]>
Commit | Line | Data |
---|---|---|
e1ab7874 | 1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
711789cc | 2 | Copyright (C) 2006-2013 Free Software Foundation, Inc. |
e1ab7874 | 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" | |
0b205f4c | 24 | #include "diagnostic-core.h" |
e1ab7874 | 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" | |
e1ab7874 | 34 | #include "recog.h" |
35 | #include "params.h" | |
36 | #include "target.h" | |
e1ab7874 | 37 | #include "sched-int.h" |
38 | #include "ggc.h" | |
39 | #include "tree.h" | |
40 | #include "vec.h" | |
41 | #include "langhooks.h" | |
42 | #include "rtlhooks-def.h" | |
06f9d6ef | 43 | #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */ |
e1ab7874 | 44 | |
45 | #ifdef INSN_SCHEDULING | |
46 | #include "sel-sched-ir.h" | |
47 | /* We don't have to use it except for sel_print_insn. */ | |
48 | #include "sel-sched-dump.h" | |
49 | ||
50 | /* A vector holding bb info for whole scheduling pass. */ | |
f1f41a6c | 51 | vec<sel_global_bb_info_def> |
1e094109 | 52 | sel_global_bb_info = vNULL; |
e1ab7874 | 53 | |
54 | /* A vector holding bb info. */ | |
f1f41a6c | 55 | vec<sel_region_bb_info_def> |
1e094109 | 56 | sel_region_bb_info = vNULL; |
e1ab7874 | 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. */ | |
1e094109 | 72 | static vec<loop_p> loop_nests = vNULL; |
e1ab7874 | 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. */ | |
48e1416a | 125 | int s; |
e1ab7874 | 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 | ||
48e1416a | 137 | /* TRUE if while scheduling current region, which is loop, its preheader |
e1ab7874 | 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); | |
f1f41a6c | 151 | static void free_history_vect (vec<expr_history_def> &); |
e1ab7874 | 152 | |
153 | static void move_bb_info (basic_block, basic_block); | |
154 | static void remove_empty_bb (basic_block, bool); | |
0424f393 | 155 | static void sel_merge_blocks (basic_block, basic_block); |
e1ab7874 | 156 | static void sel_remove_loop_preheader (void); |
49087fba | 157 | static bool bb_has_removable_jump_to_p (basic_block, basic_block); |
e1ab7874 | 158 | |
159 | static bool insn_is_the_only_one_in_bb_p (insn_t); | |
160 | static void create_initial_data_sets (basic_block); | |
161 | ||
9845d120 | 162 | static void free_av_set (basic_block); |
e1ab7874 | 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 | |
48e1416a | 264 | flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, |
f1f41a6c | 265 | insn_t last_scheduled_insn, vec<rtx, va_gc> *executing_insns, |
48e1416a | 266 | int *ready_ticks, int ready_ticks_size, insn_t sched_next, |
abb9c563 | 267 | int cycle, int cycle_issued_insns, int issue_more, |
e1ab7874 | 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; | |
abb9c563 | 292 | FENCE_ISSUE_MORE (f) = issue_more; |
e1ab7874 | 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; | |
48e1416a | 323 | |
e1ab7874 | 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 | ||
48e1416a | 334 | /* Bulk target context. It is convenient for debugging purposes to ensure |
e1ab7874 | 335 | that there are no uninitialized (null) target contexts. */ |
336 | static tc_t bulk_tc = (tc_t) 1; | |
337 | ||
48e1416a | 338 | /* Target hooks wrappers. In the future we can provide some default |
e1ab7874 | 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 | |
48e1416a | 430 | /* Functions to work with dependence contexts. |
68e419a1 | 431 | Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence |
e1ab7874 | 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 | { | |
d9ab2038 | 439 | init_deps (to, false); |
e1ab7874 | 440 | deps_join (to, from); |
441 | } | |
442 | ||
443 | /* Allocate store for dep context. */ | |
444 | static deps_t | |
445 | alloc_deps_context (void) | |
446 | { | |
68e419a1 | 447 | return XNEW (struct deps_desc); |
e1ab7874 | 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 | ||
d9ab2038 | 456 | init_deps (dc, false); |
e1ab7874 | 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); | |
d9ab2038 | 490 | init_deps (dc, false); |
e1ab7874 | 491 | } |
492 | ||
48e1416a | 493 | /* This structure describes the dependence analysis hooks for advancing |
e1ab7874 | 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 | ||
dd045aee | 583 | free (s); |
e1ab7874 | 584 | |
585 | if (dc != NULL) | |
586 | delete_deps_context (dc); | |
587 | ||
588 | if (tc != NULL) | |
589 | delete_target_context (tc); | |
f1f41a6c | 590 | vec_free (FENCE_EXECUTING_INSNS (f)); |
e1ab7874 | 591 | free (FENCE_READY_TICKS (f)); |
592 | FENCE_READY_TICKS (f) = NULL; | |
593 | } | |
594 | ||
595 | /* Init a list of fences with successors of OLD_FENCE. */ | |
596 | void | |
597 | init_fences (insn_t old_fence) | |
598 | { | |
599 | insn_t succ; | |
600 | succ_iterator si; | |
601 | bool first = true; | |
602 | int ready_ticks_size = get_max_uid () + 1; | |
48e1416a | 603 | |
604 | FOR_EACH_SUCC_1 (succ, si, old_fence, | |
e1ab7874 | 605 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
606 | { | |
48e1416a | 607 | |
e1ab7874 | 608 | if (first) |
609 | first = false; | |
610 | else | |
611 | gcc_assert (flag_sel_sched_pipelining_outer_loops); | |
612 | ||
613 | flist_add (&fences, succ, | |
614 | state_create (), | |
615 | create_deps_context () /* dc */, | |
616 | create_target_context (true) /* tc */, | |
48e1416a | 617 | NULL_RTX /* last_scheduled_insn */, |
e1ab7874 | 618 | NULL, /* executing_insns */ |
619 | XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ | |
620 | ready_ticks_size, | |
621 | NULL_RTX /* sched_next */, | |
48e1416a | 622 | 1 /* cycle */, 0 /* cycle_issued_insns */, |
abb9c563 | 623 | issue_rate, /* issue_more */ |
48e1416a | 624 | 1 /* starts_cycle_p */, 0 /* after_stall_p */); |
e1ab7874 | 625 | } |
626 | } | |
627 | ||
628 | /* Merges two fences (filling fields of fence F with resulting values) by | |
629 | following rules: 1) state, target context and last scheduled insn are | |
48e1416a | 630 | propagated from fallthrough edge if it is available; |
e1ab7874 | 631 | 2) deps context and cycle is propagated from more probable edge; |
48e1416a | 632 | 3) all other fields are set to corresponding constant values. |
e1ab7874 | 633 | |
48e1416a | 634 | INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, |
abb9c563 | 635 | READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE |
636 | and AFTER_STALL_P are the corresponding fields of the second fence. */ | |
e1ab7874 | 637 | static void |
638 | merge_fences (fence_t f, insn_t insn, | |
48e1416a | 639 | state_t state, deps_t dc, void *tc, |
f1f41a6c | 640 | rtx last_scheduled_insn, vec<rtx, va_gc> *executing_insns, |
e1ab7874 | 641 | int *ready_ticks, int ready_ticks_size, |
abb9c563 | 642 | rtx sched_next, int cycle, int issue_more, bool after_stall_p) |
e1ab7874 | 643 | { |
644 | insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f); | |
645 | ||
646 | gcc_assert (sel_bb_head_p (FENCE_INSN (f)) | |
647 | && !sched_next && !FENCE_SCHED_NEXT (f)); | |
648 | ||
48e1416a | 649 | /* Check if we can decide which path fences came. |
e1ab7874 | 650 | If we can't (or don't want to) - reset all. */ |
651 | if (last_scheduled_insn == NULL | |
652 | || last_scheduled_insn_old == NULL | |
48e1416a | 653 | /* This is a case when INSN is reachable on several paths from |
654 | one insn (this can happen when pipelining of outer loops is on and | |
655 | there are two edges: one going around of inner loop and the other - | |
e1ab7874 | 656 | right through it; in such case just reset everything). */ |
657 | || last_scheduled_insn == last_scheduled_insn_old) | |
658 | { | |
659 | state_reset (FENCE_STATE (f)); | |
660 | state_free (state); | |
48e1416a | 661 | |
e1ab7874 | 662 | reset_deps_context (FENCE_DC (f)); |
663 | delete_deps_context (dc); | |
48e1416a | 664 | |
e1ab7874 | 665 | reset_target_context (FENCE_TC (f), true); |
666 | delete_target_context (tc); | |
667 | ||
668 | if (cycle > FENCE_CYCLE (f)) | |
669 | FENCE_CYCLE (f) = cycle; | |
670 | ||
671 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; | |
abb9c563 | 672 | FENCE_ISSUE_MORE (f) = issue_rate; |
f1f41a6c | 673 | vec_free (executing_insns); |
e1ab7874 | 674 | free (ready_ticks); |
675 | if (FENCE_EXECUTING_INSNS (f)) | |
f1f41a6c | 676 | FENCE_EXECUTING_INSNS (f)->block_remove (0, |
677 | FENCE_EXECUTING_INSNS (f)->length ()); | |
e1ab7874 | 678 | if (FENCE_READY_TICKS (f)) |
679 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
680 | } | |
681 | else | |
682 | { | |
683 | edge edge_old = NULL, edge_new = NULL; | |
684 | edge candidate; | |
685 | succ_iterator si; | |
686 | insn_t succ; | |
48e1416a | 687 | |
e1ab7874 | 688 | /* Find fallthrough edge. */ |
689 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); | |
7f58c05e | 690 | candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb); |
e1ab7874 | 691 | |
692 | if (!candidate | |
693 | || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn) | |
694 | && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old))) | |
695 | { | |
696 | /* No fallthrough edge leading to basic block of INSN. */ | |
697 | state_reset (FENCE_STATE (f)); | |
698 | state_free (state); | |
48e1416a | 699 | |
e1ab7874 | 700 | reset_target_context (FENCE_TC (f), true); |
701 | delete_target_context (tc); | |
48e1416a | 702 | |
e1ab7874 | 703 | FENCE_LAST_SCHEDULED_INSN (f) = NULL; |
abb9c563 | 704 | FENCE_ISSUE_MORE (f) = issue_rate; |
e1ab7874 | 705 | } |
706 | else | |
707 | if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn)) | |
708 | { | |
48e1416a | 709 | /* Would be weird if same insn is successor of several fallthrough |
e1ab7874 | 710 | edges. */ |
711 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
712 | != BLOCK_FOR_INSN (last_scheduled_insn_old)); | |
713 | ||
714 | state_free (FENCE_STATE (f)); | |
715 | FENCE_STATE (f) = state; | |
716 | ||
717 | delete_target_context (FENCE_TC (f)); | |
718 | FENCE_TC (f) = tc; | |
719 | ||
720 | FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; | |
abb9c563 | 721 | FENCE_ISSUE_MORE (f) = issue_more; |
e1ab7874 | 722 | } |
723 | else | |
724 | { | |
725 | /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ | |
726 | state_free (state); | |
727 | delete_target_context (tc); | |
728 | ||
729 | gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb | |
730 | != BLOCK_FOR_INSN (last_scheduled_insn)); | |
731 | } | |
732 | ||
733 | /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ | |
734 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, | |
735 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
736 | { | |
737 | if (succ == insn) | |
738 | { | |
739 | /* No same successor allowed from several edges. */ | |
740 | gcc_assert (!edge_old); | |
741 | edge_old = si.e1; | |
742 | } | |
743 | } | |
744 | /* Find edge of second predecessor (last_scheduled_insn->insn). */ | |
745 | FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, | |
746 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
747 | { | |
748 | if (succ == insn) | |
749 | { | |
750 | /* No same successor allowed from several edges. */ | |
751 | gcc_assert (!edge_new); | |
752 | edge_new = si.e1; | |
753 | } | |
754 | } | |
755 | ||
756 | /* Check if we can choose most probable predecessor. */ | |
757 | if (edge_old == NULL || edge_new == NULL) | |
758 | { | |
759 | reset_deps_context (FENCE_DC (f)); | |
760 | delete_deps_context (dc); | |
f1f41a6c | 761 | vec_free (executing_insns); |
e1ab7874 | 762 | free (ready_ticks); |
48e1416a | 763 | |
e1ab7874 | 764 | FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); |
765 | if (FENCE_EXECUTING_INSNS (f)) | |
f1f41a6c | 766 | FENCE_EXECUTING_INSNS (f)->block_remove (0, |
767 | FENCE_EXECUTING_INSNS (f)->length ()); | |
e1ab7874 | 768 | if (FENCE_READY_TICKS (f)) |
769 | memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); | |
770 | } | |
771 | else | |
772 | if (edge_new->probability > edge_old->probability) | |
773 | { | |
774 | delete_deps_context (FENCE_DC (f)); | |
775 | FENCE_DC (f) = dc; | |
f1f41a6c | 776 | vec_free (FENCE_EXECUTING_INSNS (f)); |
e1ab7874 | 777 | FENCE_EXECUTING_INSNS (f) = executing_insns; |
778 | free (FENCE_READY_TICKS (f)); | |
779 | FENCE_READY_TICKS (f) = ready_ticks; | |
780 | FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; | |
781 | FENCE_CYCLE (f) = cycle; | |
782 | } | |
783 | else | |
784 | { | |
785 | /* Leave DC and CYCLE untouched. */ | |
786 | delete_deps_context (dc); | |
f1f41a6c | 787 | vec_free (executing_insns); |
e1ab7874 | 788 | free (ready_ticks); |
789 | } | |
790 | } | |
791 | ||
792 | /* Fill remaining invariant fields. */ | |
793 | if (after_stall_p) | |
794 | FENCE_AFTER_STALL_P (f) = 1; | |
795 | ||
796 | FENCE_ISSUED_INSNS (f) = 0; | |
797 | FENCE_STARTS_CYCLE_P (f) = 1; | |
798 | FENCE_SCHED_NEXT (f) = NULL; | |
799 | } | |
800 | ||
48e1416a | 801 | /* Add a new fence to NEW_FENCES list, initializing it from all |
e1ab7874 | 802 | other parameters. */ |
803 | static void | |
804 | add_to_fences (flist_tail_t new_fences, insn_t insn, | |
48e1416a | 805 | state_t state, deps_t dc, void *tc, rtx last_scheduled_insn, |
f1f41a6c | 806 | vec<rtx, va_gc> *executing_insns, int *ready_ticks, |
48e1416a | 807 | int ready_ticks_size, rtx sched_next, int cycle, |
abb9c563 | 808 | int cycle_issued_insns, int issue_rate, |
809 | bool starts_cycle_p, bool after_stall_p) | |
e1ab7874 | 810 | { |
811 | fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn); | |
812 | ||
813 | if (! f) | |
814 | { | |
815 | flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc, | |
48e1416a | 816 | last_scheduled_insn, executing_insns, ready_ticks, |
e1ab7874 | 817 | ready_ticks_size, sched_next, cycle, cycle_issued_insns, |
abb9c563 | 818 | issue_rate, starts_cycle_p, after_stall_p); |
e1ab7874 | 819 | |
820 | FLIST_TAIL_TAILP (new_fences) | |
821 | = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences)); | |
822 | } | |
823 | else | |
824 | { | |
48e1416a | 825 | merge_fences (f, insn, state, dc, tc, last_scheduled_insn, |
826 | executing_insns, ready_ticks, ready_ticks_size, | |
abb9c563 | 827 | sched_next, cycle, issue_rate, after_stall_p); |
e1ab7874 | 828 | } |
829 | } | |
830 | ||
831 | /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */ | |
832 | void | |
833 | move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences) | |
834 | { | |
835 | fence_t f, old; | |
836 | flist_t *tailp = FLIST_TAIL_TAILP (new_fences); | |
837 | ||
838 | old = FLIST_FENCE (old_fences); | |
48e1416a | 839 | f = flist_lookup (FLIST_TAIL_HEAD (new_fences), |
e1ab7874 | 840 | FENCE_INSN (FLIST_FENCE (old_fences))); |
841 | if (f) | |
842 | { | |
843 | merge_fences (f, old->insn, old->state, old->dc, old->tc, | |
844 | old->last_scheduled_insn, old->executing_insns, | |
845 | old->ready_ticks, old->ready_ticks_size, | |
abb9c563 | 846 | old->sched_next, old->cycle, old->issue_more, |
e1ab7874 | 847 | old->after_stall_p); |
848 | } | |
849 | else | |
850 | { | |
851 | _list_add (tailp); | |
852 | FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp); | |
853 | *FLIST_FENCE (*tailp) = *old; | |
854 | init_fence_for_scheduling (FLIST_FENCE (*tailp)); | |
855 | } | |
856 | FENCE_INSN (old) = NULL; | |
857 | } | |
858 | ||
48e1416a | 859 | /* Add a new fence to NEW_FENCES list and initialize most of its data |
e1ab7874 | 860 | as a clean one. */ |
861 | void | |
862 | add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
863 | { | |
864 | int ready_ticks_size = get_max_uid () + 1; | |
48e1416a | 865 | |
e1ab7874 | 866 | add_to_fences (new_fences, |
867 | succ, state_create (), create_deps_context (), | |
868 | create_target_context (true), | |
48e1416a | 869 | NULL_RTX, NULL, |
e1ab7874 | 870 | XCNEWVEC (int, ready_ticks_size), ready_ticks_size, |
871 | NULL_RTX, FENCE_CYCLE (fence) + 1, | |
abb9c563 | 872 | 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence)); |
e1ab7874 | 873 | } |
874 | ||
48e1416a | 875 | /* Add a new fence to NEW_FENCES list and initialize all of its data |
e1ab7874 | 876 | from FENCE and SUCC. */ |
877 | void | |
878 | add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) | |
879 | { | |
48e1416a | 880 | int * new_ready_ticks |
e1ab7874 | 881 | = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); |
48e1416a | 882 | |
e1ab7874 | 883 | memcpy (new_ready_ticks, FENCE_READY_TICKS (fence), |
884 | FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); | |
885 | add_to_fences (new_fences, | |
886 | succ, state_create_copy (FENCE_STATE (fence)), | |
887 | create_copy_of_deps_context (FENCE_DC (fence)), | |
888 | create_copy_of_target_context (FENCE_TC (fence)), | |
48e1416a | 889 | FENCE_LAST_SCHEDULED_INSN (fence), |
f1f41a6c | 890 | vec_safe_copy (FENCE_EXECUTING_INSNS (fence)), |
e1ab7874 | 891 | new_ready_ticks, |
892 | FENCE_READY_TICKS_SIZE (fence), | |
893 | FENCE_SCHED_NEXT (fence), | |
894 | FENCE_CYCLE (fence), | |
895 | FENCE_ISSUED_INSNS (fence), | |
abb9c563 | 896 | FENCE_ISSUE_MORE (fence), |
e1ab7874 | 897 | FENCE_STARTS_CYCLE_P (fence), |
898 | FENCE_AFTER_STALL_P (fence)); | |
899 | } | |
900 | \f | |
901 | ||
902 | /* Functions to work with regset and nop pools. */ | |
903 | ||
904 | /* Returns the new regset from pool. It might have some of the bits set | |
905 | from the previous usage. */ | |
906 | regset | |
907 | get_regset_from_pool (void) | |
908 | { | |
909 | regset rs; | |
910 | ||
911 | if (regset_pool.n != 0) | |
912 | rs = regset_pool.v[--regset_pool.n]; | |
913 | else | |
914 | /* We need to create the regset. */ | |
915 | { | |
916 | rs = ALLOC_REG_SET (®_obstack); | |
917 | ||
918 | if (regset_pool.nn == regset_pool.ss) | |
919 | regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, | |
920 | (regset_pool.ss = 2 * regset_pool.ss + 1)); | |
921 | regset_pool.vv[regset_pool.nn++] = rs; | |
922 | } | |
923 | ||
924 | regset_pool.diff++; | |
925 | ||
926 | return rs; | |
927 | } | |
928 | ||
929 | /* Same as above, but returns the empty regset. */ | |
930 | regset | |
931 | get_clear_regset_from_pool (void) | |
932 | { | |
933 | regset rs = get_regset_from_pool (); | |
934 | ||
935 | CLEAR_REG_SET (rs); | |
936 | return rs; | |
937 | } | |
938 | ||
939 | /* Return regset RS to the pool for future use. */ | |
940 | void | |
941 | return_regset_to_pool (regset rs) | |
942 | { | |
bc9cb5ed | 943 | gcc_assert (rs); |
e1ab7874 | 944 | regset_pool.diff--; |
945 | ||
946 | if (regset_pool.n == regset_pool.s) | |
947 | regset_pool.v = XRESIZEVEC (regset, regset_pool.v, | |
948 | (regset_pool.s = 2 * regset_pool.s + 1)); | |
949 | regset_pool.v[regset_pool.n++] = rs; | |
950 | } | |
951 | ||
dde7ed1e | 952 | #ifdef ENABLE_CHECKING |
e1ab7874 | 953 | /* This is used as a qsort callback for sorting regset pool stacks. |
954 | X and XX are addresses of two regsets. They are never equal. */ | |
955 | static int | |
956 | cmp_v_in_regset_pool (const void *x, const void *xx) | |
957 | { | |
c72f63ac | 958 | uintptr_t r1 = (uintptr_t) *((const regset *) x); |
959 | uintptr_t r2 = (uintptr_t) *((const regset *) xx); | |
960 | if (r1 > r2) | |
961 | return 1; | |
962 | else if (r1 < r2) | |
963 | return -1; | |
964 | gcc_unreachable (); | |
e1ab7874 | 965 | } |
dde7ed1e | 966 | #endif |
e1ab7874 | 967 | |
968 | /* Free the regset pool possibly checking for memory leaks. */ | |
969 | void | |
970 | free_regset_pool (void) | |
971 | { | |
972 | #ifdef ENABLE_CHECKING | |
973 | { | |
974 | regset *v = regset_pool.v; | |
975 | int i = 0; | |
976 | int n = regset_pool.n; | |
48e1416a | 977 | |
e1ab7874 | 978 | regset *vv = regset_pool.vv; |
979 | int ii = 0; | |
980 | int nn = regset_pool.nn; | |
48e1416a | 981 | |
e1ab7874 | 982 | int diff = 0; |
48e1416a | 983 | |
e1ab7874 | 984 | gcc_assert (n <= nn); |
48e1416a | 985 | |
e1ab7874 | 986 | /* Sort both vectors so it will be possible to compare them. */ |
987 | qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); | |
988 | qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); | |
48e1416a | 989 | |
e1ab7874 | 990 | while (ii < nn) |
991 | { | |
992 | if (v[i] == vv[ii]) | |
993 | i++; | |
994 | else | |
995 | /* VV[II] was lost. */ | |
996 | diff++; | |
48e1416a | 997 | |
e1ab7874 | 998 | ii++; |
999 | } | |
48e1416a | 1000 | |
e1ab7874 | 1001 | gcc_assert (diff == regset_pool.diff); |
1002 | } | |
1003 | #endif | |
48e1416a | 1004 | |
e1ab7874 | 1005 | /* If not true - we have a memory leak. */ |
1006 | gcc_assert (regset_pool.diff == 0); | |
48e1416a | 1007 | |
e1ab7874 | 1008 | while (regset_pool.n) |
1009 | { | |
1010 | --regset_pool.n; | |
1011 | FREE_REG_SET (regset_pool.v[regset_pool.n]); | |
1012 | } | |
1013 | ||
1014 | free (regset_pool.v); | |
1015 | regset_pool.v = NULL; | |
1016 | regset_pool.s = 0; | |
48e1416a | 1017 | |
e1ab7874 | 1018 | free (regset_pool.vv); |
1019 | regset_pool.vv = NULL; | |
1020 | regset_pool.nn = 0; | |
1021 | regset_pool.ss = 0; | |
1022 | ||
1023 | regset_pool.diff = 0; | |
1024 | } | |
1025 | \f | |
1026 | ||
48e1416a | 1027 | /* Functions to work with nop pools. NOP insns are used as temporary |
1028 | placeholders of the insns being scheduled to allow correct update of | |
e1ab7874 | 1029 | the data sets. When update is finished, NOPs are deleted. */ |
1030 | ||
1031 | /* A vinsn that is used to represent a nop. This vinsn is shared among all | |
1032 | nops sel-sched generates. */ | |
1033 | static vinsn_t nop_vinsn = NULL; | |
1034 | ||
1035 | /* Emit a nop before INSN, taking it from pool. */ | |
1036 | insn_t | |
1037 | get_nop_from_pool (insn_t insn) | |
1038 | { | |
1039 | insn_t nop; | |
1040 | bool old_p = nop_pool.n != 0; | |
1041 | int flags; | |
1042 | ||
1043 | if (old_p) | |
1044 | nop = nop_pool.v[--nop_pool.n]; | |
1045 | else | |
1046 | nop = nop_pattern; | |
1047 | ||
1048 | nop = emit_insn_before (nop, insn); | |
1049 | ||
1050 | if (old_p) | |
1051 | flags = INSN_INIT_TODO_SSID; | |
1052 | else | |
1053 | flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; | |
1054 | ||
1055 | set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); | |
1056 | sel_init_new_insn (nop, flags); | |
1057 | ||
1058 | return nop; | |
1059 | } | |
1060 | ||
1061 | /* Remove NOP from the instruction stream and return it to the pool. */ | |
1062 | void | |
9845d120 | 1063 | return_nop_to_pool (insn_t nop, bool full_tidying) |
e1ab7874 | 1064 | { |
1065 | gcc_assert (INSN_IN_STREAM_P (nop)); | |
9845d120 | 1066 | sel_remove_insn (nop, false, full_tidying); |
e1ab7874 | 1067 | |
93ff53d3 | 1068 | /* We'll recycle this nop. */ |
1069 | INSN_DELETED_P (nop) = 0; | |
1070 | ||
e1ab7874 | 1071 | if (nop_pool.n == nop_pool.s) |
48e1416a | 1072 | nop_pool.v = XRESIZEVEC (rtx, nop_pool.v, |
e1ab7874 | 1073 | (nop_pool.s = 2 * nop_pool.s + 1)); |
1074 | nop_pool.v[nop_pool.n++] = nop; | |
1075 | } | |
1076 | ||
1077 | /* Free the nop pool. */ | |
1078 | void | |
1079 | free_nop_pool (void) | |
1080 | { | |
1081 | nop_pool.n = 0; | |
1082 | nop_pool.s = 0; | |
1083 | free (nop_pool.v); | |
1084 | nop_pool.v = NULL; | |
1085 | } | |
1086 | \f | |
1087 | ||
48e1416a | 1088 | /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb. |
e1ab7874 | 1089 | The callback is given two rtxes XX and YY and writes the new rtxes |
1090 | to NX and NY in case some needs to be skipped. */ | |
1091 | static int | |
1092 | skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) | |
1093 | { | |
1094 | const_rtx x = *xx; | |
1095 | const_rtx y = *yy; | |
48e1416a | 1096 | |
e1ab7874 | 1097 | if (GET_CODE (x) == UNSPEC |
1098 | && (targetm.sched.skip_rtx_p == NULL | |
1099 | || targetm.sched.skip_rtx_p (x))) | |
1100 | { | |
1101 | *nx = XVECEXP (x, 0, 0); | |
1102 | *ny = CONST_CAST_RTX (y); | |
1103 | return 1; | |
1104 | } | |
48e1416a | 1105 | |
e1ab7874 | 1106 | if (GET_CODE (y) == UNSPEC |
1107 | && (targetm.sched.skip_rtx_p == NULL | |
1108 | || targetm.sched.skip_rtx_p (y))) | |
1109 | { | |
1110 | *nx = CONST_CAST_RTX (x); | |
1111 | *ny = XVECEXP (y, 0, 0); | |
1112 | return 1; | |
1113 | } | |
48e1416a | 1114 | |
e1ab7874 | 1115 | return 0; |
1116 | } | |
1117 | ||
48e1416a | 1118 | /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way |
e1ab7874 | 1119 | to support ia64 speculation. When changes are needed, new rtx X and new mode |
1120 | NMODE are written, and the callback returns true. */ | |
1121 | static int | |
1122 | hash_with_unspec_callback (const_rtx x, enum machine_mode mode ATTRIBUTE_UNUSED, | |
1123 | rtx *nx, enum machine_mode* nmode) | |
1124 | { | |
48e1416a | 1125 | if (GET_CODE (x) == UNSPEC |
e1ab7874 | 1126 | && targetm.sched.skip_rtx_p |
1127 | && targetm.sched.skip_rtx_p (x)) | |
1128 | { | |
1129 | *nx = XVECEXP (x, 0 ,0); | |
8458f4ca | 1130 | *nmode = VOIDmode; |
e1ab7874 | 1131 | return 1; |
1132 | } | |
48e1416a | 1133 | |
e1ab7874 | 1134 | return 0; |
1135 | } | |
1136 | ||
1137 | /* Returns LHS and RHS are ok to be scheduled separately. */ | |
1138 | static bool | |
1139 | lhs_and_rhs_separable_p (rtx lhs, rtx rhs) | |
1140 | { | |
1141 | if (lhs == NULL || rhs == NULL) | |
1142 | return false; | |
1143 | ||
48e1416a | 1144 | /* Do not schedule CONST, CONST_INT and CONST_DOUBLE etc as rhs: no point |
1145 | to use reg, if const can be used. Moreover, scheduling const as rhs may | |
1146 | lead to mode mismatch cause consts don't have modes but they could be | |
e1ab7874 | 1147 | merged from branches where the same const used in different modes. */ |
1148 | if (CONSTANT_P (rhs)) | |
1149 | return false; | |
1150 | ||
1151 | /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ | |
1152 | if (COMPARISON_P (rhs)) | |
1153 | return false; | |
1154 | ||
1155 | /* Do not allow single REG to be an rhs. */ | |
1156 | if (REG_P (rhs)) | |
1157 | return false; | |
1158 | ||
48e1416a | 1159 | /* See comment at find_used_regs_1 (*1) for explanation of this |
e1ab7874 | 1160 | restriction. */ |
1161 | /* FIXME: remove this later. */ | |
1162 | if (MEM_P (lhs)) | |
1163 | return false; | |
1164 | ||
1165 | /* This will filter all tricky things like ZERO_EXTRACT etc. | |
1166 | For now we don't handle it. */ | |
1167 | if (!REG_P (lhs) && !MEM_P (lhs)) | |
1168 | return false; | |
1169 | ||
1170 | return true; | |
1171 | } | |
1172 | ||
48e1416a | 1173 | /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When |
1174 | FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is | |
e1ab7874 | 1175 | used e.g. for insns from recovery blocks. */ |
1176 | static void | |
1177 | vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) | |
1178 | { | |
1179 | hash_rtx_callback_function hrcf; | |
1180 | int insn_class; | |
1181 | ||
1182 | VINSN_INSN_RTX (vi) = insn; | |
1183 | VINSN_COUNT (vi) = 0; | |
1184 | vi->cost = -1; | |
48e1416a | 1185 | |
bc9cb5ed | 1186 | if (INSN_NOP_P (insn)) |
1187 | return; | |
1188 | ||
e1ab7874 | 1189 | if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) |
1190 | init_id_from_df (VINSN_ID (vi), insn, force_unique_p); | |
1191 | else | |
1192 | deps_init_id (VINSN_ID (vi), insn, force_unique_p); | |
48e1416a | 1193 | |
e1ab7874 | 1194 | /* Hash vinsn depending on whether it is separable or not. */ |
