]>
Commit | Line | Data |
---|---|---|
5e962776 | 1 | /* Rtl-level loop invariant motion. |
c8d3e15a | 2 | Copyright (C) 2004, 2005 Free Software Foundation, Inc. |
5e962776 ZD |
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 2, 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 COPYING. If not, write to the Free | |
18 | Software Foundation, 59 Temple Place - Suite 330, Boston, MA | |
19 | 02111-1307, USA. */ | |
20 | ||
21 | /* This implements the loop invariant motion pass. It is very simple | |
22 | (no calls, libcalls, etc.). This should be sufficient to cleanup things like | |
23 | address arithmetics -- other more complicated invariants should be | |
24 | eliminated on tree level either in tree-ssa-loop-im.c or in tree-ssa-pre.c. | |
25 | ||
26 | We proceed loop by loop -- it is simpler than trying to handle things | |
27 | globally and should not lose much. First we inspect all sets inside loop | |
28 | and create a dependency graph on insns (saying "to move this insn, you must | |
29 | also move the following insns"). | |
30 | ||
31 | We then need to determine what to move. We estimate the number of registers | |
32 | used and move as many invariants as possible while we still have enough free | |
33 | registers. We prefer the expensive invariants. | |
34 | ||
35 | Then we move the selected invariants out of the loop, creating a new | |
36 | temporaries for them if necessary. */ | |
37 | ||
38 | #include "config.h" | |
39 | #include "system.h" | |
40 | #include "coretypes.h" | |
41 | #include "tm.h" | |
42 | #include "rtl.h" | |
43 | #include "hard-reg-set.h" | |
7932a3db | 44 | #include "obstack.h" |
5e962776 ZD |
45 | #include "basic-block.h" |
46 | #include "cfgloop.h" | |
47 | #include "expr.h" | |
48 | #include "output.h" | |
49 | #include "function.h" | |
50 | #include "flags.h" | |
51 | #include "df.h" | |
52 | ||
53 | /* The data stored for the loop. */ | |
54 | ||
55 | struct loop_data | |
56 | { | |
57 | struct loop *outermost_exit; /* The outermost exit of the loop. */ | |
58 | bool has_call; /* True if the loop contains a call. */ | |
59 | }; | |
60 | ||
61 | #define LOOP_DATA(LOOP) ((struct loop_data *) (LOOP)->aux) | |
62 | ||
63 | /* The description of an use. */ | |
64 | ||
65 | struct use | |
66 | { | |
67 | rtx *pos; /* Position of the use. */ | |
68 | rtx insn; /* The insn in that the use occurs. */ | |
69 | ||
70 | struct use *next; /* Next use in the list. */ | |
71 | }; | |
72 | ||
73 | /* The description of a def. */ | |
74 | ||
75 | struct def | |
76 | { | |
77 | struct use *uses; /* The list of uses that are uniquely reached | |
78 | by it. */ | |
79 | unsigned n_uses; /* Number of such uses. */ | |
80 | unsigned invno; /* The corresponding invariant. */ | |
81 | }; | |
82 | ||
83 | /* The data stored for each invariant. */ | |
84 | ||
85 | struct invariant | |
86 | { | |
87 | /* The number of the invariant. */ | |
88 | unsigned invno; | |
89 | ||
90 | /* Whether we already processed the invariant. */ | |
91 | bool processed; | |
92 | ||
93 | /* The definition of the invariant. */ | |
94 | struct def *def; | |
95 | ||
96 | /* The insn in that it is defined. */ | |
97 | rtx insn; | |
98 | ||
99 | /* Whether it is always executed. */ | |
100 | bool always_executed; | |
101 | ||
102 | /* Whether to move the invariant. */ | |
103 | bool move; | |
104 | ||
105 | /* Cost if the invariant. */ | |
106 | unsigned cost; | |
107 | ||
108 | /* The invariants it depends on. */ | |
109 | bitmap depends_on; | |
110 | ||
111 | /* Used for detecting already visited invariants during determining | |
112 | costs of movements. */ | |
113 | unsigned stamp; | |
114 | }; | |
115 | ||
116 | /* The actual stamp for marking already visited invariants during determining | |
117 | costs of movements. */ | |
