Merge from mainline
[official-gcc.git] / gcc / combine.c
blob96ebd208a825730ae623a5acc1e7d167a46bd652
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, 2005, 2006 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
10 version.
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
15 for more details.
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, 51 Franklin Street, Fifth Floor, Boston, MA
20 02110-1301, USA. */
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 - reg_n_refs is not adjusted in the rare case when a register is
57 no longer required in a computation
58 - there are extremely rare cases (see distribute_notes) when a
59 REG_DEAD note is lost
60 - a LOG_LINKS entry that refers to an insn with multiple SETs may be
61 removed because there is no way to know which register it was
62 linking
64 To simplify substitution, we combine only when the earlier insn(s)
65 consist of only a single assignment. To simplify updating afterward,
66 we never combine when a subroutine call appears in the middle.
68 Since we do not represent assignments to CC0 explicitly except when that
69 is all an insn does, there is no LOG_LINKS entry in an insn that uses
70 the condition code for the insn that set the condition code.
71 Fortunately, these two insns must be consecutive.
72 Therefore, every JUMP_INSN is taken to have an implicit logical link
73 to the preceding insn. This is not quite right, since non-jumps can
74 also use the condition code; but in practice such insns would not
75 combine anyway. */
77 #include "config.h"
78 #include "system.h"
79 #include "coretypes.h"
80 #include "tm.h"
81 #include "rtl.h"
82 #include "tree.h"
83 #include "tm_p.h"
84 #include "flags.h"
85 #include "regs.h"
86 #include "hard-reg-set.h"
87 #include "basic-block.h"
88 #include "insn-config.h"
89 #include "function.h"
90 /* Include expr.h after insn-config.h so we get HAVE_conditional_move. */
91 #include "expr.h"
92 #include "insn-attr.h"
93 #include "recog.h"
94 #include "real.h"
95 #include "toplev.h"
96 #include "target.h"
97 #include "optabs.h"
98 #include "insn-codes.h"
99 #include "rtlhooks-def.h"
100 /* Include output.h for dump_file. */
101 #include "output.h"
102 #include "params.h"
103 #include "timevar.h"
104 #include "tree-pass.h"
106 /* Number of attempts to combine instructions in this function. */
108 static int combine_attempts;
110 /* Number of attempts that got as far as substitution in this function. */
112 static int combine_merges;
114 /* Number of instructions combined with added SETs in this function. */
116 static int combine_extras;
118 /* Number of instructions combined in this function. */
120 static int combine_successes;
122 /* Totals over entire compilation. */
124 static int total_attempts, total_merges, total_extras, total_successes;
127 /* Vector mapping INSN_UIDs to cuids.
128 The cuids are like uids but increase monotonically always.
129 Combine always uses cuids so that it can compare them.
130 But actually renumbering the uids, which we used to do,
131 proves to be a bad idea because it makes it hard to compare
132 the dumps produced by earlier passes with those from later passes. */
134 static int *uid_cuid;
135 static int max_uid_cuid;
137 /* Get the cuid of an insn. */
139 #define INSN_CUID(INSN) \
140 (INSN_UID (INSN) > max_uid_cuid ? insn_cuid (INSN) : uid_cuid[INSN_UID (INSN)])
142 /* In case BITS_PER_WORD == HOST_BITS_PER_WIDE_INT, shifting by
143 BITS_PER_WORD would invoke undefined behavior. Work around it. */
145 #define UWIDE_SHIFT_LEFT_BY_BITS_PER_WORD(val) \
146 (((unsigned HOST_WIDE_INT) (val) << (BITS_PER_WORD - 1)) << 1)
148 /* Maximum register number, which is the size of the tables below. */
150 static unsigned int combine_max_regno;
152 struct reg_stat {
153 /* Record last point of death of (hard or pseudo) register n. */
154 rtx last_death;
156 /* Record last point of modification of (hard or pseudo) register n. */
157 rtx last_set;
159 /* The next group of fields allows the recording of the last value assigned
160 to (hard or pseudo) register n. We use this information to see if an
161 operation being processed is redundant given a prior operation performed
162 on the register. For example, an `and' with a constant is redundant if
163 all the zero bits are already known to be turned off.
