1 | /* RTL-level loop invariant motion. |
2 | Copyright (C) 2004-2024 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.cc or in tree-ssa-pre.cc. |
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 (s: &invariant_table[invariant_table_size], c: 0, |
192 | n: (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 UNSPEC_VOLATILE: |
216 | case CALL: |
217 | return false; |
218 | |
219 | case REG: |
220 | return true; |
221 | |
222 | case MEM: |
223 | /* Load/store motion is done elsewhere. ??? Perhaps also add it here? |
224 | It should not be hard, and might be faster than "elsewhere". */ |
225 | |
226 | /* Just handle the most trivial case where we load from an unchanging |
227 | location (most importantly, pic tables). */ |
228 | if (MEM_READONLY_P (x) && !MEM_VOLATILE_P (x)) |
229 | break; |
230 | |
231 | return false; |
232 | |
233 | case ASM_OPERANDS: |
234 | /* Don't mess with insns declared volatile. */ |
235 | if (MEM_VOLATILE_P (x)) |
236 | return false; |
237 | break; |
238 | |
239 | default: |
240 | break; |
241 | } |
242 | |
243 | fmt = GET_RTX_FORMAT (code); |
244 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
245 | { |
246 | if (fmt[i] == 'e') |
247 | { |
248 | if (!check_maybe_invariant (XEXP (x, i))) |
249 | return false; |
250 | } |
251 | else if (fmt[i] == 'E') |
252 | { |
253 | for (j = 0; j < XVECLEN (x, i); j++) |
254 | if (!check_maybe_invariant (XVECEXP (x, i, j))) |
255 | return false; |
256 | } |
257 | } |
258 | |
259 | return true; |
260 | } |
261 | |
262 | /* Returns the invariant definition for USE, or NULL if USE is not |
263 | invariant. */ |
264 | |
265 | static struct invariant * |
266 | invariant_for_use (df_ref use) |
267 | { |
268 | struct df_link *defs; |
269 | df_ref def; |
270 | basic_block bb = DF_REF_BB (use), def_bb; |
271 | |
272 | if (DF_REF_FLAGS (use) & DF_REF_READ_WRITE) |
273 | return NULL; |
274 | |
275 | defs = DF_REF_CHAIN (use); |
276 | if (!defs || defs->next) |
277 | return NULL; |
278 | def = defs->ref; |
279 | check_invariant_table_size (); |
280 | if (!invariant_table[DF_REF_ID (def)]) |
281 | return NULL; |
282 | |
283 | def_bb = DF_REF_BB (def); |
284 | if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb)) |
285 | return NULL; |
286 | return invariant_table[DF_REF_ID (def)]; |
287 | } |
288 | |
289 | /* Computes hash value for invariant expression X in INSN. */ |
290 | |
291 | static hashval_t |
292 | hash_invariant_expr_1 (rtx_insn *insn, rtx x) |
293 | { |
294 | enum rtx_code code = GET_CODE (x); |
295 | int i, j; |
296 | const char *fmt; |
297 | hashval_t val = code; |
298 | int do_not_record_p; |
299 | df_ref use; |
300 | struct invariant *inv; |
301 | |
302 | switch (code) |
303 | { |
304 | CASE_CONST_ANY: |
305 | case SYMBOL_REF: |
306 | case CONST: |
307 | case LABEL_REF: |
308 | return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false); |
309 | |
310 | case REG: |
311 | use = df_find_use (insn, x); |
312 | if (!use) |
313 | return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false); |
314 | inv = invariant_for_use (use); |
315 | if (!inv) |
316 | return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false); |
317 | |
318 | gcc_assert (inv->eqto != ~0u); |
319 | return inv->eqto; |
320 | |
321 | default: |
322 | break; |
323 | } |
324 | |
325 | fmt = GET_RTX_FORMAT (code); |
326 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
327 | { |
328 | if (fmt[i] == 'e') |
329 | val ^= hash_invariant_expr_1 (insn, XEXP (x, i)); |
330 | else if (fmt[i] == 'E') |
331 | { |
332 | for (j = 0; j < XVECLEN (x, i); j++) |
333 | val ^= hash_invariant_expr_1 (insn, XVECEXP (x, i, j)); |
334 | } |
335 | else if (fmt[i] == 'i' || fmt[i] == 'n') |
336 | val ^= XINT (x, i); |
337 | else if (fmt[i] == 'p') |
338 | val ^= constant_lower_bound (SUBREG_BYTE (x)); |
339 | } |
340 | |
341 | return val; |
342 | } |
343 | |
344 | /* Returns true if the invariant expressions E1 and E2 used in insns INSN1 |
345 | and INSN2 have always the same value. */ |
346 | |
347 | static bool |
348 | invariant_expr_equal_p (rtx_insn *insn1, rtx e1, rtx_insn *insn2, rtx e2) |
349 | { |
350 | enum rtx_code code = GET_CODE (e1); |
351 | int i, j; |
352 | const char *fmt; |
353 | df_ref use1, use2; |
354 | struct invariant *inv1 = NULL, *inv2 = NULL; |
355 | rtx sub1, sub2; |
356 | |
357 | /* If mode of only one of the operands is VOIDmode, it is not equivalent to |
358 | the other one. If both are VOIDmode, we rely on the caller of this |
359 | function to verify that their modes are the same. */ |
360 | if (code != GET_CODE (e2) || GET_MODE (e1) != GET_MODE (e2)) |
361 | return false; |
362 | |
363 | switch (code) |
364 | { |
365 | CASE_CONST_ANY: |
366 | case SYMBOL_REF: |
367 | case CONST: |
368 | case LABEL_REF: |
369 | return rtx_equal_p (e1, e2); |
370 | |
371 | case REG: |
372 | use1 = df_find_use (insn1, e1); |
373 | use2 = df_find_use (insn2, e2); |
374 | if (use1) |
375 | inv1 = invariant_for_use (use: use1); |
376 | if (use2) |
377 | inv2 = invariant_for_use (use: use2); |
378 | |
379 | if (!inv1 && !inv2) |
380 | return rtx_equal_p (e1, e2); |
381 | |
382 | if (!inv1 || !inv2) |
383 | return false; |
384 | |
385 | gcc_assert (inv1->eqto != ~0u); |
386 | gcc_assert (inv2->eqto != ~0u); |
387 | return inv1->eqto == inv2->eqto; |
388 | |
389 | default: |
390 | break; |
391 | } |
392 | |
393 | fmt = GET_RTX_FORMAT (code); |
394 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
395 | { |
396 | if (fmt[i] == 'e') |
397 | { |
398 | sub1 = XEXP (e1, i); |
399 | sub2 = XEXP (e2, i); |
400 | |
401 | if (!invariant_expr_equal_p (insn1, e1: sub1, insn2, e2: sub2)) |
402 | return false; |
403 | } |
404 | |
405 | else if (fmt[i] == 'E') |
406 | { |
407 | if (XVECLEN (e1, i) != XVECLEN (e2, i)) |
408 | return false; |
409 | |
410 | for (j = 0; j < XVECLEN (e1, i); j++) |
411 | { |
412 | sub1 = XVECEXP (e1, i, j); |
413 | sub2 = XVECEXP (e2, i, j); |
414 | |
415 | if (!invariant_expr_equal_p (insn1, e1: sub1, insn2, e2: sub2)) |
416 | return false; |
417 | } |
418 | } |
419 | else if (fmt[i] == 'i' || fmt[i] == 'n') |
420 | { |
421 | if (XINT (e1, i) != XINT (e2, i)) |
422 | return false; |
423 | } |
424 | else if (fmt[i] == 'p') |
425 | { |
426 | if (maybe_ne (SUBREG_BYTE (e1), SUBREG_BYTE (e2))) |
427 | return false; |
428 | } |
429 | /* Unhandled type of subexpression, we fail conservatively. */ |
430 | else |
431 | return false; |
432 | } |
433 | |
434 | return true; |
435 | } |
436 | |
437 | struct invariant_expr_hasher : free_ptr_hash <invariant_expr_entry> |
438 | { |
439 | static inline hashval_t hash (const invariant_expr_entry *); |
440 | static inline bool equal (const invariant_expr_entry *, |
441 | const invariant_expr_entry *); |
442 | }; |
443 | |
444 | /* Returns hash value for invariant expression entry ENTRY. */ |
445 | |
446 | inline hashval_t |
447 | invariant_expr_hasher::hash (const invariant_expr_entry *entry) |
448 | { |
449 | return entry->hash; |
450 | } |
451 | |
452 | /* Compares invariant expression entries ENTRY1 and ENTRY2. */ |
453 | |
454 | inline bool |
455 | invariant_expr_hasher::equal (const invariant_expr_entry *entry1, |
456 | const invariant_expr_entry *entry2) |
457 | { |
458 | if (entry1->mode != entry2->mode) |
459 | return 0; |
460 | |
461 | return invariant_expr_equal_p (insn1: entry1->inv->insn, e1: entry1->expr, |
462 | insn2: entry2->inv->insn, e2: entry2->expr); |
463 | } |
464 | |
465 | typedef hash_table<invariant_expr_hasher> invariant_htab_type; |
466 | |
467 | /* Checks whether invariant with value EXPR in machine mode MODE is |
468 | recorded in EQ. If this is the case, return the invariant. Otherwise |
469 | insert INV to the table for this expression and return INV. */ |
470 | |
471 | static struct invariant * |
472 | find_or_insert_inv (invariant_htab_type *eq, rtx expr, machine_mode mode, |
473 | struct invariant *inv) |
474 | { |
475 | hashval_t hash = hash_invariant_expr_1 (insn: inv->insn, x: expr); |
476 | struct invariant_expr_entry *entry; |
477 | struct invariant_expr_entry pentry; |
478 | invariant_expr_entry **slot; |
479 | |
480 | pentry.expr = expr; |
481 | pentry.inv = inv; |
482 | pentry.mode = mode; |
483 | slot = eq->find_slot_with_hash (comparable: &pentry, hash, insert: INSERT); |
484 | entry = *slot; |
485 | |
486 | if (entry) |
487 | return entry->inv; |
488 | |
489 | entry = XNEW (struct invariant_expr_entry); |
490 | entry->inv = inv; |
491 | entry->expr = expr; |
492 | entry->mode = mode; |
493 | entry->hash = hash; |
494 | *slot = entry; |
495 | |
496 | return inv; |
497 | } |
498 | |
499 | /* Finds invariants identical to INV and records the equivalence. EQ is the |
500 | hash table of the invariants. */ |
501 | |
502 | static void |
503 | find_identical_invariants (invariant_htab_type *eq, struct invariant *inv) |
504 | { |
505 | unsigned depno; |
506 | bitmap_iterator bi; |