118 | ||
119 | static unsigned actual_stamp; | |
120 | ||
121 | /* The invariants. */ | |
122 | ||
123 | static varray_type invariants; | |
124 | ||
125 | /* Test for possibility of invariantness of X. */ | |
126 | ||
127 | static bool | |
128 | check_maybe_invariant (rtx x) | |
129 | { | |
130 | enum rtx_code code = GET_CODE (x); | |
131 | int i, j; | |
132 | const char *fmt; | |
133 | ||
134 | switch (code) | |
135 | { | |
136 | case CONST_INT: | |
137 | case CONST_DOUBLE: | |
138 | case SYMBOL_REF: | |
139 | case CONST: | |
140 | case LABEL_REF: | |
141 | return true; | |
142 | ||
143 | case PC: | |
144 | case CC0: | |
145 | case UNSPEC_VOLATILE: | |
146 | case CALL: | |
147 | return false; | |
148 | ||
149 | case REG: | |
150 | return true; | |
151 | ||
152 | case MEM: | |
153 | /* Load/store motion is done elsewhere. ??? Perhaps also add it here? | |
154 | It should not be hard, and might be faster than "elsewhere". */ | |
155 | ||
156 | /* Just handle the most trivial case where we load from an unchanging | |
157 | location (most importantly, pic tables). */ | |
389fdba0 | 158 | if (MEM_READONLY_P (x)) |
5e962776 ZD |
159 | break; |
160 | ||
161 | return false; | |
162 | ||
163 | case ASM_OPERANDS: | |
164 | /* Don't mess with insns declared volatile. */ | |
165 | if (MEM_VOLATILE_P (x)) | |
166 | return false; | |
167 | break; | |
168 | ||
169 | default: | |
170 | break; | |
171 | } | |
172 | ||
173 | fmt = GET_RTX_FORMAT (code); | |
174 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
175 | { | |
176 | if (fmt[i] == 'e') | |
177 | { | |
178 | if (!check_maybe_invariant (XEXP (x, i))) | |
179 | return false; | |
180 | } | |
181 | else if (fmt[i] == 'E') | |
182 | { | |
183 | for (j = 0; j < XVECLEN (x, i); j++) | |
184 | if (!check_maybe_invariant (XVECEXP (x, i, j))) | |
185 | return false; | |
186 | } | |
187 | } | |
188 | ||
189 | return true; | |
190 | } | |
191 | ||
192 | /* Determines the basic blocks inside LOOP that are always executed and | |
193 | stores their bitmap to ALWAYS_REACHED. MAY_EXIT is a bitmap of | |
194 | basic blocks that may either exit the loop, or contain the call that | |
195 | does not have to return. BODY is body of the loop obtained by | |
196 | get_loop_body_in_dom_order. */ | |
197 | ||
198 | static void | |
199 | compute_always_reached (struct loop *loop, basic_block *body, | |
200 | bitmap may_exit, bitmap always_reached) | |
201 | { | |
202 | unsigned i; | |
203 | ||
204 | for (i = 0; i < loop->num_nodes; i++) | |
205 | { | |
206 | if (dominated_by_p (CDI_DOMINATORS, loop->latch, body[i])) | |
207 | bitmap_set_bit (always_reached, i); | |
208 | ||
209 | if (bitmap_bit_p (may_exit, i)) | |
210 | return; | |
211 | } | |
212 | } | |
213 | ||
214 | /* Finds exits out of the LOOP with body BODY. Marks blocks in that we may | |
215 | exit the loop by cfg edge to HAS_EXIT and MAY_EXIT. In MAY_EXIT | |
216 | additionally mark blocks that may exit due to a call. */ | |
217 | ||
218 | static void | |
219 | find_exits (struct loop *loop, basic_block *body, | |
220 | bitmap may_exit, bitmap has_exit) | |
221 | { | |
222 | unsigned i; | |
628f6a4e | 223 | edge_iterator ei; |
5e962776 ZD |
224 | edge e; |
225 | struct loop *outermost_exit = loop, *aexit; | |
226 | bool has_call = false; | |
227 | rtx insn; | |
228 | ||
229 | for (i = 0; i < loop->num_nodes; i++) | |
230 | { | |
231 | if (body[i]->loop_father == loop) | |
232 | { | |
233 | FOR_BB_INSNS (body[i], insn) | |
234 | { | |
4b4bf941 | 235 | if (CALL_P (insn) |
5e962776 ZD |
236 | && !CONST_OR_PURE_CALL_P (insn)) |
237 | { | |
238 | has_call = true; | |
239 | bitmap_set_bit (may_exit, i); | |
240 | break; | |
241 | } | |
242 | } | |
243 | ||
628f6a4e | 244 | FOR_EACH_EDGE (e, ei, body[i]->succs) |
5e962776 ZD |
245 | { |