165 We use an approach similar to that used by cse, but change it in the
166 following ways:
168 (1) We do not want to reinitialize at each label.
169 (2) It is useful, but not critical, to know the actual value assigned
170 to a register. Often just its form is helpful.
172 Therefore, we maintain the following fields:
174 last_set_value the last value assigned
175 last_set_label records the value of label_tick when the
176 register was assigned
177 last_set_table_tick records the value of label_tick when a
178 value using the register is assigned
179 last_set_invalid set to nonzero when it is not valid
180 to use the value of this register in some
181 register's value
183 To understand the usage of these tables, it is important to understand
184 the distinction between the value in last_set_value being valid and
185 the register being validly contained in some other expression in the
186 table.
188 (The next two parameters are out of date).
190 reg_stat[i].last_set_value is valid if it is nonzero, and either
191 reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick.
193 Register I may validly appear in any expression returned for the value
194 of another register if reg_n_sets[i] is 1. It may also appear in the
195 value for register J if reg_stat[j].last_set_invalid is zero, or
196 reg_stat[i].last_set_label < reg_stat[j].last_set_label.
198 If an expression is found in the table containing a register which may
199 not validly appear in an expression, the register is replaced by
200 something that won't match, (clobber (const_int 0)). */
202 /* Record last value assigned to (hard or pseudo) register n. */
204 rtx last_set_value;
206 /* Record the value of label_tick when an expression involving register n
207 is placed in last_set_value. */
209 int last_set_table_tick;
211 /* Record the value of label_tick when the value for register n is placed in
212 last_set_value. */
214 int last_set_label;
216 /* These fields are maintained in parallel with last_set_value and are
217 used to store the mode in which the register was last set, the bits
218 that were known to be zero when it was last set, and the number of
219 sign bits copies it was known to have when it was last set. */
221 unsigned HOST_WIDE_INT last_set_nonzero_bits;
222 char last_set_sign_bit_copies;
223 ENUM_BITFIELD(machine_mode) last_set_mode : 8;
225 /* Set nonzero if references to register n in expressions should not be
226 used. last_set_invalid is set nonzero when this register is being
227 assigned to and last_set_table_tick == label_tick. */
229 char last_set_invalid;
231 /* Some registers that are set more than once and used in more than one
232 basic block are nevertheless always set in similar ways. For example,
233 a QImode register may be loaded from memory in two places on a machine
234 where byte loads zero extend.
236 We record in the following fields if a register has some leading bits
237 that are always equal to the sign bit, and what we know about the
238 nonzero bits of a register, specifically which bits are known to be
239 zero.
241 If an entry is zero, it means that we don't know anything special. */
243 unsigned char sign_bit_copies;
245 unsigned HOST_WIDE_INT nonzero_bits;
247 /* Record the value of the label_tick when the last truncation
248 happened. The field truncated_to_mode is only valid if
249 truncation_label == label_tick. */
251 int truncation_label;
253 /* Record the last truncation seen for this register. If truncation
254 is not a nop to this mode we might be able to save an explicit
255 truncation if we know that value already contains a truncated
256 value. */
258 ENUM_BITFIELD(machine_mode) truncated_to_mode : 8;
261 static struct reg_stat *reg_stat;
263 /* Record the cuid of the last insn that invalidated memory
264 (anything that writes memory, and subroutine calls, but not pushes). */
266 static int mem_last_set;
268 /* Record the cuid of the last CALL_INSN
269 so we can tell whether a potential combination crosses any calls. */
271 static int last_call_cuid;
273 /* When `subst' is called, this is the insn that is being modified
274 (by combining in a previous insn). The PATTERN of this insn
275 is still the old pattern partially modified and it should not be
276 looked at, but this may be used to examine the successors of the insn
277 to judge whether a simplification is valid. */
279 static rtx subst_insn;
281 /* This is the lowest CUID that `subst' is currently dealing with.