507 | struct invariant *dep; |
508 | rtx expr, set; |
509 | machine_mode mode; |
510 | struct invariant *tmp; |
511 | |
512 | if (inv->eqto != ~0u) |
513 | return; |
514 | |
515 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi) |
516 | { |
517 | dep = invariants[depno]; |
518 | find_identical_invariants (eq, inv: dep); |
519 | } |
520 | |
521 | set = single_set (insn: inv->insn); |
522 | expr = SET_SRC (set); |
523 | mode = GET_MODE (expr); |
524 | if (mode == VOIDmode) |
525 | mode = GET_MODE (SET_DEST (set)); |
526 | |
527 | tmp = find_or_insert_inv (eq, expr, mode, inv); |
528 | inv->eqto = tmp->invno; |
529 | |
530 | if (tmp->invno != inv->invno && inv->always_executed) |
531 | tmp->eqno++; |
532 | |
533 | if (dump_file && inv->eqto != inv->invno) |
534 | fprintf (stream: dump_file, |
535 | format: "Invariant %d is equivalent to invariant %d.\n" , |
536 | inv->invno, inv->eqto); |
537 | } |
538 | |
539 | /* Find invariants with the same value and record the equivalences. */ |
540 | |
541 | static void |
542 | merge_identical_invariants (void) |
543 | { |
544 | unsigned i; |
545 | struct invariant *inv; |
546 | invariant_htab_type eq (invariants.length ()); |
547 | |
548 | FOR_EACH_VEC_ELT (invariants, i, inv) |
549 | find_identical_invariants (eq: &eq, inv); |
550 | } |
551 | |
552 | /* Determines the basic blocks inside LOOP that are always executed and |
553 | stores their bitmap to ALWAYS_REACHED. MAY_EXIT is a bitmap of |
554 | basic blocks that may either exit the loop, or contain the call that |
555 | does not have to return. BODY is body of the loop obtained by |
556 | get_loop_body_in_dom_order. */ |
557 | |
558 | static void |
559 | compute_always_reached (class loop *loop, basic_block *body, |
560 | bitmap may_exit, bitmap always_reached) |
561 | { |
562 | unsigned i; |
563 | |
564 | for (i = 0; i < loop->num_nodes; i++) |
565 | { |
566 | if (dominated_by_p (CDI_DOMINATORS, loop->latch, body[i])) |
567 | bitmap_set_bit (always_reached, i); |
568 | |
569 | if (bitmap_bit_p (may_exit, i)) |
570 | return; |
571 | } |
572 | } |
573 | |
574 | /* Finds exits out of the LOOP with body BODY. Marks blocks in that we may |
575 | exit the loop by cfg edge to HAS_EXIT and MAY_EXIT. In MAY_EXIT |
576 | additionally mark blocks that may exit due to a call. */ |
577 | |
578 | static void |
579 | find_exits (class loop *loop, basic_block *body, |
580 | bitmap may_exit, bitmap has_exit) |
581 | { |
582 | unsigned i; |
583 | edge_iterator ei; |
584 | edge e; |
585 | class loop *outermost_exit = loop, *aexit; |
586 | bool has_call = false; |
587 | rtx_insn *insn; |
588 | |
589 | for (i = 0; i < loop->num_nodes; i++) |
590 | { |
591 | if (body[i]->loop_father == loop) |
592 | { |
593 | FOR_BB_INSNS (body[i], insn) |
594 | { |
595 | if (CALL_P (insn) |
596 | && (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn) |
597 | || !RTL_CONST_OR_PURE_CALL_P (insn))) |
598 | { |
599 | has_call = true; |
600 | bitmap_set_bit (may_exit, i); |
601 | break; |
602 | } |
603 | } |
604 | |
605 | FOR_EACH_EDGE (e, ei, body[i]->succs) |
606 | { |
607 | if (! flow_bb_inside_loop_p (loop, e->dest)) |
608 | { |
609 | bitmap_set_bit (may_exit, i); |
610 | bitmap_set_bit (has_exit, i); |
611 | outermost_exit = find_common_loop (outermost_exit, |
612 | e->dest->loop_father); |
613 | } |
614 | /* If we enter a subloop that might never terminate treat |
615 | it like a possible exit. */ |
616 | if (flow_loop_nested_p (loop, e->dest->loop_father)) |
617 | bitmap_set_bit (may_exit, i); |
618 | } |
619 | continue; |
620 | } |
621 | |
622 | /* Use the data stored for the subloop to decide whether we may exit |
623 | through it. It is sufficient to do this for header of the loop, |
624 | as other basic blocks inside it must be dominated by it. */ |
625 | if (body[i]->loop_father->header != body[i]) |
626 | continue; |
627 | |
628 | if (LOOP_DATA (body[i]->loop_father)->has_call) |
629 | { |
630 | has_call = true; |
631 | bitmap_set_bit (may_exit, i); |
632 | } |
633 | aexit = LOOP_DATA (body[i]->loop_father)->outermost_exit; |
634 | if (aexit != loop) |
635 | { |
636 | bitmap_set_bit (may_exit, i); |
637 | bitmap_set_bit (has_exit, i); |
638 | |
639 | if (flow_loop_nested_p (aexit, outermost_exit)) |
640 | outermost_exit = aexit; |
641 | } |
642 | } |
643 | |
644 | if (loop->aux == NULL) |
645 | { |
646 | loop->aux = xcalloc (1, sizeof (class loop_data)); |
647 | bitmap_initialize (head: &LOOP_DATA (loop)->regs_ref, obstack: ®_obstack); |
648 | bitmap_initialize (head: &LOOP_DATA (loop)->regs_live, obstack: ®_obstack); |
649 | } |
650 | LOOP_DATA (loop)->outermost_exit = outermost_exit; |
651 | LOOP_DATA (loop)->has_call = has_call; |
652 | } |
653 | |
654 | /* Check whether we may assign a value to X from a register. */ |
655 | |
656 | static bool |
657 | may_assign_reg_p (rtx x) |
658 | { |
659 | return (GET_MODE (x) != VOIDmode |
660 | && GET_MODE (x) != BLKmode |
661 | && can_copy_p (GET_MODE (x)) |
662 | /* Do not mess with the frame pointer adjustments that can |
663 | be generated e.g. by expand_builtin_setjmp_receiver. */ |
664 | && x != frame_pointer_rtx |
665 | && (!REG_P (x) |
666 | || !HARD_REGISTER_P (x) |
667 | || REGNO_REG_CLASS (REGNO (x)) != NO_REGS)); |
668 | } |
669 | |
670 | /* Finds definitions that may correspond to invariants in LOOP with body |
671 | BODY. */ |
672 | |
673 | static void |
674 | find_defs (class loop *loop) |
675 | { |
676 | if (dump_file) |
677 | { |
678 | fprintf (stream: dump_file, |
679 | format: "*****starting processing of loop %d ******\n" , |
680 | loop->num); |
681 | } |
682 | |
683 | df_chain_add_problem (DF_UD_CHAIN); |
684 | df_set_flags (DF_RD_PRUNE_DEAD_DEFS); |
685 | df_analyze_loop (loop); |
686 | check_invariant_table_size (); |
687 | |
688 | if (dump_file) |
689 | { |
690 | df_dump_region (dump_file); |
691 | fprintf (stream: dump_file, |
692 | format: "*****ending processing of loop %d ******\n" , |
693 | loop->num); |
694 | } |
695 | } |
696 | |
697 | /* Creates a new invariant for definition DEF in INSN, depending on invariants |
698 | in DEPENDS_ON. ALWAYS_EXECUTED is true if the insn is always executed, |
699 | unless the program ends due to a function call. The newly created invariant |
700 | is returned. */ |
701 | |
702 | static struct invariant * |
703 | create_new_invariant (struct def *def, rtx_insn *insn, bitmap depends_on, |
704 | bool always_executed) |
705 | { |
706 | struct invariant *inv = XNEW (struct invariant); |
707 | rtx set = single_set (insn); |
708 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); |
709 | |
710 | inv->def = def; |
711 | inv->always_executed = always_executed; |
712 | inv->depends_on = depends_on; |
713 | |
714 | /* If the set is simple, usually by moving it we move the whole store out of |
715 | the loop. Otherwise we save only cost of the computation. */ |
716 | if (def) |
717 | { |
718 | inv->cost = set_rtx_cost (x: set, speed_p: speed); |
719 | /* ??? Try to determine cheapness of address computation. Unfortunately |
720 | the address cost is only a relative measure, we can't really compare |
721 | it with any absolute number, but only with other address costs. |
722 | But here we don't have any other addresses, so compare with a magic |
723 | number anyway. It has to be large enough to not regress PR33928 |
724 | (by avoiding to move reg+8,reg+16,reg+24 invariants), but small |
725 | enough to not regress 410.bwaves either (by still moving reg+reg |
726 | invariants). |
727 | See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ |
728 | if (SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set)))) |
729 | inv->cheap_address = address_cost (SET_SRC (set), word_mode, |
730 | ADDR_SPACE_GENERIC, speed) < 3; |
731 | else |
732 | inv->cheap_address = false; |
733 | } |
734 | else |
735 | { |
736 | inv->cost = set_src_cost (SET_SRC (set), GET_MODE (SET_DEST (set)), |
737 | speed_p: speed); |
738 | inv->cheap_address = false; |
739 | } |
740 | |
741 | inv->move = false; |
742 | inv->reg = NULL_RTX; |
743 | inv->orig_regno = -1; |
744 | inv->stamp = 0; |
745 | inv->insn = insn; |
746 | |
747 | inv->invno = invariants.length (); |
748 | inv->eqto = ~0u; |
749 | |
750 | /* Itself. */ |
751 | inv->eqno = 1; |
752 | |
753 | if (def) |
754 | def->invno = inv->invno; |
755 | invariants.safe_push (obj: inv); |
756 | |
757 | if (dump_file) |
758 | { |
759 | fprintf (stream: dump_file, |
760 | format: "Set in insn %d is invariant (%d), cost %d, depends on " , |
761 | INSN_UID (insn), inv->invno, inv->cost); |
762 | dump_bitmap (file: dump_file, map: inv->depends_on); |
763 | } |
764 | |
765 | return inv; |
766 | } |
767 | |
768 | /* Return a canonical version of X for the address, from the point of view, |
769 | that all multiplications are represented as MULT instead of the multiply |
770 | by a power of 2 being represented as ASHIFT. |
771 | |
