]>
Commit | Line | Data |
---|---|---|
e855c69d | 1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
86014d07 BS |
2 | Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 |
3 | Free Software Foundation, Inc. | |
e855c69d AB |
4 | |
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 3, or (at your option) any later | |
10 | version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
718f9c0f | 25 | #include "diagnostic-core.h" |
e855c69d AB |
26 | #include "rtl.h" |
27 | #include "tm_p.h" | |
28 | #include "hard-reg-set.h" | |
29 | #include "regs.h" | |
30 | #include "function.h" | |
31 | #include "flags.h" | |
32 | #include "insn-config.h" | |
33 | #include "insn-attr.h" | |
34 | #include "except.h" | |
e855c69d AB |
35 | #include "recog.h" |
36 | #include "params.h" | |
37 | #include "target.h" | |
38 | #include "timevar.h" | |
39 | #include "tree-pass.h" | |
40 | #include "sched-int.h" | |
41 | #include "ggc.h" | |
42 | #include "tree.h" | |
43 | #include "vec.h" | |
44 | #include "langhooks.h" | |
45 | #include "rtlhooks-def.h" | |
5936d944 | 46 | #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */ |
e855c69d AB |
47 | |
48 | #ifdef INSN_SCHEDULING | |
49 | #include "sel-sched-ir.h" | |
50 | /* We don't have to use it except for sel_print_insn. */ | |
51 | #include "sel-sched-dump.h" | |
52 | ||
53 | /* A vector holding bb info for whole scheduling pass. */ | |
54 | VEC(sel_global_bb_info_def, heap) *sel_global_bb_info = NULL; | |
55 | ||
56 | /* A vector holding bb info. */ | |
57 | VEC(sel_region_bb_info_def, heap) *sel_region_bb_info = NULL; | |
58 | ||
59 | /* A pool for allocating all lists. */ | |
60 | alloc_pool sched_lists_pool; | |
61 | ||
62 | /* This contains information about successors for compute_av_set. */ | |
63 | struct succs_info current_succs; | |
64 | ||
65 | /* Data structure to describe interaction with the generic scheduler utils. */ | |
66 | static struct common_sched_info_def sel_common_sched_info; | |
67 | ||
68 | /* The loop nest being pipelined. */ | |
69 | struct loop *current_loop_nest; | |
70 | ||
71 | /* LOOP_NESTS is a vector containing the corresponding loop nest for | |
72 | each region. */ | |
73 | static VEC(loop_p, heap) *loop_nests = NULL; | |
74 | ||
75 | /* Saves blocks already in loop regions, indexed by bb->index. */ | |
76 | static sbitmap bbs_in_loop_rgns = NULL; | |
77 | ||
78 | /* CFG hooks that are saved before changing create_basic_block hook. */ | |
79 | static struct cfg_hooks orig_cfg_hooks; | |
80 | \f | |
81 | ||
82 | /* Array containing reverse topological index of function basic blocks, | |
83 | indexed by BB->INDEX. */ | |
84 | static int *rev_top_order_index = NULL; | |
85 | ||
86 | /* Length of the above array. */ | |
87 | static int rev_top_order_index_len = -1; | |
88 | ||
89 | /* A regset pool structure. */ | |
90 | static struct | |
91 | { | |
92 | /* The stack to which regsets are returned. */ | |
93 | regset *v; | |
94 | ||
95 | /* Its pointer. */ | |
96 | int n; | |
97 | ||
98 | /* Its size. */ | |
99 | int s; | |
100 | ||
101 | /* In VV we save all generated regsets so that, when destructing the | |
102 | pool, we can compare it with V and check that every regset was returned | |
103 | back to pool. */ | |
104 | regset *vv; | |
105 | ||
106 | /* The pointer of VV stack. */ | |
107 | int nn; | |
108 | ||
109 | /* Its size. */ | |
110 | int ss; | |
111 | ||
112 | /* The difference between allocated and returned regsets. */ | |
113 | int diff; | |
114 | } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 }; | |
115 | ||
116 | /* This represents the nop pool. */ | |
117 | static struct | |
118 | { | |
119 | /* The vector which holds previously emitted nops. */ | |
120 | insn_t *v; | |
121 | ||
122 | /* Its pointer. */ | |
123 | int n; | |
124 | ||
125 | /* Its size. */ | |
b8698a0f | 126 | int s; |
e855c69d AB |
127 | } nop_pool = { NULL, 0, 0 }; |
128 | ||
129 | /* The pool for basic block notes. */ | |
130 | static rtx_vec_t bb_note_pool; | |
131 | ||
132 | /* A NOP pattern used to emit placeholder insns. */ | |
133 | rtx nop_pattern = NULL_RTX; | |
134 | /* A special instruction that resides in EXIT_BLOCK. | |
135 | EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */ | |
136 | rtx exit_insn = NULL_RTX; | |
137 | ||
b8698a0f | 138 | /* TRUE if while scheduling current region, which is loop, its preheader |
e855c69d AB |
139 | was removed. */ |
140 | bool preheader_removed = false; | |
141 | \f | |
142 | ||
143 | /* Forward static declarations. */ | |
144 | static void fence_clear (fence_t); | |
145 | ||
146 | static void deps_init_id (idata_t, insn_t, bool); | |
147 | static void init_id_from_df (idata_t, insn_t, bool); | |
148 | static expr_t set_insn_init (expr_t, vinsn_t, int); | |
149 | ||
150 | static void cfg_preds (basic_block, insn_t **, int *); | |
151 | static void prepare_insn_expr (insn_t, int); | |
152 | static void free_history_vect (VEC (expr_history_def, heap) **); | |
153 | ||
154 | static void move_bb_info (basic_block, basic_block); | |
155 | static void remove_empty_bb (basic_block, bool); | |
262d8232 | 156 | static void sel_merge_blocks (basic_block, basic_block); |
e855c69d | 157 | static void sel_remove_loop_preheader (void); |
753de8cf | 158 | static bool bb_has_removable_jump_to_p (basic_block, basic_block); |
e855c69d AB |
159 | |
160 | static bool insn_is_the_only_one_in_bb_p (insn_t); | |
161 | static void create_initial_data_sets (basic_block); | |
162 | ||
b5b8b0ac | 163 | static void free_av_set (basic_block); |
e855c69d AB |
164 | static void invalidate_av_set (basic_block); |
165 | static void extend_insn_data (void); | |
166 | static void sel_init_new_insn (insn_t, int); | |
167 | static void finish_insns (void); | |
168 | \f | |
169 | /* Various list functions. */ | |
170 | ||
171 | /* Copy an instruction list L. */ | |
172 | ilist_t | |
173 | ilist_copy (ilist_t l) | |
174 | { | |
175 | ilist_t head = NULL, *tailp = &head; | |
176 | ||
177 | while (l) | |
178 | { | |
179 | ilist_add (tailp, ILIST_INSN (l)); | |
180 | tailp = &ILIST_NEXT (*tailp); | |
181 | l = ILIST_NEXT (l); | |
182 | } | |
183 | ||
184 | return head; | |
185 | } | |
186 | ||
187 | /* Invert an instruction list L. */ | |
188 | ilist_t | |
189 | ilist_invert (ilist_t l) | |
190 | { | |
191 | ilist_t res = NULL; | |
192 | ||
193 | while (l) | |
194 | { | |
195 | ilist_add (&res, ILIST_INSN (l)); | |
196 | l = ILIST_NEXT (l); | |
197 | } | |
198 | ||
199 | return res; | |
200 | } | |
201 | ||
202 | /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */ | |
203 | void | |
204 | blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc) | |
205 | { | |
206 | bnd_t bnd; | |
207 | ||
208 | _list_add (lp); | |
209 | bnd = BLIST_BND (*lp); | |
210 | ||
211 | BND_TO (bnd) = to; | |
212 | BND_PTR (bnd) = ptr; | |
213 | BND_AV (bnd) = NULL; | |
214 | BND_AV1 (bnd) = NULL; | |
215 | BND_DC (bnd) = dc; | |
216 | } | |
217 | ||
218 | /* Remove the list note pointed to by LP. */ | |
219 | void | |
220 | blist_remove (blist_t *lp) | |
221 | { | |
222 | bnd_t b = BLIST_BND (*lp); | |
223 | ||
224 | av_set_clear (&BND_AV (b)); | |
225 | av_set_clear (&BND_AV1 (b)); | |
226 | ilist_clear (&BND_PTR (b)); | |
227 | ||
228 | _list_remove (lp); | |
229 | } | |
230 | ||
231 | /* Init a fence tail L. */ | |
232 | void | |
233 | flist_tail_init (flist_tail_t l) | |
234 | { | |
235 | FLIST_TAIL_HEAD (l) = NULL; | |
236 | FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l); | |
237 | } | |
238 | ||
239 | /* Try to find fence corresponding to INSN in L. */ | |
240 | fence_t | |
241 | flist_lookup (flist_t l, insn_t insn) | |
242 | { | |
243 | while (l) | |
244 | { | |
245 | if (FENCE_INSN (FLIST_FENCE (l)) == insn) | |
246 | return FLIST_FENCE (l); | |
247 | ||
248 | l = FLIST_NEXT (l); | |
249 | } | |
250 | ||
251 | return NULL; | |
252 | } | |
253 | ||
254 | /* Init the fields of F before running fill_insns. */ | |
255 | static void | |
256 | init_fence_for_scheduling (fence_t f) | |
257 | { | |
258 | FENCE_BNDS (f) = NULL; | |
259 | FENCE_PROCESSED_P (f) = false; | |
260 | FENCE_SCHEDULED_P (f) = false; | |
261 | } | |
262 | ||
263 | /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */ | |
264 | static void | |
b8698a0f L |
265 | flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, |
266 | insn_t last_scheduled_insn, VEC(rtx,gc) *executing_insns, | |
267 | int *ready_ticks, int ready_ticks_size, insn_t sched_next, | |
136e01a3 | 268 | int cycle, int cycle_issued_insns, int issue_more, |
e855c69d AB |
269 | bool starts_cycle_p, bool after_stall_p) |
270 | { | |
271 | fence_t f; | |
272 | ||
273 | _list_add (lp); | |
274 | f = FLIST_FENCE (*lp); | |
275 | ||
276 | FENCE_INSN (f) = insn; | |
277 | ||
278 | gcc_assert (state != NULL); | |
279 | FENCE_STATE (f) = state; | |
280 | ||
281 | FENCE_CYCLE (f) = cycle; | |
282 | FENCE_ISSUED_INSNS (f) = cycle_issued_insns; | |
283 | FENCE_STARTS_CYCLE_P (f) = starts_cycle_p; | |
284 | FENCE_AFTER_STALL_P (f) = after_stall_p; | |
285 | ||
286 | gcc_assert (dc != NULL); | |
287 | FENCE_DC (f) = dc; | |
288 | ||
289 | gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL); | |
290 | FENCE_TC (f) = tc; | |
291 | ||
292 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
136e01a3 | 293 | FENCE_ISSUE_MORE (f) = issue_more; |
e855c69d AB |
294 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
295 | FENCE_READY_TICKS (f) = ready_ticks; | |
296 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
297 | FENCE_SCHED_NEXT (f) = sched_next; | |
298 | ||
299 | init_fence_for_scheduling (f); | |
300 | } | |
301 | ||
302 | /* Remove the head node of the list pointed to by LP. */ | |
303 | static void | |
304 | flist_remove (flist_t *lp) | |
305 | { | |
306 | if (FENCE_INSN (FLIST_FENCE (*lp))) | |
307 | fence_clear (FLIST_FENCE (*lp)); | |
308 | _list_remove (lp); | |
309 | } | |
310 | ||
311 | /* Clear the fence list pointed to by LP. */ | |
312 | void | |
313 | flist_clear (flist_t *lp) | |
314 | { | |
315 | while (*lp) | |
316 | flist_remove (lp); | |
317 | } | |
318 | ||
319 | /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */ | |
320 | void | |
321 | def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call) | |
322 | { | |
323 | def_t d; | |
b8698a0f | 324 | |
e855c69d AB |
325 | _list_add (dl); |
326 | d = DEF_LIST_DEF (*dl); | |
327 | ||
328 | d->orig_insn = original_insn; | |
329 | d->crosses_call = crosses_call; | |
330 | } | |
331 | \f | |
332 | ||
333 | /* Functions to work with target contexts. */ | |
334 | ||
b8698a0f | 335 | /* Bulk target context. It is convenient for debugging purposes to ensure |
e855c69d AB |
336 | that there are no uninitialized (null) target contexts. */ |
337 | static tc_t bulk_tc = (tc_t) 1; | |
338 | ||
b8698a0f | 339 | /* Target hooks wrappers. In the future we can provide some default |
e855c69d AB |
340 | implementations for them. */ |
341 | ||
342 | /* Allocate a store for the target context. */ | |
343 | static tc_t | |
344 | alloc_target_context (void) | |
345 | { | |
346 | return (targetm.sched.alloc_sched_context | |
347 | ? targetm.sched.alloc_sched_context () : bulk_tc); | |
348 | } | |
349 | ||
350 | /* Init target context TC. | |
351 | If CLEAN_P is true, then make TC as it is beginning of the scheduler. | |
352 | Overwise, copy current backend context to TC. */ | |
353 | static void | |
354 | init_target_context (tc_t tc, bool clean_p) | |
355 | { | |
356 | if (targetm.sched.init_sched_context) | |
357 | targetm.sched.init_sched_context (tc, clean_p); | |
358 | } | |
359 | ||
360 | /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as | |
361 | int init_target_context (). */ | |
362 | tc_t | |
363 | create_target_context (bool clean_p) | |
364 | { | |
365 | tc_t tc = alloc_target_context (); | |
366 | ||
367 | init_target_context (tc, clean_p); | |
368 | return tc; | |
369 | } | |
370 | ||
371 | /* Copy TC to the current backend context. */ | |
372 | void | |
373 | set_target_context (tc_t tc) | |
374 | { | |
375 | if (targetm.sched.set_sched_context) | |
376 | targetm.sched.set_sched_context (tc); | |
377 | } | |
378 | ||
379 | /* TC is about to be destroyed. Free any internal data. */ | |
380 | static void | |
381 | clear_target_context (tc_t tc) | |
382 | { | |
383 | if (targetm.sched.clear_sched_context) | |
384 | targetm.sched.clear_sched_context (tc); | |
385 | } | |
386 | ||
387 | /* Clear and free it. */ | |
388 | static void | |
389 | delete_target_context (tc_t tc) | |
390 | { | |
391 | clear_target_context (tc); | |
392 | ||
393 | if (targetm.sched.free_sched_context) | |
394 | targetm.sched.free_sched_context (tc); | |
395 | } | |
396 | ||
397 | /* Make a copy of FROM in TO. | |
398 | NB: May be this should be a hook. */ | |
399 | static void | |
400 | copy_target_context (tc_t to, tc_t from) | |
401 | { | |
402 | tc_t tmp = create_target_context (false); | |
403 | ||
404 | set_target_context (from); | |
405 | init_target_context (to, false); | |
406 | ||
407 | set_target_context (tmp); | |
408 | delete_target_context (tmp); | |
409 | } | |
410 | ||
411 | /* Create a copy of TC. */ | |
412 | static tc_t | |
413 | create_copy_of_target_context (tc_t tc) | |
414 | { | |
415 | tc_t copy = alloc_target_context (); | |
416 | ||
417 | copy_target_context (copy, tc); | |
418 | ||
419 | return copy; | |
420 | } | |
421 | ||
422 | /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P | |
423 | is the same as in init_target_context (). */ | |
424 | void | |
425 | reset_target_context (tc_t tc, bool clean_p) | |
426 | { | |
427 | clear_target_context (tc); | |
428 | init_target_context (tc, clean_p); | |
429 | } | |
430 | \f | |
b8698a0f | 431 | /* Functions to work with dependence contexts. |
88302d54 | 432 | Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence |
e855c69d AB |
433 | context. It accumulates information about processed insns to decide if |
434 | current insn is dependent on the processed ones. */ | |
435 | ||
436 | /* Make a copy of FROM in TO. */ | |
437 | static void | |
438 | copy_deps_context (deps_t to, deps_t from) | |
439 | { | |
bcf33775 | 440 | init_deps (to, false); |
e855c69d AB |
441 | deps_join (to, from); |
442 | } | |
443 | ||
444 | /* Allocate store for dep context. */ | |
445 | static deps_t | |
446 | alloc_deps_context (void) | |
447 | { | |
88302d54 | 448 | return XNEW (struct deps_desc); |
e855c69d AB |
449 | } |
450 | ||
451 | /* Allocate and initialize dep context. */ | |
452 | static deps_t | |
453 | create_deps_context (void) | |
454 | { | |
455 | deps_t dc = alloc_deps_context (); | |
456 | ||
bcf33775 | 457 | init_deps (dc, false); |
e855c69d AB |
458 | return dc; |
459 | } | |
460 | ||
461 | /* Create a copy of FROM. */ | |
462 | static deps_t | |
463 | create_copy_of_deps_context (deps_t from) | |
464 | { | |
465 | deps_t to = alloc_deps_context (); | |
466 | ||
467 | copy_deps_context (to, from); | |
468 | return to; | |
469 | } | |
470 | ||
471 | /* Clean up internal data of DC. */ | |
472 | static void | |
473 | clear_deps_context (deps_t dc) | |
474 | { | |
475 | free_deps (dc); | |
476 | } | |
477 | ||
478 | /* Clear and free DC. */ | |
479 | static void | |
480 | delete_deps_context (deps_t dc) | |
481 | { | |
482 | clear_deps_context (dc); | |
483 | free (dc); | |
484 | } | |
485 | ||
486 | /* Clear and init DC. */ | |
487 | static void | |
488 | reset_deps_context (deps_t dc) | |
489 | { | |
490 | clear_deps_context (dc); | |
bcf33775 | 491 | init_deps (dc, false); |
e855c69d AB |
492 | } |
493 | ||
b8698a0f | 494 | /* This structure describes the dependence analysis hooks for advancing |
e855c69d AB |
495 | dependence context. */ |
496 | static struct sched_deps_info_def advance_deps_context_sched_deps_info = | |
497 | { | |
498 | NULL, | |
499 | ||
500 | NULL, /* start_insn */ | |
501 | NULL, /* finish_insn */ | |
502 | NULL, /* start_lhs */ | |
503 | NULL, /* finish_lhs */ | |
504 | NULL, /* start_rhs */ | |
505 | NULL, /* finish_rhs */ | |
506 | haifa_note_reg_set, | |
507 | haifa_note_reg_clobber, | |
508 | haifa_note_reg_use, | |
509 | NULL, /* note_mem_dep */ | |
510 | NULL, /* note_dep */ | |
511 | ||
512 | 0, 0, 0 | |
513 | }; | |
514 | ||
515 | /* Process INSN and add its impact on DC. */ | |
516 | void | |
517 | advance_deps_context (deps_t dc, insn_t insn) | |
518 | { | |
519 | sched_deps_info = &advance_deps_context_sched_deps_info; | |
520 | deps_analyze_insn (dc, insn); | |
521 | } | |
522 | \f | |
523 | ||
524 | /* Functions to work with DFA states. */ | |
525 | ||
526 | /* Allocate store for a DFA state. */ | |
527 | static state_t | |
528 | state_alloc (void) | |
529 | { | |
530 | return xmalloc (dfa_state_size); | |
531 | } | |
532 | ||
533 | /* Allocate and initialize DFA state. */ | |
534 | static state_t | |
535 | state_create (void) | |
536 | { | |
537 | state_t state = state_alloc (); | |
538 | ||
539 | state_reset (state); | |
540 | advance_state (state); | |
541 | return state; | |
542 | } | |
543 | ||
544 | /* Free DFA state. */ | |
545 | static void | |
546 | state_free (state_t state) | |
547 | { | |
548 | free (state); | |
549 | } | |
550 | ||
551 | /* Make a copy of FROM in TO. */ | |
552 | static void | |
553 | state_copy (state_t to, state_t from) | |
554 | { | |
555 | memcpy (to, from, dfa_state_size); | |
556 | } | |
557 | ||
558 | /* Create a copy of FROM. */ | |
559 | static state_t | |
560 | state_create_copy (state_t from) | |
561 | { | |
562 | state_t to = state_alloc (); | |
563 | ||
564 | state_copy (to, from); | |
565 | return to; | |
566 | } | |
567 | \f | |
568 | ||
569 | /* Functions to work with fences. */ | |
570 | ||
571 | /* Clear the fence. */ | |
572 | static void | |
573 | fence_clear (fence_t f) | |
574 | { | |
575 | state_t s = FENCE_STATE (f); | |
576 | deps_t dc = FENCE_DC (f); | |
577 | void *tc = FENCE_TC (f); | |
578 | ||
579 | ilist_clear (&FENCE_BNDS (f)); | |
580 | ||
581 | gcc_assert ((s != NULL && dc != NULL && tc != NULL) | |
582 | || (s == NULL && dc == NULL && tc == NULL)); | |
583 | ||
584 | if (s != NULL) | |
585 | free (s); | |
586 | ||
587 | if (dc != NULL) | |
588 | delete_deps_context (dc); | |
589 | ||
590 | if (tc != NULL) | |
591 | delete_target_context (tc); | |
592 | VEC_free (rtx, gc, FENCE_EXECUTING_INSNS (f)); | |
593 | free (FENCE_READY_TICKS (f)); | |
594 | FENCE_READY_TICKS (f) = NULL; | |
595 | } | |
596 | ||
597 | /* Init a list of fences with successors of OLD_FENCE. */ | |
598 | void | |
599 | init_fences (insn_t old_fence) | |
600 | { | |
601 | insn_t succ; | |
602 | succ_iterator si; | |
603 | bool first = true; | |
604 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f L |
605 | |
606 | FOR_EACH_SUCC_1 (succ, si, old_fence, | |
e855c69d AB |
607 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
608 | { | |
b8698a0f | 609 | |
e855c69d AB |
610 | if (first) |
611 | first = false; | |
612 | else | |
613 | gcc_assert (flag_sel_sched_pipelining_outer_loops); | |
614 | ||
615 | flist_add (&fences, succ, | |
616 | state_create (), | |
617 | create_deps_context () /* dc */, | |
618 | create_target_context (true) /* tc */, | |
b8698a0f | 619 | NULL_RTX /* last_scheduled_insn */, |
e855c69d AB |
620 | NULL, /* executing_insns */ |
621 | XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ | |
622 | ready_ticks_size, | |
623 | NULL_RTX /* sched_next */, | |
b8698a0f | 624 | 1 /* cycle */, 0 /* cycle_issued_insns */, |
136e01a3 | 625 | issue_rate, /* issue_more */ |
b8698a0f | 626 | 1 /* starts_cycle_p */, 0 /* after_stall_p */); |
e855c69d AB |
627 | } |
628 | } | |
629 | ||
630 | /* Merges two fences (filling fields of fence F with resulting values) by | |
631 | following rules: 1) state, target context and last scheduled insn are | |
b8698a0f | 632 | propagated from fallthrough edge if it is available; |
e855c69d | 633 | 2) deps context and cycle is propagated from more probable edge; |
b8698a0f | 634 | 3) all other fields are set to corresponding constant values. |
e855c69d | 635 | |
b8698a0f | 636 | INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, |
136e01a3 AB |
637 | READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE |
638 | and AFTER_STALL_P are the corresponding fields of the second fence. */ | |
e855c69d AB |
639 | static void |
640 | merge_fences (fence_t f, insn_t insn, | |
b8698a0f | 641 | state_t state, deps_t dc, void *tc, |
e855c69d AB |
642 | rtx last_scheduled_insn, VEC(rtx, gc) *executing_insns, |
643 | int *ready_ticks, int ready_ticks_size, | |
136e01a3 | 644 | rtx sched_next, int cycle, int issue_more, bool after_stall_p) |
e855c69d AB |
645 | { |
646 | insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f); | |
647 | ||
648 | gcc_assert (sel_bb_head_p (FENCE_INSN (f)) | |
649 | && !sched_next && !FENCE_SCHED_NEXT (f)); | |
650 | ||
b8698a0f | 651 | /* Check if we can decide which path fences came. |
e855c69d AB |
652 | If we can't (or don't want to) - reset all. */ |
653 | if (last_scheduled_insn == NULL | |
654 | || last_scheduled_insn_old == NULL | |
b8698a0f L |
655 | /* This is a case when INSN is reachable on several paths from |
656 | one insn (this can happen when pipelining of outer loops is on and | |
657 | there are two edges: one going around of inner loop and the other - | |
e855c69d AB |
658 | right through it; in such case just reset everything). */ |
659 | || last_scheduled_insn == last_scheduled_insn_old) | |
660 | { | |
661 | state_reset (FENCE_STATE (f)); | |
662 | state_free (state); | |
b8698a0f | 663 | |
e855c69d AB |
664 | reset_deps_context (FENCE_DC (f)); |
665 | delete_deps_context (dc); | |
b8698a0f | 666 | |
e855c69d AB |
667 | reset_target_context (FENCE_TC (f), true); |
668 | delete_target_context (tc); | |
669 | ||
670 | if (cycle > FENCE_CYCLE (f)) | |
671 | FENCE_CYCLE (f) = cycle; | |
672 | ||
673 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; | |
136e01a3 | 674 | FENCE_ISSUE_MORE (f) = issue_rate; |
e855c69d AB |
675 | VEC_free (rtx, gc, executing_insns); |
676 | free (ready_ticks); | |
677 | if (FENCE_EXECUTING_INSNS (f)) | |
b8698a0f | 678 | VEC_block_remove (rtx, FENCE_EXECUTING_INSNS (f), 0, |
e855c69d AB |
679 | VEC_length (rtx, FENCE_EXECUTING_INSNS (f))); |
680 | if (FENCE_READY_TICKS (f)) | |
681 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
682 | } | |
683 | else | |
684 | { | |
685 | edge edge_old = NULL, edge_new = NULL; | |
686 | edge candidate; | |
687 | succ_iterator si; | |
688 | insn_t succ; | |
b8698a0f | 689 | |
e855c69d AB |
690 | /* Find fallthrough edge. */ |
691 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); | |
0fd4b31d | 692 | candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb); |
e855c69d AB |
693 | |
694 | if (!candidate | |
695 | || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn) | |
696 | && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old))) | |
697 | { | |
698 | /* No fallthrough edge leading to basic block of INSN. */ | |
699 | state_reset (FENCE_STATE (f)); | |
700 | state_free (state); | |
b8698a0f | 701 | |
e855c69d AB |
702 | reset_target_context (FENCE_TC (f), true); |
703 | delete_target_context (tc); | |
b8698a0f | 704 | |
e855c69d | 705 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; |
136e01a3 | 706 | FENCE_ISSUE_MORE (f) = issue_rate; |
e855c69d AB |
707 | } |
708 | else | |
709 | if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn)) | |
710 | { | |
b8698a0f | 711 | /* Would be weird if same insn is successor of several fallthrough |
e855c69d AB |
712 | edges. */ |
713 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
714 | != BLOCK_FOR_INSN (last_scheduled_insn_old)); | |
715 | ||
716 | state_free (FENCE_STATE (f)); | |
717 | FENCE_STATE (f) = state; | |
718 | ||
719 | delete_target_context (FENCE_TC (f)); | |
720 | FENCE_TC (f) = tc; | |
721 | ||
722 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
136e01a3 | 723 | FENCE_ISSUE_MORE (f) = issue_more; |
e855c69d AB |
724 | } |
725 | else | |
726 | { | |
727 | /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ | |
728 | state_free (state); | |
729 | delete_target_context (tc); | |
730 | ||
731 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
732 | != BLOCK_FOR_INSN (last_scheduled_insn)); | |
733 | } | |
734 | ||
735 | /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ | |
736 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, | |
737 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
738 | { | |
739 | if (succ == insn) | |
740 | { | |
741 | /* No same successor allowed from several edges. */ | |
742 | gcc_assert (!edge_old); | |
743 | edge_old = si.e1; | |
744 | } | |
745 | } | |
746 | /* Find edge of second predecessor (last_scheduled_insn->insn). */ | |
747 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, | |
748 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
749 | { | |
750 | if (succ == insn) | |
751 | { | |
752 | /* No same successor allowed from several edges. */ | |
753 | gcc_assert (!edge_new); | |
754 | edge_new = si.e1; | |
755 | } | |
756 | } | |
757 | ||
758 | /* Check if we can choose most probable predecessor. */ | |
759 | if (edge_old == NULL || edge_new == NULL) | |
760 | { | |
761 | reset_deps_context (FENCE_DC (f)); | |
762 | delete_deps_context (dc); | |
763 | VEC_free (rtx, gc, executing_insns); | |
764 | free (ready_ticks); | |
b8698a0f | 765 | |
e855c69d AB |
766 | FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); |
767 | if (FENCE_EXECUTING_INSNS (f)) | |
b8698a0f | 768 | VEC_block_remove (rtx, FENCE_EXECUTING_INSNS (f), 0, |
e855c69d AB |
769 | VEC_length (rtx, FENCE_EXECUTING_INSNS (f))); |
770 | if (FENCE_READY_TICKS (f)) | |
771 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
772 | } | |
773 | else | |
774 | if (edge_new->probability > edge_old->probability) | |
775 | { | |
776 | delete_deps_context (FENCE_DC (f)); | |
777 | FENCE_DC (f) = dc; | |
778 | VEC_free (rtx, gc, FENCE_EXECUTING_INSNS (f)); | |
779 | FENCE_EXECUTING_INSNS (f) = executing_insns; | |
780 | free (FENCE_READY_TICKS (f)); | |
781 | FENCE_READY_TICKS (f) = ready_ticks; | |
782 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
783 | FENCE_CYCLE (f) = cycle; | |
784 | } | |
785 | else | |
786 | { | |
787 | /* Leave DC and CYCLE untouched. */ | |
788 | delete_deps_context (dc); | |
789 | VEC_free (rtx, gc, executing_insns); | |
790 | free (ready_ticks); | |
791 | } | |
792 | } | |
793 | ||
794 | /* Fill remaining invariant fields. */ | |
795 | if (after_stall_p) | |
796 | FENCE_AFTER_STALL_P (f) = 1; | |
797 | ||
798 | FENCE_ISSUED_INSNS (f) = 0; | |
799 | FENCE_STARTS_CYCLE_P (f) = 1; | |
800 | FENCE_SCHED_NEXT (f) = NULL; | |
801 | } | |
802 | ||
b8698a0f | 803 | /* Add a new fence to NEW_FENCES list, initializing it from all |
e855c69d AB |
804 | other parameters. */ |
805 | static void | |
806 | add_to_fences (flist_tail_t new_fences, insn_t insn, | |
b8698a0f L |
807 | state_t state, deps_t dc, void *tc, rtx last_scheduled_insn, |
808 | VEC(rtx, gc) *executing_insns, int *ready_ticks, | |
809 | int ready_ticks_size, rtx sched_next, int cycle, | |
136e01a3 AB |
810 | int cycle_issued_insns, int issue_rate, |
811 | bool starts_cycle_p, bool after_stall_p) | |
e855c69d AB |
812 | { |
813 | fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn); | |
814 | ||
815 | if (! f) | |
816 | { | |
817 | flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc, | |
b8698a0f | 818 | last_scheduled_insn, executing_insns, ready_ticks, |
e855c69d | 819 | ready_ticks_size, sched_next, cycle, cycle_issued_insns, |
136e01a3 | 820 | issue_rate, starts_cycle_p, after_stall_p); |
e855c69d AB |
821 | |
822 | FLIST_TAIL_TAILP (new_fences) | |
823 | = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences)); | |
824 | } | |
825 | else | |
826 | { | |
b8698a0f L |
827 | merge_fences (f, insn, state, dc, tc, last_scheduled_insn, |
828 | executing_insns, ready_ticks, ready_ticks_size, | |
136e01a3 | 829 | sched_next, cycle, issue_rate, after_stall_p); |
