pvect = NULL; | |
1549 | } | |
1550 | ||
1551 | ||
1552 | /* Compare two vinsns as rhses if possible and as vinsns otherwise. */ | |
1553 | bool | |
1554 | vinsn_equal_p (vinsn_t x, vinsn_t y) | |
1555 | { | |
1556 | rtx_equal_p_callback_function repcf; | |
1557 | ||
1558 | if (x == y) | |
1559 | return true; | |
1560 | ||
1561 | if (VINSN_TYPE (x) != VINSN_TYPE (y)) | |
1562 | return false; | |
1563 | ||
1564 | if (VINSN_HASH (x) != VINSN_HASH (y)) | |
1565 | return false; | |
1566 | ||
1567 | repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL; | |
b8698a0f | 1568 | if (VINSN_SEPARABLE_P (x)) |
e855c69d AB |
1569 | { |
1570 | /* Compare RHSes of VINSNs. */ | |
1571 | gcc_assert (VINSN_RHS (x)); | |
1572 | gcc_assert (VINSN_RHS (y)); | |
1573 | ||
1574 | return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf); | |
1575 | } | |
1576 | ||
1577 | return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); | |
1578 | } | |
1579 | \f | |
1580 | ||
1581 | /* Functions for working with expressions. */ | |
1582 | ||
1583 | /* Initialize EXPR. */ | |
1584 | static void | |
1585 | init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority, | |
1586 | int sched_times, int orig_bb_index, ds_t spec_done_ds, | |
1587 | ds_t spec_to_check_ds, int orig_sched_cycle, | |
b8698a0f | 1588 | VEC(expr_history_def, heap) *history, bool target_available, |
e855c69d AB |
1589 | bool was_substituted, bool was_renamed, bool needs_spec_check_p, |
1590 | bool cant_move) | |
1591 | { | |
1592 | vinsn_attach (vi); | |
1593 | ||
1594 | EXPR_VINSN (expr) = vi; | |
1595 | EXPR_SPEC (expr) = spec; | |
1596 | EXPR_USEFULNESS (expr) = use; | |
1597 | EXPR_PRIORITY (expr) = priority; | |
1598 | EXPR_PRIORITY_ADJ (expr) = 0; | |
1599 | EXPR_SCHED_TIMES (expr) = sched_times; | |
1600 | EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; | |
1601 | EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; | |
1602 | EXPR_SPEC_DONE_DS (expr) = spec_done_ds; | |
1603 | EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; | |
1604 | ||
1605 | if (history) | |
1606 | EXPR_HISTORY_OF_CHANGES (expr) = history; | |
1607 | else | |
1608 | EXPR_HISTORY_OF_CHANGES (expr) = NULL; | |
1609 | ||
1610 | EXPR_TARGET_AVAILABLE (expr) = target_available; | |
1611 | EXPR_WAS_SUBSTITUTED (expr) = was_substituted; | |
1612 | EXPR_WAS_RENAMED (expr) = was_renamed; | |
1613 | EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; | |
1614 | EXPR_CANT_MOVE (expr) = cant_move; | |
1615 | } | |
1616 | ||
1617 | /* Make a copy of the expr FROM into the expr TO. */ | |
1618 | void | |
1619 | copy_expr (expr_t to, expr_t from) | |
1620 | { | |
1621 | VEC(expr_history_def, heap) *temp = NULL; | |
1622 | ||
1623 | if (EXPR_HISTORY_OF_CHANGES (from)) | |
1624 | { | |
1625 | unsigned i; | |
1626 | expr_history_def *phist; | |
1627 | ||
1628 | temp = VEC_copy (expr_history_def, heap, EXPR_HISTORY_OF_CHANGES (from)); | |
b8698a0f | 1629 | for (i = 0; |
e855c69d AB |
1630 | VEC_iterate (expr_history_def, temp, i, phist); |
1631 | i++) | |
1632 | { | |
1633 | vinsn_attach (phist->old_expr_vinsn); | |
1634 | vinsn_attach (phist->new_expr_vinsn); | |
1635 | } | |
1636 | } | |
1637 | ||
b8698a0f | 1638 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), |
e855c69d AB |
1639 | EXPR_USEFULNESS (from), EXPR_PRIORITY (from), |
1640 | EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), | |
b8698a0f | 1641 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), |
e855c69d | 1642 | EXPR_ORIG_SCHED_CYCLE (from), temp, |
b8698a0f | 1643 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
e855c69d AB |
1644 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1645 | EXPR_CANT_MOVE (from)); | |
1646 | } | |
1647 | ||
b8698a0f | 1648 | /* Same, but the final expr will not ever be in av sets, so don't copy |
e855c69d AB |
1649 | "uninteresting" data such as bitmap cache. */ |
1650 | void | |
1651 | copy_expr_onside (expr_t to, expr_t from) | |
1652 | { | |
1653 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), | |
1654 | EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0, | |
1655 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0, NULL, | |
1656 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), | |
1657 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), | |
1658 | EXPR_CANT_MOVE (from)); | |
1659 | } | |
1660 | ||
1661 | /* Prepare the expr of INSN for scheduling. Used when moving insn and when | |
1662 | initializing new insns. */ | |
1663 | static void | |
1664 | prepare_insn_expr (insn_t insn, int seqno) | |
1665 | { | |
1666 | expr_t expr = INSN_EXPR (insn); | |
1667 | ds_t ds; | |
b8698a0f | 1668 | |
e855c69d AB |
1669 | INSN_SEQNO (insn) = seqno; |
1670 | EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); | |
1671 | EXPR_SPEC (expr) = 0; | |
1672 | EXPR_ORIG_SCHED_CYCLE (expr) = 0; | |
1673 | EXPR_WAS_SUBSTITUTED (expr) = 0; | |
1674 | EXPR_WAS_RENAMED (expr) = 0; | |
1675 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
1676 | INSN_LIVE_VALID_P (insn) = false; | |
1677 | ||
1678 | /* ??? If this expression is speculative, make its dependence | |
1679 | as weak as possible. We can filter this expression later | |
1680 | in process_spec_exprs, because we do not distinguish | |
1681 | between the status we got during compute_av_set and the | |
1682 | existing status. To be fixed. */ | |
1683 | ds = EXPR_SPEC_DONE_DS (expr); | |
1684 | if (ds) | |
1685 | EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); | |
1686 | ||
1687 | free_history_vect (&EXPR_HISTORY_OF_CHANGES (expr)); | |
1688 | } | |
1689 | ||
1690 | /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1691 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1692 | a split point. */ |
1693 | static void | |
1694 | update_target_availability (expr_t to, expr_t from, insn_t split_point) | |
1695 | { | |
b8698a0f | 1696 | if (EXPR_TARGET_AVAILABLE (to) < 0 |
e855c69d AB |
1697 | || EXPR_TARGET_AVAILABLE (from) < 0) |
1698 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1699 | else | |
1700 | { | |
1701 | /* We try to detect the case when one of the expressions | |
1702 | can only be reached through another one. In this case, | |
1703 | we can do better. */ | |
1704 | if (split_point == NULL) | |
1705 | { | |
1706 | int toind, fromind; | |
1707 | ||
1708 | toind = EXPR_ORIG_BB_INDEX (to); | |
1709 | fromind = EXPR_ORIG_BB_INDEX (from); | |
b8698a0f | 1710 | |
e855c69d | 1711 | if (toind && toind == fromind) |
b8698a0f | 1712 | /* Do nothing -- everything is done in |
e855c69d AB |
1713 | merge_with_other_exprs. */ |
1714 | ; | |
1715 | else | |
1716 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1717 | } | |
1718 | else | |
1719 | EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); | |
1720 | } | |
1721 | } | |
1722 | ||
1723 | /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1724 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1725 | a split point. */ |
1726 | static void | |
1727 | update_speculative_bits (expr_t to, expr_t from, insn_t split_point) | |
1728 | { | |
1729 | ds_t old_to_ds, old_from_ds; | |
1730 | ||
1731 | old_to_ds = EXPR_SPEC_DONE_DS (to); | |
1732 | old_from_ds = EXPR_SPEC_DONE_DS (from); | |
b8698a0f | 1733 | |
e855c69d AB |
1734 | EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); |
1735 | EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); | |
1736 | EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); | |
1737 | ||
1738 | /* When merging e.g. control & data speculative exprs, or a control | |
b8698a0f | 1739 | speculative with a control&data speculative one, we really have |
e855c69d AB |
1740 | to change vinsn too. Also, when speculative status is changed, |
1741 | we also need to record this as a transformation in expr's history. */ | |
1742 | if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) | |
1743 | { | |
1744 | old_to_ds = ds_get_speculation_types (old_to_ds); | |
1745 | old_from_ds = ds_get_speculation_types (old_from_ds); | |
b8698a0f | 1746 | |
e855c69d AB |
1747 | if (old_to_ds != old_from_ds) |
1748 | { | |
1749 | ds_t record_ds; | |
b8698a0f L |
1750 | |
1751 | /* When both expressions are speculative, we need to change | |
e855c69d AB |
1752 | the vinsn first. */ |
1753 | if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) | |
1754 | { | |
1755 | int res; | |
b8698a0f | 1756 | |
e855c69d AB |
1757 | res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); |
1758 | gcc_assert (res >= 0); | |
1759 | } | |
1760 | ||
1761 | if (split_point != NULL) | |
1762 | { | |
1763 | /* Record the change with proper status. */ | |
1764 | record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; | |
1765 | record_ds &= ~(old_to_ds & SPECULATIVE); | |
1766 | record_ds &= ~(old_from_ds & SPECULATIVE); | |
b8698a0f L |
1767 | |
1768 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), | |
1769 | INSN_UID (split_point), TRANS_SPECULATION, | |
e855c69d AB |
1770 | EXPR_VINSN (from), EXPR_VINSN (to), |
1771 | record_ds); | |
1772 | } | |
1773 | } | |
1774 | } | |
1775 | } | |
1776 | ||
1777 | ||
1778 | /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, | |
1779 | this is done along different paths. */ | |
1780 | void | |
1781 | merge_expr_data (expr_t to, expr_t from, insn_t split_point) | |
1782 | { | |
1783 | int i; | |
1784 | expr_history_def *phist; | |
b8698a0f | 1785 | |
e855c69d AB |
1786 | /* For now, we just set the spec of resulting expr to be minimum of the specs |
1787 | of merged exprs. */ | |
1788 | if (EXPR_SPEC (to) > EXPR_SPEC (from)) | |
1789 | EXPR_SPEC (to) = EXPR_SPEC (from); | |
1790 | ||
1791 | if (split_point) | |
1792 | EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); | |
1793 | else | |
b8698a0f | 1794 | EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), |
e855c69d AB |
1795 | EXPR_USEFULNESS (from)); |
1796 | ||
1797 | if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) | |
1798 | EXPR_PRIORITY (to) = EXPR_PRIORITY (from); | |
1799 | ||
1800 | if (EXPR_SCHED_TIMES (to) > EXPR_SCHED_TIMES (from)) | |
1801 | EXPR_SCHED_TIMES (to) = EXPR_SCHED_TIMES (from); | |
1802 | ||
1803 | if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) | |
1804 | EXPR_ORIG_BB_INDEX (to) = 0; | |
1805 | ||
b8698a0f | 1806 | EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), |
e855c69d AB |
1807 | EXPR_ORIG_SCHED_CYCLE (from)); |
1808 | ||
1809 | /* We keep this vector sorted. */ | |
b8698a0f L |
1810 | for (i = 0; |
1811 | VEC_iterate (expr_history_def, EXPR_HISTORY_OF_CHANGES (from), | |
e855c69d AB |
1812 | i, phist); |
1813 | i++) | |
b8698a0f L |
1814 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), |
1815 | phist->uid, phist->type, | |
1816 | phist->old_expr_vinsn, phist->new_expr_vinsn, | |
e855c69d AB |
1817 | phist->spec_ds); |
1818 | ||
1819 | EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); | |
1820 | EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); | |
1821 | EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); | |
1822 | ||
1823 | update_target_availability (to, from, split_point); | |
1824 | update_speculative_bits (to, from, split_point); | |
1825 | } | |
1826 | ||
1827 | /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal | |
b8698a0f | 1828 | in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions |
e855c69d AB |
1829 | are merged from different successors at a split point. */ |
1830 | void | |
1831 | merge_expr (expr_t to, expr_t from, insn_t split_point) | |
1832 | { | |
1833 | vinsn_t to_vi = EXPR_VINSN (to); | |
1834 | vinsn_t from_vi = EXPR_VINSN (from); | |
1835 | ||
1836 | gcc_assert (vinsn_equal_p (to_vi, from_vi)); | |
1837 | ||
1838 | /* Make sure that speculative pattern is propagated into exprs that | |
1839 | have non-speculative one. This will provide us with consistent | |
1840 | speculative bits and speculative patterns inside expr. */ | |
1841 | if (EXPR_SPEC_DONE_DS (to) == 0 | |
1842 | && EXPR_SPEC_DONE_DS (from) != 0) | |
1843 | change_vinsn_in_expr (to, EXPR_VINSN (from)); | |
1844 | ||
1845 | merge_expr_data (to, from, split_point); | |
1846 | gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); | |
1847 | } | |
1848 | ||
1849 | /* Clear the information of this EXPR. */ | |
1850 | void | |
1851 | clear_expr (expr_t expr) | |
1852 | { | |
b8698a0f | 1853 | |
e855c69d AB |
1854 | vinsn_detach (EXPR_VINSN (expr)); |
1855 | EXPR_VINSN (expr) = NULL; | |
1856 | ||
1857 | free_history_vect (&EXPR_HISTORY_OF_CHANGES (expr)); | |
1858 | } | |
1859 | ||
1860 | /* For a given LV_SET, mark EXPR having unavailable target register. */ | |
1861 | static void | |
1862 | set_unavailable_target_for_expr (expr_t expr, regset lv_set) | |
1863 | { | |
1864 | if (EXPR_SEPARABLE_P (expr)) | |
1865 | { | |
1866 | if (REG_P (EXPR_LHS (expr)) | |
1867 | && bitmap_bit_p (lv_set, REGNO (EXPR_LHS (expr)))) | |
1868 | { | |
b8698a0f L |
1869 | /* If it's an insn like r1 = use (r1, ...), and it exists in |
1870 | different forms in each of the av_sets being merged, we can't say | |
1871 | whether original destination register is available or not. | |
1872 | However, this still works if destination register is not used | |
e855c69d AB |
1873 | in the original expression: if the branch at which LV_SET we're |
1874 | looking here is not actually 'other branch' in sense that same | |
b8698a0f | 1875 | expression is available through it (but it can't be determined |
e855c69d | 1876 | at computation stage because of transformations on one of the |
b8698a0f L |
1877 | branches), it still won't affect the availability. |
1878 | Liveness of a register somewhere on a code motion path means | |
1879 | it's either read somewhere on a codemotion path, live on | |
e855c69d AB |
1880 | 'other' branch, live at the point immediately following |
1881 | the original operation, or is read by the original operation. | |
1882 | The latter case is filtered out in the condition below. | |
1883 | It still doesn't cover the case when register is defined and used | |
1884 | somewhere within the code motion path, and in this case we could | |
1885 | miss a unifying code motion along both branches using a renamed | |
1886 | register, but it won't affect a code correctness since upon | |
1887 | an actual code motion a bookkeeping code would be generated. */ | |
b8698a0f | 1888 | if (bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
e855c69d AB |
1889 | REGNO (EXPR_LHS (expr)))) |
1890 | EXPR_TARGET_AVAILABLE (expr) = -1; | |
1891 | else | |
1892 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1893 | } | |
1894 | } | |
1895 | else | |
1896 | { | |
1897 | unsigned regno; | |
1898 | reg_set_iterator rsi; | |
b8698a0f L |
1899 | |
1900 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), | |
e855c69d AB |
1901 | 0, regno, rsi) |
1902 | if (bitmap_bit_p (lv_set, regno)) | |
1903 | { | |
1904 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1905 | break; | |
1906 | } | |
1907 | ||
1908 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), | |
1909 | 0, regno, rsi) | |
1910 | if (bitmap_bit_p (lv_set, regno)) | |
1911 | { | |
1912 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1913 | break; | |
1914 | } | |
1915 | } | |
1916 | } | |
1917 | ||
b8698a0f | 1918 | /* Try to make EXPR speculative. Return 1 when EXPR's pattern |
e855c69d AB |
1919 | or dependence status have changed, 2 when also the target register |
1920 | became unavailable, 0 if nothing had to be changed. */ | |
1921 | int | |
1922 | speculate_expr (expr_t expr, ds_t ds) | |
1923 | { | |
1924 | int res; | |
1925 | rtx orig_insn_rtx; | |
1926 | rtx spec_pat; | |
1927 | ds_t target_ds, current_ds; | |
1928 | ||
1929 | /* Obtain the status we need to put on EXPR. */ | |
1930 | target_ds = (ds & SPECULATIVE); | |
1931 | current_ds = EXPR_SPEC_DONE_DS (expr); | |
1932 | ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); | |
1933 | ||
1934 | orig_insn_rtx = EXPR_INSN_RTX (expr); | |
1935 | ||
1936 | res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); | |
1937 | ||
1938 | switch (res) | |
1939 | { | |
1940 | case 0: | |
1941 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1942 | return current_ds != ds ? 1 : 0; | |
b8698a0f | 1943 | |
e855c69d AB |
1944 | case 1: |
1945 | { | |
1946 | rtx spec_insn_rtx = create_insn_rtx_from_pattern (spec_pat, NULL_RTX); | |
1947 | vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); | |
1948 | ||
1949 | change_vinsn_in_expr (expr, spec_vinsn); | |
1950 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1951 | EXPR_NEEDS_SPEC_CHECK_P (expr) = true; | |
1952 | ||
b8698a0f | 1953 | /* Do not allow clobbering the address register of speculative |
e855c69d | 1954 | insns. */ |
b8698a0f | 1955 | if (bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
e855c69d AB |
1956 | expr_dest_regno (expr))) |
1957 | { | |
1958 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1959 | return 2; | |
1960 | } | |
1961 | ||
1962 | return 1; | |
1963 | } | |
1964 | ||
1965 | case -1: | |
1966 | return -1; | |
1967 | ||
1968 | default: | |
1969 | gcc_unreachable (); | |
1970 | return -1; | |
1971 | } | |
1972 | } | |
1973 | ||
1974 | /* Return a destination register, if any, of EXPR. */ | |
1975 | rtx | |
1976 | expr_dest_reg (expr_t expr) | |
1977 | { | |
1978 | rtx dest = VINSN_LHS (EXPR_VINSN (expr)); | |
1979 | ||
1980 | if (dest != NULL_RTX && REG_P (dest)) | |
1981 | return dest; | |
1982 | ||
1983 | return NULL_RTX; | |
1984 | } | |
1985 | ||
1986 | /* Returns the REGNO of the R's destination. */ | |
1987 | unsigned | |
1988 | expr_dest_regno (expr_t expr) | |
1989 | { | |
1990 | rtx dest = expr_dest_reg (expr); | |
1991 | ||
1992 | gcc_assert (dest != NULL_RTX); | |
1993 | return REGNO (dest); | |
1994 | } | |
1995 | ||
b8698a0f | 1996 | /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in |
e855c69d AB |
1997 | AV_SET having unavailable target register. */ |
1998 | void | |
1999 | mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) | |
2000 | { | |
2001 | expr_t expr; | |
2002 | av_set_iterator avi; | |
2003 | ||
2004 | FOR_EACH_EXPR (expr, avi, join_set) | |
2005 | if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) | |
2006 | set_unavailable_target_for_expr (expr, lv_set); | |
2007 | } | |
2008 | \f | |
2009 | ||
2010 | /* Av set functions. */ | |
2011 | ||
2012 | /* Add a new element to av set SETP. | |
2013 | Return the element added. */ | |
2014 | static av_set_t | |
2015 | av_set_add_element (av_set_t *setp) | |
2016 | { | |
2017 | /* Insert at the beginning of the list. */ | |
2018 | _list_add (setp); | |
2019 | return *setp; | |
2020 | } | |
2021 | ||
2022 | /* Add EXPR to SETP. */ | |
2023 | void | |
2024 | av_set_add (av_set_t *setp, expr_t expr) | |
2025 | { | |
2026 | av_set_t elem; | |
b8698a0f | 2027 | |
e855c69d AB |
2028 | gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); |
2029 | elem = av_set_add_element (setp); | |
2030 | copy_expr (_AV_SET_EXPR (elem), expr); | |
2031 | } | |
2032 | ||
2033 | /* Same, but do not copy EXPR. */ | |
2034 | static void | |
2035 | av_set_add_nocopy (av_set_t *setp, expr_t expr) | |
2036 | { | |
2037 | av_set_t elem; | |
2038 | ||
2039 | elem = av_set_add_element (setp); | |
2040 | *_AV_SET_EXPR (elem) = *expr; | |
2041 | } | |
2042 | ||
2043 | /* Remove expr pointed to by IP from the av_set. */ | |
2044 | void | |
2045 | av_set_iter_remove (av_set_iterator *ip) | |
2046 | { | |
2047 | clear_expr (_AV_SET_EXPR (*ip->lp)); | |
2048 | _list_iter_remove (ip); | |
2049 | } | |
2050 | ||
2051 | /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the | |
2052 | sense of vinsn_equal_p function. Return NULL if no such expr is | |
2053 | in SET was found. */ | |
2054 | expr_t | |
2055 | av_set_lookup (av_set_t set, vinsn_t sought_vinsn) | |
2056 | { | |
2057 | expr_t expr; | |
2058 | av_set_iterator i; | |
2059 | ||
2060 | FOR_EACH_EXPR (expr, i, set) | |
2061 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2062 | return expr; | |
2063 | return NULL; | |
2064 | } | |
2065 | ||
2066 | /* Same, but also remove the EXPR found. */ | |
2067 | static expr_t | |
2068 | av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) | |
2069 | { | |
2070 | expr_t expr; | |
2071 | av_set_iterator i; | |
2072 | ||
2073 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2074 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2075 | { | |
2076 | _list_iter_remove_nofree (&i); | |
2077 | return expr; | |
2078 | } | |
2079 | return NULL; | |
2080 | } | |
2081 | ||
2082 | /* Search for an expr in SET, such that it's equivalent to EXPR in the | |
2083 | sense of vinsn_equal_p function of their vinsns, but not EXPR itself. | |
2084 | Returns NULL if no such expr is in SET was found. */ | |
2085 | static expr_t | |
2086 | av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) | |
