1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
2 | Copyright (C) 2006-2023 Free Software Foundation, Inc. |
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 "backend.h" |
24 | #include "cfghooks.h" |
25 | #include "tree.h" |
26 | #include "rtl.h" |
27 | #include "df.h" |
28 | #include "memmodel.h" |
29 | #include "tm_p.h" |
30 | #include "cfgrtl.h" |
31 | #include "cfganal.h" |
32 | #include "cfgbuild.h" |
33 | #include "insn-config.h" |
34 | #include "insn-attr.h" |
35 | #include "recog.h" |
36 | #include "target.h" |
37 | #include "sched-int.h" |
38 | #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */ |
39 | |
40 | #ifdef INSN_SCHEDULING |
41 | #include "regset.h" |
42 | #include "cfgloop.h" |
43 | #include "sel-sched-ir.h" |
44 | /* We don't have to use it except for sel_print_insn. */ |
45 | #include "sel-sched-dump.h" |
46 | |
47 | /* A vector holding bb info for whole scheduling pass. */ |
48 | vec<sel_global_bb_info_def> sel_global_bb_info; |
49 | |
50 | /* A vector holding bb info. */ |
51 | vec<sel_region_bb_info_def> sel_region_bb_info; |
52 | |
53 | /* A pool for allocating all lists. */ |
54 | object_allocator<_list_node> sched_lists_pool ("sel-sched-lists" ); |
55 | |
56 | /* This contains information about successors for compute_av_set. */ |
57 | struct succs_info current_succs; |
58 | |
59 | /* Data structure to describe interaction with the generic scheduler utils. */ |
60 | static struct common_sched_info_def sel_common_sched_info; |
61 | |
62 | /* The loop nest being pipelined. */ |
63 | class loop *current_loop_nest; |
64 | |
65 | /* LOOP_NESTS is a vector containing the corresponding loop nest for |
66 | each region. */ |
67 | static vec<loop_p> loop_nests; |
68 | |
69 | /* Saves blocks already in loop regions, indexed by bb->index. */ |
70 | static sbitmap bbs_in_loop_rgns = NULL; |
71 | |
72 | /* CFG hooks that are saved before changing create_basic_block hook. */ |
73 | static struct cfg_hooks orig_cfg_hooks; |
74 | |
75 | |
76 | /* Array containing reverse topological index of function basic blocks, |
77 | indexed by BB->INDEX. */ |
78 | static int *rev_top_order_index = NULL; |
79 | |
80 | /* Length of the above array. */ |
81 | static int rev_top_order_index_len = -1; |
82 | |
83 | /* A regset pool structure. */ |
84 | static struct |
85 | { |
86 | /* The stack to which regsets are returned. */ |
87 | regset *v; |
88 | |
89 | /* Its pointer. */ |
90 | int n; |
91 | |
92 | /* Its size. */ |
93 | int s; |
94 | |
95 | /* In VV we save all generated regsets so that, when destructing the |
96 | pool, we can compare it with V and check that every regset was returned |
97 | back to pool. */ |
98 | regset *vv; |
99 | |
100 | /* The pointer of VV stack. */ |
101 | int nn; |
102 | |
103 | /* Its size. */ |
104 | int ss; |
105 | |
106 | /* The difference between allocated and returned regsets. */ |
107 | int diff; |
108 | } regset_pool = { NULL, .n: 0, .s: 0, NULL, .nn: 0, .ss: 0, .diff: 0 }; |
109 | |
110 | /* This represents the nop pool. */ |
111 | static struct |
112 | { |
113 | /* The vector which holds previously emitted nops. */ |
114 | insn_t *v; |
115 | |
116 | /* Its pointer. */ |
117 | int n; |
118 | |
119 | /* Its size. */ |
120 | int s; |
121 | } nop_pool = { NULL, .n: 0, .s: 0 }; |
122 | |
123 | /* The pool for basic block notes. */ |
124 | static vec<rtx_note *> bb_note_pool; |
125 | |
126 | /* A NOP pattern used to emit placeholder insns. */ |
127 | rtx nop_pattern = NULL_RTX; |
128 | /* A special instruction that resides in EXIT_BLOCK. |
129 | EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */ |
130 | rtx_insn *exit_insn = NULL; |
131 | |
132 | /* TRUE if while scheduling current region, which is loop, its preheader |
133 | was removed. */ |
134 | bool = false; |
135 | |
136 | |
137 | /* Forward static declarations. */ |
138 | static void fence_clear (fence_t); |
139 | |
140 | static void deps_init_id (idata_t, insn_t, bool); |
141 | static void init_id_from_df (idata_t, insn_t, bool); |
142 | static expr_t set_insn_init (expr_t, vinsn_t, int); |
143 | |
144 | static void cfg_preds (basic_block, insn_t **, int *); |
145 | static void prepare_insn_expr (insn_t, int); |
146 | static void free_history_vect (vec<expr_history_def> &); |
147 | |
148 | static void move_bb_info (basic_block, basic_block); |
149 | static void remove_empty_bb (basic_block, bool); |
150 | static void sel_merge_blocks (basic_block, basic_block); |
151 | static void sel_remove_loop_preheader (void); |
152 | static bool bb_has_removable_jump_to_p (basic_block, basic_block); |
153 | |
154 | static bool insn_is_the_only_one_in_bb_p (insn_t); |
155 | static void create_initial_data_sets (basic_block); |
156 | |
157 | static void free_av_set (basic_block); |
158 | static void invalidate_av_set (basic_block); |
159 | static void extend_insn_data (void); |
160 | static void sel_init_new_insn (insn_t, int, int = -1); |
161 | static void finish_insns (void); |
162 | |
163 | /* Various list functions. */ |
164 | |
165 | /* Copy an instruction list L. */ |
166 | ilist_t |
167 | ilist_copy (ilist_t l) |
168 | { |
169 | ilist_t head = NULL, *tailp = &head; |
170 | |
171 | while (l) |
172 | { |
173 | ilist_add (lp: tailp, ILIST_INSN (l)); |
174 | tailp = &ILIST_NEXT (*tailp); |
175 | l = ILIST_NEXT (l); |
176 | } |
177 | |
178 | return head; |
179 | } |
180 | |
181 | /* Invert an instruction list L. */ |
182 | ilist_t |
183 | ilist_invert (ilist_t l) |
184 | { |
185 | ilist_t res = NULL; |
186 | |
187 | while (l) |
188 | { |
189 | ilist_add (lp: &res, ILIST_INSN (l)); |
190 | l = ILIST_NEXT (l); |
191 | } |
192 | |
193 | return res; |
194 | } |
195 | |
196 | /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */ |
197 | void |
198 | blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc) |
199 | { |
200 | bnd_t bnd; |
201 | |
202 | _list_add (lp); |
203 | bnd = BLIST_BND (*lp); |
204 | |
205 | BND_TO (bnd) = to; |
206 | BND_PTR (bnd) = ptr; |
207 | BND_AV (bnd) = NULL; |
208 | BND_AV1 (bnd) = NULL; |
209 | BND_DC (bnd) = dc; |
210 | } |
211 | |
212 | /* Remove the list note pointed to by LP. */ |
213 | void |
214 | blist_remove (blist_t *lp) |
215 | { |
216 | bnd_t b = BLIST_BND (*lp); |
217 | |
218 | av_set_clear (&BND_AV (b)); |
219 | av_set_clear (&BND_AV1 (b)); |
220 | ilist_clear (&BND_PTR (b)); |
221 | |
222 | _list_remove (lp); |
223 | } |
224 | |
225 | /* Init a fence tail L. */ |
226 | void |
227 | flist_tail_init (flist_tail_t l) |
228 | { |
229 | FLIST_TAIL_HEAD (l) = NULL; |
230 | FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l); |
231 | } |
232 | |
233 | /* Try to find fence corresponding to INSN in L. */ |
234 | fence_t |
235 | flist_lookup (flist_t l, insn_t insn) |
236 | { |
237 | while (l) |
238 | { |
239 | if (FENCE_INSN (FLIST_FENCE (l)) == insn) |
240 | return FLIST_FENCE (l); |
241 | |
242 | l = FLIST_NEXT (l); |
243 | } |
244 | |
245 | return NULL; |
246 | } |
247 | |
248 | /* Init the fields of F before running fill_insns. */ |
249 | static void |
250 | init_fence_for_scheduling (fence_t f) |
251 | { |
252 | FENCE_BNDS (f) = NULL; |
253 | FENCE_PROCESSED_P (f) = false; |
254 | FENCE_SCHEDULED_P (f) = false; |
255 | } |
256 | |
257 | /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */ |
258 | static void |
259 | flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, |
260 | insn_t last_scheduled_insn, vec<rtx_insn *, va_gc> *executing_insns, |
261 | int *ready_ticks, int ready_ticks_size, insn_t sched_next, |
262 | int cycle, int cycle_issued_insns, int issue_more, |
263 | bool starts_cycle_p, bool after_stall_p) |
264 | { |
265 | fence_t f; |
266 | |
267 | _list_add (lp); |
268 | f = FLIST_FENCE (*lp); |
269 | |
270 | FENCE_INSN (f) = insn; |
271 | |
272 | gcc_assert (state != NULL); |
273 | FENCE_STATE (f) = state; |
274 | |
275 | FENCE_CYCLE (f) = cycle; |
276 | FENCE_ISSUED_INSNS (f) = cycle_issued_insns; |
277 | FENCE_STARTS_CYCLE_P (f) = starts_cycle_p; |
278 | FENCE_AFTER_STALL_P (f) = after_stall_p; |
279 | |
280 | gcc_assert (dc != NULL); |
281 | FENCE_DC (f) = dc; |
282 | |
283 | gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL); |
284 | FENCE_TC (f) = tc; |
285 | |
286 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; |
287 | FENCE_ISSUE_MORE (f) = issue_more; |
288 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
289 | FENCE_READY_TICKS (f) = ready_ticks; |
290 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; |
291 | FENCE_SCHED_NEXT (f) = sched_next; |
292 | |
293 | init_fence_for_scheduling (f); |
294 | } |
295 | |
296 | /* Remove the head node of the list pointed to by LP. */ |
297 | static void |
298 | flist_remove (flist_t *lp) |
299 | { |
300 | if (FENCE_INSN (FLIST_FENCE (*lp))) |
301 | fence_clear (FLIST_FENCE (*lp)); |
302 | _list_remove (lp); |
303 | } |
304 | |
305 | /* Clear the fence list pointed to by LP. */ |
306 | void |
307 | flist_clear (flist_t *lp) |
308 | { |
309 | while (*lp) |
310 | flist_remove (lp); |
311 | } |
312 | |
313 | /* Add ORIGINAL_INSN the def list DL honoring CROSSED_CALL_ABIS. */ |
314 | void |
315 | def_list_add (def_list_t *dl, insn_t original_insn, |
316 | unsigned int crossed_call_abis) |
317 | { |
318 | def_t d; |
319 | |
320 | _list_add (lp: dl); |
321 | d = DEF_LIST_DEF (*dl); |
322 | |
323 | d->orig_insn = original_insn; |
324 | d->crossed_call_abis = crossed_call_abis; |
325 | } |
326 | |
327 | |
328 | /* Functions to work with target contexts. */ |
329 | |
330 | /* Bulk target context. It is convenient for debugging purposes to ensure |
331 | that there are no uninitialized (null) target contexts. */ |
332 | static tc_t bulk_tc = (tc_t) 1; |
333 | |
334 | /* Target hooks wrappers. In the future we can provide some default |
335 | implementations for them. */ |
336 | |
337 | /* Allocate a store for the target context. */ |
338 | static tc_t |
339 | alloc_target_context (void) |
340 | { |
341 | return (targetm.sched.alloc_sched_context |
342 | ? targetm.sched.alloc_sched_context () : bulk_tc); |
343 | } |
344 | |
345 | /* Init target context TC. |
346 | If CLEAN_P is true, then make TC as it is beginning of the scheduler. |
347 | Overwise, copy current backend context to TC. */ |
348 | static void |
349 | init_target_context (tc_t tc, bool clean_p) |
350 | { |
351 | if (targetm.sched.init_sched_context) |
352 | targetm.sched.init_sched_context (tc, clean_p); |
353 | } |
354 | |
355 | /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as |
356 | int init_target_context (). */ |
357 | tc_t |
358 | create_target_context (bool clean_p) |
359 | { |
360 | tc_t tc = alloc_target_context (); |
361 | |
362 | init_target_context (tc, clean_p); |
363 | return tc; |
364 | } |
365 | |
366 | /* Copy TC to the current backend context. */ |
367 | void |
368 | set_target_context (tc_t tc) |
369 | { |
370 | if (targetm.sched.set_sched_context) |
371 | targetm.sched.set_sched_context (tc); |
372 | } |
373 | |
374 | /* TC is about to be destroyed. Free any internal data. */ |
375 | static void |
376 | clear_target_context (tc_t tc) |
377 | { |
378 | if (targetm.sched.clear_sched_context) |
379 | targetm.sched.clear_sched_context (tc); |
380 | } |
381 | |
382 | /* Clear and free it. */ |
383 | static void |
384 | delete_target_context (tc_t tc) |
385 | { |
386 | clear_target_context (tc); |
387 | |
388 | if (targetm.sched.free_sched_context) |
389 | targetm.sched.free_sched_context (tc); |
390 | } |
391 | |
392 | /* Make a copy of FROM in TO. |
393 | NB: May be this should be a hook. */ |
394 | static void |
395 | copy_target_context (tc_t to, tc_t from) |
396 | { |
397 | tc_t tmp = create_target_context (clean_p: false); |
398 | |
399 | set_target_context (from); |
400 | init_target_context (tc: to, clean_p: false); |
401 | |
402 | set_target_context (tmp); |
403 | delete_target_context (tc: tmp); |
404 | } |
405 | |
406 | /* Create a copy of TC. */ |
407 | static tc_t |
408 | create_copy_of_target_context (tc_t tc) |
409 | { |
410 | tc_t copy = alloc_target_context (); |
411 | |
412 | copy_target_context (to: copy, from: tc); |
413 | |
414 | return copy; |
415 | } |
416 | |
417 | /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P |
418 | is the same as in init_target_context (). */ |
419 | void |
420 | reset_target_context (tc_t tc, bool clean_p) |
421 | { |
422 | clear_target_context (tc); |
423 | init_target_context (tc, clean_p); |
424 | } |
425 | |
426 | /* Functions to work with dependence contexts. |
427 | Dc (aka deps context, aka deps_t, aka class deps_desc *) is short for dependence |
428 | context. It accumulates information about processed insns to decide if |
429 | current insn is dependent on the processed ones. */ |
430 | |
431 | /* Make a copy of FROM in TO. */ |
432 | static void |
433 | copy_deps_context (deps_t to, deps_t from) |
434 | { |
435 | init_deps (to, false); |
436 | deps_join (to, from); |
437 | } |
438 | |
439 | /* Allocate store for dep context. */ |
440 | static deps_t |
441 | alloc_deps_context (void) |
442 | { |
443 | return XNEW (class deps_desc); |
444 | } |
445 | |
446 | /* Allocate and initialize dep context. */ |
447 | static deps_t |
448 | create_deps_context (void) |
449 | { |
450 | deps_t dc = alloc_deps_context (); |
451 | |
452 | init_deps (dc, false); |
453 | return dc; |
454 | } |
455 | |
456 | /* Create a copy of FROM. */ |
457 | static deps_t |
458 | create_copy_of_deps_context (deps_t from) |
459 | { |
460 | deps_t to = alloc_deps_context (); |
461 | |
462 | copy_deps_context (to, from); |
463 | return to; |
464 | } |
465 | |
466 | /* Clean up internal data of DC. */ |
467 | static void |
468 | clear_deps_context (deps_t dc) |
469 | { |
470 | free_deps (dc); |
471 | } |
472 | |
473 | /* Clear and free DC. */ |
474 | static void |
475 | delete_deps_context (deps_t dc) |
476 | { |
477 | clear_deps_context (dc); |
478 | free (ptr: dc); |
479 | } |
480 | |
481 | /* Clear and init DC. */ |
482 | static void |
483 | reset_deps_context (deps_t dc) |
484 | { |
485 | clear_deps_context (dc); |
486 | init_deps (dc, false); |
487 | } |
488 | |
489 | /* This structure describes the dependence analysis hooks for advancing |
490 | dependence context. */ |
491 | static struct sched_deps_info_def advance_deps_context_sched_deps_info = |
492 | { |
493 | NULL, |
494 | |
495 | NULL, /* start_insn */ |
496 | NULL, /* finish_insn */ |
497 | NULL, /* start_lhs */ |
498 | NULL, /* finish_lhs */ |
499 | NULL, /* start_rhs */ |
500 | NULL, /* finish_rhs */ |
501 | .note_reg_set: haifa_note_reg_set, |
502 | .note_reg_clobber: haifa_note_reg_clobber, |
503 | .note_reg_use: haifa_note_reg_use, |
504 | NULL, /* note_mem_dep */ |
505 | NULL, /* note_dep */ |
506 | |
507 | .use_cselib: 0, .use_deps_list: 0, .generate_spec_deps: 0 |
508 | }; |
509 | |
510 | /* Process INSN and add its impact on DC. */ |
511 | void |
512 | advance_deps_context (deps_t dc, insn_t insn) |
513 | { |
514 | sched_deps_info = &advance_deps_context_sched_deps_info; |
515 | deps_analyze_insn (dc, insn); |
516 | } |
517 | |
518 | |
519 | /* Functions to work with DFA states. */ |
520 | |
521 | /* Allocate store for a DFA state. */ |
522 | static state_t |
523 | state_alloc (void) |
524 | { |
525 | return xmalloc (dfa_state_size); |
526 | } |
527 | |
528 | /* Allocate and initialize DFA state. */ |
529 | static state_t |
530 | state_create (void) |
531 | { |
532 | state_t state = state_alloc (); |
533 | |
534 | state_reset (state); |
535 | advance_state (state); |
536 | return state; |
537 | } |
538 | |
539 | /* Free DFA state. */ |
540 | static void |
541 | state_free (state_t state) |
542 | { |
543 | free (ptr: state); |
544 | } |
545 | |
546 | /* Make a copy of FROM in TO. */ |
547 | static void |
548 | state_copy (state_t to, state_t from) |
549 | { |
550 | memcpy (dest: to, src: from, n: dfa_state_size); |
551 | } |
552 | |
553 | /* Create a copy of FROM. */ |
554 | static state_t |
555 | state_create_copy (state_t from) |
556 | { |
557 | state_t to = state_alloc (); |
558 | |
559 | state_copy (to, from); |
560 | return to; |
561 | } |
562 | |
563 | |
564 | /* Functions to work with fences. */ |
565 | |
566 | /* Clear the fence. */ |
567 | static void |
568 | fence_clear (fence_t f) |
569 | { |
570 | state_t s = FENCE_STATE (f); |
571 | deps_t dc = FENCE_DC (f); |
572 | void *tc = FENCE_TC (f); |
573 | |
574 | ilist_clear (&FENCE_BNDS (f)); |
575 | |
576 | gcc_assert ((s != NULL && dc != NULL && tc != NULL) |
577 | || (s == NULL && dc == NULL && tc == NULL)); |
578 | |
579 | free (ptr: s); |
580 | |
581 | if (dc != NULL) |
582 | delete_deps_context (dc); |
583 | |
584 | if (tc != NULL) |
585 | delete_target_context (tc); |
586 | vec_free (FENCE_EXECUTING_INSNS (f)); |
587 | free (FENCE_READY_TICKS (f)); |
588 | FENCE_READY_TICKS (f) = NULL; |
589 | } |
590 | |
591 | /* Init a list of fences with successors of OLD_FENCE. */ |
592 | void |
593 | init_fences (insn_t old_fence) |
594 | { |
595 | insn_t succ; |
596 | succ_iterator si; |
597 | bool first = true; |
598 | int ready_ticks_size = get_max_uid () + 1; |
599 | |
600 | FOR_EACH_SUCC_1 (succ, si, old_fence, |
601 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
602 | { |
603 | |
604 | if (first) |
605 | first = false; |
606 | else |
607 | gcc_assert (flag_sel_sched_pipelining_outer_loops); |
608 | |
609 | flist_add (lp: &fences, insn: succ, |
610 | state: state_create (), |
611 | dc: create_deps_context () /* dc */, |
612 | tc: create_target_context (clean_p: true) /* tc */, |
613 | NULL /* last_scheduled_insn */, |
614 | NULL, /* executing_insns */ |
615 | XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ |
616 | ready_ticks_size, |
617 | NULL /* sched_next */, |
618 | cycle: 1 /* cycle */, cycle_issued_insns: 0 /* cycle_issued_insns */, |
619 | issue_more: issue_rate, /* issue_more */ |
620 | starts_cycle_p: 1 /* starts_cycle_p */, after_stall_p: 0 /* after_stall_p */); |
621 | } |
622 | } |
623 | |
624 | /* Merges two fences (filling fields of fence F with resulting values) by |
625 | following rules: 1) state, target context and last scheduled insn are |
626 | propagated from fallthrough edge if it is available; |
627 | 2) deps context and cycle is propagated from more probable edge; |
628 | 3) all other fields are set to corresponding constant values. |
629 | |
630 | INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, |
631 | READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE |
632 | and AFTER_STALL_P are the corresponding fields of the second fence. */ |
633 | static void |
634 | merge_fences (fence_t f, insn_t insn, |
635 | state_t state, deps_t dc, void *tc, |
636 | rtx_insn *last_scheduled_insn, |
637 | vec<rtx_insn *, va_gc> *executing_insns, |
638 | int *ready_ticks, int ready_ticks_size, |
639 | rtx sched_next, int cycle, int issue_more, 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 | |
646 | /* Check if we can decide which path fences came. |
647 | If we can't (or don't want to) - reset all. */ |
648 | if (last_scheduled_insn == NULL |
649 | || last_scheduled_insn_old == NULL |
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 - |
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); |
658 | |
659 | reset_deps_context (FENCE_DC (f)); |
660 | delete_deps_context (dc); |
661 | |
662 | reset_target_context (FENCE_TC (f), clean_p: 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 | FENCE_ISSUE_MORE (f) = issue_rate; |
670 | vec_free (v&: executing_insns); |
671 | free (ptr: ready_ticks); |
672 | if (FENCE_EXECUTING_INSNS (f)) |
673 | FENCE_EXECUTING_INSNS (f)->block_remove (ix: 0, |
674 | FENCE_EXECUTING_INSNS (f)->length ()); |
675 | if (FENCE_READY_TICKS (f)) |
676 | memset (FENCE_READY_TICKS (f), c: 0, FENCE_READY_TICKS_SIZE (f)); |
677 | } |
678 | else |
679 | { |
680 | edge edge_old = NULL, edge_new = NULL; |
681 | edge candidate; |
682 | succ_iterator si; |
683 | insn_t succ; |
684 | |
685 | /* Find fallthrough edge. */ |
686 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); |
687 | candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb); |
688 | |
689 | if (!candidate |
690 | || (candidate->src != BLOCK_FOR_INSN (insn: last_scheduled_insn) |
691 | && candidate->src != BLOCK_FOR_INSN (insn: last_scheduled_insn_old))) |
692 | { |
693 | /* No fallthrough edge leading to basic block of INSN. */ |
694 | state_reset (FENCE_STATE (f)); |
695 | state_free (state); |
696 | |
697 | reset_target_context (FENCE_TC (f), clean_p: true); |
698 | delete_target_context (tc); |
699 | |
700 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; |
701 | FENCE_ISSUE_MORE (f) = issue_rate; |
702 | } |
703 | else |
704 | if (candidate->src == BLOCK_FOR_INSN (insn: last_scheduled_insn)) |
705 | { |
706 | state_free (FENCE_STATE (f)); |
707 | FENCE_STATE (f) = state; |
708 | |
709 | delete_target_context (FENCE_TC (f)); |
710 | FENCE_TC (f) = tc; |
711 | |
712 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; |
713 | FENCE_ISSUE_MORE (f) = issue_more; |
714 | } |
715 | else |
716 | { |
717 | /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ |
718 | state_free (state); |
719 | delete_target_context (tc); |
720 | |
721 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb |
722 | != BLOCK_FOR_INSN (last_scheduled_insn)); |
723 | } |
724 | |
725 | /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ |
726 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, |
727 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
728 | { |
729 | if (succ == insn) |
730 | { |
731 | /* No same successor allowed from several edges. */ |
732 | gcc_assert (!edge_old); |
733 | edge_old = si.e1; |
734 | } |
735 | } |
736 | /* Find edge of second predecessor (last_scheduled_insn->insn). */ |
737 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, |
738 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
739 | { |
740 | if (succ == insn) |
741 | { |
742 | /* No same successor allowed from several edges. */ |
743 | gcc_assert (!edge_new); |
744 | edge_new = si.e1; |
745 | } |
746 | } |
747 | |
748 | /* Check if we can choose most probable predecessor. */ |
749 | if (edge_old == NULL || edge_new == NULL) |
750 | { |
751 | reset_deps_context (FENCE_DC (f)); |
752 | delete_deps_context (dc); |
753 | vec_free (v&: executing_insns); |
754 | free (ptr: ready_ticks); |
755 | |
756 | FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); |
757 | if (FENCE_EXECUTING_INSNS (f)) |
758 | FENCE_EXECUTING_INSNS (f)->block_remove (ix: 0, |
759 | FENCE_EXECUTING_INSNS (f)->length ()); |
760 | if (FENCE_READY_TICKS (f)) |
761 | memset (FENCE_READY_TICKS (f), c: 0, FENCE_READY_TICKS_SIZE (f)); |
762 | } |
763 | else |
764 | if (edge_new->probability > edge_old->probability) |
765 | { |
766 | delete_deps_context (FENCE_DC (f)); |
767 | FENCE_DC (f) = dc; |
768 | vec_free (FENCE_EXECUTING_INSNS (f)); |
769 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
770 | free (FENCE_READY_TICKS (f)); |
771 | FENCE_READY_TICKS (f) = ready_ticks; |
772 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; |
773 | FENCE_CYCLE (f) = cycle; |
774 | } |
775 | else |
776 | { |
777 | /* Leave DC and CYCLE untouched. */ |
778 | delete_deps_context (dc); |
779 | vec_free (v&: executing_insns); |
780 | free (ptr: ready_ticks); |
781 | } |
782 | } |
783 | |
784 | /* Fill remaining invariant fields. */ |
785 | if (after_stall_p) |
786 | FENCE_AFTER_STALL_P (f) = 1; |
787 | |
788 | FENCE_ISSUED_INSNS (f) = 0; |
789 | FENCE_STARTS_CYCLE_P (f) = 1; |
790 | FENCE_SCHED_NEXT (f) = NULL; |
791 | } |
792 | |
793 | /* Add a new fence to NEW_FENCES list, initializing it from all |
794 | other parameters. */ |
795 | static void |
796 | add_to_fences (flist_tail_t new_fences, insn_t insn, |
797 | state_t state, deps_t dc, void *tc, |
798 | rtx_insn *last_scheduled_insn, |
799 | vec<rtx_insn *, va_gc> *executing_insns, int *ready_ticks, |
800 | int ready_ticks_size, rtx_insn *sched_next, int cycle, |
801 | int cycle_issued_insns, int issue_rate, |
802 | 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, |
809 | last_scheduled_insn, executing_insns, ready_ticks, |
810 | ready_ticks_size, sched_next, cycle, cycle_issued_insns, |
811 | issue_more: issue_rate, 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 | { |
818 | merge_fences (f, insn, state, dc, tc, last_scheduled_insn, |
819 | executing_insns, ready_ticks, ready_ticks_size, |
820 | sched_next, cycle, issue_more: issue_rate, 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); |
832 | f = flist_lookup (FLIST_TAIL_HEAD (new_fences), |
833 | FENCE_INSN (FLIST_FENCE (old_fences))); |
834 | if (f) |
835 | { |
836 | merge_fences (f, insn: old->insn, state: old->state, dc: old->dc, tc: old->tc, |
837 | last_scheduled_insn: old->last_scheduled_insn, executing_insns: old->executing_insns, |
838 | ready_ticks: old->ready_ticks, ready_ticks_size: old->ready_ticks_size, |
839 | sched_next: old->sched_next, cycle: old->cycle, issue_more: old->issue_more, |
840 | after_stall_p: old->after_stall_p); |
841 | } |
842 | else |
843 | { |
844 | _list_add (lp: 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 | |
852 | /* Add a new fence to NEW_FENCES list and initialize most of its data |
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; |
858 | |
859 | add_to_fences (new_fences, |
860 | insn: succ, state: state_create (), dc: create_deps_context (), |
861 | tc: create_target_context (clean_p: true), |
862 | NULL, NULL, |
863 | XCNEWVEC (int, ready_ticks_size), ready_ticks_size, |
864 | NULL, FENCE_CYCLE (fence) + 1, |
865 | cycle_issued_insns: 0, issue_rate, starts_cycle_p: 1, FENCE_AFTER_STALL_P (fence)); |
866 | } |
867 | |
868 | /* Add a new fence to NEW_FENCES list and initialize all of its data |
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 | { |
873 | int * new_ready_ticks |
874 | = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); |
875 | |
876 | memcpy (dest: new_ready_ticks, FENCE_READY_TICKS (fence), |
877 | FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); |
878 | add_to_fences (new_fences, |
879 | insn: succ, state: state_create_copy (FENCE_STATE (fence)), |
880 | dc: create_copy_of_deps_context (FENCE_DC (fence)), |
881 | tc: create_copy_of_target_context (FENCE_TC (fence)), |
882 | FENCE_LAST_SCHEDULED_INSN (fence), |
883 | executing_insns: vec_safe_copy (FENCE_EXECUTING_INSNS (fence)), |
884 | ready_ticks: 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_ISSUE_MORE (fence), |
890 | FENCE_STARTS_CYCLE_P (fence), |
891 | FENCE_AFTER_STALL_P (fence)); |
892 | } |
893 | |
894 | |
895 | /* Functions to work with regset and nop pools. */ |
896 | |
897 | /* Returns the new regset from pool. It might have some of the bits set |