1195 | hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; | |
1196 | if (VINSN_SEPARABLE_P (vi)) | |
1197 | { | |
1198 | rtx rhs = VINSN_RHS (vi); | |
1199 | ||
1200 | VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs), | |
1201 | NULL, NULL, false, hrcf); | |
1202 | VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi), | |
1203 | VOIDmode, NULL, NULL, | |
1204 | false, hrcf); | |
1205 | } | |
1206 | else | |
1207 | { | |
1208 | VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode, | |
1209 | NULL, NULL, false, hrcf); | |
1210 | VINSN_HASH_RTX (vi) = VINSN_HASH (vi); | |
1211 | } | |
48e1416a | 1212 | |
e1ab7874 | 1213 | insn_class = haifa_classify_insn (insn); |
1214 | if (insn_class >= 2 | |
1215 | && (!targetm.sched.get_insn_spec_ds | |
1216 | || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) | |
1217 | == 0))) | |
1218 | VINSN_MAY_TRAP_P (vi) = true; | |
1219 | else | |
1220 | VINSN_MAY_TRAP_P (vi) = false; | |
1221 | } | |
1222 | ||
1223 | /* Indicate that VI has become the part of an rtx object. */ | |
1224 | void | |
1225 | vinsn_attach (vinsn_t vi) | |
1226 | { | |
1227 | /* Assert that VI is not pending for deletion. */ | |
1228 | gcc_assert (VINSN_INSN_RTX (vi)); | |
1229 | ||
1230 | VINSN_COUNT (vi)++; | |
1231 | } | |
1232 | ||
48e1416a | 1233 | /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct |
e1ab7874 | 1234 | VINSN_TYPE (VI). */ |
1235 | static vinsn_t | |
1236 | vinsn_create (insn_t insn, bool force_unique_p) | |
1237 | { | |
1238 | vinsn_t vi = XCNEW (struct vinsn_def); | |
1239 | ||
1240 | vinsn_init (vi, insn, force_unique_p); | |
1241 | return vi; | |
1242 | } | |
1243 | ||
1244 | /* Return a copy of VI. When REATTACH_P is true, detach VI and attach | |
1245 | the copy. */ | |
48e1416a | 1246 | vinsn_t |
e1ab7874 | 1247 | vinsn_copy (vinsn_t vi, bool reattach_p) |
1248 | { | |
1249 | rtx copy; | |
1250 | bool unique = VINSN_UNIQUE_P (vi); | |
1251 | vinsn_t new_vi; | |
48e1416a | 1252 | |
e1ab7874 | 1253 | copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); |
1254 | new_vi = create_vinsn_from_insn_rtx (copy, unique); | |
1255 | if (reattach_p) | |
1256 | { | |
1257 | vinsn_detach (vi); | |
1258 | vinsn_attach (new_vi); | |
1259 | } | |
1260 | ||
1261 | return new_vi; | |
1262 | } | |
1263 | ||
1264 | /* Delete the VI vinsn and free its data. */ | |
1265 | static void | |
1266 | vinsn_delete (vinsn_t vi) | |
1267 | { | |
1268 | gcc_assert (VINSN_COUNT (vi) == 0); | |
1269 | ||
bc9cb5ed | 1270 | if (!INSN_NOP_P (VINSN_INSN_RTX (vi))) |
1271 | { | |
1272 | return_regset_to_pool (VINSN_REG_SETS (vi)); | |
1273 | return_regset_to_pool (VINSN_REG_USES (vi)); | |
1274 | return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); | |
1275 | } | |
e1ab7874 | 1276 | |
1277 | free (vi); | |
1278 | } | |
1279 | ||
48e1416a | 1280 | /* Indicate that VI is no longer a part of some rtx object. |
e1ab7874 | 1281 | Remove VI if it is no longer needed. */ |
1282 | void | |
1283 | vinsn_detach (vinsn_t vi) | |
1284 | { | |
1285 | gcc_assert (VINSN_COUNT (vi) > 0); | |
1286 | ||
1287 | if (--VINSN_COUNT (vi) == 0) | |
1288 | vinsn_delete (vi); | |
1289 | } | |
1290 | ||
1291 | /* Returns TRUE if VI is a branch. */ | |
1292 | bool | |
1293 | vinsn_cond_branch_p (vinsn_t vi) | |
1294 | { | |
1295 | insn_t insn; | |
1296 | ||
1297 | if (!VINSN_UNIQUE_P (vi)) | |
1298 | return false; | |
1299 | ||
1300 | insn = VINSN_INSN_RTX (vi); | |
1301 | if (BB_END (BLOCK_FOR_INSN (insn)) != insn) | |
1302 | return false; | |
1303 | ||
1304 | return control_flow_insn_p (insn); | |
1305 | } | |
1306 | ||
1307 | /* Return latency of INSN. */ | |
1308 | static int | |
1309 | sel_insn_rtx_cost (rtx insn) | |
1310 | { | |
1311 | int cost; | |
1312 | ||
1313 | /* A USE insn, or something else we don't need to | |
1314 | understand. We can't pass these directly to | |
1315 | result_ready_cost or insn_default_latency because it will | |
1316 | trigger a fatal error for unrecognizable insns. */ | |
1317 | if (recog_memoized (insn) < 0) | |
1318 | cost = 0; | |
1319 | else | |
1320 | { | |
1321 | cost = insn_default_latency (insn); | |
1322 | ||
1323 | if (cost < 0) | |
1324 | cost = 0; | |
1325 | } | |
1326 | ||
1327 | return cost; | |
1328 | } | |
1329 | ||
1330 | /* Return the cost of the VI. | |
1331 | !!! FIXME: Unify with haifa-sched.c: insn_cost (). */ | |
1332 | int | |
1333 | sel_vinsn_cost (vinsn_t vi) | |
1334 | { | |
1335 | int cost = vi->cost; | |
1336 | ||
1337 | if (cost < 0) | |
1338 | { | |
1339 | cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); | |
1340 | vi->cost = cost; | |
1341 | } | |
1342 | ||
1343 | return cost; | |
1344 | } | |
1345 | \f | |
1346 | ||
1347 | /* Functions for insn emitting. */ | |
1348 | ||
1349 | /* Emit new insn after AFTER based on PATTERN and initialize its data from | |
1350 | EXPR and SEQNO. */ | |
1351 | insn_t | |
1352 | sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) | |
1353 | { | |
1354 | insn_t new_insn; | |
1355 | ||
1356 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); | |
1357 | ||
1358 | new_insn = emit_insn_after (pattern, after); | |
1359 | set_insn_init (expr, NULL, seqno); | |
1360 | sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); | |
1361 | ||
1362 | return new_insn; | |
1363 | } | |
1364 | ||
1365 | /* Force newly generated vinsns to be unique. */ | |
1366 | static bool init_insn_force_unique_p = false; | |
1367 | ||
1368 | /* Emit new speculation recovery insn after AFTER based on PATTERN and | |
1369 | initialize its data from EXPR and SEQNO. */ | |
1370 | insn_t | |
1371 | sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, | |
1372 | insn_t after) | |
1373 | { | |
1374 | insn_t insn; | |
1375 | ||
1376 | gcc_assert (!init_insn_force_unique_p); | |
1377 | ||
1378 | init_insn_force_unique_p = true; | |
1379 | insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); | |
1380 | CANT_MOVE (insn) = 1; | |
1381 | init_insn_force_unique_p = false; | |
1382 | ||
1383 | return insn; | |
1384 | } | |
1385 | ||
1386 | /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, | |
48e1416a | 1387 | take it as a new vinsn instead of EXPR's vinsn. |
1388 | We simplify insns later, after scheduling region in | |
e1ab7874 | 1389 | simplify_changed_insns. */ |
1390 | insn_t | |
48e1416a | 1391 | sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
e1ab7874 | 1392 | insn_t after) |
1393 | { | |
1394 | expr_t emit_expr; | |
1395 | insn_t insn; | |
1396 | int flags; | |
48e1416a | 1397 | |
1398 | emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), | |
e1ab7874 | 1399 | seqno); |
1400 | insn = EXPR_INSN_RTX (emit_expr); | |
48e1416a | 1401 | add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); |
e1ab7874 | 1402 | |
1403 | flags = INSN_INIT_TODO_SSID; | |
1404 | if (INSN_LUID (insn) == 0) | |
1405 | flags |= INSN_INIT_TODO_LUID; | |
1406 | sel_init_new_insn (insn, flags); | |
1407 | ||
1408 | return insn; | |
1409 | } | |
1410 | ||
1411 | /* Move insn from EXPR after AFTER. */ | |
1412 | insn_t | |
1413 | sel_move_insn (expr_t expr, int seqno, insn_t after) | |
1414 | { | |
1415 | insn_t insn = EXPR_INSN_RTX (expr); | |
1416 | basic_block bb = BLOCK_FOR_INSN (after); | |
1417 | insn_t next = NEXT_INSN (after); | |
1418 | ||
1419 | /* Assert that in move_op we disconnected this insn properly. */ | |
1420 | gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); | |
1421 | PREV_INSN (insn) = after; | |
1422 | NEXT_INSN (insn) = next; | |
1423 | ||
1424 | NEXT_INSN (after) = insn; | |
1425 | PREV_INSN (next) = insn; | |
1426 | ||
1427 | /* Update links from insn to bb and vice versa. */ | |
1428 | df_insn_change_bb (insn, bb); | |
1429 | if (BB_END (bb) == after) | |
1430 | BB_END (bb) = insn; | |
48e1416a | 1431 | |
e1ab7874 | 1432 | prepare_insn_expr (insn, seqno); |
1433 | return insn; | |
1434 | } | |
1435 | ||
1436 | \f | |
1437 | /* Functions to work with right-hand sides. */ | |
1438 | ||
48e1416a | 1439 | /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector |
e1ab7874 | 1440 | VECT and return true when found. Use NEW_VINSN for comparison only when |
48e1416a | 1441 | COMPARE_VINSNS is true. Write to INDP the index on which |
1442 | the search has stopped, such that inserting the new element at INDP will | |
e1ab7874 | 1443 | retain VECT's sort order. */ |
1444 | static bool | |
f1f41a6c | 1445 | find_in_history_vect_1 (vec<expr_history_def> vect, |
48e1416a | 1446 | unsigned uid, vinsn_t new_vinsn, |
e1ab7874 | 1447 | bool compare_vinsns, int *indp) |
1448 | { | |
1449 | expr_history_def *arr; | |
f1f41a6c | 1450 | int i, j, len = vect.length (); |
e1ab7874 | 1451 | |
1452 | if (len == 0) | |
1453 | { | |
1454 | *indp = 0; | |
1455 | return false; | |
1456 | } | |
1457 | ||
f1f41a6c | 1458 | arr = vect.address (); |
e1ab7874 | 1459 | i = 0, j = len - 1; |
1460 | ||
1461 | while (i <= j) | |
1462 | { | |
1463 | unsigned auid = arr[i].uid; | |
48e1416a | 1464 | vinsn_t avinsn = arr[i].new_expr_vinsn; |
e1ab7874 | 1465 | |
1466 | if (auid == uid | |
48e1416a | 1467 | /* When undoing transformation on a bookkeeping copy, the new vinsn |
1468 | may not be exactly equal to the one that is saved in the vector. | |
e1ab7874 | 1469 | This is because the insn whose copy we're checking was possibly |
1470 | substituted itself. */ | |
48e1416a | 1471 | && (! compare_vinsns |
e1ab7874 | 1472 | || vinsn_equal_p (avinsn, new_vinsn))) |
1473 | { | |
1474 | *indp = i; | |
1475 | return true; | |
1476 | } | |
1477 | else if (auid > uid) | |
1478 | break; | |
1479 | i++; | |
1480 | } | |
1481 | ||
1482 | *indp = i; | |
1483 | return false; | |
1484 | } | |
1485 | ||
48e1416a | 1486 | /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return |
1487 | the position found or -1, if no such value is in vector. | |
e1ab7874 | 1488 | Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ |
1489 | int | |
f1f41a6c | 1490 | find_in_history_vect (vec<expr_history_def> vect, rtx insn, |
e1ab7874 | 1491 | vinsn_t new_vinsn, bool originators_p) |
1492 | { | |
1493 | int ind; | |
1494 | ||
48e1416a | 1495 | if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn, |
e1ab7874 | 1496 | false, &ind)) |
1497 | return ind; | |
1498 | ||
1499 | if (INSN_ORIGINATORS (insn) && originators_p) | |
1500 | { | |
1501 | unsigned uid; | |
1502 | bitmap_iterator bi; | |
1503 | ||
1504 | EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) | |
1505 | if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind)) | |
1506 | return ind; | |
1507 | } | |
48e1416a | 1508 | |
e1ab7874 | 1509 | return -1; |
1510 | } | |
1511 | ||
48e1416a | 1512 | /* Insert new element in a sorted history vector pointed to by PVECT, |
1513 | if it is not there already. The element is searched using | |
e1ab7874 | 1514 | UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save |
1515 | the history of a transformation. */ | |
1516 | void | |
f1f41a6c | 1517 | insert_in_history_vect (vec<expr_history_def> *pvect, |
e1ab7874 | 1518 | unsigned uid, enum local_trans_type type, |
48e1416a | 1519 | vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, |
e1ab7874 | 1520 | ds_t spec_ds) |
1521 | { | |
f1f41a6c | 1522 | vec<expr_history_def> vect = *pvect; |
e1ab7874 | 1523 | expr_history_def temp; |
1524 | bool res; | |
1525 | int ind; | |
1526 | ||
1527 | res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind); | |
1528 | ||
1529 | if (res) | |
1530 | { | |
f1f41a6c | 1531 | expr_history_def *phist = &vect[ind]; |
e1ab7874 | 1532 | |
48e1416a | 1533 | /* It is possible that speculation types of expressions that were |
e1ab7874 | 1534 | propagated through different paths will be different here. In this |
1535 | case, merge the status to get the correct check later. */ | |
1536 | if (phist->spec_ds != spec_ds) | |
1537 | phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); | |
1538 | return; | |
1539 | } | |
48e1416a | 1540 | |
e1ab7874 | 1541 | temp.uid = uid; |
1542 | temp.old_expr_vinsn = old_expr_vinsn; | |
48e1416a | 1543 | temp.new_expr_vinsn = new_expr_vinsn; |
e1ab7874 | 1544 | temp.spec_ds = spec_ds; |
1545 | temp.type = type; | |
1546 | ||
1547 | vinsn_attach (old_expr_vinsn); | |
1548 | vinsn_attach (new_expr_vinsn); | |
f1f41a6c | 1549 | vect.safe_insert (ind, temp); |
e1ab7874 | 1550 | *pvect = vect; |
1551 | } | |
1552 | ||
1553 | /* Free history vector PVECT. */ | |
1554 | static void | |
f1f41a6c | 1555 | free_history_vect (vec<expr_history_def> &pvect) |
e1ab7874 | 1556 | { |
1557 | unsigned i; | |
1558 | expr_history_def *phist; | |
1559 | ||
f1f41a6c | 1560 | if (! pvect.exists ()) |
e1ab7874 | 1561 | return; |
48e1416a | 1562 | |
f1f41a6c | 1563 | for (i = 0; pvect.iterate (i, &phist); i++) |
e1ab7874 | 1564 | { |
1565 | vinsn_detach (phist->old_expr_vinsn); | |
1566 | vinsn_detach (phist->new_expr_vinsn); | |
1567 | } | |
48e1416a | 1568 | |
f1f41a6c | 1569 | pvect.release (); |
e1ab7874 | 1570 | } |
1571 | ||
c53624fb | 1572 | /* Merge vector FROM to PVECT. */ |
1573 | static void | |
f1f41a6c | 1574 | merge_history_vect (vec<expr_history_def> *pvect, |
1575 | vec<expr_history_def> from) | |
c53624fb | 1576 | { |
1577 | expr_history_def *phist; | |
1578 | int i; | |
1579 | ||
1580 | /* We keep this vector sorted. */ | |
f1f41a6c | 1581 | for (i = 0; from.iterate (i, &phist); i++) |
c53624fb | 1582 | insert_in_history_vect (pvect, phist->uid, phist->type, |
1583 | phist->old_expr_vinsn, phist->new_expr_vinsn, | |
1584 | phist->spec_ds); | |
1585 | } | |
e1ab7874 | 1586 | |
1587 | /* Compare two vinsns as rhses if possible and as vinsns otherwise. */ | |
1588 | bool | |
1589 | vinsn_equal_p (vinsn_t x, vinsn_t y) | |
1590 | { | |
1591 | rtx_equal_p_callback_function repcf; | |
1592 | ||
1593 | if (x == y) | |
1594 | return true; | |
1595 | ||
1596 | if (VINSN_TYPE (x) != VINSN_TYPE (y)) | |
1597 | return false; | |
1598 | ||
1599 | if (VINSN_HASH (x) != VINSN_HASH (y)) | |
1600 | return false; | |
1601 | ||
1602 | repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL; | |
48e1416a | 1603 | if (VINSN_SEPARABLE_P (x)) |
e1ab7874 | 1604 | { |
1605 | /* Compare RHSes of VINSNs. */ | |
1606 | gcc_assert (VINSN_RHS (x)); | |
1607 | gcc_assert (VINSN_RHS (y)); | |
1608 | ||
1609 | return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf); | |
1610 | } | |
1611 | ||
1612 | return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); | |
1613 | } | |
1614 | \f | |
1615 | ||
1616 | /* Functions for working with expressions. */ | |
1617 | ||
1618 | /* Initialize EXPR. */ | |
1619 | static void | |
1620 | init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority, | |
1621 | int sched_times, int orig_bb_index, ds_t spec_done_ds, | |
1622 | ds_t spec_to_check_ds, int orig_sched_cycle, | |
f1f41a6c | 1623 | vec<expr_history_def> history, |
1624 | signed char target_available, | |
e1ab7874 | 1625 | bool was_substituted, bool was_renamed, bool needs_spec_check_p, |
1626 | bool cant_move) | |
1627 | { | |
1628 | vinsn_attach (vi); | |
1629 | ||
1630 | EXPR_VINSN (expr) = vi; | |
1631 | EXPR_SPEC (expr) = spec; | |
1632 | EXPR_USEFULNESS (expr) = use; | |
1633 | EXPR_PRIORITY (expr) = priority; | |
1634 | EXPR_PRIORITY_ADJ (expr) = 0; | |
1635 | EXPR_SCHED_TIMES (expr) = sched_times; | |
1636 | EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; | |
1637 | EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; | |
1638 | EXPR_SPEC_DONE_DS (expr) = spec_done_ds; | |
1639 | EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; | |
1640 | ||
f1f41a6c | 1641 | if (history.exists ()) |
e1ab7874 | 1642 | EXPR_HISTORY_OF_CHANGES (expr) = history; |
1643 | else | |
f1f41a6c | 1644 | EXPR_HISTORY_OF_CHANGES (expr).create (0); |
e1ab7874 | 1645 | |
1646 | EXPR_TARGET_AVAILABLE (expr) = target_available; | |
1647 | EXPR_WAS_SUBSTITUTED (expr) = was_substituted; | |
1648 | EXPR_WAS_RENAMED (expr) = was_renamed; | |
1649 | EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; | |
1650 | EXPR_CANT_MOVE (expr) = cant_move; | |
1651 | } | |
1652 | ||
1653 | /* Make a copy of the expr FROM into the expr TO. */ | |
1654 | void | |
1655 | copy_expr (expr_t to, expr_t from) | |
1656 | { | |
1e094109 | 1657 | vec<expr_history_def> temp = vNULL; |
e1ab7874 | 1658 | |
f1f41a6c | 1659 | if (EXPR_HISTORY_OF_CHANGES (from).exists ()) |
e1ab7874 | 1660 | { |
1661 | unsigned i; | |
1662 | expr_history_def *phist; | |
1663 | ||
f1f41a6c | 1664 | temp = EXPR_HISTORY_OF_CHANGES (from).copy (); |
48e1416a | 1665 | for (i = 0; |
f1f41a6c | 1666 | temp.iterate (i, &phist); |
e1ab7874 | 1667 | i++) |
1668 | { | |
1669 | vinsn_attach (phist->old_expr_vinsn); | |
1670 | vinsn_attach (phist->new_expr_vinsn); | |
1671 | } | |
1672 | } | |
1673 | ||
48e1416a | 1674 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), |
e1ab7874 | 1675 | EXPR_USEFULNESS (from), EXPR_PRIORITY (from), |
1676 | EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), | |
48e1416a | 1677 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), |
e1ab7874 | 1678 | EXPR_ORIG_SCHED_CYCLE (from), temp, |
48e1416a | 1679 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
e1ab7874 | 1680 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), |
1681 | EXPR_CANT_MOVE (from)); | |
1682 | } | |
1683 | ||
48e1416a | 1684 | /* Same, but the final expr will not ever be in av sets, so don't copy |
e1ab7874 | 1685 | "uninteresting" data such as bitmap cache. */ |
1686 | void | |
1687 | copy_expr_onside (expr_t to, expr_t from) | |
1688 | { | |
1689 | init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), | |
1690 | EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0, | |
f1f41a6c | 1691 | EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0, |
1e094109 | 1692 | vNULL, |
e1ab7874 | 1693 | EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), |
1694 | EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), | |
1695 | EXPR_CANT_MOVE (from)); | |
1696 | } | |
1697 | ||
1698 | /* Prepare the expr of INSN for scheduling. Used when moving insn and when | |
1699 | initializing new insns. */ | |
1700 | static void | |
1701 | prepare_insn_expr (insn_t insn, int seqno) | |
1702 | { | |
1703 | expr_t expr = INSN_EXPR (insn); | |
1704 | ds_t ds; | |
48e1416a | 1705 | |
e1ab7874 | 1706 | INSN_SEQNO (insn) = seqno; |
1707 | EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); | |
1708 | EXPR_SPEC (expr) = 0; | |
1709 | EXPR_ORIG_SCHED_CYCLE (expr) = 0; | |
1710 | EXPR_WAS_SUBSTITUTED (expr) = 0; | |
1711 | EXPR_WAS_RENAMED (expr) = 0; | |
1712 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
1713 | INSN_LIVE_VALID_P (insn) = false; | |
1714 | ||
1715 | /* ??? If this expression is speculative, make its dependence | |
1716 | as weak as possible. We can filter this expression later | |
1717 | in process_spec_exprs, because we do not distinguish | |
1718 | between the status we got during compute_av_set and the | |
1719 | existing status. To be fixed. */ | |
1720 | ds = EXPR_SPEC_DONE_DS (expr); | |
1721 | if (ds) | |
1722 | EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); | |
1723 | ||
f1f41a6c | 1724 | free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); |
e1ab7874 | 1725 | } |
1726 | ||
1727 | /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT | |
48e1416a | 1728 | is non-null when expressions are merged from different successors at |
e1ab7874 | 1729 | a split point. */ |
1730 | static void | |
1731 | update_target_availability (expr_t to, expr_t from, insn_t split_point) | |
1732 | { | |
48e1416a | 1733 | if (EXPR_TARGET_AVAILABLE (to) < 0 |
e1ab7874 | 1734 | || EXPR_TARGET_AVAILABLE (from) < 0) |
1735 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1736 | else | |
1737 | { | |
1738 | /* We try to detect the case when one of the expressions | |
1739 | can only be reached through another one. In this case, | |
1740 | we can do better. */ | |
1741 | if (split_point == NULL) | |
1742 | { | |
1743 | int toind, fromind; | |
1744 | ||
1745 | toind = EXPR_ORIG_BB_INDEX (to); | |
1746 | fromind = EXPR_ORIG_BB_INDEX (from); | |
48e1416a | 1747 | |
e1ab7874 | 1748 | if (toind && toind == fromind) |
48e1416a | 1749 | /* Do nothing -- everything is done in |
e1ab7874 | 1750 | merge_with_other_exprs. */ |
1751 | ; | |
1752 | else | |
1753 | EXPR_TARGET_AVAILABLE (to) = -1; | |
1754 | } | |
d6726470 | 1755 | else if (EXPR_TARGET_AVAILABLE (from) == 0 |
1756 | && EXPR_LHS (from) | |
1757 | && REG_P (EXPR_LHS (from)) | |
1758 | && REGNO (EXPR_LHS (to)) != REGNO (EXPR_LHS (from))) | |
1759 | EXPR_TARGET_AVAILABLE (to) = -1; | |
e1ab7874 | 1760 | else |
1761 | EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); | |
1762 | } | |
1763 | } | |
1764 | ||
1765 | /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT | |
48e1416a | 1766 | is non-null when expressions are merged from different successors at |
e1ab7874 | 1767 | a split point. */ |
1768 | static void | |
1769 | update_speculative_bits (expr_t to, expr_t from, insn_t split_point) | |
1770 | { | |
1771 | ds_t old_to_ds, old_from_ds; | |
1772 | ||
1773 | old_to_ds = EXPR_SPEC_DONE_DS (to); | |
1774 | old_from_ds = EXPR_SPEC_DONE_DS (from); | |
48e1416a | 1775 | |
e1ab7874 | 1776 | EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); |
1777 | EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); | |
1778 | EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); | |
1779 | ||
1780 | /* When merging e.g. control & data speculative exprs, or a control | |
48e1416a | 1781 | speculative with a control&data speculative one, we really have |
e1ab7874 | 1782 | to change vinsn too. Also, when speculative status is changed, |
1783 | we also need to record this as a transformation in expr's history. */ | |
1784 | if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) | |
1785 | { | |
1786 | old_to_ds = ds_get_speculation_types (old_to_ds); | |
1787 | old_from_ds = ds_get_speculation_types (old_from_ds); | |
48e1416a | 1788 | |
e1ab7874 | 1789 | if (old_to_ds != old_from_ds) |
1790 | { | |
1791 | ds_t record_ds; | |
48e1416a | 1792 | |
1793 | /* When both expressions are speculative, we need to change | |
e1ab7874 | 1794 | the vinsn first. */ |
1795 | if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) | |
1796 | { | |
1797 | int res; | |
48e1416a | 1798 | |
e1ab7874 | 1799 | res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); |
1800 | gcc_assert (res >= 0); | |
1801 | } | |
1802 | ||
1803 | if (split_point != NULL) | |
1804 | { | |
1805 | /* Record the change with proper status. */ | |
1806 | record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; | |
1807 | record_ds &= ~(old_to_ds & SPECULATIVE); | |
1808 | record_ds &= ~(old_from_ds & SPECULATIVE); | |
48e1416a | 1809 | |
1810 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), | |
1811 | INSN_UID (split_point), TRANS_SPECULATION, | |
e1ab7874 | 1812 | EXPR_VINSN (from), EXPR_VINSN (to), |
1813 | record_ds); | |
1814 | } | |
1815 | } | |
1816 | } | |
1817 | } | |
1818 | ||
1819 | ||
1820 | /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, | |
1821 | this is done along different paths. */ | |
1822 | void | |
1823 | merge_expr_data (expr_t to, expr_t from, insn_t split_point) | |
1824 | { | |
32bbc704 | 1825 | /* Choose the maximum of the specs of merged exprs. This is required |
1826 | for correctness of bookkeeping. */ | |
1827 | if (EXPR_SPEC (to) < EXPR_SPEC (from)) | |
e1ab7874 | 1828 | EXPR_SPEC (to) = EXPR_SPEC (from); |
1829 | ||
1830 | if (split_point) | |
1831 | EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); | |
1832 | else | |
48e1416a | 1833 | EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), |
e1ab7874 | 1834 | EXPR_USEFULNESS (from)); |
1835 | ||
1836 | if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) | |
1837 | EXPR_PRIORITY (to) = EXPR_PRIORITY (from); | |
1838 | ||
1839 | if (EXPR_SCHED_TIMES (to) > EXPR_SCHED_TIMES (from)) | |
1840 | EXPR_SCHED_TIMES (to) = EXPR_SCHED_TIMES (from); | |
1841 | ||
1842 | if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) | |
1843 | EXPR_ORIG_BB_INDEX (to) = 0; | |
1844 | ||
48e1416a | 1845 | EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), |
e1ab7874 | 1846 | EXPR_ORIG_SCHED_CYCLE (from)); |
1847 | ||
e1ab7874 | 1848 | EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); |
1849 | EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); | |
1850 | EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); | |
1851 | ||
c53624fb | 1852 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to), |
1853 | EXPR_HISTORY_OF_CHANGES (from)); | |
e1ab7874 | 1854 | update_target_availability (to, from, split_point); |
1855 | update_speculative_bits (to, from, split_point); | |
1856 | } | |
1857 | ||
1858 | /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal | |
48e1416a | 1859 | in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions |
e1ab7874 | 1860 | are merged from different successors at a split point. */ |
1861 | void | |
1862 | merge_expr (expr_t to, expr_t from, insn_t split_point) | |
1863 | { | |
1864 | vinsn_t to_vi = EXPR_VINSN (to); | |
1865 | vinsn_t from_vi = EXPR_VINSN (from); | |
1866 | ||
1867 | gcc_assert (vinsn_equal_p (to_vi, from_vi)); | |
1868 | ||
1869 | /* Make sure that speculative pattern is propagated into exprs that | |
1870 | have non-speculative one. This will provide us with consistent | |
1871 | speculative bits and speculative patterns inside expr. */ | |
936ab1d9 | 1872 | if ((EXPR_SPEC_DONE_DS (from) != 0 |
1873 | && EXPR_SPEC_DONE_DS (to) == 0) | |
1874 | /* Do likewise for volatile insns, so that we always retain | |
1875 | the may_trap_p bit on the resulting expression. */ | |
1876 | || (VINSN_MAY_TRAP_P (EXPR_VINSN (from)) | |
1877 | && !VINSN_MAY_TRAP_P (EXPR_VINSN (to)))) | |
e1ab7874 | 1878 | change_vinsn_in_expr (to, EXPR_VINSN (from)); |
1879 | ||
1880 | merge_expr_data (to, from, split_point); | |
1881 | gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); | |
1882 | } | |
1883 | ||
1884 | /* Clear the information of this EXPR. */ | |
1885 | void | |
1886 | clear_expr (expr_t expr) | |
1887 | { | |
48e1416a | 1888 | |
e1ab7874 | 1889 | vinsn_detach (EXPR_VINSN (expr)); |
1890 | EXPR_VINSN (expr) = NULL; | |
1891 | ||
f1f41a6c | 1892 | free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); |
e1ab7874 | 1893 | } |
1894 | ||
1895 | /* For a given LV_SET, mark EXPR having unavailable target register. */ | |
1896 | static void | |
1897 | set_unavailable_target_for_expr (expr_t expr, regset lv_set) | |
1898 | { | |
1899 | if (EXPR_SEPARABLE_P (expr)) | |
1900 | { | |
1901 | if (REG_P (EXPR_LHS (expr)) | |
1f53e226 | 1902 | && register_unavailable_p (lv_set, EXPR_LHS (expr))) |
e1ab7874 | 1903 | { |
48e1416a | 1904 | /* If it's an insn like r1 = use (r1, ...), and it exists in |
1905 | different forms in each of the av_sets being merged, we can't say | |
1906 | whether original destination register is available or not. | |
1907 | However, this still works if destination register is not used | |
e1ab7874 | 1908 | in the original expression: if the branch at which LV_SET we're |