282 get_last_value will not return a value if the register was set at or
283 after this CUID. If not for this mechanism, we could get confused if
284 I2 or I1 in try_combine were an insn that used the old value of a register
285 to obtain a new value. In that case, we might erroneously get the
286 new value of the register when we wanted the old one. */
288 static int subst_low_cuid;
290 /* This contains any hard registers that are used in newpat; reg_dead_at_p
291 must consider all these registers to be always live. */
293 static HARD_REG_SET newpat_used_regs;
295 /* This is an insn to which a LOG_LINKS entry has been added. If this
296 insn is the earlier than I2 or I3, combine should rescan starting at
297 that location. */
299 static rtx added_links_insn;
301 /* Basic block in which we are performing combines. */
302 static basic_block this_basic_block;
304 /* A bitmap indicating which blocks had registers go dead at entry.
305 After combine, we'll need to re-do global life analysis with
306 those blocks as starting points. */
307 static sbitmap refresh_blocks;
309 /* The following array records the insn_rtx_cost for every insn
310 in the instruction stream. */
312 static int *uid_insn_cost;
314 /* Length of the currently allocated uid_insn_cost array. */
316 static int last_insn_cost;
318 /* Incremented for each label. */
320 static int label_tick;
322 /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the
323 largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */
325 static enum machine_mode nonzero_bits_mode;
327 /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can
328 be safely used. It is zero while computing them and after combine has
329 completed. This former test prevents propagating values based on
330 previously set values, which can be incorrect if a variable is modified
331 in a loop. */
333 static int nonzero_sign_valid;
336 /* Record one modification to rtl structure
337 to be undone by storing old_contents into *where. */
339 struct undo
341 struct undo *next;
342 enum { UNDO_RTX, UNDO_INT, UNDO_MODE } kind;
343 union { rtx r; int i; enum machine_mode m; } old_contents;
344 union { rtx *r; int *i; } where;
347 /* Record a bunch of changes to be undone, up to MAX_UNDO of them.
348 num_undo says how many are currently recorded.
350 other_insn is nonzero if we have modified some other insn in the process
351 of working on subst_insn. It must be verified too. */
353 struct undobuf
355 struct undo *undos;
356 struct undo *frees;
357 rtx other_insn;
360 static struct undobuf undobuf;
362 /* Number of times the pseudo being substituted for
363 was found and replaced. */
365 static int n_occurrences;
367 static rtx reg_nonzero_bits_for_combine (rtx, enum machine_mode, rtx,
368 enum machine_mode,
369 unsigned HOST_WIDE_INT,
370 unsigned HOST_WIDE_INT *);
371 static rtx reg_num_sign_bit_copies_for_combine (rtx, enum machine_mode, rtx,
372 enum machine_mode,
373 unsigned int, unsigned int *);
374 static void do_SUBST (rtx *, rtx);
375 static void do_SUBST_INT (int *, int);
376 static void init_reg_last (void);
377 static void setup_incoming_promotions (void);
378 static void set_nonzero_bits_and_sign_copies (rtx, rtx, void *);
379 static int cant_combine_insn_p (rtx);
