]>
Commit | Line | Data |
---|---|---|
e855c69d | 1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
c02e2d5c | 2 | Copyright (C) 2006, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. |
e855c69d AB |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8 | Software Foundation; either version 3, or (at your option) any later | |
9 | version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING3. If not see | |
18 | <http://www.gnu.org/licenses/>. */ | |
19 | ||
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
23 | #include "tm.h" | |
718f9c0f | 24 | #include "diagnostic-core.h" |
e855c69d AB |
25 | #include "rtl.h" |
26 | #include "tm_p.h" | |
27 | #include "hard-reg-set.h" | |
28 | #include "regs.h" | |
29 | #include "function.h" | |
30 | #include "flags.h" | |
31 | #include "insn-config.h" | |
32 | #include "insn-attr.h" | |
33 | #include "except.h" | |
e855c69d AB |
34 | #include "recog.h" |
35 | #include "params.h" | |
36 | #include "target.h" | |
37 | #include "timevar.h" | |
38 | #include "tree-pass.h" | |
39 | #include "sched-int.h" | |
40 | #include "ggc.h" | |
41 | #include "tree.h" | |
42 | #include "vec.h" | |
43 | #include "langhooks.h" | |
44 | #include "rtlhooks-def.h" | |
5936d944 | 45 | #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */ |
e855c69d AB |
46 | |
47 | #ifdef INSN_SCHEDULING | |
48 | #include "sel-sched-ir.h" | |
49 | /* We don't have to use it except for sel_print_insn. */ | |
50 | #include "sel-sched-dump.h" | |
51 | ||
52 | /* A vector holding bb info for whole scheduling pass. */ | |
53 | VEC(sel_global_bb_info_def, heap) *sel_global_bb_info = NULL; | |
54 | ||
55 | /* A vector holding bb info. */ | |
56 | VEC(sel_region_bb_info_def, heap) *sel_region_bb_info = NULL; | |
57 | ||
58 | /* A pool for allocating all lists. */ | |
59 | alloc_pool sched_lists_pool; | |
60 | ||
61 | /* This contains information about successors for compute_av_set. */ | |
62 | struct succs_info current_succs; | |
63 | ||
64 | /* Data structure to describe interaction with the generic scheduler utils. */ | |
65 | static struct common_sched_info_def sel_common_sched_info; | |
66 | ||
67 | /* The loop nest being pipelined. */ | |
68 | struct loop *current_loop_nest; | |
69 | ||
70 | /* LOOP_NESTS is a vector containing the corresponding loop nest for | |
71 | each region. */ | |
72 | static VEC(loop_p, heap) *loop_nests = NULL; | |
73 | ||
74 | /* Saves blocks already in loop regions, indexed by bb->index. */ | |
75 | static sbitmap bbs_in_loop_rgns = NULL; | |
76 | ||
77 | /* CFG hooks that are saved before changing create_basic_block hook. */ | |
78 | static struct cfg_hooks orig_cfg_hooks; | |
79 | \f | |
80 | ||
81 | /* Array containing reverse topological index of function basic blocks, | |
82 | indexed by BB->INDEX. */ | |
83 | static int *rev_top_order_index = NULL; | |
84 | ||
85 | /* Length of the above array. */ | |
86 | static int rev_top_order_index_len = -1; | |
87 | ||
88 | /* A regset pool structure. */ | |
89 | static struct | |
90 | { | |
91 | /* The stack to which regsets are returned. */ | |
92 | regset *v; | |
93 | ||
94 | /* Its pointer. */ | |
95 | int n; | |
96 | ||
97 | /* Its size. */ | |
98 | int s; | |
99 | ||
100 | /* In VV we save all generated regsets so that, when destructing the | |
101 | pool, we can compare it with V and check that every regset was returned | |
102 | back to pool. */ | |
103 | regset *vv; | |
104 | ||
105 | /* The pointer of VV stack. */ | |
106 | int nn; | |
107 | ||
108 | /* Its size. */ | |
109 | int ss; | |
110 | ||
111 | /* The difference between allocated and returned regsets. */ | |
112 | int diff; | |
113 | } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 }; | |
114 | ||
115 | /* This represents the nop pool. */ | |
116 | static struct | |
117 | { | |
118 | /* The vector which holds previously emitted nops. */ | |
119 | insn_t *v; | |
120 | ||
121 | /* Its pointer. */ | |
122 | int n; | |
123 | ||
124 | /* Its size. */ | |
b8698a0f | 125 | int s; |
e855c69d AB |
126 | } nop_pool = { NULL, 0, 0 }; |
127 | ||
128 | /* The pool for basic block notes. */ | |
129 | static rtx_vec_t bb_note_pool; | |
130 | ||
131 | /* A NOP pattern used to emit placeholder insns. */ | |
132 | rtx nop_pattern = NULL_RTX; | |
133 | /* A special instruction that resides in EXIT_BLOCK. | |
134 | EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */ | |
135 | rtx exit_insn = NULL_RTX; | |
136 | ||
b8698a0f | 137 | /* TRUE if while scheduling current region, which is loop, its preheader |
e855c69d AB |
138 | was removed. */ |
139 | bool preheader_removed = false; | |
140 | \f | |
141 | ||
142 | /* Forward static declarations. */ | |
143 | static void fence_clear (fence_t); | |
144 | ||
145 | static void deps_init_id (idata_t, insn_t, bool); | |
146 | static void init_id_from_df (idata_t, insn_t, bool); | |
147 | static expr_t set_insn_init (expr_t, vinsn_t, int); | |
148 | ||
149 | static void cfg_preds (basic_block, insn_t **, int *); | |
150 | static void prepare_insn_expr (insn_t, int); | |
151 | static void free_history_vect (VEC (expr_history_def, heap) **); | |
152 | ||
153 | static void move_bb_info (basic_block, basic_block); | |
154 | static void remove_empty_bb (basic_block, bool); | |
262d8232 | 155 | static void sel_merge_blocks (basic_block, basic_block); |
e855c69d | 156 | static void sel_remove_loop_preheader (void); |
753de8cf | 157 | static bool bb_has_removable_jump_to_p (basic_block, basic_block); |
e855c69d AB |
158 | |
159 | static bool insn_is_the_only_one_in_bb_p (insn_t); | |
160 | static void create_initial_data_sets (basic_block); | |
161 | ||
b5b8b0ac | 162 | static void free_av_set (basic_block); |
e855c69d AB |
163 | static void invalidate_av_set (basic_block); |
164 | static void extend_insn_data (void); | |
165 | static void sel_init_new_insn (insn_t, int); | |
166 | static void finish_insns (void); | |
167 | \f | |
168 | /* Various list functions. */ | |
169 | ||
170 | /* Copy an instruction list L. */ | |
171 | ilist_t | |
172 | ilist_copy (ilist_t l) | |
173 | { | |
174 | ilist_t head = NULL, *tailp = &head; | |
175 | ||
176 | while (l) | |
177 | { | |
178 | ilist_add (tailp, ILIST_INSN (l)); | |
179 | tailp = &ILIST_NEXT (*tailp); | |
180 | l = ILIST_NEXT (l); | |
181 | } | |
182 | ||
183 | return head; | |
184 | } | |
185 | ||
186 | /* Invert an instruction list L. */ | |
187 | ilist_t | |
188 | ilist_invert (ilist_t l) | |
189 | { | |
190 | ilist_t res = NULL; | |
191 | ||
192 | while (l) | |
193 | { | |
194 | ilist_add (&res, ILIST_INSN (l)); | |
195 | l = ILIST_NEXT (l); | |
196 | } | |
197 | ||
198 | return res; | |
199 | } | |
200 | ||
201 | /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */ | |
202 | void | |
203 | blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc) | |
204 | { | |
205 | bnd_t bnd; | |
206 | ||
207 | _list_add (lp); | |
208 | bnd = BLIST_BND (*lp); | |
209 | ||
210 | BND_TO (bnd) = to; | |
211 | BND_PTR (bnd) = ptr; | |
212 | BND_AV (bnd) = NULL; | |
213 | BND_AV1 (bnd) = NULL; | |
214 | BND_DC (bnd) = dc; | |
215 | } | |
216 | ||
217 | /* Remove the list note pointed to by LP. */ | |
218 | void | |
219 | blist_remove (blist_t *lp) | |
220 | { | |
221 | bnd_t b = BLIST_BND (*lp); | |
222 | ||
223 | av_set_clear (&BND_AV (b)); | |
224 | av_set_clear (&BND_AV1 (b)); | |
225 | ilist_clear (&BND_PTR (b)); | |
226 | ||
227 | _list_remove (lp); | |
228 | } | |
229 | ||
230 | /* Init a fence tail L. */ | |
231 | void | |
232 | flist_tail_init (flist_tail_t l) | |
233 | { | |
234 | FLIST_TAIL_HEAD (l) = NULL; | |
235 | FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l); | |
236 | } | |
237 | ||
238 | /* Try to find fence corresponding to INSN in L. */ | |
239 | fence_t | |
240 | flist_lookup (flist_t l, insn_t insn) | |
241 | { | |
242 | while (l) | |
243 | { | |
244 | if (FENCE_INSN (FLIST_FENCE (l)) == insn) | |
245 | return FLIST_FENCE (l); | |
246 | ||
247 | l = FLIST_NEXT (l); | |
248 | } | |
249 | ||
250 | return NULL; | |
251 | } | |
252 | ||
253 | /* Init the fields of F before running fill_insns. */ | |
254 | static void | |
255 | init_fence_for_scheduling (fence_t f) | |
256 | { | |
257 | FENCE_BNDS (f) = NULL; | |
258 | FENCE_PROCESSED_P (f) = false; | |
259 | FENCE_SCHEDULED_P (f) = false; | |
260 | } | |
261 | ||
262 | /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */ | |
263 | static void | |
b8698a0f L |
264 | flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, |
265 | insn_t last_scheduled_insn, VEC(rtx,gc) *executing_insns, | |
266 | int *ready_ticks, int ready_ticks_size, insn_t sched_next, | |
136e01a3 | 267 | int cycle, int cycle_issued_insns, int issue_more, |
e855c69d AB |
268 | bool starts_cycle_p, bool after_stall_p) |
269 | { | |
270 | fence_t f; | |
271 | ||
272 | _list_add (lp); | |
273 | f = FLIST_FENCE (*lp); | |
274 | ||
275 | FENCE_INSN (f) = insn; | |
276 | ||
277 | gcc_assert (state != NULL); | |
278 | FENCE_STATE (f) = state; | |
279 | ||
280 | FENCE_CYCLE (f) = cycle; | |
281 | FENCE_ISSUED_INSNS (f) = cycle_issued_insns; | |
282 | FENCE_STARTS_CYCLE_P (f) = starts_cycle_p; | |
283 | FENCE_AFTER_STALL_P (f) = after_stall_p; | |
284 | ||
285 | gcc_assert (dc != NULL); | |
286 | FENCE_DC (f) = dc; | |
287 | ||
288 | gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL); | |
289 | FENCE_TC (f) = tc; | |
290 | ||
291 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
136e01a3 | 292 | FENCE_ISSUE_MORE (f) = issue_more; |
e855c69d AB |
293 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
294 | FENCE_READY_TICKS (f) = ready_ticks; | |
295 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
296 | FENCE_SCHED_NEXT (f) = sched_next; | |
297 | ||
298 | init_fence_for_scheduling (f); | |
299 | } | |
300 | ||
301 | /* Remove the head node of the list pointed to by LP. */ | |
302 | static void | |
303 | flist_remove (flist_t *lp) | |
304 | { | |
305 | if (FENCE_INSN (FLIST_FENCE (*lp))) | |
306 | fence_clear (FLIST_FENCE (*lp)); | |
307 | _list_remove (lp); | |
308 | } | |
309 | ||
310 | /* Clear the fence list pointed to by LP. */ | |
311 | void | |
312 | flist_clear (flist_t *lp) | |
313 | { | |
314 | while (*lp) | |
315 | flist_remove (lp); | |
316 | } | |
317 | ||
318 | /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */ | |
319 | void | |
320 | def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call) | |
321 | { | |
322 | def_t d; | |
b8698a0f | 323 | |
e855c69d AB |
324 | _list_add (dl); |
325 | d = DEF_LIST_DEF (*dl); | |
326 | ||
327 | d->orig_insn = original_insn; | |
328 | d->crosses_call = crosses_call; | |
329 | } | |
330 | \f | |
331 | ||
332 | /* Functions to work with target contexts. */ | |
333 | ||
b8698a0f | 334 | /* Bulk target context. It is convenient for debugging purposes to ensure |
e855c69d AB |
335 | that there are no uninitialized (null) target contexts. */ |
336 | static tc_t bulk_tc = (tc_t) 1; | |
337 | ||
b8698a0f | 338 | /* Target hooks wrappers. In the future we can provide some default |
e855c69d AB |
339 | implementations for them. */ |
340 | ||
341 | /* Allocate a store for the target context. */ | |
342 | static tc_t | |
343 | alloc_target_context (void) | |
344 | { | |
345 | return (targetm.sched.alloc_sched_context | |
346 | ? targetm.sched.alloc_sched_context () : bulk_tc); | |
347 | } | |
348 | ||
349 | /* Init target context TC. | |
350 | If CLEAN_P is true, then make TC as it is beginning of the scheduler. | |
351 | Overwise, copy current backend context to TC. */ | |
352 | static void | |
353 | init_target_context (tc_t tc, bool clean_p) | |
354 | { | |
355 | if (targetm.sched.init_sched_context) | |
356 | targetm.sched.init_sched_context (tc, clean_p); | |
357 | } | |
358 | ||
359 | /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as | |
360 | int init_target_context (). */ | |
361 | tc_t | |
362 | create_target_context (bool clean_p) | |
363 | { | |
364 | tc_t tc = alloc_target_context (); | |
365 | ||
366 | init_target_context (tc, clean_p); | |
367 | return tc; | |
368 | } | |
369 | ||
370 | /* Copy TC to the current backend context. */ | |
371 | void | |
372 | set_target_context (tc_t tc) | |
373 | { | |
374 | if (targetm.sched.set_sched_context) | |
375 | targetm.sched.set_sched_context (tc); | |
376 | } | |
377 | ||
378 | /* TC is about to be destroyed. Free any internal data. */ | |
379 | static void | |
380 | clear_target_context (tc_t tc) | |
381 | { | |
382 | if (targetm.sched.clear_sched_context) | |
383 | targetm.sched.clear_sched_context (tc); | |
384 | } | |
385 | ||
386 | /* Clear and free it. */ | |
387 | static void | |
388 | delete_target_context (tc_t tc) | |
389 | { | |
390 | clear_target_context (tc); | |
391 | ||
392 | if (targetm.sched.free_sched_context) | |
393 | targetm.sched.free_sched_context (tc); | |
394 | } | |
395 | ||
396 | /* Make a copy of FROM in TO. | |
397 | NB: May be this should be a hook. */ | |
398 | static void | |
399 | copy_target_context (tc_t to, tc_t from) | |
400 | { | |
401 | tc_t tmp = create_target_context (false); | |
402 | ||
403 | set_target_context (from); | |
404 | init_target_context (to, false); | |
405 | ||
406 | set_target_context (tmp); | |
407 | delete_target_context (tmp); | |
408 | } | |
409 | ||
410 | /* Create a copy of TC. */ | |
411 | static tc_t | |
412 | create_copy_of_target_context (tc_t tc) | |
413 | { | |
414 | tc_t copy = alloc_target_context (); | |
415 | ||
416 | copy_target_context (copy, tc); | |
417 | ||
418 | return copy; | |
419 | } | |
420 | ||
421 | /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P | |
422 | is the same as in init_target_context (). */ | |
423 | void | |
424 | reset_target_context (tc_t tc, bool clean_p) | |
425 | { | |
426 | clear_target_context (tc); | |
427 | init_target_context (tc, clean_p); | |
428 | } | |
429 | \f | |
b8698a0f | 430 | /* Functions to work with dependence contexts. |
88302d54 | 431 | Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence |
e855c69d AB |
432 | context. It accumulates information about processed insns to decide if |
433 | current insn is dependent on the processed ones. */ | |
434 | ||
435 | /* Make a copy of FROM in TO. */ | |
436 | static void | |
437 | copy_deps_context (deps_t to, deps_t from) | |
438 | { | |
bcf33775 | 439 | init_deps (to, false); |
e855c69d AB |
440 | deps_join (to, from); |
441 | } | |
442 | ||
443 | /* Allocate store for dep context. */ | |
444 | static deps_t | |
445 | alloc_deps_context (void) | |
446 | { | |
88302d54 | 447 | return XNEW (struct deps_desc); |
e855c69d AB |
448 | } |
449 | ||
450 | /* Allocate and initialize dep context. */ | |
451 | static deps_t | |
452 | create_deps_context (void) | |
453 | { | |
454 | deps_t dc = alloc_deps_context (); | |
455 | ||
bcf33775 | 456 | init_deps (dc, false); |
e855c69d AB |
457 | return dc; |
458 | } | |
459 | ||
460 | /* Create a copy of FROM. */ | |
461 | static deps_t | |
462 | create_copy_of_deps_context (deps_t from) | |
463 | { | |
464 | deps_t to = alloc_deps_context (); | |
465 | ||
466 | copy_deps_context (to, from); | |
467 | return to; | |
468 | } | |
469 | ||
470 | /* Clean up internal data of DC. */ | |
471 | static void | |
472 | clear_deps_context (deps_t dc) | |
473 | { | |
474 | free_deps (dc); | |
475 | } | |
476 | ||
477 | /* Clear and free DC. */ | |
478 | static void | |
479 | delete_deps_context (deps_t dc) | |
480 | { | |
481 | clear_deps_context (dc); | |
482 | free (dc); | |
483 | } | |
484 | ||
485 | /* Clear and init DC. */ | |
486 | static void | |
487 | reset_deps_context (deps_t dc) | |
488 | { | |
489 | clear_deps_context (dc); | |
bcf33775 | 490 | init_deps (dc, false); |
e855c69d AB |
491 | } |
492 | ||
b8698a0f | 493 | /* This structure describes the dependence analysis hooks for advancing |
e855c69d AB |
494 | dependence context. */ |
495 | static struct sched_deps_info_def advance_deps_context_sched_deps_info = | |
496 | { | |
497 | NULL, | |
498 | ||
499 | NULL, /* start_insn */ | |
500 | NULL, /* finish_insn */ | |
501 | NULL, /* start_lhs */ | |
502 | NULL, /* finish_lhs */ | |
503 | NULL, /* start_rhs */ | |
504 | NULL, /* finish_rhs */ | |
505 | haifa_note_reg_set, | |
506 | haifa_note_reg_clobber, | |
507 | haifa_note_reg_use, | |
508 | NULL, /* note_mem_dep */ | |
509 | NULL, /* note_dep */ | |
510 | ||
511 | 0, 0, 0 | |
512 | }; | |
513 | ||
514 | /* Process INSN and add its impact on DC. */ | |
515 | void | |
516 | advance_deps_context (deps_t dc, insn_t insn) | |
517 | { | |
518 | sched_deps_info = &advance_deps_context_sched_deps_info; | |
519 | deps_analyze_insn (dc, insn); | |
520 | } | |
521 | \f | |
522 | ||
523 | /* Functions to work with DFA states. */ | |
524 | ||
525 | /* Allocate store for a DFA state. */ | |
526 | static state_t | |
527 | state_alloc (void) | |
528 | { | |
529 | return xmalloc (dfa_state_size); | |
530 | } | |
531 | ||
532 | /* Allocate and initialize DFA state. */ | |
533 | static state_t | |
534 | state_create (void) | |
535 | { | |
536 | state_t state = state_alloc (); | |
537 | ||
538 | state_reset (state); | |
539 | advance_state (state); | |
540 | return state; | |
541 | } | |
542 | ||
543 | /* Free DFA state. */ | |
544 | static void | |
545 | state_free (state_t state) | |
546 | { | |
547 | free (state); | |
548 | } | |
549 | ||
550 | /* Make a copy of FROM in TO. */ | |
551 | static void | |
552 | state_copy (state_t to, state_t from) | |
553 | { | |
554 | memcpy (to, from, dfa_state_size); | |
555 | } | |
556 | ||
557 | /* Create a copy of FROM. */ | |
558 | static state_t | |
559 | state_create_copy (state_t from) | |
560 | { | |
561 | state_t to = state_alloc (); | |
562 | ||
563 | state_copy (to, from); | |
564 | return to; | |
565 | } | |
566 | \f | |
567 | ||
568 | /* Functions to work with fences. */ | |
569 | ||
570 | /* Clear the fence. */ | |
571 | static void | |
572 | fence_clear (fence_t f) | |
573 | { | |
574 | state_t s = FENCE_STATE (f); | |
575 | deps_t dc = FENCE_DC (f); | |
576 | void *tc = FENCE_TC (f); | |
577 | ||
578 | ilist_clear (&FENCE_BNDS (f)); | |
579 | ||
580 | gcc_assert ((s != NULL && dc != NULL && tc != NULL) | |
581 | || (s == NULL && dc == NULL && tc == NULL)); | |
582 | ||
583 | if (s != NULL) | |
584 | free (s); | |
585 | ||
586 | if (dc != NULL) | |
587 | delete_deps_context (dc); | |
588 | ||
589 | if (tc != NULL) | |
590 | delete_target_context (tc); | |
591 | VEC_free (rtx, gc, FENCE_EXECUTING_INSNS (f)); | |
592 | free (FENCE_READY_TICKS (f)); | |
593 | FENCE_READY_TICKS (f) = NULL; | |
594 | } | |
595 | ||
596 | /* Init a list of fences with successors of OLD_FENCE. */ | |
597 | void | |
598 | init_fences (insn_t old_fence) | |
599 | { | |
600 | insn_t succ; | |
601 | succ_iterator si; | |
602 | bool first = true; | |
603 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f L |
604 | |
605 | FOR_EACH_SUCC_1 (succ, si, old_fence, | |
e855c69d AB |
606 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
607 | { | |
b8698a0f | 608 | |
e855c69d AB |
609 | if (first) |
610 | first = false; | |
611 | else | |
612 | gcc_assert (flag_sel_sched_pipelining_outer_loops); | |
613 | ||
614 | flist_add (&fences, succ, | |
615 | state_create (), | |
616 | create_deps_context () /* dc */, | |
617 | create_target_context (true) /* tc */, | |
b8698a0f | 618 | NULL_RTX /* last_scheduled_insn */, |
e855c69d AB |
619 | NULL, /* executing_insns */ |
620 | XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ | |
621 | ready_ticks_size, | |
622 | NULL_RTX /* sched_next */, | |
b8698a0f | 623 | 1 /* cycle */, 0 /* cycle_issued_insns */, |
136e01a3 | 624 | issue_rate, /* issue_more */ |
b8698a0f | 625 | 1 /* starts_cycle_p */, 0 /* after_stall_p */); |
e855c69d AB |
626 | } |
627 | } | |
628 | ||
629 | /* Merges two fences (filling fields of fence F with resulting values) by | |
630 | following rules: 1) state, target context and last scheduled insn are | |
b8698a0f | 631 | propagated from fallthrough edge if it is available; |
e855c69d | 632 | 2) deps context and cycle is propagated from more probable edge; |
b8698a0f | 633 | 3) all other fields are set to corresponding constant values. |
e855c69d | 634 | |
b8698a0f | 635 | INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, |
136e01a3 AB |
636 | READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE |
637 | and AFTER_STALL_P are the corresponding fields of the second fence. */ | |
e855c69d AB |
638 | static void |
639 | merge_fences (fence_t f, insn_t insn, | |
b8698a0f | 640 | state_t state, deps_t dc, void *tc, |
e855c69d AB |
641 | rtx last_scheduled_insn, VEC(rtx, gc) *executing_insns, |
642 | int *ready_ticks, int ready_ticks_size, | |
136e01a3 | 643 | rtx sched_next, int cycle, int issue_more, bool after_stall_p) |
e855c69d AB |
644 | { |
645 | insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f); | |
646 | ||
647 | gcc_assert (sel_bb_head_p (FENCE_INSN (f)) | |
648 | && !sched_next && !FENCE_SCHED_NEXT (f)); | |
649 | ||
b8698a0f | 650 | /* Check if we can decide which path fences came. |
e855c69d AB |
651 | If we can't (or don't want to) - reset all. */ |
652 | if (last_scheduled_insn == NULL | |
653 | || last_scheduled_insn_old == NULL | |
b8698a0f L |
654 | /* This is a case when INSN is reachable on several paths from |
655 | one insn (this can happen when pipelining of outer loops is on and | |
656 | there are two edges: one going around of inner loop and the other - | |
e855c69d AB |
657 | right through it; in such case just reset everything). */ |
658 | || last_scheduled_insn == last_scheduled_insn_old) | |
659 | { | |
660 | state_reset (FENCE_STATE (f)); | |
661 | state_free (state); | |
b8698a0f | 662 | |
e855c69d AB |
663 | reset_deps_context (FENCE_DC (f)); |
664 | delete_deps_context (dc); | |
b8698a0f | 665 | |
e855c69d AB |
666 | reset_target_context (FENCE_TC (f), true); |
667 | delete_target_context (tc); | |
668 | ||
669 | if (cycle > FENCE_CYCLE (f)) | |
670 | FENCE_CYCLE (f) = cycle; | |
671 | ||
672 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; | |
136e01a3 | 673 | FENCE_ISSUE_MORE (f) = issue_rate; |
e855c69d AB |
674 | VEC_free (rtx, gc, executing_insns); |
675 | free (ready_ticks); | |
676 | if (FENCE_EXECUTING_INSNS (f)) | |
b8698a0f | 677 | VEC_block_remove (rtx, FENCE_EXECUTING_INSNS (f), 0, |
e855c69d AB |
678 | VEC_length (rtx, FENCE_EXECUTING_INSNS (f))); |
679 | if (FENCE_READY_TICKS (f)) | |
680 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
681 | } | |
682 | else | |
683 | { | |
684 | edge edge_old = NULL, edge_new = NULL; | |
685 | edge candidate; | |
686 | succ_iterator si; | |
687 | insn_t succ; | |
b8698a0f | 688 | |
e855c69d AB |
689 | /* Find fallthrough edge. */ |
690 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); | |
0fd4b31d | 691 | candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb); |
e855c69d AB |
692 | |
693 | if (!candidate | |
694 | || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn) | |
695 | && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old))) | |
696 | { | |
697 | /* No fallthrough edge leading to basic block of INSN. */ | |
698 | state_reset (FENCE_STATE (f)); | |
699 | state_free (state); | |
b8698a0f | 700 | |
e855c69d AB |
701 | reset_target_context (FENCE_TC (f), true); |
702 | delete_target_context (tc); | |
b8698a0f | 703 | |
e855c69d | 704 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; |
136e01a3 | 705 | FENCE_ISSUE_MORE (f) = issue_rate; |
e855c69d AB |
706 | } |
707 | else | |
708 | if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn)) | |
709 | { | |
b8698a0f | 710 | /* Would be weird if same insn is successor of several fallthrough |
e855c69d AB |
711 | edges. */ |
712 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
713 | != BLOCK_FOR_INSN (last_scheduled_insn_old)); | |
714 | ||
715 | state_free (FENCE_STATE (f)); | |
716 | FENCE_STATE (f) = state; | |
717 | ||
718 | delete_target_context (FENCE_TC (f)); | |
719 | FENCE_TC (f) = tc; | |
720 | ||
721 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
136e01a3 | 722 | FENCE_ISSUE_MORE (f) = issue_more; |
e855c69d AB |
723 | } |
724 | else | |
725 | { | |
726 | /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ | |
727 | state_free (state); | |
728 | delete_target_context (tc); | |
729 | ||
730 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
731 | != BLOCK_FOR_INSN (last_scheduled_insn)); | |
732 | } | |
733 | ||
734 | /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ | |
735 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, | |
736 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
737 | { | |
738 | if (succ == insn) | |
739 | { | |
740 | /* No same successor allowed from several edges. */ | |
741 | gcc_assert (!edge_old); | |
742 | edge_old = si.e1; | |
743 | } | |
744 | } | |
745 | /* Find edge of second predecessor (last_scheduled_insn->insn). */ | |
746 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, | |
747 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
748 | { | |
749 | if (succ == insn) | |
750 | { | |
751 | /* No same successor allowed from several edges. */ | |
752 | gcc_assert (!edge_new); | |
753 | edge_new = si.e1; | |
754 | } | |
755 | } | |
756 | ||
757 | /* Check if we can choose most probable predecessor. */ | |
758 | if (edge_old == NULL || edge_new == NULL) | |
759 | { | |
760 | reset_deps_context (FENCE_DC (f)); | |
761 | delete_deps_context (dc); | |
762 | VEC_free (rtx, gc, executing_insns); | |
763 | free (ready_ticks); | |
b8698a0f | 764 | |
e855c69d AB |
765 | FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); |
766 | if (FENCE_EXECUTING_INSNS (f)) | |
b8698a0f | 767 | VEC_block_remove (rtx, FENCE_EXECUTING_INSNS (f), 0, |
e855c69d AB |
768 | VEC_length (rtx, FENCE_EXECUTING_INSNS (f))); |
769 | if (FENCE_READY_TICKS (f)) | |
770 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
771 | } | |
772 | else | |
773 | if (edge_new->probability > edge_old->probability) | |
774 | { | |
775 | delete_deps_context (FENCE_DC (f)); | |
776 | FENCE_DC (f) = dc; | |
777 | VEC_free (rtx, gc, FENCE_EXECUTING_INSNS (f)); | |
778 | FENCE_EXECUTING_INSNS (f) = executing_insns; | |
779 | free (FENCE_READY_TICKS (f)); | |
780 | FENCE_READY_TICKS (f) = ready_ticks; | |
781 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
782 | FENCE_CYCLE (f) = cycle; | |
783 | } | |
784 | else | |
785 | { | |
786 | /* Leave DC and CYCLE untouched. */ | |
787 | delete_deps_context (dc); | |
788 | VEC_free (rtx, gc, executing_insns); | |
789 | free (ready_ticks); | |
790 | } | |
791 | } | |
792 | ||
793 | /* Fill remaining invariant fields. */ | |
794 | if (after_stall_p) | |
795 | FENCE_AFTER_STALL_P (f) = 1; | |
796 | ||
797 | FENCE_ISSUED_INSNS (f) = 0; | |
798 | FENCE_STARTS_CYCLE_P (f) = 1; | |
799 | FENCE_SCHED_NEXT (f) = NULL; | |
800 | } | |
801 | ||
b8698a0f | 802 | /* Add a new fence to NEW_FENCES list, initializing it from all |
e855c69d AB |
803 | other parameters. */ |
804 | static void | |
805 | add_to_fences (flist_tail_t new_fences, insn_t insn, | |
b8698a0f L |
806 | state_t state, deps_t dc, void *tc, rtx last_scheduled_insn, |
807 | VEC(rtx, gc) *executing_insns, int *ready_ticks, | |
808 | int ready_ticks_size, rtx sched_next, int cycle, | |
136e01a3 AB |
809 | int cycle_issued_insns, int issue_rate, |
810 | bool starts_cycle_p, bool after_stall_p) | |
e855c69d AB |
811 | { |
812 | fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn); | |
813 | ||
814 | if (! f) | |
815 | { | |
816 | flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc, | |
b8698a0f | 817 | last_scheduled_insn, executing_insns, ready_ticks, |
e855c69d | 818 | ready_ticks_size, sched_next, cycle, cycle_issued_insns, |
136e01a3 | 819 | issue_rate, starts_cycle_p, after_stall_p); |
e855c69d AB |
820 | |
821 | FLIST_TAIL_TAILP (new_fences) | |
822 | = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences)); | |
823 | } | |
824 | else | |
825 | { | |
b8698a0f L |
826 | merge_fences (f, insn, state, dc, tc, last_scheduled_insn, |
827 | executing_insns, ready_ticks, ready_ticks_size, | |