1909 | looking here is not actually 'other branch' in sense that same | |
48e1416a | 1910 | expression is available through it (but it can't be determined |
e1ab7874 | 1911 | at computation stage because of transformations on one of the |
48e1416a | 1912 | branches), it still won't affect the availability. |
1913 | Liveness of a register somewhere on a code motion path means | |
1914 | it's either read somewhere on a codemotion path, live on | |
e1ab7874 | 1915 | 'other' branch, live at the point immediately following |
1916 | the original operation, or is read by the original operation. | |
1917 | The latter case is filtered out in the condition below. | |
1918 | It still doesn't cover the case when register is defined and used | |
1919 | somewhere within the code motion path, and in this case we could | |
1920 | miss a unifying code motion along both branches using a renamed | |
1921 | register, but it won't affect a code correctness since upon | |
1922 | an actual code motion a bookkeeping code would be generated. */ | |
1f53e226 | 1923 | if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
1924 | EXPR_LHS (expr))) | |
e1ab7874 | 1925 | EXPR_TARGET_AVAILABLE (expr) = -1; |
1926 | else | |
1927 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1928 | } | |
1929 | } | |
1930 | else | |
1931 | { | |
1932 | unsigned regno; | |
1933 | reg_set_iterator rsi; | |
48e1416a | 1934 | |
1935 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), | |
e1ab7874 | 1936 | 0, regno, rsi) |
1937 | if (bitmap_bit_p (lv_set, regno)) | |
1938 | { | |
1939 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1940 | break; | |
1941 | } | |
1942 | ||
1943 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), | |
1944 | 0, regno, rsi) | |
1945 | if (bitmap_bit_p (lv_set, regno)) | |
1946 | { | |
1947 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1948 | break; | |
1949 | } | |
1950 | } | |
1951 | } | |
1952 | ||
48e1416a | 1953 | /* Try to make EXPR speculative. Return 1 when EXPR's pattern |
e1ab7874 | 1954 | or dependence status have changed, 2 when also the target register |
1955 | became unavailable, 0 if nothing had to be changed. */ | |
1956 | int | |
1957 | speculate_expr (expr_t expr, ds_t ds) | |
1958 | { | |
1959 | int res; | |
1960 | rtx orig_insn_rtx; | |
1961 | rtx spec_pat; | |
1962 | ds_t target_ds, current_ds; | |
1963 | ||
1964 | /* Obtain the status we need to put on EXPR. */ | |
1965 | target_ds = (ds & SPECULATIVE); | |
1966 | current_ds = EXPR_SPEC_DONE_DS (expr); | |
1967 | ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); | |
1968 | ||
1969 | orig_insn_rtx = EXPR_INSN_RTX (expr); | |
1970 | ||
1971 | res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); | |
1972 | ||
1973 | switch (res) | |
1974 | { | |
1975 | case 0: | |
1976 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1977 | return current_ds != ds ? 1 : 0; | |
48e1416a | 1978 | |
e1ab7874 | 1979 | case 1: |
1980 | { | |
1981 | rtx spec_insn_rtx = create_insn_rtx_from_pattern (spec_pat, NULL_RTX); | |
1982 | vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); | |
1983 | ||
1984 | change_vinsn_in_expr (expr, spec_vinsn); | |
1985 | EXPR_SPEC_DONE_DS (expr) = ds; | |
1986 | EXPR_NEEDS_SPEC_CHECK_P (expr) = true; | |
1987 | ||
48e1416a | 1988 | /* Do not allow clobbering the address register of speculative |
e1ab7874 | 1989 | insns. */ |
1f53e226 | 1990 | if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
1991 | expr_dest_reg (expr))) | |
e1ab7874 | 1992 | { |
1993 | EXPR_TARGET_AVAILABLE (expr) = false; | |
1994 | return 2; | |
1995 | } | |
1996 | ||
1997 | return 1; | |
1998 | } | |
1999 | ||
2000 | case -1: | |
2001 | return -1; | |
2002 | ||
2003 | default: | |
2004 | gcc_unreachable (); | |
2005 | return -1; | |
2006 | } | |
2007 | } | |
2008 | ||
2009 | /* Return a destination register, if any, of EXPR. */ | |
2010 | rtx | |
2011 | expr_dest_reg (expr_t expr) | |
2012 | { | |
2013 | rtx dest = VINSN_LHS (EXPR_VINSN (expr)); | |
2014 | ||
2015 | if (dest != NULL_RTX && REG_P (dest)) | |
2016 | return dest; | |
2017 | ||
2018 | return NULL_RTX; | |
2019 | } | |
2020 | ||
2021 | /* Returns the REGNO of the R's destination. */ | |
2022 | unsigned | |
2023 | expr_dest_regno (expr_t expr) | |
2024 | { | |
2025 | rtx dest = expr_dest_reg (expr); | |
2026 | ||
2027 | gcc_assert (dest != NULL_RTX); | |
2028 | return REGNO (dest); | |
2029 | } | |
2030 | ||
48e1416a | 2031 | /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in |
e1ab7874 | 2032 | AV_SET having unavailable target register. */ |
2033 | void | |
2034 | mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) | |
2035 | { | |
2036 | expr_t expr; | |
2037 | av_set_iterator avi; | |
2038 | ||
2039 | FOR_EACH_EXPR (expr, avi, join_set) | |
2040 | if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) | |
2041 | set_unavailable_target_for_expr (expr, lv_set); | |
2042 | } | |
2043 | \f | |
2044 | ||
1f53e226 | 2045 | /* Returns true if REG (at least partially) is present in REGS. */ |
2046 | bool | |
2047 | register_unavailable_p (regset regs, rtx reg) | |
2048 | { | |
2049 | unsigned regno, end_regno; | |
2050 | ||
2051 | regno = REGNO (reg); | |
2052 | if (bitmap_bit_p (regs, regno)) | |
2053 | return true; | |
2054 | ||
2055 | end_regno = END_REGNO (reg); | |
2056 | ||
2057 | while (++regno < end_regno) | |
2058 | if (bitmap_bit_p (regs, regno)) | |
2059 | return true; | |
2060 | ||
2061 | return false; | |
2062 | } | |
2063 | ||
e1ab7874 | 2064 | /* Av set functions. */ |
2065 | ||
2066 | /* Add a new element to av set SETP. | |
2067 | Return the element added. */ | |
2068 | static av_set_t | |
2069 | av_set_add_element (av_set_t *setp) | |
2070 | { | |
2071 | /* Insert at the beginning of the list. */ | |
2072 | _list_add (setp); | |
2073 | return *setp; | |
2074 | } | |
2075 | ||
2076 | /* Add EXPR to SETP. */ | |
2077 | void | |
2078 | av_set_add (av_set_t *setp, expr_t expr) | |
2079 | { | |
2080 | av_set_t elem; | |
48e1416a | 2081 | |
e1ab7874 | 2082 | gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); |
2083 | elem = av_set_add_element (setp); | |
2084 | copy_expr (_AV_SET_EXPR (elem), expr); | |
2085 | } | |
2086 | ||
2087 | /* Same, but do not copy EXPR. */ | |
2088 | static void | |
2089 | av_set_add_nocopy (av_set_t *setp, expr_t expr) | |
2090 | { | |
2091 | av_set_t elem; | |
2092 | ||
2093 | elem = av_set_add_element (setp); | |
2094 | *_AV_SET_EXPR (elem) = *expr; | |
2095 | } | |
2096 | ||
2097 | /* Remove expr pointed to by IP from the av_set. */ | |
2098 | void | |
2099 | av_set_iter_remove (av_set_iterator *ip) | |
2100 | { | |
2101 | clear_expr (_AV_SET_EXPR (*ip->lp)); | |
2102 | _list_iter_remove (ip); | |
2103 | } | |
2104 | ||
2105 | /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the | |
2106 | sense of vinsn_equal_p function. Return NULL if no such expr is | |
2107 | in SET was found. */ | |
2108 | expr_t | |
2109 | av_set_lookup (av_set_t set, vinsn_t sought_vinsn) | |
2110 | { | |
2111 | expr_t expr; | |
2112 | av_set_iterator i; | |
2113 | ||
2114 | FOR_EACH_EXPR (expr, i, set) | |
2115 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2116 | return expr; | |
2117 | return NULL; | |
2118 | } | |
2119 | ||
2120 | /* Same, but also remove the EXPR found. */ | |
2121 | static expr_t | |
2122 | av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) | |
2123 | { | |
2124 | expr_t expr; | |
2125 | av_set_iterator i; | |
2126 | ||
2127 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2128 | if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) | |
2129 | { | |
2130 | _list_iter_remove_nofree (&i); | |
2131 | return expr; | |
2132 | } | |
2133 | return NULL; | |
2134 | } | |
2135 | ||
2136 | /* Search for an expr in SET, such that it's equivalent to EXPR in the | |
2137 | sense of vinsn_equal_p function of their vinsns, but not EXPR itself. | |
2138 | Returns NULL if no such expr is in SET was found. */ | |
2139 | static expr_t | |
2140 | av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) | |
2141 | { | |
2142 | expr_t cur_expr; | |
2143 | av_set_iterator i; | |
2144 | ||
2145 | FOR_EACH_EXPR (cur_expr, i, set) | |
2146 | { | |
2147 | if (cur_expr == expr) | |
2148 | continue; | |
2149 | if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) | |
2150 | return cur_expr; | |
2151 | } | |
2152 | ||
2153 | return NULL; | |
2154 | } | |
2155 | ||
2156 | /* If other expression is already in AVP, remove one of them. */ | |
2157 | expr_t | |
2158 | merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) | |
2159 | { | |
2160 | expr_t expr2; | |
2161 | ||
2162 | expr2 = av_set_lookup_other_equiv_expr (*avp, expr); | |
2163 | if (expr2 != NULL) | |
2164 | { | |
2165 | /* Reset target availability on merge, since taking it only from one | |
2166 | of the exprs would be controversial for different code. */ | |
2167 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2168 | EXPR_USEFULNESS (expr2) = 0; | |
2169 | ||
2170 | merge_expr (expr2, expr, NULL); | |
48e1416a | 2171 | |
e1ab7874 | 2172 | /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ |
2173 | EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; | |
48e1416a | 2174 | |
e1ab7874 | 2175 | av_set_iter_remove (ip); |
2176 | return expr2; | |
2177 | } | |
2178 | ||
2179 | return expr; | |
2180 | } | |
2181 | ||
2182 | /* Return true if there is an expr that correlates to VI in SET. */ | |
2183 | bool | |
2184 | av_set_is_in_p (av_set_t set, vinsn_t vi) | |
2185 | { | |
2186 | return av_set_lookup (set, vi) != NULL; | |
2187 | } | |
2188 | ||
2189 | /* Return a copy of SET. */ | |
2190 | av_set_t | |
2191 | av_set_copy (av_set_t set) | |
2192 | { | |
2193 | expr_t expr; | |
2194 | av_set_iterator i; | |
2195 | av_set_t res = NULL; | |
2196 | ||
2197 | FOR_EACH_EXPR (expr, i, set) | |
2198 | av_set_add (&res, expr); | |
2199 | ||
2200 | return res; | |
2201 | } | |
2202 | ||
2203 | /* Join two av sets that do not have common elements by attaching second set | |
2204 | (pointed to by FROMP) to the end of first set (TO_TAILP must point to | |
2205 | _AV_SET_NEXT of first set's last element). */ | |
2206 | static void | |
2207 | join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) | |
2208 | { | |
2209 | gcc_assert (*to_tailp == NULL); | |
2210 | *to_tailp = *fromp; | |
2211 | *fromp = NULL; | |
2212 | } | |
2213 | ||
2214 | /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set | |
2215 | pointed to by FROMP afterwards. */ | |
2216 | void | |
2217 | av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) | |
2218 | { | |
2219 | expr_t expr1; | |
2220 | av_set_iterator i; | |
2221 | ||
2222 | /* Delete from TOP all exprs, that present in FROMP. */ | |
2223 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2224 | { | |
2225 | expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1)); | |
2226 | ||
2227 | if (expr2) | |
2228 | { | |
2229 | merge_expr (expr2, expr1, insn); | |
2230 | av_set_iter_remove (&i); | |
2231 | } | |
2232 | } | |
2233 | ||
2234 | join_distinct_sets (i.lp, fromp); | |
2235 | } | |
2236 | ||
48e1416a | 2237 | /* Same as above, but also update availability of target register in |
e1ab7874 | 2238 | TOP judging by TO_LV_SET and FROM_LV_SET. */ |
2239 | void | |
2240 | av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, | |
2241 | regset from_lv_set, insn_t insn) | |
2242 | { | |
2243 | expr_t expr1; | |
2244 | av_set_iterator i; | |
2245 | av_set_t *to_tailp, in_both_set = NULL; | |
2246 | ||
2247 | /* Delete from TOP all expres, that present in FROMP. */ | |
2248 | FOR_EACH_EXPR_1 (expr1, i, top) | |
2249 | { | |
2250 | expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1)); | |
2251 | ||
2252 | if (expr2) | |
2253 | { | |
48e1416a | 2254 | /* It may be that the expressions have different destination |
e1ab7874 | 2255 | registers, in which case we need to check liveness here. */ |
2256 | if (EXPR_SEPARABLE_P (expr1)) | |
2257 | { | |
48e1416a | 2258 | int regno1 = (REG_P (EXPR_LHS (expr1)) |
e1ab7874 | 2259 | ? (int) expr_dest_regno (expr1) : -1); |
48e1416a | 2260 | int regno2 = (REG_P (EXPR_LHS (expr2)) |
e1ab7874 | 2261 | ? (int) expr_dest_regno (expr2) : -1); |
48e1416a | 2262 | |
2263 | /* ??? We don't have a way to check restrictions for | |
e1ab7874 | 2264 | *other* register on the current path, we did it only |
2265 | for the current target register. Give up. */ | |
2266 | if (regno1 != regno2) | |
2267 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2268 | } | |
2269 | else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) | |
2270 | EXPR_TARGET_AVAILABLE (expr2) = -1; | |
2271 | ||
2272 | merge_expr (expr2, expr1, insn); | |
2273 | av_set_add_nocopy (&in_both_set, expr2); | |
2274 | av_set_iter_remove (&i); | |
2275 | } | |
2276 | else | |
48e1416a | 2277 | /* EXPR1 is present in TOP, but not in FROMP. Check it on |
e1ab7874 | 2278 | FROM_LV_SET. */ |
2279 | set_unavailable_target_for_expr (expr1, from_lv_set); | |
2280 | } | |
2281 | to_tailp = i.lp; | |
2282 | ||
2283 | /* These expressions are not present in TOP. Check liveness | |
2284 | restrictions on TO_LV_SET. */ | |
2285 | FOR_EACH_EXPR (expr1, i, *fromp) | |
2286 | set_unavailable_target_for_expr (expr1, to_lv_set); | |
2287 | ||
2288 | join_distinct_sets (i.lp, &in_both_set); | |
2289 | join_distinct_sets (to_tailp, fromp); | |
2290 | } | |
2291 | ||
2292 | /* Clear av_set pointed to by SETP. */ | |
2293 | void | |
2294 | av_set_clear (av_set_t *setp) | |
2295 | { | |
2296 | expr_t expr; | |
2297 | av_set_iterator i; | |
2298 | ||
2299 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2300 | av_set_iter_remove (&i); | |
2301 | ||
2302 | gcc_assert (*setp == NULL); | |
2303 | } | |
2304 | ||
2305 | /* Leave only one non-speculative element in the SETP. */ | |
2306 | void | |
2307 | av_set_leave_one_nonspec (av_set_t *setp) | |
2308 | { | |
2309 | expr_t expr; | |
2310 | av_set_iterator i; | |
2311 | bool has_one_nonspec = false; | |
2312 | ||
48e1416a | 2313 | /* Keep all speculative exprs, and leave one non-speculative |
e1ab7874 | 2314 | (the first one). */ |
2315 | FOR_EACH_EXPR_1 (expr, i, setp) | |
2316 | { | |
2317 | if (!EXPR_SPEC_DONE_DS (expr)) | |
2318 | { | |
2319 | if (has_one_nonspec) | |
2320 | av_set_iter_remove (&i); | |
2321 | else | |
2322 | has_one_nonspec = true; | |
2323 | } | |
2324 | } | |
2325 | } | |
2326 | ||
2327 | /* Return the N'th element of the SET. */ | |
2328 | expr_t | |
2329 | av_set_element (av_set_t set, int n) | |
2330 | { | |
2331 | expr_t expr; | |
2332 | av_set_iterator i; | |
2333 | ||
2334 | FOR_EACH_EXPR (expr, i, set) | |
2335 | if (n-- == 0) | |
2336 | return expr; | |
2337 | ||
2338 | gcc_unreachable (); | |
2339 | return NULL; | |
2340 | } | |
2341 | ||
2342 | /* Deletes all expressions from AVP that are conditional branches (IFs). */ | |
2343 | void | |
2344 | av_set_substract_cond_branches (av_set_t *avp) | |
2345 | { | |
2346 | av_set_iterator i; | |
2347 | expr_t expr; | |
2348 | ||
2349 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2350 | if (vinsn_cond_branch_p (EXPR_VINSN (expr))) | |
2351 | av_set_iter_remove (&i); | |
2352 | } | |
2353 | ||
48e1416a | 2354 | /* Multiplies usefulness attribute of each member of av-set *AVP by |
e1ab7874 | 2355 | value PROB / ALL_PROB. */ |
2356 | void | |
2357 | av_set_split_usefulness (av_set_t av, int prob, int all_prob) | |
2358 | { | |
2359 | av_set_iterator i; | |
2360 | expr_t expr; | |
2361 | ||
2362 | FOR_EACH_EXPR (expr, i, av) | |
48e1416a | 2363 | EXPR_USEFULNESS (expr) = (all_prob |
e1ab7874 | 2364 | ? (EXPR_USEFULNESS (expr) * prob) / all_prob |
2365 | : 0); | |
2366 | } | |
2367 | ||
2368 | /* Leave in AVP only those expressions, which are present in AV, | |
c53624fb | 2369 | and return it, merging history expressions. */ |
e1ab7874 | 2370 | void |
c53624fb | 2371 | av_set_code_motion_filter (av_set_t *avp, av_set_t av) |
e1ab7874 | 2372 | { |
2373 | av_set_iterator i; | |
c53624fb | 2374 | expr_t expr, expr2; |
e1ab7874 | 2375 | |
2376 | FOR_EACH_EXPR_1 (expr, i, avp) | |
c53624fb | 2377 | if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL) |
e1ab7874 | 2378 | av_set_iter_remove (&i); |
c53624fb | 2379 | else |
2380 | /* When updating av sets in bookkeeping blocks, we can add more insns | |
2381 | there which will be transformed but the upper av sets will not | |
2382 | reflect those transformations. We then fail to undo those | |
2383 | when searching for such insns. So merge the history saved | |
2384 | in the av set of the block we are processing. */ | |
2385 | merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2386 | EXPR_HISTORY_OF_CHANGES (expr2)); | |
e1ab7874 | 2387 | } |
2388 | ||
2389 | \f | |
2390 | ||
2391 | /* Dependence hooks to initialize insn data. */ | |
2392 | ||
2393 | /* This is used in hooks callable from dependence analysis when initializing | |
2394 | instruction's data. */ | |
2395 | static struct | |
2396 | { | |
2397 | /* Where the dependence was found (lhs/rhs). */ | |
2398 | deps_where_t where; | |
2399 | ||
2400 | /* The actual data object to initialize. */ | |
2401 | idata_t id; | |
2402 | ||
2403 | /* True when the insn should not be made clonable. */ | |
2404 | bool force_unique_p; | |
2405 | ||
2406 | /* True when insn should be treated as of type USE, i.e. never renamed. */ | |
2407 | bool force_use_p; | |
2408 | } deps_init_id_data; | |
2409 | ||
2410 | ||
48e1416a | 2411 | /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be |
e1ab7874 | 2412 | clonable. */ |
2413 | static void | |
2414 | setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) | |
2415 | { | |
2416 | int type; | |
48e1416a | 2417 | |
e1ab7874 | 2418 | /* Determine whether INSN could be cloned and return appropriate vinsn type. |
2419 | That clonable insns which can be separated into lhs and rhs have type SET. | |
2420 | Other clonable insns have type USE. */ | |
2421 | type = GET_CODE (insn); | |
2422 | ||
2423 | /* Only regular insns could be cloned. */ | |
2424 | if (type == INSN && !force_unique_p) | |
2425 | type = SET; | |
2426 | else if (type == JUMP_INSN && simplejump_p (insn)) | |
2427 | type = PC; | |
9845d120 | 2428 | else if (type == DEBUG_INSN) |
2429 | type = !force_unique_p ? USE : INSN; | |
48e1416a | 2430 | |
e1ab7874 | 2431 | IDATA_TYPE (id) = type; |
2432 | IDATA_REG_SETS (id) = get_clear_regset_from_pool (); | |
2433 | IDATA_REG_USES (id) = get_clear_regset_from_pool (); | |
2434 | IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); | |
2435 | } | |
2436 | ||
2437 | /* Start initializing insn data. */ | |
2438 | static void | |
2439 | deps_init_id_start_insn (insn_t insn) | |
2440 | { | |
2441 | gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); | |
2442 | ||
2443 | setup_id_for_insn (deps_init_id_data.id, insn, | |
2444 | deps_init_id_data.force_unique_p); | |
2445 | deps_init_id_data.where = DEPS_IN_INSN; | |
2446 | } | |
2447 | ||
2448 | /* Start initializing lhs data. */ | |
2449 | static void | |
2450 | deps_init_id_start_lhs (rtx lhs) | |
2451 | { | |
2452 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2453 | gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); | |
2454 | ||
2455 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2456 | { | |
2457 | IDATA_LHS (deps_init_id_data.id) = lhs; | |
2458 | deps_init_id_data.where = DEPS_IN_LHS; | |
2459 | } | |
2460 | } | |
2461 | ||
2462 | /* Finish initializing lhs data. */ | |
2463 | static void | |
2464 | deps_init_id_finish_lhs (void) | |
2465 | { | |
2466 | deps_init_id_data.where = DEPS_IN_INSN; | |
2467 | } | |
2468 | ||
2469 | /* Note a set of REGNO. */ | |
2470 | static void | |
2471 | deps_init_id_note_reg_set (int regno) | |
2472 | { | |
2473 | haifa_note_reg_set (regno); | |
2474 | ||
2475 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2476 | deps_init_id_data.force_use_p = true; | |
2477 | ||
2478 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2479 | SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); | |
2480 | ||
2481 | #ifdef STACK_REGS | |
48e1416a | 2482 | /* Make instructions that set stack registers to be ineligible for |
e1ab7874 | 2483 | renaming to avoid issues with find_used_regs. */ |
2484 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2485 | deps_init_id_data.force_use_p = true; | |
2486 | #endif | |
2487 | } | |
2488 | ||
2489 | /* Note a clobber of REGNO. */ | |
2490 | static void | |
2491 | deps_init_id_note_reg_clobber (int regno) | |
2492 | { | |
2493 | haifa_note_reg_clobber (regno); | |
2494 | ||
2495 | if (deps_init_id_data.where == DEPS_IN_RHS) | |
2496 | deps_init_id_data.force_use_p = true; | |
2497 | ||
2498 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2499 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); | |
2500 | } | |
2501 | ||
2502 | /* Note a use of REGNO. */ | |
2503 | static void | |
2504 | deps_init_id_note_reg_use (int regno) | |
2505 | { | |
2506 | haifa_note_reg_use (regno); | |
2507 | ||
2508 | if (IDATA_TYPE (deps_init_id_data.id) != PC) | |
2509 | SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); | |
2510 | } | |
2511 | ||
2512 | /* Start initializing rhs data. */ | |
2513 | static void | |
2514 | deps_init_id_start_rhs (rtx rhs) | |
2515 | { | |
2516 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2517 | ||
2518 | /* And there was no sel_deps_reset_to_insn (). */ | |
2519 | if (IDATA_LHS (deps_init_id_data.id) != NULL) | |
2520 | { | |
2521 | IDATA_RHS (deps_init_id_data.id) = rhs; | |
2522 | deps_init_id_data.where = DEPS_IN_RHS; | |
2523 | } | |
2524 | } | |
2525 | ||
2526 | /* Finish initializing rhs data. */ | |
2527 | static void | |
2528 | deps_init_id_finish_rhs (void) | |
2529 | { | |
2530 | gcc_assert (deps_init_id_data.where == DEPS_IN_RHS | |
2531 | || deps_init_id_data.where == DEPS_IN_INSN); | |
2532 | deps_init_id_data.where = DEPS_IN_INSN; | |
2533 | } | |
2534 | ||
2535 | /* Finish initializing insn data. */ | |
2536 | static void | |
2537 | deps_init_id_finish_insn (void) | |
2538 | { | |
2539 | gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); | |
2540 | ||
2541 | if (IDATA_TYPE (deps_init_id_data.id) == SET) | |
2542 | { | |
2543 | rtx lhs = IDATA_LHS (deps_init_id_data.id); | |
2544 | rtx rhs = IDATA_RHS (deps_init_id_data.id); | |
2545 | ||
2546 | if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) | |
2547 | || deps_init_id_data.force_use_p) | |
2548 | { | |
48e1416a | 2549 | /* This should be a USE, as we don't want to schedule its RHS |
e1ab7874 | 2550 | separately. However, we still want to have them recorded |
48e1416a | 2551 | for the purposes of substitution. That's why we don't |
e1ab7874 | 2552 | simply call downgrade_to_use () here. */ |
2553 | gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); | |
2554 | gcc_assert (!lhs == !rhs); | |
2555 | ||
2556 | IDATA_TYPE (deps_init_id_data.id) = USE; | |
2557 | } | |
2558 | } | |
2559 | ||
2560 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2561 | } | |
2562 | ||
2563 | /* This is dependence info used for initializing insn's data. */ | |
2564 | static struct sched_deps_info_def deps_init_id_sched_deps_info; | |
2565 | ||
2566 | /* This initializes most of the static part of the above structure. */ | |
2567 | static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = | |
2568 | { | |
2569 | NULL, | |
2570 | ||
2571 | deps_init_id_start_insn, | |
2572 | deps_init_id_finish_insn, | |
2573 | deps_init_id_start_lhs, | |
2574 | deps_init_id_finish_lhs, | |
2575 | deps_init_id_start_rhs, | |
2576 | deps_init_id_finish_rhs, | |
2577 | deps_init_id_note_reg_set, | |
2578 | deps_init_id_note_reg_clobber, | |
2579 | deps_init_id_note_reg_use, | |
2580 | NULL, /* note_mem_dep */ | |
2581 | NULL, /* note_dep */ | |
2582 | ||
2583 | 0, /* use_cselib */ | |
2584 | 0, /* use_deps_list */ | |
2585 | 0 /* generate_spec_deps */ | |
2586 | }; | |
2587 | ||
2588 | /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, | |
2589 | we don't actually need information about lhs and rhs. */ | |
2590 | static void | |
2591 | setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) | |
2592 | { | |
2593 | rtx pat = PATTERN (insn); | |
48e1416a | 2594 | |
971ba038 | 2595 | if (NONJUMP_INSN_P (insn) |
48e1416a | 2596 | && GET_CODE (pat) == SET |
e1ab7874 | 2597 | && !force_unique_p) |
2598 | { | |
2599 | IDATA_RHS (id) = SET_SRC (pat); | |
2600 | IDATA_LHS (id) = SET_DEST (pat); | |
2601 | } | |
2602 | else | |
2603 | IDATA_LHS (id) = IDATA_RHS (id) = NULL; | |
2604 | } | |
2605 | ||
2606 | /* Possibly downgrade INSN to USE. */ | |
2607 | static void | |
2608 | maybe_downgrade_id_to_use (idata_t id, insn_t insn) | |
2609 | { | |
2610 | bool must_be_use = false; | |
2611 | unsigned uid = INSN_UID (insn); | |
ed6e85ae | 2612 | df_ref *rec; |
e1ab7874 | 2613 | rtx lhs = IDATA_LHS (id); |
2614 | rtx rhs = IDATA_RHS (id); | |
48e1416a | 2615 | |
e1ab7874 | 2616 | /* We downgrade only SETs. */ |
2617 | if (IDATA_TYPE (id) != SET) | |
2618 | return; | |
2619 | ||
2620 | if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) | |
2621 | { | |
2622 | IDATA_TYPE (id) = USE; | |
2623 | return; | |
2624 | } | |
48e1416a | 2625 | |
e1ab7874 | 2626 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2627 | { | |
ed6e85ae | 2628 | df_ref def = *rec; |
48e1416a | 2629 | |
e1ab7874 | 2630 | if (DF_REF_INSN (def) |
2631 | && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) | |
2632 | && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) | |
2633 | { | |
2634 | must_be_use = true; | |
2635 | break; | |
2636 | } | |
2637 | ||
2638 | #ifdef STACK_REGS | |
48e1416a | 2639 | /* Make instructions that set stack registers to be ineligible for |
e1ab7874 | 2640 | renaming to avoid issues with find_used_regs. */ |
2641 | if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) | |
2642 | { | |
2643 | must_be_use = true; | |
2644 | break; | |
2645 | } | |
2646 | #endif | |
48e1416a | 2647 | } |
2648 | ||
e1ab7874 | 2649 | if (must_be_use) |
2650 | IDATA_TYPE (id) = USE; | |
2651 | } | |
2652 | ||
2653 | /* Setup register sets describing INSN in ID. */ | |
2654 | static void | |
2655 | setup_id_reg_sets (idata_t id, insn_t insn) | |
2656 | { | |
2657 | unsigned uid = INSN_UID (insn); | |
ed6e85ae | 2658 | df_ref *rec; |
e1ab7874 | 2659 | regset tmp = get_clear_regset_from_pool (); |
48e1416a | 2660 | |
e1ab7874 | 2661 | for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++) |
2662 | { | |
ed6e85ae | 2663 | df_ref def = *rec; |
e1ab7874 | 2664 | unsigned int regno = DF_REF_REGNO (def); |
48e1416a | 2665 | |
e1ab7874 | 2666 | /* Post modifies are treated like clobbers by sched-deps.c. */ |
2667 | if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER | |
2668 | | DF_REF_PRE_POST_MODIFY))) | |
2669 | SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); | |
2670 | else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) | |
2671 | { | |
2672 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); | |
2673 | ||
2674 | #ifdef STACK_REGS | |
48e1416a | 2675 | /* For stack registers, treat writes to them as writes |
e1ab7874 | 2676 | to the first one to be consistent with sched-deps.c. */ |
