1 /* Optimize by combining instructions for GNU compiler.
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
22 /* This module is essentially the "combiner" phase of the U. of Arizona
23 Portable Optimizer, but redone to work on our list-structured
24 representation for RTL instead of their string representation.
26 The LOG_LINKS of each insn identify the most recent assignment
27 to each REG used in the insn. It is a list of previous insns,
28 each of which contains a SET for a REG that is used in this insn
29 and not used or set in between. LOG_LINKs never cross basic blocks.
30 They were set up by the preceding pass (lifetime analysis).
32 We try to combine each pair of insns joined by a logical link.
33 We also try to combine triples of insns A, B and C when
34 C has a link back to B and B has a link back to A.
36 LOG_LINKS does not have links for use of the CC0. They don't
37 need to, because the insn that sets the CC0 is always immediately
38 before the insn that tests it. So we always regard a branch
39 insn as having a logical link to the preceding insn. The same is true
40 for an insn explicitly using CC0.
42 We check (with use_crosses_set_p) to avoid combining in such a way
43 as to move a computation to a place where its value would be different.
45 Combination is done by mathematically substituting the previous
46 insn(s) values for the regs they set into the expressions in
47 the later insns that refer to these regs. If the result is a valid insn
48 for our target machine, according to the machine description,
49 we install it, delete the earlier insns, and update the data flow
50 information (LOG_LINKS and REG_NOTES) for what we did.
52 There are a few exceptions where the dataflow information created by
53 flow.c aren't completely updated:
55 - reg_live_length is not updated
56 - a LOG_LINKS entry that refers to an insn with multiple SETs may be
57 removed because there is no way to know which register it was
60 To simplify substitution, we combine only when the earlier insn(s)
61 consist of only a single assignment. To simplify updating afterward,
62 we never combine when a subroutine call appears in the middle.
64 Since we do not represent assignments to CC0 explicitly except when that
65 is all an insn does, there is no LOG_LINKS entry in an insn that uses
66 the condition code for the insn that set the condition code.
67 Fortunately, these two insns must be consecutive.
68 Therefore, every JUMP_INSN is taken to have an implicit logical link
69 to the preceding insn. This is not quite right, since non-jumps can
70 also use the condition code; but in practice such insns would not
75 #include "coretypes.h"
82 #include "hard-reg-set.h"
83 #include "basic-block.h"
84 #include "insn-config.h"
86 /* Include expr.h after insn-config.h so we get HAVE_conditional_move. */
88 #include "insn-attr.h"
94 /* Number of attempts to combine instructions in this function. */
96 static int combine_attempts;
98 /* Number of attempts that got as far as substitution in this function. */
100 static int combine_merges;
102 /* Number of instructions combined with added SETs in this function. */
104 static int combine_extras;
106 /* Number of instructions combined in this function. */
108 static int combine_successes;
110 /* Totals over entire compilation. */
112 static int total_attempts, total_merges, total_extras, total_successes;
115 /* Vector mapping INSN_UIDs to cuids.
116 The cuids are like uids but increase monotonically always.
117 Combine always uses cuids so that it can compare them.
118 But actually renumbering the uids, which we used to do,
119 proves to be a bad idea because it makes it hard to compare
120 the dumps produced by earlier passes with those from later passes. */
122 static int *uid_cuid;
123 static int max_uid_cuid;
125 /* Get the cuid of an insn. */
127 #define INSN_CUID(INSN) \
128 (INSN_UID (INSN) > max_uid_cuid ? insn_cuid (INSN) : uid_cuid[INSN_UID (INSN)])
130 /* In case BITS_PER_WORD == HOST_BITS_PER_WIDE_INT, shifting by
131 BITS_PER_WORD would invoke undefined behavior. Work around it. */
133 #define UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD(val) \
134 (((unsigned HOST_WIDE_INT) (val) << (BITS_PER_WORD - 1)) << 1)
136 #define nonzero_bits(X, M) \
137 cached_nonzero_bits (X, M, NULL_RTX, VOIDmode, 0)
139 #define num_sign_bit_copies(X, M) \
140 cached_num_sign_bit_copies (X, M, NULL_RTX, VOIDmode, 0)
142 /* Maximum register number, which is the size of the tables below. */
144 static unsigned int combine_max_regno;
147 /* Record last point of death of (hard or pseudo) register n. */
150 /* Record last point of modification of (hard or pseudo) register n. */
153 /* The next group of fields allows the recording of the last value assigned
154 to (hard or pseudo) register n. We use this information to see if an
155 operation being processed is redundant given a prior operation performed
156 on the register. For example, an `and' with a constant is redundant if
157 all the zero bits are already known to be turned off.
159 We use an approach similar to that used by cse, but change it in the
162 (1) We do not want to reinitialize at each label.
163 (2) It is useful, but not critical, to know the actual value assigned
164 to a register. Often just its form is helpful.
166 Therefore, we maintain the following fields:
168 last_set_value the last value assigned
169 last_set_label records the value of label_tick when the
170 register was assigned
171 last_set_table_tick records the value of label_tick when a
172 value using the register is assigned
173 last_set_invalid set to nonzero when it is not valid
174 to use the value of this register in some
177 To understand the usage of these tables, it is important to understand
178 the distinction between the value in last_set_value being valid and
179 the register being validly contained in some other expression in the
182 (The next two parameters are out of date).
184 reg_stat[i].last_set_value is valid if it is nonzero, and either
185 reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick.
187 Register I may validly appear in any expression returned for the value
188 of another register if reg_n_sets[i] is 1. It may also appear in the
189 value for register J if reg_stat[j].last_set_invalid is zero, or
190 reg_stat[i].last_set_label < reg_stat[j].last_set_label.
192 If an expression is found in the table containing a register which may
193 not validly appear in an expression, the register is replaced by
194 something that won't match, (clobber (const_int 0)). */
196 /* Record last value assigned to (hard or pseudo) register n. */
200 /* Record the value of label_tick when an expression involving register n
201 is placed in last_set_value. */
203 int last_set_table_tick;
205 /* Record the value of label_tick when the value for register n is placed in
210 /* These fields are maintained in parallel with last_set_value and are
211 used to store the mode in which the register was last set, te bits
212 that were known to be zero when it was last set, and the number of
213 sign bits copies it was known to have when it was last set. */
215 unsigned HOST_WIDE_INT last_set_nonzero_bits;
216 char last_set_sign_bit_copies;
217 ENUM_BITFIELD(machine_mode) last_set_mode : 8;
219 /* Set nonzero if references to register n in expressions should not be
220 used. last_set_invalid is set nonzero when this register is being
221 assigned to and last_set_table_tick == label_tick. */
223 char last_set_invalid;
225 /* Some registers that are set more than once and used in more than one
226 basic block are nevertheless always set in similar ways. For example,
227 a QImode register may be loaded from memory in two places on a machine
228 where byte loads zero extend.
230 We record in the following fields if a register has some leading bits
231 that are always equal to the sign bit, and what we know about the
232 nonzero bits of a register, specifically which bits are known to be
235 If an entry is zero, it means that we don't know anything special. */
237 unsigned char sign_bit_copies;
239 unsigned HOST_WIDE_INT nonzero_bits;
242 static struct reg_stat *reg_stat;
244 /* Record the cuid of the last insn that invalidated memory
245 (anything that writes memory, and subroutine calls, but not pushes). */
247 static int mem_last_set;
249 /* Record the cuid of the last CALL_INSN
250 so we can tell whether a potential combination crosses any calls. */
252 static int last_call_cuid;
254 /* When `subst' is called, this is the insn that is being modified
255 (by combining in a previous insn). The PATTERN of this insn
256 is still the old pattern partially modified and it should not be
257 looked at, but this may be used to examine the successors of the insn
258 to judge whether a simplification is valid. */
260 static rtx subst_insn;
262 /* This is the lowest CUID that `subst' is currently dealing with.
263 get_last_value will not return a value if the register was set at or
264 after this CUID. If not for this mechanism, we could get confused if
265 I2 or I1 in try_combine were an insn that used the old value of a register
266 to obtain a new value. In that case, we might erroneously get the
267 new value of the register when we wanted the old one. */
269 static int subst_low_cuid;
271 /* This contains any hard registers that are used in newpat; reg_dead_at_p
272 must consider all these registers to be always live. */
274 static HARD_REG_SET newpat_used_regs;
276 /* This is an insn to which a LOG_LINKS entry has been added. If this
277 insn is the earlier than I2 or I3, combine should rescan starting at
280 static rtx added_links_insn;
282 /* Basic block in which we are performing combines. */
283 static basic_block this_basic_block;
285 /* A bitmap indicating which blocks had registers go dead at entry.
286 After combine, we'll need to re-do global life analysis with
287 those blocks as starting points. */
288 static sbitmap refresh_blocks;
290 /* Incremented for each label. */
292 static int label_tick;
294 /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the
295 largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */
297 static enum machine_mode nonzero_bits_mode;
299 /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can
300 be safely used. It is zero while computing them and after combine has
301 completed. This former test prevents propagating values based on
302 previously set values, which can be incorrect if a variable is modified
305 static int nonzero_sign_valid;
308 /* Record one modification to rtl structure
309 to be undone by storing old_contents into *where.
310 is_int is 1 if the contents are an int. */
316 union {rtx r; int i;} old_contents;
317 union {rtx *r; int *i;} where;
320 /* Record a bunch of changes to be undone, up to MAX_UNDO of them.
321 num_undo says how many are currently recorded.
323 other_insn is nonzero if we have modified some other insn in the process
324 of working on subst_insn. It must be verified too. */
333 static struct undobuf undobuf;
335 /* Number of times the pseudo being substituted for
336 was found and replaced. */
338 static int n_occurrences;
340 static void do_SUBST (rtx *, rtx);
341 static void do_SUBST_INT (int *, int);
342 static void init_reg_last (void);
343 static void setup_incoming_promotions (void);
344 static void set_nonzero_bits_and_sign_copies (rtx, rtx, void *);
345 static int cant_combine_insn_p (rtx);
346 static int can_combine_p (rtx, rtx, rtx, rtx, rtx *, rtx *);
347 static int combinable_i3pat (rtx, rtx *, rtx, rtx, int, rtx *);
348 static int contains_muldiv (rtx);
349 static rtx try_combine (rtx, rtx, rtx, int *);
350 static void undo_all (void);
351 static void undo_commit (void);
352 static rtx *find_split_point (rtx *, rtx);
353 static rtx subst (rtx, rtx, rtx, int, int);
354 static rtx combine_simplify_rtx (rtx, enum machine_mode, int);
355 static rtx simplify_if_then_else (rtx);
356 static rtx simplify_set (rtx);
357 static rtx simplify_logical (rtx);
358 static rtx expand_compound_operation (rtx);
359 static rtx expand_field_assignment (rtx);
360 static rtx make_extraction (enum machine_mode, rtx, HOST_WIDE_INT,
361 rtx, unsigned HOST_WIDE_INT, int, int, int);
362 static rtx extract_left_shift (rtx, int);
363 static rtx make_compound_operation (rtx, enum rtx_code);
364 static int get_pos_from_mask (unsigned HOST_WIDE_INT,
365 unsigned HOST_WIDE_INT *);
366 static rtx force_to_mode (rtx, enum machine_mode,
367 unsigned HOST_WIDE_INT, rtx, int);
368 static rtx if_then_else_cond (rtx, rtx *, rtx *);
369 static rtx known_cond (rtx, enum rtx_code, rtx, rtx);
370 static int rtx_equal_for_field_assignment_p (rtx, rtx);
371 static rtx make_field_assignment (rtx);
372 static rtx apply_distributive_law (rtx);
373 static rtx simplify_and_const_int (rtx, enum machine_mode, rtx,
374 unsigned HOST_WIDE_INT);
375 static unsigned HOST_WIDE_INT cached_nonzero_bits (rtx, enum machine_mode,
376 rtx, enum machine_mode,
377 unsigned HOST_WIDE_INT);
378 static unsigned HOST_WIDE_INT nonzero_bits1 (rtx, enum machine_mode, rtx,
380 unsigned HOST_WIDE_INT);
381 static unsigned int cached_num_sign_bit_copies (rtx, enum machine_mode, rtx,
384 static unsigned int num_sign_bit_copies1 (rtx, enum machine_mode, rtx,
385 enum machine_mode, unsigned int);
386 static int merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code,
387 HOST_WIDE_INT, enum machine_mode, int *);
388 static rtx simplify_shift_const (rtx, enum rtx_code, enum machine_mode, rtx,
390 static int recog_for_combine (rtx *, rtx, rtx *);
391 static rtx gen_lowpart_for_combine (enum machine_mode, rtx);
392 static rtx gen_binary (enum rtx_code, enum machine_mode, rtx, rtx);
393 static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *);
394 static void update_table_tick (rtx);
395 static void record_value_for_reg (rtx, rtx, rtx);
396 static void check_promoted_subreg (rtx, rtx);
397 static void record_dead_and_set_regs_1 (rtx, rtx, void *);
398 static void record_dead_and_set_regs (rtx);
399 static int get_last_value_validate (rtx *, rtx, int, int);
400 static rtx get_last_value (rtx);
401 static int use_crosses_set_p (rtx, int);
402 static void reg_dead_at_p_1 (rtx, rtx, void *);
403 static int reg_dead_at_p (rtx, rtx);
404 static void move_deaths (rtx, rtx, int, rtx, rtx *);
405 static int reg_bitfield_target_p (rtx, rtx);
406 static void distribute_notes (rtx, rtx, rtx, rtx);
407 static void distribute_links (rtx);
408 static void mark_used_regs_combine (rtx);
409 static int insn_cuid (rtx);
410 static void record_promoted_value (rtx, rtx);
411 static rtx reversed_comparison (rtx, enum machine_mode, rtx, rtx);
412 static enum rtx_code combine_reversed_comparison_code (rtx);
413 static int unmentioned_reg_p_1 (rtx *, void *);
414 static bool unmentioned_reg_p (rtx, rtx);
416 /* Substitute NEWVAL, an rtx expression, into INTO, a place in some
417 insn. The substitution can be undone by undo_all. If INTO is already
418 set to NEWVAL, do not record this change. Because computing NEWVAL might
419 also call SUBST, we have to compute it before we put anything into
423 do_SUBST (rtx *into, rtx newval)
428 if (oldval == newval)
431 /* We'd like to catch as many invalid transformations here as
432 possible. Unfortunately, there are way too many mode changes
433 that are perfectly valid, so we'd waste too much effort for
434 little gain doing the checks here. Focus on catching invalid
435 transformations involving integer constants. */
436 if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT
437 && GET_CODE (newval) == CONST_INT)
439 /* Sanity check that we're replacing oldval with a CONST_INT
440 that is a valid sign-extension for the original mode. */
441 if (INTVAL (newval) != trunc_int_for_mode (INTVAL (newval),
445 /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
446 CONST_INT is not valid, because after the replacement, the
447 original mode would be gone. Unfortunately, we can't tell
448 when do_SUBST is called to replace the operand thereof, so we
449 perform this test on oldval instead, checking whether an
450 invalid replacement took place before we got here. */
451 if ((GET_CODE (oldval) == SUBREG
452 && GET_CODE (SUBREG_REG (oldval)) == CONST_INT)
453 || (GET_CODE (oldval) == ZERO_EXTEND
454 && GET_CODE (XEXP (oldval, 0)) == CONST_INT))
459 buf = undobuf.frees, undobuf.frees = buf->next;
461 buf = xmalloc (sizeof (struct undo));
465 buf->old_contents.r = oldval;
468 buf->next = undobuf.undos, undobuf.undos = buf;
471 #define SUBST(INTO, NEWVAL) do_SUBST(&(INTO), (NEWVAL))
473 /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution
474 for the value of a HOST_WIDE_INT value (including CONST_INT) is
478 do_SUBST_INT (int *into, int newval)
483 if (oldval == newval)
487 buf = undobuf.frees, undobuf.frees = buf->next;
489 buf = xmalloc (sizeof (struct undo));
493 buf->old_contents.i = oldval;
496 buf->next = undobuf.undos, undobuf.undos = buf;
499 #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT(&(INTO), (NEWVAL))
501 /* Main entry point for combiner. F is the first insn of the function.