2087 | { | |
2088 | expr_t cur_expr; | |
2089 | av_set_iterator i; | |
2090 | ||
2091 | FOR_EACH_EXPR (cur_expr, i, set) | |
2092 | { | |
2093 | if (cur_expr == expr) | |
2094 | continue; | |
2095 | if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) | |
2096 | return cur_expr; | |
2097 | } | |
2098 | ||
2099 | return NULL; | |
2100 | } | |
2101 | ||
2102 | /* If other expression is already in AVP, remove one of them. */ | |
2103 | expr_t | |
2104 | merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) | |
2105 | { | |
2106 | expr_t expr2; | |
2107 | ||
2108 | expr2 = av_set_lookup_other_equiv_expr (*avp, expr); | |
2109 | if (expr2 != NULL) | |
2110 | { | |
2111 | /* Reset target availability on merge, since taking it only from one | |
2112 | of the exprs would be controversial for different code. */ | |
2113 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2114 | EXPR_USEFULNESS (expr2) = 0; | |
2115 | ||
2116 | merge_expr (expr2, expr, NULL); | |
b8698a0f | 2117 | |
e855c69d AB |
2118 | /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ |
2119 | EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; | |
b8698a0f | 2120 | |
e855c69d AB |
2121 | av_set_iter_remove (ip); |
2122 | return expr2; | |
2123 | } | |
2124 | ||
2125 | return expr; | |
2126 | } | |
2127 | ||
2128 | /* Return true if there is an expr that correlates to VI in SET. */ | |
2129 | bool | |
2130 | av_set_is_in_p (av_set_t set, vinsn_t vi) | |
2131 | { | |
2132 | return av_set_lookup (set, vi) != NULL; | |
2133 | } | |
2134 | ||
2135 | /* Return a copy of SET. */ | |
2136 | av_set_t | |
2137 | av_set_copy (av_set_t set) | |
2138 | { | |
2139 | expr_t expr; | |
2140 | av_set_iterator i; | |
2141 | av_set_t res = NULL; | |
2142 | ||
2143 | FOR_EACH_EXPR (expr, i, set) | |
2144 | av_set_add (&res, expr); | |
2145 | ||
2146 | return res; | |
2147 | } | |
2148 | ||
2149 | /* Join two av sets that do not have common elements by attaching second set | |
2150 | (pointed to by FROMP) to the end of first set (TO_TAILP must point to | |
2151 | _AV_SET_NEXT of first set's last element). */ | |
2152 | static void | |
2153 | join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) | |
2154 | { | |
2155 | gcc_assert (*to_tailp == NULL); | |
2156 | *to_tailp = *fromp; | |
2157 | *fromp = NULL; | |
2158 | } | |
2159 | ||
2160 | /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set | |
2161 | pointed to by FROMP afterwards. */ | |
2162 | void | |
2163 | av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) | |
2164 | { | |
2165 | expr_t expr1; | |
2166 | av_set_iterator i; | |
2167 | ||
2168 | /* Delete from TOP all exprs, that present in FROMP. */ | |
2169 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2170 | { | |
2171 | expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1)); | |
2172 | ||
2173 | if (expr2) | |
2174 | { | |
2175 | merge_expr (expr2, expr1, insn); | |
2176 | av_set_iter_remove (&i); | |
2177 | } | |
2178 | } | |
2179 | ||
2180 | join_distinct_sets (i.lp, fromp); | |
2181 | } | |
2182 | ||
b8698a0f | 2183 | /* Same as above, but also update availability of target register in |
e855c69d AB |
2184 | TOP judging by TO_LV_SET and FROM_LV_SET. */ |
2185 | void | |
2186 | av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, | |
2187 | regset from_lv_set, insn_t insn) | |
2188 | { | |
2189 | expr_t expr1; | |
2190 | av_set_iterator i; | |
2191 | av_set_t *to_tailp, in_both_set = NULL; | |
2192 | ||
2193 | /* Delete from TOP all expres, that present in FROMP. */ | |
2194 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2195 | { | |
2196 | expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1)); | |
2197 | ||
2198 | if (expr2) | |
2199 | { | |
b8698a0f | 2200 | /* It may be that the expressions have different destination |
e855c69d AB |
2201 | registers, in which case we need to check liveness here. */ |
2202 | if (EXPR_SEPARABLE_P (expr1)) | |
2203 | { | |
b8698a0f | 2204 | int regno1 = (REG_P (EXPR_LHS (expr1)) |
e855c69d | 2205 | ? (int) expr_dest_regno (expr1) : -1); |
b8698a0f | 2206 | int regno2 = (REG_P (EXPR_LHS (expr2)) |
e855c69d | 2207 | ? (int) expr_dest_regno (expr2) : -1); |
b8698a0f L |
2208 | |
2209 | /* ??? We don't have a way to check restrictions for | |
e855c69d AB |
2210 | *other* register on the current path, we did it only |
2211 | for the current target register. Give up. */ | |
2212 | if (regno1 != regno2) | |
2213 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2214 | } | |
2215 | else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) | |
2216 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2217 | ||
2218 | merge_expr (expr2, expr1, insn); | |
2219 | av_set_add_nocopy (&in_both_set, expr2); | |
2220 | av_set_iter_remove (&i); | |
2221 | } | |
2222 | else | |
b8698a0f | 2223 | /* EXPR1 is present in TOP, but not in FROMP. Check it on |
e855c69d AB |
2224 | FROM_LV_SET. */ |
2225 | set_unavailable_target_for_expr (expr1, from_lv_set); | |
2226 | } | |
2227 | to_tailp = i.lp; | |
2228 | ||
2229 | /* These expressions are not present in TOP. Check liveness | |
2230 | restrictions on TO_LV_SET. */ | |
2231 | FOR_EACH_EXPR (expr1, i, *fromp) | |
2232 | set_unavailable_target_for_expr (expr1, to_lv_set); | |
2233 | ||
2234 | join_distinct_sets (i.lp, &in_both_set); | |
2235 | join_distinct_sets (to_tailp, fromp); | |
2236 | } | |
2237 | ||
2238 | /* Clear av_set pointed to by SETP. */ | |
2239 | void | |
2240 | av_set_clear (av_set_t *setp) | |
2241 | { | |
2242 | expr_t expr; | |
2243 | av_set_iterator i; | |
2244 | ||
2245 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2246 | av_set_iter_remove (&i); | |
2247 | ||
2248 | gcc_assert (*setp == NULL); | |
2249 | } | |
2250 | ||
2251 | /* Leave only one non-speculative element in the SETP. */ | |
2252 | void | |
2253 | av_set_leave_one_nonspec (av_set_t *setp) | |
2254 | { | |
2255 | expr_t expr; | |
2256 | av_set_iterator i; | |
2257 | bool has_one_nonspec = false; | |
2258 | ||
b8698a0f | 2259 | /* Keep all speculative exprs, and leave one non-speculative |
e855c69d AB |
2260 | (the first one). */ |
2261 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2262 | { | |
2263 | if (!EXPR_SPEC_DONE_DS (expr)) | |
2264 | { | |
2265 | if (has_one_nonspec) | |
2266 | av_set_iter_remove (&i); | |
2267 | else | |
2268 | has_one_nonspec = true; | |
2269 | } | |
2270 | } | |
2271 | } | |
2272 | ||
2273 | /* Return the N'th element of the SET. */ | |
2274 | expr_t | |
2275 | av_set_element (av_set_t set, int n) | |
2276 | { | |
2277 | expr_t expr; | |
2278 | av_set_iterator i; | |
2279 | ||
2280 | FOR_EACH_EXPR (expr, i, set) | |
2281 | if (n-- == 0) | |
2282 | return expr; | |
2283 | ||
2284 | gcc_unreachable (); | |
2285 | return NULL; | |
2286 | } | |
2287 | ||
2288 | /* Deletes all expressions from AVP that are conditional branches (IFs). */ | |
2289 | void | |
2290 | av_set_substract_cond_branches (av_set_t *avp) | |
2291 | { | |
2292 | av_set_iterator i; | |
2293 | expr_t expr; | |
2294 | ||
2295 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2296 | if (vinsn_cond_branch_p (EXPR_VINSN (expr))) | |
2297 | av_set_iter_remove (&i); | |
2298 | } | |
2299 | ||
b8698a0f | 2300 | /* Multiplies usefulness attribute of each member of av-set *AVP by |
e855c69d AB |
2301 | value PROB / ALL_PROB. */ |
2302 | void | |
2303 | av_set_split_usefulness (av_set_t av, int prob, int all_prob) | |
2304 | { | |
2305 | av_set_iterator i; | |
2306 | expr_t expr; | |
2307 | ||
2308 | FOR_EACH_EXPR (expr, i, av) | |
b8698a0f | 2309 | EXPR_USEFULNESS (expr) = (all_prob |
e855c69d AB |
2310 | ? (EXPR_USEFULNESS (expr) * prob) / all_prob |
2311 | : 0); | |
2312 | } | |
2313 | ||
2314 | /* Leave in AVP only those expressions, which are present in AV, | |
2315 | and return it. */ | |
2316 | void | |
2317 | av_set_intersect (av_set_t *avp, av_set_t av) | |
2318 | { | |
2319 | av_set_iterator i; | |
2320 | expr_t expr; | |
2321 | ||
2322 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2323 | if (av_set_lookup (av, EXPR_VINSN (expr)) == NULL) | |
2324 | av_set_iter_remove (&i); | |
2325 | } | |
2326 | ||
2327 | \f | |
2328 | ||
2329 | /* Dependence hooks to initialize insn data. */ | |
2330 | ||
2331 | /* This is used in hooks callable from dependence analysis when initializing | |
2332 | instruction's data. */ | |
2333 | static struct | |
2334 | { | |
2335 | /* Where the dependence was found (lhs/rhs). */ | |
2336 | deps_where_t where; | |
2337 | ||
2338 | /* The actual data object to initialize. */ | |
2339 | idata_t id; | |
2340 | ||
2341 | /* True when the insn should not be made clonable. */ | |
2342 | bool force_unique_p; | |
2343 | ||
2344 | /* True when insn should be treated as of type USE, i.e. never renamed. */ | |
2345 | bool force_use_p; | |
2346 | } deps_init_id_data; | |
2347 | ||
2348 | ||
b8698a0f | 2349 | /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be |
e855c69d AB |
2350 | clonable. */ |
2351 | static void | |
2352 | setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) | |
2353 | { | |
2354 | int type; | |
b8698a0f | 2355 | |
e855c69d AB |
2356 | /* Determine whether INSN could be cloned and return appropriate vinsn type. |
2357 | That clonable insns which can be separated into lhs and rhs have type SET. | |
2358 | Other clonable insns have type USE. */ | |
2359 | type = GET_CODE (insn); | |
2360 | ||
2361 | /* Only regular insns could be cloned. */ | |
2362 | if (type == INSN && !force_unique_p) | |
2363 | type = SET; | |
2364 | else if (type == JUMP_INSN && simplejump_p (insn)) | |
2365 | type = PC; | |
b5b8b0ac AO |
2366 | else if (type == DEBUG_INSN) |
2367 | type = !force_unique_p ? USE : INSN; | |
b8698a0f | 2368 | |
e855c69d AB |
2369 | IDATA_TYPE (id) = type; |
2370 | IDATA_REG_SETS (id) = get_clear_regset_from_pool (); | |
2371 | IDATA_REG_USES (id) = get_clear_regset_from_pool (); | |
2372 | IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); | |
2373 | } | |
2374 | ||
2375 | /* Start initializing insn data. */ | |
2376 | static void | |
2377 | deps_init_id_start_insn (insn_t insn) | |
2378 | { | |
2379 | gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); | |
2380 | ||
2381 | setup_id_for_insn (deps_init_id_data.id, insn, | |
2382 | deps_init_id_data.force_unique_p); | |
2383 | deps_init_id_data.where = DEPS_IN_INSN; | |
2384 | } | |
2385 | ||
2386 | /* Start initializing lhs data. */ | |
2387 | static void | |
2388 | deps_init_id_start_lhs (rtx lhs) | |
2389 | { | |
2390 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2391 | gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); | |
2392 | ||
2393 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2394 | { | |
2395 | IDATA_LHS (deps_init_id_data.id) = lhs; | |
2396 | deps_init_id_data.where = DEPS_IN_LHS; | |
2397 | } | |
2398 | } | |
2399 | ||
2400 | /* Finish initializing lhs data. */ | |
2401 | static void | |
2402 | deps_init_id_finish_lhs (void) | |
2403 | { | |
2404 | deps_init_id_data.where = DEPS_IN_INSN; | |
2405 | } | |
2406 | ||
2407 | /* Note a set of REGNO. */ | |
2408 | static void | |
2409 | deps_init_id_note_reg_set (int regno) | |
2410 | { | |
2411 | haifa_note_reg_set (regno); | |
2412 | ||
2413 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2414 | deps_init_id_data.force_use_p = true; | |
2415 | ||
2416 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2417 | SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); | |
2418 | ||
2419 | #ifdef STACK_REGS | |
b8698a0f | 2420 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2421 | renaming to avoid issues with find_used_regs. */ |
2422 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2423 | deps_init_id_data.force_use_p = true; | |
2424 | #endif | |
2425 | } | |
2426 | ||
2427 | /* Note a clobber of REGNO. */ | |
2428 | static void | |
2429 | deps_init_id_note_reg_clobber (int regno) | |
2430 | { | |
2431 | haifa_note_reg_clobber (regno); | |
2432 | ||
2433 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2434 | deps_init_id_data.force_use_p = true; | |
2435 | ||
2436 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2437 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); | |
2438 | } | |
2439 | ||
2440 | /* Note a use of REGNO. */ | |
2441 | static void | |
2442 | deps_init_id_note_reg_use (int regno) | |
2443 | { | |
2444 | haifa_note_reg_use (regno); | |
2445 | ||
2446 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2447 | SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); | |
2448 | } | |
2449 | ||
2450 | /* Start initializing rhs data. */ | |
2451 | static void | |
2452 | deps_init_id_start_rhs (rtx rhs) | |
2453 | { | |
2454 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2455 | ||
2456 | /* And there was no sel_deps_reset_to_insn (). */ | |
2457 | if (IDATA_LHS (deps_init_id_data.id) != NULL) | |
2458 | { | |
2459 | IDATA_RHS (deps_init_id_data.id) = rhs; | |
2460 | deps_init_id_data.where = DEPS_IN_RHS; | |
2461 | } | |
2462 | } | |
2463 | ||
2464 | /* Finish initializing rhs data. */ | |
2465 | static void | |
2466 | deps_init_id_finish_rhs (void) | |
2467 | { | |
2468 | gcc_assert (deps_init_id_data.where == DEPS_IN_RHS | |
2469 | || deps_init_id_data.where == DEPS_IN_INSN); | |
2470 | deps_init_id_data.where = DEPS_IN_INSN; | |
2471 | } | |
2472 | ||
2473 | /* Finish initializing insn data. */ | |
2474 | static void | |
2475 | deps_init_id_finish_insn (void) | |
2476 | { | |
2477 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2478 | ||
2479 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2480 | { | |
2481 | rtx lhs = IDATA_LHS (deps_init_id_data.id); | |
2482 | rtx rhs = IDATA_RHS (deps_init_id_data.id); | |
2483 | ||
2484 | if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) | |
2485 | || deps_init_id_data.force_use_p) | |
2486 | { | |
b8698a0f | 2487 | /* This should be a USE, as we don't want to schedule its RHS |
e855c69d | 2488 | separately. However, we still want to have them recorded |
b8698a0f | 2489 | for the purposes of substitution. That's why we don't |
e855c69d AB |
2490 | simply call downgrade_to_use () here. */ |
2491 | gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); | |
2492 | gcc_assert (!lhs == !rhs); | |
2493 | ||
2494 | IDATA_TYPE (deps_init_id_data.id) = USE; | |
2495 | } | |
2496 | } | |
2497 | ||
2498 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2499 | } | |
2500 | ||
2501 | /* This is dependence info used for initializing insn's data. */ | |
2502 | static struct sched_deps_info_def deps_init_id_sched_deps_info; | |
2503 | ||
2504 | /* This initializes most of the static part of the above structure. */ | |
2505 | static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = | |
2506 | { | |
2507 | NULL, | |
2508 | ||
2509 | deps_init_id_start_insn, | |
2510 | deps_init_id_finish_insn, | |
2511 | deps_init_id_start_lhs, | |
2512 | deps_init_id_finish_lhs, | |
2513 | deps_init_id_start_rhs, | |
2514 | deps_init_id_finish_rhs, | |
2515 | deps_init_id_note_reg_set, | |
2516 | deps_init_id_note_reg_clobber, | |
2517 | deps_init_id_note_reg_use, | |
2518 | NULL, /* note_mem_dep */ | |
2519 | NULL, /* note_dep */ | |
2520 | ||
2521 | 0, /* use_cselib */ | |
2522 | 0, /* use_deps_list */ | |
2523 | 0 /* generate_spec_deps */ | |
2524 | }; | |
2525 | ||
2526 | /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, | |
2527 | we don't actually need information about lhs and rhs. */ | |
2528 | static void | |
2529 | setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) | |
2530 | { | |
2531 | rtx pat = PATTERN (insn); | |
b8698a0f | 2532 | |
481683e1 | 2533 | if (NONJUMP_INSN_P (insn) |
b8698a0f | 2534 | && GET_CODE (pat) == SET |
e855c69d AB |
2535 | && !force_unique_p) |
2536 | { | |
2537 | IDATA_RHS (id) = SET_SRC (pat); | |
2538 | IDATA_LHS (id) = SET_DEST (pat); | |
2539 | } | |
2540 | else | |
2541 | IDATA_LHS (id) = IDATA_RHS (id) = NULL; | |
2542 | } | |
2543 | ||
2544 | /* Possibly downgrade INSN to USE. */ | |
2545 | static void | |
2546 | maybe_downgrade_id_to_use (idata_t id, insn_t insn) | |
2547 | { | |
2548 | bool must_be_use = false; | |
2549 | unsigned uid = INSN_UID (insn); | |
57512f53 | 2550 | df_ref *rec; |
e855c69d AB |
2551 | rtx lhs = IDATA_LHS (id); |
2552 | rtx rhs = IDATA_RHS (id); | |
b8698a0f | 2553 | |
e855c69d AB |
2554 | /* We downgrade only SETs. */ |
2555 | if (IDATA_TYPE (id) != SET) | |
2556 | return; | |
2557 | ||
2558 | if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) | |
2559 | { | |
2560 | IDATA_TYPE (id) = USE; | |
2561 | return; | |
2562 | } | |
b8698a0f | 2563 | |
e855c69d AB |
2564 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2565 | { | |
57512f53 | 2566 | df_ref def = *rec; |
b8698a0f | 2567 | |
e855c69d AB |
2568 | if (DF_REF_INSN (def) |
2569 | && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) | |
2570 | && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) | |
2571 | { | |
2572 | must_be_use = true; | |
2573 | break; | |
2574 | } | |
2575 | ||
2576 | #ifdef STACK_REGS | |
b8698a0f | 2577 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2578 | renaming to avoid issues with find_used_regs. */ |
2579 | if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) | |
2580 | { | |
2581 | must_be_use = true; | |
2582 | break; | |
2583 | } | |
2584 | #endif | |
b8698a0f L |
2585 | } |
2586 | ||
e855c69d AB |
2587 | if (must_be_use) |
2588 | IDATA_TYPE (id) = USE; | |
2589 | } | |
2590 | ||
2591 | /* Setup register sets describing INSN in ID. */ | |
2592 | static void | |
2593 | setup_id_reg_sets (idata_t id, insn_t insn) | |
2594 | { | |
2595 | unsigned uid = INSN_UID (insn); | |
57512f53 | 2596 | df_ref *rec; |
e855c69d | 2597 | regset tmp = get_clear_regset_from_pool (); |
b8698a0f | 2598 | |
e855c69d AB |
2599 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2600 | { | |
57512f53 | 2601 | df_ref def = *rec; |
e855c69d | 2602 | unsigned int regno = DF_REF_REGNO (def); |
b8698a0f | 2603 | |
e855c69d AB |
2604 | /* Post modifies are treated like clobbers by sched-deps.c. */ |
2605 | if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER | |
2606 | | DF_REF_PRE_POST_MODIFY))) | |
2607 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); | |
2608 | else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) | |
2609 | { | |
2610 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); | |
2611 | ||
2612 | #ifdef STACK_REGS | |
b8698a0f | 2613 | /* For stack registers, treat writes to them as writes |
e855c69d AB |
2614 | to the first one to be consistent with sched-deps.c. */ |
2615 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2616 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); | |
2617 | #endif | |
2618 | } | |
2619 | /* Mark special refs that generate read/write def pair. */ | |
2620 | if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) | |
2621 | || regno == STACK_POINTER_REGNUM) | |
2622 | bitmap_set_bit (tmp, regno); | |
2623 | } | |
b8698a0f | 2624 | |
e855c69d AB |
2625 | for (rec = DF_INSN_UID_USES (uid); *rec; rec++) |
2626 | { | |
57512f53 | 2627 | df_ref use = *rec; |
e855c69d AB |
2628 | unsigned int regno = DF_REF_REGNO (use); |
2629 | ||
2630 | /* When these refs are met for the first time, skip them, as | |
2631 | these uses are just counterparts of some defs. */ | |
2632 | if (bitmap_bit_p (tmp, regno)) | |
2633 | bitmap_clear_bit (tmp, regno); | |
2634 | else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) | |
2635 | { | |
2636 | SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); | |
2637 | ||
2638 | #ifdef STACK_REGS | |
b8698a0f | 2639 | /* For stack registers, treat reads from them as reads from |
e855c69d AB |
2640 | the first one to be consistent with sched-deps.c. */ |
2641 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2642 | SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); | |
2643 | #endif | |
2644 | } | |
2645 | } | |
2646 | ||
2647 | return_regset_to_pool (tmp); | |
2648 | } | |
2649 | ||
2650 | /* Initialize instruction data for INSN in ID using DF's data. */ | |
2651 | static void | |
2652 | init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) | |
2653 | { | |
2654 | gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); | |
2655 | ||
2656 | setup_id_for_insn (id, insn, force_unique_p); | |
2657 | setup_id_lhs_rhs (id, insn, force_unique_p); | |
2658 | ||
2659 | if (INSN_NOP_P (insn)) | |
2660 | return; | |
2661 | ||
2662 | maybe_downgrade_id_to_use (id, insn); | |
2663 | setup_id_reg_sets (id, insn); | |
2664 | } | |
2665 | ||
2666 | /* Initialize instruction data for INSN in ID. */ | |
2667 | static void | |
2668 | deps_init_id (idata_t id, insn_t insn, bool force_unique_p) | |
2669 | { | |
2670 | struct deps _dc, *dc = &_dc; | |
2671 | ||