380 static int can_combine_p (rtx, rtx, rtx, rtx, rtx *, rtx *);
381 static int combinable_i3pat (rtx, rtx *, rtx, rtx, int, rtx *);
382 static int contains_muldiv (rtx);
383 static rtx try_combine (rtx, rtx, rtx, int *);
384 static void undo_all (void);
385 static void undo_commit (void);
386 static rtx *find_split_point (rtx *, rtx);
387 static rtx subst (rtx, rtx, rtx, int, int);
388 static rtx combine_simplify_rtx (rtx, enum machine_mode, int);
389 static rtx simplify_if_then_else (rtx);
390 static rtx simplify_set (rtx);
391 static rtx simplify_logical (rtx);
392 static rtx expand_compound_operation (rtx);
393 static rtx expand_field_assignment (rtx);
394 static rtx make_extraction (enum machine_mode, rtx, HOST_WIDE_INT,
395 rtx, unsigned HOST_WIDE_INT, int, int, int);
396 static rtx extract_left_shift (rtx, int);
397 static rtx make_compound_operation (rtx, enum rtx_code);
398 static int get_pos_from_mask (unsigned HOST_WIDE_INT,
399 unsigned HOST_WIDE_INT *);
400 static rtx canon_reg_for_combine (rtx, rtx);
401 static rtx force_to_mode (rtx, enum machine_mode,
402 unsigned HOST_WIDE_INT, int);
403 static rtx if_then_else_cond (rtx, rtx *, rtx *);
404 static rtx known_cond (rtx, enum rtx_code, rtx, rtx);
405 static int rtx_equal_for_field_assignment_p (rtx, rtx);
406 static rtx make_field_assignment (rtx);
407 static rtx apply_distributive_law (rtx);
408 static rtx distribute_and_simplify_rtx (rtx, int);
409 static rtx simplify_and_const_int_1 (enum machine_mode, rtx,
410 unsigned HOST_WIDE_INT);
411 static rtx simplify_and_const_int (rtx, enum machine_mode, rtx,
412 unsigned HOST_WIDE_INT);
413 static int merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code,
414 HOST_WIDE_INT, enum machine_mode, int *);
415 static rtx simplify_shift_const_1 (enum rtx_code, enum machine_mode, rtx, int);
416 static rtx simplify_shift_const (rtx, enum rtx_code, enum machine_mode, rtx,
417 int);
418 static int recog_for_combine (rtx *, rtx, rtx *);
419 static rtx gen_lowpart_for_combine (enum machine_mode, rtx);
420 static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *);
421 static void update_table_tick (rtx);
422 static void record_value_for_reg (rtx, rtx, rtx);
423 static void check_conversions (rtx, rtx);
424 static void record_dead_and_set_regs_1 (rtx, rtx, void *);
425 static void record_dead_and_set_regs (rtx);
426 static int get_last_value_validate (rtx *, rtx, int, int);
427 static rtx get_last_value (rtx);
428 static int use_crosses_set_p (rtx, int);
429 static void reg_dead_at_p_1 (rtx, rtx, void *);
430 static int reg_dead_at_p (rtx, rtx);
431 static void move_deaths (rtx, rtx, int, rtx, rtx *);
432 static int reg_bitfield_target_p (rtx, rtx);
433 static void distribute_notes (rtx, rtx, rtx, rtx, rtx, rtx);
434 static void distribute_links (rtx);
435 static void mark_used_regs_combine (rtx);
436 static int insn_cuid (rtx);
437 static void record_promoted_value (rtx, rtx);
438 static int unmentioned_reg_p_1 (rtx *, void *);
439 static bool unmentioned_reg_p (rtx, rtx);
440 static void record_truncated_value (rtx);
441 static bool reg_truncated_to_mode (enum machine_mode, rtx);
442 static rtx gen_lowpart_or_truncate (enum machine_mode, rtx);
445 /* It is not safe to use ordinary gen_lowpart in combine.