246 | if (flow_bb_inside_loop_p (loop, e->dest)) | |
247 | continue; | |
248 | ||
249 | bitmap_set_bit (may_exit, i); | |
250 | bitmap_set_bit (has_exit, i); | |
251 | outermost_exit = find_common_loop (outermost_exit, | |
252 | e->dest->loop_father); | |
253 | } | |
254 | continue; | |
255 | } | |
256 | ||
257 | /* Use the data stored for the subloop to decide whether we may exit | |
258 | through it. It is sufficient to do this for header of the loop, | |
259 | as other basic blocks inside it must be dominated by it. */ | |
260 | if (body[i]->loop_father->header != body[i]) | |
261 | continue; | |
262 | ||
263 | if (LOOP_DATA (body[i]->loop_father)->has_call) | |
264 | { | |
265 | has_call = true; | |
266 | bitmap_set_bit (may_exit, i); | |
267 | } | |
268 | aexit = LOOP_DATA (body[i]->loop_father)->outermost_exit; | |
269 | if (aexit != loop) | |
270 | { | |
271 | bitmap_set_bit (may_exit, i); | |
272 | bitmap_set_bit (has_exit, i); | |
273 | ||
274 | if (flow_loop_nested_p (aexit, outermost_exit)) | |
275 | outermost_exit = aexit; | |
276 | } | |
277 | } | |
278 | ||
279 | loop->aux = xcalloc (1, sizeof (struct loop_data)); | |
280 | LOOP_DATA (loop)->outermost_exit = outermost_exit; | |
281 | LOOP_DATA (loop)->has_call = has_call; | |
282 | } | |
283 | ||
284 | /* Check whether we may assign a value to X from a register. */ | |
285 | ||
286 | static bool | |
287 | may_assign_reg_p (rtx x) | |
288 | { | |
289 | return can_copy_p (GET_MODE (x)); | |
290 | } | |
291 | ||
292 | /* Finds definitions that may correspond to invariants in LOOP with body BODY. | |
293 | DF is the dataflow object. */ | |
294 | ||
295 | static void | |
296 | find_defs (struct loop *loop, basic_block *body, struct df *df) | |
297 | { | |
298 | unsigned i; | |
8bdbfff5 | 299 | bitmap blocks = BITMAP_ALLOC (NULL); |
5e962776 ZD |
300 | |
301 | for (i = 0; i < loop->num_nodes; i++) | |
302 | bitmap_set_bit (blocks, body[i]->index); | |
303 | ||
304 | df_analyze_subcfg (df, blocks, DF_UD_CHAIN | DF_HARD_REGS | DF_EQUIV_NOTES); | |
8bdbfff5 | 305 | BITMAP_FREE (blocks); |
5e962776 ZD |
306 | } |
307 | ||
308 | /* Creates a new invariant for definition DEF in INSN, depending on invariants | |
309 | in DEPENDS_ON. ALWAYS_EXECUTED is true if the insn is always executed, | |
310 | unless the program ends due to a function call. */ | |
311 | ||
312 | static void | |
313 | create_new_invariant (struct def *def, rtx insn, bitmap depends_on, | |
314 | bool always_executed) | |
315 | { | |
316 | struct invariant *inv = xmalloc (sizeof (struct invariant)); | |
317 | rtx set = single_set (insn); | |
318 | ||
319 | inv->def = def; | |
320 | inv->always_executed = always_executed; | |
321 | inv->depends_on = depends_on; | |
322 | ||
323 | /* If the set is simple, usually by moving it we move the whole store out of | |
324 | the loop. Otherwise we save only cost of the computation. */ | |
325 | if (def) | |
326 | inv->cost = rtx_cost (set, SET); | |
327 | else | |
328 | inv->cost = rtx_cost (SET_SRC (set), SET); | |
329 | ||
330 | inv->move = false; | |
331 | inv->processed = false; | |
332 | inv->stamp = 0; | |
333 | inv->insn = insn; | |
334 | ||
335 | inv->invno = VARRAY_ACTIVE_SIZE (invariants); | |
336 | if (def) | |
337 | def->invno = inv->invno; | |
338 | VARRAY_PUSH_GENERIC_PTR_NOGC (invariants, inv); | |
339 | ||
340 | if (dump_file) | |
341 | { | |
342 | fprintf (dump_file, | |
343 | "Set in insn %d is invariant (%d), cost %d, depends on ", | |
344 | INSN_UID (insn), inv->invno, inv->cost); | |
345 | dump_bitmap (dump_file, inv->depends_on); | |
346 | } | |
347 | } | |
348 | ||
349 | /* Record USE at DEF. */ | |
350 | ||
351 | static void | |
352 | record_use (struct def *def, rtx *use, rtx insn) | |
353 | { | |
354 | struct use *u = xmalloc (sizeof (struct use)); | |
