]>
Commit | Line | Data |
---|---|---|
1 | /* RTL-level loop invariant motion. | |
2 | Copyright (C) 2004-2020 Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it | |
7 | under the terms of the GNU General Public License as published by the | |
8 | Free Software Foundation; either version 3, or (at your option) any | |
9 | later version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
12 | ANY 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 | /* This implements the loop invariant motion pass. It is very simple | |
21 | (no calls, no loads/stores, etc.). This should be sufficient to cleanup | |
22 | things like address arithmetics -- other more complicated invariants should | |
23 | be eliminated on GIMPLE either in tree-ssa-loop-im.c or in tree-ssa-pre.c. | |
24 | ||
25 | We proceed loop by loop -- it is simpler than trying to handle things | |
26 | globally and should not lose much. First we inspect all sets inside loop | |
27 | and create a dependency graph on insns (saying "to move this insn, you must | |
28 | also move the following insns"). | |
29 | ||
30 | We then need to determine what to move. We estimate the number of registers | |
31 | used and move as many invariants as possible while we still have enough free | |
32 | registers. We prefer the expensive invariants. | |
33 | ||
34 | Then we move the selected invariants out of the loop, creating a new | |
35 | temporaries for them if necessary. */ | |
36 | ||
37 | #include "config.h" | |
38 | #include "system.h" | |
39 | #include "coretypes.h" | |
40 | #include "backend.h" | |
41 | #include "target.h" | |
42 | #include "rtl.h" | |
43 | #include "tree.h" | |
44 | #include "cfghooks.h" | |
45 | #include "df.h" | |
46 | #include "memmodel.h" | |
47 | #include "tm_p.h" | |
48 | #include "insn-config.h" | |
49 | #include "regs.h" | |
50 | #include "ira.h" | |
51 | #include "recog.h" | |
52 | #include "cfgrtl.h" | |
53 | #include "cfgloop.h" | |
54 | #include "expr.h" | |
55 | #include "rtl-iter.h" | |
56 | #include "dumpfile.h" | |
57 | ||
58 | /* The data stored for the loop. */ | |
59 | ||
60 | class loop_data | |
61 | { | |
62 | public: | |
63 | class loop *outermost_exit; /* The outermost exit of the loop. */ | |
64 | bool has_call; /* True if the loop contains a call. */ | |
65 | /* Maximal register pressure inside loop for given register class | |
66 | (defined only for the pressure classes). */ | |
67 | int max_reg_pressure[N_REG_CLASSES]; | |
68 | /* Loop regs referenced and live pseudo-registers. */ | |
69 | bitmap_head regs_ref; | |
70 | bitmap_head regs_live; | |
71 | }; | |
72 | ||
73 | #define LOOP_DATA(LOOP) ((class loop_data *) (LOOP)->aux) | |
74 | ||
75 | /* The description of an use. */ | |
76 | ||
77 | struct use | |
78 | { | |
79 | rtx *pos; /* Position of the use. */ | |
80 | rtx_insn *insn; /* The insn in that the use occurs. */ | |
81 | unsigned addr_use_p; /* Whether the use occurs in an address. */ | |
82 | struct use *next; /* Next use in the list. */ | |
83 | }; | |
84 | ||
85 | /* The description of a def. */ | |
86 | ||
87 | struct def | |
88 | { | |
89 | struct use *uses; /* The list of uses that are uniquely reached | |
90 | by it. */ | |
91 | unsigned n_uses; /* Number of such uses. */ | |
92 | unsigned n_addr_uses; /* Number of uses in addresses. */ | |
93 | unsigned invno; /* The corresponding invariant. */ | |
94 | bool can_prop_to_addr_uses; /* True if the corresponding inv can be | |
95 | propagated into its address uses. */ | |
96 | }; | |
97 | ||
98 | /* The data stored for each invariant. */ | |
99 | ||
100 | struct invariant | |
101 | { | |
102 | /* The number of the invariant. */ | |
103 | unsigned invno; | |
104 | ||
105 | /* The number of the invariant with the same value. */ | |
106 | unsigned eqto; | |
107 | ||
108 | /* The number of invariants which eqto this. */ | |
109 | unsigned eqno; | |
110 | ||
111 | /* If we moved the invariant out of the loop, the original regno | |
112 | that contained its value. */ | |
113 | int orig_regno; | |
114 | ||
115 | /* If we moved the invariant out of the loop, the register that contains its | |
116 | value. */ | |
117 | rtx reg; | |
118 | ||
119 | /* The definition of the invariant. */ | |
120 | struct def *def; | |
121 | ||
122 | /* The insn in that it is defined. */ | |
123 | rtx_insn *insn; | |
124 | ||
125 | /* Whether it is always executed. */ | |
126 | bool always_executed; | |
127 | ||
128 | /* Whether to move the invariant. */ | |
129 | bool move; | |
130 | ||
131 | /* Whether the invariant is cheap when used as an address. */ | |
132 | bool cheap_address; | |
133 | ||
134 | /* Cost of the invariant. */ | |
135 | unsigned cost; | |
136 | ||
137 | /* Used for detecting already visited invariants during determining | |
138 | costs of movements. */ | |
139 | unsigned stamp; | |
140 | ||
141 | /* The invariants it depends on. */ | |
142 | bitmap depends_on; | |
143 | }; | |
144 | ||
145 | /* Currently processed loop. */ | |
146 | static class loop *curr_loop; | |
147 | ||
148 | /* Table of invariants indexed by the df_ref uid field. */ | |
149 | ||
150 | static unsigned int invariant_table_size = 0; | |
151 | static struct invariant ** invariant_table; | |
152 | ||
153 | /* Entry for hash table of invariant expressions. */ | |
154 | ||
155 | struct invariant_expr_entry | |
156 | { | |
157 | /* The invariant. */ | |
158 | struct invariant *inv; | |
159 | ||
160 | /* Its value. */ | |
161 | rtx expr; | |
162 | ||
163 | /* Its mode. */ | |
164 | machine_mode mode; | |
165 | ||
166 | /* Its hash. */ | |
167 | hashval_t hash; | |
168 | }; | |
169 | ||
170 | /* The actual stamp for marking already visited invariants during determining | |
171 | costs of movements. */ | |
172 | ||
173 | static unsigned actual_stamp; | |
174 | ||
175 | typedef struct invariant *invariant_p; | |
176 | ||
177 | ||
178 | /* The invariants. */ | |
179 | ||
180 | static vec<invariant_p> invariants; | |
181 | ||
182 | /* Check the size of the invariant table and realloc if necessary. */ | |
183 | ||
184 | static void | |
185 | check_invariant_table_size (void) | |
186 | { | |
187 | if (invariant_table_size < DF_DEFS_TABLE_SIZE ()) | |
188 | { | |
189 | unsigned int new_size = DF_DEFS_TABLE_SIZE () + (DF_DEFS_TABLE_SIZE () / 4); | |
190 | invariant_table = XRESIZEVEC (struct invariant *, invariant_table, new_size); | |
191 | memset (&invariant_table[invariant_table_size], 0, | |
192 | (new_size - invariant_table_size) * sizeof (struct invariant *)); | |
193 | invariant_table_size = new_size; | |
194 | } | |
195 | } | |
196 | ||
197 | /* Test for possibility of invariantness of X. */ | |
198 | ||
199 | static bool | |
200 | check_maybe_invariant (rtx x) | |
201 | { | |
202 | enum rtx_code code = GET_CODE (x); | |
203 | int i, j; | |
204 | const char *fmt; | |
205 | ||
206 | switch (code) | |
207 | { | |
208 | CASE_CONST_ANY: | |
209 | case SYMBOL_REF: | |
210 | case CONST: | |
211 | case LABEL_REF: | |
212 | return true; | |
213 | ||
214 | case PC: | |
215 | case CC0: | |
216 | case UNSPEC_VOLATILE: | |
217 | case CALL: | |
218 | return false; | |
219 | ||
220 | case REG: | |
221 | return true; | |
222 | ||
223 | case MEM: | |
224 | /* Load/store motion is done elsewhere. ??? Perhaps also add it here? | |
225 | It should not be hard, and might be faster than "elsewhere". */ | |
226 | ||
227 | /* Just handle the most trivial case where we load from an unchanging | |
228 | location (most importantly, pic tables). */ | |
229 | if (MEM_READONLY_P (x) && !MEM_VOLATILE_P (x)) | |
230 | break; | |
231 | ||
232 | return false; | |
233 | ||
234 | case ASM_OPERANDS: | |
235 | /* Don't mess with insns declared volatile. */ | |
236 | if (MEM_VOLATILE_P (x)) | |
237 | return false; | |
238 | break; | |
239 | ||
240 | default: | |
241 | break; | |
242 | } | |
243 | ||
244 | fmt = GET_RTX_FORMAT (code); | |
245 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
246 | { | |
247 | if (fmt[i] == 'e') | |
248 | { | |
249 | if (!check_maybe_invariant (XEXP (x, i))) | |
250 | return false; | |
251 | } | |
252 | else if (fmt[i] == 'E') | |
253 | { | |
254 | for (j = 0; j < XVECLEN (x, i); j++) | |
255 | if (!check_maybe_invariant (XVECEXP (x, i, j))) | |
256 | return false; | |
257 | } | |
258 | } | |
259 | ||
260 | return true; | |
261 | } | |
262 | ||
263 | /* Returns the invariant definition for USE, or NULL if USE is not | |
264 | invariant. */ | |
265 | ||
266 | static struct invariant * | |
267 | invariant_for_use (df_ref use) | |
268 | { | |
269 | struct df_link *defs; | |
270 | df_ref def; | |
271 | basic_block bb = DF_REF_BB (use), def_bb; | |
272 | ||
273 | if (DF_REF_FLAGS (use) & DF_REF_READ_WRITE) | |
274 | return NULL; | |
275 | ||
276 | defs = DF_REF_CHAIN (use); | |
277 | if (!defs || defs->next) | |
278 | return NULL; | |
279 | def = defs->ref; | |
280 | check_invariant_table_size (); | |
281 | if (!invariant_table[DF_REF_ID (def)]) | |
282 | return NULL; | |
283 | ||
284 | def_bb = DF_REF_BB (def); | |
285 | if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb)) | |
286 | return NULL; | |
287 | return invariant_table[DF_REF_ID (def)]; | |
288 | } | |
289 | ||
290 | /* Computes hash value for invariant expression X in INSN. */ | |
291 | ||
292 | static hashval_t | |
293 | hash_invariant_expr_1 (rtx_insn *insn, rtx x) | |
294 | { | |
295 | enum rtx_code code = GET_CODE (x); | |
296 | int i, j; | |
297 | const char *fmt; | |
298 | hashval_t val = code; | |
299 | int do_not_record_p; | |
300 | df_ref use; | |
301 | struct invariant *inv; | |
302 | ||
303 | switch (code) | |
304 | { | |
305 | CASE_CONST_ANY: | |
306 | case SYMBOL_REF: | |
307 | case CONST: | |
308 | case LABEL_REF: | |
309 | return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false); | |
310 | ||
311 | case REG: | |
312 | use = df_find_use (insn, x); | |
313 | if (!use) | |
314 | return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false); | |
315 | inv = invariant_for_use (use); | |