e855c69d AB |
830 | } |
831 | } | |
832 | ||
833 | /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */ | |
834 | void | |
835 | move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences) | |
836 | { | |
837 | fence_t f, old; | |
838 | flist_t *tailp = FLIST_TAIL_TAILP (new_fences); | |
839 | ||
840 | old = FLIST_FENCE (old_fences); | |
b8698a0f | 841 | f = flist_lookup (FLIST_TAIL_HEAD (new_fences), |
e855c69d AB |
842 | FENCE_INSN (FLIST_FENCE (old_fences))); |
843 | if (f) | |
844 | { | |
845 | merge_fences (f, old->insn, old->state, old->dc, old->tc, | |
846 | old->last_scheduled_insn, old->executing_insns, | |
847 | old->ready_ticks, old->ready_ticks_size, | |
136e01a3 | 848 | old->sched_next, old->cycle, old->issue_more, |
e855c69d AB |
849 | old->after_stall_p); |
850 | } | |
851 | else | |
852 | { | |
853 | _list_add (tailp); | |
854 | FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp); | |
855 | *FLIST_FENCE (*tailp) = *old; | |
856 | init_fence_for_scheduling (FLIST_FENCE (*tailp)); | |
857 | } | |
858 | FENCE_INSN (old) = NULL; | |
859 | } | |
860 | ||
b8698a0f | 861 | /* Add a new fence to NEW_FENCES list and initialize most of its data |
e855c69d AB |
862 | as a clean one. */ |
863 | void | |
864 | add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
865 | { | |
866 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f | 867 | |
e855c69d AB |
868 | add_to_fences (new_fences, |
869 | succ, state_create (), create_deps_context (), | |
870 | create_target_context (true), | |
b8698a0f | 871 | NULL_RTX, NULL, |
e855c69d AB |
872 | XCNEWVEC (int, ready_ticks_size), ready_ticks_size, |
873 | NULL_RTX, FENCE_CYCLE (fence) + 1, | |
136e01a3 | 874 | 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence)); |
e855c69d AB |
875 | } |
876 | ||
b8698a0f | 877 | /* Add a new fence to NEW_FENCES list and initialize all of its data |
e855c69d AB |
878 | from FENCE and SUCC. */ |
879 | void | |
880 | add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
881 | { | |
b8698a0f | 882 | int * new_ready_ticks |
e855c69d | 883 | = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); |
b8698a0f | 884 | |
e855c69d AB |
885 | memcpy (new_ready_ticks, FENCE_READY_TICKS (fence), |
886 | FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); | |
887 | add_to_fences (new_fences, | |
888 | succ, state_create_copy (FENCE_STATE (fence)), | |
889 | create_copy_of_deps_context (FENCE_DC (fence)), | |
890 | create_copy_of_target_context (FENCE_TC (fence)), | |
b8698a0f | 891 | FENCE_LAST_SCHEDULED_INSN (fence), |
e855c69d AB |
892 | VEC_copy (rtx, gc, FENCE_EXECUTING_INSNS (fence)), |
893 | new_ready_ticks, | |
894 | FENCE_READY_TICKS_SIZE (fence), | |
895 | FENCE_SCHED_NEXT (fence), | |
896 | FENCE_CYCLE (fence), | |
897 | FENCE_ISSUED_INSNS (fence), | |
136e01a3 | 898 | FENCE_ISSUE_MORE (fence), |
e855c69d AB |
899 | FENCE_STARTS_CYCLE_P (fence), |
900 | FENCE_AFTER_STALL_P (fence)); | |
901 | } | |
902 | \f | |
903 | ||
904 | /* Functions to work with regset and nop pools. */ | |
905 | ||
906 | /* Returns the new regset from pool. It might have some of the bits set | |
907 | from the previous usage. */ | |
908 | regset | |
909 | get_regset_from_pool (void) | |
910 | { | |
911 | regset rs; | |
912 | ||
913 | if (regset_pool.n != 0) | |
914 | rs = regset_pool.v[--regset_pool.n]; | |
915 | else | |
916 | /* We need to create the regset. */ | |
917 | { | |
918 | rs = ALLOC_REG_SET (®_obstack); | |
919 | ||
920 | if (regset_pool.nn == regset_pool.ss) | |
921 | regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, | |
922 | (regset_pool.ss = 2 * regset_pool.ss + 1)); | |
923 | regset_pool.vv[regset_pool.nn++] = rs; | |
924 | } | |
925 | ||
926 | regset_pool.diff++; | |
927 | ||
928 | return rs; | |
929 | } | |
930 | ||
931 | /* Same as above, but returns the empty regset. */ | |
932 | regset | |
933 | get_clear_regset_from_pool (void) | |
934 | { | |
935 | regset rs = get_regset_from_pool (); | |
936 | ||
937 | CLEAR_REG_SET (rs); | |
938 | return rs; | |
939 | } | |
940 | ||
941 | /* Return regset RS to the pool for future use. */ | |
942 | void | |
943 | return_regset_to_pool (regset rs) | |
944 | { | |
9ef1bf71 | 945 | gcc_assert (rs); |
e855c69d AB |
946 | regset_pool.diff--; |
947 | ||
948 | if (regset_pool.n == regset_pool.s) | |
949 | regset_pool.v = XRESIZEVEC (regset, regset_pool.v, | |
950 | (regset_pool.s = 2 * regset_pool.s + 1)); | |
951 | regset_pool.v[regset_pool.n++] = rs; | |
952 | } | |
953 | ||
68ad446f | 954 | #ifdef ENABLE_CHECKING |
e855c69d AB |
955 | /* This is used as a qsort callback for sorting regset pool stacks. |
956 | X and XX are addresses of two regsets. They are never equal. */ | |
957 | static int | |
958 | cmp_v_in_regset_pool (const void *x, const void *xx) | |
959 | { | |
960 | return *((const regset *) x) - *((const regset *) xx); | |
961 | } | |
68ad446f | 962 | #endif |
e855c69d AB |
963 | |
964 | /* Free the regset pool possibly checking for memory leaks. */ | |
965 | void | |
966 | free_regset_pool (void) | |
967 | { | |
968 | #ifdef ENABLE_CHECKING | |
969 | { | |
970 | regset *v = regset_pool.v; | |
971 | int i = 0; | |
972 | int n = regset_pool.n; | |
b8698a0f | 973 | |
e855c69d AB |
974 | regset *vv = regset_pool.vv; |
975 | int ii = 0; | |
976 | int nn = regset_pool.nn; | |
b8698a0f | 977 | |
e855c69d | 978 | int diff = 0; |
b8698a0f | 979 | |
e855c69d | 980 | gcc_assert (n <= nn); |
b8698a0f | 981 | |
e855c69d AB |
982 | /* Sort both vectors so it will be possible to compare them. */ |
983 | qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); | |
984 | qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); | |
b8698a0f | 985 | |
e855c69d AB |
986 | while (ii < nn) |
987 | { | |
988 | if (v[i] == vv[ii]) | |
989 | i++; | |
990 | else | |
991 | /* VV[II] was lost. */ | |
992 | diff++; | |
b8698a0f | 993 | |
e855c69d AB |
994 | ii++; |
995 | } | |
b8698a0f | 996 | |
e855c69d AB |
997 | gcc_assert (diff == regset_pool.diff); |
998 | } | |
999 | #endif | |
b8698a0f | 1000 | |
e855c69d AB |
1001 | /* If not true - we have a memory leak. */ |
1002 | gcc_assert (regset_pool.diff == 0); | |
b8698a0f | 1003 | |
e855c69d AB |
1004 | while (regset_pool.n) |
1005 | { | |
1006 | --regset_pool.n; | |
1007 | FREE_REG_SET (regset_pool.v[regset_pool.n]); | |
1008 | } | |
1009 | ||
1010 | free (regset_pool.v); | |
1011 | regset_pool.v = NULL; | |
1012 | regset_pool.s = 0; | |
b8698a0f | 1013 | |
e855c69d AB |
1014 | free (regset_pool.vv); |
1015 | regset_pool.vv = NULL; | |
1016 | regset_pool.nn = 0; | |
1017 | regset_pool.ss = 0; | |
1018 | ||
1019 | regset_pool.diff = 0; | |
1020 | } | |
1021 | \f | |
1022 | ||
b8698a0f L |
1023 | /* Functions to work with nop pools. NOP insns are used as temporary |
1024 | placeholders of the insns being scheduled to allow correct update of | |
e855c69d AB |
1025 | the data sets. When update is finished, NOPs are deleted. */ |
1026 | ||
1027 | /* A vinsn that is used to represent a nop. This vinsn is shared among all | |
1028 | nops sel-sched generates. */ | |
1029 | static vinsn_t nop_vinsn = NULL; | |
1030 | ||
1031 | /* Emit a nop before INSN, taking it from pool. */ | |
1032 | insn_t | |
1033 | get_nop_from_pool (insn_t insn) | |
1034 | { | |
1035 | insn_t nop; | |
1036 | bool old_p = nop_pool.n != 0; | |
1037 | int flags; | |
1038 | ||
1039 | if (old_p) | |
1040 | nop = nop_pool.v[--nop_pool.n]; | |
1041 | else | |
1042 | nop = nop_pattern; | |
1043 | ||
1044 | nop = emit_insn_before (nop, insn); | |
1045 | ||
1046 | if (old_p) | |
1047 | flags = INSN_INIT_TODO_SSID; | |
1048 | else | |
1049 | flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; | |
1050 | ||
1051 | set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); | |
1052 | sel_init_new_insn (nop, flags); | |
1053 | ||
1054 | return nop; | |
1055 | } | |
1056 | ||
1057 | /* Remove NOP from the instruction stream and return it to the pool. */ | |
1058 | void | |
b5b8b0ac | 1059 | return_nop_to_pool (insn_t nop, bool full_tidying) |
e855c69d AB |
1060 | { |
1061 | gcc_assert (INSN_IN_STREAM_P (nop)); | |
b5b8b0ac | 1062 | sel_remove_insn (nop, false, full_tidying); |
e855c69d AB |
1063 | |
1064 | if (nop_pool.n == nop_pool.s) | |
b8698a0f | 1065 | nop_pool.v = XRESIZEVEC (rtx, nop_pool.v, |
e855c69d AB |
1066 | (nop_pool.s = 2 * nop_pool.s + 1)); |
1067 | nop_pool.v[nop_pool.n++] = nop; | |
1068 | } | |
1069 | ||
1070 | /* Free the nop pool. */ | |
1071 | void | |
1072 | free_nop_pool (void) | |
1073 | { | |
1074 | nop_pool.n = 0; | |
1075 | nop_pool.s = 0; | |
1076 | free (nop_pool.v); | |
1077 | nop_pool.v = NULL; | |
1078 | } | |
1079 | \f | |
1080 | ||
b8698a0f | 1081 | /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb. |
e855c69d AB |
1082 | The callback is given two rtxes XX and YY and writes the new rtxes |
1083 | to NX and NY in case some needs to be skipped. */ | |
1084 | static int | |
1085 | skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) | |
1086 | { | |
1087 | const_rtx x = *xx; | |
1088 | const_rtx y = *yy; | |
b8698a0f | 1089 | |
e855c69d AB |
1090 | if (GET_CODE (x) == UNSPEC |
1091 | && (targetm.sched.skip_rtx_p == NULL | |
1092 | || targetm.sched.skip_rtx_p (x))) | |
1093 | { | |
1094 | *nx = XVECEXP (x, 0, 0); | |
1095 | *ny = CONST_CAST_RTX (y); | |
1096 | return 1; | |
1097 | } | |
b8698a0f | 1098 | |
e855c69d AB |
1099 | if (GET_CODE (y) == UNSPEC |
1100 | && (targetm.sched.skip_rtx_p == NULL | |
1101 | || targetm.sched.skip_rtx_p (y))) | |
1102 | { | |
1103 | *nx = CONST_CAST_RTX (x); | |
1104 | *ny = XVECEXP (y, 0, 0); | |
1105 | return 1; | |
1106 | } | |
b8698a0f | 1107 | |
e855c69d AB |
1108 | return 0; |
1109 | } | |
1110 | ||
b8698a0f | 1111 | /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way |
e855c69d AB |
1112 | to support ia64 speculation. When changes are needed, new rtx X and new mode |
1113 | NMODE are written, and the callback returns true. */ | |
1114 | static int | |
1115 | hash_with_unspec_callback (const_rtx x, enum machine_mode mode ATTRIBUTE_UNUSED, | |
1116 | rtx *nx, enum machine_mode* nmode) | |
1117 | { | |
b8698a0f | 1118 | if (GET_CODE (x) == UNSPEC |
e855c69d AB |
1119 | && targetm.sched.skip_rtx_p |
1120 | && targetm.sched.skip_rtx_p (x)) | |
1121 | { | |
1122 | *nx = XVECEXP (x, 0 ,0); | |
32e8bb8e | 1123 | *nmode = VOIDmode; |
e855c69d AB |
1124 | return 1; |
1125 | } | |
b8698a0f | 1126 | |
e855c69d AB |
1127 | return 0; |
1128 | } | |
1129 | ||
1130 | /* Returns LHS and RHS are ok to be scheduled separately. */ | |
1131 | static bool | |
1132 | lhs_and_rhs_separable_p (rtx lhs, rtx rhs) | |
1133 | { | |
1134 | if (lhs == NULL || rhs == NULL) | |
1135 | return false; | |
1136 | ||
b8698a0f L |
1137 | /* Do not schedule CONST, CONST_INT and CONST_DOUBLE etc as rhs: no point |
1138 | to use reg, if const can be used. Moreover, scheduling const as rhs may | |
1139 | lead to mode mismatch cause consts don't have modes but they could be | |
e855c69d AB |
1140 | merged from branches where the same const used in different modes. */ |
1141 | if (CONSTANT_P (rhs)) | |
1142 | return false; | |
1143 | ||
1144 | /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ | |
1145 | if (COMPARISON_P (rhs)) | |
1146 | return false; | |
1147 | ||
1148 | /* Do not allow single REG to be an rhs. */ | |
1149 | if (REG_P (rhs)) | |
1150 | return false; | |
1151 | ||
b8698a0f | 1152 | /* See comment at find_used_regs_1 (*1) for explanation of this |
e855c69d AB |
1153 | restriction. */ |
1154 | /* FIXME: remove this later. */ | |
1155 | if (MEM_P (lhs)) | |
1156 | return false; | |
1157 | ||
1158 | /* This will filter all tricky things like ZERO_EXTRACT etc. | |
1159 | For now we don't handle it. */ | |
1160 | if (!REG_P (lhs) && !MEM_P (lhs)) | |
1161 | return false; | |
1162 | ||
1163 | return true; | |
1164 | } | |
1165 | ||
b8698a0f L |
1166 | /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When |
1167 | FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is | |
e855c69d AB |
1168 | used e.g. for insns from recovery blocks. */ |
1169 | static void | |
1170 | vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) | |
1171 | { | |
1172 | hash_rtx_callback_function hrcf; | |
1173 | int insn_class; | |
1174 | ||
1175 | VINSN_INSN_RTX (vi) = insn; | |
1176 | VINSN_COUNT (vi) = 0; | |
1177 | vi->cost = -1; | |
b8698a0f | 1178 | |
9ef1bf71 AM |
1179 | if (INSN_NOP_P (insn)) |
1180 | return; | |
1181 | ||
e855c69d AB |
1182 | if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) |
1183 | init_id_from_df (VINSN_ID (vi), insn, force_unique_p); | |
1184 | else | |
1185 | deps_init_id (VINSN_ID (vi), insn, force_unique_p); | |
b8698a0f | 1186 | |
e855c69d AB |
1187 | /* Hash vinsn depending on whether it is separable or not. */ |
1188 | hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; | |
1189 | if (VINSN_SEPARABLE_P (vi)) | |
1190 | { | |
1191 | rtx rhs = VINSN_RHS (vi); | |
1192 | ||
1193 | VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs), | |
1194 | NULL, NULL, false, hrcf); | |
1195 | VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi), | |
1196 | VOIDmode, NULL, NULL, | |
1197 | false, hrcf); | |
1198 | } | |
1199 | else | |
1200 | { | |
1201 | VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode, | |
1202 | NULL, NULL, false, hrcf); | |
1203 | VINSN_HASH_RTX (vi) = VINSN_HASH (vi); | |
1204 | } | |
b8698a0f | 1205 | |
e855c69d AB |
1206 | insn_class = haifa_classify_insn (insn); |
1207 | if (insn_class >= 2 | |
1208 | && (!targetm.sched.get_insn_spec_ds | |
1209 | || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) | |
1210 | == 0))) | |
1211 | VINSN_MAY_TRAP_P (vi) = true; | |
1212 | else | |
1213 | VINSN_MAY_TRAP_P (vi) = false; | |
1214 | } | |
1215 | ||
1216 | /* Indicate that VI has become the part of an rtx object. */ | |
1217 | void | |
1218 | vinsn_attach (vinsn_t vi) | |
1219 | { | |
1220 | /* Assert that VI is not pending for deletion. */ | |
1221 | gcc_assert (VINSN_INSN_RTX (vi)); | |
1222 | ||
1223 | VINSN_COUNT (vi)++; | |
1224 | } | |
1225 | ||
b8698a0f | 1226 | /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct |
e855c69d AB |
1227 | VINSN_TYPE (VI). */ |
1228 | static vinsn_t | |
1229 | vinsn_create (insn_t insn, bool force_unique_p) | |
1230 | { | |
1231 | vinsn_t vi = XCNEW (struct vinsn_def); | |
1232 | ||
1233 | vinsn_init (vi, insn, force_unique_p); | |
1234 | return vi; | |
1235 | } | |
1236 | ||
1237 | /* Return a copy of VI. When REATTACH_P is true, detach VI and attach | |
1238 | the copy. */ | |
b8698a0f | 1239 | vinsn_t |
e855c69d AB |
1240 | vinsn_copy (vinsn_t vi, bool reattach_p) |
1241 | { | |
1242 | rtx copy; | |
1243 | bool unique = VINSN_UNIQUE_P (vi); | |
1244 | vinsn_t new_vi; | |
b8698a0f | 1245 | |
e855c69d AB |
1246 | copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); |
1247 | new_vi = create_vinsn_from_insn_rtx (copy, unique); | |
1248 | if (reattach_p) | |
1249 | { | |
1250 | vinsn_detach (vi); | |
1251 | vinsn_attach (new_vi); | |
1252 | } | |
1253 | ||
1254 | return new_vi; | |
1255 | } | |
1256 | ||
1257 | /* Delete the VI vinsn and free its data. */ | |
1258 | static void | |
1259 | vinsn_delete (vinsn_t vi) | |
1260 | { | |
1261 | gcc_assert (VINSN_COUNT (vi) == 0); | |
1262 | ||
9ef1bf71 AM |
1263 | if (!INSN_NOP_P (VINSN_INSN_RTX (vi))) |
1264 | { | |
1265 | return_regset_to_pool (VINSN_REG_SETS (vi)); | |
1266 | return_regset_to_pool (VINSN_REG_USES (vi)); | |
1267 | return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); | |
1268 | } | |
e855c69d AB |
1269 | |
1270 | free (vi); | |
1271 | } | |
1272 | ||
b8698a0f | 1273 | /* Indicate that VI is no longer a part of some rtx object. |
e855c69d AB |
1274 | Remove VI if it is no longer needed. */ |
1275 | void | |
1276 | vinsn_detach (vinsn_t vi) | |
1277 | { | |
1278 | gcc_assert (VINSN_COUNT (vi) > 0); | |
1279 | ||
1280 | if (--VINSN_COUNT (vi) == 0) | |
1281 | vinsn_delete (vi); | |
1282 | } | |
1283 | ||
1284 | /* Returns TRUE if VI is a branch. */ | |
1285 | bool | |
1286 | vinsn_cond_branch_p (vinsn_t vi) | |
1287 | { | |
1288 | insn_t insn; | |
1289 | ||
1290 | if (!VINSN_UNIQUE_P (vi)) | |
1291 | return false; | |
1292 | ||
1293 | insn = VINSN_INSN_RTX (vi); | |
1294 | if (BB_END (BLOCK_FOR_INSN (insn)) != insn) | |
1295 | return false; | |
1296 | ||
1297 | return control_flow_insn_p (insn); | |
1298 | } | |
1299 | ||
1300 | /* Return latency of INSN. */ | |
1301 | static int | |
1302 | sel_insn_rtx_cost (rtx insn) | |
1303 | { | |
1304 | int cost; | |
1305 | ||
1306 | /* A USE insn, or something else we don't need to | |
1307 | understand. We can't pass these directly to | |
1308 | result_ready_cost or insn_default_latency because it will | |
1309 | trigger a fatal error for unrecognizable insns. */ | |
1310 | if (recog_memoized (insn) < 0) | |
1311 | cost = 0; | |
1312 | else | |
1313 | { | |
1314 | cost = insn_default_latency (insn); | |
1315 | ||
1316 | if (cost < 0) | |
1317 | cost = 0; | |
1318 | } | |
1319 | ||
1320 | return cost; | |
1321 | } | |
1322 | ||
1323 | /* Return the cost of the VI. | |
1324 | !!! FIXME: Unify with haifa-sched.c: insn_cost (). */ | |
1325 | int | |
1326 | sel_vinsn_cost (vinsn_t vi) | |
1327 | { | |
1328 | int cost = vi->cost; | |
1329 | ||
1330 | if (cost < 0) | |
1331 | { | |
1332 | cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); | |
1333 | vi->cost = cost; | |
1334 | } | |
1335 | ||
1336 | return cost; | |
1337 | } | |
1338 | \f | |
1339 | ||
1340 | /* Functions for insn emitting. */ | |
1341 | ||
1342 | /* Emit new insn after AFTER based on PATTERN and initialize its data from | |
1343 | EXPR and SEQNO. */ | |
1344 | insn_t | |
1345 | sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) | |
1346 | { | |
1347 | insn_t new_insn; | |
1348 | ||
1349 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); | |
1350 | ||
1351 | new_insn = emit_insn_after (pattern, after); | |
1352 | set_insn_init (expr, NULL, seqno); | |
1353 | sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); | |
1354 | ||
1355 | return new_insn; | |
1356 | } | |
1357 | ||
1358 | /* Force newly generated vinsns to be unique. */ | |
1359 | static bool init_insn_force_unique_p = false; | |
1360 | ||
1361 | /* Emit new speculation recovery insn after AFTER based on PATTERN and | |
1362 | initialize its data from EXPR and SEQNO. */ | |
1363 | insn_t | |
1364 | sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, | |
1365 | insn_t after) | |
1366 | { | |
1367 | insn_t insn; | |
1368 | ||
1369 | gcc_assert (!init_insn_force_unique_p); | |
1370 | ||
1371 | init_insn_force_unique_p = true; | |
1372 | insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); | |
1373 | CANT_MOVE (insn) = 1; | |
1374 | init_insn_force_unique_p = false; | |
1375 | ||
1376 | return insn; | |
1377 | } | |
1378 | ||
1379 | /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, | |
b8698a0f L |
1380 | take it as a new vinsn instead of EXPR's vinsn. |
1381 | We simplify insns later, after scheduling region in | |
e855c69d AB |
1382 | simplify_changed_insns. */ |
1383 | insn_t | |
b8698a0f | 1384 | sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
e855c69d AB |
1385 | insn_t after) |
1386 | { | |
1387 | expr_t emit_expr; | |
1388 | insn_t insn; | |
1389 | int flags; | |
b8698a0f L |
1390 | |
1391 | emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), | |
e855c69d AB |
1392 | seqno); |
1393 | insn = EXPR_INSN_RTX (emit_expr); | |
b8698a0f | 1394 | add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); |
e855c69d AB |
1395 | |
1396 | flags = INSN_INIT_TODO_SSID; | |
1397 | if (INSN_LUID (insn) == 0) | |
1398 | flags |= INSN_INIT_TODO_LUID; | |
1399 | sel_init_new_insn (insn, flags); | |
1400 | ||
1401 | return insn; | |
1402 | } | |
1403 | ||
1404 | /* Move insn from EXPR after AFTER. */ | |
1405 | insn_t | |
1406 | sel_move_insn (expr_t expr, int seqno, insn_t after) | |
1407 | { | |
1408 | insn_t insn = EXPR_INSN_RTX (expr); | |
1409 | basic_block bb = BLOCK_FOR_INSN (after); | |
1410 | insn_t next = NEXT_INSN (after); | |
1411 | ||
1412 | /* Assert that in move_op we disconnected this insn properly. */ | |
1413 | gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); | |
1414 | PREV_INSN (insn) = after; | |
1415 | NEXT_INSN (insn) = next; | |
1416 | ||
1417 | NEXT_INSN (after) = insn; | |
1418 | PREV_INSN (next) = insn; | |
1419 | ||
1420 | /* Update links from insn to bb and vice versa. */ | |
1421 | df_insn_change_bb (insn, bb); | |
1422 | if (BB_END (bb) == after) | |
1423 | BB_END (bb) = insn; | |
b8698a0f | 1424 | |
e855c69d AB |
1425 | prepare_insn_expr (insn, seqno); |
1426 | return insn; | |
1427 | } | |
1428 | ||
1429 | \f | |
1430 | /* Functions to work with right-hand sides. */ | |
1431 | ||
b8698a0f | 1432 | /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector |
e855c69d | 1433 | VECT and return true when found. Use NEW_VINSN for comparison only when |
b8698a0f L |
1434 | COMPARE_VINSNS is true. Write to INDP the index on which |
1435 | the search has stopped, such that inserting the new element at INDP will | |
e855c69d AB |
1436 | retain VECT's sort order. */ |
1437 | static bool | |
b8698a0f L |
1438 | find_in_history_vect_1 (VEC(expr_history_def, heap) *vect, |
1439 | unsigned uid, vinsn_t new_vinsn, | |
e855c69d AB |
1440 | bool compare_vinsns, int *indp) |
1441 | { | |
1442 | expr_history_def *arr; | |
1443 | int i, j, len = VEC_length (expr_history_def, vect); | |
1444 | ||
1445 | if (len == 0) | |
1446 | { | |
1447 | *indp = 0; | |
1448 | return false; | |
1449 | } | |
1450 | ||
1451 | arr = VEC_address (expr_history_def, vect); | |
1452 | i = 0, j = len - 1; | |
1453 | ||
1454 | while (i <= j) | |
1455 | { | |
1456 | unsigned auid = arr[i].uid; | |
b8698a0f | 1457 | vinsn_t avinsn = arr[i].new_expr_vinsn; |
e855c69d AB |
1458 | |
1459 | if (auid == uid | |
b8698a0f L |
1460 | /* When undoing transformation on a bookkeeping copy, the new vinsn |
1461 | may not be exactly equal to the one that is saved in the vector. | |
e855c69d AB |
1462 | This is because the insn whose copy we're checking was possibly |
1463 | substituted itself. */ | |
b8698a0f | 1464 | && (! compare_vinsns |
e855c69d AB |
1465 | || vinsn_equal_p (avinsn, new_vinsn))) |
1466 | { | |
1467 | *indp = i; | |
1468 | return true; | |
1469 | } | |
1470 | else if (auid > uid) | |
1471 | break; | |
1472 | i++; | |
1473 | } | |
1474 | ||
1475 | *indp = i; | |
1476 | return false; | |
1477 | } | |
1478 | ||
b8698a0f L |
1479 | /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return |
1480 | the position found or -1, if no such value is in vector. | |
e855c69d AB |
1481 | Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ |
1482 | int | |
b8698a0f | 1483 | find_in_history_vect (VEC(expr_history_def, heap) *vect, rtx insn, |
e855c69d AB |
1484 | vinsn_t new_vinsn, bool originators_p) |
1485 | { | |
1486 | int ind; | |
1487 | ||
b8698a0f | 1488 | if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn, |
e855c69d AB |
1489 | false, &ind)) |
1490 | return ind; | |
1491 | ||
1492 | if (INSN_ORIGINATORS (insn) && originators_p) | |
1493 | { | |
1494 | unsigned uid; | |
1495 | bitmap_iterator bi; | |
1496 | ||
1497 | EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) | |
1498 | if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind)) | |
1499 | return ind; | |
1500 | } | |
b8698a0f | 1501 | |
e855c69d AB |
1502 | return -1; |
1503 | } | |
1504 | ||
b8698a0f L |
1505 | /* Insert new element in a sorted history vector pointed to by PVECT, |
1506 | if it is not there already. The element is searched using | |
e855c69d AB |
1507 | UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save |
1508 | the history of a transformation. */ | |
1509 | void | |
1510 | insert_in_history_vect (VEC (expr_history_def, heap) **pvect, | |
1511 | unsigned uid, enum local_trans_type type, | |
b8698a0f | 1512 | vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, |
e855c69d AB |
1513 | ds_t spec_ds) |
1514 | { | |
1515 | VEC(expr_history_def, heap) *vect = *pvect; | |
1516 | expr_history_def temp; | |
1517 | bool res; | |
1518 | int ind; | |
1519 | ||
1520 | res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind); | |
1521 | ||
1522 | if (res) | |
1523 | { | |
1524 | expr_history_def *phist = VEC_index (expr_history_def, vect, ind); | |
1525 | ||
b8698a0f | 1526 | /* It is possible that speculation types of expressions that were |
e855c69d AB |
1527 | propagated through different paths will be different here. In this |
1528 | case, merge the status to get the correct check later. */ | |
1529 | if (phist->spec_ds != spec_ds) | |
1530 | phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); | |
1531 | return; | |
1532 | } | |
b8698a0f | 1533 | |
e855c69d AB |
1534 | temp.uid = uid; |
1535 | temp.old_expr_vinsn = old_expr_vinsn; | |
b8698a0f | 1536 | temp.new_expr_vinsn = new_expr_vinsn; |
e855c69d AB |
1537 | temp.spec_ds = spec_ds; |
1538 | temp.type = type; | |
1539 | ||
1540 | vinsn_attach (old_expr_vinsn); | |
1541 | vinsn_attach (new_expr_vinsn); | |
1542 | VEC_safe_insert (expr_history_def, heap, vect, ind, &temp); | |
1543 | *pvect = vect; | |
1544 | } | |
1545 | ||
1546 | /* Free history vector PVECT. */ | |
1547 | static void | |
1548 | free_history_vect (VEC (expr_history_def, heap) **pvect) | |
1549 | { | |
1550 | unsigned i; | |
1551 | expr_history_def *phist; | |
1552 | ||
1553 | if (! *pvect) | |
1554 | return; | |
b8698a0f L |
1555 | |
1556 | for (i = 0; | |
e855c69d AB |
1557 | VEC_iterate (expr_history_def, *pvect, i, phist); |
1558 | i++) | |
1559 | { | |
1560 | vinsn_detach (phist->old_expr_vinsn); | |
1561 | vinsn_detach (phist->new_expr_vinsn); | |
1562 | } | |
b8698a0f | 1563 | |
e855c69d AB |
1564 | VEC_free (expr_history_def, heap, *pvect); |
1565 | *pvect = NULL; | |
1566 | } | |
1567 | ||
5d369d58 AB |
1568 | /* Merge vector FROM to PVECT. */ |
1569 | static void | |
1570 | merge_history_vect (VEC (expr_history_def, heap) **pvect, | |
1571 | VEC (expr_history_def, heap) *from) | |
1572 | { | |
1573 | expr_history_def *phist; | |
1574 | int i; | |
1575 | ||
1576 | /* We keep this vector sorted. */ | |
1577 | for (i = 0; VEC_iterate (expr_history_def, from, i, phist); i++) | |
1578 | insert_in_history_vect (pvect, phist->uid, phist->type, | |
1579 | phist->old_expr_vinsn, phist->new_expr_vinsn, | |
1580 | phist->spec_ds); | |
1581 | } | |
e855c69d AB |
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; | |
b8698a0f | 1599 | if (VINSN_SEPARABLE_P (x)) |
e855c69d AB |
1600 | { |
1601 | /* Compare RHSes of VINSNs. */ | |
1602 | gcc_assert (VINSN_RHS (x)); | |
1603 | gcc_assert (VINSN_RHS (y)); | |
1604 | ||
1605 | return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf); | |
1606 | } | |
1607 | ||
1608 | return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); | |
1609 | } | |
1610 | \f | |
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, | |
f3764768 | 1619 | VEC(expr_history_def, heap) *history, signed char target_available, |
e855c69d AB |
1620 | bool was_substituted, bool was_renamed, bool needs_spec_check_p, |
1621 | bool cant_move) | |
1622 | { | |
1623 | vinsn_attach (vi); | |
1624 | ||
1625 | EXPR_VINSN (expr) = vi; | |
1626 | EXPR_SPEC (expr) = spec; | |
1627 | EXPR_USEFULNESS (expr) = use; | |
1628 | EXPR_PRIORITY (expr) = priority; | |
1629 | EXPR_PRIORITY_ADJ (expr) = 0; | |
1630 | EXPR_SCHED_TIMES (expr) = sched_times; | |
1631 | EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; | |
1632 | EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; | |
1633 | EXPR_SPEC_DONE_DS (expr) = spec_done_ds; | |
1634 | EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; | |
1635 | ||
1636 | if (history) | |
1637 | EXPR_HISTORY_OF_CHANGES (expr) = history; | |
1638 | else | |
1639 | EXPR_HISTORY_OF_CHANGES (expr) = NULL; | |
1640 | ||
1641 | EXPR_TARGET_AVAILABLE (expr) = target_available; | |