136e01a3 | 828 | sched_next, cycle, issue_rate, after_stall_p); |
e855c69d AB |
829 | } |
830 | } | |
831 | ||
832 | /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */ | |
833 | void | |
834 | move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences) | |
835 | { | |
836 | fence_t f, old; | |
837 | flist_t *tailp = FLIST_TAIL_TAILP (new_fences); | |
838 | ||
839 | old = FLIST_FENCE (old_fences); | |
b8698a0f | 840 | f = flist_lookup (FLIST_TAIL_HEAD (new_fences), |
e855c69d AB |
841 | FENCE_INSN (FLIST_FENCE (old_fences))); |
842 | if (f) | |
843 | { | |
844 | merge_fences (f, old->insn, old->state, old->dc, old->tc, | |
845 | old->last_scheduled_insn, old->executing_insns, | |
846 | old->ready_ticks, old->ready_ticks_size, | |
136e01a3 | 847 | old->sched_next, old->cycle, old->issue_more, |
e855c69d AB |
848 | old->after_stall_p); |
849 | } | |
850 | else | |
851 | { | |
852 | _list_add (tailp); | |
853 | FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp); | |
854 | *FLIST_FENCE (*tailp) = *old; | |
855 | init_fence_for_scheduling (FLIST_FENCE (*tailp)); | |
856 | } | |
857 | FENCE_INSN (old) = NULL; | |
858 | } | |
859 | ||
b8698a0f | 860 | /* Add a new fence to NEW_FENCES list and initialize most of its data |
e855c69d AB |
861 | as a clean one. */ |
862 | void | |
863 | add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
864 | { | |
865 | int ready_ticks_size = get_max_uid () + 1; | |
b8698a0f | 866 | |
e855c69d AB |
867 | add_to_fences (new_fences, |
868 | succ, state_create (), create_deps_context (), | |
869 | create_target_context (true), | |
b8698a0f | 870 | NULL_RTX, NULL, |
e855c69d AB |
871 | XCNEWVEC (int, ready_ticks_size), ready_ticks_size, |
872 | NULL_RTX, FENCE_CYCLE (fence) + 1, | |
136e01a3 | 873 | 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence)); |
e855c69d AB |
874 | } |
875 | ||
b8698a0f | 876 | /* Add a new fence to NEW_FENCES list and initialize all of its data |
e855c69d AB |
877 | from FENCE and SUCC. */ |
878 | void | |
879 | add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
880 | { | |
b8698a0f | 881 | int * new_ready_ticks |
e855c69d | 882 | = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); |
b8698a0f | 883 | |
e855c69d AB |
884 | memcpy (new_ready_ticks, FENCE_READY_TICKS (fence), |
885 | FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); | |
886 | add_to_fences (new_fences, | |
887 | succ, state_create_copy (FENCE_STATE (fence)), | |
888 | create_copy_of_deps_context (FENCE_DC (fence)), | |
889 | create_copy_of_target_context (FENCE_TC (fence)), | |
b8698a0f | 890 | FENCE_LAST_SCHEDULED_INSN (fence), |
e855c69d AB |
891 | VEC_copy (rtx, gc, FENCE_EXECUTING_INSNS (fence)), |
892 | new_ready_ticks, | |
893 | FENCE_READY_TICKS_SIZE (fence), | |
894 | FENCE_SCHED_NEXT (fence), | |
895 | FENCE_CYCLE (fence), | |
896 | FENCE_ISSUED_INSNS (fence), | |
136e01a3 | 897 | FENCE_ISSUE_MORE (fence), |
e855c69d AB |
898 | FENCE_STARTS_CYCLE_P (fence), |
899 | FENCE_AFTER_STALL_P (fence)); | |
900 | } | |
901 | \f | |
902 | ||
903 | /* Functions to work with regset and nop pools. */ | |
904 | ||
905 | /* Returns the new regset from pool. It might have some of the bits set | |
906 | from the previous usage. */ | |
907 | regset | |
908 | get_regset_from_pool (void) | |
909 | { | |
910 | regset rs; | |
911 | ||
912 | if (regset_pool.n != 0) | |
913 | rs = regset_pool.v[--regset_pool.n]; | |
914 | else | |
915 | /* We need to create the regset. */ | |
916 | { | |
917 | rs = ALLOC_REG_SET (®_obstack); | |
918 | ||
919 | if (regset_pool.nn == regset_pool.ss) | |
920 | regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, | |
921 | (regset_pool.ss = 2 * regset_pool.ss + 1)); | |
922 | regset_pool.vv[regset_pool.nn++] = rs; | |
923 | } | |
924 | ||
925 | regset_pool.diff++; | |
926 | ||
927 | return rs; | |
928 | } | |
929 | ||
930 | /* Same as above, but returns the empty regset. */ | |
931 | regset | |
932 | get_clear_regset_from_pool (void) | |
933 | { | |
934 | regset rs = get_regset_from_pool (); | |
935 | ||
936 | CLEAR_REG_SET (rs); | |
937 | return rs; | |
938 | } | |
939 | ||
940 | /* Return regset RS to the pool for future use. */ | |
941 | void | |
942 | return_regset_to_pool (regset rs) | |
943 | { | |
9ef1bf71 | 944 | gcc_assert (rs); |
e855c69d AB |
945 | regset_pool.diff--; |
946 | ||
947 | if (regset_pool.n == regset_pool.s) | |
948 | regset_pool.v = XRESIZEVEC (regset, regset_pool.v, | |
949 | (regset_pool.s = 2 * regset_pool.s + 1)); | |
950 | regset_pool.v[regset_pool.n++] = rs; | |
951 | } | |
952 | ||
68ad446f | 953 | #ifdef ENABLE_CHECKING |
e855c69d AB |
954 | /* This is used as a qsort callback for sorting regset pool stacks. |
955 | X and XX are addresses of two regsets. They are never equal. */ | |
956 | static int | |
957 | cmp_v_in_regset_pool (const void *x, const void *xx) | |
958 | { | |
959 | return *((const regset *) x) - *((const regset *) xx); | |
960 | } | |
68ad446f | 961 | #endif |
e855c69d AB |
962 | |
963 | /* Free the regset pool possibly checking for memory leaks. */ | |
964 | void | |
965 | free_regset_pool (void) | |
966 | { | |
967 | #ifdef ENABLE_CHECKING | |
968 | { | |
969 | regset *v = regset_pool.v; | |
970 | int i = 0; | |
971 | int n = regset_pool.n; | |
b8698a0f | 972 | |
e855c69d AB |
973 | regset *vv = regset_pool.vv; |
974 | int ii = 0; | |
975 | int nn = regset_pool.nn; | |
b8698a0f | 976 | |
e855c69d | 977 | int diff = 0; |
b8698a0f | 978 | |
e855c69d | 979 | gcc_assert (n <= nn); |
b8698a0f | 980 | |
e855c69d AB |
981 | /* Sort both vectors so it will be possible to compare them. */ |
982 | qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); | |
983 | qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); | |
b8698a0f | 984 | |
e855c69d AB |
985 | while (ii < nn) |
986 | { | |
987 | if (v[i] == vv[ii]) | |
988 | i++; | |
989 | else | |
990 | /* VV[II] was lost. */ | |
991 | diff++; | |
b8698a0f | 992 | |
e855c69d AB |
993 | ii++; |
994 | } | |
b8698a0f | 995 | |
e855c69d AB |
996 | gcc_assert (diff == regset_pool.diff); |
997 | } | |
998 | #endif | |
b8698a0f | 999 | |
e855c69d AB |
1000 | /* If not true - we have a memory leak. */ |
1001 | gcc_assert (regset_pool.diff == 0); | |
b8698a0f | 1002 | |
e855c69d AB |
1003 | while (regset_pool.n) |
1004 | { | |
1005 | --regset_pool.n; | |
1006 | FREE_REG_SET (regset_pool.v[regset_pool.n]); | |
1007 | } | |
1008 | ||
1009 | free (regset_pool.v); | |
1010 | regset_pool.v = NULL; | |
1011 | regset_pool.s = 0; | |
b8698a0f | 1012 | |
e855c69d AB |
1013 | free (regset_pool.vv); |
1014 | regset_pool.vv = NULL; | |
1015 | regset_pool.nn = 0; | |
1016 | regset_pool.ss = 0; | |
1017 | ||
1018 | regset_pool.diff = 0; | |
1019 | } | |
1020 | \f | |
1021 | ||
b8698a0f L |
1022 | /* Functions to work with nop pools. NOP insns are used as temporary |
1023 | placeholders of the insns being scheduled to allow correct update of | |
e855c69d AB |
1024 | the data sets. When update is finished, NOPs are deleted. */ |
1025 | ||
1026 | /* A vinsn that is used to represent a nop. This vinsn is shared among all | |
1027 | nops sel-sched generates. */ | |
1028 | static vinsn_t nop_vinsn = NULL; | |
1029 | ||
1030 | /* Emit a nop before INSN, taking it from pool. */ | |
1031 | insn_t | |
1032 | get_nop_from_pool (insn_t insn) | |
1033 | { | |
1034 | insn_t nop; | |
1035 | bool old_p = nop_pool.n != 0; | |
1036 | int flags; | |
1037 | ||
1038 | if (old_p) | |
1039 | nop = nop_pool.v[--nop_pool.n]; | |
1040 | else | |
1041 | nop = nop_pattern; | |
1042 | ||
1043 | nop = emit_insn_before (nop, insn); | |
1044 | ||
1045 | if (old_p) | |
1046 | flags = INSN_INIT_TODO_SSID; | |
1047 | else | |
1048 | flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; | |
1049 | ||
1050 | set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); | |
1051 | sel_init_new_insn (nop, flags); | |
1052 | ||
1053 | return nop; | |
1054 | } | |
1055 | ||
1056 | /* Remove NOP from the instruction stream and return it to the pool. */ | |
1057 | void | |
b5b8b0ac | 1058 | return_nop_to_pool (insn_t nop, bool full_tidying) |
e855c69d AB |
1059 | { |
1060 | gcc_assert (INSN_IN_STREAM_P (nop)); | |
b5b8b0ac | 1061 | sel_remove_insn (nop, false, full_tidying); |
e855c69d AB |
1062 | |
1063 | if (nop_pool.n == nop_pool.s) | |
b8698a0f | 1064 | nop_pool.v = XRESIZEVEC (rtx, nop_pool.v, |
e855c69d AB |
1065 | (nop_pool.s = 2 * nop_pool.s + 1)); |
1066 | nop_pool.v[nop_pool.n++] = nop; | |
1067 | } | |
1068 | ||
1069 | /* Free the nop pool. */ | |
1070 | void | |
1071 | free_nop_pool (void) | |
1072 | { | |
1073 | nop_pool.n = 0; | |
1074 | nop_pool.s = 0; | |
1075 | free (nop_pool.v); | |
1076 | nop_pool.v = NULL; | |
1077 | } | |
1078 | \f | |
1079 | ||
b8698a0f | 1080 | /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb. |
e855c69d AB |
1081 | The callback is given two rtxes XX and YY and writes the new rtxes |
1082 | to NX and NY in case some needs to be skipped. */ | |
1083 | static int | |
1084 | skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) | |
1085 | { | |
1086 | const_rtx x = *xx; | |
1087 | const_rtx y = *yy; | |
b8698a0f | 1088 | |
e855c69d AB |
1089 | if (GET_CODE (x) == UNSPEC |
1090 | && (targetm.sched.skip_rtx_p == NULL | |
1091 | || targetm.sched.skip_rtx_p (x))) | |
1092 | { | |
1093 | *nx = XVECEXP (x, 0, 0); | |
1094 | *ny = CONST_CAST_RTX (y); | |
1095 | return 1; | |
1096 | } | |
b8698a0f | 1097 | |
e855c69d AB |
1098 | if (GET_CODE (y) == UNSPEC |
1099 | && (targetm.sched.skip_rtx_p == NULL | |
1100 | || targetm.sched.skip_rtx_p (y))) | |
1101 | { | |
1102 | *nx = CONST_CAST_RTX (x); | |
1103 | *ny = XVECEXP (y, 0, 0); | |
1104 | return 1; | |
1105 | } | |
b8698a0f | 1106 | |
e855c69d AB |
1107 | return 0; |
1108 | } | |
1109 | ||
b8698a0f | 1110 | /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way |
e855c69d AB |
1111 | to support ia64 speculation. When changes are needed, new rtx X and new mode |
1112 | NMODE are written, and the callback returns true. */ | |
1113 | static int | |
1114 | hash_with_unspec_callback (const_rtx x, enum machine_mode mode ATTRIBUTE_UNUSED, | |
1115 | rtx *nx, enum machine_mode* nmode) | |
1116 | { | |
b8698a0f | 1117 | if (GET_CODE (x) == UNSPEC |
e855c69d AB |
1118 | && targetm.sched.skip_rtx_p |
1119 | && targetm.sched.skip_rtx_p (x)) | |
1120 | { | |
1121 | *nx = XVECEXP (x, 0 ,0); | |
32e8bb8e | 1122 | *nmode = VOIDmode; |
e855c69d AB |
1123 | return 1; |
1124 | } | |
b8698a0f | 1125 | |
e855c69d AB |
1126 | return 0; |
1127 | } | |
1128 | ||
1129 | /* Returns LHS and RHS are ok to be scheduled separately. */ | |
1130 | static bool | |
1131 | lhs_and_rhs_separable_p (rtx lhs, rtx rhs) | |
1132 | { | |
1133 | if (lhs == NULL || rhs == NULL) | |
1134 | return false; | |
1135 | ||
b8698a0f L |
1136 | /* Do not schedule CONST, CONST_INT and CONST_DOUBLE etc as rhs: no point |
1137 | to use reg, if const can be used. Moreover, scheduling const as rhs may | |
1138 | lead to mode mismatch cause consts don't have modes but they could be | |
e855c69d AB |
1139 | merged from branches where the same const used in different modes. */ |
1140 | if (CONSTANT_P (rhs)) | |
1141 | return false; | |
1142 | ||
1143 | /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ | |
1144 | if (COMPARISON_P (rhs)) | |
1145 | return false; | |
1146 | ||
1147 | /* Do not allow single REG to be an rhs. */ | |
1148 | if (REG_P (rhs)) | |
1149 | return false; | |
1150 | ||
b8698a0f | 1151 | /* See comment at find_used_regs_1 (*1) for explanation of this |
e855c69d AB |
1152 | restriction. */ |
1153 | /* FIXME: remove this later. */ | |
1154 | if (MEM_P (lhs)) | |
1155 | return false; | |
1156 | ||
1157 | /* This will filter all tricky things like ZERO_EXTRACT etc. | |
1158 | For now we don't handle it. */ | |
1159 | if (!REG_P (lhs) && !MEM_P (lhs)) | |
1160 | return false; | |
1161 | ||
1162 | return true; | |
1163 | } | |
1164 | ||
b8698a0f L |
1165 | /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When |
1166 | FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is | |
e855c69d AB |
1167 | used e.g. for insns from recovery blocks. */ |
1168 | static void | |
1169 | vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) | |
1170 | { | |
1171 | hash_rtx_callback_function hrcf; | |
1172 | int insn_class; | |
1173 | ||
1174 | VINSN_INSN_RTX (vi) = insn; | |
1175 | VINSN_COUNT (vi) = 0; | |
1176 | vi->cost = -1; | |
b8698a0f | 1177 | |
9ef1bf71 AM |
1178 | if (INSN_NOP_P (insn)) |
1179 | return; | |
1180 | ||
e855c69d AB |
1181 | if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) |
1182 | init_id_from_df (VINSN_ID (vi), insn, force_unique_p); | |
1183 | else | |
1184 | deps_init_id (VINSN_ID (vi), insn, force_unique_p); | |
b8698a0f | 1185 | |
e855c69d AB |
1186 | /* Hash vinsn depending on whether it is separable or not. */ |
1187 | hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; | |
1188 | if (VINSN_SEPARABLE_P (vi)) | |
1189 | { | |
1190 | rtx rhs = VINSN_RHS (vi); | |
1191 | ||
1192 | VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs), | |
1193 | NULL, NULL, false, hrcf); | |
1194 | VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi), | |
1195 | VOIDmode, NULL, NULL, | |
1196 | false, hrcf); | |
1197 | } | |
1198 | else | |
1199 | { | |
1200 | VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode, | |
1201 | NULL, NULL, false, hrcf); | |
1202 | VINSN_HASH_RTX (vi) = VINSN_HASH (vi); | |
1203 | } | |
b8698a0f | 1204 | |
e855c69d AB |
1205 | insn_class = haifa_classify_insn (insn); |
1206 | if (insn_class >= 2 | |
1207 | && (!targetm.sched.get_insn_spec_ds | |
1208 | || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) | |
1209 | == 0))) | |
1210 | VINSN_MAY_TRAP_P (vi) = true; | |
1211 | else | |
1212 | VINSN_MAY_TRAP_P (vi) = false; | |
1213 | } | |
1214 | ||
1215 | /* Indicate that VI has become the part of an rtx object. */ | |
1216 | void | |
1217 | vinsn_attach (vinsn_t vi) | |
1218 | { | |
1219 | /* Assert that VI is not pending for deletion. */ | |
1220 | gcc_assert (VINSN_INSN_RTX (vi)); | |
1221 | ||
1222 | VINSN_COUNT (vi)++; | |
1223 | } | |
1224 | ||
b8698a0f | 1225 | /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct |
e855c69d AB |
1226 | VINSN_TYPE (VI). */ |
1227 | static vinsn_t | |
1228 | vinsn_create (insn_t insn, bool force_unique_p) | |
1229 | { | |
1230 | vinsn_t vi = XCNEW (struct vinsn_def); | |
1231 | ||
1232 | vinsn_init (vi, insn, force_unique_p); | |
1233 | return vi; | |
1234 | } | |
1235 | ||
1236 | /* Return a copy of VI. When REATTACH_P is true, detach VI and attach | |
1237 | the copy. */ | |
b8698a0f | 1238 | vinsn_t |
e855c69d AB |
1239 | vinsn_copy (vinsn_t vi, bool reattach_p) |
1240 | { | |
1241 | rtx copy; | |
1242 | bool unique = VINSN_UNIQUE_P (vi); | |
1243 | vinsn_t new_vi; | |
b8698a0f | 1244 | |
e855c69d AB |
1245 | copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); |
1246 | new_vi = create_vinsn_from_insn_rtx (copy, unique); | |
1247 | if (reattach_p) | |
1248 | { | |
1249 | vinsn_detach (vi); | |
1250 | vinsn_attach (new_vi); | |
1251 | } | |
1252 | ||
1253 | return new_vi; | |
1254 | } | |
1255 | ||
1256 | /* Delete the VI vinsn and free its data. */ | |
1257 | static void | |
1258 | vinsn_delete (vinsn_t vi) | |
1259 | { | |
1260 | gcc_assert (VINSN_COUNT (vi) == 0); | |
1261 | ||
9ef1bf71 AM |
1262 | if (!INSN_NOP_P (VINSN_INSN_RTX (vi))) |
1263 | { | |
1264 | return_regset_to_pool (VINSN_REG_SETS (vi)); | |
1265 | return_regset_to_pool (VINSN_REG_USES (vi)); | |
1266 | return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); | |
1267 | } | |
e855c69d AB |
1268 | |
1269 | free (vi); | |
1270 | } | |
1271 | ||
b8698a0f | 1272 | /* Indicate that VI is no longer a part of some rtx object. |
e855c69d AB |
1273 | Remove VI if it is no longer needed. */ |
1274 | void | |
1275 | vinsn_detach (vinsn_t vi) | |
1276 | { | |
1277 | gcc_assert (VINSN_COUNT (vi) > 0); | |
1278 | ||
1279 | if (--VINSN_COUNT (vi) == 0) | |
1280 | vinsn_delete (vi); | |
1281 | } | |
1282 | ||
1283 | /* Returns TRUE if VI is a branch. */ | |
1284 | bool | |
1285 | vinsn_cond_branch_p (vinsn_t vi) | |
1286 | { | |
1287 | insn_t insn; | |
1288 | ||
1289 | if (!VINSN_UNIQUE_P (vi)) | |
1290 | return false; | |
1291 | ||
1292 | insn = VINSN_INSN_RTX (vi); | |
1293 | if (BB_END (BLOCK_FOR_INSN (insn)) != insn) | |
1294 | return false; | |
1295 | ||
1296 | return control_flow_insn_p (insn); | |
1297 | } | |
1298 | ||
1299 | /* Return latency of INSN. */ | |
1300 | static int | |
1301 | sel_insn_rtx_cost (rtx insn) | |
1302 | { | |
1303 | int cost; | |
1304 | ||
1305 | /* A USE insn, or something else we don't need to | |
1306 | understand. We can't pass these directly to | |
1307 | result_ready_cost or insn_default_latency because it will | |
1308 | trigger a fatal error for unrecognizable insns. */ | |
1309 | if (recog_memoized (insn) < 0) | |
1310 | cost = 0; | |
1311 | else | |
1312 | { | |
1313 | cost = insn_default_latency (insn); | |
1314 | ||
1315 | if (cost < 0) | |
1316 | cost = 0; | |
1317 | } | |
1318 | ||
1319 | return cost; | |
1320 | } | |
1321 | ||
1322 | /* Return the cost of the VI. | |
1323 | !!! FIXME: Unify with haifa-sched.c: insn_cost (). */ | |
1324 | int | |
1325 | sel_vinsn_cost (vinsn_t vi) | |
1326 | { | |
1327 | int cost = vi->cost; | |
1328 | ||
1329 | if (cost < 0) | |
1330 | { | |
1331 | cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); | |
1332 | vi->cost = cost; | |
1333 | } | |
1334 | ||
1335 | return cost; | |
1336 | } | |
1337 | \f | |
1338 | ||
1339 | /* Functions for insn emitting. */ | |
1340 | ||
1341 | /* Emit new insn after AFTER based on PATTERN and initialize its data from | |
1342 | EXPR and SEQNO. */ | |
1343 | insn_t | |
1344 | sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) | |
1345 | { | |
1346 | insn_t new_insn; | |
1347 | ||
1348 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); | |
1349 | ||
1350 | new_insn = emit_insn_after (pattern, after); | |
1351 | set_insn_init (expr, NULL, seqno); | |
1352 | sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); | |
1353 | ||
1354 | return new_insn; | |
1355 | } | |
1356 | ||
1357 | /* Force newly generated vinsns to be unique. */ | |
1358 | static bool init_insn_force_unique_p = false; | |
1359 | ||
1360 | /* Emit new speculation recovery insn after AFTER based on PATTERN and | |
1361 | initialize its data from EXPR and SEQNO. */ | |
1362 | insn_t | |
1363 | sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, | |
1364 | insn_t after) | |
1365 | { | |
1366 | insn_t insn; | |
1367 | ||
1368 | gcc_assert (!init_insn_force_unique_p); | |
1369 | ||
1370 | init_insn_force_unique_p = true; | |
1371 | insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); | |
1372 | CANT_MOVE (insn) = 1; | |
1373 | init_insn_force_unique_p = false; | |
1374 | ||
1375 | return insn; | |
1376 | } | |
1377 | ||
1378 | /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, | |
b8698a0f L |
1379 | take it as a new vinsn instead of EXPR's vinsn. |
1380 | We simplify insns later, after scheduling region in | |
e855c69d AB |
1381 | simplify_changed_insns. */ |
1382 | insn_t | |
b8698a0f | 1383 | sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
e855c69d AB |
1384 | insn_t after) |
1385 | { | |
1386 | expr_t emit_expr; | |
1387 | insn_t insn; | |
1388 | int flags; | |
b8698a0f L |
1389 | |
1390 | emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), | |
e855c69d AB |
1391 | seqno); |
1392 | insn = EXPR_INSN_RTX (emit_expr); | |
b8698a0f | 1393 | add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); |
e855c69d AB |
1394 | |
1395 | flags = INSN_INIT_TODO_SSID; | |
1396 | if (INSN_LUID (insn) == 0) | |
1397 | flags |= INSN_INIT_TODO_LUID; | |
1398 | sel_init_new_insn (insn, flags); | |
1399 | ||
1400 | return insn; | |
1401 | } | |
1402 | ||
1403 | /* Move insn from EXPR after AFTER. */ | |
1404 | insn_t | |
1405 | sel_move_insn (expr_t expr, int seqno, insn_t after) | |
1406 | { | |
1407 | insn_t insn = EXPR_INSN_RTX (expr); | |
1408 | basic_block bb = BLOCK_FOR_INSN (after); | |
1409 | insn_t next = NEXT_INSN (after); | |
1410 | ||
1411 | /* Assert that in move_op we disconnected this insn properly. */ | |
1412 | gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); | |
1413 | PREV_INSN (insn) = after; | |
1414 | NEXT_INSN (insn) = next; | |
1415 | ||
1416 | NEXT_INSN (after) = insn; | |
1417 | PREV_INSN (next) = insn; | |
1418 | ||
1419 | /* Update links from insn to bb and vice versa. */ | |
1420 | df_insn_change_bb (insn, bb); | |
1421 | if (BB_END (bb) == after) | |
1422 | BB_END (bb) = insn; | |
b8698a0f | 1423 | |
e855c69d AB |
1424 | prepare_insn_expr (insn, seqno); |
1425 | return insn; | |
1426 | } | |
1427 | ||
1428 | \f | |
1429 | /* Functions to work with right-hand sides. */ | |
1430 | ||
b8698a0f | 1431 | /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector |
e855c69d | 1432 | VECT and return true when found. Use NEW_VINSN for comparison only when |
b8698a0f L |
1433 | COMPARE_VINSNS is true. Write to INDP the index on which |
1434 | the search has stopped, such that inserting the new element at INDP will | |
e855c69d AB |
1435 | retain VECT's sort order. */ |
1436 | static bool | |
b8698a0f L |
1437 | find_in_history_vect_1 (VEC(expr_history_def, heap) *vect, |
1438 | unsigned uid, vinsn_t new_vinsn, | |
e855c69d AB |
1439 | bool compare_vinsns, int *indp) |
1440 | { | |
1441 | expr_history_def *arr; | |
1442 | int i, j, len = VEC_length (expr_history_def, vect); | |
1443 | ||
1444 | if (len == 0) | |
1445 | { | |
1446 | *indp = 0; | |
1447 | return false; | |
1448 | } | |
1449 | ||
1450 | arr = VEC_address (expr_history_def, vect); | |
1451 | i = 0, j = len - 1; | |
1452 | ||
1453 | while (i <= j) | |
1454 | { | |
1455 | unsigned auid = arr[i].uid; | |
b8698a0f | 1456 | vinsn_t avinsn = arr[i].new_expr_vinsn; |
e855c69d AB |
1457 | |
1458 | if (auid == uid | |
b8698a0f L |
1459 | /* When undoing transformation on a bookkeeping copy, the new vinsn |
1460 | may not be exactly equal to the one that is saved in the vector. | |
e855c69d AB |
1461 | This is because the insn whose copy we're checking was possibly |
1462 | substituted itself. */ | |
b8698a0f | 1463 | && (! compare_vinsns |
e855c69d AB |
1464 | || vinsn_equal_p (avinsn, new_vinsn))) |
1465 | { | |
1466 | *indp = i; | |
1467 | return true; | |
1468 | } | |
1469 | else if (auid > uid) | |
1470 | break; | |
1471 | i++; | |
1472 | } | |
1473 | ||
1474 | *indp = i; | |
1475 | return false; | |
1476 | } | |
1477 | ||
b8698a0f L |
1478 | /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return |
1479 | the position found or -1, if no such value is in vector. | |
e855c69d AB |
1480 | Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ |
1481 | int | |
b8698a0f | 1482 | find_in_history_vect (VEC(expr_history_def, heap) *vect, rtx insn, |
e855c69d AB |
1483 | vinsn_t new_vinsn, bool originators_p) |
1484 | { | |
1485 | int ind; | |
1486 | ||
b8698a0f | 1487 | if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn, |
e855c69d AB |
1488 | false, &ind)) |
1489 | return ind; | |
1490 | ||
1491 | if (INSN_ORIGINATORS (insn) && originators_p) | |
1492 | { | |
1493 | unsigned uid; | |
1494 | bitmap_iterator bi; | |
1495 | ||
1496 | EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) | |
1497 | if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind)) | |
1498 | return ind; | |
1499 | } | |
b8698a0f | 1500 | |
e855c69d AB |
1501 | return -1; |
1502 | } | |
1503 | ||
b8698a0f L |
1504 | /* Insert new element in a sorted history vector pointed to by PVECT, |
1505 | if it is not there already. The element is searched using | |
e855c69d AB |
1506 | UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save |
1507 | the history of a transformation. */ | |
1508 | void | |
1509 | insert_in_history_vect (VEC (expr_history_def, heap) **pvect, | |
1510 | unsigned uid, enum local_trans_type type, | |
b8698a0f | 1511 | vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, |
e855c69d AB |
1512 | ds_t spec_ds) |
1513 | { | |
1514 | VEC(expr_history_def, heap) *vect = *pvect; | |
1515 | expr_history_def temp; | |
1516 | bool res; | |
1517 | int ind; | |
1518 | ||
1519 | res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind); | |
1520 | ||
1521 | if (res) | |
1522 | { | |
1523 | expr_history_def *phist = VEC_index (expr_history_def, vect, ind); | |
1524 | ||
b8698a0f | 1525 | /* It is possible that speculation types of expressions that were |
e855c69d AB |
1526 | propagated through different paths will be different here. In this |
1527 | case, merge the status to get the correct check later. */ | |
1528 | if (phist->spec_ds != spec_ds) | |
1529 | phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); | |
1530 | return; | |
1531 | } | |
b8698a0f | 1532 | |
e855c69d AB |
1533 | temp.uid = uid; |
1534 | temp.old_expr_vinsn = old_expr_vinsn; | |
b8698a0f | 1535 | temp.new_expr_vinsn = new_expr_vinsn; |
e855c69d AB |
1536 | temp.spec_ds = spec_ds; |
1537 | temp.type = type; | |
1538 | ||
1539 | vinsn_attach (old_expr_vinsn); | |
1540 | vinsn_attach (new_expr_vinsn); | |
1541 | VEC_safe_insert (expr_history_def, heap, vect, ind, &temp); | |
1542 | *pvect = vect; | |
1543 | } | |
1544 | ||
1545 | /* Free history vector PVECT. */ | |
1546 | static void | |
1547 | free_history_vect (VEC (expr_history_def, heap) **pvect) | |
1548 | { | |
1549 | unsigned i; | |
1550 | expr_history_def *phist; | |
1551 | ||
1552 | if (! *pvect) | |
1553 | return; | |
b8698a0f L |
1554 | |
1555 | for (i = 0; | |
e855c69d AB |
1556 | VEC_iterate (expr_history_def, *pvect, i, phist); |
1557 | i++) | |
1558 | { | |
1559 | vinsn_detach (phist->old_expr_vinsn); | |
1560 | vinsn_detach (phist->new_expr_vinsn); | |
1561 | } | |
b8698a0f | 1562 | |
e855c69d AB |
1563 | VEC_free (expr_history_def, heap, *pvect); |
1564 | *pvect = NULL; | |
1565 | } | |
1566 | ||
5d369d58 AB |
1567 | /* Merge vector FROM to PVECT. */ |
1568 | static void | |
1569 | merge_history_vect (VEC (expr_history_def, heap) **pvect, | |
1570 | VEC (expr_history_def, heap) *from) | |
1571 | { | |
1572 | expr_history_def *phist; | |
1573 | int i; | |
1574 | ||
1575 | /* We keep this vector sorted. */ | |
1576 | for (i = 0; VEC_iterate (expr_history_def, from, i, phist); i++) | |
1577 | insert_in_history_vect (pvect, phist->uid, phist->type, | |
1578 | phist->old_expr_vinsn, phist->new_expr_vinsn, | |
1579 | phist->spec_ds); | |
1580 | } | |
e855c69d AB |
1581 | |
1582 | /* Compare two vinsns as rhses if possible and as vinsns otherwise. */ | |
1583 | bool | |
1584 | vinsn_equal_p (vinsn_t x, vinsn_t y) | |
1585 | { | |
1586 | rtx_equal_p_callback_function repcf; | |
1587 | ||
1588 | if (x == y) | |
1589 | return true; | |
1590 | ||
1591 | if (VINSN_TYPE (x) != VINSN_TYPE (y)) | |
1592 | return false; | |
1593 | ||
1594 | if (VINSN_HASH (x) != VINSN_HASH (y)) | |
1595 | return false; | |
1596 | ||
1597 | repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL; | |
b8698a0f | 1598 | if (VINSN_SEPARABLE_P (x)) |
e855c69d AB |
1599 | { |
1600 | /* Compare RHSes of VINSNs. */ | |
1601 | gcc_assert (VINSN_RHS (x)); | |
1602 | gcc_assert (VINSN_RHS (y)); | |
1603 | ||
1604 | return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf); | |
1605 | } | |
1606 | ||
1607 | return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); | |
1608 | } | |
1609 | \f | |
1610 | ||
1611 | /* Functions for working with expressions. */ | |
1612 | ||
1613 | /* Initialize EXPR. */ | |
1614 | static void | |
1615 | init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority, | |
1616 | int sched_times, int orig_bb_index, ds_t spec_done_ds, | |
1617 | ds_t spec_to_check_ds, int orig_sched_cycle, | |
f3764768 | 1618 | VEC(expr_history_def, heap) *history, signed char target_available, |
e855c69d AB |
1619 | bool was_substituted, bool was_renamed, bool needs_spec_check_p, |
1620 | bool cant_move) | |
1621 | { | |
1622 | vinsn_attach (vi); | |
1623 | ||
1624 | EXPR_VINSN (expr) = vi; | |