316 | if (!inv) | |
317 | return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false); | |
318 | ||
319 | gcc_assert (inv->eqto != ~0u); | |
320 | return inv->eqto; | |
321 | ||
322 | default: | |
323 | break; | |
324 | } | |
325 | ||
326 | fmt = GET_RTX_FORMAT (code); | |
327 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
328 | { | |
329 | if (fmt[i] == 'e') | |
330 | val ^= hash_invariant_expr_1 (insn, XEXP (x, i)); | |
331 | else if (fmt[i] == 'E') | |
332 | { | |
333 | for (j = 0; j < XVECLEN (x, i); j++) | |
334 | val ^= hash_invariant_expr_1 (insn, XVECEXP (x, i, j)); | |
335 | } | |
336 | else if (fmt[i] == 'i' || fmt[i] == 'n') | |
337 | val ^= XINT (x, i); | |
338 | else if (fmt[i] == 'p') | |
339 | val ^= constant_lower_bound (SUBREG_BYTE (x)); | |
340 | } | |
341 | ||
342 | return val; | |
343 | } | |
344 | ||
345 | /* Returns true if the invariant expressions E1 and E2 used in insns INSN1 | |
346 | and INSN2 have always the same value. */ | |
347 | ||
348 | static bool | |
349 | invariant_expr_equal_p (rtx_insn *insn1, rtx e1, rtx_insn *insn2, rtx e2) | |
350 | { | |
351 | enum rtx_code code = GET_CODE (e1); | |
352 | int i, j; | |
353 | const char *fmt; | |
354 | df_ref use1, use2; | |
355 | struct invariant *inv1 = NULL, *inv2 = NULL; | |
356 | rtx sub1, sub2; | |
357 | ||
358 | /* If mode of only one of the operands is VOIDmode, it is not equivalent to | |
359 | the other one. If both are VOIDmode, we rely on the caller of this | |
360 | function to verify that their modes are the same. */ | |
361 | if (code != GET_CODE (e2) || GET_MODE (e1) != GET_MODE (e2)) | |
362 | return false; | |
363 | ||
364 | switch (code) | |
365 | { | |
366 | CASE_CONST_ANY: | |
367 | case SYMBOL_REF: | |
368 | case CONST: | |
369 | case LABEL_REF: | |
370 | return rtx_equal_p (e1, e2); | |
371 | ||
372 | case REG: | |
373 | use1 = df_find_use (insn1, e1); | |
374 | use2 = df_find_use (insn2, e2); | |
375 | if (use1) | |
376 | inv1 = invariant_for_use (use1); | |
377 | if (use2) | |
378 | inv2 = invariant_for_use (use2); | |
379 | ||
380 | if (!inv1 && !inv2) | |
381 | return rtx_equal_p (e1, e2); | |
382 | ||
383 | if (!inv1 || !inv2) | |
384 | return false; | |
385 | ||
386 | gcc_assert (inv1->eqto != ~0u); | |
387 | gcc_assert (inv2->eqto != ~0u); | |
388 | return inv1->eqto == inv2->eqto; | |
389 | ||
390 | default: | |
391 | break; | |
392 | } | |
393 | ||
394 | fmt = GET_RTX_FORMAT (code); | |
395 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
396 | { | |
397 | if (fmt[i] == 'e') | |
398 | { | |
399 | sub1 = XEXP (e1, i); | |
400 | sub2 = XEXP (e2, i); | |
401 | ||
402 | if (!invariant_expr_equal_p (insn1, sub1, insn2, sub2)) | |
403 | return false; | |
404 | } | |
405 | ||
406 | else if (fmt[i] == 'E') | |
407 | { | |
408 | if (XVECLEN (e1, i) != XVECLEN (e2, i)) | |
409 | return false; | |
410 | ||
411 | for (j = 0; j < XVECLEN (e1, i); j++) | |
412 | { | |
413 | sub1 = XVECEXP (e1, i, j); | |
414 | sub2 = XVECEXP (e2, i, j); | |
415 | ||
416 | if (!invariant_expr_equal_p (insn1, sub1, insn2, sub2)) | |
417 | return false; | |
418 | } | |
419 | } | |
420 | else if (fmt[i] == 'i' || fmt[i] == 'n') | |
421 | { | |
422 | if (XINT (e1, i) != XINT (e2, i)) | |
423 | return false; | |
424 | } | |
425 | else if (fmt[i] == 'p') | |
426 | { | |
427 | if (maybe_ne (SUBREG_BYTE (e1), SUBREG_BYTE (e2))) | |
428 | return false; | |
429 | } | |
430 | /* Unhandled type of subexpression, we fail conservatively. */ | |
431 | else | |
432 | return false; | |
433 | } | |
434 | ||
435 | return true; | |
436 | } | |
437 | ||
438 | struct invariant_expr_hasher : free_ptr_hash <invariant_expr_entry> | |
439 | { | |
440 | static inline hashval_t hash (const invariant_expr_entry *); | |
441 | static inline bool equal (const invariant_expr_entry *, | |
442 | const invariant_expr_entry *); | |
443 | }; | |
444 | ||
445 | /* Returns hash value for invariant expression entry ENTRY. */ | |
446 | ||
447 | inline hashval_t | |
448 | invariant_expr_hasher::hash (const invariant_expr_entry *entry) | |
449 | { | |
450 | return entry->hash; | |
451 | } | |
452 | ||
453 | /* Compares invariant expression entries ENTRY1 and ENTRY2. */ | |
454 | ||
455 | inline bool | |
456 | invariant_expr_hasher::equal (const invariant_expr_entry *entry1, | |
457 | const invariant_expr_entry *entry2) | |
458 | { | |
459 | if (entry1->mode != entry2->mode) | |
460 | return 0; | |
461 | ||
462 | return invariant_expr_equal_p (entry1->inv->insn, entry1->expr, | |
463 | entry2->inv->insn, entry2->expr); | |
464 | } | |
465 | ||
466 | typedef hash_table<invariant_expr_hasher> invariant_htab_type; | |
467 | ||
468 | /* Checks whether invariant with value EXPR in machine mode MODE is | |
469 | recorded in EQ. If this is the case, return the invariant. Otherwise | |
470 | insert INV to the table for this expression and return INV. */ | |
471 | ||
472 | static struct invariant * | |
473 | find_or_insert_inv (invariant_htab_type *eq, rtx expr, machine_mode mode, | |
474 | struct invariant *inv) | |
475 | { | |
476 | hashval_t hash = hash_invariant_expr_1 (inv->insn, expr); | |
477 | struct invariant_expr_entry *entry; | |
478 | struct invariant_expr_entry pentry; | |
479 | invariant_expr_entry **slot; | |
480 | ||
481 | pentry.expr = expr; | |
482 | pentry.inv = inv; | |
483 | pentry.mode = mode; | |
484 | slot = eq->find_slot_with_hash (&pentry, hash, INSERT); | |
485 | entry = *slot; | |
486 | ||
487 | if (entry) | |
488 | return entry->inv; | |
489 | ||
490 | entry = XNEW (struct invariant_expr_entry); | |
491 | entry->inv = inv; | |
492 | entry->expr = expr; | |
493 | entry->mode = mode; | |
494 | entry->hash = hash; | |
495 | *slot = entry; | |
496 | ||
497 | return inv; | |
498 | } | |
499 | ||
500 | /* Finds invariants identical to INV and records the equivalence. EQ is the | |
501 | hash table of the invariants. */ | |
502 | ||
503 | static void | |
504 | find_identical_invariants (invariant_htab_type *eq, struct invariant *inv) | |
505 | { | |
506 | unsigned depno; | |
507 | bitmap_iterator bi; | |
508 | struct invariant *dep; | |
509 | rtx expr, set; | |
510 | machine_mode mode; | |
511 | struct invariant *tmp; | |
512 | ||
513 | if (inv->eqto != ~0u) | |
514 | return; | |
515 | ||
516 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi) | |
517 | { | |
518 | dep = invariants[depno]; | |
519 | find_identical_invariants (eq, dep); | |
520 | } | |
521 | ||
522 | set = single_set (inv->insn); | |
523 | expr = SET_SRC (set); | |
524 | mode = GET_MODE (expr); | |
525 | if (mode == VOIDmode) | |
526 | mode = GET_MODE (SET_DEST (set)); | |
527 | ||
528 | tmp = find_or_insert_inv (eq, expr, mode, inv); | |
529 | inv->eqto = tmp->invno; | |
530 | ||
531 | if (tmp->invno != inv->invno && inv->always_executed) | |
532 | tmp->eqno++; | |
533 | ||
534 | if (dump_file && inv->eqto != inv->invno) | |
535 | fprintf (dump_file, | |
536 | "Invariant %d is equivalent to invariant %d.\n", | |
537 | inv->invno, inv->eqto); | |
538 | } | |
539 | ||
540 | /* Find invariants with the same value and record the equivalences. */ | |
541 | ||
542 | static void | |
543 | merge_identical_invariants (void) | |
544 | { | |
545 | unsigned i; | |
546 | struct invariant *inv; | |
547 | invariant_htab_type eq (invariants.length ()); | |
548 | ||
549 | FOR_EACH_VEC_ELT (invariants, i, inv) | |
550 | find_identical_invariants (&eq, inv); | |
551 | } | |
552 | ||
553 | /* Determines the basic blocks inside LOOP that are always executed and | |
554 | stores their bitmap to ALWAYS_REACHED. MAY_EXIT is a bitmap of | |
555 | basic blocks that may either exit the loop, or contain the call that | |
556 | does not have to return. BODY is body of the loop obtained by | |
557 | get_loop_body_in_dom_order. */ | |
558 | ||
559 | static void | |
560 | compute_always_reached (class loop *loop, basic_block *body, | |
561 | bitmap may_exit, bitmap always_reached) | |
562 | { | |
563 | unsigned i; | |
564 | ||
565 | for (i = 0; i < loop->num_nodes; i++) | |
566 | { | |
567 | if (dominated_by_p (CDI_DOMINATORS, loop->latch, body[i])) | |
568 | bitmap_set_bit (always_reached, i); | |
569 | ||
570 | if (bitmap_bit_p (may_exit, i)) | |
571 | return; | |
572 | } | |
573 | } | |
574 | ||
575 | /* Finds exits out of the LOOP with body BODY. Marks blocks in that we may | |
576 | exit the loop by cfg edge to HAS_EXIT and MAY_EXIT. In MAY_EXIT | |
577 | additionally mark blocks that may exit due to a call. */ | |
578 | ||
579 | static void | |
580 | find_exits (class loop *loop, basic_block *body, | |
581 | bitmap may_exit, bitmap has_exit) | |
582 | { | |
583 | unsigned i; | |
584 | edge_iterator ei; | |
585 | edge e; | |
586 | class loop *outermost_exit = loop, *aexit; | |
587 | bool has_call = false; | |
588 | rtx_insn *insn; | |
589 | ||
590 | for (i = 0; i < loop->num_nodes; i++) | |
591 | { | |
592 | if (body[i]->loop_father == loop) | |
593 | { | |
594 | FOR_BB_INSNS (body[i], insn) | |
595 | { | |
596 | if (CALL_P (insn) | |
597 | && (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn) | |
598 | || !RTL_CONST_OR_PURE_CALL_P (insn))) | |
599 | { | |
600 | has_call = true; | |
601 | bitmap_set_bit (may_exit, i); | |
602 | break; | |
603 | } | |
604 | } | |
605 | ||
606 | FOR_EACH_EDGE (e, ei, body[i]->succs) | |
607 | { | |
608 | if (! flow_bb_inside_loop_p (loop, e->dest)) | |
609 | { | |
610 | bitmap_set_bit (may_exit, i); | |
611 | bitmap_set_bit (has_exit, i); | |
612 | outermost_exit = find_common_loop (outermost_exit, | |
613 | e->dest->loop_father); | |
614 | } | |
615 | /* If we enter a subloop that might never terminate treat | |
616 | it like a possible exit. */ | |
617 | if (flow_loop_nested_p (loop, e->dest->loop_father)) | |
618 | bitmap_set_bit (may_exit, i); | |
619 | } | |
620 | continue; | |
621 | } | |
622 | ||
623 | /* Use the data stored for the subloop to decide whether we may exit | |
624 | through it. It is sufficient to do this for header of the loop, | |
