Daily bump.
[official-gcc.git] / gcc / postreload.c
blob0638709639b1797563c6f1edfd67aa30cab16152
1 /* Perform simple optimizations to clean up the result of reload.
2 Copyright (C) 1987-2018 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
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/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "target.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "predict.h"
28 #include "df.h"
29 #include "memmodel.h"
30 #include "tm_p.h"
31 #include "optabs.h"
32 #include "regs.h"
33 #include "emit-rtl.h"
34 #include "recog.h"
36 #include "cfgrtl.h"
37 #include "cfgbuild.h"
38 #include "cfgcleanup.h"
39 #include "reload.h"
40 #include "cselib.h"
41 #include "tree-pass.h"
42 #include "dbgcnt.h"
44 static int reload_cse_noop_set_p (rtx);
45 static bool reload_cse_simplify (rtx_insn *, rtx);
46 static void reload_cse_regs_1 (void);
47 static int reload_cse_simplify_set (rtx, rtx_insn *);
48 static int reload_cse_simplify_operands (rtx_insn *, rtx);
50 static void reload_combine (void);
51 static void reload_combine_note_use (rtx *, rtx_insn *, int, rtx);
52 static void reload_combine_note_store (rtx, const_rtx, void *);
54 static bool reload_cse_move2add (rtx_insn *);
55 static void move2add_note_store (rtx, const_rtx, void *);
57 /* Call cse / combine like post-reload optimization phases.
58 FIRST is the first instruction. */
60 static void
61 reload_cse_regs (rtx_insn *first ATTRIBUTE_UNUSED)
63 bool moves_converted;
64 reload_cse_regs_1 ();
65 reload_combine ();
66 moves_converted = reload_cse_move2add (first);
67 if (flag_expensive_optimizations)
69 if (moves_converted)
70 reload_combine ();
71 reload_cse_regs_1 ();
75 /* See whether a single set SET is a noop. */
76 static int
77 reload_cse_noop_set_p (rtx set)
79 if (cselib_reg_set_mode (SET_DEST (set)) != GET_MODE (SET_DEST (set)))
80 return 0;
82 return rtx_equal_for_cselib_p (SET_DEST (set), SET_SRC (set));
85 /* Try to simplify INSN. Return true if the CFG may have changed. */
86 static bool
87 reload_cse_simplify (rtx_insn *insn, rtx testreg)
89 rtx body = PATTERN (insn);
90 basic_block insn_bb = BLOCK_FOR_INSN (insn);
91 unsigned insn_bb_succs = EDGE_COUNT (insn_bb->succs);
93 /* If NO_FUNCTION_CSE has been set by the target, then we should not try
94 to cse function calls. */
95 if (NO_FUNCTION_CSE && CALL_P (insn))
96 return false;
98 if (GET_CODE (body) == SET)
100 int count = 0;
102 /* Simplify even if we may think it is a no-op.
103 We may think a memory load of a value smaller than WORD_SIZE
104 is redundant because we haven't taken into account possible
105 implicit extension. reload_cse_simplify_set() will bring
106 this out, so it's safer to simplify before we delete. */
107 count += reload_cse_simplify_set (body, insn);
109 if (!count && reload_cse_noop_set_p (body))
111 if (check_for_inc_dec (insn))
112 delete_insn_and_edges (insn);
113 /* We're done with this insn. */
114 goto done;
117 if (count > 0)
118 apply_change_group ();
119 else
120 reload_cse_simplify_operands (insn, testreg);
122 else if (GET_CODE (body) == PARALLEL)
124 int i;
125 int count = 0;
126 rtx value = NULL_RTX;
128 /* Registers mentioned in the clobber list for an asm cannot be reused
129 within the body of the asm. Invalidate those registers now so that
130 we don't try to substitute values for them. */
131 if (asm_noperands (body) >= 0)
133 for (i = XVECLEN (body, 0) - 1; i >= 0; --i)
135 rtx part = XVECEXP (body, 0, i);
136 if (GET_CODE (part) == CLOBBER && REG_P (XEXP (part, 0)))
137 cselib_invalidate_rtx (XEXP (part, 0));
141 /* If every action in a PARALLEL is a noop, we can delete
142 the entire PARALLEL. */
143 for (i = XVECLEN (body, 0) - 1; i >= 0; --i)
145 rtx part = XVECEXP (body, 0, i);
146 if (GET_CODE (part) == SET)
148 if (! reload_cse_noop_set_p (part))
149 break;
150 if (REG_P (SET_DEST (part))
151 && REG_FUNCTION_VALUE_P (SET_DEST (part)))
153 if (value)
154 break;
155 value = SET_DEST (part);
158 else if (GET_CODE (part) != CLOBBER
159 && GET_CODE (part) != USE)
160 break;
163 if (i < 0)
165 if (check_for_inc_dec (insn))
166 delete_insn_and_edges (insn);
167 /* We're done with this insn. */
168 goto done;
171 /* It's not a no-op, but we can try to simplify it. */
172 for (i = XVECLEN (body, 0) - 1; i >= 0; --i)
173 if (GET_CODE (XVECEXP (body, 0, i)) == SET)
174 count += reload_cse_simplify_set (XVECEXP (body, 0, i), insn);
176 if (count > 0)
177 apply_change_group ();
178 else
179 reload_cse_simplify_operands (insn, testreg);
182 done:
183 return (EDGE_COUNT (insn_bb->succs) != insn_bb_succs);
186 /* Do a very simple CSE pass over the hard registers.
188 This function detects no-op moves where we happened to assign two
189 different pseudo-registers to the same hard register, and then
190 copied one to the other. Reload will generate a useless
191 instruction copying a register to itself.
193 This function also detects cases where we load a value from memory
194 into two different registers, and (if memory is more expensive than
195 registers) changes it to simply copy the first register into the
196 second register.
198 Another optimization is performed that scans the operands of each
199 instruction to see whether the value is already available in a
200 hard register. It then replaces the operand with the hard register
201 if possible, much like an optional reload would. */
203 static void
204 reload_cse_regs_1 (void)
206 bool cfg_changed = false;
207 basic_block bb;
208 rtx_insn *insn;
209 rtx testreg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
211 cselib_init (CSELIB_RECORD_MEMORY);
212 init_alias_analysis ();
214 FOR_EACH_BB_FN (bb, cfun)
215 FOR_BB_INSNS (bb, insn)
217 if (INSN_P (insn))
218 cfg_changed |= reload_cse_simplify (insn, testreg);
220 cselib_process_insn (insn);
223 /* Clean up. */
224 end_alias_analysis ();
225 cselib_finish ();
226 if (cfg_changed)
227 cleanup_cfg (0);
230 /* Try to simplify a single SET instruction. SET is the set pattern.
231 INSN is the instruction it came from.
232 This function only handles one case: if we set a register to a value
233 which is not a register, we try to find that value in some other register
234 and change the set into a register copy. */
236 static int
237 reload_cse_simplify_set (rtx set, rtx_insn *insn)
239 int did_change = 0;
240 int dreg;
241 rtx src;
242 reg_class_t dclass;
243 int old_cost;
244 cselib_val *val;
245 struct elt_loc_list *l;
246 enum rtx_code extend_op = UNKNOWN;
247 bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
249 dreg = true_regnum (SET_DEST (set));
250 if (dreg < 0)
251 return 0;
253 src = SET_SRC (set);
254 if (side_effects_p (src) || true_regnum (src) >= 0)
255 return 0;
257 dclass = REGNO_REG_CLASS (dreg);
259 /* When replacing a memory with a register, we need to honor assumptions
260 that combine made wrt the contents of sign bits. We'll do this by
261 generating an extend instruction instead of a reg->reg copy. Thus
262 the destination must be a register that we can widen. */
263 if (MEM_P (src)
264 && (extend_op = load_extend_op (GET_MODE (src))) != UNKNOWN
265 && !REG_P (SET_DEST (set)))
266 return 0;
268 val = cselib_lookup (src, GET_MODE (SET_DEST (set)), 0, VOIDmode);
269 if (! val)
270 return 0;
272 /* If memory loads are cheaper than register copies, don't change them. */
273 if (MEM_P (src))
274 old_cost = memory_move_cost (GET_MODE (src), dclass, true);
275 else if (REG_P (src))
276 old_cost = register_move_cost (GET_MODE (src),
277 REGNO_REG_CLASS (REGNO (src)), dclass);
278 else
279 old_cost = set_src_cost (src, GET_MODE (SET_DEST (set)), speed);
281 for (l = val->locs; l; l = l->next)
283 rtx this_rtx = l->loc;
284 int this_cost;
286 if (CONSTANT_P (this_rtx) && ! references_value_p (this_rtx, 0))
288 if (extend_op != UNKNOWN)
290 wide_int result;
292 if (!CONST_SCALAR_INT_P (this_rtx))
293 continue;
295 switch (extend_op)
297 case ZERO_EXTEND:
298 result = wide_int::from (rtx_mode_t (this_rtx,
299 GET_MODE (src)),
300 BITS_PER_WORD, UNSIGNED);
301 break;
302 case SIGN_EXTEND:
303 result = wide_int::from (rtx_mode_t (this_rtx,
304 GET_MODE (src)),
305 BITS_PER_WORD, SIGNED);
306 break;
307 default:
308 gcc_unreachable ();
310 this_rtx = immed_wide_int_const (result, word_mode);
313 this_cost = set_src_cost (this_rtx, GET_MODE (SET_DEST (set)), speed);
315 else if (REG_P (this_rtx))
317 if (extend_op != UNKNOWN)
319 this_rtx = gen_rtx_fmt_e (extend_op, word_mode, this_rtx);
320 this_cost = set_src_cost (this_rtx, word_mode, speed);
322 else
323 this_cost = register_move_cost (GET_MODE (this_rtx),
324 REGNO_REG_CLASS (REGNO (this_rtx)),
325 dclass);
327 else
328 continue;
330 /* If equal costs, prefer registers over anything else. That
331 tends to lead to smaller instructions on some machines. */
332 if (this_cost < old_cost
333 || (this_cost == old_cost
334 && REG_P (this_rtx)
335 && !REG_P (SET_SRC (set))))
337 if (extend_op != UNKNOWN
338 && REG_CAN_CHANGE_MODE_P (REGNO (SET_DEST (set)),
339 GET_MODE (SET_DEST (set)), word_mode))
341 rtx wide_dest = gen_rtx_REG (word_mode, REGNO (SET_DEST (set)));
342 ORIGINAL_REGNO (wide_dest) = ORIGINAL_REGNO (SET_DEST (set));
343 validate_change (insn, &SET_DEST (set), wide_dest, 1);
346 validate_unshare_change (insn, &SET_SRC (set), this_rtx, 1);
347 old_cost = this_cost, did_change = 1;
351 return did_change;
354 /* Try to replace operands in INSN with equivalent values that are already
355 in registers. This can be viewed as optional reloading.