1625 | EXPR_SPEC (expr) = spec; | |
1626 | EXPR_USEFULNESS (expr) = use; | |
1627 | EXPR_PRIORITY (expr) = priority; | |
1628 | EXPR_PRIORITY_ADJ (expr) = 0; | |
1629 | EXPR_SCHED_TIMES (expr) = sched_times; | |
1630 | EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; | |
1631 | EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; | |
1632 | EXPR_SPEC_DONE_DS (expr) = spec_done_ds; | |
1633 | EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; | |
1634 | ||
1635 | if (history) | |
1636 | EXPR_HISTORY_OF_CHANGES (expr) = history; | |
1637 | else | |
1638 | EXPR_HISTORY_OF_CHANGES (expr) = NULL; | |
1639 | ||
1640 | EXPR_TARGET_AVAILABLE (expr) = target_available; | |
1641 | EXPR_WAS_SUBSTITUTED (expr) = was_substituted; | |
1642 | EXPR_WAS_RENAMED (expr) = was_renamed; | |
1643 | EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; | |
1644 | EXPR_CANT_MOVE (expr) = cant_move; | |
1645 | } | |
1646 | ||
1647 | /* Make a copy of the expr FROM into the expr TO. */ | |
1648 | void | |
1649 | copy_expr (expr_t to, expr_t from) | |
1650 | { | |
1651 | VEC(expr_history_def, heap) *temp = NULL; | |
1652 | ||
1653 | if (EXPR_HISTORY_OF_CHANGES (from)) | |
1654 | { | |
1655 | unsigned i; | |
1656 | expr_history_def *phist; | |
1657 | ||
1658 | temp = VEC_copy (expr_history_def, heap, EXPR_HISTORY_OF_CHANGES (from)); | |
b8698a0f | 1659 | for (i = 0; |
e855c69d AB |
1660 | VEC_iterate (expr_history_def, temp, i, phist); |
1661 | i++) | |
1662 | { | |
1663 | vinsn_attach (phist->old_expr_vinsn); | |
1664 | vinsn_attach (phist->new_expr_vinsn); | |
1665 | } | |
1666 | } | |
1667 | ||
b8698a0f | 1668 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), |
e855c69d AB |
1669 | EXPR_USEFULNESS (from), EXPR_PRIORITY (from), |
1670 | EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), | |
b8698a0f | 1671 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), |
e855c69d | 1672 | EXPR_ORIG_SCHED_CYCLE (from), temp, |
b8698a0f | 1673 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
e855c69d AB |
1674 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1675 | EXPR_CANT_MOVE (from)); | |
1676 | } | |
1677 | ||
b8698a0f | 1678 | /* Same, but the final expr will not ever be in av sets, so don't copy |
e855c69d AB |
1679 | "uninteresting" data such as bitmap cache. */ |
1680 | void | |
1681 | copy_expr_onside (expr_t to, expr_t from) | |
1682 | { | |
1683 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), | |
1684 | EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0, | |
1685 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0, NULL, | |
1686 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), | |
1687 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), | |
1688 | EXPR_CANT_MOVE (from)); | |
1689 | } | |
1690 | ||
1691 | /* Prepare the expr of INSN for scheduling. Used when moving insn and when | |
1692 | initializing new insns. */ | |
1693 | static void | |
1694 | prepare_insn_expr (insn_t insn, int seqno) | |
1695 | { | |
1696 | expr_t expr = INSN_EXPR (insn); | |
1697 | ds_t ds; | |
b8698a0f | 1698 | |
e855c69d AB |
1699 | INSN_SEQNO (insn) = seqno; |
1700 | EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); | |
1701 | EXPR_SPEC (expr) = 0; | |
1702 | EXPR_ORIG_SCHED_CYCLE (expr) = 0; | |
1703 | EXPR_WAS_SUBSTITUTED (expr) = 0; | |
1704 | EXPR_WAS_RENAMED (expr) = 0; | |
1705 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
1706 | INSN_LIVE_VALID_P (insn) = false; | |
1707 | ||
1708 | /* ??? If this expression is speculative, make its dependence | |
1709 | as weak as possible. We can filter this expression later | |
1710 | in process_spec_exprs, because we do not distinguish | |
1711 | between the status we got during compute_av_set and the | |
1712 | existing status. To be fixed. */ | |
1713 | ds = EXPR_SPEC_DONE_DS (expr); | |
1714 | if (ds) | |
1715 | EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); | |
1716 | ||
1717 | free_history_vect (&EXPR_HISTORY_OF_CHANGES (expr)); | |
1718 | } | |
1719 | ||
1720 | /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1721 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1722 | a split point. */ |
1723 | static void | |
1724 | update_target_availability (expr_t to, expr_t from, insn_t split_point) | |
1725 | { | |
b8698a0f | 1726 | if (EXPR_TARGET_AVAILABLE (to) < 0 |
e855c69d AB |
1727 | || EXPR_TARGET_AVAILABLE (from) < 0) |
1728 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1729 | else | |
1730 | { | |
1731 | /* We try to detect the case when one of the expressions | |
1732 | can only be reached through another one. In this case, | |
1733 | we can do better. */ | |
1734 | if (split_point == NULL) | |
1735 | { | |
1736 | int toind, fromind; | |
1737 | ||
1738 | toind = EXPR_ORIG_BB_INDEX (to); | |
1739 | fromind = EXPR_ORIG_BB_INDEX (from); | |
b8698a0f | 1740 | |
e855c69d | 1741 | if (toind && toind == fromind) |
b8698a0f | 1742 | /* Do nothing -- everything is done in |
e855c69d AB |
1743 | merge_with_other_exprs. */ |
1744 | ; | |
1745 | else | |
1746 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1747 | } | |
1748 | else | |
1749 | EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); | |
1750 | } | |
1751 | } | |
1752 | ||
1753 | /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT | |
b8698a0f | 1754 | is non-null when expressions are merged from different successors at |
e855c69d AB |
1755 | a split point. */ |
1756 | static void | |
1757 | update_speculative_bits (expr_t to, expr_t from, insn_t split_point) | |
1758 | { | |
1759 | ds_t old_to_ds, old_from_ds; | |
1760 | ||
1761 | old_to_ds = EXPR_SPEC_DONE_DS (to); | |
1762 | old_from_ds = EXPR_SPEC_DONE_DS (from); | |
b8698a0f | 1763 | |
e855c69d AB |
1764 | EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); |
1765 | EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); | |
1766 | EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); | |
1767 | ||
1768 | /* When merging e.g. control & data speculative exprs, or a control | |
b8698a0f | 1769 | speculative with a control&data speculative one, we really have |
e855c69d AB |
1770 | to change vinsn too. Also, when speculative status is changed, |
1771 | we also need to record this as a transformation in expr's history. */ | |
1772 | if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) | |
1773 | { | |
1774 | old_to_ds = ds_get_speculation_types (old_to_ds); | |
1775 | old_from_ds = ds_get_speculation_types (old_from_ds); | |
b8698a0f | 1776 | |
e855c69d AB |
1777 | if (old_to_ds != old_from_ds) |
1778 | { | |
1779 | ds_t record_ds; | |
b8698a0f L |
1780 | |
1781 | /* When both expressions are speculative, we need to change | |
e855c69d AB |
1782 | the vinsn first. */ |
1783 | if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) | |
1784 | { | |
1785 | int res; | |
b8698a0f | 1786 | |
e855c69d AB |
1787 | res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); |
1788 | gcc_assert (res >= 0); | |
1789 | } | |
1790 | ||
1791 | if (split_point != NULL) | |
1792 | { | |
1793 | /* Record the change with proper status. */ | |
1794 | record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; | |
1795 | record_ds &= ~(old_to_ds & SPECULATIVE); | |
1796 | record_ds &= ~(old_from_ds & SPECULATIVE); | |
b8698a0f L |
1797 | |
1798 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), | |
1799 | INSN_UID (split_point), TRANS_SPECULATION, | |
e855c69d AB |
1800 | EXPR_VINSN (from), EXPR_VINSN (to), |
1801 | record_ds); | |
1802 | } | |
1803 | } | |
1804 | } | |
1805 | } | |
1806 | ||
1807 | ||
1808 | /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, | |
1809 | this is done along different paths. */ | |
1810 | void | |
1811 | merge_expr_data (expr_t to, expr_t from, insn_t split_point) | |
1812 | { | |
e855c69d AB |
1813 | /* For now, we just set the spec of resulting expr to be minimum of the specs |
1814 | of merged exprs. */ | |
1815 | if (EXPR_SPEC (to) > EXPR_SPEC (from)) | |
1816 | EXPR_SPEC (to) = EXPR_SPEC (from); | |
1817 | ||
1818 | if (split_point) | |
1819 | EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); | |
1820 | else | |
b8698a0f | 1821 | EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), |
e855c69d AB |
1822 | EXPR_USEFULNESS (from)); |
1823 | ||
1824 | if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) | |
1825 | EXPR_PRIORITY (to) = EXPR_PRIORITY (from); | |
1826 | ||
1827 | if (EXPR_SCHED_TIMES (to) > EXPR_SCHED_TIMES (from)) | |
1828 | EXPR_SCHED_TIMES (to) = EXPR_SCHED_TIMES (from); | |
1829 | ||
1830 | if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) | |
1831 | EXPR_ORIG_BB_INDEX (to) = 0; | |
1832 | ||
b8698a0f | 1833 | EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), |
e855c69d AB |
1834 | EXPR_ORIG_SCHED_CYCLE (from)); |
1835 | ||
e855c69d AB |
1836 | EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); |
1837 | EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); | |
1838 | EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); | |
1839 | ||
5d369d58 AB |
1840 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to), |
1841 | EXPR_HISTORY_OF_CHANGES (from)); | |
e855c69d AB |
1842 | update_target_availability (to, from, split_point); |
1843 | update_speculative_bits (to, from, split_point); | |
1844 | } | |
1845 | ||
1846 | /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal | |
b8698a0f | 1847 | in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions |
e855c69d AB |
1848 | are merged from different successors at a split point. */ |
1849 | void | |
1850 | merge_expr (expr_t to, expr_t from, insn_t split_point) | |
1851 | { | |
1852 | vinsn_t to_vi = EXPR_VINSN (to); | |
1853 | vinsn_t from_vi = EXPR_VINSN (from); | |
1854 | ||
1855 | gcc_assert (vinsn_equal_p (to_vi, from_vi)); | |
1856 | ||
1857 | /* Make sure that speculative pattern is propagated into exprs that | |
1858 | have non-speculative one. This will provide us with consistent | |
1859 | speculative bits and speculative patterns inside expr. */ | |
1860 | if (EXPR_SPEC_DONE_DS (to) == 0 | |
1861 | && EXPR_SPEC_DONE_DS (from) != 0) | |
1862 | change_vinsn_in_expr (to, EXPR_VINSN (from)); | |
1863 | ||
1864 | merge_expr_data (to, from, split_point); | |
1865 | gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); | |
1866 | } | |
1867 | ||
1868 | /* Clear the information of this EXPR. */ | |
1869 | void | |
1870 | clear_expr (expr_t expr) | |
1871 | { | |
b8698a0f | 1872 | |
e855c69d AB |
1873 | vinsn_detach (EXPR_VINSN (expr)); |
1874 | EXPR_VINSN (expr) = NULL; | |
1875 | ||
1876 | free_history_vect (&EXPR_HISTORY_OF_CHANGES (expr)); | |
1877 | } | |
1878 | ||
1879 | /* For a given LV_SET, mark EXPR having unavailable target register. */ | |
1880 | static void | |
1881 | set_unavailable_target_for_expr (expr_t expr, regset lv_set) | |
1882 | { | |
1883 | if (EXPR_SEPARABLE_P (expr)) | |
1884 | { | |
1885 | if (REG_P (EXPR_LHS (expr)) | |
1886 | && bitmap_bit_p (lv_set, REGNO (EXPR_LHS (expr)))) | |
1887 | { | |
b8698a0f L |
1888 | /* If it's an insn like r1 = use (r1, ...), and it exists in |
1889 | different forms in each of the av_sets being merged, we can't say | |
1890 | whether original destination register is available or not. | |
1891 | However, this still works if destination register is not used | |
e855c69d AB |
1892 | in the original expression: if the branch at which LV_SET we're |
1893 | looking here is not actually 'other branch' in sense that same | |
b8698a0f | 1894 | expression is available through it (but it can't be determined |
e855c69d | 1895 | at computation stage because of transformations on one of the |
b8698a0f L |
1896 | branches), it still won't affect the availability. |
1897 | Liveness of a register somewhere on a code motion path means | |
1898 | it's either read somewhere on a codemotion path, live on | |
e855c69d AB |
1899 | 'other' branch, live at the point immediately following |
1900 | the original operation, or is read by the original operation. | |
1901 | The latter case is filtered out in the condition below. | |
1902 | It still doesn't cover the case when register is defined and used | |
1903 | somewhere within the code motion path, and in this case we could | |
1904 | miss a unifying code motion along both branches using a renamed | |
1905 | register, but it won't affect a code correctness since upon | |
1906 | an actual code motion a bookkeeping code would be generated. */ | |
b8698a0f | 1907 | if (bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
e855c69d AB |
1908 | REGNO (EXPR_LHS (expr)))) |
1909 | EXPR_TARGET_AVAILABLE (expr) = -1; | |
1910 | else | |
1911 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1912 | } | |
1913 | } | |
1914 | else | |
1915 | { | |
1916 | unsigned regno; | |
1917 | reg_set_iterator rsi; | |
b8698a0f L |
1918 | |
1919 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), | |
e855c69d AB |
1920 | 0, regno, rsi) |
1921 | if (bitmap_bit_p (lv_set, regno)) | |
1922 | { | |
1923 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1924 | break; | |
1925 | } | |
1926 | ||
1927 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), | |
1928 | 0, regno, rsi) | |
1929 | if (bitmap_bit_p (lv_set, regno)) | |
1930 | { | |
1931 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1932 | break; | |
1933 | } | |
1934 | } | |
1935 | } | |
1936 | ||
b8698a0f | 1937 | /* Try to make EXPR speculative. Return 1 when EXPR's pattern |
e855c69d AB |
1938 | or dependence status have changed, 2 when also the target register |
1939 | became unavailable, 0 if nothing had to be changed. */ | |
1940 | int | |
1941 | speculate_expr (expr_t expr, ds_t ds) | |
1942 | { | |
1943 | int res; | |
1944 | rtx orig_insn_rtx; | |
1945 | rtx spec_pat; | |
1946 | ds_t target_ds, current_ds; | |
1947 | ||
1948 | /* Obtain the status we need to put on EXPR. */ | |
1949 | target_ds = (ds & SPECULATIVE); | |
1950 | current_ds = EXPR_SPEC_DONE_DS (expr); | |
1951 | ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); | |
1952 | ||
1953 | orig_insn_rtx = EXPR_INSN_RTX (expr); | |
1954 | ||
1955 | res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); | |
1956 | ||
1957 | switch (res) | |
1958 | { | |
1959 | case 0: | |
1960 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1961 | return current_ds != ds ? 1 : 0; | |
b8698a0f | 1962 | |
e855c69d AB |
1963 | case 1: |
1964 | { | |
1965 | rtx spec_insn_rtx = create_insn_rtx_from_pattern (spec_pat, NULL_RTX); | |
1966 | vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); | |
1967 | ||
1968 | change_vinsn_in_expr (expr, spec_vinsn); | |
1969 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1970 | EXPR_NEEDS_SPEC_CHECK_P (expr) = true; | |
1971 | ||
b8698a0f | 1972 | /* Do not allow clobbering the address register of speculative |
e855c69d | 1973 | insns. */ |
b8698a0f | 1974 | if (bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
e855c69d AB |
1975 | expr_dest_regno (expr))) |
1976 | { | |
1977 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1978 | return 2; | |
1979 | } | |
1980 | ||
1981 | return 1; | |
1982 | } | |
1983 | ||
1984 | case -1: | |
1985 | return -1; | |
1986 | ||
1987 | default: | |
1988 | gcc_unreachable (); | |
1989 | return -1; | |
1990 | } | |
1991 | } | |
1992 | ||
1993 | /* Return a destination register, if any, of EXPR. */ | |
1994 | rtx | |
1995 | expr_dest_reg (expr_t expr) | |
1996 | { | |
1997 | rtx dest = VINSN_LHS (EXPR_VINSN (expr)); | |
1998 | ||
1999 | if (dest != NULL_RTX && REG_P (dest)) | |
2000 | return dest; | |
2001 | ||
2002 | return NULL_RTX; | |
2003 | } | |
2004 | ||
2005 | /* Returns the REGNO of the R's destination. */ | |
2006 | unsigned | |
2007 | expr_dest_regno (expr_t expr) | |
2008 | { | |
2009 | rtx dest = expr_dest_reg (expr); | |
2010 | ||
2011 | gcc_assert (dest != NULL_RTX); | |
2012 | return REGNO (dest); | |
2013 | } | |
2014 | ||
b8698a0f | 2015 | /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in |
e855c69d AB |
2016 | AV_SET having unavailable target register. */ |
2017 | void | |
2018 | mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) | |
2019 | { | |
2020 | expr_t expr; | |
2021 | av_set_iterator avi; | |
2022 | ||
2023 | FOR_EACH_EXPR (expr, avi, join_set) | |
2024 | if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) | |
2025 | set_unavailable_target_for_expr (expr, lv_set); | |
2026 | } | |
2027 | \f | |
2028 | ||
2029 | /* Av set functions. */ | |
2030 | ||
2031 | /* Add a new element to av set SETP. | |
2032 | Return the element added. */ | |
2033 | static av_set_t | |
2034 | av_set_add_element (av_set_t *setp) | |
2035 | { | |
2036 | /* Insert at the beginning of the list. */ | |
2037 | _list_add (setp); | |
2038 | return *setp; | |
2039 | } | |
2040 | ||
2041 | /* Add EXPR to SETP. */ | |
2042 | void | |
2043 | av_set_add (av_set_t *setp, expr_t expr) | |
2044 | { | |
2045 | av_set_t elem; | |
b8698a0f | 2046 | |
e855c69d AB |
2047 | gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); |
2048 | elem = av_set_add_element (setp); | |
2049 | copy_expr (_AV_SET_EXPR (elem), expr); | |
2050 | } | |
2051 | ||
2052 | /* Same, but do not copy EXPR. */ | |
2053 | static void | |
2054 | av_set_add_nocopy (av_set_t *setp, expr_t expr) | |
2055 | { | |
2056 | av_set_t elem; | |
2057 | ||
2058 | elem = av_set_add_element (setp); | |
2059 | *_AV_SET_EXPR (elem) = *expr; | |
2060 | } | |
2061 | ||
2062 | /* Remove expr pointed to by IP from the av_set. */ | |
2063 | void | |
2064 | av_set_iter_remove (av_set_iterator *ip) | |
2065 | { | |
2066 | clear_expr (_AV_SET_EXPR (*ip->lp)); | |
2067 | _list_iter_remove (ip); | |
2068 | } | |
2069 | ||
2070 | /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the | |
2071 | sense of vinsn_equal_p function. Return NULL if no such expr is | |
2072 | in SET was found. */ | |
2073 | expr_t | |
2074 | av_set_lookup (av_set_t set, vinsn_t sought_vinsn) | |
2075 | { | |
2076 | expr_t expr; | |
2077 | av_set_iterator i; | |
2078 | ||
2079 | FOR_EACH_EXPR (expr, i, set) | |
2080 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2081 | return expr; | |
2082 | return NULL; | |
2083 | } | |
2084 | ||
2085 | /* Same, but also remove the EXPR found. */ | |
2086 | static expr_t | |
2087 | av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) | |
2088 | { | |
2089 | expr_t expr; | |
2090 | av_set_iterator i; | |
2091 | ||
2092 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2093 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2094 | { | |
2095 | _list_iter_remove_nofree (&i); | |
2096 | return expr; | |
2097 | } | |
2098 | return NULL; | |
2099 | } | |
2100 | ||
2101 | /* Search for an expr in SET, such that it's equivalent to EXPR in the | |
2102 | sense of vinsn_equal_p function of their vinsns, but not EXPR itself. | |
2103 | Returns NULL if no such expr is in SET was found. */ | |
2104 | static expr_t | |
2105 | av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) | |
2106 | { | |
2107 | expr_t cur_expr; | |
2108 | av_set_iterator i; | |
2109 | ||
2110 | FOR_EACH_EXPR (cur_expr, i, set) | |
2111 | { | |
2112 | if (cur_expr == expr) | |
2113 | continue; | |
2114 | if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) | |
2115 | return cur_expr; | |
2116 | } | |
2117 | ||
2118 | return NULL; | |
2119 | } | |
2120 | ||
2121 | /* If other expression is already in AVP, remove one of them. */ | |
2122 | expr_t | |
2123 | merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) | |
2124 | { | |
2125 | expr_t expr2; | |
2126 | ||
2127 | expr2 = av_set_lookup_other_equiv_expr (*avp, expr); | |
2128 | if (expr2 != NULL) | |
2129 | { | |
2130 | /* Reset target availability on merge, since taking it only from one | |
2131 | of the exprs would be controversial for different code. */ | |
2132 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2133 | EXPR_USEFULNESS (expr2) = 0; | |
2134 | ||
2135 | merge_expr (expr2, expr, NULL); | |
b8698a0f | 2136 | |
e855c69d AB |
2137 | /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ |
2138 | EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; | |
b8698a0f | 2139 | |
e855c69d AB |
2140 | av_set_iter_remove (ip); |
2141 | return expr2; | |
2142 | } | |
2143 | ||
2144 | return expr; | |
2145 | } | |
2146 | ||
2147 | /* Return true if there is an expr that correlates to VI in SET. */ | |
2148 | bool | |
2149 | av_set_is_in_p (av_set_t set, vinsn_t vi) | |
2150 | { | |
2151 | return av_set_lookup (set, vi) != NULL; | |
2152 | } | |
2153 | ||
2154 | /* Return a copy of SET. */ | |
2155 | av_set_t | |
2156 | av_set_copy (av_set_t set) | |
2157 | { | |
2158 | expr_t expr; | |
2159 | av_set_iterator i; | |
2160 | av_set_t res = NULL; | |
2161 | ||
2162 | FOR_EACH_EXPR (expr, i, set) | |
2163 | av_set_add (&res, expr); | |
2164 | ||
2165 | return res; | |
2166 | } | |
2167 | ||
2168 | /* Join two av sets that do not have common elements by attaching second set | |
2169 | (pointed to by FROMP) to the end of first set (TO_TAILP must point to | |
2170 | _AV_SET_NEXT of first set's last element). */ | |
2171 | static void | |
2172 | join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) | |
2173 | { | |
2174 | gcc_assert (*to_tailp == NULL); | |
2175 | *to_tailp = *fromp; | |
2176 | *fromp = NULL; | |
2177 | } | |
2178 | ||
2179 | /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set | |
2180 | pointed to by FROMP afterwards. */ | |
2181 | void | |
2182 | av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) | |
2183 | { | |
2184 | expr_t expr1; | |
2185 | av_set_iterator i; | |
2186 | ||
2187 | /* Delete from TOP all exprs, that present in FROMP. */ | |
2188 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2189 | { | |
2190 | expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1)); | |
2191 | ||
2192 | if (expr2) | |
2193 | { | |
2194 | merge_expr (expr2, expr1, insn); | |
2195 | av_set_iter_remove (&i); | |
2196 | } | |
2197 | } | |
2198 | ||
2199 | join_distinct_sets (i.lp, fromp); | |
2200 | } | |
2201 | ||
b8698a0f | 2202 | /* Same as above, but also update availability of target register in |
e855c69d AB |
2203 | TOP judging by TO_LV_SET and FROM_LV_SET. */ |
2204 | void | |
2205 | av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, | |
2206 | regset from_lv_set, insn_t insn) | |
2207 | { | |
2208 | expr_t expr1; | |
2209 | av_set_iterator i; | |
2210 | av_set_t *to_tailp, in_both_set = NULL; | |
2211 | ||
2212 | /* Delete from TOP all expres, that present in FROMP. */ | |
2213 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2214 | { | |
2215 | expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1)); | |
2216 | ||
2217 | if (expr2) | |
2218 | { | |
b8698a0f | 2219 | /* It may be that the expressions have different destination |
e855c69d AB |
2220 | registers, in which case we need to check liveness here. */ |
2221 | if (EXPR_SEPARABLE_P (expr1)) | |
2222 | { | |
b8698a0f | 2223 | int regno1 = (REG_P (EXPR_LHS (expr1)) |
e855c69d | 2224 | ? (int) expr_dest_regno (expr1) : -1); |
b8698a0f | 2225 | int regno2 = (REG_P (EXPR_LHS (expr2)) |
e855c69d | 2226 | ? (int) expr_dest_regno (expr2) : -1); |
b8698a0f L |
2227 | |
2228 | /* ??? We don't have a way to check restrictions for | |
e855c69d AB |
2229 | *other* register on the current path, we did it only |
2230 | for the current target register. Give up. */ | |
2231 | if (regno1 != regno2) | |
2232 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2233 | } | |
2234 | else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) | |
2235 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2236 | ||
2237 | merge_expr (expr2, expr1, insn); | |
2238 | av_set_add_nocopy (&in_both_set, expr2); | |
2239 | av_set_iter_remove (&i); | |
2240 | } | |
2241 | else | |
b8698a0f | 2242 | /* EXPR1 is present in TOP, but not in FROMP. Check it on |
e855c69d AB |
2243 | FROM_LV_SET. */ |
2244 | set_unavailable_target_for_expr (expr1, from_lv_set); | |
2245 | } | |
2246 | to_tailp = i.lp; | |
2247 | ||
2248 | /* These expressions are not present in TOP. Check liveness | |
2249 | restrictions on TO_LV_SET. */ | |
2250 | FOR_EACH_EXPR (expr1, i, *fromp) | |
2251 | set_unavailable_target_for_expr (expr1, to_lv_set); | |
2252 | ||
2253 | join_distinct_sets (i.lp, &in_both_set); | |
2254 | join_distinct_sets (to_tailp, fromp); | |
2255 | } | |
2256 | ||
2257 | /* Clear av_set pointed to by SETP. */ | |
2258 | void | |
2259 | av_set_clear (av_set_t *setp) | |
2260 | { | |
2261 | expr_t expr; | |
2262 | av_set_iterator i; | |
2263 | ||
2264 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2265 | av_set_iter_remove (&i); | |
2266 | ||
2267 | gcc_assert (*setp == NULL); | |
2268 | } | |
2269 | ||
2270 | /* Leave only one non-speculative element in the SETP. */ | |
2271 | void | |
2272 | av_set_leave_one_nonspec (av_set_t *setp) | |
2273 | { | |
2274 | expr_t expr; | |
2275 | av_set_iterator i; | |
2276 | bool has_one_nonspec = false; | |
2277 | ||
b8698a0f | 2278 | /* Keep all speculative exprs, and leave one non-speculative |
e855c69d AB |
2279 | (the first one). */ |
2280 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2281 | { | |
2282 | if (!EXPR_SPEC_DONE_DS (expr)) | |
2283 | { | |
2284 | if (has_one_nonspec) | |
2285 | av_set_iter_remove (&i); | |
2286 | else | |
2287 | has_one_nonspec = true; | |
2288 | } | |
2289 | } | |
2290 | } | |
2291 | ||
2292 | /* Return the N'th element of the SET. */ | |
2293 | expr_t | |
2294 | av_set_element (av_set_t set, int n) | |
2295 | { | |
2296 | expr_t expr; | |
2297 | av_set_iterator i; | |
2298 | ||
2299 | FOR_EACH_EXPR (expr, i, set) | |
2300 | if (n-- == 0) | |
2301 | return expr; | |
2302 | ||
2303 | gcc_unreachable (); | |
2304 | return NULL; | |
2305 | } | |
2306 | ||
2307 | /* Deletes all expressions from AVP that are conditional branches (IFs). */ | |
2308 | void | |
2309 | av_set_substract_cond_branches (av_set_t *avp) | |
2310 | { | |
2311 | av_set_iterator i; | |
2312 | expr_t expr; | |
2313 | ||
2314 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2315 | if (vinsn_cond_branch_p (EXPR_VINSN (expr))) | |
2316 | av_set_iter_remove (&i); | |
2317 | } | |
2318 | ||
b8698a0f | 2319 | /* Multiplies usefulness attribute of each member of av-set *AVP by |
e855c69d AB |
2320 | value PROB / ALL_PROB. */ |
2321 | void | |
2322 | av_set_split_usefulness (av_set_t av, int prob, int all_prob) | |
2323 | { | |
2324 | av_set_iterator i; | |
2325 | expr_t expr; | |
2326 | ||
2327 | FOR_EACH_EXPR (expr, i, av) | |
b8698a0f | 2328 | EXPR_USEFULNESS (expr) = (all_prob |
e855c69d AB |
2329 | ? (EXPR_USEFULNESS (expr) * prob) / all_prob |
2330 | : 0); | |
2331 | } | |
2332 | ||
2333 | /* Leave in AVP only those expressions, which are present in AV, | |
5d369d58 | 2334 | and return it, merging history expressions. */ |
e855c69d | 2335 | void |
5d369d58 | 2336 | av_set_code_motion_filter (av_set_t *avp, av_set_t av) |
e855c69d AB |
2337 | { |
2338 | av_set_iterator i; | |
5d369d58 | 2339 | expr_t expr, expr2; |
e855c69d AB |
2340 | |
2341 | FOR_EACH_EXPR_1 (expr, i, avp) | |
5d369d58 | 2342 | if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL) |
e855c69d | 2343 | av_set_iter_remove (&i); |
5d369d58 AB |
2344 | else |
2345 | /* When updating av sets in bookkeeping blocks, we can add more insns | |
2346 | there which will be transformed but the upper av sets will not | |
2347 | reflect those transformations. We then fail to undo those | |
2348 | when searching for such insns. So merge the history saved | |
2349 | in the av set of the block we are processing. */ | |
2350 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2351 | EXPR_HISTORY_OF_CHANGES (expr2)); | |
e855c69d AB |
2352 | } |
2353 | ||
2354 | \f | |
2355 | ||
2356 | /* Dependence hooks to initialize insn data. */ | |
2357 | ||
2358 | /* This is used in hooks callable from dependence analysis when initializing | |
2359 | instruction's data. */ | |
2360 | static struct | |
2361 | { | |
2362 | /* Where the dependence was found (lhs/rhs). */ | |
2363 | deps_where_t where; | |
2364 | ||
2365 | /* The actual data object to initialize. */ | |