625 | as other basic blocks inside it must be dominated by it. */ | |
626 | if (body[i]->loop_father->header != body[i]) | |
627 | continue; | |
628 | ||
629 | if (LOOP_DATA (body[i]->loop_father)->has_call) | |
630 | { | |
631 | has_call = true; | |
632 | bitmap_set_bit (may_exit, i); | |
633 | } | |
634 | aexit = LOOP_DATA (body[i]->loop_father)->outermost_exit; | |
635 | if (aexit != loop) | |
636 | { | |
637 | bitmap_set_bit (may_exit, i); | |
638 | bitmap_set_bit (has_exit, i); | |
639 | ||
640 | if (flow_loop_nested_p (aexit, outermost_exit)) | |
641 | outermost_exit = aexit; | |
642 | } | |
643 | } | |
644 | ||
645 | if (loop->aux == NULL) | |
646 | { | |
647 | loop->aux = xcalloc (1, sizeof (class loop_data)); | |
648 | bitmap_initialize (&LOOP_DATA (loop)->regs_ref, ®_obstack); | |
649 | bitmap_initialize (&LOOP_DATA (loop)->regs_live, ®_obstack); | |
650 | } | |
651 | LOOP_DATA (loop)->outermost_exit = outermost_exit; | |
652 | LOOP_DATA (loop)->has_call = has_call; | |
653 | } | |
654 | ||
655 | /* Check whether we may assign a value to X from a register. */ | |
656 | ||
657 | static bool | |
658 | may_assign_reg_p (rtx x) | |
659 | { | |
660 | return (GET_MODE (x) != VOIDmode | |
661 | && GET_MODE (x) != BLKmode | |
662 | && can_copy_p (GET_MODE (x)) | |
663 | /* Do not mess with the frame pointer adjustments that can | |
664 | be generated e.g. by expand_builtin_setjmp_receiver. */ | |
665 | && x != frame_pointer_rtx | |
666 | && (!REG_P (x) | |
667 | || !HARD_REGISTER_P (x) | |
668 | || REGNO_REG_CLASS (REGNO (x)) != NO_REGS)); | |
669 | } | |
670 | ||
671 | /* Finds definitions that may correspond to invariants in LOOP with body | |
672 | BODY. */ | |
673 | ||
674 | static void | |
675 | find_defs (class loop *loop) | |
676 | { | |
677 | if (dump_file) | |
678 | { | |
679 | fprintf (dump_file, | |
680 | "*****starting processing of loop %d ******\n", | |
681 | loop->num); | |
682 | } | |
683 | ||
684 | df_chain_add_problem (DF_UD_CHAIN); | |
685 | df_set_flags (DF_RD_PRUNE_DEAD_DEFS); | |
686 | df_analyze_loop (loop); | |
687 | check_invariant_table_size (); | |
688 | ||
689 | if (dump_file) | |
690 | { | |
691 | df_dump_region (dump_file); | |
692 | fprintf (dump_file, | |
693 | "*****ending processing of loop %d ******\n", | |
694 | loop->num); | |
695 | } | |
696 | } | |
697 | ||
698 | /* Creates a new invariant for definition DEF in INSN, depending on invariants | |
699 | in DEPENDS_ON. ALWAYS_EXECUTED is true if the insn is always executed, | |
700 | unless the program ends due to a function call. The newly created invariant | |
701 | is returned. */ | |
702 | ||
703 | static struct invariant * | |
704 | create_new_invariant (struct def *def, rtx_insn *insn, bitmap depends_on, | |
705 | bool always_executed) | |
706 | { | |
707 | struct invariant *inv = XNEW (struct invariant); | |
708 | rtx set = single_set (insn); | |
709 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); | |
710 | ||
711 | inv->def = def; | |
712 | inv->always_executed = always_executed; | |
713 | inv->depends_on = depends_on; | |
714 | ||
715 | /* If the set is simple, usually by moving it we move the whole store out of | |
716 | the loop. Otherwise we save only cost of the computation. */ | |
717 | if (def) | |
718 | { | |
719 | inv->cost = set_rtx_cost (set, speed); | |
720 | /* ??? Try to determine cheapness of address computation. Unfortunately | |
721 | the address cost is only a relative measure, we can't really compare | |
722 | it with any absolute number, but only with other address costs. | |
723 | But here we don't have any other addresses, so compare with a magic | |
724 | number anyway. It has to be large enough to not regress PR33928 | |
725 | (by avoiding to move reg+8,reg+16,reg+24 invariants), but small | |
726 | enough to not regress 410.bwaves either (by still moving reg+reg | |
727 | invariants). | |
728 | See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ | |
729 | if (SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set)))) | |
730 | inv->cheap_address = address_cost (SET_SRC (set), word_mode, | |
731 | ADDR_SPACE_GENERIC, speed) < 3; | |
732 | else | |
733 | inv->cheap_address = false; | |
734 | } | |
735 | else | |
736 | { | |
737 | inv->cost = set_src_cost (SET_SRC (set), GET_MODE (SET_DEST (set)), | |
738 | speed); | |
739 | inv->cheap_address = false; | |
740 | } | |
741 | ||
742 | inv->move = false; | |
743 | inv->reg = NULL_RTX; | |
744 | inv->orig_regno = -1; | |
745 | inv->stamp = 0; | |
746 | inv->insn = insn; | |
747 | ||
748 | inv->invno = invariants.length (); | |
749 | inv->eqto = ~0u; | |
750 | ||
751 | /* Itself. */ | |
752 | inv->eqno = 1; | |
753 | ||
754 | if (def) | |
755 | def->invno = inv->invno; | |
756 | invariants.safe_push (inv); | |
757 | ||
758 | if (dump_file) | |
759 | { | |
760 | fprintf (dump_file, | |
761 | "Set in insn %d is invariant (%d), cost %d, depends on ", | |
762 | INSN_UID (insn), inv->invno, inv->cost); | |
763 | dump_bitmap (dump_file, inv->depends_on); | |
764 | } | |
765 | ||
766 | return inv; | |
767 | } | |
768 | ||
769 | /* Return a canonical version of X for the address, from the point of view, | |
770 | that all multiplications are represented as MULT instead of the multiply | |
771 | by a power of 2 being represented as ASHIFT. | |
772 | ||
773 | Callers should prepare a copy of X because this function may modify it | |
774 | in place. */ | |
775 | ||
776 | static void | |
777 | canonicalize_address_mult (rtx x) | |
778 | { | |
779 | subrtx_var_iterator::array_type array; | |
780 | FOR_EACH_SUBRTX_VAR (iter, array, x, NONCONST) | |
781 | { | |
782 | rtx sub = *iter; | |
783 | scalar_int_mode sub_mode; | |
784 | if (is_a <scalar_int_mode> (GET_MODE (sub), &sub_mode) | |
785 | && GET_CODE (sub) == ASHIFT | |
786 | && CONST_INT_P (XEXP (sub, 1)) | |
787 | && INTVAL (XEXP (sub, 1)) < GET_MODE_BITSIZE (sub_mode) | |
788 | && INTVAL (XEXP (sub, 1)) >= 0) | |
789 | { | |
790 | HOST_WIDE_INT shift = INTVAL (XEXP (sub, 1)); | |
791 | PUT_CODE (sub, MULT); | |
792 | XEXP (sub, 1) = gen_int_mode (HOST_WIDE_INT_1 << shift, sub_mode); | |
793 | iter.skip_subrtxes (); | |
794 | } | |
795 | } | |
796 | } | |
797 | ||
798 | /* Maximum number of sub expressions in address. We set it to | |
799 | a small integer since it's unlikely to have a complicated | |
800 | address expression. */ | |
801 | ||
802 | #define MAX_CANON_ADDR_PARTS (5) | |
803 | ||
804 | /* Collect sub expressions in address X with PLUS as the seperator. | |
805 | Sub expressions are stored in vector ADDR_PARTS. */ | |
806 | ||
807 | static void | |
808 | collect_address_parts (rtx x, vec<rtx> *addr_parts) | |
809 | { | |
810 | subrtx_var_iterator::array_type array; | |
811 | FOR_EACH_SUBRTX_VAR (iter, array, x, NONCONST) | |
812 | { | |
813 | rtx sub = *iter; | |
814 | ||
815 | if (GET_CODE (sub) != PLUS) | |
816 | { | |
817 | addr_parts->safe_push (sub); | |
818 | iter.skip_subrtxes (); | |
819 | } | |
820 | } | |
821 | } | |
822 | ||
823 | /* Compare function for sorting sub expressions X and Y based on | |
824 | precedence defined for communitive operations. */ | |
825 | ||
826 | static int | |
827 | compare_address_parts (const void *x, const void *y) | |
828 | { | |
829 | const rtx *rx = (const rtx *)x; | |
830 | const rtx *ry = (const rtx *)y; | |
831 | int px = commutative_operand_precedence (*rx); | |
832 | int py = commutative_operand_precedence (*ry); | |
833 | ||
834 | return (py - px); | |
835 | } | |
836 | ||
837 | /* Return a canonical version address for X by following steps: | |
838 | 1) Rewrite ASHIFT into MULT recursively. | |
839 | 2) Divide address into sub expressions with PLUS as the | |
840 | separator. | |
841 | 3) Sort sub expressions according to precedence defined | |
842 | for communative operations. | |
843 | 4) Simplify CONST_INT_P sub expressions. | |
844 | 5) Create new canonicalized address and return. | |
845 | Callers should prepare a copy of X because this function may | |
846 | modify it in place. */ | |
847 | ||
848 | static rtx | |
849 | canonicalize_address (rtx x) | |
850 | { | |
851 | rtx res; | |
852 | unsigned int i, j; | |
853 | machine_mode mode = GET_MODE (x); | |
854 | auto_vec<rtx, MAX_CANON_ADDR_PARTS> addr_parts; | |
855 | ||
856 | /* Rewrite ASHIFT into MULT. */ | |
857 | canonicalize_address_mult (x); | |
858 | /* Divide address into sub expressions. */ | |
859 | collect_address_parts (x, &addr_parts); | |
860 | /* Unlikely to have very complicated address. */ | |
861 | if (addr_parts.length () < 2 | |
862 | || addr_parts.length () > MAX_CANON_ADDR_PARTS) | |
863 | return x; | |
864 | ||
865 | /* Sort sub expressions according to canonicalization precedence. */ | |
866 | addr_parts.qsort (compare_address_parts); | |
867 | ||
868 | /* Simplify all constant int summary if possible. */ | |
869 | for (i = 0; i < addr_parts.length (); i++) | |
870 | if (CONST_INT_P (addr_parts[i])) | |
871 | break; | |
872 | ||
873 | for (j = i + 1; j < addr_parts.length (); j++) | |
874 | { | |
875 | gcc_assert (CONST_INT_P (addr_parts[j])); | |
876 | addr_parts[i] = simplify_gen_binary (PLUS, mode, | |
877 | addr_parts[i], | |
878 | addr_parts[j]); | |
879 | } | |
880 | ||
881 | /* Chain PLUS operators to the left for !CONST_INT_P sub expressions. */ | |
882 | res = addr_parts[0]; | |
883 | for (j = 1; j < i; j++) | |
884 | res = simplify_gen_binary (PLUS, mode, res, addr_parts[j]); | |
885 | ||
886 | /* Pickup the last CONST_INT_P sub expression. */ | |
887 | if (i < addr_parts.length ()) | |
888 | res = simplify_gen_binary (PLUS, mode, res, addr_parts[i]); | |
889 | ||
890 | return res; | |
891 | } | |
892 | ||
893 | /* Given invariant DEF and its address USE, check if the corresponding | |
894 | invariant expr can be propagated into the use or not. */ | |
895 | ||
896 | static bool | |
897 | inv_can_prop_to_addr_use (struct def *def, df_ref use) | |
898 | { | |
899 | struct invariant *inv; | |