357 For each non-register operand in the insn, see if any hard regs are
358 known to be equivalent to that operand. Record the alternatives which
359 can accept these hard registers. Among all alternatives, select the
360 ones which are better or equal to the one currently matching, where
361 "better" is in terms of '?' and '!' constraints. Among the remaining
362 alternatives, select the one which replaces most operands with
363 hard registers. */
365 static int
366 reload_cse_simplify_operands (rtx_insn *insn, rtx testreg)
368 int i, j;
370 /* For each operand, all registers that are equivalent to it. */
371 HARD_REG_SET equiv_regs[MAX_RECOG_OPERANDS];
373 const char *constraints[MAX_RECOG_OPERANDS];
375 /* Vector recording how bad an alternative is. */
376 int *alternative_reject;
377 /* Vector recording how many registers can be introduced by choosing
378 this alternative. */
379 int *alternative_nregs;
380 /* Array of vectors recording, for each operand and each alternative,
381 which hard register to substitute, or -1 if the operand should be
382 left as it is. */
383 int *op_alt_regno[MAX_RECOG_OPERANDS];
384 /* Array of alternatives, sorted in order of decreasing desirability. */
385 int *alternative_order;
387 extract_constrain_insn (insn);
389 if (recog_data.n_alternatives == 0 || recog_data.n_operands == 0)
390 return 0;
392 alternative_reject = XALLOCAVEC (int, recog_data.n_alternatives);
393 alternative_nregs = XALLOCAVEC (int, recog_data.n_alternatives);
394 alternative_order = XALLOCAVEC (int, recog_data.n_alternatives);
395 memset (alternative_reject, 0, recog_data.n_alternatives * sizeof (int));
396 memset (alternative_nregs, 0, recog_data.n_alternatives * sizeof (int));
398 /* For each operand, find out which regs are equivalent. */
399 for (i = 0; i < recog_data.n_operands; i++)
401 cselib_val *v;
402 struct elt_loc_list *l;
403 rtx op;
405 CLEAR_HARD_REG_SET (equiv_regs[i]);
407 /* cselib blows up on CODE_LABELs. Trying to fix that doesn't seem
408 right, so avoid the problem here. Similarly NOTE_INSN_DELETED_LABEL.
409 Likewise if we have a constant and the insn pattern doesn't tell us
410 the mode we need. */
411 if (LABEL_P (recog_data.operand[i])
412 || (NOTE_P (recog_data.operand[i])
413 && NOTE_KIND (recog_data.operand[i]) == NOTE_INSN_DELETED_LABEL)
414 || (CONSTANT_P (recog_data.operand[i])
415 && recog_data.operand_mode[i] == VOIDmode))
416 continue;
418 op = recog_data.operand[i];
419 if (MEM_P (op) && load_extend_op (GET_MODE (op)) != UNKNOWN)
421 rtx set = single_set (insn);
423 /* We might have multiple sets, some of which do implicit
424 extension. Punt on this for now. */
425 if (! set)
426 continue;
427 /* If the destination is also a MEM or a STRICT_LOW_PART, no
428 extension applies.
429 Also, if there is an explicit extension, we don't have to
430 worry about an implicit one. */
431 else if (MEM_P (SET_DEST (set))
432 || GET_CODE (SET_DEST (set)) == STRICT_LOW_PART
433 || GET_CODE (SET_SRC (set)) == ZERO_EXTEND
434 || GET_CODE (SET_SRC (set)) == SIGN_EXTEND)
435 ; /* Continue ordinary processing. */
436 /* If the register cannot change mode to word_mode, it follows that
437 it cannot have been used in word_mode. */
438 else if (REG_P (SET_DEST (set))
439 && !REG_CAN_CHANGE_MODE_P (REGNO (SET_DEST (set)),
440 GET_MODE (SET_DEST (set)),
441 word_mode))
442 ; /* Continue ordinary processing. */
443 /* If this is a straight load, make the extension explicit. */
444 else if (REG_P (SET_DEST (set))
445 && recog_data.n_operands == 2
446 && SET_SRC (set) == op
447 && SET_DEST (set) == recog_data.operand[1-i])
449 validate_change (insn, recog_data.operand_loc[i],
450 gen_rtx_fmt_e (load_extend_op (GET_MODE (op)),
451 word_mode, op),
453 validate_change (insn, recog_data.operand_loc[1-i],
454 gen_rtx_REG (word_mode, REGNO (SET_DEST (set))),
456 if (! apply_change_group ())
457 return 0;
458 return reload_cse_simplify_operands (insn, testreg);
460 else
461 /* ??? There might be arithmetic operations with memory that are
462 safe to optimize, but is it worth the trouble? */
463 continue;
466 if (side_effects_p (op))
467 continue;
468 v = cselib_lookup (op, recog_data.operand_mode[i], 0, VOIDmode);
469 if (! v)
470 continue;
472 for (l = v->locs; l; l = l->next)
473 if (REG_P (l->loc))
474 SET_HARD_REG_BIT (equiv_regs[i], REGNO (l->loc));
477 alternative_mask preferred = get_preferred_alternatives (insn);
478 for (i = 0; i < recog_data.n_operands; i++)
480 machine_mode mode;
481 int regno;
482 const char *p;
484 op_alt_regno[i] = XALLOCAVEC (int, recog_data.n_alternatives);
485 for (j = 0; j < recog_data.n_alternatives; j++)
486 op_alt_regno[i][j] = -1;
488 p = constraints[i] = recog_data.constraints[i];
489 mode = recog_data.operand_mode[i];
491 /* Add the reject values for each alternative given by the constraints
492 for this operand. */
493 j = 0;
494 while (*p != '\0')
496 char c = *p++;
497 if (c == ',')
498 j++;
499 else if (c == '?')
500 alternative_reject[j] += 3;
501 else if (c == '!')
502 alternative_reject[j] += 300;
505 /* We won't change operands which are already registers. We
506 also don't want to modify output operands. */
507 regno = true_regnum (recog_data.operand[i]);
508 if (regno >= 0
509 || constraints[i][0] == '='
510 || constraints[i][0] == '+')
511 continue;
513 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
515 enum reg_class rclass = NO_REGS;
517 if (! TEST_HARD_REG_BIT (equiv_regs[i], regno))
518 continue;
520 set_mode_and_regno (testreg, mode, regno);
522 /* We found a register equal to this operand. Now look for all
523 alternatives that can accept this register and have not been
524 assigned a register they can use yet. */
525 j = 0;
526 p = constraints[i];
527 for (;;)
529 char c = *p;
531 switch (c)
533 case 'g':
534 rclass = reg_class_subunion[rclass][GENERAL_REGS];
535 break;
537 default:
538 rclass
539 = (reg_class_subunion
540 [rclass]
541 [reg_class_for_constraint (lookup_constraint (p))]);
542 break;
544 case ',': case '\0':
545 /* See if REGNO fits this alternative, and set it up as the
546 replacement register if we don't have one for this
547 alternative yet and the operand being replaced is not
548 a cheap CONST_INT. */
549 if (op_alt_regno[i][j] == -1
550 && TEST_BIT (preferred, j)
551 && reg_fits_class_p (testreg, rclass, 0, mode)
552 && (!CONST_INT_P (recog_data.operand[i])
553 || (set_src_cost (recog_data.operand[i], mode,
554 optimize_bb_for_speed_p
555 (BLOCK_FOR_INSN (insn)))
556 > set_src_cost (testreg, mode,
557 optimize_bb_for_speed_p
558 (BLOCK_FOR_INSN (insn))))))
560 alternative_nregs[j]++;
561 op_alt_regno[i][j] = regno;
563 j++;
564 rclass = NO_REGS;
565 break;
567 p += CONSTRAINT_LEN (c, p);
569 if (c == '\0')
570 break;
575 /* Record all alternatives which are better or equal to the currently
576 matching one in the alternative_order array. */
577 for (i = j = 0; i < recog_data.n_alternatives; i++)
578 if (alternative_reject[i] <= alternative_reject[which_alternative])
579 alternative_order[j++] = i;
580 recog_data.n_alternatives = j;
582 /* Sort it. Given a small number of alternatives, a dumb algorithm
583 won't hurt too much. */
584 for (i = 0; i < recog_data.n_alternatives - 1; i++)
586 int best = i;
587 int best_reject = alternative_reject[alternative_order[i]];
588 int best_nregs = alternative_nregs[alternative_order[i]];
590 for (j = i + 1; j < recog_data.n_alternatives; j++)
592 int this_reject = alternative_reject[alternative_order[j]];
593 int this_nregs = alternative_nregs[alternative_order[j]];
595 if (this_reject < best_reject
596 || (this_reject == best_reject && this_nregs > best_nregs))
598 best = j;
599 best_reject = this_reject;
600 best_nregs = this_nregs;
604 std::swap (alternative_order[best], alternative_order[i]);
607 /* Substitute the operands as determined by op_alt_regno for the best
608 alternative. */
609 j = alternative_order[0];
611 for (i = 0; i < recog_data.n_operands; i++)
613 machine_mode mode = recog_data.operand_mode[i];
614 if (op_alt_regno[i][j] == -1)
615 continue;
617 validate_change (insn, recog_data.operand_loc[i],
618 gen_rtx_REG (mode, op_alt_regno[i][j]), 1);
621 for (i = recog_data.n_dups - 1; i >= 0; i--)
623 int op = recog_data.dup_num[i];
624 machine_mode mode = recog_data.operand_mode[op];
626 if (op_alt_regno[op][j] == -1)
627 continue;
629 validate_change (insn, recog_data.dup_loc[i],
630 gen_rtx_REG (mode, op_alt_regno[op][j]), 1);
633 return apply_change_group ();
636 /* If reload couldn't use reg+reg+offset addressing, try to use reg+reg
637 addressing now.