446 See comments in gen_lowpart_for_combine. */
447 #undef RTL_HOOKS_GEN_LOWPART
448 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_for_combine
450 /* Our implementation of gen_lowpart never emits a new pseudo. */
451 #undef RTL_HOOKS_GEN_LOWPART_NO_EMIT
452 #define RTL_HOOKS_GEN_LOWPART_NO_EMIT gen_lowpart_for_combine
454 #undef RTL_HOOKS_REG_NONZERO_REG_BITS
455 #define RTL_HOOKS_REG_NONZERO_REG_BITS reg_nonzero_bits_for_combine
457 #undef RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES
458 #define RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES reg_num_sign_bit_copies_for_combine
460 static const struct rtl_hooks combine_rtl_hooks = RTL_HOOKS_INITIALIZER;
463 /* Substitute NEWVAL, an rtx expression, into INTO, a place in some
464 insn. The substitution can be undone by undo_all. If INTO is already
465 set to NEWVAL, do not record this change. Because computing NEWVAL might
466 also call SUBST, we have to compute it before we put anything into
467 the undo table. */
469 static void
470 do_SUBST (rtx *into, rtx newval)
472 struct undo *buf;
473 rtx oldval = *into;
475 if (oldval == newval)
476 return;
478 /* We'd like to catch as many invalid transformations here as
479 possible. Unfortunately, there are way too many mode changes
480 that are perfectly valid, so we'd waste too much effort for
481 little gain doing the checks here. Focus on catching invalid
482 transformations involving integer constants. */
483 if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT
484 && GET_CODE (newval) == CONST_INT)
486 /* Sanity check that we're replacing oldval with a CONST_INT
487 that is a valid sign-extension for the original mode. */
488 gcc_assert (INTVAL (newval)
489 == trunc_int_for_mode (INTVAL (newval), GET_MODE (oldval)));
491 /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
492 CONST_INT is not valid, because after the replacement, the
493 original mode would be gone. Unfortunately, we can't tell
494 when do_SUBST is called to replace the operand thereof, so we
495 perform this test on oldval instead, checking whether an
496 invalid replacement took place before we got here. */
497 gcc_assert (!(GET_CODE (oldval) == SUBREG
498 && GET_CODE (SUBREG_REG (oldval)) == CONST_INT));
499 gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND
500 && GET_CODE (XEXP (oldval, 0)) == CONST_INT));
503 if (undobuf.frees)
504 buf = undobuf.frees, undobuf.frees = buf->next;
505 else
506 buf = XNEW (struct undo);
508 buf->kind = UNDO_RTX;
509 buf->where.r = into;
510 buf->old_contents.r = oldval;
511 *into = newval;
513 buf->next = undobuf.undos, undobuf.undos = buf;
516 #define SUBST(INTO, NEWVAL) do_SUBST(&(INTO), (NEWVAL))
518 /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution
519 for the value of a HOST_WIDE_INT value (including CONST_INT) is
520 not safe. */
522 static void
523 do_SUBST_INT (int *into, int newval)
525 struct undo *buf;
526 int oldval = *into;
528 if (oldval == newval)
529 return;
531 if (undobuf.frees)
532 buf = undobuf.frees, undobuf.frees = buf->next;
533 else
534 buf = XNEW (struct undo);
536 buf->kind = UNDO_INT;
537 buf->where.i = into;
538 buf->old_contents.i = oldval;
539 *into = newval;
541 buf->next = undobuf.undos, undobuf.undos = buf;
544 #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT(&(INTO), (NEWVAL))
546 /* Similar to SUBST, but just substitute the mode. This is used when
547 changing the mode of a pseudo-register, so that any other
548 references to the entry in the regno_reg_rtx array will change as