900 | rtx *pos = DF_REF_REAL_LOC (use), def_set, use_set; | |
901 | rtx_insn *use_insn = DF_REF_INSN (use); | |
902 | rtx_insn *def_insn; | |
903 | bool ok; | |
904 | ||
905 | inv = invariants[def->invno]; | |
906 | /* No need to check if address expression is expensive. */ | |
907 | if (!inv->cheap_address) | |
908 | return false; | |
909 | ||
910 | def_insn = inv->insn; | |
911 | def_set = single_set (def_insn); | |
912 | if (!def_set) | |
913 | return false; | |
914 | ||
915 | validate_unshare_change (use_insn, pos, SET_SRC (def_set), true); | |
916 | ok = verify_changes (0); | |
917 | /* Try harder with canonicalization in address expression. */ | |
918 | if (!ok && (use_set = single_set (use_insn)) != NULL_RTX) | |
919 | { | |
920 | rtx src, dest, mem = NULL_RTX; | |
921 | ||
922 | src = SET_SRC (use_set); | |
923 | dest = SET_DEST (use_set); | |
924 | if (MEM_P (src)) | |
925 | mem = src; | |
926 | else if (MEM_P (dest)) | |
927 | mem = dest; | |
928 | ||
929 | if (mem != NULL_RTX | |
930 | && !memory_address_addr_space_p (GET_MODE (mem), | |
931 | XEXP (mem, 0), | |
932 | MEM_ADDR_SPACE (mem))) | |
933 | { | |
934 | rtx addr = canonicalize_address (copy_rtx (XEXP (mem, 0))); | |
935 | if (memory_address_addr_space_p (GET_MODE (mem), | |
936 | addr, MEM_ADDR_SPACE (mem))) | |
937 | ok = true; | |
938 | } | |
939 | } | |
940 | cancel_changes (0); | |
941 | return ok; | |
942 | } | |
943 | ||
944 | /* Record USE at DEF. */ | |
945 | ||
946 | static void | |
947 | record_use (struct def *def, df_ref use) | |
948 | { | |
949 | struct use *u = XNEW (struct use); | |
950 | ||
951 | u->pos = DF_REF_REAL_LOC (use); | |
952 | u->insn = DF_REF_INSN (use); | |
953 | u->addr_use_p = (DF_REF_TYPE (use) == DF_REF_REG_MEM_LOAD | |
954 | || DF_REF_TYPE (use) == DF_REF_REG_MEM_STORE); | |
955 | u->next = def->uses; | |
956 | def->uses = u; | |
957 | def->n_uses++; | |
958 | if (u->addr_use_p) | |
959 | { | |
960 | /* Initialize propagation information if this is the first addr | |
961 | use of the inv def. */ | |
962 | if (def->n_addr_uses == 0) | |
963 | def->can_prop_to_addr_uses = true; | |
964 | ||
965 | def->n_addr_uses++; | |
966 | if (def->can_prop_to_addr_uses && !inv_can_prop_to_addr_use (def, use)) | |
967 | def->can_prop_to_addr_uses = false; | |
968 | } | |
969 | } | |
970 | ||
971 | /* Finds the invariants USE depends on and store them to the DEPENDS_ON | |
972 | bitmap. Returns true if all dependencies of USE are known to be | |
973 | loop invariants, false otherwise. */ | |
974 | ||
975 | static bool | |
976 | check_dependency (basic_block bb, df_ref use, bitmap depends_on) | |
977 | { | |
978 | df_ref def; | |
979 | basic_block def_bb; | |
980 | struct df_link *defs; | |
981 | struct def *def_data; | |
982 | struct invariant *inv; | |
983 | ||
984 | if (DF_REF_FLAGS (use) & DF_REF_READ_WRITE) | |
985 | return false; | |
986 | ||
987 | defs = DF_REF_CHAIN (use); | |
988 | if (!defs) | |
989 | { | |
990 | unsigned int regno = DF_REF_REGNO (use); | |
991 | ||
992 | /* If this is the use of an uninitialized argument register that is | |
993 | likely to be spilled, do not move it lest this might extend its | |
994 | lifetime and cause reload to die. This can occur for a call to | |
995 | a function taking complex number arguments and moving the insns | |
996 | preparing the arguments without moving the call itself wouldn't | |
997 | gain much in practice. */ | |
998 | if ((DF_REF_FLAGS (use) & DF_HARD_REG_LIVE) | |
999 | && FUNCTION_ARG_REGNO_P (regno) | |
1000 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno))) | |
1001 | return false; | |
1002 | ||
1003 | return true; | |
1004 | } | |
1005 | ||
1006 | if (defs->next) | |
1007 | return false; | |
1008 | ||
1009 | def = defs->ref; | |
1010 | check_invariant_table_size (); | |
1011 | inv = invariant_table[DF_REF_ID (def)]; | |
1012 | if (!inv) | |
1013 | return false; | |
1014 | ||
1015 | def_data = inv->def; | |
1016 | gcc_assert (def_data != NULL); | |
1017 | ||
1018 | def_bb = DF_REF_BB (def); | |
1019 | /* Note that in case bb == def_bb, we know that the definition | |
1020 | dominates insn, because def has invariant_table[DF_REF_ID(def)] | |
1021 | defined and we process the insns in the basic block bb | |
1022 | sequentially. */ | |
1023 | if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb)) | |
1024 | return false; | |
1025 | ||
1026 | bitmap_set_bit (depends_on, def_data->invno); | |
1027 | return true; | |
1028 | } | |
1029 | ||
1030 | ||
1031 | /* Finds the invariants INSN depends on and store them to the DEPENDS_ON | |
1032 | bitmap. Returns true if all dependencies of INSN are known to be | |
1033 | loop invariants, false otherwise. */ | |
1034 | ||
1035 | static bool | |
1036 | check_dependencies (rtx_insn *insn, bitmap depends_on) | |
1037 | { | |
1038 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); | |
1039 | df_ref use; | |
1040 | basic_block bb = BLOCK_FOR_INSN (insn); | |
1041 | ||
1042 | FOR_EACH_INSN_INFO_USE (use, insn_info) | |
1043 | if (!check_dependency (bb, use, depends_on)) | |
1044 | return false; | |
1045 | FOR_EACH_INSN_INFO_EQ_USE (use, insn_info) | |
1046 | if (!check_dependency (bb, use, depends_on)) | |
1047 | return false; | |
1048 | ||
1049 | return true; | |
1050 | } | |
1051 | ||
1052 | /* Pre-check candidate DEST to skip the one which cannot make a valid insn | |
1053 | during move_invariant_reg. SIMPLE is to skip HARD_REGISTER. */ | |
1054 | static bool | |
1055 | pre_check_invariant_p (bool simple, rtx dest) | |
1056 | { | |
1057 | if (simple && REG_P (dest) && DF_REG_DEF_COUNT (REGNO (dest)) > 1) | |
1058 | { | |
1059 | df_ref use; | |
1060 | unsigned int i = REGNO (dest); | |
1061 | struct df_insn_info *insn_info; | |
1062 | df_ref def_rec; | |
1063 | ||
1064 | for (use = DF_REG_USE_CHAIN (i); use; use = DF_REF_NEXT_REG (use)) | |
1065 | { | |
1066 | rtx_insn *ref = DF_REF_INSN (use); | |
1067 | insn_info = DF_INSN_INFO_GET (ref); | |
1068 | ||
1069 | FOR_EACH_INSN_INFO_DEF (def_rec, insn_info) | |
1070 | if (DF_REF_REGNO (def_rec) == i) | |
1071 | { | |
1072 | /* Multi definitions at this stage, most likely are due to | |
1073 | instruction constraints, which requires both read and write | |
1074 | on the same register. Since move_invariant_reg is not | |
1075 | powerful enough to handle such cases, just ignore the INV | |
1076 | and leave the chance to others. */ | |
1077 | return false; | |
1078 | } | |
1079 | } | |
1080 | } | |
1081 | return true; | |
1082 | } | |
1083 | ||
1084 | /* Finds invariant in INSN. ALWAYS_REACHED is true if the insn is always | |
1085 | executed. ALWAYS_EXECUTED is true if the insn is always executed, | |
1086 | unless the program ends due to a function call. */ | |
1087 | ||
1088 | static void | |
1089 | find_invariant_insn (rtx_insn *insn, bool always_reached, bool always_executed) | |
1090 | { | |
1091 | df_ref ref; | |
1092 | struct def *def; | |
1093 | bitmap depends_on; | |
1094 | rtx set, dest; | |
1095 | bool simple = true; | |
1096 | struct invariant *inv; | |
1097 | ||
1098 | /* We can't move a CC0 setter without the user. */ | |
1099 | if (HAVE_cc0 && sets_cc0_p (insn)) | |
1100 | return; | |
1101 | ||
1102 | set = single_set (insn); | |
1103 | if (!set) | |
1104 | return; | |
1105 | dest = SET_DEST (set); | |
1106 | ||
1107 | if (!REG_P (dest) | |
1108 | || HARD_REGISTER_P (dest)) | |
1109 | simple = false; | |
1110 | ||
1111 | if (!may_assign_reg_p (dest) | |
1112 | || !pre_check_invariant_p (simple, dest) | |
1113 | || !check_maybe_invariant (SET_SRC (set))) | |
1114 | return; | |
1115 | ||
1116 | /* If the insn can throw exception, we cannot move it at all without changing | |
1117 | cfg. */ | |
1118 | if (can_throw_internal (insn)) | |
1119 | return; | |
1120 | ||
1121 | /* We cannot make trapping insn executed, unless it was executed before. */ | |
1122 | if (may_trap_or_fault_p (PATTERN (insn)) && !always_reached) | |
1123 | return; | |
1124 | ||
1125 | depends_on = BITMAP_ALLOC (NULL); | |
1126 | if (!check_dependencies (insn, depends_on)) | |
1127 | { | |
1128 | BITMAP_FREE (depends_on); | |
1129 | return; | |
1130 | } | |
1131 | ||
1132 | if (simple) | |
1133 | def = XCNEW (struct def); | |
1134 | else | |
1135 | def = NULL; | |
1136 | ||
1137 | inv = create_new_invariant (def, insn, depends_on, always_executed); | |
1138 | ||
1139 | if (simple) | |
1140 | { | |
1141 | ref = df_find_def (insn, dest); | |
1142 | check_invariant_table_size (); | |
1143 | invariant_table[DF_REF_ID (ref)] = inv; | |
1144 | } | |
1145 | } | |
1146 | ||
1147 | /* Record registers used in INSN that have a unique invariant definition. */ | |
1148 | ||
1149 | static void | |
1150 | record_uses (rtx_insn *insn) | |
1151 | { | |
1152 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); | |
1153 | df_ref use; | |
1154 | struct invariant *inv; | |
1155 | ||
1156 | FOR_EACH_INSN_INFO_USE (use, insn_info) | |
1157 | { | |
1158 | inv = invariant_for_use (use); | |
1159 | if (inv) | |
1160 | record_use (inv->def, use); | |
1161 | } | |
1162 | FOR_EACH_INSN_INFO_EQ_USE (use, insn_info) | |
1163 | { | |
1164 | inv = invariant_for_use (use); | |
1165 | if (inv) | |
1166 | record_use (inv->def, use); | |
1167 | } | |
1168 | } | |
1169 | ||
1170 | /* Finds invariants in INSN. ALWAYS_REACHED is true if the insn is always | |
1171 | executed. ALWAYS_EXECUTED is true if the insn is always executed, | |
1172 | unless the program ends due to a function call. */ | |
1173 | ||
1174 | static void | |
1175 | find_invariants_insn (rtx_insn *insn, bool always_reached, bool always_executed) | |
1176 | { | |
1177 | find_invariant_insn (insn, always_reached, always_executed); | |
1178 | record_uses (insn); | |
1179 | } | |
1180 | ||
1181 | /* Finds invariants in basic block BB. ALWAYS_REACHED is true if the | |
1182 | basic block is always executed. ALWAYS_EXECUTED is true if the basic | |
1183 | block is always executed, unless the program ends due to a function | |
1184 | call. */ | |