638 This code might also be useful when reload gave up on reg+reg addressing
639 because of clashes between the return register and INDEX_REG_CLASS. */
641 /* The maximum number of uses of a register we can keep track of to
642 replace them with reg+reg addressing. */
643 #define RELOAD_COMBINE_MAX_USES 16
645 /* Describes a recorded use of a register. */
646 struct reg_use
648 /* The insn where a register has been used. */
649 rtx_insn *insn;
650 /* Points to the memory reference enclosing the use, if any, NULL_RTX
651 otherwise. */
652 rtx containing_mem;
653 /* Location of the register within INSN. */
654 rtx *usep;
655 /* The reverse uid of the insn. */
656 int ruid;
659 /* If the register is used in some unknown fashion, USE_INDEX is negative.
660 If it is dead, USE_INDEX is RELOAD_COMBINE_MAX_USES, and STORE_RUID
661 indicates where it is first set or clobbered.
662 Otherwise, USE_INDEX is the index of the last encountered use of the
663 register (which is first among these we have seen since we scan backwards).
664 USE_RUID indicates the first encountered, i.e. last, of these uses.
665 If ALL_OFFSETS_MATCH is true, all encountered uses were inside a PLUS
666 with a constant offset; OFFSET contains this constant in that case.
667 STORE_RUID is always meaningful if we only want to use a value in a
668 register in a different place: it denotes the next insn in the insn
669 stream (i.e. the last encountered) that sets or clobbers the register.
670 REAL_STORE_RUID is similar, but clobbers are ignored when updating it. */
671 static struct
673 struct reg_use reg_use[RELOAD_COMBINE_MAX_USES];
674 rtx offset;
675 int use_index;
676 int store_ruid;
677 int real_store_ruid;
678 int use_ruid;
679 bool all_offsets_match;
680 } reg_state[FIRST_PSEUDO_REGISTER];
682 /* Reverse linear uid. This is increased in reload_combine while scanning
683 the instructions from last to first. It is used to set last_label_ruid
684 and the store_ruid / use_ruid fields in reg_state. */
685 static int reload_combine_ruid;
687 /* The RUID of the last label we encountered in reload_combine. */
688 static int last_label_ruid;
690 /* The RUID of the last jump we encountered in reload_combine. */
691 static int last_jump_ruid;
693 /* The register numbers of the first and last index register. A value of
694 -1 in LAST_INDEX_REG indicates that we've previously computed these
695 values and found no suitable index registers. */
696 static int first_index_reg = -1;
697 static int last_index_reg;
699 #define LABEL_LIVE(LABEL) \
700 (label_live[CODE_LABEL_NUMBER (LABEL) - min_labelno])
702 /* Subroutine of reload_combine_split_ruids, called to fix up a single
703 ruid pointed to by *PRUID if it is higher than SPLIT_RUID. */
705 static inline void
706 reload_combine_split_one_ruid (int *pruid, int split_ruid)
708 if (*pruid > split_ruid)
709 (*pruid)++;
712 /* Called when we insert a new insn in a position we've already passed in
713 the scan. Examine all our state, increasing all ruids that are higher
714 than SPLIT_RUID by one in order to make room for a new insn. */
716 static void
717 reload_combine_split_ruids (int split_ruid)
719 unsigned i;
721 reload_combine_split_one_ruid (&reload_combine_ruid, split_ruid);
722 reload_combine_split_one_ruid (&last_label_ruid, split_ruid);
723 reload_combine_split_one_ruid (&last_jump_ruid, split_ruid);
725 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
727 int j, idx = reg_state[i].use_index;
728 reload_combine_split_one_ruid (&reg_state[i].use_ruid, split_ruid);
729 reload_combine_split_one_ruid (&reg_state[i].store_ruid, split_ruid);
730 reload_combine_split_one_ruid (&reg_state[i].real_store_ruid,
731 split_ruid);
732 if (idx < 0)
733 continue;
734 for (j = idx; j < RELOAD_COMBINE_MAX_USES; j++)
736 reload_combine_split_one_ruid (&reg_state[i].reg_use[j].ruid,
737 split_ruid);
742 /* Called when we are about to rescan a previously encountered insn with
743 reload_combine_note_use after modifying some part of it. This clears all
744 information about uses in that particular insn. */
746 static void
747 reload_combine_purge_insn_uses (rtx_insn *insn)
749 unsigned i;
751 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
753 int j, k, idx = reg_state[i].use_index;
754 if (idx < 0)
755 continue;
756 j = k = RELOAD_COMBINE_MAX_USES;
757 while (j-- > idx)
759 if (reg_state[i].reg_use[j].insn != insn)
761 k--;
762 if (k != j)
763 reg_state[i].reg_use[k] = reg_state[i].reg_use[j];
766 reg_state[i].use_index = k;
770 /* Called when we need to forget about all uses of REGNO after an insn
771 which is identified by RUID. */
773 static void
774 reload_combine_purge_reg_uses_after_ruid (unsigned regno, int ruid)
776 int j, k, idx = reg_state[regno].use_index;
777 if (idx < 0)
778 return;
779 j = k = RELOAD_COMBINE_MAX_USES;
780 while (j-- > idx)
782 if (reg_state[regno].reg_use[j].ruid >= ruid)
784 k--;
785 if (k != j)
786 reg_state[regno].reg_use[k] = reg_state[regno].reg_use[j];
789 reg_state[regno].use_index = k;
792 /* Find the use of REGNO with the ruid that is highest among those
793 lower than RUID_LIMIT, and return it if it is the only use of this
794 reg in the insn. Return NULL otherwise. */
796 static struct reg_use *
797 reload_combine_closest_single_use (unsigned regno, int ruid_limit)
799 int i, best_ruid = 0;
800 int use_idx = reg_state[regno].use_index;
801 struct reg_use *retval;
803 if (use_idx < 0)
804 return NULL;
805 retval = NULL;
806 for (i = use_idx; i < RELOAD_COMBINE_MAX_USES; i++)
808 struct reg_use *use = reg_state[regno].reg_use + i;
809 int this_ruid = use->ruid;
810 if (this_ruid >= ruid_limit)
811 continue;
812 if (this_ruid > best_ruid)
814 best_ruid = this_ruid;
815 retval = use;
817 else if (this_ruid == best_ruid)
818 retval = NULL;
820 if (last_label_ruid >= best_ruid)
821 return NULL;
822 return retval;
825 /* After we've moved an add insn, fix up any debug insns that occur
826 between the old location of the add and the new location. REG is
827 the destination register of the add insn; REPLACEMENT is the
828 SET_SRC of the add. FROM and TO specify the range in which we
829 should make this change on debug insns. */
831 static void
832 fixup_debug_insns (rtx reg, rtx replacement, rtx_insn *from, rtx_insn *to)
834 rtx_insn *insn;
835 for (insn = from; insn != to; insn = NEXT_INSN (insn))
837 rtx t;
839 if (!DEBUG_BIND_INSN_P (insn))
840 continue;
842 t = INSN_VAR_LOCATION_LOC (insn);
843 t = simplify_replace_rtx (t, reg, replacement);
844 validate_change (insn, &INSN_VAR_LOCATION_LOC (insn), t, 0);
848 /* Subroutine of reload_combine_recognize_const_pattern. Try to replace REG
849 with SRC in the insn described by USE, taking costs into account. Return
850 true if we made the replacement. */
852 static bool
853 try_replace_in_use (struct reg_use *use, rtx reg, rtx src)
855 rtx_insn *use_insn = use->insn;
856 rtx mem = use->containing_mem;
857 bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (use_insn));
859 if (mem != NULL_RTX)
861 addr_space_t as = MEM_ADDR_SPACE (mem);
862 rtx oldaddr = XEXP (mem, 0);
863 rtx newaddr = NULL_RTX;
864 int old_cost = address_cost (oldaddr, GET_MODE (mem), as, speed);
865 int new_cost;
867 newaddr = simplify_replace_rtx (oldaddr, reg, src);
868 if (memory_address_addr_space_p (GET_MODE (mem), newaddr, as))
870 XEXP (mem, 0) = newaddr;
871 new_cost = address_cost (newaddr, GET_MODE (mem), as, speed);
872 XEXP (mem, 0) = oldaddr;
873 if (new_cost <= old_cost
874 && validate_change (use_insn,
875 &XEXP (mem, 0), newaddr, 0))
876 return true;
879 else
881 rtx new_set = single_set (use_insn);
882 if (new_set
883 && REG_P (SET_DEST (new_set))
884 && GET_CODE (SET_SRC (new_set)) == PLUS
885 && REG_P (XEXP (SET_SRC (new_set), 0))
886 && CONSTANT_P (XEXP (SET_SRC (new_set), 1)))
888 rtx new_src;
889 machine_mode mode = GET_MODE (SET_DEST (new_set));
890 int old_cost = set_src_cost (SET_SRC (new_set), mode, speed);
892 gcc_assert (rtx_equal_p (XEXP (SET_SRC (new_set), 0), reg));
893 new_src = simplify_replace_rtx (SET_SRC (new_set), reg, src);
895 if (set_src_cost (new_src, mode, speed) <= old_cost
896 && validate_change (use_insn, &SET_SRC (new_set),
897 new_src, 0))
898 return true;
901 return false;
904 /* Called by reload_combine when scanning INSN. This function tries to detect
905 patterns where a constant is added to a register, and the result is used
906 in an address.