549 well. */
551 static void
552 do_SUBST_MODE (rtx *into, enum machine_mode newval)
554 struct undo *buf;
555 enum machine_mode oldval = GET_MODE (*into);
557 if (oldval == newval)
558 return;
560 if (undobuf.frees)
561 buf = undobuf.frees, undobuf.frees = buf->next;
562 else
563 buf = XNEW (struct undo);
565 buf->kind = UNDO_MODE;
566 buf->where.r = into;
567 buf->old_contents.m = oldval;
568 PUT_MODE (*into, newval);
570 buf->next = undobuf.undos, undobuf.undos = buf;
573 #define SUBST_MODE(INTO, NEWVAL) do_SUBST_MODE(&(INTO), (NEWVAL))
575 /* Subroutine of try_combine. Determine whether the combine replacement
576 patterns NEWPAT and NEWI2PAT are cheaper according to insn_rtx_cost
577 that the original instruction sequence I1, I2 and I3. Note that I1
578 and/or NEWI2PAT may be NULL_RTX. This function returns false, if the
579 costs of all instructions can be estimated, and the replacements are
580 more expensive than the original sequence. */
582 static bool
583 combine_validate_cost (rtx i1, rtx i2, rtx i3, rtx newpat, rtx newi2pat)
585 int i1_cost, i2_cost, i3_cost;
586 int new_i2_cost, new_i3_cost;
587 int old_cost, new_cost;
589 /* Lookup the original insn_rtx_costs. */
590 i2_cost = INSN_UID (i2) <= last_insn_cost
591 ? uid_insn_cost[INSN_UID (i2)] : 0;
592 i3_cost = INSN_UID (i3) <= last_insn_cost
593 ? uid_insn_cost[INSN_UID (i3)] : 0;
595 if (i1)
597 i1_cost = INSN_UID (i1) <= last_insn_cost
598 ? uid_insn_cost[INSN_UID (i1)] : 0;
599 old_cost = (i1_cost > 0 && i2_cost > 0 && i3_cost > 0)
600 ? i1_cost + i2_cost + i3_cost : 0;
602 else
604 old_cost = (i2_cost > 0 && i3_cost > 0) ? i2_cost + i3_cost : 0;
605 i1_cost = 0;
608 /* Calculate the replacement insn_rtx_costs. */
609 new_i3_cost = insn_rtx_cost (newpat);
610 if (newi2pat)
612 new_i2_cost = insn_rtx_cost (newi2pat);
613 new_cost = (new_i2_cost > 0 && new_i3_cost > 0)
614 ? new_i2_cost + new_i3_cost : 0;
616 else
618 new_cost = new_i3_cost;
619 new_i2_cost = 0;
622 if (undobuf.other_insn)
624 int old_other_cost, new_other_cost;
626 old_other_cost = (INSN_UID (undobuf.other_insn) <= last_insn_cost
627 ? uid_insn_cost[INSN_UID (undobuf.other_insn)] : 0);
628 new_other_cost = insn_rtx_cost (PATTERN (undobuf.other_insn));
629 if (old_other_cost > 0 && new_other_cost > 0)
631 old_cost += old_other_cost;
632 new_cost += new_other_cost;
634 else
635 old_cost = 0;
638 /* Disallow this recombination if both new_cost and old_cost are
639 greater than zero, and new_cost is greater than old cost. */
640 if (old_cost > 0
641 && new_cost > old_cost)
643 if (dump_file)
645 if (i1)
647 fprintf (dump_file,
648 "rejecting combination of insns %d, %d and %d\n",
649 INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
650 fprintf (dump_file, "original costs %d + %d + %d = %d\n",
651 i1_cost, i2_cost, i3_cost, old_cost);
653 else
655 fprintf (dump_file,
656 "rejecting combination of insns %d and %d\n",
657 INSN_UID (i2), INSN_UID (i3));
658 fprintf (dump_file, "original costs %d + %d = %d\n",
659 i2_cost, i3_cost, old_cost);
662 if (newi2pat)
664 fprintf (dump_file, "replacement costs %d + %d = %d\n",
665 new_i2_cost, new_i3_cost, new_cost);
667 else
668 fprintf (dump_file, "replacement cost %d\n", new_cost);
671 return false;
674 /* Update the uid_insn_cost array with the replacement costs. */
675 uid_insn_cost[INSN_UID (i2)] = new_i2_cost;
676 uid_insn_cost[INSN_UID (i3)] = new_i3_cost;
677 if (i1)
678 uid_insn_cost[INSN_UID (i1)] = 0;
680 return true;
683 /* Main entry point for combiner. F is the first insn of the function.