2366 | idata_t id; | |
2367 | ||
2368 | /* True when the insn should not be made clonable. */ | |
2369 | bool force_unique_p; | |
2370 | ||
2371 | /* True when insn should be treated as of type USE, i.e. never renamed. */ | |
2372 | bool force_use_p; | |
2373 | } deps_init_id_data; | |
2374 | ||
2375 | ||
b8698a0f | 2376 | /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be |
e855c69d AB |
2377 | clonable. */ |
2378 | static void | |
2379 | setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) | |
2380 | { | |
2381 | int type; | |
b8698a0f | 2382 | |
e855c69d AB |
2383 | /* Determine whether INSN could be cloned and return appropriate vinsn type. |
2384 | That clonable insns which can be separated into lhs and rhs have type SET. | |
2385 | Other clonable insns have type USE. */ | |
2386 | type = GET_CODE (insn); | |
2387 | ||
2388 | /* Only regular insns could be cloned. */ | |
2389 | if (type == INSN && !force_unique_p) | |
2390 | type = SET; | |
2391 | else if (type == JUMP_INSN && simplejump_p (insn)) | |
2392 | type = PC; | |
b5b8b0ac AO |
2393 | else if (type == DEBUG_INSN) |
2394 | type = !force_unique_p ? USE : INSN; | |
b8698a0f | 2395 | |
e855c69d AB |
2396 | IDATA_TYPE (id) = type; |
2397 | IDATA_REG_SETS (id) = get_clear_regset_from_pool (); | |
2398 | IDATA_REG_USES (id) = get_clear_regset_from_pool (); | |
2399 | IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); | |
2400 | } | |
2401 | ||
2402 | /* Start initializing insn data. */ | |
2403 | static void | |
2404 | deps_init_id_start_insn (insn_t insn) | |
2405 | { | |
2406 | gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); | |
2407 | ||
2408 | setup_id_for_insn (deps_init_id_data.id, insn, | |
2409 | deps_init_id_data.force_unique_p); | |
2410 | deps_init_id_data.where = DEPS_IN_INSN; | |
2411 | } | |
2412 | ||
2413 | /* Start initializing lhs data. */ | |
2414 | static void | |
2415 | deps_init_id_start_lhs (rtx lhs) | |
2416 | { | |
2417 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2418 | gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); | |
2419 | ||
2420 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2421 | { | |
2422 | IDATA_LHS (deps_init_id_data.id) = lhs; | |
2423 | deps_init_id_data.where = DEPS_IN_LHS; | |
2424 | } | |
2425 | } | |
2426 | ||
2427 | /* Finish initializing lhs data. */ | |
2428 | static void | |
2429 | deps_init_id_finish_lhs (void) | |
2430 | { | |
2431 | deps_init_id_data.where = DEPS_IN_INSN; | |
2432 | } | |
2433 | ||
2434 | /* Note a set of REGNO. */ | |
2435 | static void | |
2436 | deps_init_id_note_reg_set (int regno) | |
2437 | { | |
2438 | haifa_note_reg_set (regno); | |
2439 | ||
2440 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2441 | deps_init_id_data.force_use_p = true; | |
2442 | ||
2443 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2444 | SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); | |
2445 | ||
2446 | #ifdef STACK_REGS | |
b8698a0f | 2447 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2448 | renaming to avoid issues with find_used_regs. */ |
2449 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2450 | deps_init_id_data.force_use_p = true; | |
2451 | #endif | |
2452 | } | |
2453 | ||
2454 | /* Note a clobber of REGNO. */ | |
2455 | static void | |
2456 | deps_init_id_note_reg_clobber (int regno) | |
2457 | { | |
2458 | haifa_note_reg_clobber (regno); | |
2459 | ||
2460 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2461 | deps_init_id_data.force_use_p = true; | |
2462 | ||
2463 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2464 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); | |
2465 | } | |
2466 | ||
2467 | /* Note a use of REGNO. */ | |
2468 | static void | |
2469 | deps_init_id_note_reg_use (int regno) | |
2470 | { | |
2471 | haifa_note_reg_use (regno); | |
2472 | ||
2473 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2474 | SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); | |
2475 | } | |
2476 | ||
2477 | /* Start initializing rhs data. */ | |
2478 | static void | |
2479 | deps_init_id_start_rhs (rtx rhs) | |
2480 | { | |
2481 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2482 | ||
2483 | /* And there was no sel_deps_reset_to_insn (). */ | |
2484 | if (IDATA_LHS (deps_init_id_data.id) != NULL) | |
2485 | { | |
2486 | IDATA_RHS (deps_init_id_data.id) = rhs; | |
2487 | deps_init_id_data.where = DEPS_IN_RHS; | |
2488 | } | |
2489 | } | |
2490 | ||
2491 | /* Finish initializing rhs data. */ | |
2492 | static void | |
2493 | deps_init_id_finish_rhs (void) | |
2494 | { | |
2495 | gcc_assert (deps_init_id_data.where == DEPS_IN_RHS | |
2496 | || deps_init_id_data.where == DEPS_IN_INSN); | |
2497 | deps_init_id_data.where = DEPS_IN_INSN; | |
2498 | } | |
2499 | ||
2500 | /* Finish initializing insn data. */ | |
2501 | static void | |
2502 | deps_init_id_finish_insn (void) | |
2503 | { | |
2504 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2505 | ||
2506 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2507 | { | |
2508 | rtx lhs = IDATA_LHS (deps_init_id_data.id); | |
2509 | rtx rhs = IDATA_RHS (deps_init_id_data.id); | |
2510 | ||
2511 | if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) | |
2512 | || deps_init_id_data.force_use_p) | |
2513 | { | |
b8698a0f | 2514 | /* This should be a USE, as we don't want to schedule its RHS |
e855c69d | 2515 | separately. However, we still want to have them recorded |
b8698a0f | 2516 | for the purposes of substitution. That's why we don't |
e855c69d AB |
2517 | simply call downgrade_to_use () here. */ |
2518 | gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); | |
2519 | gcc_assert (!lhs == !rhs); | |
2520 | ||
2521 | IDATA_TYPE (deps_init_id_data.id) = USE; | |
2522 | } | |
2523 | } | |
2524 | ||
2525 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2526 | } | |
2527 | ||
2528 | /* This is dependence info used for initializing insn's data. */ | |
2529 | static struct sched_deps_info_def deps_init_id_sched_deps_info; | |
2530 | ||
2531 | /* This initializes most of the static part of the above structure. */ | |
2532 | static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = | |
2533 | { | |
2534 | NULL, | |
2535 | ||
2536 | deps_init_id_start_insn, | |
2537 | deps_init_id_finish_insn, | |
2538 | deps_init_id_start_lhs, | |
2539 | deps_init_id_finish_lhs, | |
2540 | deps_init_id_start_rhs, | |
2541 | deps_init_id_finish_rhs, | |
2542 | deps_init_id_note_reg_set, | |
2543 | deps_init_id_note_reg_clobber, | |
2544 | deps_init_id_note_reg_use, | |
2545 | NULL, /* note_mem_dep */ | |
2546 | NULL, /* note_dep */ | |
2547 | ||
2548 | 0, /* use_cselib */ | |
2549 | 0, /* use_deps_list */ | |
2550 | 0 /* generate_spec_deps */ | |
2551 | }; | |
2552 | ||
2553 | /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, | |
2554 | we don't actually need information about lhs and rhs. */ | |
2555 | static void | |
2556 | setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) | |
2557 | { | |
2558 | rtx pat = PATTERN (insn); | |
b8698a0f | 2559 | |
481683e1 | 2560 | if (NONJUMP_INSN_P (insn) |
b8698a0f | 2561 | && GET_CODE (pat) == SET |
e855c69d AB |
2562 | && !force_unique_p) |
2563 | { | |
2564 | IDATA_RHS (id) = SET_SRC (pat); | |
2565 | IDATA_LHS (id) = SET_DEST (pat); | |
2566 | } | |
2567 | else | |
2568 | IDATA_LHS (id) = IDATA_RHS (id) = NULL; | |
2569 | } | |
2570 | ||
2571 | /* Possibly downgrade INSN to USE. */ | |
2572 | static void | |
2573 | maybe_downgrade_id_to_use (idata_t id, insn_t insn) | |
2574 | { | |
2575 | bool must_be_use = false; | |
2576 | unsigned uid = INSN_UID (insn); | |
57512f53 | 2577 | df_ref *rec; |
e855c69d AB |
2578 | rtx lhs = IDATA_LHS (id); |
2579 | rtx rhs = IDATA_RHS (id); | |
b8698a0f | 2580 | |
e855c69d AB |
2581 | /* We downgrade only SETs. */ |
2582 | if (IDATA_TYPE (id) != SET) | |
2583 | return; | |
2584 | ||
2585 | if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) | |
2586 | { | |
2587 | IDATA_TYPE (id) = USE; | |
2588 | return; | |
2589 | } | |
b8698a0f | 2590 | |
e855c69d AB |
2591 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2592 | { | |
57512f53 | 2593 | df_ref def = *rec; |
b8698a0f | 2594 | |
e855c69d AB |
2595 | if (DF_REF_INSN (def) |
2596 | && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) | |
2597 | && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) | |
2598 | { | |
2599 | must_be_use = true; | |
2600 | break; | |
2601 | } | |
2602 | ||
2603 | #ifdef STACK_REGS | |
b8698a0f | 2604 | /* Make instructions that set stack registers to be ineligible for |
e855c69d AB |
2605 | renaming to avoid issues with find_used_regs. */ |
2606 | if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) | |
2607 | { | |
2608 | must_be_use = true; | |
2609 | break; | |
2610 | } | |
2611 | #endif | |
b8698a0f L |
2612 | } |
2613 | ||
e855c69d AB |
2614 | if (must_be_use) |
2615 | IDATA_TYPE (id) = USE; | |
2616 | } | |
2617 | ||
2618 | /* Setup register sets describing INSN in ID. */ | |
2619 | static void | |
2620 | setup_id_reg_sets (idata_t id, insn_t insn) | |
2621 | { | |
2622 | unsigned uid = INSN_UID (insn); | |
57512f53 | 2623 | df_ref *rec; |
e855c69d | 2624 | regset tmp = get_clear_regset_from_pool (); |
b8698a0f | 2625 | |
e855c69d AB |
2626 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2627 | { | |
57512f53 | 2628 | df_ref def = *rec; |
e855c69d | 2629 | unsigned int regno = DF_REF_REGNO (def); |
b8698a0f | 2630 | |
e855c69d AB |
2631 | /* Post modifies are treated like clobbers by sched-deps.c. */ |
2632 | if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER | |
2633 | | DF_REF_PRE_POST_MODIFY))) | |
2634 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); | |
2635 | else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) | |
2636 | { | |
2637 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); | |
2638 | ||
2639 | #ifdef STACK_REGS | |
b8698a0f | 2640 | /* For stack registers, treat writes to them as writes |
e855c69d AB |
2641 | to the first one to be consistent with sched-deps.c. */ |
2642 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2643 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); | |
2644 | #endif | |
2645 | } | |
2646 | /* Mark special refs that generate read/write def pair. */ | |
2647 | if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) | |
2648 | || regno == STACK_POINTER_REGNUM) | |
2649 | bitmap_set_bit (tmp, regno); | |
2650 | } | |
b8698a0f | 2651 | |
e855c69d AB |
2652 | for (rec = DF_INSN_UID_USES (uid); *rec; rec++) |
2653 | { | |
57512f53 | 2654 | df_ref use = *rec; |
e855c69d AB |
2655 | unsigned int regno = DF_REF_REGNO (use); |
2656 | ||
2657 | /* When these refs are met for the first time, skip them, as | |
2658 | these uses are just counterparts of some defs. */ | |
2659 | if (bitmap_bit_p (tmp, regno)) | |
2660 | bitmap_clear_bit (tmp, regno); | |
2661 | else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) | |
2662 | { | |
2663 | SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); | |
2664 | ||
2665 | #ifdef STACK_REGS | |
b8698a0f | 2666 | /* For stack registers, treat reads from them as reads from |
e855c69d AB |
2667 | the first one to be consistent with sched-deps.c. */ |
2668 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2669 | SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); | |
2670 | #endif | |
2671 | } | |
2672 | } | |
2673 | ||
2674 | return_regset_to_pool (tmp); | |
2675 | } | |
2676 | ||
2677 | /* Initialize instruction data for INSN in ID using DF's data. */ | |
2678 | static void | |
2679 | init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) | |
2680 | { | |
2681 | gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); | |
2682 | ||
2683 | setup_id_for_insn (id, insn, force_unique_p); | |
2684 | setup_id_lhs_rhs (id, insn, force_unique_p); | |
2685 | ||
2686 | if (INSN_NOP_P (insn)) | |
2687 | return; | |
2688 | ||
2689 | maybe_downgrade_id_to_use (id, insn); | |
2690 | setup_id_reg_sets (id, insn); | |
2691 | } | |
2692 | ||
2693 | /* Initialize instruction data for INSN in ID. */ | |
2694 | static void | |
2695 | deps_init_id (idata_t id, insn_t insn, bool force_unique_p) | |
2696 | { | |
88302d54 | 2697 | struct deps_desc _dc, *dc = &_dc; |
e855c69d AB |
2698 | |
2699 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2700 | deps_init_id_data.id = id; | |
2701 | deps_init_id_data.force_unique_p = force_unique_p; | |
2702 | deps_init_id_data.force_use_p = false; | |
2703 | ||
bcf33775 | 2704 | init_deps (dc, false); |
e855c69d AB |
2705 | |
2706 | memcpy (&deps_init_id_sched_deps_info, | |
2707 | &const_deps_init_id_sched_deps_info, | |
2708 | sizeof (deps_init_id_sched_deps_info)); | |
2709 | ||
2710 | if (spec_info != NULL) | |
2711 | deps_init_id_sched_deps_info.generate_spec_deps = 1; | |
2712 | ||
2713 | sched_deps_info = &deps_init_id_sched_deps_info; | |
2714 | ||
2715 | deps_analyze_insn (dc, insn); | |
2716 | ||
2717 | free_deps (dc); | |
2718 | ||
2719 | deps_init_id_data.id = NULL; | |
2720 | } | |
2721 | ||
2722 | \f | |
2723 | ||
2724 | /* Implement hooks for collecting fundamental insn properties like if insn is | |
2725 | an ASM or is within a SCHED_GROUP. */ | |
2726 | ||
2727 | /* True when a "one-time init" data for INSN was already inited. */ | |
2728 | static bool | |
2729 | first_time_insn_init (insn_t insn) | |
2730 | { | |
2731 | return INSN_LIVE (insn) == NULL; | |
2732 | } | |
2733 | ||
2734 | /* Hash an entry in a transformed_insns hashtable. */ | |
2735 | static hashval_t | |
2736 | hash_transformed_insns (const void *p) | |
2737 | { | |
2738 | return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2739 | } | |
2740 | ||
2741 | /* Compare the entries in a transformed_insns hashtable. */ | |
2742 | static int | |
2743 | eq_transformed_insns (const void *p, const void *q) | |
2744 | { | |
2745 | rtx i1 = VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2746 | rtx i2 = VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); | |
2747 | ||
2748 | if (INSN_UID (i1) == INSN_UID (i2)) | |
2749 | return 1; | |
2750 | return rtx_equal_p (PATTERN (i1), PATTERN (i2)); | |
2751 | } | |
2752 | ||
2753 | /* Free an entry in a transformed_insns hashtable. */ | |
2754 | static void | |
2755 | free_transformed_insns (void *p) | |
2756 | { | |
2757 | struct transformed_insns *pti = (struct transformed_insns *) p; | |
2758 | ||
2759 | vinsn_detach (pti->vinsn_old); | |
2760 | vinsn_detach (pti->vinsn_new); | |
2761 | free (pti); | |
2762 | } | |
2763 | ||
b8698a0f | 2764 | /* Init the s_i_d data for INSN which should be inited just once, when |
e855c69d AB |
2765 | we first see the insn. */ |
2766 | static void | |
2767 | init_first_time_insn_data (insn_t insn) | |
2768 | { | |
2769 | /* This should not be set if this is the first time we init data for | |
2770 | insn. */ | |
2771 | gcc_assert (first_time_insn_init (insn)); | |
b8698a0f | 2772 | |
e855c69d AB |
2773 | /* These are needed for nops too. */ |
2774 | INSN_LIVE (insn) = get_regset_from_pool (); | |
2775 | INSN_LIVE_VALID_P (insn) = false; | |
bcf33775 | 2776 | |
e855c69d AB |
2777 | if (!INSN_NOP_P (insn)) |
2778 | { | |
2779 | INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); | |
2780 | INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); | |
b8698a0f | 2781 | INSN_TRANSFORMED_INSNS (insn) |
e855c69d AB |
2782 | = htab_create (16, hash_transformed_insns, |
2783 | eq_transformed_insns, free_transformed_insns); | |
bcf33775 | 2784 | init_deps (&INSN_DEPS_CONTEXT (insn), true); |
e855c69d AB |
2785 | } |
2786 | } | |
2787 | ||
b8698a0f | 2788 | /* Free almost all above data for INSN that is scheduled already. |
bcf33775 AB |
2789 | Used for extra-large basic blocks. */ |
2790 | void | |
2791 | free_data_for_scheduled_insn (insn_t insn) | |
e855c69d AB |
2792 | { |
2793 | gcc_assert (! first_time_insn_init (insn)); | |
b8698a0f | 2794 | |
bcf33775 AB |
2795 | if (! INSN_ANALYZED_DEPS (insn)) |
2796 | return; | |
b8698a0f | 2797 | |
e855c69d AB |
2798 | BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); |
2799 | BITMAP_FREE (INSN_FOUND_DEPS (insn)); | |
2800 | htab_delete (INSN_TRANSFORMED_INSNS (insn)); | |
b8698a0f | 2801 | |
e855c69d AB |
2802 | /* This is allocated only for bookkeeping insns. */ |
2803 | if (INSN_ORIGINATORS (insn)) | |
2804 | BITMAP_FREE (INSN_ORIGINATORS (insn)); | |
2805 | free_deps (&INSN_DEPS_CONTEXT (insn)); | |
bcf33775 AB |
2806 | |
2807 | INSN_ANALYZED_DEPS (insn) = NULL; | |
2808 | ||
b8698a0f | 2809 | /* Clear the readonly flag so we would ICE when trying to recalculate |
bcf33775 AB |
2810 | the deps context (as we believe that it should not happen). */ |
2811 | (&INSN_DEPS_CONTEXT (insn))->readonly = 0; | |
2812 | } | |
2813 | ||
2814 | /* Free the same data as above for INSN. */ | |
2815 | static void | |
2816 | free_first_time_insn_data (insn_t insn) | |
2817 | { | |
2818 | gcc_assert (! first_time_insn_init (insn)); | |
2819 | ||
2820 | free_data_for_scheduled_insn (insn); | |
2821 | return_regset_to_pool (INSN_LIVE (insn)); | |
2822 | INSN_LIVE (insn) = NULL; | |
2823 | INSN_LIVE_VALID_P (insn) = false; | |
e855c69d AB |
2824 | } |
2825 | ||
2826 | /* Initialize region-scope data structures for basic blocks. */ | |
2827 | static void | |
2828 | init_global_and_expr_for_bb (basic_block bb) | |
2829 | { | |
2830 | if (sel_bb_empty_p (bb)) | |
2831 | return; | |
2832 | ||
2833 | invalidate_av_set (bb); | |
2834 | } | |
2835 | ||
2836 | /* Data for global dependency analysis (to initialize CANT_MOVE and | |
2837 | SCHED_GROUP_P). */ | |
2838 | static struct | |
2839 | { | |
2840 | /* Previous insn. */ | |
2841 | insn_t prev_insn; | |
2842 | } init_global_data; | |
2843 | ||
2844 | /* Determine if INSN is in the sched_group, is an asm or should not be | |
2845 | cloned. After that initialize its expr. */ | |
2846 | static void | |
2847 | init_global_and_expr_for_insn (insn_t insn) | |
2848 | { | |
2849 | if (LABEL_P (insn)) | |
2850 | return; | |
2851 | ||
2852 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2853 | { | |
2854 | init_global_data.prev_insn = NULL_RTX; | |
2855 | return; | |
2856 | } | |
2857 | ||
2858 | gcc_assert (INSN_P (insn)); | |
2859 | ||
2860 | if (SCHED_GROUP_P (insn)) | |
2861 | /* Setup a sched_group. */ | |
2862 | { | |
2863 | insn_t prev_insn = init_global_data.prev_insn; | |
2864 | ||
2865 | if (prev_insn) | |
2866 | INSN_SCHED_NEXT (prev_insn) = insn; | |
2867 | ||
2868 | init_global_data.prev_insn = insn; | |
2869 | } | |
2870 | else | |
2871 | init_global_data.prev_insn = NULL_RTX; | |
2872 | ||
2873 | if (GET_CODE (PATTERN (insn)) == ASM_INPUT | |
2874 | || asm_noperands (PATTERN (insn)) >= 0) | |
2875 | /* Mark INSN as an asm. */ | |
2876 | INSN_ASM_P (insn) = true; | |
2877 | ||
2878 | { | |
2879 | bool force_unique_p; | |
2880 | ds_t spec_done_ds; | |
2881 | ||
cfeb0fa8 AB |
2882 | /* Certain instructions cannot be cloned, and frame related insns and |
2883 | the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of | |
2884 | their block. */ | |
2885 | if (prologue_epilogue_contains (insn)) | |
2886 | { | |
2887 | if (RTX_FRAME_RELATED_P (insn)) | |
2888 | CANT_MOVE (insn) = 1; | |
2889 | else | |
2890 | { | |
2891 | rtx note; | |
2892 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
2893 | if (REG_NOTE_KIND (note) == REG_SAVE_NOTE | |
2894 | && ((enum insn_note) INTVAL (XEXP (note, 0)) | |
2895 | == NOTE_INSN_EPILOGUE_BEG)) | |
2896 | { | |
2897 | CANT_MOVE (insn) = 1; | |
2898 | break; | |
2899 | } | |
2900 | } | |
2901 | force_unique_p = true; | |
2902 | } | |
e855c69d | 2903 | else |
cfeb0fa8 AB |
2904 | if (CANT_MOVE (insn) |
2905 | || INSN_ASM_P (insn) | |
2906 | || SCHED_GROUP_P (insn) | |
2907 | /* Exception handling insns are always unique. */ | |
2908 | || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) | |
2909 | /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ | |
2910 | || control_flow_insn_p (insn)) | |
2911 | force_unique_p = true; | |
2912 | else | |
2913 | force_unique_p = false; | |
e855c69d AB |
2914 | |
2915 | if (targetm.sched.get_insn_spec_ds) | |
2916 | { | |
2917 | spec_done_ds = targetm.sched.get_insn_spec_ds (insn); | |
2918 | spec_done_ds = ds_get_max_dep_weak (spec_done_ds); | |
2919 | } | |
2920 | else | |
2921 | spec_done_ds = 0; | |
2922 | ||
2923 | /* Initialize INSN's expr. */ | |
2924 | init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0, | |
2925 | REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn), | |
b8698a0f | 2926 | spec_done_ds, 0, 0, NULL, true, false, false, false, |
e855c69d AB |
2927 | CANT_MOVE (insn)); |
2928 | } | |
2929 | ||
2930 | init_first_time_insn_data (insn); | |
2931 | } | |
2932 | ||
2933 | /* Scan the region and initialize instruction data for basic blocks BBS. */ | |
2934 | void | |
2935 | sel_init_global_and_expr (bb_vec_t bbs) | |
2936 | { | |
2937 | /* ??? It would be nice to implement push / pop scheme for sched_infos. */ | |
2938 | const struct sched_scan_info_def ssi = | |
2939 | { | |
2940 | NULL, /* extend_bb */ | |
2941 | init_global_and_expr_for_bb, /* init_bb */ | |
2942 | extend_insn_data, /* extend_insn */ | |
2943 | init_global_and_expr_for_insn /* init_insn */ | |
2944 | }; | |
b8698a0f | 2945 | |
e855c69d AB |
2946 | sched_scan (&ssi, bbs, NULL, NULL, NULL); |
2947 | } | |
2948 | ||
2949 | /* Finalize region-scope data structures for basic blocks. */ | |
2950 | static void | |
2951 | finish_global_and_expr_for_bb (basic_block bb) | |
2952 | { | |
2953 | av_set_clear (&BB_AV_SET (bb)); | |
2954 | BB_AV_LEVEL (bb) = 0; | |
2955 | } | |
2956 | ||
2957 | /* Finalize INSN's data. */ | |
2958 | static void | |
2959 | finish_global_and_expr_insn (insn_t insn) | |
2960 | { | |
2961 | if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2962 | return; | |
2963 | ||
2964 | gcc_assert (INSN_P (insn)); | |
2965 | ||
2966 | if (INSN_LUID (insn) > 0) | |
2967 | { | |
2968 | free_first_time_insn_data (insn); | |
2969 | INSN_WS_LEVEL (insn) = 0; | |
2970 | CANT_MOVE (insn) = 0; | |
b8698a0f L |
2971 | |
2972 | /* We can no longer assert this, as vinsns of this insn could be | |
2973 | easily live in other insn's caches. This should be changed to | |
e855c69d AB |
2974 | a counter-like approach among all vinsns. */ |
2975 | gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); | |
2976 | clear_expr (INSN_EXPR (insn)); | |
2977 | } | |
2978 | } | |
2979 | ||
2980 | /* Finalize per instruction data for the whole region. */ | |
2981 | void | |
2982 | sel_finish_global_and_expr (void) | |
2983 | { | |
2984 | { | |
2985 | bb_vec_t bbs; | |
2986 | int i; | |
2987 | ||
2988 | bbs = VEC_alloc (basic_block, heap, current_nr_blocks); | |
2989 | ||
2990 | for (i = 0; i < current_nr_blocks; i++) | |
2991 | VEC_quick_push (basic_block, bbs, BASIC_BLOCK (BB_TO_BLOCK (i))); | |
2992 | ||
2993 | /* Clear AV_SETs and INSN_EXPRs. */ | |
2994 | { | |
2995 | const struct sched_scan_info_def ssi = | |
2996 | { | |
2997 | NULL, /* extend_bb */ | |
2998 | finish_global_and_expr_for_bb, /* init_bb */ | |
2999 | NULL, /* extend_insn */ | |
3000 | finish_global_and_expr_insn /* init_insn */ | |
3001 | }; | |
3002 | ||
3003 | sched_scan (&ssi, bbs, NULL, NULL, NULL); | |
3004 | } | |
3005 | ||
3006 | VEC_free (basic_block, heap, bbs); | |
3007 | } | |
3008 | ||
3009 | finish_insns (); | |
3010 | } | |
3011 | \f | |
3012 | ||
b8698a0f L |
3013 | /* In the below hooks, we merely calculate whether or not a dependence |
3014 | exists, and in what part of insn. However, we will need more data | |
e855c69d AB |
3015 | when we'll start caching dependence requests. */ |
3016 | ||
3017 | /* Container to hold information for dependency analysis. */ | |
3018 | static struct | |
3019 | { | |
3020 | deps_t dc; | |
3021 | ||
3022 | /* A variable to track which part of rtx we are scanning in | |
3023 | sched-deps.c: sched_analyze_insn (). */ | |
3024 | deps_where_t where; | |
3025 | ||
3026 | /* Current producer. */ | |
3027 | insn_t pro; | |
3028 | ||
3029 | /* Current consumer. */ | |
3030 | vinsn_t con; | |
3031 | ||
3032 | /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. | |
3033 | X is from { INSN, LHS, RHS }. */ | |
3034 | ds_t has_dep_p[DEPS_IN_NOWHERE]; | |
3035 | } has_dependence_data; | |
3036 | ||
3037 | /* Start analyzing dependencies of INSN. */ | |
3038 | static void | |
3039 | has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) | |
3040 | { | |
3041 | gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); | |
3042 | ||
3043 | has_dependence_data.where = DEPS_IN_INSN; | |
3044 | } | |
3045 | ||
3046 | /* Finish analyzing dependencies of an insn. */ | |
3047 | static void | |
3048 | has_dependence_finish_insn (void) | |
3049 | { | |
3050 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3051 | ||
3052 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3053 | } | |
3054 | ||
3055 | /* Start analyzing dependencies of LHS. */ | |
3056 | static void | |
3057 | has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) | |
3058 | { | |
3059 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3060 | ||
3061 | if (VINSN_LHS (has_dependence_data.con) != NULL) | |
3062 | has_dependence_data.where = DEPS_IN_LHS; | |
3063 | } | |
3064 | ||
3065 | /* Finish analyzing dependencies of an lhs. */ | |
3066 | static void | |
3067 | has_dependence_finish_lhs (void) | |
3068 | { | |
3069 | has_dependence_data.where = DEPS_IN_INSN; | |
3070 | } | |
3071 | ||
3072 | /* Start analyzing dependencies of RHS. */ | |
3073 | static void | |
3074 | has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) | |
3075 | { | |
3076 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3077 | ||
3078 | if (VINSN_RHS (has_dependence_data.con) != NULL) | |
3079 | has_dependence_data.where = DEPS_IN_RHS; | |
3080 | } | |
3081 | ||
3082 | /* Start analyzing dependencies of an rhs. */ | |
3083 | static void | |
3084 | has_dependence_finish_rhs (void) | |
3085 | { | |
3086 | gcc_assert (has_dependence_data.where == DEPS_IN_RHS | |
3087 | || has_dependence_data.where == DEPS_IN_INSN); | |
3088 | ||
3089 | has_dependence_data.where = DEPS_IN_INSN; | |
3090 | } | |
3091 | ||
3092 | /* Note a set of REGNO. */ | |
3093 | static void | |
3094 | has_dependence_note_reg_set (int regno) | |
3095 | { | |
3096 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3097 | ||
3098 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3099 | VINSN_INSN_RTX | |
3100 | (has_dependence_data.con))) | |
3101 | { | |
3102 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3103 | ||
3104 | if (reg_last->sets != NULL | |
3105 | || reg_last->clobbers != NULL) | |
3106 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
3107 | ||
3108 | if (reg_last->uses) | |
3109 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3110 | } | |
3111 | } | |
3112 | ||
3113 | /* Note a clobber of REGNO. */ | |
3114 | static void | |
3115 | has_dependence_note_reg_clobber (int regno) | |
3116 | { | |
3117 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3118 | ||
3119 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3120 | VINSN_INSN_RTX | |
3121 | (has_dependence_data.con))) | |
3122 | { | |
3123 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3124 | ||
3125 | if (reg_last->sets) | |
3126 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
b8698a0f | 3127 | |
e855c69d AB |
3128 | if (reg_last->uses) |
3129 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3130 | } | |
3131 | } | |
3132 | ||
3133 | /* Note a use of REGNO. */ | |
3134 | static void | |
3135 | has_dependence_note_reg_use (int regno) | |
3136 | { | |
3137 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3138 | ||
3139 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3140 | VINSN_INSN_RTX | |
3141 | (has_dependence_data.con))) | |
3142 | { | |
3143 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3144 | ||
3145 | if (reg_last->sets) | |
3146 | *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; | |
3147 | ||
3148 | if (reg_last->clobbers) | |