898 | from the previous usage. */ |
899 | regset |
900 | get_regset_from_pool (void) |
901 | { |
902 | regset rs; |
903 | |
904 | if (regset_pool.n != 0) |
905 | rs = regset_pool.v[--regset_pool.n]; |
906 | else |
907 | /* We need to create the regset. */ |
908 | { |
909 | rs = ALLOC_REG_SET (®_obstack); |
910 | |
911 | if (regset_pool.nn == regset_pool.ss) |
912 | regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, |
913 | (regset_pool.ss = 2 * regset_pool.ss + 1)); |
914 | regset_pool.vv[regset_pool.nn++] = rs; |
915 | } |
916 | |
917 | regset_pool.diff++; |
918 | |
919 | return rs; |
920 | } |
921 | |
922 | /* Same as above, but returns the empty regset. */ |
923 | regset |
924 | get_clear_regset_from_pool (void) |
925 | { |
926 | regset rs = get_regset_from_pool (); |
927 | |
928 | CLEAR_REG_SET (rs); |
929 | return rs; |
930 | } |
931 | |
932 | /* Return regset RS to the pool for future use. */ |
933 | void |
934 | return_regset_to_pool (regset rs) |
935 | { |
936 | gcc_assert (rs); |
937 | regset_pool.diff--; |
938 | |
939 | if (regset_pool.n == regset_pool.s) |
940 | regset_pool.v = XRESIZEVEC (regset, regset_pool.v, |
941 | (regset_pool.s = 2 * regset_pool.s + 1)); |
942 | regset_pool.v[regset_pool.n++] = rs; |
943 | } |
944 | |
945 | /* This is used as a qsort callback for sorting regset pool stacks. |
946 | X and XX are addresses of two regsets. They are never equal. */ |
947 | static int |
948 | cmp_v_in_regset_pool (const void *x, const void *xx) |
949 | { |
950 | uintptr_t r1 = (uintptr_t) *((const regset *) x); |
951 | uintptr_t r2 = (uintptr_t) *((const regset *) xx); |
952 | if (r1 > r2) |
953 | return 1; |
954 | else if (r1 < r2) |
955 | return -1; |
956 | gcc_unreachable (); |
957 | } |
958 | |
959 | /* Free the regset pool possibly checking for memory leaks. */ |
960 | void |
961 | free_regset_pool (void) |
962 | { |
963 | if (flag_checking) |
964 | { |
965 | regset *v = regset_pool.v; |
966 | int i = 0; |
967 | int n = regset_pool.n; |
968 | |
969 | regset *vv = regset_pool.vv; |
970 | int ii = 0; |
971 | int nn = regset_pool.nn; |
972 | |
973 | int diff = 0; |
974 | |
975 | gcc_assert (n <= nn); |
976 | |
977 | /* Sort both vectors so it will be possible to compare them. */ |
978 | qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); |
979 | qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); |
980 | |
981 | while (ii < nn) |
982 | { |
983 | if (v[i] == vv[ii]) |
984 | i++; |
985 | else |
986 | /* VV[II] was lost. */ |
987 | diff++; |
988 | |
989 | ii++; |
990 | } |
991 | |
992 | gcc_assert (diff == regset_pool.diff); |
993 | } |
994 | |
995 | /* If not true - we have a memory leak. */ |
996 | gcc_assert (regset_pool.diff == 0); |
997 | |
998 | while (regset_pool.n) |
999 | { |
1000 | --regset_pool.n; |
1001 | FREE_REG_SET (regset_pool.v[regset_pool.n]); |
1002 | } |
1003 | |
1004 | free (ptr: regset_pool.v); |
1005 | regset_pool.v = NULL; |
1006 | regset_pool.s = 0; |
1007 | |
1008 | free (ptr: regset_pool.vv); |
1009 | regset_pool.vv = NULL; |
1010 | regset_pool.nn = 0; |
1011 | regset_pool.ss = 0; |
1012 | |
1013 | regset_pool.diff = 0; |
1014 | } |
1015 | |
1016 | |
1017 | /* Functions to work with nop pools. NOP insns are used as temporary |
1018 | placeholders of the insns being scheduled to allow correct update of |
1019 | the data sets. When update is finished, NOPs are deleted. */ |
1020 | |
1021 | /* A vinsn that is used to represent a nop. This vinsn is shared among all |
1022 | nops sel-sched generates. */ |
1023 | static vinsn_t nop_vinsn = NULL; |
1024 | |
1025 | /* Emit a nop before INSN, taking it from pool. */ |
1026 | insn_t |
1027 | get_nop_from_pool (insn_t insn) |
1028 | { |
1029 | rtx nop_pat; |
1030 | insn_t nop; |
1031 | bool old_p = nop_pool.n != 0; |
1032 | int flags; |
1033 | |
1034 | if (old_p) |
1035 | nop_pat = nop_pool.v[--nop_pool.n]; |
1036 | else |
1037 | nop_pat = nop_pattern; |
1038 | |
1039 | nop = emit_insn_before (nop_pat, insn); |
1040 | |
1041 | if (old_p) |
1042 | flags = INSN_INIT_TODO_SSID; |
1043 | else |
1044 | flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; |
1045 | |
1046 | set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); |
1047 | sel_init_new_insn (nop, flags); |
1048 | |
1049 | return nop; |
1050 | } |
1051 | |
1052 | /* Remove NOP from the instruction stream and return it to the pool. */ |
1053 | void |
1054 | return_nop_to_pool (insn_t nop, bool full_tidying) |
1055 | { |
1056 | gcc_assert (INSN_IN_STREAM_P (nop)); |
1057 | sel_remove_insn (nop, false, full_tidying); |
1058 | |
1059 | /* We'll recycle this nop. */ |
1060 | nop->set_undeleted (); |
1061 | |
1062 | if (nop_pool.n == nop_pool.s) |
1063 | nop_pool.v = XRESIZEVEC (rtx_insn *, nop_pool.v, |
1064 | (nop_pool.s = 2 * nop_pool.s + 1)); |
1065 | nop_pool.v[nop_pool.n++] = nop; |
1066 | } |
1067 | |
1068 | /* Free the nop pool. */ |
1069 | void |
1070 | free_nop_pool (void) |
1071 | { |
1072 | nop_pool.n = 0; |
1073 | nop_pool.s = 0; |
1074 | free (ptr: nop_pool.v); |
1075 | nop_pool.v = NULL; |
1076 | } |
1077 | |
1078 | |
1079 | /* Skip unspec to support ia64 speculation. Called from rtx_equal_p. |
1080 | The callback is given two rtxes XX and YY and writes the new rtxes |
1081 | to NX and NY in case some needs to be skipped. */ |
1082 | static bool |
1083 | skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) |
1084 | { |
1085 | const_rtx x = *xx; |
1086 | const_rtx y = *yy; |
1087 | |
1088 | if (GET_CODE (x) == UNSPEC |
1089 | && (targetm.sched.skip_rtx_p == NULL |
1090 | || targetm.sched.skip_rtx_p (x))) |
1091 | { |
1092 | *nx = XVECEXP (x, 0, 0); |
1093 | *ny = CONST_CAST_RTX (y); |
1094 | return true; |
1095 | } |
1096 | |
1097 | if (GET_CODE (y) == UNSPEC |
1098 | && (targetm.sched.skip_rtx_p == NULL |
1099 | || targetm.sched.skip_rtx_p (y))) |
1100 | { |
1101 | *nx = CONST_CAST_RTX (x); |
1102 | *ny = XVECEXP (y, 0, 0); |
1103 | return true; |
1104 | } |
1105 | |
1106 | return false; |
1107 | } |
1108 | |
1109 | /* Callback, called from hash_rtx. Helps to hash UNSPEC rtx X in a correct way |
1110 | to support ia64 speculation. When changes are needed, new rtx X and new mode |
1111 | NMODE are written, and the callback returns true. */ |
1112 | static bool |
1113 | hash_with_unspec_callback (const_rtx x, machine_mode mode ATTRIBUTE_UNUSED, |
1114 | rtx *nx, machine_mode* nmode) |
1115 | { |
1116 | if (GET_CODE (x) == UNSPEC |
1117 | && targetm.sched.skip_rtx_p |
1118 | && targetm.sched.skip_rtx_p (x)) |
1119 | { |
1120 | *nx = XVECEXP (x, 0 ,0); |
1121 | *nmode = VOIDmode; |
1122 | return true; |
1123 | } |
1124 | |
1125 | return false; |
1126 | } |
1127 | |
1128 | /* Returns LHS and RHS are ok to be scheduled separately. */ |
1129 | static bool |
1130 | lhs_and_rhs_separable_p (rtx lhs, rtx rhs) |
1131 | { |
1132 | if (lhs == NULL || rhs == NULL) |
1133 | return false; |
1134 | |
1135 | /* Do not schedule constants as rhs: no point to use reg, if const |
1136 | can be used. Moreover, scheduling const as rhs may lead to mode |
1137 | mismatch cause consts don't have modes but they could be merged |
1138 | from branches where the same const used in different modes. */ |
1139 | if (CONSTANT_P (rhs)) |
1140 | return false; |
1141 | |
1142 | /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ |
1143 | if (COMPARISON_P (rhs)) |
1144 | return false; |
1145 | |
1146 | /* Do not allow single REG to be an rhs. */ |
1147 | if (REG_P (rhs)) |
1148 | return false; |
1149 | |
1150 | /* See comment at find_used_regs_1 (*1) for explanation of this |
1151 | restriction. */ |
1152 | /* FIXME: remove this later. */ |
1153 | if (MEM_P (lhs)) |
1154 | return false; |
1155 | |
1156 | /* This will filter all tricky things like ZERO_EXTRACT etc. |
1157 | For now we don't handle it. */ |
1158 | if (!REG_P (lhs) && !MEM_P (lhs)) |
1159 | return false; |
1160 | |
1161 | return true; |
1162 | } |
1163 | |
1164 | /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When |
1165 | FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is |
1166 | used e.g. for insns from recovery blocks. */ |
1167 | static void |
1168 | vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) |
1169 | { |
1170 | hash_rtx_callback_function hrcf; |
1171 | int insn_class; |
1172 | |
1173 | VINSN_INSN_RTX (vi) = insn; |
1174 | VINSN_COUNT (vi) = 0; |
1175 | vi->cost = -1; |
1176 | |
1177 | if (INSN_NOP_P (insn)) |
1178 | return; |
1179 | |
1180 | if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) |
1181 | init_id_from_df (VINSN_ID (vi), insn, force_unique_p); |
1182 | else |
1183 | deps_init_id (VINSN_ID (vi), insn, force_unique_p); |
1184 | |
1185 | /* Hash vinsn depending on whether it is separable or not. */ |
1186 | hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; |
1187 | if (VINSN_SEPARABLE_P (vi)) |
1188 | { |
1189 | rtx rhs = VINSN_RHS (vi); |
1190 | |
1191 | VINSN_HASH (vi) = hash_rtx (rhs, GET_MODE (rhs), |
1192 | NULL, NULL, false, hrcf); |
1193 | VINSN_HASH_RTX (vi) = hash_rtx (VINSN_PATTERN (vi), |
1194 | VOIDmode, NULL, NULL, |
1195 | false, hrcf); |
1196 | } |
1197 | else |
1198 | { |
1199 | VINSN_HASH (vi) = hash_rtx (VINSN_PATTERN (vi), VOIDmode, |
1200 | NULL, NULL, false, hrcf); |
1201 | VINSN_HASH_RTX (vi) = VINSN_HASH (vi); |
1202 | } |
1203 | |
1204 | insn_class = haifa_classify_insn (insn); |
1205 | if (insn_class >= 2 |
1206 | && (!targetm.sched.get_insn_spec_ds |
1207 | || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) |
1208 | == 0))) |
1209 | VINSN_MAY_TRAP_P (vi) = true; |
1210 | else |
1211 | VINSN_MAY_TRAP_P (vi) = false; |
1212 | } |
1213 | |
1214 | /* Indicate that VI has become the part of an rtx object. */ |
1215 | void |
1216 | vinsn_attach (vinsn_t vi) |
1217 | { |
1218 | /* Assert that VI is not pending for deletion. */ |
1219 | gcc_assert (VINSN_INSN_RTX (vi)); |
1220 | |
1221 | VINSN_COUNT (vi)++; |
1222 | } |
1223 | |
1224 | /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct |
1225 | VINSN_TYPE (VI). */ |
1226 | static vinsn_t |
1227 | vinsn_create (insn_t insn, bool force_unique_p) |
1228 | { |
1229 | vinsn_t vi = XCNEW (struct vinsn_def); |
1230 | |
1231 | vinsn_init (vi, insn, force_unique_p); |
1232 | return vi; |
1233 | } |
1234 | |
1235 | /* Return a copy of VI. When REATTACH_P is true, detach VI and attach |
1236 | the copy. */ |
1237 | vinsn_t |
1238 | vinsn_copy (vinsn_t vi, bool reattach_p) |
1239 | { |
1240 | rtx_insn *copy; |
1241 | bool unique = VINSN_UNIQUE_P (vi); |
1242 | vinsn_t new_vi; |
1243 | |
1244 | copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); |
1245 | new_vi = create_vinsn_from_insn_rtx (copy, unique); |
1246 | if (reattach_p) |
1247 | { |
1248 | vinsn_detach (vi); |
1249 | vinsn_attach (vi: new_vi); |
1250 | } |
1251 | |
1252 | return new_vi; |
1253 | } |
1254 | |
1255 | /* Delete the VI vinsn and free its data. */ |
1256 | static void |
1257 | vinsn_delete (vinsn_t vi) |
1258 | { |
1259 | gcc_assert (VINSN_COUNT (vi) == 0); |
1260 | |
1261 | if (!INSN_NOP_P (VINSN_INSN_RTX (vi))) |
1262 | { |
1263 | return_regset_to_pool (VINSN_REG_SETS (vi)); |
1264 | return_regset_to_pool (VINSN_REG_USES (vi)); |
1265 | return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); |
1266 | } |
1267 | |
1268 | free (ptr: vi); |
1269 | } |
1270 | |
1271 | /* Indicate that VI is no longer a part of some rtx object. |
1272 | Remove VI if it is no longer needed. */ |
1273 | void |
1274 | vinsn_detach (vinsn_t vi) |
1275 | { |
1276 | gcc_assert (VINSN_COUNT (vi) > 0); |
1277 | |
1278 | if (--VINSN_COUNT (vi) == 0) |
1279 | vinsn_delete (vi); |
1280 | } |
1281 | |
1282 | /* Returns TRUE if VI is a branch. */ |
1283 | bool |
1284 | vinsn_cond_branch_p (vinsn_t vi) |
1285 | { |
1286 | insn_t insn; |
1287 | |
1288 | if (!VINSN_UNIQUE_P (vi)) |
1289 | return false; |
1290 | |
1291 | insn = VINSN_INSN_RTX (vi); |
1292 | if (BB_END (BLOCK_FOR_INSN (insn)) != insn) |
1293 | return false; |
1294 | |
1295 | return control_flow_insn_p (insn); |
1296 | } |
1297 | |
1298 | /* Return latency of INSN. */ |
1299 | static int |
1300 | sel_insn_rtx_cost (rtx_insn *insn) |
1301 | { |
1302 | int cost; |
1303 | |
1304 | /* A USE insn, or something else we don't need to |
1305 | understand. We can't pass these directly to |
1306 | result_ready_cost or insn_default_latency because it will |
1307 | trigger a fatal error for unrecognizable insns. */ |
1308 | if (recog_memoized (insn) < 0) |
1309 | cost = 0; |
1310 | else |
1311 | { |
1312 | cost = insn_default_latency (insn); |
1313 | |
1314 | if (cost < 0) |
1315 | cost = 0; |
1316 | } |
1317 | |
1318 | return cost; |
1319 | } |
1320 | |
1321 | /* Return the cost of the VI. |
1322 | !!! FIXME: Unify with haifa-sched.cc: insn_sched_cost (). */ |
1323 | int |
1324 | sel_vinsn_cost (vinsn_t vi) |
1325 | { |
1326 | int cost = vi->cost; |
1327 | |
1328 | if (cost < 0) |
1329 | { |
1330 | cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); |
1331 | vi->cost = cost; |
1332 | } |
1333 | |
1334 | return cost; |
1335 | } |
1336 | |
1337 | |
1338 | /* Functions for insn emitting. */ |
1339 | |
1340 | /* Emit new insn after AFTER based on PATTERN and initialize its data from |
1341 | EXPR and SEQNO. */ |
1342 | insn_t |
1343 | sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) |
1344 | { |
1345 | insn_t new_insn; |
1346 | |
1347 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); |
1348 | |
1349 | new_insn = emit_insn_after (pattern, after); |
1350 | set_insn_init (expr, NULL, seqno); |
1351 | sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); |
1352 | |
1353 | return new_insn; |
1354 | } |
1355 | |
1356 | /* Force newly generated vinsns to be unique. */ |
1357 | static bool init_insn_force_unique_p = false; |
1358 | |
1359 | /* Emit new speculation recovery insn after AFTER based on PATTERN and |
1360 | initialize its data from EXPR and SEQNO. */ |
1361 | insn_t |
1362 | sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, |
1363 | insn_t after) |
1364 | { |
1365 | insn_t insn; |
1366 | |
1367 | gcc_assert (!init_insn_force_unique_p); |
1368 | |
1369 | init_insn_force_unique_p = true; |
1370 | insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); |
1371 | CANT_MOVE (insn) = 1; |
1372 | init_insn_force_unique_p = false; |
1373 | |
1374 | return insn; |
1375 | } |
1376 | |
1377 | /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, |
1378 | take it as a new vinsn instead of EXPR's vinsn. |
1379 | We simplify insns later, after scheduling region in |
1380 | simplify_changed_insns. */ |
1381 | insn_t |
1382 | sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
1383 | insn_t after) |
1384 | { |
1385 | expr_t emit_expr; |
1386 | insn_t insn; |
1387 | int flags; |
1388 | |
1389 | emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), |
1390 | seqno); |
1391 | insn = EXPR_INSN_RTX (emit_expr); |
1392 | |
1393 | /* The insn may come from the transformation cache, which may hold already |
1394 | deleted insns, so mark it as not deleted. */ |
1395 | insn->set_undeleted (); |
1396 | |
1397 | add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); |
1398 | |
1399 | flags = INSN_INIT_TODO_SSID; |
1400 | if (INSN_LUID (insn) == 0) |
1401 | flags |= INSN_INIT_TODO_LUID; |
1402 | sel_init_new_insn (insn, flags); |
1403 | |
1404 | return insn; |
1405 | } |
1406 | |
1407 | /* Move insn from EXPR after AFTER. */ |
1408 | insn_t |
1409 | sel_move_insn (expr_t expr, int seqno, insn_t after) |
1410 | { |
1411 | insn_t insn = EXPR_INSN_RTX (expr); |
1412 | basic_block bb = BLOCK_FOR_INSN (insn: after); |
1413 | insn_t next = NEXT_INSN (insn: after); |
1414 | |
1415 | /* Assert that in move_op we disconnected this insn properly. */ |
1416 | gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); |
1417 | SET_PREV_INSN (insn) = after; |
1418 | SET_NEXT_INSN (insn) = next; |
1419 | |
1420 | SET_NEXT_INSN (after) = insn; |
1421 | SET_PREV_INSN (next) = insn; |
1422 | |
1423 | /* Update links from insn to bb and vice versa. */ |
1424 | df_insn_change_bb (insn, bb); |
1425 | if (BB_END (bb) == after) |
1426 | BB_END (bb) = insn; |
1427 | |
1428 | prepare_insn_expr (insn, seqno); |
1429 | return insn; |
1430 | } |
1431 | |
1432 | |
1433 | /* Functions to work with right-hand sides. */ |
1434 | |
1435 | /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector |
1436 | VECT and return true when found. Use NEW_VINSN for comparison only when |
1437 | COMPARE_VINSNS is true. Write to INDP the index on which |
1438 | the search has stopped, such that inserting the new element at INDP will |
1439 | retain VECT's sort order. */ |
1440 | static bool |
1441 | find_in_history_vect_1 (vec<expr_history_def> vect, |
1442 | unsigned uid, vinsn_t new_vinsn, |
1443 | bool compare_vinsns, int *indp) |
1444 | { |
1445 | expr_history_def *arr; |
1446 | int i, j, len = vect.length (); |
1447 | |
1448 | if (len == 0) |
1449 | { |
1450 | *indp = 0; |
1451 | return false; |
1452 | } |
1453 | |
1454 | arr = vect.address (); |
1455 | i = 0, j = len - 1; |
1456 | |
1457 | while (i <= j) |
1458 | { |
1459 | unsigned auid = arr[i].uid; |
1460 | vinsn_t avinsn = arr[i].new_expr_vinsn; |
1461 | |
1462 | if (auid == uid |
1463 | /* When undoing transformation on a bookkeeping copy, the new vinsn |
1464 | may not be exactly equal to the one that is saved in the vector. |
1465 | This is because the insn whose copy we're checking was possibly |
1466 | substituted itself. */ |
1467 | && (! compare_vinsns |
1468 | || vinsn_equal_p (avinsn, new_vinsn))) |
1469 | { |
1470 | *indp = i; |
1471 | return true; |
1472 | } |
1473 | else if (auid > uid) |
1474 | break; |
1475 | i++; |
1476 | } |
1477 | |
1478 | *indp = i; |
1479 | return false; |
1480 | } |
1481 | |
1482 | /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return |
1483 | the position found or -1, if no such value is in vector. |
1484 | Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ |
1485 | int |
1486 | find_in_history_vect (vec<expr_history_def> vect, rtx insn, |
1487 | vinsn_t new_vinsn, bool originators_p) |
1488 | { |
1489 | int ind; |
1490 | |
1491 | if (find_in_history_vect_1 (vect, uid: INSN_UID (insn), new_vinsn, |
1492 | compare_vinsns: false, indp: &ind)) |
1493 | return ind; |
1494 | |
1495 | if (INSN_ORIGINATORS (insn) && originators_p) |
1496 | { |
1497 | unsigned uid; |
1498 | bitmap_iterator bi; |
1499 | |
1500 | EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) |
1501 | if (find_in_history_vect_1 (vect, uid, new_vinsn, compare_vinsns: false, indp: &ind)) |
1502 | return ind; |
1503 | } |
1504 | |
1505 | return -1; |
1506 | } |
1507 | |
1508 | /* Insert new element in a sorted history vector pointed to by PVECT, |
1509 | if it is not there already. The element is searched using |
1510 | UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save |
1511 | the history of a transformation. */ |
1512 | void |
1513 | insert_in_history_vect (vec<expr_history_def> *pvect, |
1514 | unsigned uid, enum local_trans_type type, |
1515 | vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, |
1516 | ds_t spec_ds) |
1517 | { |
1518 | vec<expr_history_def> vect = *pvect; |
1519 | expr_history_def temp; |
1520 | bool res; |
1521 | int ind; |
1522 | |
1523 | res = find_in_history_vect_1 (vect, uid, new_vinsn: new_expr_vinsn, compare_vinsns: true, indp: &ind); |
1524 | |
1525 | if (res) |
1526 | { |
1527 | expr_history_def *phist = &vect[ind]; |
1528 | |
1529 | /* It is possible that speculation types of expressions that were |
1530 | propagated through different paths will be different here. In this |
1531 | case, merge the status to get the correct check later. */ |
1532 | if (phist->spec_ds != spec_ds) |
1533 | phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); |
1534 | return; |
1535 | } |
1536 | |
1537 | temp.uid = uid; |
1538 | temp.old_expr_vinsn = old_expr_vinsn; |
1539 | temp.new_expr_vinsn = new_expr_vinsn; |
1540 | temp.spec_ds = spec_ds; |
1541 | temp.type = type; |
1542 | |
1543 | vinsn_attach (vi: old_expr_vinsn); |
1544 | vinsn_attach (vi: new_expr_vinsn); |
1545 | vect.safe_insert (ix: ind, obj: temp); |
1546 | *pvect = vect; |
1547 | } |
1548 | |
1549 | /* Free history vector PVECT. */ |
1550 | static void |
1551 | free_history_vect (vec<expr_history_def> &pvect) |
1552 | { |
1553 | unsigned i; |
1554 | expr_history_def *phist; |
1555 | |
1556 | if (! pvect.exists ()) |
1557 | return; |
1558 | |
1559 | for (i = 0; pvect.iterate (ix: i, ptr: &phist); i++) |
1560 | { |
1561 | vinsn_detach (vi: phist->old_expr_vinsn); |
1562 | vinsn_detach (vi: phist->new_expr_vinsn); |
1563 | } |
1564 | |
1565 | pvect.release (); |
1566 | } |
1567 | |
1568 | /* Merge vector FROM to PVECT. */ |
1569 | static void |
1570 | merge_history_vect (vec<expr_history_def> *pvect, |
1571 | vec<expr_history_def> from) |
1572 | { |
1573 | expr_history_def *phist; |
1574 | int i; |
1575 | |
1576 | /* We keep this vector sorted. */ |
1577 | for (i = 0; from.iterate (ix: i, ptr: &phist); i++) |
1578 | insert_in_history_vect (pvect, uid: phist->uid, type: phist->type, |
1579 | old_expr_vinsn: phist->old_expr_vinsn, new_expr_vinsn: phist->new_expr_vinsn, |
1580 | spec_ds: phist->spec_ds); |
1581 | } |
1582 | |
1583 | /* Compare two vinsns as rhses if possible and as vinsns otherwise. */ |
1584 | bool |
1585 | vinsn_equal_p (vinsn_t x, vinsn_t y) |
1586 | { |
1587 | rtx_equal_p_callback_function repcf; |
1588 | |
1589 | if (x == y) |
1590 | return true; |
1591 | |
1592 | if (VINSN_TYPE (x) != VINSN_TYPE (y)) |
1593 | return false; |
1594 | |
1595 | if (VINSN_HASH (x) != VINSN_HASH (y)) |
1596 | return false; |
1597 | |
1598 | repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL; |
1599 | if (VINSN_SEPARABLE_P (x)) |
1600 | { |
1601 | /* Compare RHSes of VINSNs. */ |
1602 | gcc_assert (VINSN_RHS (x)); |
1603 | gcc_assert (VINSN_RHS (y)); |
1604 | |
1605 | return rtx_equal_p (VINSN_RHS (x), VINSN_RHS (y), repcf); |
1606 | } |
1607 | |
1608 | return rtx_equal_p (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); |
1609 | } |
1610 | |
1611 | |
1612 | /* Functions for working with expressions. */ |
1613 | |
1614 | /* Initialize EXPR. */ |
1615 | static void |
1616 | init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority, |
1617 | int sched_times, int orig_bb_index, ds_t spec_done_ds, |
1618 | ds_t spec_to_check_ds, int orig_sched_cycle, |
1619 | vec<expr_history_def> history, |
1620 | signed char target_available, |
1621 | bool was_substituted, bool was_renamed, bool needs_spec_check_p, |
1622 | bool cant_move) |
1623 | { |
1624 | vinsn_attach (vi); |
1625 | |
1626 | EXPR_VINSN (expr) = vi; |
1627 | EXPR_SPEC (expr) = spec; |
1628 | EXPR_USEFULNESS (expr) = use; |
1629 | EXPR_PRIORITY (expr) = priority; |
1630 | EXPR_PRIORITY_ADJ (expr) = 0; |
1631 | EXPR_SCHED_TIMES (expr) = sched_times; |
1632 | EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; |
1633 | EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; |
1634 | EXPR_SPEC_DONE_DS (expr) = spec_done_ds; |
1635 | EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; |
1636 | |
1637 | if (history.exists ()) |
1638 | EXPR_HISTORY_OF_CHANGES (expr) = history; |
1639 | else |
1640 | EXPR_HISTORY_OF_CHANGES (expr).create (nelems: 0); |
1641 | |
1642 | EXPR_TARGET_AVAILABLE (expr) = target_available; |
1643 | EXPR_WAS_SUBSTITUTED (expr) = was_substituted; |
1644 | EXPR_WAS_RENAMED (expr) = was_renamed; |
1645 | EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; |
1646 | EXPR_CANT_MOVE (expr) = cant_move; |
1647 | } |
1648 | |
1649 | /* Make a copy of the expr FROM into the expr TO. */ |
1650 | void |
1651 | copy_expr (expr_t to, expr_t from) |
1652 | { |
1653 | vec<expr_history_def> temp = vNULL; |
1654 | |
1655 | if (EXPR_HISTORY_OF_CHANGES (from).exists ()) |
1656 | { |
1657 | unsigned i; |
1658 | expr_history_def *phist; |
1659 | |
1660 | temp = EXPR_HISTORY_OF_CHANGES (from).copy (); |
1661 | for (i = 0; |
1662 | temp.iterate (ix: i, ptr: &phist); |
1663 | i++) |
1664 | { |
1665 | vinsn_attach (vi: phist->old_expr_vinsn); |
1666 | vinsn_attach (vi: phist->new_expr_vinsn); |
1667 | } |
1668 | } |
1669 | |
1670 | init_expr (expr: to, EXPR_VINSN (from), EXPR_SPEC (from), |
1671 | EXPR_USEFULNESS (from), EXPR_PRIORITY (from), |
1672 | EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), |
1673 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), |
1674 | EXPR_ORIG_SCHED_CYCLE (from), history: temp, |
1675 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
1676 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1677 | EXPR_CANT_MOVE (from)); |
1678 | } |
1679 | |
1680 | /* Same, but the final expr will not ever be in av sets, so don't copy |
1681 | "uninteresting" data such as bitmap cache. */ |
1682 | void |
1683 | copy_expr_onside (expr_t to, expr_t from) |
1684 | { |
1685 | init_expr (expr: to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), |
1686 | EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), orig_bb_index: 0, |
1687 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), orig_sched_cycle: 0, |
1688 | history: vNULL, |
1689 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
1690 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1691 | EXPR_CANT_MOVE (from)); |
1692 | } |
1693 | |
1694 | /* Prepare the expr of INSN for scheduling. Used when moving insn and when |
1695 | initializing new insns. */ |
1696 | static void |
1697 | prepare_insn_expr (insn_t insn, int seqno) |
1698 | { |
1699 | expr_t expr = INSN_EXPR (insn); |
1700 | ds_t ds; |
1701 | |
1702 | INSN_SEQNO (insn) = seqno; |
1703 | EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); |
1704 | EXPR_SPEC (expr) = 0; |
1705 | EXPR_ORIG_SCHED_CYCLE (expr) = 0; |
1706 | EXPR_WAS_SUBSTITUTED (expr) = 0; |
1707 | EXPR_WAS_RENAMED (expr) = 0; |
1708 | EXPR_TARGET_AVAILABLE (expr) = 1; |
1709 | INSN_LIVE_VALID_P (insn) = false; |
1710 | |
1711 | /* ??? If this expression is speculative, make its dependence |
1712 | as weak as possible. We can filter this expression later |
1713 | in process_spec_exprs, because we do not distinguish |
1714 | between the status we got during compute_av_set and the |
1715 | existing status. To be fixed. */ |
1716 | ds = EXPR_SPEC_DONE_DS (expr); |
1717 | if (ds) |
1718 | EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); |
1719 | |
1720 | free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); |
1721 | } |
1722 | |
1723 | /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT |
1724 | is non-null when expressions are merged from different successors at |
1725 | a split point. */ |
1726 | static void |
1727 | update_target_availability (expr_t to, expr_t from, insn_t split_point) |
1728 | { |
1729 | if (EXPR_TARGET_AVAILABLE (to) < 0 |