1185 | ||
1186 | static void | |
1187 | find_invariants_bb (basic_block bb, bool always_reached, bool always_executed) | |
1188 | { | |
1189 | rtx_insn *insn; | |
1190 | ||
1191 | FOR_BB_INSNS (bb, insn) | |
1192 | { | |
1193 | if (!NONDEBUG_INSN_P (insn)) | |
1194 | continue; | |
1195 | ||
1196 | find_invariants_insn (insn, always_reached, always_executed); | |
1197 | ||
1198 | if (always_reached | |
1199 | && CALL_P (insn) | |
1200 | && (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn) | |
1201 | || ! RTL_CONST_OR_PURE_CALL_P (insn))) | |
1202 | always_reached = false; | |
1203 | } | |
1204 | } | |
1205 | ||
1206 | /* Finds invariants in LOOP with body BODY. ALWAYS_REACHED is the bitmap of | |
1207 | basic blocks in BODY that are always executed. ALWAYS_EXECUTED is the | |
1208 | bitmap of basic blocks in BODY that are always executed unless the program | |
1209 | ends due to a function call. */ | |
1210 | ||
1211 | static void | |
1212 | find_invariants_body (class loop *loop, basic_block *body, | |
1213 | bitmap always_reached, bitmap always_executed) | |
1214 | { | |
1215 | unsigned i; | |
1216 | ||
1217 | for (i = 0; i < loop->num_nodes; i++) | |
1218 | find_invariants_bb (body[i], | |
1219 | bitmap_bit_p (always_reached, i), | |
1220 | bitmap_bit_p (always_executed, i)); | |
1221 | } | |
1222 | ||
1223 | /* Finds invariants in LOOP. */ | |
1224 | ||
1225 | static void | |
1226 | find_invariants (class loop *loop) | |
1227 | { | |
1228 | auto_bitmap may_exit; | |
1229 | auto_bitmap always_reached; | |
1230 | auto_bitmap has_exit; | |
1231 | auto_bitmap always_executed; | |
1232 | basic_block *body = get_loop_body_in_dom_order (loop); | |
1233 | ||
1234 | find_exits (loop, body, may_exit, has_exit); | |
1235 | compute_always_reached (loop, body, may_exit, always_reached); | |
1236 | compute_always_reached (loop, body, has_exit, always_executed); | |
1237 | ||
1238 | find_defs (loop); | |
1239 | find_invariants_body (loop, body, always_reached, always_executed); | |
1240 | merge_identical_invariants (); | |
1241 | ||
1242 | free (body); | |
1243 | } | |
1244 | ||
1245 | /* Frees a list of uses USE. */ | |
1246 | ||
1247 | static void | |
1248 | free_use_list (struct use *use) | |
1249 | { | |
1250 | struct use *next; | |
1251 | ||
1252 | for (; use; use = next) | |
1253 | { | |
1254 | next = use->next; | |
1255 | free (use); | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | /* Return pressure class and number of hard registers (through *NREGS) | |
1260 | for destination of INSN. */ | |
1261 | static enum reg_class | |
1262 | get_pressure_class_and_nregs (rtx_insn *insn, int *nregs) | |
1263 | { | |
1264 | rtx reg; | |
1265 | enum reg_class pressure_class; | |
1266 | rtx set = single_set (insn); | |
1267 | ||
1268 | /* Considered invariant insns have only one set. */ | |
1269 | gcc_assert (set != NULL_RTX); | |
1270 | reg = SET_DEST (set); | |
1271 | if (GET_CODE (reg) == SUBREG) | |
1272 | reg = SUBREG_REG (reg); | |
1273 | if (MEM_P (reg)) | |
1274 | { | |
1275 | *nregs = 0; | |
1276 | pressure_class = NO_REGS; | |
1277 | } | |
1278 | else | |
1279 | { | |
1280 | if (! REG_P (reg)) | |
1281 | reg = NULL_RTX; | |
1282 | if (reg == NULL_RTX) | |
1283 | pressure_class = GENERAL_REGS; | |
1284 | else | |
1285 | { | |
1286 | pressure_class = reg_allocno_class (REGNO (reg)); | |
1287 | pressure_class = ira_pressure_class_translate[pressure_class]; | |
1288 | } | |
1289 | *nregs | |
1290 | = ira_reg_class_max_nregs[pressure_class][GET_MODE (SET_SRC (set))]; | |
1291 | } | |
1292 | return pressure_class; | |
1293 | } | |
1294 | ||
1295 | /* Calculates cost and number of registers needed for moving invariant INV | |
1296 | out of the loop and stores them to *COST and *REGS_NEEDED. *CL will be | |
1297 | the REG_CLASS of INV. Return | |
1298 | -1: if INV is invalid. | |
1299 | 0: if INV and its depends_on have same reg_class | |
1300 | 1: if INV and its depends_on have different reg_classes. */ | |
1301 | ||
1302 | static int | |
1303 | get_inv_cost (struct invariant *inv, int *comp_cost, unsigned *regs_needed, | |
1304 | enum reg_class *cl) | |
1305 | { | |
1306 | int i, acomp_cost; | |
1307 | unsigned aregs_needed[N_REG_CLASSES]; | |
1308 | unsigned depno; | |
1309 | struct invariant *dep; | |
1310 | bitmap_iterator bi; | |
1311 | int ret = 1; | |
1312 | ||
1313 | /* Find the representative of the class of the equivalent invariants. */ | |
1314 | inv = invariants[inv->eqto]; | |
1315 | ||
1316 | *comp_cost = 0; | |
1317 | if (! flag_ira_loop_pressure) | |
1318 | regs_needed[0] = 0; | |
1319 | else | |
1320 | { | |
1321 | for (i = 0; i < ira_pressure_classes_num; i++) | |
1322 | regs_needed[ira_pressure_classes[i]] = 0; | |
1323 | } | |
1324 | ||
1325 | if (inv->move | |
1326 | || inv->stamp == actual_stamp) | |
1327 | return -1; | |
1328 | inv->stamp = actual_stamp; | |
1329 | ||
1330 | if (! flag_ira_loop_pressure) | |
1331 | regs_needed[0]++; | |
1332 | else | |
1333 | { | |
1334 | int nregs; | |
1335 | enum reg_class pressure_class; | |
1336 | ||
1337 | pressure_class = get_pressure_class_and_nregs (inv->insn, &nregs); | |
1338 | regs_needed[pressure_class] += nregs; | |
1339 | *cl = pressure_class; | |
1340 | ret = 0; | |
1341 | } | |
1342 | ||
1343 | if (!inv->cheap_address | |
1344 | || inv->def->n_uses == 0 | |
1345 | || inv->def->n_addr_uses < inv->def->n_uses | |
1346 | /* Count cost if the inv can't be propagated into address uses. */ | |
1347 | || !inv->def->can_prop_to_addr_uses) | |
1348 | (*comp_cost) += inv->cost * inv->eqno; | |
1349 | ||
1350 | #ifdef STACK_REGS | |
1351 | { | |
1352 | /* Hoisting constant pool constants into stack regs may cost more than | |
1353 | just single register. On x87, the balance is affected both by the | |
1354 | small number of FP registers, and by its register stack organization, | |
1355 | that forces us to add compensation code in and around the loop to | |
1356 | shuffle the operands to the top of stack before use, and pop them | |
1357 | from the stack after the loop finishes. | |
1358 | ||
1359 | To model this effect, we increase the number of registers needed for | |
1360 | stack registers by two: one register push, and one register pop. | |
1361 | This usually has the effect that FP constant loads from the constant | |
1362 | pool are not moved out of the loop. | |
1363 | ||
1364 | Note that this also means that dependent invariants cannot be moved. | |
1365 | However, the primary purpose of this pass is to move loop invariant | |
1366 | address arithmetic out of loops, and address arithmetic that depends | |
1367 | on floating point constants is unlikely to ever occur. */ | |
1368 | rtx set = single_set (inv->insn); | |
1369 | if (set | |
1370 | && IS_STACK_MODE (GET_MODE (SET_SRC (set))) | |
1371 | && constant_pool_constant_p (SET_SRC (set))) | |
1372 | { | |
1373 | if (flag_ira_loop_pressure) | |
1374 | regs_needed[ira_stack_reg_pressure_class] += 2; | |
1375 | else | |
1376 | regs_needed[0] += 2; | |
1377 | } | |
1378 | } | |
1379 | #endif | |
1380 | ||
1381 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi) | |
1382 | { | |
1383 | bool check_p; | |
1384 | enum reg_class dep_cl = ALL_REGS; | |
1385 | int dep_ret; | |
1386 | ||
1387 | dep = invariants[depno]; | |
1388 | ||
1389 | /* If DEP is moved out of the loop, it is not a depends_on any more. */ | |
1390 | if (dep->move) | |
1391 | continue; | |
1392 | ||
1393 | dep_ret = get_inv_cost (dep, &acomp_cost, aregs_needed, &dep_cl); | |
1394 | ||
1395 | if (! flag_ira_loop_pressure) | |
1396 | check_p = aregs_needed[0] != 0; | |
1397 | else | |
1398 | { | |
1399 | for (i = 0; i < ira_pressure_classes_num; i++) | |
1400 | if (aregs_needed[ira_pressure_classes[i]] != 0) | |
1401 | break; | |
1402 | check_p = i < ira_pressure_classes_num; | |
1403 | ||
1404 | if ((dep_ret == 1) || ((dep_ret == 0) && (*cl != dep_cl))) | |
1405 | { | |
1406 | *cl = ALL_REGS; | |
1407 | ret = 1; | |
1408 | } | |
1409 | } | |
1410 | if (check_p | |
1411 | /* We need to check always_executed, since if the original value of | |
1412 | the invariant may be preserved, we may need to keep it in a | |
1413 | separate register. TODO check whether the register has an | |
1414 | use outside of the loop. */ | |
1415 | && dep->always_executed | |
1416 | && !dep->def->uses->next) | |
1417 | { | |
1418 | /* If this is a single use, after moving the dependency we will not | |
1419 | need a new register. */ | |
1420 | if (! flag_ira_loop_pressure) | |
1421 | aregs_needed[0]--; | |
1422 | else | |
1423 | { | |
1424 | int nregs; | |
1425 | enum reg_class pressure_class; | |
1426 | ||
1427 | pressure_class = get_pressure_class_and_nregs (inv->insn, &nregs); | |
1428 | aregs_needed[pressure_class] -= nregs; | |
1429 | } | |
1430 | } | |
1431 | ||
1432 | if (! flag_ira_loop_pressure) | |
1433 | regs_needed[0] += aregs_needed[0]; | |
1434 | else | |
1435 | { | |
1436 | for (i = 0; i < ira_pressure_classes_num; i++) | |
1437 | regs_needed[ira_pressure_classes[i]] | |
1438 | += aregs_needed[ira_pressure_classes[i]]; | |
1439 | } | |
1440 | (*comp_cost) += acomp_cost; | |
1441 | } | |
1442 | return ret; | |
1443 | } | |
1444 | ||
1445 | /* Calculates gain for eliminating invariant INV. REGS_USED is the number | |
1446 | of registers used in the loop, NEW_REGS is the number of new variables | |
1447 | already added due to the invariant motion. The number of registers needed | |
1448 | for it is stored in *REGS_NEEDED. SPEED and CALL_P are flags passed | |
1449 | through to estimate_reg_pressure_cost. */ | |
1450 | ||
1451 | static int | |
1452 | gain_for_invariant (struct invariant *inv, unsigned *regs_needed, | |
1453 | unsigned *new_regs, unsigned regs_used, | |
1454 | bool speed, bool call_p) | |
1455 | { | |
1456 | int comp_cost, size_cost; | |
1457 | /* Workaround -Wmaybe-uninitialized false positive during | |
1458 | profiledbootstrap by initializing it. */ | |
1459 | enum reg_class cl = NO_REGS; | |