1642 | EXPR_WAS_SUBSTITUTED (expr) = was_substituted; | |
1643 | EXPR_WAS_RENAMED (expr) = was_renamed; | |
1644 | EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; | |
1645 | EXPR_CANT_MOVE (expr) = cant_move; | |
1646 | } | |
1647 | ||
1648 | /* Make a copy of the expr FROM into the expr TO. */ | |
1649 | void | |
1650 | copy_expr (expr_t to, expr_t from) | |
1651 | { | |
1652 | VEC(expr_history_def, heap) *temp = NULL; | |
1653 | ||
1654 | if (EXPR_HISTORY_OF_CHANGES (from)) | |
1655 | { | |
1656 | unsigned i; | |
1657 | expr_history_def *phist; | |
1658 | ||
1659 | temp = VEC_copy (expr_history_def, heap, EXPR_HISTORY_OF_CHANGES (from)); | |
b8698a0f | 1660 | for (i = 0; |
e855c69d AB |
1661 | VEC_iterate (expr_history_def, temp, i, phist); |
1662 | i++) | |
1663 | { | |
1664 | vinsn_attach (phist->old_expr_vinsn); | |
1665 | vinsn_attach (phist->new_expr_vinsn); | |
1666 | } | |
1667 | } | |
1668 | ||
b8698a0f | 1669 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), |
e855c69d AB |
1670 | EXPR_USEFULNESS (from), EXPR_PRIORITY (from), |
1671 | EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), | |
b8698a0f | 1672 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), |
e855c69d | 1673 | EXPR_ORIG_SCHED_CYCLE (from), temp, |
b8698a0f | 1674 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
e855c69d AB |
1675 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1676 | EXPR_CANT_MOVE (from)); | |
1677 | } | |
1678 | ||
b8698a0f | 1679 | /* Same, but the final expr will not ever be in av sets, so don't copy |
e855c69d AB |
1680 | "uninteresting" data such as bitmap cache. */ |
1681 | void | |
1682 | copy_expr_onside (expr_t to, expr_t from) | |
1683 | { | |
1684 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), | |
1685 | EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0, | |
1686 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0, NULL, | |
1687 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), | |
1688 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), | |
1689 | EXPR_CANT_MOVE (from)); | |
1690 | } | |
1691 | ||
1692 | /* Prepare the expr of INSN for scheduling. Used when moving insn and when | |
1693 | initializing new insns. */ | |
1694 | static void | |
1695 | prepare_insn_expr (insn_t insn, int seqno) | |
1696 | { | |
1697 | expr_t expr = INSN_EXPR (insn); | |
1698 | ds_t ds; | |
b8698a0f | 1699 | |
e855c69d AB |
1700 | INSN_SEQNO (insn) = seqno; |
1701 | EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); | |
1702 | EXPR_SPEC (expr) = 0; | |
1703 | EXPR_ORIG_SCHED_CYCLE (expr) = 0; | |
1704 | EXPR_WAS_SUBSTITUTED (expr) = 0; | |
1705 | EXPR_WAS_RENAMED (expr) = 0; | |
1706 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
1707 | INSN_LIVE_VALID_P (insn) = false; | |
1708 | ||
1709 | /* ??? If this expression is speculative, make its dependence | |
1710 | as weak as possible. We can filter this expression later | |
1711 | in process_spec_exprs, because we do not distinguish | |
1712 | between the status we got during compute_av_set and the | |
1713 | existing status. To be fixed. */ | |
1714 | ds = EXPR_SPEC_DONE_DS (expr); | |
1715 | if (ds) | |
1716 | EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); | |
1717 | ||
1718 | free_history_vect (&EXPR_HISTORY_OF_CHANGES (expr)); | |
1719 | } | |
1720 | ||
1721 | /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1722 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1723 | a split point. */ |
1724 | static void | |
1725 | update_target_availability (expr_t to, expr_t from, insn_t split_point) | |
1726 | { | |
b8698a0f | 1727 | if (EXPR_TARGET_AVAILABLE (to) < 0 |
e855c69d AB |
1728 | || EXPR_TARGET_AVAILABLE (from) < 0) |
1729 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1730 | else | |
1731 | { | |
1732 | /* We try to detect the case when one of the expressions | |
1733 | can only be reached through another one. In this case, | |
1734 | we can do better. */ | |
1735 | if (split_point == NULL) | |
1736 | { | |
1737 | int toind, fromind; | |
1738 | ||
1739 | toind = EXPR_ORIG_BB_INDEX (to); | |
1740 | fromind = EXPR_ORIG_BB_INDEX (from); | |
b8698a0f | 1741 | |
e855c69d | 1742 | if (toind && toind == fromind) |
b8698a0f | 1743 | /* Do nothing -- everything is done in |
e855c69d AB |
1744 | merge_with_other_exprs. */ |
1745 | ; | |
1746 | else | |
1747 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1748 | } | |
1749 | else | |
1750 | EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); | |
1751 | } | |
1752 | } | |
1753 | ||
1754 | /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1755 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1756 | a split point. */ |
1757 | static void | |
1758 | update_speculative_bits (expr_t to, expr_t from, insn_t split_point) | |
1759 | { | |
1760 | ds_t old_to_ds, old_from_ds; | |
1761 | ||
1762 | old_to_ds = EXPR_SPEC_DONE_DS (to); | |
1763 | old_from_ds = EXPR_SPEC_DONE_DS (from); | |
b8698a0f | 1764 | |
e855c69d AB |
1765 | EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); |
1766 | EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); | |
1767 | EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); | |
1768 | ||
1769 | /* When merging e.g. control & data speculative exprs, or a control | |
b8698a0f | 1770 | speculative with a control&data speculative one, we really have |
e855c69d AB |
1771 | to change vinsn too. Also, when speculative status is changed, |
1772 | we also need to record this as a transformation in expr's history. */ | |
1773 | if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) | |
1774 | { | |
1775 | old_to_ds = ds_get_speculation_types (old_to_ds); | |
1776 | old_from_ds = ds_get_speculation_types (old_from_ds); | |
b8698a0f | 1777 | |
e855c69d AB |
1778 | if (old_to_ds != old_from_ds) |
1779 | { | |
1780 | ds_t record_ds; | |
b8698a0f L |
1781 | |
1782 | /* When both expressions are speculative, we need to change | |
e855c69d AB |
1783 | the vinsn first. */ |
1784 | if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) | |
1785 | { | |
1786 | int res; | |
b8698a0f | 1787 | |
e855c69d AB |
1788 | res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); |
1789 | gcc_assert (res >= 0); | |
1790 | } | |
1791 | ||
1792 | if (split_point != NULL) | |
1793 | { | |
1794 | /* Record the change with proper status. */ | |
1795 | record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; | |
1796 | record_ds &= ~(old_to_ds & SPECULATIVE); | |
1797 | record_ds &= ~(old_from_ds & SPECULATIVE); | |
b8698a0f L |
1798 | |
1799 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), | |
1800 | INSN_UID (split_point), TRANS_SPECULATION, | |
e855c69d AB |
1801 | EXPR_VINSN (from), EXPR_VINSN (to), |
1802 | record_ds); | |
1803 | } | |
1804 | } | |
1805 | } | |
1806 | } | |
1807 | ||
1808 | ||
1809 | /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, | |
1810 | this is done along different paths. */ | |
1811 | void | |
1812 | merge_expr_data (expr_t to, expr_t from, insn_t split_point) | |
1813 | { | |
e855c69d AB |
1814 | /* For now, we just set the spec of resulting expr to be minimum of the specs |
1815 | of merged exprs. */ | |
1816 | if (EXPR_SPEC (to) > EXPR_SPEC (from)) | |
1817 | EXPR_SPEC (to) = EXPR_SPEC (from); | |
1818 | ||
1819 | if (split_point) | |
1820 | EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); | |
1821 | else | |
b8698a0f | 1822 | EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), |
e855c69d AB |
1823 | EXPR_USEFULNESS (from)); |
1824 | ||
1825 | if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) | |
1826 | EXPR_PRIORITY (to) = EXPR_PRIORITY (from); | |
1827 | ||
1828 | if (EXPR_SCHED_TIMES (to) > EXPR_SCHED_TIMES (from)) | |
1829 | EXPR_SCHED_TIMES (to) = EXPR_SCHED_TIMES (from); | |
1830 | ||
1831 | if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) | |
1832 | EXPR_ORIG_BB_INDEX (to) = 0; | |
1833 | ||
b8698a0f | 1834 | EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), |
e855c69d AB |
1835 | EXPR_ORIG_SCHED_CYCLE (from)); |
1836 | ||
e855c69d AB |
1837 | EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); |
1838 | EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); | |
1839 | EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); | |
1840 | ||
5d369d58 AB |
1841 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to), |
1842 | EXPR_HISTORY_OF_CHANGES (from)); | |
e855c69d AB |
1843 | update_target_availability (to, from, split_point); |
1844 | update_speculative_bits (to, from, split_point); | |
1845 | } | |
1846 | ||
1847 | /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal | |
b8698a0f | 1848 | in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions |
e855c69d AB |
1849 | are merged from different successors at a split point. */ |
1850 | void | |
1851 | merge_expr (expr_t to, expr_t from, insn_t split_point) | |
1852 | { | |
1853 | vinsn_t to_vi = EXPR_VINSN (to); | |
1854 | vinsn_t from_vi = EXPR_VINSN (from); | |
1855 | ||
1856 | gcc_assert (vinsn_equal_p (to_vi, from_vi)); | |
1857 | ||
1858 | /* Make sure that speculative pattern is propagated into exprs that | |
1859 | have non-speculative one. This will provide us with consistent | |
1860 | speculative bits and speculative patterns inside expr. */ | |
1861 | if (EXPR_SPEC_DONE_DS (to) == 0 | |
1862 | && EXPR_SPEC_DONE_DS (from) != 0) | |
1863 | change_vinsn_in_expr (to, EXPR_VINSN (from)); | |
1864 | ||
1865 | merge_expr_data (to, from, split_point); | |
1866 | gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); | |
1867 | } | |
1868 | ||
1869 | /* Clear the information of this EXPR. */ | |
1870 | void | |
1871 | clear_expr (expr_t expr) | |
1872 | { | |
b8698a0f | 1873 | |
e855c69d AB |
1874 | vinsn_detach (EXPR_VINSN (expr)); |
1875 | EXPR_VINSN (expr) = NULL; | |
1876 | ||
1877 | free_history_vect (&EXPR_HISTORY_OF_CHANGES (expr)); | |
1878 | } | |
1879 | ||
1880 | /* For a given LV_SET, mark EXPR having unavailable target register. */ | |
1881 | static void | |
1882 | set_unavailable_target_for_expr (expr_t expr, regset lv_set) | |
1883 | { | |
1884 | if (EXPR_SEPARABLE_P (expr)) | |
1885 | { | |
1886 | if (REG_P (EXPR_LHS (expr)) | |
1887 | && bitmap_bit_p (lv_set, REGNO (EXPR_LHS (expr)))) | |
1888 | { | |
b8698a0f L |
1889 | /* If it's an insn like r1 = use (r1, ...), and it exists in |
1890 | different forms in each of the av_sets being merged, we can't say | |
1891 | whether original destination register is available or not. | |
1892 | However, this still works if destination register is not used | |
e855c69d AB |
1893 | in the original expression: if the branch at which LV_SET we're |
1894 | looking here is not actually 'other branch' in sense that same | |
b8698a0f | 1895 | expression is available through it (but it can't be determined |
e855c69d | 1896 | at computation stage because of transformations on one of the |
b8698a0f L |
1897 | branches), it still won't affect the availability. |
1898 | Liveness of a register somewhere on a code motion path means | |
1899 | it's either read somewhere on a codemotion path, live on | |
e855c69d AB |
1900 | 'other' branch, live at the point immediately following |
1901 | the original operation, or is read by the original operation. | |
1902 | The latter case is filtered out in the condition below. | |
1903 | It still doesn't cover the case when register is defined and used | |
1904 | somewhere within the code motion path, and in this case we could | |
1905 | miss a unifying code motion along both branches using a renamed | |
1906 | register, but it won't affect a code correctness since upon | |
1907 | an actual code motion a bookkeeping code would be generated. */ | |
b8698a0f | 1908 | if (bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
e855c69d AB |
1909 | REGNO (EXPR_LHS (expr)))) |
1910 | EXPR_TARGET_AVAILABLE (expr) = -1; | |
1911 | else | |
1912 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1913 | } | |
1914 | } | |
1915 | else | |
1916 | { | |
1917 | unsigned regno; | |
1918 | reg_set_iterator rsi; | |
b8698a0f L |
1919 | |
1920 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), | |
e855c69d AB |
1921 | 0, regno, rsi) |
1922 | if (bitmap_bit_p (lv_set, regno)) | |
1923 | { | |
1924 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1925 | break; | |
1926 | } | |
1927 | ||
1928 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), | |
1929 | 0, regno, rsi) | |
1930 | if (bitmap_bit_p (lv_set, regno)) | |
1931 | { | |
1932 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1933 | break; | |
1934 | } | |
1935 | } | |
1936 | } | |
1937 | ||
b8698a0f | 1938 | /* Try to make EXPR speculative. Return 1 when EXPR's pattern |
e855c69d AB |
1939 | or dependence status have changed, 2 when also the target register |
1940 | became unavailable, 0 if nothing had to be changed. */ | |
1941 | int | |
1942 | speculate_expr (expr_t expr, ds_t ds) | |
1943 | { | |
1944 | int res; | |
1945 | rtx orig_insn_rtx; | |
1946 | rtx spec_pat; | |
1947 | ds_t target_ds, current_ds; | |
1948 | ||
1949 | /* Obtain the status we need to put on EXPR. */ | |
1950 | target_ds = (ds & SPECULATIVE); | |
1951 | current_ds = EXPR_SPEC_DONE_DS (expr); | |
1952 | ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); | |
1953 | ||
1954 | orig_insn_rtx = EXPR_INSN_RTX (expr); | |
1955 | ||
1956 | res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); | |
1957 | ||
1958 | switch (res) | |
1959 | { | |
1960 | case 0: | |
1961 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1962 | return current_ds != ds ? 1 : 0; | |
b8698a0f | 1963 | |
e855c69d AB |
1964 | case 1: |
1965 | { | |
1966 | rtx spec_insn_rtx = create_insn_rtx_from_pattern (spec_pat, NULL_RTX); | |
1967 | vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); | |
1968 | ||
1969 | change_vinsn_in_expr (expr, spec_vinsn); | |
1970 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1971 | EXPR_NEEDS_SPEC_CHECK_P (expr) = true; | |
1972 | ||
b8698a0f | 1973 | /* Do not allow clobbering the address register of speculative |
e855c69d | 1974 | insns. */ |
b8698a0f | 1975 | if (bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
e855c69d AB |
1976 | expr_dest_regno (expr))) |
1977 | { | |
1978 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1979 | return 2; | |
1980 | } | |
1981 | ||
1982 | return 1; | |
1983 | } | |
1984 | ||
1985 | case -1: | |
1986 | return -1; | |
1987 | ||
1988 | default: | |
1989 | gcc_unreachable (); | |
1990 | return -1; | |
1991 | } | |
1992 | } | |
1993 | ||
1994 | /* Return a destination register, if any, of EXPR. */ | |
1995 | rtx | |
1996 | expr_dest_reg (expr_t expr) | |
1997 | { | |
1998 | rtx dest = VINSN_LHS (EXPR_VINSN (expr)); | |
1999 | ||
2000 | if (dest != NULL_RTX && REG_P (dest)) | |
2001 | return dest; | |
2002 | ||
2003 | return NULL_RTX; | |
2004 | } | |
2005 | ||
2006 | /* Returns the REGNO of the R's destination. */ | |
2007 | unsigned | |
2008 | expr_dest_regno (expr_t expr) | |
2009 | { | |
2010 | rtx dest = expr_dest_reg (expr); | |
2011 | ||
2012 | gcc_assert (dest != NULL_RTX); | |
2013 | return REGNO (dest); | |
2014 | } | |
2015 | ||
b8698a0f | 2016 | /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in |
e855c69d AB |
2017 | AV_SET having unavailable target register. */ |
2018 | void | |
2019 | mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) | |
2020 | { | |
2021 | expr_t expr; | |
2022 | av_set_iterator avi; | |
2023 | ||
2024 | FOR_EACH_EXPR (expr, avi, join_set) | |
2025 | if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) | |
2026 | set_unavailable_target_for_expr (expr, lv_set); | |
2027 | } | |
2028 | \f | |
2029 | ||
2030 | /* Av set functions. */ | |
2031 | ||
2032 | /* Add a new element to av set SETP. | |
2033 | Return the element added. */ | |
2034 | static av_set_t | |
2035 | av_set_add_element (av_set_t *setp) | |
2036 | { | |
2037 | /* Insert at the beginning of the list. */ | |
2038 | _list_add (setp); | |
2039 | return *setp; | |
2040 | } | |
2041 | ||
2042 | /* Add EXPR to SETP. */ | |
2043 | void | |
2044 | av_set_add (av_set_t *setp, expr_t expr) | |
2045 | { | |
2046 | av_set_t elem; | |
b8698a0f | 2047 | |
e855c69d AB |
2048 | gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); |
2049 | elem = av_set_add_element (setp); | |
2050 | copy_expr (_AV_SET_EXPR (elem), expr); | |
2051 | } | |
2052 | ||
2053 | /* Same, but do not copy EXPR. */ | |
2054 | static void | |
2055 | av_set_add_nocopy (av_set_t *setp, expr_t expr) | |
2056 | { | |
2057 | av_set_t elem; | |
2058 | ||
2059 | elem = av_set_add_element (setp); | |
2060 | *_AV_SET_EXPR (elem) = *expr; | |
2061 | } | |
2062 | ||
2063 | /* Remove expr pointed to by IP from the av_set. */ | |
2064 | void | |
2065 | av_set_iter_remove (av_set_iterator *ip) | |
2066 | { | |
2067 | clear_expr (_AV_SET_EXPR (*ip->lp)); | |
2068 | _list_iter_remove (ip); | |
2069 | } | |
2070 | ||
2071 | /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the | |
2072 | sense of vinsn_equal_p function. Return NULL if no such expr is | |
2073 | in SET was found. */ | |
2074 | expr_t | |
2075 | av_set_lookup (av_set_t set, vinsn_t sought_vinsn) | |
2076 | { | |
2077 | expr_t expr; | |
2078 | av_set_iterator i; | |
2079 | ||
2080 | FOR_EACH_EXPR (expr, i, set) | |
2081 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2082 | return expr; | |
2083 | return NULL; | |
2084 | } | |
2085 | ||
2086 | /* Same, but also remove the EXPR found. */ | |
2087 | static expr_t | |
2088 | av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) | |
2089 | { | |
2090 | expr_t expr; | |
2091 | av_set_iterator i; | |
2092 | ||
2093 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2094 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2095 | { | |
2096 | _list_iter_remove_nofree (&i); | |
2097 | return expr; | |
2098 | } | |
2099 | return NULL; | |
2100 | } | |
2101 | ||
2102 | /* Search for an expr in SET, such that it's equivalent to EXPR in the | |
2103 | sense of vinsn_equal_p function of their vinsns, but not EXPR itself. | |
2104 | Returns NULL if no such expr is in SET was found. */ | |
2105 | static expr_t | |
2106 | av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) | |
2107 | { | |
2108 | expr_t cur_expr; | |
2109 | av_set_iterator i; | |
2110 | ||
2111 | FOR_EACH_EXPR (cur_expr, i, set) | |
2112 | { | |
2113 | if (cur_expr == expr) | |
2114 | continue; | |
2115 | if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) | |
2116 | return cur_expr; | |
2117 | } | |
2118 | ||
2119 | return NULL; | |
2120 | } | |
2121 | ||
2122 | /* If other expression is already in AVP, remove one of them. */ | |
2123 | expr_t | |
2124 | merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) | |
2125 | { | |
2126 | expr_t expr2; | |
2127 | ||
2128 | expr2 = av_set_lookup_other_equiv_expr (*avp, expr); | |
2129 | if (expr2 != NULL) | |
2130 | { | |
2131 | /* Reset target availability on merge, since taking it only from one | |
2132 | of the exprs would be controversial for different code. */ | |
2133 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2134 | EXPR_USEFULNESS (expr2) = 0; | |
2135 | ||
2136 | merge_expr (expr2, expr, NULL); | |
b8698a0f | 2137 | |
e855c69d AB |
2138 | /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ |
2139 | EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; | |
b8698a0f | 2140 | |
e855c69d AB |
2141 | av_set_iter_remove (ip); |
2142 | return expr2; | |
2143 | } | |
2144 | ||
2145 | return expr; | |
2146 | } | |
2147 | ||
2148 | /* Return true if there is an expr that correlates to VI in SET. */ | |
2149 | bool | |
2150 | av_set_is_in_p (av_set_t set, vinsn_t vi) | |
2151 | { | |
2152 | return av_set_lookup (set, vi) != NULL; | |
2153 | } | |
2154 | ||
2155 | /* Return a copy of SET. */ | |
2156 | av_set_t | |
2157 | av_set_copy (av_set_t set) | |
2158 | { | |
2159 | expr_t expr; | |
2160 | av_set_iterator i; | |
2161 | av_set_t res = NULL; | |
2162 | ||
2163 | FOR_EACH_EXPR (expr, i, set) | |
2164 | av_set_add (&res, expr); | |
2165 | ||
2166 | return res; | |
2167 | } | |
2168 | ||
2169 | /* Join two av sets that do not have common elements by attaching second set | |
2170 | (pointed to by FROMP) to the end of first set (TO_TAILP must point to | |
2171 | _AV_SET_NEXT of first set's last element). */ | |
2172 | static void | |
2173 | join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) | |
2174 | { | |
2175 | gcc_assert (*to_tailp == NULL); | |
2176 | *to_tailp = *fromp; | |
2177 | *fromp = NULL; | |
2178 | } | |
2179 | ||
2180 | /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set | |
2181 | pointed to by FROMP afterwards. */ | |
2182 | void | |
2183 | av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) | |
2184 | { | |
2185 | expr_t expr1; | |
2186 | av_set_iterator i; | |
2187 | ||
2188 | /* Delete from TOP all exprs, that present in FROMP. */ | |
2189 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2190 | { | |
2191 | expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1)); | |
2192 | ||
2193 | if (expr2) | |
2194 | { | |
2195 | merge_expr (expr2, expr1, insn); | |
2196 | av_set_iter_remove (&i); | |
2197 | } | |
2198 | } | |
2199 | ||
2200 | join_distinct_sets (i.lp, fromp); | |
2201 | } | |
2202 | ||
b8698a0f | 2203 | /* Same as above, but also update availability of target register in |
e855c69d AB |
2204 | TOP judging by TO_LV_SET and FROM_LV_SET. */ |
2205 | void | |
2206 | av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, | |
2207 | regset from_lv_set, insn_t insn) | |
2208 | { | |
2209 | expr_t expr1; | |
2210 | av_set_iterator i; | |
2211 | av_set_t *to_tailp, in_both_set = NULL; | |
2212 | ||
2213 | /* Delete from TOP all expres, that present in FROMP. */ | |
2214 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2215 | { | |
2216 | expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1)); | |
2217 | ||
2218 | if (expr2) | |
2219 | { | |
b8698a0f | 2220 | /* It may be that the expressions have different destination |
e855c69d AB |
2221 | registers, in which case we need to check liveness here. */ |
2222 | if (EXPR_SEPARABLE_P (expr1)) | |
2223 | { | |
b8698a0f | 2224 | int regno1 = (REG_P (EXPR_LHS (expr1)) |
e855c69d | 2225 | ? (int) expr_dest_regno (expr1) : -1); |
b8698a0f | 2226 | int regno2 = (REG_P (EXPR_LHS (expr2)) |
e855c69d | 2227 | ? (int) expr_dest_regno (expr2) : -1); |
b8698a0f L |
2228 | |
2229 | /* ??? We don't have a way to check restrictions for | |
e855c69d AB |
2230 | *other* register on the current path, we did it only |
2231 | for the current target register. Give up. */ | |
2232 | if (regno1 != regno2) | |
2233 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2234 | } | |
2235 | else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) | |
2236 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2237 | ||
2238 | merge_expr (expr2, expr1, insn); | |
2239 | av_set_add_nocopy (&in_both_set, expr2); | |
2240 | av_set_iter_remove (&i); | |
2241 | } | |
2242 | else | |
b8698a0f | 2243 | /* EXPR1 is present in TOP, but not in FROMP. Check it on |
e855c69d AB |
2244 | FROM_LV_SET. */ |
2245 | set_unavailable_target_for_expr (expr1, from_lv_set); | |
2246 | } | |
2247 | to_tailp = i.lp; | |
2248 | ||
2249 | /* These expressions are not present in TOP. Check liveness | |
2250 | restrictions on TO_LV_SET. */ | |
2251 | FOR_EACH_EXPR (expr1, i, *fromp) | |
2252 | set_unavailable_target_for_expr (expr1, to_lv_set); | |
2253 | ||
2254 | join_distinct_sets (i.lp, &in_both_set); | |
2255 | join_distinct_sets (to_tailp, fromp); | |
2256 | } | |
2257 | ||
2258 | /* Clear av_set pointed to by SETP. */ | |
2259 | void | |
2260 | av_set_clear (av_set_t *setp) | |
2261 | { | |
2262 | expr_t expr; | |
2263 | av_set_iterator i; | |
2264 | ||
2265 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2266 | av_set_iter_remove (&i); | |
2267 | ||
2268 | gcc_assert (*setp == NULL); | |
2269 | } | |
2270 | ||
2271 | /* Leave only one non-speculative element in the SETP. */ | |
2272 | void | |
2273 | av_set_leave_one_nonspec (av_set_t *setp) | |
2274 | { | |
2275 | expr_t expr; | |
2276 | av_set_iterator i; | |
2277 | bool has_one_nonspec = false; | |
2278 | ||
b8698a0f | 2279 | /* Keep all speculative exprs, and leave one non-speculative |
e855c69d AB |
2280 | (the first one). */ |
2281 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2282 | { | |
2283 | if (!EXPR_SPEC_DONE_DS (expr)) | |
2284 | { | |
2285 | if (has_one_nonspec) | |
2286 | av_set_iter_remove (&i); | |
2287 | else | |
2288 | has_one_nonspec = true; | |
2289 | } | |
2290 | } | |
2291 | } | |
2292 | ||
2293 | /* Return the N'th element of the SET. */ | |
2294 | expr_t | |
2295 | av_set_element (av_set_t set, int n) | |
2296 | { | |
2297 | expr_t expr; | |
2298 | av_set_iterator i; | |
2299 | ||
2300 | FOR_EACH_EXPR (expr, i, set) | |
2301 | if (n-- == 0) | |
2302 | return expr; | |
2303 | ||
2304 | gcc_unreachable (); | |
2305 | return NULL; | |
2306 | } | |
2307 | ||
2308 | /* Deletes all expressions from AVP that are conditional branches (IFs). */ | |
2309 | void | |
2310 | av_set_substract_cond_branches (av_set_t *avp) | |
2311 | { | |
2312 | av_set_iterator i; | |
2313 | expr_t expr; | |
2314 | ||
2315 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2316 | if (vinsn_cond_branch_p (EXPR_VINSN (expr))) | |
2317 | av_set_iter_remove (&i); | |
2318 | } | |
2319 | ||
b8698a0f | 2320 | /* Multiplies usefulness attribute of each member of av-set *AVP by |
e855c69d AB |
2321 | value PROB / ALL_PROB. */ |
2322 | void | |
2323 | av_set_split_usefulness (av_set_t av, int prob, int all_prob) | |
2324 | { | |
2325 | av_set_iterator i; | |
2326 | expr_t expr; | |
2327 | ||
2328 | FOR_EACH_EXPR (expr, i, av) | |
b8698a0f | 2329 | EXPR_USEFULNESS (expr) = (all_prob |
e855c69d AB |
2330 | ? (EXPR_USEFULNESS (expr) * prob) / all_prob |
2331 | : 0); | |
2332 | } | |
2333 | ||
2334 | /* Leave in AVP only those expressions, which are present in AV, | |
5d369d58 | 2335 | and return it, merging history expressions. */ |
e855c69d | 2336 | void |
5d369d58 | 2337 | av_set_code_motion_filter (av_set_t *avp, av_set_t av) |
e855c69d AB |
2338 | { |
2339 | av_set_iterator i; | |
5d369d58 | 2340 | expr_t expr, expr2; |
e855c69d AB |
2341 | |
2342 | FOR_EACH_EXPR_1 (expr, i, avp) | |
5d369d58 | 2343 | if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL) |
e855c69d | 2344 | av_set_iter_remove (&i); |
5d369d58 AB |
2345 | else |
2346 | /* When updating av sets in bookkeeping blocks, we can add more insns | |
2347 | there which will be transformed but the upper av sets will not | |
2348 | reflect those transformations. We then fail to undo those | |
2349 | when searching for such insns. So merge the history saved | |
2350 | in the av set of the block we are processing. */ | |
2351 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2352 | EXPR_HISTORY_OF_CHANGES (expr2)); | |
e855c69d AB |
2353 | } |
2354 | ||
2355 | \f | |
2356 | ||
2357 | /* Dependence hooks to initialize insn data. */ | |
2358 | ||
2359 | /* This is used in hooks callable from dependence analysis when initializing | |
2360 | instruction's data. */ | |
2361 | static struct | |
2362 | { | |
2363 | /* Where the dependence was found (lhs/rhs). */ | |
2364 | deps_where_t where; | |
2365 | ||
2366 | /* The actual data object to initialize. */ | |
2367 | idata_t id; | |
2368 | ||
2369 | /* True when the insn should not be made clonable. */ | |
2370 | bool force_unique_p; | |
2371 | ||
2372 | /* True when insn should be treated as of type USE, i.e. never renamed. */ | |
2373 | bool force_use_p; | |
2374 | } deps_init_id_data; | |
2375 | ||
2376 | ||
b8698a0f | 2377 | /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be |
e855c69d AB |
2378 | clonable. */ |
2379 | static void | |
2380 | setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) | |
2381 | { | |
2382 | int type; | |
b8698a0f | 2383 | |
e855c69d AB |
2384 | /* Determine whether INSN could be cloned and return appropriate vinsn type. |