2677 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2678 | SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); | |
2679 | #endif | |
2680 | } | |
2681 | /* Mark special refs that generate read/write def pair. */ | |
2682 | if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) | |
2683 | || regno == STACK_POINTER_REGNUM) | |
2684 | bitmap_set_bit (tmp, regno); | |
2685 | } | |
48e1416a | 2686 | |
e1ab7874 | 2687 | for (rec = DF_INSN_UID_USES (uid); *rec; rec++) |
2688 | { | |
ed6e85ae | 2689 | df_ref use = *rec; |
e1ab7874 | 2690 | unsigned int regno = DF_REF_REGNO (use); |
2691 | ||
2692 | /* When these refs are met for the first time, skip them, as | |
2693 | these uses are just counterparts of some defs. */ | |
2694 | if (bitmap_bit_p (tmp, regno)) | |
2695 | bitmap_clear_bit (tmp, regno); | |
2696 | else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) | |
2697 | { | |
2698 | SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); | |
2699 | ||
2700 | #ifdef STACK_REGS | |
48e1416a | 2701 | /* For stack registers, treat reads from them as reads from |
e1ab7874 | 2702 | the first one to be consistent with sched-deps.c. */ |
2703 | if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) | |
2704 | SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); | |
2705 | #endif | |
2706 | } | |
2707 | } | |
2708 | ||
2709 | return_regset_to_pool (tmp); | |
2710 | } | |
2711 | ||
2712 | /* Initialize instruction data for INSN in ID using DF's data. */ | |
2713 | static void | |
2714 | init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) | |
2715 | { | |
2716 | gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); | |
2717 | ||
2718 | setup_id_for_insn (id, insn, force_unique_p); | |
2719 | setup_id_lhs_rhs (id, insn, force_unique_p); | |
2720 | ||
2721 | if (INSN_NOP_P (insn)) | |
2722 | return; | |
2723 | ||
2724 | maybe_downgrade_id_to_use (id, insn); | |
2725 | setup_id_reg_sets (id, insn); | |
2726 | } | |
2727 | ||
2728 | /* Initialize instruction data for INSN in ID. */ | |
2729 | static void | |
2730 | deps_init_id (idata_t id, insn_t insn, bool force_unique_p) | |
2731 | { | |
68e419a1 | 2732 | struct deps_desc _dc, *dc = &_dc; |
e1ab7874 | 2733 | |
2734 | deps_init_id_data.where = DEPS_IN_NOWHERE; | |
2735 | deps_init_id_data.id = id; | |
2736 | deps_init_id_data.force_unique_p = force_unique_p; | |
2737 | deps_init_id_data.force_use_p = false; | |
2738 | ||
d9ab2038 | 2739 | init_deps (dc, false); |
e1ab7874 | 2740 | |
2741 | memcpy (&deps_init_id_sched_deps_info, | |
2742 | &const_deps_init_id_sched_deps_info, | |
2743 | sizeof (deps_init_id_sched_deps_info)); | |
2744 | ||
2745 | if (spec_info != NULL) | |
2746 | deps_init_id_sched_deps_info.generate_spec_deps = 1; | |
2747 | ||
2748 | sched_deps_info = &deps_init_id_sched_deps_info; | |
2749 | ||
2750 | deps_analyze_insn (dc, insn); | |
2751 | ||
2752 | free_deps (dc); | |
2753 | ||
2754 | deps_init_id_data.id = NULL; | |
2755 | } | |
2756 | ||
2757 | \f | |
52d7e28c | 2758 | struct sched_scan_info_def |
2759 | { | |
2760 | /* This hook notifies scheduler frontend to extend its internal per basic | |
2761 | block data structures. This hook should be called once before a series of | |
2762 | calls to bb_init (). */ | |
2763 | void (*extend_bb) (void); | |
2764 | ||
2765 | /* This hook makes scheduler frontend to initialize its internal data | |
2766 | structures for the passed basic block. */ | |
2767 | void (*init_bb) (basic_block); | |
2768 | ||
2769 | /* This hook notifies scheduler frontend to extend its internal per insn data | |
2770 | structures. This hook should be called once before a series of calls to | |
2771 | insn_init (). */ | |
2772 | void (*extend_insn) (void); | |
2773 | ||
2774 | /* This hook makes scheduler frontend to initialize its internal data | |
2775 | structures for the passed insn. */ | |
2776 | void (*init_insn) (rtx); | |
2777 | }; | |
2778 | ||
2779 | /* A driver function to add a set of basic blocks (BBS) to the | |
2780 | scheduling region. */ | |
2781 | static void | |
2782 | sched_scan (const struct sched_scan_info_def *ssi, bb_vec_t bbs) | |
2783 | { | |
2784 | unsigned i; | |
2785 | basic_block bb; | |
2786 | ||
2787 | if (ssi->extend_bb) | |
2788 | ssi->extend_bb (); | |
2789 | ||
2790 | if (ssi->init_bb) | |
f1f41a6c | 2791 | FOR_EACH_VEC_ELT (bbs, i, bb) |
52d7e28c | 2792 | ssi->init_bb (bb); |
2793 | ||
2794 | if (ssi->extend_insn) | |
2795 | ssi->extend_insn (); | |
2796 | ||
2797 | if (ssi->init_insn) | |
f1f41a6c | 2798 | FOR_EACH_VEC_ELT (bbs, i, bb) |
52d7e28c | 2799 | { |
2800 | rtx insn; | |
2801 | ||
2802 | FOR_BB_INSNS (bb, insn) | |
2803 | ssi->init_insn (insn); | |
2804 | } | |
2805 | } | |
e1ab7874 | 2806 | |
2807 | /* Implement hooks for collecting fundamental insn properties like if insn is | |
2808 | an ASM or is within a SCHED_GROUP. */ | |
2809 | ||
2810 | /* True when a "one-time init" data for INSN was already inited. */ | |
2811 | static bool | |
2812 | first_time_insn_init (insn_t insn) | |
2813 | { | |
2814 | return INSN_LIVE (insn) == NULL; | |
2815 | } | |
2816 | ||
2817 | /* Hash an entry in a transformed_insns hashtable. */ | |
2818 | static hashval_t | |
2819 | hash_transformed_insns (const void *p) | |
2820 | { | |
2821 | return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2822 | } | |
2823 | ||
2824 | /* Compare the entries in a transformed_insns hashtable. */ | |
2825 | static int | |
2826 | eq_transformed_insns (const void *p, const void *q) | |
2827 | { | |
2828 | rtx i1 = VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); | |
2829 | rtx i2 = VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); | |
2830 | ||
2831 | if (INSN_UID (i1) == INSN_UID (i2)) | |
2832 | return 1; | |
2833 | return rtx_equal_p (PATTERN (i1), PATTERN (i2)); | |
2834 | } | |
2835 | ||
2836 | /* Free an entry in a transformed_insns hashtable. */ | |
2837 | static void | |
2838 | free_transformed_insns (void *p) | |
2839 | { | |
2840 | struct transformed_insns *pti = (struct transformed_insns *) p; | |
2841 | ||
2842 | vinsn_detach (pti->vinsn_old); | |
2843 | vinsn_detach (pti->vinsn_new); | |
2844 | free (pti); | |
2845 | } | |
2846 | ||
48e1416a | 2847 | /* Init the s_i_d data for INSN which should be inited just once, when |
e1ab7874 | 2848 | we first see the insn. */ |
2849 | static void | |
2850 | init_first_time_insn_data (insn_t insn) | |
2851 | { | |
2852 | /* This should not be set if this is the first time we init data for | |
2853 | insn. */ | |
2854 | gcc_assert (first_time_insn_init (insn)); | |
48e1416a | 2855 | |
e1ab7874 | 2856 | /* These are needed for nops too. */ |
2857 | INSN_LIVE (insn) = get_regset_from_pool (); | |
2858 | INSN_LIVE_VALID_P (insn) = false; | |
d9ab2038 | 2859 | |
e1ab7874 | 2860 | if (!INSN_NOP_P (insn)) |
2861 | { | |
2862 | INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); | |
2863 | INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); | |
48e1416a | 2864 | INSN_TRANSFORMED_INSNS (insn) |
e1ab7874 | 2865 | = htab_create (16, hash_transformed_insns, |
2866 | eq_transformed_insns, free_transformed_insns); | |
d9ab2038 | 2867 | init_deps (&INSN_DEPS_CONTEXT (insn), true); |
e1ab7874 | 2868 | } |
2869 | } | |
2870 | ||
48e1416a | 2871 | /* Free almost all above data for INSN that is scheduled already. |
d9ab2038 | 2872 | Used for extra-large basic blocks. */ |
2873 | void | |
2874 | free_data_for_scheduled_insn (insn_t insn) | |
e1ab7874 | 2875 | { |
2876 | gcc_assert (! first_time_insn_init (insn)); | |
48e1416a | 2877 | |
d9ab2038 | 2878 | if (! INSN_ANALYZED_DEPS (insn)) |
2879 | return; | |
48e1416a | 2880 | |
e1ab7874 | 2881 | BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); |
2882 | BITMAP_FREE (INSN_FOUND_DEPS (insn)); | |
2883 | htab_delete (INSN_TRANSFORMED_INSNS (insn)); | |
48e1416a | 2884 | |
e1ab7874 | 2885 | /* This is allocated only for bookkeeping insns. */ |
2886 | if (INSN_ORIGINATORS (insn)) | |
2887 | BITMAP_FREE (INSN_ORIGINATORS (insn)); | |
2888 | free_deps (&INSN_DEPS_CONTEXT (insn)); | |
d9ab2038 | 2889 | |
2890 | INSN_ANALYZED_DEPS (insn) = NULL; | |
2891 | ||
48e1416a | 2892 | /* Clear the readonly flag so we would ICE when trying to recalculate |
d9ab2038 | 2893 | the deps context (as we believe that it should not happen). */ |
2894 | (&INSN_DEPS_CONTEXT (insn))->readonly = 0; | |
2895 | } | |
2896 | ||
2897 | /* Free the same data as above for INSN. */ | |
2898 | static void | |
2899 | free_first_time_insn_data (insn_t insn) | |
2900 | { | |
2901 | gcc_assert (! first_time_insn_init (insn)); | |
2902 | ||
2903 | free_data_for_scheduled_insn (insn); | |
2904 | return_regset_to_pool (INSN_LIVE (insn)); | |
2905 | INSN_LIVE (insn) = NULL; | |
2906 | INSN_LIVE_VALID_P (insn) = false; | |
e1ab7874 | 2907 | } |
2908 | ||
2909 | /* Initialize region-scope data structures for basic blocks. */ | |
2910 | static void | |
2911 | init_global_and_expr_for_bb (basic_block bb) | |
2912 | { | |
2913 | if (sel_bb_empty_p (bb)) | |
2914 | return; | |
2915 | ||
2916 | invalidate_av_set (bb); | |
2917 | } | |
2918 | ||
2919 | /* Data for global dependency analysis (to initialize CANT_MOVE and | |
2920 | SCHED_GROUP_P). */ | |
2921 | static struct | |
2922 | { | |
2923 | /* Previous insn. */ | |
2924 | insn_t prev_insn; | |
2925 | } init_global_data; | |
2926 | ||
2927 | /* Determine if INSN is in the sched_group, is an asm or should not be | |
2928 | cloned. After that initialize its expr. */ | |
2929 | static void | |
2930 | init_global_and_expr_for_insn (insn_t insn) | |
2931 | { | |
2932 | if (LABEL_P (insn)) | |
2933 | return; | |
2934 | ||
2935 | if (NOTE_INSN_BASIC_BLOCK_P (insn)) | |
2936 | { | |
2937 | init_global_data.prev_insn = NULL_RTX; | |
2938 | return; | |
2939 | } | |
2940 | ||
2941 | gcc_assert (INSN_P (insn)); | |
2942 | ||
2943 | if (SCHED_GROUP_P (insn)) | |
2944 | /* Setup a sched_group. */ | |
2945 | { | |
2946 | insn_t prev_insn = init_global_data.prev_insn; | |
2947 | ||
2948 | if (prev_insn) | |
2949 | INSN_SCHED_NEXT (prev_insn) = insn; | |
2950 | ||
2951 | init_global_data.prev_insn = insn; | |
2952 | } | |
2953 | else | |
2954 | init_global_data.prev_insn = NULL_RTX; | |
2955 | ||
2956 | if (GET_CODE (PATTERN (insn)) == ASM_INPUT | |
2957 | || asm_noperands (PATTERN (insn)) >= 0) | |
2958 | /* Mark INSN as an asm. */ | |
2959 | INSN_ASM_P (insn) = true; | |
2960 | ||
2961 | { | |
2962 | bool force_unique_p; | |
2963 | ds_t spec_done_ds; | |
2964 | ||
982b0787 | 2965 | /* Certain instructions cannot be cloned, and frame related insns and |
2966 | the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of | |
2967 | their block. */ | |
2968 | if (prologue_epilogue_contains (insn)) | |
2969 | { | |
2970 | if (RTX_FRAME_RELATED_P (insn)) | |
2971 | CANT_MOVE (insn) = 1; | |
2972 | else | |
2973 | { | |
2974 | rtx note; | |
2975 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
2976 | if (REG_NOTE_KIND (note) == REG_SAVE_NOTE | |
2977 | && ((enum insn_note) INTVAL (XEXP (note, 0)) | |
2978 | == NOTE_INSN_EPILOGUE_BEG)) | |
2979 | { | |
2980 | CANT_MOVE (insn) = 1; | |
2981 | break; | |
2982 | } | |
2983 | } | |
2984 | force_unique_p = true; | |
2985 | } | |
e1ab7874 | 2986 | else |
982b0787 | 2987 | if (CANT_MOVE (insn) |
2988 | || INSN_ASM_P (insn) | |
2989 | || SCHED_GROUP_P (insn) | |
a8d6ade3 | 2990 | || CALL_P (insn) |
982b0787 | 2991 | /* Exception handling insns are always unique. */ |
2992 | || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) | |
2993 | /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ | |
13434dcb | 2994 | || control_flow_insn_p (insn) |
2995 | || volatile_insn_p (PATTERN (insn)) | |
2996 | || (targetm.cannot_copy_insn_p | |
2997 | && targetm.cannot_copy_insn_p (insn))) | |
982b0787 | 2998 | force_unique_p = true; |
2999 | else | |
3000 | force_unique_p = false; | |
e1ab7874 | 3001 | |
3002 | if (targetm.sched.get_insn_spec_ds) | |
3003 | { | |
3004 | spec_done_ds = targetm.sched.get_insn_spec_ds (insn); | |
3005 | spec_done_ds = ds_get_max_dep_weak (spec_done_ds); | |
3006 | } | |
3007 | else | |
3008 | spec_done_ds = 0; | |
3009 | ||
3010 | /* Initialize INSN's expr. */ | |
3011 | init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0, | |
3012 | REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn), | |
1e094109 | 3013 | spec_done_ds, 0, 0, vNULL, true, |
f1f41a6c | 3014 | false, false, false, CANT_MOVE (insn)); |
e1ab7874 | 3015 | } |
3016 | ||
3017 | init_first_time_insn_data (insn); | |
3018 | } | |
3019 | ||
3020 | /* Scan the region and initialize instruction data for basic blocks BBS. */ | |
3021 | void | |
3022 | sel_init_global_and_expr (bb_vec_t bbs) | |
3023 | { | |
3024 | /* ??? It would be nice to implement push / pop scheme for sched_infos. */ | |
3025 | const struct sched_scan_info_def ssi = | |
3026 | { | |
3027 | NULL, /* extend_bb */ | |
3028 | init_global_and_expr_for_bb, /* init_bb */ | |
3029 | extend_insn_data, /* extend_insn */ | |
3030 | init_global_and_expr_for_insn /* init_insn */ | |
3031 | }; | |
48e1416a | 3032 | |
52d7e28c | 3033 | sched_scan (&ssi, bbs); |
e1ab7874 | 3034 | } |
3035 | ||
3036 | /* Finalize region-scope data structures for basic blocks. */ | |
3037 | static void | |
3038 | finish_global_and_expr_for_bb (basic_block bb) | |
3039 | { | |
3040 | av_set_clear (&BB_AV_SET (bb)); | |
3041 | BB_AV_LEVEL (bb) = 0; | |
3042 | } | |
3043 | ||
3044 | /* Finalize INSN's data. */ | |
3045 | static void | |
3046 | finish_global_and_expr_insn (insn_t insn) | |
3047 | { | |
3048 | if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) | |
3049 | return; | |
3050 | ||
3051 | gcc_assert (INSN_P (insn)); | |
3052 | ||
3053 | if (INSN_LUID (insn) > 0) | |
3054 | { | |
3055 | free_first_time_insn_data (insn); | |
3056 | INSN_WS_LEVEL (insn) = 0; | |
3057 | CANT_MOVE (insn) = 0; | |
48e1416a | 3058 | |
3059 | /* We can no longer assert this, as vinsns of this insn could be | |
3060 | easily live in other insn's caches. This should be changed to | |
e1ab7874 | 3061 | a counter-like approach among all vinsns. */ |
3062 | gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); | |
3063 | clear_expr (INSN_EXPR (insn)); | |
3064 | } | |
3065 | } | |
3066 | ||
3067 | /* Finalize per instruction data for the whole region. */ | |
3068 | void | |
3069 | sel_finish_global_and_expr (void) | |
3070 | { | |
3071 | { | |
3072 | bb_vec_t bbs; | |
3073 | int i; | |
3074 | ||
f1f41a6c | 3075 | bbs.create (current_nr_blocks); |
e1ab7874 | 3076 | |
3077 | for (i = 0; i < current_nr_blocks; i++) | |
f5a6b05f | 3078 | bbs.quick_push (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))); |
e1ab7874 | 3079 | |
3080 | /* Clear AV_SETs and INSN_EXPRs. */ | |
3081 | { | |
3082 | const struct sched_scan_info_def ssi = | |
3083 | { | |
3084 | NULL, /* extend_bb */ | |
3085 | finish_global_and_expr_for_bb, /* init_bb */ | |
3086 | NULL, /* extend_insn */ | |
3087 | finish_global_and_expr_insn /* init_insn */ | |
3088 | }; | |
3089 | ||
52d7e28c | 3090 | sched_scan (&ssi, bbs); |
e1ab7874 | 3091 | } |
3092 | ||
f1f41a6c | 3093 | bbs.release (); |
e1ab7874 | 3094 | } |
3095 | ||
3096 | finish_insns (); | |
3097 | } | |
3098 | \f | |
3099 | ||
48e1416a | 3100 | /* In the below hooks, we merely calculate whether or not a dependence |
3101 | exists, and in what part of insn. However, we will need more data | |
e1ab7874 | 3102 | when we'll start caching dependence requests. */ |
3103 | ||
3104 | /* Container to hold information for dependency analysis. */ | |
3105 | static struct | |
3106 | { | |
3107 | deps_t dc; | |
3108 | ||
3109 | /* A variable to track which part of rtx we are scanning in | |
3110 | sched-deps.c: sched_analyze_insn (). */ | |
3111 | deps_where_t where; | |
3112 | ||
3113 | /* Current producer. */ | |
3114 | insn_t pro; | |
3115 | ||
3116 | /* Current consumer. */ | |
3117 | vinsn_t con; | |
3118 | ||
3119 | /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. | |
3120 | X is from { INSN, LHS, RHS }. */ | |
3121 | ds_t has_dep_p[DEPS_IN_NOWHERE]; | |
3122 | } has_dependence_data; | |
3123 | ||
3124 | /* Start analyzing dependencies of INSN. */ | |
3125 | static void | |
3126 | has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) | |
3127 | { | |
3128 | gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); | |
3129 | ||
3130 | has_dependence_data.where = DEPS_IN_INSN; | |
3131 | } | |
3132 | ||
3133 | /* Finish analyzing dependencies of an insn. */ | |
3134 | static void | |
3135 | has_dependence_finish_insn (void) | |
3136 | { | |
3137 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3138 | ||
3139 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3140 | } | |
3141 | ||
3142 | /* Start analyzing dependencies of LHS. */ | |
3143 | static void | |
3144 | has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) | |
3145 | { | |
3146 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3147 | ||
3148 | if (VINSN_LHS (has_dependence_data.con) != NULL) | |
3149 | has_dependence_data.where = DEPS_IN_LHS; | |
3150 | } | |
3151 | ||
3152 | /* Finish analyzing dependencies of an lhs. */ | |
3153 | static void | |
3154 | has_dependence_finish_lhs (void) | |
3155 | { | |
3156 | has_dependence_data.where = DEPS_IN_INSN; | |
3157 | } | |
3158 | ||
3159 | /* Start analyzing dependencies of RHS. */ | |
3160 | static void | |
3161 | has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) | |
3162 | { | |
3163 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3164 | ||
3165 | if (VINSN_RHS (has_dependence_data.con) != NULL) | |
3166 | has_dependence_data.where = DEPS_IN_RHS; | |
3167 | } | |
3168 | ||
3169 | /* Start analyzing dependencies of an rhs. */ | |
3170 | static void | |
3171 | has_dependence_finish_rhs (void) | |
3172 | { | |
3173 | gcc_assert (has_dependence_data.where == DEPS_IN_RHS | |
3174 | || has_dependence_data.where == DEPS_IN_INSN); | |
3175 | ||
3176 | has_dependence_data.where = DEPS_IN_INSN; | |
3177 | } | |
3178 | ||
3179 | /* Note a set of REGNO. */ | |
3180 | static void | |
3181 | has_dependence_note_reg_set (int regno) | |
3182 | { | |
3183 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3184 | ||
3185 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3186 | VINSN_INSN_RTX | |
3187 | (has_dependence_data.con))) | |
3188 | { | |
3189 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3190 | ||
3191 | if (reg_last->sets != NULL | |
3192 | || reg_last->clobbers != NULL) | |
3193 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
3194 | ||
a9bfd373 | 3195 | if (reg_last->uses || reg_last->implicit_sets) |
e1ab7874 | 3196 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3197 | } | |
3198 | } | |
3199 | ||
3200 | /* Note a clobber of REGNO. */ | |
3201 | static void | |
3202 | has_dependence_note_reg_clobber (int regno) | |
3203 | { | |
3204 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3205 | ||
3206 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3207 | VINSN_INSN_RTX | |
3208 | (has_dependence_data.con))) | |
3209 | { | |
3210 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3211 | ||
3212 | if (reg_last->sets) | |
3213 | *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; | |
48e1416a | 3214 | |
a9bfd373 | 3215 | if (reg_last->uses || reg_last->implicit_sets) |
e1ab7874 | 3216 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3217 | } | |
3218 | } | |
3219 | ||
3220 | /* Note a use of REGNO. */ | |
3221 | static void | |
3222 | has_dependence_note_reg_use (int regno) | |
3223 | { | |
3224 | struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; | |
3225 | ||
3226 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3227 | VINSN_INSN_RTX | |
3228 | (has_dependence_data.con))) | |
3229 | { | |
3230 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3231 | ||
3232 | if (reg_last->sets) | |
3233 | *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; | |
3234 | ||
a9bfd373 | 3235 | if (reg_last->clobbers || reg_last->implicit_sets) |
e1ab7874 | 3236 | *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; |
3237 | ||
b0691607 | 3238 | /* Merge BE_IN_SPEC bits into *DSP when the dependency producer |
3239 | is actually a check insn. We need to do this for any register | |
3240 | read-read dependency with the check unless we track properly | |
3241 | all registers written by BE_IN_SPEC-speculated insns, as | |
3242 | we don't have explicit dependence lists. See PR 53975. */ | |
e1ab7874 | 3243 | if (reg_last->uses) |
3244 | { | |
3245 | ds_t pro_spec_checked_ds; | |
3246 | ||
3247 | pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); | |
3248 | pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); | |
3249 | ||
b0691607 | 3250 | if (pro_spec_checked_ds != 0) |
e1ab7874 | 3251 | *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, |
3252 | NULL_RTX, NULL_RTX); | |
3253 | } | |
3254 | } | |
3255 | } | |
3256 | ||
3257 | /* Note a memory dependence. */ | |
3258 | static void | |
3259 | has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, | |
3260 | rtx pending_mem ATTRIBUTE_UNUSED, | |
3261 | insn_t pending_insn ATTRIBUTE_UNUSED, | |
3262 | ds_t ds ATTRIBUTE_UNUSED) | |
3263 | { | |
3264 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3265 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3266 | { | |
3267 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3268 | ||
3269 | *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); | |
3270 | } | |
3271 | } | |
3272 | ||
3273 | /* Note a dependence. */ | |
3274 | static void | |
3275 | has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED, | |
3276 | ds_t ds ATTRIBUTE_UNUSED) | |
3277 | { | |
3278 | if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, | |
3279 | VINSN_INSN_RTX (has_dependence_data.con))) | |
3280 | { | |
3281 | ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; | |
3282 | ||
3283 | *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); | |
3284 | } | |
3285 | } | |
3286 | ||
3287 | /* Mark the insn as having a hard dependence that prevents speculation. */ | |
3288 | void | |
3289 | sel_mark_hard_insn (rtx insn) | |
3290 | { | |
3291 | int i; | |
3292 | ||
3293 | /* Only work when we're in has_dependence_p mode. | |
3294 | ??? This is a hack, this should actually be a hook. */ | |
3295 | if (!has_dependence_data.dc || !has_dependence_data.pro) | |
3296 | return; | |
3297 | ||
3298 | gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); | |
3299 | gcc_assert (has_dependence_data.where == DEPS_IN_INSN); | |
3300 | ||
3301 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3302 | has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; | |
3303 | } | |
3304 | ||
3305 | /* This structure holds the hooks for the dependency analysis used when | |
3306 | actually processing dependencies in the scheduler. */ | |
3307 | static struct sched_deps_info_def has_dependence_sched_deps_info; | |
3308 | ||
3309 | /* This initializes most of the fields of the above structure. */ | |
3310 | static const struct sched_deps_info_def const_has_dependence_sched_deps_info = | |
3311 | { | |
3312 | NULL, | |
3313 | ||
3314 | has_dependence_start_insn, | |
3315 | has_dependence_finish_insn, | |
3316 | has_dependence_start_lhs, | |
3317 | has_dependence_finish_lhs, | |
3318 | has_dependence_start_rhs, | |
3319 | has_dependence_finish_rhs, | |
3320 | has_dependence_note_reg_set, | |
3321 | has_dependence_note_reg_clobber, | |
3322 | has_dependence_note_reg_use, | |
3323 | has_dependence_note_mem_dep, | |
3324 | has_dependence_note_dep, | |
3325 | ||
3326 | 0, /* use_cselib */ | |
3327 | 0, /* use_deps_list */ | |
3328 | 0 /* generate_spec_deps */ | |
3329 | }; | |
3330 | ||
3331 | /* Initialize has_dependence_sched_deps_info with extra spec field. */ | |
3332 | static void | |
3333 | setup_has_dependence_sched_deps_info (void) | |
3334 | { | |
3335 | memcpy (&has_dependence_sched_deps_info, | |
3336 | &const_has_dependence_sched_deps_info, | |
3337 | sizeof (has_dependence_sched_deps_info)); | |
3338 | ||
3339 | if (spec_info != NULL) | |
3340 | has_dependence_sched_deps_info.generate_spec_deps = 1; | |
3341 | ||
3342 | sched_deps_info = &has_dependence_sched_deps_info; | |
3343 | } | |
3344 | ||
3345 | /* Remove all dependences found and recorded in has_dependence_data array. */ | |
3346 | void | |
3347 | sel_clear_has_dependence (void) | |
3348 | { | |
3349 | int i; | |
3350 | ||
3351 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3352 | has_dependence_data.has_dep_p[i] = 0; | |
3353 | } | |
3354 | ||
3355 | /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer | |
3356 | to the dependence information array in HAS_DEP_PP. */ | |
3357 | ds_t | |
3358 | has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) | |
3359 | { | |
3360 | int i; | |
3361 | ds_t ds; | |
68e419a1 | 3362 | struct deps_desc *dc; |
e1ab7874 | 3363 | |
3364 | if (INSN_SIMPLEJUMP_P (pred)) | |
3365 | /* Unconditional jump is just a transfer of control flow. | |
3366 | Ignore it. */ | |
3367 | return false; | |
3368 | ||
3369 | dc = &INSN_DEPS_CONTEXT (pred); | |
d9ab2038 | 3370 | |
3371 | /* We init this field lazily. */ | |
3372 | if (dc->reg_last == NULL) | |
3373 | init_deps_reg_last (dc); | |
48e1416a | 3374 | |
e1ab7874 | 3375 | if (!dc->readonly) |
3376 | { | |
3377 | has_dependence_data.pro = NULL; | |
3378 | /* Initialize empty dep context with information about PRED. */ | |
3379 | advance_deps_context (dc, pred); | |
3380 | dc->readonly = 1; | |
3381 | } | |
3382 | ||
3383 | has_dependence_data.where = DEPS_IN_NOWHERE; | |
3384 | has_dependence_data.pro = pred; | |
3385 | has_dependence_data.con = EXPR_VINSN (expr); | |
3386 | has_dependence_data.dc = dc; | |
3387 | ||
3388 | sel_clear_has_dependence (); | |
3389 | ||
3390 | /* Now catch all dependencies that would be generated between PRED and | |
3391 | INSN. */ | |
3392 | setup_has_dependence_sched_deps_info (); | |
3393 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3394 | has_dependence_data.dc = NULL; | |
3395 | ||
3396 | /* When a barrier was found, set DEPS_IN_INSN bits. */ | |
3397 | if (dc->last_reg_pending_barrier == TRUE_BARRIER) | |
3398 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; | |
3399 | else if (dc->last_reg_pending_barrier == MOVE_BARRIER) | |
3400 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
3401 | ||
3402 | /* Do not allow stores to memory to move through checks. Currently | |
3403 | we don't move this to sched-deps.c as the check doesn't have | |
48e1416a | 3404 | obvious places to which this dependence can be attached. |
e1ab7874 | 3405 | FIMXE: this should go to a hook. */ |
3406 | if (EXPR_LHS (expr) | |
3407 | && MEM_P (EXPR_LHS (expr)) | |
3408 | && sel_insn_is_speculation_check (pred)) | |
3409 | has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; | |
48e1416a | 3410 | |
e1ab7874 | 3411 | *has_dep_pp = has_dependence_data.has_dep_p; |
3412 | ds = 0; | |
3413 | for (i = 0; i < DEPS_IN_NOWHERE; i++) | |
3414 | ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], | |
3415 | NULL_RTX, NULL_RTX); | |
3416 | ||
3417 | return ds; | |
3418 | } | |
3419 | \f | |