1460 | int ret; | |
1461 | ||
1462 | actual_stamp++; | |
1463 | ||
1464 | ret = get_inv_cost (inv, &comp_cost, regs_needed, &cl); | |
1465 | ||
1466 | if (! flag_ira_loop_pressure) | |
1467 | { | |
1468 | size_cost = (estimate_reg_pressure_cost (new_regs[0] + regs_needed[0], | |
1469 | regs_used, speed, call_p) | |
1470 | - estimate_reg_pressure_cost (new_regs[0], | |
1471 | regs_used, speed, call_p)); | |
1472 | } | |
1473 | else if (ret < 0) | |
1474 | return -1; | |
1475 | else if ((ret == 0) && (cl == NO_REGS)) | |
1476 | /* Hoist it anyway since it does not impact register pressure. */ | |
1477 | return 1; | |
1478 | else | |
1479 | { | |
1480 | int i; | |
1481 | enum reg_class pressure_class; | |
1482 | ||
1483 | for (i = 0; i < ira_pressure_classes_num; i++) | |
1484 | { | |
1485 | pressure_class = ira_pressure_classes[i]; | |
1486 | ||
1487 | if (!reg_classes_intersect_p (pressure_class, cl)) | |
1488 | continue; | |
1489 | ||
1490 | if ((int) new_regs[pressure_class] | |
1491 | + (int) regs_needed[pressure_class] | |
1492 | + LOOP_DATA (curr_loop)->max_reg_pressure[pressure_class] | |
1493 | + param_ira_loop_reserved_regs | |
1494 | > ira_class_hard_regs_num[pressure_class]) | |
1495 | break; | |
1496 | } | |
1497 | if (i < ira_pressure_classes_num) | |
1498 | /* There will be register pressure excess and we want not to | |
1499 | make this loop invariant motion. All loop invariants with | |
1500 | non-positive gains will be rejected in function | |
1501 | find_invariants_to_move. Therefore we return the negative | |
1502 | number here. | |
1503 | ||
1504 | One could think that this rejects also expensive loop | |
1505 | invariant motions and this will hurt code performance. | |
1506 | However numerous experiments with different heuristics | |
1507 | taking invariant cost into account did not confirm this | |
1508 | assumption. There are possible explanations for this | |
1509 | result: | |
1510 | o probably all expensive invariants were already moved out | |
1511 | of the loop by PRE and gimple invariant motion pass. | |
1512 | o expensive invariant execution will be hidden by insn | |
1513 | scheduling or OOO processor hardware because usually such | |
1514 | invariants have a lot of freedom to be executed | |
1515 | out-of-order. | |
1516 | Another reason for ignoring invariant cost vs spilling cost | |
1517 | heuristics is also in difficulties to evaluate accurately | |
1518 | spill cost at this stage. */ | |
1519 | return -1; | |
1520 | else | |
1521 | size_cost = 0; | |
1522 | } | |
1523 | ||
1524 | return comp_cost - size_cost; | |
1525 | } | |
1526 | ||
1527 | /* Finds invariant with best gain for moving. Returns the gain, stores | |
1528 | the invariant in *BEST and number of registers needed for it to | |
1529 | *REGS_NEEDED. REGS_USED is the number of registers used in the loop. | |
1530 | NEW_REGS is the number of new variables already added due to invariant | |
1531 | motion. */ | |
1532 | ||
1533 | static int | |
1534 | best_gain_for_invariant (struct invariant **best, unsigned *regs_needed, | |
1535 | unsigned *new_regs, unsigned regs_used, | |
1536 | bool speed, bool call_p) | |
1537 | { | |
1538 | struct invariant *inv; | |
1539 | int i, gain = 0, again; | |
1540 | unsigned aregs_needed[N_REG_CLASSES], invno; | |
1541 | ||
1542 | FOR_EACH_VEC_ELT (invariants, invno, inv) | |
1543 | { | |
1544 | if (inv->move) | |
1545 | continue; | |
1546 | ||
1547 | /* Only consider the "representatives" of equivalent invariants. */ | |
1548 | if (inv->eqto != inv->invno) | |
1549 | continue; | |
1550 | ||
1551 | again = gain_for_invariant (inv, aregs_needed, new_regs, regs_used, | |
1552 | speed, call_p); | |
1553 | if (again > gain) | |
1554 | { | |
1555 | gain = again; | |
1556 | *best = inv; | |
1557 | if (! flag_ira_loop_pressure) | |
1558 | regs_needed[0] = aregs_needed[0]; | |
1559 | else | |
1560 | { | |
1561 | for (i = 0; i < ira_pressure_classes_num; i++) | |
1562 | regs_needed[ira_pressure_classes[i]] | |
1563 | = aregs_needed[ira_pressure_classes[i]]; | |
1564 | } | |
1565 | } | |
1566 | } | |
1567 | ||
1568 | return gain; | |
1569 | } | |
1570 | ||
1571 | /* Marks invariant INVNO and all its dependencies for moving. */ | |
1572 | ||
1573 | static void | |
1574 | set_move_mark (unsigned invno, int gain) | |
1575 | { | |
1576 | struct invariant *inv = invariants[invno]; | |
1577 | bitmap_iterator bi; | |
1578 | ||
1579 | /* Find the representative of the class of the equivalent invariants. */ | |
1580 | inv = invariants[inv->eqto]; | |
1581 | ||
1582 | if (inv->move) | |
1583 | return; | |
1584 | inv->move = true; | |
1585 | ||
1586 | if (dump_file) | |
1587 | { | |
1588 | if (gain >= 0) | |
1589 | fprintf (dump_file, "Decided to move invariant %d -- gain %d\n", | |
1590 | invno, gain); | |
1591 | else | |
1592 | fprintf (dump_file, "Decided to move dependent invariant %d\n", | |
1593 | invno); | |
1594 | }; | |
1595 | ||
1596 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, invno, bi) | |
1597 | { | |
1598 | set_move_mark (invno, -1); | |
1599 | } | |
1600 | } | |
1601 | ||
1602 | /* Determines which invariants to move. */ | |
1603 | ||
1604 | static void | |
1605 | find_invariants_to_move (bool speed, bool call_p) | |
1606 | { | |
1607 | int gain; | |
1608 | unsigned i, regs_used, regs_needed[N_REG_CLASSES], new_regs[N_REG_CLASSES]; | |
1609 | struct invariant *inv = NULL; | |
1610 | ||
1611 | if (!invariants.length ()) | |
1612 | return; | |
1613 | ||
1614 | if (flag_ira_loop_pressure) | |
1615 | /* REGS_USED is actually never used when the flag is on. */ | |
1616 | regs_used = 0; | |
1617 | else | |
1618 | /* We do not really do a good job in estimating number of | |
1619 | registers used; we put some initial bound here to stand for | |
1620 | induction variables etc. that we do not detect. */ | |
1621 | { | |
1622 | unsigned int n_regs = DF_REG_SIZE (df); | |
1623 | ||
1624 | regs_used = 2; | |
1625 | ||
1626 | for (i = 0; i < n_regs; i++) | |
1627 | { | |
1628 | if (!DF_REGNO_FIRST_DEF (i) && DF_REGNO_LAST_USE (i)) | |
1629 | { | |
1630 | /* This is a value that is used but not changed inside loop. */ | |
1631 | regs_used++; | |
1632 | } | |
1633 | } | |
1634 | } | |
1635 | ||
1636 | if (! flag_ira_loop_pressure) | |
1637 | new_regs[0] = regs_needed[0] = 0; | |
1638 | else | |
1639 | { | |
1640 | for (i = 0; (int) i < ira_pressure_classes_num; i++) | |
1641 | new_regs[ira_pressure_classes[i]] = 0; | |
1642 | } | |
1643 | while ((gain = best_gain_for_invariant (&inv, regs_needed, | |
1644 | new_regs, regs_used, | |
1645 | speed, call_p)) > 0) | |
1646 | { | |
1647 | set_move_mark (inv->invno, gain); | |
1648 | if (! flag_ira_loop_pressure) | |
1649 | new_regs[0] += regs_needed[0]; | |
1650 | else | |
1651 | { | |
1652 | for (i = 0; (int) i < ira_pressure_classes_num; i++) | |
1653 | new_regs[ira_pressure_classes[i]] | |
1654 | += regs_needed[ira_pressure_classes[i]]; | |
1655 | } | |
1656 | } | |
1657 | } | |
1658 | ||
1659 | /* Replace the uses, reached by the definition of invariant INV, by REG. | |
1660 | ||
1661 | IN_GROUP is nonzero if this is part of a group of changes that must be | |
1662 | performed as a group. In that case, the changes will be stored. The | |
1663 | function `apply_change_group' will validate and apply the changes. */ | |
1664 | ||
1665 | static int | |
1666 | replace_uses (struct invariant *inv, rtx reg, bool in_group) | |
1667 | { | |
1668 | /* Replace the uses we know to be dominated. It saves work for copy | |
1669 | propagation, and also it is necessary so that dependent invariants | |
1670 | are computed right. */ | |
1671 | if (inv->def) | |
1672 | { | |
1673 | struct use *use; | |
1674 | for (use = inv->def->uses; use; use = use->next) | |
1675 | validate_change (use->insn, use->pos, reg, true); | |
1676 | ||
1677 | /* If we aren't part of a larger group, apply the changes now. */ | |
1678 | if (!in_group) | |
1679 | return apply_change_group (); | |
1680 | } | |
1681 | ||
1682 | return 1; | |
1683 | } | |
1684 | ||
1685 | /* Whether invariant INV setting REG can be moved out of LOOP, at the end of | |
1686 | the block preceding its header. */ | |
1687 | ||
1688 | static bool | |
1689 | can_move_invariant_reg (class loop *loop, struct invariant *inv, rtx reg) | |
1690 | { | |
1691 | df_ref def, use; | |
1692 | unsigned int dest_regno, defs_in_loop_count = 0; | |
1693 | rtx_insn *insn = inv->insn; | |
1694 | basic_block bb = BLOCK_FOR_INSN (inv->insn); | |
1695 | ||
1696 | /* We ignore hard register and memory access for cost and complexity reasons. | |
1697 | Hard register are few at this stage and expensive to consider as they | |
1698 | require building a separate data flow. Memory access would require using | |
1699 | df_simulate_* and can_move_insns_across functions and is more complex. */ | |
1700 | if (!REG_P (reg) || HARD_REGISTER_P (reg)) | |
1701 | return false; | |
1702 | ||
1703 | /* Check whether the set is always executed. We could omit this condition if | |
1704 | we know that the register is unused outside of the loop, but it does not | |
1705 | seem worth finding out. */ | |
1706 | if (!inv->always_executed) | |
1707 | return false; | |
1708 | ||
1709 | /* Check that all uses that would be dominated by def are already dominated | |
1710 | by it. */ | |
1711 | dest_regno = REGNO (reg); | |
1712 | for (use = DF_REG_USE_CHAIN (dest_regno); use; use = DF_REF_NEXT_REG (use)) | |
1713 | { | |
1714 | rtx_insn *use_insn; | |
1715 | basic_block use_bb; | |
1716 | ||
1717 | use_insn = DF_REF_INSN (use); | |
1718 | use_bb = BLOCK_FOR_INSN (use_insn); | |
1719 | ||
1720 | /* Ignore instruction considered for moving. */ | |
1721 | if (use_insn == insn) | |
1722 | continue; | |
1723 | ||
1724 | /* Don't consider uses outside loop. */ | |
1725 | if (!flow_bb_inside_loop_p (loop, use_bb)) | |
1726 | continue; | |
1727 | ||
1728 | /* Don't move if a use is not dominated by def in insn. */ | |