2385 | That clonable insns which can be separated into lhs and rhs have type SET. | |
2386 | Other clonable insns have type USE. */ | |
2387 | type = GET_CODE (insn); | |
2388 | ||
2389 | /* Only regular insns could be cloned. */ | |
2390 | if (type == INSN && !force_unique_p) | |
2391 | type = SET; | |
2392 | else if (type == JUMP_INSN && simplejump_p (insn)) | |
2393 | type = PC; | |
b5b8b0ac AO |
2394 | else if (type == DEBUG_INSN) |
2395 | type = !force_unique_p ? USE : INSN; | |
b8698a0f | 2396 | |
e855c69d AB |
2397 | IDATA_TYPE (id) = type; |
2398 | IDATA_REG_SETS (id) = get_clear_regset_from_pool (); | |
2399 | IDATA_REG_USES (id) = get_clear_regset_from_pool (); | |
2400 | IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); | |
2401 | } | |
2402 | ||
2403 | /* Start initializing insn data. */ | |
2404 | static void | |
2405 | deps_init_id_start_insn (insn_t insn) | |
2406 | { | |
2407 | gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); | |
2408 | ||
2409 | setup_id_for_insn (deps_init_id_data.id, insn, | |
2410 | deps_init_id_data.force_unique_p); | |
2411 | deps_init_id_data.where = DEPS_IN_INSN; | |
2412 | } | |
2413 | ||
2414 | /* Start initializing lhs data. */ | |
2415 | static void | |
2416 | deps_init_id_start_lhs (rtx lhs) | |
2417 | { | |
2418 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2419 | gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); | |
2420 | ||
2421 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2422 | { | |
2423 | IDATA_LHS (deps_init_id_data.id) = lhs; | |
2424 | deps_init_id_data.where = DEPS_IN_LHS; | |
2425 | } | |
2426 | } | |
2427 | ||
2428 | /* Finish initializing lhs data. */ | |
2429 | static void | |
2430 | deps_init_id_finish_lhs (void) | |
2431 | { | |
2432 | deps_init_id_data.where = DEPS_IN_INSN; | |
2433 | } | |
2434 | ||
2435 | /* Note a set of REGNO. */ | |
2436 | static void | |
2437 | deps_init_id_note_reg_set (int regno) | |
2438 | { | |
2439 | haifa_note_reg_set (regno); | |
2440 | ||
2441 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2442 | deps_init_id_data.force_use_p = true; | |
2443 | ||
2444 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2445 | SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); | |
2446 | ||
2447 | #ifdef STACK_REGS | |
b8698a0f | 2448 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2449 | renaming to avoid issues with find_used_regs. */ |
2450 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2451 | deps_init_id_data.force_use_p = true; | |
2452 | #endif | |
2453 | } | |
2454 | ||
2455 | /* Note a clobber of REGNO. */ | |
2456 | static void | |
2457 | deps_init_id_note_reg_clobber (int regno) | |
2458 | { | |
2459 | haifa_note_reg_clobber (regno); | |
2460 | ||
2461 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2462 | deps_init_id_data.force_use_p = true; | |
2463 | ||
2464 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2465 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); | |
2466 | } | |
2467 | ||
2468 | /* Note a use of REGNO. */ | |
2469 | static void | |
2470 | deps_init_id_note_reg_use (int regno) | |
2471 | { | |
2472 | haifa_note_reg_use (regno); | |
2473 | ||
2474 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2475 | SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); | |
2476 | } | |
2477 | ||
2478 | /* Start initializing rhs data. */ | |
2479 | static void | |
2480 | deps_init_id_start_rhs (rtx rhs) | |
2481 | { | |
2482 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2483 | ||
2484 | /* And there was no sel_deps_reset_to_insn (). */ | |
2485 | if (IDATA_LHS (deps_init_id_data.id) != NULL) | |
2486 | { | |
2487 | IDATA_RHS (deps_init_id_data.id) = rhs; | |
2488 | deps_init_id_data.where = DEPS_IN_RHS; | |
2489 | } | |
2490 | } | |
2491 | ||
2492 | /* Finish initializing rhs data. */ | |
2493 | static void | |
2494 | deps_init_id_finish_rhs (void) | |
2495 | { | |
2496 | gcc_assert (deps_init_id_data.where == DEPS_IN_RHS | |
2497 | || deps_init_id_data.where == DEPS_IN_INSN); | |
2498 | deps_init_id_data.where = DEPS_IN_INSN; | |
2499 | } | |
2500 | ||
2501 | /* Finish initializing insn data. */ | |
2502 | static void | |
2503 | deps_init_id_finish_insn (void) | |
2504 | { | |
2505 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2506 | ||
2507 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2508 | { | |
2509 | rtx lhs = IDATA_LHS (deps_init_id_data.id); | |
2510 | rtx rhs = IDATA_RHS (deps_init_id_data.id); | |
2511 | ||
2512 | if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) | |
2513 | || deps_init_id_data.force_use_p) | |
2514 | { | |
b8698a0f | 2515 | /* This should be a USE, as we don't want to schedule its RHS |
e855c69d | 2516 | separately. However, we still want to have them recorded |
b8698a0f | 2517 | for the purposes of substitution. That's why we don't |
e855c69d AB |
2518 | simply call downgrade_to_use () here. */ |
2519 | gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); | |
2520 | gcc_assert (!lhs == !rhs); | |
2521 | ||
2522 | IDATA_TYPE (deps_init_id_data.id) = USE; | |
2523 | } | |
2524 | } | |
2525 | ||
2526 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2527 | } | |
2528 | ||
2529 | /* This is dependence info used for initializing insn's data. */ | |
2530 | static struct sched_deps_info_def deps_init_id_sched_deps_info; | |
2531 | ||
2532 | /* This initializes most of the static part of the above structure. */ | |
2533 | static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = | |
2534 | { | |
2535 | NULL, | |
2536 | ||
2537 | deps_init_id_start_insn, | |
2538 | deps_init_id_finish_insn, | |
2539 | deps_init_id_start_lhs, | |
2540 | deps_init_id_finish_lhs, | |
2541 | deps_init_id_start_rhs, | |
2542 | deps_init_id_finish_rhs, | |
2543 | deps_init_id_note_reg_set, | |
2544 | deps_init_id_note_reg_clobber, | |
2545 | deps_init_id_note_reg_use, | |
2546 | NULL, /* note_mem_dep */ | |
2547 | NULL, /* note_dep */ | |
2548 | ||
2549 | 0, /* use_cselib */ | |
2550 | 0, /* use_deps_list */ | |
2551 | 0 /* generate_spec_deps */ | |
2552 | }; | |
2553 | ||
2554 | /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, | |
2555 | we don't actually need information about lhs and rhs. */ | |
2556 | static void | |
2557 | setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) | |
2558 | { | |
2559 | rtx pat = PATTERN (insn); | |
b8698a0f | 2560 | |
481683e1 | 2561 | if (NONJUMP_INSN_P (insn) |
b8698a0f | 2562 | && GET_CODE (pat) == SET |
e855c69d AB |
2563 | && !force_unique_p) |
2564 | { | |
2565 | IDATA_RHS (id) = SET_SRC (pat); | |
2566 | IDATA_LHS (id) = SET_DEST (pat); | |
2567 | } | |
2568 | else | |
2569 | IDATA_LHS (id) = IDATA_RHS (id) = NULL; | |
2570 | } | |
2571 | ||
2572 | /* Possibly downgrade INSN to USE. */ | |
2573 | static void | |
2574 | maybe_downgrade_id_to_use (idata_t id, insn_t insn) | |
2575 | { | |
2576 | bool must_be_use = false; | |
2577 | unsigned uid = INSN_UID (insn); | |
57512f53 | 2578 | df_ref *rec; |
e855c69d AB |
2579 | rtx lhs = IDATA_LHS (id); |
2580 | rtx rhs = IDATA_RHS (id); | |
b8698a0f | 2581 | |
e855c69d AB |
2582 | /* We downgrade only SETs. */ |
2583 | if (IDATA_TYPE (id) != SET) | |
2584 | return; | |
2585 | ||
2586 | if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) | |
2587 | { | |
2588 | IDATA_TYPE (id) = USE; | |
2589 | return; | |
2590 | } | |
b8698a0f | 2591 | |
e855c69d AB |
2592 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2593 | { | |
57512f53 | 2594 | df_ref def = *rec; |
b8698a0f | 2595 | |
e855c69d AB |
2596 | if (DF_REF_INSN (def) |
2597 | && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) | |
2598 | && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) | |
2599 | { | |
2600 | must_be_use = true; | |
2601 | break; | |
2602 | } | |
2603 | ||
2604 | #ifdef STACK_REGS | |
b8698a0f | 2605 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2606 | renaming to avoid issues with find_used_regs. */ |
2607 | if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) | |
2608 | { | |
2609 | must_be_use = true; | |
2610 | break; | |
2611 | } | |
2612 | #endif | |
b8698a0f L |
2613 | } |
2614 | ||
e855c69d AB |
2615 | if (must_be_use) |
2616 | IDATA_TYPE (id) = USE; | |
2617 | } | |
2618 | ||
2619 | /* Setup register sets describing INSN in ID. */ | |
2620 | static void | |
2621 | setup_id_reg_sets (idata_t id, insn_t insn) | |
2622 | { | |
2623 | unsigned uid = INSN_UID (insn); | |
57512f53 | 2624 | df_ref *rec; |
e855c69d | 2625 | regset tmp = get_clear_regset_from_pool (); |
b8698a0f | 2626 | |
e855c69d AB |
2627 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2628 | { | |
57512f53 | 2629 | df_ref def = *rec; |
e855c69d | 2630 | unsigned int regno = DF_REF_REGNO (def); |
b8698a0f | 2631 | |
e855c69d AB |
2632 | /* Post modifies are treated like clobbers by sched-deps.c. */ |
2633 | if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER | |
2634 | | DF_REF_PRE_POST_MODIFY))) | |
2635 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); | |
2636 | else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) | |
2637 | { | |
2638 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); | |
2639 | ||
2640 | #ifdef STACK_REGS | |
b8698a0f | 2641 | /* For stack registers, treat writes to them as writes |
e855c69d AB |
2642 | to the first one to be consistent with sched-deps.c. */ |
2643 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2644 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); | |
2645 | #endif | |
2646 | } | |
2647 | /* Mark special refs that generate read/write def pair. */ | |
2648 | if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) | |
2649 | || regno == STACK_POINTER_REGNUM) | |
2650 | bitmap_set_bit (tmp, regno); | |
2651 | } | |
b8698a0f | 2652 | |
e855c69d AB |
2653 | for (rec = DF_INSN_UID_USES (uid); *rec; rec++) |
2654 | { | |
57512f53 | 2655 | df_ref use = *rec; |
e855c69d AB |
2656 | unsigned int regno = DF_REF_REGNO (use); |
2657 | ||
2658 | /* When these refs are met for the first time, skip them, as | |
2659 | these uses are just counterparts of some defs. */ | |
2660 | if (bitmap_bit_p (tmp, regno)) | |
2661 | bitmap_clear_bit (tmp, regno); | |
2662 | else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) | |
2663 | { | |
2664 | SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); | |
2665 | ||
2666 | #ifdef STACK_REGS | |
b8698a0f | 2667 | /* For stack registers, treat reads from them as reads from |
e855c69d AB |
2668 | the first one to be consistent with sched-deps.c. */ |
2669 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2670 | SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); | |
2671 | #endif | |
2672 | } | |
2673 | } | |
2674 | ||
2675 | return_regset_to_pool (tmp); | |
2676 | } | |
2677 | ||
2678 | /* Initialize instruction data for INSN in ID using DF's data. */ | |
2679 | static void | |
2680 | init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) | |
2681 | { | |
2682 | gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); | |
2683 | ||
2684 | setup_id_for_insn (id, insn, force_unique_p); | |
2685 | setup_id_lhs_rhs (id, insn, force_unique_p); | |
2686 | ||
2687 | if (INSN_NOP_P (insn)) | |
2688 | return; | |
2689 | ||
2690 | maybe_downgrade_id_to_use (id, insn); | |
2691 | setup_id_reg_sets (id, insn); | |
2692 | } | |
2693 | ||
2694 | /* Initialize instruction data for INSN in ID. */ | |
2695 | static void | |
2696 | deps_init_id (idata_t id, insn_t insn, bool force_unique_p) | |
2697 | { | |
88302d54 | 2698 | struct deps_desc _dc, *dc = &_dc; |
e855c69d AB |
2699 | |
2700 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2701 | deps_init_id_data.id = id; | |
2702 | deps_init_id_data.force_unique_p = force_unique_p; | |
2703 | deps_init_id_data.force_use_p = false; | |
2704 | ||
bcf33775 | 2705 | init_deps (dc, false); |
e855c69d AB |
2706 | |
2707 | memcpy (&deps_init_id_sched_deps_info, | |
2708 | &const_deps_init_id_sched_deps_info, | |
2709 | sizeof (deps_init_id_sched_deps_info)); | |
2710 | ||
2711 | if (spec_info != NULL) | |
2712 | deps_init_id_sched_deps_info.generate_spec_deps = 1; | |
2713 | ||
2714 | sched_deps_info = &deps_init_id_sched_deps_info; | |
2715 | ||
2716 | deps_analyze_insn (dc, insn); | |
2717 | ||
2718 | free_deps (dc); | |
2719 | ||
2720 | deps_init_id_data.id = NULL; | |
2721 | } | |
2722 | ||
2723 | \f | |
2724 | ||
2725 | /* Implement hooks for collecting fundamental insn properties like if insn is | |
2726 | an ASM or is within a SCHED_GROUP. */ | |
2727 | ||
2728 | /* True when a "one-time init" data for INSN was already inited. */ | |
2729 | static bool | |
2730 | first_time_insn_init (insn_t insn) | |
2731 | { | |
2732 | return INSN_LIVE (insn) == NULL; | |
2733 | } | |
2734 | ||
2735 | /* Hash an entry in a transformed_insns hashtable. */ | |
2736 | static hashval_t | |
2737 | hash_transformed_insns (const void *p) | |
2738 | { | |
2739 | return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2740 | } | |
2741 | ||
2742 | /* Compare the entries in a transformed_insns hashtable. */ | |
2743 | static int | |
2744 | eq_transformed_insns (const void *p, const void *q) | |
2745 | { | |
2746 | rtx i1 = VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2747 | rtx i2 = VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); | |
2748 | ||
2749 | if (INSN_UID (i1) == INSN_UID (i2)) | |
2750 | return 1; | |
2751 | return rtx_equal_p (PATTERN (i1), PATTERN (i2)); | |
2752 | } | |
2753 | ||
2754 | /* Free an entry in a transformed_insns hashtable. */ | |
2755 | static void | |
2756 | free_transformed_insns (void *p) | |
2757 | { | |
2758 | struct transformed_insns *pti = (struct transformed_insns *) p; | |
2759 | ||
2760 | vinsn_detach (pti->vinsn_old); | |
2761 | vinsn_detach (pti->vinsn_new); | |
2762 | free (pti); | |
2763 | } | |
2764 | ||
b8698a0f | 2765 | /* Init the s_i_d data for INSN which should be inited just once, when |
e855c69d AB |
2766 | we first see the insn. */ |
2767 | static void | |
2768 | init_first_time_insn_data (insn_t insn) | |
2769 | { | |
2770 | /* This should not be set if this is the first time we init data for | |
2771 | insn. */ | |
2772 | gcc_assert (first_time_insn_init (insn)); | |
b8698a0f | 2773 | |
e855c69d AB |
2774 | /* These are needed for nops too. */ |
2775 | INSN_LIVE (insn) = get_regset_from_pool (); | |
2776 | INSN_LIVE_VALID_P (insn) = false; | |
bcf33775 | 2777 | |
e855c69d AB |
2778 | if (!INSN_NOP_P (insn)) |
2779 | { | |
2780 | INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); | |
2781 | INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); | |
b8698a0f | 2782 | INSN_TRANSFORMED_INSNS (insn) |
e855c69d AB |
2783 | = htab_create (16, hash_transformed_insns, |
2784 | eq_transformed_insns, free_transformed_insns); | |
bcf33775 | 2785 | init_deps (&INSN_DEPS_CONTEXT (insn), true); |
e855c69d AB |
2786 | } |
2787 | } | |
2788 | ||
b8698a0f | 2789 | /* Free almost all above data for INSN that is scheduled already. |
bcf33775 AB |
2790 | Used for extra-large basic blocks. */ |
2791 | void | |
2792 | free_data_for_scheduled_insn (insn_t insn) | |
e855c69d AB |
2793 | { |
2794 | gcc_assert (! first_time_insn_init (insn)); | |
b8698a0f | 2795 | |
bcf33775 AB |
2796 | if (! INSN_ANALYZED_DEPS (insn)) |
2797 | return; | |
b8698a0f | 2798 | |
e855c69d AB |
2799 | BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); |
2800 | BITMAP_FREE (INSN_FOUND_DEPS (insn)); | |
2801 | htab_delete (INSN_TRANSFORMED_INSNS (insn)); | |
b8698a0f | 2802 | |
e855c69d AB |
2803 | /* This is allocated only for bookkeeping insns. */ |
2804 | if (INSN_ORIGINATORS (insn)) | |
2805 | BITMAP_FREE (INSN_ORIGINATORS (insn)); | |
2806 | free_deps (&INSN_DEPS_CONTEXT (insn)); | |
bcf33775 AB |
2807 | |
2808 | INSN_ANALYZED_DEPS (insn) = NULL; | |
2809 | ||
b8698a0f | 2810 | /* Clear the readonly flag so we would ICE when trying to recalculate |
bcf33775 AB |
2811 | the deps context (as we believe that it should not happen). */ |
2812 | (&INSN_DEPS_CONTEXT (insn))->readonly = 0; | |
2813 | } | |
2814 | ||
2815 | /* Free the same data as above for INSN. */ | |
2816 | static void | |
2817 | free_first_time_insn_data (insn_t insn) | |
2818 | { | |
2819 | gcc_assert (! first_time_insn_init (insn)); | |
2820 | ||
2821 | free_data_for_scheduled_insn (insn); | |
2822 | return_regset_to_pool (INSN_LIVE (insn)); | |
2823 | INSN_LIVE (insn) = NULL; | |
2824 | INSN_LIVE_VALID_P (insn) = false; | |
e855c69d AB |
2825 | } |
2826 | ||
2827 | /* Initialize region-scope data structures for basic blocks. */ | |
2828 | static void | |
2829 | init_global_and_expr_for_bb (basic_block bb) | |
2830 | { | |
2831 | if (sel_bb_empty_p (bb)) | |
2832 | return; | |
2833 | ||
2834 | invalidate_av_set (bb); | |
2835 | } | |
2836 | ||
2837 | /* Data for global dependency analysis (to initialize CANT_MOVE and | |
2838 | SCHED_GROUP_P). */ | |
2839 | static struct | |
2840 | { | |
2841 | /* Previous insn. */ | |
2842 | insn_t prev_insn; | |
2843 | } init_global_data; | |
2844 | ||
2845 | /* Determine if INSN is in the sched_group, is an asm or should not be | |
2846 | cloned. After that initialize its expr. */ | |
2847 | static void | |
2848 | init_global_and_expr_for_insn (insn_t insn) | |
2849 | { | |
2850 | if (LABEL_P (insn)) | |
2851 | return; | |
2852 | ||
2853 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2854 | { | |
2855 | init_global_data.prev_insn = NULL_RTX; | |
2856 | return; | |
2857 | } | |
2858 | ||
2859 | gcc_assert (INSN_P (insn)); | |
2860 | ||
2861 | if (SCHED_GROUP_P (insn)) | |
2862 | /* Setup a sched_group. */ | |
2863 | { | |
2864 | insn_t prev_insn = init_global_data.prev_insn; | |
2865 | ||
2866 | if (prev_insn) | |
2867 | INSN_SCHED_NEXT (prev_insn) = insn; | |
2868 | ||
2869 | init_global_data.prev_insn = insn; | |
2870 | } | |
2871 | else | |
2872 | init_global_data.prev_insn = NULL_RTX; | |
2873 | ||
2874 | if (GET_CODE (PATTERN (insn)) == ASM_INPUT | |
2875 | || asm_noperands (PATTERN (insn)) >= 0) | |
2876 | /* Mark INSN as an asm. */ | |
2877 | INSN_ASM_P (insn) = true; | |
2878 | ||
2879 | { | |
2880 | bool force_unique_p; | |
2881 | ds_t spec_done_ds; | |
2882 | ||
cfeb0fa8 AB |
2883 | /* Certain instructions cannot be cloned, and frame related insns and |
2884 | the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of | |
2885 | their block. */ | |
2886 | if (prologue_epilogue_contains (insn)) | |
2887 | { | |
2888 | if (RTX_FRAME_RELATED_P (insn)) | |
2889 | CANT_MOVE (insn) = 1; | |
2890 | else | |
2891 | { | |
2892 | rtx note; | |
2893 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
2894 | if (REG_NOTE_KIND (note) == REG_SAVE_NOTE | |
2895 | && ((enum insn_note) INTVAL (XEXP (note, 0)) | |
2896 | == NOTE_INSN_EPILOGUE_BEG)) | |
2897 | { | |
2898 | CANT_MOVE (insn) = 1; | |
2899 | break; | |
2900 | } | |
2901 | } | |
2902 | force_unique_p = true; | |
2903 | } | |
e855c69d | 2904 | else |
cfeb0fa8 AB |
2905 | if (CANT_MOVE (insn) |
2906 | || INSN_ASM_P (insn) | |
2907 | || SCHED_GROUP_P (insn) | |
07643d76 | 2908 | || CALL_P (insn) |
cfeb0fa8 AB |
2909 | /* Exception handling insns are always unique. */ |
2910 | || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) | |
2911 | /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ | |
2912 | || control_flow_insn_p (insn)) | |
2913 | force_unique_p = true; | |
2914 | else | |
2915 | force_unique_p = false; | |
e855c69d AB |
2916 | |
2917 | if (targetm.sched.get_insn_spec_ds) | |
2918 | { | |
2919 | spec_done_ds = targetm.sched.get_insn_spec_ds (insn); | |
2920 | spec_done_ds = ds_get_max_dep_weak (spec_done_ds); | |
2921 | } | |
2922 | else | |
2923 | spec_done_ds = 0; | |
2924 | ||
2925 | /* Initialize INSN's expr. */ | |
2926 | init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0, | |
2927 | REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn), | |
b8698a0f | 2928 | spec_done_ds, 0, 0, NULL, true, false, false, false, |
e855c69d AB |
2929 | CANT_MOVE (insn)); |
2930 | } | |
2931 | ||
2932 | init_first_time_insn_data (insn); | |
2933 | } | |
2934 | ||
2935 | /* Scan the region and initialize instruction data for basic blocks BBS. */ | |
2936 | void | |
2937 | sel_init_global_and_expr (bb_vec_t bbs) | |
2938 | { | |
2939 | /* ??? It would be nice to implement push / pop scheme for sched_infos. */ | |
2940 | const struct sched_scan_info_def ssi = | |
2941 | { | |
2942 | NULL, /* extend_bb */ | |
2943 | init_global_and_expr_for_bb, /* init_bb */ | |
2944 | extend_insn_data, /* extend_insn */ | |
2945 | init_global_and_expr_for_insn /* init_insn */ | |
2946 | }; | |
b8698a0f | 2947 | |
e855c69d AB |
2948 | sched_scan (&ssi, bbs, NULL, NULL, NULL); |
2949 | } | |
2950 | ||
2951 | /* Finalize region-scope data structures for basic blocks. */ | |
2952 | static void | |
2953 | finish_global_and_expr_for_bb (basic_block bb) | |
2954 | { | |
2955 | av_set_clear (&BB_AV_SET (bb)); | |
2956 | BB_AV_LEVEL (bb) = 0; | |
2957 | } | |
2958 | ||
2959 | /* Finalize INSN's data. */ | |
2960 | static void | |
2961 | finish_global_and_expr_insn (insn_t insn) | |
2962 | { | |
2963 | if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2964 | return; | |
2965 | ||
2966 | gcc_assert (INSN_P (insn)); | |
2967 | ||
2968 | if (INSN_LUID (insn) > 0) | |
2969 | { | |
2970 | free_first_time_insn_data (insn); | |
2971 | INSN_WS_LEVEL (insn) = 0; | |
2972 | CANT_MOVE (insn) = 0; | |
b8698a0f L |
2973 | |
2974 | /* We can no longer assert this, as vinsns of this insn could be | |
2975 | easily live in other insn's caches. This should be changed to | |
e855c69d AB |
2976 | a counter-like approach among all vinsns. */ |
2977 | gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); | |
2978 | clear_expr (INSN_EXPR (insn)); | |
2979 | } | |
2980 | } | |
2981 | ||
2982 | /* Finalize per instruction data for the whole region. */ | |
2983 | void | |
2984 | sel_finish_global_and_expr (void) | |
2985 | { | |
2986 | { | |
2987 | bb_vec_t bbs; | |
2988 | int i; | |
2989 | ||
2990 | bbs = VEC_alloc (basic_block, heap, current_nr_blocks); | |
2991 | ||
2992 | for (i = 0; i < current_nr_blocks; i++) | |
2993 | VEC_quick_push (basic_block, bbs, BASIC_BLOCK (BB_TO_BLOCK (i))); | |
2994 | ||
2995 | /* Clear AV_SETs and INSN_EXPRs. */ | |
2996 | { | |
2997 | const struct sched_scan_info_def ssi = | |
2998 | { | |
2999 | NULL, /* extend_bb */ | |
3000 | finish_global_and_expr_for_bb, /* init_bb */ | |
3001 | NULL, /* extend_insn */ | |
3002 | finish_global_and_expr_insn /* init_insn */ | |
3003 | }; | |
3004 | ||
3005 | sched_scan (&ssi, bbs, NULL, NULL, NULL); | |
3006 | } | |
3007 | ||
3008 | VEC_free (basic_block, heap, bbs); | |
3009 | } | |
3010 | ||
3011 | finish_insns (); | |
3012 | } | |
3013 | \f | |
3014 | ||
b8698a0f L |
3015 | /* In the below hooks, we merely calculate whether or not a dependence |
3016 | exists, and in what part of insn. However, we will need more data | |
e855c69d AB |
3017 | when we'll start caching dependence requests. */ |
3018 | ||
3019 | /* Container to hold information for dependency analysis. */ | |
3020 | static struct | |
3021 | { | |
3022 | deps_t dc; | |
3023 | ||
3024 | /* A variable to track which part of rtx we are scanning in | |
3025 | sched-deps.c: sched_analyze_insn (). */ | |
3026 | deps_where_t where; | |
3027 | ||
3028 | /* Current producer. */ | |
3029 | insn_t pro; | |
3030 | ||
3031 | /* Current consumer. */ | |
3032 | vinsn_t con; | |
3033 | ||
3034 | /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. | |
3035 | X is from { INSN, LHS, RHS }. */ | |
3036 | ds_t has_dep_p[DEPS_IN_NOWHERE]; | |
3037 | } has_dependence_data; | |
3038 | ||
3039 | /* Start analyzing dependencies of INSN. */ | |
3040 | static void | |
3041 | has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) | |
3042 | { | |
3043 | gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); | |
3044 | ||
3045 | has_dependence_data.where = DEPS_IN_INSN; | |
3046 | } | |
3047 | ||
3048 | /* Finish analyzing dependencies of an insn. */ | |
3049 | static void | |
3050 | has_dependence_finish_insn (void) | |
3051 | { | |
3052 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3053 | ||
3054 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3055 | } | |
3056 | ||
3057 | /* Start analyzing dependencies of LHS. */ | |
3058 | static void | |
3059 | has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) | |
3060 | { | |
3061 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3062 | ||
3063 | if (VINSN_LHS (has_dependence_data.con) != NULL) | |
3064 | has_dependence_data.where = DEPS_IN_LHS; | |
3065 | } | |
3066 | ||
3067 | /* Finish analyzing dependencies of an lhs. */ | |
3068 | static void | |
3069 | has_dependence_finish_lhs (void) | |
3070 | { | |
3071 | has_dependence_data.where = DEPS_IN_INSN; | |
3072 | } | |
3073 | ||
3074 | /* Start analyzing dependencies of RHS. */ | |
3075 | static void | |
3076 | has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) | |
3077 | { | |
3078 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3079 | ||
3080 | if (VINSN_RHS (has_dependence_data.con) != NULL) | |
3081 | has_dependence_data.where = DEPS_IN_RHS; | |
3082 | } | |
3083 | ||
3084 | /* Start analyzing dependencies of an rhs. */ | |
3085 | static void | |
3086 | has_dependence_finish_rhs (void) | |
3087 | { | |
3088 | gcc_assert (has_dependence_data.where == DEPS_IN_RHS | |
3089 | || has_dependence_data.where == DEPS_IN_INSN); | |
3090 | ||
3091 | has_dependence_data.where = DEPS_IN_INSN; | |
3092 | } | |
3093 | ||
3094 | /* Note a set of REGNO. */ | |
3095 | static void | |
3096 | has_dependence_note_reg_set (int regno) | |
3097 | { | |
3098 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3099 | ||
3100 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3101 | VINSN_INSN_RTX | |
3102 | (has_dependence_data.con))) | |
3103 | { | |
3104 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3105 | ||
3106 | if (reg_last->sets != NULL | |
3107 | || reg_last->clobbers != NULL) | |
3108 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
3109 | ||
3110 | if (reg_last->uses) | |
3111 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3112 | } | |
3113 | } | |
3114 | ||
3115 | /* Note a clobber of REGNO. */ | |
3116 | static void | |
3117 | has_dependence_note_reg_clobber (int regno) | |
3118 | { | |
3119 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3120 | ||
3121 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3122 | VINSN_INSN_RTX | |
3123 | (has_dependence_data.con))) | |
3124 | { | |
3125 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3126 | ||
3127 | if (reg_last->sets) | |
3128 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
b8698a0f | 3129 | |
e855c69d AB |
3130 | if (reg_last->uses) |
3131 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3132 | } | |
3133 | } | |
3134 | ||
3135 | /* Note a use of REGNO. */ | |
3136 | static void | |
3137 | has_dependence_note_reg_use (int regno) | |
3138 | { | |
3139 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3140 | ||
3141 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3142 | VINSN_INSN_RTX | |
3143 | (has_dependence_data.con))) | |
3144 | { | |
3145 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3146 | ||
3147 | if (reg_last->sets) | |
3148 | *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; | |
3149 | ||
3150 | if (reg_last->clobbers) | |
3151 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3152 | ||
3153 | /* Handle BE_IN_SPEC. */ | |
3154 | if (reg_last->uses) | |
3155 | { | |
3156 | ds_t pro_spec_checked_ds; | |
3157 | ||
3158 | pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); | |
3159 | pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); | |
3160 | ||
3161 | if (pro_spec_checked_ds != 0) | |
3162 | /* Merge BE_IN_SPEC bits into *DSP. */ | |
3163 | *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, | |
3164 | NULL_RTX, NULL_RTX); | |
3165 | } | |
3166 | } | |