3420 | ||
48e1416a | 3421 | /* Dependence hooks implementation that checks dependence latency constraints |
3422 | on the insns being scheduled. The entry point for these routines is | |
3423 | tick_check_p predicate. */ | |
e1ab7874 | 3424 | |
3425 | static struct | |
3426 | { | |
3427 | /* An expr we are currently checking. */ | |
3428 | expr_t expr; | |
3429 | ||
3430 | /* A minimal cycle for its scheduling. */ | |
3431 | int cycle; | |
3432 | ||
3433 | /* Whether we have seen a true dependence while checking. */ | |
3434 | bool seen_true_dep_p; | |
3435 | } tick_check_data; | |
3436 | ||
3437 | /* Update minimal scheduling cycle for tick_check_insn given that it depends | |
3438 | on PRO with status DS and weight DW. */ | |
3439 | static void | |
3440 | tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) | |
3441 | { | |
3442 | expr_t con_expr = tick_check_data.expr; | |
3443 | insn_t con_insn = EXPR_INSN_RTX (con_expr); | |
3444 | ||
3445 | if (con_insn != pro_insn) | |
3446 | { | |
3447 | enum reg_note dt; | |
3448 | int tick; | |
3449 | ||
3450 | if (/* PROducer was removed from above due to pipelining. */ | |
3451 | !INSN_IN_STREAM_P (pro_insn) | |
3452 | /* Or PROducer was originally on the next iteration regarding the | |
3453 | CONsumer. */ | |
3454 | || (INSN_SCHED_TIMES (pro_insn) | |
3455 | - EXPR_SCHED_TIMES (con_expr)) > 1) | |
3456 | /* Don't count this dependence. */ | |
3457 | return; | |
3458 | ||
3459 | dt = ds_to_dt (ds); | |
3460 | if (dt == REG_DEP_TRUE) | |
3461 | tick_check_data.seen_true_dep_p = true; | |
3462 | ||
3463 | gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); | |
3464 | ||
3465 | { | |
3466 | dep_def _dep, *dep = &_dep; | |
3467 | ||
3468 | init_dep (dep, pro_insn, con_insn, dt); | |
3469 | ||
3470 | tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); | |
3471 | } | |
3472 | ||
3473 | /* When there are several kinds of dependencies between pro and con, | |
3474 | only REG_DEP_TRUE should be taken into account. */ | |
3475 | if (tick > tick_check_data.cycle | |
3476 | && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) | |
3477 | tick_check_data.cycle = tick; | |
3478 | } | |
3479 | } | |
3480 | ||
3481 | /* An implementation of note_dep hook. */ | |
3482 | static void | |
3483 | tick_check_note_dep (insn_t pro, ds_t ds) | |
3484 | { | |
3485 | tick_check_dep_with_dw (pro, ds, 0); | |
3486 | } | |
3487 | ||
3488 | /* An implementation of note_mem_dep hook. */ | |
3489 | static void | |
3490 | tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) | |
3491 | { | |
3492 | dw_t dw; | |
3493 | ||
3494 | dw = (ds_to_dt (ds) == REG_DEP_TRUE | |
3495 | ? estimate_dep_weak (mem1, mem2) | |
3496 | : 0); | |
3497 | ||
3498 | tick_check_dep_with_dw (pro, ds, dw); | |
3499 | } | |
3500 | ||
3501 | /* This structure contains hooks for dependence analysis used when determining | |
3502 | whether an insn is ready for scheduling. */ | |
3503 | static struct sched_deps_info_def tick_check_sched_deps_info = | |
3504 | { | |
3505 | NULL, | |
3506 | ||
3507 | NULL, | |
3508 | NULL, | |
3509 | NULL, | |
3510 | NULL, | |
3511 | NULL, | |
3512 | NULL, | |
3513 | haifa_note_reg_set, | |
3514 | haifa_note_reg_clobber, | |
3515 | haifa_note_reg_use, | |
3516 | tick_check_note_mem_dep, | |
3517 | tick_check_note_dep, | |
3518 | ||
3519 | 0, 0, 0 | |
3520 | }; | |
3521 | ||
3522 | /* Estimate number of cycles from the current cycle of FENCE until EXPR can be | |
3523 | scheduled. Return 0 if all data from producers in DC is ready. */ | |
3524 | int | |
3525 | tick_check_p (expr_t expr, deps_t dc, fence_t fence) | |
3526 | { | |
3527 | int cycles_left; | |
3528 | /* Initialize variables. */ | |
3529 | tick_check_data.expr = expr; | |
3530 | tick_check_data.cycle = 0; | |
3531 | tick_check_data.seen_true_dep_p = false; | |
3532 | sched_deps_info = &tick_check_sched_deps_info; | |
48e1416a | 3533 | |
e1ab7874 | 3534 | gcc_assert (!dc->readonly); |
3535 | dc->readonly = 1; | |
3536 | deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); | |
3537 | dc->readonly = 0; | |
3538 | ||
3539 | cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); | |
3540 | ||
3541 | return cycles_left >= 0 ? cycles_left : 0; | |
3542 | } | |
3543 | \f | |
3544 | ||
3545 | /* Functions to work with insns. */ | |
3546 | ||
3547 | /* Returns true if LHS of INSN is the same as DEST of an insn | |
3548 | being moved. */ | |
3549 | bool | |
3550 | lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) | |
3551 | { | |
3552 | rtx lhs = INSN_LHS (insn); | |
3553 | ||
3554 | if (lhs == NULL || dest == NULL) | |
3555 | return false; | |
48e1416a | 3556 | |
e1ab7874 | 3557 | return rtx_equal_p (lhs, dest); |
3558 | } | |
3559 | ||
3560 | /* Return s_i_d entry of INSN. Callable from debugger. */ | |
3561 | sel_insn_data_def | |
3562 | insn_sid (insn_t insn) | |
3563 | { | |
3564 | return *SID (insn); | |
3565 | } | |
3566 | ||
3567 | /* True when INSN is a speculative check. We can tell this by looking | |
3568 | at the data structures of the selective scheduler, not by examining | |
3569 | the pattern. */ | |
3570 | bool | |
3571 | sel_insn_is_speculation_check (rtx insn) | |
3572 | { | |
f1f41a6c | 3573 | return s_i_d.exists () && !! INSN_SPEC_CHECKED_DS (insn); |
e1ab7874 | 3574 | } |
3575 | ||
48e1416a | 3576 | /* Extracts machine mode MODE and destination location DST_LOC |
e1ab7874 | 3577 | for given INSN. */ |
3578 | void | |
3579 | get_dest_and_mode (rtx insn, rtx *dst_loc, enum machine_mode *mode) | |
3580 | { | |
3581 | rtx pat = PATTERN (insn); | |
3582 | ||
3583 | gcc_assert (dst_loc); | |
3584 | gcc_assert (GET_CODE (pat) == SET); | |
3585 | ||
3586 | *dst_loc = SET_DEST (pat); | |
3587 | ||
3588 | gcc_assert (*dst_loc); | |
3589 | gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); | |
3590 | ||
3591 | if (mode) | |
3592 | *mode = GET_MODE (*dst_loc); | |
3593 | } | |
3594 | ||
48e1416a | 3595 | /* Returns true when moving through JUMP will result in bookkeeping |
e1ab7874 | 3596 | creation. */ |
3597 | bool | |
3598 | bookkeeping_can_be_created_if_moved_through_p (insn_t jump) | |
3599 | { | |
3600 | insn_t succ; | |
3601 | succ_iterator si; | |
3602 | ||
3603 | FOR_EACH_SUCC (succ, si, jump) | |
3604 | if (sel_num_cfg_preds_gt_1 (succ)) | |
3605 | return true; | |
3606 | ||
3607 | return false; | |
3608 | } | |
3609 | ||
3610 | /* Return 'true' if INSN is the only one in its basic block. */ | |
3611 | static bool | |
3612 | insn_is_the_only_one_in_bb_p (insn_t insn) | |
3613 | { | |
3614 | return sel_bb_head_p (insn) && sel_bb_end_p (insn); | |
3615 | } | |
3616 | ||
3617 | #ifdef ENABLE_CHECKING | |
48e1416a | 3618 | /* Check that the region we're scheduling still has at most one |
e1ab7874 | 3619 | backedge. */ |
3620 | static void | |
3621 | verify_backedges (void) | |
3622 | { | |
3623 | if (pipelining_p) | |
3624 | { | |
3625 | int i, n = 0; | |
3626 | edge e; | |
3627 | edge_iterator ei; | |
48e1416a | 3628 | |
e1ab7874 | 3629 | for (i = 0; i < current_nr_blocks; i++) |
f5a6b05f | 3630 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))->succs) |
e1ab7874 | 3631 | if (in_current_region_p (e->dest) |
3632 | && BLOCK_TO_BB (e->dest->index) < i) | |
3633 | n++; | |
48e1416a | 3634 | |
e1ab7874 | 3635 | gcc_assert (n <= 1); |
3636 | } | |
3637 | } | |
3638 | #endif | |
3639 | \f | |
3640 | ||
3641 | /* Functions to work with control flow. */ | |
3642 | ||
93919afc | 3643 | /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks |
3644 | are sorted in topological order (it might have been invalidated by | |
3645 | redirecting an edge). */ | |
3646 | static void | |
3647 | sel_recompute_toporder (void) | |
3648 | { | |
3649 | int i, n, rgn; | |
3650 | int *postorder, n_blocks; | |
3651 | ||
a28770e1 | 3652 | postorder = XALLOCAVEC (int, n_basic_blocks_for_fn (cfun)); |
93919afc | 3653 | n_blocks = post_order_compute (postorder, false, false); |
3654 | ||
3655 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
3656 | for (n = 0, i = n_blocks - 1; i >= 0; i--) | |
3657 | if (CONTAINING_RGN (postorder[i]) == rgn) | |
3658 | { | |
3659 | BLOCK_TO_BB (postorder[i]) = n; | |
3660 | BB_TO_BLOCK (n) = postorder[i]; | |
3661 | n++; | |
3662 | } | |
3663 | ||
3664 | /* Assert that we updated info for all blocks. We may miss some blocks if | |
3665 | this function is called when redirecting an edge made a block | |
3666 | unreachable, but that block is not deleted yet. */ | |
3667 | gcc_assert (n == RGN_NR_BLOCKS (rgn)); | |
3668 | } | |
3669 | ||
e1ab7874 | 3670 | /* Tidy the possibly empty block BB. */ |
81d1ad0f | 3671 | static bool |
6f0e7980 | 3672 | maybe_tidy_empty_bb (basic_block bb) |
e1ab7874 | 3673 | { |
ef4cf572 | 3674 | basic_block succ_bb, pred_bb, note_bb; |
f1f41a6c | 3675 | vec<basic_block> dom_bbs; |
df6266b9 | 3676 | edge e; |
3677 | edge_iterator ei; | |
e1ab7874 | 3678 | bool rescan_p; |
3679 | ||
3680 | /* Keep empty bb only if this block immediately precedes EXIT and | |
61e213e2 | 3681 | has incoming non-fallthrough edge, or it has no predecessors or |
3682 | successors. Otherwise remove it. */ | |
9845d120 | 3683 | if (!sel_bb_empty_p (bb) |
48e1416a | 3684 | || (single_succ_p (bb) |
34154e27 | 3685 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
48e1416a | 3686 | && (!single_pred_p (bb) |
61e213e2 | 3687 | || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU))) |
3688 | || EDGE_COUNT (bb->preds) == 0 | |
3689 | || EDGE_COUNT (bb->succs) == 0) | |
e1ab7874 | 3690 | return false; |
3691 | ||
df6266b9 | 3692 | /* Do not attempt to redirect complex edges. */ |
3693 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3694 | if (e->flags & EDGE_COMPLEX) | |
3695 | return false; | |
a62f9dca | 3696 | else if (e->flags & EDGE_FALLTHRU) |
3697 | { | |
3698 | rtx note; | |
3699 | /* If prev bb ends with asm goto, see if any of the | |
3700 | ASM_OPERANDS_LABELs don't point to the fallthru | |
3701 | label. Do not attempt to redirect it in that case. */ | |
3702 | if (JUMP_P (BB_END (e->src)) | |
3703 | && (note = extract_asm_operands (PATTERN (BB_END (e->src))))) | |
3704 | { | |
3705 | int i, n = ASM_OPERANDS_LABEL_LENGTH (note); | |
3706 | ||
3707 | for (i = 0; i < n; ++i) | |
3708 | if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (bb)) | |
3709 | return false; | |
3710 | } | |
3711 | } | |
df6266b9 | 3712 | |
e1ab7874 | 3713 | free_data_sets (bb); |
3714 | ||
3715 | /* Do not delete BB if it has more than one successor. | |
3716 | That can occur when we moving a jump. */ | |
3717 | if (!single_succ_p (bb)) | |
3718 | { | |
3719 | gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); | |
3720 | sel_merge_blocks (bb->prev_bb, bb); | |
3721 | return true; | |
3722 | } | |
3723 | ||
3724 | succ_bb = single_succ (bb); | |
3725 | rescan_p = true; | |
3726 | pred_bb = NULL; | |
f1f41a6c | 3727 | dom_bbs.create (0); |
e1ab7874 | 3728 | |
ef4cf572 | 3729 | /* Save a pred/succ from the current region to attach the notes to. */ |
3730 | note_bb = NULL; | |
3731 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3732 | if (in_current_region_p (e->src)) | |
3733 | { | |
3734 | note_bb = e->src; | |
3735 | break; | |
3736 | } | |
3737 | if (note_bb == NULL) | |
3738 | note_bb = succ_bb; | |
3739 | ||
e1ab7874 | 3740 | /* Redirect all non-fallthru edges to the next bb. */ |
3741 | while (rescan_p) | |
3742 | { | |
e1ab7874 | 3743 | rescan_p = false; |
3744 | ||
3745 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3746 | { | |
3747 | pred_bb = e->src; | |
3748 | ||
3749 | if (!(e->flags & EDGE_FALLTHRU)) | |
3750 | { | |
6f0e7980 | 3751 | /* We can not invalidate computed topological order by moving |
1a5dbaab | 3752 | the edge destination block (E->SUCC) along a fallthru edge. |
3753 | ||
3754 | We will update dominators here only when we'll get | |
3755 | an unreachable block when redirecting, otherwise | |
3756 | sel_redirect_edge_and_branch will take care of it. */ | |
3757 | if (e->dest != bb | |
3758 | && single_pred_p (e->dest)) | |
f1f41a6c | 3759 | dom_bbs.safe_push (e->dest); |
6f0e7980 | 3760 | sel_redirect_edge_and_branch (e, succ_bb); |
e1ab7874 | 3761 | rescan_p = true; |
3762 | break; | |
3763 | } | |
6f0e7980 | 3764 | /* If the edge is fallthru, but PRED_BB ends in a conditional jump |
3765 | to BB (so there is no non-fallthru edge from PRED_BB to BB), we | |
3766 | still have to adjust it. */ | |
3767 | else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb))) | |
3768 | { | |
3769 | /* If possible, try to remove the unneeded conditional jump. */ | |
3770 | if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0 | |
3771 | && !IN_CURRENT_FENCE_P (BB_END (pred_bb))) | |
3772 | { | |
3773 | if (!sel_remove_insn (BB_END (pred_bb), false, false)) | |
3774 | tidy_fallthru_edge (e); | |
3775 | } | |
3776 | else | |
3777 | sel_redirect_edge_and_branch (e, succ_bb); | |
3778 | rescan_p = true; | |
3779 | break; | |
3780 | } | |
e1ab7874 | 3781 | } |
3782 | } | |
3783 | ||
e1ab7874 | 3784 | if (can_merge_blocks_p (bb->prev_bb, bb)) |
3785 | sel_merge_blocks (bb->prev_bb, bb); | |
3786 | else | |
e1ab7874 | 3787 | { |
0424f393 | 3788 | /* This is a block without fallthru predecessor. Just delete it. */ |
ef4cf572 | 3789 | gcc_assert (note_bb); |
3790 | move_bb_info (note_bb, bb); | |
e1ab7874 | 3791 | remove_empty_bb (bb, true); |
3792 | } | |
3793 | ||
f1f41a6c | 3794 | if (!dom_bbs.is_empty ()) |
1a5dbaab | 3795 | { |
f1f41a6c | 3796 | dom_bbs.safe_push (succ_bb); |
1a5dbaab | 3797 | iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); |
f1f41a6c | 3798 | dom_bbs.release (); |
1a5dbaab | 3799 | } |
3800 | ||
e1ab7874 | 3801 | return true; |
3802 | } | |
3803 | ||
48e1416a | 3804 | /* Tidy the control flow after we have removed original insn from |
e1ab7874 | 3805 | XBB. Return true if we have removed some blocks. When FULL_TIDYING |
3806 | is true, also try to optimize control flow on non-empty blocks. */ | |
3807 | bool | |
3808 | tidy_control_flow (basic_block xbb, bool full_tidying) | |
3809 | { | |
3810 | bool changed = true; | |
9845d120 | 3811 | insn_t first, last; |
48e1416a | 3812 | |
e1ab7874 | 3813 | /* First check whether XBB is empty. */ |
6f0e7980 | 3814 | changed = maybe_tidy_empty_bb (xbb); |
e1ab7874 | 3815 | if (changed || !full_tidying) |
3816 | return changed; | |
48e1416a | 3817 | |
e1ab7874 | 3818 | /* Check if there is a unnecessary jump after insn left. */ |
49087fba | 3819 | if (bb_has_removable_jump_to_p (xbb, xbb->next_bb) |
e1ab7874 | 3820 | && INSN_SCHED_TIMES (BB_END (xbb)) == 0 |
3821 | && !IN_CURRENT_FENCE_P (BB_END (xbb))) | |
3822 | { | |
3823 | if (sel_remove_insn (BB_END (xbb), false, false)) | |
3824 | return true; | |
3825 | tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); | |
3826 | } | |
3827 | ||
9845d120 | 3828 | first = sel_bb_head (xbb); |
3829 | last = sel_bb_end (xbb); | |
3830 | if (MAY_HAVE_DEBUG_INSNS) | |
3831 | { | |
3832 | if (first != last && DEBUG_INSN_P (first)) | |
3833 | do | |
3834 | first = NEXT_INSN (first); | |
3835 | while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); | |
3836 | ||
3837 | if (first != last && DEBUG_INSN_P (last)) | |
3838 | do | |
3839 | last = PREV_INSN (last); | |
3840 | while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); | |
3841 | } | |
e1ab7874 | 3842 | /* Check if there is an unnecessary jump in previous basic block leading |
48e1416a | 3843 | to next basic block left after removing INSN from stream. |
3844 | If it is so, remove that jump and redirect edge to current | |
3845 | basic block (where there was INSN before deletion). This way | |
3846 | when NOP will be deleted several instructions later with its | |
3847 | basic block we will not get a jump to next instruction, which | |
e1ab7874 | 3848 | can be harmful. */ |
9845d120 | 3849 | if (first == last |
e1ab7874 | 3850 | && !sel_bb_empty_p (xbb) |
9845d120 | 3851 | && INSN_NOP_P (last) |
e1ab7874 | 3852 | /* Flow goes fallthru from current block to the next. */ |
3853 | && EDGE_COUNT (xbb->succs) == 1 | |
3854 | && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) | |
3855 | /* When successor is an EXIT block, it may not be the next block. */ | |
34154e27 | 3856 | && single_succ (xbb) != EXIT_BLOCK_PTR_FOR_FN (cfun) |
e1ab7874 | 3857 | /* And unconditional jump in previous basic block leads to |
3858 | next basic block of XBB and this jump can be safely removed. */ | |
3859 | && in_current_region_p (xbb->prev_bb) | |
49087fba | 3860 | && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb) |
e1ab7874 | 3861 | && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 |
3862 | /* Also this jump is not at the scheduling boundary. */ | |
3863 | && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) | |
3864 | { | |
93919afc | 3865 | bool recompute_toporder_p; |
e1ab7874 | 3866 | /* Clear data structures of jump - jump itself will be removed |
3867 | by sel_redirect_edge_and_branch. */ | |
3868 | clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); | |
93919afc | 3869 | recompute_toporder_p |
3870 | = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); | |
3871 | ||
e1ab7874 | 3872 | gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); |
3873 | ||
3874 | /* It can turn out that after removing unused jump, basic block | |
3875 | that contained that jump, becomes empty too. In such case | |
3876 | remove it too. */ | |
3877 | if (sel_bb_empty_p (xbb->prev_bb)) | |
6f0e7980 | 3878 | changed = maybe_tidy_empty_bb (xbb->prev_bb); |
3879 | if (recompute_toporder_p) | |
93919afc | 3880 | sel_recompute_toporder (); |
e1ab7874 | 3881 | } |
7af466ad | 3882 | |
3883 | #ifdef ENABLE_CHECKING | |
3884 | verify_backedges (); | |
1a5dbaab | 3885 | verify_dominators (CDI_DOMINATORS); |
7af466ad | 3886 | #endif |
3887 | ||
e1ab7874 | 3888 | return changed; |
3889 | } | |
3890 | ||
93919afc | 3891 | /* Purge meaningless empty blocks in the middle of a region. */ |
3892 | void | |
3893 | purge_empty_blocks (void) | |
3894 | { | |
a6e634c6 | 3895 | int i; |
93919afc | 3896 | |
a6e634c6 | 3897 | /* Do not attempt to delete the first basic block in the region. */ |
3898 | for (i = 1; i < current_nr_blocks; ) | |
93919afc | 3899 | { |
f5a6b05f | 3900 | basic_block b = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
93919afc | 3901 | |
6f0e7980 | 3902 | if (maybe_tidy_empty_bb (b)) |
93919afc | 3903 | continue; |
3904 | ||
3905 | i++; | |
3906 | } | |
3907 | } | |
3908 | ||
48e1416a | 3909 | /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, |
3910 | do not delete insn's data, because it will be later re-emitted. | |
e1ab7874 | 3911 | Return true if we have removed some blocks afterwards. */ |
3912 | bool | |
3913 | sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) | |
3914 | { | |
3915 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3916 | ||
3917 | gcc_assert (INSN_IN_STREAM_P (insn)); | |
3918 | ||
9845d120 | 3919 | if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) |
3920 | { | |
3921 | expr_t expr; | |
3922 | av_set_iterator i; | |
3923 | ||
3924 | /* When we remove a debug insn that is head of a BB, it remains | |
3925 | in the AV_SET of the block, but it shouldn't. */ | |
3926 | FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) | |
3927 | if (EXPR_INSN_RTX (expr) == insn) | |
3928 | { | |
3929 | av_set_iter_remove (&i); | |
3930 | break; | |
3931 | } | |
3932 | } | |
3933 | ||
e1ab7874 | 3934 | if (only_disconnect) |
93ff53d3 | 3935 | remove_insn (insn); |
e1ab7874 | 3936 | else |
3937 | { | |
93ff53d3 | 3938 | delete_insn (insn); |
e1ab7874 | 3939 | clear_expr (INSN_EXPR (insn)); |
3940 | } | |
3941 | ||
93ff53d3 | 3942 | /* It is necessary to NULL these fields in case we are going to re-insert |
3943 | INSN into the insns stream, as will usually happen in the ONLY_DISCONNECT | |
3944 | case, but also for NOPs that we will return to the nop pool. */ | |
e1ab7874 | 3945 | PREV_INSN (insn) = NULL_RTX; |
3946 | NEXT_INSN (insn) = NULL_RTX; | |
93ff53d3 | 3947 | set_block_for_insn (insn, NULL); |
e1ab7874 | 3948 | |
3949 | return tidy_control_flow (bb, full_tidying); | |
3950 | } | |
3951 | ||
3952 | /* Estimate number of the insns in BB. */ | |
3953 | static int | |
3954 | sel_estimate_number_of_insns (basic_block bb) | |
3955 | { | |
3956 | int res = 0; | |
3957 | insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); | |
3958 | ||
3959 | for (; insn != next_tail; insn = NEXT_INSN (insn)) | |
9845d120 | 3960 | if (NONDEBUG_INSN_P (insn)) |
e1ab7874 | 3961 | res++; |
3962 | ||
3963 | return res; | |
3964 | } | |
3965 | ||
3966 | /* We don't need separate luids for notes or labels. */ | |
3967 | static int | |
3968 | sel_luid_for_non_insn (rtx x) | |
3969 | { | |
3970 | gcc_assert (NOTE_P (x) || LABEL_P (x)); | |
3971 | ||
3972 | return -1; | |
3973 | } | |
3974 | ||
bdcc104c | 3975 | /* Find the proper seqno for inserting at INSN by successors. |
3976 | Return -1 if no successors with positive seqno exist. */ | |
e1ab7874 | 3977 | static int |
bdcc104c | 3978 | get_seqno_by_succs (rtx insn) |
3979 | { | |
3980 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3981 | rtx tmp = insn, end = BB_END (bb); | |
3982 | int seqno; | |
3983 | insn_t succ = NULL; | |
3984 | succ_iterator si; | |
3985 | ||
3986 | while (tmp != end) | |
3987 | { | |
3988 | tmp = NEXT_INSN (tmp); | |
3989 | if (INSN_P (tmp)) | |
3990 | return INSN_SEQNO (tmp); | |
3991 | } | |
3992 | ||
3993 | seqno = INT_MAX; | |
3994 | ||
3995 | FOR_EACH_SUCC_1 (succ, si, end, SUCCS_NORMAL) | |
3996 | if (INSN_SEQNO (succ) > 0) | |
3997 | seqno = MIN (seqno, INSN_SEQNO (succ)); | |
3998 | ||
3999 | if (seqno == INT_MAX) | |
4000 | return -1; | |
4001 | ||
4002 | return seqno; | |
4003 | } | |
4004 | ||
4005 | /* Compute seqno for INSN by its preds or succs. */ | |
4006 | static int | |
4007 | get_seqno_for_a_jump (insn_t insn) | |
e1ab7874 | 4008 | { |
4009 | int seqno; | |
4010 | ||
4011 | gcc_assert (INSN_SIMPLEJUMP_P (insn)); | |
4012 | ||
4013 | if (!sel_bb_head_p (insn)) | |
4014 | seqno = INSN_SEQNO (PREV_INSN (insn)); | |
4015 | else | |
4016 | { | |
4017 | basic_block bb = BLOCK_FOR_INSN (insn); | |
4018 | ||
4019 | if (single_pred_p (bb) | |
4020 | && !in_current_region_p (single_pred (bb))) | |
4021 | { | |
4022 | /* We can have preds outside a region when splitting edges | |
48e1416a | 4023 | for pipelining of an outer loop. Use succ instead. |
e1ab7874 | 4024 | There should be only one of them. */ |
4025 | insn_t succ = NULL; | |
4026 | succ_iterator si; | |
4027 | bool first = true; | |
48e1416a | 4028 | |
e1ab7874 | 4029 | gcc_assert (flag_sel_sched_pipelining_outer_loops |
4030 | && current_loop_nest); | |
48e1416a | 4031 | FOR_EACH_SUCC_1 (succ, si, insn, |
e1ab7874 | 4032 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
4033 | { | |
4034 | gcc_assert (first); | |
4035 | first = false; | |
4036 | } | |
4037 | ||
4038 | gcc_assert (succ != NULL); | |
4039 | seqno = INSN_SEQNO (succ); | |
4040 | } | |
4041 | else | |
4042 | { | |
4043 | insn_t *preds; | |
4044 | int n; | |
4045 | ||
4046 | cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); | |
e1ab7874 | 4047 | |
bdcc104c | 4048 | gcc_assert (n > 0); |
4049 | /* For one predecessor, use simple method. */ | |
4050 | if (n == 1) | |
4051 | seqno = INSN_SEQNO (preds[0]); | |
4052 | else | |
4053 | seqno = get_seqno_by_preds (insn); | |
48e1416a | 4054 | |
e1ab7874 | 4055 | free (preds); |
4056 | } | |
4057 | } | |
4058 | ||
bdcc104c | 4059 | /* We were unable to find a good seqno among preds. */ |
4060 | if (seqno < 0) | |
4061 | seqno = get_seqno_by_succs (insn); | |
4062 | ||
4063 | gcc_assert (seqno >= 0); | |
4064 | ||
e1ab7874 | 4065 | return seqno; |
4066 | } | |
4067 | ||
961d3eb8 | 4068 | /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors |
4069 | with positive seqno exist. */ | |
e1ab7874 | 4070 | int |
4071 | get_seqno_by_preds (rtx insn) | |
4072 | { | |
4073 | basic_block bb = BLOCK_FOR_INSN (insn); | |
4074 | rtx tmp = insn, head = BB_HEAD (bb); | |
4075 | insn_t *preds; | |
4076 | int n, i, seqno; | |
4077 | ||
4078 | while (tmp != head) | |
bdcc104c | 4079 | { |
e1ab7874 | 4080 | tmp = PREV_INSN (tmp); |
bdcc104c | 4081 | if (INSN_P (tmp)) |
4082 | return INSN_SEQNO (tmp); | |
4083 | } | |
48e1416a | 4084 | |
e1ab7874 | 4085 | cfg_preds (bb, &preds, &n); |
4086 | for (i = 0, seqno = -1; i < n; i++) | |
4087 | seqno = MAX (seqno, INSN_SEQNO (preds[i])); | |
4088 | ||
e1ab7874 | 4089 | return seqno; |
4090 | } | |
4091 | ||
4092 | \f | |
4093 | ||
4094 | /* Extend pass-scope data structures for basic blocks. */ | |
4095 | void | |
4096 | sel_extend_global_bb_info (void) | |
4097 | { | |
fe672ac0 | 4098 | sel_global_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun)); |
e1ab7874 | 4099 | } |
4100 | ||
4101 | /* Extend region-scope data structures for basic blocks. */ | |
4102 | static void | |
4103 | extend_region_bb_info (void) | |
4104 | { | |
fe672ac0 | 4105 | sel_region_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun)); |
e1ab7874 | 4106 | } |
4107 | ||
4108 | /* Extend all data structures to fit for all basic blocks. */ | |
4109 | static void | |
4110 | extend_bb_info (void) | |
4111 | { | |
4112 | sel_extend_global_bb_info (); | |
4113 | extend_region_bb_info (); | |
4114 | } | |
4115 | ||
4116 | /* Finalize pass-scope data structures for basic blocks. */ | |
4117 | void | |
4118 | sel_finish_global_bb_info (void) | |
4119 | { | |
f1f41a6c | 4120 | sel_global_bb_info.release (); |
e1ab7874 | 4121 | } |
4122 | ||
4123 | /* Finalize region-scope data structures for basic blocks. */ | |
4124 | static void | |
4125 | finish_region_bb_info (void) | |
4126 | { | |
f1f41a6c | 4127 | sel_region_bb_info.release (); |
e1ab7874 | 4128 | } |
4129 | \f | |
4130 | ||
4131 | /* Data for each insn in current region. */ | |
1e094109 | 4132 | vec<sel_insn_data_def> s_i_d = vNULL; |
e1ab7874 | 4133 | |
e1ab7874 | 4134 | /* Extend data structures for insns from current region. */ |
4135 | static void | |
4136 | extend_insn_data (void) | |
4137 | { | |
4138 | int reserve; | |
48e1416a | 4139 | |
e1ab7874 | 4140 | sched_extend_target (); |
4141 | sched_deps_init (false); | |
4142 | ||
4143 | /* Extend data structures for insns from current region. */ | |
f1f41a6c | 4144 | reserve = (sched_max_luid + 1 - s_i_d.length ()); |
4145 | if (reserve > 0 && ! s_i_d.space (reserve)) | |
d9ab2038 | 4146 | { |
4147 | int size; | |
4148 | ||
4149 | if (sched_max_luid / 2 > 1024) | |
4150 | size = sched_max_luid + 1024; | |
4151 | else | |
4152 | size = 3 * sched_max_luid / 2; | |
48e1416a | 4153 | |
d9ab2038 | 4154 | |
f1f41a6c | 4155 | s_i_d.safe_grow_cleared (size); |
d9ab2038 | 4156 | } |
e1ab7874 | 4157 | } |
4158 | ||
4159 | /* Finalize data structures for insns from current region. */ | |
4160 | static void | |
4161 | finish_insns (void) | |
4162 | { | |
4163 | unsigned i; | |
4164 | ||
4165 | /* Clear here all dependence contexts that may have left from insns that were | |
4166 | removed during the scheduling. */ | |
f1f41a6c | 4167 | for (i = 0; i < s_i_d.length (); i++) |
e1ab7874 | 4168 | { |