355 | ||
356 | if (GET_CODE (*use) == SUBREG) | |
357 | use = &SUBREG_REG (*use); | |
b5e624c6 | 358 | gcc_assert (REG_P (*use)); |
5e962776 ZD |
359 | |
360 | u->pos = use; | |
361 | u->insn = insn; | |
362 | u->next = def->uses; | |
363 | def->uses = u; | |
364 | def->n_uses++; | |
365 | } | |
366 | ||
367 | /* Finds the invariants INSN depends on and store them to the DEPENDS_ON | |
368 | bitmap. DF is the dataflow object. */ | |
369 | ||
370 | static bool | |
371 | check_dependencies (rtx insn, struct df *df, bitmap depends_on) | |
372 | { | |
373 | struct df_link *uses, *defs; | |
374 | struct ref *use, *def; | |
375 | basic_block bb = BLOCK_FOR_INSN (insn), def_bb; | |
376 | struct def *def_data; | |
377 | ||
378 | for (uses = DF_INSN_USES (df, insn); uses; uses = uses->next) | |
379 | { | |
380 | use = uses->ref; | |
381 | ||
382 | defs = DF_REF_CHAIN (use); | |
383 | if (!defs) | |
384 | continue; | |
385 | ||
386 | if (defs->next) | |
387 | return false; | |
388 | ||
389 | def = defs->ref; | |
390 | def_data = DF_REF_DATA (def); | |
391 | if (!def_data) | |
392 | return false; | |
393 | ||
394 | def_bb = DF_REF_BB (def); | |
395 | if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb)) | |
396 | return false; | |
397 | ||
398 | bitmap_set_bit (depends_on, def_data->invno); | |
399 | } | |
400 | ||
401 | return true; | |
402 | } | |
403 | ||
404 | /* Finds invariant in INSN. ALWAYS_REACHED is true if the insn is always | |
405 | executed. ALWAYS_EXECUTED is true if the insn is always executed, | |
406 | unless the program ends due to a function call. DF is the dataflow | |
407 | object. */ | |
408 | ||
409 | static void | |
410 | find_invariant_insn (rtx insn, bool always_reached, bool always_executed, | |
411 | struct df *df) | |
412 | { | |
413 | struct ref *ref; | |
414 | struct def *def; | |
415 | bitmap depends_on; | |
416 | rtx set, dest; | |
417 | bool simple = true; | |
418 | ||
419 | /* Until we get rid of LIBCALLS. */ | |
420 | if (find_reg_note (insn, REG_RETVAL, NULL_RTX) | |
421 | || find_reg_note (insn, REG_LIBCALL, NULL_RTX) | |
422 | || find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX)) | |
423 | return; | |
424 | ||
425 | set = single_set (insn); | |
426 | if (!set) | |
427 | return; | |
428 | dest = SET_DEST (set); | |
429 | ||
430 | if (GET_CODE (dest) != REG | |
431 | || HARD_REGISTER_P (dest)) | |
432 | simple = false; | |
433 | ||
434 | if (!check_maybe_invariant (SET_SRC (set)) | |
435 | || !may_assign_reg_p (SET_DEST (set))) | |
436 | return; | |
437 | ||
438 | if (may_trap_p (PATTERN (insn))) | |
439 | { | |
440 | if (!always_reached) | |
441 | return; | |
442 | ||
443 | /* Unless the exceptions are handled, the behavior is undefined | |
444 | if the trap occurs. */ | |
445 | if (flag_non_call_exceptions) | |
446 | return; | |
447 | } | |
448 | ||
8bdbfff5 | 449 | depends_on = BITMAP_ALLOC (NULL); |
5e962776 ZD |
450 | if (!check_dependencies (insn, df, depends_on)) |
451 | { | |
8bdbfff5 | 452 | BITMAP_FREE (depends_on); |
5e962776 ZD |
453 | return; |
454 | } | |
455 | ||
456 | if (simple) | |
457 | { | |
458 | ref = df_find_def (df, insn, dest); | |
459 | def = xcalloc (1, sizeof (struct def)); | |
460 | DF_REF_DATA (ref) = def; | |
461 | } | |
462 | else | |
463 | def = NULL; | |
464 | ||
465 | create_new_invariant (def, insn, depends_on, always_executed); | |
466 | } | |
467 | ||
468 | /* Record registers used in INSN that have an unique invariant definition. | |
469 | DF is the dataflow object. */ | |
470 | ||
471 | static void | |
472 | record_uses (rtx insn, struct df *df) | |
473 | { | |
474 | struct df_link *uses, *defs; | |
475 | struct ref *use, *def; | |
476 | basic_block bb = BLOCK_FOR_INSN (insn), def_bb; | |
477 | ||
478 | for (uses = DF_INSN_USES (df, insn); uses; uses = uses->next) | |