3167 | } | |
3168 | ||
3169 | /* Note a memory dependence. */ | |
3170 | static void | |
3171 | has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, | |
3172 | rtx pending_mem ATTRIBUTE_UNUSED, | |
3173 | insn_t pending_insn ATTRIBUTE_UNUSED, | |
3174 | ds_t ds ATTRIBUTE_UNUSED) | |
3175 | { | |
3176 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3177 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3178 | { | |
3179 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3180 | ||
3181 | *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); | |
3182 | } | |
3183 | } | |
3184 | ||
3185 | /* Note a dependence. */ | |
3186 | static void | |
3187 | has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED, | |
3188 | ds_t ds ATTRIBUTE_UNUSED) | |
3189 | { | |
3190 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3191 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3192 | { | |
3193 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3194 | ||
3195 | *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); | |
3196 | } | |
3197 | } | |
3198 | ||
3199 | /* Mark the insn as having a hard dependence that prevents speculation. */ | |
3200 | void | |
3201 | sel_mark_hard_insn (rtx insn) | |
3202 | { | |
3203 | int i; | |
3204 | ||
3205 | /* Only work when we're in has_dependence_p mode. | |
3206 | ??? This is a hack, this should actually be a hook. */ | |
3207 | if (!has_dependence_data.dc || !has_dependence_data.pro) | |
3208 | return; | |
3209 | ||
3210 | gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); | |
3211 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3212 | ||
3213 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3214 | has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; | |
3215 | } | |
3216 | ||
3217 | /* This structure holds the hooks for the dependency analysis used when | |
3218 | actually processing dependencies in the scheduler. */ | |
3219 | static struct sched_deps_info_def has_dependence_sched_deps_info; | |
3220 | ||
3221 | /* This initializes most of the fields of the above structure. */ | |
3222 | static const struct sched_deps_info_def const_has_dependence_sched_deps_info = | |
3223 | { | |
3224 | NULL, | |
3225 | ||
3226 | has_dependence_start_insn, | |
3227 | has_dependence_finish_insn, | |
3228 | has_dependence_start_lhs, | |
3229 | has_dependence_finish_lhs, | |
3230 | has_dependence_start_rhs, | |
3231 | has_dependence_finish_rhs, | |
3232 | has_dependence_note_reg_set, | |
3233 | has_dependence_note_reg_clobber, | |
3234 | has_dependence_note_reg_use, | |
3235 | has_dependence_note_mem_dep, | |
3236 | has_dependence_note_dep, | |
3237 | ||
3238 | 0, /* use_cselib */ | |
3239 | 0, /* use_deps_list */ | |
3240 | 0 /* generate_spec_deps */ | |
3241 | }; | |
3242 | ||
3243 | /* Initialize has_dependence_sched_deps_info with extra spec field. */ | |
3244 | static void | |
3245 | setup_has_dependence_sched_deps_info (void) | |
3246 | { | |
3247 | memcpy (&has_dependence_sched_deps_info, | |
3248 | &const_has_dependence_sched_deps_info, | |
3249 | sizeof (has_dependence_sched_deps_info)); | |
3250 | ||
3251 | if (spec_info != NULL) | |
3252 | has_dependence_sched_deps_info.generate_spec_deps = 1; | |
3253 | ||
3254 | sched_deps_info = &has_dependence_sched_deps_info; | |
3255 | } | |
3256 | ||
3257 | /* Remove all dependences found and recorded in has_dependence_data array. */ | |
3258 | void | |
3259 | sel_clear_has_dependence (void) | |
3260 | { | |
3261 | int i; | |
3262 | ||
3263 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3264 | has_dependence_data.has_dep_p[i] = 0; | |
3265 | } | |
3266 | ||
3267 | /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer | |
3268 | to the dependence information array in HAS_DEP_PP. */ | |
3269 | ds_t | |
3270 | has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) | |
3271 | { | |
3272 | int i; | |
3273 | ds_t ds; | |
88302d54 | 3274 | struct deps_desc *dc; |
e855c69d AB |
3275 | |
3276 | if (INSN_SIMPLEJUMP_P (pred)) | |
3277 | /* Unconditional jump is just a transfer of control flow. | |
3278 | Ignore it. */ | |
3279 | return false; | |
3280 | ||
3281 | dc = &INSN_DEPS_CONTEXT (pred); | |
bcf33775 AB |
3282 | |
3283 | /* We init this field lazily. */ | |
3284 | if (dc->reg_last == NULL) | |
3285 | init_deps_reg_last (dc); | |
b8698a0f | 3286 | |
e855c69d AB |
3287 | if (!dc->readonly) |
3288 | { | |
3289 | has_dependence_data.pro = NULL; | |
3290 | /* Initialize empty dep context with information about PRED. */ | |
3291 | advance_deps_context (dc, pred); | |
3292 | dc->readonly = 1; | |
3293 | } | |
3294 | ||
3295 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3296 | has_dependence_data.pro = pred; | |
3297 | has_dependence_data.con = EXPR_VINSN (expr); | |
3298 | has_dependence_data.dc = dc; | |
3299 | ||
3300 | sel_clear_has_dependence (); | |
3301 | ||
3302 | /* Now catch all dependencies that would be generated between PRED and | |
3303 | INSN. */ | |
3304 | setup_has_dependence_sched_deps_info (); | |
3305 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3306 | has_dependence_data.dc = NULL; | |
3307 | ||
3308 | /* When a barrier was found, set DEPS_IN_INSN bits. */ | |
3309 | if (dc->last_reg_pending_barrier == TRUE_BARRIER) | |
3310 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; | |
3311 | else if (dc->last_reg_pending_barrier == MOVE_BARRIER) | |
3312 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
3313 | ||
3314 | /* Do not allow stores to memory to move through checks. Currently | |
3315 | we don't move this to sched-deps.c as the check doesn't have | |
b8698a0f | 3316 | obvious places to which this dependence can be attached. |
e855c69d AB |
3317 | FIMXE: this should go to a hook. */ |
3318 | if (EXPR_LHS (expr) | |
3319 | && MEM_P (EXPR_LHS (expr)) | |
3320 | && sel_insn_is_speculation_check (pred)) | |
3321 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
b8698a0f | 3322 | |
e855c69d AB |
3323 | *has_dep_pp = has_dependence_data.has_dep_p; |
3324 | ds = 0; | |
3325 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3326 | ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], | |
3327 | NULL_RTX, NULL_RTX); | |
3328 | ||
3329 | return ds; | |
3330 | } | |
3331 | \f | |
3332 | ||
b8698a0f L |
3333 | /* Dependence hooks implementation that checks dependence latency constraints |
3334 | on the insns being scheduled. The entry point for these routines is | |
3335 | tick_check_p predicate. */ | |
e855c69d AB |
3336 | |
3337 | static struct | |
3338 | { | |
3339 | /* An expr we are currently checking. */ | |
3340 | expr_t expr; | |
3341 | ||
3342 | /* A minimal cycle for its scheduling. */ | |
3343 | int cycle; | |
3344 | ||
3345 | /* Whether we have seen a true dependence while checking. */ | |
3346 | bool seen_true_dep_p; | |
3347 | } tick_check_data; | |
3348 | ||
3349 | /* Update minimal scheduling cycle for tick_check_insn given that it depends | |
3350 | on PRO with status DS and weight DW. */ | |
3351 | static void | |
3352 | tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) | |
3353 | { | |
3354 | expr_t con_expr = tick_check_data.expr; | |
3355 | insn_t con_insn = EXPR_INSN_RTX (con_expr); | |
3356 | ||
3357 | if (con_insn != pro_insn) | |
3358 | { | |
3359 | enum reg_note dt; | |
3360 | int tick; | |
3361 | ||
3362 | if (/* PROducer was removed from above due to pipelining. */ | |
3363 | !INSN_IN_STREAM_P (pro_insn) | |
3364 | /* Or PROducer was originally on the next iteration regarding the | |
3365 | CONsumer. */ | |
3366 | || (INSN_SCHED_TIMES (pro_insn) | |
3367 | - EXPR_SCHED_TIMES (con_expr)) > 1) | |
3368 | /* Don't count this dependence. */ | |
3369 | return; | |
3370 | ||
3371 | dt = ds_to_dt (ds); | |
3372 | if (dt == REG_DEP_TRUE) | |
3373 | tick_check_data.seen_true_dep_p = true; | |
3374 | ||
3375 | gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); | |
3376 | ||
3377 | { | |
3378 | dep_def _dep, *dep = &_dep; | |
3379 | ||
3380 | init_dep (dep, pro_insn, con_insn, dt); | |
3381 | ||
3382 | tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); | |
3383 | } | |
3384 | ||
3385 | /* When there are several kinds of dependencies between pro and con, | |
3386 | only REG_DEP_TRUE should be taken into account. */ | |
3387 | if (tick > tick_check_data.cycle | |
3388 | && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) | |
3389 | tick_check_data.cycle = tick; | |
3390 | } | |
3391 | } | |
3392 | ||
3393 | /* An implementation of note_dep hook. */ | |
3394 | static void | |
3395 | tick_check_note_dep (insn_t pro, ds_t ds) | |
3396 | { | |
3397 | tick_check_dep_with_dw (pro, ds, 0); | |
3398 | } | |
3399 | ||
3400 | /* An implementation of note_mem_dep hook. */ | |
3401 | static void | |
3402 | tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) | |
3403 | { | |
3404 | dw_t dw; | |
3405 | ||
3406 | dw = (ds_to_dt (ds) == REG_DEP_TRUE | |
3407 | ? estimate_dep_weak (mem1, mem2) | |
3408 | : 0); | |
3409 | ||
3410 | tick_check_dep_with_dw (pro, ds, dw); | |
3411 | } | |
3412 | ||
3413 | /* This structure contains hooks for dependence analysis used when determining | |
3414 | whether an insn is ready for scheduling. */ | |
3415 | static struct sched_deps_info_def tick_check_sched_deps_info = | |
3416 | { | |
3417 | NULL, | |
3418 | ||
3419 | NULL, | |
3420 | NULL, | |
3421 | NULL, | |
3422 | NULL, | |
3423 | NULL, | |
3424 | NULL, | |
3425 | haifa_note_reg_set, | |
3426 | haifa_note_reg_clobber, | |
3427 | haifa_note_reg_use, | |
3428 | tick_check_note_mem_dep, | |
3429 | tick_check_note_dep, | |
3430 | ||
3431 | 0, 0, 0 | |
3432 | }; | |
3433 | ||
3434 | /* Estimate number of cycles from the current cycle of FENCE until EXPR can be | |
3435 | scheduled. Return 0 if all data from producers in DC is ready. */ | |
3436 | int | |
3437 | tick_check_p (expr_t expr, deps_t dc, fence_t fence) | |
3438 | { | |
3439 | int cycles_left; | |
3440 | /* Initialize variables. */ | |
3441 | tick_check_data.expr = expr; | |
3442 | tick_check_data.cycle = 0; | |
3443 | tick_check_data.seen_true_dep_p = false; | |
3444 | sched_deps_info = &tick_check_sched_deps_info; | |
b8698a0f | 3445 | |
e855c69d AB |
3446 | gcc_assert (!dc->readonly); |
3447 | dc->readonly = 1; | |
3448 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3449 | dc->readonly = 0; | |
3450 | ||
3451 | cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); | |
3452 | ||
3453 | return cycles_left >= 0 ? cycles_left : 0; | |
3454 | } | |
3455 | \f | |
3456 | ||
3457 | /* Functions to work with insns. */ | |
3458 | ||
3459 | /* Returns true if LHS of INSN is the same as DEST of an insn | |
3460 | being moved. */ | |
3461 | bool | |
3462 | lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) | |
3463 | { | |
3464 | rtx lhs = INSN_LHS (insn); | |
3465 | ||
3466 | if (lhs == NULL || dest == NULL) | |
3467 | return false; | |
b8698a0f | 3468 | |
e855c69d AB |
3469 | return rtx_equal_p (lhs, dest); |
3470 | } | |
3471 | ||
3472 | /* Return s_i_d entry of INSN. Callable from debugger. */ | |
3473 | sel_insn_data_def | |
3474 | insn_sid (insn_t insn) | |
3475 | { | |
3476 | return *SID (insn); | |
3477 | } | |
3478 | ||
3479 | /* True when INSN is a speculative check. We can tell this by looking | |
3480 | at the data structures of the selective scheduler, not by examining | |
3481 | the pattern. */ | |
3482 | bool | |
3483 | sel_insn_is_speculation_check (rtx insn) | |
3484 | { | |
3485 | return s_i_d && !! INSN_SPEC_CHECKED_DS (insn); | |
3486 | } | |
3487 | ||
b8698a0f | 3488 | /* Extracts machine mode MODE and destination location DST_LOC |
e855c69d AB |
3489 | for given INSN. */ |
3490 | void | |
3491 | get_dest_and_mode (rtx insn, rtx *dst_loc, enum machine_mode *mode) | |
3492 | { | |
3493 | rtx pat = PATTERN (insn); | |
3494 | ||
3495 | gcc_assert (dst_loc); | |
3496 | gcc_assert (GET_CODE (pat) == SET); | |
3497 | ||
3498 | *dst_loc = SET_DEST (pat); | |
3499 | ||
3500 | gcc_assert (*dst_loc); | |
3501 | gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); | |
3502 | ||
3503 | if (mode) | |
3504 | *mode = GET_MODE (*dst_loc); | |
3505 | } | |
3506 | ||
b8698a0f | 3507 | /* Returns true when moving through JUMP will result in bookkeeping |
e855c69d AB |
3508 | creation. */ |
3509 | bool | |
3510 | bookkeeping_can_be_created_if_moved_through_p (insn_t jump) | |
3511 | { | |
3512 | insn_t succ; | |
3513 | succ_iterator si; | |
3514 | ||
3515 | FOR_EACH_SUCC (succ, si, jump) | |
3516 | if (sel_num_cfg_preds_gt_1 (succ)) | |
3517 | return true; | |
3518 | ||
3519 | return false; | |
3520 | } | |
3521 | ||
3522 | /* Return 'true' if INSN is the only one in its basic block. */ | |
3523 | static bool | |
3524 | insn_is_the_only_one_in_bb_p (insn_t insn) | |
3525 | { | |
3526 | return sel_bb_head_p (insn) && sel_bb_end_p (insn); | |
3527 | } | |
3528 | ||
3529 | #ifdef ENABLE_CHECKING | |
b8698a0f | 3530 | /* Check that the region we're scheduling still has at most one |
e855c69d AB |
3531 | backedge. */ |
3532 | static void | |
3533 | verify_backedges (void) | |
3534 | { | |
3535 | if (pipelining_p) | |
3536 | { | |
3537 | int i, n = 0; | |
3538 | edge e; | |
3539 | edge_iterator ei; | |
b8698a0f | 3540 | |
e855c69d AB |
3541 | for (i = 0; i < current_nr_blocks; i++) |
3542 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (BB_TO_BLOCK (i))->succs) | |
3543 | if (in_current_region_p (e->dest) | |
3544 | && BLOCK_TO_BB (e->dest->index) < i) | |
3545 | n++; | |
b8698a0f | 3546 | |
e855c69d AB |
3547 | gcc_assert (n <= 1); |
3548 | } | |
3549 | } | |
3550 | #endif | |
3551 | \f | |
3552 | ||
3553 | /* Functions to work with control flow. */ | |
3554 | ||
b59ab570 AM |
3555 | /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks |
3556 | are sorted in topological order (it might have been invalidated by | |
3557 | redirecting an edge). */ | |
3558 | static void | |
3559 | sel_recompute_toporder (void) | |
3560 | { | |
3561 | int i, n, rgn; | |
3562 | int *postorder, n_blocks; | |
3563 | ||
3564 | postorder = XALLOCAVEC (int, n_basic_blocks); | |
3565 | n_blocks = post_order_compute (postorder, false, false); | |
3566 | ||
3567 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
3568 | for (n = 0, i = n_blocks - 1; i >= 0; i--) | |
3569 | if (CONTAINING_RGN (postorder[i]) == rgn) | |
3570 | { | |
3571 | BLOCK_TO_BB (postorder[i]) = n; | |
3572 | BB_TO_BLOCK (n) = postorder[i]; | |
3573 | n++; | |
3574 | } | |
3575 | ||
3576 | /* Assert that we updated info for all blocks. We may miss some blocks if | |
3577 | this function is called when redirecting an edge made a block | |
3578 | unreachable, but that block is not deleted yet. */ | |
3579 | gcc_assert (n == RGN_NR_BLOCKS (rgn)); | |
3580 | } | |
3581 | ||
e855c69d | 3582 | /* Tidy the possibly empty block BB. */ |
65592aad | 3583 | static bool |
5f33b972 | 3584 | maybe_tidy_empty_bb (basic_block bb) |
e855c69d AB |
3585 | { |
3586 | basic_block succ_bb, pred_bb; | |
00c4e97c | 3587 | VEC (basic_block, heap) *dom_bbs; |
f2c45f08 AM |
3588 | edge e; |
3589 | edge_iterator ei; | |
e855c69d AB |
3590 | bool rescan_p; |
3591 | ||
3592 | /* Keep empty bb only if this block immediately precedes EXIT and | |
762bffba AB |
3593 | has incoming non-fallthrough edge, or it has no predecessors or |
3594 | successors. Otherwise remove it. */ | |
b5b8b0ac | 3595 | if (!sel_bb_empty_p (bb) |
b8698a0f | 3596 | || (single_succ_p (bb) |
e855c69d | 3597 | && single_succ (bb) == EXIT_BLOCK_PTR |
b8698a0f | 3598 | && (!single_pred_p (bb) |
762bffba AB |
3599 | || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU))) |
3600 | || EDGE_COUNT (bb->preds) == 0 | |
3601 | || EDGE_COUNT (bb->succs) == 0) | |
e855c69d AB |
3602 | return false; |
3603 | ||
f2c45f08 AM |
3604 | /* Do not attempt to redirect complex edges. */ |
3605 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3606 | if (e->flags & EDGE_COMPLEX) | |
3607 | return false; | |
3608 | ||
e855c69d AB |
3609 | free_data_sets (bb); |
3610 | ||
3611 | /* Do not delete BB if it has more than one successor. | |
3612 | That can occur when we moving a jump. */ | |
3613 | if (!single_succ_p (bb)) | |
3614 | { | |
3615 | gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); | |
3616 | sel_merge_blocks (bb->prev_bb, bb); | |
3617 | return true; | |
3618 | } | |
3619 | ||
3620 | succ_bb = single_succ (bb); | |
3621 | rescan_p = true; | |
3622 | pred_bb = NULL; | |
00c4e97c | 3623 | dom_bbs = NULL; |
e855c69d AB |
3624 | |
3625 | /* Redirect all non-fallthru edges to the next bb. */ | |
3626 | while (rescan_p) | |
3627 | { | |
e855c69d AB |
3628 | rescan_p = false; |
3629 | ||
3630 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3631 | { | |
3632 | pred_bb = e->src; | |
3633 | ||
3634 | if (!(e->flags & EDGE_FALLTHRU)) | |
3635 | { | |
5f33b972 | 3636 | /* We can not invalidate computed topological order by moving |
00c4e97c AB |
3637 | the edge destination block (E->SUCC) along a fallthru edge. |
3638 | ||
3639 | We will update dominators here only when we'll get | |
3640 | an unreachable block when redirecting, otherwise | |
3641 | sel_redirect_edge_and_branch will take care of it. */ | |
3642 | if (e->dest != bb | |
3643 | && single_pred_p (e->dest)) | |
3644 | VEC_safe_push (basic_block, heap, dom_bbs, e->dest); | |
5f33b972 | 3645 | sel_redirect_edge_and_branch (e, succ_bb); |
e855c69d AB |
3646 | rescan_p = true; |
3647 | break; | |
3648 | } | |
5f33b972 AM |
3649 | /* If the edge is fallthru, but PRED_BB ends in a conditional jump |
3650 | to BB (so there is no non-fallthru edge from PRED_BB to BB), we | |
3651 | still have to adjust it. */ | |
3652 | else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb))) | |
3653 | { | |
3654 | /* If possible, try to remove the unneeded conditional jump. */ | |
3655 | if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0 | |
3656 | && !IN_CURRENT_FENCE_P (BB_END (pred_bb))) | |
3657 | { | |
3658 | if (!sel_remove_insn (BB_END (pred_bb), false, false)) | |
3659 | tidy_fallthru_edge (e); | |
3660 | } | |
3661 | else | |
3662 | sel_redirect_edge_and_branch (e, succ_bb); | |
3663 | rescan_p = true; | |
3664 | break; | |
3665 | } | |
e855c69d AB |
3666 | } |
3667 | } | |
3668 | ||
e855c69d AB |
3669 | if (can_merge_blocks_p (bb->prev_bb, bb)) |
3670 | sel_merge_blocks (bb->prev_bb, bb); | |
3671 | else | |
e855c69d | 3672 | { |
262d8232 | 3673 | /* This is a block without fallthru predecessor. Just delete it. */ |
e855c69d AB |
3674 | gcc_assert (pred_bb != NULL); |
3675 | ||
762bffba AB |
3676 | if (in_current_region_p (pred_bb)) |
3677 | move_bb_info (pred_bb, bb); | |
e855c69d AB |
3678 | remove_empty_bb (bb, true); |
3679 | } | |
3680 | ||
00c4e97c AB |
3681 | if (!VEC_empty (basic_block, dom_bbs)) |
3682 | { | |
3683 | VEC_safe_push (basic_block, heap, dom_bbs, succ_bb); | |
3684 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
3685 | VEC_free (basic_block, heap, dom_bbs); | |
3686 | } | |
3687 | ||
e855c69d AB |
3688 | return true; |
3689 | } | |
3690 | ||
b8698a0f | 3691 | /* Tidy the control flow after we have removed original insn from |
e855c69d AB |
3692 | XBB. Return true if we have removed some blocks. When FULL_TIDYING |
3693 | is true, also try to optimize control flow on non-empty blocks. */ | |
3694 | bool | |
3695 | tidy_control_flow (basic_block xbb, bool full_tidying) | |
3696 | { | |
3697 | bool changed = true; | |
b5b8b0ac | 3698 | insn_t first, last; |
b8698a0f | 3699 | |
e855c69d | 3700 | /* First check whether XBB is empty. */ |
5f33b972 | 3701 | changed = maybe_tidy_empty_bb (xbb); |
e855c69d AB |
3702 | if (changed || !full_tidying) |
3703 | return changed; | |
b8698a0f | 3704 | |
e855c69d | 3705 | /* Check if there is a unnecessary jump after insn left. */ |
753de8cf | 3706 | if (bb_has_removable_jump_to_p (xbb, xbb->next_bb) |
e855c69d AB |
3707 | && INSN_SCHED_TIMES (BB_END (xbb)) == 0 |
3708 | && !IN_CURRENT_FENCE_P (BB_END (xbb))) | |
3709 | { | |
3710 | if (sel_remove_insn (BB_END (xbb), false, false)) | |
3711 | return true; | |
3712 | tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); | |
3713 | } | |
3714 | ||
b5b8b0ac AO |
3715 | first = sel_bb_head (xbb); |
3716 | last = sel_bb_end (xbb); | |
3717 | if (MAY_HAVE_DEBUG_INSNS) | |
3718 | { | |
3719 | if (first != last && DEBUG_INSN_P (first)) | |
3720 | do | |
3721 | first = NEXT_INSN (first); | |
3722 | while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); | |
3723 | ||
3724 | if (first != last && DEBUG_INSN_P (last)) | |
3725 | do | |
3726 | last = PREV_INSN (last); | |
3727 | while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); | |
3728 | } | |
e855c69d | 3729 | /* Check if there is an unnecessary jump in previous basic block leading |
b8698a0f L |
3730 | to next basic block left after removing INSN from stream. |
3731 | If it is so, remove that jump and redirect edge to current | |
3732 | basic block (where there was INSN before deletion). This way | |
3733 | when NOP will be deleted several instructions later with its | |
3734 | basic block we will not get a jump to next instruction, which | |
e855c69d | 3735 | can be harmful. */ |
b5b8b0ac | 3736 | if (first == last |
e855c69d | 3737 | && !sel_bb_empty_p (xbb) |
b5b8b0ac | 3738 | && INSN_NOP_P (last) |
e855c69d AB |
3739 | /* Flow goes fallthru from current block to the next. */ |
3740 | && EDGE_COUNT (xbb->succs) == 1 | |
3741 | && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) | |
3742 | /* When successor is an EXIT block, it may not be the next block. */ | |
3743 | && single_succ (xbb) != EXIT_BLOCK_PTR | |
3744 | /* And unconditional jump in previous basic block leads to | |
3745 | next basic block of XBB and this jump can be safely removed. */ | |
3746 | && in_current_region_p (xbb->prev_bb) | |
753de8cf | 3747 | && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb) |
e855c69d AB |
3748 | && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 |
3749 | /* Also this jump is not at the scheduling boundary. */ | |
3750 | && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) | |
3751 | { | |
b59ab570 | 3752 | bool recompute_toporder_p; |
e855c69d AB |
3753 | /* Clear data structures of jump - jump itself will be removed |
3754 | by sel_redirect_edge_and_branch. */ | |
3755 | clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); | |
b59ab570 AM |
3756 | recompute_toporder_p |
3757 | = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); | |
3758 | ||
e855c69d AB |
3759 | gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); |
3760 | ||
3761 | /* It can turn out that after removing unused jump, basic block | |
3762 | that contained that jump, becomes empty too. In such case | |
3763 | remove it too. */ | |
3764 | if (sel_bb_empty_p (xbb->prev_bb)) | |
5f33b972 AM |
3765 | changed = maybe_tidy_empty_bb (xbb->prev_bb); |
3766 | if (recompute_toporder_p) | |
b59ab570 | 3767 | sel_recompute_toporder (); |
e855c69d | 3768 | } |
d787f788 AM |
3769 | |
3770 | #ifdef ENABLE_CHECKING | |
3771 | verify_backedges (); | |
00c4e97c | 3772 | verify_dominators (CDI_DOMINATORS); |
d787f788 AM |
3773 | #endif |
3774 | ||
e855c69d AB |
3775 | return changed; |
3776 | } | |
3777 | ||
b59ab570 AM |
3778 | /* Purge meaningless empty blocks in the middle of a region. */ |
3779 | void | |
3780 | purge_empty_blocks (void) | |
3781 | { | |
9d0dcda1 | 3782 | int i; |
b59ab570 | 3783 | |
9d0dcda1 AM |
3784 | /* Do not attempt to delete the first basic block in the region. */ |
3785 | for (i = 1; i < current_nr_blocks; ) | |
b59ab570 AM |
3786 | { |
3787 | basic_block b = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
3788 | ||
5f33b972 | 3789 | if (maybe_tidy_empty_bb (b)) |
b59ab570 AM |
3790 | continue; |
3791 | ||
3792 | i++; | |
3793 | } | |
3794 | } | |
3795 | ||
b8698a0f L |
3796 | /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, |
3797 | do not delete insn's data, because it will be later re-emitted. | |
e855c69d AB |
3798 | Return true if we have removed some blocks afterwards. */ |
3799 | bool | |
3800 | sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) | |
3801 | { | |
3802 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3803 | ||
3804 | gcc_assert (INSN_IN_STREAM_P (insn)); | |
3805 | ||
b5b8b0ac AO |
3806 | if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) |
3807 | { | |
3808 | expr_t expr; | |
3809 | av_set_iterator i; | |
3810 | ||
3811 | /* When we remove a debug insn that is head of a BB, it remains | |
3812 | in the AV_SET of the block, but it shouldn't. */ | |
3813 | FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) | |
3814 | if (EXPR_INSN_RTX (expr) == insn) | |
3815 | { | |
3816 | av_set_iter_remove (&i); | |
3817 | break; | |
3818 | } | |
3819 | } | |
3820 | ||
e855c69d AB |
3821 | if (only_disconnect) |
3822 | { | |
3823 | insn_t prev = PREV_INSN (insn); | |
3824 | insn_t next = NEXT_INSN (insn); | |
3825 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3826 | ||
3827 | NEXT_INSN (prev) = next; | |
3828 | PREV_INSN (next) = prev; | |
3829 | ||
3830 | if (BB_HEAD (bb) == insn) | |
3831 | { | |
3832 | gcc_assert (BLOCK_FOR_INSN (prev) == bb); | |
3833 | BB_HEAD (bb) = prev; | |
3834 | } | |
3835 | if (BB_END (bb) == insn) | |
3836 | BB_END (bb) = prev; | |
3837 | } | |
3838 | else | |
3839 | { | |
3840 | remove_insn (insn); | |
3841 | clear_expr (INSN_EXPR (insn)); | |
3842 | } | |
3843 | ||
3844 | /* It is necessary to null this fields before calling add_insn (). */ | |
3845 | PREV_INSN (insn) = NULL_RTX; | |
3846 | NEXT_INSN (insn) = NULL_RTX; | |
3847 | ||
3848 | return tidy_control_flow (bb, full_tidying); | |
3849 | } | |
3850 | ||
3851 | /* Estimate number of the insns in BB. */ | |
3852 | static int | |
3853 | sel_estimate_number_of_insns (basic_block bb) | |
3854 | { | |
3855 | int res = 0; | |
3856 | insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); | |
3857 | ||
3858 | for (; insn != next_tail; insn = NEXT_INSN (insn)) | |
b5b8b0ac | 3859 | if (NONDEBUG_INSN_P (insn)) |
e855c69d AB |
3860 | res++; |
3861 | ||
3862 | return res; | |
3863 | } | |
3864 | ||
3865 | /* We don't need separate luids for notes or labels. */ | |
3866 | static int | |
3867 | sel_luid_for_non_insn (rtx x) | |
3868 | { | |
3869 | gcc_assert (NOTE_P (x) || LABEL_P (x)); | |
3870 | ||
3871 | return -1; | |
3872 | } | |
3873 | ||
3874 | /* Return seqno of the only predecessor of INSN. */ | |
3875 | static int | |
3876 | get_seqno_of_a_pred (insn_t insn) | |
3877 | { | |
3878 | int seqno; | |
3879 | ||
3880 | gcc_assert (INSN_SIMPLEJUMP_P (insn)); | |
3881 | ||
3882 | if (!sel_bb_head_p (insn)) | |
3883 | seqno = INSN_SEQNO (PREV_INSN (insn)); | |
3884 | else | |
3885 | { | |
3886 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3887 | ||
3888 | if (single_pred_p (bb) | |
3889 | && !in_current_region_p (single_pred (bb))) | |
3890 | { | |
3891 | /* We can have preds outside a region when splitting edges | |
b8698a0f | 3892 | for pipelining of an outer loop. Use succ instead. |
e855c69d AB |
3893 | There should be only one of them. */ |
3894 | insn_t succ = NULL; | |
3895 | succ_iterator si; | |
3896 | bool first = true; | |
b8698a0f | 3897 | |
e855c69d AB |
3898 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
3899 | && current_loop_nest); | |
b8698a0f | 3900 | FOR_EACH_SUCC_1 (succ, si, insn, |
e855c69d AB |
3901 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
3902 | { | |
3903 | gcc_assert (first); | |
3904 | first = false; | |
3905 | } | |
3906 | ||
3907 | gcc_assert (succ != NULL); | |
3908 | seqno = INSN_SEQNO (succ); | |
3909 | } | |
3910 | else | |
3911 | { | |
3912 | insn_t *preds; | |
3913 | int n; | |
3914 | ||
3915 | cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); | |
3916 | gcc_assert (n == 1); | |
3917 | ||
3918 | seqno = INSN_SEQNO (preds[0]); | |
b8698a0f | 3919 | |
e855c69d AB |
3920 | free (preds); |
3921 | } | |
3922 | } | |
3923 | ||
3924 | return seqno; | |
3925 | } | |
3926 | ||
da7ba240 AB |