1730 | || EXPR_TARGET_AVAILABLE (from) < 0) |
1731 | EXPR_TARGET_AVAILABLE (to) = -1; |
1732 | else |
1733 | { |
1734 | /* We try to detect the case when one of the expressions |
1735 | can only be reached through another one. In this case, |
1736 | we can do better. */ |
1737 | if (split_point == NULL) |
1738 | { |
1739 | int toind, fromind; |
1740 | |
1741 | toind = EXPR_ORIG_BB_INDEX (to); |
1742 | fromind = EXPR_ORIG_BB_INDEX (from); |
1743 | |
1744 | if (toind && toind == fromind) |
1745 | /* Do nothing -- everything is done in |
1746 | merge_with_other_exprs. */ |
1747 | ; |
1748 | else |
1749 | EXPR_TARGET_AVAILABLE (to) = -1; |
1750 | } |
1751 | else if (EXPR_TARGET_AVAILABLE (from) == 0 |
1752 | && EXPR_LHS (from) |
1753 | && REG_P (EXPR_LHS (from)) |
1754 | && REGNO (EXPR_LHS (to)) != REGNO (EXPR_LHS (from))) |
1755 | EXPR_TARGET_AVAILABLE (to) = -1; |
1756 | else |
1757 | EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); |
1758 | } |
1759 | } |
1760 | |
1761 | /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT |
1762 | is non-null when expressions are merged from different successors at |
1763 | a split point. */ |
1764 | static void |
1765 | update_speculative_bits (expr_t to, expr_t from, insn_t split_point) |
1766 | { |
1767 | ds_t old_to_ds, old_from_ds; |
1768 | |
1769 | old_to_ds = EXPR_SPEC_DONE_DS (to); |
1770 | old_from_ds = EXPR_SPEC_DONE_DS (from); |
1771 | |
1772 | EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); |
1773 | EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); |
1774 | EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); |
1775 | |
1776 | /* When merging e.g. control & data speculative exprs, or a control |
1777 | speculative with a control&data speculative one, we really have |
1778 | to change vinsn too. Also, when speculative status is changed, |
1779 | we also need to record this as a transformation in expr's history. */ |
1780 | if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) |
1781 | { |
1782 | old_to_ds = ds_get_speculation_types (old_to_ds); |
1783 | old_from_ds = ds_get_speculation_types (old_from_ds); |
1784 | |
1785 | if (old_to_ds != old_from_ds) |
1786 | { |
1787 | ds_t record_ds; |
1788 | |
1789 | /* When both expressions are speculative, we need to change |
1790 | the vinsn first. */ |
1791 | if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) |
1792 | { |
1793 | int res; |
1794 | |
1795 | res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); |
1796 | gcc_assert (res >= 0); |
1797 | } |
1798 | |
1799 | if (split_point != NULL) |
1800 | { |
1801 | /* Record the change with proper status. */ |
1802 | record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; |
1803 | record_ds &= ~(old_to_ds & SPECULATIVE); |
1804 | record_ds &= ~(old_from_ds & SPECULATIVE); |
1805 | |
1806 | insert_in_history_vect (pvect: &EXPR_HISTORY_OF_CHANGES (to), |
1807 | uid: INSN_UID (insn: split_point), type: TRANS_SPECULATION, |
1808 | EXPR_VINSN (from), EXPR_VINSN (to), |
1809 | spec_ds: record_ds); |
1810 | } |
1811 | } |
1812 | } |
1813 | } |
1814 | |
1815 | |
1816 | /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, |
1817 | this is done along different paths. */ |
1818 | void |
1819 | merge_expr_data (expr_t to, expr_t from, insn_t split_point) |
1820 | { |
1821 | /* Choose the maximum of the specs of merged exprs. This is required |
1822 | for correctness of bookkeeping. */ |
1823 | if (EXPR_SPEC (to) < EXPR_SPEC (from)) |
1824 | EXPR_SPEC (to) = EXPR_SPEC (from); |
1825 | |
1826 | if (split_point) |
1827 | EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); |
1828 | else |
1829 | EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), |
1830 | EXPR_USEFULNESS (from)); |
1831 | |
1832 | if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) |
1833 | EXPR_PRIORITY (to) = EXPR_PRIORITY (from); |
1834 | |
1835 | /* We merge sched-times half-way to the larger value to avoid the endless |
1836 | pipelining of unneeded insns. The average seems to be good compromise |
1837 | between pipelining opportunities and avoiding extra work. */ |
1838 | if (EXPR_SCHED_TIMES (to) != EXPR_SCHED_TIMES (from)) |
1839 | EXPR_SCHED_TIMES (to) = ((EXPR_SCHED_TIMES (from) + EXPR_SCHED_TIMES (to) |
1840 | + 1) / 2); |
1841 | |
1842 | if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) |
1843 | EXPR_ORIG_BB_INDEX (to) = 0; |
1844 | |
1845 | EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), |
1846 | EXPR_ORIG_SCHED_CYCLE (from)); |
1847 | |
1848 | EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); |
1849 | EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); |
1850 | EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); |
1851 | |
1852 | merge_history_vect (pvect: &EXPR_HISTORY_OF_CHANGES (to), |
1853 | EXPR_HISTORY_OF_CHANGES (from)); |
1854 | update_target_availability (to, from, split_point); |
1855 | update_speculative_bits (to, from, split_point); |
1856 | } |
1857 | |
1858 | /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal |
1859 | in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions |
1860 | are merged from different successors at a split point. */ |
1861 | void |
1862 | merge_expr (expr_t to, expr_t from, insn_t split_point) |
1863 | { |
1864 | vinsn_t to_vi = EXPR_VINSN (to); |
1865 | vinsn_t from_vi = EXPR_VINSN (from); |
1866 | |
1867 | gcc_assert (vinsn_equal_p (to_vi, from_vi)); |
1868 | |
1869 | /* Make sure that speculative pattern is propagated into exprs that |
1870 | have non-speculative one. This will provide us with consistent |
1871 | speculative bits and speculative patterns inside expr. */ |
1872 | if (EXPR_SPEC_DONE_DS (to) == 0 |
1873 | && (EXPR_SPEC_DONE_DS (from) != 0 |
1874 | /* Do likewise for volatile insns, so that we always retain |
1875 | the may_trap_p bit on the resulting expression. However, |
1876 | avoid propagating the trapping bit into the instructions |
1877 | already speculated. This would result in replacing the |
1878 | speculative pattern with the non-speculative one and breaking |
1879 | the speculation support. */ |
1880 | || (!VINSN_MAY_TRAP_P (EXPR_VINSN (to)) |
1881 | && VINSN_MAY_TRAP_P (EXPR_VINSN (from))))) |
1882 | change_vinsn_in_expr (to, EXPR_VINSN (from)); |
1883 | |
1884 | merge_expr_data (to, from, split_point); |
1885 | gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); |
1886 | } |
1887 | |
1888 | /* Clear the information of this EXPR. */ |
1889 | void |
1890 | clear_expr (expr_t expr) |
1891 | { |
1892 | |
1893 | vinsn_detach (EXPR_VINSN (expr)); |
1894 | EXPR_VINSN (expr) = NULL; |
1895 | |
1896 | free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); |
1897 | } |
1898 | |
1899 | /* For a given LV_SET, mark EXPR having unavailable target register. */ |
1900 | static void |
1901 | set_unavailable_target_for_expr (expr_t expr, regset lv_set) |
1902 | { |
1903 | if (EXPR_SEPARABLE_P (expr)) |
1904 | { |
1905 | if (REG_P (EXPR_LHS (expr)) |
1906 | && register_unavailable_p (lv_set, EXPR_LHS (expr))) |
1907 | { |
1908 | /* If it's an insn like r1 = use (r1, ...), and it exists in |
1909 | different forms in each of the av_sets being merged, we can't say |
1910 | whether original destination register is available or not. |
1911 | However, this still works if destination register is not used |
1912 | in the original expression: if the branch at which LV_SET we're |
1913 | looking here is not actually 'other branch' in sense that same |
1914 | expression is available through it (but it can't be determined |
1915 | at computation stage because of transformations on one of the |
1916 | branches), it still won't affect the availability. |
1917 | Liveness of a register somewhere on a code motion path means |
1918 | it's either read somewhere on a codemotion path, live on |
1919 | 'other' branch, live at the point immediately following |
1920 | the original operation, or is read by the original operation. |
1921 | The latter case is filtered out in the condition below. |
1922 | It still doesn't cover the case when register is defined and used |
1923 | somewhere within the code motion path, and in this case we could |
1924 | miss a unifying code motion along both branches using a renamed |
1925 | register, but it won't affect a code correctness since upon |
1926 | an actual code motion a bookkeeping code would be generated. */ |
1927 | if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
1928 | EXPR_LHS (expr))) |
1929 | EXPR_TARGET_AVAILABLE (expr) = -1; |
1930 | else |
1931 | EXPR_TARGET_AVAILABLE (expr) = false; |
1932 | } |
1933 | } |
1934 | else |
1935 | { |
1936 | unsigned regno; |
1937 | reg_set_iterator rsi; |
1938 | |
1939 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), |
1940 | 0, regno, rsi) |
1941 | if (bitmap_bit_p (lv_set, regno)) |
1942 | { |
1943 | EXPR_TARGET_AVAILABLE (expr) = false; |
1944 | break; |
1945 | } |
1946 | |
1947 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), |
1948 | 0, regno, rsi) |
1949 | if (bitmap_bit_p (lv_set, regno)) |
1950 | { |
1951 | EXPR_TARGET_AVAILABLE (expr) = false; |
1952 | break; |
1953 | } |
1954 | } |
1955 | } |
1956 | |
1957 | /* Try to make EXPR speculative. Return 1 when EXPR's pattern |
1958 | or dependence status have changed, 2 when also the target register |
1959 | became unavailable, 0 if nothing had to be changed. */ |
1960 | int |
1961 | speculate_expr (expr_t expr, ds_t ds) |
1962 | { |
1963 | int res; |
1964 | rtx_insn *orig_insn_rtx; |
1965 | rtx spec_pat; |
1966 | ds_t target_ds, current_ds; |
1967 | |
1968 | /* Obtain the status we need to put on EXPR. */ |
1969 | target_ds = (ds & SPECULATIVE); |
1970 | current_ds = EXPR_SPEC_DONE_DS (expr); |
1971 | ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); |
1972 | |
1973 | orig_insn_rtx = EXPR_INSN_RTX (expr); |
1974 | |
1975 | res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); |
1976 | |
1977 | switch (res) |
1978 | { |
1979 | case 0: |
1980 | EXPR_SPEC_DONE_DS (expr) = ds; |
1981 | return current_ds != ds ? 1 : 0; |
1982 | |
1983 | case 1: |
1984 | { |
1985 | rtx_insn *spec_insn_rtx = |
1986 | create_insn_rtx_from_pattern (spec_pat, NULL_RTX); |
1987 | vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); |
1988 | |
1989 | change_vinsn_in_expr (expr, spec_vinsn); |
1990 | EXPR_SPEC_DONE_DS (expr) = ds; |
1991 | EXPR_NEEDS_SPEC_CHECK_P (expr) = true; |
1992 | |
1993 | /* Do not allow clobbering the address register of speculative |
1994 | insns. */ |
1995 | if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
1996 | expr_dest_reg (expr))) |
1997 | { |
1998 | EXPR_TARGET_AVAILABLE (expr) = false; |
1999 | return 2; |
2000 | } |
2001 | |
2002 | return 1; |
2003 | } |
2004 | |
2005 | case -1: |
2006 | return -1; |
2007 | |
2008 | default: |
2009 | gcc_unreachable (); |
2010 | return -1; |
2011 | } |
2012 | } |
2013 | |
2014 | /* Return a destination register, if any, of EXPR. */ |
2015 | rtx |
2016 | expr_dest_reg (expr_t expr) |
2017 | { |
2018 | rtx dest = VINSN_LHS (EXPR_VINSN (expr)); |
2019 | |
2020 | if (dest != NULL_RTX && REG_P (dest)) |
2021 | return dest; |
2022 | |
2023 | return NULL_RTX; |
2024 | } |
2025 | |
2026 | /* Returns the REGNO of the R's destination. */ |
2027 | unsigned |
2028 | expr_dest_regno (expr_t expr) |
2029 | { |
2030 | rtx dest = expr_dest_reg (expr); |
2031 | |
2032 | gcc_assert (dest != NULL_RTX); |
2033 | return REGNO (dest); |
2034 | } |
2035 | |
2036 | /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in |
2037 | AV_SET having unavailable target register. */ |
2038 | void |
2039 | mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) |
2040 | { |
2041 | expr_t expr; |
2042 | av_set_iterator avi; |
2043 | |
2044 | FOR_EACH_EXPR (expr, avi, join_set) |
2045 | if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) |
2046 | set_unavailable_target_for_expr (expr, lv_set); |
2047 | } |
2048 | |
2049 | |
2050 | /* Returns true if REG (at least partially) is present in REGS. */ |
2051 | bool |
2052 | register_unavailable_p (regset regs, rtx reg) |
2053 | { |
2054 | unsigned regno, end_regno; |
2055 | |
2056 | regno = REGNO (reg); |
2057 | if (bitmap_bit_p (regs, regno)) |
2058 | return true; |
2059 | |
2060 | end_regno = END_REGNO (x: reg); |
2061 | |
2062 | while (++regno < end_regno) |
2063 | if (bitmap_bit_p (regs, regno)) |
2064 | return true; |
2065 | |
2066 | return false; |
2067 | } |
2068 | |
2069 | /* Av set functions. */ |
2070 | |
2071 | /* Add a new element to av set SETP. |
2072 | Return the element added. */ |
2073 | static av_set_t |
2074 | av_set_add_element (av_set_t *setp) |
2075 | { |
2076 | /* Insert at the beginning of the list. */ |
2077 | _list_add (lp: setp); |
2078 | return *setp; |
2079 | } |
2080 | |
2081 | /* Add EXPR to SETP. */ |
2082 | void |
2083 | av_set_add (av_set_t *setp, expr_t expr) |
2084 | { |
2085 | av_set_t elem; |
2086 | |
2087 | gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); |
2088 | elem = av_set_add_element (setp); |
2089 | copy_expr (_AV_SET_EXPR (elem), from: expr); |
2090 | } |
2091 | |
2092 | /* Same, but do not copy EXPR. */ |
2093 | static void |
2094 | av_set_add_nocopy (av_set_t *setp, expr_t expr) |
2095 | { |
2096 | av_set_t elem; |
2097 | |
2098 | elem = av_set_add_element (setp); |
2099 | *_AV_SET_EXPR (elem) = *expr; |
2100 | } |
2101 | |
2102 | /* Remove expr pointed to by IP from the av_set. */ |
2103 | void |
2104 | av_set_iter_remove (av_set_iterator *ip) |
2105 | { |
2106 | clear_expr (_AV_SET_EXPR (*ip->lp)); |
2107 | _list_iter_remove (ip); |
2108 | } |
2109 | |
2110 | /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the |
2111 | sense of vinsn_equal_p function. Return NULL if no such expr is |
2112 | in SET was found. */ |
2113 | expr_t |
2114 | av_set_lookup (av_set_t set, vinsn_t sought_vinsn) |
2115 | { |
2116 | expr_t expr; |
2117 | av_set_iterator i; |
2118 | |
2119 | FOR_EACH_EXPR (expr, i, set) |
2120 | if (vinsn_equal_p (EXPR_VINSN (expr), y: sought_vinsn)) |
2121 | return expr; |
2122 | return NULL; |
2123 | } |
2124 | |
2125 | /* Same, but also remove the EXPR found. */ |
2126 | static expr_t |
2127 | av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) |
2128 | { |
2129 | expr_t expr; |
2130 | av_set_iterator i; |
2131 | |
2132 | FOR_EACH_EXPR_1 (expr, i, setp) |
2133 | if (vinsn_equal_p (EXPR_VINSN (expr), y: sought_vinsn)) |
2134 | { |
2135 | _list_iter_remove_nofree (ip: &i); |
2136 | return expr; |
2137 | } |
2138 | return NULL; |
2139 | } |
2140 | |
2141 | /* Search for an expr in SET, such that it's equivalent to EXPR in the |
2142 | sense of vinsn_equal_p function of their vinsns, but not EXPR itself. |
2143 | Returns NULL if no such expr is in SET was found. */ |
2144 | static expr_t |
2145 | av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) |
2146 | { |
2147 | expr_t cur_expr; |
2148 | av_set_iterator i; |
2149 | |
2150 | FOR_EACH_EXPR (cur_expr, i, set) |
2151 | { |
2152 | if (cur_expr == expr) |
2153 | continue; |
2154 | if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) |
2155 | return cur_expr; |
2156 | } |
2157 | |
2158 | return NULL; |
2159 | } |
2160 | |
2161 | /* If other expression is already in AVP, remove one of them. */ |
2162 | expr_t |
2163 | merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) |
2164 | { |
2165 | expr_t expr2; |
2166 | |
2167 | expr2 = av_set_lookup_other_equiv_expr (set: *avp, expr); |
2168 | if (expr2 != NULL) |
2169 | { |
2170 | /* Reset target availability on merge, since taking it only from one |
2171 | of the exprs would be controversial for different code. */ |
2172 | EXPR_TARGET_AVAILABLE (expr2) = -1; |
2173 | EXPR_USEFULNESS (expr2) = 0; |
2174 | |
2175 | merge_expr (to: expr2, from: expr, NULL); |
2176 | |
2177 | /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ |
2178 | EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; |
2179 | |
2180 | av_set_iter_remove (ip); |
2181 | return expr2; |
2182 | } |
2183 | |
2184 | return expr; |
2185 | } |
2186 | |
2187 | /* Return true if there is an expr that correlates to VI in SET. */ |
2188 | bool |
2189 | av_set_is_in_p (av_set_t set, vinsn_t vi) |
2190 | { |
2191 | return av_set_lookup (set, sought_vinsn: vi) != NULL; |
2192 | } |
2193 | |
2194 | /* Return a copy of SET. */ |
2195 | av_set_t |
2196 | av_set_copy (av_set_t set) |
2197 | { |
2198 | expr_t expr; |
2199 | av_set_iterator i; |
2200 | av_set_t res = NULL; |
2201 | |
2202 | FOR_EACH_EXPR (expr, i, set) |
2203 | av_set_add (setp: &res, expr); |
2204 | |
2205 | return res; |
2206 | } |
2207 | |
2208 | /* Join two av sets that do not have common elements by attaching second set |
2209 | (pointed to by FROMP) to the end of first set (TO_TAILP must point to |
2210 | _AV_SET_NEXT of first set's last element). */ |
2211 | static void |
2212 | join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) |
2213 | { |
2214 | gcc_assert (*to_tailp == NULL); |
2215 | *to_tailp = *fromp; |
2216 | *fromp = NULL; |
2217 | } |
2218 | |
2219 | /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set |
2220 | pointed to by FROMP afterwards. */ |
2221 | void |
2222 | av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) |
2223 | { |
2224 | expr_t expr1; |
2225 | av_set_iterator i; |
2226 | |
2227 | /* Delete from TOP all exprs, that present in FROMP. */ |
2228 | FOR_EACH_EXPR_1 (expr1, i, top) |
2229 | { |
2230 | expr_t expr2 = av_set_lookup (set: *fromp, EXPR_VINSN (expr1)); |
2231 | |
2232 | if (expr2) |
2233 | { |
2234 | merge_expr (to: expr2, from: expr1, split_point: insn); |
2235 | av_set_iter_remove (ip: &i); |
2236 | } |
2237 | } |
2238 | |
2239 | join_distinct_sets (to_tailp: i.lp, fromp); |
2240 | } |
2241 | |
2242 | /* Same as above, but also update availability of target register in |
2243 | TOP judging by TO_LV_SET and FROM_LV_SET. */ |
2244 | void |
2245 | av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, |
2246 | regset from_lv_set, insn_t insn) |
2247 | { |
2248 | expr_t expr1; |
2249 | av_set_iterator i; |
2250 | av_set_t *to_tailp, in_both_set = NULL; |
2251 | |
2252 | /* Delete from TOP all expres, that present in FROMP. */ |
2253 | FOR_EACH_EXPR_1 (expr1, i, top) |
2254 | { |
2255 | expr_t expr2 = av_set_lookup_and_remove (setp: fromp, EXPR_VINSN (expr1)); |
2256 | |
2257 | if (expr2) |
2258 | { |
2259 | /* It may be that the expressions have different destination |
2260 | registers, in which case we need to check liveness here. */ |
2261 | if (EXPR_SEPARABLE_P (expr1)) |
2262 | { |
2263 | int regno1 = (REG_P (EXPR_LHS (expr1)) |
2264 | ? (int) expr_dest_regno (expr: expr1) : -1); |
2265 | int regno2 = (REG_P (EXPR_LHS (expr2)) |
2266 | ? (int) expr_dest_regno (expr: expr2) : -1); |
2267 | |
2268 | /* ??? We don't have a way to check restrictions for |
2269 | *other* register on the current path, we did it only |
2270 | for the current target register. Give up. */ |
2271 | if (regno1 != regno2) |
2272 | EXPR_TARGET_AVAILABLE (expr2) = -1; |
2273 | } |
2274 | else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) |
2275 | EXPR_TARGET_AVAILABLE (expr2) = -1; |
2276 | |
2277 | merge_expr (to: expr2, from: expr1, split_point: insn); |
2278 | av_set_add_nocopy (setp: &in_both_set, expr: expr2); |
2279 | av_set_iter_remove (ip: &i); |
2280 | } |
2281 | else |
2282 | /* EXPR1 is present in TOP, but not in FROMP. Check it on |
2283 | FROM_LV_SET. */ |
2284 | set_unavailable_target_for_expr (expr: expr1, lv_set: from_lv_set); |
2285 | } |
2286 | to_tailp = i.lp; |
2287 | |
2288 | /* These expressions are not present in TOP. Check liveness |
2289 | restrictions on TO_LV_SET. */ |
2290 | FOR_EACH_EXPR (expr1, i, *fromp) |
2291 | set_unavailable_target_for_expr (expr: expr1, lv_set: to_lv_set); |
2292 | |
2293 | join_distinct_sets (to_tailp: i.lp, fromp: &in_both_set); |
2294 | join_distinct_sets (to_tailp, fromp); |
2295 | } |
2296 | |
2297 | /* Clear av_set pointed to by SETP. */ |
2298 | void |
2299 | av_set_clear (av_set_t *setp) |
2300 | { |
2301 | expr_t expr; |
2302 | av_set_iterator i; |
2303 | |
2304 | FOR_EACH_EXPR_1 (expr, i, setp) |
2305 | av_set_iter_remove (ip: &i); |
2306 | |
2307 | gcc_assert (*setp == NULL); |
2308 | } |
2309 | |
2310 | /* Leave only one non-speculative element in the SETP. */ |
2311 | void |
2312 | av_set_leave_one_nonspec (av_set_t *setp) |
2313 | { |
2314 | expr_t expr; |
2315 | av_set_iterator i; |
2316 | bool has_one_nonspec = false; |
2317 | |
2318 | /* Keep all speculative exprs, and leave one non-speculative |
2319 | (the first one). */ |
2320 | FOR_EACH_EXPR_1 (expr, i, setp) |
2321 | { |
2322 | if (!EXPR_SPEC_DONE_DS (expr)) |
2323 | { |
2324 | if (has_one_nonspec) |
2325 | av_set_iter_remove (ip: &i); |
2326 | else |
2327 | has_one_nonspec = true; |
2328 | } |
2329 | } |
2330 | } |
2331 | |
2332 | /* Return the N'th element of the SET. */ |
2333 | expr_t |
2334 | av_set_element (av_set_t set, int n) |
2335 | { |
2336 | expr_t expr; |
2337 | av_set_iterator i; |
2338 | |
2339 | FOR_EACH_EXPR (expr, i, set) |
2340 | if (n-- == 0) |
2341 | return expr; |
2342 | |
2343 | gcc_unreachable (); |
2344 | return NULL; |
2345 | } |
2346 | |
2347 | /* Deletes all expressions from AVP that are conditional branches (IFs). */ |
2348 | void |
2349 | av_set_substract_cond_branches (av_set_t *avp) |
2350 | { |
2351 | av_set_iterator i; |
2352 | expr_t expr; |
2353 | |
2354 | FOR_EACH_EXPR_1 (expr, i, avp) |
2355 | if (vinsn_cond_branch_p (EXPR_VINSN (expr))) |
2356 | av_set_iter_remove (ip: &i); |
2357 | } |
2358 | |
2359 | /* Multiplies usefulness attribute of each member of av-set *AVP by |
2360 | value PROB / ALL_PROB. */ |
2361 | void |
2362 | av_set_split_usefulness (av_set_t av, int prob, int all_prob) |
2363 | { |
2364 | av_set_iterator i; |
2365 | expr_t expr; |
2366 | |
2367 | FOR_EACH_EXPR (expr, i, av) |
2368 | EXPR_USEFULNESS (expr) = (all_prob |
2369 | ? (EXPR_USEFULNESS (expr) * prob) / all_prob |
2370 | : 0); |
2371 | } |
2372 | |
2373 | /* Leave in AVP only those expressions, which are present in AV, |
2374 | and return it, merging history expressions. */ |
2375 | void |
2376 | av_set_code_motion_filter (av_set_t *avp, av_set_t av) |
2377 | { |
2378 | av_set_iterator i; |
2379 | expr_t expr, expr2; |
2380 | |
2381 | FOR_EACH_EXPR_1 (expr, i, avp) |
2382 | if ((expr2 = av_set_lookup (set: av, EXPR_VINSN (expr))) == NULL) |
2383 | av_set_iter_remove (ip: &i); |
2384 | else |
2385 | /* When updating av sets in bookkeeping blocks, we can add more insns |
2386 | there which will be transformed but the upper av sets will not |
2387 | reflect those transformations. We then fail to undo those |
2388 | when searching for such insns. So merge the history saved |
2389 | in the av set of the block we are processing. */ |
2390 | merge_history_vect (pvect: &EXPR_HISTORY_OF_CHANGES (expr), |
2391 | EXPR_HISTORY_OF_CHANGES (expr2)); |
2392 | } |
2393 | |
2394 | |
2395 | |
2396 | /* Dependence hooks to initialize insn data. */ |
2397 | |
2398 | /* This is used in hooks callable from dependence analysis when initializing |
2399 | instruction's data. */ |
2400 | static struct |
2401 | { |
2402 | /* Where the dependence was found (lhs/rhs). */ |
2403 | deps_where_t where; |
2404 | |
2405 | /* The actual data object to initialize. */ |
2406 | idata_t id; |
2407 | |
2408 | /* True when the insn should not be made clonable. */ |
2409 | bool force_unique_p; |
2410 | |
2411 | /* True when insn should be treated as of type USE, i.e. never renamed. */ |
2412 | bool force_use_p; |
2413 | } deps_init_id_data; |
2414 | |
2415 | |
2416 | /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be |
2417 | clonable. */ |
2418 | static void |
2419 | setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) |
2420 | { |
2421 | int type; |
2422 | |
2423 | /* Determine whether INSN could be cloned and return appropriate vinsn type. |
2424 | That clonable insns which can be separated into lhs and rhs have type SET. |
2425 | Other clonable insns have type USE. */ |
2426 | type = GET_CODE (insn); |
2427 | |
2428 | /* Only regular insns could be cloned. */ |
2429 | if (type == INSN && !force_unique_p) |
2430 | type = SET; |
2431 | else if (type == JUMP_INSN && simplejump_p (insn)) |
2432 | type = PC; |
2433 | else if (type == DEBUG_INSN) |
2434 | type = !force_unique_p ? USE : INSN; |
2435 | |
2436 | IDATA_TYPE (id) = type; |
2437 | IDATA_REG_SETS (id) = get_clear_regset_from_pool (); |
2438 | IDATA_REG_USES (id) = get_clear_regset_from_pool (); |
2439 | IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); |
2440 | } |
2441 | |
2442 | /* Start initializing insn data. */ |
2443 | static void |
2444 | deps_init_id_start_insn (insn_t insn) |
2445 | { |
2446 | gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); |
2447 | |
2448 | setup_id_for_insn (id: deps_init_id_data.id, insn, |
2449 | force_unique_p: deps_init_id_data.force_unique_p); |
2450 | deps_init_id_data.where = DEPS_IN_INSN; |
2451 | } |
2452 | |
2453 | /* Start initializing lhs data. */ |
2454 | static void |
2455 | deps_init_id_start_lhs (rtx lhs) |
2456 | { |
2457 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); |
2458 | gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); |
2459 | |
2460 | if (IDATA_TYPE (deps_init_id_data.id) == SET) |
2461 | { |
2462 | IDATA_LHS (deps_init_id_data.id) = lhs; |
2463 | deps_init_id_data.where = DEPS_IN_LHS; |
2464 | } |
2465 | } |
2466 | |
2467 | /* Finish initializing lhs data. */ |
2468 | static void |
2469 | deps_init_id_finish_lhs (void) |
2470 | { |
2471 | deps_init_id_data.where = DEPS_IN_INSN; |
2472 | } |
2473 | |
2474 | /* Note a set of REGNO. */ |
2475 | static void |
2476 | deps_init_id_note_reg_set (int regno) |
2477 | { |
2478 | haifa_note_reg_set (regno); |
2479 | |
2480 | if (deps_init_id_data.where == DEPS_IN_RHS) |
2481 | deps_init_id_data.force_use_p = true; |
2482 | |
2483 | if (IDATA_TYPE (deps_init_id_data.id) != PC) |
2484 | SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); |
2485 | |
2486 | #ifdef STACK_REGS |
2487 | /* Make instructions that set stack registers to be ineligible for |
2488 | renaming to avoid issues with find_used_regs. */ |
2489 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) |
2490 | deps_init_id_data.force_use_p = true; |
2491 | #endif |
2492 | } |
2493 | |
2494 | /* Note a clobber of REGNO. */ |
2495 | static void |
2496 | deps_init_id_note_reg_clobber (int regno) |
2497 | { |
2498 | haifa_note_reg_clobber (regno); |
2499 | |
2500 | if (deps_init_id_data.where == DEPS_IN_RHS) |
2501 | deps_init_id_data.force_use_p = true; |
2502 | |
2503 | if (IDATA_TYPE (deps_init_id_data.id) != PC) |
2504 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); |
2505 | } |
2506 | |