479 | { | |
480 | use = uses->ref; | |
481 | ||
482 | defs = DF_REF_CHAIN (use); | |
483 | if (!defs || defs->next) | |
484 | continue; | |
485 | def = defs->ref; | |
486 | if (!DF_REF_DATA (def)) | |
487 | continue; | |
488 | ||
489 | def_bb = DF_REF_BB (def); | |
490 | if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb)) | |
491 | continue; | |
492 | ||
493 | record_use (DF_REF_DATA (def), DF_REF_LOC (use), DF_REF_INSN (use)); | |
494 | } | |
495 | } | |
496 | ||
497 | /* Finds invariants in INSN. ALWAYS_REACHED is true if the insn is always | |
498 | executed. ALWAYS_EXECUTED is true if the insn is always executed, | |
499 | unless the program ends due to a function call. DF is the dataflow | |
500 | object. */ | |
501 | ||
502 | static void | |
503 | find_invariants_insn (rtx insn, bool always_reached, bool always_executed, | |
504 | struct df *df) | |
505 | { | |
506 | find_invariant_insn (insn, always_reached, always_executed, df); | |
507 | record_uses (insn, df); | |
508 | } | |
509 | ||
510 | /* Finds invariants in basic block BB. ALWAYS_REACHED is true if the | |
511 | basic block is always executed. ALWAYS_EXECUTED is true if the basic | |
512 | block is always executed, unless the program ends due to a function | |
513 | call. DF is the dataflow object. */ | |
514 | ||
515 | static void | |
516 | find_invariants_bb (basic_block bb, bool always_reached, bool always_executed, | |
517 | struct df *df) | |
518 | { | |
519 | rtx insn; | |
520 | ||
521 | FOR_BB_INSNS (bb, insn) | |
522 | { | |
523 | if (!INSN_P (insn)) | |
524 | continue; | |
525 | ||
526 | find_invariants_insn (insn, always_reached, always_executed, df); | |
527 | ||
528 | if (always_reached | |
4b4bf941 | 529 | && CALL_P (insn) |
5e962776 ZD |
530 | && !CONST_OR_PURE_CALL_P (insn)) |
531 | always_reached = false; | |
532 | } | |
533 | } | |
534 | ||
535 | /* Finds invariants in LOOP with body BODY. ALWAYS_REACHED is the bitmap of | |
536 | basic blocks in BODY that are always executed. ALWAYS_EXECUTED is the | |
537 | bitmap of basic blocks in BODY that are always executed unless the program | |
538 | ends due to a function call. DF is the dataflow object. */ | |
539 | ||
540 | static void | |
541 | find_invariants_body (struct loop *loop, basic_block *body, | |
542 | bitmap always_reached, bitmap always_executed, | |
543 | struct df *df) | |
544 | { | |
545 | unsigned i; | |
546 | ||
547 | for (i = 0; i < loop->num_nodes; i++) | |
548 | find_invariants_bb (body[i], | |
549 | bitmap_bit_p (always_reached, i), | |
550 | bitmap_bit_p (always_executed, i), | |
551 | df); | |
552 | } | |
553 | ||
554 | /* Finds invariants in LOOP. DF is the dataflow object. */ | |
555 | ||
556 | static void | |
557 | find_invariants (struct loop *loop, struct df *df) | |
558 | { | |
8bdbfff5 NS |
559 | bitmap may_exit = BITMAP_ALLOC (NULL); |
560 | bitmap always_reached = BITMAP_ALLOC (NULL); | |
561 | bitmap has_exit = BITMAP_ALLOC (NULL); | |
562 | bitmap always_executed = BITMAP_ALLOC (NULL); | |
5e962776 ZD |
563 | basic_block *body = get_loop_body_in_dom_order (loop); |
564 | ||
565 | find_exits (loop, body, may_exit, has_exit); | |
566 | compute_always_reached (loop, body, may_exit, always_reached); | |
567 | compute_always_reached (loop, body, has_exit, always_executed); | |
568 | ||
569 | find_defs (loop, body, df); | |
570 | find_invariants_body (loop, body, always_reached, always_executed, df); | |
571 | ||
8bdbfff5 NS |
572 | BITMAP_FREE (always_reached); |
573 | BITMAP_FREE (always_executed); | |
574 | BITMAP_FREE (may_exit); | |
575 | BITMAP_FREE (has_exit); | |
5e962776 ZD |
576 | free (body); |
577 | } | |
578 | ||
579 | /* Frees a list of uses USE. */ | |
580 | ||
581 | static void | |
582 | free_use_list (struct use *use) | |