684 NREGS is the first unused pseudo-reg number.
686 Return nonzero if the combiner has turned an indirect jump
687 instruction into a direct jump. */
688 static int
689 combine_instructions (rtx f, unsigned int nregs)
691 rtx insn, next;
692 #ifdef HAVE_cc0
693 rtx prev;
694 #endif
695 int i;
696 unsigned int j = 0;
697 rtx links, nextlinks;
698 sbitmap_iterator sbi;
700 int new_direct_jump_p = 0;
702 combine_attempts = 0;
703 combine_merges = 0;
704 combine_extras = 0;
705 combine_successes = 0;
707 combine_max_regno = nregs;
709 rtl_hooks = combine_rtl_hooks;
711 reg_stat = XCNEWVEC (struct reg_stat, nregs);
713 init_recog_no_volatile ();
715 /* Compute maximum uid value so uid_cuid can be allocated. */
717 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
718 if (INSN_UID (insn) > i)
719 i = INSN_UID (insn);
721 uid_cuid = XNEWVEC (int, i + 1);
722 max_uid_cuid = i;
724 nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0);
726 /* Don't use reg_stat[].nonzero_bits when computing it. This can cause
727 problems when, for example, we have j <<= 1 in a loop. */
729 nonzero_sign_valid = 0;
731 /* Compute the mapping from uids to cuids.
732 Cuids are numbers assigned to insns, like uids,
733 except that cuids increase monotonically through the code.
735 Scan all SETs and see if we can deduce anything about what
736 bits are known to be zero for some registers and how many copies
737 of the sign bit are known to exist for those registers.
739 Also set any known values so that we can use it while searching
740 for what bits are known to be set. */
742 label_tick = 1;
744 setup_incoming_promotions ();
746 refresh_blocks = sbitmap_alloc (last_basic_block);
747 sbitmap_zero (refresh_blocks);
749 /* Allocate array of current insn_rtx_costs. */
750 uid_insn_cost = XCNEWVEC (int, max_uid_cuid + 1);
751 last_insn_cost = max_uid_cuid;
753 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
755 uid_cuid[INSN_UID (insn)] = ++i;
756 subst_low_cuid = i;
757 subst_insn = insn;
759 if (INSN_P (insn))
761 note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies,
762 NULL);
763 record_dead_and_set_regs (insn);
765 #ifdef AUTO_INC_DEC
766 for (links = REG_NOTES (insn); links; links = XEXP (links, 1))
767 if (REG_NOTE_KIND (links) == REG_INC)
768 set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX,
769 NULL);
770 #endif
772 /* Record the current insn_rtx_cost of this instruction. */
773 if (NONJUMP_INSN_P (insn))
774 uid_insn_cost[INSN_UID (insn)] = insn_rtx_cost (PATTERN (insn));
775 if (dump_file)
776 fprintf(dump_file, "insn_cost %d: %d\n",
777 INSN_UID (insn), uid_insn_cost[INSN_UID (insn)]);
780 if (LABEL_P (insn))
781 label_tick++;
784 nonzero_sign_valid = 1;
786 /* Now scan all the insns in forward order. */
788 label_tick = 1;
789 last_call_cuid = 0;
790 mem_last_set = 0;
791 init_reg_last ();
792 setup_incoming_promotions ();
794 FOR_EACH_BB (this_basic_block)
796 for (insn = BB_HEAD (this_basic_block);
797 insn != NEXT_INSN (BB_END (this_basic_block));
798 insn = next ? next : NEXT_INSN (insn))
800 next = 0;
802 if (LABEL_P (insn))
803 label_tick++;
805 else if (INSN_P (insn))
807 /* See if we know about function return values before this
808 insn based upon SUBREG flags. */
809 check_conversions (insn, PATTERN (insn));
811 /* Try this insn with each insn it links back to. */
813 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
814 if ((next = try_combine (insn, XEXP (links, 0),
815 NULL_RTX, &new_direct_jump_p)) != 0)
816 goto retry;
818 /* Try each sequence of three linked insns ending with this one. */
820 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
822 rtx link = XEXP (links, 0);
824 /* If the linked insn has been replaced by a note, then there
825 is no point in pursuing this chain any further. */
826 if (NOTE_P (link))
827 continue;
829 for (nextlinks = LOG_LINKS (link);
830 nextlinks;
831 nextlinks = XEXP (nextlinks, 1))
832 if ((next = try_combine (insn, link,
833 XEXP (nextlinks, 0),
834 &new_direct_jump_p)) != 0)
835 goto retry;
838 #ifdef HAVE_cc0
839 /* Try to combine a jump insn that uses CC0
840 with a preceding insn that sets CC0, and maybe with its
841 logical predecessor as well.