2507 | /* Note a use of REGNO. */ |
2508 | static void |
2509 | deps_init_id_note_reg_use (int regno) |
2510 | { |
2511 | haifa_note_reg_use (regno); |
2512 | |
2513 | if (IDATA_TYPE (deps_init_id_data.id) != PC) |
2514 | SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); |
2515 | } |
2516 | |
2517 | /* Start initializing rhs data. */ |
2518 | static void |
2519 | deps_init_id_start_rhs (rtx rhs) |
2520 | { |
2521 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); |
2522 | |
2523 | /* And there was no sel_deps_reset_to_insn (). */ |
2524 | if (IDATA_LHS (deps_init_id_data.id) != NULL) |
2525 | { |
2526 | IDATA_RHS (deps_init_id_data.id) = rhs; |
2527 | deps_init_id_data.where = DEPS_IN_RHS; |
2528 | } |
2529 | } |
2530 | |
2531 | /* Finish initializing rhs data. */ |
2532 | static void |
2533 | deps_init_id_finish_rhs (void) |
2534 | { |
2535 | gcc_assert (deps_init_id_data.where == DEPS_IN_RHS |
2536 | || deps_init_id_data.where == DEPS_IN_INSN); |
2537 | deps_init_id_data.where = DEPS_IN_INSN; |
2538 | } |
2539 | |
2540 | /* Finish initializing insn data. */ |
2541 | static void |
2542 | deps_init_id_finish_insn (void) |
2543 | { |
2544 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); |
2545 | |
2546 | if (IDATA_TYPE (deps_init_id_data.id) == SET) |
2547 | { |
2548 | rtx lhs = IDATA_LHS (deps_init_id_data.id); |
2549 | rtx rhs = IDATA_RHS (deps_init_id_data.id); |
2550 | |
2551 | if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) |
2552 | || deps_init_id_data.force_use_p) |
2553 | { |
2554 | /* This should be a USE, as we don't want to schedule its RHS |
2555 | separately. However, we still want to have them recorded |
2556 | for the purposes of substitution. That's why we don't |
2557 | simply call downgrade_to_use () here. */ |
2558 | gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); |
2559 | gcc_assert (!lhs == !rhs); |
2560 | |
2561 | IDATA_TYPE (deps_init_id_data.id) = USE; |
2562 | } |
2563 | } |
2564 | |
2565 | deps_init_id_data.where = DEPS_IN_NOWHERE; |
2566 | } |
2567 | |
2568 | /* This is dependence info used for initializing insn's data. */ |
2569 | static struct sched_deps_info_def deps_init_id_sched_deps_info; |
2570 | |
2571 | /* This initializes most of the static part of the above structure. */ |
2572 | static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = |
2573 | { |
2574 | NULL, |
2575 | |
2576 | .start_insn: deps_init_id_start_insn, |
2577 | .finish_insn: deps_init_id_finish_insn, |
2578 | .start_lhs: deps_init_id_start_lhs, |
2579 | .finish_lhs: deps_init_id_finish_lhs, |
2580 | .start_rhs: deps_init_id_start_rhs, |
2581 | .finish_rhs: deps_init_id_finish_rhs, |
2582 | .note_reg_set: deps_init_id_note_reg_set, |
2583 | .note_reg_clobber: deps_init_id_note_reg_clobber, |
2584 | .note_reg_use: deps_init_id_note_reg_use, |
2585 | NULL, /* note_mem_dep */ |
2586 | NULL, /* note_dep */ |
2587 | |
2588 | .use_cselib: 0, /* use_cselib */ |
2589 | .use_deps_list: 0, /* use_deps_list */ |
2590 | .generate_spec_deps: 0 /* generate_spec_deps */ |
2591 | }; |
2592 | |
2593 | /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, |
2594 | we don't actually need information about lhs and rhs. */ |
2595 | static void |
2596 | setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) |
2597 | { |
2598 | rtx pat = PATTERN (insn); |
2599 | |
2600 | if (NONJUMP_INSN_P (insn) |
2601 | && GET_CODE (pat) == SET |
2602 | && !force_unique_p) |
2603 | { |
2604 | IDATA_RHS (id) = SET_SRC (pat); |
2605 | IDATA_LHS (id) = SET_DEST (pat); |
2606 | } |
2607 | else |
2608 | IDATA_LHS (id) = IDATA_RHS (id) = NULL; |
2609 | } |
2610 | |
2611 | /* Possibly downgrade INSN to USE. */ |
2612 | static void |
2613 | maybe_downgrade_id_to_use (idata_t id, insn_t insn) |
2614 | { |
2615 | bool must_be_use = false; |
2616 | df_ref def; |
2617 | rtx lhs = IDATA_LHS (id); |
2618 | rtx rhs = IDATA_RHS (id); |
2619 | |
2620 | /* We downgrade only SETs. */ |
2621 | if (IDATA_TYPE (id) != SET) |
2622 | return; |
2623 | |
2624 | if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) |
2625 | { |
2626 | IDATA_TYPE (id) = USE; |
2627 | return; |
2628 | } |
2629 | |
2630 | FOR_EACH_INSN_DEF (def, insn) |
2631 | { |
2632 | if (DF_REF_INSN (def) |
2633 | && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) |
2634 | && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) |
2635 | { |
2636 | must_be_use = true; |
2637 | break; |
2638 | } |
2639 | |
2640 | #ifdef STACK_REGS |
2641 | /* Make instructions that set stack registers to be ineligible for |
2642 | renaming to avoid issues with find_used_regs. */ |
2643 | if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) |
2644 | { |
2645 | must_be_use = true; |
2646 | break; |
2647 | } |
2648 | #endif |
2649 | } |
2650 | |
2651 | if (must_be_use) |
2652 | IDATA_TYPE (id) = USE; |
2653 | } |
2654 | |
2655 | /* Setup implicit register clobbers calculated by sched-deps for INSN |
2656 | before reload and save them in ID. */ |
2657 | static void |
2658 | setup_id_implicit_regs (idata_t id, insn_t insn) |
2659 | { |
2660 | if (reload_completed) |
2661 | return; |
2662 | |
2663 | HARD_REG_SET temp; |
2664 | |
2665 | get_implicit_reg_pending_clobbers (&temp, insn); |
2666 | IOR_REG_SET_HRS (IDATA_REG_SETS (id), temp); |
2667 | } |
2668 | |
2669 | /* Setup register sets describing INSN in ID. */ |
2670 | static void |
2671 | setup_id_reg_sets (idata_t id, insn_t insn) |
2672 | { |
2673 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
2674 | df_ref def, use; |
2675 | regset tmp = get_clear_regset_from_pool (); |
2676 | |
2677 | FOR_EACH_INSN_INFO_DEF (def, insn_info) |
2678 | { |
2679 | unsigned int regno = DF_REF_REGNO (def); |
2680 | |
2681 | /* Post modifies are treated like clobbers by sched-deps.cc. */ |
2682 | if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER |
2683 | | DF_REF_PRE_POST_MODIFY))) |
2684 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); |
2685 | else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) |
2686 | { |
2687 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); |
2688 | |
2689 | #ifdef STACK_REGS |
2690 | /* For stack registers, treat writes to them as writes |
2691 | to the first one to be consistent with sched-deps.cc. */ |
2692 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) |
2693 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); |
2694 | #endif |
2695 | } |
2696 | /* Mark special refs that generate read/write def pair. */ |
2697 | if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) |
2698 | || regno == STACK_POINTER_REGNUM) |
2699 | bitmap_set_bit (tmp, regno); |
2700 | } |
2701 | |
2702 | FOR_EACH_INSN_INFO_USE (use, insn_info) |
2703 | { |
2704 | unsigned int regno = DF_REF_REGNO (use); |
2705 | |
2706 | /* When these refs are met for the first time, skip them, as |
2707 | these uses are just counterparts of some defs. */ |
2708 | if (bitmap_bit_p (tmp, regno)) |
2709 | bitmap_clear_bit (tmp, regno); |
2710 | else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) |
2711 | { |
2712 | SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); |
2713 | |
2714 | #ifdef STACK_REGS |
2715 | /* For stack registers, treat reads from them as reads from |
2716 | the first one to be consistent with sched-deps.cc. */ |
2717 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) |
2718 | SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); |
2719 | #endif |
2720 | } |
2721 | } |
2722 | |
2723 | /* Also get implicit reg clobbers from sched-deps. */ |
2724 | setup_id_implicit_regs (id, insn); |
2725 | |
2726 | return_regset_to_pool (rs: tmp); |
2727 | } |
2728 | |
2729 | /* Initialize instruction data for INSN in ID using DF's data. */ |
2730 | static void |
2731 | init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) |
2732 | { |
2733 | gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); |
2734 | |
2735 | setup_id_for_insn (id, insn, force_unique_p); |
2736 | setup_id_lhs_rhs (id, insn, force_unique_p); |
2737 | |
2738 | if (INSN_NOP_P (insn)) |
2739 | return; |
2740 | |
2741 | maybe_downgrade_id_to_use (id, insn); |
2742 | setup_id_reg_sets (id, insn); |
2743 | } |
2744 | |
2745 | /* Initialize instruction data for INSN in ID. */ |
2746 | static void |
2747 | deps_init_id (idata_t id, insn_t insn, bool force_unique_p) |
2748 | { |
2749 | class deps_desc _dc, *dc = &_dc; |
2750 | |
2751 | deps_init_id_data.where = DEPS_IN_NOWHERE; |
2752 | deps_init_id_data.id = id; |
2753 | deps_init_id_data.force_unique_p = force_unique_p; |
2754 | deps_init_id_data.force_use_p = false; |
2755 | |
2756 | init_deps (dc, false); |
2757 | memcpy (dest: &deps_init_id_sched_deps_info, |
2758 | src: &const_deps_init_id_sched_deps_info, |
2759 | n: sizeof (deps_init_id_sched_deps_info)); |
2760 | if (spec_info != NULL) |
2761 | deps_init_id_sched_deps_info.generate_spec_deps = 1; |
2762 | sched_deps_info = &deps_init_id_sched_deps_info; |
2763 | |
2764 | deps_analyze_insn (dc, insn); |
2765 | /* Implicit reg clobbers received from sched-deps separately. */ |
2766 | setup_id_implicit_regs (id, insn); |
2767 | |
2768 | free_deps (dc); |
2769 | deps_init_id_data.id = NULL; |
2770 | } |
2771 | |
2772 | |
2773 | struct sched_scan_info_def |
2774 | { |
2775 | /* This hook notifies scheduler frontend to extend its internal per basic |
2776 | block data structures. This hook should be called once before a series of |
2777 | calls to bb_init (). */ |
2778 | void (*extend_bb) (void); |
2779 | |
2780 | /* This hook makes scheduler frontend to initialize its internal data |
2781 | structures for the passed basic block. */ |
2782 | void (*init_bb) (basic_block); |
2783 | |
2784 | /* This hook notifies scheduler frontend to extend its internal per insn data |
2785 | structures. This hook should be called once before a series of calls to |
2786 | insn_init (). */ |
2787 | void (*extend_insn) (void); |
2788 | |
2789 | /* This hook makes scheduler frontend to initialize its internal data |
2790 | structures for the passed insn. */ |
2791 | void (*init_insn) (insn_t); |
2792 | }; |
2793 | |
2794 | /* A driver function to add a set of basic blocks (BBS) to the |
2795 | scheduling region. */ |
2796 | static void |
2797 | sched_scan (const struct sched_scan_info_def *ssi, bb_vec_t bbs) |
2798 | { |
2799 | unsigned i; |
2800 | basic_block bb; |
2801 | |
2802 | if (ssi->extend_bb) |
2803 | ssi->extend_bb (); |
2804 | |
2805 | if (ssi->init_bb) |
2806 | FOR_EACH_VEC_ELT (bbs, i, bb) |
2807 | ssi->init_bb (bb); |
2808 | |
2809 | if (ssi->extend_insn) |
2810 | ssi->extend_insn (); |
2811 | |
2812 | if (ssi->init_insn) |
2813 | FOR_EACH_VEC_ELT (bbs, i, bb) |
2814 | { |
2815 | rtx_insn *insn; |
2816 | |
2817 | FOR_BB_INSNS (bb, insn) |
2818 | ssi->init_insn (insn); |
2819 | } |
2820 | } |
2821 | |
2822 | /* Implement hooks for collecting fundamental insn properties like if insn is |
2823 | an ASM or is within a SCHED_GROUP. */ |
2824 | |
2825 | /* True when a "one-time init" data for INSN was already inited. */ |
2826 | static bool |
2827 | first_time_insn_init (insn_t insn) |
2828 | { |
2829 | return INSN_LIVE (insn) == NULL; |
2830 | } |
2831 | |
2832 | /* Hash an entry in a transformed_insns hashtable. */ |
2833 | static hashval_t |
2834 | hash_transformed_insns (const void *p) |
2835 | { |
2836 | return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); |
2837 | } |
2838 | |
2839 | /* Compare the entries in a transformed_insns hashtable. */ |
2840 | static int |
2841 | eq_transformed_insns (const void *p, const void *q) |
2842 | { |
2843 | rtx_insn *i1 = |
2844 | VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); |
2845 | rtx_insn *i2 = |
2846 | VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); |
2847 | |
2848 | if (INSN_UID (insn: i1) == INSN_UID (insn: i2)) |
2849 | return 1; |
2850 | return rtx_equal_p (PATTERN (insn: i1), PATTERN (insn: i2)); |
2851 | } |
2852 | |
2853 | /* Free an entry in a transformed_insns hashtable. */ |
2854 | static void |
2855 | free_transformed_insns (void *p) |
2856 | { |
2857 | struct transformed_insns *pti = (struct transformed_insns *) p; |
2858 | |
2859 | vinsn_detach (vi: pti->vinsn_old); |
2860 | vinsn_detach (vi: pti->vinsn_new); |
2861 | free (ptr: pti); |
2862 | } |
2863 | |
2864 | /* Init the s_i_d data for INSN which should be inited just once, when |
2865 | we first see the insn. */ |
2866 | static void |
2867 | init_first_time_insn_data (insn_t insn) |
2868 | { |
2869 | /* This should not be set if this is the first time we init data for |
2870 | insn. */ |
2871 | gcc_assert (first_time_insn_init (insn)); |
2872 | |
2873 | /* These are needed for nops too. */ |
2874 | INSN_LIVE (insn) = get_regset_from_pool (); |
2875 | INSN_LIVE_VALID_P (insn) = false; |
2876 | |
2877 | if (!INSN_NOP_P (insn)) |
2878 | { |
2879 | INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); |
2880 | INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); |
2881 | INSN_TRANSFORMED_INSNS (insn) |
2882 | = htab_create (16, hash_transformed_insns, |
2883 | eq_transformed_insns, free_transformed_insns); |
2884 | init_deps (&INSN_DEPS_CONTEXT (insn), true); |
2885 | } |
2886 | } |
2887 | |
2888 | /* Free almost all above data for INSN that is scheduled already. |
2889 | Used for extra-large basic blocks. */ |
2890 | void |
2891 | free_data_for_scheduled_insn (insn_t insn) |
2892 | { |
2893 | gcc_assert (! first_time_insn_init (insn)); |
2894 | |
2895 | if (! INSN_ANALYZED_DEPS (insn)) |
2896 | return; |
2897 | |
2898 | BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); |
2899 | BITMAP_FREE (INSN_FOUND_DEPS (insn)); |
2900 | htab_delete (INSN_TRANSFORMED_INSNS (insn)); |
2901 | |
2902 | /* This is allocated only for bookkeeping insns. */ |
2903 | if (INSN_ORIGINATORS (insn)) |
2904 | BITMAP_FREE (INSN_ORIGINATORS (insn)); |
2905 | free_deps (&INSN_DEPS_CONTEXT (insn)); |
2906 | |
2907 | INSN_ANALYZED_DEPS (insn) = NULL; |
2908 | |
2909 | /* Clear the readonly flag so we would ICE when trying to recalculate |
2910 | the deps context (as we believe that it should not happen). */ |
2911 | (&INSN_DEPS_CONTEXT (insn))->readonly = 0; |
2912 | } |
2913 | |
2914 | /* Free the same data as above for INSN. */ |
2915 | static void |
2916 | free_first_time_insn_data (insn_t insn) |
2917 | { |
2918 | gcc_assert (! first_time_insn_init (insn)); |
2919 | |
2920 | free_data_for_scheduled_insn (insn); |
2921 | return_regset_to_pool (INSN_LIVE (insn)); |
2922 | INSN_LIVE (insn) = NULL; |
2923 | INSN_LIVE_VALID_P (insn) = false; |
2924 | } |
2925 | |
2926 | /* Initialize region-scope data structures for basic blocks. */ |
2927 | static void |
2928 | init_global_and_expr_for_bb (basic_block bb) |
2929 | { |
2930 | if (sel_bb_empty_p (bb)) |
2931 | return; |
2932 | |
2933 | invalidate_av_set (bb); |
2934 | } |
2935 | |
2936 | /* Data for global dependency analysis (to initialize CANT_MOVE and |
2937 | SCHED_GROUP_P). */ |
2938 | static struct |
2939 | { |
2940 | /* Previous insn. */ |
2941 | insn_t prev_insn; |
2942 | } init_global_data; |
2943 | |
2944 | /* Determine if INSN is in the sched_group, is an asm or should not be |
2945 | cloned. After that initialize its expr. */ |
2946 | static void |
2947 | init_global_and_expr_for_insn (insn_t insn) |
2948 | { |
2949 | if (LABEL_P (insn)) |
2950 | return; |
2951 | |
2952 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) |
2953 | { |
2954 | init_global_data.prev_insn = NULL; |
2955 | return; |
2956 | } |
2957 | |
2958 | gcc_assert (INSN_P (insn)); |
2959 | |
2960 | if (SCHED_GROUP_P (insn)) |
2961 | /* Setup a sched_group. */ |
2962 | { |
2963 | insn_t prev_insn = init_global_data.prev_insn; |
2964 | |
2965 | if (prev_insn) |
2966 | INSN_SCHED_NEXT (prev_insn) = insn; |
2967 | |
2968 | init_global_data.prev_insn = insn; |
2969 | } |
2970 | else |
2971 | init_global_data.prev_insn = NULL; |
2972 | |
2973 | if (GET_CODE (PATTERN (insn)) == ASM_INPUT |
2974 | || asm_noperands (PATTERN (insn)) >= 0) |
2975 | /* Mark INSN as an asm. */ |
2976 | INSN_ASM_P (insn) = true; |
2977 | |
2978 | { |
2979 | bool force_unique_p; |
2980 | ds_t spec_done_ds; |
2981 | |
2982 | /* Certain instructions cannot be cloned, and frame related insns and |
2983 | the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of |
2984 | their block. */ |
2985 | if (prologue_epilogue_contains (insn)) |
2986 | { |
2987 | if (RTX_FRAME_RELATED_P (insn)) |
2988 | CANT_MOVE (insn) = 1; |
2989 | else |
2990 | { |
2991 | rtx note; |
2992 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) |
2993 | if (REG_NOTE_KIND (note) == REG_SAVE_NOTE |
2994 | && ((enum insn_note) INTVAL (XEXP (note, 0)) |
2995 | == NOTE_INSN_EPILOGUE_BEG)) |
2996 | { |
2997 | CANT_MOVE (insn) = 1; |
2998 | break; |
2999 | } |
3000 | } |
3001 | force_unique_p = true; |
3002 | } |
3003 | else |
3004 | if (CANT_MOVE (insn) |
3005 | || INSN_ASM_P (insn) |
3006 | || SCHED_GROUP_P (insn) |
3007 | || CALL_P (insn) |
3008 | /* Exception handling insns are always unique. */ |
3009 | || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) |
3010 | /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ |
3011 | || control_flow_insn_p (insn) |
3012 | || volatile_insn_p (PATTERN (insn)) |
3013 | || (targetm.cannot_copy_insn_p |
3014 | && targetm.cannot_copy_insn_p (insn))) |
3015 | force_unique_p = true; |
3016 | else |
3017 | force_unique_p = false; |
3018 | |
3019 | if (targetm.sched.get_insn_spec_ds) |
3020 | { |
3021 | spec_done_ds = targetm.sched.get_insn_spec_ds (insn); |
3022 | spec_done_ds = ds_get_max_dep_weak (spec_done_ds); |
3023 | } |
3024 | else |
3025 | spec_done_ds = 0; |
3026 | |
3027 | /* Initialize INSN's expr. */ |
3028 | init_expr (INSN_EXPR (insn), vi: vinsn_create (insn, force_unique_p), spec: 0, |
3029 | REG_BR_PROB_BASE, INSN_PRIORITY (insn), sched_times: 0, BLOCK_NUM (insn), |
3030 | spec_done_ds, spec_to_check_ds: 0, orig_sched_cycle: 0, history: vNULL, target_available: true, |
3031 | was_substituted: false, was_renamed: false, needs_spec_check_p: false, CANT_MOVE (insn)); |
3032 | } |
3033 | |
3034 | init_first_time_insn_data (insn); |
3035 | } |
3036 | |
3037 | /* Scan the region and initialize instruction data for basic blocks BBS. */ |
3038 | void |
3039 | sel_init_global_and_expr (bb_vec_t bbs) |
3040 | { |
3041 | /* ??? It would be nice to implement push / pop scheme for sched_infos. */ |
3042 | const struct sched_scan_info_def ssi = |
3043 | { |
3044 | NULL, /* extend_bb */ |
3045 | .init_bb: init_global_and_expr_for_bb, /* init_bb */ |
3046 | .extend_insn: extend_insn_data, /* extend_insn */ |
3047 | .init_insn: init_global_and_expr_for_insn /* init_insn */ |
3048 | }; |
3049 | |
3050 | sched_scan (ssi: &ssi, bbs); |
3051 | } |
3052 | |
3053 | /* Finalize region-scope data structures for basic blocks. */ |
3054 | static void |
3055 | finish_global_and_expr_for_bb (basic_block bb) |
3056 | { |
3057 | av_set_clear (setp: &BB_AV_SET (bb)); |
3058 | BB_AV_LEVEL (bb) = 0; |
3059 | } |
3060 | |
3061 | /* Finalize INSN's data. */ |
3062 | static void |
3063 | finish_global_and_expr_insn (insn_t insn) |
3064 | { |
3065 | if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) |
3066 | return; |
3067 | |
3068 | gcc_assert (INSN_P (insn)); |
3069 | |
3070 | if (INSN_LUID (insn) > 0) |
3071 | { |
3072 | free_first_time_insn_data (insn); |
3073 | INSN_WS_LEVEL (insn) = 0; |
3074 | CANT_MOVE (insn) = 0; |
3075 | |
3076 | /* We can no longer assert this, as vinsns of this insn could be |
3077 | easily live in other insn's caches. This should be changed to |
3078 | a counter-like approach among all vinsns. */ |
3079 | gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); |
3080 | clear_expr (INSN_EXPR (insn)); |
3081 | } |
3082 | } |
3083 | |
3084 | /* Finalize per instruction data for the whole region. */ |
3085 | void |
3086 | sel_finish_global_and_expr (void) |
3087 | { |
3088 | { |
3089 | bb_vec_t bbs; |
3090 | int i; |
3091 | |
3092 | bbs.create (nelems: current_nr_blocks); |
3093 | |
3094 | for (i = 0; i < current_nr_blocks; i++) |
3095 | bbs.quick_push (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))); |
3096 | |
3097 | /* Clear AV_SETs and INSN_EXPRs. */ |
3098 | { |
3099 | const struct sched_scan_info_def ssi = |
3100 | { |
3101 | NULL, /* extend_bb */ |
3102 | .init_bb: finish_global_and_expr_for_bb, /* init_bb */ |
3103 | NULL, /* extend_insn */ |
3104 | .init_insn: finish_global_and_expr_insn /* init_insn */ |
3105 | }; |
3106 | |
3107 | sched_scan (ssi: &ssi, bbs); |
3108 | } |
3109 | |
3110 | bbs.release (); |
3111 | } |
3112 | |
3113 | finish_insns (); |
3114 | } |
3115 | |
3116 | |
3117 | /* In the below hooks, we merely calculate whether or not a dependence |
3118 | exists, and in what part of insn. However, we will need more data |
3119 | when we'll start caching dependence requests. */ |
3120 | |
3121 | /* Container to hold information for dependency analysis. */ |
3122 | static struct |
3123 | { |
3124 | deps_t dc; |
3125 | |
3126 | /* A variable to track which part of rtx we are scanning in |
3127 | sched-deps.cc: sched_analyze_insn (). */ |
3128 | deps_where_t where; |
3129 | |
3130 | /* Current producer. */ |
3131 | insn_t pro; |
3132 | |
3133 | /* Current consumer. */ |
3134 | vinsn_t con; |
3135 | |
3136 | /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. |
3137 | X is from { INSN, LHS, RHS }. */ |
3138 | ds_t has_dep_p[DEPS_IN_NOWHERE]; |
3139 | } has_dependence_data; |
3140 | |
3141 | /* Start analyzing dependencies of INSN. */ |
3142 | static void |
3143 | has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) |
3144 | { |
3145 | gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); |
3146 | |
3147 | has_dependence_data.where = DEPS_IN_INSN; |
3148 | } |
3149 | |
3150 | /* Finish analyzing dependencies of an insn. */ |
3151 | static void |
3152 | has_dependence_finish_insn (void) |
3153 | { |
3154 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); |
3155 | |
3156 | has_dependence_data.where = DEPS_IN_NOWHERE; |
3157 | } |
3158 | |
3159 | /* Start analyzing dependencies of LHS. */ |
3160 | static void |
3161 | has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) |
3162 | { |
3163 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); |
3164 | |
3165 | if (VINSN_LHS (has_dependence_data.con) != NULL) |
3166 | has_dependence_data.where = DEPS_IN_LHS; |
3167 | } |
3168 | |
3169 | /* Finish analyzing dependencies of an lhs. */ |
3170 | static void |
3171 | has_dependence_finish_lhs (void) |
3172 | { |
3173 | has_dependence_data.where = DEPS_IN_INSN; |
3174 | } |
3175 | |
3176 | /* Start analyzing dependencies of RHS. */ |
3177 | static void |
3178 | has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) |
3179 | { |
3180 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); |
3181 | |
3182 | if (VINSN_RHS (has_dependence_data.con) != NULL) |
3183 | has_dependence_data.where = DEPS_IN_RHS; |
3184 | } |
3185 | |
3186 | /* Start analyzing dependencies of an rhs. */ |
3187 | static void |
3188 | has_dependence_finish_rhs (void) |
3189 | { |
3190 | gcc_assert (has_dependence_data.where == DEPS_IN_RHS |
3191 | || has_dependence_data.where == DEPS_IN_INSN); |
3192 | |
3193 | has_dependence_data.where = DEPS_IN_INSN; |
3194 | } |
3195 | |
3196 | /* Note a set of REGNO. */ |
3197 | static void |
3198 | has_dependence_note_reg_set (int regno) |
3199 | { |
3200 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; |
3201 | |
3202 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, |
3203 | VINSN_INSN_RTX |
3204 | (has_dependence_data.con))) |
3205 | { |
3206 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; |
3207 | |
3208 | if (reg_last->sets != NULL |
3209 | || reg_last->clobbers != NULL) |
3210 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; |
3211 | |
3212 | if (reg_last->uses || reg_last->implicit_sets) |
3213 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3214 | } |
3215 | } |
3216 | |
3217 | /* Note a clobber of REGNO. */ |
3218 | static void |
3219 | has_dependence_note_reg_clobber (int regno) |
3220 | { |
3221 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; |
3222 | |
3223 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, |
3224 | VINSN_INSN_RTX |
3225 | (has_dependence_data.con))) |
3226 | { |
3227 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; |
3228 | |
3229 | if (reg_last->sets) |
3230 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; |
3231 | |
3232 | if (reg_last->uses || reg_last->implicit_sets) |
3233 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3234 | } |
3235 | } |
3236 | |
3237 | /* Note a use of REGNO. */ |
3238 | static void |
3239 | has_dependence_note_reg_use (int regno) |
3240 | { |
3241 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; |
3242 | |
3243 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, |
3244 | VINSN_INSN_RTX |
3245 | (has_dependence_data.con))) |
3246 | { |
3247 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; |
3248 | |
3249 | if (reg_last->sets) |
3250 | *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; |
3251 | |
3252 | if (reg_last->clobbers || reg_last->implicit_sets) |
3253 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3254 | |
3255 | /* Merge BE_IN_SPEC bits into *DSP when the dependency producer |
3256 | is actually a check insn. We need to do this for any register |
3257 | read-read dependency with the check unless we track properly |
3258 | all registers written by BE_IN_SPEC-speculated insns, as |
3259 | we don't have explicit dependence lists. See PR 53975. */ |
3260 | if (reg_last->uses) |
3261 | { |
3262 | ds_t pro_spec_checked_ds; |
3263 | |
3264 | pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); |
3265 | pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); |
3266 | |
3267 | if (pro_spec_checked_ds != 0) |
3268 | *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, |
3269 | NULL_RTX, NULL_RTX); |
3270 | } |
3271 | } |
3272 | } |
3273 | |
3274 | /* Note a memory dependence. */ |
3275 | static void |
3276 | has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, |
3277 | rtx pending_mem ATTRIBUTE_UNUSED, |
3278 | insn_t pending_insn ATTRIBUTE_UNUSED, |
3279 | ds_t ds ATTRIBUTE_UNUSED) |
3280 | { |
3281 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, |
3282 | VINSN_INSN_RTX (has_dependence_data.con))) |
3283 | { |
3284 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; |
3285 | |
3286 | *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); |
3287 | } |
3288 | } |
3289 | |
3290 | /* Note a dependence. */ |
3291 | static void |
3292 | has_dependence_note_dep (insn_t pro, ds_t ds ATTRIBUTE_UNUSED) |
3293 | { |
3294 | insn_t real_pro = has_dependence_data.pro; |