3927 | /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors |
3928 | with positive seqno exist. */ | |
e855c69d AB |
3929 | int |
3930 | get_seqno_by_preds (rtx insn) | |
3931 | { | |
3932 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3933 | rtx tmp = insn, head = BB_HEAD (bb); | |
3934 | insn_t *preds; | |
3935 | int n, i, seqno; | |
3936 | ||
3937 | while (tmp != head) | |
3938 | if (INSN_P (tmp)) | |
3939 | return INSN_SEQNO (tmp); | |
3940 | else | |
3941 | tmp = PREV_INSN (tmp); | |
b8698a0f | 3942 | |
e855c69d AB |
3943 | cfg_preds (bb, &preds, &n); |
3944 | for (i = 0, seqno = -1; i < n; i++) | |
3945 | seqno = MAX (seqno, INSN_SEQNO (preds[i])); | |
3946 | ||
e855c69d AB |
3947 | return seqno; |
3948 | } | |
3949 | ||
3950 | \f | |
3951 | ||
3952 | /* Extend pass-scope data structures for basic blocks. */ | |
3953 | void | |
3954 | sel_extend_global_bb_info (void) | |
3955 | { | |
3956 | VEC_safe_grow_cleared (sel_global_bb_info_def, heap, sel_global_bb_info, | |
3957 | last_basic_block); | |
3958 | } | |
3959 | ||
3960 | /* Extend region-scope data structures for basic blocks. */ | |
3961 | static void | |
3962 | extend_region_bb_info (void) | |
3963 | { | |
3964 | VEC_safe_grow_cleared (sel_region_bb_info_def, heap, sel_region_bb_info, | |
3965 | last_basic_block); | |
3966 | } | |
3967 | ||
3968 | /* Extend all data structures to fit for all basic blocks. */ | |
3969 | static void | |
3970 | extend_bb_info (void) | |
3971 | { | |
3972 | sel_extend_global_bb_info (); | |
3973 | extend_region_bb_info (); | |
3974 | } | |
3975 | ||
3976 | /* Finalize pass-scope data structures for basic blocks. */ | |
3977 | void | |
3978 | sel_finish_global_bb_info (void) | |
3979 | { | |
3980 | VEC_free (sel_global_bb_info_def, heap, sel_global_bb_info); | |
3981 | } | |
3982 | ||
3983 | /* Finalize region-scope data structures for basic blocks. */ | |
3984 | static void | |
3985 | finish_region_bb_info (void) | |
3986 | { | |
3987 | VEC_free (sel_region_bb_info_def, heap, sel_region_bb_info); | |
3988 | } | |
3989 | \f | |
3990 | ||
3991 | /* Data for each insn in current region. */ | |
3992 | VEC (sel_insn_data_def, heap) *s_i_d = NULL; | |
3993 | ||
3994 | /* A vector for the insns we've emitted. */ | |
3995 | static insn_vec_t new_insns = NULL; | |
3996 | ||
3997 | /* Extend data structures for insns from current region. */ | |
3998 | static void | |
3999 | extend_insn_data (void) | |
4000 | { | |
4001 | int reserve; | |
b8698a0f | 4002 | |
e855c69d AB |
4003 | sched_extend_target (); |
4004 | sched_deps_init (false); | |
4005 | ||
4006 | /* Extend data structures for insns from current region. */ | |
b8698a0f | 4007 | reserve = (sched_max_luid + 1 |
e855c69d | 4008 | - VEC_length (sel_insn_data_def, s_i_d)); |
b8698a0f | 4009 | if (reserve > 0 |
e855c69d | 4010 | && ! VEC_space (sel_insn_data_def, s_i_d, reserve)) |
bcf33775 AB |
4011 | { |
4012 | int size; | |
4013 | ||
4014 | if (sched_max_luid / 2 > 1024) | |
4015 | size = sched_max_luid + 1024; | |
4016 | else | |
4017 | size = 3 * sched_max_luid / 2; | |
b8698a0f | 4018 | |
bcf33775 AB |
4019 | |
4020 | VEC_safe_grow_cleared (sel_insn_data_def, heap, s_i_d, size); | |
4021 | } | |
e855c69d AB |
4022 | } |
4023 | ||
4024 | /* Finalize data structures for insns from current region. */ | |
4025 | static void | |
4026 | finish_insns (void) | |
4027 | { | |
4028 | unsigned i; | |
4029 | ||
4030 | /* Clear here all dependence contexts that may have left from insns that were | |
4031 | removed during the scheduling. */ | |
4032 | for (i = 0; i < VEC_length (sel_insn_data_def, s_i_d); i++) | |
4033 | { | |
4034 | sel_insn_data_def *sid_entry = VEC_index (sel_insn_data_def, s_i_d, i); | |
b8698a0f | 4035 | |
e855c69d AB |
4036 | if (sid_entry->live) |
4037 | return_regset_to_pool (sid_entry->live); | |
4038 | if (sid_entry->analyzed_deps) | |
4039 | { | |
4040 | BITMAP_FREE (sid_entry->analyzed_deps); | |
4041 | BITMAP_FREE (sid_entry->found_deps); | |
4042 | htab_delete (sid_entry->transformed_insns); | |
4043 | free_deps (&sid_entry->deps_context); | |
4044 | } | |
4045 | if (EXPR_VINSN (&sid_entry->expr)) | |
4046 | { | |
4047 | clear_expr (&sid_entry->expr); | |
b8698a0f | 4048 | |
e855c69d AB |
4049 | /* Also, clear CANT_MOVE bit here, because we really don't want it |
4050 | to be passed to the next region. */ | |
4051 | CANT_MOVE_BY_LUID (i) = 0; | |
4052 | } | |
4053 | } | |
b8698a0f | 4054 | |
e855c69d AB |
4055 | VEC_free (sel_insn_data_def, heap, s_i_d); |
4056 | } | |
4057 | ||
4058 | /* A proxy to pass initialization data to init_insn (). */ | |
4059 | static sel_insn_data_def _insn_init_ssid; | |
4060 | static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; | |
4061 | ||
4062 | /* If true create a new vinsn. Otherwise use the one from EXPR. */ | |
4063 | static bool insn_init_create_new_vinsn_p; | |
4064 | ||
4065 | /* Set all necessary data for initialization of the new insn[s]. */ | |
4066 | static expr_t | |
4067 | set_insn_init (expr_t expr, vinsn_t vi, int seqno) | |
4068 | { | |
4069 | expr_t x = &insn_init_ssid->expr; | |
4070 | ||
4071 | copy_expr_onside (x, expr); | |
4072 | if (vi != NULL) | |
4073 | { | |
4074 | insn_init_create_new_vinsn_p = false; | |
4075 | change_vinsn_in_expr (x, vi); | |
4076 | } | |
4077 | else | |
4078 | insn_init_create_new_vinsn_p = true; | |
4079 | ||
4080 | insn_init_ssid->seqno = seqno; | |
4081 | return x; | |
4082 | } | |
4083 | ||
4084 | /* Init data for INSN. */ | |
4085 | static void | |
4086 | init_insn_data (insn_t insn) | |
4087 | { | |
4088 | expr_t expr; | |
4089 | sel_insn_data_t ssid = insn_init_ssid; | |
4090 | ||
4091 | /* The fields mentioned below are special and hence are not being | |
4092 | propagated to the new insns. */ | |
4093 | gcc_assert (!ssid->asm_p && ssid->sched_next == NULL | |
4094 | && !ssid->after_stall_p && ssid->sched_cycle == 0); | |
4095 | gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); | |
4096 | ||
4097 | expr = INSN_EXPR (insn); | |
4098 | copy_expr (expr, &ssid->expr); | |
4099 | prepare_insn_expr (insn, ssid->seqno); | |
4100 | ||
4101 | if (insn_init_create_new_vinsn_p) | |
4102 | change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p)); | |
b8698a0f | 4103 | |
e855c69d AB |
4104 | if (first_time_insn_init (insn)) |
4105 | init_first_time_insn_data (insn); | |
4106 | } | |
4107 | ||
4108 | /* This is used to initialize spurious jumps generated by | |
4109 | sel_redirect_edge (). */ | |
4110 | static void | |
4111 | init_simplejump_data (insn_t insn) | |
4112 | { | |
4113 | init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0, | |
b8698a0f | 4114 | REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0, NULL, true, false, false, |
e855c69d AB |
4115 | false, true); |
4116 | INSN_SEQNO (insn) = get_seqno_of_a_pred (insn); | |
4117 | init_first_time_insn_data (insn); | |
4118 | } | |
4119 | ||
b8698a0f | 4120 | /* Perform deferred initialization of insns. This is used to process |
e855c69d AB |
4121 | a new jump that may be created by redirect_edge. */ |
4122 | void | |
4123 | sel_init_new_insn (insn_t insn, int flags) | |
4124 | { | |
4125 | /* We create data structures for bb when the first insn is emitted in it. */ | |
4126 | if (INSN_P (insn) | |
4127 | && INSN_IN_STREAM_P (insn) | |
4128 | && insn_is_the_only_one_in_bb_p (insn)) | |
4129 | { | |
4130 | extend_bb_info (); | |
4131 | create_initial_data_sets (BLOCK_FOR_INSN (insn)); | |
4132 | } | |
b8698a0f | 4133 | |
e855c69d AB |
4134 | if (flags & INSN_INIT_TODO_LUID) |
4135 | sched_init_luids (NULL, NULL, NULL, insn); | |
4136 | ||
4137 | if (flags & INSN_INIT_TODO_SSID) | |
4138 | { | |
4139 | extend_insn_data (); | |
4140 | init_insn_data (insn); | |
4141 | clear_expr (&insn_init_ssid->expr); | |
4142 | } | |
4143 | ||
4144 | if (flags & INSN_INIT_TODO_SIMPLEJUMP) | |
4145 | { | |
4146 | extend_insn_data (); | |
4147 | init_simplejump_data (insn); | |
4148 | } | |
b8698a0f | 4149 | |
e855c69d AB |
4150 | gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) |
4151 | == CONTAINING_RGN (BB_TO_BLOCK (0))); | |
4152 | } | |
4153 | \f | |
4154 | ||
4155 | /* Functions to init/finish work with lv sets. */ | |
4156 | ||
4157 | /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ | |
4158 | static void | |
4159 | init_lv_set (basic_block bb) | |
4160 | { | |
4161 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
4162 | ||
4163 | BB_LV_SET (bb) = get_regset_from_pool (); | |
b8698a0f | 4164 | COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); |
e855c69d AB |
4165 | BB_LV_SET_VALID_P (bb) = true; |
4166 | } | |
4167 | ||
4168 | /* Copy liveness information to BB from FROM_BB. */ | |
4169 | static void | |
4170 | copy_lv_set_from (basic_block bb, basic_block from_bb) | |
4171 | { | |
4172 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
b8698a0f | 4173 | |
e855c69d AB |
4174 | COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); |
4175 | BB_LV_SET_VALID_P (bb) = true; | |
b8698a0f | 4176 | } |
e855c69d AB |
4177 | |
4178 | /* Initialize lv set of all bb headers. */ | |
4179 | void | |
4180 | init_lv_sets (void) | |
4181 | { | |
4182 | basic_block bb; | |
4183 | ||
4184 | /* Initialize of LV sets. */ | |
4185 | FOR_EACH_BB (bb) | |
4186 | init_lv_set (bb); | |
4187 | ||
4188 | /* Don't forget EXIT_BLOCK. */ | |
4189 | init_lv_set (EXIT_BLOCK_PTR); | |
4190 | } | |
4191 | ||
4192 | /* Release lv set of HEAD. */ | |
4193 | static void | |
4194 | free_lv_set (basic_block bb) | |
4195 | { | |
4196 | gcc_assert (BB_LV_SET (bb) != NULL); | |
4197 | ||
4198 | return_regset_to_pool (BB_LV_SET (bb)); | |
4199 | BB_LV_SET (bb) = NULL; | |
4200 | BB_LV_SET_VALID_P (bb) = false; | |
4201 | } | |
4202 | ||
4203 | /* Finalize lv sets of all bb headers. */ | |
4204 | void | |
4205 | free_lv_sets (void) | |
4206 | { | |
4207 | basic_block bb; | |
4208 | ||
4209 | /* Don't forget EXIT_BLOCK. */ | |
4210 | free_lv_set (EXIT_BLOCK_PTR); | |
4211 | ||
4212 | /* Free LV sets. */ | |
4213 | FOR_EACH_BB (bb) | |
4214 | if (BB_LV_SET (bb)) | |
4215 | free_lv_set (bb); | |
4216 | } | |
4217 | ||
4218 | /* Initialize an invalid AV_SET for BB. | |
4219 | This set will be updated next time compute_av () process BB. */ | |
4220 | static void | |
4221 | invalidate_av_set (basic_block bb) | |
4222 | { | |
4223 | gcc_assert (BB_AV_LEVEL (bb) <= 0 | |
4224 | && BB_AV_SET (bb) == NULL); | |
4225 | ||
4226 | BB_AV_LEVEL (bb) = -1; | |
4227 | } | |
4228 | ||
4229 | /* Create initial data sets for BB (they will be invalid). */ | |
4230 | static void | |
4231 | create_initial_data_sets (basic_block bb) | |
4232 | { | |
4233 | if (BB_LV_SET (bb)) | |
4234 | BB_LV_SET_VALID_P (bb) = false; | |
4235 | else | |
4236 | BB_LV_SET (bb) = get_regset_from_pool (); | |
4237 | invalidate_av_set (bb); | |
4238 | } | |
4239 | ||
4240 | /* Free av set of BB. */ | |
4241 | static void | |
4242 | free_av_set (basic_block bb) | |
4243 | { | |
4244 | av_set_clear (&BB_AV_SET (bb)); | |
4245 | BB_AV_LEVEL (bb) = 0; | |
4246 | } | |
4247 | ||
4248 | /* Free data sets of BB. */ | |
4249 | void | |
4250 | free_data_sets (basic_block bb) | |
4251 | { | |
4252 | free_lv_set (bb); | |
4253 | free_av_set (bb); | |
4254 | } | |
4255 | ||
4256 | /* Exchange lv sets of TO and FROM. */ | |
4257 | static void | |
4258 | exchange_lv_sets (basic_block to, basic_block from) | |
4259 | { | |
4260 | { | |
4261 | regset to_lv_set = BB_LV_SET (to); | |
4262 | ||
4263 | BB_LV_SET (to) = BB_LV_SET (from); | |
4264 | BB_LV_SET (from) = to_lv_set; | |
4265 | } | |
4266 | ||
4267 | { | |
4268 | bool to_lv_set_valid_p = BB_LV_SET_VALID_P (to); | |
4269 | ||
4270 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4271 | BB_LV_SET_VALID_P (from) = to_lv_set_valid_p; | |
4272 | } | |
4273 | } | |
4274 | ||
4275 | ||
4276 | /* Exchange av sets of TO and FROM. */ | |
4277 | static void | |
4278 | exchange_av_sets (basic_block to, basic_block from) | |
4279 | { | |
4280 | { | |
4281 | av_set_t to_av_set = BB_AV_SET (to); | |
4282 | ||
4283 | BB_AV_SET (to) = BB_AV_SET (from); | |
4284 | BB_AV_SET (from) = to_av_set; | |
4285 | } | |
4286 | ||
4287 | { | |
4288 | int to_av_level = BB_AV_LEVEL (to); | |
4289 | ||
4290 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4291 | BB_AV_LEVEL (from) = to_av_level; | |
4292 | } | |
4293 | } | |
4294 | ||
4295 | /* Exchange data sets of TO and FROM. */ | |
4296 | void | |
4297 | exchange_data_sets (basic_block to, basic_block from) | |
4298 | { | |
4299 | exchange_lv_sets (to, from); | |
4300 | exchange_av_sets (to, from); | |
4301 | } | |
4302 | ||
4303 | /* Copy data sets of FROM to TO. */ | |
4304 | void | |
4305 | copy_data_sets (basic_block to, basic_block from) | |
4306 | { | |
4307 | gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); | |
4308 | gcc_assert (BB_AV_SET (to) == NULL); | |
4309 | ||
4310 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4311 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4312 | ||
4313 | if (BB_AV_SET_VALID_P (from)) | |
4314 | { | |
4315 | BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); | |
4316 | } | |
4317 | if (BB_LV_SET_VALID_P (from)) | |
4318 | { | |
4319 | gcc_assert (BB_LV_SET (to) != NULL); | |
4320 | COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); | |
4321 | } | |
4322 | } | |
4323 | ||
4324 | /* Return an av set for INSN, if any. */ | |
4325 | av_set_t | |
4326 | get_av_set (insn_t insn) | |
4327 | { | |
4328 | av_set_t av_set; | |
4329 | ||
4330 | gcc_assert (AV_SET_VALID_P (insn)); | |
4331 | ||
4332 | if (sel_bb_head_p (insn)) | |
4333 | av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); | |
4334 | else | |
4335 | av_set = NULL; | |
4336 | ||
4337 | return av_set; | |
4338 | } | |
4339 | ||
4340 | /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ | |
4341 | int | |
4342 | get_av_level (insn_t insn) | |
4343 | { | |
4344 | int av_level; | |
4345 | ||
4346 | gcc_assert (INSN_P (insn)); | |
4347 | ||
4348 | if (sel_bb_head_p (insn)) | |
4349 | av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); | |
4350 | else | |
4351 | av_level = INSN_WS_LEVEL (insn); | |
4352 | ||
4353 | return av_level; | |
4354 | } | |
4355 | ||
4356 | \f | |
4357 | ||
4358 | /* Variables to work with control-flow graph. */ | |
4359 | ||
4360 | /* The basic block that already has been processed by the sched_data_update (), | |
4361 | but hasn't been in sel_add_bb () yet. */ | |
4362 | static VEC (basic_block, heap) *last_added_blocks = NULL; | |
4363 | ||
4364 | /* A pool for allocating successor infos. */ | |
4365 | static struct | |
4366 | { | |
4367 | /* A stack for saving succs_info structures. */ | |
4368 | struct succs_info *stack; | |
4369 | ||
4370 | /* Its size. */ | |
4371 | int size; | |
4372 | ||
4373 | /* Top of the stack. */ | |
4374 | int top; | |
4375 | ||
4376 | /* Maximal value of the top. */ | |
4377 | int max_top; | |
4378 | } succs_info_pool; | |
4379 | ||
4380 | /* Functions to work with control-flow graph. */ | |
4381 | ||
4382 | /* Return basic block note of BB. */ | |
4383 | insn_t | |
4384 | sel_bb_head (basic_block bb) | |
4385 | { | |
4386 | insn_t head; | |
4387 | ||
4388 | if (bb == EXIT_BLOCK_PTR) | |
4389 | { | |
4390 | gcc_assert (exit_insn != NULL_RTX); | |
4391 | head = exit_insn; | |
4392 | } | |
4393 | else | |
4394 | { | |
4395 | insn_t note; | |
4396 | ||
4397 | note = bb_note (bb); | |
4398 | head = next_nonnote_insn (note); | |
4399 | ||
89ad0f25 | 4400 | if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb)) |
e855c69d AB |
4401 | head = NULL_RTX; |
4402 | } | |
4403 | ||
4404 | return head; | |
4405 | } | |
4406 | ||
4407 | /* Return true if INSN is a basic block header. */ | |
4408 | bool | |
4409 | sel_bb_head_p (insn_t insn) | |
4410 | { | |
4411 | return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn; | |
4412 | } | |
4413 | ||
4414 | /* Return last insn of BB. */ | |
4415 | insn_t | |
4416 | sel_bb_end (basic_block bb) | |
4417 | { | |
4418 | if (sel_bb_empty_p (bb)) | |
4419 | return NULL_RTX; | |
4420 | ||
4421 | gcc_assert (bb != EXIT_BLOCK_PTR); | |
4422 | ||
4423 | return BB_END (bb); | |
4424 | } | |
4425 | ||
4426 | /* Return true if INSN is the last insn in its basic block. */ | |
4427 | bool | |
4428 | sel_bb_end_p (insn_t insn) | |
4429 | { | |
4430 | return insn == sel_bb_end (BLOCK_FOR_INSN (insn)); | |
4431 | } | |
4432 | ||
4433 | /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ | |
4434 | bool | |
4435 | sel_bb_empty_p (basic_block bb) | |
4436 | { | |
4437 | return sel_bb_head (bb) == NULL; | |
4438 | } | |
4439 | ||
4440 | /* True when BB belongs to the current scheduling region. */ | |
4441 | bool | |
4442 | in_current_region_p (basic_block bb) | |
4443 | { | |
4444 | if (bb->index < NUM_FIXED_BLOCKS) | |
4445 | return false; | |
4446 | ||
4447 | return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4448 | } | |
4449 | ||
4450 | /* Return the block which is a fallthru bb of a conditional jump JUMP. */ | |
4451 | basic_block | |
4452 | fallthru_bb_of_jump (rtx jump) | |
4453 | { | |
4454 | if (!JUMP_P (jump)) | |
4455 | return NULL; | |
4456 | ||
e855c69d AB |
4457 | if (!any_condjump_p (jump)) |
4458 | return NULL; | |
4459 | ||
268bab85 AB |
4460 | /* A basic block that ends with a conditional jump may still have one successor |
4461 | (and be followed by a barrier), we are not interested. */ | |
4462 | if (single_succ_p (BLOCK_FOR_INSN (jump))) | |
4463 | return NULL; | |
4464 | ||
e855c69d AB |
4465 | return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; |
4466 | } | |
4467 | ||
4468 | /* Remove all notes from BB. */ | |
4469 | static void | |
4470 | init_bb (basic_block bb) | |
4471 | { | |
4472 | remove_notes (bb_note (bb), BB_END (bb)); | |
4473 | BB_NOTE_LIST (bb) = note_list; | |
4474 | } | |
4475 | ||
4476 | void | |
4477 | sel_init_bbs (bb_vec_t bbs, basic_block bb) | |
4478 | { | |
4479 | const struct sched_scan_info_def ssi = | |
4480 | { | |
4481 | extend_bb_info, /* extend_bb */ | |
4482 | init_bb, /* init_bb */ | |
4483 | NULL, /* extend_insn */ | |
4484 | NULL /* init_insn */ | |
4485 | }; | |
4486 | ||
4487 | sched_scan (&ssi, bbs, bb, new_insns, NULL); | |
4488 | } | |
4489 | ||
7898b93b | 4490 | /* Restore notes for the whole region. */ |
e855c69d | 4491 | static void |
7898b93b | 4492 | sel_restore_notes (void) |
e855c69d AB |
4493 | { |
4494 | int bb; | |
7898b93b | 4495 | insn_t insn; |
e855c69d AB |
4496 | |
4497 | for (bb = 0; bb < current_nr_blocks; bb++) | |
4498 | { | |
4499 | basic_block first, last; | |
4500 | ||
4501 | first = EBB_FIRST_BB (bb); | |
4502 | last = EBB_LAST_BB (bb)->next_bb; | |
4503 | ||
4504 | do | |
4505 | { | |
4506 | note_list = BB_NOTE_LIST (first); | |
4507 | restore_other_notes (NULL, first); | |
4508 | BB_NOTE_LIST (first) = NULL_RTX; | |
4509 | ||
7898b93b AM |
4510 | FOR_BB_INSNS (first, insn) |
4511 | if (NONDEBUG_INSN_P (insn)) | |
4512 | reemit_notes (insn); | |
4513 | ||
e855c69d AB |
4514 | first = first->next_bb; |
4515 | } | |
4516 | while (first != last); | |
4517 | } | |
4518 | } | |
4519 | ||
4520 | /* Free per-bb data structures. */ | |
4521 | void | |
4522 | sel_finish_bbs (void) | |
4523 | { | |
7898b93b | 4524 | sel_restore_notes (); |
e855c69d AB |
4525 | |
4526 | /* Remove current loop preheader from this loop. */ | |
4527 | if (current_loop_nest) | |
4528 | sel_remove_loop_preheader (); | |
4529 | ||
4530 | finish_region_bb_info (); | |
4531 | } | |
4532 | ||
4533 | /* Return true if INSN has a single successor of type FLAGS. */ | |
4534 | bool | |
4535 | sel_insn_has_single_succ_p (insn_t insn, int flags) | |
4536 | { | |
4537 | insn_t succ; | |
4538 | succ_iterator si; | |
4539 | bool first_p = true; | |
4540 | ||
4541 | FOR_EACH_SUCC_1 (succ, si, insn, flags) | |
4542 | { | |
4543 | if (first_p) | |
4544 | first_p = false; | |
4545 | else | |
4546 | return false; | |
4547 | } | |
4548 | ||
4549 | return true; | |
4550 | } | |
4551 | ||
4552 | /* Allocate successor's info. */ | |
4553 | static struct succs_info * | |
4554 | alloc_succs_info (void) | |
4555 | { | |
4556 | if (succs_info_pool.top == succs_info_pool.max_top) | |
4557 | { | |
4558 | int i; | |
b8698a0f | 4559 | |
e855c69d AB |
4560 | if (++succs_info_pool.max_top >= succs_info_pool.size) |
4561 | gcc_unreachable (); | |
4562 | ||
4563 | i = ++succs_info_pool.top; | |
4564 | succs_info_pool.stack[i].succs_ok = VEC_alloc (rtx, heap, 10); | |
4565 | succs_info_pool.stack[i].succs_other = VEC_alloc (rtx, heap, 10); | |
4566 | succs_info_pool.stack[i].probs_ok = VEC_alloc (int, heap, 10); | |
4567 | } | |
4568 | else | |
4569 | succs_info_pool.top++; | |
4570 | ||
4571 | return &succs_info_pool.stack[succs_info_pool.top]; | |
4572 | } | |
4573 | ||
4574 | /* Free successor's info. */ | |
4575 | void | |
4576 | free_succs_info (struct succs_info * sinfo) | |
4577 | { | |
b8698a0f | 4578 | gcc_assert (succs_info_pool.top >= 0 |
e855c69d AB |
4579 | && &succs_info_pool.stack[succs_info_pool.top] == sinfo); |
4580 | succs_info_pool.top--; | |
4581 | ||
4582 | /* Clear stale info. */ | |
b8698a0f | 4583 | VEC_block_remove (rtx, sinfo->succs_ok, |
e855c69d | 4584 | 0, VEC_length (rtx, sinfo->succs_ok)); |
b8698a0f | 4585 | VEC_block_remove (rtx, sinfo->succs_other, |
e855c69d | 4586 | 0, VEC_length (rtx, sinfo->succs_other)); |
b8698a0f | 4587 | VEC_block_remove (int, sinfo->probs_ok, |
e855c69d AB |
4588 | 0, VEC_length (int, sinfo->probs_ok)); |
4589 | sinfo->all_prob = 0; | |
4590 | sinfo->succs_ok_n = 0; | |
4591 | sinfo->all_succs_n = 0; | |
4592 | } | |
4593 | ||
b8698a0f | 4594 | /* Compute successor info for INSN. FLAGS are the flags passed |
e855c69d AB |
4595 | to the FOR_EACH_SUCC_1 iterator. */ |
4596 | struct succs_info * | |
4597 | compute_succs_info (insn_t insn, short flags) | |
4598 | { | |
4599 | succ_iterator si; | |
4600 | insn_t succ; | |
4601 | struct succs_info *sinfo = alloc_succs_info (); | |
4602 | ||
4603 | /* Traverse *all* successors and decide what to do with each. */ | |
4604 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
4605 | { | |
4606 | /* FIXME: this doesn't work for skipping to loop exits, as we don't | |
4607 | perform code motion through inner loops. */ | |
4608 | short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; | |
4609 | ||
4610 | if (current_flags & flags) | |
4611 | { | |
4612 | VEC_safe_push (rtx, heap, sinfo->succs_ok, succ); | |
4613 | VEC_safe_push (int, heap, sinfo->probs_ok, | |
b8698a0f | 4614 | /* FIXME: Improve calculation when skipping |
e855c69d | 4615 | inner loop to exits. */ |
b8698a0f L |
4616 | (si.bb_end |
4617 | ? si.e1->probability | |
e855c69d AB |
4618 | : REG_BR_PROB_BASE)); |
4619 | sinfo->succs_ok_n++; | |
4620 | } | |
4621 | else | |
4622 | VEC_safe_push (rtx, heap, sinfo->succs_other, succ); | |
4623 | ||
4624 | /* Compute all_prob. */ | |
4625 | if (!si.bb_end) | |
4626 | sinfo->all_prob = REG_BR_PROB_BASE; | |
4627 | else | |
4628 | sinfo->all_prob += si.e1->probability; | |
4629 | ||
4630 | sinfo->all_succs_n++; | |
4631 | } | |
4632 | ||
4633 | return sinfo; | |
4634 | } | |
4635 | ||
b8698a0f | 4636 | /* Return the predecessors of BB in PREDS and their number in N. |
e855c69d AB |
4637 | Empty blocks are skipped. SIZE is used to allocate PREDS. */ |
4638 | static void | |
4639 | cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) | |
4640 | { | |
4641 | edge e; | |
4642 | edge_iterator ei; | |
4643 | ||
4644 | gcc_assert (BLOCK_TO_BB (bb->index) != 0); | |
4645 | ||
4646 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4647 | { | |
4648 | basic_block pred_bb = e->src; | |
4649 | insn_t bb_end = BB_END (pred_bb); | |
4650 | ||
3e6a3f6f AB |
4651 | if (!in_current_region_p (pred_bb)) |
4652 | { | |
4653 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
4654 | && current_loop_nest); | |
4655 | continue; | |
4656 | } | |
e855c69d AB |
4657 | |
4658 | if (sel_bb_empty_p (pred_bb)) | |
4659 | cfg_preds_1 (pred_bb, preds, n, size); | |
4660 | else | |
4661 | { | |
4662 | if (*n == *size) | |
b8698a0f | 4663 | *preds = XRESIZEVEC (insn_t, *preds, |
e855c69d AB |
4664 | (*size = 2 * *size + 1)); |
4665 | (*preds)[(*n)++] = bb_end; | |
4666 | } | |
4667 | } | |
4668 | ||
3e6a3f6f AB |
4669 | gcc_assert (*n != 0 |
4670 | || (flag_sel_sched_pipelining_outer_loops | |
4671 | && current_loop_nest)); | |
e855c69d AB |
4672 | } |
4673 | ||
b8698a0f L |
4674 | /* Find all predecessors of BB and record them in PREDS and their number |
4675 | in N. Empty blocks are skipped, and only normal (forward in-region) | |
e855c69d AB |
4676 | edges are processed. */ |
4677 | static void | |
4678 | cfg_preds (basic_block bb, insn_t **preds, int *n) | |
4679 | { | |
4680 | int size = 0; | |
4681 | ||
4682 | *preds = NULL; | |
4683 | *n = 0; | |
4684 | cfg_preds_1 (bb, preds, n, &size); | |
4685 | } | |
4686 | ||
4687 | /* Returns true if we are moving INSN through join point. */ | |
4688 | bool | |
4689 | sel_num_cfg_preds_gt_1 (insn_t insn) | |
4690 | { | |
4691 | basic_block bb; | |
4692 | ||
4693 | if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) | |
4694 | return false; | |
4695 | ||
4696 | bb = BLOCK_FOR_INSN (insn); | |
4697 | ||
4698 | while (1) | |
4699 | { | |
4700 | if (EDGE_COUNT (bb->preds) > 1) | |
4701 | return true; | |
4702 | ||
4703 | gcc_assert (EDGE_PRED (bb, 0)->dest == bb); | |
4704 | bb = EDGE_PRED (bb, 0)->src; | |
4705 | ||
4706 | if (!sel_bb_empty_p (bb)) | |
4707 | break; | |
4708 | } | |
4709 | ||
4710 | return false; | |
4711 | } | |
4712 | ||
b8698a0f | 4713 | /* Returns true when BB should be the end of an ebb. Adapted from the |
e855c69d AB |
4714 | code in sched-ebb.c. */ |
4715 | bool | |
4716 | bb_ends_ebb_p (basic_block bb) | |
4717 | { | |
4718 | basic_block next_bb = bb_next_bb (bb); | |
4719 | edge e; | |
b8698a0f | 4720 | |
e855c69d AB |
4721 | if (next_bb == EXIT_BLOCK_PTR |
4722 | || bitmap_bit_p (forced_ebb_heads, next_bb->index) | |
4723 | || (LABEL_P (BB_HEAD (next_bb)) | |
4724 | /* NB: LABEL_NUSES () is not maintained outside of jump.c. | |
4725 | Work around that. */ | |
4726 | && !single_pred_p (next_bb))) | |
4727 | return true; | |
4728 | ||
4729 | if (!in_current_region_p (next_bb)) | |
4730 | return true; | |
4731 | ||
0fd4b31d NF |
4732 | e = find_fallthru_edge (bb->succs); |
4733 | if (e) | |
4734 | { | |
4735 | gcc_assert (e->dest == next_bb); | |
4736 | ||
4737 | return false; | |
4738 | } | |
e855c69d AB |
4739 | |
4740 | return true; | |
4741 | } | |
4742 | ||
4743 | /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a | |
4744 | successor of INSN. */ | |
4745 | bool | |
4746 | in_same_ebb_p (insn_t insn, insn_t succ) | |
4747 | { | |
4748 | basic_block ptr = BLOCK_FOR_INSN (insn); | |
4749 | ||
4750 | for(;;) | |
4751 | { | |
4752 | if (ptr == BLOCK_FOR_INSN (succ)) | |
4753 | return true; | |
b8698a0f | 4754 | |
e855c69d AB |
4755 | if (bb_ends_ebb_p (ptr)) |
4756 | return false; | |
4757 | ||
4758 | ptr = bb_next_bb (ptr); | |
4759 | } | |
4760 | ||
4761 | gcc_unreachable (); | |
4762 | return false; | |
4763 | } | |
4764 | ||
4765 | /* Recomputes the reverse topological order for the function and | |
4766 | saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also | |
4767 | modified appropriately. */ | |
4768 | static void | |
4769 | recompute_rev_top_order (void) | |
4770 | { | |
4771 | int *postorder; | |
4772 | int n_blocks, i; | |
4773 | ||
4774 | if (!rev_top_order_index || rev_top_order_index_len < last_basic_block) | |
4775 | { | |
b8698a0f | 4776 | rev_top_order_index_len = last_basic_block; |
e855c69d AB |
4777 | rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, |
4778 | rev_top_order_index_len); | |
4779 | } | |
4780 | ||
4781 | postorder = XNEWVEC (int, n_basic_blocks); | |
4782 | ||
4783 | n_blocks = post_order_compute (postorder, true, false); | |
4784 | gcc_assert (n_basic_blocks == n_blocks); | |
4785 | ||
4786 | /* Build reverse function: for each basic block with BB->INDEX == K | |