502 NREGS is the first unused pseudo-reg number.
504 Return nonzero if the combiner has turned an indirect jump
505 instruction into a direct jump. */
507 combine_instructions (rtx f, unsigned int nregs)
514 rtx links, nextlinks;
516 int new_direct_jump_p = 0;
518 combine_attempts = 0;
521 combine_successes = 0;
523 combine_max_regno = nregs;
525 /* It is not safe to use ordinary gen_lowpart in combine.
526 See comments in gen_lowpart_for_combine. */
527 gen_lowpart = gen_lowpart_for_combine;
529 reg_stat = xcalloc (nregs, sizeof (struct reg_stat));
531 init_recog_no_volatile ();
533 /* Compute maximum uid value so uid_cuid can be allocated. */
535 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
536 if (INSN_UID (insn) > i)
539 uid_cuid = xmalloc ((i + 1) * sizeof (int));
542 nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0);
544 /* Don't use reg_stat[].nonzero_bits when computing it. This can cause
545 problems when, for example, we have j <<= 1 in a loop. */
547 nonzero_sign_valid = 0;
549 /* Compute the mapping from uids to cuids.
550 Cuids are numbers assigned to insns, like uids,
551 except that cuids increase monotonically through the code.
553 Scan all SETs and see if we can deduce anything about what
554 bits are known to be zero for some registers and how many copies
555 of the sign bit are known to exist for those registers.
557 Also set any known values so that we can use it while searching
558 for what bits are known to be set. */
562 setup_incoming_promotions ();
564 refresh_blocks = sbitmap_alloc (last_basic_block);
565 sbitmap_zero (refresh_blocks);
567 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
569 uid_cuid[INSN_UID (insn)] = ++i;
575 note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies,
577 record_dead_and_set_regs (insn);
580 for (links = REG_NOTES (insn); links; links = XEXP (links, 1))
581 if (REG_NOTE_KIND (links) == REG_INC)
582 set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX,
587 if (GET_CODE (insn) == CODE_LABEL)
591 nonzero_sign_valid = 1;
593 /* Now scan all the insns in forward order. */
599 setup_incoming_promotions ();
601 FOR_EACH_BB (this_basic_block)
603 for (insn = BB_HEAD (this_basic_block);
604 insn != NEXT_INSN (BB_END (this_basic_block));
605 insn = next ? next : NEXT_INSN (insn))
609 if (GET_CODE (insn) == CODE_LABEL)
612 else if (INSN_P (insn))
614 /* See if we know about function return values before this
615 insn based upon SUBREG flags. */
616 check_promoted_subreg (insn, PATTERN (insn));
618 /* Try this insn with each insn it links back to. */
620 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
621 if ((next = try_combine (insn, XEXP (links, 0),
622 NULL_RTX, &new_direct_jump_p)) != 0)
625 /* Try each sequence of three linked insns ending with this one. */
627 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
629 rtx link = XEXP (links, 0);
631 /* If the linked insn has been replaced by a note, then there
632 is no point in pursuing this chain any further. */
633 if (GET_CODE (link) == NOTE)
636 for (nextlinks = LOG_LINKS (link);
638 nextlinks = XEXP (nextlinks, 1))
639 if ((next = try_combine (insn, link,
641 &new_direct_jump_p)) != 0)
646 /* Try to combine a jump insn that uses CC0
647 with a preceding insn that sets CC0, and maybe with its
648 logical predecessor as well.
649 This is how we make decrement-and-branch insns.
650 We need this special code because data flow connections
651 via CC0 do not get entered in LOG_LINKS. */
653 if (GET_CODE (insn) == JUMP_INSN
654 && (prev = prev_nonnote_insn (insn)) != 0
655 && GET_CODE (prev) == INSN
656 && sets_cc0_p (PATTERN (prev)))
658 if ((next = try_combine (insn, prev,
659 NULL_RTX, &new_direct_jump_p)) != 0)
662 for (nextlinks = LOG_LINKS (prev); nextlinks;
663 nextlinks = XEXP (nextlinks, 1))
664 if ((next = try_combine (insn, prev,
666 &new_direct_jump_p)) != 0)
670 /* Do the same for an insn that explicitly references CC0. */
671 if (GET_CODE (insn) == INSN
672 && (prev = prev_nonnote_insn (insn)) != 0
673 && GET_CODE (prev) == INSN
674 && sets_cc0_p (PATTERN (prev))
675 && GET_CODE (PATTERN (insn)) == SET
676 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn))))
678 if ((next = try_combine (insn, prev,
679 NULL_RTX, &new_direct_jump_p)) != 0)
682 for (nextlinks = LOG_LINKS (prev); nextlinks;
683 nextlinks = XEXP (nextlinks, 1))
684 if ((next = try_combine (insn, prev,
686 &new_direct_jump_p)) != 0)
690 /* Finally, see if any of the insns that this insn links to
691 explicitly references CC0. If so, try this insn, that insn,
692 and its predecessor if it sets CC0. */
693 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
694 if (GET_CODE (XEXP (links, 0)) == INSN
695 && GET_CODE (PATTERN (XEXP (links, 0))) == SET
696 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0))))
697 && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0
698 && GET_CODE (prev) == INSN
699 && sets_cc0_p (PATTERN (prev))
700 && (next = try_combine (insn, XEXP (links, 0),
701 prev, &new_direct_jump_p)) != 0)
705 /* Try combining an insn with two different insns whose results it
707 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
708 for (nextlinks = XEXP (links, 1); nextlinks;
709 nextlinks = XEXP (nextlinks, 1))
710 if ((next = try_combine (insn, XEXP (links, 0),
712 &new_direct_jump_p)) != 0)
715 /* Try this insn with each REG_EQUAL note it links back to. */
716 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
719 rtx temp = XEXP (links, 0);
720 if ((set = single_set (temp)) != 0
721 && (note = find_reg_equal_equiv_note (temp)) != 0
722 && GET_CODE (XEXP (note, 0)) != EXPR_LIST
723 /* Avoid using a register that may already been marked
724 dead by an earlier instruction. */
725 && ! unmentioned_reg_p (XEXP (note, 0), SET_SRC (set)))
727 /* Temporarily replace the set's source with the
728 contents of the REG_EQUAL note. The insn will
729 be deleted or recognized by try_combine. */
730 rtx orig = SET_SRC (set);
731 SET_SRC (set) = XEXP (note, 0);
732 next = try_combine (insn, temp, NULL_RTX,
736 SET_SRC (set) = orig;
740 if (GET_CODE (insn) != NOTE)
741 record_dead_and_set_regs (insn);
750 EXECUTE_IF_SET_IN_SBITMAP (refresh_blocks, 0, i,
751 BASIC_BLOCK (i)->flags |= BB_DIRTY);
752 new_direct_jump_p |= purge_all_dead_edges (0);
753 delete_noop_moves (f);
755 update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
756 PROP_DEATH_NOTES | PROP_SCAN_DEAD_CODE
757 | PROP_KILL_DEAD_CODE);
760 sbitmap_free (refresh_blocks);
765 struct undo *undo, *next;
766 for (undo = undobuf.frees; undo; undo = next)
774 total_attempts += combine_attempts;
775 total_merges += combine_merges;
776 total_extras += combine_extras;
777 total_successes += combine_successes;
779 nonzero_sign_valid = 0;
780 gen_lowpart = gen_lowpart_general;
782 /* Make recognizer allow volatile MEMs again. */
785 return new_direct_jump_p;
788 /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */
794 for (i = 0; i < combine_max_regno; i++)
795 memset (reg_stat + i, 0, offsetof (struct reg_stat, sign_bit_copies));
798 /* Set up any promoted values for incoming argument registers. */
801 setup_incoming_promotions (void)
805 enum machine_mode mode;
807 rtx first = get_insns ();
809 if (targetm.calls.promote_function_args (TREE_TYPE (cfun->decl)))
811 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
812 /* Check whether this register can hold an incoming pointer
813 argument. FUNCTION_ARG_REGNO_P tests outgoing register
814 numbers, so translate if necessary due to register windows. */
815 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (regno))
816 && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0)
819 (reg, first, gen_rtx_fmt_e ((unsignedp ? ZERO_EXTEND
822 gen_rtx_CLOBBER (mode, const0_rtx)));
827 /* Called via note_stores. If X is a pseudo that is narrower than
828 HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero.
830 If we are setting only a portion of X and we can't figure out what
831 portion, assume all bits will be used since we don't know what will
834 Similarly, set how many bits of X are known to be copies of the sign bit
835 at all locations in the function. This is the smallest number implied
839 set_nonzero_bits_and_sign_copies (rtx x, rtx set,
840 void *data ATTRIBUTE_UNUSED)
844 if (GET_CODE (x) == REG
845 && REGNO (x) >= FIRST_PSEUDO_REGISTER
846 /* If this register is undefined at the start of the file, we can't
847 say what its contents were. */
848 && ! REGNO_REG_SET_P (ENTRY_BLOCK_PTR->next_bb->global_live_at_start, REGNO (x))
849 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
851 if (set == 0 || GET_CODE (set) == CLOBBER)
853 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
854 reg_stat[REGNO (x)].sign_bit_copies = 1;
858 /* If this is a complex assignment, see if we can convert it into a
859 simple assignment. */
860 set = expand_field_assignment (set);
862 /* If this is a simple assignment, or we have a paradoxical SUBREG,
863 set what we know about X. */
865 if (SET_DEST (set) == x
866 || (GET_CODE (SET_DEST (set)) == SUBREG
867 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
868 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set)))))
869 && SUBREG_REG (SET_DEST (set)) == x))
871 rtx src = SET_SRC (set);
873 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
874 /* If X is narrower than a word and SRC is a non-negative
875 constant that would appear negative in the mode of X,
876 sign-extend it for use in reg_stat[].nonzero_bits because some
877 machines (maybe most) will actually do the sign-extension
878 and this is the conservative approach.
880 ??? For 2.5, try to tighten up the MD files in this regard
881 instead of this kludge. */
883 if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD
884 && GET_CODE (src) == CONST_INT
886 && 0 != (INTVAL (src)
888 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
889 src = GEN_INT (INTVAL (src)
890 | ((HOST_WIDE_INT) (-1)
891 << GET_MODE_BITSIZE (GET_MODE (x))));
894 /* Don't call nonzero_bits if it cannot change anything. */
895 if (reg_stat[REGNO (x)].nonzero_bits != ~(unsigned HOST_WIDE_INT) 0)
896 reg_stat[REGNO (x)].nonzero_bits
897 |= nonzero_bits (src, nonzero_bits_mode);
898 num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x));
899 if (reg_stat[REGNO (x)].sign_bit_copies == 0
900 || reg_stat[REGNO (x)].sign_bit_copies > num)
901 reg_stat[REGNO (x)].sign_bit_copies = num;
905 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
906 reg_stat[REGNO (x)].sign_bit_copies = 1;
911 /* See if INSN can be combined into I3. PRED and SUCC are optionally
912 insns that were previously combined into I3 or that will be combined
913 into the merger of INSN and I3.
915 Return 0 if the combination is not allowed for any reason.
917 If the combination is allowed, *PDEST will be set to the single
918 destination of INSN and *PSRC to the single source, and this function
922 can_combine_p (rtx insn, rtx i3, rtx pred ATTRIBUTE_UNUSED, rtx succ,
923 rtx *pdest, rtx *psrc)
926 rtx set = 0, src, dest;
931 int all_adjacent = (succ ? (next_active_insn (insn) == succ
932 && next_active_insn (succ) == i3)
933 : next_active_insn (insn) == i3);
935 /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0.
936 or a PARALLEL consisting of such a SET and CLOBBERs.
938 If INSN has CLOBBER parallel parts, ignore them for our processing.
939 By definition, these happen during the execution of the insn. When it
940 is merged with another insn, all bets are off. If they are, in fact,
941 needed and aren't also supplied in I3, they may be added by
942 recog_for_combine. Otherwise, it won't match.
944 We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED
947 Get the source and destination of INSN. If more than one, can't
950 if (GET_CODE (PATTERN (insn)) == SET)
951 set = PATTERN (insn);
952 else if (GET_CODE (PATTERN (insn)) == PARALLEL
953 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET)
955 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
957 rtx elt = XVECEXP (PATTERN (insn), 0, i);
960 switch (GET_CODE (elt))
962 /* This is important to combine floating point insns
965 /* Combining an isolated USE doesn't make sense.