3295 | insn_t real_con = VINSN_INSN_RTX (has_dependence_data.con); |
3296 | |
3297 | /* We do not allow for debug insns to move through others unless they |
3298 | are at the start of bb. This movement may create bookkeeping copies |
3299 | that later would not be able to move up, violating the invariant |
3300 | that a bookkeeping copy should be movable as the original insn. |
3301 | Detect that here and allow that movement if we allowed it before |
3302 | in the first place. */ |
3303 | if (DEBUG_INSN_P (real_con) && !DEBUG_INSN_P (real_pro) |
3304 | && INSN_UID (insn: NEXT_INSN (insn: pro)) == INSN_UID (insn: real_con)) |
3305 | return; |
3306 | |
3307 | if (!sched_insns_conditions_mutex_p (real_pro, real_con)) |
3308 | { |
3309 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; |
3310 | |
3311 | *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); |
3312 | } |
3313 | } |
3314 | |
3315 | /* Mark the insn as having a hard dependence that prevents speculation. */ |
3316 | void |
3317 | sel_mark_hard_insn (rtx insn) |
3318 | { |
3319 | int i; |
3320 | |
3321 | /* Only work when we're in has_dependence_p mode. |
3322 | ??? This is a hack, this should actually be a hook. */ |
3323 | if (!has_dependence_data.dc || !has_dependence_data.pro) |
3324 | return; |
3325 | |
3326 | gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); |
3327 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); |
3328 | |
3329 | for (i = 0; i < DEPS_IN_NOWHERE; i++) |
3330 | has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; |
3331 | } |
3332 | |
3333 | /* This structure holds the hooks for the dependency analysis used when |
3334 | actually processing dependencies in the scheduler. */ |
3335 | static struct sched_deps_info_def has_dependence_sched_deps_info; |
3336 | |
3337 | /* This initializes most of the fields of the above structure. */ |
3338 | static const struct sched_deps_info_def const_has_dependence_sched_deps_info = |
3339 | { |
3340 | NULL, |
3341 | |
3342 | .start_insn: has_dependence_start_insn, |
3343 | .finish_insn: has_dependence_finish_insn, |
3344 | .start_lhs: has_dependence_start_lhs, |
3345 | .finish_lhs: has_dependence_finish_lhs, |
3346 | .start_rhs: has_dependence_start_rhs, |
3347 | .finish_rhs: has_dependence_finish_rhs, |
3348 | .note_reg_set: has_dependence_note_reg_set, |
3349 | .note_reg_clobber: has_dependence_note_reg_clobber, |
3350 | .note_reg_use: has_dependence_note_reg_use, |
3351 | .note_mem_dep: has_dependence_note_mem_dep, |
3352 | .note_dep: has_dependence_note_dep, |
3353 | |
3354 | .use_cselib: 0, /* use_cselib */ |
3355 | .use_deps_list: 0, /* use_deps_list */ |
3356 | .generate_spec_deps: 0 /* generate_spec_deps */ |
3357 | }; |
3358 | |
3359 | /* Initialize has_dependence_sched_deps_info with extra spec field. */ |
3360 | static void |
3361 | setup_has_dependence_sched_deps_info (void) |
3362 | { |
3363 | memcpy (dest: &has_dependence_sched_deps_info, |
3364 | src: &const_has_dependence_sched_deps_info, |
3365 | n: sizeof (has_dependence_sched_deps_info)); |
3366 | |
3367 | if (spec_info != NULL) |
3368 | has_dependence_sched_deps_info.generate_spec_deps = 1; |
3369 | |
3370 | sched_deps_info = &has_dependence_sched_deps_info; |
3371 | } |
3372 | |
3373 | /* Remove all dependences found and recorded in has_dependence_data array. */ |
3374 | void |
3375 | sel_clear_has_dependence (void) |
3376 | { |
3377 | int i; |
3378 | |
3379 | for (i = 0; i < DEPS_IN_NOWHERE; i++) |
3380 | has_dependence_data.has_dep_p[i] = 0; |
3381 | } |
3382 | |
3383 | /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer |
3384 | to the dependence information array in HAS_DEP_PP. */ |
3385 | ds_t |
3386 | has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) |
3387 | { |
3388 | int i; |
3389 | ds_t ds; |
3390 | class deps_desc *dc; |
3391 | |
3392 | if (INSN_SIMPLEJUMP_P (pred)) |
3393 | /* Unconditional jump is just a transfer of control flow. |
3394 | Ignore it. */ |
3395 | return false; |
3396 | |
3397 | dc = &INSN_DEPS_CONTEXT (pred); |
3398 | |
3399 | /* We init this field lazily. */ |
3400 | if (dc->reg_last == NULL) |
3401 | init_deps_reg_last (dc); |
3402 | |
3403 | if (!dc->readonly) |
3404 | { |
3405 | has_dependence_data.pro = NULL; |
3406 | /* Initialize empty dep context with information about PRED. */ |
3407 | advance_deps_context (dc, insn: pred); |
3408 | dc->readonly = 1; |
3409 | } |
3410 | |
3411 | has_dependence_data.where = DEPS_IN_NOWHERE; |
3412 | has_dependence_data.pro = pred; |
3413 | has_dependence_data.con = EXPR_VINSN (expr); |
3414 | has_dependence_data.dc = dc; |
3415 | |
3416 | sel_clear_has_dependence (); |
3417 | |
3418 | /* Now catch all dependencies that would be generated between PRED and |
3419 | INSN. */ |
3420 | setup_has_dependence_sched_deps_info (); |
3421 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); |
3422 | has_dependence_data.dc = NULL; |
3423 | |
3424 | /* When a barrier was found, set DEPS_IN_INSN bits. */ |
3425 | if (dc->last_reg_pending_barrier == TRUE_BARRIER) |
3426 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; |
3427 | else if (dc->last_reg_pending_barrier == MOVE_BARRIER) |
3428 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; |
3429 | |
3430 | /* Do not allow stores to memory to move through checks. Currently |
3431 | we don't move this to sched-deps.cc as the check doesn't have |
3432 | obvious places to which this dependence can be attached. |
3433 | FIMXE: this should go to a hook. */ |
3434 | if (EXPR_LHS (expr) |
3435 | && MEM_P (EXPR_LHS (expr)) |
3436 | && sel_insn_is_speculation_check (pred)) |
3437 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; |
3438 | |
3439 | *has_dep_pp = has_dependence_data.has_dep_p; |
3440 | ds = 0; |
3441 | for (i = 0; i < DEPS_IN_NOWHERE; i++) |
3442 | ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], |
3443 | NULL_RTX, NULL_RTX); |
3444 | |
3445 | return ds; |
3446 | } |
3447 | |
3448 | |
3449 | /* Dependence hooks implementation that checks dependence latency constraints |
3450 | on the insns being scheduled. The entry point for these routines is |
3451 | tick_check_p predicate. */ |
3452 | |
3453 | static struct |
3454 | { |
3455 | /* An expr we are currently checking. */ |
3456 | expr_t expr; |
3457 | |
3458 | /* A minimal cycle for its scheduling. */ |
3459 | int cycle; |
3460 | |
3461 | /* Whether we have seen a true dependence while checking. */ |
3462 | bool seen_true_dep_p; |
3463 | } tick_check_data; |
3464 | |
3465 | /* Update minimal scheduling cycle for tick_check_insn given that it depends |
3466 | on PRO with status DS and weight DW. */ |
3467 | static void |
3468 | tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) |
3469 | { |
3470 | expr_t con_expr = tick_check_data.expr; |
3471 | insn_t con_insn = EXPR_INSN_RTX (con_expr); |
3472 | |
3473 | if (con_insn != pro_insn) |
3474 | { |
3475 | enum reg_note dt; |
3476 | int tick; |
3477 | |
3478 | if (/* PROducer was removed from above due to pipelining. */ |
3479 | !INSN_IN_STREAM_P (pro_insn) |
3480 | /* Or PROducer was originally on the next iteration regarding the |
3481 | CONsumer. */ |
3482 | || (INSN_SCHED_TIMES (pro_insn) |
3483 | - EXPR_SCHED_TIMES (con_expr)) > 1) |
3484 | /* Don't count this dependence. */ |
3485 | return; |
3486 | |
3487 | dt = ds_to_dt (ds); |
3488 | if (dt == REG_DEP_TRUE) |
3489 | tick_check_data.seen_true_dep_p = true; |
3490 | |
3491 | gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); |
3492 | |
3493 | { |
3494 | dep_def _dep, *dep = &_dep; |
3495 | |
3496 | init_dep (dep, pro_insn, con_insn, dt); |
3497 | |
3498 | tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); |
3499 | } |
3500 | |
3501 | /* When there are several kinds of dependencies between pro and con, |
3502 | only REG_DEP_TRUE should be taken into account. */ |
3503 | if (tick > tick_check_data.cycle |
3504 | && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) |
3505 | tick_check_data.cycle = tick; |
3506 | } |
3507 | } |
3508 | |
3509 | /* An implementation of note_dep hook. */ |
3510 | static void |
3511 | tick_check_note_dep (insn_t pro, ds_t ds) |
3512 | { |
3513 | tick_check_dep_with_dw (pro_insn: pro, ds, dw: 0); |
3514 | } |
3515 | |
3516 | /* An implementation of note_mem_dep hook. */ |
3517 | static void |
3518 | tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) |
3519 | { |
3520 | dw_t dw; |
3521 | |
3522 | dw = (ds_to_dt (ds) == REG_DEP_TRUE |
3523 | ? estimate_dep_weak (mem1, mem2) |
3524 | : 0); |
3525 | |
3526 | tick_check_dep_with_dw (pro_insn: pro, ds, dw); |
3527 | } |
3528 | |
3529 | /* This structure contains hooks for dependence analysis used when determining |
3530 | whether an insn is ready for scheduling. */ |
3531 | static struct sched_deps_info_def tick_check_sched_deps_info = |
3532 | { |
3533 | NULL, |
3534 | |
3535 | NULL, |
3536 | NULL, |
3537 | NULL, |
3538 | NULL, |
3539 | NULL, |
3540 | NULL, |
3541 | .note_reg_set: haifa_note_reg_set, |
3542 | .note_reg_clobber: haifa_note_reg_clobber, |
3543 | .note_reg_use: haifa_note_reg_use, |
3544 | .note_mem_dep: tick_check_note_mem_dep, |
3545 | .note_dep: tick_check_note_dep, |
3546 | |
3547 | .use_cselib: 0, .use_deps_list: 0, .generate_spec_deps: 0 |
3548 | }; |
3549 | |
3550 | /* Estimate number of cycles from the current cycle of FENCE until EXPR can be |
3551 | scheduled. Return 0 if all data from producers in DC is ready. */ |
3552 | int |
3553 | tick_check_p (expr_t expr, deps_t dc, fence_t fence) |
3554 | { |
3555 | int cycles_left; |
3556 | /* Initialize variables. */ |
3557 | tick_check_data.expr = expr; |
3558 | tick_check_data.cycle = 0; |
3559 | tick_check_data.seen_true_dep_p = false; |
3560 | sched_deps_info = &tick_check_sched_deps_info; |
3561 | |
3562 | gcc_assert (!dc->readonly); |
3563 | dc->readonly = 1; |
3564 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); |
3565 | dc->readonly = 0; |
3566 | |
3567 | cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); |
3568 | |
3569 | return cycles_left >= 0 ? cycles_left : 0; |
3570 | } |
3571 | |
3572 | |
3573 | /* Functions to work with insns. */ |
3574 | |
3575 | /* Returns true if LHS of INSN is the same as DEST of an insn |
3576 | being moved. */ |
3577 | bool |
3578 | lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) |
3579 | { |
3580 | rtx lhs = INSN_LHS (insn); |
3581 | |
3582 | if (lhs == NULL || dest == NULL) |
3583 | return false; |
3584 | |
3585 | return rtx_equal_p (lhs, dest); |
3586 | } |
3587 | |
3588 | /* Return s_i_d entry of INSN. Callable from debugger. */ |
3589 | sel_insn_data_def |
3590 | insn_sid (insn_t insn) |
3591 | { |
3592 | return *SID (insn); |
3593 | } |
3594 | |
3595 | /* True when INSN is a speculative check. We can tell this by looking |
3596 | at the data structures of the selective scheduler, not by examining |
3597 | the pattern. */ |
3598 | bool |
3599 | sel_insn_is_speculation_check (rtx insn) |
3600 | { |
3601 | return s_i_d.exists () && !! INSN_SPEC_CHECKED_DS (insn); |
3602 | } |
3603 | |
3604 | /* Extracts machine mode MODE and destination location DST_LOC |
3605 | for given INSN. */ |
3606 | void |
3607 | get_dest_and_mode (rtx insn, rtx *dst_loc, machine_mode *mode) |
3608 | { |
3609 | rtx pat = PATTERN (insn); |
3610 | |
3611 | gcc_assert (dst_loc); |
3612 | gcc_assert (GET_CODE (pat) == SET); |
3613 | |
3614 | *dst_loc = SET_DEST (pat); |
3615 | |
3616 | gcc_assert (*dst_loc); |
3617 | gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); |
3618 | |
3619 | if (mode) |
3620 | *mode = GET_MODE (*dst_loc); |
3621 | } |
3622 | |
3623 | /* Returns true when moving through JUMP will result in bookkeeping |
3624 | creation. */ |
3625 | bool |
3626 | bookkeeping_can_be_created_if_moved_through_p (insn_t jump) |
3627 | { |
3628 | insn_t succ; |
3629 | succ_iterator si; |
3630 | |
3631 | FOR_EACH_SUCC (succ, si, jump) |
3632 | if (sel_num_cfg_preds_gt_1 (succ)) |
3633 | return true; |
3634 | |
3635 | return false; |
3636 | } |
3637 | |
3638 | /* Return 'true' if INSN is the only one in its basic block. */ |
3639 | static bool |
3640 | insn_is_the_only_one_in_bb_p (insn_t insn) |
3641 | { |
3642 | return sel_bb_head_p (insn) && sel_bb_end_p (insn); |
3643 | } |
3644 | |
3645 | /* Check that the region we're scheduling still has at most one |
3646 | backedge. */ |
3647 | static void |
3648 | verify_backedges (void) |
3649 | { |
3650 | if (pipelining_p) |
3651 | { |
3652 | int i, n = 0; |
3653 | edge e; |
3654 | edge_iterator ei; |
3655 | |
3656 | for (i = 0; i < current_nr_blocks; i++) |
3657 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))->succs) |
3658 | if (in_current_region_p (e->dest) |
3659 | && BLOCK_TO_BB (e->dest->index) < i) |
3660 | n++; |
3661 | |
3662 | gcc_assert (n <= 1); |
3663 | } |
3664 | } |
3665 | |
3666 | |
3667 | /* Functions to work with control flow. */ |
3668 | |
3669 | /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks |
3670 | are sorted in topological order (it might have been invalidated by |
3671 | redirecting an edge). */ |
3672 | static void |
3673 | sel_recompute_toporder (void) |
3674 | { |
3675 | int i, n, rgn; |
3676 | int *postorder, n_blocks; |
3677 | |
3678 | postorder = XALLOCAVEC (int, n_basic_blocks_for_fn (cfun)); |
3679 | n_blocks = post_order_compute (postorder, false, false); |
3680 | |
3681 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); |
3682 | for (n = 0, i = n_blocks - 1; i >= 0; i--) |
3683 | if (CONTAINING_RGN (postorder[i]) == rgn) |
3684 | { |
3685 | BLOCK_TO_BB (postorder[i]) = n; |
3686 | BB_TO_BLOCK (n) = postorder[i]; |
3687 | n++; |
3688 | } |
3689 | |
3690 | /* Assert that we updated info for all blocks. We may miss some blocks if |
3691 | this function is called when redirecting an edge made a block |
3692 | unreachable, but that block is not deleted yet. */ |
3693 | gcc_assert (n == RGN_NR_BLOCKS (rgn)); |
3694 | } |
3695 | |
3696 | /* Tidy the possibly empty block BB. */ |
3697 | static bool |
3698 | maybe_tidy_empty_bb (basic_block bb) |
3699 | { |
3700 | basic_block succ_bb, pred_bb, note_bb; |
3701 | vec<basic_block> dom_bbs; |
3702 | edge e; |
3703 | edge_iterator ei; |
3704 | bool rescan_p; |
3705 | |
3706 | /* Keep empty bb only if this block immediately precedes EXIT and |
3707 | has incoming non-fallthrough edge, or it has no predecessors or |
3708 | successors. Otherwise remove it. */ |
3709 | if (!sel_bb_empty_p (bb) |
3710 | || (single_succ_p (bb) |
3711 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
3712 | && (!single_pred_p (bb) |
3713 | || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU))) |
3714 | || EDGE_COUNT (bb->preds) == 0 |
3715 | || EDGE_COUNT (bb->succs) == 0) |
3716 | return false; |
3717 | |
3718 | /* Do not attempt to redirect complex edges. */ |
3719 | FOR_EACH_EDGE (e, ei, bb->preds) |
3720 | if (e->flags & EDGE_COMPLEX) |
3721 | return false; |
3722 | else if (e->flags & EDGE_FALLTHRU) |
3723 | { |
3724 | rtx note; |
3725 | /* If prev bb ends with asm goto, see if any of the |
3726 | ASM_OPERANDS_LABELs don't point to the fallthru |
3727 | label. Do not attempt to redirect it in that case. */ |
3728 | if (JUMP_P (BB_END (e->src)) |
3729 | && (note = extract_asm_operands (PATTERN (BB_END (e->src))))) |
3730 | { |
3731 | int i, n = ASM_OPERANDS_LABEL_LENGTH (note); |
3732 | |
3733 | for (i = 0; i < n; ++i) |
3734 | if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (bb)) |
3735 | return false; |
3736 | } |
3737 | } |
3738 | |
3739 | free_data_sets (bb); |
3740 | |
3741 | /* Do not delete BB if it has more than one successor. |
3742 | That can occur when we moving a jump. */ |
3743 | if (!single_succ_p (bb)) |
3744 | { |
3745 | gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); |
3746 | sel_merge_blocks (bb->prev_bb, bb); |
3747 | return true; |
3748 | } |
3749 | |
3750 | succ_bb = single_succ (bb); |
3751 | rescan_p = true; |
3752 | pred_bb = NULL; |
3753 | dom_bbs.create (nelems: 0); |
3754 | |
3755 | /* Save a pred/succ from the current region to attach the notes to. */ |
3756 | note_bb = NULL; |
3757 | FOR_EACH_EDGE (e, ei, bb->preds) |
3758 | if (in_current_region_p (e->src)) |
3759 | { |
3760 | note_bb = e->src; |
3761 | break; |
3762 | } |
3763 | if (note_bb == NULL) |
3764 | note_bb = succ_bb; |
3765 | |
3766 | /* Redirect all non-fallthru edges to the next bb. */ |
3767 | while (rescan_p) |
3768 | { |
3769 | rescan_p = false; |
3770 | |
3771 | FOR_EACH_EDGE (e, ei, bb->preds) |
3772 | { |
3773 | pred_bb = e->src; |
3774 | |
3775 | if (!(e->flags & EDGE_FALLTHRU)) |
3776 | { |
3777 | /* We cannot invalidate computed topological order by moving |
3778 | the edge destination block (E->SUCC) along a fallthru edge. |
3779 | |
3780 | We will update dominators here only when we'll get |
3781 | an unreachable block when redirecting, otherwise |
3782 | sel_redirect_edge_and_branch will take care of it. */ |
3783 | if (e->dest != bb |
3784 | && single_pred_p (bb: e->dest)) |
3785 | dom_bbs.safe_push (obj: e->dest); |
3786 | sel_redirect_edge_and_branch (e, succ_bb); |
3787 | rescan_p = true; |
3788 | break; |
3789 | } |
3790 | /* If the edge is fallthru, but PRED_BB ends in a conditional jump |
3791 | to BB (so there is no non-fallthru edge from PRED_BB to BB), we |
3792 | still have to adjust it. */ |
3793 | else if (single_succ_p (bb: pred_bb) && any_condjump_p (BB_END (pred_bb))) |
3794 | { |
3795 | /* If possible, try to remove the unneeded conditional jump. */ |
3796 | if (onlyjump_p (BB_END (pred_bb)) |
3797 | && INSN_SCHED_TIMES (BB_END (pred_bb)) == 0 |
3798 | && !IN_CURRENT_FENCE_P (BB_END (pred_bb))) |
3799 | { |
3800 | if (!sel_remove_insn (BB_END (pred_bb), false, false)) |
3801 | tidy_fallthru_edge (e); |
3802 | } |
3803 | else |
3804 | sel_redirect_edge_and_branch (e, succ_bb); |
3805 | rescan_p = true; |
3806 | break; |
3807 | } |
3808 | } |
3809 | } |
3810 | |
3811 | if (can_merge_blocks_p (bb->prev_bb, bb)) |
3812 | sel_merge_blocks (bb->prev_bb, bb); |
3813 | else |
3814 | { |
3815 | /* This is a block without fallthru predecessor. Just delete it. */ |
3816 | gcc_assert (note_bb); |
3817 | move_bb_info (note_bb, bb); |
3818 | remove_empty_bb (bb, true); |
3819 | } |
3820 | |
3821 | if (!dom_bbs.is_empty ()) |
3822 | { |
3823 | dom_bbs.safe_push (obj: succ_bb); |
3824 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); |
3825 | dom_bbs.release (); |
3826 | } |
3827 | |
3828 | return true; |
3829 | } |
3830 | |
3831 | /* Tidy the control flow after we have removed original insn from |
3832 | XBB. Return true if we have removed some blocks. When FULL_TIDYING |
3833 | is true, also try to optimize control flow on non-empty blocks. */ |
3834 | bool |
3835 | tidy_control_flow (basic_block xbb, bool full_tidying) |
3836 | { |
3837 | bool changed = true; |
3838 | insn_t first, last; |
3839 | |
3840 | /* First check whether XBB is empty. */ |
3841 | changed = maybe_tidy_empty_bb (bb: xbb); |
3842 | if (changed || !full_tidying) |
3843 | return changed; |
3844 | |
3845 | /* Check if there is a unnecessary jump after insn left. */ |
3846 | if (bb_has_removable_jump_to_p (xbb, xbb->next_bb) |
3847 | && INSN_SCHED_TIMES (BB_END (xbb)) == 0 |
3848 | && !IN_CURRENT_FENCE_P (BB_END (xbb))) |
3849 | { |
3850 | /* We used to call sel_remove_insn here that can trigger tidy_control_flow |
3851 | before we fix up the fallthru edge. Correct that ordering by |
3852 | explicitly doing the latter before the former. */ |
3853 | clear_expr (INSN_EXPR (BB_END (xbb))); |
3854 | tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); |
3855 | if (tidy_control_flow (xbb, full_tidying: false)) |
3856 | return true; |
3857 | } |
3858 | |
3859 | first = sel_bb_head (xbb); |
3860 | last = sel_bb_end (xbb); |
3861 | if (MAY_HAVE_DEBUG_INSNS) |
3862 | { |
3863 | if (first != last && DEBUG_INSN_P (first)) |
3864 | do |
3865 | first = NEXT_INSN (insn: first); |
3866 | while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); |
3867 | |
3868 | if (first != last && DEBUG_INSN_P (last)) |
3869 | do |
3870 | last = PREV_INSN (insn: last); |
3871 | while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); |
3872 | } |
3873 | /* Check if there is an unnecessary jump in previous basic block leading |
3874 | to next basic block left after removing INSN from stream. |
3875 | If it is so, remove that jump and redirect edge to current |
3876 | basic block (where there was INSN before deletion). This way |
3877 | when NOP will be deleted several instructions later with its |
3878 | basic block we will not get a jump to next instruction, which |
3879 | can be harmful. */ |
3880 | if (first == last |
3881 | && !sel_bb_empty_p (xbb) |
3882 | && INSN_NOP_P (last) |
3883 | /* Flow goes fallthru from current block to the next. */ |
3884 | && EDGE_COUNT (xbb->succs) == 1 |
3885 | && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) |
3886 | /* When successor is an EXIT block, it may not be the next block. */ |
3887 | && single_succ (bb: xbb) != EXIT_BLOCK_PTR_FOR_FN (cfun) |
3888 | /* And unconditional jump in previous basic block leads to |
3889 | next basic block of XBB and this jump can be safely removed. */ |
3890 | && in_current_region_p (xbb->prev_bb) |
3891 | && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb) |
3892 | && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 |
3893 | /* Also this jump is not at the scheduling boundary. */ |
3894 | && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) |
3895 | { |
3896 | bool recompute_toporder_p; |
3897 | /* Clear data structures of jump - jump itself will be removed |
3898 | by sel_redirect_edge_and_branch. */ |
3899 | clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); |
3900 | recompute_toporder_p |
3901 | = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); |
3902 | |
3903 | gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); |
3904 | |
3905 | /* We could have skipped some debug insns which did not get removed with the block, |
3906 | and the seqnos could become incorrect. Fix them up here. */ |
3907 | if (MAY_HAVE_DEBUG_INSNS && (sel_bb_head (xbb) != first || sel_bb_end (xbb) != last)) |
3908 | { |
3909 | if (!sel_bb_empty_p (xbb->prev_bb)) |
3910 | { |
3911 | int prev_seqno = INSN_SEQNO (sel_bb_end (xbb->prev_bb)); |
3912 | if (prev_seqno > INSN_SEQNO (sel_bb_head (xbb))) |
3913 | for (insn_t insn = sel_bb_head (xbb); insn != first; insn = NEXT_INSN (insn)) |
3914 | INSN_SEQNO (insn) = prev_seqno + 1; |
3915 | } |
3916 | } |
3917 | |
3918 | /* It can turn out that after removing unused jump, basic block |
3919 | that contained that jump, becomes empty too. In such case |
3920 | remove it too. */ |
3921 | if (sel_bb_empty_p (xbb->prev_bb)) |
3922 | changed = maybe_tidy_empty_bb (bb: xbb->prev_bb); |
3923 | if (recompute_toporder_p) |
3924 | sel_recompute_toporder (); |
3925 | } |
3926 | |
3927 | /* TODO: use separate flag for CFG checking. */ |
3928 | if (flag_checking) |
3929 | { |
3930 | verify_backedges (); |
3931 | verify_dominators (CDI_DOMINATORS); |
3932 | } |
3933 | |
3934 | return changed; |
3935 | } |
3936 | |
3937 | /* Purge meaningless empty blocks in the middle of a region. */ |
3938 | void |
3939 | purge_empty_blocks (void) |
3940 | { |
3941 | int i; |
3942 | |
3943 | /* Do not attempt to delete the first basic block in the region. */ |
3944 | for (i = 1; i < current_nr_blocks; ) |
3945 | { |
3946 | basic_block b = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
3947 | |
3948 | if (maybe_tidy_empty_bb (bb: b)) |
3949 | continue; |
3950 | |
3951 | i++; |
3952 | } |
3953 | } |
3954 | |
3955 | /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, |
3956 | do not delete insn's data, because it will be later re-emitted. |
3957 | Return true if we have removed some blocks afterwards. */ |
3958 | bool |
3959 | sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) |
3960 | { |
3961 | basic_block bb = BLOCK_FOR_INSN (insn); |
3962 | |
3963 | gcc_assert (INSN_IN_STREAM_P (insn)); |
3964 | |
3965 | if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) |
3966 | { |
3967 | expr_t expr; |
3968 | av_set_iterator i; |
3969 | |
3970 | /* When we remove a debug insn that is head of a BB, it remains |
3971 | in the AV_SET of the block, but it shouldn't. */ |
3972 | FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) |
3973 | if (EXPR_INSN_RTX (expr) == insn) |
3974 | { |
3975 | av_set_iter_remove (ip: &i); |
3976 | break; |
3977 | } |
3978 | } |
3979 | |
3980 | if (only_disconnect) |
3981 | remove_insn (insn); |
3982 | else |
3983 | { |
3984 | delete_insn (insn); |
3985 | clear_expr (INSN_EXPR (insn)); |
3986 | } |
3987 | |
3988 | /* It is necessary to NULL these fields in case we are going to re-insert |
3989 | INSN into the insns stream, as will usually happen in the ONLY_DISCONNECT |
3990 | case, but also for NOPs that we will return to the nop pool. */ |
3991 | SET_PREV_INSN (insn) = NULL_RTX; |
3992 | SET_NEXT_INSN (insn) = NULL_RTX; |
3993 | set_block_for_insn (insn, NULL); |
3994 | |
3995 | return tidy_control_flow (xbb: bb, full_tidying); |
3996 | } |
3997 | |
3998 | /* Estimate number of the insns in BB. */ |
3999 | static int |
4000 | sel_estimate_number_of_insns (basic_block bb) |
4001 | { |
4002 | int res = 0; |
4003 | insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); |
4004 | |
4005 | for (; insn != next_tail; insn = NEXT_INSN (insn)) |
4006 | if (NONDEBUG_INSN_P (insn)) |
4007 | res++; |
4008 | |
4009 | return res; |
4010 | } |
4011 | |
4012 | /* We don't need separate luids for notes or labels. */ |
4013 | static int |
4014 | sel_luid_for_non_insn (rtx x) |
4015 | { |
4016 | gcc_assert (NOTE_P (x) || LABEL_P (x)); |
4017 | |
4018 | return -1; |
4019 | } |
4020 | |
4021 | /* Find the proper seqno for inserting at INSN by successors. |
4022 | Return -1 if no successors with positive seqno exist. */ |
4023 | static int |
4024 | get_seqno_by_succs (rtx_insn *insn) |
4025 | { |
4026 | basic_block bb = BLOCK_FOR_INSN (insn); |
4027 | rtx_insn *tmp = insn, *end = BB_END (bb); |
4028 | int seqno; |
4029 | insn_t succ = NULL; |
4030 | succ_iterator si; |
4031 | |
4032 | while (tmp != end) |
4033 | { |
4034 | tmp = NEXT_INSN (insn: tmp); |
4035 | if (INSN_P (tmp)) |
4036 | return INSN_SEQNO (tmp); |
4037 | } |
4038 | |
4039 | seqno = INT_MAX; |
4040 | |
4041 | FOR_EACH_SUCC_1 (succ, si, end, SUCCS_NORMAL) |
4042 | if (INSN_SEQNO (succ) > 0) |
4043 | seqno = MIN (seqno, INSN_SEQNO (succ)); |
4044 | |
4045 | if (seqno == INT_MAX) |
4046 | return -1; |
4047 | |
4048 | return seqno; |
4049 | } |
4050 | |
4051 | /* Compute seqno for INSN by its preds or succs. Use OLD_SEQNO to compute |
4052 | seqno in corner cases. */ |
4053 | static int |
4054 | get_seqno_for_a_jump (insn_t insn, int old_seqno) |