3149 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; | |
3150 | ||
3151 | /* Handle BE_IN_SPEC. */ | |
3152 | if (reg_last->uses) | |
3153 | { | |
3154 | ds_t pro_spec_checked_ds; | |
3155 | ||
3156 | pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); | |
3157 | pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); | |
3158 | ||
3159 | if (pro_spec_checked_ds != 0) | |
3160 | /* Merge BE_IN_SPEC bits into *DSP. */ | |
3161 | *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, | |
3162 | NULL_RTX, NULL_RTX); | |
3163 | } | |
3164 | } | |
3165 | } | |
3166 | ||
3167 | /* Note a memory dependence. */ | |
3168 | static void | |
3169 | has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, | |
3170 | rtx pending_mem ATTRIBUTE_UNUSED, | |
3171 | insn_t pending_insn ATTRIBUTE_UNUSED, | |
3172 | ds_t ds ATTRIBUTE_UNUSED) | |
3173 | { | |
3174 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3175 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3176 | { | |
3177 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3178 | ||
3179 | *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); | |
3180 | } | |
3181 | } | |
3182 | ||
3183 | /* Note a dependence. */ | |
3184 | static void | |
3185 | has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED, | |
3186 | ds_t ds ATTRIBUTE_UNUSED) | |
3187 | { | |
3188 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3189 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3190 | { | |
3191 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3192 | ||
3193 | *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); | |
3194 | } | |
3195 | } | |
3196 | ||
3197 | /* Mark the insn as having a hard dependence that prevents speculation. */ | |
3198 | void | |
3199 | sel_mark_hard_insn (rtx insn) | |
3200 | { | |
3201 | int i; | |
3202 | ||
3203 | /* Only work when we're in has_dependence_p mode. | |
3204 | ??? This is a hack, this should actually be a hook. */ | |
3205 | if (!has_dependence_data.dc || !has_dependence_data.pro) | |
3206 | return; | |
3207 | ||
3208 | gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); | |
3209 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3210 | ||
3211 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3212 | has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; | |
3213 | } | |
3214 | ||
3215 | /* This structure holds the hooks for the dependency analysis used when | |
3216 | actually processing dependencies in the scheduler. */ | |
3217 | static struct sched_deps_info_def has_dependence_sched_deps_info; | |
3218 | ||
3219 | /* This initializes most of the fields of the above structure. */ | |
3220 | static const struct sched_deps_info_def const_has_dependence_sched_deps_info = | |
3221 | { | |
3222 | NULL, | |
3223 | ||
3224 | has_dependence_start_insn, | |
3225 | has_dependence_finish_insn, | |
3226 | has_dependence_start_lhs, | |
3227 | has_dependence_finish_lhs, | |
3228 | has_dependence_start_rhs, | |
3229 | has_dependence_finish_rhs, | |
3230 | has_dependence_note_reg_set, | |
3231 | has_dependence_note_reg_clobber, | |
3232 | has_dependence_note_reg_use, | |
3233 | has_dependence_note_mem_dep, | |
3234 | has_dependence_note_dep, | |
3235 | ||
3236 | 0, /* use_cselib */ | |
3237 | 0, /* use_deps_list */ | |
3238 | 0 /* generate_spec_deps */ | |
3239 | }; | |
3240 | ||
3241 | /* Initialize has_dependence_sched_deps_info with extra spec field. */ | |
3242 | static void | |
3243 | setup_has_dependence_sched_deps_info (void) | |
3244 | { | |
3245 | memcpy (&has_dependence_sched_deps_info, | |
3246 | &const_has_dependence_sched_deps_info, | |
3247 | sizeof (has_dependence_sched_deps_info)); | |
3248 | ||
3249 | if (spec_info != NULL) | |
3250 | has_dependence_sched_deps_info.generate_spec_deps = 1; | |
3251 | ||
3252 | sched_deps_info = &has_dependence_sched_deps_info; | |
3253 | } | |
3254 | ||
3255 | /* Remove all dependences found and recorded in has_dependence_data array. */ | |
3256 | void | |
3257 | sel_clear_has_dependence (void) | |
3258 | { | |
3259 | int i; | |
3260 | ||
3261 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3262 | has_dependence_data.has_dep_p[i] = 0; | |
3263 | } | |
3264 | ||
3265 | /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer | |
3266 | to the dependence information array in HAS_DEP_PP. */ | |
3267 | ds_t | |
3268 | has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) | |
3269 | { | |
3270 | int i; | |
3271 | ds_t ds; | |
88302d54 | 3272 | struct deps_desc *dc; |
e855c69d AB |
3273 | |
3274 | if (INSN_SIMPLEJUMP_P (pred)) | |
3275 | /* Unconditional jump is just a transfer of control flow. | |
3276 | Ignore it. */ | |
3277 | return false; | |
3278 | ||
3279 | dc = &INSN_DEPS_CONTEXT (pred); | |
bcf33775 AB |
3280 | |
3281 | /* We init this field lazily. */ | |
3282 | if (dc->reg_last == NULL) | |
3283 | init_deps_reg_last (dc); | |
b8698a0f | 3284 | |
e855c69d AB |
3285 | if (!dc->readonly) |
3286 | { | |
3287 | has_dependence_data.pro = NULL; | |
3288 | /* Initialize empty dep context with information about PRED. */ | |
3289 | advance_deps_context (dc, pred); | |
3290 | dc->readonly = 1; | |
3291 | } | |
3292 | ||
3293 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3294 | has_dependence_data.pro = pred; | |
3295 | has_dependence_data.con = EXPR_VINSN (expr); | |
3296 | has_dependence_data.dc = dc; | |
3297 | ||
3298 | sel_clear_has_dependence (); | |
3299 | ||
3300 | /* Now catch all dependencies that would be generated between PRED and | |
3301 | INSN. */ | |
3302 | setup_has_dependence_sched_deps_info (); | |
3303 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3304 | has_dependence_data.dc = NULL; | |
3305 | ||
3306 | /* When a barrier was found, set DEPS_IN_INSN bits. */ | |
3307 | if (dc->last_reg_pending_barrier == TRUE_BARRIER) | |
3308 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; | |
3309 | else if (dc->last_reg_pending_barrier == MOVE_BARRIER) | |
3310 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
3311 | ||
3312 | /* Do not allow stores to memory to move through checks. Currently | |
3313 | we don't move this to sched-deps.c as the check doesn't have | |
b8698a0f | 3314 | obvious places to which this dependence can be attached. |
e855c69d AB |
3315 | FIMXE: this should go to a hook. */ |
3316 | if (EXPR_LHS (expr) | |
3317 | && MEM_P (EXPR_LHS (expr)) | |
3318 | && sel_insn_is_speculation_check (pred)) | |
3319 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
b8698a0f | 3320 | |
e855c69d AB |
3321 | *has_dep_pp = has_dependence_data.has_dep_p; |
3322 | ds = 0; | |
3323 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3324 | ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], | |
3325 | NULL_RTX, NULL_RTX); | |
3326 | ||
3327 | return ds; | |
3328 | } | |
3329 | \f | |
3330 | ||
b8698a0f L |
3331 | /* Dependence hooks implementation that checks dependence latency constraints |
3332 | on the insns being scheduled. The entry point for these routines is | |
3333 | tick_check_p predicate. */ | |
e855c69d AB |
3334 | |
3335 | static struct | |
3336 | { | |
3337 | /* An expr we are currently checking. */ | |
3338 | expr_t expr; | |
3339 | ||
3340 | /* A minimal cycle for its scheduling. */ | |
3341 | int cycle; | |
3342 | ||
3343 | /* Whether we have seen a true dependence while checking. */ | |
3344 | bool seen_true_dep_p; | |
3345 | } tick_check_data; | |
3346 | ||
3347 | /* Update minimal scheduling cycle for tick_check_insn given that it depends | |
3348 | on PRO with status DS and weight DW. */ | |
3349 | static void | |
3350 | tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) | |
3351 | { | |
3352 | expr_t con_expr = tick_check_data.expr; | |
3353 | insn_t con_insn = EXPR_INSN_RTX (con_expr); | |
3354 | ||
3355 | if (con_insn != pro_insn) | |
3356 | { | |
3357 | enum reg_note dt; | |
3358 | int tick; | |
3359 | ||
3360 | if (/* PROducer was removed from above due to pipelining. */ | |
3361 | !INSN_IN_STREAM_P (pro_insn) | |
3362 | /* Or PROducer was originally on the next iteration regarding the | |
3363 | CONsumer. */ | |
3364 | || (INSN_SCHED_TIMES (pro_insn) | |
3365 | - EXPR_SCHED_TIMES (con_expr)) > 1) | |
3366 | /* Don't count this dependence. */ | |
3367 | return; | |
3368 | ||
3369 | dt = ds_to_dt (ds); | |
3370 | if (dt == REG_DEP_TRUE) | |
3371 | tick_check_data.seen_true_dep_p = true; | |
3372 | ||
3373 | gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); | |
3374 | ||
3375 | { | |
3376 | dep_def _dep, *dep = &_dep; | |
3377 | ||
3378 | init_dep (dep, pro_insn, con_insn, dt); | |
3379 | ||
3380 | tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); | |
3381 | } | |
3382 | ||
3383 | /* When there are several kinds of dependencies between pro and con, | |
3384 | only REG_DEP_TRUE should be taken into account. */ | |
3385 | if (tick > tick_check_data.cycle | |
3386 | && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) | |
3387 | tick_check_data.cycle = tick; | |
3388 | } | |
3389 | } | |
3390 | ||
3391 | /* An implementation of note_dep hook. */ | |
3392 | static void | |
3393 | tick_check_note_dep (insn_t pro, ds_t ds) | |
3394 | { | |
3395 | tick_check_dep_with_dw (pro, ds, 0); | |
3396 | } | |
3397 | ||
3398 | /* An implementation of note_mem_dep hook. */ | |
3399 | static void | |
3400 | tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) | |
3401 | { | |
3402 | dw_t dw; | |
3403 | ||
3404 | dw = (ds_to_dt (ds) == REG_DEP_TRUE | |
3405 | ? estimate_dep_weak (mem1, mem2) | |
3406 | : 0); | |
3407 | ||
3408 | tick_check_dep_with_dw (pro, ds, dw); | |
3409 | } | |
3410 | ||
3411 | /* This structure contains hooks for dependence analysis used when determining | |
3412 | whether an insn is ready for scheduling. */ | |
3413 | static struct sched_deps_info_def tick_check_sched_deps_info = | |
3414 | { | |
3415 | NULL, | |
3416 | ||
3417 | NULL, | |
3418 | NULL, | |
3419 | NULL, | |
3420 | NULL, | |
3421 | NULL, | |
3422 | NULL, | |
3423 | haifa_note_reg_set, | |
3424 | haifa_note_reg_clobber, | |
3425 | haifa_note_reg_use, | |
3426 | tick_check_note_mem_dep, | |
3427 | tick_check_note_dep, | |
3428 | ||
3429 | 0, 0, 0 | |
3430 | }; | |
3431 | ||
3432 | /* Estimate number of cycles from the current cycle of FENCE until EXPR can be | |
3433 | scheduled. Return 0 if all data from producers in DC is ready. */ | |
3434 | int | |
3435 | tick_check_p (expr_t expr, deps_t dc, fence_t fence) | |
3436 | { | |
3437 | int cycles_left; | |
3438 | /* Initialize variables. */ | |
3439 | tick_check_data.expr = expr; | |
3440 | tick_check_data.cycle = 0; | |
3441 | tick_check_data.seen_true_dep_p = false; | |
3442 | sched_deps_info = &tick_check_sched_deps_info; | |
b8698a0f | 3443 | |
e855c69d AB |
3444 | gcc_assert (!dc->readonly); |
3445 | dc->readonly = 1; | |
3446 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3447 | dc->readonly = 0; | |
3448 | ||
3449 | cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); | |
3450 | ||
3451 | return cycles_left >= 0 ? cycles_left : 0; | |
3452 | } | |
3453 | \f | |
3454 | ||
3455 | /* Functions to work with insns. */ | |
3456 | ||
3457 | /* Returns true if LHS of INSN is the same as DEST of an insn | |
3458 | being moved. */ | |
3459 | bool | |
3460 | lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) | |
3461 | { | |
3462 | rtx lhs = INSN_LHS (insn); | |
3463 | ||
3464 | if (lhs == NULL || dest == NULL) | |
3465 | return false; | |
b8698a0f | 3466 | |
e855c69d AB |
3467 | return rtx_equal_p (lhs, dest); |
3468 | } | |
3469 | ||
3470 | /* Return s_i_d entry of INSN. Callable from debugger. */ | |
3471 | sel_insn_data_def | |
3472 | insn_sid (insn_t insn) | |
3473 | { | |
3474 | return *SID (insn); | |
3475 | } | |
3476 | ||
3477 | /* True when INSN is a speculative check. We can tell this by looking | |
3478 | at the data structures of the selective scheduler, not by examining | |
3479 | the pattern. */ | |
3480 | bool | |
3481 | sel_insn_is_speculation_check (rtx insn) | |
3482 | { | |
3483 | return s_i_d && !! INSN_SPEC_CHECKED_DS (insn); | |
3484 | } | |
3485 | ||
b8698a0f | 3486 | /* Extracts machine mode MODE and destination location DST_LOC |
e855c69d AB |
3487 | for given INSN. */ |
3488 | void | |
3489 | get_dest_and_mode (rtx insn, rtx *dst_loc, enum machine_mode *mode) | |
3490 | { | |
3491 | rtx pat = PATTERN (insn); | |
3492 | ||
3493 | gcc_assert (dst_loc); | |
3494 | gcc_assert (GET_CODE (pat) == SET); | |
3495 | ||
3496 | *dst_loc = SET_DEST (pat); | |
3497 | ||
3498 | gcc_assert (*dst_loc); | |
3499 | gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); | |
3500 | ||
3501 | if (mode) | |
3502 | *mode = GET_MODE (*dst_loc); | |
3503 | } | |
3504 | ||
b8698a0f | 3505 | /* Returns true when moving through JUMP will result in bookkeeping |
e855c69d AB |
3506 | creation. */ |
3507 | bool | |
3508 | bookkeeping_can_be_created_if_moved_through_p (insn_t jump) | |
3509 | { | |
3510 | insn_t succ; | |
3511 | succ_iterator si; | |
3512 | ||
3513 | FOR_EACH_SUCC (succ, si, jump) | |
3514 | if (sel_num_cfg_preds_gt_1 (succ)) | |
3515 | return true; | |
3516 | ||
3517 | return false; | |
3518 | } | |
3519 | ||
3520 | /* Return 'true' if INSN is the only one in its basic block. */ | |
3521 | static bool | |
3522 | insn_is_the_only_one_in_bb_p (insn_t insn) | |
3523 | { | |
3524 | return sel_bb_head_p (insn) && sel_bb_end_p (insn); | |
3525 | } | |
3526 | ||
3527 | #ifdef ENABLE_CHECKING | |
b8698a0f | 3528 | /* Check that the region we're scheduling still has at most one |
e855c69d AB |
3529 | backedge. */ |
3530 | static void | |
3531 | verify_backedges (void) | |
3532 | { | |
3533 | if (pipelining_p) | |
3534 | { | |
3535 | int i, n = 0; | |
3536 | edge e; | |
3537 | edge_iterator ei; | |
b8698a0f | 3538 | |
e855c69d AB |
3539 | for (i = 0; i < current_nr_blocks; i++) |
3540 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (BB_TO_BLOCK (i))->succs) | |
3541 | if (in_current_region_p (e->dest) | |
3542 | && BLOCK_TO_BB (e->dest->index) < i) | |
3543 | n++; | |
b8698a0f | 3544 | |
e855c69d AB |
3545 | gcc_assert (n <= 1); |
3546 | } | |
3547 | } | |
3548 | #endif | |
3549 | \f | |
3550 | ||
3551 | /* Functions to work with control flow. */ | |
3552 | ||
b59ab570 AM |
3553 | /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks |
3554 | are sorted in topological order (it might have been invalidated by | |
3555 | redirecting an edge). */ | |
3556 | static void | |
3557 | sel_recompute_toporder (void) | |
3558 | { | |
3559 | int i, n, rgn; | |
3560 | int *postorder, n_blocks; | |
3561 | ||
3562 | postorder = XALLOCAVEC (int, n_basic_blocks); | |
3563 | n_blocks = post_order_compute (postorder, false, false); | |
3564 | ||
3565 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
3566 | for (n = 0, i = n_blocks - 1; i >= 0; i--) | |
3567 | if (CONTAINING_RGN (postorder[i]) == rgn) | |
3568 | { | |
3569 | BLOCK_TO_BB (postorder[i]) = n; | |
3570 | BB_TO_BLOCK (n) = postorder[i]; | |
3571 | n++; | |
3572 | } | |
3573 | ||
3574 | /* Assert that we updated info for all blocks. We may miss some blocks if | |
3575 | this function is called when redirecting an edge made a block | |
3576 | unreachable, but that block is not deleted yet. */ | |
3577 | gcc_assert (n == RGN_NR_BLOCKS (rgn)); | |
3578 | } | |
3579 | ||
e855c69d | 3580 | /* Tidy the possibly empty block BB. */ |
65592aad | 3581 | static bool |
5f33b972 | 3582 | maybe_tidy_empty_bb (basic_block bb) |
e855c69d AB |
3583 | { |
3584 | basic_block succ_bb, pred_bb; | |
00c4e97c | 3585 | VEC (basic_block, heap) *dom_bbs; |
f2c45f08 AM |
3586 | edge e; |
3587 | edge_iterator ei; | |
e855c69d AB |
3588 | bool rescan_p; |
3589 | ||
3590 | /* Keep empty bb only if this block immediately precedes EXIT and | |
762bffba AB |
3591 | has incoming non-fallthrough edge, or it has no predecessors or |
3592 | successors. Otherwise remove it. */ | |
b5b8b0ac | 3593 | if (!sel_bb_empty_p (bb) |
b8698a0f | 3594 | || (single_succ_p (bb) |
e855c69d | 3595 | && single_succ (bb) == EXIT_BLOCK_PTR |
b8698a0f | 3596 | && (!single_pred_p (bb) |
762bffba AB |
3597 | || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU))) |
3598 | || EDGE_COUNT (bb->preds) == 0 | |
3599 | || EDGE_COUNT (bb->succs) == 0) | |
e855c69d AB |
3600 | return false; |
3601 | ||
f2c45f08 AM |
3602 | /* Do not attempt to redirect complex edges. */ |
3603 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3604 | if (e->flags & EDGE_COMPLEX) | |
3605 | return false; | |
3606 | ||
e855c69d AB |
3607 | free_data_sets (bb); |
3608 | ||
3609 | /* Do not delete BB if it has more than one successor. | |
3610 | That can occur when we moving a jump. */ | |
3611 | if (!single_succ_p (bb)) | |
3612 | { | |
3613 | gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); | |
3614 | sel_merge_blocks (bb->prev_bb, bb); | |
3615 | return true; | |
3616 | } | |
3617 | ||
3618 | succ_bb = single_succ (bb); | |
3619 | rescan_p = true; | |
3620 | pred_bb = NULL; | |
00c4e97c | 3621 | dom_bbs = NULL; |
e855c69d AB |
3622 | |
3623 | /* Redirect all non-fallthru edges to the next bb. */ | |
3624 | while (rescan_p) | |
3625 | { | |
e855c69d AB |
3626 | rescan_p = false; |
3627 | ||
3628 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3629 | { | |
3630 | pred_bb = e->src; | |
3631 | ||
3632 | if (!(e->flags & EDGE_FALLTHRU)) | |
3633 | { | |
5f33b972 | 3634 | /* We can not invalidate computed topological order by moving |
00c4e97c AB |
3635 | the edge destination block (E->SUCC) along a fallthru edge. |
3636 | ||
3637 | We will update dominators here only when we'll get | |
3638 | an unreachable block when redirecting, otherwise | |
3639 | sel_redirect_edge_and_branch will take care of it. */ | |
3640 | if (e->dest != bb | |
3641 | && single_pred_p (e->dest)) | |
3642 | VEC_safe_push (basic_block, heap, dom_bbs, e->dest); | |
5f33b972 | 3643 | sel_redirect_edge_and_branch (e, succ_bb); |
e855c69d AB |
3644 | rescan_p = true; |
3645 | break; | |
3646 | } | |
5f33b972 AM |
3647 | /* If the edge is fallthru, but PRED_BB ends in a conditional jump |
3648 | to BB (so there is no non-fallthru edge from PRED_BB to BB), we | |
3649 | still have to adjust it. */ | |
3650 | else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb))) | |
3651 | { | |
3652 | /* If possible, try to remove the unneeded conditional jump. */ | |
3653 | if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0 | |
3654 | && !IN_CURRENT_FENCE_P (BB_END (pred_bb))) | |
3655 | { | |
3656 | if (!sel_remove_insn (BB_END (pred_bb), false, false)) | |
3657 | tidy_fallthru_edge (e); | |
3658 | } | |
3659 | else | |
3660 | sel_redirect_edge_and_branch (e, succ_bb); | |
3661 | rescan_p = true; | |
3662 | break; | |
3663 | } | |
e855c69d AB |
3664 | } |
3665 | } | |
3666 | ||
e855c69d AB |
3667 | if (can_merge_blocks_p (bb->prev_bb, bb)) |
3668 | sel_merge_blocks (bb->prev_bb, bb); | |
3669 | else | |
e855c69d | 3670 | { |
262d8232 | 3671 | /* This is a block without fallthru predecessor. Just delete it. */ |
e855c69d AB |
3672 | gcc_assert (pred_bb != NULL); |
3673 | ||
762bffba AB |
3674 | if (in_current_region_p (pred_bb)) |
3675 | move_bb_info (pred_bb, bb); | |
e855c69d AB |
3676 | remove_empty_bb (bb, true); |
3677 | } | |
3678 | ||
00c4e97c AB |
3679 | if (!VEC_empty (basic_block, dom_bbs)) |
3680 | { | |
3681 | VEC_safe_push (basic_block, heap, dom_bbs, succ_bb); | |
3682 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); | |
3683 | VEC_free (basic_block, heap, dom_bbs); | |
3684 | } | |
3685 | ||
e855c69d AB |
3686 | return true; |
3687 | } | |
3688 | ||
b8698a0f | 3689 | /* Tidy the control flow after we have removed original insn from |
e855c69d AB |
3690 | XBB. Return true if we have removed some blocks. When FULL_TIDYING |
3691 | is true, also try to optimize control flow on non-empty blocks. */ | |
3692 | bool | |
3693 | tidy_control_flow (basic_block xbb, bool full_tidying) | |
3694 | { | |
3695 | bool changed = true; | |
b5b8b0ac | 3696 | insn_t first, last; |
b8698a0f | 3697 | |
e855c69d | 3698 | /* First check whether XBB is empty. */ |
5f33b972 | 3699 | changed = maybe_tidy_empty_bb (xbb); |
e855c69d AB |
3700 | if (changed || !full_tidying) |
3701 | return changed; | |
b8698a0f | 3702 | |
e855c69d | 3703 | /* Check if there is a unnecessary jump after insn left. */ |
753de8cf | 3704 | if (bb_has_removable_jump_to_p (xbb, xbb->next_bb) |
e855c69d AB |
3705 | && INSN_SCHED_TIMES (BB_END (xbb)) == 0 |
3706 | && !IN_CURRENT_FENCE_P (BB_END (xbb))) | |
3707 | { | |
3708 | if (sel_remove_insn (BB_END (xbb), false, false)) | |
3709 | return true; | |
3710 | tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); | |
3711 | } | |
3712 | ||
b5b8b0ac AO |
3713 | first = sel_bb_head (xbb); |
3714 | last = sel_bb_end (xbb); | |
3715 | if (MAY_HAVE_DEBUG_INSNS) | |
3716 | { | |
3717 | if (first != last && DEBUG_INSN_P (first)) | |
3718 | do | |
3719 | first = NEXT_INSN (first); | |
3720 | while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); | |
3721 | ||
3722 | if (first != last && DEBUG_INSN_P (last)) | |
3723 | do | |
3724 | last = PREV_INSN (last); | |
3725 | while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); | |
3726 | } | |
e855c69d | 3727 | /* Check if there is an unnecessary jump in previous basic block leading |
b8698a0f L |
3728 | to next basic block left after removing INSN from stream. |
3729 | If it is so, remove that jump and redirect edge to current | |
3730 | basic block (where there was INSN before deletion). This way | |
3731 | when NOP will be deleted several instructions later with its | |
3732 | basic block we will not get a jump to next instruction, which | |
e855c69d | 3733 | can be harmful. */ |
b5b8b0ac | 3734 | if (first == last |
e855c69d | 3735 | && !sel_bb_empty_p (xbb) |
b5b8b0ac | 3736 | && INSN_NOP_P (last) |
e855c69d AB |
3737 | /* Flow goes fallthru from current block to the next. */ |
3738 | && EDGE_COUNT (xbb->succs) == 1 | |
3739 | && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) | |
3740 | /* When successor is an EXIT block, it may not be the next block. */ | |
3741 | && single_succ (xbb) != EXIT_BLOCK_PTR | |
3742 | /* And unconditional jump in previous basic block leads to | |
3743 | next basic block of XBB and this jump can be safely removed. */ | |
3744 | && in_current_region_p (xbb->prev_bb) | |
753de8cf | 3745 | && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb) |
e855c69d AB |
3746 | && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 |
3747 | /* Also this jump is not at the scheduling boundary. */ | |
3748 | && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) | |
3749 | { | |
b59ab570 | 3750 | bool recompute_toporder_p; |
e855c69d AB |
3751 | /* Clear data structures of jump - jump itself will be removed |
3752 | by sel_redirect_edge_and_branch. */ | |
3753 | clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); | |
b59ab570 AM |
3754 | recompute_toporder_p |
3755 | = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); | |
3756 | ||
e855c69d AB |
3757 | gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); |
3758 | ||
3759 | /* It can turn out that after removing unused jump, basic block | |
3760 | that contained that jump, becomes empty too. In such case | |
3761 | remove it too. */ | |
3762 | if (sel_bb_empty_p (xbb->prev_bb)) | |
5f33b972 AM |
3763 | changed = maybe_tidy_empty_bb (xbb->prev_bb); |
3764 | if (recompute_toporder_p) | |
b59ab570 | 3765 | sel_recompute_toporder (); |
e855c69d | 3766 | } |
d787f788 AM |
3767 | |
3768 | #ifdef ENABLE_CHECKING | |
3769 | verify_backedges (); | |
00c4e97c | 3770 | verify_dominators (CDI_DOMINATORS); |
d787f788 AM |
3771 | #endif |
3772 | ||
e855c69d AB |
3773 | return changed; |
3774 | } | |
3775 | ||
b59ab570 AM |
3776 | /* Purge meaningless empty blocks in the middle of a region. */ |
3777 | void | |
3778 | purge_empty_blocks (void) | |
3779 | { | |
9d0dcda1 | 3780 | int i; |
b59ab570 | 3781 | |
9d0dcda1 AM |
3782 | /* Do not attempt to delete the first basic block in the region. */ |
3783 | for (i = 1; i < current_nr_blocks; ) | |
b59ab570 AM |
3784 | { |
3785 | basic_block b = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
3786 | ||
5f33b972 | 3787 | if (maybe_tidy_empty_bb (b)) |
b59ab570 AM |
3788 | continue; |
3789 | ||
3790 | i++; | |
3791 | } | |
3792 | } | |
3793 | ||
b8698a0f L |
3794 | /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, |
3795 | do not delete insn's data, because it will be later re-emitted. | |
e855c69d AB |
3796 | Return true if we have removed some blocks afterwards. */ |
3797 | bool | |
3798 | sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) | |
3799 | { | |
3800 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3801 | ||
3802 | gcc_assert (INSN_IN_STREAM_P (insn)); | |
3803 | ||
b5b8b0ac AO |
3804 | if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) |
3805 | { | |
3806 | expr_t expr; | |
3807 | av_set_iterator i; | |
3808 | ||
3809 | /* When we remove a debug insn that is head of a BB, it remains | |
3810 | in the AV_SET of the block, but it shouldn't. */ | |
3811 | FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) | |
3812 | if (EXPR_INSN_RTX (expr) == insn) | |
3813 | { | |
3814 | av_set_iter_remove (&i); | |
3815 | break; | |
3816 | } | |
3817 | } | |
3818 | ||
e855c69d AB |
3819 | if (only_disconnect) |
3820 | { | |
3821 | insn_t prev = PREV_INSN (insn); | |
3822 | insn_t next = NEXT_INSN (insn); | |
3823 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3824 | ||
3825 | NEXT_INSN (prev) = next; | |
3826 | PREV_INSN (next) = prev; | |
3827 | ||
3828 | if (BB_HEAD (bb) == insn) | |
3829 | { | |
3830 | gcc_assert (BLOCK_FOR_INSN (prev) == bb); | |
3831 | BB_HEAD (bb) = prev; | |
3832 | } | |
3833 | if (BB_END (bb) == insn) | |
3834 | BB_END (bb) = prev; | |
3835 | } | |
3836 | else | |
3837 | { | |
3838 | remove_insn (insn); | |
3839 | clear_expr (INSN_EXPR (insn)); | |
3840 | } | |
3841 | ||
3842 | /* It is necessary to null this fields before calling add_insn (). */ | |
3843 | PREV_INSN (insn) = NULL_RTX; | |
3844 | NEXT_INSN (insn) = NULL_RTX; | |
3845 | ||
3846 | return tidy_control_flow (bb, full_tidying); | |
3847 | } | |
3848 | ||
3849 | /* Estimate number of the insns in BB. */ | |
3850 | static int | |
3851 | sel_estimate_number_of_insns (basic_block bb) | |
3852 | { | |
3853 | int res = 0; | |
3854 | insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); | |
3855 | ||
3856 | for (; insn != next_tail; insn = NEXT_INSN (insn)) | |
b5b8b0ac | 3857 | if (NONDEBUG_INSN_P (insn)) |
e855c69d AB |
3858 | res++; |
3859 | ||
3860 | return res; | |
3861 | } | |
3862 | ||
3863 | /* We don't need separate luids for notes or labels. */ | |
3864 | static int | |
3865 | sel_luid_for_non_insn (rtx x) | |
3866 | { | |
3867 | gcc_assert (NOTE_P (x) || LABEL_P (x)); | |
3868 | ||
3869 | return -1; | |
3870 | } | |
3871 | ||
3872 | /* Return seqno of the only predecessor of INSN. */ | |
3873 | static int | |
3874 | get_seqno_of_a_pred (insn_t insn) | |
3875 | { | |
3876 | int seqno; | |
3877 | ||
3878 | gcc_assert (INSN_SIMPLEJUMP_P (insn)); | |
3879 | ||
3880 | if (!sel_bb_head_p (insn)) | |
3881 | seqno = INSN_SEQNO (PREV_INSN (insn)); | |
3882 | else | |
3883 | { | |
3884 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3885 | ||
3886 | if (single_pred_p (bb) | |
3887 | && !in_current_region_p (single_pred (bb))) | |
3888 | { | |
3889 | /* We can have preds outside a region when splitting edges | |
b8698a0f | 3890 | for pipelining of an outer loop. Use succ instead. |
e855c69d AB |
3891 | There should be only one of them. */ |
3892 | insn_t succ = NULL; | |
3893 | succ_iterator si; | |
3894 | bool first = true; | |
b8698a0f | 3895 | |
e855c69d AB |
3896 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
3897 | && current_loop_nest); | |
b8698a0f | 3898 | FOR_EACH_SUCC_1 (succ, si, insn, |
e855c69d AB |
3899 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
3900 | { | |
3901 | gcc_assert (first); | |
3902 | first = false; | |
3903 | } | |
3904 | ||
3905 | gcc_assert (succ != NULL); | |