772 | Callers should prepare a copy of X because this function may modify it |
773 | in place. */ |
774 | |
775 | static void |
776 | canonicalize_address_mult (rtx x) |
777 | { |
778 | subrtx_var_iterator::array_type array; |
779 | FOR_EACH_SUBRTX_VAR (iter, array, x, NONCONST) |
780 | { |
781 | rtx sub = *iter; |
782 | scalar_int_mode sub_mode; |
783 | if (is_a <scalar_int_mode> (GET_MODE (sub), result: &sub_mode) |
784 | && GET_CODE (sub) == ASHIFT |
785 | && CONST_INT_P (XEXP (sub, 1)) |
786 | && INTVAL (XEXP (sub, 1)) < GET_MODE_BITSIZE (mode: sub_mode) |
787 | && INTVAL (XEXP (sub, 1)) >= 0) |
788 | { |
789 | HOST_WIDE_INT shift = INTVAL (XEXP (sub, 1)); |
790 | PUT_CODE (sub, MULT); |
791 | XEXP (sub, 1) = gen_int_mode (HOST_WIDE_INT_1 << shift, sub_mode); |
792 | iter.skip_subrtxes (); |
793 | } |
794 | } |
795 | } |
796 | |
797 | /* Maximum number of sub expressions in address. We set it to |
798 | a small integer since it's unlikely to have a complicated |
799 | address expression. */ |
800 | |
801 | #define MAX_CANON_ADDR_PARTS (5) |
802 | |
803 | /* Collect sub expressions in address X with PLUS as the seperator. |
804 | Sub expressions are stored in vector ADDR_PARTS. */ |
805 | |
806 | static void |
807 | collect_address_parts (rtx x, vec<rtx> *addr_parts) |
808 | { |
809 | subrtx_var_iterator::array_type array; |
810 | FOR_EACH_SUBRTX_VAR (iter, array, x, NONCONST) |
811 | { |
812 | rtx sub = *iter; |
813 | |
814 | if (GET_CODE (sub) != PLUS) |
815 | { |
816 | addr_parts->safe_push (obj: sub); |
817 | iter.skip_subrtxes (); |
818 | } |
819 | } |
820 | } |
821 | |
822 | /* Compare function for sorting sub expressions X and Y based on |
823 | precedence defined for communitive operations. */ |
824 | |
825 | static int |
826 | compare_address_parts (const void *x, const void *y) |
827 | { |
828 | const rtx *rx = (const rtx *)x; |
829 | const rtx *ry = (const rtx *)y; |
830 | int px = commutative_operand_precedence (*rx); |
831 | int py = commutative_operand_precedence (*ry); |
832 | |
833 | return (py - px); |
834 | } |
835 | |
836 | /* Return a canonical version address for X by following steps: |
837 | 1) Rewrite ASHIFT into MULT recursively. |
838 | 2) Divide address into sub expressions with PLUS as the |
839 | separator. |
840 | 3) Sort sub expressions according to precedence defined |
841 | for communative operations. |
842 | 4) Simplify CONST_INT_P sub expressions. |
843 | 5) Create new canonicalized address and return. |
844 | Callers should prepare a copy of X because this function may |
845 | modify it in place. */ |
846 | |
847 | static rtx |
848 | canonicalize_address (rtx x) |
849 | { |
850 | rtx res; |
851 | unsigned int i, j; |
852 | machine_mode mode = GET_MODE (x); |
853 | auto_vec<rtx, MAX_CANON_ADDR_PARTS> addr_parts; |
854 | |
855 | /* Rewrite ASHIFT into MULT. */ |
856 | canonicalize_address_mult (x); |
857 | /* Divide address into sub expressions. */ |
858 | collect_address_parts (x, addr_parts: &addr_parts); |
859 | /* Unlikely to have very complicated address. */ |
860 | if (addr_parts.length () < 2 |
861 | || addr_parts.length () > MAX_CANON_ADDR_PARTS) |
862 | return x; |
863 | |
864 | /* Sort sub expressions according to canonicalization precedence. */ |
865 | addr_parts.qsort (compare_address_parts); |
866 | |
867 | /* Simplify all constant int summary if possible. */ |
868 | for (i = 0; i < addr_parts.length (); i++) |
869 | if (CONST_INT_P (addr_parts[i])) |
870 | break; |
871 | |
872 | for (j = i + 1; j < addr_parts.length (); j++) |
873 | { |
874 | gcc_assert (CONST_INT_P (addr_parts[j])); |
875 | addr_parts[i] = simplify_gen_binary (code: PLUS, mode, |
876 | op0: addr_parts[i], |
877 | op1: addr_parts[j]); |
878 | } |
879 | |
880 | /* Chain PLUS operators to the left for !CONST_INT_P sub expressions. */ |
881 | res = addr_parts[0]; |
882 | for (j = 1; j < i; j++) |
883 | res = simplify_gen_binary (code: PLUS, mode, op0: res, op1: addr_parts[j]); |
884 | |
885 | /* Pickup the last CONST_INT_P sub expression. */ |
886 | if (i < addr_parts.length ()) |
887 | res = simplify_gen_binary (code: PLUS, mode, op0: res, op1: addr_parts[i]); |
888 | |
889 | return res; |
890 | } |
891 | |
892 | /* Given invariant DEF and its address USE, check if the corresponding |
893 | invariant expr can be propagated into the use or not. */ |
894 | |
895 | static bool |
896 | inv_can_prop_to_addr_use (struct def *def, df_ref use) |
897 | { |
898 | struct invariant *inv; |
899 | rtx *pos = DF_REF_REAL_LOC (use), def_set, use_set; |
900 | rtx_insn *use_insn = DF_REF_INSN (use); |
901 | rtx_insn *def_insn; |
902 | bool ok; |
903 | |
904 | inv = invariants[def->invno]; |
905 | /* No need to check if address expression is expensive. */ |
906 | if (!inv->cheap_address) |
907 | return false; |
908 | |
909 | def_insn = inv->insn; |
910 | def_set = single_set (insn: def_insn); |
911 | if (!def_set) |
912 | return false; |
913 | |
914 | validate_unshare_change (use_insn, pos, SET_SRC (def_set), true); |
915 | ok = verify_changes (0); |
916 | /* Try harder with canonicalization in address expression. */ |
917 | if (!ok && (use_set = single_set (insn: use_insn)) != NULL_RTX) |
918 | { |
919 | rtx src, dest, mem = NULL_RTX; |
920 | |
921 | src = SET_SRC (use_set); |
922 | dest = SET_DEST (use_set); |
923 | if (MEM_P (src)) |
924 | mem = src; |
925 | else if (MEM_P (dest)) |
926 | mem = dest; |
927 | |
928 | if (mem != NULL_RTX |
929 | && !memory_address_addr_space_p (GET_MODE (mem), |
930 | XEXP (mem, 0), |
931 | MEM_ADDR_SPACE (mem))) |
932 | { |
933 | rtx addr = canonicalize_address (x: copy_rtx (XEXP (mem, 0))); |
934 | if (memory_address_addr_space_p (GET_MODE (mem), |
935 | addr, MEM_ADDR_SPACE (mem))) |
936 | ok = true; |
937 | } |
938 | } |
939 | cancel_changes (0); |
940 | return ok; |
941 | } |
942 | |
943 | /* Record USE at DEF. */ |
944 | |
945 | static void |
946 | record_use (struct def *def, df_ref use) |
947 | { |
948 | struct use *u = XNEW (struct use); |
949 | |
950 | u->pos = DF_REF_REAL_LOC (use); |
951 | u->insn = DF_REF_INSN (use); |
952 | u->addr_use_p = (DF_REF_TYPE (use) == DF_REF_REG_MEM_LOAD |
953 | || DF_REF_TYPE (use) == DF_REF_REG_MEM_STORE); |
954 | u->next = def->uses; |
955 | def->uses = u; |
956 | def->n_uses++; |
957 | if (u->addr_use_p) |
958 | { |
959 | /* Initialize propagation information if this is the first addr |
960 | use of the inv def. */ |
961 | if (def->n_addr_uses == 0) |
962 | def->can_prop_to_addr_uses = true; |
963 | |
964 | def->n_addr_uses++; |
965 | if (def->can_prop_to_addr_uses && !inv_can_prop_to_addr_use (def, use)) |
966 | def->can_prop_to_addr_uses = false; |
967 | } |
968 | } |
969 | |
970 | /* Finds the invariants USE depends on and store them to the DEPENDS_ON |
971 | bitmap. Returns true if all dependencies of USE are known to be |
972 | loop invariants, false otherwise. */ |
973 | |
974 | static bool |
975 | check_dependency (basic_block bb, df_ref use, bitmap depends_on) |
976 | { |
977 | df_ref def; |
978 | basic_block def_bb; |
979 | struct df_link *defs; |
980 | struct def *def_data; |
981 | struct invariant *inv; |
982 | |
983 | if (DF_REF_FLAGS (use) & DF_REF_READ_WRITE) |
984 | return false; |
985 | |
986 | defs = DF_REF_CHAIN (use); |
987 | if (!defs) |
988 | { |
989 | unsigned int regno = DF_REF_REGNO (use); |
990 | |
991 | /* If this is the use of an uninitialized argument register that is |
992 | likely to be spilled, do not move it lest this might extend its |
993 | lifetime and cause reload to die. This can occur for a call to |
994 | a function taking complex number arguments and moving the insns |
995 | preparing the arguments without moving the call itself wouldn't |
996 | gain much in practice. */ |
997 | if ((DF_REF_FLAGS (use) & DF_HARD_REG_LIVE) |
998 | && FUNCTION_ARG_REGNO_P (regno) |
999 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno))) |
1000 | return false; |
1001 | |
1002 | return true; |
1003 | } |
1004 | |
1005 | if (defs->next) |
1006 | return false; |
1007 | |
1008 | def = defs->ref; |
1009 | check_invariant_table_size (); |
1010 | inv = invariant_table[DF_REF_ID (def)]; |
1011 | if (!inv) |
1012 | return false; |
1013 | |
1014 | def_data = inv->def; |
1015 | gcc_assert (def_data != NULL); |
1016 | |
1017 | def_bb = DF_REF_BB (def); |
1018 | /* Note that in case bb == def_bb, we know that the definition |
1019 | dominates insn, because def has invariant_table[DF_REF_ID(def)] |
1020 | defined and we process the insns in the basic block bb |
1021 | sequentially. */ |
1022 | if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb)) |
1023 | return false; |
1024 | |
1025 | bitmap_set_bit (depends_on, def_data->invno); |
1026 | return true; |
1027 | } |
1028 | |
1029 | |
1030 | /* Finds the invariants INSN depends on and store them to the DEPENDS_ON |
1031 | bitmap. Returns true if all dependencies of INSN are known to be |
1032 | loop invariants, false otherwise. */ |