907 Return true if no further processing is needed on INSN; false if it wasn't
908 recognized and should be handled normally. */
910 static bool
911 reload_combine_recognize_const_pattern (rtx_insn *insn)
913 int from_ruid = reload_combine_ruid;
914 rtx set, pat, reg, src, addreg;
915 unsigned int regno;
916 struct reg_use *use;
917 bool must_move_add;
918 rtx_insn *add_moved_after_insn = NULL;
919 int add_moved_after_ruid = 0;
920 int clobbered_regno = -1;
922 set = single_set (insn);
923 if (set == NULL_RTX)
924 return false;
926 reg = SET_DEST (set);
927 src = SET_SRC (set);
928 if (!REG_P (reg)
929 || REG_NREGS (reg) != 1
930 || GET_MODE (reg) != Pmode
931 || reg == stack_pointer_rtx)
932 return false;
934 regno = REGNO (reg);
936 /* We look for a REG1 = REG2 + CONSTANT insn, followed by either
937 uses of REG1 inside an address, or inside another add insn. If
938 possible and profitable, merge the addition into subsequent
939 uses. */
940 if (GET_CODE (src) != PLUS
941 || !REG_P (XEXP (src, 0))
942 || !CONSTANT_P (XEXP (src, 1)))
943 return false;
945 addreg = XEXP (src, 0);
946 must_move_add = rtx_equal_p (reg, addreg);
948 pat = PATTERN (insn);
949 if (must_move_add && set != pat)
951 /* We have to be careful when moving the add; apart from the
952 single_set there may also be clobbers. Recognize one special
953 case, that of one clobber alongside the set (likely a clobber
954 of the CC register). */
955 gcc_assert (GET_CODE (PATTERN (insn)) == PARALLEL);
956 if (XVECLEN (pat, 0) != 2 || XVECEXP (pat, 0, 0) != set
957 || GET_CODE (XVECEXP (pat, 0, 1)) != CLOBBER
958 || !REG_P (XEXP (XVECEXP (pat, 0, 1), 0)))
959 return false;
960 clobbered_regno = REGNO (XEXP (XVECEXP (pat, 0, 1), 0));
965 use = reload_combine_closest_single_use (regno, from_ruid);
967 if (use)
968 /* Start the search for the next use from here. */
969 from_ruid = use->ruid;
971 if (use && GET_MODE (*use->usep) == Pmode)
973 bool delete_add = false;
974 rtx_insn *use_insn = use->insn;
975 int use_ruid = use->ruid;
977 /* Avoid moving the add insn past a jump. */
978 if (must_move_add && use_ruid <= last_jump_ruid)
979 break;
981 /* If the add clobbers another hard reg in parallel, don't move
982 it past a real set of this hard reg. */
983 if (must_move_add && clobbered_regno >= 0
984 && reg_state[clobbered_regno].real_store_ruid >= use_ruid)
985 break;
987 /* Do not separate cc0 setter and cc0 user on HAVE_cc0 targets. */
988 if (HAVE_cc0 && must_move_add && sets_cc0_p (PATTERN (use_insn)))
989 break;
991 gcc_assert (reg_state[regno].store_ruid <= use_ruid);
992 /* Avoid moving a use of ADDREG past a point where it is stored. */
993 if (reg_state[REGNO (addreg)].store_ruid > use_ruid)
994 break;
996 /* We also must not move the addition past an insn that sets
997 the same register, unless we can combine two add insns. */
998 if (must_move_add && reg_state[regno].store_ruid == use_ruid)
1000 if (use->containing_mem == NULL_RTX)
1001 delete_add = true;
1002 else
1003 break;
1006 if (try_replace_in_use (use, reg, src))
1008 reload_combine_purge_insn_uses (use_insn);
1009 reload_combine_note_use (&PATTERN (use_insn), use_insn,
1010 use_ruid, NULL_RTX);
1012 if (delete_add)
1014 fixup_debug_insns (reg, src, insn, use_insn);
1015 delete_insn (insn);
1016 return true;
1018 if (must_move_add)
1020 add_moved_after_insn = use_insn;
1021 add_moved_after_ruid = use_ruid;
1023 continue;
1026 /* If we get here, we couldn't handle this use. */
1027 if (must_move_add)
1028 break;
1030 while (use);
1032 if (!must_move_add || add_moved_after_insn == NULL_RTX)
1033 /* Process the add normally. */
1034 return false;
1036 fixup_debug_insns (reg, src, insn, add_moved_after_insn);
1038 reorder_insns (insn, insn, add_moved_after_insn);
1039 reload_combine_purge_reg_uses_after_ruid (regno, add_moved_after_ruid);
1040 reload_combine_split_ruids (add_moved_after_ruid - 1);
1041 reload_combine_note_use (&PATTERN (insn), insn,
1042 add_moved_after_ruid, NULL_RTX);
1043 reg_state[regno].store_ruid = add_moved_after_ruid;
1045 return true;
1048 /* Called by reload_combine when scanning INSN. Try to detect a pattern we
1049 can handle and improve. Return true if no further processing is needed on
1050 INSN; false if it wasn't recognized and should be handled normally. */
1052 static bool
1053 reload_combine_recognize_pattern (rtx_insn *insn)
1055 rtx set, reg, src;
1057 set = single_set (insn);
1058 if (set == NULL_RTX)
1059 return false;
1061 reg = SET_DEST (set);
1062 src = SET_SRC (set);
1063 if (!REG_P (reg) || REG_NREGS (reg) != 1)
1064 return false;
1066 unsigned int regno = REGNO (reg);
1067 machine_mode mode = GET_MODE (reg);
1069 if (reg_state[regno].use_index < 0
1070 || reg_state[regno].use_index >= RELOAD_COMBINE_MAX_USES)
1071 return false;
1073 for (int i = reg_state[regno].use_index;
1074 i < RELOAD_COMBINE_MAX_USES; i++)
1076 struct reg_use *use = reg_state[regno].reg_use + i;
1077 if (GET_MODE (*use->usep) != mode)
1078 return false;
1081 /* Look for (set (REGX) (CONST_INT))
1082 (set (REGX) (PLUS (REGX) (REGY)))
1084 ... (MEM (REGX)) ...
1085 and convert it to
1086 (set (REGZ) (CONST_INT))
1088 ... (MEM (PLUS (REGZ) (REGY)))... .
1090 First, check that we have (set (REGX) (PLUS (REGX) (REGY)))
1091 and that we know all uses of REGX before it dies.
1092 Also, explicitly check that REGX != REGY; our life information
1093 does not yet show whether REGY changes in this insn. */
1095 if (GET_CODE (src) == PLUS
1096 && reg_state[regno].all_offsets_match
1097 && last_index_reg != -1
1098 && REG_P (XEXP (src, 1))
1099 && rtx_equal_p (XEXP (src, 0), reg)
1100 && !rtx_equal_p (XEXP (src, 1), reg)
1101 && last_label_ruid < reg_state[regno].use_ruid)
1103 rtx base = XEXP (src, 1);
1104 rtx_insn *prev = prev_nonnote_nondebug_insn (insn);
1105 rtx prev_set = prev ? single_set (prev) : NULL_RTX;
1106 rtx index_reg = NULL_RTX;
1107 rtx reg_sum = NULL_RTX;
1108 int i;
1110 /* Now we need to set INDEX_REG to an index register (denoted as
1111 REGZ in the illustration above) and REG_SUM to the expression
1112 register+register that we want to use to substitute uses of REG
1113 (typically in MEMs) with. First check REG and BASE for being
1114 index registers; we can use them even if they are not dead. */
1115 if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], regno)
1116 || TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS],
1117 REGNO (base)))
1119 index_reg = reg;
1120 reg_sum = src;
1122 else
1124 /* Otherwise, look for a free index register. Since we have
1125 checked above that neither REG nor BASE are index registers,
1126 if we find anything at all, it will be different from these
1127 two registers. */
1128 for (i = first_index_reg; i <= last_index_reg; i++)
1130 if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], i)
1131 && reg_state[i].use_index == RELOAD_COMBINE_MAX_USES
1132 && reg_state[i].store_ruid <= reg_state[regno].use_ruid
1133 && (call_used_regs[i] || df_regs_ever_live_p (i))
1134 && (!frame_pointer_needed || i != HARD_FRAME_POINTER_REGNUM)
1135 && !fixed_regs[i] && !global_regs[i]
1136 && hard_regno_nregs (i, GET_MODE (reg)) == 1
1137 && targetm.hard_regno_scratch_ok (i))
1139 index_reg = gen_rtx_REG (GET_MODE (reg), i);
1140 reg_sum = gen_rtx_PLUS (GET_MODE (reg), index_reg, base);
1141 break;
1146 /* Check that PREV_SET is indeed (set (REGX) (CONST_INT)) and that
1147 (REGY), i.e. BASE, is not clobbered before the last use we'll
1148 create. */
1149 if (reg_sum
1150 && prev_set
1151 && CONST_INT_P (SET_SRC (prev_set))
1152 && rtx_equal_p (SET_DEST (prev_set), reg)
1153 && (reg_state[REGNO (base)].store_ruid
1154 <= reg_state[regno].use_ruid))
1156 /* Change destination register and, if necessary, the constant
1157 value in PREV, the constant loading instruction. */
1158 validate_change (prev, &SET_DEST (prev_set), index_reg, 1);
1159 if (reg_state[regno].offset != const0_rtx)
1161 HOST_WIDE_INT c
1162 = trunc_int_for_mode (UINTVAL (SET_SRC (prev_set))
1163 + UINTVAL (reg_state[regno].offset),
1164 GET_MODE (index_reg));
1165 validate_change (prev, &SET_SRC (prev_set), GEN_INT (c), 1);
1168 /* Now for every use of REG that we have recorded, replace REG
1169 with REG_SUM. */
1170 for (i = reg_state[regno].use_index;
1171 i < RELOAD_COMBINE_MAX_USES; i++)
1172 validate_unshare_change (reg_state[regno].reg_use[i].insn,
1173 reg_state[regno].reg_use[i].usep,
1174 /* Each change must have its own
1175 replacement. */
1176 reg_sum, 1);
1178 if (apply_change_group ())
1180 struct reg_use *lowest_ruid = NULL;
1182 /* For every new use of REG_SUM, we have to record the use