2672 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2673 | deps_init_id_data.id = id; | |
2674 | deps_init_id_data.force_unique_p = force_unique_p; | |
2675 | deps_init_id_data.force_use_p = false; | |
2676 | ||
bcf33775 | 2677 | init_deps (dc, false); |
e855c69d AB |
2678 | |
2679 | memcpy (&deps_init_id_sched_deps_info, | |
2680 | &const_deps_init_id_sched_deps_info, | |
2681 | sizeof (deps_init_id_sched_deps_info)); | |
2682 | ||
2683 | if (spec_info != NULL) | |
2684 | deps_init_id_sched_deps_info.generate_spec_deps = 1; | |
2685 | ||
2686 | sched_deps_info = &deps_init_id_sched_deps_info; | |
2687 | ||
2688 | deps_analyze_insn (dc, insn); | |
2689 | ||
2690 | free_deps (dc); | |
2691 | ||
2692 | deps_init_id_data.id = NULL; | |
2693 | } | |
2694 | ||
2695 | \f | |
2696 | ||
2697 | /* Implement hooks for collecting fundamental insn properties like if insn is | |
2698 | an ASM or is within a SCHED_GROUP. */ | |
2699 | ||
2700 | /* True when a "one-time init" data for INSN was already inited. */ | |
2701 | static bool | |
2702 | first_time_insn_init (insn_t insn) | |
2703 | { | |
2704 | return INSN_LIVE (insn) == NULL; | |
2705 | } | |
2706 | ||
2707 | /* Hash an entry in a transformed_insns hashtable. */ | |
2708 | static hashval_t | |
2709 | hash_transformed_insns (const void *p) | |
2710 | { | |
2711 | return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2712 | } | |
2713 | ||
2714 | /* Compare the entries in a transformed_insns hashtable. */ | |
2715 | static int | |
2716 | eq_transformed_insns (const void *p, const void *q) | |
2717 | { | |
2718 | rtx i1 = VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2719 | rtx i2 = VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); | |
2720 | ||
2721 | if (INSN_UID (i1) == INSN_UID (i2)) | |
2722 | return 1; | |
2723 | return rtx_equal_p (PATTERN (i1), PATTERN (i2)); | |
2724 | } | |
2725 | ||
2726 | /* Free an entry in a transformed_insns hashtable. */ | |
2727 | static void | |
2728 | free_transformed_insns (void *p) | |
2729 | { | |
2730 | struct transformed_insns *pti = (struct transformed_insns *) p; | |
2731 | ||
2732 | vinsn_detach (pti->vinsn_old); | |
2733 | vinsn_detach (pti->vinsn_new); | |
2734 | free (pti); | |
2735 | } | |
2736 | ||
b8698a0f | 2737 | /* Init the s_i_d data for INSN which should be inited just once, when |
e855c69d AB |
2738 | we first see the insn. */ |
2739 | static void | |
2740 | init_first_time_insn_data (insn_t insn) | |
2741 | { | |
2742 | /* This should not be set if this is the first time we init data for | |
2743 | insn. */ | |
2744 | gcc_assert (first_time_insn_init (insn)); | |
b8698a0f | 2745 | |
e855c69d AB |
2746 | /* These are needed for nops too. */ |
2747 | INSN_LIVE (insn) = get_regset_from_pool (); | |
2748 | INSN_LIVE_VALID_P (insn) = false; | |
bcf33775 | 2749 | |
e855c69d AB |
2750 | if (!INSN_NOP_P (insn)) |
2751 | { | |
2752 | INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); | |
2753 | INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); | |
b8698a0f | 2754 | INSN_TRANSFORMED_INSNS (insn) |
e855c69d AB |
2755 | = htab_create (16, hash_transformed_insns, |
2756 | eq_transformed_insns, free_transformed_insns); | |
bcf33775 | 2757 | init_deps (&INSN_DEPS_CONTEXT (insn), true); |
e855c69d AB |
2758 | } |
2759 | } | |
2760 | ||
b8698a0f | 2761 | /* Free almost all above data for INSN that is scheduled already. |
bcf33775 AB |
2762 | Used for extra-large basic blocks. */ |
2763 | void | |
2764 | free_data_for_scheduled_insn (insn_t insn) | |
e855c69d AB |
2765 | { |
2766 | gcc_assert (! first_time_insn_init (insn)); | |
b8698a0f | 2767 | |
bcf33775 AB |
2768 | if (! INSN_ANALYZED_DEPS (insn)) |
2769 | return; | |
b8698a0f | 2770 | |
e855c69d AB |
2771 | BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); |
2772 | BITMAP_FREE (INSN_FOUND_DEPS (insn)); | |
2773 | htab_delete (INSN_TRANSFORMED_INSNS (insn)); | |
b8698a0f | 2774 | |
e855c69d AB |
2775 | /* This is allocated only for bookkeeping insns. */ |
2776 | if (INSN_ORIGINATORS (insn)) | |
2777 | BITMAP_FREE (INSN_ORIGINATORS (insn)); | |
2778 | free_deps (&INSN_DEPS_CONTEXT (insn)); | |
bcf33775 AB |
2779 | |
2780 | INSN_ANALYZED_DEPS (insn) = NULL; | |
2781 | ||
b8698a0f | 2782 | /* Clear the readonly flag so we would ICE when trying to recalculate |
bcf33775 AB |
2783 | the deps context (as we believe that it should not happen). */ |
2784 | (&INSN_DEPS_CONTEXT (insn))->readonly = 0; | |
2785 | } | |
2786 | ||
2787 | /* Free the same data as above for INSN. */ | |
2788 | static void | |
2789 | free_first_time_insn_data (insn_t insn) | |
2790 | { | |
2791 | gcc_assert (! first_time_insn_init (insn)); | |
2792 | ||
2793 | free_data_for_scheduled_insn (insn); | |
2794 | return_regset_to_pool (INSN_LIVE (insn)); | |
2795 | INSN_LIVE (insn) = NULL; | |
2796 | INSN_LIVE_VALID_P (insn) = false; | |
e855c69d AB |
2797 | } |
2798 | ||
2799 | /* Initialize region-scope data structures for basic blocks. */ | |
2800 | static void | |
2801 | init_global_and_expr_for_bb (basic_block bb) | |
2802 | { | |
2803 | if (sel_bb_empty_p (bb)) | |
2804 | return; | |
2805 | ||
2806 | invalidate_av_set (bb); | |
2807 | } | |
2808 | ||
2809 | /* Data for global dependency analysis (to initialize CANT_MOVE and | |
2810 | SCHED_GROUP_P). */ | |
2811 | static struct | |
2812 | { | |
2813 | /* Previous insn. */ | |
2814 | insn_t prev_insn; | |
2815 | } init_global_data; | |
2816 | ||
2817 | /* Determine if INSN is in the sched_group, is an asm or should not be | |
2818 | cloned. After that initialize its expr. */ | |
2819 | static void | |
2820 | init_global_and_expr_for_insn (insn_t insn) | |
2821 | { | |
2822 | if (LABEL_P (insn)) | |
2823 | return; | |
2824 | ||
2825 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2826 | { | |
2827 | init_global_data.prev_insn = NULL_RTX; | |
2828 | return; | |
2829 | } | |
2830 | ||
2831 | gcc_assert (INSN_P (insn)); | |
2832 | ||
2833 | if (SCHED_GROUP_P (insn)) | |
2834 | /* Setup a sched_group. */ | |
2835 | { | |
2836 | insn_t prev_insn = init_global_data.prev_insn; | |
2837 | ||
2838 | if (prev_insn) | |
2839 | INSN_SCHED_NEXT (prev_insn) = insn; | |
2840 | ||
2841 | init_global_data.prev_insn = insn; | |
2842 | } | |
2843 | else | |
2844 | init_global_data.prev_insn = NULL_RTX; | |
2845 | ||
2846 | if (GET_CODE (PATTERN (insn)) == ASM_INPUT | |
2847 | || asm_noperands (PATTERN (insn)) >= 0) | |
2848 | /* Mark INSN as an asm. */ | |
2849 | INSN_ASM_P (insn) = true; | |
2850 | ||
2851 | { | |
2852 | bool force_unique_p; | |
2853 | ds_t spec_done_ds; | |
2854 | ||
2855 | /* Certain instructions cannot be cloned. */ | |
2856 | if (CANT_MOVE (insn) | |
2857 | || INSN_ASM_P (insn) | |
2858 | || SCHED_GROUP_P (insn) | |
b8698a0f | 2859 | || prologue_epilogue_contains (insn) |
e855c69d AB |
2860 | /* Exception handling insns are always unique. */ |
2861 | || (flag_non_call_exceptions && can_throw_internal (insn)) | |
2862 | /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ | |
2863 | || control_flow_insn_p (insn)) | |
2864 | force_unique_p = true; | |
2865 | else | |
2866 | force_unique_p = false; | |
2867 | ||
2868 | if (targetm.sched.get_insn_spec_ds) | |
2869 | { | |
2870 | spec_done_ds = targetm.sched.get_insn_spec_ds (insn); | |
2871 | spec_done_ds = ds_get_max_dep_weak (spec_done_ds); | |
2872 | } | |
2873 | else | |
2874 | spec_done_ds = 0; | |
2875 | ||
2876 | /* Initialize INSN's expr. */ | |
2877 | init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0, | |
2878 | REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn), | |
b8698a0f | 2879 | spec_done_ds, 0, 0, NULL, true, false, false, false, |
e855c69d AB |
2880 | CANT_MOVE (insn)); |
2881 | } | |
2882 | ||
2883 | init_first_time_insn_data (insn); | |
2884 | } | |
2885 | ||
2886 | /* Scan the region and initialize instruction data for basic blocks BBS. */ | |
2887 | void | |
2888 | sel_init_global_and_expr (bb_vec_t bbs) | |
2889 | { | |
2890 | /* ??? It would be nice to implement push / pop scheme for sched_infos. */ | |
2891 | const struct sched_scan_info_def ssi = | |
2892 | { | |
2893 | NULL, /* extend_bb */ | |
2894 | init_global_and_expr_for_bb, /* init_bb */ | |
2895 | extend_insn_data, /* extend_insn */ | |
2896 | init_global_and_expr_for_insn /* init_insn */ | |
2897 | }; | |
b8698a0f | 2898 | |
e855c69d AB |
2899 | sched_scan (&ssi, bbs, NULL, NULL, NULL); |
2900 | } | |
2901 | ||
2902 | /* Finalize region-scope data structures for basic blocks. */ | |
2903 | static void | |
2904 | finish_global_and_expr_for_bb (basic_block bb) | |
2905 | { | |
2906 | av_set_clear (&BB_AV_SET (bb)); | |
2907 | BB_AV_LEVEL (bb) = 0; | |
2908 | } | |
2909 | ||
2910 | /* Finalize INSN's data. */ | |
2911 | static void | |
2912 | finish_global_and_expr_insn (insn_t insn) | |
2913 | { | |
2914 | if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2915 | return; | |
2916 | ||
2917 | gcc_assert (INSN_P (insn)); | |
2918 | ||
2919 | if (INSN_LUID (insn) > 0) | |
2920 | { | |
2921 | free_first_time_insn_data (insn); | |
2922 | INSN_WS_LEVEL (insn) = 0; | |
2923 | CANT_MOVE (insn) = 0; | |
b8698a0f L |
2924 | |
2925 | /* We can no longer assert this, as vinsns of this insn could be | |
2926 | easily live in other insn's caches. This should be changed to | |
e855c69d AB |
2927 | a counter-like approach among all vinsns. */ |
2928 | gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); | |
2929 | clear_expr (INSN_EXPR (insn)); | |
2930 | } | |
2931 | } | |
2932 | ||
2933 | /* Finalize per instruction data for the whole region. */ | |
2934 | void | |
2935 | sel_finish_global_and_expr (void) | |
2936 | { | |
2937 | { | |
2938 | bb_vec_t bbs; | |
2939 | int i; | |
2940 | ||
2941 | bbs = VEC_alloc (basic_block, heap, current_nr_blocks); | |
2942 | ||
2943 | for (i = 0; i < current_nr_blocks; i++) | |
2944 | VEC_quick_push (basic_block, bbs, BASIC_BLOCK (BB_TO_BLOCK (i))); | |
2945 | ||
2946 | /* Clear AV_SETs and INSN_EXPRs. */ | |
2947 | { | |
2948 | const struct sched_scan_info_def ssi = | |
2949 | { | |
2950 | NULL, /* extend_bb */ | |
2951 | finish_global_and_expr_for_bb, /* init_bb */ | |
2952 | NULL, /* extend_insn */ | |
2953 | finish_global_and_expr_insn /* init_insn */ | |
2954 | }; | |
2955 | ||
2956 | sched_scan (&ssi, bbs, NULL, NULL, NULL); | |
2957 | } | |
2958 | ||
2959 | VEC_free (basic_block, heap, bbs); | |
2960 | } | |
2961 | ||
2962 | finish_insns (); | |
2963 | } | |
2964 | \f | |
2965 | ||
b8698a0f L |
2966 | /* In the below hooks, we merely calculate whether or not a dependence |
2967 | exists, and in what part of insn. However, we will need more data | |
e855c69d AB |
2968 | when we'll start caching dependence requests. */ |
2969 | ||
2970 | /* Container to hold information for dependency analysis. */ | |
2971 | static struct | |
2972 | { | |
2973 | deps_t dc; | |
2974 | ||
2975 | /* A variable to track which part of rtx we are scanning in | |
2976 | sched-deps.c: sched_analyze_insn (). */ | |
2977 | deps_where_t where; | |
2978 | ||
2979 | /* Current producer. */ | |
2980 | insn_t pro; | |
2981 | ||
2982 | /* Current consumer. */ | |
2983 | vinsn_t con; | |
2984 | ||
2985 | /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. | |
2986 | X is from { INSN, LHS, RHS }. */ | |
2987 | ds_t has_dep_p[DEPS_IN_NOWHERE]; | |
2988 | } has_dependence_data; | |
2989 | ||
2990 | /* Start analyzing dependencies of INSN. */ | |
2991 | static void | |
2992 | has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) | |
2993 | { | |
2994 | gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); | |
2995 | ||
2996 | has_dependence_data.where = DEPS_IN_INSN; | |
2997 | } | |
2998 | ||
2999 | /* Finish analyzing dependencies of an insn. */ | |
3000 | static void | |
3001 | has_dependence_finish_insn (void) | |
3002 | { | |
3003 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3004 | ||
3005 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3006 | } | |
3007 | ||
3008 | /* Start analyzing dependencies of LHS. */ | |
3009 | static void | |
3010 | has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) | |
3011 | { | |
3012 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3013 | ||
3014 | if (VINSN_LHS (has_dependence_data.con) != NULL) | |
3015 | has_dependence_data.where = DEPS_IN_LHS; | |
3016 | } | |
3017 | ||
3018 | /* Finish analyzing dependencies of an lhs. */ | |
3019 | static void | |
3020 | has_dependence_finish_lhs (void) | |
3021 | { | |
3022 | has_dependence_data.where = DEPS_IN_INSN; | |
3023 | } | |
3024 | ||
3025 | /* Start analyzing dependencies of RHS. */ | |
3026 | static void | |
3027 | has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) | |
3028 | { | |
3029 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3030 | ||
3031 | if (VINSN_RHS (has_dependence_data.con) != NULL) | |
3032 | has_dependence_data.where = DEPS_IN_RHS; | |
3033 | } | |
3034 | ||
3035 | /* Start analyzing dependencies of an rhs. */ | |
3036 | static void | |
3037 | has_dependence_finish_rhs (void) | |
3038 | { | |
3039 | gcc_assert (has_dependence_data.where == DEPS_IN_RHS | |
3040 | || has_dependence_data.where == DEPS_IN_INSN); | |
3041 | ||
3042 | has_dependence_data.where = DEPS_IN_INSN; | |
3043 | } | |
3044 | ||
3045 | /* Note a set of REGNO. */ | |
3046 | static void | |
3047 | has_dependence_note_reg_set (int regno) | |
3048 | { | |
3049 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3050 | ||
3051 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3052 | VINSN_INSN_RTX | |
3053 | (has_dependence_data.con))) | |
3054 | { | |
3055 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3056 | ||
3057 | if (reg_last->sets != NULL | |
3058 | || reg_last->clobbers != NULL) | |
3059 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
3060 | ||
3061 | if (reg_last->uses) | |
3062 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3063 | } | |
3064 | } | |
3065 | ||
3066 | /* Note a clobber of REGNO. */ | |
3067 | static void | |
3068 | has_dependence_note_reg_clobber (int regno) | |
3069 | { | |
3070 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3071 | ||
3072 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3073 | VINSN_INSN_RTX | |
3074 | (has_dependence_data.con))) | |
3075 | { | |
3076 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3077 | ||
3078 | if (reg_last->sets) | |
3079 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
b8698a0f | 3080 | |
e855c69d AB |
3081 | if (reg_last->uses) |
3082 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3083 | } | |
3084 | } | |
3085 | ||
3086 | /* Note a use of REGNO. */ | |
3087 | static void | |
3088 | has_dependence_note_reg_use (int regno) | |
3089 | { | |
3090 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3091 | ||
3092 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3093 | VINSN_INSN_RTX | |
3094 | (has_dependence_data.con))) | |
3095 | { | |
3096 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3097 | ||
3098 | if (reg_last->sets) | |
3099 | *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; | |
3100 | ||
3101 | if (reg_last->clobbers) | |
3102 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3103 | ||
3104 | /* Handle BE_IN_SPEC. */ | |
3105 | if (reg_last->uses) | |
3106 | { | |
3107 | ds_t pro_spec_checked_ds; | |
3108 | ||
3109 | pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); | |
3110 | pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); | |
3111 | ||
3112 | if (pro_spec_checked_ds != 0) | |
3113 | /* Merge BE_IN_SPEC bits into *DSP. */ | |
3114 | *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, | |
3115 | NULL_RTX, NULL_RTX); | |
3116 | } | |
3117 | } | |
3118 | } | |
3119 | ||
3120 | /* Note a memory dependence. */ | |
3121 | static void | |
3122 | has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, | |
3123 | rtx pending_mem ATTRIBUTE_UNUSED, | |
3124 | insn_t pending_insn ATTRIBUTE_UNUSED, | |
3125 | ds_t ds ATTRIBUTE_UNUSED) | |
3126 | { | |
3127 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3128 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3129 | { | |
3130 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3131 | ||
3132 | *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); | |
3133 | } | |
3134 | } | |
3135 | ||
3136 | /* Note a dependence. */ | |
3137 | static void | |
3138 | has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED, | |
3139 | ds_t ds ATTRIBUTE_UNUSED) | |
3140 | { | |
3141 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3142 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3143 | { | |
3144 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3145 | ||
3146 | *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); | |
3147 | } | |
3148 | } | |
3149 | ||
3150 | /* Mark the insn as having a hard dependence that prevents speculation. */ | |
3151 | void | |
3152 | sel_mark_hard_insn (rtx insn) | |
3153 | { | |
3154 | int i; | |
3155 | ||
3156 | /* Only work when we're in has_dependence_p mode. | |
3157 | ??? This is a hack, this should actually be a hook. */ | |
3158 | if (!has_dependence_data.dc || !has_dependence_data.pro) | |
3159 | return; | |
3160 | ||
3161 | gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); | |
3162 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3163 | ||
3164 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3165 | has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; | |
3166 | } | |
3167 | ||
3168 | /* This structure holds the hooks for the dependency analysis used when | |
3169 | actually processing dependencies in the scheduler. */ | |
3170 | static struct sched_deps_info_def has_dependence_sched_deps_info; | |
3171 | ||
3172 | /* This initializes most of the fields of the above structure. */ | |
3173 | static const struct sched_deps_info_def const_has_dependence_sched_deps_info = | |
3174 | { | |
3175 | NULL, | |
3176 | ||
3177 | has_dependence_start_insn, | |
3178 | has_dependence_finish_insn, | |
3179 | has_dependence_start_lhs, | |
3180 | has_dependence_finish_lhs, | |
3181 | has_dependence_start_rhs, | |
3182 | has_dependence_finish_rhs, | |
3183 | has_dependence_note_reg_set, | |
3184 | has_dependence_note_reg_clobber, | |
3185 | has_dependence_note_reg_use, | |
3186 | has_dependence_note_mem_dep, | |
3187 | has_dependence_note_dep, | |
3188 | ||
3189 | 0, /* use_cselib */ | |
3190 | 0, /* use_deps_list */ | |
3191 | 0 /* generate_spec_deps */ | |
3192 | }; | |
3193 | ||
3194 | /* Initialize has_dependence_sched_deps_info with extra spec field. */ | |
3195 | static void | |
3196 | setup_has_dependence_sched_deps_info (void) | |
3197 | { | |
3198 | memcpy (&has_dependence_sched_deps_info, | |
3199 | &const_has_dependence_sched_deps_info, | |
3200 | sizeof (has_dependence_sched_deps_info)); | |
3201 | ||
3202 | if (spec_info != NULL) | |
3203 | has_dependence_sched_deps_info.generate_spec_deps = 1; | |
3204 | ||
3205 | sched_deps_info = &has_dependence_sched_deps_info; | |
3206 | } | |
3207 | ||
3208 | /* Remove all dependences found and recorded in has_dependence_data array. */ | |
3209 | void | |
3210 | sel_clear_has_dependence (void) | |
3211 | { | |
3212 | int i; | |
3213 | ||
3214 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3215 | has_dependence_data.has_dep_p[i] = 0; | |
3216 | } | |
3217 | ||
3218 | /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer | |
3219 | to the dependence information array in HAS_DEP_PP. */ | |
3220 | ds_t | |
3221 | has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) | |
3222 | { | |
3223 | int i; | |
3224 | ds_t ds; | |
3225 | struct deps *dc; | |
3226 | ||
3227 | if (INSN_SIMPLEJUMP_P (pred)) | |
3228 | /* Unconditional jump is just a transfer of control flow. | |
3229 | Ignore it. */ | |
3230 | return false; | |
3231 | ||
3232 | dc = &INSN_DEPS_CONTEXT (pred); | |
bcf33775 AB |
3233 | |
3234 | /* We init this field lazily. */ | |
3235 | if (dc->reg_last == NULL) | |
3236 | init_deps_reg_last (dc); | |
b8698a0f | 3237 | |
e855c69d AB |
3238 | if (!dc->readonly) |
3239 | { | |
3240 | has_dependence_data.pro = NULL; | |
3241 | /* Initialize empty dep context with information about PRED. */ | |
3242 | advance_deps_context (dc, pred); | |
3243 | dc->readonly = 1; | |
3244 | } | |
3245 | ||
3246 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3247 | has_dependence_data.pro = pred; | |
3248 | has_dependence_data.con = EXPR_VINSN (expr); | |
3249 | has_dependence_data.dc = dc; | |
3250 | ||
3251 | sel_clear_has_dependence (); | |
3252 | ||
3253 | /* Now catch all dependencies that would be generated between PRED and | |
3254 | INSN. */ | |
3255 | setup_has_dependence_sched_deps_info (); | |
3256 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3257 | has_dependence_data.dc = NULL; | |
3258 | ||
3259 | /* When a barrier was found, set DEPS_IN_INSN bits. */ | |