966 We depend here on combinable_i3pat to reject them. */
967 /* The code below this loop only verifies that the inputs of
968 the SET in INSN do not change. We call reg_set_between_p
969 to verify that the REG in the USE does not change between
971 If the USE in INSN was for a pseudo register, the matching
972 insn pattern will likely match any register; combining this
973 with any other USE would only be safe if we knew that the
974 used registers have identical values, or if there was
975 something to tell them apart, e.g. different modes. For
976 now, we forgo such complicated tests and simply disallow
977 combining of USES of pseudo registers with any other USE. */
978 if (GET_CODE (XEXP (elt, 0)) == REG
979 && GET_CODE (PATTERN (i3)) == PARALLEL)
981 rtx i3pat = PATTERN (i3);
982 int i = XVECLEN (i3pat, 0) - 1;
983 unsigned int regno = REGNO (XEXP (elt, 0));
987 rtx i3elt = XVECEXP (i3pat, 0, i);
989 if (GET_CODE (i3elt) == USE
990 && GET_CODE (XEXP (i3elt, 0)) == REG
991 && (REGNO (XEXP (i3elt, 0)) == regno
992 ? reg_set_between_p (XEXP (elt, 0),
993 PREV_INSN (insn), i3)
994 : regno >= FIRST_PSEUDO_REGISTER))
1001 /* We can ignore CLOBBERs. */
1006 /* Ignore SETs whose result isn't used but not those that
1007 have side-effects. */
1008 if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt))
1009 && (!(note = find_reg_note (insn, REG_EH_REGION, NULL_RTX))
1010 || INTVAL (XEXP (note, 0)) <= 0)
1011 && ! side_effects_p (elt))
1014 /* If we have already found a SET, this is a second one and
1015 so we cannot combine with this insn. */
1023 /* Anything else means we can't combine. */
1029 /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs,
1030 so don't do anything with it. */
1031 || GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1040 set = expand_field_assignment (set);
1041 src = SET_SRC (set), dest = SET_DEST (set);
1043 /* Don't eliminate a store in the stack pointer. */
1044 if (dest == stack_pointer_rtx
1045 /* Don't combine with an insn that sets a register to itself if it has
1046 a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */
1047 || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX))
1048 /* Can't merge an ASM_OPERANDS. */
1049 || GET_CODE (src) == ASM_OPERANDS
1050 /* Can't merge a function call. */
1051 || GET_CODE (src) == CALL
1052 /* Don't eliminate a function call argument. */
1053 || (GET_CODE (i3) == CALL_INSN
1054 && (find_reg_fusage (i3, USE, dest)
1055 || (GET_CODE (dest) == REG
1056 && REGNO (dest) < FIRST_PSEUDO_REGISTER
1057 && global_regs[REGNO (dest)])))
1058 /* Don't substitute into an incremented register. */
1059 || FIND_REG_INC_NOTE (i3, dest)
1060 || (succ && FIND_REG_INC_NOTE (succ, dest))
1062 /* Don't combine the end of a libcall into anything. */
1063 /* ??? This gives worse code, and appears to be unnecessary, since no
1064 pass after flow uses REG_LIBCALL/REG_RETVAL notes. Local-alloc does
1065 use REG_RETVAL notes for noconflict blocks, but other code here
1066 makes sure that those insns don't disappear. */
1067 || find_reg_note (insn, REG_RETVAL, NULL_RTX)
1069 /* Make sure that DEST is not used after SUCC but before I3. */
1070 || (succ && ! all_adjacent
1071 && reg_used_between_p (dest, succ, i3))
1072 /* Make sure that the value that is to be substituted for the register
1073 does not use any registers whose values alter in between. However,
1074 If the insns are adjacent, a use can't cross a set even though we
1075 think it might (this can happen for a sequence of insns each setting
1076 the same destination; last_set of that register might point to
1077 a NOTE). If INSN has a REG_EQUIV note, the register is always
1078 equivalent to the memory so the substitution is valid even if there
1079 are intervening stores. Also, don't move a volatile asm or
1080 UNSPEC_VOLATILE across any other insns. */
1082 && (((GET_CODE (src) != MEM
1083 || ! find_reg_note (insn, REG_EQUIV, src))
1084 && use_crosses_set_p (src, INSN_CUID (insn)))
1085 || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src))
1086 || GET_CODE (src) == UNSPEC_VOLATILE))
1087 /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get
1088 better register allocation by not doing the combine. */
1089 || find_reg_note (i3, REG_NO_CONFLICT, dest)
1090 || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest))
1091 /* Don't combine across a CALL_INSN, because that would possibly
1092 change whether the life span of some REGs crosses calls or not,
1093 and it is a pain to update that information.
1094 Exception: if source is a constant, moving it later can't hurt.
1095 Accept that special case, because it helps -fforce-addr a lot. */
1096 || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src)))
1099 /* DEST must either be a REG or CC0. */
1100 if (GET_CODE (dest) == REG)
1102 /* If register alignment is being enforced for multi-word items in all
1103 cases except for parameters, it is possible to have a register copy
1104 insn referencing a hard register that is not allowed to contain the
1105 mode being copied and which would not be valid as an operand of most
1106 insns. Eliminate this problem by not combining with such an insn.
1108 Also, on some machines we don't want to extend the life of a hard
1111 if (GET_CODE (src) == REG
1112 && ((REGNO (dest) < FIRST_PSEUDO_REGISTER
1113 && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest)))
1114 /* Don't extend the life of a hard register unless it is
1115 user variable (if we have few registers) or it can't
1116 fit into the desired register (meaning something special
1118 Also avoid substituting a return register into I3, because
1119 reload can't handle a conflict with constraints of other
1121 || (REGNO (src) < FIRST_PSEUDO_REGISTER
1122 && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)))))
1125 else if (GET_CODE (dest) != CC0)
1128 /* Don't substitute for a register intended as a clobberable operand.
1129 Similarly, don't substitute an expression containing a register that
1130 will be clobbered in I3. */
1131 if (GET_CODE (PATTERN (i3)) == PARALLEL)
1132 for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--)
1133 if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER
1134 && (reg_overlap_mentioned_p (XEXP (XVECEXP (PATTERN (i3), 0, i), 0),
1136 || rtx_equal_p (XEXP (XVECEXP (PATTERN (i3), 0, i), 0), dest)))
1139 /* If INSN contains anything volatile, or is an `asm' (whether volatile
1140 or not), reject, unless nothing volatile comes between it and I3 */
1142 if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src))
1144 /* Make sure succ doesn't contain a volatile reference. */
1145 if (succ != 0 && volatile_refs_p (PATTERN (succ)))
1148 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1149 if (INSN_P (p) && p != succ && volatile_refs_p (PATTERN (p)))
1153 /* If INSN is an asm, and DEST is a hard register, reject, since it has
1154 to be an explicit register variable, and was chosen for a reason. */
1156 if (GET_CODE (src) == ASM_OPERANDS
1157 && GET_CODE (dest) == REG && REGNO (dest) < FIRST_PSEUDO_REGISTER)
1160 /* If there are any volatile insns between INSN and I3, reject, because
1161 they might affect machine state. */
1163 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1164 if (INSN_P (p) && p != succ && volatile_insn_p (PATTERN (p)))
1167 /* If INSN or I2 contains an autoincrement or autodecrement,
1168 make sure that register is not used between there and I3,
1169 and not already used in I3 either.
1170 Also insist that I3 not be a jump; if it were one
1171 and the incremented register were spilled, we would lose. */
1174 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
1175 if (REG_NOTE_KIND (link) == REG_INC
1176 && (GET_CODE (i3) == JUMP_INSN
1177 || reg_used_between_p (XEXP (link, 0), insn, i3)
1178 || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3))))
1183 /* Don't combine an insn that follows a CC0-setting insn.
1184 An insn that uses CC0 must not be separated from the one that sets it.
1185 We do, however, allow I2 to follow a CC0-setting insn if that insn
1186 is passed as I1; in that case it will be deleted also.
1187 We also allow combining in this case if all the insns are adjacent
1188 because that would leave the two CC0 insns adjacent as well.
1189 It would be more logical to test whether CC0 occurs inside I1 or I2,
1190 but that would be much slower, and this ought to be equivalent. */
1192 p = prev_nonnote_insn (insn);
1193 if (p && p != pred && GET_CODE (p) == INSN && sets_cc0_p (PATTERN (p))
1198 /* If we get here, we have passed all the tests and the combination is
1207 /* LOC is the location within I3 that contains its pattern or the component
1208 of a PARALLEL of the pattern. We validate that it is valid for combining.
1210 One problem is if I3 modifies its output, as opposed to replacing it
1211 entirely, we can't allow the output to contain I2DEST or I1DEST as doing
1212 so would produce an insn that is not equivalent to the original insns.
1216 (set (reg:DI 101) (reg:DI 100))
1217 (set (subreg:SI (reg:DI 101) 0) <foo>)
1219 This is NOT equivalent to:
1221 (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>)
1222 (set (reg:DI 101) (reg:DI 100))])
1224 Not only does this modify 100 (in which case it might still be valid
1225 if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100.
1227 We can also run into a problem if I2 sets a register that I1
1228 uses and I1 gets directly substituted into I3 (not via I2). In that
1229 case, we would be getting the wrong value of I2DEST into I3, so we
1230 must reject the combination. This case occurs when I2 and I1 both
1231 feed into I3, rather than when I1 feeds into I2, which feeds into I3.
1232 If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source
1233 of a SET must prevent combination from occurring.
1235 Before doing the above check, we first try to expand a field assignment
1236 into a set of logical operations.
1238 If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which
1239 we place a register that is both set and used within I3. If more than one
1240 such register is detected, we fail.
1242 Return 1 if the combination is valid, zero otherwise. */
1245 combinable_i3pat (rtx i3, rtx *loc, rtx i2dest, rtx i1dest,
1246 int i1_not_in_src, rtx *pi3dest_killed)
1250 if (GET_CODE (x) == SET)
1253 rtx dest = SET_DEST (set);
1254 rtx src = SET_SRC (set);
1255 rtx inner_dest = dest;
1257 while (GET_CODE (inner_dest) == STRICT_LOW_PART
1258 || GET_CODE (inner_dest) == SUBREG
1259 || GET_CODE (inner_dest) == ZERO_EXTRACT)
1260 inner_dest = XEXP (inner_dest, 0);
1262 /* Check for the case where I3 modifies its output, as discussed
1263 above. We don't want to prevent pseudos from being combined
1264 into the address of a MEM, so only prevent the combination if
1265 i1 or i2 set the same MEM. */
1266 if ((inner_dest != dest &&
1267 (GET_CODE (inner_dest) != MEM
1268 || rtx_equal_p (i2dest, inner_dest)
1269 || (i1dest && rtx_equal_p (i1dest, inner_dest)))
1270 && (reg_overlap_mentioned_p (i2dest, inner_dest)
1271 || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest))))
1273 /* This is the same test done in can_combine_p except we can't test
1274 all_adjacent; we don't have to, since this instruction will stay
1275 in place, thus we are not considering increasing the lifetime of
1278 Also, if this insn sets a function argument, combining it with
1279 something that might need a spill could clobber a previous
1280 function argument; the all_adjacent test in can_combine_p also
1281 checks this; here, we do a more specific test for this case. */
1283 || (GET_CODE (inner_dest) == REG
1284 && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER
1285 && (! HARD_REGNO_MODE_OK (REGNO (inner_dest),
1286 GET_MODE (inner_dest))))
1287 || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src)))
1290 /* If DEST is used in I3, it is being killed in this insn,
1291 so record that for later.
1292 Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the
1293 STACK_POINTER_REGNUM, since these are always considered to be
1294 live. Similarly for ARG_POINTER_REGNUM if it is fixed. */
1295 if (pi3dest_killed && GET_CODE (dest) == REG
1296 && reg_referenced_p (dest, PATTERN (i3))
1297 && REGNO (dest) != FRAME_POINTER_REGNUM
1298 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1299 && REGNO (dest) != HARD_FRAME_POINTER_REGNUM
1301 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
1302 && (REGNO (dest) != ARG_POINTER_REGNUM
1303 || ! fixed_regs [REGNO (dest)])
1305 && REGNO (dest) != STACK_POINTER_REGNUM)
1307 if (*pi3dest_killed)
1310 *pi3dest_killed = dest;
1314 else if (GET_CODE (x) == PARALLEL)
1318 for (i = 0; i < XVECLEN (x, 0); i++)
1319 if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest,
1320 i1_not_in_src, pi3dest_killed))
1327 /* Return 1 if X is an arithmetic expression that contains a multiplication
1328 and division. We don't count multiplications by powers of two here. */
1331 contains_muldiv (rtx x)
1333 switch (GET_CODE (x))
1335 case MOD: case DIV: case UMOD: case UDIV:
1339 return ! (GET_CODE (XEXP (x, 1)) == CONST_INT
1340 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0);
1343 return contains_muldiv (XEXP (x, 0))
1344 || contains_muldiv (XEXP (x, 1));
1347 return contains_muldiv (XEXP (x, 0));
1353 /* Determine whether INSN can be used in a combination. Return nonzero if
1354 not. This is used in try_combine to detect early some cases where we
1355 can't perform combinations. */
1358 cant_combine_insn_p (rtx insn)
1363 /* If this isn't really an insn, we can't do anything.
1364 This can occur when flow deletes an insn that it has merged into an
1365 auto-increment address. */
1366 if (! INSN_P (insn))
1369 /* Never combine loads and stores involving hard regs that are likely
1370 to be spilled. The register allocator can usually handle such
1371 reg-reg moves by tying. If we allow the combiner to make
1372 substitutions of likely-spilled regs, we may abort in reload.
1373 As an exception, we allow combinations involving fixed regs; these are
1374 not available to the register allocator so there's no risk involved. */
1376 set = single_set (insn);
1379 src = SET_SRC (set);
1380 dest = SET_DEST (set);
1381 if (GET_CODE (src) == SUBREG)
1382 src = SUBREG_REG (src);
1383 if (GET_CODE (dest) == SUBREG)
1384 dest = SUBREG_REG (dest);
1385 if (REG_P (src) && REG_P (dest)
1386 && ((REGNO (src) < FIRST_PSEUDO_REGISTER
1387 && ! fixed_regs[REGNO (src)]
1388 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (src))))
1389 || (REGNO (dest) < FIRST_PSEUDO_REGISTER
1390 && ! fixed_regs[REGNO (dest)]
1391 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (dest))))))
1397 /* Adjust INSN after we made a change to its destination.
1399 Changing the destination can invalidate notes that say something about
1400 the results of the insn and a LOG_LINK pointing to the insn. */
1403 adjust_for_new_dest (rtx insn)
1407 /* For notes, be conservative and simply remove them. */
1408 loc = ®_NOTES (insn);
1411 enum reg_note kind = REG_NOTE_KIND (*loc);
1412 if (kind == REG_EQUAL || kind == REG_EQUIV)
1413 *loc = XEXP (*loc, 1);
1415 loc = &XEXP (*loc, 1);
1418 /* The new insn will have a destination that was previously the destination
1419 of an insn just above it. Call distribute_links to make a LOG_LINK from
1420 the next use of that destination. */
1421 distribute_links (gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX));
1424 /* Try to combine the insns I1 and I2 into I3.