1183 of BASE therein, i.e. operand 1. */
1184 for (i = reg_state[regno].use_index;
1185 i < RELOAD_COMBINE_MAX_USES; i++)
1187 struct reg_use *use = reg_state[regno].reg_use + i;
1188 reload_combine_note_use (&XEXP (*use->usep, 1), use->insn,
1189 use->ruid, use->containing_mem);
1190 if (lowest_ruid == NULL || use->ruid < lowest_ruid->ruid)
1191 lowest_ruid = use;
1194 fixup_debug_insns (reg, reg_sum, insn, lowest_ruid->insn);
1196 /* Delete the reg-reg addition. */
1197 delete_insn (insn);
1199 if (reg_state[regno].offset != const0_rtx
1200 /* Previous REG_EQUIV / REG_EQUAL notes for PREV
1201 are now invalid. */
1202 && remove_reg_equal_equiv_notes (prev))
1203 df_notes_rescan (prev);
1205 reg_state[regno].use_index = RELOAD_COMBINE_MAX_USES;
1206 return true;
1210 return false;
1213 static void
1214 reload_combine (void)
1216 rtx_insn *insn, *prev;
1217 basic_block bb;
1218 unsigned int r;
1219 int min_labelno, n_labels;
1220 HARD_REG_SET ever_live_at_start, *label_live;
1222 /* To avoid wasting too much time later searching for an index register,
1223 determine the minimum and maximum index register numbers. */
1224 if (INDEX_REG_CLASS == NO_REGS)
1225 last_index_reg = -1;
1226 else if (first_index_reg == -1 && last_index_reg == 0)
1228 for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
1229 if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], r))
1231 if (first_index_reg == -1)
1232 first_index_reg = r;
1234 last_index_reg = r;
1237 /* If no index register is available, we can quit now. Set LAST_INDEX_REG
1238 to -1 so we'll know to quit early the next time we get here. */
1239 if (first_index_reg == -1)
1241 last_index_reg = -1;
1242 return;
1246 /* Set up LABEL_LIVE and EVER_LIVE_AT_START. The register lifetime
1247 information is a bit fuzzy immediately after reload, but it's
1248 still good enough to determine which registers are live at a jump
1249 destination. */
1250 min_labelno = get_first_label_num ();
1251 n_labels = max_label_num () - min_labelno;
1252 label_live = XNEWVEC (HARD_REG_SET, n_labels);
1253 CLEAR_HARD_REG_SET (ever_live_at_start);
1255 FOR_EACH_BB_REVERSE_FN (bb, cfun)
1257 insn = BB_HEAD (bb);
1258 if (LABEL_P (insn))
1260 HARD_REG_SET live;
1261 bitmap live_in = df_get_live_in (bb);
1263 REG_SET_TO_HARD_REG_SET (live, live_in);
1264 compute_use_by_pseudos (&live, live_in);
1265 COPY_HARD_REG_SET (LABEL_LIVE (insn), live);
1266 IOR_HARD_REG_SET (ever_live_at_start, live);
1270 /* Initialize last_label_ruid, reload_combine_ruid and reg_state. */
1271 last_label_ruid = last_jump_ruid = reload_combine_ruid = 0;
1272 for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
1274 reg_state[r].store_ruid = 0;
1275 reg_state[r].real_store_ruid = 0;
1276 if (fixed_regs[r])
1277 reg_state[r].use_index = -1;
1278 else
1279 reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
1282 for (insn = get_last_insn (); insn; insn = prev)
1284 bool control_flow_insn;
1285 rtx note;
1287 prev = PREV_INSN (insn);
1289 /* We cannot do our optimization across labels. Invalidating all the use
1290 information we have would be costly, so we just note where the label
1291 is and then later disable any optimization that would cross it. */
1292 if (LABEL_P (insn))
1293 last_label_ruid = reload_combine_ruid;
1294 else if (BARRIER_P (insn))
1296 /* Crossing a barrier resets all the use information. */
1297 for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
1298 if (! fixed_regs[r])
1299 reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
1301 else if (INSN_P (insn) && volatile_insn_p (PATTERN (insn)))
1302 /* Optimizations across insns being marked as volatile must be
1303 prevented. All the usage information is invalidated
1304 here. */
1305 for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
1306 if (! fixed_regs[r]
1307 && reg_state[r].use_index != RELOAD_COMBINE_MAX_USES)
1308 reg_state[r].use_index = -1;
1310 if (! NONDEBUG_INSN_P (insn))
1311 continue;
1313 reload_combine_ruid++;
1315 control_flow_insn = control_flow_insn_p (insn);
1316 if (control_flow_insn)
1317 last_jump_ruid = reload_combine_ruid;
1319 if (reload_combine_recognize_const_pattern (insn)
1320 || reload_combine_recognize_pattern (insn))
1321 continue;
1323 note_stores (PATTERN (insn), reload_combine_note_store, NULL);
1325 if (CALL_P (insn))
1327 rtx link;
1328 HARD_REG_SET used_regs;
1330 get_call_reg_set_usage (insn, &used_regs, call_used_reg_set);
1332 for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
1333 if (TEST_HARD_REG_BIT (used_regs, r))
1335 reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
1336 reg_state[r].store_ruid = reload_combine_ruid;
1339 for (link = CALL_INSN_FUNCTION_USAGE (insn); link;
1340 link = XEXP (link, 1))
1342 rtx setuse = XEXP (link, 0);
1343 rtx usage_rtx = XEXP (setuse, 0);
1344 if ((GET_CODE (setuse) == USE || GET_CODE (setuse) == CLOBBER)
1345 && REG_P (usage_rtx))
1347 unsigned int end_regno = END_REGNO (usage_rtx);
1348 for (unsigned int i = REGNO (usage_rtx); i < end_regno; ++i)
1349 if (GET_CODE (XEXP (link, 0)) == CLOBBER)
1351 reg_state[i].use_index = RELOAD_COMBINE_MAX_USES;
1352 reg_state[i].store_ruid = reload_combine_ruid;
1354 else
1355 reg_state[i].use_index = -1;
1360 if (control_flow_insn && !ANY_RETURN_P (PATTERN (insn)))
1362 /* Non-spill registers might be used at the call destination in
1363 some unknown fashion, so we have to mark the unknown use. */
1364 HARD_REG_SET *live;
1366 if ((condjump_p (insn) || condjump_in_parallel_p (insn))
1367 && JUMP_LABEL (insn))
1369 if (ANY_RETURN_P (JUMP_LABEL (insn)))
1370 live = NULL;
1371 else
1372 live = &LABEL_LIVE (JUMP_LABEL (insn));
1374 else
1375 live = &ever_live_at_start;
1377 if (live)
1378 for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
1379 if (TEST_HARD_REG_BIT (*live, r))
1380 reg_state[r].use_index = -1;
1383 reload_combine_note_use (&PATTERN (insn), insn, reload_combine_ruid,
1384 NULL_RTX);
1386 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
1388 if (REG_NOTE_KIND (note) == REG_INC && REG_P (XEXP (note, 0)))
1390 int regno = REGNO (XEXP (note, 0));
1391 reg_state[regno].store_ruid = reload_combine_ruid;
1392 reg_state[regno].real_store_ruid = reload_combine_ruid;
1393 reg_state[regno].use_index = -1;
1398 free (label_live);
1401 /* Check if DST is a register or a subreg of a register; if it is,
1402 update store_ruid, real_store_ruid and use_index in the reg_state
1403 structure accordingly. Called via note_stores from reload_combine. */
1405 static void
1406 reload_combine_note_store (rtx dst, const_rtx set, void *data ATTRIBUTE_UNUSED)
1408 int regno = 0;
1409 int i;
1410 machine_mode mode = GET_MODE (dst);
1412 if (GET_CODE (dst) == SUBREG)
1414 regno = subreg_regno_offset (REGNO (SUBREG_REG (dst)),
1415 GET_MODE (SUBREG_REG (dst)),
1416 SUBREG_BYTE (dst),
1417 GET_MODE (dst));
1418 dst = SUBREG_REG (dst);
1421 /* Some targets do argument pushes without adding REG_INC notes. */
1423 if (MEM_P (dst))
1425 dst = XEXP (dst, 0);
1426 if (GET_CODE (dst) == PRE_INC || GET_CODE (dst) == POST_INC
1427 || GET_CODE (dst) == PRE_DEC || GET_CODE (dst) == POST_DEC
1428 || GET_CODE (dst) == PRE_MODIFY || GET_CODE (dst) == POST_MODIFY)
1430 unsigned int end_regno = END_REGNO (XEXP (dst, 0));
1431 for (unsigned int i = REGNO (XEXP (dst, 0)); i < end_regno; ++i)
1433 /* We could probably do better, but for now mark the register
1434 as used in an unknown fashion and set/clobbered at this
1435 insn. */
1436 reg_state[i].use_index = -1;
1437 reg_state[i].store_ruid = reload_combine_ruid;
1438 reg_state[i].real_store_ruid = reload_combine_ruid;
1441 else
1442 return;
1445 if (!REG_P (dst))
1446 return;
1447 regno += REGNO (dst);
1449 /* note_stores might have stripped a STRICT_LOW_PART, so we have to be
1450 careful with registers / register parts that are not full words.
1451 Similarly for ZERO_EXTRACT. */
1452 if (GET_CODE (SET_DEST (set)) == ZERO_EXTRACT
1453 || GET_CODE (SET_DEST (set)) == STRICT_LOW_PART)
1455 for (i = end_hard_regno (mode, regno) - 1; i >= regno; i--)
1457 reg_state[i].use_index = -1;
1458 reg_state[i].store_ruid = reload_combine_ruid;
1459 reg_state[i].real_store_ruid = reload_combine_ruid;
1462 else
1464 for (i = end_hard_regno (mode, regno) - 1; i >= regno; i--)
1466 reg_state[i].store_ruid = reload_combine_ruid;
1467 if (GET_CODE (set) == SET)
1468 reg_state[i].real_store_ruid = reload_combine_ruid;
1469 reg_state[i].use_index = RELOAD_COMBINE_MAX_USES;
1474 /* XP points to a piece of rtl that has to be checked for any uses of
1475 registers.
1476 *XP is the pattern of INSN, or a part of it.