1033 | |
1034 | static bool |
1035 | check_dependencies (rtx_insn *insn, bitmap depends_on) |
1036 | { |
1037 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
1038 | df_ref use; |
1039 | basic_block bb = BLOCK_FOR_INSN (insn); |
1040 | |
1041 | FOR_EACH_INSN_INFO_USE (use, insn_info) |
1042 | if (!check_dependency (bb, use, depends_on)) |
1043 | return false; |
1044 | FOR_EACH_INSN_INFO_EQ_USE (use, insn_info) |
1045 | if (!check_dependency (bb, use, depends_on)) |
1046 | return false; |
1047 | |
1048 | return true; |
1049 | } |
1050 | |
1051 | /* Pre-check candidate DEST to skip the one which cannot make a valid insn |
1052 | during move_invariant_reg. SIMPLE is to skip HARD_REGISTER. */ |
1053 | static bool |
1054 | pre_check_invariant_p (bool simple, rtx dest) |
1055 | { |
1056 | if (simple && REG_P (dest) && DF_REG_DEF_COUNT (REGNO (dest)) > 1) |
1057 | { |
1058 | df_ref use; |
1059 | unsigned int i = REGNO (dest); |
1060 | struct df_insn_info *insn_info; |
1061 | df_ref def_rec; |
1062 | |
1063 | for (use = DF_REG_USE_CHAIN (i); use; use = DF_REF_NEXT_REG (use)) |
1064 | { |
1065 | rtx_insn *ref = DF_REF_INSN (use); |
1066 | insn_info = DF_INSN_INFO_GET (ref); |
1067 | |
1068 | FOR_EACH_INSN_INFO_DEF (def_rec, insn_info) |
1069 | if (DF_REF_REGNO (def_rec) == i) |
1070 | { |
1071 | /* Multi definitions at this stage, most likely are due to |
1072 | instruction constraints, which requires both read and write |
1073 | on the same register. Since move_invariant_reg is not |
1074 | powerful enough to handle such cases, just ignore the INV |
1075 | and leave the chance to others. */ |
1076 | return false; |
1077 | } |
1078 | } |
1079 | } |
1080 | return true; |
1081 | } |
1082 | |
1083 | /* Finds invariant in INSN. ALWAYS_REACHED is true if the insn is always |
1084 | executed. ALWAYS_EXECUTED is true if the insn is always executed, |
1085 | unless the program ends due to a function call. */ |
1086 | |
1087 | static void |
1088 | find_invariant_insn (rtx_insn *insn, bool always_reached, bool always_executed) |
1089 | { |
1090 | df_ref ref; |
1091 | struct def *def; |
1092 | bitmap depends_on; |
1093 | rtx set, dest; |
1094 | bool simple = true; |
1095 | struct invariant *inv; |
1096 | |
1097 | /* Jumps have control flow side-effects. */ |
1098 | if (JUMP_P (insn)) |
1099 | return; |
1100 | |
1101 | set = single_set (insn); |
1102 | if (!set) |
1103 | return; |
1104 | dest = SET_DEST (set); |
1105 | |
1106 | if (!REG_P (dest) |
1107 | || HARD_REGISTER_P (dest)) |
1108 | simple = false; |
1109 | |
1110 | if (!may_assign_reg_p (x: dest) |
1111 | || !pre_check_invariant_p (simple, dest) |
1112 | || !check_maybe_invariant (SET_SRC (set))) |
1113 | return; |
1114 | |
1115 | /* If the insn can throw exception, we cannot move it at all without changing |
1116 | cfg. */ |
1117 | if (can_throw_internal (insn)) |
1118 | return; |
1119 | |
1120 | /* We cannot make trapping insn executed, unless it was executed before. */ |
1121 | if (may_trap_or_fault_p (PATTERN (insn)) && !always_reached) |
1122 | return; |
1123 | |
1124 | depends_on = BITMAP_ALLOC (NULL); |
1125 | if (!check_dependencies (insn, depends_on)) |
1126 | { |
1127 | BITMAP_FREE (depends_on); |
1128 | return; |
1129 | } |
1130 | |
1131 | if (simple) |
1132 | def = XCNEW (struct def); |
1133 | else |
1134 | def = NULL; |
1135 | |
1136 | inv = create_new_invariant (def, insn, depends_on, always_executed); |
1137 | |
1138 | if (simple) |
1139 | { |
1140 | ref = df_find_def (insn, dest); |
1141 | check_invariant_table_size (); |
1142 | invariant_table[DF_REF_ID (ref)] = inv; |
1143 | } |
1144 | } |
1145 | |
1146 | /* Record registers used in INSN that have a unique invariant definition. */ |
1147 | |
1148 | static void |
1149 | record_uses (rtx_insn *insn) |
1150 | { |
1151 | struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); |
1152 | df_ref use; |
1153 | struct invariant *inv; |
1154 | |
1155 | FOR_EACH_INSN_INFO_USE (use, insn_info) |
1156 | { |
1157 | inv = invariant_for_use (use); |
1158 | if (inv) |
1159 | record_use (def: inv->def, use); |
1160 | } |
1161 | FOR_EACH_INSN_INFO_EQ_USE (use, insn_info) |
1162 | { |
1163 | inv = invariant_for_use (use); |
1164 | if (inv) |
1165 | record_use (def: inv->def, use); |
1166 | } |
1167 | } |
1168 | |
1169 | /* Finds invariants in INSN. ALWAYS_REACHED is true if the insn is always |
1170 | executed. ALWAYS_EXECUTED is true if the insn is always executed, |
1171 | unless the program ends due to a function call. */ |
1172 | |
1173 | static void |
1174 | find_invariants_insn (rtx_insn *insn, bool always_reached, bool always_executed) |
1175 | { |
1176 | find_invariant_insn (insn, always_reached, always_executed); |
1177 | record_uses (insn); |
1178 | } |
1179 | |
1180 | /* Finds invariants in basic block BB. ALWAYS_REACHED is true if the |
1181 | basic block is always executed. ALWAYS_EXECUTED is true if the basic |
1182 | block is always executed, unless the program ends due to a function |
1183 | call. */ |
1184 | |
1185 | static void |
1186 | find_invariants_bb (class loop *loop, basic_block bb, bool always_reached, |
1187 | bool always_executed) |
1188 | { |
1189 | rtx_insn *insn; |
1190 | basic_block = loop_preheader_edge (loop)->src; |
1191 | |
1192 | /* Don't move insn of cold BB out of loop to preheader to reduce calculations |
1193 | and register live range in hot loop with cold BB. */ |
1194 | if (!always_executed && preheader->count > bb->count) |
1195 | { |
1196 | if (dump_file) |
1197 | fprintf (stream: dump_file, format: "Don't move invariant from bb: %d out of loop %d\n" , |
1198 | bb->index, loop->num); |
1199 | return; |
1200 | } |
1201 | |
1202 | FOR_BB_INSNS (bb, insn) |
1203 | { |
1204 | if (!NONDEBUG_INSN_P (insn)) |
1205 | continue; |
1206 | |
1207 | find_invariants_insn (insn, always_reached, always_executed); |
1208 | |
1209 | if (always_reached |
1210 | && CALL_P (insn) |
1211 | && (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn) |
1212 | || ! RTL_CONST_OR_PURE_CALL_P (insn))) |
1213 | always_reached = false; |
1214 | } |
1215 | } |
1216 | |
1217 | /* Finds invariants in LOOP with body BODY. ALWAYS_REACHED is the bitmap of |
1218 | basic blocks in BODY that are always executed. ALWAYS_EXECUTED is the |
1219 | bitmap of basic blocks in BODY that are always executed unless the program |
1220 | ends due to a function call. */ |
1221 | |
1222 | static void |
1223 | find_invariants_body (class loop *loop, basic_block *body, |
1224 | bitmap always_reached, bitmap always_executed) |
1225 | { |
1226 | unsigned i; |
1227 | |
1228 | for (i = 0; i < loop->num_nodes; i++) |
1229 | find_invariants_bb (loop, bb: body[i], always_reached: bitmap_bit_p (always_reached, i), |
1230 | always_executed: bitmap_bit_p (always_executed, i)); |
1231 | } |
1232 | |
1233 | /* Finds invariants in LOOP. */ |
1234 | |
1235 | static void |
1236 | find_invariants (class loop *loop) |
1237 | { |
1238 | auto_bitmap may_exit; |
1239 | auto_bitmap always_reached; |
1240 | auto_bitmap has_exit; |
1241 | auto_bitmap always_executed; |
1242 | basic_block *body = get_loop_body_in_dom_order (loop); |
1243 | |
1244 | find_exits (loop, body, may_exit, has_exit); |
1245 | compute_always_reached (loop, body, may_exit, always_reached); |
1246 | compute_always_reached (loop, body, may_exit: has_exit, always_reached: always_executed); |
1247 | |
1248 | find_defs (loop); |
1249 | find_invariants_body (loop, body, always_reached, always_executed); |
1250 | merge_identical_invariants (); |
1251 | |
1252 | free (ptr: body); |
1253 | } |
1254 | |
1255 | /* Frees a list of uses USE. */ |
1256 | |
1257 | static void |
1258 | free_use_list (struct use *use) |
1259 | { |
1260 | struct use *next; |
1261 | |
1262 | for (; use; use = next) |
1263 | { |
1264 | next = use->next; |
1265 | free (ptr: use); |
1266 | } |
1267 | } |
1268 | |
1269 | /* Return pressure class and number of hard registers (through *NREGS) |
1270 | for destination of INSN. */ |
1271 | static enum reg_class |
1272 | get_pressure_class_and_nregs (rtx_insn *insn, int *nregs) |
1273 | { |
1274 | rtx reg; |
1275 | enum reg_class pressure_class; |
1276 | rtx set = single_set (insn); |
1277 | |
1278 | /* Considered invariant insns have only one set. */ |
1279 | gcc_assert (set != NULL_RTX); |
1280 | reg = SET_DEST (set); |
1281 | if (GET_CODE (reg) == SUBREG) |
1282 | reg = SUBREG_REG (reg); |
1283 | if (MEM_P (reg)) |
1284 | { |
1285 | *nregs = 0; |
1286 | pressure_class = NO_REGS; |
1287 | } |
1288 | else |
1289 | { |
1290 | if (! REG_P (reg)) |
1291 | reg = NULL_RTX; |
1292 | if (reg == NULL_RTX) |
1293 | pressure_class = GENERAL_REGS; |
1294 | else |
1295 | { |
1296 | pressure_class = reg_allocno_class (REGNO (reg)); |
1297 | pressure_class = ira_pressure_class_translate[pressure_class]; |
1298 | } |
1299 | *nregs |
1300 | = ira_reg_class_max_nregs[pressure_class][GET_MODE (SET_SRC (set))]; |