4787 | rev_top_order_index[K] is it's reverse topological sort number. */ | |
4788 | for (i = 0; i < n_blocks; i++) | |
4789 | { | |
4790 | gcc_assert (postorder[i] < rev_top_order_index_len); | |
4791 | rev_top_order_index[postorder[i]] = i; | |
4792 | } | |
4793 | ||
4794 | free (postorder); | |
4795 | } | |
4796 | ||
4797 | /* Clear all flags from insns in BB that could spoil its rescheduling. */ | |
4798 | void | |
4799 | clear_outdated_rtx_info (basic_block bb) | |
4800 | { | |
4801 | rtx insn; | |
4802 | ||
4803 | FOR_BB_INSNS (bb, insn) | |
4804 | if (INSN_P (insn)) | |
4805 | { | |
4806 | SCHED_GROUP_P (insn) = 0; | |
4807 | INSN_AFTER_STALL_P (insn) = 0; | |
4808 | INSN_SCHED_TIMES (insn) = 0; | |
4809 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; | |
4810 | ||
4811 | /* We cannot use the changed caches, as previously we could ignore | |
b8698a0f | 4812 | the LHS dependence due to enabled renaming and transform |
e855c69d AB |
4813 | the expression, and currently we'll be unable to do this. */ |
4814 | htab_empty (INSN_TRANSFORMED_INSNS (insn)); | |
4815 | } | |
4816 | } | |
4817 | ||
4818 | /* Add BB_NOTE to the pool of available basic block notes. */ | |
4819 | static void | |
4820 | return_bb_to_pool (basic_block bb) | |
4821 | { | |
4822 | rtx note = bb_note (bb); | |
4823 | ||
4824 | gcc_assert (NOTE_BASIC_BLOCK (note) == bb | |
4825 | && bb->aux == NULL); | |
4826 | ||
4827 | /* It turns out that current cfg infrastructure does not support | |
4828 | reuse of basic blocks. Don't bother for now. */ | |
4829 | /*VEC_safe_push (rtx, heap, bb_note_pool, note);*/ | |
4830 | } | |
4831 | ||
4832 | /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ | |
4833 | static rtx | |
4834 | get_bb_note_from_pool (void) | |
4835 | { | |
4836 | if (VEC_empty (rtx, bb_note_pool)) | |
4837 | return NULL_RTX; | |
4838 | else | |
4839 | { | |
4840 | rtx note = VEC_pop (rtx, bb_note_pool); | |
4841 | ||
4842 | PREV_INSN (note) = NULL_RTX; | |
4843 | NEXT_INSN (note) = NULL_RTX; | |
4844 | ||
4845 | return note; | |
4846 | } | |
4847 | } | |
4848 | ||
4849 | /* Free bb_note_pool. */ | |
4850 | void | |
4851 | free_bb_note_pool (void) | |
4852 | { | |
4853 | VEC_free (rtx, heap, bb_note_pool); | |
4854 | } | |
4855 | ||
4856 | /* Setup scheduler pool and successor structure. */ | |
4857 | void | |
4858 | alloc_sched_pools (void) | |
4859 | { | |
4860 | int succs_size; | |
4861 | ||
4862 | succs_size = MAX_WS + 1; | |
b8698a0f | 4863 | succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); |
e855c69d AB |
4864 | succs_info_pool.size = succs_size; |
4865 | succs_info_pool.top = -1; | |
4866 | succs_info_pool.max_top = -1; | |
4867 | ||
b8698a0f | 4868 | sched_lists_pool = create_alloc_pool ("sel-sched-lists", |
e855c69d AB |
4869 | sizeof (struct _list_node), 500); |
4870 | } | |
4871 | ||
4872 | /* Free the pools. */ | |
4873 | void | |
4874 | free_sched_pools (void) | |
4875 | { | |
4876 | int i; | |
b8698a0f | 4877 | |
e855c69d AB |
4878 | free_alloc_pool (sched_lists_pool); |
4879 | gcc_assert (succs_info_pool.top == -1); | |
4880 | for (i = 0; i < succs_info_pool.max_top; i++) | |
4881 | { | |
4882 | VEC_free (rtx, heap, succs_info_pool.stack[i].succs_ok); | |
4883 | VEC_free (rtx, heap, succs_info_pool.stack[i].succs_other); | |
4884 | VEC_free (int, heap, succs_info_pool.stack[i].probs_ok); | |
4885 | } | |
4886 | free (succs_info_pool.stack); | |
4887 | } | |
4888 | \f | |
4889 | ||
b8698a0f | 4890 | /* Returns a position in RGN where BB can be inserted retaining |
e855c69d AB |
4891 | topological order. */ |
4892 | static int | |
4893 | find_place_to_insert_bb (basic_block bb, int rgn) | |
4894 | { | |
4895 | bool has_preds_outside_rgn = false; | |
4896 | edge e; | |
4897 | edge_iterator ei; | |
b8698a0f | 4898 | |
e855c69d AB |
4899 | /* Find whether we have preds outside the region. */ |
4900 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4901 | if (!in_current_region_p (e->src)) | |
4902 | { | |
4903 | has_preds_outside_rgn = true; | |
4904 | break; | |
4905 | } | |
b8698a0f | 4906 | |
e855c69d AB |
4907 | /* Recompute the top order -- needed when we have > 1 pred |
4908 | and in case we don't have preds outside. */ | |
4909 | if (flag_sel_sched_pipelining_outer_loops | |
4910 | && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) | |
4911 | { | |
4912 | int i, bbi = bb->index, cur_bbi; | |
4913 | ||
4914 | recompute_rev_top_order (); | |
4915 | for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) | |
4916 | { | |
4917 | cur_bbi = BB_TO_BLOCK (i); | |
b8698a0f | 4918 | if (rev_top_order_index[bbi] |
e855c69d AB |
4919 | < rev_top_order_index[cur_bbi]) |
4920 | break; | |
4921 | } | |
b8698a0f | 4922 | |
e855c69d AB |
4923 | /* We skipped the right block, so we increase i. We accomodate |
4924 | it for increasing by step later, so we decrease i. */ | |
4925 | return (i + 1) - 1; | |
4926 | } | |
4927 | else if (has_preds_outside_rgn) | |
4928 | { | |
4929 | /* This is the case when we generate an extra empty block | |
4930 | to serve as region head during pipelining. */ | |
4931 | e = EDGE_SUCC (bb, 0); | |
4932 | gcc_assert (EDGE_COUNT (bb->succs) == 1 | |
4933 | && in_current_region_p (EDGE_SUCC (bb, 0)->dest) | |
4934 | && (BLOCK_TO_BB (e->dest->index) == 0)); | |
4935 | return -1; | |
4936 | } | |
4937 | ||
4938 | /* We don't have preds outside the region. We should have | |
4939 | the only pred, because the multiple preds case comes from | |
4940 | the pipelining of outer loops, and that is handled above. | |
4941 | Just take the bbi of this single pred. */ | |
4942 | if (EDGE_COUNT (bb->succs) > 0) | |
4943 | { | |
4944 | int pred_bbi; | |
b8698a0f | 4945 | |
e855c69d | 4946 | gcc_assert (EDGE_COUNT (bb->preds) == 1); |
b8698a0f | 4947 | |
e855c69d AB |
4948 | pred_bbi = EDGE_PRED (bb, 0)->src->index; |
4949 | return BLOCK_TO_BB (pred_bbi); | |
4950 | } | |
4951 | else | |
4952 | /* BB has no successors. It is safe to put it in the end. */ | |
4953 | return current_nr_blocks - 1; | |
4954 | } | |
4955 | ||
4956 | /* Deletes an empty basic block freeing its data. */ | |
4957 | static void | |
4958 | delete_and_free_basic_block (basic_block bb) | |
4959 | { | |
4960 | gcc_assert (sel_bb_empty_p (bb)); | |
4961 | ||
4962 | if (BB_LV_SET (bb)) | |
4963 | free_lv_set (bb); | |
4964 | ||
4965 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
4966 | ||
b8698a0f L |
4967 | /* Can't assert av_set properties because we use sel_aremove_bb |
4968 | when removing loop preheader from the region. At the point of | |
e855c69d AB |
4969 | removing the preheader we already have deallocated sel_region_bb_info. */ |
4970 | gcc_assert (BB_LV_SET (bb) == NULL | |
4971 | && !BB_LV_SET_VALID_P (bb) | |
4972 | && BB_AV_LEVEL (bb) == 0 | |
4973 | && BB_AV_SET (bb) == NULL); | |
b8698a0f | 4974 | |
e855c69d AB |
4975 | delete_basic_block (bb); |
4976 | } | |
4977 | ||
4978 | /* Add BB to the current region and update the region data. */ | |
4979 | static void | |
4980 | add_block_to_current_region (basic_block bb) | |
4981 | { | |
4982 | int i, pos, bbi = -2, rgn; | |
4983 | ||
4984 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4985 | bbi = find_place_to_insert_bb (bb, rgn); | |
4986 | bbi += 1; | |
4987 | pos = RGN_BLOCKS (rgn) + bbi; | |
4988 | ||
4989 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
4990 | && ebb_head[bbi] == pos); | |
b8698a0f | 4991 | |
e855c69d AB |
4992 | /* Make a place for the new block. */ |
4993 | extend_regions (); | |
4994 | ||
4995 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
4996 | BLOCK_TO_BB (rgn_bb_table[i])++; | |
b8698a0f | 4997 | |
e855c69d AB |
4998 | memmove (rgn_bb_table + pos + 1, |
4999 | rgn_bb_table + pos, | |
5000 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5001 | ||
5002 | /* Initialize data for BB. */ | |
5003 | rgn_bb_table[pos] = bb->index; | |
5004 | BLOCK_TO_BB (bb->index) = bbi; | |
5005 | CONTAINING_RGN (bb->index) = rgn; | |
5006 | ||
5007 | RGN_NR_BLOCKS (rgn)++; | |
b8698a0f | 5008 | |
e855c69d AB |
5009 | for (i = rgn + 1; i <= nr_regions; i++) |
5010 | RGN_BLOCKS (i)++; | |
5011 | } | |
5012 | ||
5013 | /* Remove BB from the current region and update the region data. */ | |
5014 | static void | |
5015 | remove_bb_from_region (basic_block bb) | |
5016 | { | |
5017 | int i, pos, bbi = -2, rgn; | |
5018 | ||
5019 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
5020 | bbi = BLOCK_TO_BB (bb->index); | |
5021 | pos = RGN_BLOCKS (rgn) + bbi; | |
5022 | ||
5023 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
5024 | && ebb_head[bbi] == pos); | |
5025 | ||
5026 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
5027 | BLOCK_TO_BB (rgn_bb_table[i])--; | |
5028 | ||
5029 | memmove (rgn_bb_table + pos, | |
5030 | rgn_bb_table + pos + 1, | |
5031 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5032 | ||
5033 | RGN_NR_BLOCKS (rgn)--; | |
5034 | for (i = rgn + 1; i <= nr_regions; i++) | |
5035 | RGN_BLOCKS (i)--; | |
5036 | } | |
5037 | ||
b8698a0f | 5038 | /* Add BB to the current region and update all data. If BB is NULL, add all |
e855c69d AB |
5039 | blocks from last_added_blocks vector. */ |
5040 | static void | |
5041 | sel_add_bb (basic_block bb) | |
5042 | { | |
5043 | /* Extend luids so that new notes will receive zero luids. */ | |
5044 | sched_init_luids (NULL, NULL, NULL, NULL); | |
5045 | sched_init_bbs (); | |
5046 | sel_init_bbs (last_added_blocks, NULL); | |
5047 | ||
b8698a0f | 5048 | /* When bb is passed explicitly, the vector should contain |
e855c69d AB |
5049 | the only element that equals to bb; otherwise, the vector |
5050 | should not be NULL. */ | |
5051 | gcc_assert (last_added_blocks != NULL); | |
b8698a0f | 5052 | |
e855c69d AB |
5053 | if (bb != NULL) |
5054 | { | |
5055 | gcc_assert (VEC_length (basic_block, last_added_blocks) == 1 | |
b8698a0f L |
5056 | && VEC_index (basic_block, |
5057 | last_added_blocks, 0) == bb); | |
e855c69d AB |
5058 | add_block_to_current_region (bb); |
5059 | ||
5060 | /* We associate creating/deleting data sets with the first insn | |
5061 | appearing / disappearing in the bb. */ | |
5062 | if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) | |
5063 | create_initial_data_sets (bb); | |
b8698a0f | 5064 | |
e855c69d AB |
5065 | VEC_free (basic_block, heap, last_added_blocks); |
5066 | } | |
5067 | else | |
5068 | /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ | |
5069 | { | |
5070 | int i; | |
5071 | basic_block temp_bb = NULL; | |
5072 | ||
b8698a0f | 5073 | for (i = 0; |
e855c69d AB |
5074 | VEC_iterate (basic_block, last_added_blocks, i, bb); i++) |
5075 | { | |
5076 | add_block_to_current_region (bb); | |
5077 | temp_bb = bb; | |
5078 | } | |
5079 | ||
b8698a0f | 5080 | /* We need to fetch at least one bb so we know the region |
e855c69d AB |
5081 | to update. */ |
5082 | gcc_assert (temp_bb != NULL); | |
5083 | bb = temp_bb; | |
5084 | ||
5085 | VEC_free (basic_block, heap, last_added_blocks); | |
5086 | } | |
5087 | ||
5088 | rgn_setup_region (CONTAINING_RGN (bb->index)); | |
5089 | } | |
5090 | ||
b8698a0f | 5091 | /* Remove BB from the current region and update all data. |
e855c69d AB |
5092 | If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ |
5093 | static void | |
5094 | sel_remove_bb (basic_block bb, bool remove_from_cfg_p) | |
5095 | { | |
262d8232 AB |
5096 | unsigned idx = bb->index; |
5097 | ||
e855c69d | 5098 | gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); |
b8698a0f | 5099 | |
e855c69d AB |
5100 | remove_bb_from_region (bb); |
5101 | return_bb_to_pool (bb); | |
262d8232 | 5102 | bitmap_clear_bit (blocks_to_reschedule, idx); |
b8698a0f | 5103 | |
e855c69d | 5104 | if (remove_from_cfg_p) |
00c4e97c AB |
5105 | { |
5106 | basic_block succ = single_succ (bb); | |
5107 | delete_and_free_basic_block (bb); | |
5108 | set_immediate_dominator (CDI_DOMINATORS, succ, | |
5109 | recompute_dominator (CDI_DOMINATORS, succ)); | |
5110 | } | |
e855c69d | 5111 | |
262d8232 | 5112 | rgn_setup_region (CONTAINING_RGN (idx)); |
e855c69d AB |
5113 | } |
5114 | ||
5115 | /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ | |
5116 | static void | |
5117 | move_bb_info (basic_block merge_bb, basic_block empty_bb) | |
5118 | { | |
5119 | gcc_assert (in_current_region_p (merge_bb)); | |
5120 | ||
b8698a0f | 5121 | concat_note_lists (BB_NOTE_LIST (empty_bb), |
e855c69d AB |
5122 | &BB_NOTE_LIST (merge_bb)); |
5123 | BB_NOTE_LIST (empty_bb) = NULL_RTX; | |
5124 | ||
5125 | } | |
5126 | ||
e855c69d AB |
5127 | /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from |
5128 | region, but keep it in CFG. */ | |
5129 | static void | |
5130 | remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) | |
5131 | { | |
5132 | /* The block should contain just a note or a label. | |
5133 | We try to check whether it is unused below. */ | |
5134 | gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) | |
5135 | || LABEL_P (BB_HEAD (empty_bb))); | |
5136 | ||
5137 | /* If basic block has predecessors or successors, redirect them. */ | |
5138 | if (remove_from_cfg_p | |
5139 | && (EDGE_COUNT (empty_bb->preds) > 0 | |
5140 | || EDGE_COUNT (empty_bb->succs) > 0)) | |
5141 | { | |
5142 | basic_block pred; | |
5143 | basic_block succ; | |
5144 | ||
5145 | /* We need to init PRED and SUCC before redirecting edges. */ | |
5146 | if (EDGE_COUNT (empty_bb->preds) > 0) | |
5147 | { | |
5148 | edge e; | |
5149 | ||
5150 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); | |
5151 | ||
5152 | e = EDGE_PRED (empty_bb, 0); | |
5153 | gcc_assert (e->src == empty_bb->prev_bb | |
5154 | && (e->flags & EDGE_FALLTHRU)); | |
5155 | ||
5156 | pred = empty_bb->prev_bb; | |
5157 | } | |
5158 | else | |
5159 | pred = NULL; | |
5160 | ||
5161 | if (EDGE_COUNT (empty_bb->succs) > 0) | |
5162 | { | |
5163 | /* We do not check fallthruness here as above, because | |
5164 | after removing a jump the edge may actually be not fallthru. */ | |
5165 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); | |
5166 | succ = EDGE_SUCC (empty_bb, 0)->dest; | |
5167 | } | |
5168 | else | |
5169 | succ = NULL; | |
5170 | ||
5171 | if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) | |
5172 | { | |
5173 | edge e = EDGE_PRED (empty_bb, 0); | |
5174 | ||
5175 | if (e->flags & EDGE_FALLTHRU) | |
5176 | redirect_edge_succ_nodup (e, succ); | |
5177 | else | |
5178 | sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); | |
5179 | } | |
5180 | ||
5181 | if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) | |
5182 | { | |
5183 | edge e = EDGE_SUCC (empty_bb, 0); | |
5184 | ||
5185 | if (find_edge (pred, e->dest) == NULL) | |
5186 | redirect_edge_pred (e, pred); | |
5187 | } | |
5188 | } | |
5189 | ||
5190 | /* Finish removing. */ | |
5191 | sel_remove_bb (empty_bb, remove_from_cfg_p); | |
5192 | } | |
5193 | ||
b8698a0f | 5194 | /* An implementation of create_basic_block hook, which additionally updates |
e855c69d AB |
5195 | per-bb data structures. */ |
5196 | static basic_block | |
5197 | sel_create_basic_block (void *headp, void *endp, basic_block after) | |
5198 | { | |
5199 | basic_block new_bb; | |
5200 | insn_t new_bb_note; | |
b8698a0f L |
5201 | |
5202 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
e855c69d AB |
5203 | || last_added_blocks == NULL); |
5204 | ||
5205 | new_bb_note = get_bb_note_from_pool (); | |
5206 | ||
5207 | if (new_bb_note == NULL_RTX) | |
5208 | new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); | |
5209 | else | |
5210 | { | |
5211 | new_bb = create_basic_block_structure ((rtx) headp, (rtx) endp, | |
5212 | new_bb_note, after); | |
5213 | new_bb->aux = NULL; | |
5214 | } | |
5215 | ||
5216 | VEC_safe_push (basic_block, heap, last_added_blocks, new_bb); | |
5217 | ||
5218 | return new_bb; | |
5219 | } | |
5220 | ||
5221 | /* Implement sched_init_only_bb (). */ | |
5222 | static void | |
5223 | sel_init_only_bb (basic_block bb, basic_block after) | |
5224 | { | |
5225 | gcc_assert (after == NULL); | |
5226 | ||
5227 | extend_regions (); | |
5228 | rgn_make_new_region_out_of_new_block (bb); | |
5229 | } | |
5230 | ||
5231 | /* Update the latch when we've splitted or merged it from FROM block to TO. | |
5232 | This should be checked for all outer loops, too. */ | |
5233 | static void | |
5234 | change_loops_latches (basic_block from, basic_block to) | |
5235 | { | |
5236 | gcc_assert (from != to); | |
5237 | ||
5238 | if (current_loop_nest) | |
5239 | { | |
5240 | struct loop *loop; | |
5241 | ||
5242 | for (loop = current_loop_nest; loop; loop = loop_outer (loop)) | |
5243 | if (considered_for_pipelining_p (loop) && loop->latch == from) | |
5244 | { | |
5245 | gcc_assert (loop == current_loop_nest); | |
5246 | loop->latch = to; | |
5247 | gcc_assert (loop_latch_edge (loop)); | |
5248 | } | |
5249 | } | |
5250 | } | |
5251 | ||
b8698a0f | 5252 | /* Splits BB on two basic blocks, adding it to the region and extending |
e855c69d AB |
5253 | per-bb data structures. Returns the newly created bb. */ |
5254 | static basic_block | |
5255 | sel_split_block (basic_block bb, rtx after) | |
5256 | { | |
5257 | basic_block new_bb; | |
5258 | insn_t insn; | |
5259 | ||
5260 | new_bb = sched_split_block_1 (bb, after); | |
5261 | sel_add_bb (new_bb); | |
5262 | ||
5263 | /* This should be called after sel_add_bb, because this uses | |
b8698a0f | 5264 | CONTAINING_RGN for the new block, which is not yet initialized. |
e855c69d AB |
5265 | FIXME: this function may be a no-op now. */ |
5266 | change_loops_latches (bb, new_bb); | |
5267 | ||
5268 | /* Update ORIG_BB_INDEX for insns moved into the new block. */ | |
5269 | FOR_BB_INSNS (new_bb, insn) | |
5270 | if (INSN_P (insn)) | |
5271 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; | |
5272 | ||
5273 | if (sel_bb_empty_p (bb)) | |
5274 | { | |
5275 | gcc_assert (!sel_bb_empty_p (new_bb)); | |
5276 | ||
5277 | /* NEW_BB has data sets that need to be updated and BB holds | |
5278 | data sets that should be removed. Exchange these data sets | |
5279 | so that we won't lose BB's valid data sets. */ | |
5280 | exchange_data_sets (new_bb, bb); | |
5281 | free_data_sets (bb); | |
5282 | } | |
5283 | ||
5284 | if (!sel_bb_empty_p (new_bb) | |
5285 | && bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
5286 | bitmap_set_bit (blocks_to_reschedule, new_bb->index); | |
5287 | ||
5288 | return new_bb; | |
5289 | } | |
5290 | ||
5291 | /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. | |
5292 | Otherwise returns NULL. */ | |
5293 | static rtx | |
5294 | check_for_new_jump (basic_block bb, int prev_max_uid) | |
5295 | { | |
5296 | rtx end; | |
5297 | ||
5298 | end = sel_bb_end (bb); | |
5299 | if (end && INSN_UID (end) >= prev_max_uid) | |
5300 | return end; | |
5301 | return NULL; | |
5302 | } | |
5303 | ||
b8698a0f | 5304 | /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. |
e855c69d AB |
5305 | New means having UID at least equal to PREV_MAX_UID. */ |
5306 | static rtx | |
5307 | find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) | |
5308 | { | |
5309 | rtx jump; | |
5310 | ||
5311 | /* Return immediately if no new insns were emitted. */ | |
5312 | if (get_max_uid () == prev_max_uid) | |
5313 | return NULL; | |
b8698a0f | 5314 | |
e855c69d AB |
5315 | /* Now check both blocks for new jumps. It will ever be only one. */ |
5316 | if ((jump = check_for_new_jump (from, prev_max_uid))) | |
5317 | return jump; | |
5318 | ||
5319 | if (jump_bb != NULL | |
5320 | && (jump = check_for_new_jump (jump_bb, prev_max_uid))) | |
5321 | return jump; | |
5322 | return NULL; | |
5323 | } | |
5324 | ||
5325 | /* Splits E and adds the newly created basic block to the current region. | |
5326 | Returns this basic block. */ | |
5327 | basic_block | |
5328 | sel_split_edge (edge e) | |
5329 | { | |
5330 | basic_block new_bb, src, other_bb = NULL; | |
5331 | int prev_max_uid; | |
5332 | rtx jump; | |
5333 | ||
5334 | src = e->src; | |
5335 | prev_max_uid = get_max_uid (); | |
5336 | new_bb = split_edge (e); | |
5337 | ||
b8698a0f | 5338 | if (flag_sel_sched_pipelining_outer_loops |
e855c69d AB |
5339 | && current_loop_nest) |
5340 | { | |
5341 | int i; | |
5342 | basic_block bb; | |
5343 | ||
b8698a0f | 5344 | /* Some of the basic blocks might not have been added to the loop. |
e855c69d | 5345 | Add them here, until this is fixed in force_fallthru. */ |
b8698a0f | 5346 | for (i = 0; |
e855c69d AB |
5347 | VEC_iterate (basic_block, last_added_blocks, i, bb); i++) |
5348 | if (!bb->loop_father) | |
5349 | { | |
5350 | add_bb_to_loop (bb, e->dest->loop_father); | |
5351 | ||
5352 | gcc_assert (!other_bb && (new_bb->index != bb->index)); | |
5353 | other_bb = bb; | |
5354 | } | |
5355 | } | |
5356 | ||
5357 | /* Add all last_added_blocks to the region. */ | |
5358 | sel_add_bb (NULL); | |
5359 | ||
5360 | jump = find_new_jump (src, new_bb, prev_max_uid); | |
5361 | if (jump) | |
5362 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5363 | ||
5364 | /* Put the correct lv set on this block. */ | |
5365 | if (other_bb && !sel_bb_empty_p (other_bb)) | |
5366 | compute_live (sel_bb_head (other_bb)); | |
5367 | ||
5368 | return new_bb; | |
5369 | } | |
5370 | ||
5371 | /* Implement sched_create_empty_bb (). */ | |
5372 | static basic_block | |
5373 | sel_create_empty_bb (basic_block after) | |
5374 | { | |
5375 | basic_block new_bb; | |
5376 | ||
5377 | new_bb = sched_create_empty_bb_1 (after); | |
5378 | ||
5379 | /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit | |
5380 | later. */ | |
5381 | gcc_assert (VEC_length (basic_block, last_added_blocks) == 1 | |
5382 | && VEC_index (basic_block, last_added_blocks, 0) == new_bb); | |
5383 | ||
5384 | VEC_free (basic_block, heap, last_added_blocks); | |
5385 | return new_bb; | |
5386 | } | |
5387 | ||
5388 | /* Implement sched_create_recovery_block. ORIG_INSN is where block | |
5389 | will be splitted to insert a check. */ | |
5390 | basic_block | |
5391 | sel_create_recovery_block (insn_t orig_insn) | |
5392 | { | |
5393 | basic_block first_bb, second_bb, recovery_block; | |
5394 | basic_block before_recovery = NULL; | |
5395 | rtx jump; | |
5396 | ||
5397 | first_bb = BLOCK_FOR_INSN (orig_insn); | |
5398 | if (sel_bb_end_p (orig_insn)) | |
5399 | { | |
5400 | /* Avoid introducing an empty block while splitting. */ | |
5401 | gcc_assert (single_succ_p (first_bb)); | |
5402 | second_bb = single_succ (first_bb); | |
5403 | } | |
5404 | else | |
5405 | second_bb = sched_split_block (first_bb, orig_insn); | |
5406 | ||
5407 | recovery_block = sched_create_recovery_block (&before_recovery); | |
5408 | if (before_recovery) | |
5409 | copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR); | |
5410 | ||
5411 | gcc_assert (sel_bb_empty_p (recovery_block)); | |
5412 | sched_create_recovery_edges (first_bb, recovery_block, second_bb); | |
5413 | if (current_loops != NULL) | |
5414 | add_bb_to_loop (recovery_block, first_bb->loop_father); | |
b8698a0f | 5415 | |
e855c69d | 5416 | sel_add_bb (recovery_block); |
b8698a0f | 5417 | |
e855c69d AB |
5418 | jump = BB_END (recovery_block); |
5419 | gcc_assert (sel_bb_head (recovery_block) == jump); | |
5420 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5421 | ||
5422 | return recovery_block; | |
5423 | } | |
5424 | ||
5425 | /* Merge basic block B into basic block A. */ | |
262d8232 | 5426 | static void |
e855c69d AB |
5427 | sel_merge_blocks (basic_block a, basic_block b) |
5428 | { | |
262d8232 AB |
5429 | gcc_assert (sel_bb_empty_p (b) |
5430 | && EDGE_COUNT (b->preds) == 1 | |
5431 | && EDGE_PRED (b, 0)->src == b->prev_bb); | |
e855c69d | 5432 | |
262d8232 AB |
5433 | move_bb_info (b->prev_bb, b); |
5434 | remove_empty_bb (b, false); | |
5435 | merge_blocks (a, b); | |
e855c69d AB |
5436 | change_loops_latches (b, a); |
5437 | } | |
5438 | ||
5439 | /* A wrapper for redirect_edge_and_branch_force, which also initializes | |
5440 | data structures for possibly created bb and insns. Returns the newly | |
5441 | added bb or NULL, when a bb was not needed. */ | |
5442 | void | |
5443 | sel_redirect_edge_and_branch_force (edge e, basic_block to) | |
5444 | { | |
00c4e97c | 5445 | basic_block jump_bb, src, orig_dest = e->dest; |
e855c69d AB |
5446 | int prev_max_uid; |
5447 | rtx jump; | |
b8698a0f | 5448 | |
00c4e97c AB |
5449 | /* This function is now used only for bookkeeping code creation, where |
5450 | we'll never get the single pred of orig_dest block and thus will not | |
5451 | hit unreachable blocks when updating dominator info. */ | |
5452 | gcc_assert (!sel_bb_empty_p (e->src) | |
5453 | && !single_pred_p (orig_dest)); | |
e855c69d AB |
5454 | src = e->src; |
5455 | prev_max_uid = get_max_uid (); | |
5456 | jump_bb = redirect_edge_and_branch_force (e, to); | |
5457 | ||
5458 | if (jump_bb != NULL) | |
5459 | sel_add_bb (jump_bb); | |
5460 | ||
5461 | /* This function could not be used to spoil the loop structure by now, | |
5462 | thus we don't care to update anything. But check it to be sure. */ | |
5463 | if (current_loop_nest | |
5464 | && pipelining_p) | |
5465 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
b8698a0f | 5466 | |
e855c69d AB |
5467 | jump = find_new_jump (src, jump_bb, prev_max_uid); |
5468 | if (jump) | |
5469 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
00c4e97c AB |
5470 | set_immediate_dominator (CDI_DOMINATORS, to, |
5471 | recompute_dominator (CDI_DOMINATORS, to)); | |
5472 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5473 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
e855c69d AB |
5474 | } |
5475 | ||
b59ab570 AM |
5476 | /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by |
5477 | redirected edge are in reverse topological order. */ | |
5478 | bool | |
e855c69d AB |
5479 | sel_redirect_edge_and_branch (edge e, basic_block to) |
5480 | { | |
5481 | bool latch_edge_p; | |
00c4e97c | 5482 | basic_block src, orig_dest = e->dest; |
e855c69d AB |
5483 | int prev_max_uid; |
5484 | rtx jump; | |
f2c45f08 | 5485 | edge redirected; |
b59ab570 | 5486 | bool recompute_toporder_p = false; |
00c4e97c | 5487 | bool maybe_unreachable = single_pred_p (orig_dest); |
e855c69d AB |
5488 | |
5489 | latch_edge_p = (pipelining_p | |
5490 | && current_loop_nest | |
5491 | && e == loop_latch_edge (current_loop_nest)); | |
5492 | ||
5493 | src = e->src; | |
5494 | prev_max_uid = get_max_uid (); | |
f2c45f08 AM |
5495 | |
5496 | redirected = redirect_edge_and_branch (e, to); | |
5497 | ||
5498 | gcc_assert (redirected && last_added_blocks == NULL); | |
e855c69d AB |
5499 | |
5500 | /* When we've redirected a latch edge, update the header. */ | |
5501 | if (latch_edge_p) | |
5502 | { | |
5503 | current_loop_nest->header = to; | |
5504 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
5505 | } | |
5506 | ||
b59ab570 AM |
5507 | /* In rare situations, the topological relation between the blocks connected |
5508 | by the redirected edge can change (see PR42245 for an example). Update | |
5509 | block_to_bb/bb_to_block. */ | |
5510 | if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) | |
5511 | && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) | |
5512 | recompute_toporder_p = true; | |
5513 | ||
e855c69d AB |
5514 | jump = find_new_jump (src, NULL, prev_max_uid); |
5515 | if (jump) | |
5516 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
b59ab570 | 5517 | |
00c4e97c AB |
5518 | /* Only update dominator info when we don't have unreachable blocks. |
5519 | Otherwise we'll update in maybe_tidy_empty_bb. */ | |
5520 | if (!maybe_unreachable) | |
5521 | { | |
5522 | set_immediate_dominator (CDI_DOMINATORS, to, | |
5523 | recompute_dominator (CDI_DOMINATORS, to)); | |
5524 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5525 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
5526 | } | |
b59ab570 | 5527 | return recompute_toporder_p; |
e855c69d AB |
5528 | } |
5529 | ||
5530 | /* This variable holds the cfg hooks used by the selective scheduler. */ | |