f1f41a6c | 4169 | sel_insn_data_def *sid_entry = &s_i_d[i]; |
48e1416a | 4170 | |
e1ab7874 | 4171 | if (sid_entry->live) |
4172 | return_regset_to_pool (sid_entry->live); | |
4173 | if (sid_entry->analyzed_deps) | |
4174 | { | |
4175 | BITMAP_FREE (sid_entry->analyzed_deps); | |
4176 | BITMAP_FREE (sid_entry->found_deps); | |
4177 | htab_delete (sid_entry->transformed_insns); | |
4178 | free_deps (&sid_entry->deps_context); | |
4179 | } | |
4180 | if (EXPR_VINSN (&sid_entry->expr)) | |
4181 | { | |
4182 | clear_expr (&sid_entry->expr); | |
48e1416a | 4183 | |
e1ab7874 | 4184 | /* Also, clear CANT_MOVE bit here, because we really don't want it |
4185 | to be passed to the next region. */ | |
4186 | CANT_MOVE_BY_LUID (i) = 0; | |
4187 | } | |
4188 | } | |
48e1416a | 4189 | |
f1f41a6c | 4190 | s_i_d.release (); |
e1ab7874 | 4191 | } |
4192 | ||
4193 | /* A proxy to pass initialization data to init_insn (). */ | |
4194 | static sel_insn_data_def _insn_init_ssid; | |
4195 | static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; | |
4196 | ||
4197 | /* If true create a new vinsn. Otherwise use the one from EXPR. */ | |
4198 | static bool insn_init_create_new_vinsn_p; | |
4199 | ||
4200 | /* Set all necessary data for initialization of the new insn[s]. */ | |
4201 | static expr_t | |
4202 | set_insn_init (expr_t expr, vinsn_t vi, int seqno) | |
4203 | { | |
4204 | expr_t x = &insn_init_ssid->expr; | |
4205 | ||
4206 | copy_expr_onside (x, expr); | |
4207 | if (vi != NULL) | |
4208 | { | |
4209 | insn_init_create_new_vinsn_p = false; | |
4210 | change_vinsn_in_expr (x, vi); | |
4211 | } | |
4212 | else | |
4213 | insn_init_create_new_vinsn_p = true; | |
4214 | ||
4215 | insn_init_ssid->seqno = seqno; | |
4216 | return x; | |
4217 | } | |
4218 | ||
4219 | /* Init data for INSN. */ | |
4220 | static void | |
4221 | init_insn_data (insn_t insn) | |
4222 | { | |
4223 | expr_t expr; | |
4224 | sel_insn_data_t ssid = insn_init_ssid; | |
4225 | ||
4226 | /* The fields mentioned below are special and hence are not being | |
4227 | propagated to the new insns. */ | |
4228 | gcc_assert (!ssid->asm_p && ssid->sched_next == NULL | |
4229 | && !ssid->after_stall_p && ssid->sched_cycle == 0); | |
4230 | gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); | |
4231 | ||
4232 | expr = INSN_EXPR (insn); | |
4233 | copy_expr (expr, &ssid->expr); | |
4234 | prepare_insn_expr (insn, ssid->seqno); | |
4235 | ||
4236 | if (insn_init_create_new_vinsn_p) | |
4237 | change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p)); | |
48e1416a | 4238 | |
e1ab7874 | 4239 | if (first_time_insn_init (insn)) |
4240 | init_first_time_insn_data (insn); | |
4241 | } | |
4242 | ||
4243 | /* This is used to initialize spurious jumps generated by | |
4244 | sel_redirect_edge (). */ | |
4245 | static void | |
4246 | init_simplejump_data (insn_t insn) | |
4247 | { | |
4248 | init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0, | |
f1f41a6c | 4249 | REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0, |
1e094109 | 4250 | vNULL, true, false, false, |
e1ab7874 | 4251 | false, true); |
bdcc104c | 4252 | INSN_SEQNO (insn) = get_seqno_for_a_jump (insn); |
e1ab7874 | 4253 | init_first_time_insn_data (insn); |
4254 | } | |
4255 | ||
48e1416a | 4256 | /* Perform deferred initialization of insns. This is used to process |
e1ab7874 | 4257 | a new jump that may be created by redirect_edge. */ |
4258 | void | |
4259 | sel_init_new_insn (insn_t insn, int flags) | |
4260 | { | |
4261 | /* We create data structures for bb when the first insn is emitted in it. */ | |
4262 | if (INSN_P (insn) | |
4263 | && INSN_IN_STREAM_P (insn) | |
4264 | && insn_is_the_only_one_in_bb_p (insn)) | |
4265 | { | |
4266 | extend_bb_info (); | |
4267 | create_initial_data_sets (BLOCK_FOR_INSN (insn)); | |
4268 | } | |
48e1416a | 4269 | |
e1ab7874 | 4270 | if (flags & INSN_INIT_TODO_LUID) |
52d7e28c | 4271 | { |
4272 | sched_extend_luids (); | |
4273 | sched_init_insn_luid (insn); | |
4274 | } | |
e1ab7874 | 4275 | |
4276 | if (flags & INSN_INIT_TODO_SSID) | |
4277 | { | |
4278 | extend_insn_data (); | |
4279 | init_insn_data (insn); | |
4280 | clear_expr (&insn_init_ssid->expr); | |
4281 | } | |
4282 | ||
4283 | if (flags & INSN_INIT_TODO_SIMPLEJUMP) | |
4284 | { | |
4285 | extend_insn_data (); | |
4286 | init_simplejump_data (insn); | |
4287 | } | |
48e1416a | 4288 | |
e1ab7874 | 4289 | gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) |
4290 | == CONTAINING_RGN (BB_TO_BLOCK (0))); | |
4291 | } | |
4292 | \f | |
4293 | ||
4294 | /* Functions to init/finish work with lv sets. */ | |
4295 | ||
4296 | /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ | |
4297 | static void | |
4298 | init_lv_set (basic_block bb) | |
4299 | { | |
4300 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
4301 | ||
4302 | BB_LV_SET (bb) = get_regset_from_pool (); | |
48e1416a | 4303 | COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); |
e1ab7874 | 4304 | BB_LV_SET_VALID_P (bb) = true; |
4305 | } | |
4306 | ||
4307 | /* Copy liveness information to BB from FROM_BB. */ | |
4308 | static void | |
4309 | copy_lv_set_from (basic_block bb, basic_block from_bb) | |
4310 | { | |
4311 | gcc_assert (!BB_LV_SET_VALID_P (bb)); | |
48e1416a | 4312 | |
e1ab7874 | 4313 | COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); |
4314 | BB_LV_SET_VALID_P (bb) = true; | |
48e1416a | 4315 | } |
e1ab7874 | 4316 | |
4317 | /* Initialize lv set of all bb headers. */ | |
4318 | void | |
4319 | init_lv_sets (void) | |
4320 | { | |
4321 | basic_block bb; | |
4322 | ||
4323 | /* Initialize of LV sets. */ | |
fc00614f | 4324 | FOR_EACH_BB_FN (bb, cfun) |
e1ab7874 | 4325 | init_lv_set (bb); |
4326 | ||
4327 | /* Don't forget EXIT_BLOCK. */ | |
34154e27 | 4328 | init_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e1ab7874 | 4329 | } |
4330 | ||
4331 | /* Release lv set of HEAD. */ | |
4332 | static void | |
4333 | free_lv_set (basic_block bb) | |
4334 | { | |
4335 | gcc_assert (BB_LV_SET (bb) != NULL); | |
4336 | ||
4337 | return_regset_to_pool (BB_LV_SET (bb)); | |
4338 | BB_LV_SET (bb) = NULL; | |
4339 | BB_LV_SET_VALID_P (bb) = false; | |
4340 | } | |
4341 | ||
4342 | /* Finalize lv sets of all bb headers. */ | |
4343 | void | |
4344 | free_lv_sets (void) | |
4345 | { | |
4346 | basic_block bb; | |
4347 | ||
4348 | /* Don't forget EXIT_BLOCK. */ | |
34154e27 | 4349 | free_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e1ab7874 | 4350 | |
4351 | /* Free LV sets. */ | |
fc00614f | 4352 | FOR_EACH_BB_FN (bb, cfun) |
e1ab7874 | 4353 | if (BB_LV_SET (bb)) |
4354 | free_lv_set (bb); | |
4355 | } | |
4356 | ||
c1c8a3d0 | 4357 | /* Mark AV_SET for BB as invalid, so this set will be updated the next time |
4358 | compute_av() processes BB. This function is called when creating new basic | |
4359 | blocks, as well as for blocks (either new or existing) where new jumps are | |
4360 | created when the control flow is being updated. */ | |
e1ab7874 | 4361 | static void |
4362 | invalidate_av_set (basic_block bb) | |
4363 | { | |
e1ab7874 | 4364 | BB_AV_LEVEL (bb) = -1; |
4365 | } | |
4366 | ||
4367 | /* Create initial data sets for BB (they will be invalid). */ | |
4368 | static void | |
4369 | create_initial_data_sets (basic_block bb) | |
4370 | { | |
4371 | if (BB_LV_SET (bb)) | |
4372 | BB_LV_SET_VALID_P (bb) = false; | |
4373 | else | |
4374 | BB_LV_SET (bb) = get_regset_from_pool (); | |
4375 | invalidate_av_set (bb); | |
4376 | } | |
4377 | ||
4378 | /* Free av set of BB. */ | |
4379 | static void | |
4380 | free_av_set (basic_block bb) | |
4381 | { | |
4382 | av_set_clear (&BB_AV_SET (bb)); | |
4383 | BB_AV_LEVEL (bb) = 0; | |
4384 | } | |
4385 | ||
4386 | /* Free data sets of BB. */ | |
4387 | void | |
4388 | free_data_sets (basic_block bb) | |
4389 | { | |
4390 | free_lv_set (bb); | |
4391 | free_av_set (bb); | |
4392 | } | |
4393 | ||
4394 | /* Exchange lv sets of TO and FROM. */ | |
4395 | static void | |
4396 | exchange_lv_sets (basic_block to, basic_block from) | |
4397 | { | |
4398 | { | |
4399 | regset to_lv_set = BB_LV_SET (to); | |
4400 | ||
4401 | BB_LV_SET (to) = BB_LV_SET (from); | |
4402 | BB_LV_SET (from) = to_lv_set; | |
4403 | } | |
4404 | ||
4405 | { | |
4406 | bool to_lv_set_valid_p = BB_LV_SET_VALID_P (to); | |
4407 | ||
4408 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4409 | BB_LV_SET_VALID_P (from) = to_lv_set_valid_p; | |
4410 | } | |
4411 | } | |
4412 | ||
4413 | ||
4414 | /* Exchange av sets of TO and FROM. */ | |
4415 | static void | |
4416 | exchange_av_sets (basic_block to, basic_block from) | |
4417 | { | |
4418 | { | |
4419 | av_set_t to_av_set = BB_AV_SET (to); | |
4420 | ||
4421 | BB_AV_SET (to) = BB_AV_SET (from); | |
4422 | BB_AV_SET (from) = to_av_set; | |
4423 | } | |
4424 | ||
4425 | { | |
4426 | int to_av_level = BB_AV_LEVEL (to); | |
4427 | ||
4428 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4429 | BB_AV_LEVEL (from) = to_av_level; | |
4430 | } | |
4431 | } | |
4432 | ||
4433 | /* Exchange data sets of TO and FROM. */ | |
4434 | void | |
4435 | exchange_data_sets (basic_block to, basic_block from) | |
4436 | { | |
4437 | exchange_lv_sets (to, from); | |
4438 | exchange_av_sets (to, from); | |
4439 | } | |
4440 | ||
4441 | /* Copy data sets of FROM to TO. */ | |
4442 | void | |
4443 | copy_data_sets (basic_block to, basic_block from) | |
4444 | { | |
4445 | gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); | |
4446 | gcc_assert (BB_AV_SET (to) == NULL); | |
4447 | ||
4448 | BB_AV_LEVEL (to) = BB_AV_LEVEL (from); | |
4449 | BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); | |
4450 | ||
4451 | if (BB_AV_SET_VALID_P (from)) | |
4452 | { | |
4453 | BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); | |
4454 | } | |
4455 | if (BB_LV_SET_VALID_P (from)) | |
4456 | { | |
4457 | gcc_assert (BB_LV_SET (to) != NULL); | |
4458 | COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); | |
4459 | } | |
4460 | } | |
4461 | ||
4462 | /* Return an av set for INSN, if any. */ | |
4463 | av_set_t | |
4464 | get_av_set (insn_t insn) | |
4465 | { | |
4466 | av_set_t av_set; | |
4467 | ||
4468 | gcc_assert (AV_SET_VALID_P (insn)); | |
4469 | ||
4470 | if (sel_bb_head_p (insn)) | |
4471 | av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); | |
4472 | else | |
4473 | av_set = NULL; | |
4474 | ||
4475 | return av_set; | |
4476 | } | |
4477 | ||
4478 | /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ | |
4479 | int | |
4480 | get_av_level (insn_t insn) | |
4481 | { | |
4482 | int av_level; | |
4483 | ||
4484 | gcc_assert (INSN_P (insn)); | |
4485 | ||
4486 | if (sel_bb_head_p (insn)) | |
4487 | av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); | |
4488 | else | |
4489 | av_level = INSN_WS_LEVEL (insn); | |
4490 | ||
4491 | return av_level; | |
4492 | } | |
4493 | ||
4494 | \f | |
4495 | ||
4496 | /* Variables to work with control-flow graph. */ | |
4497 | ||
4498 | /* The basic block that already has been processed by the sched_data_update (), | |
4499 | but hasn't been in sel_add_bb () yet. */ | |
f1f41a6c | 4500 | static vec<basic_block> |
1e094109 | 4501 | last_added_blocks = vNULL; |
e1ab7874 | 4502 | |
4503 | /* A pool for allocating successor infos. */ | |
4504 | static struct | |
4505 | { | |
4506 | /* A stack for saving succs_info structures. */ | |
4507 | struct succs_info *stack; | |
4508 | ||
4509 | /* Its size. */ | |
4510 | int size; | |
4511 | ||
4512 | /* Top of the stack. */ | |
4513 | int top; | |
4514 | ||
4515 | /* Maximal value of the top. */ | |
4516 | int max_top; | |
4517 | } succs_info_pool; | |
4518 | ||
4519 | /* Functions to work with control-flow graph. */ | |
4520 | ||
4521 | /* Return basic block note of BB. */ | |
4522 | insn_t | |
4523 | sel_bb_head (basic_block bb) | |
4524 | { | |
4525 | insn_t head; | |
4526 | ||
34154e27 | 4527 | if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
e1ab7874 | 4528 | { |
4529 | gcc_assert (exit_insn != NULL_RTX); | |
4530 | head = exit_insn; | |
4531 | } | |
4532 | else | |
4533 | { | |
4534 | insn_t note; | |
4535 | ||
4536 | note = bb_note (bb); | |
4537 | head = next_nonnote_insn (note); | |
4538 | ||
cabd2128 | 4539 | if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb)) |
e1ab7874 | 4540 | head = NULL_RTX; |
4541 | } | |
4542 | ||
4543 | return head; | |
4544 | } | |
4545 | ||
4546 | /* Return true if INSN is a basic block header. */ | |
4547 | bool | |
4548 | sel_bb_head_p (insn_t insn) | |
4549 | { | |
4550 | return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn; | |
4551 | } | |
4552 | ||
4553 | /* Return last insn of BB. */ | |
4554 | insn_t | |
4555 | sel_bb_end (basic_block bb) | |
4556 | { | |
4557 | if (sel_bb_empty_p (bb)) | |
4558 | return NULL_RTX; | |
4559 | ||
34154e27 | 4560 | gcc_assert (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e1ab7874 | 4561 | |
4562 | return BB_END (bb); | |
4563 | } | |
4564 | ||
4565 | /* Return true if INSN is the last insn in its basic block. */ | |
4566 | bool | |
4567 | sel_bb_end_p (insn_t insn) | |
4568 | { | |
4569 | return insn == sel_bb_end (BLOCK_FOR_INSN (insn)); | |
4570 | } | |
4571 | ||
4572 | /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ | |
4573 | bool | |
4574 | sel_bb_empty_p (basic_block bb) | |
4575 | { | |
4576 | return sel_bb_head (bb) == NULL; | |
4577 | } | |
4578 | ||
4579 | /* True when BB belongs to the current scheduling region. */ | |
4580 | bool | |
4581 | in_current_region_p (basic_block bb) | |
4582 | { | |
4583 | if (bb->index < NUM_FIXED_BLOCKS) | |
4584 | return false; | |
4585 | ||
4586 | return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); | |
4587 | } | |
4588 | ||
4589 | /* Return the block which is a fallthru bb of a conditional jump JUMP. */ | |
4590 | basic_block | |
4591 | fallthru_bb_of_jump (rtx jump) | |
4592 | { | |
4593 | if (!JUMP_P (jump)) | |
4594 | return NULL; | |
4595 | ||
e1ab7874 | 4596 | if (!any_condjump_p (jump)) |
4597 | return NULL; | |
4598 | ||
bf19734b | 4599 | /* A basic block that ends with a conditional jump may still have one successor |
4600 | (and be followed by a barrier), we are not interested. */ | |
4601 | if (single_succ_p (BLOCK_FOR_INSN (jump))) | |
4602 | return NULL; | |
4603 | ||
e1ab7874 | 4604 | return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; |
4605 | } | |
4606 | ||
4607 | /* Remove all notes from BB. */ | |
4608 | static void | |
4609 | init_bb (basic_block bb) | |
4610 | { | |
4611 | remove_notes (bb_note (bb), BB_END (bb)); | |
4612 | BB_NOTE_LIST (bb) = note_list; | |
4613 | } | |
4614 | ||
4615 | void | |
52d7e28c | 4616 | sel_init_bbs (bb_vec_t bbs) |
e1ab7874 | 4617 | { |
4618 | const struct sched_scan_info_def ssi = | |
4619 | { | |
4620 | extend_bb_info, /* extend_bb */ | |
4621 | init_bb, /* init_bb */ | |
4622 | NULL, /* extend_insn */ | |
4623 | NULL /* init_insn */ | |
4624 | }; | |
4625 | ||
52d7e28c | 4626 | sched_scan (&ssi, bbs); |
e1ab7874 | 4627 | } |
4628 | ||
3baa98a0 | 4629 | /* Restore notes for the whole region. */ |
e1ab7874 | 4630 | static void |
3baa98a0 | 4631 | sel_restore_notes (void) |
e1ab7874 | 4632 | { |
4633 | int bb; | |
3baa98a0 | 4634 | insn_t insn; |
e1ab7874 | 4635 | |
4636 | for (bb = 0; bb < current_nr_blocks; bb++) | |
4637 | { | |
4638 | basic_block first, last; | |
4639 | ||
4640 | first = EBB_FIRST_BB (bb); | |
4641 | last = EBB_LAST_BB (bb)->next_bb; | |
4642 | ||
4643 | do | |
4644 | { | |
4645 | note_list = BB_NOTE_LIST (first); | |
4646 | restore_other_notes (NULL, first); | |
4647 | BB_NOTE_LIST (first) = NULL_RTX; | |
4648 | ||
3baa98a0 | 4649 | FOR_BB_INSNS (first, insn) |
4650 | if (NONDEBUG_INSN_P (insn)) | |
4651 | reemit_notes (insn); | |
4652 | ||
e1ab7874 | 4653 | first = first->next_bb; |
4654 | } | |
4655 | while (first != last); | |
4656 | } | |
4657 | } | |
4658 | ||
4659 | /* Free per-bb data structures. */ | |
4660 | void | |
4661 | sel_finish_bbs (void) | |
4662 | { | |
3baa98a0 | 4663 | sel_restore_notes (); |
e1ab7874 | 4664 | |
4665 | /* Remove current loop preheader from this loop. */ | |
4666 | if (current_loop_nest) | |
4667 | sel_remove_loop_preheader (); | |
4668 | ||
4669 | finish_region_bb_info (); | |
4670 | } | |
4671 | ||
4672 | /* Return true if INSN has a single successor of type FLAGS. */ | |
4673 | bool | |
4674 | sel_insn_has_single_succ_p (insn_t insn, int flags) | |
4675 | { | |
4676 | insn_t succ; | |
4677 | succ_iterator si; | |
4678 | bool first_p = true; | |
4679 | ||
4680 | FOR_EACH_SUCC_1 (succ, si, insn, flags) | |
4681 | { | |
4682 | if (first_p) | |
4683 | first_p = false; | |
4684 | else | |
4685 | return false; | |
4686 | } | |
4687 | ||
4688 | return true; | |
4689 | } | |
4690 | ||
4691 | /* Allocate successor's info. */ | |
4692 | static struct succs_info * | |
4693 | alloc_succs_info (void) | |
4694 | { | |
4695 | if (succs_info_pool.top == succs_info_pool.max_top) | |
4696 | { | |
4697 | int i; | |
48e1416a | 4698 | |
e1ab7874 | 4699 | if (++succs_info_pool.max_top >= succs_info_pool.size) |
4700 | gcc_unreachable (); | |
4701 | ||
4702 | i = ++succs_info_pool.top; | |
f1f41a6c | 4703 | succs_info_pool.stack[i].succs_ok.create (10); |
4704 | succs_info_pool.stack[i].succs_other.create (10); | |
4705 | succs_info_pool.stack[i].probs_ok.create (10); | |
e1ab7874 | 4706 | } |
4707 | else | |
4708 | succs_info_pool.top++; | |
4709 | ||
4710 | return &succs_info_pool.stack[succs_info_pool.top]; | |
4711 | } | |
4712 | ||
4713 | /* Free successor's info. */ | |
4714 | void | |
4715 | free_succs_info (struct succs_info * sinfo) | |
4716 | { | |
48e1416a | 4717 | gcc_assert (succs_info_pool.top >= 0 |
e1ab7874 | 4718 | && &succs_info_pool.stack[succs_info_pool.top] == sinfo); |
4719 | succs_info_pool.top--; | |
4720 | ||
4721 | /* Clear stale info. */ | |
f1f41a6c | 4722 | sinfo->succs_ok.block_remove (0, sinfo->succs_ok.length ()); |
4723 | sinfo->succs_other.block_remove (0, sinfo->succs_other.length ()); | |
4724 | sinfo->probs_ok.block_remove (0, sinfo->probs_ok.length ()); | |
e1ab7874 | 4725 | sinfo->all_prob = 0; |
4726 | sinfo->succs_ok_n = 0; | |
4727 | sinfo->all_succs_n = 0; | |
4728 | } | |
4729 | ||
48e1416a | 4730 | /* Compute successor info for INSN. FLAGS are the flags passed |
e1ab7874 | 4731 | to the FOR_EACH_SUCC_1 iterator. */ |
4732 | struct succs_info * | |
4733 | compute_succs_info (insn_t insn, short flags) | |
4734 | { | |
4735 | succ_iterator si; | |
4736 | insn_t succ; | |
4737 | struct succs_info *sinfo = alloc_succs_info (); | |
4738 | ||
4739 | /* Traverse *all* successors and decide what to do with each. */ | |
4740 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
4741 | { | |
4742 | /* FIXME: this doesn't work for skipping to loop exits, as we don't | |
4743 | perform code motion through inner loops. */ | |
4744 | short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; | |
4745 | ||
4746 | if (current_flags & flags) | |
4747 | { | |
f1f41a6c | 4748 | sinfo->succs_ok.safe_push (succ); |
4749 | sinfo->probs_ok.safe_push ( | |
4750 | /* FIXME: Improve calculation when skipping | |
4751 | inner loop to exits. */ | |
4752 | si.bb_end ? si.e1->probability : REG_BR_PROB_BASE); | |
e1ab7874 | 4753 | sinfo->succs_ok_n++; |
4754 | } | |
4755 | else | |
f1f41a6c | 4756 | sinfo->succs_other.safe_push (succ); |
e1ab7874 | 4757 | |
4758 | /* Compute all_prob. */ | |
4759 | if (!si.bb_end) | |
4760 | sinfo->all_prob = REG_BR_PROB_BASE; | |
4761 | else | |
4762 | sinfo->all_prob += si.e1->probability; | |
4763 | ||
4764 | sinfo->all_succs_n++; | |
4765 | } | |
4766 | ||
4767 | return sinfo; | |
4768 | } | |
4769 | ||
48e1416a | 4770 | /* Return the predecessors of BB in PREDS and their number in N. |
e1ab7874 | 4771 | Empty blocks are skipped. SIZE is used to allocate PREDS. */ |
4772 | static void | |
4773 | cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) | |
4774 | { | |
4775 | edge e; | |
4776 | edge_iterator ei; | |
4777 | ||
4778 | gcc_assert (BLOCK_TO_BB (bb->index) != 0); | |
4779 | ||
4780 | FOR_EACH_EDGE (e, ei, bb->preds) | |
4781 | { | |
4782 | basic_block pred_bb = e->src; | |
4783 | insn_t bb_end = BB_END (pred_bb); | |
4784 | ||
f1ec9c64 | 4785 | if (!in_current_region_p (pred_bb)) |
4786 | { | |
4787 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
4788 | && current_loop_nest); | |
4789 | continue; | |
4790 | } | |
e1ab7874 | 4791 | |
4792 | if (sel_bb_empty_p (pred_bb)) | |
4793 | cfg_preds_1 (pred_bb, preds, n, size); | |
4794 | else | |
4795 | { | |
4796 | if (*n == *size) | |
48e1416a | 4797 | *preds = XRESIZEVEC (insn_t, *preds, |
e1ab7874 | 4798 | (*size = 2 * *size + 1)); |
4799 | (*preds)[(*n)++] = bb_end; | |
4800 | } | |
4801 | } | |
4802 | ||
f1ec9c64 | 4803 | gcc_assert (*n != 0 |
4804 | || (flag_sel_sched_pipelining_outer_loops | |
4805 | && current_loop_nest)); | |
e1ab7874 | 4806 | } |
4807 | ||
48e1416a | 4808 | /* Find all predecessors of BB and record them in PREDS and their number |
4809 | in N. Empty blocks are skipped, and only normal (forward in-region) | |
e1ab7874 | 4810 | edges are processed. */ |
4811 | static void | |
4812 | cfg_preds (basic_block bb, insn_t **preds, int *n) | |
4813 | { | |
4814 | int size = 0; | |
4815 | ||
4816 | *preds = NULL; | |
4817 | *n = 0; | |
4818 | cfg_preds_1 (bb, preds, n, &size); | |
4819 | } | |
4820 | ||
4821 | /* Returns true if we are moving INSN through join point. */ | |
4822 | bool | |
4823 | sel_num_cfg_preds_gt_1 (insn_t insn) | |
4824 | { | |
4825 | basic_block bb; | |
4826 | ||
4827 | if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) | |
4828 | return false; | |
4829 | ||
4830 | bb = BLOCK_FOR_INSN (insn); | |
4831 | ||
4832 | while (1) | |
4833 | { | |
4834 | if (EDGE_COUNT (bb->preds) > 1) | |
4835 | return true; | |
4836 | ||
4837 | gcc_assert (EDGE_PRED (bb, 0)->dest == bb); | |
4838 | bb = EDGE_PRED (bb, 0)->src; | |
4839 | ||
4840 | if (!sel_bb_empty_p (bb)) | |
4841 | break; | |
4842 | } | |
4843 | ||
4844 | return false; | |
4845 | } | |
4846 | ||
48e1416a | 4847 | /* Returns true when BB should be the end of an ebb. Adapted from the |
e1ab7874 | 4848 | code in sched-ebb.c. */ |
4849 | bool | |
4850 | bb_ends_ebb_p (basic_block bb) | |
4851 | { | |
4852 | basic_block next_bb = bb_next_bb (bb); | |
4853 | edge e; | |
48e1416a | 4854 | |
34154e27 | 4855 | if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e1ab7874 | 4856 | || bitmap_bit_p (forced_ebb_heads, next_bb->index) |
4857 | || (LABEL_P (BB_HEAD (next_bb)) | |
4858 | /* NB: LABEL_NUSES () is not maintained outside of jump.c. | |
4859 | Work around that. */ | |
4860 | && !single_pred_p (next_bb))) | |
4861 | return true; | |
4862 | ||
4863 | if (!in_current_region_p (next_bb)) | |
4864 | return true; | |
4865 | ||
7f58c05e | 4866 | e = find_fallthru_edge (bb->succs); |
4867 | if (e) | |
4868 | { | |
4869 | gcc_assert (e->dest == next_bb); | |
4870 | ||
4871 | return false; | |
4872 | } | |
e1ab7874 | 4873 | |
4874 | return true; | |
4875 | } | |
4876 | ||
4877 | /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a | |
4878 | successor of INSN. */ | |
4879 | bool | |
4880 | in_same_ebb_p (insn_t insn, insn_t succ) | |
4881 | { | |
4882 | basic_block ptr = BLOCK_FOR_INSN (insn); | |
4883 | ||
9af5ce0c | 4884 | for (;;) |
e1ab7874 | 4885 | { |
4886 | if (ptr == BLOCK_FOR_INSN (succ)) | |
4887 | return true; | |
48e1416a | 4888 | |
e1ab7874 | 4889 | if (bb_ends_ebb_p (ptr)) |
4890 | return false; | |
4891 | ||
4892 | ptr = bb_next_bb (ptr); | |
4893 | } | |
4894 | ||
4895 | gcc_unreachable (); | |
4896 | return false; | |
4897 | } | |
4898 | ||
4899 | /* Recomputes the reverse topological order for the function and | |
4900 | saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also | |
4901 | modified appropriately. */ | |
4902 | static void | |
4903 | recompute_rev_top_order (void) | |
4904 | { | |
4905 | int *postorder; | |
4906 | int n_blocks, i; | |
4907 | ||
fe672ac0 | 4908 | if (!rev_top_order_index |
4909 | || rev_top_order_index_len < last_basic_block_for_fn (cfun)) | |
e1ab7874 | 4910 | { |
fe672ac0 | 4911 | rev_top_order_index_len = last_basic_block_for_fn (cfun); |
e1ab7874 | 4912 | rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, |
4913 | rev_top_order_index_len); | |
4914 | } | |
4915 | ||
a28770e1 | 4916 | postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
e1ab7874 | 4917 | |
4918 | n_blocks = post_order_compute (postorder, true, false); | |
a28770e1 | 4919 | gcc_assert (n_basic_blocks_for_fn (cfun) == n_blocks); |
e1ab7874 | 4920 | |
4921 | /* Build reverse function: for each basic block with BB->INDEX == K | |
4922 | rev_top_order_index[K] is it's reverse topological sort number. */ | |
4923 | for (i = 0; i < n_blocks; i++) | |
4924 | { | |
4925 | gcc_assert (postorder[i] < rev_top_order_index_len); | |
4926 | rev_top_order_index[postorder[i]] = i; | |
4927 | } | |
4928 | ||
4929 | free (postorder); | |
4930 | } | |
4931 | ||
4932 | /* Clear all flags from insns in BB that could spoil its rescheduling. */ | |
4933 | void | |
4934 | clear_outdated_rtx_info (basic_block bb) | |
4935 | { | |
4936 | rtx insn; | |
4937 | ||
4938 | FOR_BB_INSNS (bb, insn) | |
4939 | if (INSN_P (insn)) | |
4940 | { | |
4941 | SCHED_GROUP_P (insn) = 0; | |
4942 | INSN_AFTER_STALL_P (insn) = 0; | |
4943 | INSN_SCHED_TIMES (insn) = 0; | |
4944 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; | |
4945 | ||
4946 | /* We cannot use the changed caches, as previously we could ignore | |
48e1416a | 4947 | the LHS dependence due to enabled renaming and transform |
e1ab7874 | 4948 | the expression, and currently we'll be unable to do this. */ |
4949 | htab_empty (INSN_TRANSFORMED_INSNS (insn)); | |
4950 | } | |
4951 | } | |
4952 | ||
4953 | /* Add BB_NOTE to the pool of available basic block notes. */ | |
4954 | static void | |
4955 | return_bb_to_pool (basic_block bb) | |
4956 | { | |
4957 | rtx note = bb_note (bb); | |
4958 | ||
4959 | gcc_assert (NOTE_BASIC_BLOCK (note) == bb | |
4960 | && bb->aux == NULL); | |
4961 | ||
4962 | /* It turns out that current cfg infrastructure does not support | |
4963 | reuse of basic blocks. Don't bother for now. */ | |
f1f41a6c | 4964 | /*bb_note_pool.safe_push (note);*/ |
e1ab7874 | 4965 | } |
4966 | ||
4967 | /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ | |
4968 | static rtx | |
4969 | get_bb_note_from_pool (void) | |
4970 | { | |
f1f41a6c | 4971 | if (bb_note_pool.is_empty ()) |
e1ab7874 | 4972 | return NULL_RTX; |
4973 | else | |
4974 | { | |
f1f41a6c | 4975 | rtx note = bb_note_pool.pop (); |
e1ab7874 | 4976 | |
4977 | PREV_INSN (note) = NULL_RTX; | |
4978 | NEXT_INSN (note) = NULL_RTX; | |
4979 | ||
4980 | return note; | |
4981 | } | |
4982 | } | |
4983 | ||
4984 | /* Free bb_note_pool. */ | |
4985 | void | |
4986 | free_bb_note_pool (void) | |
4987 | { | |
f1f41a6c | 4988 | bb_note_pool.release (); |
e1ab7874 | 4989 | } |
4990 | ||
4991 | /* Setup scheduler pool and successor structure. */ | |
4992 | void | |
4993 | alloc_sched_pools (void) | |
4994 | { | |
4995 | int succs_size; | |
4996 | ||
4997 | succs_size = MAX_WS + 1; | |