1729 | if (use_bb == bb && DF_INSN_LUID (insn) >= DF_INSN_LUID (use_insn)) | |
1730 | return false; | |
1731 | if (!dominated_by_p (CDI_DOMINATORS, use_bb, bb)) | |
1732 | return false; | |
1733 | } | |
1734 | ||
1735 | /* Check for other defs. Any other def in the loop might reach a use | |
1736 | currently reached by the def in insn. */ | |
1737 | for (def = DF_REG_DEF_CHAIN (dest_regno); def; def = DF_REF_NEXT_REG (def)) | |
1738 | { | |
1739 | basic_block def_bb = DF_REF_BB (def); | |
1740 | ||
1741 | /* Defs in exit block cannot reach a use they weren't already. */ | |
1742 | if (single_succ_p (def_bb)) | |
1743 | { | |
1744 | basic_block def_bb_succ; | |
1745 | ||
1746 | def_bb_succ = single_succ (def_bb); | |
1747 | if (!flow_bb_inside_loop_p (loop, def_bb_succ)) | |
1748 | continue; | |
1749 | } | |
1750 | ||
1751 | if (++defs_in_loop_count > 1) | |
1752 | return false; | |
1753 | } | |
1754 | ||
1755 | return true; | |
1756 | } | |
1757 | ||
1758 | /* Move invariant INVNO out of the LOOP. Returns true if this succeeds, false | |
1759 | otherwise. */ | |
1760 | ||
1761 | static bool | |
1762 | move_invariant_reg (class loop *loop, unsigned invno) | |
1763 | { | |
1764 | struct invariant *inv = invariants[invno]; | |
1765 | struct invariant *repr = invariants[inv->eqto]; | |
1766 | unsigned i; | |
1767 | basic_block preheader = loop_preheader_edge (loop)->src; | |
1768 | rtx reg, set, dest, note; | |
1769 | bitmap_iterator bi; | |
1770 | int regno = -1; | |
1771 | ||
1772 | if (inv->reg) | |
1773 | return true; | |
1774 | if (!repr->move) | |
1775 | return false; | |
1776 | ||
1777 | /* If this is a representative of the class of equivalent invariants, | |
1778 | really move the invariant. Otherwise just replace its use with | |
1779 | the register used for the representative. */ | |
1780 | if (inv == repr) | |
1781 | { | |
1782 | if (inv->depends_on) | |
1783 | { | |
1784 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, i, bi) | |
1785 | { | |
1786 | if (!move_invariant_reg (loop, i)) | |
1787 | goto fail; | |
1788 | } | |
1789 | } | |
1790 | ||
1791 | /* If possible, just move the set out of the loop. Otherwise, we | |
1792 | need to create a temporary register. */ | |
1793 | set = single_set (inv->insn); | |
1794 | reg = dest = SET_DEST (set); | |
1795 | if (GET_CODE (reg) == SUBREG) | |
1796 | reg = SUBREG_REG (reg); | |
1797 | if (REG_P (reg)) | |
1798 | regno = REGNO (reg); | |
1799 | ||
1800 | if (!can_move_invariant_reg (loop, inv, dest)) | |
1801 | { | |
1802 | reg = gen_reg_rtx_and_attrs (dest); | |
1803 | ||
1804 | /* Try replacing the destination by a new pseudoregister. */ | |
1805 | validate_change (inv->insn, &SET_DEST (set), reg, true); | |
1806 | ||
1807 | /* As well as all the dominated uses. */ | |
1808 | replace_uses (inv, reg, true); | |
1809 | ||
1810 | /* And validate all the changes. */ | |
1811 | if (!apply_change_group ()) | |
1812 | goto fail; | |
1813 | ||
1814 | emit_insn_after (gen_move_insn (dest, reg), inv->insn); | |
1815 | } | |
1816 | else if (dump_file) | |
1817 | fprintf (dump_file, "Invariant %d moved without introducing a new " | |
1818 | "temporary register\n", invno); | |
1819 | reorder_insns (inv->insn, inv->insn, BB_END (preheader)); | |
1820 | df_recompute_luids (preheader); | |
1821 | ||
1822 | /* If there is a REG_EQUAL note on the insn we just moved, and the | |
1823 | insn is in a basic block that is not always executed or the note | |
1824 | contains something for which we don't know the invariant status, | |
1825 | the note may no longer be valid after we move the insn. Note that | |
1826 | uses in REG_EQUAL notes are taken into account in the computation | |
1827 | of invariants, so it is safe to retain the note even if it contains | |
1828 | register references for which we know the invariant status. */ | |
1829 | if ((note = find_reg_note (inv->insn, REG_EQUAL, NULL_RTX)) | |
1830 | && (!inv->always_executed | |
1831 | || !check_maybe_invariant (XEXP (note, 0)))) | |
1832 | remove_note (inv->insn, note); | |
1833 | } | |
1834 | else | |
1835 | { | |
1836 | if (!move_invariant_reg (loop, repr->invno)) | |
1837 | goto fail; | |
1838 | reg = repr->reg; | |
1839 | regno = repr->orig_regno; | |
1840 | if (!replace_uses (inv, reg, false)) | |
1841 | goto fail; | |
1842 | set = single_set (inv->insn); | |
1843 | emit_insn_after (gen_move_insn (SET_DEST (set), reg), inv->insn); | |
1844 | delete_insn (inv->insn); | |
1845 | } | |
1846 | ||
1847 | inv->reg = reg; | |
1848 | inv->orig_regno = regno; | |
1849 | ||
1850 | return true; | |
1851 | ||
1852 | fail: | |
1853 | /* If we failed, clear move flag, so that we do not try to move inv | |
1854 | again. */ | |
1855 | if (dump_file) | |
1856 | fprintf (dump_file, "Failed to move invariant %d\n", invno); | |
1857 | inv->move = false; | |
1858 | inv->reg = NULL_RTX; | |
1859 | inv->orig_regno = -1; | |
1860 | ||
1861 | return false; | |
1862 | } | |
1863 | ||
1864 | /* Move selected invariant out of the LOOP. Newly created regs are marked | |
1865 | in TEMPORARY_REGS. */ | |
1866 | ||
1867 | static void | |
1868 | move_invariants (class loop *loop) | |
1869 | { | |
1870 | struct invariant *inv; | |
1871 | unsigned i; | |
1872 | ||
1873 | FOR_EACH_VEC_ELT (invariants, i, inv) | |
1874 | move_invariant_reg (loop, i); | |
1875 | if (flag_ira_loop_pressure && resize_reg_info ()) | |
1876 | { | |
1877 | FOR_EACH_VEC_ELT (invariants, i, inv) | |
1878 | if (inv->reg != NULL_RTX) | |
1879 | { | |
1880 | if (inv->orig_regno >= 0) | |
1881 | setup_reg_classes (REGNO (inv->reg), | |
1882 | reg_preferred_class (inv->orig_regno), | |
1883 | reg_alternate_class (inv->orig_regno), | |
1884 | reg_allocno_class (inv->orig_regno)); | |
1885 | else | |
1886 | setup_reg_classes (REGNO (inv->reg), | |
1887 | GENERAL_REGS, NO_REGS, GENERAL_REGS); | |
1888 | } | |
1889 | } | |
1890 | /* Remove the DF_UD_CHAIN problem added in find_defs before rescanning, | |
1891 | to save a bit of compile time. */ | |
1892 | df_remove_problem (df_chain); | |
1893 | df_process_deferred_rescans (); | |
1894 | } | |
1895 | ||
1896 | /* Initializes invariant motion data. */ | |
1897 | ||
1898 | static void | |
1899 | init_inv_motion_data (void) | |
1900 | { | |
1901 | actual_stamp = 1; | |
1902 | ||
1903 | invariants.create (100); | |
1904 | } | |
1905 | ||
1906 | /* Frees the data allocated by invariant motion. */ | |
1907 | ||
1908 | static void | |
1909 | free_inv_motion_data (void) | |
1910 | { | |
1911 | unsigned i; | |
1912 | struct def *def; | |
1913 | struct invariant *inv; | |
1914 | ||
1915 | check_invariant_table_size (); | |
1916 | for (i = 0; i < DF_DEFS_TABLE_SIZE (); i++) | |
1917 | { | |
1918 | inv = invariant_table[i]; | |
1919 | if (inv) | |
1920 | { | |
1921 | def = inv->def; | |
1922 | gcc_assert (def != NULL); | |
1923 | ||
1924 | free_use_list (def->uses); | |
1925 | free (def); | |
1926 | invariant_table[i] = NULL; | |
1927 | } | |
1928 | } | |
1929 | ||
1930 | FOR_EACH_VEC_ELT (invariants, i, inv) | |
1931 | { | |
1932 | BITMAP_FREE (inv->depends_on); | |
1933 | free (inv); | |
1934 | } | |
1935 | invariants.release (); | |
1936 | } | |
1937 | ||
1938 | /* Move the invariants out of the LOOP. */ | |
1939 | ||
1940 | static void | |
1941 | move_single_loop_invariants (class loop *loop) | |
1942 | { | |
1943 | init_inv_motion_data (); | |
1944 | ||
1945 | find_invariants (loop); | |
1946 | find_invariants_to_move (optimize_loop_for_speed_p (loop), | |
1947 | LOOP_DATA (loop)->has_call); | |
1948 | move_invariants (loop); | |
1949 | ||
1950 | free_inv_motion_data (); | |
1951 | } | |
1952 | ||
1953 | /* Releases the auxiliary data for LOOP. */ | |
1954 | ||
1955 | static void | |
1956 | free_loop_data (class loop *loop) | |
1957 | { | |
1958 | class loop_data *data = LOOP_DATA (loop); | |
1959 | if (!data) | |
1960 | return; | |
1961 | ||
1962 | bitmap_clear (&LOOP_DATA (loop)->regs_ref); | |
1963 | bitmap_clear (&LOOP_DATA (loop)->regs_live); | |
1964 | free (data); | |
1965 | loop->aux = NULL; | |
1966 | } | |
1967 | ||
1968 | \f | |
1969 | ||
1970 | /* Registers currently living. */ | |
1971 | static bitmap_head curr_regs_live; | |
1972 | ||
1973 | /* Current reg pressure for each pressure class. */ | |
1974 | static int curr_reg_pressure[N_REG_CLASSES]; | |
1975 | ||
1976 | /* Record all regs that are set in any one insn. Communication from | |
1977 | mark_reg_{store,clobber} and global_conflicts. Asm can refer to | |
1978 | all hard-registers. */ | |
1979 | static rtx regs_set[(FIRST_PSEUDO_REGISTER > MAX_RECOG_OPERANDS | |
1980 | ? FIRST_PSEUDO_REGISTER : MAX_RECOG_OPERANDS) * 2]; | |
1981 | /* Number of regs stored in the previous array. */ | |
1982 | static int n_regs_set; | |
1983 | ||
1984 | /* Return pressure class and number of needed hard registers (through | |
1985 | *NREGS) of register REGNO. */ | |
1986 | static enum reg_class | |
1987 | get_regno_pressure_class (int regno, int *nregs) | |
1988 | { | |
1989 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1990 | { | |
1991 | enum reg_class pressure_class; | |
1992 | ||
1993 | pressure_class = reg_allocno_class (regno); | |
1994 | pressure_class = ira_pressure_class_translate[pressure_class]; | |
1995 | *nregs | |
1996 | = ira_reg_class_max_nregs[pressure_class][PSEUDO_REGNO_MODE (regno)]; | |
1997 | return pressure_class; | |
1998 | } | |
1999 | else if (! TEST_HARD_REG_BIT (ira_no_alloc_regs, regno) | |
2000 | && ! TEST_HARD_REG_BIT (eliminable_regset, regno)) | |
2001 | { | |
2002 | *nregs = 1; | |
2003 | return ira_pressure_class_translate[REGNO_REG_CLASS (regno)]; | |
2004 | } | |
2005 | else | |
2006 | { | |
2007 | *nregs = 0; | |
2008 | return NO_REGS; | |
2009 | } | |
2010 | } | |
2011 | ||
2012 | /* Increase (if INCR_P) or decrease current register pressure for | |
2013 | register REGNO. */ | |
2014 | static void | |
2015 | change_pressure (int regno, bool incr_p) | |
2016 | { | |
2017 | int nregs; | |
2018 | enum reg_class pressure_class; | |
2019 | ||