5531 | static struct cfg_hooks sel_cfg_hooks; | |
5532 | ||
5533 | /* Register sel-sched cfg hooks. */ | |
5534 | void | |
5535 | sel_register_cfg_hooks (void) | |
5536 | { | |
5537 | sched_split_block = sel_split_block; | |
5538 | ||
5539 | orig_cfg_hooks = get_cfg_hooks (); | |
5540 | sel_cfg_hooks = orig_cfg_hooks; | |
5541 | ||
5542 | sel_cfg_hooks.create_basic_block = sel_create_basic_block; | |
5543 | ||
5544 | set_cfg_hooks (sel_cfg_hooks); | |
5545 | ||
5546 | sched_init_only_bb = sel_init_only_bb; | |
5547 | sched_split_block = sel_split_block; | |
5548 | sched_create_empty_bb = sel_create_empty_bb; | |
5549 | } | |
5550 | ||
5551 | /* Unregister sel-sched cfg hooks. */ | |
5552 | void | |
5553 | sel_unregister_cfg_hooks (void) | |
5554 | { | |
5555 | sched_create_empty_bb = NULL; | |
5556 | sched_split_block = NULL; | |
5557 | sched_init_only_bb = NULL; | |
5558 | ||
5559 | set_cfg_hooks (orig_cfg_hooks); | |
5560 | } | |
5561 | \f | |
5562 | ||
5563 | /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, | |
5564 | LABEL is where this jump should be directed. */ | |
5565 | rtx | |
5566 | create_insn_rtx_from_pattern (rtx pattern, rtx label) | |
5567 | { | |
5568 | rtx insn_rtx; | |
5569 | ||
5570 | gcc_assert (!INSN_P (pattern)); | |
5571 | ||
5572 | start_sequence (); | |
5573 | ||
5574 | if (label == NULL_RTX) | |
5575 | insn_rtx = emit_insn (pattern); | |
b5b8b0ac AO |
5576 | else if (DEBUG_INSN_P (label)) |
5577 | insn_rtx = emit_debug_insn (pattern); | |
e855c69d AB |
5578 | else |
5579 | { | |
5580 | insn_rtx = emit_jump_insn (pattern); | |
5581 | JUMP_LABEL (insn_rtx) = label; | |
5582 | ++LABEL_NUSES (label); | |
5583 | } | |
5584 | ||
5585 | end_sequence (); | |
5586 | ||
5587 | sched_init_luids (NULL, NULL, NULL, NULL); | |
5588 | sched_extend_target (); | |
5589 | sched_deps_init (false); | |
5590 | ||
5591 | /* Initialize INSN_CODE now. */ | |
5592 | recog_memoized (insn_rtx); | |
5593 | return insn_rtx; | |
5594 | } | |
5595 | ||
5596 | /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn | |
5597 | must not be clonable. */ | |
5598 | vinsn_t | |
5599 | create_vinsn_from_insn_rtx (rtx insn_rtx, bool force_unique_p) | |
5600 | { | |
5601 | gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); | |
5602 | ||
5603 | /* If VINSN_TYPE is not USE, retain its uniqueness. */ | |
5604 | return vinsn_create (insn_rtx, force_unique_p); | |
5605 | } | |
5606 | ||
5607 | /* Create a copy of INSN_RTX. */ | |
5608 | rtx | |
5609 | create_copy_of_insn_rtx (rtx insn_rtx) | |
5610 | { | |
5611 | rtx res; | |
5612 | ||
b5b8b0ac AO |
5613 | if (DEBUG_INSN_P (insn_rtx)) |
5614 | return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5615 | insn_rtx); | |
5616 | ||
e855c69d AB |
5617 | gcc_assert (NONJUMP_INSN_P (insn_rtx)); |
5618 | ||
5619 | res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5620 | NULL_RTX); | |
5621 | return res; | |
5622 | } | |
5623 | ||
5624 | /* Change vinsn field of EXPR to hold NEW_VINSN. */ | |
5625 | void | |
5626 | change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) | |
5627 | { | |
5628 | vinsn_detach (EXPR_VINSN (expr)); | |
5629 | ||
5630 | EXPR_VINSN (expr) = new_vinsn; | |
5631 | vinsn_attach (new_vinsn); | |
5632 | } | |
5633 | ||
5634 | /* Helpers for global init. */ | |
5635 | /* This structure is used to be able to call existing bundling mechanism | |
5636 | and calculate insn priorities. */ | |
b8698a0f | 5637 | static struct haifa_sched_info sched_sel_haifa_sched_info = |
e855c69d AB |
5638 | { |
5639 | NULL, /* init_ready_list */ | |
5640 | NULL, /* can_schedule_ready_p */ | |
5641 | NULL, /* schedule_more_p */ | |
5642 | NULL, /* new_ready */ | |
5643 | NULL, /* rgn_rank */ | |
5644 | sel_print_insn, /* rgn_print_insn */ | |
5645 | contributes_to_priority, | |
356c23b3 | 5646 | NULL, /* insn_finishes_block_p */ |
e855c69d AB |
5647 | |
5648 | NULL, NULL, | |
5649 | NULL, NULL, | |
5650 | 0, 0, | |
5651 | ||
5652 | NULL, /* add_remove_insn */ | |
5653 | NULL, /* begin_schedule_ready */ | |
86014d07 | 5654 | NULL, /* begin_move_insn */ |
e855c69d AB |
5655 | NULL, /* advance_target_bb */ |
5656 | SEL_SCHED | NEW_BBS | |
5657 | }; | |
5658 | ||
5659 | /* Setup special insns used in the scheduler. */ | |
b8698a0f | 5660 | void |
e855c69d AB |
5661 | setup_nop_and_exit_insns (void) |
5662 | { | |
5663 | gcc_assert (nop_pattern == NULL_RTX | |
5664 | && exit_insn == NULL_RTX); | |
5665 | ||
9ef1bf71 | 5666 | nop_pattern = constm1_rtx; |
e855c69d AB |
5667 | |
5668 | start_sequence (); | |
5669 | emit_insn (nop_pattern); | |
5670 | exit_insn = get_insns (); | |
5671 | end_sequence (); | |
5672 | set_block_for_insn (exit_insn, EXIT_BLOCK_PTR); | |
5673 | } | |
5674 | ||
5675 | /* Free special insns used in the scheduler. */ | |
5676 | void | |
5677 | free_nop_and_exit_insns (void) | |
5678 | { | |
5679 | exit_insn = NULL_RTX; | |
5680 | nop_pattern = NULL_RTX; | |
5681 | } | |
5682 | ||
5683 | /* Setup a special vinsn used in new insns initialization. */ | |
5684 | void | |
5685 | setup_nop_vinsn (void) | |
5686 | { | |
5687 | nop_vinsn = vinsn_create (exit_insn, false); | |
5688 | vinsn_attach (nop_vinsn); | |
5689 | } | |
5690 | ||
5691 | /* Free a special vinsn used in new insns initialization. */ | |
5692 | void | |
5693 | free_nop_vinsn (void) | |
5694 | { | |
5695 | gcc_assert (VINSN_COUNT (nop_vinsn) == 1); | |
5696 | vinsn_detach (nop_vinsn); | |
5697 | nop_vinsn = NULL; | |
5698 | } | |
5699 | ||
5700 | /* Call a set_sched_flags hook. */ | |
5701 | void | |
5702 | sel_set_sched_flags (void) | |
5703 | { | |
b8698a0f | 5704 | /* ??? This means that set_sched_flags were called, and we decided to |
e855c69d | 5705 | support speculation. However, set_sched_flags also modifies flags |
b8698a0f | 5706 | on current_sched_info, doing this only at global init. And we |
e855c69d AB |
5707 | sometimes change c_s_i later. So put the correct flags again. */ |
5708 | if (spec_info && targetm.sched.set_sched_flags) | |
5709 | targetm.sched.set_sched_flags (spec_info); | |
5710 | } | |
5711 | ||
5712 | /* Setup pointers to global sched info structures. */ | |
5713 | void | |
5714 | sel_setup_sched_infos (void) | |
5715 | { | |
5716 | rgn_setup_common_sched_info (); | |
5717 | ||
5718 | memcpy (&sel_common_sched_info, common_sched_info, | |
5719 | sizeof (sel_common_sched_info)); | |
5720 | ||
5721 | sel_common_sched_info.fix_recovery_cfg = NULL; | |
5722 | sel_common_sched_info.add_block = NULL; | |
5723 | sel_common_sched_info.estimate_number_of_insns | |
5724 | = sel_estimate_number_of_insns; | |
5725 | sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; | |
5726 | sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; | |
5727 | ||
5728 | common_sched_info = &sel_common_sched_info; | |
5729 | ||
5730 | current_sched_info = &sched_sel_haifa_sched_info; | |
b8698a0f | 5731 | current_sched_info->sched_max_insns_priority = |
e855c69d | 5732 | get_rgn_sched_max_insns_priority (); |
b8698a0f | 5733 | |
e855c69d AB |
5734 | sel_set_sched_flags (); |
5735 | } | |
5736 | \f | |
5737 | ||
5738 | /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, | |
5739 | *BB_ORD_INDEX after that is increased. */ | |
5740 | static void | |
5741 | sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) | |
5742 | { | |
5743 | RGN_NR_BLOCKS (rgn) += 1; | |
5744 | RGN_DONT_CALC_DEPS (rgn) = 0; | |
5745 | RGN_HAS_REAL_EBB (rgn) = 0; | |
5746 | CONTAINING_RGN (bb->index) = rgn; | |
5747 | BLOCK_TO_BB (bb->index) = *bb_ord_index; | |
5748 | rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; | |
5749 | (*bb_ord_index)++; | |
5750 | ||
5751 | /* FIXME: it is true only when not scheduling ebbs. */ | |
5752 | RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); | |
5753 | } | |
5754 | ||
5755 | /* Functions to support pipelining of outer loops. */ | |
5756 | ||
5757 | /* Creates a new empty region and returns it's number. */ | |
5758 | static int | |
5759 | sel_create_new_region (void) | |
5760 | { | |
5761 | int new_rgn_number = nr_regions; | |
5762 | ||
5763 | RGN_NR_BLOCKS (new_rgn_number) = 0; | |
5764 | ||
5765 | /* FIXME: This will work only when EBBs are not created. */ | |
5766 | if (new_rgn_number != 0) | |
b8698a0f | 5767 | RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + |
e855c69d AB |
5768 | RGN_NR_BLOCKS (new_rgn_number - 1); |
5769 | else | |
5770 | RGN_BLOCKS (new_rgn_number) = 0; | |
5771 | ||
5772 | /* Set the blocks of the next region so the other functions may | |
5773 | calculate the number of blocks in the region. */ | |
b8698a0f | 5774 | RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + |
e855c69d AB |
5775 | RGN_NR_BLOCKS (new_rgn_number); |
5776 | ||
5777 | nr_regions++; | |
5778 | ||
5779 | return new_rgn_number; | |
5780 | } | |
5781 | ||
5782 | /* If X has a smaller topological sort number than Y, returns -1; | |
5783 | if greater, returns 1. */ | |
5784 | static int | |
5785 | bb_top_order_comparator (const void *x, const void *y) | |
5786 | { | |
5787 | basic_block bb1 = *(const basic_block *) x; | |
5788 | basic_block bb2 = *(const basic_block *) y; | |
5789 | ||
b8698a0f L |
5790 | gcc_assert (bb1 == bb2 |
5791 | || rev_top_order_index[bb1->index] | |
e855c69d AB |
5792 | != rev_top_order_index[bb2->index]); |
5793 | ||
5794 | /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so | |
5795 | bbs with greater number should go earlier. */ | |
5796 | if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) | |
5797 | return -1; | |
5798 | else | |
5799 | return 1; | |
5800 | } | |
5801 | ||
b8698a0f | 5802 | /* Create a region for LOOP and return its number. If we don't want |
e855c69d AB |
5803 | to pipeline LOOP, return -1. */ |
5804 | static int | |
5805 | make_region_from_loop (struct loop *loop) | |
5806 | { | |
5807 | unsigned int i; | |
5808 | int new_rgn_number = -1; | |
5809 | struct loop *inner; | |
5810 | ||
5811 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5812 | int bb_ord_index = 0; | |
5813 | basic_block *loop_blocks; | |
5814 | basic_block preheader_block; | |
5815 | ||
b8698a0f | 5816 | if (loop->num_nodes |
e855c69d AB |
5817 | > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS)) |
5818 | return -1; | |
b8698a0f | 5819 | |
e855c69d AB |
5820 | /* Don't pipeline loops whose latch belongs to some of its inner loops. */ |
5821 | for (inner = loop->inner; inner; inner = inner->inner) | |
5822 | if (flow_bb_inside_loop_p (inner, loop->latch)) | |
5823 | return -1; | |
5824 | ||
5825 | loop->ninsns = num_loop_insns (loop); | |
5826 | if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS)) | |
5827 | return -1; | |
5828 | ||
5829 | loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); | |
5830 | ||
5831 | for (i = 0; i < loop->num_nodes; i++) | |
5832 | if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) | |
5833 | { | |
5834 | free (loop_blocks); | |
5835 | return -1; | |
5836 | } | |
5837 | ||
5838 | preheader_block = loop_preheader_edge (loop)->src; | |
5839 | gcc_assert (preheader_block); | |
5840 | gcc_assert (loop_blocks[0] == loop->header); | |
5841 | ||
5842 | new_rgn_number = sel_create_new_region (); | |
5843 | ||
5844 | sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number); | |
5845 | SET_BIT (bbs_in_loop_rgns, preheader_block->index); | |
5846 | ||
5847 | for (i = 0; i < loop->num_nodes; i++) | |
5848 | { | |
5849 | /* Add only those blocks that haven't been scheduled in the inner loop. | |
5850 | The exception is the basic blocks with bookkeeping code - they should | |
b8698a0f | 5851 | be added to the region (and they actually don't belong to the loop |
e855c69d AB |
5852 | body, but to the region containing that loop body). */ |
5853 | ||
5854 | gcc_assert (new_rgn_number >= 0); | |
5855 | ||
5856 | if (! TEST_BIT (bbs_in_loop_rgns, loop_blocks[i]->index)) | |
5857 | { | |
b8698a0f | 5858 | sel_add_block_to_region (loop_blocks[i], &bb_ord_index, |
e855c69d AB |
5859 | new_rgn_number); |
5860 | SET_BIT (bbs_in_loop_rgns, loop_blocks[i]->index); | |
5861 | } | |
5862 | } | |
5863 | ||
5864 | free (loop_blocks); | |
5865 | MARK_LOOP_FOR_PIPELINING (loop); | |
5866 | ||
5867 | return new_rgn_number; | |
5868 | } | |
5869 | ||
5870 | /* Create a new region from preheader blocks LOOP_BLOCKS. */ | |
5871 | void | |
5872 | make_region_from_loop_preheader (VEC(basic_block, heap) **loop_blocks) | |
5873 | { | |
5874 | unsigned int i; | |
5875 | int new_rgn_number = -1; | |
5876 | basic_block bb; | |
5877 | ||
5878 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5879 | int bb_ord_index = 0; | |
5880 | ||
5881 | new_rgn_number = sel_create_new_region (); | |
5882 | ||
ac47786e | 5883 | FOR_EACH_VEC_ELT (basic_block, *loop_blocks, i, bb) |
e855c69d AB |
5884 | { |
5885 | gcc_assert (new_rgn_number >= 0); | |
5886 | ||
5887 | sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number); | |
5888 | } | |
5889 | ||
5890 | VEC_free (basic_block, heap, *loop_blocks); | |
5891 | gcc_assert (*loop_blocks == NULL); | |
5892 | } | |
5893 | ||
5894 | ||
5895 | /* Create region(s) from loop nest LOOP, such that inner loops will be | |
b8698a0f | 5896 | pipelined before outer loops. Returns true when a region for LOOP |
e855c69d AB |
5897 | is created. */ |
5898 | static bool | |
5899 | make_regions_from_loop_nest (struct loop *loop) | |
b8698a0f | 5900 | { |
e855c69d AB |
5901 | struct loop *cur_loop; |
5902 | int rgn_number; | |
5903 | ||
5904 | /* Traverse all inner nodes of the loop. */ | |
5905 | for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) | |
5906 | if (! TEST_BIT (bbs_in_loop_rgns, cur_loop->header->index)) | |
5907 | return false; | |
5908 | ||
5909 | /* At this moment all regular inner loops should have been pipelined. | |
5910 | Try to create a region from this loop. */ | |
5911 | rgn_number = make_region_from_loop (loop); | |
5912 | ||
5913 | if (rgn_number < 0) | |
5914 | return false; | |
5915 | ||
5916 | VEC_safe_push (loop_p, heap, loop_nests, loop); | |
5917 | return true; | |
5918 | } | |
5919 | ||
5920 | /* Initalize data structures needed. */ | |
5921 | void | |
5922 | sel_init_pipelining (void) | |
5923 | { | |
5924 | /* Collect loop information to be used in outer loops pipelining. */ | |
5925 | loop_optimizer_init (LOOPS_HAVE_PREHEADERS | |
5926 | | LOOPS_HAVE_FALLTHRU_PREHEADERS | |
5927 | | LOOPS_HAVE_RECORDED_EXITS | |
5928 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); | |
5929 | current_loop_nest = NULL; | |
5930 | ||
5931 | bbs_in_loop_rgns = sbitmap_alloc (last_basic_block); | |
5932 | sbitmap_zero (bbs_in_loop_rgns); | |
5933 | ||
5934 | recompute_rev_top_order (); | |
5935 | } | |
5936 | ||
5937 | /* Returns a struct loop for region RGN. */ | |
5938 | loop_p | |
5939 | get_loop_nest_for_rgn (unsigned int rgn) | |
5940 | { | |
5941 | /* Regions created with extend_rgns don't have corresponding loop nests, | |
5942 | because they don't represent loops. */ | |
5943 | if (rgn < VEC_length (loop_p, loop_nests)) | |
5944 | return VEC_index (loop_p, loop_nests, rgn); | |
5945 | else | |
5946 | return NULL; | |
5947 | } | |
5948 | ||
5949 | /* True when LOOP was included into pipelining regions. */ | |
5950 | bool | |
5951 | considered_for_pipelining_p (struct loop *loop) | |
5952 | { | |
5953 | if (loop_depth (loop) == 0) | |
5954 | return false; | |
5955 | ||
b8698a0f L |
5956 | /* Now, the loop could be too large or irreducible. Check whether its |
5957 | region is in LOOP_NESTS. | |
5958 | We determine the region number of LOOP as the region number of its | |
5959 | latch. We can't use header here, because this header could be | |
e855c69d AB |
5960 | just removed preheader and it will give us the wrong region number. |
5961 | Latch can't be used because it could be in the inner loop too. */ | |
8ec4d0ad | 5962 | if (LOOP_MARKED_FOR_PIPELINING_P (loop)) |
e855c69d AB |
5963 | { |
5964 | int rgn = CONTAINING_RGN (loop->latch->index); | |
5965 | ||
5966 | gcc_assert ((unsigned) rgn < VEC_length (loop_p, loop_nests)); | |
5967 | return true; | |
5968 | } | |
b8698a0f | 5969 | |
e855c69d AB |
5970 | return false; |
5971 | } | |
5972 | ||
b8698a0f | 5973 | /* Makes regions from the rest of the blocks, after loops are chosen |
e855c69d AB |
5974 | for pipelining. */ |
5975 | static void | |
5976 | make_regions_from_the_rest (void) | |
5977 | { | |
5978 | int cur_rgn_blocks; | |
5979 | int *loop_hdr; | |
5980 | int i; | |
5981 | ||
5982 | basic_block bb; | |
5983 | edge e; | |
5984 | edge_iterator ei; | |
5985 | int *degree; | |
e855c69d AB |
5986 | |
5987 | /* Index in rgn_bb_table where to start allocating new regions. */ | |
5988 | cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; | |
e855c69d | 5989 | |
b8698a0f | 5990 | /* Make regions from all the rest basic blocks - those that don't belong to |
e855c69d AB |
5991 | any loop or belong to irreducible loops. Prepare the data structures |
5992 | for extend_rgns. */ | |
5993 | ||
5994 | /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, | |
5995 | LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same | |
5996 | loop. */ | |
5997 | loop_hdr = XNEWVEC (int, last_basic_block); | |
5998 | degree = XCNEWVEC (int, last_basic_block); | |
5999 | ||
6000 | ||
6001 | /* For each basic block that belongs to some loop assign the number | |
6002 | of innermost loop it belongs to. */ | |
6003 | for (i = 0; i < last_basic_block; i++) | |
6004 | loop_hdr[i] = -1; | |
6005 | ||
6006 | FOR_EACH_BB (bb) | |
6007 | { | |
6008 | if (bb->loop_father && !bb->loop_father->num == 0 | |
6009 | && !(bb->flags & BB_IRREDUCIBLE_LOOP)) | |
6010 | loop_hdr[bb->index] = bb->loop_father->num; | |
6011 | } | |
6012 | ||
b8698a0f | 6013 | /* For each basic block degree is calculated as the number of incoming |
e855c69d AB |
6014 | edges, that are going out of bbs that are not yet scheduled. |
6015 | The basic blocks that are scheduled have degree value of zero. */ | |
b8698a0f | 6016 | FOR_EACH_BB (bb) |
e855c69d AB |
6017 | { |
6018 | degree[bb->index] = 0; | |
6019 | ||
6020 | if (!TEST_BIT (bbs_in_loop_rgns, bb->index)) | |
6021 | { | |
6022 | FOR_EACH_EDGE (e, ei, bb->preds) | |
6023 | if (!TEST_BIT (bbs_in_loop_rgns, e->src->index)) | |
6024 | degree[bb->index]++; | |
6025 | } | |
6026 | else | |
6027 | degree[bb->index] = -1; | |
6028 | } | |
6029 | ||
6030 | extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); | |
6031 | ||
6032 | /* Any block that did not end up in a region is placed into a region | |
6033 | by itself. */ | |
6034 | FOR_EACH_BB (bb) | |
6035 | if (degree[bb->index] >= 0) | |
6036 | { | |
6037 | rgn_bb_table[cur_rgn_blocks] = bb->index; | |
6038 | RGN_NR_BLOCKS (nr_regions) = 1; | |
6039 | RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; | |
6040 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
6041 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
6042 | CONTAINING_RGN (bb->index) = nr_regions++; | |
6043 | BLOCK_TO_BB (bb->index) = 0; | |
6044 | } | |
6045 | ||
6046 | free (degree); | |
6047 | free (loop_hdr); | |
6048 | } | |
6049 | ||
6050 | /* Free data structures used in pipelining of loops. */ | |
6051 | void sel_finish_pipelining (void) | |
6052 | { | |
6053 | loop_iterator li; | |
6054 | struct loop *loop; | |
6055 | ||
6056 | /* Release aux fields so we don't free them later by mistake. */ | |
6057 | FOR_EACH_LOOP (li, loop, 0) | |
6058 | loop->aux = NULL; | |
6059 | ||
6060 | loop_optimizer_finalize (); | |
6061 | ||
6062 | VEC_free (loop_p, heap, loop_nests); | |
6063 | ||
6064 | free (rev_top_order_index); | |
6065 | rev_top_order_index = NULL; | |
6066 | } | |
6067 | ||
b8698a0f | 6068 | /* This function replaces the find_rgns when |
e855c69d | 6069 | FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ |
b8698a0f | 6070 | void |
e855c69d AB |
6071 | sel_find_rgns (void) |
6072 | { | |
6073 | sel_init_pipelining (); | |
6074 | extend_regions (); | |
6075 | ||
6076 | if (current_loops) | |
6077 | { | |
6078 | loop_p loop; | |
6079 | loop_iterator li; | |
6080 | ||
6081 | FOR_EACH_LOOP (li, loop, (flag_sel_sched_pipelining_outer_loops | |
6082 | ? LI_FROM_INNERMOST | |
6083 | : LI_ONLY_INNERMOST)) | |
6084 | make_regions_from_loop_nest (loop); | |
6085 | } | |
6086 | ||
6087 | /* Make regions from all the rest basic blocks and schedule them. | |
b8698a0f | 6088 | These blocks include blocks that don't belong to any loop or belong |
e855c69d AB |
6089 | to irreducible loops. */ |
6090 | make_regions_from_the_rest (); | |
6091 | ||
6092 | /* We don't need bbs_in_loop_rgns anymore. */ | |
6093 | sbitmap_free (bbs_in_loop_rgns); | |
6094 | bbs_in_loop_rgns = NULL; | |
6095 | } | |
6096 | ||
ea4d630f AM |
6097 | /* Add the preheader blocks from previous loop to current region taking |
6098 | it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS. | |
e855c69d AB |
6099 | This function is only used with -fsel-sched-pipelining-outer-loops. */ |
6100 | void | |
ea4d630f | 6101 | sel_add_loop_preheaders (bb_vec_t *bbs) |
e855c69d AB |
6102 | { |
6103 | int i; | |
6104 | basic_block bb; | |
b8698a0f | 6105 | VEC(basic_block, heap) *preheader_blocks |
e855c69d AB |
6106 | = LOOP_PREHEADER_BLOCKS (current_loop_nest); |
6107 | ||
6108 | for (i = 0; | |
6109 | VEC_iterate (basic_block, preheader_blocks, i, bb); | |
6110 | i++) | |
8ec4d0ad | 6111 | { |
ea4d630f | 6112 | VEC_safe_push (basic_block, heap, *bbs, bb); |
8ec4d0ad | 6113 | VEC_safe_push (basic_block, heap, last_added_blocks, bb); |
e855c69d | 6114 | sel_add_bb (bb); |
8ec4d0ad | 6115 | } |
e855c69d AB |
6116 | |
6117 | VEC_free (basic_block, heap, preheader_blocks); | |
6118 | } | |
6119 | ||
b8698a0f L |
6120 | /* While pipelining outer loops, returns TRUE if BB is a loop preheader. |
6121 | Please note that the function should also work when pipelining_p is | |
6122 | false, because it is used when deciding whether we should or should | |
e855c69d AB |
6123 | not reschedule pipelined code. */ |
6124 | bool | |
6125 | sel_is_loop_preheader_p (basic_block bb) | |
6126 | { | |
6127 | if (current_loop_nest) | |
6128 | { | |
6129 | struct loop *outer; | |
6130 | ||
6131 | if (preheader_removed) | |
6132 | return false; | |
6133 | ||
6134 | /* Preheader is the first block in the region. */ | |
6135 | if (BLOCK_TO_BB (bb->index) == 0) | |
6136 | return true; | |
6137 | ||
6138 | /* We used to find a preheader with the topological information. | |
6139 | Check that the above code is equivalent to what we did before. */ | |
6140 | ||
6141 | if (in_current_region_p (current_loop_nest->header)) | |
b8698a0f | 6142 | gcc_assert (!(BLOCK_TO_BB (bb->index) |
e855c69d AB |
6143 | < BLOCK_TO_BB (current_loop_nest->header->index))); |
6144 | ||
6145 | /* Support the situation when the latch block of outer loop | |
6146 | could be from here. */ | |
6147 | for (outer = loop_outer (current_loop_nest); | |
6148 | outer; | |
6149 | outer = loop_outer (outer)) | |
6150 | if (considered_for_pipelining_p (outer) && outer->latch == bb) | |
6151 | gcc_unreachable (); | |
6152 | } | |
6153 | ||
6154 | return false; | |
6155 | } | |
6156 | ||
753de8cf AM |
6157 | /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and |
6158 | can be removed, making the corresponding edge fallthrough (assuming that | |
6159 | all basic blocks between JUMP_BB and DEST_BB are empty). */ | |
6160 | static bool | |
6161 | bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb) | |
e855c69d | 6162 | { |
b4550bf7 AM |
6163 | if (!onlyjump_p (BB_END (jump_bb)) |
6164 | || tablejump_p (BB_END (jump_bb), NULL, NULL)) | |
e855c69d AB |
6165 | return false; |
6166 | ||
b8698a0f | 6167 | /* Several outgoing edges, abnormal edge or destination of jump is |
e855c69d AB |
6168 | not DEST_BB. */ |
6169 | if (EDGE_COUNT (jump_bb->succs) != 1 | |
753de8cf | 6170 | || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING) |
e855c69d AB |
6171 | || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) |
6172 | return false; | |
6173 | ||
6174 | /* If not anything of the upper. */ | |
6175 | return true; | |
6176 | } | |
6177 | ||
6178 | /* Removes the loop preheader from the current region and saves it in | |
b8698a0f | 6179 | PREHEADER_BLOCKS of the father loop, so they will be added later to |
e855c69d AB |
6180 | region that represents an outer loop. */ |
6181 | static void | |
6182 | sel_remove_loop_preheader (void) | |
6183 | { | |
6184 | int i, old_len; | |
6185 | int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
6186 | basic_block bb; | |
6187 | bool all_empty_p = true; | |
b8698a0f | 6188 | VEC(basic_block, heap) *preheader_blocks |
e855c69d AB |
6189 | = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); |
6190 | ||
6191 | gcc_assert (current_loop_nest); | |
6192 | old_len = VEC_length (basic_block, preheader_blocks); | |
6193 | ||
6194 | /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ | |
6195 | for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) | |
6196 | { | |
6197 | bb = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6198 | ||
b8698a0f | 6199 | /* If the basic block belongs to region, but doesn't belong to |
e855c69d AB |
6200 | corresponding loop, then it should be a preheader. */ |
6201 | if (sel_is_loop_preheader_p (bb)) | |
6202 | { | |
6203 | VEC_safe_push (basic_block, heap, preheader_blocks, bb); | |
6204 | if (BB_END (bb) != bb_note (bb)) | |
6205 | all_empty_p = false; | |
6206 | } | |
6207 | } | |
b8698a0f | 6208 | |
e855c69d AB |
6209 | /* Remove these blocks only after iterating over the whole region. */ |
6210 | for (i = VEC_length (basic_block, preheader_blocks) - 1; | |
6211 | i >= old_len; | |
6212 | i--) | |
6213 | { | |
b8698a0f | 6214 | bb = VEC_index (basic_block, preheader_blocks, i); |
e855c69d AB |
6215 | sel_remove_bb (bb, false); |
6216 | } | |
6217 | ||
6218 | if (!considered_for_pipelining_p (loop_outer (current_loop_nest))) | |
6219 | { | |
6220 | if (!all_empty_p) | |
6221 | /* Immediately create new region from preheader. */ | |
6222 | make_region_from_loop_preheader (&preheader_blocks); | |
6223 | else | |
6224 | { | |
6225 | /* If all preheader blocks are empty - dont create new empty region. | |
6226 | Instead, remove them completely. */ | |
ac47786e | 6227 | FOR_EACH_VEC_ELT (basic_block, preheader_blocks, i, bb) |
e855c69d AB |
6228 | { |
6229 | edge e; | |
6230 | edge_iterator ei; | |
6231 | basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; | |
6232 | ||
6233 | /* Redirect all incoming edges to next basic block. */ | |
6234 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) | |
6235 | { | |
6236 | if (! (e->flags & EDGE_FALLTHRU)) | |
6237 | redirect_edge_and_branch (e, bb->next_bb); | |
6238 | else | |
6239 | redirect_edge_succ (e, bb->next_bb); | |
6240 | } | |
6241 | gcc_assert (BB_NOTE_LIST (bb) == NULL); | |
6242 | delete_and_free_basic_block (bb); | |
6243 | ||
b8698a0f L |
6244 | /* Check if after deleting preheader there is a nonconditional |
6245 | jump in PREV_BB that leads to the next basic block NEXT_BB. | |
6246 | If it is so - delete this jump and clear data sets of its | |
e855c69d AB |
6247 | basic block if it becomes empty. */ |
6248 | if (next_bb->prev_bb == prev_bb | |
6249 | && prev_bb != ENTRY_BLOCK_PTR | |
753de8cf | 6250 | && bb_has_removable_jump_to_p (prev_bb, next_bb)) |
e855c69d AB |
6251 | { |
6252 | redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); | |
6253 | if (BB_END (prev_bb) == bb_note (prev_bb)) | |
6254 | free_data_sets (prev_bb); | |
6255 | } | |
00c4e97c AB |
6256 | |
6257 | set_immediate_dominator (CDI_DOMINATORS, next_bb, | |
6258 | recompute_dominator (CDI_DOMINATORS, | |
6259 | next_bb)); | |
e855c69d AB |
6260 | } |
6261 | } | |
6262 | VEC_free (basic_block, heap, preheader_blocks); | |
6263 | } | |
6264 | else | |
6265 | /* Store preheader within the father's loop structure. */ | |
6266 | SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), | |
6267 | preheader_blocks); | |
6268 | } | |
6269 | #endif |