1425 Here I1 and I2 appear earlier than I3.
1426 I1 can be zero; then we combine just I2 into I3.
1428 If we are combining three insns and the resulting insn is not recognized,
1429 try splitting it into two insns. If that happens, I2 and I3 are retained
1430 and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2
1433 Return 0 if the combination does not work. Then nothing is changed.
1434 If we did the combination, return the insn at which combine should
1437 Set NEW_DIRECT_JUMP_P to a nonzero value if try_combine creates a
1438 new direct jump instruction. */
1441 try_combine (rtx i3, rtx i2, rtx i1, int *new_direct_jump_p)
1443 /* New patterns for I3 and I2, respectively. */
1444 rtx newpat, newi2pat = 0;
1445 int substed_i2 = 0, substed_i1 = 0;
1446 /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */
1447 int added_sets_1, added_sets_2;
1448 /* Total number of SETs to put into I3. */
1450 /* Nonzero if I2's body now appears in I3. */
1452 /* INSN_CODEs for new I3, new I2, and user of condition code. */
1453 int insn_code_number, i2_code_number = 0, other_code_number = 0;
1454 /* Contains I3 if the destination of I3 is used in its source, which means
1455 that the old life of I3 is being killed. If that usage is placed into
1456 I2 and not in I3, a REG_DEAD note must be made. */
1457 rtx i3dest_killed = 0;
1458 /* SET_DEST and SET_SRC of I2 and I1. */
1459 rtx i2dest, i2src, i1dest = 0, i1src = 0;
1460 /* PATTERN (I2), or a copy of it in certain cases. */
1462 /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */
1463 int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0;
1464 int i1_feeds_i3 = 0;
1465 /* Notes that must be added to REG_NOTES in I3 and I2. */
1466 rtx new_i3_notes, new_i2_notes;
1467 /* Notes that we substituted I3 into I2 instead of the normal case. */
1468 int i3_subst_into_i2 = 0;
1469 /* Notes that I1, I2 or I3 is a MULT operation. */
1477 /* Exit early if one of the insns involved can't be used for
1479 if (cant_combine_insn_p (i3)
1480 || cant_combine_insn_p (i2)
1481 || (i1 && cant_combine_insn_p (i1))
1482 /* We also can't do anything if I3 has a
1483 REG_LIBCALL note since we don't want to disrupt the contiguity of a
1486 /* ??? This gives worse code, and appears to be unnecessary, since no
1487 pass after flow uses REG_LIBCALL/REG_RETVAL notes. */
1488 || find_reg_note (i3, REG_LIBCALL, NULL_RTX)
1494 undobuf.other_insn = 0;
1496 /* Reset the hard register usage information. */
1497 CLEAR_HARD_REG_SET (newpat_used_regs);
1499 /* If I1 and I2 both feed I3, they can be in any order. To simplify the
1500 code below, set I1 to be the earlier of the two insns. */
1501 if (i1 && INSN_CUID (i1) > INSN_CUID (i2))
1502 temp = i1, i1 = i2, i2 = temp;
1504 added_links_insn = 0;
1506 /* First check for one important special-case that the code below will
1507 not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
1508 and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case,
1509 we may be able to replace that destination with the destination of I3.
1510 This occurs in the common code where we compute both a quotient and
1511 remainder into a structure, in which case we want to do the computation
1512 directly into the structure to avoid register-register copies.
1514 Note that this case handles both multiple sets in I2 and also
1515 cases where I2 has a number of CLOBBER or PARALLELs.
1517 We make very conservative checks below and only try to handle the
1518 most common cases of this. For example, we only handle the case
1519 where I2 and I3 are adjacent to avoid making difficult register
1522 if (i1 == 0 && GET_CODE (i3) == INSN && GET_CODE (PATTERN (i3)) == SET
1523 && GET_CODE (SET_SRC (PATTERN (i3))) == REG
1524 && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER
1525 && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3)))
1526 && GET_CODE (PATTERN (i2)) == PARALLEL
1527 && ! side_effects_p (SET_DEST (PATTERN (i3)))
1528 /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code
1529 below would need to check what is inside (and reg_overlap_mentioned_p
1530 doesn't support those codes anyway). Don't allow those destinations;
1531 the resulting insn isn't likely to be recognized anyway. */
1532 && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT
1533 && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART
1534 && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)),
1535 SET_DEST (PATTERN (i3)))
1536 && next_real_insn (i2) == i3)
1538 rtx p2 = PATTERN (i2);
1540 /* Make sure that the destination of I3,
1541 which we are going to substitute into one output of I2,
1542 is not used within another output of I2. We must avoid making this:
1543 (parallel [(set (mem (reg 69)) ...)
1544 (set (reg 69) ...)])
1545 which is not well-defined as to order of actions.
1546 (Besides, reload can't handle output reloads for this.)
1548 The problem can also happen if the dest of I3 is a memory ref,
1549 if another dest in I2 is an indirect memory ref. */
1550 for (i = 0; i < XVECLEN (p2, 0); i++)
1551 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1552 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1553 && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)),
1554 SET_DEST (XVECEXP (p2, 0, i))))
1557 if (i == XVECLEN (p2, 0))
1558 for (i = 0; i < XVECLEN (p2, 0); i++)
1559 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1560 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1561 && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3)))
1566 subst_low_cuid = INSN_CUID (i2);
1568 added_sets_2 = added_sets_1 = 0;
1569 i2dest = SET_SRC (PATTERN (i3));
1571 /* Replace the dest in I2 with our dest and make the resulting
1572 insn the new pattern for I3. Then skip to where we
1573 validate the pattern. Everything was set up above. */
1574 SUBST (SET_DEST (XVECEXP (p2, 0, i)),
1575 SET_DEST (PATTERN (i3)));
1578 i3_subst_into_i2 = 1;
1579 goto validate_replacement;
1583 /* If I2 is setting a double-word pseudo to a constant and I3 is setting
1584 one of those words to another constant, merge them by making a new
1587 && (temp = single_set (i2)) != 0
1588 && (GET_CODE (SET_SRC (temp)) == CONST_INT
1589 || GET_CODE (SET_SRC (temp)) == CONST_DOUBLE)
1590 && GET_CODE (SET_DEST (temp)) == REG
1591 && GET_MODE_CLASS (GET_MODE (SET_DEST (temp))) == MODE_INT
1592 && GET_MODE_SIZE (GET_MODE (SET_DEST (temp))) == 2 * UNITS_PER_WORD
1593 && GET_CODE (PATTERN (i3)) == SET
1594 && GET_CODE (SET_DEST (PATTERN (i3))) == SUBREG
1595 && SUBREG_REG (SET_DEST (PATTERN (i3))) == SET_DEST (temp)
1596 && GET_MODE_CLASS (GET_MODE (SET_DEST (PATTERN (i3)))) == MODE_INT
1597 && GET_MODE_SIZE (GET_MODE (SET_DEST (PATTERN (i3)))) == UNITS_PER_WORD
1598 && GET_CODE (SET_SRC (PATTERN (i3))) == CONST_INT)
1600 HOST_WIDE_INT lo, hi;
1602 if (GET_CODE (SET_SRC (temp)) == CONST_INT)
1603 lo = INTVAL (SET_SRC (temp)), hi = lo < 0 ? -1 : 0;
1606 lo = CONST_DOUBLE_LOW (SET_SRC (temp));
1607 hi = CONST_DOUBLE_HIGH (SET_SRC (temp));
1610 if (subreg_lowpart_p (SET_DEST (PATTERN (i3))))
1612 /* We don't handle the case of the target word being wider
1613 than a host wide int. */
1614 if (HOST_BITS_PER_WIDE_INT < BITS_PER_WORD)
1617 lo &= ~(UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD (1) - 1);
1618 lo |= (INTVAL (SET_SRC (PATTERN (i3)))
1619 & (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD (1) - 1));
1621 else if (HOST_BITS_PER_WIDE_INT == BITS_PER_WORD)
1622 hi = INTVAL (SET_SRC (PATTERN (i3)));
1623 else if (HOST_BITS_PER_WIDE_INT >= 2 * BITS_PER_WORD)
1625 int sign = -(int) ((unsigned HOST_WIDE_INT) lo
1626 >> (HOST_BITS_PER_WIDE_INT - 1));
1628 lo &= ~ (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD
1629 (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD (1) - 1));
1630 lo |= (UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD
1631 (INTVAL (SET_SRC (PATTERN (i3)))));
1633 hi = lo < 0 ? -1 : 0;
1636 /* We don't handle the case of the higher word not fitting
1637 entirely in either hi or lo. */
1642 subst_low_cuid = INSN_CUID (i2);
1643 added_sets_2 = added_sets_1 = 0;
1644 i2dest = SET_DEST (temp);
1646 SUBST (SET_SRC (temp),
1647 immed_double_const (lo, hi, GET_MODE (SET_DEST (temp))));
1649 newpat = PATTERN (i2);
1650 goto validate_replacement;
1654 /* If we have no I1 and I2 looks like:
1655 (parallel [(set (reg:CC X) (compare:CC OP (const_int 0)))
1657 make up a dummy I1 that is
1660 (set (reg:CC X) (compare:CC Y (const_int 0)))
1662 (We can ignore any trailing CLOBBERs.)
1664 This undoes a previous combination and allows us to match a branch-and-
1667 if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL
1668 && XVECLEN (PATTERN (i2), 0) >= 2
1669 && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET
1670 && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0))))
1672 && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE
1673 && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx
1674 && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET
1675 && GET_CODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 1))) == REG
1676 && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0),
1677 SET_SRC (XVECEXP (PATTERN (i2), 0, 1))))
1679 for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--)
1680 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER)
1685 /* We make I1 with the same INSN_UID as I2. This gives it
1686 the same INSN_CUID for value tracking. Our fake I1 will
1687 never appear in the insn stream so giving it the same INSN_UID
1688 as I2 will not cause a problem. */
1690 i1 = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2,
1691 BLOCK_FOR_INSN (i2), INSN_LOCATOR (i2),
1692 XVECEXP (PATTERN (i2), 0, 1), -1, NULL_RTX,
1695 SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0));
1696 SUBST (XEXP (SET_SRC (PATTERN (i2)), 0),
1697 SET_DEST (PATTERN (i1)));
1702 /* Verify that I2 and I1 are valid for combining. */
1703 if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src)
1704 || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src)))
1710 /* Record whether I2DEST is used in I2SRC and similarly for the other
1711 cases. Knowing this will help in register status updating below. */
1712 i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src);
1713 i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src);
1714 i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src);
1716 /* See if I1 directly feeds into I3. It does if I1DEST is not used
1718 i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src);
1720 /* Ensure that I3's pattern can be the destination of combines. */
1721 if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest,
1722 i1 && i2dest_in_i1src && i1_feeds_i3,
1729 /* See if any of the insns is a MULT operation. Unless one is, we will
1730 reject a combination that is, since it must be slower. Be conservative
1732 if (GET_CODE (i2src) == MULT
1733 || (i1 != 0 && GET_CODE (i1src) == MULT)
1734 || (GET_CODE (PATTERN (i3)) == SET
1735 && GET_CODE (SET_SRC (PATTERN (i3))) == MULT))
1738 /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd.
1739 We used to do this EXCEPT in one case: I3 has a post-inc in an
1740 output operand. However, that exception can give rise to insns like
1742 which is a famous insn on the PDP-11 where the value of r3 used as the
1743 source was model-dependent. Avoid this sort of thing. */
1746 if (!(GET_CODE (PATTERN (i3)) == SET
1747 && GET_CODE (SET_SRC (PATTERN (i3))) == REG
1748 && GET_CODE (SET_DEST (PATTERN (i3))) == MEM
1749 && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC
1750 || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC)))
1751 /* It's not the exception. */
1754 for (link = REG_NOTES (i3); link; link = XEXP (link, 1))
1755 if (REG_NOTE_KIND (link) == REG_INC
1756 && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2))
1758 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1)))))
1765 /* See if the SETs in I1 or I2 need to be kept around in the merged
1766 instruction: whenever the value set there is still needed past I3.
1767 For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3.
1769 For the SET in I1, we have two cases: If I1 and I2 independently
1770 feed into I3, the set in I1 needs to be kept around if I1DEST dies
1771 or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set
1772 in I1 needs to be kept around unless I1DEST dies or is set in either
1773 I2 or I3. We can distinguish these cases by seeing if I2SRC mentions
1774 I1DEST. If so, we know I1 feeds into I2. */
1776 added_sets_2 = ! dead_or_set_p (i3, i2dest);
1779 = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest)
1780 : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest)));
1782 /* If the set in I2 needs to be kept around, we must make a copy of
1783 PATTERN (I2), so that when we substitute I1SRC for I1DEST in
1784 PATTERN (I2), we are only substituting for the original I1DEST, not into
1785 an already-substituted copy. This also prevents making self-referential
1786 rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to
1789 i2pat = (GET_CODE (PATTERN (i2)) == PARALLEL
1790 ? gen_rtx_SET (VOIDmode, i2dest, i2src)
1794 i2pat = copy_rtx (i2pat);
1798 /* Substitute in the latest insn for the regs set by the earlier ones. */
1800 maxreg = max_reg_num ();
1804 /* It is possible that the source of I2 or I1 may be performing an
1805 unneeded operation, such as a ZERO_EXTEND of something that is known
1806 to have the high part zero. Handle that case by letting subst look at
1807 the innermost one of them.
1809 Another way to do this would be to have a function that tries to
1810 simplify a single insn instead of merging two or more insns. We don't
1811 do this because of the potential of infinite loops and because
1812 of the potential extra memory required. However, doing it the way
1813 we are is a bit of a kludge and doesn't catch all cases.
1815 But only do this if -fexpensive-optimizations since it slows things down
1816 and doesn't usually win. */
1818 if (flag_expensive_optimizations)
1820 /* Pass pc_rtx so no substitutions are done, just simplifications. */
1823 subst_low_cuid = INSN_CUID (i1);
1824 i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0);
1828 subst_low_cuid = INSN_CUID (i2);
1829 i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0);
1834 /* Many machines that don't use CC0 have insns that can both perform an
1835 arithmetic operation and set the condition code. These operations will
1836 be represented as a PARALLEL with the first element of the vector
1837 being a COMPARE of an arithmetic operation with the constant zero.