583 | { | |
584 | struct use *next; | |
585 | ||
586 | for (; use; use = next) | |
587 | { | |
588 | next = use->next; | |
589 | free (use); | |
590 | } | |
591 | } | |
592 | ||
593 | /* Calculates cost and number of registers needed for moving invariant INV | |
594 | out of the loop and stores them to *COST and *REGS_NEEDED. */ | |
595 | ||
596 | static void | |
597 | get_inv_cost (struct invariant *inv, int *comp_cost, unsigned *regs_needed) | |
598 | { | |
599 | int acomp_cost; | |
600 | unsigned aregs_needed; | |
601 | unsigned depno; | |
602 | struct invariant *dep; | |
87c476a2 | 603 | bitmap_iterator bi; |
5e962776 ZD |
604 | |
605 | *comp_cost = 0; | |
606 | *regs_needed = 0; | |
607 | if (inv->move | |
608 | || inv->stamp == actual_stamp) | |
609 | return; | |
610 | inv->stamp = actual_stamp; | |
611 | ||
612 | (*regs_needed)++; | |
613 | (*comp_cost) += inv->cost; | |
614 | ||
87c476a2 | 615 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi) |
5e962776 ZD |
616 | { |
617 | dep = VARRAY_GENERIC_PTR_NOGC (invariants, depno); | |
618 | ||
619 | get_inv_cost (dep, &acomp_cost, &aregs_needed); | |
620 | ||
621 | if (aregs_needed | |
622 | /* We need to check always_executed, since if the original value of | |
623 | the invariant may be preserved, we may need to keep it in a | |
624 | separate register. TODO check whether the register has an | |
625 | use outside of the loop. */ | |
626 | && dep->always_executed | |
627 | && !dep->def->uses->next) | |
628 | { | |
629 | /* If this is a single use, after moving the dependency we will not | |
630 | need a new register. */ | |
631 | aregs_needed--; | |
632 | } | |
633 | ||
634 | (*regs_needed) += aregs_needed; | |
635 | (*comp_cost) += acomp_cost; | |
87c476a2 | 636 | } |
5e962776 ZD |
637 | } |
638 | ||
639 | /* Calculates gain for eliminating invariant INV. REGS_USED is the number | |
640 | of registers used in the loop, N_INV_USES is the number of uses of | |
641 | invariants, NEW_REGS is the number of new variables already added due to | |
642 | the invariant motion. The number of registers needed for it is stored in | |
643 | *REGS_NEEDED. */ | |
644 | ||
645 | static int | |
646 | gain_for_invariant (struct invariant *inv, unsigned *regs_needed, | |
647 | unsigned new_regs, unsigned regs_used, unsigned n_inv_uses) | |
648 | { | |
649 | int comp_cost, size_cost; | |
650 | ||
651 | get_inv_cost (inv, &comp_cost, regs_needed); | |
652 | actual_stamp++; | |
653 | ||
654 | size_cost = (global_cost_for_size (new_regs + *regs_needed, | |
655 | regs_used, n_inv_uses) | |
656 | - global_cost_for_size (new_regs, regs_used, n_inv_uses)); | |
657 | ||
658 | return comp_cost - size_cost; | |
659 | } | |
660 | ||
661 | /* Finds invariant with best gain for moving. Returns the gain, stores | |
662 | the invariant in *BEST and number of registers needed for it to | |
663 | *REGS_NEEDED. REGS_USED is the number of registers used in | |
664 | the loop, N_INV_USES is the number of uses of invariants. NEW_REGS | |
665 | is the number of new variables already added due to invariant motion. */ | |
666 | ||
667 | static int | |
668 | best_gain_for_invariant (struct invariant **best, unsigned *regs_needed, | |
669 | unsigned new_regs, unsigned regs_used, | |
670 | unsigned n_inv_uses) | |
671 | { | |
672 | struct invariant *inv; | |
673 | int gain = 0, again; | |
674 | unsigned aregs_needed, invno; | |
675 | ||
676 | for (invno = 0; invno < VARRAY_ACTIVE_SIZE (invariants); invno++) | |
677 | { | |
678 | inv = VARRAY_GENERIC_PTR_NOGC (invariants, invno); | |
679 | if (inv->move) | |
680 | continue; | |
681 | ||
682 | again = gain_for_invariant (inv, &aregs_needed, | |
683 | new_regs, regs_used, n_inv_uses); | |
684 | if (again > gain) | |
685 | { | |
686 | gain = again; | |