4055 | { |
4056 | int seqno; |
4057 | |
4058 | gcc_assert (INSN_SIMPLEJUMP_P (insn)); |
4059 | |
4060 | if (!sel_bb_head_p (insn)) |
4061 | seqno = INSN_SEQNO (PREV_INSN (insn)); |
4062 | else |
4063 | { |
4064 | basic_block bb = BLOCK_FOR_INSN (insn); |
4065 | |
4066 | if (single_pred_p (bb) |
4067 | && !in_current_region_p (single_pred (bb))) |
4068 | { |
4069 | /* We can have preds outside a region when splitting edges |
4070 | for pipelining of an outer loop. Use succ instead. |
4071 | There should be only one of them. */ |
4072 | insn_t succ = NULL; |
4073 | succ_iterator si; |
4074 | bool first = true; |
4075 | |
4076 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
4077 | && current_loop_nest); |
4078 | FOR_EACH_SUCC_1 (succ, si, insn, |
4079 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
4080 | { |
4081 | gcc_assert (first); |
4082 | first = false; |
4083 | } |
4084 | |
4085 | gcc_assert (succ != NULL); |
4086 | seqno = INSN_SEQNO (succ); |
4087 | } |
4088 | else |
4089 | { |
4090 | insn_t *preds; |
4091 | int n; |
4092 | |
4093 | cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); |
4094 | |
4095 | gcc_assert (n > 0); |
4096 | /* For one predecessor, use simple method. */ |
4097 | if (n == 1) |
4098 | seqno = INSN_SEQNO (preds[0]); |
4099 | else |
4100 | seqno = get_seqno_by_preds (insn); |
4101 | |
4102 | free (ptr: preds); |
4103 | } |
4104 | } |
4105 | |
4106 | /* We were unable to find a good seqno among preds. */ |
4107 | if (seqno < 0) |
4108 | seqno = get_seqno_by_succs (insn); |
4109 | |
4110 | if (seqno < 0) |
4111 | { |
4112 | /* The only case where this could be here legally is that the only |
4113 | unscheduled insn was a conditional jump that got removed and turned |
4114 | into this unconditional one. Initialize from the old seqno |
4115 | of that jump passed down to here. */ |
4116 | seqno = old_seqno; |
4117 | } |
4118 | |
4119 | gcc_assert (seqno >= 0); |
4120 | return seqno; |
4121 | } |
4122 | |
4123 | /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors |
4124 | with positive seqno exist. */ |
4125 | int |
4126 | get_seqno_by_preds (rtx_insn *insn) |
4127 | { |
4128 | basic_block bb = BLOCK_FOR_INSN (insn); |
4129 | rtx_insn *tmp = insn, *head = BB_HEAD (bb); |
4130 | insn_t *preds; |
4131 | int n, i, seqno; |
4132 | |
4133 | /* Loop backwards from INSN to HEAD including both. */ |
4134 | while (1) |
4135 | { |
4136 | if (INSN_P (tmp)) |
4137 | return INSN_SEQNO (tmp); |
4138 | if (tmp == head) |
4139 | break; |
4140 | tmp = PREV_INSN (insn: tmp); |
4141 | } |
4142 | |
4143 | cfg_preds (bb, &preds, &n); |
4144 | for (i = 0, seqno = -1; i < n; i++) |
4145 | seqno = MAX (seqno, INSN_SEQNO (preds[i])); |
4146 | |
4147 | return seqno; |
4148 | } |
4149 | |
4150 | |
4151 | |
4152 | /* Extend pass-scope data structures for basic blocks. */ |
4153 | void |
4154 | sel_extend_global_bb_info (void) |
4155 | { |
4156 | sel_global_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun), exact: true); |
4157 | } |
4158 | |
4159 | /* Extend region-scope data structures for basic blocks. */ |
4160 | static void |
4161 | extend_region_bb_info (void) |
4162 | { |
4163 | sel_region_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun), exact: true); |
4164 | } |
4165 | |
4166 | /* Extend all data structures to fit for all basic blocks. */ |
4167 | static void |
4168 | extend_bb_info (void) |
4169 | { |
4170 | sel_extend_global_bb_info (); |
4171 | extend_region_bb_info (); |
4172 | } |
4173 | |
4174 | /* Finalize pass-scope data structures for basic blocks. */ |
4175 | void |
4176 | sel_finish_global_bb_info (void) |
4177 | { |
4178 | sel_global_bb_info.release (); |
4179 | } |
4180 | |
4181 | /* Finalize region-scope data structures for basic blocks. */ |
4182 | static void |
4183 | finish_region_bb_info (void) |
4184 | { |
4185 | sel_region_bb_info.release (); |
4186 | } |
4187 | |
4188 | |
4189 | /* Data for each insn in current region. */ |
4190 | vec<sel_insn_data_def> s_i_d; |
4191 | |
4192 | /* Extend data structures for insns from current region. */ |
4193 | static void |
4194 | extend_insn_data (void) |
4195 | { |
4196 | int reserve; |
4197 | |
4198 | sched_extend_target (); |
4199 | sched_deps_init (false); |
4200 | |
4201 | /* Extend data structures for insns from current region. */ |
4202 | reserve = (sched_max_luid + 1 - s_i_d.length ()); |
4203 | if (reserve > 0 && ! s_i_d.space (nelems: reserve)) |
4204 | { |
4205 | int size; |
4206 | |
4207 | if (sched_max_luid / 2 > 1024) |
4208 | size = sched_max_luid + 1024; |
4209 | else |
4210 | size = 3 * sched_max_luid / 2; |
4211 | |
4212 | |
4213 | s_i_d.safe_grow_cleared (len: size, exact: true); |
4214 | } |
4215 | } |
4216 | |
4217 | /* Finalize data structures for insns from current region. */ |
4218 | static void |
4219 | finish_insns (void) |
4220 | { |
4221 | unsigned i; |
4222 | |
4223 | /* Clear here all dependence contexts that may have left from insns that were |
4224 | removed during the scheduling. */ |
4225 | for (i = 0; i < s_i_d.length (); i++) |
4226 | { |
4227 | sel_insn_data_def *sid_entry = &s_i_d[i]; |
4228 | |
4229 | if (sid_entry->live) |
4230 | return_regset_to_pool (rs: sid_entry->live); |
4231 | if (sid_entry->analyzed_deps) |
4232 | { |
4233 | BITMAP_FREE (sid_entry->analyzed_deps); |
4234 | BITMAP_FREE (sid_entry->found_deps); |
4235 | htab_delete (sid_entry->transformed_insns); |
4236 | free_deps (&sid_entry->deps_context); |
4237 | } |
4238 | if (EXPR_VINSN (&sid_entry->expr)) |
4239 | { |
4240 | clear_expr (expr: &sid_entry->expr); |
4241 | |
4242 | /* Also, clear CANT_MOVE bit here, because we really don't want it |
4243 | to be passed to the next region. */ |
4244 | CANT_MOVE_BY_LUID (i) = 0; |
4245 | } |
4246 | } |
4247 | |
4248 | s_i_d.release (); |
4249 | } |
4250 | |
4251 | /* A proxy to pass initialization data to init_insn (). */ |
4252 | static sel_insn_data_def _insn_init_ssid; |
4253 | static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; |
4254 | |
4255 | /* If true create a new vinsn. Otherwise use the one from EXPR. */ |
4256 | static bool insn_init_create_new_vinsn_p; |
4257 | |
4258 | /* Set all necessary data for initialization of the new insn[s]. */ |
4259 | static expr_t |
4260 | set_insn_init (expr_t expr, vinsn_t vi, int seqno) |
4261 | { |
4262 | expr_t x = &insn_init_ssid->expr; |
4263 | |
4264 | copy_expr_onside (to: x, from: expr); |
4265 | if (vi != NULL) |
4266 | { |
4267 | insn_init_create_new_vinsn_p = false; |
4268 | change_vinsn_in_expr (x, vi); |
4269 | } |
4270 | else |
4271 | insn_init_create_new_vinsn_p = true; |
4272 | |
4273 | insn_init_ssid->seqno = seqno; |
4274 | return x; |
4275 | } |
4276 | |
4277 | /* Init data for INSN. */ |
4278 | static void |
4279 | init_insn_data (insn_t insn) |
4280 | { |
4281 | expr_t expr; |
4282 | sel_insn_data_t ssid = insn_init_ssid; |
4283 | |
4284 | /* The fields mentioned below are special and hence are not being |
4285 | propagated to the new insns. */ |
4286 | gcc_assert (!ssid->asm_p && ssid->sched_next == NULL |
4287 | && !ssid->after_stall_p && ssid->sched_cycle == 0); |
4288 | gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); |
4289 | |
4290 | expr = INSN_EXPR (insn); |
4291 | copy_expr (to: expr, from: &ssid->expr); |
4292 | prepare_insn_expr (insn, seqno: ssid->seqno); |
4293 | |
4294 | if (insn_init_create_new_vinsn_p) |
4295 | change_vinsn_in_expr (expr, vinsn_create (insn, force_unique_p: init_insn_force_unique_p)); |
4296 | |
4297 | if (first_time_insn_init (insn)) |
4298 | init_first_time_insn_data (insn); |
4299 | } |
4300 | |
4301 | /* This is used to initialize spurious jumps generated by |
4302 | sel_redirect_edge (). OLD_SEQNO is used for initializing seqnos |
4303 | in corner cases within get_seqno_for_a_jump. */ |
4304 | static void |
4305 | init_simplejump_data (insn_t insn, int old_seqno) |
4306 | { |
4307 | init_expr (INSN_EXPR (insn), vi: vinsn_create (insn, force_unique_p: false), spec: 0, |
4308 | REG_BR_PROB_BASE, priority: 0, sched_times: 0, orig_bb_index: 0, spec_done_ds: 0, spec_to_check_ds: 0, orig_sched_cycle: 0, |
4309 | history: vNULL, target_available: true, was_substituted: false, was_renamed: false, |
4310 | needs_spec_check_p: false, cant_move: true); |
4311 | INSN_SEQNO (insn) = get_seqno_for_a_jump (insn, old_seqno); |
4312 | init_first_time_insn_data (insn); |
4313 | } |
4314 | |
4315 | /* Perform deferred initialization of insns. This is used to process |
4316 | a new jump that may be created by redirect_edge. OLD_SEQNO is used |
4317 | for initializing simplejumps in init_simplejump_data. */ |
4318 | static void |
4319 | sel_init_new_insn (insn_t insn, int flags, int old_seqno) |
4320 | { |
4321 | /* We create data structures for bb when the first insn is emitted in it. */ |
4322 | if (INSN_P (insn) |
4323 | && INSN_IN_STREAM_P (insn) |
4324 | && insn_is_the_only_one_in_bb_p (insn)) |
4325 | { |
4326 | extend_bb_info (); |
4327 | create_initial_data_sets (BLOCK_FOR_INSN (insn)); |
4328 | } |
4329 | |
4330 | if (flags & INSN_INIT_TODO_LUID) |
4331 | { |
4332 | sched_extend_luids (); |
4333 | sched_init_insn_luid (insn); |
4334 | } |
4335 | |
4336 | if (flags & INSN_INIT_TODO_SSID) |
4337 | { |
4338 | extend_insn_data (); |
4339 | init_insn_data (insn); |
4340 | clear_expr (expr: &insn_init_ssid->expr); |
4341 | } |
4342 | |
4343 | if (flags & INSN_INIT_TODO_SIMPLEJUMP) |
4344 | { |
4345 | extend_insn_data (); |
4346 | init_simplejump_data (insn, old_seqno); |
4347 | } |
4348 | |
4349 | gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) |
4350 | == CONTAINING_RGN (BB_TO_BLOCK (0))); |
4351 | } |
4352 | |
4353 | |
4354 | /* Functions to init/finish work with lv sets. */ |
4355 | |
4356 | /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ |
4357 | static void |
4358 | init_lv_set (basic_block bb) |
4359 | { |
4360 | gcc_assert (!BB_LV_SET_VALID_P (bb)); |
4361 | |
4362 | BB_LV_SET (bb) = get_regset_from_pool (); |
4363 | COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); |
4364 | BB_LV_SET_VALID_P (bb) = true; |
4365 | } |
4366 | |
4367 | /* Copy liveness information to BB from FROM_BB. */ |
4368 | static void |
4369 | copy_lv_set_from (basic_block bb, basic_block from_bb) |
4370 | { |
4371 | gcc_assert (!BB_LV_SET_VALID_P (bb)); |
4372 | |
4373 | COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); |
4374 | BB_LV_SET_VALID_P (bb) = true; |
4375 | } |
4376 | |
4377 | /* Initialize lv set of all bb headers. */ |
4378 | void |
4379 | init_lv_sets (void) |
4380 | { |
4381 | basic_block bb; |
4382 | |
4383 | /* Initialize of LV sets. */ |
4384 | FOR_EACH_BB_FN (bb, cfun) |
4385 | init_lv_set (bb); |
4386 | |
4387 | /* Don't forget EXIT_BLOCK. */ |
4388 | init_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
4389 | } |
4390 | |
4391 | /* Release lv set of HEAD. */ |
4392 | static void |
4393 | free_lv_set (basic_block bb) |
4394 | { |
4395 | gcc_assert (BB_LV_SET (bb) != NULL); |
4396 | |
4397 | return_regset_to_pool (BB_LV_SET (bb)); |
4398 | BB_LV_SET (bb) = NULL; |
4399 | BB_LV_SET_VALID_P (bb) = false; |
4400 | } |
4401 | |
4402 | /* Finalize lv sets of all bb headers. */ |
4403 | void |
4404 | free_lv_sets (void) |
4405 | { |
4406 | basic_block bb; |
4407 | |
4408 | /* Don't forget EXIT_BLOCK. */ |
4409 | free_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
4410 | |
4411 | /* Free LV sets. */ |
4412 | FOR_EACH_BB_FN (bb, cfun) |
4413 | if (BB_LV_SET (bb)) |
4414 | free_lv_set (bb); |
4415 | } |
4416 | |
4417 | /* Mark AV_SET for BB as invalid, so this set will be updated the next time |
4418 | compute_av() processes BB. This function is called when creating new basic |
4419 | blocks, as well as for blocks (either new or existing) where new jumps are |
4420 | created when the control flow is being updated. */ |
4421 | static void |
4422 | invalidate_av_set (basic_block bb) |
4423 | { |
4424 | BB_AV_LEVEL (bb) = -1; |
4425 | } |
4426 | |
4427 | /* Create initial data sets for BB (they will be invalid). */ |
4428 | static void |
4429 | create_initial_data_sets (basic_block bb) |
4430 | { |
4431 | if (BB_LV_SET (bb)) |
4432 | BB_LV_SET_VALID_P (bb) = false; |
4433 | else |
4434 | BB_LV_SET (bb) = get_regset_from_pool (); |
4435 | invalidate_av_set (bb); |
4436 | } |
4437 | |
4438 | /* Free av set of BB. */ |
4439 | static void |
4440 | free_av_set (basic_block bb) |
4441 | { |
4442 | av_set_clear (setp: &BB_AV_SET (bb)); |
4443 | BB_AV_LEVEL (bb) = 0; |
4444 | } |
4445 | |
4446 | /* Free data sets of BB. */ |
4447 | void |
4448 | free_data_sets (basic_block bb) |
4449 | { |
4450 | free_lv_set (bb); |
4451 | free_av_set (bb); |
4452 | } |
4453 | |
4454 | /* Exchange data sets of TO and FROM. */ |
4455 | void |
4456 | exchange_data_sets (basic_block to, basic_block from) |
4457 | { |
4458 | /* Exchange lv sets of TO and FROM. */ |
4459 | std::swap (BB_LV_SET (from), BB_LV_SET (to)); |
4460 | std::swap (BB_LV_SET_VALID_P (from), BB_LV_SET_VALID_P (to)); |
4461 | |
4462 | /* Exchange av sets of TO and FROM. */ |
4463 | std::swap (BB_AV_SET (from), BB_AV_SET (to)); |
4464 | std::swap (BB_AV_LEVEL (from), BB_AV_LEVEL (to)); |
4465 | } |
4466 | |
4467 | /* Copy data sets of FROM to TO. */ |
4468 | void |
4469 | copy_data_sets (basic_block to, basic_block from) |
4470 | { |
4471 | gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); |
4472 | gcc_assert (BB_AV_SET (to) == NULL); |
4473 | |
4474 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); |
4475 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); |
4476 | |
4477 | if (BB_AV_SET_VALID_P (from)) |
4478 | { |
4479 | BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); |
4480 | } |
4481 | if (BB_LV_SET_VALID_P (from)) |
4482 | { |
4483 | gcc_assert (BB_LV_SET (to) != NULL); |
4484 | COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); |
4485 | } |
4486 | } |
4487 | |
4488 | /* Return an av set for INSN, if any. */ |
4489 | av_set_t |
4490 | get_av_set (insn_t insn) |
4491 | { |
4492 | av_set_t av_set; |
4493 | |
4494 | gcc_assert (AV_SET_VALID_P (insn)); |
4495 | |
4496 | if (sel_bb_head_p (insn)) |
4497 | av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); |
4498 | else |
4499 | av_set = NULL; |
4500 | |
4501 | return av_set; |
4502 | } |
4503 | |
4504 | /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ |
4505 | int |
4506 | get_av_level (insn_t insn) |
4507 | { |
4508 | int av_level; |
4509 | |
4510 | gcc_assert (INSN_P (insn)); |
4511 | |
4512 | if (sel_bb_head_p (insn)) |
4513 | av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); |
4514 | else |
4515 | av_level = INSN_WS_LEVEL (insn); |
4516 | |
4517 | return av_level; |
4518 | } |
4519 | |
4520 | |
4521 | |
4522 | /* Variables to work with control-flow graph. */ |
4523 | |
4524 | /* The basic block that already has been processed by the sched_data_update (), |
4525 | but hasn't been in sel_add_bb () yet. */ |
4526 | static vec<basic_block> last_added_blocks; |
4527 | |
4528 | /* A pool for allocating successor infos. */ |
4529 | static struct |
4530 | { |
4531 | /* A stack for saving succs_info structures. */ |
4532 | struct succs_info *stack; |
4533 | |
4534 | /* Its size. */ |
4535 | int size; |
4536 | |
4537 | /* Top of the stack. */ |
4538 | int top; |
4539 | |
4540 | /* Maximal value of the top. */ |
4541 | int max_top; |
4542 | } succs_info_pool; |
4543 | |
4544 | /* Functions to work with control-flow graph. */ |
4545 | |
4546 | /* Return basic block note of BB. */ |
4547 | rtx_insn * |
4548 | sel_bb_head (basic_block bb) |
4549 | { |
4550 | rtx_insn *head; |
4551 | |
4552 | if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
4553 | { |
4554 | gcc_assert (exit_insn != NULL_RTX); |
4555 | head = exit_insn; |
4556 | } |
4557 | else |
4558 | { |
4559 | rtx_note *note = bb_note (bb); |
4560 | head = next_nonnote_insn (note); |
4561 | |
4562 | if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (insn: head) != bb)) |
4563 | head = NULL; |
4564 | } |
4565 | |
4566 | return head; |
4567 | } |
4568 | |
4569 | /* Return true if INSN is a basic block header. */ |
4570 | bool |
4571 | sel_bb_head_p (insn_t insn) |
4572 | { |
4573 | return sel_bb_head (bb: BLOCK_FOR_INSN (insn)) == insn; |
4574 | } |
4575 | |
4576 | /* Return last insn of BB. */ |
4577 | rtx_insn * |
4578 | sel_bb_end (basic_block bb) |
4579 | { |
4580 | if (sel_bb_empty_p (bb)) |
4581 | return NULL; |
4582 | |
4583 | gcc_assert (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)); |
4584 | |
4585 | return BB_END (bb); |
4586 | } |
4587 | |
4588 | /* Return true if INSN is the last insn in its basic block. */ |
4589 | bool |
4590 | sel_bb_end_p (insn_t insn) |
4591 | { |
4592 | return insn == sel_bb_end (bb: BLOCK_FOR_INSN (insn)); |
4593 | } |
4594 | |
4595 | /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ |
4596 | bool |
4597 | sel_bb_empty_p (basic_block bb) |
4598 | { |
4599 | return sel_bb_head (bb) == NULL; |
4600 | } |
4601 | |
4602 | /* True when BB belongs to the current scheduling region. */ |
4603 | bool |
4604 | in_current_region_p (basic_block bb) |
4605 | { |
4606 | if (bb->index < NUM_FIXED_BLOCKS) |
4607 | return false; |
4608 | |
4609 | return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); |
4610 | } |
4611 | |
4612 | /* Return the block which is a fallthru bb of a conditional jump JUMP. */ |
4613 | basic_block |
4614 | fallthru_bb_of_jump (const rtx_insn *jump) |
4615 | { |
4616 | if (!JUMP_P (jump)) |
4617 | return NULL; |
4618 | |
4619 | if (!any_condjump_p (jump)) |
4620 | return NULL; |
4621 | |
4622 | /* A basic block that ends with a conditional jump may still have one successor |
4623 | (and be followed by a barrier), we are not interested. */ |
4624 | if (single_succ_p (bb: BLOCK_FOR_INSN (insn: jump))) |
4625 | return NULL; |
4626 | |
4627 | return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; |
4628 | } |
4629 | |
4630 | /* Remove all notes from BB. */ |
4631 | static void |
4632 | init_bb (basic_block bb) |
4633 | { |
4634 | remove_notes (bb_note (bb), BB_END (bb)); |
4635 | BB_NOTE_LIST (bb) = note_list; |
4636 | } |
4637 | |
4638 | void |
4639 | sel_init_bbs (bb_vec_t bbs) |
4640 | { |
4641 | const struct sched_scan_info_def ssi = |
4642 | { |
4643 | .extend_bb: extend_bb_info, /* extend_bb */ |
4644 | .init_bb: init_bb, /* init_bb */ |
4645 | NULL, /* extend_insn */ |
4646 | NULL /* init_insn */ |
4647 | }; |
4648 | |
4649 | sched_scan (ssi: &ssi, bbs); |
4650 | } |
4651 | |
4652 | /* Restore notes for the whole region. */ |
4653 | static void |
4654 | sel_restore_notes (void) |
4655 | { |
4656 | int bb; |
4657 | insn_t insn; |
4658 | |
4659 | for (bb = 0; bb < current_nr_blocks; bb++) |
4660 | { |
4661 | basic_block first, last; |
4662 | |
4663 | first = EBB_FIRST_BB (bb); |
4664 | last = EBB_LAST_BB (bb)->next_bb; |
4665 | |
4666 | do |
4667 | { |
4668 | note_list = BB_NOTE_LIST (first); |
4669 | restore_other_notes (NULL, first); |
4670 | BB_NOTE_LIST (first) = NULL; |
4671 | |
4672 | FOR_BB_INSNS (first, insn) |
4673 | if (NONDEBUG_INSN_P (insn)) |
4674 | reemit_notes (insn); |
4675 | |
4676 | first = first->next_bb; |
4677 | } |
4678 | while (first != last); |
4679 | } |
4680 | } |
4681 | |
4682 | /* Free per-bb data structures. */ |
4683 | void |
4684 | sel_finish_bbs (void) |
4685 | { |
4686 | sel_restore_notes (); |
4687 | |
4688 | /* Remove current loop preheader from this loop. */ |
4689 | if (current_loop_nest) |
4690 | sel_remove_loop_preheader (); |
4691 | |
4692 | finish_region_bb_info (); |
4693 | } |
4694 | |
4695 | /* Return true if INSN has a single successor of type FLAGS. */ |
4696 | bool |
4697 | sel_insn_has_single_succ_p (insn_t insn, int flags) |
4698 | { |
4699 | insn_t succ; |
4700 | succ_iterator si; |
4701 | bool first_p = true; |
4702 | |
4703 | FOR_EACH_SUCC_1 (succ, si, insn, flags) |
4704 | { |
4705 | if (first_p) |
4706 | first_p = false; |
4707 | else |
4708 | return false; |
4709 | } |
4710 | |
4711 | return true; |
4712 | } |
4713 | |
4714 | /* Allocate successor's info. */ |
4715 | static struct succs_info * |
4716 | alloc_succs_info (void) |
4717 | { |
4718 | if (succs_info_pool.top == succs_info_pool.max_top) |
4719 | { |
4720 | int i; |
4721 | |
4722 | if (++succs_info_pool.max_top >= succs_info_pool.size) |
4723 | gcc_unreachable (); |
4724 | |
4725 | i = ++succs_info_pool.top; |
4726 | succs_info_pool.stack[i].succs_ok.create (nelems: 10); |
4727 | succs_info_pool.stack[i].succs_other.create (nelems: 10); |
4728 | succs_info_pool.stack[i].probs_ok.create (nelems: 10); |
4729 | } |
4730 | else |
4731 | succs_info_pool.top++; |
4732 | |
4733 | return &succs_info_pool.stack[succs_info_pool.top]; |
4734 | } |
4735 | |
4736 | /* Free successor's info. */ |
4737 | void |
4738 | free_succs_info (struct succs_info * sinfo) |
4739 | { |
4740 | gcc_assert (succs_info_pool.top >= 0 |
4741 | && &succs_info_pool.stack[succs_info_pool.top] == sinfo); |
4742 | succs_info_pool.top--; |
4743 | |
4744 | /* Clear stale info. */ |
4745 | sinfo->succs_ok.block_remove (ix: 0, len: sinfo->succs_ok.length ()); |
4746 | sinfo->succs_other.block_remove (ix: 0, len: sinfo->succs_other.length ()); |
4747 | sinfo->probs_ok.block_remove (ix: 0, len: sinfo->probs_ok.length ()); |
4748 | sinfo->all_prob = 0; |
4749 | sinfo->succs_ok_n = 0; |
4750 | sinfo->all_succs_n = 0; |
4751 | } |
4752 | |
4753 | /* Compute successor info for INSN. FLAGS are the flags passed |
4754 | to the FOR_EACH_SUCC_1 iterator. */ |
4755 | struct succs_info * |
4756 | compute_succs_info (insn_t insn, short flags) |
4757 | { |
4758 | succ_iterator si; |
4759 | insn_t succ; |
4760 | struct succs_info *sinfo = alloc_succs_info (); |
4761 | |
4762 | /* Traverse *all* successors and decide what to do with each. */ |
4763 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) |
4764 | { |
4765 | /* FIXME: this doesn't work for skipping to loop exits, as we don't |
4766 | perform code motion through inner loops. */ |
4767 | short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; |
4768 | |
4769 | if (current_flags & flags) |
4770 | { |
4771 | sinfo->succs_ok.safe_push (obj: succ); |
4772 | sinfo->probs_ok.safe_push ( |
4773 | /* FIXME: Improve calculation when skipping |
4774 | inner loop to exits. */ |
4775 | obj: si.bb_end |
4776 | ? (si.e1->probability.initialized_p () |
4777 | ? si.e1->probability.to_reg_br_prob_base () |
4778 | : 0) |
4779 | : REG_BR_PROB_BASE); |
4780 | sinfo->succs_ok_n++; |
4781 | } |
4782 | else |
4783 | sinfo->succs_other.safe_push (obj: succ); |
4784 | |
4785 | /* Compute all_prob. */ |
4786 | if (!si.bb_end) |
4787 | sinfo->all_prob = REG_BR_PROB_BASE; |
4788 | else if (si.e1->probability.initialized_p ()) |
4789 | sinfo->all_prob += si.e1->probability.to_reg_br_prob_base (); |
4790 | |
4791 | sinfo->all_succs_n++; |
4792 | } |
4793 | |
4794 | return sinfo; |
4795 | } |
4796 | |
4797 | /* Return the predecessors of BB in PREDS and their number in N. |
4798 | Empty blocks are skipped. SIZE is used to allocate PREDS. */ |
4799 | static void |
4800 | cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) |
4801 | { |
4802 | edge e; |
4803 | edge_iterator ei; |
4804 | |
4805 | gcc_assert (BLOCK_TO_BB (bb->index) != 0); |
4806 | |
4807 | FOR_EACH_EDGE (e, ei, bb->preds) |
4808 | { |
4809 | basic_block pred_bb = e->src; |
4810 | insn_t bb_end = BB_END (pred_bb); |
4811 | |
4812 | if (!in_current_region_p (bb: pred_bb)) |
4813 | { |
4814 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
4815 | && current_loop_nest); |
4816 | continue; |
4817 | } |
4818 | |
4819 | if (sel_bb_empty_p (bb: pred_bb)) |
4820 | cfg_preds_1 (bb: pred_bb, preds, n, size); |
4821 | else |
4822 | { |
4823 | if (*n == *size) |
4824 | *preds = XRESIZEVEC (insn_t, *preds, |
4825 | (*size = 2 * *size + 1)); |
4826 | (*preds)[(*n)++] = bb_end; |
4827 | } |
4828 | } |
4829 | |
4830 | gcc_assert (*n != 0 |
4831 | || (flag_sel_sched_pipelining_outer_loops |
4832 | && current_loop_nest)); |
4833 | } |
4834 | |
4835 | /* Find all predecessors of BB and record them in PREDS and their number |
4836 | in N. Empty blocks are skipped, and only normal (forward in-region) |
4837 | edges are processed. */ |
4838 | static void |
4839 | cfg_preds (basic_block bb, insn_t **preds, int *n) |
4840 | { |
4841 | int size = 0; |
4842 | |
4843 | *preds = NULL; |
4844 | *n = 0; |
4845 | cfg_preds_1 (bb, preds, n, size: &size); |
4846 | } |
4847 | |
4848 | /* Returns true if we are moving INSN through join point. */ |
4849 | bool |
4850 | sel_num_cfg_preds_gt_1 (insn_t insn) |
4851 | { |
4852 | basic_block bb; |
4853 | |
4854 | if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) |
4855 | return false; |
4856 | |
4857 | bb = BLOCK_FOR_INSN (insn); |
4858 | |
4859 | while (1) |
4860 | { |
4861 | if (EDGE_COUNT (bb->preds) > 1) |
4862 | return true; |
4863 | |
4864 | gcc_assert (EDGE_PRED (bb, 0)->dest == bb); |
4865 | bb = EDGE_PRED (bb, 0)->src; |
4866 | |
4867 | if (!sel_bb_empty_p (bb)) |
4868 | break; |
4869 | } |
4870 | |
4871 | return false; |
4872 | } |
4873 | |
4874 | /* Returns true when BB should be the end of an ebb. Adapted from the |
4875 | code in sched-ebb.cc. */ |
4876 | bool |
4877 | bb_ends_ebb_p (basic_block bb) |
4878 | { |
4879 | basic_block next_bb = bb_next_bb (bb); |
4880 | edge e; |
4881 | |
4882 | if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
4883 | || bitmap_bit_p (forced_ebb_heads, next_bb->index) |
4884 | || (LABEL_P (BB_HEAD (next_bb)) |
4885 | /* NB: LABEL_NUSES () is not maintained outside of jump.cc. |
4886 | Work around that. */ |
4887 | && !single_pred_p (bb: next_bb))) |
4888 | return true; |
4889 | |
4890 | if (!in_current_region_p (bb: next_bb)) |
4891 | return true; |
4892 | |
4893 | e = find_fallthru_edge (edges: bb->succs); |
4894 | if (e) |
4895 | { |
4896 | gcc_assert (e->dest == next_bb); |
4897 | |
4898 | return false; |
4899 | } |
4900 | |
4901 | return true; |
4902 | } |
4903 | |
4904 | /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a |
4905 | successor of INSN. */ |
4906 | bool |
4907 | in_same_ebb_p (insn_t insn, insn_t succ) |
4908 | { |
4909 | basic_block ptr = BLOCK_FOR_INSN (insn); |
4910 | |
4911 | for (;;) |
4912 | { |
4913 | if (ptr == BLOCK_FOR_INSN (insn: succ)) |
4914 | return true; |
4915 | |
4916 | if (bb_ends_ebb_p (bb: ptr)) |
4917 | return false; |
4918 | |
4919 | ptr = bb_next_bb (bb: ptr); |
4920 | } |
4921 | } |
4922 | |
4923 | /* Recomputes the reverse topological order for the function and |
4924 | saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also |
4925 | modified appropriately. */ |
4926 | static void |
4927 | recompute_rev_top_order (void) |
4928 | { |