1838 The second element of the vector will set some pseudo to the result
1839 of the same arithmetic operation. If we simplify the COMPARE, we won't
1840 match such a pattern and so will generate an extra insn. Here we test
1841 for this case, where both the comparison and the operation result are
1842 needed, and make the PARALLEL by just replacing I2DEST in I3SRC with
1843 I2SRC. Later we will make the PARALLEL that contains I2. */
1845 if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET
1846 && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE
1847 && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx
1848 && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest))
1850 #ifdef SELECT_CC_MODE
1852 enum machine_mode compare_mode;
1855 newpat = PATTERN (i3);
1856 SUBST (XEXP (SET_SRC (newpat), 0), i2src);
1860 #ifdef SELECT_CC_MODE
1861 /* See if a COMPARE with the operand we substituted in should be done
1862 with the mode that is currently being used. If not, do the same
1863 processing we do in `subst' for a SET; namely, if the destination
1864 is used only once, try to replace it with a register of the proper
1865 mode and also replace the COMPARE. */
1866 if (undobuf.other_insn == 0
1867 && (cc_use = find_single_use (SET_DEST (newpat), i3,
1868 &undobuf.other_insn))
1869 && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use),
1871 != GET_MODE (SET_DEST (newpat))))
1873 unsigned int regno = REGNO (SET_DEST (newpat));
1874 rtx new_dest = gen_rtx_REG (compare_mode, regno);
1876 if (regno < FIRST_PSEUDO_REGISTER
1877 || (REG_N_SETS (regno) == 1 && ! added_sets_2
1878 && ! REG_USERVAR_P (SET_DEST (newpat))))
1880 if (regno >= FIRST_PSEUDO_REGISTER)
1881 SUBST (regno_reg_rtx[regno], new_dest);
1883 SUBST (SET_DEST (newpat), new_dest);
1884 SUBST (XEXP (*cc_use, 0), new_dest);
1885 SUBST (SET_SRC (newpat),
1886 gen_rtx_COMPARE (compare_mode, i2src, const0_rtx));
1889 undobuf.other_insn = 0;
1896 n_occurrences = 0; /* `subst' counts here */
1898 /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we
1899 need to make a unique copy of I2SRC each time we substitute it
1900 to avoid self-referential rtl. */
1902 subst_low_cuid = INSN_CUID (i2);
1903 newpat = subst (PATTERN (i3), i2dest, i2src, 0,
1904 ! i1_feeds_i3 && i1dest_in_i1src);
1907 /* Record whether i2's body now appears within i3's body. */
1908 i2_is_used = n_occurrences;
1911 /* If we already got a failure, don't try to do more. Otherwise,
1912 try to substitute in I1 if we have it. */
1914 if (i1 && GET_CODE (newpat) != CLOBBER)
1916 /* Before we can do this substitution, we must redo the test done
1917 above (see detailed comments there) that ensures that I1DEST
1918 isn't mentioned in any SETs in NEWPAT that are field assignments. */
1920 if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX,
1928 subst_low_cuid = INSN_CUID (i1);
1929 newpat = subst (newpat, i1dest, i1src, 0, 0);
1933 /* Fail if an autoincrement side-effect has been duplicated. Be careful
1934 to count all the ways that I2SRC and I1SRC can be used. */
1935 if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0
1936 && i2_is_used + added_sets_2 > 1)
1937 || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
1938 && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3)
1940 /* Fail if we tried to make a new register (we used to abort, but there's
1941 really no reason to). */
1942 || max_reg_num () != maxreg
1943 /* Fail if we couldn't do something and have a CLOBBER. */
1944 || GET_CODE (newpat) == CLOBBER
1945 /* Fail if this new pattern is a MULT and we didn't have one before
1946 at the outer level. */
1947 || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT
1954 /* If the actions of the earlier insns must be kept
1955 in addition to substituting them into the latest one,
1956 we must make a new PARALLEL for the latest insn
1957 to hold additional the SETs. */
1959 if (added_sets_1 || added_sets_2)
1963 if (GET_CODE (newpat) == PARALLEL)
1965 rtvec old = XVEC (newpat, 0);
1966 total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2;
1967 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
1968 memcpy (XVEC (newpat, 0)->elem, &old->elem[0],
1969 sizeof (old->elem[0]) * old->num_elem);
1974 total_sets = 1 + added_sets_1 + added_sets_2;
1975 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
1976 XVECEXP (newpat, 0, 0) = old;
1980 XVECEXP (newpat, 0, --total_sets)
1981 = (GET_CODE (PATTERN (i1)) == PARALLEL
1982 ? gen_rtx_SET (VOIDmode, i1dest, i1src) : PATTERN (i1));
1986 /* If there is no I1, use I2's body as is. We used to also not do
1987 the subst call below if I2 was substituted into I3,
1988 but that could lose a simplification. */
1990 XVECEXP (newpat, 0, --total_sets) = i2pat;
1992 /* See comment where i2pat is assigned. */
1993 XVECEXP (newpat, 0, --total_sets)
1994 = subst (i2pat, i1dest, i1src, 0, 0);
1998 /* We come here when we are replacing a destination in I2 with the
1999 destination of I3. */
2000 validate_replacement:
2002 /* Note which hard regs this insn has as inputs. */
2003 mark_used_regs_combine (newpat);
2005 /* Is the result of combination a valid instruction? */
2006 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2008 /* If the result isn't valid, see if it is a PARALLEL of two SETs where
2009 the second SET's destination is a register that is unused and isn't
2010 marked as an instruction that might trap in an EH region. In that case,
2011 we just need the first SET. This can occur when simplifying a divmod
2012 insn. We *must* test for this case here because the code below that
2013 splits two independent SETs doesn't handle this case correctly when it
2014 updates the register status.
2016 It's pointless doing this if we originally had two sets, one from
2017 i3, and one from i2. Combining then splitting the parallel results
2018 in the original i2 again plus an invalid insn (which we delete).
2019 The net effect is only to move instructions around, which makes
2020 debug info less accurate.
2022 Also check the case where the first SET's destination is unused.
2023 That would not cause incorrect code, but does cause an unneeded
2026 if (insn_code_number < 0
2027 && !(added_sets_2 && i1 == 0)
2028 && GET_CODE (newpat) == PARALLEL
2029 && XVECLEN (newpat, 0) == 2
2030 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2031 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2032 && asm_noperands (newpat) < 0)
2034 rtx set0 = XVECEXP (newpat, 0, 0);
2035 rtx set1 = XVECEXP (newpat, 0, 1);
2038 if (((GET_CODE (SET_DEST (set1)) == REG
2039 && find_reg_note (i3, REG_UNUSED, SET_DEST (set1)))
2040 || (GET_CODE (SET_DEST (set1)) == SUBREG
2041 && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1)))))
2042 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2043 || INTVAL (XEXP (note, 0)) <= 0)
2044 && ! side_effects_p (SET_SRC (set1)))
2047 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2050 else if (((GET_CODE (SET_DEST (set0)) == REG
2051 && find_reg_note (i3, REG_UNUSED, SET_DEST (set0)))
2052 || (GET_CODE (SET_DEST (set0)) == SUBREG
2053 && find_reg_note (i3, REG_UNUSED,
2054 SUBREG_REG (SET_DEST (set0)))))
2055 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2056 || INTVAL (XEXP (note, 0)) <= 0)
2057 && ! side_effects_p (SET_SRC (set0)))
2060 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2062 if (insn_code_number >= 0)
2064 /* If we will be able to accept this, we have made a
2065 change to the destination of I3. This requires us to
2066 do a few adjustments. */
2068 PATTERN (i3) = newpat;
2069 adjust_for_new_dest (i3);
2074 /* If we were combining three insns and the result is a simple SET
2075 with no ASM_OPERANDS that wasn't recognized, try to split it into two
2076 insns. There are two ways to do this. It can be split using a
2077 machine-specific method (like when you have an addition of a large
2078 constant) or by combine in the function find_split_point. */
2080 if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET
2081 && asm_noperands (newpat) < 0)
2083 rtx m_split, *split;
2084 rtx ni2dest = i2dest;
2086 /* See if the MD file can split NEWPAT. If it can't, see if letting it
2087 use I2DEST as a scratch register will help. In the latter case,
2088 convert I2DEST to the mode of the source of NEWPAT if we can. */
2090 m_split = split_insns (newpat, i3);
2092 /* We can only use I2DEST as a scratch reg if it doesn't overlap any
2093 inputs of NEWPAT. */
2095 /* ??? If I2DEST is not safe, and I1DEST exists, then it would be
2096 possible to try that as a scratch reg. This would require adding
2097 more code to make it work though. */
2099 if (m_split == 0 && ! reg_overlap_mentioned_p (ni2dest, newpat))
2101 /* If I2DEST is a hard register or the only use of a pseudo,
2102 we can change its mode. */
2103 if (GET_MODE (SET_DEST (newpat)) != GET_MODE (i2dest)
2104 && GET_MODE (SET_DEST (newpat)) != VOIDmode
2105 && GET_CODE (i2dest) == REG
2106 && (REGNO (i2dest) < FIRST_PSEUDO_REGISTER
2107 || (REG_N_SETS (REGNO (i2dest)) == 1 && ! added_sets_2
2108 && ! REG_USERVAR_P (i2dest))))
2109 ni2dest = gen_rtx_REG (GET_MODE (SET_DEST (newpat)),
2112 m_split = split_insns (gen_rtx_PARALLEL
2114 gen_rtvec (2, newpat,
2115 gen_rtx_CLOBBER (VOIDmode,
2118 /* If the split with the mode-changed register didn't work, try
2119 the original register. */
2120 if (! m_split && ni2dest != i2dest)
2123 m_split = split_insns (gen_rtx_PARALLEL
2125 gen_rtvec (2, newpat,
2126 gen_rtx_CLOBBER (VOIDmode,
2132 if (m_split && NEXT_INSN (m_split) == NULL_RTX)
2134 m_split = PATTERN (m_split);
2135 insn_code_number = recog_for_combine (&m_split, i3, &new_i3_notes);
2136 if (insn_code_number >= 0)
2139 else if (m_split && NEXT_INSN (NEXT_INSN (m_split)) == NULL_RTX
2140 && (next_real_insn (i2) == i3
2141 || ! use_crosses_set_p (PATTERN (m_split), INSN_CUID (i2))))
2144 rtx newi3pat = PATTERN (NEXT_INSN (m_split));
2145 newi2pat = PATTERN (m_split);
2147 i3set = single_set (NEXT_INSN (m_split));
2148 i2set = single_set (m_split);
2150 /* In case we changed the mode of I2DEST, replace it in the
2151 pseudo-register table here. We can't do it above in case this
2152 code doesn't get executed and we do a split the other way. */
2154 if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2155 SUBST (regno_reg_rtx[REGNO (i2dest)], ni2dest);
2157 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2159 /* If I2 or I3 has multiple SETs, we won't know how to track
2160 register status, so don't use these insns. If I2's destination
2161 is used between I2 and I3, we also can't use these insns. */
2163 if (i2_code_number >= 0 && i2set && i3set
2164 && (next_real_insn (i2) == i3
2165 || ! reg_used_between_p (SET_DEST (i2set), i2, i3)))
2166 insn_code_number = recog_for_combine (&newi3pat, i3,
2168 if (insn_code_number >= 0)
2171 /* It is possible that both insns now set the destination of I3.
2172 If so, we must show an extra use of it. */
2174 if (insn_code_number >= 0)
2176 rtx new_i3_dest = SET_DEST (i3set);
2177 rtx new_i2_dest = SET_DEST (i2set);
2179 while (GET_CODE (new_i3_dest) == ZERO_EXTRACT
2180 || GET_CODE (new_i3_dest) == STRICT_LOW_PART
2181 || GET_CODE (new_i3_dest) == SUBREG)
2182 new_i3_dest = XEXP (new_i3_dest, 0);
2184 while (GET_CODE (new_i2_dest) == ZERO_EXTRACT
2185 || GET_CODE (new_i2_dest) == STRICT_LOW_PART
2186 || GET_CODE (new_i2_dest) == SUBREG)
2187 new_i2_dest = XEXP (new_i2_dest, 0);
2189 if (GET_CODE (new_i3_dest) == REG
2190 && GET_CODE (new_i2_dest) == REG
2191 && REGNO (new_i3_dest) == REGNO (new_i2_dest))
2192 REG_N_SETS (REGNO (new_i2_dest))++;
2196 /* If we can split it and use I2DEST, go ahead and see if that
2197 helps things be recognized. Verify that none of the registers
2198 are set between I2 and I3. */
2199 if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0
2201 && GET_CODE (i2dest) == REG
2203 /* We need I2DEST in the proper mode. If it is a hard register
2204 or the only use of a pseudo, we can change its mode. */
2205 && (GET_MODE (*split) == GET_MODE (i2dest)
2206 || GET_MODE (*split) == VOIDmode
2207 || REGNO (i2dest) < FIRST_PSEUDO_REGISTER
2208 || (REG_N_SETS (REGNO (i2dest)) == 1 && ! added_sets_2
2209 && ! REG_USERVAR_P (i2dest)))
2210 && (next_real_insn (i2) == i3
2211 || ! use_crosses_set_p (*split, INSN_CUID (i2)))
2212 /* We can't overwrite I2DEST if its value is still used by
2214 && ! reg_referenced_p (i2dest, newpat))
2216 rtx newdest = i2dest;
2217 enum rtx_code split_code = GET_CODE (*split);
2218 enum machine_mode split_mode = GET_MODE (*split);
2220 /* Get NEWDEST as a register in the proper mode. We have already
2221 validated that we can do this. */
2222 if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode)
2224 newdest = gen_rtx_REG (split_mode, REGNO (i2dest));
2226 if (REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2227 SUBST (regno_reg_rtx[REGNO (i2dest)], newdest);
2230 /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to
2231 an ASHIFT. This can occur if it was inside a PLUS and hence
2232 appeared to be a memory address. This is a kludge. */
2233 if (split_code == MULT
2234 && GET_CODE (XEXP (*split, 1)) == CONST_INT
2235 && INTVAL (XEXP (*split, 1)) > 0
2236 && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0)
2238 SUBST (*split, gen_rtx_ASHIFT (split_mode,
2239 XEXP (*split, 0), GEN_INT (i)));
2240 /* Update split_code because we may not have a multiply
2242 split_code = GET_CODE (*split);
2245 #ifdef INSN_SCHEDULING
2246 /* If *SPLIT is a paradoxical SUBREG, when we split it, it should
2247 be written as a ZERO_EXTEND. */
2248 if (split_code == SUBREG && GET_CODE (SUBREG_REG (*split)) == MEM)
2250 #ifdef LOAD_EXTEND_OP
2251 /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's
2252 what it really is. */
2253 if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (*split)))
2255 SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode,
2256 SUBREG_REG (*split)));
2259 SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode,
2260 SUBREG_REG (*split)));
2264 newi2pat = gen_rtx_SET (VOIDmode, newdest, *split);
2265 SUBST (*split, newdest);
2266 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2268 /* If the split point was a MULT and we didn't have one before,
2269 don't use one now. */
2270 if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult))
2271 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2275 /* Check for a case where we loaded from memory in a narrow mode and
2276 then sign extended it, but we need both registers. In that case,
2277 we have a PARALLEL with both loads from the same memory location.
2278 We can split this into a load from memory followed by a register-register
2279 copy. This saves at least one insn, more if register allocation can
2282 We cannot do this if the destination of the first assignment is a
2283 condition code register or cc0. We eliminate this case by making sure
2284 the SET_DEST and SET_SRC have the same mode.
2286 We cannot do this if the destination of the second assignment is
2287 a register that we have already assumed is zero-extended. Similarly
2288 for a SUBREG of such a register. */
2290 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2291 && GET_CODE (newpat) == PARALLEL
2292 && XVECLEN (newpat, 0) == 2
2293 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2294 && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND
2295 && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0)))
2296 == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0))))
2297 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2298 && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2299 XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0))
2300 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2302 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2303 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2304 && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)),
2305 (GET_CODE (temp) == REG
2306 && reg_stat[REGNO (temp)].nonzero_bits != 0
2307 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2308 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2309 && (reg_stat[REGNO (temp)].nonzero_bits
2310 != GET_MODE_MASK (word_mode))))
2311 && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG
2312 && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))),
2313 (GET_CODE (temp) == REG
2314 && reg_stat[REGNO (temp)].nonzero_bits != 0
2315 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2316 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2317 && (reg_stat[REGNO (temp)].nonzero_bits
2318 != GET_MODE_MASK (word_mode)))))
2319 && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2320 SET_SRC (XVECEXP (newpat, 0, 1)))
2321 && ! find_reg_note (i3, REG_UNUSED,
2322 SET_DEST (XVECEXP (newpat, 0, 0))))
2326 newi2pat = XVECEXP (newpat, 0, 0);
2327 ni2dest = SET_DEST (XVECEXP (newpat, 0, 0));
2328 newpat = XVECEXP (newpat, 0, 1);
2329 SUBST (SET_SRC (newpat),
2330 gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest));
2331 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2333 if (i2_code_number >= 0)
2334 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2336 if (insn_code_number >= 0)
2341 /* If we will be able to accept this, we have made a change to the
2342 destination of I3. This requires us to do a few adjustments. */
2343 PATTERN (i3) = newpat;
2344 adjust_for_new_dest (i3);
2346 /* I3 now uses what used to be its destination and which is
2347 now I2's destination. That means we need a LOG_LINK from
2348 I3 to I2. But we used to have one, so we still will.
2350 However, some later insn might be using I2's dest and have
2351 a LOG_LINK pointing at I3. We must remove this link.