2020 | pressure_class = get_regno_pressure_class (regno, &nregs); | |
2021 | if (! incr_p) | |
2022 | curr_reg_pressure[pressure_class] -= nregs; | |
2023 | else | |
2024 | { | |
2025 | curr_reg_pressure[pressure_class] += nregs; | |
2026 | if (LOOP_DATA (curr_loop)->max_reg_pressure[pressure_class] | |
2027 | < curr_reg_pressure[pressure_class]) | |
2028 | LOOP_DATA (curr_loop)->max_reg_pressure[pressure_class] | |
2029 | = curr_reg_pressure[pressure_class]; | |
2030 | } | |
2031 | } | |
2032 | ||
2033 | /* Mark REGNO birth. */ | |
2034 | static void | |
2035 | mark_regno_live (int regno) | |
2036 | { | |
2037 | class loop *loop; | |
2038 | ||
2039 | for (loop = curr_loop; | |
2040 | loop != current_loops->tree_root; | |
2041 | loop = loop_outer (loop)) | |
2042 | bitmap_set_bit (&LOOP_DATA (loop)->regs_live, regno); | |
2043 | if (!bitmap_set_bit (&curr_regs_live, regno)) | |
2044 | return; | |
2045 | change_pressure (regno, true); | |
2046 | } | |
2047 | ||
2048 | /* Mark REGNO death. */ | |
2049 | static void | |
2050 | mark_regno_death (int regno) | |
2051 | { | |
2052 | if (! bitmap_clear_bit (&curr_regs_live, regno)) | |
2053 | return; | |
2054 | change_pressure (regno, false); | |
2055 | } | |
2056 | ||
2057 | /* Mark setting register REG. */ | |
2058 | static void | |
2059 | mark_reg_store (rtx reg, const_rtx setter ATTRIBUTE_UNUSED, | |
2060 | void *data ATTRIBUTE_UNUSED) | |
2061 | { | |
2062 | if (GET_CODE (reg) == SUBREG) | |
2063 | reg = SUBREG_REG (reg); | |
2064 | ||
2065 | if (! REG_P (reg)) | |
2066 | return; | |
2067 | ||
2068 | regs_set[n_regs_set++] = reg; | |
2069 | ||
2070 | unsigned int end_regno = END_REGNO (reg); | |
2071 | for (unsigned int regno = REGNO (reg); regno < end_regno; ++regno) | |
2072 | mark_regno_live (regno); | |
2073 | } | |
2074 | ||
2075 | /* Mark clobbering register REG. */ | |
2076 | static void | |
2077 | mark_reg_clobber (rtx reg, const_rtx setter, void *data) | |
2078 | { | |
2079 | if (GET_CODE (setter) == CLOBBER) | |
2080 | mark_reg_store (reg, setter, data); | |
2081 | } | |
2082 | ||
2083 | /* Mark register REG death. */ | |
2084 | static void | |
2085 | mark_reg_death (rtx reg) | |
2086 | { | |
2087 | unsigned int end_regno = END_REGNO (reg); | |
2088 | for (unsigned int regno = REGNO (reg); regno < end_regno; ++regno) | |
2089 | mark_regno_death (regno); | |
2090 | } | |
2091 | ||
2092 | /* Mark occurrence of registers in X for the current loop. */ | |
2093 | static void | |
2094 | mark_ref_regs (rtx x) | |
2095 | { | |
2096 | RTX_CODE code; | |
2097 | int i; | |
2098 | const char *fmt; | |
2099 | ||
2100 | if (!x) | |
2101 | return; | |
2102 | ||
2103 | code = GET_CODE (x); | |
2104 | if (code == REG) | |
2105 | { | |
2106 | class loop *loop; | |
2107 | ||
2108 | for (loop = curr_loop; | |
2109 | loop != current_loops->tree_root; | |
2110 | loop = loop_outer (loop)) | |
2111 | bitmap_set_bit (&LOOP_DATA (loop)->regs_ref, REGNO (x)); | |
2112 | return; | |
2113 | } | |
2114 | ||
2115 | fmt = GET_RTX_FORMAT (code); | |
2116 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
2117 | if (fmt[i] == 'e') | |
2118 | mark_ref_regs (XEXP (x, i)); | |
2119 | else if (fmt[i] == 'E') | |
2120 | { | |
2121 | int j; | |
2122 | ||
2123 | for (j = 0; j < XVECLEN (x, i); j++) | |
2124 | mark_ref_regs (XVECEXP (x, i, j)); | |
2125 | } | |
2126 | } | |
2127 | ||
2128 | /* Calculate register pressure in the loops. */ | |
2129 | static void | |
2130 | calculate_loop_reg_pressure (void) | |
2131 | { | |
2132 | int i; | |
2133 | unsigned int j; | |
2134 | bitmap_iterator bi; | |
2135 | basic_block bb; | |
2136 | rtx_insn *insn; | |
2137 | rtx link; | |
2138 | class loop *loop, *parent; | |
2139 | ||
2140 | FOR_EACH_LOOP (loop, 0) | |
2141 | if (loop->aux == NULL) | |
2142 | { | |
2143 | loop->aux = xcalloc (1, sizeof (class loop_data)); | |
2144 | bitmap_initialize (&LOOP_DATA (loop)->regs_ref, ®_obstack); | |
2145 | bitmap_initialize (&LOOP_DATA (loop)->regs_live, ®_obstack); | |
2146 | } | |
2147 | ira_setup_eliminable_regset (); | |
2148 | bitmap_initialize (&curr_regs_live, ®_obstack); | |
2149 | FOR_EACH_BB_FN (bb, cfun) | |
2150 | { | |
2151 | curr_loop = bb->loop_father; | |
2152 | if (curr_loop == current_loops->tree_root) | |
2153 | continue; | |
2154 | ||
2155 | for (loop = curr_loop; | |
2156 | loop != current_loops->tree_root; | |
2157 | loop = loop_outer (loop)) | |
2158 | bitmap_ior_into (&LOOP_DATA (loop)->regs_live, DF_LR_IN (bb)); | |
2159 | ||
2160 | bitmap_copy (&curr_regs_live, DF_LR_IN (bb)); | |
2161 | for (i = 0; i < ira_pressure_classes_num; i++) | |
2162 | curr_reg_pressure[ira_pressure_classes[i]] = 0; | |
2163 | EXECUTE_IF_SET_IN_BITMAP (&curr_regs_live, 0, j, bi) | |
2164 | change_pressure (j, true); | |
2165 | ||
2166 | FOR_BB_INSNS (bb, insn) | |
2167 | { | |
2168 | if (! NONDEBUG_INSN_P (insn)) | |
2169 | continue; | |
2170 | ||
2171 | mark_ref_regs (PATTERN (insn)); | |
2172 | n_regs_set = 0; | |
2173 | note_stores (insn, mark_reg_clobber, NULL); | |
2174 | ||
2175 | /* Mark any registers dead after INSN as dead now. */ | |
2176 | ||
2177 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
2178 | if (REG_NOTE_KIND (link) == REG_DEAD) | |
2179 | mark_reg_death (XEXP (link, 0)); | |
2180 | ||
2181 | /* Mark any registers set in INSN as live, | |
2182 | and mark them as conflicting with all other live regs. | |
2183 | Clobbers are processed again, so they conflict with | |
2184 | the registers that are set. */ | |
2185 | ||
2186 | note_stores (insn, mark_reg_store, NULL); | |
2187 | ||
2188 | if (AUTO_INC_DEC) | |
2189 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
2190 | if (REG_NOTE_KIND (link) == REG_INC) | |
2191 | mark_reg_store (XEXP (link, 0), NULL_RTX, NULL); | |
2192 | ||
2193 | while (n_regs_set-- > 0) | |
2194 | { | |
2195 | rtx note = find_regno_note (insn, REG_UNUSED, | |
2196 | REGNO (regs_set[n_regs_set])); | |
2197 | if (! note) | |
2198 | continue; | |
2199 | ||
2200 | mark_reg_death (XEXP (note, 0)); | |
2201 | } | |
2202 | } | |
2203 | } | |
2204 | bitmap_release (&curr_regs_live); | |
2205 | if (flag_ira_region == IRA_REGION_MIXED | |
2206 | || flag_ira_region == IRA_REGION_ALL) | |
2207 | FOR_EACH_LOOP (loop, 0) | |
2208 | { | |
2209 | EXECUTE_IF_SET_IN_BITMAP (&LOOP_DATA (loop)->regs_live, 0, j, bi) | |
2210 | if (! bitmap_bit_p (&LOOP_DATA (loop)->regs_ref, j)) | |
2211 | { | |
2212 | enum reg_class pressure_class; | |
2213 | int nregs; | |
2214 | ||
2215 | pressure_class = get_regno_pressure_class (j, &nregs); | |
2216 | LOOP_DATA (loop)->max_reg_pressure[pressure_class] -= nregs; | |
2217 | } | |
2218 | } | |
2219 | if (dump_file == NULL) | |
2220 | return; | |
2221 | FOR_EACH_LOOP (loop, 0) | |
2222 | { | |
2223 | parent = loop_outer (loop); | |
2224 | fprintf (dump_file, "\n Loop %d (parent %d, header bb%d, depth %d)\n", | |
2225 | loop->num, (parent == NULL ? -1 : parent->num), | |
2226 | loop->header->index, loop_depth (loop)); | |
2227 | fprintf (dump_file, "\n ref. regnos:"); | |
2228 | EXECUTE_IF_SET_IN_BITMAP (&LOOP_DATA (loop)->regs_ref, 0, j, bi) | |
2229 | fprintf (dump_file, " %d", j); | |
2230 | fprintf (dump_file, "\n live regnos:"); | |
2231 | EXECUTE_IF_SET_IN_BITMAP (&LOOP_DATA (loop)->regs_live, 0, j, bi) | |
2232 | fprintf (dump_file, " %d", j); | |
2233 | fprintf (dump_file, "\n Pressure:"); | |
2234 | for (i = 0; (int) i < ira_pressure_classes_num; i++) | |
2235 | { | |
2236 | enum reg_class pressure_class; | |
2237 | ||
2238 | pressure_class = ira_pressure_classes[i]; | |
2239 | if (LOOP_DATA (loop)->max_reg_pressure[pressure_class] == 0) | |
2240 | continue; | |
2241 | fprintf (dump_file, " %s=%d", reg_class_names[pressure_class], | |
2242 | LOOP_DATA (loop)->max_reg_pressure[pressure_class]); | |
2243 | } | |
2244 | fprintf (dump_file, "\n"); | |
2245 | } | |
2246 | } | |
2247 | ||
2248 | \f | |
2249 | ||
2250 | /* Move the invariants out of the loops. */ | |
2251 | ||
2252 | void | |
2253 | move_loop_invariants (void) | |
2254 | { | |
2255 | class loop *loop; | |
2256 | ||
2257 | if (optimize == 1) | |
2258 | df_live_add_problem (); | |
2259 | /* ??? This is a hack. We should only need to call df_live_set_all_dirty | |
2260 | for optimize == 1, but can_move_invariant_reg relies on DF_INSN_LUID | |
2261 | being up-to-date. That isn't always true (even after df_analyze) | |
2262 | because df_process_deferred_rescans doesn't necessarily cause | |
2263 | blocks to be rescanned. */ | |
2264 | df_live_set_all_dirty (); | |
2265 | if (flag_ira_loop_pressure) | |
2266 | { | |
2267 | df_analyze (); | |
2268 | regstat_init_n_sets_and_refs (); | |
2269 | ira_set_pseudo_classes (true, dump_file); | |
2270 | calculate_loop_reg_pressure (); | |
2271 | regstat_free_n_sets_and_refs (); | |
2272 | } | |
2273 | df_set_flags (DF_EQ_NOTES + DF_DEFER_INSN_RESCAN); | |
2274 | /* Process the loops, innermost first. */ | |
2275 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) | |
2276 | { | |
2277 | curr_loop = loop; | |
2278 | /* move_single_loop_invariants for very large loops is time consuming | |
2279 | and might need a lot of memory. For -O1 only do loop invariant | |
2280 | motion for very small loops. */ | |
2281 | unsigned max_bbs = param_loop_invariant_max_bbs_in_loop; | |
2282 | if (optimize < 2) | |
2283 | max_bbs /= 10; | |
2284 | if (loop->num_nodes <= max_bbs) | |
2285 | move_single_loop_invariants (loop); | |
2286 | } | |
2287 | ||
2288 | FOR_EACH_LOOP (loop, 0) | |
2289 | { | |
2290 | free_loop_data (loop); | |
2291 | } | |
2292 | ||
2293 | if (flag_ira_loop_pressure) | |
2294 | /* There is no sense to keep this info because it was most | |
2295 | probably outdated by subsequent passes. */ | |
2296 | free_reg_info (); | |
2297 | free (invariant_table); | |
2298 | invariant_table = NULL; | |
2299 | invariant_table_size = 0; | |
2300 | ||
2301 | if (optimize == 1) | |
2302 | df_remove_problem (df_live); | |
2303 | ||
2304 | checking_verify_flow_info (); | |
2305 | } |