3260 | if (dc->last_reg_pending_barrier == TRUE_BARRIER) | |
3261 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; | |
3262 | else if (dc->last_reg_pending_barrier == MOVE_BARRIER) | |
3263 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
3264 | ||
3265 | /* Do not allow stores to memory to move through checks. Currently | |
3266 | we don't move this to sched-deps.c as the check doesn't have | |
b8698a0f | 3267 | obvious places to which this dependence can be attached. |
e855c69d AB |
3268 | FIMXE: this should go to a hook. */ |
3269 | if (EXPR_LHS (expr) | |
3270 | && MEM_P (EXPR_LHS (expr)) | |
3271 | && sel_insn_is_speculation_check (pred)) | |
3272 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
b8698a0f | 3273 | |
e855c69d AB |
3274 | *has_dep_pp = has_dependence_data.has_dep_p; |
3275 | ds = 0; | |
3276 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3277 | ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], | |
3278 | NULL_RTX, NULL_RTX); | |
3279 | ||
3280 | return ds; | |
3281 | } | |
3282 | \f | |
3283 | ||
b8698a0f L |
3284 | /* Dependence hooks implementation that checks dependence latency constraints |
3285 | on the insns being scheduled. The entry point for these routines is | |
3286 | tick_check_p predicate. */ | |
e855c69d AB |
3287 | |
3288 | static struct | |
3289 | { | |
3290 | /* An expr we are currently checking. */ | |
3291 | expr_t expr; | |
3292 | ||
3293 | /* A minimal cycle for its scheduling. */ | |
3294 | int cycle; | |
3295 | ||
3296 | /* Whether we have seen a true dependence while checking. */ | |
3297 | bool seen_true_dep_p; | |
3298 | } tick_check_data; | |
3299 | ||
3300 | /* Update minimal scheduling cycle for tick_check_insn given that it depends | |
3301 | on PRO with status DS and weight DW. */ | |
3302 | static void | |
3303 | tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) | |
3304 | { | |
3305 | expr_t con_expr = tick_check_data.expr; | |
3306 | insn_t con_insn = EXPR_INSN_RTX (con_expr); | |
3307 | ||
3308 | if (con_insn != pro_insn) | |
3309 | { | |
3310 | enum reg_note dt; | |
3311 | int tick; | |
3312 | ||
3313 | if (/* PROducer was removed from above due to pipelining. */ | |
3314 | !INSN_IN_STREAM_P (pro_insn) | |
3315 | /* Or PROducer was originally on the next iteration regarding the | |
3316 | CONsumer. */ | |
3317 | || (INSN_SCHED_TIMES (pro_insn) | |
3318 | - EXPR_SCHED_TIMES (con_expr)) > 1) | |
3319 | /* Don't count this dependence. */ | |
3320 | return; | |
3321 | ||
3322 | dt = ds_to_dt (ds); | |
3323 | if (dt == REG_DEP_TRUE) | |
3324 | tick_check_data.seen_true_dep_p = true; | |
3325 | ||
3326 | gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); | |
3327 | ||
3328 | { | |
3329 | dep_def _dep, *dep = &_dep; | |
3330 | ||
3331 | init_dep (dep, pro_insn, con_insn, dt); | |
3332 | ||
3333 | tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); | |
3334 | } | |
3335 | ||
3336 | /* When there are several kinds of dependencies between pro and con, | |
3337 | only REG_DEP_TRUE should be taken into account. */ | |
3338 | if (tick > tick_check_data.cycle | |
3339 | && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) | |
3340 | tick_check_data.cycle = tick; | |
3341 | } | |
3342 | } | |
3343 | ||
3344 | /* An implementation of note_dep hook. */ | |
3345 | static void | |
3346 | tick_check_note_dep (insn_t pro, ds_t ds) | |
3347 | { | |
3348 | tick_check_dep_with_dw (pro, ds, 0); | |
3349 | } | |
3350 | ||
3351 | /* An implementation of note_mem_dep hook. */ | |
3352 | static void | |
3353 | tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) | |
3354 | { | |
3355 | dw_t dw; | |
3356 | ||
3357 | dw = (ds_to_dt (ds) == REG_DEP_TRUE | |
3358 | ? estimate_dep_weak (mem1, mem2) | |
3359 | : 0); | |
3360 | ||
3361 | tick_check_dep_with_dw (pro, ds, dw); | |
3362 | } | |
3363 | ||
3364 | /* This structure contains hooks for dependence analysis used when determining | |
3365 | whether an insn is ready for scheduling. */ | |
3366 | static struct sched_deps_info_def tick_check_sched_deps_info = | |
3367 | { | |
3368 | NULL, | |
3369 | ||
3370 | NULL, | |
3371 | NULL, | |
3372 | NULL, | |
3373 | NULL, | |
3374 | NULL, | |
3375 | NULL, | |
3376 | haifa_note_reg_set, | |
3377 | haifa_note_reg_clobber, | |
3378 | haifa_note_reg_use, | |
3379 | tick_check_note_mem_dep, | |
3380 | tick_check_note_dep, | |
3381 | ||
3382 | 0, 0, 0 | |
3383 | }; | |
3384 | ||
3385 | /* Estimate number of cycles from the current cycle of FENCE until EXPR can be | |
3386 | scheduled. Return 0 if all data from producers in DC is ready. */ | |
3387 | int | |
3388 | tick_check_p (expr_t expr, deps_t dc, fence_t fence) | |
3389 | { | |
3390 | int cycles_left; | |
3391 | /* Initialize variables. */ | |
3392 | tick_check_data.expr = expr; | |
3393 | tick_check_data.cycle = 0; | |
3394 | tick_check_data.seen_true_dep_p = false; | |
3395 | sched_deps_info = &tick_check_sched_deps_info; | |
b8698a0f | 3396 | |
e855c69d AB |
3397 | gcc_assert (!dc->readonly); |
3398 | dc->readonly = 1; | |
3399 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3400 | dc->readonly = 0; | |
3401 | ||
3402 | cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); | |
3403 | ||
3404 | return cycles_left >= 0 ? cycles_left : 0; | |
3405 | } | |
3406 | \f | |
3407 | ||
3408 | /* Functions to work with insns. */ | |
3409 | ||
3410 | /* Returns true if LHS of INSN is the same as DEST of an insn | |
3411 | being moved. */ | |
3412 | bool | |
3413 | lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) | |
3414 | { | |
3415 | rtx lhs = INSN_LHS (insn); | |
3416 | ||
3417 | if (lhs == NULL || dest == NULL) | |
3418 | return false; | |
b8698a0f | 3419 | |
e855c69d AB |
3420 | return rtx_equal_p (lhs, dest); |
3421 | } | |
3422 | ||
3423 | /* Return s_i_d entry of INSN. Callable from debugger. */ | |
3424 | sel_insn_data_def | |
3425 | insn_sid (insn_t insn) | |
3426 | { | |
3427 | return *SID (insn); | |
3428 | } | |
3429 | ||
3430 | /* True when INSN is a speculative check. We can tell this by looking | |
3431 | at the data structures of the selective scheduler, not by examining | |
3432 | the pattern. */ | |
3433 | bool | |
3434 | sel_insn_is_speculation_check (rtx insn) | |
3435 | { | |
3436 | return s_i_d && !! INSN_SPEC_CHECKED_DS (insn); | |
3437 | } | |
3438 | ||
b8698a0f | 3439 | /* Extracts machine mode MODE and destination location DST_LOC |
e855c69d AB |
3440 | for given INSN. */ |
3441 | void | |
3442 | get_dest_and_mode (rtx insn, rtx *dst_loc, enum machine_mode *mode) | |
3443 | { | |
3444 | rtx pat = PATTERN (insn); | |
3445 | ||
3446 | gcc_assert (dst_loc); | |
3447 | gcc_assert (GET_CODE (pat) == SET); | |
3448 | ||
3449 | *dst_loc = SET_DEST (pat); | |
3450 | ||
3451 | gcc_assert (*dst_loc); | |
3452 | gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); | |
3453 | ||
3454 | if (mode) | |
3455 | *mode = GET_MODE (*dst_loc); | |
3456 | } | |
3457 | ||
b8698a0f | 3458 | /* Returns true when moving through JUMP will result in bookkeeping |
e855c69d AB |
3459 | creation. */ |
3460 | bool | |
3461 | bookkeeping_can_be_created_if_moved_through_p (insn_t jump) | |
3462 | { | |
3463 | insn_t succ; | |
3464 | succ_iterator si; | |
3465 | ||
3466 | FOR_EACH_SUCC (succ, si, jump) | |
3467 | if (sel_num_cfg_preds_gt_1 (succ)) | |
3468 | return true; | |
3469 | ||
3470 | return false; | |
3471 | } | |
3472 | ||
3473 | /* Return 'true' if INSN is the only one in its basic block. */ | |
3474 | static bool | |
3475 | insn_is_the_only_one_in_bb_p (insn_t insn) | |
3476 | { | |
3477 | return sel_bb_head_p (insn) && sel_bb_end_p (insn); | |
3478 | } | |
3479 | ||
3480 | #ifdef ENABLE_CHECKING | |
b8698a0f | 3481 | /* Check that the region we're scheduling still has at most one |
e855c69d AB |
3482 | backedge. */ |
3483 | static void | |
3484 | verify_backedges (void) | |
3485 | { | |
3486 | if (pipelining_p) | |
3487 | { | |
3488 | int i, n = 0; | |
3489 | edge e; | |
3490 | edge_iterator ei; | |
b8698a0f | 3491 | |
e855c69d AB |
3492 | for (i = 0; i < current_nr_blocks; i++) |
3493 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (BB_TO_BLOCK (i))->succs) | |
3494 | if (in_current_region_p (e->dest) | |
3495 | && BLOCK_TO_BB (e->dest->index) < i) | |
3496 | n++; | |
b8698a0f | 3497 | |
e855c69d AB |
3498 | gcc_assert (n <= 1); |
3499 | } | |
3500 | } | |
3501 | #endif | |
3502 | \f | |
3503 | ||
3504 | /* Functions to work with control flow. */ | |
3505 | ||
b59ab570 AM |
3506 | /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks |
3507 | are sorted in topological order (it might have been invalidated by | |
3508 | redirecting an edge). */ | |
3509 | static void | |
3510 | sel_recompute_toporder (void) | |
3511 | { | |
3512 | int i, n, rgn; | |
3513 | int *postorder, n_blocks; | |
3514 | ||
3515 | postorder = XALLOCAVEC (int, n_basic_blocks); | |
3516 | n_blocks = post_order_compute (postorder, false, false); | |
3517 | ||
3518 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
3519 | for (n = 0, i = n_blocks - 1; i >= 0; i--) | |
3520 | if (CONTAINING_RGN (postorder[i]) == rgn) | |
3521 | { | |
3522 | BLOCK_TO_BB (postorder[i]) = n; | |
3523 | BB_TO_BLOCK (n) = postorder[i]; | |
3524 | n++; | |
3525 | } | |
3526 | ||
3527 | /* Assert that we updated info for all blocks. We may miss some blocks if | |
3528 | this function is called when redirecting an edge made a block | |
3529 | unreachable, but that block is not deleted yet. */ | |
3530 | gcc_assert (n == RGN_NR_BLOCKS (rgn)); | |
3531 | } | |
3532 | ||
e855c69d | 3533 | /* Tidy the possibly empty block BB. */ |
b59ab570 AM |
3534 | static bool |
3535 | maybe_tidy_empty_bb (basic_block bb, bool recompute_toporder_p) | |
e855c69d AB |
3536 | { |
3537 | basic_block succ_bb, pred_bb; | |
f2c45f08 AM |
3538 | edge e; |
3539 | edge_iterator ei; | |
e855c69d AB |
3540 | bool rescan_p; |
3541 | ||
3542 | /* Keep empty bb only if this block immediately precedes EXIT and | |
3543 | has incoming non-fallthrough edge. Otherwise remove it. */ | |
b5b8b0ac | 3544 | if (!sel_bb_empty_p (bb) |
b8698a0f | 3545 | || (single_succ_p (bb) |
e855c69d | 3546 | && single_succ (bb) == EXIT_BLOCK_PTR |
b8698a0f | 3547 | && (!single_pred_p (bb) |
e855c69d AB |
3548 | || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU)))) |
3549 | return false; | |
3550 | ||
f2c45f08 AM |
3551 | /* Do not attempt to redirect complex edges. */ |
3552 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3553 | if (e->flags & EDGE_COMPLEX) | |
3554 | return false; | |
3555 | ||
e855c69d AB |
3556 | free_data_sets (bb); |
3557 | ||
3558 | /* Do not delete BB if it has more than one successor. | |
3559 | That can occur when we moving a jump. */ | |
3560 | if (!single_succ_p (bb)) | |
3561 | { | |
3562 | gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); | |
3563 | sel_merge_blocks (bb->prev_bb, bb); | |
3564 | return true; | |
3565 | } | |
3566 | ||
3567 | succ_bb = single_succ (bb); | |
3568 | rescan_p = true; | |
3569 | pred_bb = NULL; | |
3570 | ||
3571 | /* Redirect all non-fallthru edges to the next bb. */ | |
3572 | while (rescan_p) | |
3573 | { | |
e855c69d AB |
3574 | rescan_p = false; |
3575 | ||
3576 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3577 | { | |
3578 | pred_bb = e->src; | |
3579 | ||
3580 | if (!(e->flags & EDGE_FALLTHRU)) | |
3581 | { | |
b59ab570 | 3582 | recompute_toporder_p |= sel_redirect_edge_and_branch (e, succ_bb); |
e855c69d AB |
3583 | rescan_p = true; |
3584 | break; | |
3585 | } | |
3586 | } | |
3587 | } | |
3588 | ||
3589 | /* If it is possible - merge BB with its predecessor. */ | |
3590 | if (can_merge_blocks_p (bb->prev_bb, bb)) | |
3591 | sel_merge_blocks (bb->prev_bb, bb); | |
3592 | else | |
3593 | /* Otherwise this is a block without fallthru predecessor. | |
3594 | Just delete it. */ | |
3595 | { | |
3596 | gcc_assert (pred_bb != NULL); | |
3597 | ||
3598 | move_bb_info (pred_bb, bb); | |
3599 | remove_empty_bb (bb, true); | |
3600 | } | |
3601 | ||
b59ab570 AM |
3602 | if (recompute_toporder_p) |
3603 | sel_recompute_toporder (); | |
3604 | ||
e855c69d AB |
3605 | #ifdef ENABLE_CHECKING |
3606 | verify_backedges (); | |
3607 | #endif | |
3608 | ||
3609 | return true; | |
3610 | } | |
3611 | ||
b8698a0f | 3612 | /* Tidy the control flow after we have removed original insn from |
e855c69d AB |
3613 | XBB. Return true if we have removed some blocks. When FULL_TIDYING |
3614 | is true, also try to optimize control flow on non-empty blocks. */ | |
3615 | bool | |
3616 | tidy_control_flow (basic_block xbb, bool full_tidying) | |
3617 | { | |
3618 | bool changed = true; | |
b5b8b0ac | 3619 | insn_t first, last; |
b8698a0f | 3620 | |
e855c69d | 3621 | /* First check whether XBB is empty. */ |
b59ab570 | 3622 | changed = maybe_tidy_empty_bb (xbb, false); |
e855c69d AB |
3623 | if (changed || !full_tidying) |
3624 | return changed; | |
b8698a0f | 3625 | |
e855c69d AB |
3626 | /* Check if there is a unnecessary jump after insn left. */ |
3627 | if (jump_leads_only_to_bb_p (BB_END (xbb), xbb->next_bb) | |
3628 | && INSN_SCHED_TIMES (BB_END (xbb)) == 0 | |
3629 | && !IN_CURRENT_FENCE_P (BB_END (xbb))) | |
3630 | { | |
3631 | if (sel_remove_insn (BB_END (xbb), false, false)) | |
3632 | return true; | |
3633 | tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); | |
3634 | } | |
3635 | ||
b5b8b0ac AO |
3636 | first = sel_bb_head (xbb); |
3637 | last = sel_bb_end (xbb); | |
3638 | if (MAY_HAVE_DEBUG_INSNS) | |
3639 | { | |
3640 | if (first != last && DEBUG_INSN_P (first)) | |
3641 | do | |
3642 | first = NEXT_INSN (first); | |
3643 | while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); | |
3644 | ||
3645 | if (first != last && DEBUG_INSN_P (last)) | |
3646 | do | |
3647 | last = PREV_INSN (last); | |
3648 | while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); | |
3649 | } | |
e855c69d | 3650 | /* Check if there is an unnecessary jump in previous basic block leading |
b8698a0f L |
3651 | to next basic block left after removing INSN from stream. |
3652 | If it is so, remove that jump and redirect edge to current | |
3653 | basic block (where there was INSN before deletion). This way | |
3654 | when NOP will be deleted several instructions later with its | |
3655 | basic block we will not get a jump to next instruction, which | |
e855c69d | 3656 | can be harmful. */ |
b5b8b0ac | 3657 | if (first == last |
e855c69d | 3658 | && !sel_bb_empty_p (xbb) |
b5b8b0ac | 3659 | && INSN_NOP_P (last) |
e855c69d AB |
3660 | /* Flow goes fallthru from current block to the next. */ |
3661 | && EDGE_COUNT (xbb->succs) == 1 | |
3662 | && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) | |
3663 | /* When successor is an EXIT block, it may not be the next block. */ | |
3664 | && single_succ (xbb) != EXIT_BLOCK_PTR | |
3665 | /* And unconditional jump in previous basic block leads to | |
3666 | next basic block of XBB and this jump can be safely removed. */ | |
3667 | && in_current_region_p (xbb->prev_bb) | |
3668 | && jump_leads_only_to_bb_p (BB_END (xbb->prev_bb), xbb->next_bb) | |
3669 | && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 | |
3670 | /* Also this jump is not at the scheduling boundary. */ | |
3671 | && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) | |
3672 | { | |
b59ab570 | 3673 | bool recompute_toporder_p; |
e855c69d AB |
3674 | /* Clear data structures of jump - jump itself will be removed |
3675 | by sel_redirect_edge_and_branch. */ | |
3676 | clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); | |
b59ab570 AM |
3677 | recompute_toporder_p |
3678 | = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); | |
3679 | ||
e855c69d AB |
3680 | gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); |
3681 | ||
3682 | /* It can turn out that after removing unused jump, basic block | |
3683 | that contained that jump, becomes empty too. In such case | |
3684 | remove it too. */ | |
3685 | if (sel_bb_empty_p (xbb->prev_bb)) | |
b59ab570 AM |
3686 | changed = maybe_tidy_empty_bb (xbb->prev_bb, recompute_toporder_p); |
3687 | else if (recompute_toporder_p) | |
3688 | sel_recompute_toporder (); | |
e855c69d AB |
3689 | } |
3690 | ||
3691 | return changed; | |
3692 | } | |
3693 | ||
b59ab570 AM |
3694 | /* Purge meaningless empty blocks in the middle of a region. */ |
3695 | void | |
3696 | purge_empty_blocks (void) | |
3697 | { | |
3698 | /* Do not attempt to delete preheader. */ | |
3699 | int i = sel_is_loop_preheader_p (BASIC_BLOCK (BB_TO_BLOCK (0))) ? 1 : 0; | |
3700 | ||
3701 | while (i < current_nr_blocks) | |
3702 | { | |
3703 | basic_block b = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
3704 | ||
3705 | if (maybe_tidy_empty_bb (b, false)) | |
3706 | continue; | |
3707 | ||
3708 | i++; | |
3709 | } | |
3710 | } | |
3711 | ||
b8698a0f L |
3712 | /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, |
3713 | do not delete insn's data, because it will be later re-emitted. | |
e855c69d AB |
3714 | Return true if we have removed some blocks afterwards. */ |
3715 | bool | |
3716 | sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) | |
3717 | { | |
3718 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3719 | ||
3720 | gcc_assert (INSN_IN_STREAM_P (insn)); | |
3721 | ||
b5b8b0ac AO |
3722 | if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) |
3723 | { | |
3724 | expr_t expr; | |
3725 | av_set_iterator i; | |
3726 | ||
3727 | /* When we remove a debug insn that is head of a BB, it remains | |
3728 | in the AV_SET of the block, but it shouldn't. */ | |
3729 | FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) | |
3730 | if (EXPR_INSN_RTX (expr) == insn) | |
3731 | { | |
3732 | av_set_iter_remove (&i); | |
3733 | break; | |
3734 | } | |
3735 | } | |
3736 | ||
e855c69d AB |
3737 | if (only_disconnect) |
3738 | { | |
3739 | insn_t prev = PREV_INSN (insn); | |
3740 | insn_t next = NEXT_INSN (insn); | |
3741 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3742 | ||
3743 | NEXT_INSN (prev) = next; | |
3744 | PREV_INSN (next) = prev; | |
3745 | ||
3746 | if (BB_HEAD (bb) == insn) | |
3747 | { | |
3748 | gcc_assert (BLOCK_FOR_INSN (prev) == bb); | |
3749 | BB_HEAD (bb) = prev; | |
3750 | } | |
3751 | if (BB_END (bb) == insn) | |
3752 | BB_END (bb) = prev; | |
3753 | } | |
3754 | else | |
3755 | { | |
3756 | remove_insn (insn); | |
3757 | clear_expr (INSN_EXPR (insn)); | |
3758 | } | |
3759 | ||
3760 | /* It is necessary to null this fields before calling add_insn (). */ | |
3761 | PREV_INSN (insn) = NULL_RTX; | |
3762 | NEXT_INSN (insn) = NULL_RTX; | |
3763 | ||
3764 | return tidy_control_flow (bb, full_tidying); | |
3765 | } | |
3766 | ||
3767 | /* Estimate number of the insns in BB. */ | |
3768 | static int | |
3769 | sel_estimate_number_of_insns (basic_block bb) | |
3770 | { | |
3771 | int res = 0; | |
3772 | insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); | |
3773 | ||
3774 | for (; insn != next_tail; insn = NEXT_INSN (insn)) | |
b5b8b0ac | 3775 | if (NONDEBUG_INSN_P (insn)) |
e855c69d AB |
3776 | res++; |
3777 | ||
3778 | return res; | |
3779 | } | |
3780 | ||
3781 | /* We don't need separate luids for notes or labels. */ | |
3782 | static int | |
3783 | sel_luid_for_non_insn (rtx x) | |
3784 | { | |
3785 | gcc_assert (NOTE_P (x) || LABEL_P (x)); | |
3786 | ||
3787 | return -1; | |
3788 | } | |
3789 | ||
3790 | /* Return seqno of the only predecessor of INSN. */ | |
3791 | static int | |
3792 | get_seqno_of_a_pred (insn_t insn) | |
3793 | { | |
3794 | int seqno; | |
3795 | ||
3796 | gcc_assert (INSN_SIMPLEJUMP_P (insn)); | |
3797 | ||
3798 | if (!sel_bb_head_p (insn)) | |
3799 | seqno = INSN_SEQNO (PREV_INSN (insn)); | |
3800 | else | |
3801 | { | |
3802 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3803 | ||
3804 | if (single_pred_p (bb) | |
3805 | && !in_current_region_p (single_pred (bb))) | |
3806 | { | |
3807 | /* We can have preds outside a region when splitting edges | |
b8698a0f | 3808 | for pipelining of an outer loop. Use succ instead. |
e855c69d AB |
3809 | There should be only one of them. */ |
3810 | insn_t succ = NULL; | |
3811 | succ_iterator si; | |
3812 | bool first = true; | |
b8698a0f | 3813 | |
e855c69d AB |
3814 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
3815 | && current_loop_nest); | |
b8698a0f | 3816 | FOR_EACH_SUCC_1 (succ, si, insn, |
e855c69d AB |
3817 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
3818 | { | |
3819 | gcc_assert (first); | |
3820 | first = false; | |
3821 | } | |
3822 | ||
3823 | gcc_assert (succ != NULL); | |
3824 | seqno = INSN_SEQNO (succ); | |
3825 | } | |
3826 | else | |
3827 | { | |
3828 | insn_t *preds; | |
3829 | int n; | |
3830 | ||
3831 | cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); | |
3832 | gcc_assert (n == 1); | |
3833 | ||