1477 Called from reload_combine, and recursively by itself. */
1478 static void
1479 reload_combine_note_use (rtx *xp, rtx_insn *insn, int ruid, rtx containing_mem)
1481 rtx x = *xp;
1482 enum rtx_code code = x->code;
1483 const char *fmt;
1484 int i, j;
1485 rtx offset = const0_rtx; /* For the REG case below. */
1487 switch (code)
1489 case SET:
1490 if (REG_P (SET_DEST (x)))
1492 reload_combine_note_use (&SET_SRC (x), insn, ruid, NULL_RTX);
1493 return;
1495 break;
1497 case USE:
1498 /* If this is the USE of a return value, we can't change it. */
1499 if (REG_P (XEXP (x, 0)) && REG_FUNCTION_VALUE_P (XEXP (x, 0)))
1501 /* Mark the return register as used in an unknown fashion. */
1502 rtx reg = XEXP (x, 0);
1503 unsigned int end_regno = END_REGNO (reg);
1504 for (unsigned int regno = REGNO (reg); regno < end_regno; ++regno)
1505 reg_state[regno].use_index = -1;
1506 return;
1508 break;
1510 case CLOBBER:
1511 if (REG_P (SET_DEST (x)))
1513 /* No spurious CLOBBERs of pseudo registers may remain. */
1514 gcc_assert (REGNO (SET_DEST (x)) < FIRST_PSEUDO_REGISTER);
1515 return;
1517 break;
1519 case PLUS:
1520 /* We are interested in (plus (reg) (const_int)) . */
1521 if (!REG_P (XEXP (x, 0))
1522 || !CONST_INT_P (XEXP (x, 1)))
1523 break;
1524 offset = XEXP (x, 1);
1525 x = XEXP (x, 0);
1526 /* Fall through. */
1527 case REG:
1529 int regno = REGNO (x);
1530 int use_index;
1531 int nregs;
1533 /* No spurious USEs of pseudo registers may remain. */
1534 gcc_assert (regno < FIRST_PSEUDO_REGISTER);
1536 nregs = REG_NREGS (x);
1538 /* We can't substitute into multi-hard-reg uses. */
1539 if (nregs > 1)
1541 while (--nregs >= 0)
1542 reg_state[regno + nregs].use_index = -1;
1543 return;
1546 /* We may be called to update uses in previously seen insns.
1547 Don't add uses beyond the last store we saw. */
1548 if (ruid < reg_state[regno].store_ruid)
1549 return;
1551 /* If this register is already used in some unknown fashion, we
1552 can't do anything.
1553 If we decrement the index from zero to -1, we can't store more
1554 uses, so this register becomes used in an unknown fashion. */
1555 use_index = --reg_state[regno].use_index;
1556 if (use_index < 0)
1557 return;
1559 if (use_index == RELOAD_COMBINE_MAX_USES - 1)
1561 /* This is the first use of this register we have seen since we
1562 marked it as dead. */
1563 reg_state[regno].offset = offset;
1564 reg_state[regno].all_offsets_match = true;
1565 reg_state[regno].use_ruid = ruid;
1567 else
1569 if (reg_state[regno].use_ruid > ruid)
1570 reg_state[regno].use_ruid = ruid;
1572 if (! rtx_equal_p (offset, reg_state[regno].offset))
1573 reg_state[regno].all_offsets_match = false;
1576 reg_state[regno].reg_use[use_index].insn = insn;
1577 reg_state[regno].reg_use[use_index].ruid = ruid;
1578 reg_state[regno].reg_use[use_index].containing_mem = containing_mem;
1579 reg_state[regno].reg_use[use_index].usep = xp;
1580 return;
1583 case MEM:
1584 containing_mem = x;
1585 break;
1587 default:
1588 break;
1591 /* Recursively process the components of X. */
1592 fmt = GET_RTX_FORMAT (code);
1593 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1595 if (fmt[i] == 'e')
1596 reload_combine_note_use (&XEXP (x, i), insn, ruid, containing_mem);
1597 else if (fmt[i] == 'E')
1599 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1600 reload_combine_note_use (&XVECEXP (x, i, j), insn, ruid,
1601 containing_mem);
1606 /* See if we can reduce the cost of a constant by replacing a move
1607 with an add. We track situations in which a register is set to a
1608 constant or to a register plus a constant. */
1609 /* We cannot do our optimization across labels. Invalidating all the
1610 information about register contents we have would be costly, so we
1611 use move2add_last_label_luid to note where the label is and then
1612 later disable any optimization that would cross it.
1613 reg_offset[n] / reg_base_reg[n] / reg_symbol_ref[n] / reg_mode[n]
1614 are only valid if reg_set_luid[n] is greater than
1615 move2add_last_label_luid.
1616 For a set that established a new (potential) base register with
1617 non-constant value, we use move2add_luid from the place where the
1618 setting insn is encountered; registers based off that base then
1619 get the same reg_set_luid. Constants all get
1620 move2add_last_label_luid + 1 as their reg_set_luid. */
1621 static int reg_set_luid[FIRST_PSEUDO_REGISTER];
1623 /* If reg_base_reg[n] is negative, register n has been set to
1624 reg_offset[n] or reg_symbol_ref[n] + reg_offset[n] in mode reg_mode[n].
1625 If reg_base_reg[n] is non-negative, register n has been set to the
1626 sum of reg_offset[n] and the value of register reg_base_reg[n]
1627 before reg_set_luid[n], calculated in mode reg_mode[n] .
1628 For multi-hard-register registers, all but the first one are
1629 recorded as BLKmode in reg_mode. Setting reg_mode to VOIDmode
1630 marks it as invalid. */
1631 static HOST_WIDE_INT reg_offset[FIRST_PSEUDO_REGISTER];
1632 static int reg_base_reg[FIRST_PSEUDO_REGISTER];
1633 static rtx reg_symbol_ref[FIRST_PSEUDO_REGISTER];
1634 static machine_mode reg_mode[FIRST_PSEUDO_REGISTER];
1636 /* move2add_luid is linearly increased while scanning the instructions
1637 from first to last. It is used to set reg_set_luid in
1638 reload_cse_move2add and move2add_note_store. */
1639 static int move2add_luid;
1641 /* move2add_last_label_luid is set whenever a label is found. Labels
1642 invalidate all previously collected reg_offset data. */
1643 static int move2add_last_label_luid;
1645 /* ??? We don't know how zero / sign extension is handled, hence we
1646 can't go from a narrower to a wider mode. */
1647 #define MODES_OK_FOR_MOVE2ADD(OUTMODE, INMODE) \
1648 (GET_MODE_SIZE (OUTMODE) == GET_MODE_SIZE (INMODE) \
1649 || (GET_MODE_SIZE (OUTMODE) <= GET_MODE_SIZE (INMODE) \
1650 && TRULY_NOOP_TRUNCATION_MODES_P (OUTMODE, INMODE)))
1652 /* Record that REG is being set to a value with the mode of REG. */
1654 static void
1655 move2add_record_mode (rtx reg)
1657 int regno, nregs;
1658 machine_mode mode = GET_MODE (reg);
1660 if (GET_CODE (reg) == SUBREG)
1662 regno = subreg_regno (reg);
1663 nregs = subreg_nregs (reg);
1665 else if (REG_P (reg))
1667 regno = REGNO (reg);
1668 nregs = REG_NREGS (reg);
1670 else
1671 gcc_unreachable ();
1672 for (int i = nregs - 1; i > 0; i--)
1673 reg_mode[regno + i] = BLKmode;
1674 reg_mode[regno] = mode;
1677 /* Record that REG is being set to the sum of SYM and OFF. */
1679 static void
1680 move2add_record_sym_value (rtx reg, rtx sym, rtx off)
1682 int regno = REGNO (reg);
1684 move2add_record_mode (reg);
1685 reg_set_luid[regno] = move2add_luid;
1686 reg_base_reg[regno] = -1;
1687 reg_symbol_ref[regno] = sym;
1688 reg_offset[regno] = INTVAL (off);
1691 /* Check if REGNO contains a valid value in MODE. */
1693 static bool
1694 move2add_valid_value_p (int regno, scalar_int_mode mode)
1696 if (reg_set_luid[regno] <= move2add_last_label_luid)
1697 return false;
1699 if (mode != reg_mode[regno])
1701 scalar_int_mode old_mode;
1702 if (!is_a <scalar_int_mode> (reg_mode[regno], &old_mode)
1703 || !MODES_OK_FOR_MOVE2ADD (mode, old_mode))
1704 return false;
1705 /* The value loaded into regno in reg_mode[regno] is also valid in
1706 mode after truncation only if (REG:mode regno) is the lowpart of
1707 (REG:reg_mode[regno] regno). Now, for big endian, the starting
1708 regno of the lowpart might be different. */
1709 poly_int64 s_off = subreg_lowpart_offset (mode, old_mode);
1710 s_off = subreg_regno_offset (regno, old_mode, s_off, mode);
1711 if (maybe_ne (s_off, 0))
1712 /* We could in principle adjust regno, check reg_mode[regno] to be
1713 BLKmode, and return s_off to the caller (vs. -1 for failure),
1714 but we currently have no callers that could make use of this
1715 information. */
1716 return false;
1719 for (int i = end_hard_regno (mode, regno) - 1; i > regno; i--)
1720 if (reg_mode[i] != BLKmode)
1721 return false;
1722 return true;
1725 /* This function is called with INSN that sets REG (of mode MODE)
1726 to (SYM + OFF), while REG is known to already have value (SYM + offset).
1727 This function tries to change INSN into an add instruction
1728 (set (REG) (plus (REG) (OFF - offset))) using the known value.
1729 It also updates the information about REG's known value.
1730 Return true if we made a change. */
1732 static bool
1733 move2add_use_add2_insn (scalar_int_mode mode, rtx reg, rtx sym, rtx off,
1734 rtx_insn *insn)
1736 rtx pat = PATTERN (insn);
1737 rtx src = SET_SRC (pat);
1738 int regno = REGNO (reg);
1739 rtx new_src = gen_int_mode (UINTVAL (off) - reg_offset[regno], mode);
1740 bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
1741 bool changed = false;
1743 /* (set (reg) (plus (reg) (const_int 0))) is not canonical;
1744 use (set (reg) (reg)) instead.