687 | *best = inv; | |
688 | *regs_needed = aregs_needed; | |
689 | } | |
690 | } | |
691 | ||
692 | return gain; | |
693 | } | |
694 | ||
695 | /* Marks invariant INVNO and all its dependencies for moving. */ | |
696 | ||
697 | static void | |
698 | set_move_mark (unsigned invno) | |
699 | { | |
700 | struct invariant *inv = VARRAY_GENERIC_PTR_NOGC (invariants, invno); | |
87c476a2 | 701 | bitmap_iterator bi; |
5e962776 ZD |
702 | |
703 | if (inv->move) | |
704 | return; | |
705 | inv->move = true; | |
706 | ||
707 | if (dump_file) | |
708 | fprintf (dump_file, "Decided to move invariant %d\n", invno); | |
709 | ||
87c476a2 ZD |
710 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, invno, bi) |
711 | { | |
712 | set_move_mark (invno); | |
713 | } | |
5e962776 ZD |
714 | } |
715 | ||
716 | /* Determines which invariants to move. DF is the dataflow object. */ | |
717 | ||
718 | static void | |
719 | find_invariants_to_move (struct df *df) | |
720 | { | |
721 | unsigned i, regs_used, n_inv_uses, regs_needed = 0, new_regs; | |
722 | struct invariant *inv = NULL; | |
723 | ||
5e962776 ZD |
724 | if (!VARRAY_ACTIVE_SIZE (invariants)) |
725 | return; | |
726 | ||
727 | /* Now something slightly more involved. First estimate the number of used | |
728 | registers. */ | |
729 | n_inv_uses = 0; | |
730 | ||
731 | /* We do not really do a good job in this estimation; put some initial bound | |
732 | here to stand for induction variables etc. that we do not detect. */ | |
733 | regs_used = 2; | |
734 | ||
735 | for (i = 0; i < df->n_regs; i++) | |
736 | { | |
737 | if (!DF_REGNO_FIRST_DEF (df, i) && DF_REGNO_LAST_USE (df, i)) | |
738 | { | |
739 | /* This is a value that is used but not changed inside loop. */ | |
740 | regs_used++; | |
741 | } | |
742 | } | |
743 | ||
744 | for (i = 0; i < VARRAY_ACTIVE_SIZE (invariants); i++) | |
745 | { | |
746 | inv = VARRAY_GENERIC_PTR_NOGC (invariants, i); | |
747 | if (inv->def) | |
748 | n_inv_uses += inv->def->n_uses; | |
749 | } | |
750 | ||
751 | new_regs = 0; | |
752 | while (best_gain_for_invariant (&inv, ®s_needed, | |
753 | new_regs, regs_used, n_inv_uses) > 0) | |
754 | { | |
755 | set_move_mark (inv->invno); | |
756 | new_regs += regs_needed; | |
757 | } | |
758 | } | |
759 | ||
760 | /* Move invariant INVNO out of the LOOP. DF is the dataflow object. */ | |
761 | ||
762 | static void | |
763 | move_invariant_reg (struct loop *loop, unsigned invno, struct df *df) | |
764 | { | |
765 | struct invariant *inv = VARRAY_GENERIC_PTR_NOGC (invariants, invno); | |
766 | unsigned i; | |
767 | basic_block preheader = loop_preheader_edge (loop)->src; | |
768 | rtx reg, set; | |
769 | struct use *use; | |
87c476a2 | 770 | bitmap_iterator bi; |
5e962776 ZD |
771 | |
772 | if (inv->processed) | |
773 | return; | |
774 | inv->processed = true; | |
775 | ||
776 | if (inv->depends_on) | |
777 | { | |
87c476a2 | 778 | EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, i, bi) |
5e962776 ZD |
779 | { |
780 | move_invariant_reg (loop, i, df); | |
87c476a2 | 781 | } |
5e962776 ZD |
782 | } |
783 | ||
784 | /* Move the set out of the loop. If the set is always executed (we could | |
785 | omit this condition if we know that the register is unused outside of the | |
786 | loop, but it does not seem worth finding out) and it has no uses that | |
787 | would not be dominated by it, we may just move it (TODO). Otherwise we | |
788 | need to create a temporary register. */ | |
789 | set = single_set (inv->insn); | |
790 | reg = gen_reg_rtx (GET_MODE (SET_DEST (set))); | |
791 | df_pattern_emit_after (df, gen_move_insn (SET_DEST (set), reg), | |
792 | BLOCK_FOR_INSN (inv->insn), inv->insn); | |
793 | SET_DEST (set) = reg; | |
794 | reorder_insns (inv->insn, inv->insn, BB_END (preheader)); | |