3834 | seqno = INSN_SEQNO (preds[0]); | |
b8698a0f | 3835 | |
e855c69d AB |
3836 | free (preds); |
3837 | } | |
3838 | } | |
3839 | ||
3840 | return seqno; | |
3841 | } | |
3842 | ||
da7ba240 AB |
3843 | /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors |
3844 | with positive seqno exist. */ | |
e855c69d AB |
3845 | int |
3846 | get_seqno_by_preds (rtx insn) | |
3847 | { | |
3848 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3849 | rtx tmp = insn, head = BB_HEAD (bb); | |
3850 | insn_t *preds; | |
3851 | int n, i, seqno; | |
3852 | ||
3853 | while (tmp != head) | |
3854 | if (INSN_P (tmp)) | |
3855 | return INSN_SEQNO (tmp); | |
3856 | else | |
3857 | tmp = PREV_INSN (tmp); | |
b8698a0f | 3858 | |
e855c69d AB |
3859 | cfg_preds (bb, &preds, &n); |
3860 | for (i = 0, seqno = -1; i < n; i++) | |
3861 | seqno = MAX (seqno, INSN_SEQNO (preds[i])); | |
3862 | ||
e855c69d AB |
3863 | return seqno; |
3864 | } | |
3865 | ||
3866 | \f | |
3867 | ||
3868 | /* Extend pass-scope data structures for basic blocks. */ | |
3869 | void | |
3870 | sel_extend_global_bb_info (void) | |
3871 | { | |
3872 | VEC_safe_grow_cleared (sel_global_bb_info_def, heap, sel_global_bb_info, | |
3873 | last_basic_block); | |
3874 | } | |
3875 | ||
3876 | /* Extend region-scope data structures for basic blocks. */ | |
3877 | static void | |
3878 | extend_region_bb_info (void) | |
3879 | { | |
3880 | VEC_safe_grow_cleared (sel_region_bb_info_def, heap, sel_region_bb_info, | |
3881 | last_basic_block); | |
3882 | } | |
3883 | ||
3884 | /* Extend all data structures to fit for all basic blocks. */ | |
3885 | static void | |
3886 | extend_bb_info (void) | |
3887 | { | |
3888 | sel_extend_global_bb_info (); | |
3889 | extend_region_bb_info (); | |
3890 | } | |
3891 | ||
3892 | /* Finalize pass-scope data structures for basic blocks. */ | |
3893 | void | |
3894 | sel_finish_global_bb_info (void) | |
3895 | { | |
3896 | VEC_free (sel_global_bb_info_def, heap, sel_global_bb_info); | |
3897 | } | |
3898 | ||
3899 | /* Finalize region-scope data structures for basic blocks. */ | |
3900 | static void | |
3901 | finish_region_bb_info (void) | |
3902 | { | |
3903 | VEC_free (sel_region_bb_info_def, heap, sel_region_bb_info); | |
3904 | } | |
3905 | \f | |
3906 | ||
3907 | /* Data for each insn in current region. */ | |
3908 | VEC (sel_insn_data_def, heap) *s_i_d = NULL; | |
3909 | ||
3910 | /* A vector for the insns we've emitted. */ | |
3911 | static insn_vec_t new_insns = NULL; | |
3912 | ||
3913 | /* Extend data structures for insns from current region. */ | |
3914 | static void | |
3915 | extend_insn_data (void) | |
3916 | { | |
3917 | int reserve; | |
b8698a0f | 3918 | |
e855c69d AB |
3919 | sched_extend_target (); |
3920 | sched_deps_init (false); | |
3921 | ||
3922 | /* Extend data structures for insns from current region. */ | |
b8698a0f | 3923 | reserve = (sched_max_luid + 1 |
e855c69d | 3924 | - VEC_length (sel_insn_data_def, s_i_d)); |
b8698a0f | 3925 | if (reserve > 0 |
e855c69d | 3926 | && ! VEC_space (sel_insn_data_def, s_i_d, reserve)) |
bcf33775 AB |
3927 | { |
3928 | int size; | |
3929 | ||
3930 | if (sched_max_luid / 2 > 1024) | |
3931 | size = sched_max_luid + 1024; | |
3932 | else | |
3933 | size = 3 * sched_max_luid / 2; | |
b8698a0f | 3934 | |
bcf33775 AB |
3935 | |
3936 | VEC_safe_grow_cleared (sel_insn_data_def, heap, s_i_d, size); | |
3937 | } | |
e855c69d AB |
3938 | } |
3939 | ||
3940 | /* Finalize data structures for insns from current region. */ | |
3941 | static void | |
3942 | finish_insns (void) | |
3943 | { | |
3944 | unsigned i; | |
3945 | ||
3946 | /* Clear here all dependence contexts that may have left from insns that were | |
3947 | removed during the scheduling. */ | |
3948 | for (i = 0; i < VEC_length (sel_insn_data_def, s_i_d); i++) | |
3949 | { | |
3950 | sel_insn_data_def *sid_entry = VEC_index (sel_insn_data_def, s_i_d, i); | |
b8698a0f | 3951 | |
e855c69d AB |
3952 | if (sid_entry->live) |
3953 | return_regset_to_pool (sid_entry->live); | |
3954 | if (sid_entry->analyzed_deps) | |
3955 | { | |
3956 | BITMAP_FREE (sid_entry->analyzed_deps); | |
3957 | BITMAP_FREE (sid_entry->found_deps); | |
3958 | htab_delete (sid_entry->transformed_insns); | |
3959 | free_deps (&sid_entry->deps_context); | |
3960 | } | |
3961 | if (EXPR_VINSN (&sid_entry->expr)) | |
3962 | { | |
3963 | clear_expr (&sid_entry->expr); | |
b8698a0f | 3964 | |
e855c69d AB |
3965 | /* Also, clear CANT_MOVE bit here, because we really don't want it |
3966 | to be passed to the next region. */ | |
3967 | CANT_MOVE_BY_LUID (i) = 0; | |
3968 | } | |
3969 | } | |
b8698a0f | 3970 | |
e855c69d AB |
3971 | VEC_free (sel_insn_data_def, heap, s_i_d); |
3972 | } | |
3973 | ||
3974 | /* A proxy to pass initialization data to init_insn (). */ | |
3975 | static sel_insn_data_def _insn_init_ssid; | |
3976 | static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; | |
3977 | ||
3978 | /* If true create a new vinsn. Otherwise use the one from EXPR. */ | |
3979 | static bool insn_init_create_new_vinsn_p; | |
3980 | ||
3981 | /* Set all necessary data for initialization of the new insn[s]. */ | |
3982 | static expr_t | |
3983 | set_insn_init (expr_t expr, vinsn_t vi, int seqno) | |
3984 | { | |
3985 | expr_t x = &insn_init_ssid->expr; | |
3986 | ||
3987 | copy_expr_onside (x, expr); | |
3988 | if (vi != NULL) | |
3989 | { | |
3990 | insn_init_create_new_vinsn_p = false; | |
3991 | change_vinsn_in_expr (x, vi); | |
3992 | } | |
3993 | else | |
3994 | insn_init_create_new_vinsn_p = true; | |
3995 | ||
3996 | insn_init_ssid->seqno = seqno; | |
3997 | return x; | |
3998 | } | |
3999 | ||
4000 | /* Init data for INSN. */ | |
4001 | static void | |
4002 | init_insn_data (insn_t insn) | |
4003 | { | |
4004 | expr_t expr; | |
4005 | sel_insn_data_t ssid = insn_init_ssid; | |
4006 | ||
4007 | /* The fields mentioned below are special and hence are not being | |
4008 | propagated to the new insns. */ | |
4009 | gcc_assert (!ssid->asm_p && ssid->sched_next == NULL | |
4010 | && !ssid->after_stall_p && ssid->sched_cycle == 0); | |
4011 | gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); | |
4012 | ||
4013 | expr = INSN_EXPR (insn); | |
4014 | copy_expr (expr, &ssid->expr); | |
4015 | prepare_insn_expr (insn, ssid->seqno); | |
4016 | ||
4017 | if (insn_init_create_new_vinsn_p) | |
4018 | change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p)); | |
b8698a0f | 4019 | |
e855c69d AB |
4020 | if (first_time_insn_init (insn)) |
4021 | init_first_time_insn_data (insn); | |
4022 | } | |
4023 | ||
4024 | /* This is used to initialize spurious jumps generated by | |
4025 | sel_redirect_edge (). */ | |
4026 | static void | |
4027 | init_simplejump_data (insn_t insn) | |
4028 | { | |
4029 | init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0, | |
b8698a0f | 4030 | REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0, NULL, true, false, false, |
e855c69d AB |
4031 | false, true); |
4032 | INSN_SEQNO (insn) = get_seqno_of_a_pred (insn); | |
4033 | init_first_time_insn_data (insn); | |
4034 | } | |
4035 | ||
b8698a0f | 4036 | /* Perform deferred initialization of insns. This is used to process |
e855c69d AB |
4037 | a new jump that may be created by redirect_edge. */ |
4038 | void | |
4039 | sel_init_new_insn (insn_t insn, int flags) | |
4040 | { | |
4041 | /* We create data structures for bb when the first insn is emitted in it. */ | |
4042 | if (INSN_P (insn) | |
4043 | && INSN_IN_STREAM_P (insn) | |
4044 | && insn_is_the_only_one_in_bb_p (insn)) | |
4045 | { | |
4046 | extend_bb_info (); | |
4047 | create_initial_data_sets (BLOCK_FOR_INSN (insn)); | |
4048 | } | |
b8698a0f | 4049 | |
e855c69d AB |
4050 | if (flags & INSN_INIT_TODO_LUID) |
4051 | sched_init_luids (NULL, NULL, NULL, insn); | |
4052 | ||
4053 | if (flags & INSN_INIT_TODO_SSID) | |
4054 | { | |
4055 | extend_insn_data (); | |
4056 | init_insn_data (insn); | |
4057 | clear_expr (&insn_init_ssid->expr); | |
4058 | } | |
4059 | ||
4060 | if (flags & INSN_INIT_TODO_SIMPLEJUMP) | |
4061 | { | |
4062 | extend_insn_data (); | |
4063 | init_simplejump_data (insn); | |
4064 | } | |
b8698a0f | 4065 | |
e855c69d AB |
4066 | gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) |
4067 | == CONTAINING_RGN (BB_TO_BLOCK (0))); | |
4068 | } | |
4069 | \f | |
4070 | ||
4071 | /* Functions to init/finish work with lv sets. */ | |
4072 | ||
4073 | /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ | |
4074 | static void | |
4075 | init_lv_set (basic_block bb) | |
4076 | { | |
4077 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
4078 | ||
4079 | BB_LV_SET (bb) = get_regset_from_pool (); | |
b8698a0f | 4080 | COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); |
e855c69d AB |
4081 | BB_LV_SET_VALID_P (bb) = true; |
4082 | } | |
4083 | ||
4084 | /* Copy liveness information to BB from FROM_BB. */ | |
4085 | static void | |
4086 | copy_lv_set_from (basic_block bb, basic_block from_bb) | |
4087 | { | |
4088 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
b8698a0f | 4089 | |
e855c69d AB |
4090 | COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); |
4091 | BB_LV_SET_VALID_P (bb) = true; | |
b8698a0f | 4092 | } |
e855c69d AB |
4093 | |
4094 | /* Initialize lv set of all bb headers. */ | |
4095 | void | |
4096 | init_lv_sets (void) | |
4097 | { | |
4098 | basic_block bb; | |
4099 | ||
4100 | /* Initialize of LV sets. */ | |
4101 | FOR_EACH_BB (bb) | |
4102 | init_lv_set (bb); | |
4103 | ||
4104 | /* Don't forget EXIT_BLOCK. */ | |
4105 | init_lv_set (EXIT_BLOCK_PTR); | |
4106 | } | |
4107 | ||
4108 | /* Release lv set of HEAD. */ | |
4109 | static void | |
4110 | free_lv_set (basic_block bb) | |
4111 | { | |
4112 | gcc_assert (BB_LV_SET (bb) != NULL); | |
4113 | ||
4114 | return_regset_to_pool (BB_LV_SET (bb)); | |
4115 | BB_LV_SET (bb) = NULL; | |
4116 | BB_LV_SET_VALID_P (bb) = false; | |
4117 | } | |
4118 | ||
4119 | /* Finalize lv sets of all bb headers. */ | |
4120 | void | |
4121 | free_lv_sets (void) | |
4122 | { | |
4123 | basic_block bb; | |
4124 | ||
4125 | /* Don't forget EXIT_BLOCK. */ | |
4126 | free_lv_set (EXIT_BLOCK_PTR); | |
4127 | ||
4128 | /* Free LV sets. */ | |
4129 | FOR_EACH_BB (bb) | |
4130 | if (BB_LV_SET (bb)) | |
4131 | free_lv_set (bb); | |
4132 | } | |
4133 | ||
4134 | /* Initialize an invalid AV_SET for BB. | |
4135 | This set will be updated next time compute_av () process BB. */ | |
4136 | static void | |
4137 | invalidate_av_set (basic_block bb) | |
4138 | { | |
4139 | gcc_assert (BB_AV_LEVEL (bb) <= 0 | |
4140 | && BB_AV_SET (bb) == NULL); | |
4141 | ||
4142 | BB_AV_LEVEL (bb) = -1; | |
4143 | } | |
4144 | ||
4145 | /* Create initial data sets for BB (they will be invalid). */ | |
4146 | static void | |
4147 | create_initial_data_sets (basic_block bb) | |
4148 | { | |
4149 | if (BB_LV_SET (bb)) | |
4150 | BB_LV_SET_VALID_P (bb) = false; | |
4151 | else | |
4152 | BB_LV_SET (bb) = get_regset_from_pool (); | |
4153 | invalidate_av_set (bb); | |
4154 | } | |
4155 | ||
4156 | /* Free av set of BB. */ | |
4157 | static void | |
4158 | free_av_set (basic_block bb) | |
4159 | { | |
4160 | av_set_clear (&BB_AV_SET (bb)); | |
4161 | BB_AV_LEVEL (bb) = 0; | |
4162 | } | |
4163 | ||
4164 | /* Free data sets of BB. */ | |
4165 | void | |
4166 | free_data_sets (basic_block bb) | |
4167 | { | |
4168 | free_lv_set (bb); | |
4169 | free_av_set (bb); | |
4170 | } | |
4171 | ||
4172 | /* Exchange lv sets of TO and FROM. */ | |
4173 | static void | |
4174 | exchange_lv_sets (basic_block to, basic_block from) | |
4175 | { | |
4176 | { | |
4177 | regset to_lv_set = BB_LV_SET (to); | |
4178 | ||
4179 | BB_LV_SET (to) = BB_LV_SET (from); | |
4180 | BB_LV_SET (from) = to_lv_set; | |
4181 | } | |
4182 | ||
4183 | { | |
4184 | bool to_lv_set_valid_p = BB_LV_SET_VALID_P (to); | |
4185 | ||
4186 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4187 | BB_LV_SET_VALID_P (from) = to_lv_set_valid_p; | |
4188 | } | |
4189 | } | |
4190 | ||
4191 | ||
4192 | /* Exchange av sets of TO and FROM. */ | |
4193 | static void | |
4194 | exchange_av_sets (basic_block to, basic_block from) | |
4195 | { | |
4196 | { | |
4197 | av_set_t to_av_set = BB_AV_SET (to); | |
4198 | ||
4199 | BB_AV_SET (to) = BB_AV_SET (from); | |
4200 | BB_AV_SET (from) = to_av_set; | |
4201 | } | |
4202 | ||
4203 | { | |
4204 | int to_av_level = BB_AV_LEVEL (to); | |
4205 | ||
4206 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4207 | BB_AV_LEVEL (from) = to_av_level; | |
4208 | } | |
4209 | } | |
4210 | ||
4211 | /* Exchange data sets of TO and FROM. */ | |
4212 | void | |
4213 | exchange_data_sets (basic_block to, basic_block from) | |
4214 | { | |
4215 | exchange_lv_sets (to, from); | |
4216 | exchange_av_sets (to, from); | |
4217 | } | |
4218 | ||
4219 | /* Copy data sets of FROM to TO. */ | |
4220 | void | |
4221 | copy_data_sets (basic_block to, basic_block from) | |
4222 | { | |
4223 | gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); | |
4224 | gcc_assert (BB_AV_SET (to) == NULL); | |
4225 | ||
4226 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4227 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4228 | ||
4229 | if (BB_AV_SET_VALID_P (from)) | |
4230 | { | |
4231 | BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); | |
4232 | } | |
4233 | if (BB_LV_SET_VALID_P (from)) | |
4234 | { | |
4235 | gcc_assert (BB_LV_SET (to) != NULL); | |
4236 | COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); | |
4237 | } | |
4238 | } | |
4239 | ||
4240 | /* Return an av set for INSN, if any. */ | |
4241 | av_set_t | |
4242 | get_av_set (insn_t insn) | |
4243 | { | |
4244 | av_set_t av_set; | |
4245 | ||
4246 | gcc_assert (AV_SET_VALID_P (insn)); | |
4247 | ||
4248 | if (sel_bb_head_p (insn)) | |
4249 | av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); | |
4250 | else | |
4251 | av_set = NULL; | |
4252 | ||
4253 | return av_set; | |
4254 | } | |
4255 | ||
4256 | /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ | |
4257 | int | |
4258 | get_av_level (insn_t insn) | |
4259 | { | |
4260 | int av_level; | |
4261 | ||
4262 | gcc_assert (INSN_P (insn)); | |
4263 | ||
4264 | if (sel_bb_head_p (insn)) | |
4265 | av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); | |
4266 | else | |
4267 | av_level = INSN_WS_LEVEL (insn); | |
4268 | ||
4269 | return av_level; | |
4270 | } | |
4271 | ||
4272 | \f | |
4273 | ||
4274 | /* Variables to work with control-flow graph. */ | |
4275 | ||
4276 | /* The basic block that already has been processed by the sched_data_update (), | |
4277 | but hasn't been in sel_add_bb () yet. */ | |
4278 | static VEC (basic_block, heap) *last_added_blocks = NULL; | |
4279 | ||
4280 | /* A pool for allocating successor infos. */ | |
4281 | static struct | |
4282 | { | |
4283 | /* A stack for saving succs_info structures. */ | |
4284 | struct succs_info *stack; | |
4285 | ||
4286 | /* Its size. */ | |
4287 | int size; | |
4288 | ||
4289 | /* Top of the stack. */ | |
4290 | int top; | |
4291 | ||
4292 | /* Maximal value of the top. */ | |
4293 | int max_top; | |
4294 | } succs_info_pool; | |
4295 | ||
4296 | /* Functions to work with control-flow graph. */ | |
4297 | ||
4298 | /* Return basic block note of BB. */ | |
4299 | insn_t | |
4300 | sel_bb_head (basic_block bb) | |
4301 | { | |
4302 | insn_t head; | |
4303 | ||
4304 | if (bb == EXIT_BLOCK_PTR) | |
4305 | { | |
4306 | gcc_assert (exit_insn != NULL_RTX); | |
4307 | head = exit_insn; | |
4308 | } | |
4309 | else | |
4310 | { | |
4311 | insn_t note; | |
4312 | ||
4313 | note = bb_note (bb); | |
4314 | head = next_nonnote_insn (note); | |
4315 | ||
4316 | if (head && BLOCK_FOR_INSN (head) != bb) | |
4317 | head = NULL_RTX; | |
4318 | } | |
4319 | ||
4320 | return head; | |
4321 | } | |
4322 | ||
4323 | /* Return true if INSN is a basic block header. */ | |
4324 | bool | |
4325 | sel_bb_head_p (insn_t insn) | |
4326 | { | |
4327 | return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn; | |
4328 | } | |
4329 | ||
4330 | /* Return last insn of BB. */ | |
4331 | insn_t | |
4332 | sel_bb_end (basic_block bb) | |
4333 | { | |
4334 | if (sel_bb_empty_p (bb)) | |
4335 | return NULL_RTX; | |
4336 | ||
4337 | gcc_assert (bb != EXIT_BLOCK_PTR); | |
4338 | ||
4339 | return BB_END (bb); | |
4340 | } | |
4341 | ||
4342 | /* Return true if INSN is the last insn in its basic block. */ | |
4343 | bool | |
4344 | sel_bb_end_p (insn_t insn) | |
4345 | { | |
4346 | return insn == sel_bb_end (BLOCK_FOR_INSN (insn)); | |
4347 | } | |
4348 | ||
4349 | /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ | |
4350 | bool | |
4351 | sel_bb_empty_p (basic_block bb) | |
4352 | { | |
4353 | return sel_bb_head (bb) == NULL; | |
4354 | } | |
4355 | ||
4356 | /* True when BB belongs to the current scheduling region. */ | |
4357 | bool | |
4358 | in_current_region_p (basic_block bb) | |
4359 | { | |
4360 | if (bb->index < NUM_FIXED_BLOCKS) | |
4361 | return false; | |
4362 | ||
4363 | return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4364 | } | |
4365 | ||
4366 | /* Return the block which is a fallthru bb of a conditional jump JUMP. */ | |
4367 | basic_block | |
4368 | fallthru_bb_of_jump (rtx jump) | |
4369 | { | |
4370 | if (!JUMP_P (jump)) | |
4371 | return NULL; | |
4372 | ||
4373 | if (any_uncondjump_p (jump)) | |
4374 | return single_succ (BLOCK_FOR_INSN (jump)); | |
4375 | ||
4376 | if (!any_condjump_p (jump)) | |
4377 | return NULL; | |
4378 | ||
268bab85 AB |
4379 | /* A basic block that ends with a conditional jump may still have one successor |
4380 | (and be followed by a barrier), we are not interested. */ | |
4381 | if (single_succ_p (BLOCK_FOR_INSN (jump))) | |
4382 | return NULL; | |
4383 | ||
e855c69d AB |
4384 | return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; |
4385 | } | |
4386 | ||
4387 | /* Remove all notes from BB. */ | |
4388 | static void | |
4389 | init_bb (basic_block bb) | |
4390 | { | |
4391 | remove_notes (bb_note (bb), BB_END (bb)); | |
4392 | BB_NOTE_LIST (bb) = note_list; | |
4393 | } | |
4394 | ||
4395 | void | |
4396 | sel_init_bbs (bb_vec_t bbs, basic_block bb) | |
4397 | { | |
4398 | const struct sched_scan_info_def ssi = | |
4399 | { | |
4400 | extend_bb_info, /* extend_bb */ | |
4401 | init_bb, /* init_bb */ | |
4402 | NULL, /* extend_insn */ | |
4403 | NULL /* init_insn */ | |
4404 | }; | |
4405 | ||
4406 | sched_scan (&ssi, bbs, bb, new_insns, NULL); | |
4407 | } | |
4408 | ||
4409 | /* Restore other notes for the whole region. */ | |
4410 | static void | |
4411 | sel_restore_other_notes (void) | |
4412 | { | |
4413 | int bb; | |
4414 | ||
4415 | for (bb = 0; bb < current_nr_blocks; bb++) | |
4416 | { | |
4417 | basic_block first, last; | |
4418 | ||
4419 | first = EBB_FIRST_BB (bb); | |
4420 | last = EBB_LAST_BB (bb)->next_bb; | |
4421 | ||
4422 | do | |
4423 | { | |
4424 | note_list = BB_NOTE_LIST (first); | |
4425 | restore_other_notes (NULL, first); | |
4426 | BB_NOTE_LIST (first) = NULL_RTX; | |
4427 | ||
4428 | first = first->next_bb; | |
4429 | } | |
4430 | while (first != last); | |
4431 | } | |
4432 | } | |
4433 | ||
4434 | /* Free per-bb data structures. */ | |
4435 | void | |
4436 | sel_finish_bbs (void) | |
4437 | { | |
4438 | sel_restore_other_notes (); | |
4439 | ||
4440 | /* Remove current loop preheader from this loop. */ | |
4441 | if (current_loop_nest) | |
4442 | sel_remove_loop_preheader (); | |
4443 | ||
4444 | finish_region_bb_info (); | |
4445 | } | |
4446 | ||
4447 | /* Return true if INSN has a single successor of type FLAGS. */ | |
4448 | bool | |
4449 | sel_insn_has_single_succ_p (insn_t insn, int flags) | |
4450 | { | |
4451 | insn_t succ; | |
4452 | succ_iterator si; | |
4453 | bool first_p = true; | |
4454 | ||
4455 | FOR_EACH_SUCC_1 (succ, si, insn, flags) | |
4456 | { | |
4457 | if (first_p) | |
4458 | first_p = false; | |
4459 | else | |
4460 | return false; | |
4461 | } | |
4462 | ||
4463 | return true; | |
4464 | } | |
4465 | ||
4466 | /* Allocate successor's info. */ | |
4467 | static struct succs_info * | |
4468 | alloc_succs_info (void) | |
4469 | { | |
4470 | if (succs_info_pool.top == succs_info_pool.max_top) | |
4471 | { | |
4472 | int i; | |
b8698a0f | 4473 | |
e855c69d AB |
4474 | if (++succs_info_pool.max_top >= succs_info_pool.size) |
4475 | gcc_unreachable (); | |
4476 | ||
4477 | i = ++succs_info_pool.top; | |
4478 | succs_info_pool.stack[i].succs_ok = VEC_alloc (rtx, heap, 10); | |
4479 | succs_info_pool.stack[i].succs_other = VEC_alloc (rtx, heap, 10); | |
4480 | succs_info_pool.stack[i].probs_ok = VEC_alloc (int, heap, 10); | |
4481 | } | |
4482 | else | |
4483 | succs_info_pool.top++; | |
4484 | ||
4485 | return &succs_info_pool.stack[succs_info_pool.top]; | |
4486 | } | |
4487 | ||
4488 | /* Free successor's info. */ | |