2352 The simplest way to remove the link is to point it at I1,
2353 which we know will be a NOTE. */
2355 for (insn = NEXT_INSN (i3);
2356 insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
2357 || insn != BB_HEAD (this_basic_block->next_bb));
2358 insn = NEXT_INSN (insn))
2360 if (INSN_P (insn) && reg_referenced_p (ni2dest, PATTERN (insn)))
2362 for (link = LOG_LINKS (insn); link;
2363 link = XEXP (link, 1))
2364 if (XEXP (link, 0) == i3)
2365 XEXP (link, 0) = i1;
2373 /* Similarly, check for a case where we have a PARALLEL of two independent
2374 SETs but we started with three insns. In this case, we can do the sets
2375 as two separate insns. This case occurs when some SET allows two
2376 other insns to combine, but the destination of that SET is still live. */
2378 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2379 && GET_CODE (newpat) == PARALLEL
2380 && XVECLEN (newpat, 0) == 2
2381 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2382 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT
2383 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART
2384 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2385 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2386 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2387 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2389 /* Don't pass sets with (USE (MEM ...)) dests to the following. */
2390 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != USE
2391 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != USE
2392 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2393 XVECEXP (newpat, 0, 0))
2394 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)),
2395 XVECEXP (newpat, 0, 1))
2396 && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0)))
2397 && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1)))))
2399 /* Normally, it doesn't matter which of the two is done first,
2400 but it does if one references cc0. In that case, it has to
2403 if (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0)))
2405 newi2pat = XVECEXP (newpat, 0, 0);
2406 newpat = XVECEXP (newpat, 0, 1);
2411 newi2pat = XVECEXP (newpat, 0, 1);
2412 newpat = XVECEXP (newpat, 0, 0);
2415 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2417 if (i2_code_number >= 0)
2418 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2421 /* If it still isn't recognized, fail and change things back the way they
2423 if ((insn_code_number < 0
2424 /* Is the result a reasonable ASM_OPERANDS? */
2425 && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2)))
2431 /* If we had to change another insn, make sure it is valid also. */
2432 if (undobuf.other_insn)
2434 rtx other_pat = PATTERN (undobuf.other_insn);
2435 rtx new_other_notes;
2438 CLEAR_HARD_REG_SET (newpat_used_regs);
2440 other_code_number = recog_for_combine (&other_pat, undobuf.other_insn,
2443 if (other_code_number < 0 && ! check_asm_operands (other_pat))
2449 PATTERN (undobuf.other_insn) = other_pat;
2451 /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they
2452 are still valid. Then add any non-duplicate notes added by
2453 recog_for_combine. */
2454 for (note = REG_NOTES (undobuf.other_insn); note; note = next)
2456 next = XEXP (note, 1);
2458 if (REG_NOTE_KIND (note) == REG_UNUSED
2459 && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn)))
2461 if (GET_CODE (XEXP (note, 0)) == REG)
2462 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
2464 remove_note (undobuf.other_insn, note);
2468 for (note = new_other_notes; note; note = XEXP (note, 1))
2469 if (GET_CODE (XEXP (note, 0)) == REG)
2470 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
2472 distribute_notes (new_other_notes, undobuf.other_insn,
2473 undobuf.other_insn, NULL_RTX);
2476 /* If I2 is the CC0 setter and I3 is the CC0 user then check whether
2477 they are adjacent to each other or not. */
2479 rtx p = prev_nonnote_insn (i3);
2480 if (p && p != i2 && GET_CODE (p) == INSN && newi2pat
2481 && sets_cc0_p (newi2pat))
2489 /* We now know that we can do this combination. Merge the insns and
2490 update the status of registers and LOG_LINKS. */
2493 rtx i3notes, i2notes, i1notes = 0;
2494 rtx i3links, i2links, i1links = 0;
2498 /* Get the old REG_NOTES and LOG_LINKS from all our insns and
2500 i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3);
2501 i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2);
2503 i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1);
2505 /* Ensure that we do not have something that should not be shared but
2506 occurs multiple times in the new insns. Check this by first
2507 resetting all the `used' flags and then copying anything is shared. */
2509 reset_used_flags (i3notes);
2510 reset_used_flags (i2notes);
2511 reset_used_flags (i1notes);
2512 reset_used_flags (newpat);
2513 reset_used_flags (newi2pat);
2514 if (undobuf.other_insn)
2515 reset_used_flags (PATTERN (undobuf.other_insn));
2517 i3notes = copy_rtx_if_shared (i3notes);
2518 i2notes = copy_rtx_if_shared (i2notes);
2519 i1notes = copy_rtx_if_shared (i1notes);
2520 newpat = copy_rtx_if_shared (newpat);
2521 newi2pat = copy_rtx_if_shared (newi2pat);
2522 if (undobuf.other_insn)
2523 reset_used_flags (PATTERN (undobuf.other_insn));
2525 INSN_CODE (i3) = insn_code_number;
2526 PATTERN (i3) = newpat;
2528 if (GET_CODE (i3) == CALL_INSN && CALL_INSN_FUNCTION_USAGE (i3))
2530 rtx call_usage = CALL_INSN_FUNCTION_USAGE (i3);
2532 reset_used_flags (call_usage);
2533 call_usage = copy_rtx (call_usage);
2536 replace_rtx (call_usage, i2dest, i2src);
2539 replace_rtx (call_usage, i1dest, i1src);
2541 CALL_INSN_FUNCTION_USAGE (i3) = call_usage;
2544 if (undobuf.other_insn)
2545 INSN_CODE (undobuf.other_insn) = other_code_number;
2547 /* We had one special case above where I2 had more than one set and
2548 we replaced a destination of one of those sets with the destination
2549 of I3. In that case, we have to update LOG_LINKS of insns later
2550 in this basic block. Note that this (expensive) case is rare.
2552 Also, in this case, we must pretend that all REG_NOTEs for I2
2553 actually came from I3, so that REG_UNUSED notes from I2 will be
2554 properly handled. */
2556 if (i3_subst_into_i2)
2558 for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++)
2559 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != USE
2560 && GET_CODE (SET_DEST (XVECEXP (PATTERN (i2), 0, i))) == REG
2561 && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest
2562 && ! find_reg_note (i2, REG_UNUSED,
2563 SET_DEST (XVECEXP (PATTERN (i2), 0, i))))
2564 for (temp = NEXT_INSN (i2);
2565 temp && (this_basic_block->next_bb == EXIT_BLOCK_PTR
2566 || BB_HEAD (this_basic_block) != temp);
2567 temp = NEXT_INSN (temp))
2568 if (temp != i3 && INSN_P (temp))
2569 for (link = LOG_LINKS (temp); link; link = XEXP (link, 1))
2570 if (XEXP (link, 0) == i2)
2571 XEXP (link, 0) = i3;
2576 while (XEXP (link, 1))
2577 link = XEXP (link, 1);
2578 XEXP (link, 1) = i2notes;
2592 INSN_CODE (i2) = i2_code_number;
2593 PATTERN (i2) = newi2pat;
2597 PUT_CODE (i2, NOTE);
2598 NOTE_LINE_NUMBER (i2) = NOTE_INSN_DELETED;
2599 NOTE_SOURCE_FILE (i2) = 0;
2606 PUT_CODE (i1, NOTE);
2607 NOTE_LINE_NUMBER (i1) = NOTE_INSN_DELETED;
2608 NOTE_SOURCE_FILE (i1) = 0;
2611 /* Get death notes for everything that is now used in either I3 or
2612 I2 and used to die in a previous insn. If we built two new
2613 patterns, move from I1 to I2 then I2 to I3 so that we get the
2614 proper movement on registers that I2 modifies. */
2618 move_deaths (newi2pat, NULL_RTX, INSN_CUID (i1), i2, &midnotes);
2619 move_deaths (newpat, newi2pat, INSN_CUID (i1), i3, &midnotes);
2622 move_deaths (newpat, NULL_RTX, i1 ? INSN_CUID (i1) : INSN_CUID (i2),
2625 /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */
2627 distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX);
2629 distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX);
2631 distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX);
2633 distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2635 /* Distribute any notes added to I2 or I3 by recog_for_combine. We
2636 know these are REG_UNUSED and want them to go to the desired insn,
2637 so we always pass it as i3. We have not counted the notes in
2638 reg_n_deaths yet, so we need to do so now. */
2640 if (newi2pat && new_i2_notes)
2642 for (temp = new_i2_notes; temp; temp = XEXP (temp, 1))
2643 if (GET_CODE (XEXP (temp, 0)) == REG)
2644 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
2646 distribute_notes (new_i2_notes, i2, i2, NULL_RTX);
2651 for (temp = new_i3_notes; temp; temp = XEXP (temp, 1))
2652 if (GET_CODE (XEXP (temp, 0)) == REG)
2653 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
2655 distribute_notes (new_i3_notes, i3, i3, NULL_RTX);
2658 /* If I3DEST was used in I3SRC, it really died in I3. We may need to
2659 put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets
2660 I3DEST, the death must be somewhere before I2, not I3. If we passed I3
2661 in that case, it might delete I2. Similarly for I2 and I1.
2662 Show an additional death due to the REG_DEAD note we make here. If
2663 we discard it in distribute_notes, we will decrement it again. */
2667 if (GET_CODE (i3dest_killed) == REG)
2668 REG_N_DEATHS (REGNO (i3dest_killed))++;
2670 if (newi2pat && reg_set_p (i3dest_killed, newi2pat))
2671 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
2673 NULL_RTX, i2, NULL_RTX);
2675 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
2677 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2680 if (i2dest_in_i2src)
2682 if (GET_CODE (i2dest) == REG)
2683 REG_N_DEATHS (REGNO (i2dest))++;
2685 if (newi2pat && reg_set_p (i2dest, newi2pat))
2686 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
2687 NULL_RTX, i2, NULL_RTX);
2689 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
2690 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2693 if (i1dest_in_i1src)
2695 if (GET_CODE (i1dest) == REG)
2696 REG_N_DEATHS (REGNO (i1dest))++;
2698 if (newi2pat && reg_set_p (i1dest, newi2pat))
2699 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
2700 NULL_RTX, i2, NULL_RTX);
2702 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
2703 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX);
2706 distribute_links (i3links);
2707 distribute_links (i2links);
2708 distribute_links (i1links);
2710 if (GET_CODE (i2dest) == REG)
2713 rtx i2_insn = 0, i2_val = 0, set;
2715 /* The insn that used to set this register doesn't exist, and
2716 this life of the register may not exist either. See if one of
2717 I3's links points to an insn that sets I2DEST. If it does,
2718 that is now the last known value for I2DEST. If we don't update
2719 this and I2 set the register to a value that depended on its old
2720 contents, we will get confused. If this insn is used, thing
2721 will be set correctly in combine_instructions. */
2723 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
2724 if ((set = single_set (XEXP (link, 0))) != 0
2725 && rtx_equal_p (i2dest, SET_DEST (set)))
2726 i2_insn = XEXP (link, 0), i2_val = SET_SRC (set);
2728 record_value_for_reg (i2dest, i2_insn, i2_val);
2730 /* If the reg formerly set in I2 died only once and that was in I3,
2731 zero its use count so it won't make `reload' do any work. */
2733 && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat))
2734 && ! i2dest_in_i2src)
2736 regno = REGNO (i2dest);
2737 REG_N_SETS (regno)--;
2741 if (i1 && GET_CODE (i1dest) == REG)
2744 rtx i1_insn = 0, i1_val = 0, set;
2746 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
2747 if ((set = single_set (XEXP (link, 0))) != 0
2748 && rtx_equal_p (i1dest, SET_DEST (set)))
2749 i1_insn = XEXP (link, 0), i1_val = SET_SRC (set);
2751 record_value_for_reg (i1dest, i1_insn, i1_val);
2753 regno = REGNO (i1dest);
2754 if (! added_sets_1 && ! i1dest_in_i1src)
2755 REG_N_SETS (regno)--;
2758 /* Update reg_stat[].nonzero_bits et al for any changes that may have
2759 been made to this insn. The order of
2760 set_nonzero_bits_and_sign_copies() is important. Because newi2pat
2761 can affect nonzero_bits of newpat */
2763 note_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL);
2764 note_stores (newpat, set_nonzero_bits_and_sign_copies, NULL);
2766 /* Set new_direct_jump_p if a new return or simple jump instruction
2769 If I3 is now an unconditional jump, ensure that it has a
2770 BARRIER following it since it may have initially been a
2771 conditional jump. It may also be the last nonnote insn. */
2773 if (returnjump_p (i3) || any_uncondjump_p (i3))
2775 *new_direct_jump_p = 1;
2776 mark_jump_label (PATTERN (i3), i3, 0);
2778 if ((temp = next_nonnote_insn (i3)) == NULL_RTX
2779 || GET_CODE (temp) != BARRIER)
2780 emit_barrier_after (i3);
2783 if (undobuf.other_insn != NULL_RTX
2784 && (returnjump_p (undobuf.other_insn)
2785 || any_uncondjump_p (undobuf.other_insn)))
2787 *new_direct_jump_p = 1;
2789 if ((temp = next_nonnote_insn (undobuf.other_insn)) == NULL_RTX
2790 || GET_CODE (temp) != BARRIER)
2791 emit_barrier_after (undobuf.other_insn);
2794 /* An NOOP jump does not need barrier, but it does need cleaning up
2796 if (GET_CODE (newpat) == SET
2797 && SET_SRC (newpat) == pc_rtx
2798 && SET_DEST (newpat) == pc_rtx)
2799 *new_direct_jump_p = 1;
2802 combine_successes++;
2805 if (added_links_insn
2806 && (newi2pat == 0 || INSN_CUID (added_links_insn) < INSN_CUID (i2))
2807 && INSN_CUID (added_links_insn) < INSN_CUID (i3))
2808 return added_links_insn;
2810 return newi2pat ? i2 : i3;
2813 /* Undo all the modifications recorded in undobuf. */
2818 struct undo *undo, *next;
2820 for (undo = undobuf.undos; undo; undo = next)
2824 *undo->where.i = undo->old_contents.i;
2826 *undo->where.r = undo->old_contents.r;
2828 undo->next = undobuf.frees;
2829 undobuf.frees = undo;
2835 /* We've committed to accepting the changes we made. Move all
2836 of the undos to the free list. */
2841 struct undo *undo, *next;
2843 for (undo = undobuf.undos; undo; undo = next)
2846 undo->next = undobuf.frees;
2847 undobuf.frees = undo;
2853 /* Find the innermost point within the rtx at LOC, possibly LOC itself,
2854 where we have an arithmetic expression and return that point. LOC will
2857 try_combine will call this function to see if an insn can be split into
2861 find_split_point (rtx *loc, rtx insn)
2864 enum rtx_code code = GET_CODE (x);
2866 unsigned HOST_WIDE_INT len = 0;
2867 HOST_WIDE_INT pos = 0;
2869 rtx inner = NULL_RTX;
2871 /* First special-case some codes. */
2875 #ifdef INSN_SCHEDULING
2876 /* If we are making a paradoxical SUBREG invalid, it becomes a split
2878 if (GET_CODE (SUBREG_REG (x)) == MEM)
2881 return find_split_point (&SUBREG_REG (x), insn);
2885 /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it
2886 using LO_SUM and HIGH. */
2887 if (GET_CODE (XEXP (x, 0)) == CONST
2888 || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
2891 gen_rtx_LO_SUM (Pmode,
2892 gen_rtx_HIGH (Pmode, XEXP (x, 0)),
2894 return &XEXP (XEXP (x, 0), 0);
2898 /* If we have a PLUS whose second operand is a constant and the
2899 address is not valid, perhaps will can split it up using
2900 the machine-specific way to split large constants. We use
2901 the first pseudo-reg (one of the virtual regs) as a placeholder;
2902 it will not remain in the result. */
2903 if (GET_CODE (XEXP (x, 0)) == PLUS
2904 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
2905 && ! memory_address_p (GET_MODE (x), XEXP (x, 0)))
2907 rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER];
2908 rtx seq = split_insns (gen_rtx_SET (VOIDmode, reg, XEXP (x, 0)),
2911 /* This should have produced two insns, each of which sets our
2912 placeholder. If the source of the second is a valid address,
2913 we can make put both sources together and make a split point
2917 && NEXT_INSN (seq) != NULL_RTX
2918 && NEXT_INSN (NEXT_INSN (seq)) == NULL_RTX
2919 && GET_CODE (seq) == INSN
2920 && GET_CODE (PATTERN (seq)) == SET
2921 && SET_DEST (PATTERN (seq)) == reg
2922 && ! reg_mentioned_p (reg,
2923 SET_SRC (PATTERN (seq)))
2924 && GET_CODE (NEXT_INSN (seq)) == INSN
2925 && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET
2926 && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg
2927 && memory_address_p (GET_MODE (x),
2928 SET_SRC (PATTERN (NEXT_INSN (seq)))))
2930 rtx src1 = SET_SRC (PATTERN (seq));
2931 rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq)));
2933 /* Replace the placeholder in SRC2 with SRC1. If we can
2934 find where in SRC2 it was placed, that can become our
2935 split point and we can replace this address with SRC2.