48e1416a | 4998 | succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); |
e1ab7874 | 4999 | succs_info_pool.size = succs_size; |
5000 | succs_info_pool.top = -1; | |
5001 | succs_info_pool.max_top = -1; | |
5002 | ||
48e1416a | 5003 | sched_lists_pool = create_alloc_pool ("sel-sched-lists", |
e1ab7874 | 5004 | sizeof (struct _list_node), 500); |
5005 | } | |
5006 | ||
5007 | /* Free the pools. */ | |
5008 | void | |
5009 | free_sched_pools (void) | |
5010 | { | |
5011 | int i; | |
48e1416a | 5012 | |
e1ab7874 | 5013 | free_alloc_pool (sched_lists_pool); |
5014 | gcc_assert (succs_info_pool.top == -1); | |
862c1934 | 5015 | for (i = 0; i <= succs_info_pool.max_top; i++) |
e1ab7874 | 5016 | { |
f1f41a6c | 5017 | succs_info_pool.stack[i].succs_ok.release (); |
5018 | succs_info_pool.stack[i].succs_other.release (); | |
5019 | succs_info_pool.stack[i].probs_ok.release (); | |
e1ab7874 | 5020 | } |
5021 | free (succs_info_pool.stack); | |
5022 | } | |
5023 | \f | |
5024 | ||
48e1416a | 5025 | /* Returns a position in RGN where BB can be inserted retaining |
e1ab7874 | 5026 | topological order. */ |
5027 | static int | |
5028 | find_place_to_insert_bb (basic_block bb, int rgn) | |
5029 | { | |
5030 | bool has_preds_outside_rgn = false; | |
5031 | edge e; | |
5032 | edge_iterator ei; | |
48e1416a | 5033 | |
e1ab7874 | 5034 | /* Find whether we have preds outside the region. */ |
5035 | FOR_EACH_EDGE (e, ei, bb->preds) | |
5036 | if (!in_current_region_p (e->src)) | |
5037 | { | |
5038 | has_preds_outside_rgn = true; | |
5039 | break; | |
5040 | } | |
48e1416a | 5041 | |
e1ab7874 | 5042 | /* Recompute the top order -- needed when we have > 1 pred |
5043 | and in case we don't have preds outside. */ | |
5044 | if (flag_sel_sched_pipelining_outer_loops | |
5045 | && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) | |
5046 | { | |
5047 | int i, bbi = bb->index, cur_bbi; | |
5048 | ||
5049 | recompute_rev_top_order (); | |
5050 | for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) | |
5051 | { | |
5052 | cur_bbi = BB_TO_BLOCK (i); | |
48e1416a | 5053 | if (rev_top_order_index[bbi] |
e1ab7874 | 5054 | < rev_top_order_index[cur_bbi]) |
5055 | break; | |
5056 | } | |
48e1416a | 5057 | |
9d75589a | 5058 | /* We skipped the right block, so we increase i. We accommodate |
e1ab7874 | 5059 | it for increasing by step later, so we decrease i. */ |
5060 | return (i + 1) - 1; | |
5061 | } | |
5062 | else if (has_preds_outside_rgn) | |
5063 | { | |
5064 | /* This is the case when we generate an extra empty block | |
5065 | to serve as region head during pipelining. */ | |
5066 | e = EDGE_SUCC (bb, 0); | |
5067 | gcc_assert (EDGE_COUNT (bb->succs) == 1 | |
5068 | && in_current_region_p (EDGE_SUCC (bb, 0)->dest) | |
5069 | && (BLOCK_TO_BB (e->dest->index) == 0)); | |
5070 | return -1; | |
5071 | } | |
5072 | ||
5073 | /* We don't have preds outside the region. We should have | |
5074 | the only pred, because the multiple preds case comes from | |
5075 | the pipelining of outer loops, and that is handled above. | |
5076 | Just take the bbi of this single pred. */ | |
5077 | if (EDGE_COUNT (bb->succs) > 0) | |
5078 | { | |
5079 | int pred_bbi; | |
48e1416a | 5080 | |
e1ab7874 | 5081 | gcc_assert (EDGE_COUNT (bb->preds) == 1); |
48e1416a | 5082 | |
e1ab7874 | 5083 | pred_bbi = EDGE_PRED (bb, 0)->src->index; |
5084 | return BLOCK_TO_BB (pred_bbi); | |
5085 | } | |
5086 | else | |
5087 | /* BB has no successors. It is safe to put it in the end. */ | |
5088 | return current_nr_blocks - 1; | |
5089 | } | |
5090 | ||
5091 | /* Deletes an empty basic block freeing its data. */ | |
5092 | static void | |
5093 | delete_and_free_basic_block (basic_block bb) | |
5094 | { | |
5095 | gcc_assert (sel_bb_empty_p (bb)); | |
5096 | ||
5097 | if (BB_LV_SET (bb)) | |
5098 | free_lv_set (bb); | |
5099 | ||
5100 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
5101 | ||
48e1416a | 5102 | /* Can't assert av_set properties because we use sel_aremove_bb |
5103 | when removing loop preheader from the region. At the point of | |
e1ab7874 | 5104 | removing the preheader we already have deallocated sel_region_bb_info. */ |
5105 | gcc_assert (BB_LV_SET (bb) == NULL | |
5106 | && !BB_LV_SET_VALID_P (bb) | |
5107 | && BB_AV_LEVEL (bb) == 0 | |
5108 | && BB_AV_SET (bb) == NULL); | |
48e1416a | 5109 | |
e1ab7874 | 5110 | delete_basic_block (bb); |
5111 | } | |
5112 | ||
5113 | /* Add BB to the current region and update the region data. */ | |
5114 | static void | |
5115 | add_block_to_current_region (basic_block bb) | |
5116 | { | |
5117 | int i, pos, bbi = -2, rgn; | |
5118 | ||
5119 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
5120 | bbi = find_place_to_insert_bb (bb, rgn); | |
5121 | bbi += 1; | |
5122 | pos = RGN_BLOCKS (rgn) + bbi; | |
5123 | ||
5124 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
5125 | && ebb_head[bbi] == pos); | |
48e1416a | 5126 | |
e1ab7874 | 5127 | /* Make a place for the new block. */ |
5128 | extend_regions (); | |
5129 | ||
5130 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
5131 | BLOCK_TO_BB (rgn_bb_table[i])++; | |
48e1416a | 5132 | |
e1ab7874 | 5133 | memmove (rgn_bb_table + pos + 1, |
5134 | rgn_bb_table + pos, | |
5135 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5136 | ||
5137 | /* Initialize data for BB. */ | |
5138 | rgn_bb_table[pos] = bb->index; | |
5139 | BLOCK_TO_BB (bb->index) = bbi; | |
5140 | CONTAINING_RGN (bb->index) = rgn; | |
5141 | ||
5142 | RGN_NR_BLOCKS (rgn)++; | |
48e1416a | 5143 | |
e1ab7874 | 5144 | for (i = rgn + 1; i <= nr_regions; i++) |
5145 | RGN_BLOCKS (i)++; | |
5146 | } | |
5147 | ||
5148 | /* Remove BB from the current region and update the region data. */ | |
5149 | static void | |
5150 | remove_bb_from_region (basic_block bb) | |
5151 | { | |
5152 | int i, pos, bbi = -2, rgn; | |
5153 | ||
5154 | rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
5155 | bbi = BLOCK_TO_BB (bb->index); | |
5156 | pos = RGN_BLOCKS (rgn) + bbi; | |
5157 | ||
5158 | gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 | |
5159 | && ebb_head[bbi] == pos); | |
5160 | ||
5161 | for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) | |
5162 | BLOCK_TO_BB (rgn_bb_table[i])--; | |
5163 | ||
5164 | memmove (rgn_bb_table + pos, | |
5165 | rgn_bb_table + pos + 1, | |
5166 | (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); | |
5167 | ||
5168 | RGN_NR_BLOCKS (rgn)--; | |
5169 | for (i = rgn + 1; i <= nr_regions; i++) | |
5170 | RGN_BLOCKS (i)--; | |
5171 | } | |
5172 | ||
48e1416a | 5173 | /* Add BB to the current region and update all data. If BB is NULL, add all |
e1ab7874 | 5174 | blocks from last_added_blocks vector. */ |
5175 | static void | |
5176 | sel_add_bb (basic_block bb) | |
5177 | { | |
5178 | /* Extend luids so that new notes will receive zero luids. */ | |
52d7e28c | 5179 | sched_extend_luids (); |
e1ab7874 | 5180 | sched_init_bbs (); |
52d7e28c | 5181 | sel_init_bbs (last_added_blocks); |
e1ab7874 | 5182 | |
48e1416a | 5183 | /* When bb is passed explicitly, the vector should contain |
e1ab7874 | 5184 | the only element that equals to bb; otherwise, the vector |
5185 | should not be NULL. */ | |
f1f41a6c | 5186 | gcc_assert (last_added_blocks.exists ()); |
48e1416a | 5187 | |
e1ab7874 | 5188 | if (bb != NULL) |
5189 | { | |
f1f41a6c | 5190 | gcc_assert (last_added_blocks.length () == 1 |
5191 | && last_added_blocks[0] == bb); | |
e1ab7874 | 5192 | add_block_to_current_region (bb); |
5193 | ||
5194 | /* We associate creating/deleting data sets with the first insn | |
5195 | appearing / disappearing in the bb. */ | |
5196 | if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) | |
5197 | create_initial_data_sets (bb); | |
48e1416a | 5198 | |
f1f41a6c | 5199 | last_added_blocks.release (); |
e1ab7874 | 5200 | } |
5201 | else | |
5202 | /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ | |
5203 | { | |
5204 | int i; | |
5205 | basic_block temp_bb = NULL; | |
5206 | ||
48e1416a | 5207 | for (i = 0; |
f1f41a6c | 5208 | last_added_blocks.iterate (i, &bb); i++) |
e1ab7874 | 5209 | { |
5210 | add_block_to_current_region (bb); | |
5211 | temp_bb = bb; | |
5212 | } | |
5213 | ||
48e1416a | 5214 | /* We need to fetch at least one bb so we know the region |
e1ab7874 | 5215 | to update. */ |
5216 | gcc_assert (temp_bb != NULL); | |
5217 | bb = temp_bb; | |
5218 | ||
f1f41a6c | 5219 | last_added_blocks.release (); |
e1ab7874 | 5220 | } |
5221 | ||
5222 | rgn_setup_region (CONTAINING_RGN (bb->index)); | |
5223 | } | |
5224 | ||
48e1416a | 5225 | /* Remove BB from the current region and update all data. |
e1ab7874 | 5226 | If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ |
5227 | static void | |
5228 | sel_remove_bb (basic_block bb, bool remove_from_cfg_p) | |
5229 | { | |
0424f393 | 5230 | unsigned idx = bb->index; |
5231 | ||
e1ab7874 | 5232 | gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); |
48e1416a | 5233 | |
e1ab7874 | 5234 | remove_bb_from_region (bb); |
5235 | return_bb_to_pool (bb); | |
0424f393 | 5236 | bitmap_clear_bit (blocks_to_reschedule, idx); |
48e1416a | 5237 | |
e1ab7874 | 5238 | if (remove_from_cfg_p) |
1a5dbaab | 5239 | { |
5240 | basic_block succ = single_succ (bb); | |
5241 | delete_and_free_basic_block (bb); | |
5242 | set_immediate_dominator (CDI_DOMINATORS, succ, | |
5243 | recompute_dominator (CDI_DOMINATORS, succ)); | |
5244 | } | |
e1ab7874 | 5245 | |
0424f393 | 5246 | rgn_setup_region (CONTAINING_RGN (idx)); |
e1ab7874 | 5247 | } |
5248 | ||
5249 | /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ | |
5250 | static void | |
5251 | move_bb_info (basic_block merge_bb, basic_block empty_bb) | |
5252 | { | |
ef4cf572 | 5253 | if (in_current_region_p (merge_bb)) |
5254 | concat_note_lists (BB_NOTE_LIST (empty_bb), | |
5255 | &BB_NOTE_LIST (merge_bb)); | |
e1ab7874 | 5256 | BB_NOTE_LIST (empty_bb) = NULL_RTX; |
5257 | ||
5258 | } | |
5259 | ||
e1ab7874 | 5260 | /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from |
5261 | region, but keep it in CFG. */ | |
5262 | static void | |
5263 | remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) | |
5264 | { | |
5265 | /* The block should contain just a note or a label. | |
5266 | We try to check whether it is unused below. */ | |
5267 | gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) | |
5268 | || LABEL_P (BB_HEAD (empty_bb))); | |
5269 | ||
5270 | /* If basic block has predecessors or successors, redirect them. */ | |
5271 | if (remove_from_cfg_p | |
5272 | && (EDGE_COUNT (empty_bb->preds) > 0 | |
5273 | || EDGE_COUNT (empty_bb->succs) > 0)) | |
5274 | { | |
5275 | basic_block pred; | |
5276 | basic_block succ; | |
5277 | ||
5278 | /* We need to init PRED and SUCC before redirecting edges. */ | |
5279 | if (EDGE_COUNT (empty_bb->preds) > 0) | |
5280 | { | |
5281 | edge e; | |
5282 | ||
5283 | gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); | |
5284 | ||
5285 | e = EDGE_PRED (empty_bb, 0); | |
5286 | gcc_assert (e->src == empty_bb->prev_bb | |
5287 | && (e->flags & EDGE_FALLTHRU)); | |
5288 | ||
5289 | pred = empty_bb->prev_bb; | |
5290 | } | |
5291 | else | |
5292 | pred = NULL; | |
5293 | ||
5294 | if (EDGE_COUNT (empty_bb->succs) > 0) | |
5295 | { | |
5296 | /* We do not check fallthruness here as above, because | |
5297 | after removing a jump the edge may actually be not fallthru. */ | |
5298 | gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); | |
5299 | succ = EDGE_SUCC (empty_bb, 0)->dest; | |
5300 | } | |
5301 | else | |
5302 | succ = NULL; | |
5303 | ||
5304 | if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) | |
5305 | { | |
5306 | edge e = EDGE_PRED (empty_bb, 0); | |
5307 | ||
5308 | if (e->flags & EDGE_FALLTHRU) | |
5309 | redirect_edge_succ_nodup (e, succ); | |
5310 | else | |
5311 | sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); | |
5312 | } | |
5313 | ||
5314 | if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) | |
5315 | { | |
5316 | edge e = EDGE_SUCC (empty_bb, 0); | |
5317 | ||
5318 | if (find_edge (pred, e->dest) == NULL) | |
5319 | redirect_edge_pred (e, pred); | |
5320 | } | |
5321 | } | |
5322 | ||
5323 | /* Finish removing. */ | |
5324 | sel_remove_bb (empty_bb, remove_from_cfg_p); | |
5325 | } | |
5326 | ||
48e1416a | 5327 | /* An implementation of create_basic_block hook, which additionally updates |
e1ab7874 | 5328 | per-bb data structures. */ |
5329 | static basic_block | |
5330 | sel_create_basic_block (void *headp, void *endp, basic_block after) | |
5331 | { | |
5332 | basic_block new_bb; | |
5333 | insn_t new_bb_note; | |
48e1416a | 5334 | |
5335 | gcc_assert (flag_sel_sched_pipelining_outer_loops | |
f1f41a6c | 5336 | || !last_added_blocks.exists ()); |
e1ab7874 | 5337 | |
5338 | new_bb_note = get_bb_note_from_pool (); | |
5339 | ||
5340 | if (new_bb_note == NULL_RTX) | |
5341 | new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); | |
5342 | else | |
5343 | { | |
5344 | new_bb = create_basic_block_structure ((rtx) headp, (rtx) endp, | |
5345 | new_bb_note, after); | |
5346 | new_bb->aux = NULL; | |
5347 | } | |
5348 | ||
f1f41a6c | 5349 | last_added_blocks.safe_push (new_bb); |
e1ab7874 | 5350 | |
5351 | return new_bb; | |
5352 | } | |
5353 | ||
5354 | /* Implement sched_init_only_bb (). */ | |
5355 | static void | |
5356 | sel_init_only_bb (basic_block bb, basic_block after) | |
5357 | { | |
5358 | gcc_assert (after == NULL); | |
5359 | ||
5360 | extend_regions (); | |
5361 | rgn_make_new_region_out_of_new_block (bb); | |
5362 | } | |
5363 | ||
5364 | /* Update the latch when we've splitted or merged it from FROM block to TO. | |
5365 | This should be checked for all outer loops, too. */ | |
5366 | static void | |
5367 | change_loops_latches (basic_block from, basic_block to) | |
5368 | { | |
5369 | gcc_assert (from != to); | |
5370 | ||
5371 | if (current_loop_nest) | |
5372 | { | |
5373 | struct loop *loop; | |
5374 | ||
5375 | for (loop = current_loop_nest; loop; loop = loop_outer (loop)) | |
5376 | if (considered_for_pipelining_p (loop) && loop->latch == from) | |
5377 | { | |
5378 | gcc_assert (loop == current_loop_nest); | |
5379 | loop->latch = to; | |
5380 | gcc_assert (loop_latch_edge (loop)); | |
5381 | } | |
5382 | } | |
5383 | } | |
5384 | ||
48e1416a | 5385 | /* Splits BB on two basic blocks, adding it to the region and extending |
e1ab7874 | 5386 | per-bb data structures. Returns the newly created bb. */ |
5387 | static basic_block | |
5388 | sel_split_block (basic_block bb, rtx after) | |
5389 | { | |
5390 | basic_block new_bb; | |
5391 | insn_t insn; | |
5392 | ||
5393 | new_bb = sched_split_block_1 (bb, after); | |
5394 | sel_add_bb (new_bb); | |
5395 | ||
5396 | /* This should be called after sel_add_bb, because this uses | |
48e1416a | 5397 | CONTAINING_RGN for the new block, which is not yet initialized. |
e1ab7874 | 5398 | FIXME: this function may be a no-op now. */ |
5399 | change_loops_latches (bb, new_bb); | |
5400 | ||
5401 | /* Update ORIG_BB_INDEX for insns moved into the new block. */ | |
5402 | FOR_BB_INSNS (new_bb, insn) | |
5403 | if (INSN_P (insn)) | |
5404 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; | |
5405 | ||
5406 | if (sel_bb_empty_p (bb)) | |
5407 | { | |
5408 | gcc_assert (!sel_bb_empty_p (new_bb)); | |
5409 | ||
5410 | /* NEW_BB has data sets that need to be updated and BB holds | |
5411 | data sets that should be removed. Exchange these data sets | |
5412 | so that we won't lose BB's valid data sets. */ | |
5413 | exchange_data_sets (new_bb, bb); | |
5414 | free_data_sets (bb); | |
5415 | } | |
5416 | ||
5417 | if (!sel_bb_empty_p (new_bb) | |
5418 | && bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
5419 | bitmap_set_bit (blocks_to_reschedule, new_bb->index); | |
5420 | ||
5421 | return new_bb; | |
5422 | } | |
5423 | ||
5424 | /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. | |
5425 | Otherwise returns NULL. */ | |
5426 | static rtx | |
5427 | check_for_new_jump (basic_block bb, int prev_max_uid) | |
5428 | { | |
5429 | rtx end; | |
5430 | ||
5431 | end = sel_bb_end (bb); | |
5432 | if (end && INSN_UID (end) >= prev_max_uid) | |
5433 | return end; | |
5434 | return NULL; | |
5435 | } | |
5436 | ||
48e1416a | 5437 | /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. |
e1ab7874 | 5438 | New means having UID at least equal to PREV_MAX_UID. */ |
5439 | static rtx | |
5440 | find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) | |
5441 | { | |
5442 | rtx jump; | |
5443 | ||
5444 | /* Return immediately if no new insns were emitted. */ | |
5445 | if (get_max_uid () == prev_max_uid) | |
5446 | return NULL; | |
48e1416a | 5447 | |
e1ab7874 | 5448 | /* Now check both blocks for new jumps. It will ever be only one. */ |
5449 | if ((jump = check_for_new_jump (from, prev_max_uid))) | |
5450 | return jump; | |
5451 | ||
5452 | if (jump_bb != NULL | |
5453 | && (jump = check_for_new_jump (jump_bb, prev_max_uid))) | |
5454 | return jump; | |
5455 | return NULL; | |
5456 | } | |
5457 | ||
5458 | /* Splits E and adds the newly created basic block to the current region. | |
5459 | Returns this basic block. */ | |
5460 | basic_block | |
5461 | sel_split_edge (edge e) | |
5462 | { | |
5463 | basic_block new_bb, src, other_bb = NULL; | |
5464 | int prev_max_uid; | |
5465 | rtx jump; | |
5466 | ||
5467 | src = e->src; | |
5468 | prev_max_uid = get_max_uid (); | |
5469 | new_bb = split_edge (e); | |
5470 | ||
48e1416a | 5471 | if (flag_sel_sched_pipelining_outer_loops |
e1ab7874 | 5472 | && current_loop_nest) |
5473 | { | |
5474 | int i; | |
5475 | basic_block bb; | |
5476 | ||
48e1416a | 5477 | /* Some of the basic blocks might not have been added to the loop. |
e1ab7874 | 5478 | Add them here, until this is fixed in force_fallthru. */ |
48e1416a | 5479 | for (i = 0; |
f1f41a6c | 5480 | last_added_blocks.iterate (i, &bb); i++) |
e1ab7874 | 5481 | if (!bb->loop_father) |
5482 | { | |
5483 | add_bb_to_loop (bb, e->dest->loop_father); | |
5484 | ||
5485 | gcc_assert (!other_bb && (new_bb->index != bb->index)); | |
5486 | other_bb = bb; | |
5487 | } | |
5488 | } | |
5489 | ||
5490 | /* Add all last_added_blocks to the region. */ | |
5491 | sel_add_bb (NULL); | |
5492 | ||
5493 | jump = find_new_jump (src, new_bb, prev_max_uid); | |
5494 | if (jump) | |
5495 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5496 | ||
5497 | /* Put the correct lv set on this block. */ | |
5498 | if (other_bb && !sel_bb_empty_p (other_bb)) | |
5499 | compute_live (sel_bb_head (other_bb)); | |
5500 | ||
5501 | return new_bb; | |
5502 | } | |
5503 | ||
5504 | /* Implement sched_create_empty_bb (). */ | |
5505 | static basic_block | |
5506 | sel_create_empty_bb (basic_block after) | |
5507 | { | |
5508 | basic_block new_bb; | |
5509 | ||
5510 | new_bb = sched_create_empty_bb_1 (after); | |
5511 | ||
5512 | /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit | |
5513 | later. */ | |
f1f41a6c | 5514 | gcc_assert (last_added_blocks.length () == 1 |
5515 | && last_added_blocks[0] == new_bb); | |
e1ab7874 | 5516 | |
f1f41a6c | 5517 | last_added_blocks.release (); |
e1ab7874 | 5518 | return new_bb; |
5519 | } | |
5520 | ||
5521 | /* Implement sched_create_recovery_block. ORIG_INSN is where block | |
5522 | will be splitted to insert a check. */ | |
5523 | basic_block | |
5524 | sel_create_recovery_block (insn_t orig_insn) | |
5525 | { | |
5526 | basic_block first_bb, second_bb, recovery_block; | |
5527 | basic_block before_recovery = NULL; | |
5528 | rtx jump; | |
5529 | ||
5530 | first_bb = BLOCK_FOR_INSN (orig_insn); | |
5531 | if (sel_bb_end_p (orig_insn)) | |
5532 | { | |
5533 | /* Avoid introducing an empty block while splitting. */ | |
5534 | gcc_assert (single_succ_p (first_bb)); | |
5535 | second_bb = single_succ (first_bb); | |
5536 | } | |
5537 | else | |
5538 | second_bb = sched_split_block (first_bb, orig_insn); | |
5539 | ||
5540 | recovery_block = sched_create_recovery_block (&before_recovery); | |
5541 | if (before_recovery) | |
34154e27 | 5542 | copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e1ab7874 | 5543 | |
5544 | gcc_assert (sel_bb_empty_p (recovery_block)); | |
5545 | sched_create_recovery_edges (first_bb, recovery_block, second_bb); | |
5546 | if (current_loops != NULL) | |
5547 | add_bb_to_loop (recovery_block, first_bb->loop_father); | |
48e1416a | 5548 | |
e1ab7874 | 5549 | sel_add_bb (recovery_block); |
48e1416a | 5550 | |
e1ab7874 | 5551 | jump = BB_END (recovery_block); |
5552 | gcc_assert (sel_bb_head (recovery_block) == jump); | |
5553 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
5554 | ||
5555 | return recovery_block; | |
5556 | } | |
5557 | ||
5558 | /* Merge basic block B into basic block A. */ | |
0424f393 | 5559 | static void |
e1ab7874 | 5560 | sel_merge_blocks (basic_block a, basic_block b) |
5561 | { | |
0424f393 | 5562 | gcc_assert (sel_bb_empty_p (b) |
5563 | && EDGE_COUNT (b->preds) == 1 | |
5564 | && EDGE_PRED (b, 0)->src == b->prev_bb); | |
e1ab7874 | 5565 | |
0424f393 | 5566 | move_bb_info (b->prev_bb, b); |
5567 | remove_empty_bb (b, false); | |
5568 | merge_blocks (a, b); | |
e1ab7874 | 5569 | change_loops_latches (b, a); |
5570 | } | |
5571 | ||
5572 | /* A wrapper for redirect_edge_and_branch_force, which also initializes | |
5573 | data structures for possibly created bb and insns. Returns the newly | |
5574 | added bb or NULL, when a bb was not needed. */ | |
5575 | void | |
5576 | sel_redirect_edge_and_branch_force (edge e, basic_block to) | |
5577 | { | |
1a5dbaab | 5578 | basic_block jump_bb, src, orig_dest = e->dest; |
e1ab7874 | 5579 | int prev_max_uid; |
5580 | rtx jump; | |
48e1416a | 5581 | |
1a5dbaab | 5582 | /* This function is now used only for bookkeeping code creation, where |
5583 | we'll never get the single pred of orig_dest block and thus will not | |
5584 | hit unreachable blocks when updating dominator info. */ | |
5585 | gcc_assert (!sel_bb_empty_p (e->src) | |
5586 | && !single_pred_p (orig_dest)); | |
e1ab7874 | 5587 | src = e->src; |
5588 | prev_max_uid = get_max_uid (); | |
5589 | jump_bb = redirect_edge_and_branch_force (e, to); | |
5590 | ||
5591 | if (jump_bb != NULL) | |
5592 | sel_add_bb (jump_bb); | |
5593 | ||
5594 | /* This function could not be used to spoil the loop structure by now, | |
5595 | thus we don't care to update anything. But check it to be sure. */ | |
5596 | if (current_loop_nest | |
5597 | && pipelining_p) | |
5598 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
48e1416a | 5599 | |
e1ab7874 | 5600 | jump = find_new_jump (src, jump_bb, prev_max_uid); |
5601 | if (jump) | |
5602 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
1a5dbaab | 5603 | set_immediate_dominator (CDI_DOMINATORS, to, |
5604 | recompute_dominator (CDI_DOMINATORS, to)); | |
5605 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5606 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
e1ab7874 | 5607 | } |
5608 | ||
93919afc | 5609 | /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by |
5610 | redirected edge are in reverse topological order. */ | |
5611 | bool | |
e1ab7874 | 5612 | sel_redirect_edge_and_branch (edge e, basic_block to) |
5613 | { | |
5614 | bool latch_edge_p; | |
1a5dbaab | 5615 | basic_block src, orig_dest = e->dest; |
e1ab7874 | 5616 | int prev_max_uid; |
5617 | rtx jump; | |
df6266b9 | 5618 | edge redirected; |
93919afc | 5619 | bool recompute_toporder_p = false; |
1a5dbaab | 5620 | bool maybe_unreachable = single_pred_p (orig_dest); |
e1ab7874 | 5621 | |
5622 | latch_edge_p = (pipelining_p | |
5623 | && current_loop_nest | |
5624 | && e == loop_latch_edge (current_loop_nest)); | |
5625 | ||
5626 | src = e->src; | |
5627 | prev_max_uid = get_max_uid (); | |
df6266b9 | 5628 | |
5629 | redirected = redirect_edge_and_branch (e, to); | |
5630 | ||
f1f41a6c | 5631 | gcc_assert (redirected && !last_added_blocks.exists ()); |
e1ab7874 | 5632 | |
5633 | /* When we've redirected a latch edge, update the header. */ | |
5634 | if (latch_edge_p) | |
5635 | { | |
5636 | current_loop_nest->header = to; | |
5637 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
5638 | } | |
5639 | ||
93919afc | 5640 | /* In rare situations, the topological relation between the blocks connected |
5641 | by the redirected edge can change (see PR42245 for an example). Update | |
5642 | block_to_bb/bb_to_block. */ | |
5643 | if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) | |
5644 | && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) | |
5645 | recompute_toporder_p = true; | |
5646 | ||
e1ab7874 | 5647 | jump = find_new_jump (src, NULL, prev_max_uid); |
5648 | if (jump) | |
5649 | sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); | |
93919afc | 5650 | |
1a5dbaab | 5651 | /* Only update dominator info when we don't have unreachable blocks. |
5652 | Otherwise we'll update in maybe_tidy_empty_bb. */ | |
5653 | if (!maybe_unreachable) | |
5654 | { | |
5655 | set_immediate_dominator (CDI_DOMINATORS, to, | |
5656 | recompute_dominator (CDI_DOMINATORS, to)); | |
5657 | set_immediate_dominator (CDI_DOMINATORS, orig_dest, | |
5658 | recompute_dominator (CDI_DOMINATORS, orig_dest)); | |
5659 | } | |
93919afc | 5660 | return recompute_toporder_p; |
e1ab7874 | 5661 | } |
5662 | ||
5663 | /* This variable holds the cfg hooks used by the selective scheduler. */ | |
5664 | static struct cfg_hooks sel_cfg_hooks; | |
5665 | ||
5666 | /* Register sel-sched cfg hooks. */ | |
5667 | void | |
5668 | sel_register_cfg_hooks (void) | |
5669 | { | |
5670 | sched_split_block = sel_split_block; | |
5671 | ||
5672 | orig_cfg_hooks = get_cfg_hooks (); | |
5673 | sel_cfg_hooks = orig_cfg_hooks; | |
5674 | ||
5675 | sel_cfg_hooks.create_basic_block = sel_create_basic_block; | |
5676 | ||
5677 | set_cfg_hooks (sel_cfg_hooks); | |
5678 | ||
5679 | sched_init_only_bb = sel_init_only_bb; | |
5680 | sched_split_block = sel_split_block; | |
5681 | sched_create_empty_bb = sel_create_empty_bb; | |
5682 | } | |
5683 | ||
5684 | /* Unregister sel-sched cfg hooks. */ | |
5685 | void | |
5686 | sel_unregister_cfg_hooks (void) | |
5687 | { | |
5688 | sched_create_empty_bb = NULL; | |
5689 | sched_split_block = NULL; | |
5690 | sched_init_only_bb = NULL; | |
5691 | ||
5692 | set_cfg_hooks (orig_cfg_hooks); | |
5693 | } | |
5694 | \f | |
5695 | ||
5696 | /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, | |
5697 | LABEL is where this jump should be directed. */ | |
5698 | rtx | |
5699 | create_insn_rtx_from_pattern (rtx pattern, rtx label) | |
5700 | { | |
5701 | rtx insn_rtx; | |
5702 | ||
5703 | gcc_assert (!INSN_P (pattern)); | |
5704 | ||
5705 | start_sequence (); | |
5706 | ||
5707 | if (label == NULL_RTX) | |
5708 | insn_rtx = emit_insn (pattern); | |
9845d120 | 5709 | else if (DEBUG_INSN_P (label)) |
5710 | insn_rtx = emit_debug_insn (pattern); | |
e1ab7874 | 5711 | else |
5712 | { | |
5713 | insn_rtx = emit_jump_insn (pattern); | |
5714 | JUMP_LABEL (insn_rtx) = label; | |
5715 | ++LABEL_NUSES (label); | |
5716 | } | |
5717 | ||
5718 | end_sequence (); | |
5719 | ||
52d7e28c | 5720 | sched_extend_luids (); |
e1ab7874 | 5721 | sched_extend_target (); |
5722 | sched_deps_init (false); | |
5723 | ||