3906 | seqno = INSN_SEQNO (succ); | |
3907 | } | |
3908 | else | |
3909 | { | |
3910 | insn_t *preds; | |
3911 | int n; | |
3912 | ||
3913 | cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); | |
3914 | gcc_assert (n == 1); | |
3915 | ||
3916 | seqno = INSN_SEQNO (preds[0]); | |
b8698a0f | 3917 | |
e855c69d AB |
3918 | free (preds); |
3919 | } | |
3920 | } | |
3921 | ||
3922 | return seqno; | |
3923 | } | |
3924 | ||
da7ba240 AB |
3925 | /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors |
3926 | with positive seqno exist. */ | |
e855c69d AB |
3927 | int |
3928 | get_seqno_by_preds (rtx insn) | |
3929 | { | |
3930 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3931 | rtx tmp = insn, head = BB_HEAD (bb); | |
3932 | insn_t *preds; | |
3933 | int n, i, seqno; | |
3934 | ||
3935 | while (tmp != head) | |
3936 | if (INSN_P (tmp)) | |
3937 | return INSN_SEQNO (tmp); | |
3938 | else | |
3939 | tmp = PREV_INSN (tmp); | |
b8698a0f | 3940 | |
e855c69d AB |
3941 | cfg_preds (bb, &preds, &n); |
3942 | for (i = 0, seqno = -1; i < n; i++) | |
3943 | seqno = MAX (seqno, INSN_SEQNO (preds[i])); | |
3944 | ||
e855c69d AB |
3945 | return seqno; |
3946 | } | |
3947 | ||
3948 | \f | |
3949 | ||
3950 | /* Extend pass-scope data structures for basic blocks. */ | |
3951 | void | |
3952 | sel_extend_global_bb_info (void) | |
3953 | { | |
3954 | VEC_safe_grow_cleared (sel_global_bb_info_def, heap, sel_global_bb_info, | |
3955 | last_basic_block); | |
3956 | } | |
3957 | ||
3958 | /* Extend region-scope data structures for basic blocks. */ | |
3959 | static void | |
3960 | extend_region_bb_info (void) | |
3961 | { | |
3962 | VEC_safe_grow_cleared (sel_region_bb_info_def, heap, sel_region_bb_info, | |
3963 | last_basic_block); | |
3964 | } | |
3965 | ||
3966 | /* Extend all data structures to fit for all basic blocks. */ | |
3967 | static void | |
3968 | extend_bb_info (void) | |
3969 | { | |
3970 | sel_extend_global_bb_info (); | |
3971 | extend_region_bb_info (); | |
3972 | } | |
3973 | ||
3974 | /* Finalize pass-scope data structures for basic blocks. */ | |
3975 | void | |
3976 | sel_finish_global_bb_info (void) | |
3977 | { | |
3978 | VEC_free (sel_global_bb_info_def, heap, sel_global_bb_info); | |
3979 | } | |
3980 | ||
3981 | /* Finalize region-scope data structures for basic blocks. */ | |
3982 | static void | |
3983 | finish_region_bb_info (void) | |
3984 | { | |
3985 | VEC_free (sel_region_bb_info_def, heap, sel_region_bb_info); | |
3986 | } | |
3987 | \f | |
3988 | ||
3989 | /* Data for each insn in current region. */ | |
3990 | VEC (sel_insn_data_def, heap) *s_i_d = NULL; | |
3991 | ||
3992 | /* A vector for the insns we've emitted. */ | |
3993 | static insn_vec_t new_insns = NULL; | |
3994 | ||
3995 | /* Extend data structures for insns from current region. */ | |
3996 | static void | |
3997 | extend_insn_data (void) | |
3998 | { | |
3999 | int reserve; | |
b8698a0f | 4000 | |
e855c69d AB |
4001 | sched_extend_target (); |
4002 | sched_deps_init (false); | |
4003 | ||
4004 | /* Extend data structures for insns from current region. */ | |
b8698a0f | 4005 | reserve = (sched_max_luid + 1 |
e855c69d | 4006 | - VEC_length (sel_insn_data_def, s_i_d)); |
b8698a0f | 4007 | if (reserve > 0 |
e855c69d | 4008 | && ! VEC_space (sel_insn_data_def, s_i_d, reserve)) |
bcf33775 AB |
4009 | { |
4010 | int size; | |
4011 | ||
4012 | if (sched_max_luid / 2 > 1024) | |
4013 | size = sched_max_luid + 1024; | |
4014 | else | |
4015 | size = 3 * sched_max_luid / 2; | |
b8698a0f | 4016 | |
bcf33775 AB |
4017 | |
4018 | VEC_safe_grow_cleared (sel_insn_data_def, heap, s_i_d, size); | |
4019 | } | |
e855c69d AB |
4020 | } |
4021 | ||
4022 | /* Finalize data structures for insns from current region. */ | |
4023 | static void | |
4024 | finish_insns (void) | |
4025 | { | |
4026 | unsigned i; | |
4027 | ||
4028 | /* Clear here all dependence contexts that may have left from insns that were | |
4029 | removed during the scheduling. */ | |
4030 | for (i = 0; i < VEC_length (sel_insn_data_def, s_i_d); i++) | |
4031 | { | |
4032 | sel_insn_data_def *sid_entry = VEC_index (sel_insn_data_def, s_i_d, i); | |
b8698a0f | 4033 | |
e855c69d AB |
4034 | if (sid_entry->live) |
4035 | return_regset_to_pool (sid_entry->live); | |
4036 | if (sid_entry->analyzed_deps) | |
4037 | { | |
4038 | BITMAP_FREE (sid_entry->analyzed_deps); | |
4039 | BITMAP_FREE (sid_entry->found_deps); | |
4040 | htab_delete (sid_entry->transformed_insns); | |
4041 | free_deps (&sid_entry->deps_context); | |
4042 | } | |
4043 | if (EXPR_VINSN (&sid_entry->expr)) | |
4044 | { | |
4045 | clear_expr (&sid_entry->expr); | |
b8698a0f | 4046 | |
e855c69d AB |
4047 | /* Also, clear CANT_MOVE bit here, because we really don't want it |
4048 | to be passed to the next region. */ | |
4049 | CANT_MOVE_BY_LUID (i) = 0; | |
4050 | } | |
4051 | } | |
b8698a0f | 4052 | |
e855c69d AB |
4053 | VEC_free (sel_insn_data_def, heap, s_i_d); |
4054 | } | |
4055 | ||
4056 | /* A proxy to pass initialization data to init_insn (). */ | |
4057 | static sel_insn_data_def _insn_init_ssid; | |
4058 | static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; | |
4059 | ||
4060 | /* If true create a new vinsn. Otherwise use the one from EXPR. */ | |
4061 | static bool insn_init_create_new_vinsn_p; | |
4062 | ||
4063 | /* Set all necessary data for initialization of the new insn[s]. */ | |
4064 | static expr_t | |
4065 | set_insn_init (expr_t expr, vinsn_t vi, int seqno) | |
4066 | { | |
4067 | expr_t x = &insn_init_ssid->expr; | |
4068 | ||
4069 | copy_expr_onside (x, expr); | |
4070 | if (vi != NULL) | |
4071 | { | |
4072 | insn_init_create_new_vinsn_p = false; | |
4073 | change_vinsn_in_expr (x, vi); | |
4074 | } | |
4075 | else | |
4076 | insn_init_create_new_vinsn_p = true; | |
4077 | ||
4078 | insn_init_ssid->seqno = seqno; | |
4079 | return x; | |
4080 | } | |
4081 | ||
4082 | /* Init data for INSN. */ | |
4083 | static void | |
4084 | init_insn_data (insn_t insn) | |
4085 | { | |
4086 | expr_t expr; | |
4087 | sel_insn_data_t ssid = insn_init_ssid; | |
4088 | ||
4089 | /* The fields mentioned below are special and hence are not being | |
4090 | propagated to the new insns. */ | |
4091 | gcc_assert (!ssid->asm_p && ssid->sched_next == NULL | |
4092 | && !ssid->after_stall_p && ssid->sched_cycle == 0); | |
4093 | gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); | |
4094 | ||
4095 | expr = INSN_EXPR (insn); | |
4096 | copy_expr (expr, &ssid->expr); | |
4097 | prepare_insn_expr (insn, ssid->seqno); | |
4098 | ||
4099 | if (insn_init_create_new_vinsn_p) | |
4100 | change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p)); | |
b8698a0f | 4101 | |
e855c69d AB |
4102 | if (first_time_insn_init (insn)) |
4103 | init_first_time_insn_data (insn); | |
4104 | } | |
4105 | ||
4106 | /* This is used to initialize spurious jumps generated by | |
4107 | sel_redirect_edge (). */ | |
4108 | static void | |
4109 | init_simplejump_data (insn_t insn) | |
4110 | { | |
4111 | init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0, | |
b8698a0f | 4112 | REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0, NULL, true, false, false, |
e855c69d AB |
4113 | false, true); |
4114 | INSN_SEQNO (insn) = get_seqno_of_a_pred (insn); | |
4115 | init_first_time_insn_data (insn); | |
4116 | } | |
4117 | ||
b8698a0f | 4118 | /* Perform deferred initialization of insns. This is used to process |
e855c69d AB |
4119 | a new jump that may be created by redirect_edge. */ |
4120 | void | |
4121 | sel_init_new_insn (insn_t insn, int flags) | |
4122 | { | |
4123 | /* We create data structures for bb when the first insn is emitted in it. */ | |
4124 | if (INSN_P (insn) | |
4125 | && INSN_IN_STREAM_P (insn) | |
4126 | && insn_is_the_only_one_in_bb_p (insn)) | |
4127 | { | |
4128 | extend_bb_info (); | |
4129 | create_initial_data_sets (BLOCK_FOR_INSN (insn)); | |
4130 | } | |
b8698a0f | 4131 | |
e855c69d AB |
4132 | if (flags & INSN_INIT_TODO_LUID) |
4133 | sched_init_luids (NULL, NULL, NULL, insn); | |
4134 | ||
4135 | if (flags & INSN_INIT_TODO_SSID) | |
4136 | { | |
4137 | extend_insn_data (); | |
4138 | init_insn_data (insn); | |
4139 | clear_expr (&insn_init_ssid->expr); | |
4140 | } | |
4141 | ||
4142 | if (flags & INSN_INIT_TODO_SIMPLEJUMP) | |
4143 | { | |
4144 | extend_insn_data (); | |
4145 | init_simplejump_data (insn); | |
4146 | } | |
b8698a0f | 4147 | |
e855c69d AB |
4148 | gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) |
4149 | == CONTAINING_RGN (BB_TO_BLOCK (0))); | |
4150 | } | |
4151 | \f | |
4152 | ||
4153 | /* Functions to init/finish work with lv sets. */ | |
4154 | ||
4155 | /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ | |
4156 | static void | |
4157 | init_lv_set (basic_block bb) | |
4158 | { | |
4159 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
4160 | ||
4161 | BB_LV_SET (bb) = get_regset_from_pool (); | |
b8698a0f | 4162 | COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); |
e855c69d AB |
4163 | BB_LV_SET_VALID_P (bb) = true; |
4164 | } | |
4165 | ||
4166 | /* Copy liveness information to BB from FROM_BB. */ | |
4167 | static void | |
4168 | copy_lv_set_from (basic_block bb, basic_block from_bb) | |
4169 | { | |
4170 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
b8698a0f | 4171 | |
e855c69d AB |
4172 | COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); |
4173 | BB_LV_SET_VALID_P (bb) = true; | |
b8698a0f | 4174 | } |
e855c69d AB |
4175 | |
4176 | /* Initialize lv set of all bb headers. */ | |
4177 | void | |
4178 | init_lv_sets (void) | |
4179 | { | |
4180 | basic_block bb; | |
4181 | ||
4182 | /* Initialize of LV sets. */ | |
4183 | FOR_EACH_BB (bb) | |
4184 | init_lv_set (bb); | |
4185 | ||
4186 | /* Don't forget EXIT_BLOCK. */ | |
4187 | init_lv_set (EXIT_BLOCK_PTR); | |
4188 | } | |
4189 | ||
4190 | /* Release lv set of HEAD. */ | |
4191 | static void | |
4192 | free_lv_set (basic_block bb) | |
4193 | { | |
4194 | gcc_assert (BB_LV_SET (bb) != NULL); | |
4195 | ||
4196 | return_regset_to_pool (BB_LV_SET (bb)); | |
4197 | BB_LV_SET (bb) = NULL; | |
4198 | BB_LV_SET_VALID_P (bb) = false; | |
4199 | } | |
4200 | ||
4201 | /* Finalize lv sets of all bb headers. */ | |
4202 | void | |
4203 | free_lv_sets (void) | |
4204 | { | |
4205 | basic_block bb; | |
4206 | ||
4207 | /* Don't forget EXIT_BLOCK. */ | |
4208 | free_lv_set (EXIT_BLOCK_PTR); | |
4209 | ||
4210 | /* Free LV sets. */ | |
4211 | FOR_EACH_BB (bb) | |
4212 | if (BB_LV_SET (bb)) | |
4213 | free_lv_set (bb); | |
4214 | } | |
4215 | ||
4216 | /* Initialize an invalid AV_SET for BB. | |
4217 | This set will be updated next time compute_av () process BB. */ | |
4218 | static void | |
4219 | invalidate_av_set (basic_block bb) | |
4220 | { | |
4221 | gcc_assert (BB_AV_LEVEL (bb) <= 0 | |
4222 | && BB_AV_SET (bb) == NULL); | |
4223 | ||
4224 | BB_AV_LEVEL (bb) = -1; | |
4225 | } | |
4226 | ||
4227 | /* Create initial data sets for BB (they will be invalid). */ | |
4228 | static void | |
4229 | create_initial_data_sets (basic_block bb) | |
4230 | { | |
4231 | if (BB_LV_SET (bb)) | |
4232 | BB_LV_SET_VALID_P (bb) = false; | |
4233 | else | |
4234 | BB_LV_SET (bb) = get_regset_from_pool (); | |
4235 | invalidate_av_set (bb); | |
4236 | } | |
4237 | ||
4238 | /* Free av set of BB. */ | |
4239 | static void | |
4240 | free_av_set (basic_block bb) | |
4241 | { | |
4242 | av_set_clear (&BB_AV_SET (bb)); | |
4243 | BB_AV_LEVEL (bb) = 0; | |
4244 | } | |
4245 | ||
4246 | /* Free data sets of BB. */ | |
4247 | void | |
4248 | free_data_sets (basic_block bb) | |
4249 | { | |
4250 | free_lv_set (bb); | |
4251 | free_av_set (bb); | |
4252 | } | |
4253 | ||
4254 | /* Exchange lv sets of TO and FROM. */ | |
4255 | static void | |
4256 | exchange_lv_sets (basic_block to, basic_block from) | |
4257 | { | |
4258 | { | |
4259 | regset to_lv_set = BB_LV_SET (to); | |
4260 | ||
4261 | BB_LV_SET (to) = BB_LV_SET (from); | |
4262 | BB_LV_SET (from) = to_lv_set; | |
4263 | } | |
4264 | ||
4265 | { | |
4266 | bool to_lv_set_valid_p = BB_LV_SET_VALID_P (to); | |
4267 | ||
4268 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4269 | BB_LV_SET_VALID_P (from) = to_lv_set_valid_p; | |
4270 | } | |
4271 | } | |
4272 | ||
4273 | ||
4274 | /* Exchange av sets of TO and FROM. */ | |
4275 | static void | |
4276 | exchange_av_sets (basic_block to, basic_block from) | |
4277 | { | |
4278 | { | |
4279 | av_set_t to_av_set = BB_AV_SET (to); | |
4280 | ||
4281 | BB_AV_SET (to) = BB_AV_SET (from); | |
4282 | BB_AV_SET (from) = to_av_set; | |
4283 | } | |
4284 | ||
4285 | { | |
4286 | int to_av_level = BB_AV_LEVEL (to); | |
4287 | ||
4288 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4289 | BB_AV_LEVEL (from) = to_av_level; | |
4290 | } | |
4291 | } | |
4292 | ||
4293 | /* Exchange data sets of TO and FROM. */ | |
4294 | void | |
4295 | exchange_data_sets (basic_block to, basic_block from) | |
4296 | { | |
4297 | exchange_lv_sets (to, from); | |
4298 | exchange_av_sets (to, from); | |
4299 | } | |
4300 | ||
4301 | /* Copy data sets of FROM to TO. */ | |
4302 | void | |
4303 | copy_data_sets (basic_block to, basic_block from) | |
4304 | { | |
4305 | gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); | |
4306 | gcc_assert (BB_AV_SET (to) == NULL); | |
4307 | ||
4308 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4309 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4310 | ||
4311 | if (BB_AV_SET_VALID_P (from)) | |
4312 | { | |
4313 | BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); | |
4314 | } | |
4315 | if (BB_LV_SET_VALID_P (from)) | |
4316 | { | |
4317 | gcc_assert (BB_LV_SET (to) != NULL); | |
4318 | COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); | |
4319 | } | |
4320 | } | |
4321 | ||
4322 | /* Return an av set for INSN, if any. */ | |
4323 | av_set_t | |
4324 | get_av_set (insn_t insn) | |
4325 | { | |
4326 | av_set_t av_set; | |
4327 | ||
4328 | gcc_assert (AV_SET_VALID_P (insn)); | |
4329 | ||
4330 | if (sel_bb_head_p (insn)) | |
4331 | av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); | |
4332 | else | |
4333 | av_set = NULL; | |
4334 | ||
4335 | return av_set; | |
4336 | } | |
4337 | ||
4338 | /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ | |
4339 | int | |
4340 | get_av_level (insn_t insn) | |
4341 | { | |
4342 | int av_level; | |
4343 | ||
4344 | gcc_assert (INSN_P (insn)); | |
4345 | ||
4346 | if (sel_bb_head_p (insn)) | |
4347 | av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); | |
4348 | else | |
4349 | av_level = INSN_WS_LEVEL (insn); | |
4350 | ||
4351 | return av_level; | |
4352 | } | |
4353 | ||
4354 | \f | |
4355 | ||
4356 | /* Variables to work with control-flow graph. */ | |
4357 | ||
4358 | /* The basic block that already has been processed by the sched_data_update (), | |
4359 | but hasn't been in sel_add_bb () yet. */ | |
4360 | static VEC (basic_block, heap) *last_added_blocks = NULL; | |
4361 | ||
4362 | /* A pool for allocating successor infos. */ | |
4363 | static struct | |
4364 | { | |
4365 | /* A stack for saving succs_info structures. */ | |
4366 | struct succs_info *stack; | |
4367 | ||
4368 | /* Its size. */ | |
4369 | int size; | |
4370 | ||
4371 | /* Top of the stack. */ | |
4372 | int top; | |
4373 | ||
4374 | /* Maximal value of the top. */ | |
4375 | int max_top; | |
4376 | } succs_info_pool; | |
4377 | ||
4378 | /* Functions to work with control-flow graph. */ | |
4379 | ||
4380 | /* Return basic block note of BB. */ | |
4381 | insn_t | |
4382 | sel_bb_head (basic_block bb) | |
4383 | { | |
4384 | insn_t head; | |
4385 | ||
4386 | if (bb == EXIT_BLOCK_PTR) | |
4387 | { | |
4388 | gcc_assert (exit_insn != NULL_RTX); | |
4389 | head = exit_insn; | |
4390 | } | |
4391 | else | |
4392 | { | |
4393 | insn_t note; | |
4394 | ||
4395 | note = bb_note (bb); | |
4396 | head = next_nonnote_insn (note); | |
4397 | ||
89ad0f25 | 4398 | if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb)) |
e855c69d AB |
4399 | head = NULL_RTX; |
4400 | } | |
4401 | ||
4402 | return head; | |
4403 | } | |
4404 | ||
4405 | /* Return true if INSN is a basic block header. */ | |
4406 | bool | |
4407 | sel_bb_head_p (insn_t insn) | |
4408 | { | |
4409 | return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn; | |
4410 | } | |
4411 | ||
4412 | /* Return last insn of BB. */ | |
4413 | insn_t | |
4414 | sel_bb_end (basic_block bb) | |
4415 | { | |
4416 | if (sel_bb_empty_p (bb)) | |
4417 | return NULL_RTX; | |
4418 | ||
4419 | gcc_assert (bb != EXIT_BLOCK_PTR); | |
4420 | ||
4421 | return BB_END (bb); | |
4422 | } | |
4423 | ||
4424 | /* Return true if INSN is the last insn in its basic block. */ | |
4425 | bool | |
4426 | sel_bb_end_p (insn_t insn) | |
4427 | { | |
4428 | return insn == sel_bb_end (BLOCK_FOR_INSN (insn)); | |
4429 | } | |
4430 | ||
4431 | /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ | |
4432 | bool | |
4433 | sel_bb_empty_p (basic_block bb) | |
4434 | { | |
4435 | return sel_bb_head (bb) == NULL; | |
4436 | } | |
4437 | ||
4438 | /* True when BB belongs to the current scheduling region. */ | |
4439 | bool | |
4440 | in_current_region_p (basic_block bb) | |
4441 | { | |
4442 | if (bb->index < NUM_FIXED_BLOCKS) | |
4443 | return false; | |
4444 | ||
4445 | return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4446 | } | |
4447 | ||
4448 | /* Return the block which is a fallthru bb of a conditional jump JUMP. */ | |
4449 | basic_block | |
4450 | fallthru_bb_of_jump (rtx jump) | |
4451 | { | |
4452 | if (!JUMP_P (jump)) | |
4453 | return NULL; | |
4454 | ||
e855c69d AB |
4455 | if (!any_condjump_p (jump)) |
4456 | return NULL; | |
4457 | ||
268bab85 AB |
4458 | /* A basic block that ends with a conditional jump may still have one successor |
4459 | (and be followed by a barrier), we are not interested. */ | |
4460 | if (single_succ_p (BLOCK_FOR_INSN (jump))) | |
4461 | return NULL; | |
4462 | ||
e855c69d AB |
4463 | return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; |
4464 | } | |
4465 | ||
4466 | /* Remove all notes from BB. */ | |
4467 | static void | |
4468 | init_bb (basic_block bb) | |
4469 | { | |
4470 | remove_notes (bb_note (bb), BB_END (bb)); | |
4471 | BB_NOTE_LIST (bb) = note_list; | |
4472 | } | |
4473 | ||
4474 | void | |
4475 | sel_init_bbs (bb_vec_t bbs, basic_block bb) | |
4476 | { | |
4477 | const struct sched_scan_info_def ssi = | |
4478 | { | |
4479 | extend_bb_info, /* extend_bb */ | |
4480 | init_bb, /* init_bb */ | |
4481 | NULL, /* extend_insn */ | |
4482 | NULL /* init_insn */ | |
4483 | }; | |
4484 | ||
4485 | sched_scan (&ssi, bbs, bb, new_insns, NULL); | |
4486 | } | |
4487 | ||
7898b93b | 4488 | /* Restore notes for the whole region. */ |
e855c69d | 4489 | static void |
7898b93b | 4490 | sel_restore_notes (void) |
e855c69d AB |
4491 | { |
4492 | int bb; | |
7898b93b | 4493 | insn_t insn; |
e855c69d AB |
4494 | |
4495 | for (bb = 0; bb < current_nr_blocks; bb++) | |
4496 | { | |
4497 | basic_block first, last; | |
4498 | ||
4499 | first = EBB_FIRST_BB (bb); | |
4500 | last = EBB_LAST_BB (bb)->next_bb; | |
4501 | ||
4502 | do | |
4503 | { | |
4504 | note_list = BB_NOTE_LIST (first); | |
4505 | restore_other_notes (NULL, first); | |
4506 | BB_NOTE_LIST (first) = NULL_RTX; | |
4507 | ||
7898b93b AM |
4508 | FOR_BB_INSNS (first, insn) |
4509 | if (NONDEBUG_INSN_P (insn)) | |
4510 | reemit_notes (insn); | |
4511 | ||
e855c69d AB |
4512 | first = first->next_bb; |
4513 | } | |
4514 | while (first != last); | |
4515 | } | |
4516 | } | |
4517 | ||
4518 | /* Free per-bb data structures. */ | |
4519 | void | |
4520 | sel_finish_bbs (void) | |
4521 | { | |
7898b93b | 4522 | sel_restore_notes (); |
e855c69d AB |
4523 | |
4524 | /* Remove current loop preheader from this loop. */ | |
4525 | if (current_loop_nest) | |
4526 | sel_remove_loop_preheader (); | |
4527 | ||
4528 | finish_region_bb_info (); | |
4529 | } | |
4530 | ||
4531 | /* Return true if INSN has a single successor of type FLAGS. */ | |
4532 | bool | |
4533 | sel_insn_has_single_succ_p (insn_t insn, int flags) | |
4534 | { | |
4535 | insn_t succ; | |
4536 | succ_iterator si; | |
4537 | bool first_p = true; | |
4538 | ||
4539 | FOR_EACH_SUCC_1 (succ, si, insn, flags) | |
4540 | { | |
4541 | if (first_p) | |
4542 | first_p = false; | |
4543 | else | |
4544 | return false; | |
4545 | } | |
4546 | ||
4547 | return true; | |
4548 | } | |
4549 | ||
4550 | /* Allocate successor's info. */ | |
4551 | static struct succs_info * | |
4552 | alloc_succs_info (void) | |
4553 | { | |
4554 | if (succs_info_pool.top == succs_info_pool.max_top) | |
4555 | { | |
4556 | int i; | |
b8698a0f | 4557 | |
e855c69d AB |
4558 | if (++succs_info_pool.max_top >= succs_info_pool.size) |
4559 | gcc_unreachable (); | |
4560 | ||
4561 | i = ++succs_info_pool.top; | |
4562 | succs_info_pool.stack[i].succs_ok = VEC_alloc (rtx, heap, 10); | |
4563 | succs_info_pool.stack[i].succs_other = VEC_alloc (rtx, heap, 10); | |
4564 | succs_info_pool.stack[i].probs_ok = VEC_alloc (int, heap, 10); | |
4565 | } | |
4566 | else | |
4567 | succs_info_pool.top++; | |
4568 | ||
4569 | return &succs_info_pool.stack[succs_info_pool.top]; | |
4570 | } | |
4571 | ||
4572 | /* Free successor's info. */ | |
4573 | void | |
4574 | free_succs_info (struct succs_info * sinfo) | |
4575 | { | |
b8698a0f | 4576 | gcc_assert (succs_info_pool.top >= 0 |
e855c69d AB |
4577 | && &succs_info_pool.stack[succs_info_pool.top] == sinfo); |
4578 | succs_info_pool.top--; | |
4579 | ||
4580 | /* Clear stale info. */ | |
b8698a0f | 4581 | VEC_block_remove (rtx, sinfo->succs_ok, |
e855c69d | 4582 | 0, VEC_length (rtx, sinfo->succs_ok)); |
b8698a0f | 4583 | VEC_block_remove (rtx, sinfo->succs_other, |
e855c69d | 4584 | 0, VEC_length (rtx, sinfo->succs_other)); |
b8698a0f | 4585 | VEC_block_remove (int, sinfo->probs_ok, |
e855c69d AB |
4586 | 0, VEC_length (int, sinfo->probs_ok)); |
4587 | sinfo->all_prob = 0; | |
4588 | sinfo->succs_ok_n = 0; | |
4589 | sinfo->all_succs_n = 0; | |
4590 | } | |
4591 | ||
b8698a0f | 4592 | /* Compute successor info for INSN. FLAGS are the flags passed |
e855c69d AB |
4593 | to the FOR_EACH_SUCC_1 iterator. */ |
4594 | struct succs_info * | |
4595 | compute_succs_info (insn_t insn, short flags) | |
4596 | { | |
4597 | succ_iterator si; | |
4598 | insn_t succ; | |
4599 | struct succs_info *sinfo = alloc_succs_info (); | |
4600 | ||
4601 | /* Traverse *all* successors and decide what to do with each. */ | |
4602 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
4603 | { | |
4604 | /* FIXME: this doesn't work for skipping to loop exits, as we don't | |
4605 | perform code motion through inner loops. */ | |
4606 | short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; | |
4607 | ||
4608 | if (current_flags & flags) | |
4609 | { | |
4610 | VEC_safe_push (rtx, heap, sinfo->succs_ok, succ); | |
4611 | VEC_safe_push (int, heap, sinfo->probs_ok, | |
b8698a0f | 4612 | /* FIXME: Improve calculation when skipping |
e855c69d | 4613 | inner loop to exits. */ |
b8698a0f L |
4614 | (si.bb_end |
4615 | ? si.e1->probability | |
e855c69d AB |
4616 | : REG_BR_PROB_BASE)); |
4617 | sinfo->succs_ok_n++; | |
4618 | } | |
4619 | else | |
4620 | VEC_safe_push (rtx, heap, sinfo->succs_other, succ); | |
4621 | ||
4622 | /* Compute all_prob. */ | |
4623 | if (!si.bb_end) | |
4624 | sinfo->all_prob = REG_BR_PROB_BASE; | |
4625 | else | |
4626 | sinfo->all_prob += si.e1->probability; | |
4627 | ||
4628 | sinfo->all_succs_n++; | |
4629 | } | |
4630 | ||
4631 | return sinfo; | |
4632 | } | |
4633 | ||
b8698a0f | 4634 | /* Return the predecessors of BB in PREDS and their number in N. |
e855c69d AB |
4635 | Empty blocks are skipped. SIZE is used to allocate PREDS. */ |
4636 | static void | |
4637 | cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) | |
4638 | { | |
4639 | edge e; | |
4640 | edge_iterator ei; | |
4641 | ||
4642 | gcc_assert (BLOCK_TO_BB (bb->index) != 0); | |
4643 | ||
4644 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4645 | { | |
4646 | basic_block pred_bb = e->src; | |
4647 | insn_t bb_end = BB_END (pred_bb); | |
4648 | ||
3e6a3f6f AB |
4649 | if (!in_current_region_p (pred_bb)) |
4650 | { | |
4651 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
4652 | && current_loop_nest); | |
4653 | continue; | |
4654 | } | |
e855c69d AB |
4655 | |
4656 | if (sel_bb_empty_p (pred_bb)) | |
4657 | cfg_preds_1 (pred_bb, preds, n, size); | |
4658 | else | |
4659 | { | |
4660 | if (*n == *size) | |
b8698a0f | 4661 | *preds = XRESIZEVEC (insn_t, *preds, |
e855c69d AB |
4662 | (*size = 2 * *size + 1)); |
4663 | (*preds)[(*n)++] = bb_end; | |
4664 | } | |
4665 | } | |
4666 | ||
3e6a3f6f AB |
4667 | gcc_assert (*n != 0 |
4668 | || (flag_sel_sched_pipelining_outer_loops | |
4669 | && current_loop_nest)); | |
e855c69d AB |
4670 | } |
4671 | ||
b8698a0f L |
4672 | /* Find all predecessors of BB and record them in PREDS and their number |
4673 | in N. Empty blocks are skipped, and only normal (forward in-region) | |
e855c69d AB |
4674 | edges are processed. */ |
4675 | static void | |
4676 | cfg_preds (basic_block bb, insn_t **preds, int *n) | |
4677 | { | |
4678 | int size = 0; | |
4679 | ||
4680 | *preds = NULL; | |
4681 | *n = 0; | |
4682 | cfg_preds_1 (bb, preds, n, &size); | |
4683 | } | |
4684 | ||
4685 | /* Returns true if we are moving INSN through join point. */ | |
4686 | bool | |
4687 | sel_num_cfg_preds_gt_1 (insn_t insn) | |
4688 | { | |
4689 | basic_block bb; | |
4690 | ||
4691 | if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) | |
4692 | return false; | |
4693 | ||
4694 | bb = BLOCK_FOR_INSN (insn); | |
4695 | ||
4696 | while (1) | |
4697 | { | |
4698 | if (EDGE_COUNT (bb->preds) > 1) | |
4699 | return true; | |
4700 | ||
4701 | gcc_assert (EDGE_PRED (bb, 0)->dest == bb); | |
4702 | bb = EDGE_PRED (bb, 0)->src; | |
4703 | ||
4704 | if (!sel_bb_empty_p (bb)) | |
4705 | break; | |
4706 | } | |
4707 | ||
4708 | return false; | |
4709 | } | |
4710 | ||
b8698a0f | 4711 | /* Returns true when BB should be the end of an ebb. Adapted from the |
e855c69d AB |
4712 | code in sched-ebb.c. */ |
4713 | bool | |
4714 | bb_ends_ebb_p (basic_block bb) | |
4715 | { | |
4716 | basic_block next_bb = bb_next_bb (bb); | |
4717 | edge e; | |
b8698a0f | 4718 | |
e855c69d AB |
4719 | if (next_bb == EXIT_BLOCK_PTR |
4720 | || bitmap_bit_p (forced_ebb_heads, next_bb->index) | |
4721 | || (LABEL_P (BB_HEAD (next_bb)) | |
4722 | /* NB: LABEL_NUSES () is not maintained outside of jump.c. | |
4723 | Work around that. */ | |
4724 | && !single_pred_p (next_bb))) | |
4725 | return true; | |
4726 | ||
4727 | if (!in_current_region_p (next_bb)) | |
4728 | return true; | |
4729 | ||
0fd4b31d NF |
4730 | e = find_fallthru_edge (bb->succs); |
4731 | if (e) | |
4732 | { | |
4733 | gcc_assert (e->dest == next_bb); | |
4734 | ||
4735 | return false; | |
4736 | } | |
e855c69d AB |
4737 | |
4738 | return true; | |
4739 | } | |
4740 | ||
4741 | /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a | |
4742 | successor of INSN. */ | |
4743 | bool | |
4744 | in_same_ebb_p (insn_t insn, insn_t succ) | |
4745 | { | |
4746 | basic_block ptr = BLOCK_FOR_INSN (insn); | |
4747 | ||
4748 | for(;;) | |
4749 | { | |
4750 | if (ptr == BLOCK_FOR_INSN (succ)) | |
4751 | return true; | |
b8698a0f | 4752 | |
e855c69d AB |
4753 | if (bb_ends_ebb_p (ptr)) |
4754 | return false; | |
4755 | ||
4756 | ptr = bb_next_bb (ptr); | |
4757 | } | |
4758 | ||
4759 | gcc_unreachable (); | |
4760 | return false; | |
4761 | } | |
4762 | ||
4763 | /* Recomputes the reverse topological order for the function and | |
4764 | saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also | |
4765 | modified appropriately. */ | |
4766 | static void | |
4767 | recompute_rev_top_order (void) | |
4768 | { | |
4769 | int *postorder; | |
4770 | int n_blocks, i; | |
4771 | ||
4772 | if (!rev_top_order_index || rev_top_order_index_len < last_basic_block) | |
4773 | { | |
b8698a0f | 4774 | rev_top_order_index_len = last_basic_block; |
e855c69d AB |