4929 | int *postorder; |
4930 | int n_blocks, i; |
4931 | |
4932 | if (!rev_top_order_index |
4933 | || rev_top_order_index_len < last_basic_block_for_fn (cfun)) |
4934 | { |
4935 | rev_top_order_index_len = last_basic_block_for_fn (cfun); |
4936 | rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, |
4937 | rev_top_order_index_len); |
4938 | } |
4939 | |
4940 | postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
4941 | |
4942 | n_blocks = post_order_compute (postorder, true, false); |
4943 | gcc_assert (n_basic_blocks_for_fn (cfun) == n_blocks); |
4944 | |
4945 | /* Build reverse function: for each basic block with BB->INDEX == K |
4946 | rev_top_order_index[K] is it's reverse topological sort number. */ |
4947 | for (i = 0; i < n_blocks; i++) |
4948 | { |
4949 | gcc_assert (postorder[i] < rev_top_order_index_len); |
4950 | rev_top_order_index[postorder[i]] = i; |
4951 | } |
4952 | |
4953 | free (ptr: postorder); |
4954 | } |
4955 | |
4956 | /* Clear all flags from insns in BB that could spoil its rescheduling. */ |
4957 | void |
4958 | clear_outdated_rtx_info (basic_block bb) |
4959 | { |
4960 | rtx_insn *insn; |
4961 | |
4962 | FOR_BB_INSNS (bb, insn) |
4963 | if (INSN_P (insn)) |
4964 | { |
4965 | SCHED_GROUP_P (insn) = 0; |
4966 | INSN_AFTER_STALL_P (insn) = 0; |
4967 | INSN_SCHED_TIMES (insn) = 0; |
4968 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; |
4969 | |
4970 | /* We cannot use the changed caches, as previously we could ignore |
4971 | the LHS dependence due to enabled renaming and transform |
4972 | the expression, and currently we'll be unable to do this. */ |
4973 | htab_empty (INSN_TRANSFORMED_INSNS (insn)); |
4974 | } |
4975 | } |
4976 | |
4977 | /* Add BB_NOTE to the pool of available basic block notes. */ |
4978 | static void |
4979 | return_bb_to_pool (basic_block bb) |
4980 | { |
4981 | rtx_note *note = bb_note (bb); |
4982 | |
4983 | gcc_assert (NOTE_BASIC_BLOCK (note) == bb |
4984 | && bb->aux == NULL); |
4985 | |
4986 | /* It turns out that current cfg infrastructure does not support |
4987 | reuse of basic blocks. Don't bother for now. */ |
4988 | /*bb_note_pool.safe_push (note);*/ |
4989 | } |
4990 | |
4991 | /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ |
4992 | static rtx_note * |
4993 | get_bb_note_from_pool (void) |
4994 | { |
4995 | if (bb_note_pool.is_empty ()) |
4996 | return NULL; |
4997 | else |
4998 | { |
4999 | rtx_note *note = bb_note_pool.pop (); |
5000 | |
5001 | SET_PREV_INSN (note) = NULL_RTX; |
5002 | SET_NEXT_INSN (note) = NULL_RTX; |
5003 | |
5004 | return note; |
5005 | } |
5006 | } |
5007 | |
5008 | /* Free bb_note_pool. */ |
5009 | void |
5010 | free_bb_note_pool (void) |
5011 | { |
5012 | bb_note_pool.release (); |
5013 | } |
5014 | |
5015 | /* Setup scheduler pool and successor structure. */ |
5016 | void |
5017 | alloc_sched_pools (void) |
5018 | { |
5019 | int succs_size; |
5020 | |
5021 | succs_size = MAX_WS + 1; |
5022 | succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); |
5023 | succs_info_pool.size = succs_size; |
5024 | succs_info_pool.top = -1; |
5025 | succs_info_pool.max_top = -1; |
5026 | } |
5027 | |
5028 | /* Free the pools. */ |
5029 | void |
5030 | free_sched_pools (void) |
5031 | { |
5032 | int i; |
5033 | |
5034 | sched_lists_pool.release (); |
5035 | gcc_assert (succs_info_pool.top == -1); |
5036 | for (i = 0; i <= succs_info_pool.max_top; i++) |
5037 | { |
5038 | succs_info_pool.stack[i].succs_ok.release (); |
5039 | succs_info_pool.stack[i].succs_other.release (); |
5040 | succs_info_pool.stack[i].probs_ok.release (); |
5041 | } |
5042 | free (ptr: succs_info_pool.stack); |
5043 | } |
5044 | |
5045 | |
5046 | /* Returns a position in RGN where BB can be inserted retaining |
5047 | topological order. */ |
5048 | static int |
5049 | find_place_to_insert_bb (basic_block bb, int rgn) |
5050 | { |
5051 | bool has_preds_outside_rgn = false; |
5052 | edge e; |
5053 | edge_iterator ei; |
5054 | |
5055 | /* Find whether we have preds outside the region. */ |
5056 | FOR_EACH_EDGE (e, ei, bb->preds) |
5057 | if (!in_current_region_p (bb: e->src)) |
5058 | { |
5059 | has_preds_outside_rgn = true; |
5060 | break; |
5061 | } |
5062 | |
5063 | /* Recompute the top order -- needed when we have > 1 pred |
5064 | and in case we don't have preds outside. */ |
5065 | if (flag_sel_sched_pipelining_outer_loops |
5066 | && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) |
5067 | { |
5068 | int i, bbi = bb->index, cur_bbi; |
5069 | |
5070 | recompute_rev_top_order (); |
5071 | for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) |
5072 | { |
5073 | cur_bbi = BB_TO_BLOCK (i); |
5074 | if (rev_top_order_index[bbi] |
5075 | < rev_top_order_index[cur_bbi]) |
5076 | break; |
5077 | } |
5078 | |
5079 | /* We skipped the right block, so we increase i. We accommodate |
5080 | it for increasing by step later, so we decrease i. */ |
5081 | return (i + 1) - 1; |
5082 | } |
5083 | else if (has_preds_outside_rgn) |
5084 | { |
5085 | /* This is the case when we generate an extra empty block |
5086 | to serve as region head during pipelining. */ |
5087 | e = EDGE_SUCC (bb, 0); |
5088 | gcc_assert (EDGE_COUNT (bb->succs) == 1 |
5089 | && in_current_region_p (EDGE_SUCC (bb, 0)->dest) |
5090 | && (BLOCK_TO_BB (e->dest->index) == 0)); |
5091 | return -1; |
5092 | } |
5093 | |
5094 | /* We don't have preds outside the region. We should have |
5095 | the only pred, because the multiple preds case comes from |
5096 | the pipelining of outer loops, and that is handled above. |
5097 | Just take the bbi of this single pred. */ |
5098 | if (EDGE_COUNT (bb->succs) > 0) |
5099 | { |
5100 | int pred_bbi; |
5101 | |
5102 | gcc_assert (EDGE_COUNT (bb->preds) == 1); |
5103 | |
5104 | pred_bbi = EDGE_PRED (bb, 0)->src->index; |
5105 | return BLOCK_TO_BB (pred_bbi); |
5106 | } |
5107 | else |
5108 | /* BB has no successors. It is safe to put it in the end. */ |
5109 | return current_nr_blocks - 1; |
5110 | } |
5111 | |
5112 | /* Deletes an empty basic block freeing its data. */ |
5113 | static void |
5114 | delete_and_free_basic_block (basic_block bb) |
5115 | { |
5116 | gcc_assert (sel_bb_empty_p (bb)); |
5117 | |
5118 | if (BB_LV_SET (bb)) |
5119 | free_lv_set (bb); |
5120 | |
5121 | bitmap_clear_bit (blocks_to_reschedule, bb->index); |
5122 | |
5123 | /* Can't assert av_set properties because we use sel_aremove_bb |
5124 | when removing loop preheader from the region. At the point of |
5125 | removing the preheader we already have deallocated sel_region_bb_info. */ |
5126 | gcc_assert (BB_LV_SET (bb) == NULL |
5127 | && !BB_LV_SET_VALID_P (bb) |
5128 | && BB_AV_LEVEL (bb) == 0 |
5129 | && BB_AV_SET (bb) == NULL); |
5130 | |
5131 | delete_basic_block (bb); |
5132 | } |
5133 | |
5134 | /* Add BB to the current region and update the region data. */ |
5135 | static void |
5136 | add_block_to_current_region (basic_block bb) |
5137 | { |
5138 | int i, pos, bbi = -2, rgn; |
5139 | |
5140 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); |
5141 | bbi = find_place_to_insert_bb (bb, rgn); |
5142 | bbi += 1; |
5143 | pos = RGN_BLOCKS (rgn) + bbi; |
5144 | |
5145 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 |
5146 | && ebb_head[bbi] == pos); |
5147 | |
5148 | /* Make a place for the new block. */ |
5149 | extend_regions (); |
5150 | |
5151 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) |
5152 | BLOCK_TO_BB (rgn_bb_table[i])++; |
5153 | |
5154 | memmove (dest: rgn_bb_table + pos + 1, |
5155 | src: rgn_bb_table + pos, |
5156 | n: (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); |
5157 | |
5158 | /* Initialize data for BB. */ |
5159 | rgn_bb_table[pos] = bb->index; |
5160 | BLOCK_TO_BB (bb->index) = bbi; |
5161 | CONTAINING_RGN (bb->index) = rgn; |
5162 | |
5163 | RGN_NR_BLOCKS (rgn)++; |
5164 | |
5165 | for (i = rgn + 1; i <= nr_regions; i++) |
5166 | RGN_BLOCKS (i)++; |
5167 | } |
5168 | |
5169 | /* Remove BB from the current region and update the region data. */ |
5170 | static void |
5171 | remove_bb_from_region (basic_block bb) |
5172 | { |
5173 | int i, pos, bbi = -2, rgn; |
5174 | |
5175 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); |
5176 | bbi = BLOCK_TO_BB (bb->index); |
5177 | pos = RGN_BLOCKS (rgn) + bbi; |
5178 | |
5179 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 |
5180 | && ebb_head[bbi] == pos); |
5181 | |
5182 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) |
5183 | BLOCK_TO_BB (rgn_bb_table[i])--; |
5184 | |
5185 | memmove (dest: rgn_bb_table + pos, |
5186 | src: rgn_bb_table + pos + 1, |
5187 | n: (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); |
5188 | |
5189 | RGN_NR_BLOCKS (rgn)--; |
5190 | for (i = rgn + 1; i <= nr_regions; i++) |
5191 | RGN_BLOCKS (i)--; |
5192 | } |
5193 | |
5194 | /* Add BB to the current region and update all data. If BB is NULL, add all |
5195 | blocks from last_added_blocks vector. */ |
5196 | static void |
5197 | sel_add_bb (basic_block bb) |
5198 | { |
5199 | /* Extend luids so that new notes will receive zero luids. */ |
5200 | sched_extend_luids (); |
5201 | sched_init_bbs (); |
5202 | sel_init_bbs (bbs: last_added_blocks); |
5203 | |
5204 | /* When bb is passed explicitly, the vector should contain |
5205 | the only element that equals to bb; otherwise, the vector |
5206 | should not be NULL. */ |
5207 | gcc_assert (last_added_blocks.exists ()); |
5208 | |
5209 | if (bb != NULL) |
5210 | { |
5211 | gcc_assert (last_added_blocks.length () == 1 |
5212 | && last_added_blocks[0] == bb); |
5213 | add_block_to_current_region (bb); |
5214 | |
5215 | /* We associate creating/deleting data sets with the first insn |
5216 | appearing / disappearing in the bb. */ |
5217 | if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) |
5218 | create_initial_data_sets (bb); |
5219 | |
5220 | last_added_blocks.release (); |
5221 | } |
5222 | else |
5223 | /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ |
5224 | { |
5225 | int i; |
5226 | basic_block temp_bb = NULL; |
5227 | |
5228 | for (i = 0; |
5229 | last_added_blocks.iterate (ix: i, ptr: &bb); i++) |
5230 | { |
5231 | add_block_to_current_region (bb); |
5232 | temp_bb = bb; |
5233 | } |
5234 | |
5235 | /* We need to fetch at least one bb so we know the region |
5236 | to update. */ |
5237 | gcc_assert (temp_bb != NULL); |
5238 | bb = temp_bb; |
5239 | |
5240 | last_added_blocks.release (); |
5241 | } |
5242 | |
5243 | rgn_setup_region (CONTAINING_RGN (bb->index)); |
5244 | } |
5245 | |
5246 | /* Remove BB from the current region and update all data. |
5247 | If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ |
5248 | static void |
5249 | sel_remove_bb (basic_block bb, bool remove_from_cfg_p) |
5250 | { |
5251 | unsigned idx = bb->index; |
5252 | |
5253 | gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); |
5254 | |
5255 | remove_bb_from_region (bb); |
5256 | return_bb_to_pool (bb); |
5257 | bitmap_clear_bit (blocks_to_reschedule, idx); |
5258 | |
5259 | if (remove_from_cfg_p) |
5260 | { |
5261 | basic_block succ = single_succ (bb); |
5262 | delete_and_free_basic_block (bb); |
5263 | set_immediate_dominator (CDI_DOMINATORS, succ, |
5264 | recompute_dominator (CDI_DOMINATORS, succ)); |
5265 | } |
5266 | |
5267 | rgn_setup_region (CONTAINING_RGN (idx)); |
5268 | } |
5269 | |
5270 | /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ |
5271 | static void |
5272 | move_bb_info (basic_block merge_bb, basic_block empty_bb) |
5273 | { |
5274 | if (in_current_region_p (bb: merge_bb)) |
5275 | concat_note_lists (BB_NOTE_LIST (empty_bb), |
5276 | &BB_NOTE_LIST (merge_bb)); |
5277 | BB_NOTE_LIST (empty_bb) = NULL; |
5278 | |
5279 | } |
5280 | |
5281 | /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from |
5282 | region, but keep it in CFG. */ |
5283 | static void |
5284 | remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) |
5285 | { |
5286 | /* The block should contain just a note or a label. |
5287 | We try to check whether it is unused below. */ |
5288 | gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) |
5289 | || LABEL_P (BB_HEAD (empty_bb))); |
5290 | |
5291 | /* If basic block has predecessors or successors, redirect them. */ |
5292 | if (remove_from_cfg_p |
5293 | && (EDGE_COUNT (empty_bb->preds) > 0 |
5294 | || EDGE_COUNT (empty_bb->succs) > 0)) |
5295 | { |
5296 | basic_block pred; |
5297 | basic_block succ; |
5298 | |
5299 | /* We need to init PRED and SUCC before redirecting edges. */ |
5300 | if (EDGE_COUNT (empty_bb->preds) > 0) |
5301 | { |
5302 | edge e; |
5303 | |
5304 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); |
5305 | |
5306 | e = EDGE_PRED (empty_bb, 0); |
5307 | gcc_assert (e->src == empty_bb->prev_bb |
5308 | && (e->flags & EDGE_FALLTHRU)); |
5309 | |
5310 | pred = empty_bb->prev_bb; |
5311 | } |
5312 | else |
5313 | pred = NULL; |
5314 | |
5315 | if (EDGE_COUNT (empty_bb->succs) > 0) |
5316 | { |
5317 | /* We do not check fallthruness here as above, because |
5318 | after removing a jump the edge may actually be not fallthru. */ |
5319 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); |
5320 | succ = EDGE_SUCC (empty_bb, 0)->dest; |
5321 | } |
5322 | else |
5323 | succ = NULL; |
5324 | |
5325 | if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) |
5326 | { |
5327 | edge e = EDGE_PRED (empty_bb, 0); |
5328 | |
5329 | if (e->flags & EDGE_FALLTHRU) |
5330 | redirect_edge_succ_nodup (e, succ); |
5331 | else |
5332 | sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); |
5333 | } |
5334 | |
5335 | if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) |
5336 | { |
5337 | edge e = EDGE_SUCC (empty_bb, 0); |
5338 | |
5339 | if (find_edge (pred, e->dest) == NULL) |
5340 | redirect_edge_pred (e, pred); |
5341 | } |
5342 | } |
5343 | |
5344 | /* Finish removing. */ |
5345 | sel_remove_bb (bb: empty_bb, remove_from_cfg_p); |
5346 | } |
5347 | |
5348 | /* An implementation of create_basic_block hook, which additionally updates |
5349 | per-bb data structures. */ |
5350 | static basic_block |
5351 | sel_create_basic_block (void *headp, void *endp, basic_block after) |
5352 | { |
5353 | basic_block new_bb; |
5354 | rtx_note *new_bb_note; |
5355 | |
5356 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
5357 | || !last_added_blocks.exists ()); |
5358 | |
5359 | new_bb_note = get_bb_note_from_pool (); |
5360 | |
5361 | if (new_bb_note == NULL_RTX) |
5362 | new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); |
5363 | else |
5364 | { |
5365 | new_bb = create_basic_block_structure ((rtx_insn *) headp, |
5366 | (rtx_insn *) endp, |
5367 | new_bb_note, after); |
5368 | new_bb->aux = NULL; |
5369 | } |
5370 | |
5371 | last_added_blocks.safe_push (obj: new_bb); |
5372 | |
5373 | return new_bb; |
5374 | } |
5375 | |
5376 | /* Implement sched_init_only_bb (). */ |
5377 | static void |
5378 | sel_init_only_bb (basic_block bb, basic_block after) |
5379 | { |
5380 | gcc_assert (after == NULL); |
5381 | |
5382 | extend_regions (); |
5383 | rgn_make_new_region_out_of_new_block (bb); |
5384 | } |
5385 | |
5386 | /* Update the latch when we've splitted or merged it from FROM block to TO. |
5387 | This should be checked for all outer loops, too. */ |
5388 | static void |
5389 | change_loops_latches (basic_block from, basic_block to) |
5390 | { |
5391 | gcc_assert (from != to); |
5392 | |
5393 | if (current_loop_nest) |
5394 | { |
5395 | class loop *loop; |
5396 | |
5397 | for (loop = current_loop_nest; loop; loop = loop_outer (loop)) |
5398 | if (considered_for_pipelining_p (loop) && loop->latch == from) |
5399 | { |
5400 | gcc_assert (loop == current_loop_nest); |
5401 | loop->latch = to; |
5402 | gcc_assert (loop_latch_edge (loop)); |
5403 | } |
5404 | } |
5405 | } |
5406 | |
5407 | /* Splits BB on two basic blocks, adding it to the region and extending |
5408 | per-bb data structures. Returns the newly created bb. */ |
5409 | static basic_block |
5410 | sel_split_block (basic_block bb, rtx after) |
5411 | { |
5412 | basic_block new_bb; |
5413 | insn_t insn; |
5414 | |
5415 | new_bb = sched_split_block_1 (bb, after); |
5416 | sel_add_bb (bb: new_bb); |
5417 | |
5418 | /* This should be called after sel_add_bb, because this uses |
5419 | CONTAINING_RGN for the new block, which is not yet initialized. |
5420 | FIXME: this function may be a no-op now. */ |
5421 | change_loops_latches (from: bb, to: new_bb); |
5422 | |
5423 | /* Update ORIG_BB_INDEX for insns moved into the new block. */ |
5424 | FOR_BB_INSNS (new_bb, insn) |
5425 | if (INSN_P (insn)) |
5426 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; |
5427 | |
5428 | if (sel_bb_empty_p (bb)) |
5429 | { |
5430 | gcc_assert (!sel_bb_empty_p (new_bb)); |
5431 | |
5432 | /* NEW_BB has data sets that need to be updated and BB holds |
5433 | data sets that should be removed. Exchange these data sets |
5434 | so that we won't lose BB's valid data sets. */ |
5435 | exchange_data_sets (to: new_bb, from: bb); |
5436 | free_data_sets (bb); |
5437 | } |
5438 | |
5439 | if (!sel_bb_empty_p (bb: new_bb) |
5440 | && bitmap_bit_p (blocks_to_reschedule, bb->index)) |
5441 | bitmap_set_bit (blocks_to_reschedule, new_bb->index); |
5442 | |
5443 | return new_bb; |
5444 | } |
5445 | |
5446 | /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. |
5447 | Otherwise returns NULL. */ |
5448 | static rtx_insn * |
5449 | check_for_new_jump (basic_block bb, int prev_max_uid) |
5450 | { |
5451 | rtx_insn *end; |
5452 | |
5453 | end = sel_bb_end (bb); |
5454 | if (end && INSN_UID (insn: end) >= prev_max_uid) |
5455 | return end; |
5456 | return NULL; |
5457 | } |
5458 | |
5459 | /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. |
5460 | New means having UID at least equal to PREV_MAX_UID. */ |
5461 | static rtx_insn * |
5462 | find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) |
5463 | { |
5464 | rtx_insn *jump; |
5465 | |
5466 | /* Return immediately if no new insns were emitted. */ |
5467 | if (get_max_uid () == prev_max_uid) |
5468 | return NULL; |
5469 | |
5470 | /* Now check both blocks for new jumps. It will ever be only one. */ |
5471 | if ((jump = check_for_new_jump (bb: from, prev_max_uid))) |
5472 | return jump; |
5473 | |
5474 | if (jump_bb != NULL |
5475 | && (jump = check_for_new_jump (bb: jump_bb, prev_max_uid))) |
5476 | return jump; |
5477 | return NULL; |
5478 | } |
5479 | |
5480 | /* Splits E and adds the newly created basic block to the current region. |
5481 | Returns this basic block. */ |
5482 | basic_block |
5483 | sel_split_edge (edge e) |
5484 | { |
5485 | basic_block new_bb, src, other_bb = NULL; |
5486 | int prev_max_uid; |
5487 | rtx_insn *jump; |
5488 | |
5489 | src = e->src; |
5490 | prev_max_uid = get_max_uid (); |
5491 | new_bb = split_edge (e); |
5492 | |
5493 | if (flag_sel_sched_pipelining_outer_loops |
5494 | && current_loop_nest) |
5495 | { |
5496 | int i; |
5497 | basic_block bb; |
5498 | |
5499 | /* Some of the basic blocks might not have been added to the loop. |
5500 | Add them here, until this is fixed in force_fallthru. */ |
5501 | for (i = 0; |
5502 | last_added_blocks.iterate (ix: i, ptr: &bb); i++) |
5503 | if (!bb->loop_father) |
5504 | { |
5505 | add_bb_to_loop (bb, e->dest->loop_father); |
5506 | |
5507 | gcc_assert (!other_bb && (new_bb->index != bb->index)); |
5508 | other_bb = bb; |
5509 | } |
5510 | } |
5511 | |
5512 | /* Add all last_added_blocks to the region. */ |
5513 | sel_add_bb (NULL); |
5514 | |
5515 | jump = find_new_jump (from: src, jump_bb: new_bb, prev_max_uid); |
5516 | if (jump) |
5517 | sel_init_new_insn (insn: jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); |
5518 | |
5519 | /* Put the correct lv set on this block. */ |
5520 | if (other_bb && !sel_bb_empty_p (bb: other_bb)) |
5521 | compute_live (sel_bb_head (bb: other_bb)); |
5522 | |
5523 | return new_bb; |
5524 | } |
5525 | |
5526 | /* Implement sched_create_empty_bb (). */ |
5527 | static basic_block |
5528 | sel_create_empty_bb (basic_block after) |
5529 | { |
5530 | basic_block new_bb; |
5531 | |
5532 | new_bb = sched_create_empty_bb_1 (after); |
5533 | |
5534 | /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit |
5535 | later. */ |
5536 | gcc_assert (last_added_blocks.length () == 1 |
5537 | && last_added_blocks[0] == new_bb); |
5538 | |
5539 | last_added_blocks.release (); |
5540 | return new_bb; |
5541 | } |
5542 | |
5543 | /* Implement sched_create_recovery_block. ORIG_INSN is where block |
5544 | will be splitted to insert a check. */ |
5545 | basic_block |
5546 | sel_create_recovery_block (insn_t orig_insn) |
5547 | { |
5548 | basic_block first_bb, second_bb, recovery_block; |
5549 | basic_block before_recovery = NULL; |
5550 | rtx_insn *jump; |
5551 | |
5552 | first_bb = BLOCK_FOR_INSN (insn: orig_insn); |
5553 | if (sel_bb_end_p (insn: orig_insn)) |
5554 | { |
5555 | /* Avoid introducing an empty block while splitting. */ |
5556 | gcc_assert (single_succ_p (first_bb)); |
5557 | second_bb = single_succ (bb: first_bb); |
5558 | } |
5559 | else |
5560 | second_bb = sched_split_block (first_bb, orig_insn); |
5561 | |
5562 | recovery_block = sched_create_recovery_block (&before_recovery); |
5563 | if (before_recovery) |
5564 | copy_lv_set_from (bb: before_recovery, EXIT_BLOCK_PTR_FOR_FN (cfun)); |
5565 | |
5566 | gcc_assert (sel_bb_empty_p (recovery_block)); |
5567 | sched_create_recovery_edges (first_bb, recovery_block, second_bb); |
5568 | if (current_loops != NULL) |
5569 | add_bb_to_loop (recovery_block, first_bb->loop_father); |
5570 | |
5571 | sel_add_bb (bb: recovery_block); |
5572 | |
5573 | jump = BB_END (recovery_block); |
5574 | gcc_assert (sel_bb_head (recovery_block) == jump); |
5575 | sel_init_new_insn (insn: jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); |
5576 | |
5577 | return recovery_block; |
5578 | } |
5579 | |
5580 | /* Merge basic block B into basic block A. */ |
5581 | static void |
5582 | sel_merge_blocks (basic_block a, basic_block b) |
5583 | { |
5584 | gcc_assert (sel_bb_empty_p (b) |
5585 | && EDGE_COUNT (b->preds) == 1 |
5586 | && EDGE_PRED (b, 0)->src == b->prev_bb); |
5587 | |
5588 | move_bb_info (merge_bb: b->prev_bb, empty_bb: b); |
5589 | remove_empty_bb (empty_bb: b, remove_from_cfg_p: false); |
5590 | merge_blocks (a, b); |
5591 | change_loops_latches (from: b, to: a); |
5592 | } |
5593 | |
5594 | /* A wrapper for redirect_edge_and_branch_force, which also initializes |
5595 | data structures for possibly created bb and insns. */ |
5596 | void |
5597 | sel_redirect_edge_and_branch_force (edge e, basic_block to) |
5598 | { |
5599 | basic_block jump_bb, src, orig_dest = e->dest; |
5600 | int prev_max_uid; |
5601 | rtx_insn *jump; |
5602 | int old_seqno = -1; |
5603 | |
5604 | /* This function is now used only for bookkeeping code creation, where |
5605 | we'll never get the single pred of orig_dest block and thus will not |
5606 | hit unreachable blocks when updating dominator info. */ |
5607 | gcc_assert (!sel_bb_empty_p (e->src) |
5608 | && !single_pred_p (orig_dest)); |
5609 | src = e->src; |
5610 | prev_max_uid = get_max_uid (); |
5611 | /* Compute and pass old_seqno down to sel_init_new_insn only for the case |
5612 | when the conditional jump being redirected may become unconditional. */ |
5613 | if (any_condjump_p (BB_END (src)) |
5614 | && INSN_SEQNO (BB_END (src)) >= 0) |
5615 | old_seqno = INSN_SEQNO (BB_END (src)); |
5616 | |
5617 | jump_bb = redirect_edge_and_branch_force (e, to); |
5618 | if (jump_bb != NULL) |
5619 | sel_add_bb (bb: jump_bb); |
5620 | |
5621 | /* This function could not be used to spoil the loop structure by now, |
5622 | thus we don't care to update anything. But check it to be sure. */ |
5623 | if (current_loop_nest |
5624 | && pipelining_p) |
5625 | gcc_assert (loop_latch_edge (current_loop_nest)); |
5626 | |
5627 | jump = find_new_jump (from: src, jump_bb, prev_max_uid); |
5628 | if (jump) |
5629 | sel_init_new_insn (insn: jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, |
5630 | old_seqno); |
5631 | set_immediate_dominator (CDI_DOMINATORS, to, |
5632 | recompute_dominator (CDI_DOMINATORS, to)); |
5633 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, |
5634 | recompute_dominator (CDI_DOMINATORS, orig_dest)); |
5635 | if (jump && sel_bb_head_p (insn: jump)) |
5636 | compute_live (jump); |
5637 | } |
5638 | |
5639 | /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by |
5640 | redirected edge are in reverse topological order. */ |
5641 | bool |
5642 | sel_redirect_edge_and_branch (edge e, basic_block to) |
5643 | { |
5644 | bool latch_edge_p; |
5645 | basic_block src, orig_dest = e->dest; |
5646 | int prev_max_uid; |
5647 | rtx_insn *jump; |
5648 | edge redirected; |
5649 | bool recompute_toporder_p = false; |
5650 | bool maybe_unreachable = single_pred_p (bb: orig_dest); |
5651 | int old_seqno = -1; |
5652 | |
5653 | latch_edge_p = (pipelining_p |
5654 | && current_loop_nest |
5655 | && e == loop_latch_edge (current_loop_nest)); |
5656 | |
5657 | src = e->src; |
5658 | prev_max_uid = get_max_uid (); |
5659 | |
5660 | /* Compute and pass old_seqno down to sel_init_new_insn only for the case |
5661 | when the conditional jump being redirected may become unconditional. */ |
5662 | if (any_condjump_p (BB_END (src)) |
5663 | && INSN_SEQNO (BB_END (src)) >= 0) |
5664 | old_seqno = INSN_SEQNO (BB_END (src)); |
5665 | |
5666 | redirected = redirect_edge_and_branch (e, to); |
5667 | |
5668 | gcc_assert (redirected && !last_added_blocks.exists ()); |
5669 | |
5670 | /* When we've redirected a latch edge, update the header. */ |
5671 | if (latch_edge_p) |
5672 | { |
5673 | current_loop_nest->header = to; |
5674 | gcc_assert (loop_latch_edge (current_loop_nest)); |
5675 | } |
5676 | |
5677 | /* In rare situations, the topological relation between the blocks connected |
5678 | by the redirected edge can change (see PR42245 for an example). Update |
5679 | block_to_bb/bb_to_block. */ |
5680 | if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) |
5681 | && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) |
5682 | recompute_toporder_p = true; |
5683 | |
5684 | jump = find_new_jump (from: src, NULL, prev_max_uid); |
5685 | if (jump) |
5686 | sel_init_new_insn (insn: jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, old_seqno); |
5687 | |
5688 | /* Only update dominator info when we don't have unreachable blocks. |
5689 | Otherwise we'll update in maybe_tidy_empty_bb. */ |
5690 | if (!maybe_unreachable) |
5691 | { |
5692 | set_immediate_dominator (CDI_DOMINATORS, to, |
5693 | recompute_dominator (CDI_DOMINATORS, to)); |
5694 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, |
5695 | recompute_dominator (CDI_DOMINATORS, orig_dest)); |
5696 | } |