5724 | /* Initialize INSN_CODE now. */ | |
5725 | recog_memoized (insn_rtx); | |
5726 | return insn_rtx; | |
5727 | } | |
5728 | ||
5729 | /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn | |
5730 | must not be clonable. */ | |
5731 | vinsn_t | |
5732 | create_vinsn_from_insn_rtx (rtx insn_rtx, bool force_unique_p) | |
5733 | { | |
5734 | gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); | |
5735 | ||
5736 | /* If VINSN_TYPE is not USE, retain its uniqueness. */ | |
5737 | return vinsn_create (insn_rtx, force_unique_p); | |
5738 | } | |
5739 | ||
5740 | /* Create a copy of INSN_RTX. */ | |
5741 | rtx | |
5742 | create_copy_of_insn_rtx (rtx insn_rtx) | |
5743 | { | |
114c1eb1 | 5744 | rtx res, link; |
e1ab7874 | 5745 | |
9845d120 | 5746 | if (DEBUG_INSN_P (insn_rtx)) |
5747 | return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5748 | insn_rtx); | |
5749 | ||
e1ab7874 | 5750 | gcc_assert (NONJUMP_INSN_P (insn_rtx)); |
5751 | ||
5752 | res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), | |
5753 | NULL_RTX); | |
114c1eb1 | 5754 | |
5755 | /* Copy all REG_NOTES except REG_EQUAL/REG_EQUIV and REG_LABEL_OPERAND | |
5756 | since mark_jump_label will make them. REG_LABEL_TARGETs are created | |
5757 | there too, but are supposed to be sticky, so we copy them. */ | |
5758 | for (link = REG_NOTES (insn_rtx); link; link = XEXP (link, 1)) | |
5759 | if (REG_NOTE_KIND (link) != REG_LABEL_OPERAND | |
5760 | && REG_NOTE_KIND (link) != REG_EQUAL | |
5761 | && REG_NOTE_KIND (link) != REG_EQUIV) | |
5762 | { | |
5763 | if (GET_CODE (link) == EXPR_LIST) | |
5764 | add_reg_note (res, REG_NOTE_KIND (link), | |
5765 | copy_insn_1 (XEXP (link, 0))); | |
5766 | else | |
5767 | add_reg_note (res, REG_NOTE_KIND (link), XEXP (link, 0)); | |
5768 | } | |
5769 | ||
e1ab7874 | 5770 | return res; |
5771 | } | |
5772 | ||
5773 | /* Change vinsn field of EXPR to hold NEW_VINSN. */ | |
5774 | void | |
5775 | change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) | |
5776 | { | |
5777 | vinsn_detach (EXPR_VINSN (expr)); | |
5778 | ||
5779 | EXPR_VINSN (expr) = new_vinsn; | |
5780 | vinsn_attach (new_vinsn); | |
5781 | } | |
5782 | ||
5783 | /* Helpers for global init. */ | |
5784 | /* This structure is used to be able to call existing bundling mechanism | |
5785 | and calculate insn priorities. */ | |
48e1416a | 5786 | static struct haifa_sched_info sched_sel_haifa_sched_info = |
e1ab7874 | 5787 | { |
5788 | NULL, /* init_ready_list */ | |
5789 | NULL, /* can_schedule_ready_p */ | |
5790 | NULL, /* schedule_more_p */ | |
5791 | NULL, /* new_ready */ | |
5792 | NULL, /* rgn_rank */ | |
5793 | sel_print_insn, /* rgn_print_insn */ | |
5794 | contributes_to_priority, | |
4db82bc9 | 5795 | NULL, /* insn_finishes_block_p */ |
e1ab7874 | 5796 | |
5797 | NULL, NULL, | |
5798 | NULL, NULL, | |
5799 | 0, 0, | |
5800 | ||
5801 | NULL, /* add_remove_insn */ | |
5802 | NULL, /* begin_schedule_ready */ | |
d2412f57 | 5803 | NULL, /* begin_move_insn */ |
e1ab7874 | 5804 | NULL, /* advance_target_bb */ |
e2f4a6ff | 5805 | |
5806 | NULL, | |
5807 | NULL, | |
5808 | ||
e1ab7874 | 5809 | SEL_SCHED | NEW_BBS |
5810 | }; | |
5811 | ||
5812 | /* Setup special insns used in the scheduler. */ | |
48e1416a | 5813 | void |
e1ab7874 | 5814 | setup_nop_and_exit_insns (void) |
5815 | { | |
5816 | gcc_assert (nop_pattern == NULL_RTX | |
5817 | && exit_insn == NULL_RTX); | |
5818 | ||
bc9cb5ed | 5819 | nop_pattern = constm1_rtx; |
e1ab7874 | 5820 | |
5821 | start_sequence (); | |
5822 | emit_insn (nop_pattern); | |
5823 | exit_insn = get_insns (); | |
5824 | end_sequence (); | |
34154e27 | 5825 | set_block_for_insn (exit_insn, EXIT_BLOCK_PTR_FOR_FN (cfun)); |
e1ab7874 | 5826 | } |
5827 | ||
5828 | /* Free special insns used in the scheduler. */ | |
5829 | void | |
5830 | free_nop_and_exit_insns (void) | |
5831 | { | |
5832 | exit_insn = NULL_RTX; | |
5833 | nop_pattern = NULL_RTX; | |
5834 | } | |
5835 | ||
5836 | /* Setup a special vinsn used in new insns initialization. */ | |
5837 | void | |
5838 | setup_nop_vinsn (void) | |
5839 | { | |
5840 | nop_vinsn = vinsn_create (exit_insn, false); | |
5841 | vinsn_attach (nop_vinsn); | |
5842 | } | |
5843 | ||
5844 | /* Free a special vinsn used in new insns initialization. */ | |
5845 | void | |
5846 | free_nop_vinsn (void) | |
5847 | { | |
5848 | gcc_assert (VINSN_COUNT (nop_vinsn) == 1); | |
5849 | vinsn_detach (nop_vinsn); | |
5850 | nop_vinsn = NULL; | |
5851 | } | |
5852 | ||
5853 | /* Call a set_sched_flags hook. */ | |
5854 | void | |
5855 | sel_set_sched_flags (void) | |
5856 | { | |
48e1416a | 5857 | /* ??? This means that set_sched_flags were called, and we decided to |
e1ab7874 | 5858 | support speculation. However, set_sched_flags also modifies flags |
48e1416a | 5859 | on current_sched_info, doing this only at global init. And we |
e1ab7874 | 5860 | sometimes change c_s_i later. So put the correct flags again. */ |
5861 | if (spec_info && targetm.sched.set_sched_flags) | |
5862 | targetm.sched.set_sched_flags (spec_info); | |
5863 | } | |
5864 | ||
5865 | /* Setup pointers to global sched info structures. */ | |
5866 | void | |
5867 | sel_setup_sched_infos (void) | |
5868 | { | |
5869 | rgn_setup_common_sched_info (); | |
5870 | ||
5871 | memcpy (&sel_common_sched_info, common_sched_info, | |
5872 | sizeof (sel_common_sched_info)); | |
5873 | ||
5874 | sel_common_sched_info.fix_recovery_cfg = NULL; | |
5875 | sel_common_sched_info.add_block = NULL; | |
5876 | sel_common_sched_info.estimate_number_of_insns | |
5877 | = sel_estimate_number_of_insns; | |
5878 | sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; | |
5879 | sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; | |
5880 | ||
5881 | common_sched_info = &sel_common_sched_info; | |
5882 | ||
5883 | current_sched_info = &sched_sel_haifa_sched_info; | |
48e1416a | 5884 | current_sched_info->sched_max_insns_priority = |
e1ab7874 | 5885 | get_rgn_sched_max_insns_priority (); |
48e1416a | 5886 | |
e1ab7874 | 5887 | sel_set_sched_flags (); |
5888 | } | |
5889 | \f | |
5890 | ||
5891 | /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, | |
5892 | *BB_ORD_INDEX after that is increased. */ | |
5893 | static void | |
5894 | sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) | |
5895 | { | |
5896 | RGN_NR_BLOCKS (rgn) += 1; | |
5897 | RGN_DONT_CALC_DEPS (rgn) = 0; | |
5898 | RGN_HAS_REAL_EBB (rgn) = 0; | |
5899 | CONTAINING_RGN (bb->index) = rgn; | |
5900 | BLOCK_TO_BB (bb->index) = *bb_ord_index; | |
5901 | rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; | |
5902 | (*bb_ord_index)++; | |
5903 | ||
5904 | /* FIXME: it is true only when not scheduling ebbs. */ | |
5905 | RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); | |
5906 | } | |
5907 | ||
5908 | /* Functions to support pipelining of outer loops. */ | |
5909 | ||
5910 | /* Creates a new empty region and returns it's number. */ | |
5911 | static int | |
5912 | sel_create_new_region (void) | |
5913 | { | |
5914 | int new_rgn_number = nr_regions; | |
5915 | ||
5916 | RGN_NR_BLOCKS (new_rgn_number) = 0; | |
5917 | ||
5918 | /* FIXME: This will work only when EBBs are not created. */ | |
5919 | if (new_rgn_number != 0) | |
48e1416a | 5920 | RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + |
e1ab7874 | 5921 | RGN_NR_BLOCKS (new_rgn_number - 1); |
5922 | else | |
5923 | RGN_BLOCKS (new_rgn_number) = 0; | |
5924 | ||
5925 | /* Set the blocks of the next region so the other functions may | |
5926 | calculate the number of blocks in the region. */ | |
48e1416a | 5927 | RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + |
e1ab7874 | 5928 | RGN_NR_BLOCKS (new_rgn_number); |
5929 | ||
5930 | nr_regions++; | |
5931 | ||
5932 | return new_rgn_number; | |
5933 | } | |
5934 | ||
5935 | /* If X has a smaller topological sort number than Y, returns -1; | |
5936 | if greater, returns 1. */ | |
5937 | static int | |
5938 | bb_top_order_comparator (const void *x, const void *y) | |
5939 | { | |
5940 | basic_block bb1 = *(const basic_block *) x; | |
5941 | basic_block bb2 = *(const basic_block *) y; | |
5942 | ||
48e1416a | 5943 | gcc_assert (bb1 == bb2 |
5944 | || rev_top_order_index[bb1->index] | |
e1ab7874 | 5945 | != rev_top_order_index[bb2->index]); |
5946 | ||
5947 | /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so | |
5948 | bbs with greater number should go earlier. */ | |
5949 | if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) | |
5950 | return -1; | |
5951 | else | |
5952 | return 1; | |
5953 | } | |
5954 | ||
48e1416a | 5955 | /* Create a region for LOOP and return its number. If we don't want |
e1ab7874 | 5956 | to pipeline LOOP, return -1. */ |
5957 | static int | |
5958 | make_region_from_loop (struct loop *loop) | |
5959 | { | |
5960 | unsigned int i; | |
5961 | int new_rgn_number = -1; | |
5962 | struct loop *inner; | |
5963 | ||
5964 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
5965 | int bb_ord_index = 0; | |
5966 | basic_block *loop_blocks; | |
5967 | basic_block preheader_block; | |
5968 | ||
48e1416a | 5969 | if (loop->num_nodes |
e1ab7874 | 5970 | > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS)) |
5971 | return -1; | |
48e1416a | 5972 | |
e1ab7874 | 5973 | /* Don't pipeline loops whose latch belongs to some of its inner loops. */ |
5974 | for (inner = loop->inner; inner; inner = inner->inner) | |
5975 | if (flow_bb_inside_loop_p (inner, loop->latch)) | |
5976 | return -1; | |
5977 | ||
5978 | loop->ninsns = num_loop_insns (loop); | |
5979 | if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS)) | |
5980 | return -1; | |
5981 | ||
5982 | loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); | |
5983 | ||
5984 | for (i = 0; i < loop->num_nodes; i++) | |
5985 | if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) | |
5986 | { | |
5987 | free (loop_blocks); | |
5988 | return -1; | |
5989 | } | |
5990 | ||
5991 | preheader_block = loop_preheader_edge (loop)->src; | |
5992 | gcc_assert (preheader_block); | |
5993 | gcc_assert (loop_blocks[0] == loop->header); | |
5994 | ||
5995 | new_rgn_number = sel_create_new_region (); | |
5996 | ||
5997 | sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number); | |
08b7917c | 5998 | bitmap_set_bit (bbs_in_loop_rgns, preheader_block->index); |
e1ab7874 | 5999 | |
6000 | for (i = 0; i < loop->num_nodes; i++) | |
6001 | { | |
6002 | /* Add only those blocks that haven't been scheduled in the inner loop. | |
6003 | The exception is the basic blocks with bookkeeping code - they should | |
48e1416a | 6004 | be added to the region (and they actually don't belong to the loop |
e1ab7874 | 6005 | body, but to the region containing that loop body). */ |
6006 | ||
6007 | gcc_assert (new_rgn_number >= 0); | |
6008 | ||
08b7917c | 6009 | if (! bitmap_bit_p (bbs_in_loop_rgns, loop_blocks[i]->index)) |
e1ab7874 | 6010 | { |
48e1416a | 6011 | sel_add_block_to_region (loop_blocks[i], &bb_ord_index, |
e1ab7874 | 6012 | new_rgn_number); |
08b7917c | 6013 | bitmap_set_bit (bbs_in_loop_rgns, loop_blocks[i]->index); |
e1ab7874 | 6014 | } |
6015 | } | |
6016 | ||
6017 | free (loop_blocks); | |
6018 | MARK_LOOP_FOR_PIPELINING (loop); | |
6019 | ||
6020 | return new_rgn_number; | |
6021 | } | |
6022 | ||
6023 | /* Create a new region from preheader blocks LOOP_BLOCKS. */ | |
6024 | void | |
f1f41a6c | 6025 | make_region_from_loop_preheader (vec<basic_block> *&loop_blocks) |
e1ab7874 | 6026 | { |
6027 | unsigned int i; | |
6028 | int new_rgn_number = -1; | |
6029 | basic_block bb; | |
6030 | ||
6031 | /* Basic block index, to be assigned to BLOCK_TO_BB. */ | |
6032 | int bb_ord_index = 0; | |
6033 | ||
6034 | new_rgn_number = sel_create_new_region (); | |
6035 | ||
f1f41a6c | 6036 | FOR_EACH_VEC_ELT (*loop_blocks, i, bb) |
e1ab7874 | 6037 | { |
6038 | gcc_assert (new_rgn_number >= 0); | |
6039 | ||
6040 | sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number); | |
6041 | } | |
6042 | ||
f1f41a6c | 6043 | vec_free (loop_blocks); |
e1ab7874 | 6044 | } |
6045 | ||
6046 | ||
6047 | /* Create region(s) from loop nest LOOP, such that inner loops will be | |
48e1416a | 6048 | pipelined before outer loops. Returns true when a region for LOOP |
e1ab7874 | 6049 | is created. */ |
6050 | static bool | |
6051 | make_regions_from_loop_nest (struct loop *loop) | |
48e1416a | 6052 | { |
e1ab7874 | 6053 | struct loop *cur_loop; |
6054 | int rgn_number; | |
6055 | ||
6056 | /* Traverse all inner nodes of the loop. */ | |
6057 | for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) | |
08b7917c | 6058 | if (! bitmap_bit_p (bbs_in_loop_rgns, cur_loop->header->index)) |
e1ab7874 | 6059 | return false; |
6060 | ||
6061 | /* At this moment all regular inner loops should have been pipelined. | |
6062 | Try to create a region from this loop. */ | |
6063 | rgn_number = make_region_from_loop (loop); | |
6064 | ||
6065 | if (rgn_number < 0) | |
6066 | return false; | |
6067 | ||
f1f41a6c | 6068 | loop_nests.safe_push (loop); |
e1ab7874 | 6069 | return true; |
6070 | } | |
6071 | ||
6072 | /* Initalize data structures needed. */ | |
6073 | void | |
6074 | sel_init_pipelining (void) | |
6075 | { | |
6076 | /* Collect loop information to be used in outer loops pipelining. */ | |
6077 | loop_optimizer_init (LOOPS_HAVE_PREHEADERS | |
6078 | | LOOPS_HAVE_FALLTHRU_PREHEADERS | |
6079 | | LOOPS_HAVE_RECORDED_EXITS | |
6080 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); | |
6081 | current_loop_nest = NULL; | |
6082 | ||
fe672ac0 | 6083 | bbs_in_loop_rgns = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
53c5d9d4 | 6084 | bitmap_clear (bbs_in_loop_rgns); |
e1ab7874 | 6085 | |
6086 | recompute_rev_top_order (); | |
6087 | } | |
6088 | ||
6089 | /* Returns a struct loop for region RGN. */ | |
6090 | loop_p | |
6091 | get_loop_nest_for_rgn (unsigned int rgn) | |
6092 | { | |
6093 | /* Regions created with extend_rgns don't have corresponding loop nests, | |
6094 | because they don't represent loops. */ | |
f1f41a6c | 6095 | if (rgn < loop_nests.length ()) |
6096 | return loop_nests[rgn]; | |
e1ab7874 | 6097 | else |
6098 | return NULL; | |
6099 | } | |
6100 | ||
6101 | /* True when LOOP was included into pipelining regions. */ | |
6102 | bool | |
6103 | considered_for_pipelining_p (struct loop *loop) | |
6104 | { | |
6105 | if (loop_depth (loop) == 0) | |
6106 | return false; | |
6107 | ||
48e1416a | 6108 | /* Now, the loop could be too large or irreducible. Check whether its |
6109 | region is in LOOP_NESTS. | |
6110 | We determine the region number of LOOP as the region number of its | |
6111 | latch. We can't use header here, because this header could be | |
e1ab7874 | 6112 | just removed preheader and it will give us the wrong region number. |
6113 | Latch can't be used because it could be in the inner loop too. */ | |
a2d56a0e | 6114 | if (LOOP_MARKED_FOR_PIPELINING_P (loop)) |
e1ab7874 | 6115 | { |
6116 | int rgn = CONTAINING_RGN (loop->latch->index); | |
6117 | ||
f1f41a6c | 6118 | gcc_assert ((unsigned) rgn < loop_nests.length ()); |
e1ab7874 | 6119 | return true; |
6120 | } | |
48e1416a | 6121 | |
e1ab7874 | 6122 | return false; |
6123 | } | |
6124 | ||
48e1416a | 6125 | /* Makes regions from the rest of the blocks, after loops are chosen |
e1ab7874 | 6126 | for pipelining. */ |
6127 | static void | |
6128 | make_regions_from_the_rest (void) | |
6129 | { | |
6130 | int cur_rgn_blocks; | |
6131 | int *loop_hdr; | |
6132 | int i; | |
6133 | ||
6134 | basic_block bb; | |
6135 | edge e; | |
6136 | edge_iterator ei; | |
6137 | int *degree; | |
e1ab7874 | 6138 | |
6139 | /* Index in rgn_bb_table where to start allocating new regions. */ | |
6140 | cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; | |
e1ab7874 | 6141 | |
48e1416a | 6142 | /* Make regions from all the rest basic blocks - those that don't belong to |
e1ab7874 | 6143 | any loop or belong to irreducible loops. Prepare the data structures |
6144 | for extend_rgns. */ | |
6145 | ||
6146 | /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, | |
6147 | LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same | |
6148 | loop. */ | |
fe672ac0 | 6149 | loop_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
6150 | degree = XCNEWVEC (int, last_basic_block_for_fn (cfun)); | |
e1ab7874 | 6151 | |
6152 | ||
6153 | /* For each basic block that belongs to some loop assign the number | |
6154 | of innermost loop it belongs to. */ | |
fe672ac0 | 6155 | for (i = 0; i < last_basic_block_for_fn (cfun); i++) |
e1ab7874 | 6156 | loop_hdr[i] = -1; |
6157 | ||
fc00614f | 6158 | FOR_EACH_BB_FN (bb, cfun) |
e1ab7874 | 6159 | { |
6160 | if (bb->loop_father && !bb->loop_father->num == 0 | |
6161 | && !(bb->flags & BB_IRREDUCIBLE_LOOP)) | |
6162 | loop_hdr[bb->index] = bb->loop_father->num; | |
6163 | } | |
6164 | ||
48e1416a | 6165 | /* For each basic block degree is calculated as the number of incoming |
e1ab7874 | 6166 | edges, that are going out of bbs that are not yet scheduled. |
6167 | The basic blocks that are scheduled have degree value of zero. */ | |
fc00614f | 6168 | FOR_EACH_BB_FN (bb, cfun) |
e1ab7874 | 6169 | { |
6170 | degree[bb->index] = 0; | |
6171 | ||
08b7917c | 6172 | if (!bitmap_bit_p (bbs_in_loop_rgns, bb->index)) |
e1ab7874 | 6173 | { |
6174 | FOR_EACH_EDGE (e, ei, bb->preds) | |
08b7917c | 6175 | if (!bitmap_bit_p (bbs_in_loop_rgns, e->src->index)) |
e1ab7874 | 6176 | degree[bb->index]++; |
6177 | } | |
6178 | else | |
6179 | degree[bb->index] = -1; | |
6180 | } | |
6181 | ||
6182 | extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); | |
6183 | ||
6184 | /* Any block that did not end up in a region is placed into a region | |
6185 | by itself. */ | |
fc00614f | 6186 | FOR_EACH_BB_FN (bb, cfun) |
e1ab7874 | 6187 | if (degree[bb->index] >= 0) |
6188 | { | |
6189 | rgn_bb_table[cur_rgn_blocks] = bb->index; | |
6190 | RGN_NR_BLOCKS (nr_regions) = 1; | |
6191 | RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; | |
6192 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
6193 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
6194 | CONTAINING_RGN (bb->index) = nr_regions++; | |
6195 | BLOCK_TO_BB (bb->index) = 0; | |
6196 | } | |
6197 | ||
6198 | free (degree); | |
6199 | free (loop_hdr); | |
6200 | } | |
6201 | ||
6202 | /* Free data structures used in pipelining of loops. */ | |
6203 | void sel_finish_pipelining (void) | |
6204 | { | |
e1ab7874 | 6205 | struct loop *loop; |
6206 | ||
6207 | /* Release aux fields so we don't free them later by mistake. */ | |
f21d4d00 | 6208 | FOR_EACH_LOOP (loop, 0) |
e1ab7874 | 6209 | loop->aux = NULL; |
6210 | ||
6211 | loop_optimizer_finalize (); | |
6212 | ||
f1f41a6c | 6213 | loop_nests.release (); |
e1ab7874 | 6214 | |
6215 | free (rev_top_order_index); | |
6216 | rev_top_order_index = NULL; | |
6217 | } | |
6218 | ||
48e1416a | 6219 | /* This function replaces the find_rgns when |
e1ab7874 | 6220 | FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ |
48e1416a | 6221 | void |
e1ab7874 | 6222 | sel_find_rgns (void) |
6223 | { | |
6224 | sel_init_pipelining (); | |
6225 | extend_regions (); | |
6226 | ||
6227 | if (current_loops) | |
6228 | { | |
6229 | loop_p loop; | |
e1ab7874 | 6230 | |
f21d4d00 | 6231 | FOR_EACH_LOOP (loop, (flag_sel_sched_pipelining_outer_loops |
6232 | ? LI_FROM_INNERMOST | |
6233 | : LI_ONLY_INNERMOST)) | |
e1ab7874 | 6234 | make_regions_from_loop_nest (loop); |
6235 | } | |
6236 | ||
6237 | /* Make regions from all the rest basic blocks and schedule them. | |
48e1416a | 6238 | These blocks include blocks that don't belong to any loop or belong |
e1ab7874 | 6239 | to irreducible loops. */ |
6240 | make_regions_from_the_rest (); | |
6241 | ||
6242 | /* We don't need bbs_in_loop_rgns anymore. */ | |
6243 | sbitmap_free (bbs_in_loop_rgns); | |
6244 | bbs_in_loop_rgns = NULL; | |
6245 | } | |
6246 | ||
b73edd22 | 6247 | /* Add the preheader blocks from previous loop to current region taking |
6248 | it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS. | |
e1ab7874 | 6249 | This function is only used with -fsel-sched-pipelining-outer-loops. */ |
6250 | void | |
b73edd22 | 6251 | sel_add_loop_preheaders (bb_vec_t *bbs) |
e1ab7874 | 6252 | { |
6253 | int i; | |
6254 | basic_block bb; | |
f1f41a6c | 6255 | vec<basic_block> *preheader_blocks |
e1ab7874 | 6256 | = LOOP_PREHEADER_BLOCKS (current_loop_nest); |
6257 | ||
f1f41a6c | 6258 | if (!preheader_blocks) |
6259 | return; | |
6260 | ||
6261 | for (i = 0; preheader_blocks->iterate (i, &bb); i++) | |
a2d56a0e | 6262 | { |
f1f41a6c | 6263 | bbs->safe_push (bb); |
6264 | last_added_blocks.safe_push (bb); | |
e1ab7874 | 6265 | sel_add_bb (bb); |
a2d56a0e | 6266 | } |
e1ab7874 | 6267 | |
f1f41a6c | 6268 | vec_free (preheader_blocks); |
e1ab7874 | 6269 | } |
6270 | ||
48e1416a | 6271 | /* While pipelining outer loops, returns TRUE if BB is a loop preheader. |
6272 | Please note that the function should also work when pipelining_p is | |
6273 | false, because it is used when deciding whether we should or should | |
e1ab7874 | 6274 | not reschedule pipelined code. */ |
6275 | bool | |
6276 | sel_is_loop_preheader_p (basic_block bb) | |
6277 | { | |
6278 | if (current_loop_nest) | |
6279 | { | |
6280 | struct loop *outer; | |
6281 | ||
6282 | if (preheader_removed) | |
6283 | return false; | |
6284 | ||
6285 | /* Preheader is the first block in the region. */ | |
6286 | if (BLOCK_TO_BB (bb->index) == 0) | |
6287 | return true; | |
6288 | ||
6289 | /* We used to find a preheader with the topological information. | |
6290 | Check that the above code is equivalent to what we did before. */ | |
6291 | ||
6292 | if (in_current_region_p (current_loop_nest->header)) | |
48e1416a | 6293 | gcc_assert (!(BLOCK_TO_BB (bb->index) |
e1ab7874 | 6294 | < BLOCK_TO_BB (current_loop_nest->header->index))); |
6295 | ||
6296 | /* Support the situation when the latch block of outer loop | |
6297 | could be from here. */ | |
6298 | for (outer = loop_outer (current_loop_nest); | |
6299 | outer; | |
6300 | outer = loop_outer (outer)) | |
6301 | if (considered_for_pipelining_p (outer) && outer->latch == bb) | |
6302 | gcc_unreachable (); | |
6303 | } | |
6304 | ||
6305 | return false; | |
6306 | } | |
6307 | ||
49087fba | 6308 | /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and |
6309 | can be removed, making the corresponding edge fallthrough (assuming that | |
6310 | all basic blocks between JUMP_BB and DEST_BB are empty). */ | |
6311 | static bool | |
6312 | bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb) | |
e1ab7874 | 6313 | { |
4b816303 | 6314 | if (!onlyjump_p (BB_END (jump_bb)) |
6315 | || tablejump_p (BB_END (jump_bb), NULL, NULL)) | |
e1ab7874 | 6316 | return false; |
6317 | ||
48e1416a | 6318 | /* Several outgoing edges, abnormal edge or destination of jump is |
e1ab7874 | 6319 | not DEST_BB. */ |
6320 | if (EDGE_COUNT (jump_bb->succs) != 1 | |
49087fba | 6321 | || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING) |
e1ab7874 | 6322 | || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) |
6323 | return false; | |
6324 | ||
6325 | /* If not anything of the upper. */ | |
6326 | return true; | |
6327 | } | |
6328 | ||
6329 | /* Removes the loop preheader from the current region and saves it in | |
48e1416a | 6330 | PREHEADER_BLOCKS of the father loop, so they will be added later to |
e1ab7874 | 6331 | region that represents an outer loop. */ |
6332 | static void | |
6333 | sel_remove_loop_preheader (void) | |
6334 | { | |
6335 | int i, old_len; | |
6336 | int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); | |
6337 | basic_block bb; | |
6338 | bool all_empty_p = true; | |
f1f41a6c | 6339 | vec<basic_block> *preheader_blocks |
e1ab7874 | 6340 | = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); |
6341 | ||
f1f41a6c | 6342 | vec_check_alloc (preheader_blocks, 0); |
6343 | ||
e1ab7874 | 6344 | gcc_assert (current_loop_nest); |
f1f41a6c | 6345 | old_len = preheader_blocks->length (); |
e1ab7874 | 6346 | |
6347 | /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ | |
6348 | for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) | |
6349 | { | |
f5a6b05f | 6350 | bb = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
e1ab7874 | 6351 | |
48e1416a | 6352 | /* If the basic block belongs to region, but doesn't belong to |
e1ab7874 | 6353 | corresponding loop, then it should be a preheader. */ |
6354 | if (sel_is_loop_preheader_p (bb)) | |
6355 | { | |
f1f41a6c | 6356 | preheader_blocks->safe_push (bb); |
e1ab7874 | 6357 | if (BB_END (bb) != bb_note (bb)) |
6358 | all_empty_p = false; | |
6359 | } | |
6360 | } | |
48e1416a | 6361 | |
e1ab7874 | 6362 | /* Remove these blocks only after iterating over the whole region. */ |
f1f41a6c | 6363 | for (i = preheader_blocks->length () - 1; i >= old_len; i--) |
e1ab7874 | 6364 | { |
f1f41a6c | 6365 | bb = (*preheader_blocks)[i]; |
e1ab7874 | 6366 | sel_remove_bb (bb, false); |
6367 | } | |
6368 | ||
6369 | if (!considered_for_pipelining_p (loop_outer (current_loop_nest))) | |
6370 | { | |
6371 | if (!all_empty_p) | |
6372 | /* Immediately create new region from preheader. */ | |
f1f41a6c | 6373 | make_region_from_loop_preheader (preheader_blocks); |
e1ab7874 | 6374 | else |
6375 | { | |
6376 | /* If all preheader blocks are empty - dont create new empty region. | |
6377 | Instead, remove them completely. */ | |
f1f41a6c | 6378 | FOR_EACH_VEC_ELT (*preheader_blocks, i, bb) |
e1ab7874 | 6379 | { |
6380 | edge e; | |
6381 | edge_iterator ei; | |
6382 | basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; | |
6383 | ||
6384 | /* Redirect all incoming edges to next basic block. */ | |
6385 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) | |
6386 | { | |
6387 | if (! (e->flags & EDGE_FALLTHRU)) | |
6388 | redirect_edge_and_branch (e, bb->next_bb); | |
6389 | else | |
6390 | redirect_edge_succ (e, bb->next_bb); | |
6391 | } | |
6392 | gcc_assert (BB_NOTE_LIST (bb) == NULL); | |
6393 | delete_and_free_basic_block (bb); | |
6394 | ||
48e1416a | 6395 | /* Check if after deleting preheader there is a nonconditional |
6396 | jump in PREV_BB that leads to the next basic block NEXT_BB. | |
6397 | If it is so - delete this jump and clear data sets of its | |
e1ab7874 | 6398 | basic block if it becomes empty. */ |
6399 | if (next_bb->prev_bb == prev_bb | |
34154e27 | 6400 | && prev_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
49087fba | 6401 | && bb_has_removable_jump_to_p (prev_bb, next_bb)) |
e1ab7874 | 6402 | { |
6403 | redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); | |
6404 | if (BB_END (prev_bb) == bb_note (prev_bb)) | |
6405 | free_data_sets (prev_bb); | |
6406 | } | |
1a5dbaab | 6407 | |
6408 | set_immediate_dominator (CDI_DOMINATORS, next_bb, | |
6409 | recompute_dominator (CDI_DOMINATORS, | |
6410 | next_bb)); | |
e1ab7874 | 6411 | } |
6412 | } | |
f1f41a6c | 6413 | vec_free (preheader_blocks); |
e1ab7874 | 6414 | } |
6415 | else | |
6416 | /* Store preheader within the father's loop structure. */ | |
6417 | SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), | |
6418 | preheader_blocks); | |
6419 | } | |
6420 | #endif |