2936 Just try two obvious places. */
2938 src2 = replace_rtx (src2, reg, src1);
2940 if (XEXP (src2, 0) == src1)
2941 split = &XEXP (src2, 0);
2942 else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e'
2943 && XEXP (XEXP (src2, 0), 0) == src1)
2944 split = &XEXP (XEXP (src2, 0), 0);
2948 SUBST (XEXP (x, 0), src2);
2953 /* If that didn't work, perhaps the first operand is complex and
2954 needs to be computed separately, so make a split point there.
2955 This will occur on machines that just support REG + CONST
2956 and have a constant moved through some previous computation. */
2958 else if (!OBJECT_P (XEXP (XEXP (x, 0), 0))
2959 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
2960 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
2961 return &XEXP (XEXP (x, 0), 0);
2967 /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a
2968 ZERO_EXTRACT, the most likely reason why this doesn't match is that
2969 we need to put the operand into a register. So split at that
2972 if (SET_DEST (x) == cc0_rtx
2973 && GET_CODE (SET_SRC (x)) != COMPARE
2974 && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT
2975 && !OBJECT_P (SET_SRC (x))
2976 && ! (GET_CODE (SET_SRC (x)) == SUBREG
2977 && OBJECT_P (SUBREG_REG (SET_SRC (x)))))
2978 return &SET_SRC (x);
2981 /* See if we can split SET_SRC as it stands. */
2982 split = find_split_point (&SET_SRC (x), insn);
2983 if (split && split != &SET_SRC (x))
2986 /* See if we can split SET_DEST as it stands. */
2987 split = find_split_point (&SET_DEST (x), insn);
2988 if (split && split != &SET_DEST (x))
2991 /* See if this is a bitfield assignment with everything constant. If
2992 so, this is an IOR of an AND, so split it into that. */
2993 if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
2994 && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))
2995 <= HOST_BITS_PER_WIDE_INT)
2996 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT
2997 && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT
2998 && GET_CODE (SET_SRC (x)) == CONST_INT
2999 && ((INTVAL (XEXP (SET_DEST (x), 1))
3000 + INTVAL (XEXP (SET_DEST (x), 2)))
3001 <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))))
3002 && ! side_effects_p (XEXP (SET_DEST (x), 0)))
3004 HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2));
3005 unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1));
3006 unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x));
3007 rtx dest = XEXP (SET_DEST (x), 0);
3008 enum machine_mode mode = GET_MODE (dest);
3009 unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1;
3011 if (BITS_BIG_ENDIAN)
3012 pos = GET_MODE_BITSIZE (mode) - len - pos;
3016 gen_binary (IOR, mode, dest, GEN_INT (src << pos)));
3019 gen_binary (IOR, mode,
3020 gen_binary (AND, mode, dest,
3021 gen_int_mode (~(mask << pos),
3023 GEN_INT (src << pos)));
3025 SUBST (SET_DEST (x), dest);
3027 split = find_split_point (&SET_SRC (x), insn);
3028 if (split && split != &SET_SRC (x))
3032 /* Otherwise, see if this is an operation that we can split into two.
3033 If so, try to split that. */
3034 code = GET_CODE (SET_SRC (x));
3039 /* If we are AND'ing with a large constant that is only a single
3040 bit and the result is only being used in a context where we
3041 need to know if it is zero or nonzero, replace it with a bit
3042 extraction. This will avoid the large constant, which might
3043 have taken more than one insn to make. If the constant were
3044 not a valid argument to the AND but took only one insn to make,
3045 this is no worse, but if it took more than one insn, it will
3048 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3049 && GET_CODE (XEXP (SET_SRC (x), 0)) == REG
3050 && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7
3051 && GET_CODE (SET_DEST (x)) == REG
3052 && (split = find_single_use (SET_DEST (x), insn, (rtx*) 0)) != 0
3053 && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE)
3054 && XEXP (*split, 0) == SET_DEST (x)
3055 && XEXP (*split, 1) == const0_rtx)
3057 rtx extraction = make_extraction (GET_MODE (SET_DEST (x)),
3058 XEXP (SET_SRC (x), 0),
3059 pos, NULL_RTX, 1, 1, 0, 0);
3060 if (extraction != 0)
3062 SUBST (SET_SRC (x), extraction);
3063 return find_split_point (loc, insn);
3069 /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X
3070 is known to be on, this can be converted into a NEG of a shift. */
3071 if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx
3072 && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0))
3073 && 1 <= (pos = exact_log2
3074 (nonzero_bits (XEXP (SET_SRC (x), 0),
3075 GET_MODE (XEXP (SET_SRC (x), 0))))))
3077 enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0));
3081 gen_rtx_LSHIFTRT (mode,
3082 XEXP (SET_SRC (x), 0),
3085 split = find_split_point (&SET_SRC (x), insn);
3086 if (split && split != &SET_SRC (x))
3092 inner = XEXP (SET_SRC (x), 0);
3094 /* We can't optimize if either mode is a partial integer
3095 mode as we don't know how many bits are significant
3097 if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT
3098 || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT)
3102 len = GET_MODE_BITSIZE (GET_MODE (inner));
3108 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3109 && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT)
3111 inner = XEXP (SET_SRC (x), 0);
3112 len = INTVAL (XEXP (SET_SRC (x), 1));
3113 pos = INTVAL (XEXP (SET_SRC (x), 2));
3115 if (BITS_BIG_ENDIAN)
3116 pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos;
3117 unsignedp = (code == ZERO_EXTRACT);
3125 if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner)))
3127 enum machine_mode mode = GET_MODE (SET_SRC (x));
3129 /* For unsigned, we have a choice of a shift followed by an
3130 AND or two shifts. Use two shifts for field sizes where the
3131 constant might be too large. We assume here that we can
3132 always at least get 8-bit constants in an AND insn, which is
3133 true for every current RISC. */
3135 if (unsignedp && len <= 8)
3140 (mode, gen_lowpart (mode, inner),
3142 GEN_INT (((HOST_WIDE_INT) 1 << len) - 1)));
3144 split = find_split_point (&SET_SRC (x), insn);
3145 if (split && split != &SET_SRC (x))
3152 (unsignedp ? LSHIFTRT : ASHIFTRT, mode,
3153 gen_rtx_ASHIFT (mode,
3154 gen_lowpart (mode, inner),
3155 GEN_INT (GET_MODE_BITSIZE (mode)
3157 GEN_INT (GET_MODE_BITSIZE (mode) - len)));
3159 split = find_split_point (&SET_SRC (x), insn);
3160 if (split && split != &SET_SRC (x))
3165 /* See if this is a simple operation with a constant as the second
3166 operand. It might be that this constant is out of range and hence
3167 could be used as a split point. */
3168 if (BINARY_P (SET_SRC (x))
3169 && CONSTANT_P (XEXP (SET_SRC (x), 1))
3170 && (OBJECT_P (XEXP (SET_SRC (x), 0))
3171 || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG
3172 && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0))))))
3173 return &XEXP (SET_SRC (x), 1);
3175 /* Finally, see if this is a simple operation with its first operand
3176 not in a register. The operation might require this operand in a
3177 register, so return it as a split point. We can always do this
3178 because if the first operand were another operation, we would have
3179 already found it as a split point. */
3180 if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x)))
3181 && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode))
3182 return &XEXP (SET_SRC (x), 0);
3188 /* We write NOR as (and (not A) (not B)), but if we don't have a NOR,
3189 it is better to write this as (not (ior A B)) so we can split it.
3190 Similarly for IOR. */
3191 if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT)
3194 gen_rtx_NOT (GET_MODE (x),
3195 gen_rtx_fmt_ee (code == IOR ? AND : IOR,
3197 XEXP (XEXP (x, 0), 0),
3198 XEXP (XEXP (x, 1), 0))));
3199 return find_split_point (loc, insn);
3202 /* Many RISC machines have a large set of logical insns. If the
3203 second operand is a NOT, put it first so we will try to split the
3204 other operand first. */
3205 if (GET_CODE (XEXP (x, 1)) == NOT)
3207 rtx tem = XEXP (x, 0);
3208 SUBST (XEXP (x, 0), XEXP (x, 1));
3209 SUBST (XEXP (x, 1), tem);
3217 /* Otherwise, select our actions depending on our rtx class. */
3218 switch (GET_RTX_CLASS (code))
3220 case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */
3222 split = find_split_point (&XEXP (x, 2), insn);
3225 /* ... fall through ... */
3227 case RTX_COMM_ARITH:
3229 case RTX_COMM_COMPARE:
3230 split = find_split_point (&XEXP (x, 1), insn);
3233 /* ... fall through ... */
3235 /* Some machines have (and (shift ...) ...) insns. If X is not
3236 an AND, but XEXP (X, 0) is, use it as our split point. */
3237 if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND)
3238 return &XEXP (x, 0);
3240 split = find_split_point (&XEXP (x, 0), insn);
3246 /* Otherwise, we don't have a split point. */
3251 /* Throughout X, replace FROM with TO, and return the result.
3252 The result is TO if X is FROM;
3253 otherwise the result is X, but its contents may have been modified.
3254 If they were modified, a record was made in undobuf so that
3255 undo_all will (among other things) return X to its original state.
3257 If the number of changes necessary is too much to record to undo,
3258 the excess changes are not made, so the result is invalid.
3259 The changes already made can still be undone.
3260 undobuf.num_undo is incremented for such changes, so by testing that
3261 the caller can tell whether the result is valid.
3263 `n_occurrences' is incremented each time FROM is replaced.
3265 IN_DEST is nonzero if we are processing the SET_DEST of a SET.
3267 UNIQUE_COPY is nonzero if each substitution must be unique. We do this
3268 by copying if `n_occurrences' is nonzero. */
3271 subst (rtx x, rtx from, rtx to, int in_dest, int unique_copy)
3273 enum rtx_code code = GET_CODE (x);
3274 enum machine_mode op0_mode = VOIDmode;
3279 /* Two expressions are equal if they are identical copies of a shared
3280 RTX or if they are both registers with the same register number
3283 #define COMBINE_RTX_EQUAL_P(X,Y) \
3285 || (GET_CODE (X) == REG && GET_CODE (Y) == REG \
3286 && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y)))
3288 if (! in_dest && COMBINE_RTX_EQUAL_P (x, from))
3291 return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to);
3294 /* If X and FROM are the same register but different modes, they will
3295 not have been seen as equal above. However, flow.c will make a
3296 LOG_LINKS entry for that case. If we do nothing, we will try to
3297 rerecognize our original insn and, when it succeeds, we will
3298 delete the feeding insn, which is incorrect.
3300 So force this insn not to match in this (rare) case. */
3301 if (! in_dest && code == REG && GET_CODE (from) == REG
3302 && REGNO (x) == REGNO (from))
3303 return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
3305 /* If this is an object, we are done unless it is a MEM or LO_SUM, both
3306 of which may contain things that can be combined. */
3307 if (code != MEM && code != LO_SUM && OBJECT_P (x))
3310 /* It is possible to have a subexpression appear twice in the insn.
3311 Suppose that FROM is a register that appears within TO.
3312 Then, after that subexpression has been scanned once by `subst',
3313 the second time it is scanned, TO may be found. If we were
3314 to scan TO here, we would find FROM within it and create a
3315 self-referent rtl structure which is completely wrong. */
3316 if (COMBINE_RTX_EQUAL_P (x, to))
3319 /* Parallel asm_operands need special attention because all of the
3320 inputs are shared across the arms. Furthermore, unsharing the
3321 rtl results in recognition failures. Failure to handle this case
3322 specially can result in circular rtl.