4775 | rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, |
4776 | rev_top_order_index_len); | |
4777 | } | |
4778 | ||
4779 | postorder = XNEWVEC (int, n_basic_blocks); | |
4780 | ||
4781 | n_blocks = post_order_compute (postorder, true, false); | |
4782 | gcc_assert (n_basic_blocks == n_blocks); | |
4783 | ||
4784 | /* Build reverse function: for each basic block with BB->INDEX == K | |
4785 | rev_top_order_index[K] is it's reverse topological sort number. */ | |
4786 | for (i = 0; i < n_blocks; i++) | |
4787 | { | |
4788 | gcc_assert (postorder[i] < rev_top_order_index_len); | |
4789 | rev_top_order_index[postorder[i]] = i; | |
4790 | } | |
4791 | ||
4792 | free (postorder); | |
4793 | } | |
4794 | ||
4795 | /* Clear all flags from insns in BB that could spoil its rescheduling. */ | |
4796 | void | |
4797 | clear_outdated_rtx_info (basic_block bb) | |
4798 | { | |
4799 | rtx insn; | |
4800 | ||
4801 | FOR_BB_INSNS (bb, insn) | |
4802 | if (INSN_P (insn)) | |
4803 | { | |
4804 | SCHED_GROUP_P (insn) = 0; | |
4805 | INSN_AFTER_STALL_P (insn) = 0; | |
4806 | INSN_SCHED_TIMES (insn) = 0; | |
4807 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; | |
4808 | ||
4809 | /* We cannot use the changed caches, as previously we could ignore | |
b8698a0f | 4810 | the LHS dependence due to enabled renaming and transform |
e855c69d AB |
4811 | the expression, and currently we'll be unable to do this. */ |
4812 | htab_empty (INSN_TRANSFORMED_INSNS (insn)); | |
4813 | } | |
4814 | } | |
4815 | ||
4816 | /* Add BB_NOTE to the pool of available basic block notes. */ | |
4817 | static void | |
4818 | return_bb_to_pool (basic_block bb) | |
4819 | { | |
4820 | rtx note = bb_note (bb); | |
4821 | ||
4822 | gcc_assert (NOTE_BASIC_BLOCK (note) == bb | |
4823 | && bb->aux == NULL); | |
4824 | ||
4825 | /* It turns out that current cfg infrastructure does not support | |
4826 | reuse of basic blocks. Don't bother for now. */ | |
4827 | /*VEC_safe_push (rtx, heap, bb_note_pool, note);*/ | |
4828 | } | |
4829 | ||
4830 | /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ | |
4831 | static rtx | |
4832 | get_bb_note_from_pool (void) | |
4833 | { | |
4834 | if (VEC_empty (rtx, bb_note_pool)) | |
4835 | return NULL_RTX; | |
4836 | else | |
4837 | { | |
4838 | rtx note = VEC_pop (rtx, bb_note_pool); | |
4839 | ||
4840 | PREV_INSN (note) = NULL_RTX; | |
4841 | NEXT_INSN (note) = NULL_RTX; | |
4842 | ||
4843 | return note; | |
4844 | } | |
4845 | } | |
4846 | ||
4847 | /* Free bb_note_pool. */ | |
4848 | void | |
4849 | free_bb_note_pool (void) | |
4850 | { | |
4851 | VEC_free (rtx, heap, bb_note_pool); | |
4852 | } | |
4853 | ||
4854 | /* Setup scheduler pool and successor structure. */ | |
4855 | void | |
4856 | alloc_sched_pools (void) | |
4857 | { | |
4858 | int succs_size; | |
4859 | ||
4860 | succs_size = MAX_WS + 1; | |
b8698a0f | 4861 | succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); |
e855c69d AB |
4862 | succs_info_pool.size = succs_size; |
4863 | succs_info_pool.top = -1; | |
4864 | succs_info_pool.max_top = -1; | |
4865 | ||
b8698a0f | 4866 | sched_lists_pool = create_alloc_pool ("sel-sched-lists", |
e855c69d AB |
4867 | sizeof (struct _list_node), 500); |
4868 | } | |
4869 | ||
4870 | /* Free the pools. */ | |
4871 | void | |
4872 | free_sched_pools (void) | |
4873 | { | |
4874 | int i; | |
b8698a0f | 4875 | |
e855c69d AB |
4876 | free_alloc_pool (sched_lists_pool); |
4877 | gcc_assert (succs_info_pool.top == -1); | |
4878 | for (i = 0; i < succs_info_pool.max_top; i++) | |
4879 | { | |
4880 | VEC_free (rtx, heap, succs_info_pool.stack[i].succs_ok); | |
4881 | VEC_free (rtx, heap, succs_info_pool.stack[i].succs_other); | |
4882 | VEC_free (int, heap, succs_info_pool.stack[i].probs_ok); | |
4883 | } | |
4884 | free (succs_info_pool.stack); | |
4885 | } | |
4886 | \f | |
4887 | ||
b8698a0f | 4888 | /* Returns a position in RGN where BB can be inserted retaining |
e855c69d AB |
4889 | topological order. */ |
4890 | static int | |
4891 | find_place_to_insert_bb (basic_block bb, int rgn) | |
4892 | { | |
4893 | bool has_preds_outside_rgn = false; | |
4894 | edge e; | |
4895 | edge_iterator ei; | |
b8698a0f | 4896 | |
e855c69d AB |
4897 | /* Find whether we have preds outside the region. */ |
4898 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4899 | if (!in_current_region_p (e->src)) | |
4900 | { | |
4901 | has_preds_outside_rgn = true; | |
4902 | break; | |
4903 | } | |
b8698a0f | 4904 | |
e855c69d AB |
4905 | /* Recompute the top order -- needed when we have > 1 pred |
4906 | and in case we don't have preds outside. */ | |
4907 | if (flag_sel_sched_pipelining_outer_loops | |
4908 | && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) | |
4909 | { | |
4910 | int i, bbi = bb->index, cur_bbi; | |
4911 | ||
4912 | recompute_rev_top_order (); | |
4913 | for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) | |
4914 | { | |
4915 | cur_bbi = BB_TO_BLOCK (i); | |
b8698a0f | 4916 | if (rev_top_order_index[bbi] |
e855c69d AB |
4917 | < rev_top_order_index[cur_bbi]) |
4918 | break; | |
4919 | } | |
b8698a0f | 4920 | |
e855c69d AB |
4921 | /* We skipped the right block, so we increase i. We accomodate |
4922 | it for increasing by step later, so we decrease i. */ | |
4923 | return (i + 1) - 1; | |
4924 | } | |
4925 | else if (has_preds_outside_rgn) | |
4926 | { | |
4927 | /* This is the case when we generate an extra empty block | |
4928 | to serve as region head during pipelining. */ | |
4929 | e = EDGE_SUCC (bb, 0); | |
4930 | gcc_assert (EDGE_COUNT (bb->succs) == 1 | |
4931 | && in_current_region_p (EDGE_SUCC (bb, 0)->dest) | |
4932 | && (BLOCK_TO_BB (e->dest->index) == 0)); | |
4933 | return -1; | |
4934 | } | |
4935 | ||
4936 | /* We don't have preds outside the region. We should have | |
4937 | the only pred, because the multiple preds case comes from | |
4938 | the pipelining of outer loops, and that is handled above. | |
4939 | Just take the bbi of this single pred. */ | |
4940 | if (EDGE_COUNT (bb->succs) > 0) | |
4941 | { | |
4942 | int pred_bbi; | |
b8698a0f | 4943 | |
e855c69d | 4944 | gcc_assert (EDGE_COUNT (bb->preds) == 1); |
b8698a0f | 4945 | |
e855c69d AB |
4946 | pred_bbi = EDGE_PRED (bb, 0)->src->index; |
4947 | return BLOCK_TO_BB (pred_bbi); | |
4948 | } | |
4949 | else | |
4950 | /* BB has no successors. It is safe to put it in the end. */ | |
4951 | return current_nr_blocks - 1; | |
4952 | } | |
4953 | ||
4954 | /* Deletes an empty basic block freeing its data. */ | |
4955 | static void | |
4956 | delete_and_free_basic_block (basic_block bb) | |
4957 | { | |
4958 | gcc_assert (sel_bb_empty_p (bb)); | |
4959 | ||
4960 | if (BB_LV_SET (bb)) | |
4961 | free_lv_set (bb); | |
4962 | ||
4963 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
4964 | ||
b8698a0f L |
4965 | /* Can't assert av_set properties because we use sel_aremove_bb |
4966 | when removing loop preheader from the region. At the point of | |
e855c69d AB |
4967 | removing the preheader we already have deallocated sel_region_bb_info. */ |
4968 | gcc_assert (BB_LV_SET (bb) == NULL | |
4969 | && !BB_LV_SET_VALID_P (bb) | |
4970 | && BB_AV_LEVEL (bb) == 0 | |
4971 | && BB_AV_SET (bb) == NULL); | |
b8698a0f | 4972 | |
e855c69d AB |
4973 | delete_basic_block (bb); |
4974 | } | |
4975 | ||
4976 | /* Add BB to the current region and update the region data. */ | |
4977 | static void | |
4978 | add_block_to_current_region (basic_block bb) | |
4979 | { | |
4980 | int i, pos, bbi = -2, rgn; | |
4981 | ||
4982 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4983 | bbi = find_place_to_insert_bb (bb, rgn); | |
4984 | bbi += 1; | |
4985 | pos = RGN_BLOCKS (rgn) + bbi; | |
4986 | ||
4987 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
4988 | && ebb_head[bbi] == pos); | |
b8698a0f | 4989 | |
e855c69d AB |
4990 | /* Make a place for the new block. */ |
4991 | extend_regions (); | |
4992 | ||
4993 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
4994 | BLOCK_TO_BB (rgn_bb_table[i])++; | |
b8698a0f | 4995 | |
e855c69d AB |
4996 | memmove (rgn_bb_table + pos + 1, |
4997 | rgn_bb_table + pos, | |
4998 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
4999 | ||
5000 | /* Initialize data for BB. */ | |
5001 | rgn_bb_table[pos] = bb->index; | |
5002 | BLOCK_TO_BB (bb->index) = bbi; | |
5003 | CONTAINING_RGN (bb->index) = rgn; | |
5004 | ||
5005 | RGN_NR_BLOCKS (rgn)++; | |
b8698a0f | 5006 | |
e855c69d AB |
5007 | for (i = rgn + 1; i <= nr_regions; i++) |
5008 | RGN_BLOCKS (i)++; | |
5009 | } | |
5010 | ||
5011 | /* Remove BB from the current region and update the region data. */ | |
5012 | static void | |
5013 | remove_bb_from_region (basic_block bb) | |
5014 | { | |
5015 | int i, pos, bbi = -2, rgn; | |
5016 | ||
5017 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
5018 | bbi = BLOCK_TO_BB (bb->index); | |
5019 | pos = RGN_BLOCKS (rgn) + bbi; | |
5020 | ||
5021 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
5022 | && ebb_head[bbi] == pos); | |
5023 | ||
5024 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
5025 | BLOCK_TO_BB (rgn_bb_table[i])--; | |
5026 | ||
5027 | memmove (rgn_bb_table + pos, | |
5028 | rgn_bb_table + pos + 1, | |
5029 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5030 | ||
5031 | RGN_NR_BLOCKS (rgn)--; | |
5032 | for (i = rgn + 1; i <= nr_regions; i++) | |
5033 | RGN_BLOCKS (i)--; | |
5034 | } | |
5035 | ||
b8698a0f | 5036 | /* Add BB to the current region and update all data. If BB is NULL, add all |
e855c69d AB |
5037 | blocks from last_added_blocks vector. */ |
5038 | static void | |
5039 | sel_add_bb (basic_block bb) | |
5040 | { | |
5041 | /* Extend luids so that new notes will receive zero luids. */ | |
5042 | sched_init_luids (NULL, NULL, NULL, NULL); | |
5043 | sched_init_bbs (); | |
5044 | sel_init_bbs (last_added_blocks, NULL); | |
5045 | ||
b8698a0f | 5046 | /* When bb is passed explicitly, the vector should contain |
e855c69d AB |
5047 | the only element that equals to bb; otherwise, the vector |
5048 | should not be NULL. */ | |
5049 | gcc_assert (last_added_blocks != NULL); | |
b8698a0f | 5050 | |
e855c69d AB |
5051 | if (bb != NULL) |
5052 | { | |
5053 | gcc_assert (VEC_length (basic_block, last_added_blocks) == 1 | |
b8698a0f L |
5054 | && VEC_index (basic_block, |
5055 | last_added_blocks, 0) == bb); | |
e855c69d AB |
5056 | add_block_to_current_region (bb); |
5057 | ||
5058 | /* We associate creating/deleting data sets with the first insn | |
5059 | appearing / disappearing in the bb. */ | |
5060 | if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) | |
5061 | create_initial_data_sets (bb); | |
b8698a0f | 5062 | |
e855c69d AB |
5063 | VEC_free (basic_block, heap, last_added_blocks); |
5064 | } | |
5065 | else | |
5066 | /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ | |
5067 | { | |
5068 | int i; | |
5069 | basic_block temp_bb = NULL; | |
5070 | ||
b8698a0f | 5071 | for (i = 0; |
e855c69d AB |
5072 | VEC_iterate (basic_block, last_added_blocks, i, bb); i++) |
5073 | { | |
5074 | add_block_to_current_region (bb); | |
5075 | temp_bb = bb; | |
5076 | } | |
5077 | ||
b8698a0f | 5078 | /* We need to fetch at least one bb so we know the region |
e855c69d AB |
5079 | to update. */ |
5080 | gcc_assert (temp_bb != NULL); | |
5081 | bb = temp_bb; | |
5082 | ||
5083 | VEC_free (basic_block, heap, last_added_blocks); | |
5084 | } | |
5085 | ||
5086 | rgn_setup_region (CONTAINING_RGN (bb->index)); | |
5087 | } | |
5088 | ||
b8698a0f | 5089 | /* Remove BB from the current region and update all data. |
e855c69d AB |
5090 | If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ |
5091 | static void | |
5092 | sel_remove_bb (basic_block bb, bool remove_from_cfg_p) | |
5093 | { | |
262d8232 AB |
5094 | unsigned idx = bb->index; |
5095 | ||
e855c69d | 5096 | gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); |
b8698a0f | 5097 | |
e855c69d AB |
5098 | remove_bb_from_region (bb); |
5099 | return_bb_to_pool (bb); | |
262d8232 | 5100 | bitmap_clear_bit (blocks_to_reschedule, idx); |
b8698a0f | 5101 | |
e855c69d | 5102 | if (remove_from_cfg_p) |
00c4e97c AB |
5103 | { |
5104 | basic_block succ = single_succ (bb); | |
5105 | delete_and_free_basic_block (bb); | |
5106 | set_immediate_dominator (CDI_DOMINATORS, succ, | |
5107 | recompute_dominator (CDI_DOMINATORS, succ)); | |
5108 | } | |
e855c69d | 5109 | |
262d8232 | 5110 | rgn_setup_region (CONTAINING_RGN (idx)); |
e855c69d AB |
5111 | } |
5112 | ||
5113 | /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ | |
5114 | static void | |
5115 | move_bb_info (basic_block merge_bb, basic_block empty_bb) | |
5116 | { | |
5117 | gcc_assert (in_current_region_p (merge_bb)); | |
5118 | ||
b8698a0f | 5119 | concat_note_lists (BB_NOTE_LIST (empty_bb), |
e855c69d AB |
5120 | &BB_NOTE_LIST (merge_bb)); |
5121 | BB_NOTE_LIST (empty_bb) = NULL_RTX; | |
5122 | ||
5123 | } | |
5124 | ||
e855c69d AB |
5125 | /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from |
5126 | region, but keep it in CFG. */ | |
5127 | static void | |
5128 | remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) | |
5129 | { | |
5130 | /* The block should contain just a note or a label. | |
5131 | We try to check whether it is unused below. */ | |
5132 | gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) | |
5133 | || LABEL_P (BB_HEAD (empty_bb))); | |
5134 | ||
5135 | /* If basic block has predecessors or successors, redirect them. */ | |
5136 | if (remove_from_cfg_p | |
5137 | && (EDGE_COUNT (empty_bb->preds) > 0 | |
5138 | || EDGE_COUNT (empty_bb->succs) > 0)) | |
5139 | { | |
5140 | basic_block pred; | |
5141 | basic_block succ; | |
5142 | ||
5143 | /* We need to init PRED and SUCC before redirecting edges. */ | |
5144 | if (EDGE_COUNT (empty_bb->preds) > 0) | |
5145 | { | |
5146 | edge e; | |
5147 | ||
5148 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); | |
5149 | ||
5150 | e = EDGE_PRED (empty_bb, 0); | |
5151 | gcc_assert (e->src == empty_bb->prev_bb | |
5152 | && (e->flags & EDGE_FALLTHRU)); | |
5153 | ||
5154 | pred = empty_bb->prev_bb; | |
5155 | } | |
5156 | else | |
5157 | pred = NULL; | |
5158 | ||
5159 | if (EDGE_COUNT (empty_bb->succs) > 0) | |
5160 | { | |
5161 | /* We do not check fallthruness here as above, because | |
5162 | after removing a jump the edge may actually be not fallthru. */ | |
5163 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); | |
5164 | succ = EDGE_SUCC (empty_bb, 0)->dest; | |
5165 | } | |
5166 | else | |
5167 | succ = NULL; | |
5168 | ||
5169 | if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) | |
5170 | { | |
5171 | edge e = EDGE_PRED (empty_bb, 0); | |
5172 | ||
5173 | if (e->flags & EDGE_FALLTHRU) | |
5174 | redirect_edge_succ_nodup (e, succ); | |
5175 | else | |
5176 | sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); | |
5177 | } | |
5178 | ||
5179 | if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) | |
5180 | { | |
5181 | edge e = EDGE_SUCC (empty_bb, 0); | |
5182 | ||
5183 | if (find_edge (pred, e->dest) == NULL) | |
5184 | redirect_edge_pred (e, pred); | |
5185 | } | |
5186 | } | |
5187 | ||
5188 | /* Finish removing. */ | |
5189 | sel_remove_bb (empty_bb, remove_from_cfg_p); | |
5190 | } | |
5191 | ||
b8698a0f | 5192 | /* An implementation of create_basic_block hook, which additionally updates |
e855c69d AB |
5193 | per-bb data structures. */ |
5194 | static basic_block | |
5195 | sel_create_basic_block (void *headp, void *endp, basic_block after) | |
5196 | { | |
5197 | basic_block new_bb; | |
5198 | insn_t new_bb_note; | |
b8698a0f L |
5199 | |
5200 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
e855c69d AB |
5201 | || last_added_blocks == NULL); |
5202 | ||
5203 | new_bb_note = get_bb_note_from_pool (); | |
5204 | ||
5205 | if (new_bb_note == NULL_RTX) | |
5206 | new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); | |
5207 | else | |
5208 | { | |
5209 | new_bb = create_basic_block_structure ((rtx) headp, (rtx) endp, | |
5210 | new_bb_note, after); | |
5211 | new_bb->aux = NULL; | |
5212 | } | |
5213 | ||
5214 | VEC_safe_push (basic_block, heap, last_added_blocks, new_bb); | |
5215 | ||
5216 | return new_bb; | |
5217 | } | |
5218 | ||
5219 | /* Implement sched_init_only_bb (). */ | |
5220 | static void | |
5221 | sel_init_only_bb (basic_block bb, basic_block after) | |
5222 | { | |
5223 | gcc_assert (after == NULL); | |
5224 | ||
5225 | extend_regions (); | |
5226 | rgn_make_new_region_out_of_new_block (bb); | |
5227 | } | |
5228 | ||
5229 | /* Update the latch when we've splitted or merged it from FROM block to TO. | |
5230 | This should be checked for all outer loops, too. */ | |
5231 | static void | |
5232 | change_loops_latches (basic_block from, basic_block to) | |
5233 | { | |
5234 | gcc_assert (from != to); | |
5235 | ||
5236 | if (current_loop_nest) | |
5237 | { | |
5238 | struct loop *loop; | |
5239 | ||
5240 | for (loop = current_loop_nest; loop; loop = loop_outer (loop)) | |
5241 | if (considered_for_pipelining_p (loop) && loop->latch == from) | |
5242 | { | |
5243 | gcc_assert (loop == current_loop_nest); | |
5244 | loop->latch = to; | |
5245 | gcc_assert (loop_latch_edge (loop)); | |
5246 | } | |
5247 | } | |
5248 | } | |
5249 | ||
b8698a0f | 5250 | /* Splits BB on two basic blocks, adding it to the region and extending |
e855c69d AB |
5251 | per-bb data structures. Returns the newly created bb. */ |
5252 | static basic_block | |
5253 | sel_split_block (basic_block bb, rtx after) | |
5254 | { | |
5255 | basic_block new_bb; | |
5256 | insn_t insn; | |
5257 | ||
5258 | new_bb = sched_split_block_1 (bb, after); | |
5259 | sel_add_bb (new_bb); | |
5260 | ||
5261 | /* This should be called after sel_add_bb, because this uses | |
b8698a0f | 5262 | CONTAINING_RGN for the new block, which is not yet initialized. |
e855c69d AB |
5263 | FIXME: this function may be a no-op now. */ |
5264 | change_loops_latches (bb, new_bb); | |
5265 | ||
5266 | /* Update ORIG_BB_INDEX for insns moved into the new block. */ | |
5267 | FOR_BB_INSNS (new_bb, insn) | |
5268 | if (INSN_P (insn)) | |
5269 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; | |
5270 | ||
5271 | if (sel_bb_empty_p (bb)) | |
5272 | { | |
5273 | gcc_assert (!sel_bb_empty_p (new_bb)); | |
5274 | ||
5275 | /* NEW_BB has data sets that need to be updated and BB holds | |
5276 | data sets that should be removed. Exchange these data sets | |
5277 | so that we won't lose BB's valid data sets. */ | |
5278 | exchange_data_sets (new_bb, bb); | |
5279 | free_data_sets (bb); | |
5280 | } | |
5281 | ||
5282 | if (!sel_bb_empty_p (new_bb) | |
5283 | && bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
5284 | bitmap_set_bit (blocks_to_reschedule, new_bb->index); | |
5285 | ||
5286 | return new_bb; | |
5287 | } | |
5288 | ||
5289 | /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. | |
5290 | Otherwise returns NULL. */ | |
5291 | static rtx | |
5292 | check_for_new_jump (basic_block bb, int prev_max_uid) | |
5293 | { | |
5294 | rtx end; | |
5295 | ||
5296 | end = sel_bb_end (bb); | |
5297 | if (end && INSN_UID (end) >= prev_max_uid) | |
5298 | return end; | |
5299 | return NULL; | |
5300 | } | |
5301 | ||
b8698a0f | 5302 | /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. |
e855c69d AB |
5303 | New means having UID at least equal to PREV_MAX_UID. */ |
5304 | static rtx | |
5305 | find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) | |
5306 | { | |
5307 | rtx jump; | |
5308 | ||
5309 | /* Return immediately if no new insns were emitted. */ | |
5310 | if (get_max_uid () == prev_max_uid) | |
5311 | return NULL; | |
b8698a0f | 5312 | |
e855c69d AB |
5313 | /* Now check both blocks for new jumps. It will ever be only one. */ |
5314 | if ((jump = check_for_new_jump (from, prev_max_uid))) | |
5315 | return jump; | |
5316 | ||
5317 | if (jump_bb != NULL | |
5318 | && (jump = check_for_new_jump (jump_bb, prev_max_uid))) | |
5319 | return jump; | |
5320 | return NULL; | |
5321 | } | |
5322 | ||
5323 | /* Splits E and adds the newly created basic block to the current region. | |
5324 | Returns this basic block. */ | |
5325 | basic_block | |
5326 | sel_split_edge (edge e) | |
5327 | { | |
5328 | basic_block new_bb, src, other_bb = NULL; | |
5329 | int prev_max_uid; | |
5330 | rtx jump; | |
5331 | ||
5332 | src = e->src; | |
5333 | prev_max_uid = get_max_uid (); | |
5334 | new_bb = split_edge (e); | |
5335 | ||
b8698a0f | 5336 | if (flag_sel_sched_pipelining_outer_loops |
e855c69d AB |
5337 | && current_loop_nest) |
5338 | { | |
5339 | int i; | |
5340 | basic_block bb; | |
5341 | ||
b8698a0f | 5342 | /* Some of the basic blocks might not have been added to the loop. |
e855c69d | 5343 | Add them here, until this is fixed in force_fallthru. */ |
b8698a0f | 5344 | for (i = 0; |
e855c69d AB |
5345 | VEC_iterate (basic_block, last_added_blocks, i, bb); i++) |
5346 | if (!bb->loop_father) | |
5347 | { | |
5348 | add_bb_to_loop (bb, e->dest->loop_father); | |
5349 | ||
5350 | gcc_assert (!other_bb && (new_bb->index != bb->index)); | |
5351 | other_bb = bb; | |
5352 | } | |
5353 | } | |
5354 | ||
5355 | /* Add all last_added_blocks to the region. */ | |
5356 | sel_add_bb (NULL); | |
5357 | ||
5358 | jump = find_new_jump (src, new_bb, prev_max_uid); | |
5359 | if (jump) | |
5360 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5361 | ||
5362 | /* Put the correct lv set on this block. */ | |
5363 | if (other_bb && !sel_bb_empty_p (other_bb)) | |
5364 | compute_live (sel_bb_head (other_bb)); | |
5365 | ||
5366 | return new_bb; | |
5367 | } | |
5368 | ||
5369 | /* Implement sched_create_empty_bb (). */ | |
5370 | static basic_block | |
5371 | sel_create_empty_bb (basic_block after) | |
5372 | { | |
5373 | basic_block new_bb; | |
5374 | ||
5375 | new_bb = sched_create_empty_bb_1 (after); | |
5376 | ||
5377 | /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit | |
5378 | later. */ | |
5379 | gcc_assert (VEC_length (basic_block, last_added_blocks) == 1 | |
5380 | && VEC_index (basic_block, last_added_blocks, 0) == new_bb); | |
5381 | ||
5382 | VEC_free (basic_block, heap, last_added_blocks); | |
5383 | return new_bb; | |
5384 | } | |
5385 | ||
5386 | /* Implement sched_create_recovery_block. ORIG_INSN is where block | |
5387 | will be splitted to insert a check. */ | |
5388 | basic_block | |
5389 | sel_create_recovery_block (insn_t orig_insn) | |
5390 | { | |
5391 | basic_block first_bb, second_bb, recovery_block; | |
5392 | basic_block before_recovery = NULL; | |
5393 | rtx jump; | |
5394 | ||
5395 | first_bb = BLOCK_FOR_INSN (orig_insn); | |
5396 | if (sel_bb_end_p (orig_insn)) | |
5397 | { | |
5398 | /* Avoid introducing an empty block while splitting. */ | |
5399 | gcc_assert (single_succ_p (first_bb)); | |
5400 | second_bb = single_succ (first_bb); | |
5401 | } | |
5402 | else | |
5403 | second_bb = sched_split_block (first_bb, orig_insn); | |
5404 | ||
5405 | recovery_block = sched_create_recovery_block (&before_recovery); | |
5406 | if (before_recovery) | |
5407 | copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR); | |
5408 | ||
5409 | gcc_assert (sel_bb_empty_p (recovery_block)); | |
5410 | sched_create_recovery_edges (first_bb, recovery_block, second_bb); | |
5411 | if (current_loops != NULL) | |
5412 | add_bb_to_loop (recovery_block, first_bb->loop_father); | |
b8698a0f | 5413 | |
e855c69d | 5414 | sel_add_bb (recovery_block); |
b8698a0f | 5415 | |
e855c69d AB |
5416 | jump = BB_END (recovery_block); |
5417 | gcc_assert (sel_bb_head (recovery_block) == jump); | |
5418 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5419 | ||
5420 | return recovery_block; | |
5421 | } | |
5422 | ||
5423 | /* Merge basic block B into basic block A. */ | |
262d8232 | 5424 | static void |
e855c69d AB |
5425 | sel_merge_blocks (basic_block a, basic_block b) |
5426 | { | |
262d8232 AB |
5427 | gcc_assert (sel_bb_empty_p (b) |
5428 | && EDGE_COUNT (b->preds) == 1 | |
5429 | && EDGE_PRED (b, 0)->src == b->prev_bb); | |
e855c69d | 5430 | |
262d8232 AB |
5431 | move_bb_info (b->prev_bb, b); |
5432 | remove_empty_bb (b, false); | |
5433 | merge_blocks (a, b); | |
e855c69d AB |
5434 | change_loops_latches (b, a); |
5435 | } | |
5436 | ||
5437 | /* A wrapper for redirect_edge_and_branch_force, which also initializes | |
5438 | data structures for possibly created bb and insns. Returns the newly | |
5439 | added bb or NULL, when a bb was not needed. */ | |
5440 | void | |
5441 | sel_redirect_edge_and_branch_force (edge e, basic_block to) | |
5442 | { | |
00c4e97c | 5443 | basic_block jump_bb, src, orig_dest = e->dest; |
e855c69d AB |
5444 | int prev_max_uid; |
5445 | rtx jump; | |
b8698a0f | 5446 | |
00c4e97c AB |
5447 | /* This function is now used only for bookkeeping code creation, where |
5448 | we'll never get the single pred of orig_dest block and thus will not | |
5449 | hit unreachable blocks when updating dominator info. */ | |
5450 | gcc_assert (!sel_bb_empty_p (e->src) | |
5451 | && !single_pred_p (orig_dest)); | |
e855c69d AB |
5452 | src = e->src; |
5453 | prev_max_uid = get_max_uid (); | |
5454 | jump_bb = redirect_edge_and_branch_force (e, to); | |
5455 | ||
5456 | if (jump_bb != NULL) | |
5457 | sel_add_bb (jump_bb); | |
5458 | ||
5459 | /* This function could not be used to spoil the loop structure by now, | |
5460 | thus we don't care to update anything. But check it to be sure. */ | |
5461 | if (current_loop_nest | |
5462 | && pipelining_p) | |
5463 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
b8698a0f | 5464 | |
e855c69d AB |
5465 | jump = find_new_jump (src, jump_bb, prev_max_uid); |
5466 | if (jump) | |
5467 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
00c4e97c AB |
5468 | set_immediate_dominator (CDI_DOMINATORS, to, |
5469 | recompute_dominator (CDI_DOMINATORS, to)); | |
5470 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5471 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
e855c69d AB |
5472 | } |
5473 | ||
b59ab570 AM |
5474 | /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by |
5475 | redirected edge are in reverse topological order. */ | |
5476 | bool | |
e855c69d AB |
5477 | sel_redirect_edge_and_branch (edge e, basic_block to) |
5478 | { | |
5479 | bool latch_edge_p; | |
00c4e97c | 5480 | basic_block src, orig_dest = e->dest; |
e855c69d AB |
5481 | int prev_max_uid; |
5482 | rtx jump; | |
f2c45f08 | 5483 | edge redirected; |
b59ab570 | 5484 | bool recompute_toporder_p = false; |
00c4e97c | 5485 | bool maybe_unreachable = single_pred_p (orig_dest); |
e855c69d AB |
5486 | |
5487 | latch_edge_p = (pipelining_p | |
5488 | && current_loop_nest | |
5489 | && e == loop_latch_edge (current_loop_nest)); | |
5490 | ||
5491 | src = e->src; | |
5492 | prev_max_uid = get_max_uid (); | |
f2c45f08 AM |
5493 | |
5494 | redirected = redirect_edge_and_branch (e, to); | |
5495 | ||
5496 | gcc_assert (redirected && last_added_blocks == NULL); | |
e855c69d AB |
5497 | |
5498 | /* When we've redirected a latch edge, update the header. */ | |
5499 | if (latch_edge_p) | |
5500 | { | |
5501 | current_loop_nest->header = to; | |
5502 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
5503 | } | |
5504 | ||
b59ab570 AM |
5505 | /* In rare situations, the topological relation between the blocks connected |
5506 | by the redirected edge can change (see PR42245 for an example). Update | |
5507 | block_to_bb/bb_to_block. */ | |
5508 | if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) | |
5509 | && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) | |
5510 | recompute_toporder_p = true; | |
5511 | ||
e855c69d AB |
5512 | jump = find_new_jump (src, NULL, prev_max_uid); |
5513 | if (jump) | |
5514 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
b59ab570 | 5515 | |
00c4e97c AB |
5516 | /* Only update dominator info when we don't have unreachable blocks. |
5517 | Otherwise we'll update in maybe_tidy_empty_bb. */ | |
5518 | if (!maybe_unreachable) | |
5519 | { | |
5520 | set_immediate_dominator (CDI_DOMINATORS, to, | |
5521 | recompute_dominator (CDI_DOMINATORS, to)); | |