1301 | } |
1302 | return pressure_class; |
1303 | } |
1304 | |
1305 | /* Calculates cost and number of registers needed for moving invariant INV |
1306 | out of the loop and stores them to *COST and *REGS_NEEDED. *CL will be |
1307 | the REG_CLASS of INV. Return |
1308 | -1: if INV is invalid. |
1309 | 0: if INV and its depends_on have same reg_class |
1310 | 1: if INV and its depends_on have different reg_classes. */ |
1311 | |
1312 | static int |
1313 | get_inv_cost (struct invariant *inv, int *comp_cost, unsigned *regs_needed, |
1314 | enum reg_class *cl) |
1315 | { |
1316 | int i, acomp_cost; |
1317 | unsigned aregs_needed[N_REG_CLASSES]; |
1318 | unsigned depno; |
1319 | struct invariant *dep; |
1320 | bitmap_iterator bi; |
1321 | int ret = 1; |
1322 | |
1323 | /* Find the representative of the class of the equivalent invariants. */ |
1324 | inv = invariants[inv->eqto]; |
1325 | |
1326 | *comp_cost = 0; |
1327 | if (! flag_ira_loop_pressure) |
1328 | regs_needed[0] = 0; |
1329 | else |
1330 | { |
1331 | for (i = 0; i < ira_pressure_classes_num; i++) |
1332 | regs_needed[ira_pressure_classes[i]] = 0; |
1333 | } |
1334 | |
1335 | if (inv->move |
1336 | || inv->stamp == actual_stamp) |
1337 | return -1; |
1338 | inv->stamp = actual_stamp; |
1339 | |
1340 | if (! flag_ira_loop_pressure) |
1341 | regs_needed[0]++; |
1342 | else |
1343 | { |
1344 | int nregs; |
1345 | enum reg_class pressure_class; |
1346 | |
1347 | pressure_class = get_pressure_class_and_nregs (insn: inv->insn, nregs: &nregs); |
1348 | regs_needed[pressure_class] += nregs; |
1349 | *cl = pressure_class; |
1350 | ret = 0; |
1351 | } |
1352 | |
1353 | if (!inv->cheap_address |
1354 | || inv->def->n_uses == 0 |
1355 | || inv->def->n_addr_uses < inv->def->n_uses |
1356 | /* Count cost if the inv can't be propagated into address uses. */ |
1357 | || !inv->def->can_prop_to_addr_uses) |
1358 | (*comp_cost) += inv->cost * inv->eqno; |
1359 | |
1360 | #ifdef STACK_REGS |
1361 | { |
1362 | /* Hoisting constant pool constants into stack regs may cost more than |
1363 | just single register. On x87, the balance is affected both by the |
1364 | small number of FP registers, and by its register stack organization, |
1365 | that forces us to add compensation code in and around the loop to |
1366 | shuffle the operands to the top of stack before use, and pop them |
1367 | from the stack after the loop finishes. |
1368 | |
1369 | To model this effect, we increase the number of registers needed for |
1370 | stack registers by two: one register push, and one register pop. |
1371 | This usually has the effect that FP constant loads from the constant |
1372 | pool are not moved out of the loop. |
1373 | |
1374 | Note that this also means that dependent invariants cannot be moved. |
1375 | However, the primary purpose of this pass is to move loop invariant |
1376 | address arithmetic out of loops, and address arithmetic that depends |
1377 | on floating point constants is unlikely to ever occur. */ |
1378 | rtx set = single_set (insn: inv->insn); |
1379 | if (set |
1380 | && IS_STACK_MODE (GET_MODE (SET_SRC (set))) |
1381 | && constant_pool_constant_p (SET_SRC (set))) |
1382 | { |
1383 | if (flag_ira_loop_pressure) |
1384 | regs_needed[ira_stack_reg_pressure_class] += 2; |
1385 | else |
1386 | regs_needed[0] += 2; |
1387 | } |
1388 | } |
1389 | #endif |
1390 | |
1391 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi) |
1392 | { |
1393 | bool check_p; |
1394 | enum reg_class dep_cl = ALL_REGS; |
1395 | int dep_ret; |
1396 | |
1397 | dep = invariants[depno]; |
1398 | |
1399 | /* If DEP is moved out of the loop, it is not a depends_on any more. */ |
1400 | if (dep->move) |
1401 | continue; |
1402 | |
1403 | dep_ret = get_inv_cost (inv: dep, comp_cost: &acomp_cost, regs_needed: aregs_needed, cl: &dep_cl); |
1404 | |
1405 | if (! flag_ira_loop_pressure) |
1406 | check_p = aregs_needed[0] != 0; |
1407 | else |
1408 | { |
1409 | for (i = 0; i < ira_pressure_classes_num; i++) |
1410 | if (aregs_needed[ira_pressure_classes[i]] != 0) |
1411 | break; |
1412 | check_p = i < ira_pressure_classes_num; |
1413 | |
1414 | if ((dep_ret == 1) || ((dep_ret == 0) && (*cl != dep_cl))) |
1415 | { |
1416 | *cl = ALL_REGS; |
1417 | ret = 1; |
1418 | } |
1419 | } |
1420 | if (check_p |
1421 | /* We need to check always_executed, since if the original value of |
1422 | the invariant may be preserved, we may need to keep it in a |
1423 | separate register. TODO check whether the register has an |
1424 | use outside of the loop. */ |
1425 | && dep->always_executed |
1426 | && !dep->def->uses->next) |
1427 | { |
1428 | /* If this is a single use, after moving the dependency we will not |
1429 | need a new register. */ |
1430 | if (! flag_ira_loop_pressure) |
1431 | aregs_needed[0]--; |
1432 | else |
1433 | { |
1434 | int nregs; |
1435 | enum reg_class pressure_class; |
1436 | |
1437 | pressure_class = get_pressure_class_and_nregs (insn: inv->insn, nregs: &nregs); |
1438 | aregs_needed[pressure_class] -= nregs; |
1439 | } |
1440 | } |
1441 | |
1442 | if (! flag_ira_loop_pressure) |
1443 | regs_needed[0] += aregs_needed[0]; |
1444 | else |
1445 | { |
1446 | for (i = 0; i < ira_pressure_classes_num; i++) |
1447 | regs_needed[ira_pressure_classes[i]] |
1448 | += aregs_needed[ira_pressure_classes[i]]; |
1449 | } |
1450 | (*comp_cost) += acomp_cost; |
1451 | } |
1452 | return ret; |
1453 | } |
1454 | |
1455 | /* Calculates gain for eliminating invariant INV. REGS_USED is the number |
1456 | of registers used in the loop, NEW_REGS is the number of new variables |
1457 | already added due to the invariant motion. The number of registers needed |
1458 | for it is stored in *REGS_NEEDED. SPEED and CALL_P are flags passed |
1459 | through to estimate_reg_pressure_cost. */ |
1460 | |
1461 | static int |
1462 | gain_for_invariant (struct invariant *inv, unsigned *regs_needed, |
1463 | unsigned *new_regs, unsigned regs_used, |
1464 | bool speed, bool call_p) |
1465 | { |
1466 | int comp_cost, size_cost; |
1467 | /* Workaround -Wmaybe-uninitialized false positive during |
1468 | profiledbootstrap by initializing it. */ |
1469 | enum reg_class cl = NO_REGS; |
1470 | int ret; |
1471 | |
1472 | actual_stamp++; |
1473 | |
1474 | ret = get_inv_cost (inv, comp_cost: &comp_cost, regs_needed, cl: &cl); |
1475 | |
1476 | if (! flag_ira_loop_pressure) |
1477 | { |
1478 | size_cost = (estimate_reg_pressure_cost (new_regs[0] + regs_needed[0], |
1479 | regs_used, speed, call_p) |
1480 | - estimate_reg_pressure_cost (new_regs[0], |
1481 | regs_used, speed, call_p)); |
1482 | } |
1483 | else if (ret < 0) |
1484 | return -1; |
1485 | else if ((ret == 0) && (cl == NO_REGS)) |
1486 | /* Hoist it anyway since it does not impact register pressure. */ |
1487 | return 1; |
1488 | else |
1489 | { |
1490 | int i; |
1491 | enum reg_class pressure_class; |
1492 | |
1493 | for (i = 0; i < ira_pressure_classes_num; i++) |
1494 | { |
1495 | pressure_class = ira_pressure_classes[i]; |
1496 | |
1497 | if (!reg_classes_intersect_p (pressure_class, cl)) |
1498 | continue; |
1499 | |
1500 | if ((int) new_regs[pressure_class] |
1501 | + (int) regs_needed[pressure_class] |
1502 | + LOOP_DATA (curr_loop)->max_reg_pressure[pressure_class] |
1503 | + param_ira_loop_reserved_regs |
1504 | > ira_class_hard_regs_num[pressure_class]) |
1505 | break; |
1506 | } |
1507 | if (i < ira_pressure_classes_num) |
1508 | /* There will be register pressure excess and we want not to |
1509 | make this loop invariant motion. All loop invariants with |
1510 | non-positive gains will be rejected in function |
1511 | find_invariants_to_move. Therefore we return the negative |
1512 | number here. |
1513 | |
1514 | One could think that this rejects also expensive loop |
1515 | invariant motions and this will hurt code performance. |
1516 | However numerous experiments with different heuristics |
1517 | taking invariant cost into account did not confirm this |
1518 | assumption. There are possible explanations for this |
1519 | result: |
1520 | o probably all expensive invariants were already moved out |
1521 | of the loop by PRE and gimple invariant motion pass. |
1522 | o expensive invariant execution will be hidden by insn |
1523 | scheduling or OOO processor hardware because usually such |
1524 | invariants have a lot of freedom to be executed |
1525 | out-of-order. |
1526 | Another reason for ignoring invariant cost vs spilling cost |
1527 | heuristics is also in difficulties to evaluate accurately |
1528 | spill cost at this stage. */ |
1529 | return -1; |
1530 | else |
1531 | size_cost = 0; |
1532 | } |
1533 | |
1534 | return comp_cost - size_cost; |
1535 | } |
1536 | |
1537 | /* Finds invariant with best gain for moving. Returns the gain, stores |
1538 | the invariant in *BEST and number of registers needed for it to |