795 | df_insn_modify (df, preheader, inv->insn); | |
796 | ||
797 | /* Replace the uses we know to be dominated. It saves work for copy | |
798 | propagation, and also it is necessary so that dependent invariants | |
799 | are computed right. */ | |
800 | if (inv->def) | |
801 | { | |
802 | for (use = inv->def->uses; use; use = use->next) | |
803 | { | |
804 | *use->pos = reg; | |
805 | df_insn_modify (df, BLOCK_FOR_INSN (use->insn), use->insn); | |
806 | } | |
807 | } | |
808 | } | |
809 | ||
810 | /* Move selected invariant out of the LOOP. Newly created regs are marked | |
811 | in TEMPORARY_REGS. DF is the dataflow object. */ | |
812 | ||
813 | static void | |
814 | move_invariants (struct loop *loop, struct df *df) | |
815 | { | |
816 | struct invariant *inv; | |
817 | unsigned i; | |
818 | ||
819 | for (i = 0; i < VARRAY_ACTIVE_SIZE (invariants); i++) | |
820 | { | |
821 | inv = VARRAY_GENERIC_PTR_NOGC (invariants, i); | |
822 | if (inv->move) | |
823 | move_invariant_reg (loop, i, df); | |
824 | } | |
825 | } | |
826 | ||
827 | /* Initializes invariant motion data. */ | |
828 | ||
829 | static void | |
830 | init_inv_motion_data (void) | |
831 | { | |
832 | actual_stamp = 1; | |
833 | ||
834 | if (!invariants) | |
835 | VARRAY_GENERIC_PTR_NOGC_INIT (invariants, 100, "invariants"); | |
836 | } | |
837 | ||
838 | /* Frees the data allocated by invariant motion. DF is the dataflow | |
839 | object. */ | |
840 | ||
841 | static void | |
842 | free_inv_motion_data (struct df *df) | |
843 | { | |
844 | unsigned i; | |
845 | struct def *def; | |
846 | struct invariant *inv; | |
847 | ||
848 | for (i = 0; i < df->n_defs; i++) | |
849 | { | |
850 | if (!df->defs[i]) | |
851 | continue; | |
852 | ||
853 | def = DF_REF_DATA (df->defs[i]); | |
854 | if (!def) | |
855 | continue; | |
856 | ||
857 | free_use_list (def->uses); | |
858 | free (def); | |
859 | DF_REF_DATA (df->defs[i]) = NULL; | |
860 | } | |
861 | ||
862 | for (i = 0; i < VARRAY_ACTIVE_SIZE (invariants); i++) | |
863 | { | |
864 | inv = VARRAY_GENERIC_PTR_NOGC (invariants, i); | |
8bdbfff5 | 865 | BITMAP_FREE (inv->depends_on); |
5e962776 ZD |
866 | free (inv); |
867 | } | |
868 | VARRAY_POP_ALL (invariants); | |
869 | } | |
870 | ||
871 | /* Move the invariants out of the LOOP. DF is the dataflow object. */ | |
872 | ||
873 | static void | |
874 | move_single_loop_invariants (struct loop *loop, struct df *df) | |
875 | { | |
876 | init_inv_motion_data (); | |
877 | ||
878 | find_invariants (loop, df); | |
879 | find_invariants_to_move (df); | |
880 | move_invariants (loop, df); | |
881 | ||
882 | free_inv_motion_data (df); | |
883 | } | |
884 | ||
885 | /* Releases the auxiliary data for LOOP. */ | |
886 | ||
887 | static void | |
888 | free_loop_data (struct loop *loop) | |
889 | { | |
890 | struct loop_data *data = LOOP_DATA (loop); | |
891 | ||
892 | free (data); | |
893 | loop->aux = NULL; | |
894 | } | |
895 | ||
896 | /* Move the invariants out of the LOOPS. */ | |
897 | ||
898 | void | |
899 | move_loop_invariants (struct loops *loops) | |
900 | { | |
901 | struct loop *loop; | |
902 | unsigned i; | |
903 | struct df *df = df_init (); | |
904 | ||
905 | /* Process the loops, innermost first. */ | |
906 | loop = loops->tree_root; | |
907 | while (loop->inner) | |
908 | loop = loop->inner; | |
909 | ||
910 | while (loop != loops->tree_root) | |
911 | { | |
912 | move_single_loop_invariants (loop, df); | |
913 | ||
914 | if (loop->next) | |
915 | { | |
916 | loop = loop->next; | |
917 | while (loop->inner) | |
918 | loop = loop->inner; | |
919 | } | |
920 | else | |
921 | loop = loop->outer; | |
922 | } | |
923 | ||
924 | for (i = 1; i < loops->num; i++) | |
925 | if (loops->parray[i]) | |
926 | free_loop_data (loops->parray[i]); | |
927 | ||
928 | df_finish (df); | |
929 | } |