5522 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5523 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
5524 | } | |
b59ab570 | 5525 | return recompute_toporder_p; |
e855c69d AB |
5526 | } |
5527 | ||
5528 | /* This variable holds the cfg hooks used by the selective scheduler. */ | |
5529 | static struct cfg_hooks sel_cfg_hooks; | |
5530 | ||
5531 | /* Register sel-sched cfg hooks. */ | |
5532 | void | |
5533 | sel_register_cfg_hooks (void) | |
5534 | { | |
5535 | sched_split_block = sel_split_block; | |
5536 | ||
5537 | orig_cfg_hooks = get_cfg_hooks (); | |
5538 | sel_cfg_hooks = orig_cfg_hooks; | |
5539 | ||
5540 | sel_cfg_hooks.create_basic_block = sel_create_basic_block; | |
5541 | ||
5542 | set_cfg_hooks (sel_cfg_hooks); | |
5543 | ||
5544 | sched_init_only_bb = sel_init_only_bb; | |
5545 | sched_split_block = sel_split_block; | |
5546 | sched_create_empty_bb = sel_create_empty_bb; | |
5547 | } | |
5548 | ||
5549 | /* Unregister sel-sched cfg hooks. */ | |
5550 | void | |
5551 | sel_unregister_cfg_hooks (void) | |
5552 | { | |
5553 | sched_create_empty_bb = NULL; | |
5554 | sched_split_block = NULL; | |
5555 | sched_init_only_bb = NULL; | |
5556 | ||
5557 | set_cfg_hooks (orig_cfg_hooks); | |
5558 | } | |
5559 | \f | |
5560 | ||
5561 | /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, | |
5562 | LABEL is where this jump should be directed. */ | |
5563 | rtx | |
5564 | create_insn_rtx_from_pattern (rtx pattern, rtx label) | |
5565 | { | |
5566 | rtx insn_rtx; | |
5567 | ||
5568 | gcc_assert (!INSN_P (pattern)); | |
5569 | ||
5570 | start_sequence (); | |
5571 | ||
5572 | if (label == NULL_RTX) | |
5573 | insn_rtx = emit_insn (pattern); | |
b5b8b0ac AO |
5574 | else if (DEBUG_INSN_P (label)) |
5575 | insn_rtx = emit_debug_insn (pattern); | |
e855c69d AB |
5576 | else |
5577 | { | |
5578 | insn_rtx = emit_jump_insn (pattern); | |
5579 | JUMP_LABEL (insn_rtx) = label; | |
5580 | ++LABEL_NUSES (label); | |
5581 | } | |
5582 | ||
5583 | end_sequence (); | |
5584 | ||
5585 | sched_init_luids (NULL, NULL, NULL, NULL); | |
5586 | sched_extend_target (); | |
5587 | sched_deps_init (false); | |
5588 | ||
5589 | /* Initialize INSN_CODE now. */ | |
5590 | recog_memoized (insn_rtx); | |
5591 | return insn_rtx; | |
5592 | } | |
5593 | ||
5594 | /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn | |
5595 | must not be clonable. */ | |
5596 | vinsn_t | |
5597 | create_vinsn_from_insn_rtx (rtx insn_rtx, bool force_unique_p) | |
5598 | { | |
5599 | gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); | |
5600 | ||
5601 | /* If VINSN_TYPE is not USE, retain its uniqueness. */ | |
5602 | return vinsn_create (insn_rtx, force_unique_p); | |
5603 | } | |
5604 | ||
5605 | /* Create a copy of INSN_RTX. */ | |
5606 | rtx | |
5607 | create_copy_of_insn_rtx (rtx insn_rtx) | |
5608 | { | |
5609 | rtx res; | |
5610 | ||
b5b8b0ac AO |
5611 | if (DEBUG_INSN_P (insn_rtx)) |
5612 | return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5613 | insn_rtx); | |
5614 | ||
e855c69d AB |
5615 | gcc_assert (NONJUMP_INSN_P (insn_rtx)); |
5616 | ||
5617 | res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5618 | NULL_RTX); | |
5619 | return res; | |
5620 | } | |
5621 | ||
5622 | /* Change vinsn field of EXPR to hold NEW_VINSN. */ | |
5623 | void | |
5624 | change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) | |
5625 | { | |
5626 | vinsn_detach (EXPR_VINSN (expr)); | |
5627 | ||
5628 | EXPR_VINSN (expr) = new_vinsn; | |
5629 | vinsn_attach (new_vinsn); | |
5630 | } | |
5631 | ||
5632 | /* Helpers for global init. */ | |
5633 | /* This structure is used to be able to call existing bundling mechanism | |
5634 | and calculate insn priorities. */ | |
b8698a0f | 5635 | static struct haifa_sched_info sched_sel_haifa_sched_info = |
e855c69d AB |
5636 | { |
5637 | NULL, /* init_ready_list */ | |
5638 | NULL, /* can_schedule_ready_p */ | |
5639 | NULL, /* schedule_more_p */ | |
5640 | NULL, /* new_ready */ | |
5641 | NULL, /* rgn_rank */ | |
5642 | sel_print_insn, /* rgn_print_insn */ | |
5643 | contributes_to_priority, | |
356c23b3 | 5644 | NULL, /* insn_finishes_block_p */ |
e855c69d AB |
5645 | |
5646 | NULL, NULL, | |
5647 | NULL, NULL, | |
5648 | 0, 0, | |
5649 | ||
5650 | NULL, /* add_remove_insn */ | |
5651 | NULL, /* begin_schedule_ready */ | |
5652 | NULL, /* advance_target_bb */ | |
5653 | SEL_SCHED | NEW_BBS | |
5654 | }; | |
5655 | ||
5656 | /* Setup special insns used in the scheduler. */ | |
b8698a0f | 5657 | void |
e855c69d AB |
5658 | setup_nop_and_exit_insns (void) |
5659 | { | |
5660 | gcc_assert (nop_pattern == NULL_RTX | |
5661 | && exit_insn == NULL_RTX); | |
5662 | ||
9ef1bf71 | 5663 | nop_pattern = constm1_rtx; |
e855c69d AB |
5664 | |
5665 | start_sequence (); | |
5666 | emit_insn (nop_pattern); | |
5667 | exit_insn = get_insns (); | |
5668 | end_sequence (); | |
5669 | set_block_for_insn (exit_insn, EXIT_BLOCK_PTR); | |
5670 | } | |
5671 | ||
5672 | /* Free special insns used in the scheduler. */ | |
5673 | void | |
5674 | free_nop_and_exit_insns (void) | |
5675 | { | |
5676 | exit_insn = NULL_RTX; | |
5677 | nop_pattern = NULL_RTX; | |
5678 | } | |
5679 | ||
5680 | /* Setup a special vinsn used in new insns initialization. */ | |
5681 | void | |
5682 | setup_nop_vinsn (void) | |
5683 | { | |
5684 | nop_vinsn = vinsn_create (exit_insn, false); | |
5685 | vinsn_attach (nop_vinsn); | |
5686 | } | |
5687 | ||
5688 | /* Free a special vinsn used in new insns initialization. */ | |
5689 | void | |
5690 | free_nop_vinsn (void) | |
5691 | { | |
5692 | gcc_assert (VINSN_COUNT (nop_vinsn) == 1); | |
5693 | vinsn_detach (nop_vinsn); | |
5694 | nop_vinsn = NULL; | |
5695 | } | |
5696 | ||
5697 | /* Call a set_sched_flags hook. */ | |
5698 | void | |
5699 | sel_set_sched_flags (void) | |
5700 | { | |
b8698a0f | 5701 | /* ??? This means that set_sched_flags were called, and we decided to |
e855c69d | 5702 | support speculation. However, set_sched_flags also modifies flags |
b8698a0f | 5703 | on current_sched_info, doing this only at global init. And we |
e855c69d AB |
5704 | sometimes change c_s_i later. So put the correct flags again. */ |
5705 | if (spec_info && targetm.sched.set_sched_flags) | |
5706 | targetm.sched.set_sched_flags (spec_info); | |
5707 | } | |
5708 | ||
5709 | /* Setup pointers to global sched info structures. */ | |
5710 | void | |
5711 | sel_setup_sched_infos (void) | |
5712 | { | |
5713 | rgn_setup_common_sched_info (); | |
5714 | ||
5715 | memcpy (&sel_common_sched_info, common_sched_info, | |
5716 | sizeof (sel_common_sched_info)); | |
5717 | ||
5718 | sel_common_sched_info.fix_recovery_cfg = NULL; | |
5719 | sel_common_sched_info.add_block = NULL; | |
5720 | sel_common_sched_info.estimate_number_of_insns | |
5721 | = sel_estimate_number_of_insns; | |
5722 | sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; | |
5723 | sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; | |
5724 | ||
5725 | common_sched_info = &sel_common_sched_info; | |
5726 | ||
5727 | current_sched_info = &sched_sel_haifa_sched_info; | |
b8698a0f | 5728 | current_sched_info->sched_max_insns_priority = |
e855c69d | 5729 | get_rgn_sched_max_insns_priority (); |
b8698a0f | 5730 | |
e855c69d AB |
5731 | sel_set_sched_flags (); |
5732 | } | |
5733 | \f | |
5734 | ||
5735 | /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, | |
5736 | *BB_ORD_INDEX after that is increased. */ | |
5737 | static void | |
5738 | sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) | |
5739 | { | |
5740 | RGN_NR_BLOCKS (rgn) += 1; | |
5741 | RGN_DONT_CALC_DEPS (rgn) = 0; | |
5742 | RGN_HAS_REAL_EBB (rgn) = 0; | |
5743 | CONTAINING_RGN (bb->index) = rgn; | |
5744 | BLOCK_TO_BB (bb->index) = *bb_ord_index; | |
5745 | rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; | |
5746 | (*bb_ord_index)++; | |
5747 | ||
5748 | /* FIXME: it is true only when not scheduling ebbs. */ | |
5749 | RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); | |
5750 | } | |
5751 | ||
5752 | /* Functions to support pipelining of outer loops. */ | |
5753 | ||
5754 | /* Creates a new empty region and returns it's number. */ | |
5755 | static int | |
5756 | sel_create_new_region (void) | |
5757 | { | |
5758 | int new_rgn_number = nr_regions; | |
5759 | ||
5760 | RGN_NR_BLOCKS (new_rgn_number) = 0; | |
5761 | ||
5762 | /* FIXME: This will work only when EBBs are not created. */ | |
5763 | if (new_rgn_number != 0) | |
b8698a0f | 5764 | RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + |
e855c69d AB |
5765 | RGN_NR_BLOCKS (new_rgn_number - 1); |
5766 | else | |
5767 | RGN_BLOCKS (new_rgn_number) = 0; | |
5768 | ||
5769 | /* Set the blocks of the next region so the other functions may | |
5770 | calculate the number of blocks in the region. */ | |
b8698a0f | 5771 | RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + |
e855c69d AB |
5772 | RGN_NR_BLOCKS (new_rgn_number); |
5773 | ||
5774 | nr_regions++; | |
5775 | ||
5776 | return new_rgn_number; | |
5777 | } | |
5778 | ||
5779 | /* If X has a smaller topological sort number than Y, returns -1; | |
5780 | if greater, returns 1. */ | |
5781 | static int | |
5782 | bb_top_order_comparator (const void *x, const void *y) | |
5783 | { | |
5784 | basic_block bb1 = *(const basic_block *) x; | |
5785 | basic_block bb2 = *(const basic_block *) y; | |
5786 | ||
b8698a0f L |
5787 | gcc_assert (bb1 == bb2 |
5788 | || rev_top_order_index[bb1->index] | |
e855c69d AB |
5789 | != rev_top_order_index[bb2->index]); |
5790 | ||
5791 | /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so | |
5792 | bbs with greater number should go earlier. */ | |
5793 | if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) | |
5794 | return -1; | |
5795 | else | |
5796 | return 1; | |
5797 | } | |
5798 | ||
b8698a0f | 5799 | /* Create a region for LOOP and return its number. If we don't want |
e855c69d AB |
5800 | to pipeline LOOP, return -1. */ |
5801 | static int | |
5802 | make_region_from_loop (struct loop *loop) | |
5803 | { | |
5804 | unsigned int i; | |
5805 | int new_rgn_number = -1; | |
5806 | struct loop *inner; | |
5807 | ||
5808 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5809 | int bb_ord_index = 0; | |
5810 | basic_block *loop_blocks; | |
5811 | basic_block preheader_block; | |
5812 | ||
b8698a0f | 5813 | if (loop->num_nodes |
e855c69d AB |
5814 | > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS)) |
5815 | return -1; | |
b8698a0f | 5816 | |
e855c69d AB |
5817 | /* Don't pipeline loops whose latch belongs to some of its inner loops. */ |
5818 | for (inner = loop->inner; inner; inner = inner->inner) | |
5819 | if (flow_bb_inside_loop_p (inner, loop->latch)) | |
5820 | return -1; | |
5821 | ||
5822 | loop->ninsns = num_loop_insns (loop); | |
5823 | if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS)) | |
5824 | return -1; | |
5825 | ||
5826 | loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); | |
5827 | ||
5828 | for (i = 0; i < loop->num_nodes; i++) | |
5829 | if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) | |
5830 | { | |
5831 | free (loop_blocks); | |
5832 | return -1; | |
5833 | } | |
5834 | ||
5835 | preheader_block = loop_preheader_edge (loop)->src; | |
5836 | gcc_assert (preheader_block); | |
5837 | gcc_assert (loop_blocks[0] == loop->header); | |
5838 | ||
5839 | new_rgn_number = sel_create_new_region (); | |
5840 | ||
5841 | sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number); | |
5842 | SET_BIT (bbs_in_loop_rgns, preheader_block->index); | |
5843 | ||
5844 | for (i = 0; i < loop->num_nodes; i++) | |
5845 | { | |
5846 | /* Add only those blocks that haven't been scheduled in the inner loop. | |
5847 | The exception is the basic blocks with bookkeeping code - they should | |
b8698a0f | 5848 | be added to the region (and they actually don't belong to the loop |
e855c69d AB |
5849 | body, but to the region containing that loop body). */ |
5850 | ||
5851 | gcc_assert (new_rgn_number >= 0); | |
5852 | ||
5853 | if (! TEST_BIT (bbs_in_loop_rgns, loop_blocks[i]->index)) | |
5854 | { | |
b8698a0f | 5855 | sel_add_block_to_region (loop_blocks[i], &bb_ord_index, |
e855c69d AB |
5856 | new_rgn_number); |
5857 | SET_BIT (bbs_in_loop_rgns, loop_blocks[i]->index); | |
5858 | } | |
5859 | } | |
5860 | ||
5861 | free (loop_blocks); | |
5862 | MARK_LOOP_FOR_PIPELINING (loop); | |
5863 | ||
5864 | return new_rgn_number; | |
5865 | } | |
5866 | ||
5867 | /* Create a new region from preheader blocks LOOP_BLOCKS. */ | |
5868 | void | |
5869 | make_region_from_loop_preheader (VEC(basic_block, heap) **loop_blocks) | |
5870 | { | |
5871 | unsigned int i; | |
5872 | int new_rgn_number = -1; | |
5873 | basic_block bb; | |
5874 | ||
5875 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5876 | int bb_ord_index = 0; | |
5877 | ||
5878 | new_rgn_number = sel_create_new_region (); | |
5879 | ||
ac47786e | 5880 | FOR_EACH_VEC_ELT (basic_block, *loop_blocks, i, bb) |
e855c69d AB |
5881 | { |
5882 | gcc_assert (new_rgn_number >= 0); | |
5883 | ||
5884 | sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number); | |
5885 | } | |
5886 | ||
5887 | VEC_free (basic_block, heap, *loop_blocks); | |
5888 | gcc_assert (*loop_blocks == NULL); | |
5889 | } | |
5890 | ||
5891 | ||
5892 | /* Create region(s) from loop nest LOOP, such that inner loops will be | |
b8698a0f | 5893 | pipelined before outer loops. Returns true when a region for LOOP |
e855c69d AB |
5894 | is created. */ |
5895 | static bool | |
5896 | make_regions_from_loop_nest (struct loop *loop) | |
b8698a0f | 5897 | { |
e855c69d AB |
5898 | struct loop *cur_loop; |
5899 | int rgn_number; | |
5900 | ||
5901 | /* Traverse all inner nodes of the loop. */ | |
5902 | for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) | |
5903 | if (! TEST_BIT (bbs_in_loop_rgns, cur_loop->header->index)) | |
5904 | return false; | |
5905 | ||
5906 | /* At this moment all regular inner loops should have been pipelined. | |
5907 | Try to create a region from this loop. */ | |
5908 | rgn_number = make_region_from_loop (loop); | |
5909 | ||
5910 | if (rgn_number < 0) | |
5911 | return false; | |
5912 | ||
5913 | VEC_safe_push (loop_p, heap, loop_nests, loop); | |
5914 | return true; | |
5915 | } | |
5916 | ||
5917 | /* Initalize data structures needed. */ | |
5918 | void | |
5919 | sel_init_pipelining (void) | |
5920 | { | |
5921 | /* Collect loop information to be used in outer loops pipelining. */ | |
5922 | loop_optimizer_init (LOOPS_HAVE_PREHEADERS | |
5923 | | LOOPS_HAVE_FALLTHRU_PREHEADERS | |
5924 | | LOOPS_HAVE_RECORDED_EXITS | |
5925 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); | |
5926 | current_loop_nest = NULL; | |
5927 | ||
5928 | bbs_in_loop_rgns = sbitmap_alloc (last_basic_block); | |
5929 | sbitmap_zero (bbs_in_loop_rgns); | |
5930 | ||
5931 | recompute_rev_top_order (); | |
5932 | } | |
5933 | ||
5934 | /* Returns a struct loop for region RGN. */ | |
5935 | loop_p | |
5936 | get_loop_nest_for_rgn (unsigned int rgn) | |
5937 | { | |
5938 | /* Regions created with extend_rgns don't have corresponding loop nests, | |
5939 | because they don't represent loops. */ | |
5940 | if (rgn < VEC_length (loop_p, loop_nests)) | |
5941 | return VEC_index (loop_p, loop_nests, rgn); | |
5942 | else | |
5943 | return NULL; | |
5944 | } | |
5945 | ||
5946 | /* True when LOOP was included into pipelining regions. */ | |
5947 | bool | |
5948 | considered_for_pipelining_p (struct loop *loop) | |
5949 | { | |
5950 | if (loop_depth (loop) == 0) | |
5951 | return false; | |
5952 | ||
b8698a0f L |
5953 | /* Now, the loop could be too large or irreducible. Check whether its |
5954 | region is in LOOP_NESTS. | |
5955 | We determine the region number of LOOP as the region number of its | |
5956 | latch. We can't use header here, because this header could be | |
e855c69d AB |
5957 | just removed preheader and it will give us the wrong region number. |
5958 | Latch can't be used because it could be in the inner loop too. */ | |
8ec4d0ad | 5959 | if (LOOP_MARKED_FOR_PIPELINING_P (loop)) |
e855c69d AB |
5960 | { |
5961 | int rgn = CONTAINING_RGN (loop->latch->index); | |
5962 | ||
5963 | gcc_assert ((unsigned) rgn < VEC_length (loop_p, loop_nests)); | |
5964 | return true; | |
5965 | } | |
b8698a0f | 5966 | |
e855c69d AB |
5967 | return false; |
5968 | } | |
5969 | ||
b8698a0f | 5970 | /* Makes regions from the rest of the blocks, after loops are chosen |
e855c69d AB |
5971 | for pipelining. */ |
5972 | static void | |
5973 | make_regions_from_the_rest (void) | |
5974 | { | |
5975 | int cur_rgn_blocks; | |
5976 | int *loop_hdr; | |
5977 | int i; | |
5978 | ||
5979 | basic_block bb; | |
5980 | edge e; | |
5981 | edge_iterator ei; | |
5982 | int *degree; | |
e855c69d AB |
5983 | |
5984 | /* Index in rgn_bb_table where to start allocating new regions. */ | |
5985 | cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; | |
e855c69d | 5986 | |
b8698a0f | 5987 | /* Make regions from all the rest basic blocks - those that don't belong to |
e855c69d AB |
5988 | any loop or belong to irreducible loops. Prepare the data structures |
5989 | for extend_rgns. */ | |
5990 | ||
5991 | /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, | |
5992 | LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same | |
5993 | loop. */ | |
5994 | loop_hdr = XNEWVEC (int, last_basic_block); | |
5995 | degree = XCNEWVEC (int, last_basic_block); | |
5996 | ||
5997 | ||
5998 | /* For each basic block that belongs to some loop assign the number | |
5999 | of innermost loop it belongs to. */ | |
6000 | for (i = 0; i < last_basic_block; i++) | |
6001 | loop_hdr[i] = -1; | |
6002 | ||
6003 | FOR_EACH_BB (bb) | |
6004 | { | |
6005 | if (bb->loop_father && !bb->loop_father->num == 0 | |
6006 | && !(bb->flags & BB_IRREDUCIBLE_LOOP)) | |
6007 | loop_hdr[bb->index] = bb->loop_father->num; | |
6008 | } | |
6009 | ||
b8698a0f | 6010 | /* For each basic block degree is calculated as the number of incoming |
e855c69d AB |
6011 | edges, that are going out of bbs that are not yet scheduled. |
6012 | The basic blocks that are scheduled have degree value of zero. */ | |
b8698a0f | 6013 | FOR_EACH_BB (bb) |
e855c69d AB |
6014 | { |
6015 | degree[bb->index] = 0; | |
6016 | ||
6017 | if (!TEST_BIT (bbs_in_loop_rgns, bb->index)) | |
6018 | { | |
6019 | FOR_EACH_EDGE (e, ei, bb->preds) | |
6020 | if (!TEST_BIT (bbs_in_loop_rgns, e->src->index)) | |
6021 | degree[bb->index]++; | |
6022 | } | |
6023 | else | |
6024 | degree[bb->index] = -1; | |
6025 | } | |
6026 | ||
6027 | extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); | |
6028 | ||
6029 | /* Any block that did not end up in a region is placed into a region | |
6030 | by itself. */ | |
6031 | FOR_EACH_BB (bb) | |
6032 | if (degree[bb->index] >= 0) | |
6033 | { | |
6034 | rgn_bb_table[cur_rgn_blocks] = bb->index; | |
6035 | RGN_NR_BLOCKS (nr_regions) = 1; | |
6036 | RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; | |
6037 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
6038 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
6039 | CONTAINING_RGN (bb->index) = nr_regions++; | |
6040 | BLOCK_TO_BB (bb->index) = 0; | |
6041 | } | |
6042 | ||
6043 | free (degree); | |
6044 | free (loop_hdr); | |
6045 | } | |
6046 | ||
6047 | /* Free data structures used in pipelining of loops. */ | |
6048 | void sel_finish_pipelining (void) | |
6049 | { | |
6050 | loop_iterator li; | |
6051 | struct loop *loop; | |
6052 | ||
6053 | /* Release aux fields so we don't free them later by mistake. */ | |
6054 | FOR_EACH_LOOP (li, loop, 0) | |
6055 | loop->aux = NULL; | |
6056 | ||
6057 | loop_optimizer_finalize (); | |
6058 | ||
6059 | VEC_free (loop_p, heap, loop_nests); | |
6060 | ||
6061 | free (rev_top_order_index); | |
6062 | rev_top_order_index = NULL; | |
6063 | } | |
6064 | ||
b8698a0f | 6065 | /* This function replaces the find_rgns when |
e855c69d | 6066 | FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ |
b8698a0f | 6067 | void |
e855c69d AB |
6068 | sel_find_rgns (void) |
6069 | { | |
6070 | sel_init_pipelining (); | |
6071 | extend_regions (); | |
6072 | ||
6073 | if (current_loops) | |
6074 | { | |
6075 | loop_p loop; | |
6076 | loop_iterator li; | |
6077 | ||
6078 | FOR_EACH_LOOP (li, loop, (flag_sel_sched_pipelining_outer_loops | |
6079 | ? LI_FROM_INNERMOST | |
6080 | : LI_ONLY_INNERMOST)) | |
6081 | make_regions_from_loop_nest (loop); | |
6082 | } | |
6083 | ||
6084 | /* Make regions from all the rest basic blocks and schedule them. | |
b8698a0f | 6085 | These blocks include blocks that don't belong to any loop or belong |
e855c69d AB |
6086 | to irreducible loops. */ |
6087 | make_regions_from_the_rest (); | |
6088 | ||
6089 | /* We don't need bbs_in_loop_rgns anymore. */ | |
6090 | sbitmap_free (bbs_in_loop_rgns); | |
6091 | bbs_in_loop_rgns = NULL; | |
6092 | } | |
6093 | ||
b8698a0f L |
6094 | /* Adds the preheader blocks from previous loop to current region taking |
6095 | it from LOOP_PREHEADER_BLOCKS (current_loop_nest). | |
e855c69d AB |
6096 | This function is only used with -fsel-sched-pipelining-outer-loops. */ |
6097 | void | |
6098 | sel_add_loop_preheaders (void) | |
6099 | { | |
6100 | int i; | |
6101 | basic_block bb; | |
b8698a0f | 6102 | VEC(basic_block, heap) *preheader_blocks |
e855c69d AB |
6103 | = LOOP_PREHEADER_BLOCKS (current_loop_nest); |
6104 | ||
6105 | for (i = 0; | |
6106 | VEC_iterate (basic_block, preheader_blocks, i, bb); | |
6107 | i++) | |
8ec4d0ad AM |
6108 | { |
6109 | VEC_safe_push (basic_block, heap, last_added_blocks, bb); | |
e855c69d | 6110 | sel_add_bb (bb); |
8ec4d0ad | 6111 | } |
e855c69d AB |
6112 | |
6113 | VEC_free (basic_block, heap, preheader_blocks); | |
6114 | } | |
6115 | ||
b8698a0f L |
6116 | /* While pipelining outer loops, returns TRUE if BB is a loop preheader. |
6117 | Please note that the function should also work when pipelining_p is | |
6118 | false, because it is used when deciding whether we should or should | |
e855c69d AB |
6119 | not reschedule pipelined code. */ |
6120 | bool | |
6121 | sel_is_loop_preheader_p (basic_block bb) | |
6122 | { | |
6123 | if (current_loop_nest) | |
6124 | { | |
6125 | struct loop *outer; | |
6126 | ||
6127 | if (preheader_removed) | |
6128 | return false; | |
6129 | ||
6130 | /* Preheader is the first block in the region. */ | |
6131 | if (BLOCK_TO_BB (bb->index) == 0) | |
6132 | return true; | |
6133 | ||
6134 | /* We used to find a preheader with the topological information. | |
6135 | Check that the above code is equivalent to what we did before. */ | |
6136 | ||
6137 | if (in_current_region_p (current_loop_nest->header)) | |
b8698a0f | 6138 | gcc_assert (!(BLOCK_TO_BB (bb->index) |
e855c69d AB |
6139 | < BLOCK_TO_BB (current_loop_nest->header->index))); |
6140 | ||
6141 | /* Support the situation when the latch block of outer loop | |
6142 | could be from here. */ | |
6143 | for (outer = loop_outer (current_loop_nest); | |
6144 | outer; | |
6145 | outer = loop_outer (outer)) | |
6146 | if (considered_for_pipelining_p (outer) && outer->latch == bb) | |
6147 | gcc_unreachable (); | |
6148 | } | |
6149 | ||
6150 | return false; | |
6151 | } | |
6152 | ||
753de8cf AM |
6153 | /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and |
6154 | can be removed, making the corresponding edge fallthrough (assuming that | |
6155 | all basic blocks between JUMP_BB and DEST_BB are empty). */ | |
6156 | static bool | |
6157 | bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb) | |
e855c69d | 6158 | { |
b4550bf7 AM |
6159 | if (!onlyjump_p (BB_END (jump_bb)) |
6160 | || tablejump_p (BB_END (jump_bb), NULL, NULL)) | |
e855c69d AB |
6161 | return false; |
6162 | ||
b8698a0f | 6163 | /* Several outgoing edges, abnormal edge or destination of jump is |
e855c69d AB |
6164 | not DEST_BB. */ |
6165 | if (EDGE_COUNT (jump_bb->succs) != 1 | |
753de8cf | 6166 | || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING) |
e855c69d AB |
6167 | || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) |
6168 | return false; | |
6169 | ||
6170 | /* If not anything of the upper. */ | |
6171 | return true; | |
6172 | } | |
6173 | ||
6174 | /* Removes the loop preheader from the current region and saves it in | |
b8698a0f | 6175 | PREHEADER_BLOCKS of the father loop, so they will be added later to |
e855c69d AB |
6176 | region that represents an outer loop. */ |
6177 | static void | |
6178 | sel_remove_loop_preheader (void) | |
6179 | { | |
6180 | int i, old_len; | |
6181 | int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
6182 | basic_block bb; | |
6183 | bool all_empty_p = true; | |
b8698a0f | 6184 | VEC(basic_block, heap) *preheader_blocks |
e855c69d AB |
6185 | = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); |
6186 | ||
6187 | gcc_assert (current_loop_nest); | |
6188 | old_len = VEC_length (basic_block, preheader_blocks); | |
6189 | ||
6190 | /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ | |
6191 | for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) | |
6192 | { | |
6193 | bb = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6194 | ||
b8698a0f | 6195 | /* If the basic block belongs to region, but doesn't belong to |
e855c69d AB |
6196 | corresponding loop, then it should be a preheader. */ |
6197 | if (sel_is_loop_preheader_p (bb)) | |
6198 | { | |
6199 | VEC_safe_push (basic_block, heap, preheader_blocks, bb); | |
6200 | if (BB_END (bb) != bb_note (bb)) | |
6201 | all_empty_p = false; | |
6202 | } | |
6203 | } | |
b8698a0f | 6204 | |
e855c69d AB |
6205 | /* Remove these blocks only after iterating over the whole region. */ |
6206 | for (i = VEC_length (basic_block, preheader_blocks) - 1; | |
6207 | i >= old_len; | |
6208 | i--) | |
6209 | { | |
b8698a0f | 6210 | bb = VEC_index (basic_block, preheader_blocks, i); |
e855c69d AB |
6211 | sel_remove_bb (bb, false); |
6212 | } | |
6213 | ||
6214 | if (!considered_for_pipelining_p (loop_outer (current_loop_nest))) | |
6215 | { | |
6216 | if (!all_empty_p) | |
6217 | /* Immediately create new region from preheader. */ | |
6218 | make_region_from_loop_preheader (&preheader_blocks); | |
6219 | else | |
6220 | { | |
6221 | /* If all preheader blocks are empty - dont create new empty region. | |
6222 | Instead, remove them completely. */ | |
ac47786e | 6223 | FOR_EACH_VEC_ELT (basic_block, preheader_blocks, i, bb) |
e855c69d AB |
6224 | { |
6225 | edge e; | |
6226 | edge_iterator ei; | |
6227 | basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; | |
6228 | ||
6229 | /* Redirect all incoming edges to next basic block. */ | |
6230 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) | |
6231 | { | |
6232 | if (! (e->flags & EDGE_FALLTHRU)) | |
6233 | redirect_edge_and_branch (e, bb->next_bb); | |
6234 | else | |
6235 | redirect_edge_succ (e, bb->next_bb); | |
6236 | } | |
6237 | gcc_assert (BB_NOTE_LIST (bb) == NULL); | |
6238 | delete_and_free_basic_block (bb); | |
6239 | ||
b8698a0f L |
6240 | /* Check if after deleting preheader there is a nonconditional |
6241 | jump in PREV_BB that leads to the next basic block NEXT_BB. | |
6242 | If it is so - delete this jump and clear data sets of its | |
e855c69d AB |
6243 | basic block if it becomes empty. */ |
6244 | if (next_bb->prev_bb == prev_bb | |
6245 | && prev_bb != ENTRY_BLOCK_PTR | |
753de8cf | 6246 | && bb_has_removable_jump_to_p (prev_bb, next_bb)) |
e855c69d AB |
6247 | { |
6248 | redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); | |
6249 | if (BB_END (prev_bb) == bb_note (prev_bb)) | |
6250 | free_data_sets (prev_bb); | |
6251 | } | |
00c4e97c AB |
6252 | |
6253 | set_immediate_dominator (CDI_DOMINATORS, next_bb, | |
6254 | recompute_dominator (CDI_DOMINATORS, | |
6255 | next_bb)); | |
e855c69d AB |
6256 | } |
6257 | } | |
6258 | VEC_free (basic_block, heap, preheader_blocks); | |
6259 | } | |
6260 | else | |
6261 | /* Store preheader within the father's loop structure. */ | |
6262 | SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), | |
6263 | preheader_blocks); | |
6264 | } | |
6265 | #endif |