1745 We don't delete this insn, nor do we convert it into a
1746 note, to avoid losing register notes or the return
1747 value flag. jump2 already knows how to get rid of
1748 no-op moves. */
1749 if (new_src == const0_rtx)
1751 /* If the constants are different, this is a
1752 truncation, that, if turned into (set (reg)
1753 (reg)), would be discarded. Maybe we should
1754 try a truncMN pattern? */
1755 if (INTVAL (off) == reg_offset [regno])
1756 changed = validate_change (insn, &SET_SRC (pat), reg, 0);
1758 else
1760 struct full_rtx_costs oldcst, newcst;
1761 rtx tem = gen_rtx_PLUS (mode, reg, new_src);
1763 get_full_set_rtx_cost (pat, &oldcst);
1764 SET_SRC (pat) = tem;
1765 get_full_set_rtx_cost (pat, &newcst);
1766 SET_SRC (pat) = src;
1768 if (costs_lt_p (&newcst, &oldcst, speed)
1769 && have_add2_insn (reg, new_src))
1770 changed = validate_change (insn, &SET_SRC (pat), tem, 0);
1771 else if (sym == NULL_RTX && mode != BImode)
1773 scalar_int_mode narrow_mode;
1774 FOR_EACH_MODE_UNTIL (narrow_mode, mode)
1776 if (have_insn_for (STRICT_LOW_PART, narrow_mode)
1777 && ((reg_offset[regno] & ~GET_MODE_MASK (narrow_mode))
1778 == (INTVAL (off) & ~GET_MODE_MASK (narrow_mode))))
1780 rtx narrow_reg = gen_lowpart_common (narrow_mode, reg);
1781 rtx narrow_src = gen_int_mode (INTVAL (off),
1782 narrow_mode);
1783 rtx new_set
1784 = gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode,
1785 narrow_reg),
1786 narrow_src);
1787 get_full_set_rtx_cost (new_set, &newcst);
1788 if (costs_lt_p (&newcst, &oldcst, speed))
1790 changed = validate_change (insn, &PATTERN (insn),
1791 new_set, 0);
1792 if (changed)
1793 break;
1799 move2add_record_sym_value (reg, sym, off);
1800 return changed;
1804 /* This function is called with INSN that sets REG (of mode MODE) to
1805 (SYM + OFF), but REG doesn't have known value (SYM + offset). This
1806 function tries to find another register which is known to already have
1807 value (SYM + offset) and change INSN into an add instruction
1808 (set (REG) (plus (the found register) (OFF - offset))) if such
1809 a register is found. It also updates the information about
1810 REG's known value.
1811 Return true iff we made a change. */
1813 static bool
1814 move2add_use_add3_insn (scalar_int_mode mode, rtx reg, rtx sym, rtx off,
1815 rtx_insn *insn)
1817 rtx pat = PATTERN (insn);
1818 rtx src = SET_SRC (pat);
1819 int regno = REGNO (reg);
1820 int min_regno = 0;
1821 bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
1822 int i;
1823 bool changed = false;
1824 struct full_rtx_costs oldcst, newcst, mincst;
1825 rtx plus_expr;
1827 init_costs_to_max (&mincst);
1828 get_full_set_rtx_cost (pat, &oldcst);
1830 plus_expr = gen_rtx_PLUS (GET_MODE (reg), reg, const0_rtx);
1831 SET_SRC (pat) = plus_expr;
1833 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1834 if (move2add_valid_value_p (i, mode)
1835 && reg_base_reg[i] < 0
1836 && reg_symbol_ref[i] != NULL_RTX
1837 && rtx_equal_p (sym, reg_symbol_ref[i]))
1839 rtx new_src = gen_int_mode (UINTVAL (off) - reg_offset[i],
1840 GET_MODE (reg));
1841 /* (set (reg) (plus (reg) (const_int 0))) is not canonical;
1842 use (set (reg) (reg)) instead.
1843 We don't delete this insn, nor do we convert it into a
1844 note, to avoid losing register notes or the return
1845 value flag. jump2 already knows how to get rid of
1846 no-op moves. */
1847 if (new_src == const0_rtx)
1849 init_costs_to_zero (&mincst);
1850 min_regno = i;
1851 break;
1853 else
1855 XEXP (plus_expr, 1) = new_src;
1856 get_full_set_rtx_cost (pat, &newcst);
1858 if (costs_lt_p (&newcst, &mincst, speed))
1860 mincst = newcst;
1861 min_regno = i;
1865 SET_SRC (pat) = src;
1867 if (costs_lt_p (&mincst, &oldcst, speed))
1869 rtx tem;
1871 tem = gen_rtx_REG (GET_MODE (reg), min_regno);
1872 if (i != min_regno)
1874 rtx new_src = gen_int_mode (UINTVAL (off) - reg_offset[min_regno],
1875 GET_MODE (reg));
1876 tem = gen_rtx_PLUS (GET_MODE (reg), tem, new_src);
1878 if (validate_change (insn, &SET_SRC (pat), tem, 0))
1879 changed = true;
1881 reg_set_luid[regno] = move2add_luid;
1882 move2add_record_sym_value (reg, sym, off);
1883 return changed;
1886 /* Convert move insns with constant inputs to additions if they are cheaper.
1887 Return true if any changes were made. */
1888 static bool
1889 reload_cse_move2add (rtx_insn *first)
1891 int i;
1892 rtx_insn *insn;
1893 bool changed = false;
1895 for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--)
1897 reg_set_luid[i] = 0;
1898 reg_offset[i] = 0;
1899 reg_base_reg[i] = 0;
1900 reg_symbol_ref[i] = NULL_RTX;
1901 reg_mode[i] = VOIDmode;
1904 move2add_last_label_luid = 0;
1905 move2add_luid = 2;
1906 for (insn = first; insn; insn = NEXT_INSN (insn), move2add_luid++)
1908 rtx pat, note;
1910 if (LABEL_P (insn))
1912 move2add_last_label_luid = move2add_luid;
1913 /* We're going to increment move2add_luid twice after a
1914 label, so that we can use move2add_last_label_luid + 1 as
1915 the luid for constants. */
1916 move2add_luid++;
1917 continue;
1919 if (! INSN_P (insn))
1920 continue;
1921 pat = PATTERN (insn);
1922 /* For simplicity, we only perform this optimization on
1923 straightforward SETs. */
1924 scalar_int_mode mode;
1925 if (GET_CODE (pat) == SET
1926 && REG_P (SET_DEST (pat))
1927 && is_a <scalar_int_mode> (GET_MODE (SET_DEST (pat)), &mode))
1929 rtx reg = SET_DEST (pat);
1930 int regno = REGNO (reg);
1931 rtx src = SET_SRC (pat);
1933 /* Check if we have valid information on the contents of this
1934 register in the mode of REG. */
1935 if (move2add_valid_value_p (regno, mode)
1936 && dbg_cnt (cse2_move2add))
1938 /* Try to transform (set (REGX) (CONST_INT A))
1940 (set (REGX) (CONST_INT B))
1942 (set (REGX) (CONST_INT A))
1944 (set (REGX) (plus (REGX) (CONST_INT B-A)))
1946 (set (REGX) (CONST_INT A))
1948 (set (STRICT_LOW_PART (REGX)) (CONST_INT B))
1951 if (CONST_INT_P (src)
1952 && reg_base_reg[regno] < 0
1953 && reg_symbol_ref[regno] == NULL_RTX)
1955 changed |= move2add_use_add2_insn (mode, reg, NULL_RTX,
1956 src, insn);
1957 continue;
1960 /* Try to transform (set (REGX) (REGY))
1961 (set (REGX) (PLUS (REGX) (CONST_INT A)))
1963 (set (REGX) (REGY))
1964 (set (REGX) (PLUS (REGX) (CONST_INT B)))
1966 (set (REGX) (REGY))
1967 (set (REGX) (PLUS (REGX) (CONST_INT A)))
1969 (set (REGX) (plus (REGX) (CONST_INT B-A))) */
1970 else if (REG_P (src)
1971 && reg_set_luid[regno] == reg_set_luid[REGNO (src)]
1972 && reg_base_reg[regno] == reg_base_reg[REGNO (src)]
1973 && move2add_valid_value_p (REGNO (src), mode))
1975 rtx_insn *next = next_nonnote_nondebug_insn (insn);
1976 rtx set = NULL_RTX;
1977 if (next)
1978 set = single_set (next);
1979 if (set
1980 && SET_DEST (set) == reg
1981 && GET_CODE (SET_SRC (set)) == PLUS
1982 && XEXP (SET_SRC (set), 0) == reg
1983 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
1985 rtx src3 = XEXP (SET_SRC (set), 1);
1986 unsigned HOST_WIDE_INT added_offset = UINTVAL (src3);
1987 HOST_WIDE_INT base_offset = reg_offset[REGNO (src)];
1988 HOST_WIDE_INT regno_offset = reg_offset[regno];
1989 rtx new_src =
1990 gen_int_mode (added_offset
1991 + base_offset
1992 - regno_offset,
1993 mode);
1994 bool success = false;
1995 bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn));
1997 if (new_src == const0_rtx)
1998 /* See above why we create (set (reg) (reg)) here. */
1999 success
2000 = validate_change (next, &SET_SRC (set), reg, 0);
2001 else
2003 rtx old_src = SET_SRC (set);
2004 struct full_rtx_costs oldcst, newcst;
2005 rtx tem = gen_rtx_PLUS (mode, reg, new_src);
2007 get_full_set_rtx_cost (set, &oldcst);
2008 SET_SRC (set) = tem;
2009 get_full_set_src_cost (tem, mode, &newcst);
2010 SET_SRC (set) = old_src;
2011 costs_add_n_insns (&oldcst, 1);
2013 if (costs_lt_p (&newcst, &oldcst, speed)
2014 && have_add2_insn (reg, new_src))
2016 rtx newpat = gen_rtx_SET (reg, tem);
2017 success
2018 = validate_change (next, &PATTERN (next),
2019 newpat, 0);
2022 if (success)
2023 delete_insn (insn);
2024 changed |= success;
2025 insn = next;
2026 move2add_record_mode (reg);
2027 reg_offset[regno]
2028 = trunc_int_for_mode (added_offset + base_offset,
2029 mode);
2030 continue;
2035 /* Try to transform
2036 (set (REGX) (CONST (PLUS (SYMBOL_REF) (CONST_INT A))))
2038 (set (REGY) (CONST (PLUS (SYMBOL_REF) (CONST_INT B))))
2040 (set (REGX) (CONST (PLUS (SYMBOL_REF) (CONST_INT A))))
2042 (set (REGY) (CONST (PLUS (REGX) (CONST_INT B-A)))) */