4489 | void | |
4490 | free_succs_info (struct succs_info * sinfo) | |
4491 | { | |
b8698a0f | 4492 | gcc_assert (succs_info_pool.top >= 0 |
e855c69d AB |
4493 | && &succs_info_pool.stack[succs_info_pool.top] == sinfo); |
4494 | succs_info_pool.top--; | |
4495 | ||
4496 | /* Clear stale info. */ | |
b8698a0f | 4497 | VEC_block_remove (rtx, sinfo->succs_ok, |
e855c69d | 4498 | 0, VEC_length (rtx, sinfo->succs_ok)); |
b8698a0f | 4499 | VEC_block_remove (rtx, sinfo->succs_other, |
e855c69d | 4500 | 0, VEC_length (rtx, sinfo->succs_other)); |
b8698a0f | 4501 | VEC_block_remove (int, sinfo->probs_ok, |
e855c69d AB |
4502 | 0, VEC_length (int, sinfo->probs_ok)); |
4503 | sinfo->all_prob = 0; | |
4504 | sinfo->succs_ok_n = 0; | |
4505 | sinfo->all_succs_n = 0; | |
4506 | } | |
4507 | ||
b8698a0f | 4508 | /* Compute successor info for INSN. FLAGS are the flags passed |
e855c69d AB |
4509 | to the FOR_EACH_SUCC_1 iterator. */ |
4510 | struct succs_info * | |
4511 | compute_succs_info (insn_t insn, short flags) | |
4512 | { | |
4513 | succ_iterator si; | |
4514 | insn_t succ; | |
4515 | struct succs_info *sinfo = alloc_succs_info (); | |
4516 | ||
4517 | /* Traverse *all* successors and decide what to do with each. */ | |
4518 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
4519 | { | |
4520 | /* FIXME: this doesn't work for skipping to loop exits, as we don't | |
4521 | perform code motion through inner loops. */ | |
4522 | short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; | |
4523 | ||
4524 | if (current_flags & flags) | |
4525 | { | |
4526 | VEC_safe_push (rtx, heap, sinfo->succs_ok, succ); | |
4527 | VEC_safe_push (int, heap, sinfo->probs_ok, | |
b8698a0f | 4528 | /* FIXME: Improve calculation when skipping |
e855c69d | 4529 | inner loop to exits. */ |
b8698a0f L |
4530 | (si.bb_end |
4531 | ? si.e1->probability | |
e855c69d AB |
4532 | : REG_BR_PROB_BASE)); |
4533 | sinfo->succs_ok_n++; | |
4534 | } | |
4535 | else | |
4536 | VEC_safe_push (rtx, heap, sinfo->succs_other, succ); | |
4537 | ||
4538 | /* Compute all_prob. */ | |
4539 | if (!si.bb_end) | |
4540 | sinfo->all_prob = REG_BR_PROB_BASE; | |
4541 | else | |
4542 | sinfo->all_prob += si.e1->probability; | |
4543 | ||
4544 | sinfo->all_succs_n++; | |
4545 | } | |
4546 | ||
4547 | return sinfo; | |
4548 | } | |
4549 | ||
b8698a0f | 4550 | /* Return the predecessors of BB in PREDS and their number in N. |
e855c69d AB |
4551 | Empty blocks are skipped. SIZE is used to allocate PREDS. */ |
4552 | static void | |
4553 | cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) | |
4554 | { | |
4555 | edge e; | |
4556 | edge_iterator ei; | |
4557 | ||
4558 | gcc_assert (BLOCK_TO_BB (bb->index) != 0); | |
4559 | ||
4560 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4561 | { | |
4562 | basic_block pred_bb = e->src; | |
4563 | insn_t bb_end = BB_END (pred_bb); | |
4564 | ||
4565 | /* ??? This code is not supposed to walk out of a region. */ | |
4566 | gcc_assert (in_current_region_p (pred_bb)); | |
4567 | ||
4568 | if (sel_bb_empty_p (pred_bb)) | |
4569 | cfg_preds_1 (pred_bb, preds, n, size); | |
4570 | else | |
4571 | { | |
4572 | if (*n == *size) | |
b8698a0f | 4573 | *preds = XRESIZEVEC (insn_t, *preds, |
e855c69d AB |
4574 | (*size = 2 * *size + 1)); |
4575 | (*preds)[(*n)++] = bb_end; | |
4576 | } | |
4577 | } | |
4578 | ||
4579 | gcc_assert (*n != 0); | |
4580 | } | |
4581 | ||
b8698a0f L |
4582 | /* Find all predecessors of BB and record them in PREDS and their number |
4583 | in N. Empty blocks are skipped, and only normal (forward in-region) | |
e855c69d AB |
4584 | edges are processed. */ |
4585 | static void | |
4586 | cfg_preds (basic_block bb, insn_t **preds, int *n) | |
4587 | { | |
4588 | int size = 0; | |
4589 | ||
4590 | *preds = NULL; | |
4591 | *n = 0; | |
4592 | cfg_preds_1 (bb, preds, n, &size); | |
4593 | } | |
4594 | ||
4595 | /* Returns true if we are moving INSN through join point. */ | |
4596 | bool | |
4597 | sel_num_cfg_preds_gt_1 (insn_t insn) | |
4598 | { | |
4599 | basic_block bb; | |
4600 | ||
4601 | if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) | |
4602 | return false; | |
4603 | ||
4604 | bb = BLOCK_FOR_INSN (insn); | |
4605 | ||
4606 | while (1) | |
4607 | { | |
4608 | if (EDGE_COUNT (bb->preds) > 1) | |
4609 | return true; | |
4610 | ||
4611 | gcc_assert (EDGE_PRED (bb, 0)->dest == bb); | |
4612 | bb = EDGE_PRED (bb, 0)->src; | |
4613 | ||
4614 | if (!sel_bb_empty_p (bb)) | |
4615 | break; | |
4616 | } | |
4617 | ||
4618 | return false; | |
4619 | } | |
4620 | ||
b8698a0f | 4621 | /* Returns true when BB should be the end of an ebb. Adapted from the |
e855c69d AB |
4622 | code in sched-ebb.c. */ |
4623 | bool | |
4624 | bb_ends_ebb_p (basic_block bb) | |
4625 | { | |
4626 | basic_block next_bb = bb_next_bb (bb); | |
4627 | edge e; | |
4628 | edge_iterator ei; | |
b8698a0f | 4629 | |
e855c69d AB |
4630 | if (next_bb == EXIT_BLOCK_PTR |
4631 | || bitmap_bit_p (forced_ebb_heads, next_bb->index) | |
4632 | || (LABEL_P (BB_HEAD (next_bb)) | |
4633 | /* NB: LABEL_NUSES () is not maintained outside of jump.c. | |
4634 | Work around that. */ | |
4635 | && !single_pred_p (next_bb))) | |
4636 | return true; | |
4637 | ||
4638 | if (!in_current_region_p (next_bb)) | |
4639 | return true; | |
4640 | ||
4641 | FOR_EACH_EDGE (e, ei, bb->succs) | |
4642 | if ((e->flags & EDGE_FALLTHRU) != 0) | |
4643 | { | |
4644 | gcc_assert (e->dest == next_bb); | |
4645 | ||
4646 | return false; | |
4647 | } | |
4648 | ||
4649 | return true; | |
4650 | } | |
4651 | ||
4652 | /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a | |
4653 | successor of INSN. */ | |
4654 | bool | |
4655 | in_same_ebb_p (insn_t insn, insn_t succ) | |
4656 | { | |
4657 | basic_block ptr = BLOCK_FOR_INSN (insn); | |
4658 | ||
4659 | for(;;) | |
4660 | { | |
4661 | if (ptr == BLOCK_FOR_INSN (succ)) | |
4662 | return true; | |
b8698a0f | 4663 | |
e855c69d AB |
4664 | if (bb_ends_ebb_p (ptr)) |
4665 | return false; | |
4666 | ||
4667 | ptr = bb_next_bb (ptr); | |
4668 | } | |
4669 | ||
4670 | gcc_unreachable (); | |
4671 | return false; | |
4672 | } | |
4673 | ||
4674 | /* Recomputes the reverse topological order for the function and | |
4675 | saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also | |
4676 | modified appropriately. */ | |
4677 | static void | |
4678 | recompute_rev_top_order (void) | |
4679 | { | |
4680 | int *postorder; | |
4681 | int n_blocks, i; | |
4682 | ||
4683 | if (!rev_top_order_index || rev_top_order_index_len < last_basic_block) | |
4684 | { | |
b8698a0f | 4685 | rev_top_order_index_len = last_basic_block; |
e855c69d AB |
4686 | rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, |
4687 | rev_top_order_index_len); | |
4688 | } | |
4689 | ||
4690 | postorder = XNEWVEC (int, n_basic_blocks); | |
4691 | ||
4692 | n_blocks = post_order_compute (postorder, true, false); | |
4693 | gcc_assert (n_basic_blocks == n_blocks); | |
4694 | ||
4695 | /* Build reverse function: for each basic block with BB->INDEX == K | |
4696 | rev_top_order_index[K] is it's reverse topological sort number. */ | |
4697 | for (i = 0; i < n_blocks; i++) | |
4698 | { | |
4699 | gcc_assert (postorder[i] < rev_top_order_index_len); | |
4700 | rev_top_order_index[postorder[i]] = i; | |
4701 | } | |
4702 | ||
4703 | free (postorder); | |
4704 | } | |
4705 | ||
4706 | /* Clear all flags from insns in BB that could spoil its rescheduling. */ | |
4707 | void | |
4708 | clear_outdated_rtx_info (basic_block bb) | |
4709 | { | |
4710 | rtx insn; | |
4711 | ||
4712 | FOR_BB_INSNS (bb, insn) | |
4713 | if (INSN_P (insn)) | |
4714 | { | |
4715 | SCHED_GROUP_P (insn) = 0; | |
4716 | INSN_AFTER_STALL_P (insn) = 0; | |
4717 | INSN_SCHED_TIMES (insn) = 0; | |
4718 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; | |
4719 | ||
4720 | /* We cannot use the changed caches, as previously we could ignore | |
b8698a0f | 4721 | the LHS dependence due to enabled renaming and transform |
e855c69d AB |
4722 | the expression, and currently we'll be unable to do this. */ |
4723 | htab_empty (INSN_TRANSFORMED_INSNS (insn)); | |
4724 | } | |
4725 | } | |
4726 | ||
4727 | /* Add BB_NOTE to the pool of available basic block notes. */ | |
4728 | static void | |
4729 | return_bb_to_pool (basic_block bb) | |
4730 | { | |
4731 | rtx note = bb_note (bb); | |
4732 | ||
4733 | gcc_assert (NOTE_BASIC_BLOCK (note) == bb | |
4734 | && bb->aux == NULL); | |
4735 | ||
4736 | /* It turns out that current cfg infrastructure does not support | |
4737 | reuse of basic blocks. Don't bother for now. */ | |
4738 | /*VEC_safe_push (rtx, heap, bb_note_pool, note);*/ | |
4739 | } | |
4740 | ||
4741 | /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ | |
4742 | static rtx | |
4743 | get_bb_note_from_pool (void) | |
4744 | { | |
4745 | if (VEC_empty (rtx, bb_note_pool)) | |
4746 | return NULL_RTX; | |
4747 | else | |
4748 | { | |
4749 | rtx note = VEC_pop (rtx, bb_note_pool); | |
4750 | ||
4751 | PREV_INSN (note) = NULL_RTX; | |
4752 | NEXT_INSN (note) = NULL_RTX; | |
4753 | ||
4754 | return note; | |
4755 | } | |
4756 | } | |
4757 | ||
4758 | /* Free bb_note_pool. */ | |
4759 | void | |
4760 | free_bb_note_pool (void) | |
4761 | { | |
4762 | VEC_free (rtx, heap, bb_note_pool); | |
4763 | } | |
4764 | ||
4765 | /* Setup scheduler pool and successor structure. */ | |
4766 | void | |
4767 | alloc_sched_pools (void) | |
4768 | { | |
4769 | int succs_size; | |
4770 | ||
4771 | succs_size = MAX_WS + 1; | |
b8698a0f | 4772 | succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); |
e855c69d AB |
4773 | succs_info_pool.size = succs_size; |
4774 | succs_info_pool.top = -1; | |
4775 | succs_info_pool.max_top = -1; | |
4776 | ||
b8698a0f | 4777 | sched_lists_pool = create_alloc_pool ("sel-sched-lists", |
e855c69d AB |
4778 | sizeof (struct _list_node), 500); |
4779 | } | |
4780 | ||
4781 | /* Free the pools. */ | |
4782 | void | |
4783 | free_sched_pools (void) | |
4784 | { | |
4785 | int i; | |
b8698a0f | 4786 | |
e855c69d AB |
4787 | free_alloc_pool (sched_lists_pool); |
4788 | gcc_assert (succs_info_pool.top == -1); | |
4789 | for (i = 0; i < succs_info_pool.max_top; i++) | |
4790 | { | |
4791 | VEC_free (rtx, heap, succs_info_pool.stack[i].succs_ok); | |
4792 | VEC_free (rtx, heap, succs_info_pool.stack[i].succs_other); | |
4793 | VEC_free (int, heap, succs_info_pool.stack[i].probs_ok); | |
4794 | } | |
4795 | free (succs_info_pool.stack); | |
4796 | } | |
4797 | \f | |
4798 | ||
b8698a0f | 4799 | /* Returns a position in RGN where BB can be inserted retaining |
e855c69d AB |
4800 | topological order. */ |
4801 | static int | |
4802 | find_place_to_insert_bb (basic_block bb, int rgn) | |
4803 | { | |
4804 | bool has_preds_outside_rgn = false; | |
4805 | edge e; | |
4806 | edge_iterator ei; | |
b8698a0f | 4807 | |
e855c69d AB |
4808 | /* Find whether we have preds outside the region. */ |
4809 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4810 | if (!in_current_region_p (e->src)) | |
4811 | { | |
4812 | has_preds_outside_rgn = true; | |
4813 | break; | |
4814 | } | |
b8698a0f | 4815 | |
e855c69d AB |
4816 | /* Recompute the top order -- needed when we have > 1 pred |
4817 | and in case we don't have preds outside. */ | |
4818 | if (flag_sel_sched_pipelining_outer_loops | |
4819 | && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) | |
4820 | { | |
4821 | int i, bbi = bb->index, cur_bbi; | |
4822 | ||
4823 | recompute_rev_top_order (); | |
4824 | for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) | |
4825 | { | |
4826 | cur_bbi = BB_TO_BLOCK (i); | |
b8698a0f | 4827 | if (rev_top_order_index[bbi] |
e855c69d AB |
4828 | < rev_top_order_index[cur_bbi]) |
4829 | break; | |
4830 | } | |
b8698a0f | 4831 | |
e855c69d AB |
4832 | /* We skipped the right block, so we increase i. We accomodate |
4833 | it for increasing by step later, so we decrease i. */ | |
4834 | return (i + 1) - 1; | |
4835 | } | |
4836 | else if (has_preds_outside_rgn) | |
4837 | { | |
4838 | /* This is the case when we generate an extra empty block | |
4839 | to serve as region head during pipelining. */ | |
4840 | e = EDGE_SUCC (bb, 0); | |
4841 | gcc_assert (EDGE_COUNT (bb->succs) == 1 | |
4842 | && in_current_region_p (EDGE_SUCC (bb, 0)->dest) | |
4843 | && (BLOCK_TO_BB (e->dest->index) == 0)); | |
4844 | return -1; | |
4845 | } | |
4846 | ||
4847 | /* We don't have preds outside the region. We should have | |
4848 | the only pred, because the multiple preds case comes from | |
4849 | the pipelining of outer loops, and that is handled above. | |
4850 | Just take the bbi of this single pred. */ | |
4851 | if (EDGE_COUNT (bb->succs) > 0) | |
4852 | { | |
4853 | int pred_bbi; | |
b8698a0f | 4854 | |
e855c69d | 4855 | gcc_assert (EDGE_COUNT (bb->preds) == 1); |
b8698a0f | 4856 | |
e855c69d AB |
4857 | pred_bbi = EDGE_PRED (bb, 0)->src->index; |
4858 | return BLOCK_TO_BB (pred_bbi); | |
4859 | } | |
4860 | else | |
4861 | /* BB has no successors. It is safe to put it in the end. */ | |
4862 | return current_nr_blocks - 1; | |
4863 | } | |
4864 | ||
4865 | /* Deletes an empty basic block freeing its data. */ | |
4866 | static void | |
4867 | delete_and_free_basic_block (basic_block bb) | |
4868 | { | |
4869 | gcc_assert (sel_bb_empty_p (bb)); | |
4870 | ||
4871 | if (BB_LV_SET (bb)) | |
4872 | free_lv_set (bb); | |
4873 | ||
4874 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
4875 | ||
b8698a0f L |
4876 | /* Can't assert av_set properties because we use sel_aremove_bb |
4877 | when removing loop preheader from the region. At the point of | |
e855c69d AB |
4878 | removing the preheader we already have deallocated sel_region_bb_info. */ |
4879 | gcc_assert (BB_LV_SET (bb) == NULL | |
4880 | && !BB_LV_SET_VALID_P (bb) | |
4881 | && BB_AV_LEVEL (bb) == 0 | |
4882 | && BB_AV_SET (bb) == NULL); | |
b8698a0f | 4883 | |
e855c69d AB |
4884 | delete_basic_block (bb); |
4885 | } | |
4886 | ||
4887 | /* Add BB to the current region and update the region data. */ | |
4888 | static void | |
4889 | add_block_to_current_region (basic_block bb) | |
4890 | { | |
4891 | int i, pos, bbi = -2, rgn; | |
4892 | ||
4893 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4894 | bbi = find_place_to_insert_bb (bb, rgn); | |
4895 | bbi += 1; | |
4896 | pos = RGN_BLOCKS (rgn) + bbi; | |
4897 | ||
4898 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
4899 | && ebb_head[bbi] == pos); | |
b8698a0f | 4900 | |
e855c69d AB |
4901 | /* Make a place for the new block. */ |
4902 | extend_regions (); | |
4903 | ||
4904 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
4905 | BLOCK_TO_BB (rgn_bb_table[i])++; | |
b8698a0f | 4906 | |
e855c69d AB |
4907 | memmove (rgn_bb_table + pos + 1, |
4908 | rgn_bb_table + pos, | |
4909 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
4910 | ||
4911 | /* Initialize data for BB. */ | |
4912 | rgn_bb_table[pos] = bb->index; | |
4913 | BLOCK_TO_BB (bb->index) = bbi; | |
4914 | CONTAINING_RGN (bb->index) = rgn; | |
4915 | ||
4916 | RGN_NR_BLOCKS (rgn)++; | |
b8698a0f | 4917 | |
e855c69d AB |
4918 | for (i = rgn + 1; i <= nr_regions; i++) |
4919 | RGN_BLOCKS (i)++; | |
4920 | } | |
4921 | ||
4922 | /* Remove BB from the current region and update the region data. */ | |
4923 | static void | |
4924 | remove_bb_from_region (basic_block bb) | |
4925 | { | |
4926 | int i, pos, bbi = -2, rgn; | |
4927 | ||
4928 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4929 | bbi = BLOCK_TO_BB (bb->index); | |
4930 | pos = RGN_BLOCKS (rgn) + bbi; | |
4931 | ||
4932 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
4933 | && ebb_head[bbi] == pos); | |
4934 | ||
4935 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
4936 | BLOCK_TO_BB (rgn_bb_table[i])--; | |
4937 | ||
4938 | memmove (rgn_bb_table + pos, | |
4939 | rgn_bb_table + pos + 1, | |
4940 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
4941 | ||
4942 | RGN_NR_BLOCKS (rgn)--; | |
4943 | for (i = rgn + 1; i <= nr_regions; i++) | |
4944 | RGN_BLOCKS (i)--; | |
4945 | } | |
4946 | ||
b8698a0f | 4947 | /* Add BB to the current region and update all data. If BB is NULL, add all |
e855c69d AB |
4948 | blocks from last_added_blocks vector. */ |
4949 | static void | |
4950 | sel_add_bb (basic_block bb) | |
4951 | { | |
4952 | /* Extend luids so that new notes will receive zero luids. */ | |
4953 | sched_init_luids (NULL, NULL, NULL, NULL); | |
4954 | sched_init_bbs (); | |
4955 | sel_init_bbs (last_added_blocks, NULL); | |
4956 | ||
b8698a0f | 4957 | /* When bb is passed explicitly, the vector should contain |
e855c69d AB |
4958 | the only element that equals to bb; otherwise, the vector |
4959 | should not be NULL. */ | |
4960 | gcc_assert (last_added_blocks != NULL); | |
b8698a0f | 4961 | |
e855c69d AB |
4962 | if (bb != NULL) |
4963 | { | |
4964 | gcc_assert (VEC_length (basic_block, last_added_blocks) == 1 | |
b8698a0f L |
4965 | && VEC_index (basic_block, |
4966 | last_added_blocks, 0) == bb); | |
e855c69d AB |
4967 | add_block_to_current_region (bb); |
4968 | ||
4969 | /* We associate creating/deleting data sets with the first insn | |
4970 | appearing / disappearing in the bb. */ | |
4971 | if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) | |
4972 | create_initial_data_sets (bb); | |
b8698a0f | 4973 | |
e855c69d AB |
4974 | VEC_free (basic_block, heap, last_added_blocks); |
4975 | } | |
4976 | else | |
4977 | /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ | |
4978 | { | |
4979 | int i; | |
4980 | basic_block temp_bb = NULL; | |
4981 | ||
b8698a0f | 4982 | for (i = 0; |
e855c69d AB |
4983 | VEC_iterate (basic_block, last_added_blocks, i, bb); i++) |
4984 | { | |
4985 | add_block_to_current_region (bb); | |
4986 | temp_bb = bb; | |
4987 | } | |
4988 | ||
b8698a0f | 4989 | /* We need to fetch at least one bb so we know the region |
e855c69d AB |
4990 | to update. */ |
4991 | gcc_assert (temp_bb != NULL); | |
4992 | bb = temp_bb; | |
4993 | ||
4994 | VEC_free (basic_block, heap, last_added_blocks); | |
4995 | } | |
4996 | ||
4997 | rgn_setup_region (CONTAINING_RGN (bb->index)); | |
4998 | } | |
4999 | ||
b8698a0f | 5000 | /* Remove BB from the current region and update all data. |
e855c69d AB |
5001 | If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ |
5002 | static void | |
5003 | sel_remove_bb (basic_block bb, bool remove_from_cfg_p) | |
5004 | { | |
5005 | gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); | |
b8698a0f | 5006 | |
e855c69d AB |
5007 | remove_bb_from_region (bb); |
5008 | return_bb_to_pool (bb); | |
5009 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
b8698a0f | 5010 | |
e855c69d AB |
5011 | if (remove_from_cfg_p) |
5012 | delete_and_free_basic_block (bb); | |
5013 | ||
5014 | rgn_setup_region (CONTAINING_RGN (bb->index)); | |
5015 | } | |
5016 | ||
5017 | /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ | |
5018 | static void | |
5019 | move_bb_info (basic_block merge_bb, basic_block empty_bb) | |
5020 | { | |
5021 | gcc_assert (in_current_region_p (merge_bb)); | |
5022 | ||
b8698a0f | 5023 | concat_note_lists (BB_NOTE_LIST (empty_bb), |
e855c69d AB |
5024 | &BB_NOTE_LIST (merge_bb)); |
5025 | BB_NOTE_LIST (empty_bb) = NULL_RTX; | |
5026 | ||
5027 | } | |
5028 | ||
b8698a0f L |
5029 | /* Remove an empty basic block EMPTY_BB. When MERGE_UP_P is true, we put |
5030 | EMPTY_BB's note lists into its predecessor instead of putting them | |
5031 | into the successor. When REMOVE_FROM_CFG_P is true, also remove | |
e855c69d AB |
5032 | the empty block. */ |
5033 | void | |
5034 | sel_remove_empty_bb (basic_block empty_bb, bool merge_up_p, | |
5035 | bool remove_from_cfg_p) | |
5036 | { | |
5037 | basic_block merge_bb; | |
5038 | ||
5039 | gcc_assert (sel_bb_empty_p (empty_bb)); | |
5040 | ||
5041 | if (merge_up_p) | |
5042 | { | |
5043 | merge_bb = empty_bb->prev_bb; | |
5044 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1 | |
5045 | && EDGE_PRED (empty_bb, 0)->src == merge_bb); | |
5046 | } | |
5047 | else | |
5048 | { | |
5049 | edge e; | |
5050 | edge_iterator ei; | |
5051 | ||
5052 | merge_bb = bb_next_bb (empty_bb); | |
5053 | ||
b8698a0f | 5054 | /* Redirect incoming edges (except fallthrough one) of EMPTY_BB to its |
e855c69d AB |
5055 | successor block. */ |
5056 | for (ei = ei_start (empty_bb->preds); | |
5057 | (e = ei_safe_edge (ei)); ) | |
5058 | { | |
5059 | if (! (e->flags & EDGE_FALLTHRU)) | |
5060 | sel_redirect_edge_and_branch (e, merge_bb); | |
5061 | else | |
5062 | ei_next (&ei); | |
5063 | } | |
5064 | ||
5065 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1 | |
5066 | && EDGE_SUCC (empty_bb, 0)->dest == merge_bb); | |
5067 | } | |
5068 | ||
5069 | move_bb_info (merge_bb, empty_bb); | |
5070 | remove_empty_bb (empty_bb, remove_from_cfg_p); | |
5071 | } | |
5072 | ||