1539 | *REGS_NEEDED. REGS_USED is the number of registers used in the loop. |
1540 | NEW_REGS is the number of new variables already added due to invariant |
1541 | motion. */ |
1542 | |
1543 | static int |
1544 | best_gain_for_invariant (struct invariant **best, unsigned *regs_needed, |
1545 | unsigned *new_regs, unsigned regs_used, |
1546 | bool speed, bool call_p) |
1547 | { |
1548 | struct invariant *inv; |
1549 | int i, gain = 0, again; |
1550 | unsigned aregs_needed[N_REG_CLASSES], invno; |
1551 | |
1552 | FOR_EACH_VEC_ELT (invariants, invno, inv) |
1553 | { |
1554 | if (inv->move) |
1555 | continue; |
1556 | |
1557 | /* Only consider the "representatives" of equivalent invariants. */ |
1558 | if (inv->eqto != inv->invno) |
1559 | continue; |
1560 | |
1561 | again = gain_for_invariant (inv, regs_needed: aregs_needed, new_regs, regs_used, |
1562 | speed, call_p); |
1563 | if (again > gain) |
1564 | { |
1565 | gain = again; |
1566 | *best = inv; |
1567 | if (! flag_ira_loop_pressure) |
1568 | regs_needed[0] = aregs_needed[0]; |
1569 | else |
1570 | { |
1571 | for (i = 0; i < ira_pressure_classes_num; i++) |
1572 | regs_needed[ira_pressure_classes[i]] |
1573 | = aregs_needed[ira_pressure_classes[i]]; |
1574 | } |
1575 | } |
1576 | } |
1577 | |
1578 | return gain; |
1579 | } |
1580 | |
1581 | /* Marks invariant INVNO and all its dependencies for moving. */ |
1582 | |
1583 | static void |
1584 | set_move_mark (unsigned invno, int gain) |
1585 | { |
1586 | struct invariant *inv = invariants[invno]; |
1587 | bitmap_iterator bi; |
1588 | |
1589 | /* Find the representative of the class of the equivalent invariants. */ |
1590 | inv = invariants[inv->eqto]; |
1591 | |
1592 | if (inv->move) |
1593 | return; |
1594 | inv->move = true; |
1595 | |
1596 | if (dump_file) |
1597 | { |
1598 | if (gain >= 0) |
1599 | fprintf (stream: dump_file, format: "Decided to move invariant %d -- gain %d\n" , |
1600 | invno, gain); |
1601 | else |
1602 | fprintf (stream: dump_file, format: "Decided to move dependent invariant %d\n" , |
1603 | invno); |
1604 | }; |
1605 | |
1606 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, invno, bi) |
1607 | { |
1608 | set_move_mark (invno, gain: -1); |
1609 | } |
1610 | } |
1611 | |
1612 | /* Determines which invariants to move. */ |
1613 | |
1614 | static void |
1615 | find_invariants_to_move (bool speed, bool call_p) |
1616 | { |
1617 | int gain; |
1618 | unsigned i, regs_used, regs_needed[N_REG_CLASSES], new_regs[N_REG_CLASSES]; |
1619 | struct invariant *inv = NULL; |
1620 | |
1621 | if (!invariants.length ()) |
1622 | return; |
1623 | |
1624 | if (flag_ira_loop_pressure) |
1625 | /* REGS_USED is actually never used when the flag is on. */ |
1626 | regs_used = 0; |
1627 | else |
1628 | /* We do not really do a good job in estimating number of |
1629 | registers used; we put some initial bound here to stand for |
1630 | induction variables etc. that we do not detect. */ |
1631 | { |
1632 | unsigned int n_regs = DF_REG_SIZE (df); |
1633 | |
1634 | regs_used = 2; |
1635 | |
1636 | for (i = 0; i < n_regs; i++) |
1637 | { |
1638 | if (!DF_REGNO_FIRST_DEF (i) && DF_REGNO_LAST_USE (i)) |
1639 | { |
1640 | /* This is a value that is used but not changed inside loop. */ |
1641 | regs_used++; |
1642 | } |
1643 | } |
1644 | } |
1645 | |
1646 | if (! flag_ira_loop_pressure) |
1647 | new_regs[0] = regs_needed[0] = 0; |
1648 | else |
1649 | { |
1650 | for (i = 0; (int) i < ira_pressure_classes_num; i++) |
1651 | new_regs[ira_pressure_classes[i]] = 0; |
1652 | } |
1653 | while ((gain = best_gain_for_invariant (best: &inv, regs_needed, |
1654 | new_regs, regs_used, |
1655 | speed, call_p)) > 0) |
1656 | { |
1657 | set_move_mark (invno: inv->invno, gain); |
1658 | if (! flag_ira_loop_pressure) |
1659 | new_regs[0] += regs_needed[0]; |
1660 | else |
1661 | { |
1662 | for (i = 0; (int) i < ira_pressure_classes_num; i++) |
1663 | new_regs[ira_pressure_classes[i]] |
1664 | += regs_needed[ira_pressure_classes[i]]; |
1665 | } |
1666 | } |
1667 | } |
1668 | |
1669 | /* Replace the uses, reached by the definition of invariant INV, by REG. |
1670 | |
1671 | IN_GROUP is nonzero if this is part of a group of changes that must be |
1672 | performed as a group. In that case, the changes will be stored. The |
1673 | function `apply_change_group' will validate and apply the changes. */ |
1674 | |
1675 | static int |
1676 | replace_uses (struct invariant *inv, rtx reg, bool in_group) |
1677 | { |
1678 | /* Replace the uses we know to be dominated. It saves work for copy |
1679 | propagation, and also it is necessary so that dependent invariants |
1680 | are computed right. */ |
1681 | if (inv->def) |
1682 | { |
1683 | struct use *use; |
1684 | for (use = inv->def->uses; use; use = use->next) |
1685 | validate_change (use->insn, use->pos, reg, true); |
1686 | |
1687 | /* If we aren't part of a larger group, apply the changes now. */ |
1688 | if (!in_group) |
1689 | return apply_change_group (); |
1690 | } |
1691 | |
1692 | return 1; |
1693 | } |
1694 | |
1695 | /* Whether invariant INV setting REG can be moved out of LOOP, at the end of |
1696 | the block preceding its header. */ |
1697 | |
1698 | static bool |
1699 | can_move_invariant_reg (class loop *loop, struct invariant *inv, rtx reg) |
1700 | { |
1701 | df_ref def, use; |
1702 | unsigned int dest_regno, defs_in_loop_count = 0; |
1703 | rtx_insn *insn = inv->insn; |
1704 | basic_block bb = BLOCK_FOR_INSN (insn: inv->insn); |
1705 | auto_vec <rtx_insn *, 16> debug_insns_to_reset; |
1706 | |
1707 | /* We ignore hard register and memory access for cost and complexity reasons. |
1708 | Hard register are few at this stage and expensive to consider as they |
1709 | require building a separate data flow. Memory access would require using |
1710 | df_simulate_* and can_move_insns_across functions and is more complex. */ |
1711 | if (!REG_P (reg) || HARD_REGISTER_P (reg)) |
1712 | return false; |
1713 | |
1714 | /* Check whether the set is always executed. We could omit this condition if |
1715 | we know that the register is unused outside of the loop, but it does not |
1716 | seem worth finding out. */ |
1717 | if (!inv->always_executed) |
1718 | return false; |
1719 | |
1720 | /* Check that all uses that would be dominated by def are already dominated |
1721 | by it. */ |
1722 | dest_regno = REGNO (reg); |
1723 | for (use = DF_REG_USE_CHAIN (dest_regno); use; use = DF_REF_NEXT_REG (use)) |
1724 | { |
1725 | rtx_insn *use_insn; |
1726 | basic_block use_bb; |
1727 | |
1728 | use_insn = DF_REF_INSN (use); |
1729 | use_bb = BLOCK_FOR_INSN (insn: use_insn); |
1730 | |
1731 | /* Ignore instruction considered for moving. */ |
1732 | if (use_insn == insn) |
1733 | continue; |
1734 | |
1735 | /* Don't consider uses outside loop. */ |
1736 | if (!flow_bb_inside_loop_p (loop, use_bb)) |
1737 | continue; |
1738 | |
1739 | /* Don't move if a use is not dominated by def in insn. */ |
1740 | if ((use_bb == bb && DF_INSN_LUID (insn) >= DF_INSN_LUID (use_insn)) |
1741 | || !dominated_by_p (CDI_DOMINATORS, use_bb, bb)) |
1742 | { |
1743 | if (!DEBUG_INSN_P (use_insn)) |
1744 | return false; |
1745 | debug_insns_to_reset.safe_push (obj: use_insn); |
1746 | } |
1747 | } |
1748 | |
1749 | /* Check for other defs. Any other def in the loop might reach a use |
1750 | currently reached by the def in insn. */ |
1751 | for (def = DF_REG_DEF_CHAIN (dest_regno); def; def = DF_REF_NEXT_REG (def)) |
1752 | { |
1753 | basic_block def_bb = DF_REF_BB (def); |
1754 | |
1755 | /* Defs in exit block cannot reach a use they weren't already. */ |
1756 | if (single_succ_p (bb: def_bb)) |
1757 | { |
1758 | basic_block def_bb_succ; |
1759 | |
1760 | def_bb_succ = single_succ (bb: def_bb); |
1761 | if (!flow_bb_inside_loop_p (loop, def_bb_succ)) |
1762 | continue; |
1763 | } |
1764 | |
1765 | if (++defs_in_loop_count > 1) |
1766 | return false; |
1767 | } |
1768 | |
1769 | /* Reset debug uses if a use is not dominated by def in insn. */ |
1770 | for (auto use_insn : debug_insns_to_reset) |
1771 | { |
1772 | INSN_VAR_LOCATION_LOC (use_insn) = gen_rtx_UNKNOWN_VAR_LOC (); |
1773 | df_insn_rescan (use_insn); |
1774 | } |
1775 | |
1776 | return true; |
1777 | } |
1778 | |
1779 | /* Move invariant INVNO out of the LOOP. Returns true if this succeeds, false |
1780 | otherwise. */ |
1781 | |
1782 | static bool |
1783 | move_invariant_reg (class loop *loop, unsigned invno) |
1784 | { |
1785 | struct invariant *inv = invariants[invno]; |
1786 | struct invariant *repr = invariants[inv->eqto]; |
1787 | unsigned i; |
1788 | basic_block = loop_preheader_edge (loop)->src; |
1789 | rtx reg, set, dest, note; |
1790 | bitmap_iterator bi; |
1791 | int regno = -1; |
1792 | |
1793 | if (inv->reg) |
1794 | return true; |
1795 | if (!repr->move) |
1796 | return false; |
1797 | |
1798 | /* If this is a representative of the class of equivalent invariants, |