5697 | if (jump && sel_bb_head_p (insn: jump)) |
5698 | compute_live (jump); |
5699 | return recompute_toporder_p; |
5700 | } |
5701 | |
5702 | /* This variable holds the cfg hooks used by the selective scheduler. */ |
5703 | static struct cfg_hooks sel_cfg_hooks; |
5704 | |
5705 | /* Register sel-sched cfg hooks. */ |
5706 | void |
5707 | sel_register_cfg_hooks (void) |
5708 | { |
5709 | sched_split_block = sel_split_block; |
5710 | |
5711 | orig_cfg_hooks = get_cfg_hooks (); |
5712 | sel_cfg_hooks = orig_cfg_hooks; |
5713 | |
5714 | sel_cfg_hooks.create_basic_block = sel_create_basic_block; |
5715 | |
5716 | set_cfg_hooks (sel_cfg_hooks); |
5717 | |
5718 | sched_init_only_bb = sel_init_only_bb; |
5719 | sched_split_block = sel_split_block; |
5720 | sched_create_empty_bb = sel_create_empty_bb; |
5721 | } |
5722 | |
5723 | /* Unregister sel-sched cfg hooks. */ |
5724 | void |
5725 | sel_unregister_cfg_hooks (void) |
5726 | { |
5727 | sched_create_empty_bb = NULL; |
5728 | sched_split_block = NULL; |
5729 | sched_init_only_bb = NULL; |
5730 | |
5731 | set_cfg_hooks (orig_cfg_hooks); |
5732 | } |
5733 | |
5734 | |
5735 | /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, |
5736 | LABEL is where this jump should be directed. */ |
5737 | rtx_insn * |
5738 | create_insn_rtx_from_pattern (rtx pattern, rtx label) |
5739 | { |
5740 | rtx_insn *insn_rtx; |
5741 | |
5742 | gcc_assert (!INSN_P (pattern)); |
5743 | |
5744 | start_sequence (); |
5745 | |
5746 | if (label == NULL_RTX) |
5747 | insn_rtx = emit_insn (pattern); |
5748 | else if (DEBUG_INSN_P (label)) |
5749 | insn_rtx = emit_debug_insn (pattern); |
5750 | else |
5751 | { |
5752 | insn_rtx = emit_jump_insn (pattern); |
5753 | JUMP_LABEL (insn_rtx) = label; |
5754 | ++LABEL_NUSES (label); |
5755 | } |
5756 | |
5757 | end_sequence (); |
5758 | |
5759 | sched_extend_luids (); |
5760 | sched_extend_target (); |
5761 | sched_deps_init (false); |
5762 | |
5763 | /* Initialize INSN_CODE now. */ |
5764 | recog_memoized (insn: insn_rtx); |
5765 | return insn_rtx; |
5766 | } |
5767 | |
5768 | /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn |
5769 | must not be clonable. */ |
5770 | vinsn_t |
5771 | create_vinsn_from_insn_rtx (rtx_insn *insn_rtx, bool force_unique_p) |
5772 | { |
5773 | gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); |
5774 | |
5775 | /* If VINSN_TYPE is not USE, retain its uniqueness. */ |
5776 | return vinsn_create (insn: insn_rtx, force_unique_p); |
5777 | } |
5778 | |
5779 | /* Create a copy of INSN_RTX. */ |
5780 | rtx_insn * |
5781 | create_copy_of_insn_rtx (rtx insn_rtx) |
5782 | { |
5783 | rtx_insn *res; |
5784 | rtx link; |
5785 | |
5786 | if (DEBUG_INSN_P (insn_rtx)) |
5787 | return create_insn_rtx_from_pattern (pattern: copy_rtx (PATTERN (insn: insn_rtx)), |
5788 | label: insn_rtx); |
5789 | |
5790 | gcc_assert (NONJUMP_INSN_P (insn_rtx)); |
5791 | |
5792 | res = create_insn_rtx_from_pattern (pattern: copy_rtx (PATTERN (insn: insn_rtx)), |
5793 | NULL_RTX); |
5794 | |
5795 | /* Locate the end of existing REG_NOTES in NEW_RTX. */ |
5796 | rtx *ptail = ®_NOTES (res); |
5797 | while (*ptail != NULL_RTX) |
5798 | ptail = &XEXP (*ptail, 1); |
5799 | |
5800 | /* Copy all REG_NOTES except REG_EQUAL/REG_EQUIV and REG_LABEL_OPERAND |
5801 | since mark_jump_label will make them. REG_LABEL_TARGETs are created |
5802 | there too, but are supposed to be sticky, so we copy them. */ |
5803 | for (link = REG_NOTES (insn_rtx); link; link = XEXP (link, 1)) |
5804 | if (REG_NOTE_KIND (link) != REG_LABEL_OPERAND |
5805 | && REG_NOTE_KIND (link) != REG_EQUAL |
5806 | && REG_NOTE_KIND (link) != REG_EQUIV) |
5807 | { |
5808 | *ptail = duplicate_reg_note (link); |
5809 | ptail = &XEXP (*ptail, 1); |
5810 | } |
5811 | |
5812 | return res; |
5813 | } |
5814 | |
5815 | /* Change vinsn field of EXPR to hold NEW_VINSN. */ |
5816 | void |
5817 | change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) |
5818 | { |
5819 | vinsn_detach (EXPR_VINSN (expr)); |
5820 | |
5821 | EXPR_VINSN (expr) = new_vinsn; |
5822 | vinsn_attach (vi: new_vinsn); |
5823 | } |
5824 | |
5825 | /* Helpers for global init. */ |
5826 | /* This structure is used to be able to call existing bundling mechanism |
5827 | and calculate insn priorities. */ |
5828 | static struct haifa_sched_info sched_sel_haifa_sched_info = |
5829 | { |
5830 | NULL, /* init_ready_list */ |
5831 | NULL, /* can_schedule_ready_p */ |
5832 | NULL, /* schedule_more_p */ |
5833 | NULL, /* new_ready */ |
5834 | NULL, /* rgn_rank */ |
5835 | .print_insn: sel_print_insn, /* rgn_print_insn */ |
5836 | .contributes_to_priority: contributes_to_priority, |
5837 | NULL, /* insn_finishes_block_p */ |
5838 | |
5839 | NULL, NULL, |
5840 | NULL, NULL, |
5841 | .queue_must_finish_empty: 0, .sched_max_insns_priority: 0, |
5842 | |
5843 | NULL, /* add_remove_insn */ |
5844 | NULL, /* begin_schedule_ready */ |
5845 | NULL, /* begin_move_insn */ |
5846 | NULL, /* advance_target_bb */ |
5847 | |
5848 | NULL, |
5849 | NULL, |
5850 | |
5851 | .flags: SEL_SCHED | NEW_BBS |
5852 | }; |
5853 | |
5854 | /* Setup special insns used in the scheduler. */ |
5855 | void |
5856 | setup_nop_and_exit_insns (void) |
5857 | { |
5858 | gcc_assert (nop_pattern == NULL_RTX |
5859 | && exit_insn == NULL_RTX); |
5860 | |
5861 | nop_pattern = constm1_rtx; |
5862 | |
5863 | start_sequence (); |
5864 | emit_insn (nop_pattern); |
5865 | exit_insn = get_insns (); |
5866 | end_sequence (); |
5867 | set_block_for_insn (insn: exit_insn, EXIT_BLOCK_PTR_FOR_FN (cfun)); |
5868 | } |
5869 | |
5870 | /* Free special insns used in the scheduler. */ |
5871 | void |
5872 | free_nop_and_exit_insns (void) |
5873 | { |
5874 | exit_insn = NULL; |
5875 | nop_pattern = NULL_RTX; |
5876 | } |
5877 | |
5878 | /* Setup a special vinsn used in new insns initialization. */ |
5879 | void |
5880 | setup_nop_vinsn (void) |
5881 | { |
5882 | nop_vinsn = vinsn_create (insn: exit_insn, force_unique_p: false); |
5883 | vinsn_attach (vi: nop_vinsn); |
5884 | } |
5885 | |
5886 | /* Free a special vinsn used in new insns initialization. */ |
5887 | void |
5888 | free_nop_vinsn (void) |
5889 | { |
5890 | gcc_assert (VINSN_COUNT (nop_vinsn) == 1); |
5891 | vinsn_detach (vi: nop_vinsn); |
5892 | nop_vinsn = NULL; |
5893 | } |
5894 | |
5895 | /* Call a set_sched_flags hook. */ |
5896 | void |
5897 | sel_set_sched_flags (void) |
5898 | { |
5899 | /* ??? This means that set_sched_flags were called, and we decided to |
5900 | support speculation. However, set_sched_flags also modifies flags |
5901 | on current_sched_info, doing this only at global init. And we |
5902 | sometimes change c_s_i later. So put the correct flags again. */ |
5903 | if (spec_info && targetm.sched.set_sched_flags) |
5904 | targetm.sched.set_sched_flags (spec_info); |
5905 | } |
5906 | |
5907 | /* Setup pointers to global sched info structures. */ |
5908 | void |
5909 | sel_setup_sched_infos (void) |
5910 | { |
5911 | rgn_setup_common_sched_info (); |
5912 | |
5913 | memcpy (dest: &sel_common_sched_info, src: common_sched_info, |
5914 | n: sizeof (sel_common_sched_info)); |
5915 | |
5916 | sel_common_sched_info.fix_recovery_cfg = NULL; |
5917 | sel_common_sched_info.add_block = NULL; |
5918 | sel_common_sched_info.estimate_number_of_insns |
5919 | = sel_estimate_number_of_insns; |
5920 | sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; |
5921 | sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; |
5922 | |
5923 | common_sched_info = &sel_common_sched_info; |
5924 | |
5925 | current_sched_info = &sched_sel_haifa_sched_info; |
5926 | current_sched_info->sched_max_insns_priority = |
5927 | get_rgn_sched_max_insns_priority (); |
5928 | |
5929 | sel_set_sched_flags (); |
5930 | } |
5931 | |
5932 | |
5933 | /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, |
5934 | *BB_ORD_INDEX after that is increased. */ |
5935 | static void |
5936 | sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) |
5937 | { |
5938 | RGN_NR_BLOCKS (rgn) += 1; |
5939 | RGN_DONT_CALC_DEPS (rgn) = 0; |
5940 | RGN_HAS_REAL_EBB (rgn) = 0; |
5941 | CONTAINING_RGN (bb->index) = rgn; |
5942 | BLOCK_TO_BB (bb->index) = *bb_ord_index; |
5943 | rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; |
5944 | (*bb_ord_index)++; |
5945 | |
5946 | /* FIXME: it is true only when not scheduling ebbs. */ |
5947 | RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); |
5948 | } |
5949 | |
5950 | /* Functions to support pipelining of outer loops. */ |
5951 | |
5952 | /* Creates a new empty region and returns it's number. */ |
5953 | static int |
5954 | sel_create_new_region (void) |
5955 | { |
5956 | int new_rgn_number = nr_regions; |
5957 | |
5958 | RGN_NR_BLOCKS (new_rgn_number) = 0; |
5959 | |
5960 | /* FIXME: This will work only when EBBs are not created. */ |
5961 | if (new_rgn_number != 0) |
5962 | RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + |
5963 | RGN_NR_BLOCKS (new_rgn_number - 1); |
5964 | else |
5965 | RGN_BLOCKS (new_rgn_number) = 0; |
5966 | |
5967 | /* Set the blocks of the next region so the other functions may |
5968 | calculate the number of blocks in the region. */ |
5969 | RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + |
5970 | RGN_NR_BLOCKS (new_rgn_number); |
5971 | |
5972 | nr_regions++; |
5973 | |
5974 | return new_rgn_number; |
5975 | } |
5976 | |
5977 | /* If X has a smaller topological sort number than Y, returns -1; |
5978 | if greater, returns 1. */ |
5979 | static int |
5980 | bb_top_order_comparator (const void *x, const void *y) |
5981 | { |
5982 | basic_block bb1 = *(const basic_block *) x; |
5983 | basic_block bb2 = *(const basic_block *) y; |
5984 | |
5985 | gcc_assert (bb1 == bb2 |
5986 | || rev_top_order_index[bb1->index] |
5987 | != rev_top_order_index[bb2->index]); |
5988 | |
5989 | /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so |
5990 | bbs with greater number should go earlier. */ |
5991 | if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) |
5992 | return -1; |
5993 | else |
5994 | return 1; |
5995 | } |
5996 | |
5997 | /* Create a region for LOOP and return its number. If we don't want |
5998 | to pipeline LOOP, return -1. */ |
5999 | static int |
6000 | make_region_from_loop (class loop *loop) |
6001 | { |
6002 | unsigned int i; |
6003 | int new_rgn_number = -1; |
6004 | class loop *inner; |
6005 | |
6006 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ |
6007 | int bb_ord_index = 0; |
6008 | basic_block *loop_blocks; |
6009 | basic_block ; |
6010 | |
6011 | if (loop->num_nodes |
6012 | > (unsigned) param_max_pipeline_region_blocks) |
6013 | return -1; |
6014 | |
6015 | /* Don't pipeline loops whose latch belongs to some of its inner loops. */ |
6016 | for (inner = loop->inner; inner; inner = inner->inner) |
6017 | if (flow_bb_inside_loop_p (inner, loop->latch)) |
6018 | return -1; |
6019 | |
6020 | loop->ninsns = num_loop_insns (loop); |
6021 | if ((int) loop->ninsns > param_max_pipeline_region_insns) |
6022 | return -1; |
6023 | |
6024 | loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); |
6025 | |
6026 | for (i = 0; i < loop->num_nodes; i++) |
6027 | if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) |
6028 | { |
6029 | free (ptr: loop_blocks); |
6030 | return -1; |
6031 | } |
6032 | |
6033 | preheader_block = loop_preheader_edge (loop)->src; |
6034 | gcc_assert (preheader_block); |
6035 | gcc_assert (loop_blocks[0] == loop->header); |
6036 | |
6037 | new_rgn_number = sel_create_new_region (); |
6038 | |
6039 | sel_add_block_to_region (bb: preheader_block, bb_ord_index: &bb_ord_index, rgn: new_rgn_number); |
6040 | bitmap_set_bit (map: bbs_in_loop_rgns, bitno: preheader_block->index); |
6041 | |
6042 | for (i = 0; i < loop->num_nodes; i++) |
6043 | { |
6044 | /* Add only those blocks that haven't been scheduled in the inner loop. |
6045 | The exception is the basic blocks with bookkeeping code - they should |
6046 | be added to the region (and they actually don't belong to the loop |
6047 | body, but to the region containing that loop body). */ |
6048 | |
6049 | gcc_assert (new_rgn_number >= 0); |
6050 | |
6051 | if (! bitmap_bit_p (map: bbs_in_loop_rgns, bitno: loop_blocks[i]->index)) |
6052 | { |
6053 | sel_add_block_to_region (bb: loop_blocks[i], bb_ord_index: &bb_ord_index, |
6054 | rgn: new_rgn_number); |
6055 | bitmap_set_bit (map: bbs_in_loop_rgns, bitno: loop_blocks[i]->index); |
6056 | } |
6057 | } |
6058 | |
6059 | free (ptr: loop_blocks); |
6060 | MARK_LOOP_FOR_PIPELINING (loop); |
6061 | |
6062 | return new_rgn_number; |
6063 | } |
6064 | |
6065 | /* Create a new region from preheader blocks LOOP_BLOCKS. */ |
6066 | void |
6067 | (vec<basic_block> *&loop_blocks) |
6068 | { |
6069 | unsigned int i; |
6070 | int new_rgn_number = -1; |
6071 | basic_block bb; |
6072 | |
6073 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ |
6074 | int bb_ord_index = 0; |
6075 | |
6076 | new_rgn_number = sel_create_new_region (); |
6077 | |
6078 | FOR_EACH_VEC_ELT (*loop_blocks, i, bb) |
6079 | { |
6080 | gcc_assert (new_rgn_number >= 0); |
6081 | |
6082 | sel_add_block_to_region (bb, bb_ord_index: &bb_ord_index, rgn: new_rgn_number); |
6083 | } |
6084 | |
6085 | vec_free (v&: loop_blocks); |
6086 | } |
6087 | |
6088 | |
6089 | /* Create region(s) from loop nest LOOP, such that inner loops will be |
6090 | pipelined before outer loops. Returns true when a region for LOOP |
6091 | is created. */ |
6092 | static bool |
6093 | make_regions_from_loop_nest (class loop *loop) |
6094 | { |
6095 | class loop *cur_loop; |
6096 | int rgn_number; |
6097 | |
6098 | /* Traverse all inner nodes of the loop. */ |
6099 | for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) |
6100 | if (! bitmap_bit_p (map: bbs_in_loop_rgns, bitno: cur_loop->header->index)) |
6101 | return false; |
6102 | |
6103 | /* At this moment all regular inner loops should have been pipelined. |
6104 | Try to create a region from this loop. */ |
6105 | rgn_number = make_region_from_loop (loop); |
6106 | |
6107 | if (rgn_number < 0) |
6108 | return false; |
6109 | |
6110 | loop_nests.safe_push (obj: loop); |
6111 | return true; |
6112 | } |
6113 | |
6114 | /* Initalize data structures needed. */ |
6115 | void |
6116 | sel_init_pipelining (void) |
6117 | { |
6118 | /* Collect loop information to be used in outer loops pipelining. */ |
6119 | loop_optimizer_init (LOOPS_HAVE_PREHEADERS |
6120 | | LOOPS_HAVE_FALLTHRU_PREHEADERS |
6121 | | LOOPS_HAVE_RECORDED_EXITS |
6122 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); |
6123 | current_loop_nest = NULL; |
6124 | |
6125 | bbs_in_loop_rgns = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
6126 | bitmap_clear (bbs_in_loop_rgns); |
6127 | |
6128 | recompute_rev_top_order (); |
6129 | } |
6130 | |
6131 | /* Returns a class loop for region RGN. */ |
6132 | loop_p |
6133 | get_loop_nest_for_rgn (unsigned int rgn) |
6134 | { |
6135 | /* Regions created with extend_rgns don't have corresponding loop nests, |
6136 | because they don't represent loops. */ |
6137 | if (rgn < loop_nests.length ()) |
6138 | return loop_nests[rgn]; |
6139 | else |
6140 | return NULL; |
6141 | } |
6142 | |
6143 | /* True when LOOP was included into pipelining regions. */ |
6144 | bool |
6145 | considered_for_pipelining_p (class loop *loop) |
6146 | { |
6147 | if (loop_depth (loop) == 0) |
6148 | return false; |
6149 | |
6150 | /* Now, the loop could be too large or irreducible. Check whether its |
6151 | region is in LOOP_NESTS. |
6152 | We determine the region number of LOOP as the region number of its |
6153 | latch. We can't use header here, because this header could be |
6154 | just removed preheader and it will give us the wrong region number. |
6155 | Latch can't be used because it could be in the inner loop too. */ |
6156 | if (LOOP_MARKED_FOR_PIPELINING_P (loop)) |
6157 | { |
6158 | int rgn = CONTAINING_RGN (loop->latch->index); |
6159 | |
6160 | gcc_assert ((unsigned) rgn < loop_nests.length ()); |
6161 | return true; |
6162 | } |
6163 | |
6164 | return false; |
6165 | } |
6166 | |
6167 | /* Makes regions from the rest of the blocks, after loops are chosen |
6168 | for pipelining. */ |
6169 | static void |
6170 | make_regions_from_the_rest (void) |
6171 | { |
6172 | int cur_rgn_blocks; |
6173 | int *loop_hdr; |
6174 | int i; |
6175 | |
6176 | basic_block bb; |
6177 | edge e; |
6178 | edge_iterator ei; |
6179 | int *degree; |
6180 | |
6181 | /* Index in rgn_bb_table where to start allocating new regions. */ |
6182 | cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; |
6183 | |
6184 | /* Make regions from all the rest basic blocks - those that don't belong to |
6185 | any loop or belong to irreducible loops. Prepare the data structures |
6186 | for extend_rgns. */ |
6187 | |
6188 | /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, |
6189 | LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same |
6190 | loop. */ |
6191 | loop_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
6192 | degree = XCNEWVEC (int, last_basic_block_for_fn (cfun)); |
6193 | |
6194 | |
6195 | /* For each basic block that belongs to some loop assign the number |
6196 | of innermost loop it belongs to. */ |
6197 | for (i = 0; i < last_basic_block_for_fn (cfun); i++) |
6198 | loop_hdr[i] = -1; |
6199 | |
6200 | FOR_EACH_BB_FN (bb, cfun) |
6201 | { |
6202 | if (bb->loop_father && bb->loop_father->num != 0 |
6203 | && !(bb->flags & BB_IRREDUCIBLE_LOOP)) |
6204 | loop_hdr[bb->index] = bb->loop_father->num; |
6205 | } |
6206 | |
6207 | /* For each basic block degree is calculated as the number of incoming |
6208 | edges, that are going out of bbs that are not yet scheduled. |
6209 | The basic blocks that are scheduled have degree value of zero. */ |
6210 | FOR_EACH_BB_FN (bb, cfun) |
6211 | { |
6212 | degree[bb->index] = 0; |
6213 | |
6214 | if (!bitmap_bit_p (map: bbs_in_loop_rgns, bitno: bb->index)) |
6215 | { |
6216 | FOR_EACH_EDGE (e, ei, bb->preds) |
6217 | if (!bitmap_bit_p (map: bbs_in_loop_rgns, bitno: e->src->index)) |
6218 | degree[bb->index]++; |
6219 | } |
6220 | else |
6221 | degree[bb->index] = -1; |
6222 | } |
6223 | |
6224 | extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); |
6225 | |
6226 | /* Any block that did not end up in a region is placed into a region |
6227 | by itself. */ |
6228 | FOR_EACH_BB_FN (bb, cfun) |
6229 | if (degree[bb->index] >= 0) |
6230 | { |
6231 | rgn_bb_table[cur_rgn_blocks] = bb->index; |
6232 | RGN_NR_BLOCKS (nr_regions) = 1; |
6233 | RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; |
6234 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
6235 | RGN_HAS_REAL_EBB (nr_regions) = 0; |
6236 | CONTAINING_RGN (bb->index) = nr_regions++; |
6237 | BLOCK_TO_BB (bb->index) = 0; |
6238 | } |
6239 | |
6240 | free (ptr: degree); |
6241 | free (ptr: loop_hdr); |
6242 | } |
6243 | |
6244 | /* Free data structures used in pipelining of loops. */ |
6245 | void sel_finish_pipelining (void) |
6246 | { |
6247 | /* Release aux fields so we don't free them later by mistake. */ |
6248 | for (auto loop : loops_list (cfun, 0)) |
6249 | loop->aux = NULL; |
6250 | |
6251 | loop_optimizer_finalize (); |
6252 | |
6253 | loop_nests.release (); |
6254 | |
6255 | free (ptr: rev_top_order_index); |
6256 | rev_top_order_index = NULL; |
6257 | } |
6258 | |
6259 | /* This function replaces the find_rgns when |
6260 | FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ |
6261 | void |
6262 | sel_find_rgns (void) |
6263 | { |
6264 | sel_init_pipelining (); |
6265 | extend_regions (); |
6266 | |
6267 | if (current_loops) |
6268 | { |
6269 | unsigned flags = flag_sel_sched_pipelining_outer_loops |
6270 | ? LI_FROM_INNERMOST |
6271 | : LI_ONLY_INNERMOST; |
6272 | |
6273 | for (auto loop : loops_list (cfun, flags)) |
6274 | make_regions_from_loop_nest (loop); |
6275 | } |
6276 | |
6277 | /* Make regions from all the rest basic blocks and schedule them. |
6278 | These blocks include blocks that don't belong to any loop or belong |
6279 | to irreducible loops. */ |
6280 | make_regions_from_the_rest (); |
6281 | |
6282 | /* We don't need bbs_in_loop_rgns anymore. */ |
6283 | sbitmap_free (map: bbs_in_loop_rgns); |
6284 | bbs_in_loop_rgns = NULL; |
6285 | } |
6286 | |
6287 | /* Add the preheader blocks from previous loop to current region taking |
6288 | it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS. |
6289 | This function is only used with -fsel-sched-pipelining-outer-loops. */ |
6290 | void |
6291 | (bb_vec_t *bbs) |
6292 | { |
6293 | int i; |
6294 | basic_block bb; |
6295 | vec<basic_block> * |
6296 | = LOOP_PREHEADER_BLOCKS (current_loop_nest); |
6297 | |
6298 | if (!preheader_blocks) |
6299 | return; |
6300 | |
6301 | for (i = 0; preheader_blocks->iterate (ix: i, ptr: &bb); i++) |
6302 | { |
6303 | bbs->safe_push (obj: bb); |
6304 | last_added_blocks.safe_push (obj: bb); |
6305 | sel_add_bb (bb); |
6306 | } |
6307 | |
6308 | vec_free (v&: preheader_blocks); |
6309 | } |
6310 | |
6311 | /* While pipelining outer loops, returns TRUE if BB is a loop preheader. |
6312 | Please note that the function should also work when pipelining_p is |
6313 | false, because it is used when deciding whether we should or should |
6314 | not reschedule pipelined code. */ |
6315 | bool |
6316 | (basic_block bb) |
6317 | { |
6318 | if (current_loop_nest) |
6319 | { |
6320 | class loop *outer; |
6321 | |
6322 | if (preheader_removed) |
6323 | return false; |
6324 | |
6325 | /* Preheader is the first block in the region. */ |
6326 | if (BLOCK_TO_BB (bb->index) == 0) |
6327 | return true; |
6328 | |
6329 | /* We used to find a preheader with the topological information. |
6330 | Check that the above code is equivalent to what we did before. */ |
6331 | |
6332 | if (in_current_region_p (bb: current_loop_nest->header)) |
6333 | gcc_assert (!(BLOCK_TO_BB (bb->index) |
6334 | < BLOCK_TO_BB (current_loop_nest->header->index))); |
6335 | |
6336 | /* Support the situation when the latch block of outer loop |
6337 | could be from here. */ |
6338 | for (outer = loop_outer (loop: current_loop_nest); |
6339 | outer; |
6340 | outer = loop_outer (loop: outer)) |
6341 | if (considered_for_pipelining_p (loop: outer) && outer->latch == bb) |
6342 | gcc_unreachable (); |
6343 | } |
6344 | |
6345 | return false; |
6346 | } |
6347 | |
6348 | /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and |
6349 | can be removed, making the corresponding edge fallthrough (assuming that |
6350 | all basic blocks between JUMP_BB and DEST_BB are empty). */ |
6351 | static bool |
6352 | bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb) |
6353 | { |
6354 | if (!onlyjump_p (BB_END (jump_bb)) |
6355 | || tablejump_p (BB_END (jump_bb), NULL, NULL)) |
6356 | return false; |
6357 | |
6358 | /* Several outgoing edges, abnormal edge or destination of jump is |
6359 | not DEST_BB. */ |
6360 | if (EDGE_COUNT (jump_bb->succs) != 1 |
6361 | || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING) |
6362 | || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) |
6363 | return false; |
6364 | |
6365 | /* If not anything of the upper. */ |
6366 | return true; |
6367 | } |
6368 | |
6369 | /* Removes the loop preheader from the current region and saves it in |
6370 | PREHEADER_BLOCKS of the father loop, so they will be added later to |
6371 | region that represents an outer loop. */ |
6372 | static void |
6373 | (void) |
6374 | { |
6375 | int i, old_len; |
6376 | int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); |
6377 | basic_block bb; |
6378 | bool all_empty_p = true; |
6379 | vec<basic_block> * |
6380 | = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); |
6381 | |
6382 | vec_check_alloc (vec&: preheader_blocks, nelems: 0); |
6383 | |
6384 | gcc_assert (current_loop_nest); |
6385 | old_len = preheader_blocks->length (); |
6386 | |
6387 | /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ |
6388 | for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) |
6389 | { |
6390 | bb = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
6391 | |
6392 | /* If the basic block belongs to region, but doesn't belong to |
6393 | corresponding loop, then it should be a preheader. */ |
6394 | if (sel_is_loop_preheader_p (bb)) |
6395 | { |
6396 | preheader_blocks->safe_push (obj: bb); |
6397 | if (BB_END (bb) != bb_note (bb)) |
6398 | all_empty_p = false; |
6399 | } |
6400 | } |
6401 | |
6402 | /* Remove these blocks only after iterating over the whole region. */ |
6403 | for (i = preheader_blocks->length () - 1; i >= old_len; i--) |
6404 | { |
6405 | bb = (*preheader_blocks)[i]; |
6406 | sel_remove_bb (bb, remove_from_cfg_p: false); |
6407 | } |
6408 | |
6409 | if (!considered_for_pipelining_p (loop: loop_outer (loop: current_loop_nest))) |
6410 | { |
6411 | if (!all_empty_p) |
6412 | /* Immediately create new region from preheader. */ |
6413 | make_region_from_loop_preheader (loop_blocks&: preheader_blocks); |
6414 | else |
6415 | { |
6416 | /* If all preheader blocks are empty - dont create new empty region. |
6417 | Instead, remove them completely. */ |
6418 | FOR_EACH_VEC_ELT (*preheader_blocks, i, bb) |
6419 | { |
6420 | edge e; |
6421 | edge_iterator ei; |
6422 | basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; |
6423 | |
6424 | /* Redirect all incoming edges to next basic block. */ |
6425 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (i: ei)); ) |
6426 | { |
6427 | if (! (e->flags & EDGE_FALLTHRU)) |
6428 | redirect_edge_and_branch (e, bb->next_bb); |
6429 | else |
6430 | redirect_edge_succ (e, bb->next_bb); |
6431 | } |
6432 | gcc_assert (BB_NOTE_LIST (bb) == NULL); |
6433 | delete_and_free_basic_block (bb); |
6434 | |
6435 | /* Check if after deleting preheader there is a nonconditional |
6436 | jump in PREV_BB that leads to the next basic block NEXT_BB. |
6437 | If it is so - delete this jump and clear data sets of its |
6438 | basic block if it becomes empty. */ |
6439 | if (next_bb->prev_bb == prev_bb |
6440 | && prev_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
6441 | && bb_has_removable_jump_to_p (jump_bb: prev_bb, dest_bb: next_bb)) |
6442 | { |
6443 | redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); |
6444 | if (BB_END (prev_bb) == bb_note (prev_bb)) |
6445 | free_data_sets (bb: prev_bb); |
6446 | } |
6447 | |
6448 | set_immediate_dominator (CDI_DOMINATORS, next_bb, |
6449 | recompute_dominator (CDI_DOMINATORS, |
6450 | next_bb)); |
6451 | } |
6452 | } |
6453 | vec_free (v&: preheader_blocks); |
6454 | } |
6455 | else |
6456 | /* Store preheader within the father's loop structure. */ |
6457 | SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), |
6458 | preheader_blocks); |
6459 | } |
6460 | |
6461 | #endif |
6462 | |