2043 if ((GET_CODE (src) == SYMBOL_REF
2044 || (GET_CODE (src) == CONST
2045 && GET_CODE (XEXP (src, 0)) == PLUS
2046 && GET_CODE (XEXP (XEXP (src, 0), 0)) == SYMBOL_REF
2047 && CONST_INT_P (XEXP (XEXP (src, 0), 1))))
2048 && dbg_cnt (cse2_move2add))
2050 rtx sym, off;
2052 if (GET_CODE (src) == SYMBOL_REF)
2054 sym = src;
2055 off = const0_rtx;
2057 else
2059 sym = XEXP (XEXP (src, 0), 0);
2060 off = XEXP (XEXP (src, 0), 1);
2063 /* If the reg already contains the value which is sum of
2064 sym and some constant value, we can use an add2 insn. */
2065 if (move2add_valid_value_p (regno, mode)
2066 && reg_base_reg[regno] < 0
2067 && reg_symbol_ref[regno] != NULL_RTX
2068 && rtx_equal_p (sym, reg_symbol_ref[regno]))
2069 changed |= move2add_use_add2_insn (mode, reg, sym, off, insn);
2071 /* Otherwise, we have to find a register whose value is sum
2072 of sym and some constant value. */
2073 else
2074 changed |= move2add_use_add3_insn (mode, reg, sym, off, insn);
2076 continue;
2080 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
2082 if (REG_NOTE_KIND (note) == REG_INC
2083 && REG_P (XEXP (note, 0)))
2085 /* Reset the information about this register. */
2086 int regno = REGNO (XEXP (note, 0));
2087 if (regno < FIRST_PSEUDO_REGISTER)
2089 move2add_record_mode (XEXP (note, 0));
2090 reg_mode[regno] = VOIDmode;
2094 note_stores (PATTERN (insn), move2add_note_store, insn);
2096 /* If INSN is a conditional branch, we try to extract an
2097 implicit set out of it. */
2098 if (any_condjump_p (insn))
2100 rtx cnd = fis_get_condition (insn);
2102 if (cnd != NULL_RTX
2103 && GET_CODE (cnd) == NE
2104 && REG_P (XEXP (cnd, 0))
2105 && !reg_set_p (XEXP (cnd, 0), insn)
2106 /* The following two checks, which are also in
2107 move2add_note_store, are intended to reduce the
2108 number of calls to gen_rtx_SET to avoid memory
2109 allocation if possible. */
2110 && SCALAR_INT_MODE_P (GET_MODE (XEXP (cnd, 0)))
2111 && REG_NREGS (XEXP (cnd, 0)) == 1
2112 && CONST_INT_P (XEXP (cnd, 1)))
2114 rtx implicit_set =
2115 gen_rtx_SET (XEXP (cnd, 0), XEXP (cnd, 1));
2116 move2add_note_store (SET_DEST (implicit_set), implicit_set, insn);
2120 /* If this is a CALL_INSN, all call used registers are stored with
2121 unknown values. */
2122 if (CALL_P (insn))
2124 rtx link;
2126 for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--)
2128 if (call_used_regs[i])
2129 /* Reset the information about this register. */
2130 reg_mode[i] = VOIDmode;
2133 for (link = CALL_INSN_FUNCTION_USAGE (insn); link;
2134 link = XEXP (link, 1))
2136 rtx setuse = XEXP (link, 0);
2137 rtx usage_rtx = XEXP (setuse, 0);
2138 if (GET_CODE (setuse) == CLOBBER
2139 && REG_P (usage_rtx))
2141 unsigned int end_regno = END_REGNO (usage_rtx);
2142 for (unsigned int r = REGNO (usage_rtx); r < end_regno; ++r)
2143 /* Reset the information about this register. */
2144 reg_mode[r] = VOIDmode;
2149 return changed;
2152 /* SET is a SET or CLOBBER that sets DST. DATA is the insn which
2153 contains SET.
2154 Update reg_set_luid, reg_offset and reg_base_reg accordingly.
2155 Called from reload_cse_move2add via note_stores. */
2157 static void
2158 move2add_note_store (rtx dst, const_rtx set, void *data)
2160 rtx_insn *insn = (rtx_insn *) data;
2161 unsigned int regno = 0;
2162 scalar_int_mode mode;
2164 /* Some targets do argument pushes without adding REG_INC notes. */
2166 if (MEM_P (dst))
2168 dst = XEXP (dst, 0);
2169 if (GET_CODE (dst) == PRE_INC || GET_CODE (dst) == POST_INC
2170 || GET_CODE (dst) == PRE_DEC || GET_CODE (dst) == POST_DEC)
2171 reg_mode[REGNO (XEXP (dst, 0))] = VOIDmode;
2172 return;
2175 if (GET_CODE (dst) == SUBREG)
2176 regno = subreg_regno (dst);
2177 else if (REG_P (dst))
2178 regno = REGNO (dst);
2179 else
2180 return;
2182 if (!is_a <scalar_int_mode> (GET_MODE (dst), &mode))
2183 goto invalidate;
2185 if (GET_CODE (set) == SET)
2187 rtx note, sym = NULL_RTX;
2188 rtx off;
2190 note = find_reg_equal_equiv_note (insn);
2191 if (note && GET_CODE (XEXP (note, 0)) == SYMBOL_REF)
2193 sym = XEXP (note, 0);
2194 off = const0_rtx;
2196 else if (note && GET_CODE (XEXP (note, 0)) == CONST
2197 && GET_CODE (XEXP (XEXP (note, 0), 0)) == PLUS
2198 && GET_CODE (XEXP (XEXP (XEXP (note, 0), 0), 0)) == SYMBOL_REF
2199 && CONST_INT_P (XEXP (XEXP (XEXP (note, 0), 0), 1)))
2201 sym = XEXP (XEXP (XEXP (note, 0), 0), 0);
2202 off = XEXP (XEXP (XEXP (note, 0), 0), 1);
2205 if (sym != NULL_RTX)
2207 move2add_record_sym_value (dst, sym, off);
2208 return;
2212 if (GET_CODE (set) == SET
2213 && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT
2214 && GET_CODE (SET_DEST (set)) != STRICT_LOW_PART)
2216 rtx src = SET_SRC (set);
2217 rtx base_reg;
2218 unsigned HOST_WIDE_INT offset;
2219 int base_regno;
2221 switch (GET_CODE (src))
2223 case PLUS:
2224 if (REG_P (XEXP (src, 0)))
2226 base_reg = XEXP (src, 0);
2228 if (CONST_INT_P (XEXP (src, 1)))
2229 offset = UINTVAL (XEXP (src, 1));
2230 else if (REG_P (XEXP (src, 1))
2231 && move2add_valid_value_p (REGNO (XEXP (src, 1)), mode))
2233 if (reg_base_reg[REGNO (XEXP (src, 1))] < 0
2234 && reg_symbol_ref[REGNO (XEXP (src, 1))] == NULL_RTX)
2235 offset = reg_offset[REGNO (XEXP (src, 1))];
2236 /* Maybe the first register is known to be a
2237 constant. */
2238 else if (move2add_valid_value_p (REGNO (base_reg), mode)
2239 && reg_base_reg[REGNO (base_reg)] < 0
2240 && reg_symbol_ref[REGNO (base_reg)] == NULL_RTX)
2242 offset = reg_offset[REGNO (base_reg)];
2243 base_reg = XEXP (src, 1);
2245 else
2246 goto invalidate;
2248 else
2249 goto invalidate;
2251 break;
2254 goto invalidate;
2256 case REG:
2257 base_reg = src;
2258 offset = 0;
2259 break;
2261 case CONST_INT:
2262 /* Start tracking the register as a constant. */
2263 reg_base_reg[regno] = -1;
2264 reg_symbol_ref[regno] = NULL_RTX;
2265 reg_offset[regno] = INTVAL (SET_SRC (set));
2266 /* We assign the same luid to all registers set to constants. */
2267 reg_set_luid[regno] = move2add_last_label_luid + 1;
2268 move2add_record_mode (dst);
2269 return;
2271 default:
2272 goto invalidate;
2275 base_regno = REGNO (base_reg);
2276 /* If information about the base register is not valid, set it
2277 up as a new base register, pretending its value is known
2278 starting from the current insn. */
2279 if (!move2add_valid_value_p (base_regno, mode))
2281 reg_base_reg[base_regno] = base_regno;
2282 reg_symbol_ref[base_regno] = NULL_RTX;
2283 reg_offset[base_regno] = 0;
2284 reg_set_luid[base_regno] = move2add_luid;
2285 gcc_assert (GET_MODE (base_reg) == mode);
2286 move2add_record_mode (base_reg);
2289 /* Copy base information from our base register. */
2290 reg_set_luid[regno] = reg_set_luid[base_regno];
2291 reg_base_reg[regno] = reg_base_reg[base_regno];
2292 reg_symbol_ref[regno] = reg_symbol_ref[base_regno];
2294 /* Compute the sum of the offsets or constants. */
2295 reg_offset[regno]
2296 = trunc_int_for_mode (offset + reg_offset[base_regno], mode);
2298 move2add_record_mode (dst);
2300 else
2302 invalidate:
2303 /* Invalidate the contents of the register. */
2304 move2add_record_mode (dst);
2305 reg_mode[regno] = VOIDmode;
2309 namespace {
2311 const pass_data pass_data_postreload_cse =
2313 RTL_PASS, /* type */
2314 "postreload", /* name */
2315 OPTGROUP_NONE, /* optinfo_flags */
2316 TV_RELOAD_CSE_REGS, /* tv_id */
2317 0, /* properties_required */
2318 0, /* properties_provided */
2319 0, /* properties_destroyed */
2320 0, /* todo_flags_start */
2321 TODO_df_finish, /* todo_flags_finish */
2324 class pass_postreload_cse : public rtl_opt_pass
2326 public:
2327 pass_postreload_cse (gcc::context *ctxt)
2328 : rtl_opt_pass (pass_data_postreload_cse, ctxt)
2331 /* opt_pass methods: */
2332 virtual bool gate (function *) { return (optimize > 0 && reload_completed); }
2334 virtual unsigned int execute (function *);
2336 }; // class pass_postreload_cse
2338 unsigned int
2339 pass_postreload_cse::execute (function *fun)
2341 if (!dbg_cnt (postreload_cse))
2342 return 0;
2344 /* Do a very simple CSE pass over just the hard registers. */
2345 reload_cse_regs (get_insns ());
2346 /* Reload_cse_regs can eliminate potentially-trapping MEMs.
2347 Remove any EH edges associated with them. */
2348 if (fun->can_throw_non_call_exceptions
2349 && purge_all_dead_edges ())
2350 cleanup_cfg (0);
2352 return 0;
2355 } // anon namespace
2357 rtl_opt_pass *
2358 make_pass_postreload_cse (gcc::context *ctxt)
2360 return new pass_postreload_cse (ctxt);