1799 | really move the invariant. Otherwise just replace its use with |
1800 | the register used for the representative. */ |
1801 | if (inv == repr) |
1802 | { |
1803 | if (inv->depends_on) |
1804 | { |
1805 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, i, bi) |
1806 | { |
1807 | if (!move_invariant_reg (loop, invno: i)) |
1808 | goto fail; |
1809 | } |
1810 | } |
1811 | |
1812 | /* If possible, just move the set out of the loop. Otherwise, we |
1813 | need to create a temporary register. */ |
1814 | set = single_set (insn: inv->insn); |
1815 | reg = dest = SET_DEST (set); |
1816 | if (GET_CODE (reg) == SUBREG) |
1817 | reg = SUBREG_REG (reg); |
1818 | if (REG_P (reg)) |
1819 | regno = REGNO (reg); |
1820 | |
1821 | if (!can_move_invariant_reg (loop, inv, reg: dest)) |
1822 | { |
1823 | reg = gen_reg_rtx_and_attrs (dest); |
1824 | |
1825 | /* Try replacing the destination by a new pseudoregister. */ |
1826 | validate_change (inv->insn, &SET_DEST (set), reg, true); |
1827 | |
1828 | /* As well as all the dominated uses. */ |
1829 | replace_uses (inv, reg, in_group: true); |
1830 | |
1831 | /* And validate all the changes. */ |
1832 | if (!apply_change_group ()) |
1833 | goto fail; |
1834 | |
1835 | emit_insn_after (gen_move_insn (dest, reg), inv->insn); |
1836 | } |
1837 | else if (dump_file) |
1838 | fprintf (stream: dump_file, format: "Invariant %d moved without introducing a new " |
1839 | "temporary register\n" , invno); |
1840 | if (JUMP_P (BB_END (preheader))) |
1841 | preheader = split_edge (loop_preheader_edge (loop)); |
1842 | reorder_insns (inv->insn, inv->insn, BB_END (preheader)); |
1843 | df_recompute_luids (preheader); |
1844 | |
1845 | /* If there is a REG_EQUAL note on the insn we just moved, and the |
1846 | insn is in a basic block that is not always executed or the note |
1847 | contains something for which we don't know the invariant status, |
1848 | the note may no longer be valid after we move the insn. Note that |
1849 | uses in REG_EQUAL notes are taken into account in the computation |
1850 | of invariants, so it is safe to retain the note even if it contains |
1851 | register references for which we know the invariant status. */ |
1852 | if ((note = find_reg_note (inv->insn, REG_EQUAL, NULL_RTX)) |
1853 | && (!inv->always_executed |
1854 | || !check_maybe_invariant (XEXP (note, 0)))) |
1855 | remove_note (inv->insn, note); |
1856 | } |
1857 | else |
1858 | { |
1859 | if (!move_invariant_reg (loop, invno: repr->invno)) |
1860 | goto fail; |
1861 | reg = repr->reg; |
1862 | regno = repr->orig_regno; |
1863 | if (!replace_uses (inv, reg, in_group: false)) |
1864 | goto fail; |
1865 | set = single_set (insn: inv->insn); |
1866 | emit_insn_after (gen_move_insn (SET_DEST (set), reg), inv->insn); |
1867 | delete_insn (inv->insn); |
1868 | } |
1869 | |
1870 | inv->reg = reg; |
1871 | inv->orig_regno = regno; |
1872 | |
1873 | return true; |
1874 | |
1875 | fail: |
1876 | /* If we failed, clear move flag, so that we do not try to move inv |
1877 | again. */ |
1878 | if (dump_file) |
1879 | fprintf (stream: dump_file, format: "Failed to move invariant %d\n" , invno); |
1880 | inv->move = false; |
1881 | inv->reg = NULL_RTX; |
1882 | inv->orig_regno = -1; |
1883 | |
1884 | return false; |
1885 | } |
1886 | |
1887 | /* Move selected invariant out of the LOOP. Newly created regs are marked |
1888 | in TEMPORARY_REGS. */ |
1889 | |
1890 | static void |
1891 | move_invariants (class loop *loop) |
1892 | { |
1893 | struct invariant *inv; |
1894 | unsigned i; |
1895 | |
1896 | FOR_EACH_VEC_ELT (invariants, i, inv) |
1897 | move_invariant_reg (loop, invno: i); |
1898 | if (flag_ira_loop_pressure && resize_reg_info ()) |
1899 | { |
1900 | FOR_EACH_VEC_ELT (invariants, i, inv) |
1901 | if (inv->reg != NULL_RTX) |
1902 | { |
1903 | if (inv->orig_regno >= 0) |
1904 | setup_reg_classes (REGNO (inv->reg), |
1905 | reg_preferred_class (inv->orig_regno), |
1906 | reg_alternate_class (inv->orig_regno), |
1907 | reg_allocno_class (inv->orig_regno)); |
1908 | else |
1909 | setup_reg_classes (REGNO (inv->reg), |
1910 | GENERAL_REGS, NO_REGS, GENERAL_REGS); |
1911 | } |
1912 | } |
1913 | /* Remove the DF_UD_CHAIN problem added in find_defs before rescanning, |
1914 | to save a bit of compile time. */ |
1915 | df_remove_problem (df_chain); |
1916 | df_process_deferred_rescans (); |
1917 | } |
1918 | |
1919 | /* Initializes invariant motion data. */ |
1920 | |
1921 | static void |
1922 | init_inv_motion_data (void) |
1923 | { |
1924 | actual_stamp = 1; |
1925 | |
1926 | invariants.create (nelems: 100); |
1927 | } |
1928 | |
1929 | /* Frees the data allocated by invariant motion. */ |
1930 | |
1931 | static void |
1932 | free_inv_motion_data (void) |
1933 | { |
1934 | unsigned i; |
1935 | struct def *def; |
1936 | struct invariant *inv; |
1937 | |
1938 | check_invariant_table_size (); |
1939 | for (i = 0; i < DF_DEFS_TABLE_SIZE (); i++) |
1940 | { |
1941 | inv = invariant_table[i]; |
1942 | if (inv) |
1943 | { |
1944 | def = inv->def; |
1945 | gcc_assert (def != NULL); |
1946 | |
1947 | free_use_list (use: def->uses); |
1948 | free (ptr: def); |
1949 | invariant_table[i] = NULL; |
1950 | } |
1951 | } |
1952 | |
1953 | FOR_EACH_VEC_ELT (invariants, i, inv) |
1954 | { |
1955 | BITMAP_FREE (inv->depends_on); |
1956 | free (ptr: inv); |
1957 | } |
1958 | invariants.release (); |
1959 | } |
1960 | |
1961 | /* Move the invariants out of the LOOP. */ |
1962 | |
1963 | static void |
1964 | move_single_loop_invariants (class loop *loop) |
1965 | { |
1966 | init_inv_motion_data (); |
1967 | |
1968 | find_invariants (loop); |
1969 | find_invariants_to_move (speed: optimize_loop_for_speed_p (loop), |
1970 | LOOP_DATA (loop)->has_call); |
1971 | move_invariants (loop); |
1972 | |
1973 | free_inv_motion_data (); |
1974 | } |
1975 | |
1976 | /* Releases the auxiliary data for LOOP. */ |
1977 | |
1978 | static void |
1979 | free_loop_data (class loop *loop) |
1980 | { |
1981 | class loop_data *data = LOOP_DATA (loop); |
1982 | if (!data) |
1983 | return; |
1984 | |
1985 | bitmap_clear (&LOOP_DATA (loop)->regs_ref); |
1986 | bitmap_clear (&LOOP_DATA (loop)->regs_live); |
1987 | free (ptr: data); |
1988 | loop->aux = NULL; |
1989 | } |
1990 | |
1991 | |
1992 | |
1993 | /* Registers currently living. */ |
1994 | static bitmap_head curr_regs_live; |
1995 | |
1996 | /* Current reg pressure for each pressure class. */ |
1997 | static int curr_reg_pressure[N_REG_CLASSES]; |
1998 | |
1999 | /* Record all regs that are set in any one insn. Communication from |
2000 | mark_reg_{store,clobber} and global_conflicts. Asm can refer to |
2001 | all hard-registers. */ |
2002 | static rtx regs_set[(FIRST_PSEUDO_REGISTER > MAX_RECOG_OPERANDS |
2003 | ? FIRST_PSEUDO_REGISTER : MAX_RECOG_OPERANDS) * 2]; |
2004 | /* Number of regs stored in the previous array. */ |
2005 | static int n_regs_set; |
2006 | |
2007 | /* Return pressure class and number of needed hard registers (through |
2008 | *NREGS) of register REGNO. */ |
2009 | static enum reg_class |
2010 | get_regno_pressure_class (int regno, int *nregs) |
2011 | { |
2012 | if (regno >= FIRST_PSEUDO_REGISTER) |
2013 | { |
2014 | enum reg_class pressure_class; |
2015 | |
2016 | pressure_class = reg_allocno_class (regno); |
2017 | pressure_class = ira_pressure_class_translate[pressure_class]; |
2018 | *nregs |
2019 | = ira_reg_class_max_nregs[pressure_class][PSEUDO_REGNO_MODE (regno)]; |
2020 | return pressure_class; |
2021 | } |
2022 | else if (! TEST_HARD_REG_BIT (ira_no_alloc_regs, bit: regno) |
2023 | && ! TEST_HARD_REG_BIT (set: eliminable_regset, bit: regno)) |
2024 | { |
2025 | *nregs = 1; |
2026 | return ira_pressure_class_translate[REGNO_REG_CLASS (regno)]; |
2027 | } |
2028 | else |
2029 | { |
2030 | *nregs = 0; |
2031 | return NO_REGS; |
2032 | } |
2033 | } |
2034 | |
2035 | /* Increase (if INCR_P) or decrease current register pressure for |
2036 | register REGNO. */ |
2037 | static void |
2038 | change_pressure (int regno, bool incr_p) |
2039 | { |
2040 | int nregs; |
2041 | enum reg_class pressure_class; |
2042 | |
2043 | pressure_class = get_regno_pressure_class (regno, nregs: &nregs); |
2044 | if (! incr_p) |
2045 | curr_reg_pressure[pressure_class] -= nregs; |
2046 | else |
2047 | { |
2048 | curr_reg_pressure[pressure_class] += nregs; |
2049 | if (LOOP_DATA (curr_loop)->max_reg_pressure[pressure_class] |
2050 | < curr_reg_pressure[pressure_class]) |
2051 | LOOP_DATA (curr_loop)->max_reg_pressure[pressure_class] |
2052 | = curr_reg_pressure[pressure_class]; |
2053 | } |
2054 | } |
2055 | |
2056 | /* Mark REGNO birth. */ |
2057 | static void |
2058 | mark_regno_live (int regno) |
2059 | { |
2060 | class loop *loop; |
2061 | |
2062 | for (loop = curr_loop; |
2063 | loop != current_loops->tree_root; |
2064 | loop = loop_outer (loop)) |
2065 | bitmap_set_bit (&LOOP_DATA (loop)->regs_live, regno); |
2066 | if (!bitmap_set_bit (&curr_regs_live, regno)) |