3324 Solve this by doing a normal pass across the first entry of the
3325 parallel, and only processing the SET_DESTs of the subsequent
3328 if (code == PARALLEL
3329 && GET_CODE (XVECEXP (x, 0, 0)) == SET
3330 && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS)
3332 new = subst (XVECEXP (x, 0, 0), from, to, 0, unique_copy);
3334 /* If this substitution failed, this whole thing fails. */
3335 if (GET_CODE (new) == CLOBBER
3336 && XEXP (new, 0) == const0_rtx)
3339 SUBST (XVECEXP (x, 0, 0), new);
3341 for (i = XVECLEN (x, 0) - 1; i >= 1; i--)
3343 rtx dest = SET_DEST (XVECEXP (x, 0, i));
3345 if (GET_CODE (dest) != REG
3346 && GET_CODE (dest) != CC0
3347 && GET_CODE (dest) != PC)
3349 new = subst (dest, from, to, 0, unique_copy);
3351 /* If this substitution failed, this whole thing fails. */
3352 if (GET_CODE (new) == CLOBBER
3353 && XEXP (new, 0) == const0_rtx)
3356 SUBST (SET_DEST (XVECEXP (x, 0, i)), new);
3362 len = GET_RTX_LENGTH (code);
3363 fmt = GET_RTX_FORMAT (code);
3365 /* We don't need to process a SET_DEST that is a register, CC0,
3366 or PC, so set up to skip this common case. All other cases
3367 where we want to suppress replacing something inside a
3368 SET_SRC are handled via the IN_DEST operand. */
3370 && (GET_CODE (SET_DEST (x)) == REG
3371 || GET_CODE (SET_DEST (x)) == CC0
3372 || GET_CODE (SET_DEST (x)) == PC))
3375 /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a
3378 op0_mode = GET_MODE (XEXP (x, 0));
3380 for (i = 0; i < len; i++)
3385 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3387 if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from))
3389 new = (unique_copy && n_occurrences
3390 ? copy_rtx (to) : to);
3395 new = subst (XVECEXP (x, i, j), from, to, 0,
3398 /* If this substitution failed, this whole thing
3400 if (GET_CODE (new) == CLOBBER
3401 && XEXP (new, 0) == const0_rtx)
3405 SUBST (XVECEXP (x, i, j), new);
3408 else if (fmt[i] == 'e')
3410 /* If this is a register being set, ignore it. */
3413 && (code == SUBREG || code == STRICT_LOW_PART
3414 || code == ZERO_EXTRACT)
3416 && GET_CODE (new) == REG)
3419 else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from))
3421 /* In general, don't install a subreg involving two
3422 modes not tieable. It can worsen register
3423 allocation, and can even make invalid reload
3424 insns, since the reg inside may need to be copied
3425 from in the outside mode, and that may be invalid
3426 if it is an fp reg copied in integer mode.
3428 We allow two exceptions to this: It is valid if
3429 it is inside another SUBREG and the mode of that
3430 SUBREG and the mode of the inside of TO is
3431 tieable and it is valid if X is a SET that copies
3434 if (GET_CODE (to) == SUBREG
3435 && ! MODES_TIEABLE_P (GET_MODE (to),
3436 GET_MODE (SUBREG_REG (to)))
3437 && ! (code == SUBREG
3438 && MODES_TIEABLE_P (GET_MODE (x),
3439 GET_MODE (SUBREG_REG (to))))
3441 && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx)
3444 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
3446 #ifdef CANNOT_CHANGE_MODE_CLASS
3448 && GET_CODE (to) == REG
3449 && REGNO (to) < FIRST_PSEUDO_REGISTER
3450 && REG_CANNOT_CHANGE_MODE_P (REGNO (to),
3453 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
3456 new = (unique_copy && n_occurrences ? copy_rtx (to) : to);
3460 /* If we are in a SET_DEST, suppress most cases unless we
3461 have gone inside a MEM, in which case we want to
3462 simplify the address. We assume here that things that
3463 are actually part of the destination have their inner
3464 parts in the first expression. This is true for SUBREG,
3465 STRICT_LOW_PART, and ZERO_EXTRACT, which are the only
3466 things aside from REG and MEM that should appear in a
3468 new = subst (XEXP (x, i), from, to,
3470 && (code == SUBREG || code == STRICT_LOW_PART
3471 || code == ZERO_EXTRACT))
3473 && i == 0), unique_copy);
3475 /* If we found that we will have to reject this combination,
3476 indicate that by returning the CLOBBER ourselves, rather than
3477 an expression containing it. This will speed things up as
3478 well as prevent accidents where two CLOBBERs are considered
3479 to be equal, thus producing an incorrect simplification. */
3481 if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx)
3484 if (GET_CODE (x) == SUBREG
3485 && (GET_CODE (new) == CONST_INT
3486 || GET_CODE (new) == CONST_DOUBLE))
3488 enum machine_mode mode = GET_MODE (x);
3490 x = simplify_subreg (GET_MODE (x), new,
3491 GET_MODE (SUBREG_REG (x)),
3494 x = gen_rtx_CLOBBER (mode, const0_rtx);
3496 else if (GET_CODE (new) == CONST_INT
3497 && GET_CODE (x) == ZERO_EXTEND)
3499 x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
3500 new, GET_MODE (XEXP (x, 0)));
3505 SUBST (XEXP (x, i), new);
3510 /* Try to simplify X. If the simplification changed the code, it is likely
3511 that further simplification will help, so loop, but limit the number
3512 of repetitions that will be performed. */
3514 for (i = 0; i < 4; i++)
3516 /* If X is sufficiently simple, don't bother trying to do anything
3518 if (code != CONST_INT && code != REG && code != CLOBBER)
3519 x = combine_simplify_rtx (x, op0_mode, in_dest);
3521 if (GET_CODE (x) == code)
3524 code = GET_CODE (x);
3526 /* We no longer know the original mode of operand 0 since we
3527 have changed the form of X) */
3528 op0_mode = VOIDmode;
3534 /* Simplify X, a piece of RTL. We just operate on the expression at the
3535 outer level; call `subst' to simplify recursively. Return the new
3538 OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero
3539 if we are inside a SET_DEST. */
3542 combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest)
3544 enum rtx_code code = GET_CODE (x);
3545 enum machine_mode mode = GET_MODE (x);
3550 /* If this is a commutative operation, put a constant last and a complex
3551 expression first. We don't need to do this for comparisons here. */
3552 if (COMMUTATIVE_ARITH_P (x)
3553 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
3556 SUBST (XEXP (x, 0), XEXP (x, 1));
3557 SUBST (XEXP (x, 1), temp);
3560 /* If this is a PLUS, MINUS, or MULT, and the first operand is the
3561 sign extension of a PLUS with a constant, reverse the order of the sign
3562 extension and the addition. Note that this not the same as the original
3563 code, but overflow is undefined for signed values. Also note that the
3564 PLUS will have been partially moved "inside" the sign-extension, so that
3565 the first operand of X will really look like:
3566 (ashiftrt (plus (ashift A C4) C5) C4).
3568 (plus (ashiftrt (ashift A C4) C2) C4)
3569 and replace the first operand of X with that expression. Later parts
3570 of this function may simplify the expression further.
3572 For example, if we start with (mult (sign_extend (plus A C1)) C2),
3573 we swap the SIGN_EXTEND and PLUS. Later code will apply the
3574 distributive law to produce (plus (mult (sign_extend X) C1) C3).
3576 We do this to simplify address expressions. */
3578 if ((code == PLUS || code == MINUS || code == MULT)
3579 && GET_CODE (XEXP (x, 0)) == ASHIFTRT
3580 && GET_CODE (XEXP (XEXP (x, 0), 0)) == PLUS
3581 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ASHIFT
3582 && GET_CODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1)) == CONST_INT
3583 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3584 && XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 1) == XEXP (XEXP (x, 0), 1)
3585 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
3586 && (temp = simplify_binary_operation (ASHIFTRT, mode,
3587 XEXP (XEXP (XEXP (x, 0), 0), 1),
3588 XEXP (XEXP (x, 0), 1))) != 0)
3591 = simplify_shift_const (NULL_RTX, ASHIFT, mode,
3592 XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0),
3593 INTVAL (XEXP (XEXP (x, 0), 1)));
3595 new = simplify_shift_const (NULL_RTX, ASHIFTRT, mode, new,
3596 INTVAL (XEXP (XEXP (x, 0), 1)));
3598 SUBST (XEXP (x, 0), gen_binary (PLUS, mode, new, temp));
3601 /* If this is a simple operation applied to an IF_THEN_ELSE, try
3602 applying it to the arms of the IF_THEN_ELSE. This often simplifies
3603 things. Check for cases where both arms are testing the same
3606 Don't do anything if all operands are very simple. */
3609 && ((!OBJECT_P (XEXP (x, 0))
3610 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
3611 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))
3612 || (!OBJECT_P (XEXP (x, 1))
3613 && ! (GET_CODE (XEXP (x, 1)) == SUBREG
3614 && OBJECT_P (SUBREG_REG (XEXP (x, 1)))))))
3616 && (!OBJECT_P (XEXP (x, 0))
3617 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
3618 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))))
3620 rtx cond, true_rtx, false_rtx;
3622 cond = if_then_else_cond (x, &true_rtx, &false_rtx);
3624 /* If everything is a comparison, what we have is highly unlikely
3625 to be simpler, so don't use it. */
3626 && ! (COMPARISON_P (x)
3627 && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))))
3629 rtx cop1 = const0_rtx;
3630 enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1);
3632 if (cond_code == NE && COMPARISON_P (cond))
3635 /* Simplify the alternative arms; this may collapse the true and
3636 false arms to store-flag values. Be careful to use copy_rtx
3637 here since true_rtx or false_rtx might share RTL with x as a
3638 result of the if_then_else_cond call above. */
3639 true_rtx = subst (copy_rtx (true_rtx), pc_rtx, pc_rtx, 0, 0);
3640 false_rtx = subst (copy_rtx (false_rtx), pc_rtx, pc_rtx, 0, 0);
3642 /* If true_rtx and false_rtx are not general_operands, an if_then_else
3643 is unlikely to be simpler. */
3644 if (general_operand (true_rtx, VOIDmode)
3645 && general_operand (false_rtx, VOIDmode))
3647 enum rtx_code reversed;
3649 /* Restarting if we generate a store-flag expression will cause
3650 us to loop. Just drop through in this case. */
3652 /* If the result values are STORE_FLAG_VALUE and zero, we can
3653 just make the comparison operation. */
3654 if (true_rtx == const_true_rtx && false_rtx == const0_rtx)
3655 x = gen_binary (cond_code, mode, cond, cop1);
3656 else if (true_rtx == const0_rtx && false_rtx == const_true_rtx
3657 && ((reversed = reversed_comparison_code_parts
3658 (cond_code, cond, cop1, NULL))
3660 x = gen_binary (reversed, mode, cond, cop1);
3662 /* Likewise, we can make the negate of a comparison operation
3663 if the result values are - STORE_FLAG_VALUE and zero. */
3664 else if (GET_CODE (true_rtx) == CONST_INT
3665 && INTVAL (true_rtx) == - STORE_FLAG_VALUE
3666 && false_rtx == const0_rtx)
3667 x = simplify_gen_unary (NEG, mode,
3668 gen_binary (cond_code, mode, cond,
3671 else if (GET_CODE (false_rtx) == CONST_INT
3672 && INTVAL (false_rtx) == - STORE_FLAG_VALUE
3673 && true_rtx == const0_rtx
3674 && ((reversed = reversed_comparison_code_parts
3675 (cond_code, cond, cop1, NULL))
3677 x = simplify_gen_unary (NEG, mode,
3678 gen_binary (reversed, mode,
3682 return gen_rtx_IF_THEN_ELSE (mode,
3683 gen_binary (cond_code, VOIDmode,
3685 true_rtx, false_rtx);
3687 code = GET_CODE (x);
3688 op0_mode = VOIDmode;
3693 /* Try to fold this expression in case we have constants that weren't
3696 switch (GET_RTX_CLASS (code))
3699 if (op0_mode == VOIDmode)
3700 op0_mode = GET_MODE (XEXP (x, 0));
3701 temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode);
3704 case RTX_COMM_COMPARE:
3706 enum machine_mode cmp_mode = GET_MODE (XEXP (x, 0));
3707 if (cmp_mode == VOIDmode)
3709 cmp_mode = GET_MODE (XEXP (x, 1));
3710 if (cmp_mode == VOIDmode)
3711 cmp_mode = op0_mode;
3713 temp = simplify_relational_operation (code, mode, cmp_mode,
3714 XEXP (x, 0), XEXP (x, 1));
3717 case RTX_COMM_ARITH:
3719 temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1));
3721 case RTX_BITFIELD_OPS:
3723 temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0),
3724 XEXP (x, 1), XEXP (x, 2));
3733 code = GET_CODE (temp);
3734 op0_mode = VOIDmode;
3735 mode = GET_MODE (temp);
3738 /* First see if we can apply the inverse distributive law. */
3739 if (code == PLUS || code == MINUS
3740 || code == AND || code == IOR || code == XOR)
3742 x = apply_distributive_law (x);
3743 code = GET_CODE (x);
3744 op0_mode = VOIDmode;
3747 /* If CODE is an associative operation not otherwise handled, see if we
3748 can associate some operands. This can win if they are constants or
3749 if they are logically related (i.e. (a & b) & a). */
3750 if ((code == PLUS || code == MINUS || code == MULT || code == DIV
3751 || code == AND || code == IOR || code == XOR
3752 || code == SMAX || code == SMIN || code == UMAX || code == UMIN)
3753 && ((INTEGRAL_MODE_P (mode) && code != DIV)
3754 || (flag_unsafe_math_optimizations && FLOAT_MODE_P (mode))))
3756 if (GET_CODE (XEXP (x, 0)) == code)
3758 rtx other = XEXP (XEXP (x, 0), 0);
3759 rtx inner_op0 = XEXP (XEXP (x, 0), 1);
3760 rtx inner_op1 = XEXP (x, 1);
3763 /* Make sure we pass the constant operand if any as the second
3764 one if this is a commutative operation. */
3765 if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x))
3767 rtx tem = inner_op0;
3768 inner_op0 = inner_op1;
3771 inner = simplify_binary_operation (code == MINUS ? PLUS
3772 : code == DIV ? MULT
3774 mode, inner_op0, inner_op1);
3776 /* For commutative operations, try the other pair if that one
3778 if (inner == 0 && COMMUTATIVE_ARITH_P (x))
3780 other = XEXP (XEXP (x, 0), 1);
3781 inner = simplify_binary_operation (code, mode,
3782 XEXP (XEXP (x, 0), 0),
3787 return gen_binary (code, mode, other, inner);
3791 /* A little bit of algebraic simplification here. */
3795 /* Ensure that our address has any ASHIFTs converted to MULT in case
3796 address-recognizing predicates are called later. */
3797 temp = make_compound_operation (XEXP (x, 0), MEM);
3798 SUBST (XEXP (x, 0), temp);
3802 if (op0_mode == VOIDmode)
3803 op0_mode = GET_MODE (SUBREG_REG (x));
3805 /* See if this can be moved to simplify_subreg. */
3806 if (CONSTANT_P (SUBREG_REG (x))
3807 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
3808 /* Don't call gen_lowpart if the inner mode
3809 is VOIDmode and we cannot simplify it, as SUBREG without
3810 inner mode is invalid. */
3811 && (GET_MODE (SUBREG_REG (x)) != VOIDmode
3812 || gen_lowpart_common (mode, SUBREG_REG (x))))
3813 return gen_lowpart (mode, SUBREG_REG (x));
3815 if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC)
3819 temp = simplify_subreg (mode, SUBREG_REG (x), op0_mode,