5073 | /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from | |
5074 | region, but keep it in CFG. */ | |
5075 | static void | |
5076 | remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) | |
5077 | { | |
5078 | /* The block should contain just a note or a label. | |
5079 | We try to check whether it is unused below. */ | |
5080 | gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) | |
5081 | || LABEL_P (BB_HEAD (empty_bb))); | |
5082 | ||
5083 | /* If basic block has predecessors or successors, redirect them. */ | |
5084 | if (remove_from_cfg_p | |
5085 | && (EDGE_COUNT (empty_bb->preds) > 0 | |
5086 | || EDGE_COUNT (empty_bb->succs) > 0)) | |
5087 | { | |
5088 | basic_block pred; | |
5089 | basic_block succ; | |
5090 | ||
5091 | /* We need to init PRED and SUCC before redirecting edges. */ | |
5092 | if (EDGE_COUNT (empty_bb->preds) > 0) | |
5093 | { | |
5094 | edge e; | |
5095 | ||
5096 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); | |
5097 | ||
5098 | e = EDGE_PRED (empty_bb, 0); | |
5099 | gcc_assert (e->src == empty_bb->prev_bb | |
5100 | && (e->flags & EDGE_FALLTHRU)); | |
5101 | ||
5102 | pred = empty_bb->prev_bb; | |
5103 | } | |
5104 | else | |
5105 | pred = NULL; | |
5106 | ||
5107 | if (EDGE_COUNT (empty_bb->succs) > 0) | |
5108 | { | |
5109 | /* We do not check fallthruness here as above, because | |
5110 | after removing a jump the edge may actually be not fallthru. */ | |
5111 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); | |
5112 | succ = EDGE_SUCC (empty_bb, 0)->dest; | |
5113 | } | |
5114 | else | |
5115 | succ = NULL; | |
5116 | ||
5117 | if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) | |
5118 | { | |
5119 | edge e = EDGE_PRED (empty_bb, 0); | |
5120 | ||
5121 | if (e->flags & EDGE_FALLTHRU) | |
5122 | redirect_edge_succ_nodup (e, succ); | |
5123 | else | |
5124 | sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); | |
5125 | } | |
5126 | ||
5127 | if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) | |
5128 | { | |
5129 | edge e = EDGE_SUCC (empty_bb, 0); | |
5130 | ||
5131 | if (find_edge (pred, e->dest) == NULL) | |
5132 | redirect_edge_pred (e, pred); | |
5133 | } | |
5134 | } | |
5135 | ||
5136 | /* Finish removing. */ | |
5137 | sel_remove_bb (empty_bb, remove_from_cfg_p); | |
5138 | } | |
5139 | ||
b8698a0f | 5140 | /* An implementation of create_basic_block hook, which additionally updates |
e855c69d AB |
5141 | per-bb data structures. */ |
5142 | static basic_block | |
5143 | sel_create_basic_block (void *headp, void *endp, basic_block after) | |
5144 | { | |
5145 | basic_block new_bb; | |
5146 | insn_t new_bb_note; | |
b8698a0f L |
5147 | |
5148 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
e855c69d AB |
5149 | || last_added_blocks == NULL); |
5150 | ||
5151 | new_bb_note = get_bb_note_from_pool (); | |
5152 | ||
5153 | if (new_bb_note == NULL_RTX) | |
5154 | new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); | |
5155 | else | |
5156 | { | |
5157 | new_bb = create_basic_block_structure ((rtx) headp, (rtx) endp, | |
5158 | new_bb_note, after); | |
5159 | new_bb->aux = NULL; | |
5160 | } | |
5161 | ||
5162 | VEC_safe_push (basic_block, heap, last_added_blocks, new_bb); | |
5163 | ||
5164 | return new_bb; | |
5165 | } | |
5166 | ||
5167 | /* Implement sched_init_only_bb (). */ | |
5168 | static void | |
5169 | sel_init_only_bb (basic_block bb, basic_block after) | |
5170 | { | |
5171 | gcc_assert (after == NULL); | |
5172 | ||
5173 | extend_regions (); | |
5174 | rgn_make_new_region_out_of_new_block (bb); | |
5175 | } | |
5176 | ||
5177 | /* Update the latch when we've splitted or merged it from FROM block to TO. | |
5178 | This should be checked for all outer loops, too. */ | |
5179 | static void | |
5180 | change_loops_latches (basic_block from, basic_block to) | |
5181 | { | |
5182 | gcc_assert (from != to); | |
5183 | ||
5184 | if (current_loop_nest) | |
5185 | { | |
5186 | struct loop *loop; | |
5187 | ||
5188 | for (loop = current_loop_nest; loop; loop = loop_outer (loop)) | |
5189 | if (considered_for_pipelining_p (loop) && loop->latch == from) | |
5190 | { | |
5191 | gcc_assert (loop == current_loop_nest); | |
5192 | loop->latch = to; | |
5193 | gcc_assert (loop_latch_edge (loop)); | |
5194 | } | |
5195 | } | |
5196 | } | |
5197 | ||
b8698a0f | 5198 | /* Splits BB on two basic blocks, adding it to the region and extending |
e855c69d AB |
5199 | per-bb data structures. Returns the newly created bb. */ |
5200 | static basic_block | |
5201 | sel_split_block (basic_block bb, rtx after) | |
5202 | { | |
5203 | basic_block new_bb; | |
5204 | insn_t insn; | |
5205 | ||
5206 | new_bb = sched_split_block_1 (bb, after); | |
5207 | sel_add_bb (new_bb); | |
5208 | ||
5209 | /* This should be called after sel_add_bb, because this uses | |
b8698a0f | 5210 | CONTAINING_RGN for the new block, which is not yet initialized. |
e855c69d AB |
5211 | FIXME: this function may be a no-op now. */ |
5212 | change_loops_latches (bb, new_bb); | |
5213 | ||
5214 | /* Update ORIG_BB_INDEX for insns moved into the new block. */ | |
5215 | FOR_BB_INSNS (new_bb, insn) | |
5216 | if (INSN_P (insn)) | |
5217 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; | |
5218 | ||
5219 | if (sel_bb_empty_p (bb)) | |
5220 | { | |
5221 | gcc_assert (!sel_bb_empty_p (new_bb)); | |
5222 | ||
5223 | /* NEW_BB has data sets that need to be updated and BB holds | |
5224 | data sets that should be removed. Exchange these data sets | |
5225 | so that we won't lose BB's valid data sets. */ | |
5226 | exchange_data_sets (new_bb, bb); | |
5227 | free_data_sets (bb); | |
5228 | } | |
5229 | ||
5230 | if (!sel_bb_empty_p (new_bb) | |
5231 | && bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
5232 | bitmap_set_bit (blocks_to_reschedule, new_bb->index); | |
5233 | ||
5234 | return new_bb; | |
5235 | } | |
5236 | ||
5237 | /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. | |
5238 | Otherwise returns NULL. */ | |
5239 | static rtx | |
5240 | check_for_new_jump (basic_block bb, int prev_max_uid) | |
5241 | { | |
5242 | rtx end; | |
5243 | ||
5244 | end = sel_bb_end (bb); | |
5245 | if (end && INSN_UID (end) >= prev_max_uid) | |
5246 | return end; | |
5247 | return NULL; | |
5248 | } | |
5249 | ||
b8698a0f | 5250 | /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. |
e855c69d AB |
5251 | New means having UID at least equal to PREV_MAX_UID. */ |
5252 | static rtx | |
5253 | find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) | |
5254 | { | |
5255 | rtx jump; | |
5256 | ||
5257 | /* Return immediately if no new insns were emitted. */ | |
5258 | if (get_max_uid () == prev_max_uid) | |
5259 | return NULL; | |
b8698a0f | 5260 | |
e855c69d AB |
5261 | /* Now check both blocks for new jumps. It will ever be only one. */ |
5262 | if ((jump = check_for_new_jump (from, prev_max_uid))) | |
5263 | return jump; | |
5264 | ||
5265 | if (jump_bb != NULL | |
5266 | && (jump = check_for_new_jump (jump_bb, prev_max_uid))) | |
5267 | return jump; | |
5268 | return NULL; | |
5269 | } | |
5270 | ||
5271 | /* Splits E and adds the newly created basic block to the current region. | |
5272 | Returns this basic block. */ | |
5273 | basic_block | |
5274 | sel_split_edge (edge e) | |
5275 | { | |
5276 | basic_block new_bb, src, other_bb = NULL; | |
5277 | int prev_max_uid; | |
5278 | rtx jump; | |
5279 | ||
5280 | src = e->src; | |
5281 | prev_max_uid = get_max_uid (); | |
5282 | new_bb = split_edge (e); | |
5283 | ||
b8698a0f | 5284 | if (flag_sel_sched_pipelining_outer_loops |
e855c69d AB |
5285 | && current_loop_nest) |
5286 | { | |
5287 | int i; | |
5288 | basic_block bb; | |
5289 | ||
b8698a0f | 5290 | /* Some of the basic blocks might not have been added to the loop. |
e855c69d | 5291 | Add them here, until this is fixed in force_fallthru. */ |
b8698a0f | 5292 | for (i = 0; |
e855c69d AB |
5293 | VEC_iterate (basic_block, last_added_blocks, i, bb); i++) |
5294 | if (!bb->loop_father) | |
5295 | { | |
5296 | add_bb_to_loop (bb, e->dest->loop_father); | |
5297 | ||
5298 | gcc_assert (!other_bb && (new_bb->index != bb->index)); | |
5299 | other_bb = bb; | |
5300 | } | |
5301 | } | |
5302 | ||
5303 | /* Add all last_added_blocks to the region. */ | |
5304 | sel_add_bb (NULL); | |
5305 | ||
5306 | jump = find_new_jump (src, new_bb, prev_max_uid); | |
5307 | if (jump) | |
5308 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5309 | ||
5310 | /* Put the correct lv set on this block. */ | |
5311 | if (other_bb && !sel_bb_empty_p (other_bb)) | |
5312 | compute_live (sel_bb_head (other_bb)); | |
5313 | ||
5314 | return new_bb; | |
5315 | } | |
5316 | ||
5317 | /* Implement sched_create_empty_bb (). */ | |
5318 | static basic_block | |
5319 | sel_create_empty_bb (basic_block after) | |
5320 | { | |
5321 | basic_block new_bb; | |
5322 | ||
5323 | new_bb = sched_create_empty_bb_1 (after); | |
5324 | ||
5325 | /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit | |
5326 | later. */ | |
5327 | gcc_assert (VEC_length (basic_block, last_added_blocks) == 1 | |
5328 | && VEC_index (basic_block, last_added_blocks, 0) == new_bb); | |
5329 | ||
5330 | VEC_free (basic_block, heap, last_added_blocks); | |
5331 | return new_bb; | |
5332 | } | |
5333 | ||
5334 | /* Implement sched_create_recovery_block. ORIG_INSN is where block | |
5335 | will be splitted to insert a check. */ | |
5336 | basic_block | |
5337 | sel_create_recovery_block (insn_t orig_insn) | |
5338 | { | |
5339 | basic_block first_bb, second_bb, recovery_block; | |
5340 | basic_block before_recovery = NULL; | |
5341 | rtx jump; | |
5342 | ||
5343 | first_bb = BLOCK_FOR_INSN (orig_insn); | |
5344 | if (sel_bb_end_p (orig_insn)) | |
5345 | { | |
5346 | /* Avoid introducing an empty block while splitting. */ | |
5347 | gcc_assert (single_succ_p (first_bb)); | |
5348 | second_bb = single_succ (first_bb); | |
5349 | } | |
5350 | else | |
5351 | second_bb = sched_split_block (first_bb, orig_insn); | |
5352 | ||
5353 | recovery_block = sched_create_recovery_block (&before_recovery); | |
5354 | if (before_recovery) | |
5355 | copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR); | |
5356 | ||
5357 | gcc_assert (sel_bb_empty_p (recovery_block)); | |
5358 | sched_create_recovery_edges (first_bb, recovery_block, second_bb); | |
5359 | if (current_loops != NULL) | |
5360 | add_bb_to_loop (recovery_block, first_bb->loop_father); | |
b8698a0f | 5361 | |
e855c69d | 5362 | sel_add_bb (recovery_block); |
b8698a0f | 5363 | |
e855c69d AB |
5364 | jump = BB_END (recovery_block); |
5365 | gcc_assert (sel_bb_head (recovery_block) == jump); | |
5366 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5367 | ||
5368 | return recovery_block; | |
5369 | } | |
5370 | ||
5371 | /* Merge basic block B into basic block A. */ | |
5372 | void | |
5373 | sel_merge_blocks (basic_block a, basic_block b) | |
5374 | { | |
e855c69d AB |
5375 | sel_remove_empty_bb (b, true, false); |
5376 | merge_blocks (a, b); | |
5377 | ||
5378 | change_loops_latches (b, a); | |
5379 | } | |
5380 | ||
5381 | /* A wrapper for redirect_edge_and_branch_force, which also initializes | |
5382 | data structures for possibly created bb and insns. Returns the newly | |
5383 | added bb or NULL, when a bb was not needed. */ | |
5384 | void | |
5385 | sel_redirect_edge_and_branch_force (edge e, basic_block to) | |
5386 | { | |
5387 | basic_block jump_bb, src; | |
5388 | int prev_max_uid; | |
5389 | rtx jump; | |
b8698a0f | 5390 | |
e855c69d | 5391 | gcc_assert (!sel_bb_empty_p (e->src)); |
b8698a0f | 5392 | |
e855c69d AB |
5393 | src = e->src; |
5394 | prev_max_uid = get_max_uid (); | |
5395 | jump_bb = redirect_edge_and_branch_force (e, to); | |
5396 | ||
5397 | if (jump_bb != NULL) | |
5398 | sel_add_bb (jump_bb); | |
5399 | ||
5400 | /* This function could not be used to spoil the loop structure by now, | |
5401 | thus we don't care to update anything. But check it to be sure. */ | |
5402 | if (current_loop_nest | |
5403 | && pipelining_p) | |
5404 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
b8698a0f | 5405 | |
e855c69d AB |
5406 | jump = find_new_jump (src, jump_bb, prev_max_uid); |
5407 | if (jump) | |
5408 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5409 | } | |
5410 | ||
b59ab570 AM |
5411 | /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by |
5412 | redirected edge are in reverse topological order. */ | |
5413 | bool | |
e855c69d AB |
5414 | sel_redirect_edge_and_branch (edge e, basic_block to) |
5415 | { | |
5416 | bool latch_edge_p; | |
5417 | basic_block src; | |
5418 | int prev_max_uid; | |
5419 | rtx jump; | |
f2c45f08 | 5420 | edge redirected; |
b59ab570 | 5421 | bool recompute_toporder_p = false; |
e855c69d AB |
5422 | |
5423 | latch_edge_p = (pipelining_p | |
5424 | && current_loop_nest | |
5425 | && e == loop_latch_edge (current_loop_nest)); | |
5426 | ||
5427 | src = e->src; | |
5428 | prev_max_uid = get_max_uid (); | |
f2c45f08 AM |
5429 | |
5430 | redirected = redirect_edge_and_branch (e, to); | |
5431 | ||
5432 | gcc_assert (redirected && last_added_blocks == NULL); | |
e855c69d AB |
5433 | |
5434 | /* When we've redirected a latch edge, update the header. */ | |
5435 | if (latch_edge_p) | |
5436 | { | |
5437 | current_loop_nest->header = to; | |
5438 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
5439 | } | |
5440 | ||
b59ab570 AM |
5441 | /* In rare situations, the topological relation between the blocks connected |
5442 | by the redirected edge can change (see PR42245 for an example). Update | |
5443 | block_to_bb/bb_to_block. */ | |
5444 | if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) | |
5445 | && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) | |
5446 | recompute_toporder_p = true; | |
5447 | ||
e855c69d AB |
5448 | jump = find_new_jump (src, NULL, prev_max_uid); |
5449 | if (jump) | |
5450 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
b59ab570 AM |
5451 | |
5452 | return recompute_toporder_p; | |
e855c69d AB |
5453 | } |
5454 | ||
5455 | /* This variable holds the cfg hooks used by the selective scheduler. */ | |
5456 | static struct cfg_hooks sel_cfg_hooks; | |
5457 | ||
5458 | /* Register sel-sched cfg hooks. */ | |
5459 | void | |
5460 | sel_register_cfg_hooks (void) | |
5461 | { | |
5462 | sched_split_block = sel_split_block; | |
5463 | ||
5464 | orig_cfg_hooks = get_cfg_hooks (); | |
5465 | sel_cfg_hooks = orig_cfg_hooks; | |
5466 | ||
5467 | sel_cfg_hooks.create_basic_block = sel_create_basic_block; | |
5468 | ||
5469 | set_cfg_hooks (sel_cfg_hooks); | |
5470 | ||
5471 | sched_init_only_bb = sel_init_only_bb; | |
5472 | sched_split_block = sel_split_block; | |
5473 | sched_create_empty_bb = sel_create_empty_bb; | |
5474 | } | |
5475 | ||
5476 | /* Unregister sel-sched cfg hooks. */ | |
5477 | void | |
5478 | sel_unregister_cfg_hooks (void) | |
5479 | { | |
5480 | sched_create_empty_bb = NULL; | |
5481 | sched_split_block = NULL; | |
5482 | sched_init_only_bb = NULL; | |
5483 | ||
5484 | set_cfg_hooks (orig_cfg_hooks); | |
5485 | } | |
5486 | \f | |
5487 | ||
5488 | /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, | |
5489 | LABEL is where this jump should be directed. */ | |
5490 | rtx | |
5491 | create_insn_rtx_from_pattern (rtx pattern, rtx label) | |
5492 | { | |
5493 | rtx insn_rtx; | |
5494 | ||
5495 | gcc_assert (!INSN_P (pattern)); | |
5496 | ||
5497 | start_sequence (); | |
5498 | ||
5499 | if (label == NULL_RTX) | |
5500 | insn_rtx = emit_insn (pattern); | |
b5b8b0ac AO |
5501 | else if (DEBUG_INSN_P (label)) |
5502 | insn_rtx = emit_debug_insn (pattern); | |
e855c69d AB |
5503 | else |
5504 | { | |
5505 | insn_rtx = emit_jump_insn (pattern); | |
5506 | JUMP_LABEL (insn_rtx) = label; | |
5507 | ++LABEL_NUSES (label); | |
5508 | } | |
5509 | ||
5510 | end_sequence (); | |
5511 | ||
5512 | sched_init_luids (NULL, NULL, NULL, NULL); | |
5513 | sched_extend_target (); | |
5514 | sched_deps_init (false); | |
5515 | ||
5516 | /* Initialize INSN_CODE now. */ | |
5517 | recog_memoized (insn_rtx); | |
5518 | return insn_rtx; | |
5519 | } | |
5520 | ||
5521 | /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn | |
5522 | must not be clonable. */ | |
5523 | vinsn_t | |
5524 | create_vinsn_from_insn_rtx (rtx insn_rtx, bool force_unique_p) | |
5525 | { | |
5526 | gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); | |
5527 | ||
5528 | /* If VINSN_TYPE is not USE, retain its uniqueness. */ | |
5529 | return vinsn_create (insn_rtx, force_unique_p); | |
5530 | } | |
5531 | ||
5532 | /* Create a copy of INSN_RTX. */ | |
5533 | rtx | |
5534 | create_copy_of_insn_rtx (rtx insn_rtx) | |
5535 | { | |
5536 | rtx res; | |
5537 | ||
b5b8b0ac AO |
5538 | if (DEBUG_INSN_P (insn_rtx)) |
5539 | return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5540 | insn_rtx); | |
5541 | ||
e855c69d AB |
5542 | gcc_assert (NONJUMP_INSN_P (insn_rtx)); |
5543 | ||
5544 | res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5545 | NULL_RTX); | |
5546 | return res; | |
5547 | } | |
5548 | ||
5549 | /* Change vinsn field of EXPR to hold NEW_VINSN. */ | |
5550 | void | |
5551 | change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) | |
5552 | { | |
5553 | vinsn_detach (EXPR_VINSN (expr)); | |
5554 | ||
5555 | EXPR_VINSN (expr) = new_vinsn; | |
5556 | vinsn_attach (new_vinsn); | |
5557 | } | |
5558 | ||
5559 | /* Helpers for global init. */ | |
5560 | /* This structure is used to be able to call existing bundling mechanism | |
5561 | and calculate insn priorities. */ | |
b8698a0f | 5562 | static struct haifa_sched_info sched_sel_haifa_sched_info = |
e855c69d AB |
5563 | { |
5564 | NULL, /* init_ready_list */ | |
5565 | NULL, /* can_schedule_ready_p */ | |
5566 | NULL, /* schedule_more_p */ | |
5567 | NULL, /* new_ready */ | |
5568 | NULL, /* rgn_rank */ | |
5569 | sel_print_insn, /* rgn_print_insn */ | |
5570 | contributes_to_priority, | |
356c23b3 | 5571 | NULL, /* insn_finishes_block_p */ |
e855c69d AB |
5572 | |
5573 | NULL, NULL, | |
5574 | NULL, NULL, | |
5575 | 0, 0, | |
5576 | ||
5577 | NULL, /* add_remove_insn */ | |
5578 | NULL, /* begin_schedule_ready */ | |
5579 | NULL, /* advance_target_bb */ | |
5580 | SEL_SCHED | NEW_BBS | |
5581 | }; | |
5582 | ||
5583 | /* Setup special insns used in the scheduler. */ | |
b8698a0f | 5584 | void |
e855c69d AB |
5585 | setup_nop_and_exit_insns (void) |
5586 | { | |
5587 | gcc_assert (nop_pattern == NULL_RTX | |
5588 | && exit_insn == NULL_RTX); | |
5589 | ||
5590 | nop_pattern = gen_nop (); | |
5591 | ||
5592 | start_sequence (); | |
5593 | emit_insn (nop_pattern); | |
5594 | exit_insn = get_insns (); | |
5595 | end_sequence (); | |
5596 | set_block_for_insn (exit_insn, EXIT_BLOCK_PTR); | |
5597 | } | |
5598 | ||
5599 | /* Free special insns used in the scheduler. */ | |
5600 | void | |
5601 | free_nop_and_exit_insns (void) | |
5602 | { | |
5603 | exit_insn = NULL_RTX; | |
5604 | nop_pattern = NULL_RTX; | |
5605 | } | |
5606 | ||
5607 | /* Setup a special vinsn used in new insns initialization. */ | |
5608 | void | |
5609 | setup_nop_vinsn (void) | |
5610 | { | |
5611 | nop_vinsn = vinsn_create (exit_insn, false); | |
5612 | vinsn_attach (nop_vinsn); | |
5613 | } | |
5614 | ||
5615 | /* Free a special vinsn used in new insns initialization. */ | |
5616 | void | |
5617 | free_nop_vinsn (void) | |
5618 | { | |
5619 | gcc_assert (VINSN_COUNT (nop_vinsn) == 1); | |
5620 | vinsn_detach (nop_vinsn); | |
5621 | nop_vinsn = NULL; | |
5622 | } | |
5623 | ||
5624 | /* Call a set_sched_flags hook. */ | |
5625 | void | |
5626 | sel_set_sched_flags (void) | |
5627 | { | |
b8698a0f | 5628 | /* ??? This means that set_sched_flags were called, and we decided to |
e855c69d | 5629 | support speculation. However, set_sched_flags also modifies flags |
b8698a0f | 5630 | on current_sched_info, doing this only at global init. And we |
e855c69d AB |
5631 | sometimes change c_s_i later. So put the correct flags again. */ |
5632 | if (spec_info && targetm.sched.set_sched_flags) | |
5633 | targetm.sched.set_sched_flags (spec_info); | |
5634 | } | |
5635 | ||
5636 | /* Setup pointers to global sched info structures. */ | |
5637 | void | |
5638 | sel_setup_sched_infos (void) | |
5639 | { | |
5640 | rgn_setup_common_sched_info (); | |
5641 | ||
5642 | memcpy (&sel_common_sched_info, common_sched_info, | |
5643 | sizeof (sel_common_sched_info)); | |
5644 | ||
5645 | sel_common_sched_info.fix_recovery_cfg = NULL; | |
5646 | sel_common_sched_info.add_block = NULL; | |
5647 | sel_common_sched_info.estimate_number_of_insns | |
5648 | = sel_estimate_number_of_insns; | |
5649 | sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; | |
5650 | sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; | |
5651 | ||
5652 | common_sched_info = &sel_common_sched_info; | |