2067 | return; |
2068 | change_pressure (regno, incr_p: true); |
2069 | } |
2070 | |
2071 | /* Mark REGNO death. */ |
2072 | static void |
2073 | mark_regno_death (int regno) |
2074 | { |
2075 | if (! bitmap_clear_bit (&curr_regs_live, regno)) |
2076 | return; |
2077 | change_pressure (regno, incr_p: false); |
2078 | } |
2079 | |
2080 | /* Mark setting register REG. */ |
2081 | static void |
2082 | mark_reg_store (rtx reg, const_rtx setter ATTRIBUTE_UNUSED, |
2083 | void *data ATTRIBUTE_UNUSED) |
2084 | { |
2085 | if (GET_CODE (reg) == SUBREG) |
2086 | reg = SUBREG_REG (reg); |
2087 | |
2088 | if (! REG_P (reg)) |
2089 | return; |
2090 | |
2091 | regs_set[n_regs_set++] = reg; |
2092 | |
2093 | unsigned int end_regno = END_REGNO (x: reg); |
2094 | for (unsigned int regno = REGNO (reg); regno < end_regno; ++regno) |
2095 | mark_regno_live (regno); |
2096 | } |
2097 | |
2098 | /* Mark clobbering register REG. */ |
2099 | static void |
2100 | mark_reg_clobber (rtx reg, const_rtx setter, void *data) |
2101 | { |
2102 | if (GET_CODE (setter) == CLOBBER) |
2103 | mark_reg_store (reg, setter, data); |
2104 | } |
2105 | |
2106 | /* Mark register REG death. */ |
2107 | static void |
2108 | mark_reg_death (rtx reg) |
2109 | { |
2110 | unsigned int end_regno = END_REGNO (x: reg); |
2111 | for (unsigned int regno = REGNO (reg); regno < end_regno; ++regno) |
2112 | mark_regno_death (regno); |
2113 | } |
2114 | |
2115 | /* Mark occurrence of registers in X for the current loop. */ |
2116 | static void |
2117 | mark_ref_regs (rtx x) |
2118 | { |
2119 | RTX_CODE code; |
2120 | int i; |
2121 | const char *fmt; |
2122 | |
2123 | if (!x) |
2124 | return; |
2125 | |
2126 | code = GET_CODE (x); |
2127 | if (code == REG) |
2128 | { |
2129 | class loop *loop; |
2130 | |
2131 | for (loop = curr_loop; |
2132 | loop != current_loops->tree_root; |
2133 | loop = loop_outer (loop)) |
2134 | bitmap_set_bit (&LOOP_DATA (loop)->regs_ref, REGNO (x)); |
2135 | return; |
2136 | } |
2137 | |
2138 | fmt = GET_RTX_FORMAT (code); |
2139 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
2140 | if (fmt[i] == 'e') |
2141 | mark_ref_regs (XEXP (x, i)); |
2142 | else if (fmt[i] == 'E') |
2143 | { |
2144 | int j; |
2145 | |
2146 | for (j = 0; j < XVECLEN (x, i); j++) |
2147 | mark_ref_regs (XVECEXP (x, i, j)); |
2148 | } |
2149 | } |
2150 | |
2151 | /* Calculate register pressure in the loops. */ |
2152 | static void |
2153 | calculate_loop_reg_pressure (void) |
2154 | { |
2155 | int i; |
2156 | unsigned int j; |
2157 | bitmap_iterator bi; |
2158 | basic_block bb; |
2159 | rtx_insn *insn; |
2160 | rtx link; |
2161 | class loop *parent; |
2162 | |
2163 | for (auto loop : loops_list (cfun, 0)) |
2164 | if (loop->aux == NULL) |
2165 | { |
2166 | loop->aux = xcalloc (1, sizeof (class loop_data)); |
2167 | bitmap_initialize (head: &LOOP_DATA (loop)->regs_ref, obstack: ®_obstack); |
2168 | bitmap_initialize (head: &LOOP_DATA (loop)->regs_live, obstack: ®_obstack); |
2169 | } |
2170 | ira_setup_eliminable_regset (); |
2171 | bitmap_initialize (head: &curr_regs_live, obstack: ®_obstack); |
2172 | FOR_EACH_BB_FN (bb, cfun) |
2173 | { |
2174 | curr_loop = bb->loop_father; |
2175 | if (curr_loop == current_loops->tree_root) |
2176 | continue; |
2177 | |
2178 | for (class loop *loop = curr_loop; |
2179 | loop != current_loops->tree_root; |
2180 | loop = loop_outer (loop)) |
2181 | bitmap_ior_into (&LOOP_DATA (loop)->regs_live, DF_LR_IN (bb)); |
2182 | |
2183 | bitmap_copy (&curr_regs_live, DF_LR_IN (bb)); |
2184 | for (i = 0; i < ira_pressure_classes_num; i++) |
2185 | curr_reg_pressure[ira_pressure_classes[i]] = 0; |
2186 | EXECUTE_IF_SET_IN_BITMAP (&curr_regs_live, 0, j, bi) |
2187 | change_pressure (regno: j, incr_p: true); |
2188 | |
2189 | FOR_BB_INSNS (bb, insn) |
2190 | { |
2191 | if (! NONDEBUG_INSN_P (insn)) |
2192 | continue; |
2193 | |
2194 | mark_ref_regs (x: PATTERN (insn)); |
2195 | n_regs_set = 0; |
2196 | note_stores (insn, mark_reg_clobber, NULL); |
2197 | |
2198 | /* Mark any registers dead after INSN as dead now. */ |
2199 | |
2200 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) |
2201 | if (REG_NOTE_KIND (link) == REG_DEAD) |
2202 | mark_reg_death (XEXP (link, 0)); |
2203 | |
2204 | /* Mark any registers set in INSN as live, |
2205 | and mark them as conflicting with all other live regs. |
2206 | Clobbers are processed again, so they conflict with |
2207 | the registers that are set. */ |
2208 | |
2209 | note_stores (insn, mark_reg_store, NULL); |
2210 | |
2211 | if (AUTO_INC_DEC) |
2212 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) |
2213 | if (REG_NOTE_KIND (link) == REG_INC) |
2214 | mark_reg_store (XEXP (link, 0), NULL_RTX, NULL); |
2215 | |
2216 | while (n_regs_set-- > 0) |
2217 | { |
2218 | rtx note = find_regno_note (insn, REG_UNUSED, |
2219 | REGNO (regs_set[n_regs_set])); |
2220 | if (! note) |
2221 | continue; |
2222 | |
2223 | mark_reg_death (XEXP (note, 0)); |
2224 | } |
2225 | } |
2226 | } |
2227 | bitmap_release (head: &curr_regs_live); |
2228 | if (flag_ira_region == IRA_REGION_MIXED |
2229 | || flag_ira_region == IRA_REGION_ALL) |
2230 | for (auto loop : loops_list (cfun, 0)) |
2231 | { |
2232 | EXECUTE_IF_SET_IN_BITMAP (&LOOP_DATA (loop)->regs_live, 0, j, bi) |
2233 | if (! bitmap_bit_p (&LOOP_DATA (loop)->regs_ref, j)) |
2234 | { |
2235 | enum reg_class pressure_class; |
2236 | int nregs; |
2237 | |
2238 | pressure_class = get_regno_pressure_class (regno: j, nregs: &nregs); |
2239 | LOOP_DATA (loop)->max_reg_pressure[pressure_class] -= nregs; |
2240 | } |
2241 | } |
2242 | if (dump_file == NULL) |
2243 | return; |
2244 | for (auto loop : loops_list (cfun, 0)) |
2245 | { |
2246 | parent = loop_outer (loop); |
2247 | fprintf (stream: dump_file, format: "\n Loop %d (parent %d, header bb%d, depth %d)\n" , |
2248 | loop->num, (parent == NULL ? -1 : parent->num), |
2249 | loop->header->index, loop_depth (loop)); |
2250 | fprintf (stream: dump_file, format: "\n ref. regnos:" ); |
2251 | EXECUTE_IF_SET_IN_BITMAP (&LOOP_DATA (loop)->regs_ref, 0, j, bi) |
2252 | fprintf (stream: dump_file, format: " %d" , j); |
2253 | fprintf (stream: dump_file, format: "\n live regnos:" ); |
2254 | EXECUTE_IF_SET_IN_BITMAP (&LOOP_DATA (loop)->regs_live, 0, j, bi) |
2255 | fprintf (stream: dump_file, format: " %d" , j); |
2256 | fprintf (stream: dump_file, format: "\n Pressure:" ); |
2257 | for (i = 0; (int) i < ira_pressure_classes_num; i++) |
2258 | { |
2259 | enum reg_class pressure_class; |
2260 | |
2261 | pressure_class = ira_pressure_classes[i]; |
2262 | if (LOOP_DATA (loop)->max_reg_pressure[pressure_class] == 0) |
2263 | continue; |
2264 | fprintf (stream: dump_file, format: " %s=%d" , reg_class_names[pressure_class], |
2265 | LOOP_DATA (loop)->max_reg_pressure[pressure_class]); |
2266 | } |
2267 | fprintf (stream: dump_file, format: "\n" ); |
2268 | } |
2269 | } |
2270 | |
2271 | |
2272 | |
2273 | /* Move the invariants out of the loops. */ |
2274 | |
2275 | void |
2276 | move_loop_invariants (void) |
2277 | { |
2278 | if (optimize == 1) |
2279 | df_live_add_problem (); |
2280 | /* ??? This is a hack. We should only need to call df_live_set_all_dirty |
2281 | for optimize == 1, but can_move_invariant_reg relies on DF_INSN_LUID |
2282 | being up-to-date. That isn't always true (even after df_analyze) |
2283 | because df_process_deferred_rescans doesn't necessarily cause |
2284 | blocks to be rescanned. */ |
2285 | df_live_set_all_dirty (); |
2286 | if (flag_ira_loop_pressure) |
2287 | { |
2288 | df_analyze (); |
2289 | regstat_init_n_sets_and_refs (); |
2290 | ira_set_pseudo_classes (true, dump_file); |
2291 | calculate_loop_reg_pressure (); |
2292 | regstat_free_n_sets_and_refs (); |
2293 | } |
2294 | df_set_flags (DF_EQ_NOTES + DF_DEFER_INSN_RESCAN); |
2295 | /* Process the loops, innermost first. */ |
2296 | for (auto loop : loops_list (cfun, LI_FROM_INNERMOST)) |
2297 | { |
2298 | curr_loop = loop; |
2299 | /* move_single_loop_invariants for very large loops is time consuming |
2300 | and might need a lot of memory. For -O1 only do loop invariant |
2301 | motion for very small loops. */ |
2302 | unsigned max_bbs = param_loop_invariant_max_bbs_in_loop; |
2303 | if (optimize < 2) |
2304 | max_bbs /= 10; |
2305 | if (loop->num_nodes <= max_bbs) |
2306 | move_single_loop_invariants (loop); |
2307 | } |
2308 | |
2309 | for (auto loop : loops_list (cfun, 0)) |
2310 | free_loop_data (loop); |
2311 | |
2312 | if (flag_ira_loop_pressure) |
2313 | /* There is no sense to keep this info because it was most |
2314 | probably outdated by subsequent passes. */ |
2315 | free_reg_info (); |
2316 | free (ptr: invariant_table); |
2317 | invariant_table = NULL; |
2318 | invariant_table_size = 0; |
2319 | |
2320 | if (optimize == 1) |
2321 | df_remove_problem (df_live); |
2322 | |
2323 | checking_verify_flow_info (); |
2324 | } |
2325 | |