5653 | ||
5654 | current_sched_info = &sched_sel_haifa_sched_info; | |
b8698a0f | 5655 | current_sched_info->sched_max_insns_priority = |
e855c69d | 5656 | get_rgn_sched_max_insns_priority (); |
b8698a0f | 5657 | |
e855c69d AB |
5658 | sel_set_sched_flags (); |
5659 | } | |
5660 | \f | |
5661 | ||
5662 | /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, | |
5663 | *BB_ORD_INDEX after that is increased. */ | |
5664 | static void | |
5665 | sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) | |
5666 | { | |
5667 | RGN_NR_BLOCKS (rgn) += 1; | |
5668 | RGN_DONT_CALC_DEPS (rgn) = 0; | |
5669 | RGN_HAS_REAL_EBB (rgn) = 0; | |
5670 | CONTAINING_RGN (bb->index) = rgn; | |
5671 | BLOCK_TO_BB (bb->index) = *bb_ord_index; | |
5672 | rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; | |
5673 | (*bb_ord_index)++; | |
5674 | ||
5675 | /* FIXME: it is true only when not scheduling ebbs. */ | |
5676 | RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); | |
5677 | } | |
5678 | ||
5679 | /* Functions to support pipelining of outer loops. */ | |
5680 | ||
5681 | /* Creates a new empty region and returns it's number. */ | |
5682 | static int | |
5683 | sel_create_new_region (void) | |
5684 | { | |
5685 | int new_rgn_number = nr_regions; | |
5686 | ||
5687 | RGN_NR_BLOCKS (new_rgn_number) = 0; | |
5688 | ||
5689 | /* FIXME: This will work only when EBBs are not created. */ | |
5690 | if (new_rgn_number != 0) | |
b8698a0f | 5691 | RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + |
e855c69d AB |
5692 | RGN_NR_BLOCKS (new_rgn_number - 1); |
5693 | else | |
5694 | RGN_BLOCKS (new_rgn_number) = 0; | |
5695 | ||
5696 | /* Set the blocks of the next region so the other functions may | |
5697 | calculate the number of blocks in the region. */ | |
b8698a0f | 5698 | RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + |
e855c69d AB |
5699 | RGN_NR_BLOCKS (new_rgn_number); |
5700 | ||
5701 | nr_regions++; | |
5702 | ||
5703 | return new_rgn_number; | |
5704 | } | |
5705 | ||
5706 | /* If X has a smaller topological sort number than Y, returns -1; | |
5707 | if greater, returns 1. */ | |
5708 | static int | |
5709 | bb_top_order_comparator (const void *x, const void *y) | |
5710 | { | |
5711 | basic_block bb1 = *(const basic_block *) x; | |
5712 | basic_block bb2 = *(const basic_block *) y; | |
5713 | ||
b8698a0f L |
5714 | gcc_assert (bb1 == bb2 |
5715 | || rev_top_order_index[bb1->index] | |
e855c69d AB |
5716 | != rev_top_order_index[bb2->index]); |
5717 | ||
5718 | /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so | |
5719 | bbs with greater number should go earlier. */ | |
5720 | if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) | |
5721 | return -1; | |
5722 | else | |
5723 | return 1; | |
5724 | } | |
5725 | ||
b8698a0f | 5726 | /* Create a region for LOOP and return its number. If we don't want |
e855c69d AB |
5727 | to pipeline LOOP, return -1. */ |
5728 | static int | |
5729 | make_region_from_loop (struct loop *loop) | |
5730 | { | |
5731 | unsigned int i; | |
5732 | int new_rgn_number = -1; | |
5733 | struct loop *inner; | |
5734 | ||
5735 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5736 | int bb_ord_index = 0; | |
5737 | basic_block *loop_blocks; | |
5738 | basic_block preheader_block; | |
5739 | ||
b8698a0f | 5740 | if (loop->num_nodes |
e855c69d AB |
5741 | > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS)) |
5742 | return -1; | |
b8698a0f | 5743 | |
e855c69d AB |
5744 | /* Don't pipeline loops whose latch belongs to some of its inner loops. */ |
5745 | for (inner = loop->inner; inner; inner = inner->inner) | |
5746 | if (flow_bb_inside_loop_p (inner, loop->latch)) | |
5747 | return -1; | |
5748 | ||
5749 | loop->ninsns = num_loop_insns (loop); | |
5750 | if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS)) | |
5751 | return -1; | |
5752 | ||
5753 | loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); | |
5754 | ||
5755 | for (i = 0; i < loop->num_nodes; i++) | |
5756 | if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) | |
5757 | { | |
5758 | free (loop_blocks); | |
5759 | return -1; | |
5760 | } | |
5761 | ||
5762 | preheader_block = loop_preheader_edge (loop)->src; | |
5763 | gcc_assert (preheader_block); | |
5764 | gcc_assert (loop_blocks[0] == loop->header); | |
5765 | ||
5766 | new_rgn_number = sel_create_new_region (); | |
5767 | ||
5768 | sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number); | |
5769 | SET_BIT (bbs_in_loop_rgns, preheader_block->index); | |
5770 | ||
5771 | for (i = 0; i < loop->num_nodes; i++) | |
5772 | { | |
5773 | /* Add only those blocks that haven't been scheduled in the inner loop. | |
5774 | The exception is the basic blocks with bookkeeping code - they should | |
b8698a0f | 5775 | be added to the region (and they actually don't belong to the loop |
e855c69d AB |
5776 | body, but to the region containing that loop body). */ |
5777 | ||
5778 | gcc_assert (new_rgn_number >= 0); | |
5779 | ||
5780 | if (! TEST_BIT (bbs_in_loop_rgns, loop_blocks[i]->index)) | |
5781 | { | |
b8698a0f | 5782 | sel_add_block_to_region (loop_blocks[i], &bb_ord_index, |
e855c69d AB |
5783 | new_rgn_number); |
5784 | SET_BIT (bbs_in_loop_rgns, loop_blocks[i]->index); | |
5785 | } | |
5786 | } | |
5787 | ||
5788 | free (loop_blocks); | |
5789 | MARK_LOOP_FOR_PIPELINING (loop); | |
5790 | ||
5791 | return new_rgn_number; | |
5792 | } | |
5793 | ||
5794 | /* Create a new region from preheader blocks LOOP_BLOCKS. */ | |
5795 | void | |
5796 | make_region_from_loop_preheader (VEC(basic_block, heap) **loop_blocks) | |
5797 | { | |
5798 | unsigned int i; | |
5799 | int new_rgn_number = -1; | |
5800 | basic_block bb; | |
5801 | ||
5802 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5803 | int bb_ord_index = 0; | |
5804 | ||
5805 | new_rgn_number = sel_create_new_region (); | |
5806 | ||
5807 | for (i = 0; VEC_iterate (basic_block, *loop_blocks, i, bb); i++) | |
5808 | { | |
5809 | gcc_assert (new_rgn_number >= 0); | |
5810 | ||
5811 | sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number); | |
5812 | } | |
5813 | ||
5814 | VEC_free (basic_block, heap, *loop_blocks); | |
5815 | gcc_assert (*loop_blocks == NULL); | |
5816 | } | |
5817 | ||
5818 | ||
5819 | /* Create region(s) from loop nest LOOP, such that inner loops will be | |
b8698a0f | 5820 | pipelined before outer loops. Returns true when a region for LOOP |
e855c69d AB |
5821 | is created. */ |
5822 | static bool | |
5823 | make_regions_from_loop_nest (struct loop *loop) | |
b8698a0f | 5824 | { |
e855c69d AB |
5825 | struct loop *cur_loop; |
5826 | int rgn_number; | |
5827 | ||
5828 | /* Traverse all inner nodes of the loop. */ | |
5829 | for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) | |
5830 | if (! TEST_BIT (bbs_in_loop_rgns, cur_loop->header->index)) | |
5831 | return false; | |
5832 | ||
5833 | /* At this moment all regular inner loops should have been pipelined. | |
5834 | Try to create a region from this loop. */ | |
5835 | rgn_number = make_region_from_loop (loop); | |
5836 | ||
5837 | if (rgn_number < 0) | |
5838 | return false; | |
5839 | ||
5840 | VEC_safe_push (loop_p, heap, loop_nests, loop); | |
5841 | return true; | |
5842 | } | |
5843 | ||
5844 | /* Initalize data structures needed. */ | |
5845 | void | |
5846 | sel_init_pipelining (void) | |
5847 | { | |
5848 | /* Collect loop information to be used in outer loops pipelining. */ | |
5849 | loop_optimizer_init (LOOPS_HAVE_PREHEADERS | |
5850 | | LOOPS_HAVE_FALLTHRU_PREHEADERS | |
5851 | | LOOPS_HAVE_RECORDED_EXITS | |
5852 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); | |
5853 | current_loop_nest = NULL; | |
5854 | ||
5855 | bbs_in_loop_rgns = sbitmap_alloc (last_basic_block); | |
5856 | sbitmap_zero (bbs_in_loop_rgns); | |
5857 | ||
5858 | recompute_rev_top_order (); | |
5859 | } | |
5860 | ||
5861 | /* Returns a struct loop for region RGN. */ | |
5862 | loop_p | |
5863 | get_loop_nest_for_rgn (unsigned int rgn) | |
5864 | { | |
5865 | /* Regions created with extend_rgns don't have corresponding loop nests, | |
5866 | because they don't represent loops. */ | |
5867 | if (rgn < VEC_length (loop_p, loop_nests)) | |
5868 | return VEC_index (loop_p, loop_nests, rgn); | |
5869 | else | |
5870 | return NULL; | |
5871 | } | |
5872 | ||
5873 | /* True when LOOP was included into pipelining regions. */ | |
5874 | bool | |
5875 | considered_for_pipelining_p (struct loop *loop) | |
5876 | { | |
5877 | if (loop_depth (loop) == 0) | |
5878 | return false; | |
5879 | ||
b8698a0f L |
5880 | /* Now, the loop could be too large or irreducible. Check whether its |
5881 | region is in LOOP_NESTS. | |
5882 | We determine the region number of LOOP as the region number of its | |
5883 | latch. We can't use header here, because this header could be | |
e855c69d AB |
5884 | just removed preheader and it will give us the wrong region number. |
5885 | Latch can't be used because it could be in the inner loop too. */ | |
8ec4d0ad | 5886 | if (LOOP_MARKED_FOR_PIPELINING_P (loop)) |
e855c69d AB |
5887 | { |
5888 | int rgn = CONTAINING_RGN (loop->latch->index); | |
5889 | ||
5890 | gcc_assert ((unsigned) rgn < VEC_length (loop_p, loop_nests)); | |
5891 | return true; | |
5892 | } | |
b8698a0f | 5893 | |
e855c69d AB |
5894 | return false; |
5895 | } | |
5896 | ||
b8698a0f | 5897 | /* Makes regions from the rest of the blocks, after loops are chosen |
e855c69d AB |
5898 | for pipelining. */ |
5899 | static void | |
5900 | make_regions_from_the_rest (void) | |
5901 | { | |
5902 | int cur_rgn_blocks; | |
5903 | int *loop_hdr; | |
5904 | int i; | |
5905 | ||
5906 | basic_block bb; | |
5907 | edge e; | |
5908 | edge_iterator ei; | |
5909 | int *degree; | |
e855c69d AB |
5910 | |
5911 | /* Index in rgn_bb_table where to start allocating new regions. */ | |
5912 | cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; | |
e855c69d | 5913 | |
b8698a0f | 5914 | /* Make regions from all the rest basic blocks - those that don't belong to |
e855c69d AB |
5915 | any loop or belong to irreducible loops. Prepare the data structures |
5916 | for extend_rgns. */ | |
5917 | ||
5918 | /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, | |
5919 | LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same | |
5920 | loop. */ | |
5921 | loop_hdr = XNEWVEC (int, last_basic_block); | |
5922 | degree = XCNEWVEC (int, last_basic_block); | |
5923 | ||
5924 | ||
5925 | /* For each basic block that belongs to some loop assign the number | |
5926 | of innermost loop it belongs to. */ | |
5927 | for (i = 0; i < last_basic_block; i++) | |
5928 | loop_hdr[i] = -1; | |
5929 | ||
5930 | FOR_EACH_BB (bb) | |
5931 | { | |
5932 | if (bb->loop_father && !bb->loop_father->num == 0 | |
5933 | && !(bb->flags & BB_IRREDUCIBLE_LOOP)) | |
5934 | loop_hdr[bb->index] = bb->loop_father->num; | |
5935 | } | |
5936 | ||
b8698a0f | 5937 | /* For each basic block degree is calculated as the number of incoming |
e855c69d AB |
5938 | edges, that are going out of bbs that are not yet scheduled. |
5939 | The basic blocks that are scheduled have degree value of zero. */ | |
b8698a0f | 5940 | FOR_EACH_BB (bb) |
e855c69d AB |
5941 | { |
5942 | degree[bb->index] = 0; | |
5943 | ||
5944 | if (!TEST_BIT (bbs_in_loop_rgns, bb->index)) | |
5945 | { | |
5946 | FOR_EACH_EDGE (e, ei, bb->preds) | |
5947 | if (!TEST_BIT (bbs_in_loop_rgns, e->src->index)) | |
5948 | degree[bb->index]++; | |
5949 | } | |
5950 | else | |
5951 | degree[bb->index] = -1; | |
5952 | } | |
5953 | ||
5954 | extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); | |
5955 | ||
5956 | /* Any block that did not end up in a region is placed into a region | |
5957 | by itself. */ | |
5958 | FOR_EACH_BB (bb) | |
5959 | if (degree[bb->index] >= 0) | |
5960 | { | |
5961 | rgn_bb_table[cur_rgn_blocks] = bb->index; | |
5962 | RGN_NR_BLOCKS (nr_regions) = 1; | |
5963 | RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; | |
5964 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
5965 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
5966 | CONTAINING_RGN (bb->index) = nr_regions++; | |
5967 | BLOCK_TO_BB (bb->index) = 0; | |
5968 | } | |
5969 | ||
5970 | free (degree); | |
5971 | free (loop_hdr); | |
5972 | } | |
5973 | ||
5974 | /* Free data structures used in pipelining of loops. */ | |
5975 | void sel_finish_pipelining (void) | |
5976 | { | |
5977 | loop_iterator li; | |
5978 | struct loop *loop; | |
5979 | ||
5980 | /* Release aux fields so we don't free them later by mistake. */ | |
5981 | FOR_EACH_LOOP (li, loop, 0) | |
5982 | loop->aux = NULL; | |
5983 | ||
5984 | loop_optimizer_finalize (); | |
5985 | ||
5986 | VEC_free (loop_p, heap, loop_nests); | |
5987 | ||
5988 | free (rev_top_order_index); | |
5989 | rev_top_order_index = NULL; | |
5990 | } | |
5991 | ||
b8698a0f | 5992 | /* This function replaces the find_rgns when |
e855c69d | 5993 | FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ |
b8698a0f | 5994 | void |
e855c69d AB |
5995 | sel_find_rgns (void) |
5996 | { | |
5997 | sel_init_pipelining (); | |
5998 | extend_regions (); | |
5999 | ||
6000 | if (current_loops) | |
6001 | { | |
6002 | loop_p loop; | |
6003 | loop_iterator li; | |
6004 | ||
6005 | FOR_EACH_LOOP (li, loop, (flag_sel_sched_pipelining_outer_loops | |
6006 | ? LI_FROM_INNERMOST | |
6007 | : LI_ONLY_INNERMOST)) | |
6008 | make_regions_from_loop_nest (loop); | |
6009 | } | |
6010 | ||
6011 | /* Make regions from all the rest basic blocks and schedule them. | |
b8698a0f | 6012 | These blocks include blocks that don't belong to any loop or belong |
e855c69d AB |
6013 | to irreducible loops. */ |
6014 | make_regions_from_the_rest (); | |
6015 | ||
6016 | /* We don't need bbs_in_loop_rgns anymore. */ | |
6017 | sbitmap_free (bbs_in_loop_rgns); | |
6018 | bbs_in_loop_rgns = NULL; | |
6019 | } | |
6020 | ||
b8698a0f L |
6021 | /* Adds the preheader blocks from previous loop to current region taking |
6022 | it from LOOP_PREHEADER_BLOCKS (current_loop_nest). | |
e855c69d AB |
6023 | This function is only used with -fsel-sched-pipelining-outer-loops. */ |
6024 | void | |
6025 | sel_add_loop_preheaders (void) | |
6026 | { | |
6027 | int i; | |
6028 | basic_block bb; | |
b8698a0f | 6029 | VEC(basic_block, heap) *preheader_blocks |
e855c69d AB |
6030 | = LOOP_PREHEADER_BLOCKS (current_loop_nest); |
6031 | ||
6032 | for (i = 0; | |
6033 | VEC_iterate (basic_block, preheader_blocks, i, bb); | |
6034 | i++) | |
8ec4d0ad AM |
6035 | { |
6036 | VEC_safe_push (basic_block, heap, last_added_blocks, bb); | |
e855c69d | 6037 | sel_add_bb (bb); |
8ec4d0ad | 6038 | } |
e855c69d AB |
6039 | |
6040 | VEC_free (basic_block, heap, preheader_blocks); | |
6041 | } | |
6042 | ||
b8698a0f L |
6043 | /* While pipelining outer loops, returns TRUE if BB is a loop preheader. |
6044 | Please note that the function should also work when pipelining_p is | |
6045 | false, because it is used when deciding whether we should or should | |
e855c69d AB |
6046 | not reschedule pipelined code. */ |
6047 | bool | |
6048 | sel_is_loop_preheader_p (basic_block bb) | |
6049 | { | |
6050 | if (current_loop_nest) | |
6051 | { | |
6052 | struct loop *outer; | |
6053 | ||
6054 | if (preheader_removed) | |
6055 | return false; | |
6056 | ||
6057 | /* Preheader is the first block in the region. */ | |
6058 | if (BLOCK_TO_BB (bb->index) == 0) | |
6059 | return true; | |
6060 | ||
6061 | /* We used to find a preheader with the topological information. | |
6062 | Check that the above code is equivalent to what we did before. */ | |
6063 | ||
6064 | if (in_current_region_p (current_loop_nest->header)) | |
b8698a0f | 6065 | gcc_assert (!(BLOCK_TO_BB (bb->index) |
e855c69d AB |
6066 | < BLOCK_TO_BB (current_loop_nest->header->index))); |
6067 | ||
6068 | /* Support the situation when the latch block of outer loop | |
6069 | could be from here. */ | |
6070 | for (outer = loop_outer (current_loop_nest); | |
6071 | outer; | |
6072 | outer = loop_outer (outer)) | |
6073 | if (considered_for_pipelining_p (outer) && outer->latch == bb) | |
6074 | gcc_unreachable (); | |
6075 | } | |
6076 | ||
6077 | return false; | |
6078 | } | |
6079 | ||
6080 | /* Checks whether JUMP leads to basic block DEST_BB and no other blocks. */ | |
6081 | bool | |
6082 | jump_leads_only_to_bb_p (insn_t jump, basic_block dest_bb) | |
6083 | { | |
6084 | basic_block jump_bb = BLOCK_FOR_INSN (jump); | |
6085 | ||
b8698a0f | 6086 | /* It is not jump, jump with side-effects or jump can lead to several |
e855c69d AB |
6087 | basic blocks. */ |
6088 | if (!onlyjump_p (jump) | |
6089 | || !any_uncondjump_p (jump)) | |
6090 | return false; | |
6091 | ||
b8698a0f | 6092 | /* Several outgoing edges, abnormal edge or destination of jump is |
e855c69d AB |
6093 | not DEST_BB. */ |
6094 | if (EDGE_COUNT (jump_bb->succs) != 1 | |
6095 | || EDGE_SUCC (jump_bb, 0)->flags & EDGE_ABNORMAL | |
6096 | || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) | |
6097 | return false; | |
6098 | ||
6099 | /* If not anything of the upper. */ | |
6100 | return true; | |
6101 | } | |
6102 | ||
6103 | /* Removes the loop preheader from the current region and saves it in | |
b8698a0f | 6104 | PREHEADER_BLOCKS of the father loop, so they will be added later to |
e855c69d AB |
6105 | region that represents an outer loop. */ |
6106 | static void | |
6107 | sel_remove_loop_preheader (void) | |
6108 | { | |
6109 | int i, old_len; | |
6110 | int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
6111 | basic_block bb; | |
6112 | bool all_empty_p = true; | |
b8698a0f | 6113 | VEC(basic_block, heap) *preheader_blocks |
e855c69d AB |
6114 | = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); |
6115 | ||
6116 | gcc_assert (current_loop_nest); | |
6117 | old_len = VEC_length (basic_block, preheader_blocks); | |
6118 | ||
6119 | /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ | |
6120 | for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) | |
6121 | { | |
6122 | bb = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6123 | ||
b8698a0f | 6124 | /* If the basic block belongs to region, but doesn't belong to |
e855c69d AB |
6125 | corresponding loop, then it should be a preheader. */ |
6126 | if (sel_is_loop_preheader_p (bb)) | |
6127 | { | |
6128 | VEC_safe_push (basic_block, heap, preheader_blocks, bb); | |
6129 | if (BB_END (bb) != bb_note (bb)) | |
6130 | all_empty_p = false; | |
6131 | } | |
6132 | } | |
b8698a0f | 6133 | |
e855c69d AB |
6134 | /* Remove these blocks only after iterating over the whole region. */ |
6135 | for (i = VEC_length (basic_block, preheader_blocks) - 1; | |
6136 | i >= old_len; | |
6137 | i--) | |
6138 | { | |
b8698a0f | 6139 | bb = VEC_index (basic_block, preheader_blocks, i); |
e855c69d AB |
6140 | sel_remove_bb (bb, false); |
6141 | } | |
6142 | ||
6143 | if (!considered_for_pipelining_p (loop_outer (current_loop_nest))) | |
6144 | { | |
6145 | if (!all_empty_p) | |
6146 | /* Immediately create new region from preheader. */ | |
6147 | make_region_from_loop_preheader (&preheader_blocks); | |
6148 | else | |
6149 | { | |
6150 | /* If all preheader blocks are empty - dont create new empty region. | |
6151 | Instead, remove them completely. */ | |
6152 | for (i = 0; VEC_iterate (basic_block, preheader_blocks, i, bb); i++) | |
6153 | { | |
6154 | edge e; | |
6155 | edge_iterator ei; | |
6156 | basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; | |
6157 | ||
6158 | /* Redirect all incoming edges to next basic block. */ | |
6159 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) | |
6160 | { | |
6161 | if (! (e->flags & EDGE_FALLTHRU)) | |
6162 | redirect_edge_and_branch (e, bb->next_bb); | |
6163 | else | |
6164 | redirect_edge_succ (e, bb->next_bb); | |
6165 | } | |
6166 | gcc_assert (BB_NOTE_LIST (bb) == NULL); | |
6167 | delete_and_free_basic_block (bb); | |
6168 | ||
b8698a0f L |
6169 | /* Check if after deleting preheader there is a nonconditional |
6170 | jump in PREV_BB that leads to the next basic block NEXT_BB. | |
6171 | If it is so - delete this jump and clear data sets of its | |
e855c69d AB |
6172 | basic block if it becomes empty. */ |
6173 | if (next_bb->prev_bb == prev_bb | |
6174 | && prev_bb != ENTRY_BLOCK_PTR | |
6175 | && jump_leads_only_to_bb_p (BB_END (prev_bb), next_bb)) | |
6176 | { | |
6177 | redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); | |
6178 | if (BB_END (prev_bb) == bb_note (prev_bb)) | |
6179 | free_data_sets (prev_bb); | |
6180 | } | |
6181 | } | |
6182 | } | |
6183 | VEC_free (basic_block, heap, preheader_blocks); | |
6184 | } | |
6185 | else | |
6186 | /* Store preheader within the father's loop structure. */ | |
6187 | SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), | |
6188 | preheader_blocks); | |
6189 | } | |
6190 | #endif |