842 This is how we make decrement-and-branch insns.
843 We need this special code because data flow connections
844 via CC0 do not get entered in LOG_LINKS. */
846 if (JUMP_P (insn)
847 && (prev = prev_nonnote_insn (insn)) != 0
848 && NONJUMP_INSN_P (prev)
849 && sets_cc0_p (PATTERN (prev)))
851 if ((next = try_combine (insn, prev,
852 NULL_RTX, &new_direct_jump_p)) != 0)
853 goto retry;
855 for (nextlinks = LOG_LINKS (prev); nextlinks;
856 nextlinks = XEXP (nextlinks, 1))
857 if ((next = try_combine (insn, prev,
858 XEXP (nextlinks, 0),
859 &new_direct_jump_p)) != 0)
860 goto retry;
863 /* Do the same for an insn that explicitly references CC0. */
864 if (NONJUMP_INSN_P (insn)
865 && (prev = prev_nonnote_insn (insn)) != 0
866 && NONJUMP_INSN_P (prev)
867 && sets_cc0_p (PATTERN (prev))
868 && GET_CODE (PATTERN (insn)) == SET
869 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn))))
871 if ((next = try_combine (insn, prev,
872 NULL_RTX, &new_direct_jump_p)) != 0)
873 goto retry;
875 for (nextlinks = LOG_LINKS (prev); nextlinks;
876 nextlinks = XEXP (nextlinks, 1))
877 if ((next = try_combine (insn, prev,
878 XEXP (nextlinks, 0),
879 &new_direct_jump_p)) != 0)
880 goto retry;
883 /* Finally, see if any of the insns that this insn links to
884 explicitly references CC0. If so, try this insn, that insn,
885 and its predecessor if it sets CC0. */
886 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
887 if (NONJUMP_INSN_P (XEXP (links, 0))
888 && GET_CODE (PATTERN (XEXP (links, 0))) == SET
889 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0))))
890 && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0
891 && NONJUMP_INSN_P (prev)
892 && sets_cc0_p (PATTERN (prev))
893 && (next = try_combine (insn, XEXP (links, 0),
894 prev, &new_direct_jump_p)) != 0)
895 goto retry;
896 #endif
898 /* Try combining an insn with two different insns whose results it
899 uses. */
900 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
901 for (nextlinks = XEXP (links, 1); nextlinks;
902 nextlinks = XEXP (nextlinks, 1))
903 if ((next = try_combine (insn, XEXP (links, 0),
904 XEXP (nextlinks, 0),
905 &new_direct_jump_p)) != 0)
906 goto retry;
908 /* Try this insn with each REG_EQUAL note it links back to. */
909 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
911 rtx set, note;
912 rtx temp = XEXP (links, 0);
913 if ((set = single_set (temp)) != 0
914 && (note = find_reg_equal_equiv_note (temp)) != 0
915 && (note = XEXP (note, 0), GET_CODE (note)) != EXPR_LIST
916 /* Avoid using a register that may already been marked
917 dead by an earlier instruction. */
918 && ! unmentioned_reg_p (note, SET_SRC (set))
919 && (GET_MODE (note) == VOIDmode
920 ? SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set)))
921 : GET_MODE (SET_DEST (set)) == GET_MODE (note)))
923 /* Temporarily replace the set's source with the
924 contents of the REG_EQUAL note. The insn will
925 be deleted or recognized by try_combine. */
926 rtx orig = SET_SRC (set);
927 SET_SRC (set) = note;
928 next = try_combine (insn, temp, NULL_RTX,
929 &new_direct_jump_p);
930 if (next)
931 goto retry;
932 SET_SRC (set) = orig;
936 if (!NOTE_P (insn))
937 record_dead_and_set_regs (insn);
939 retry:
944 clear_bb_flags ();
946 EXECUTE_IF_SET_IN_SBITMAP (refresh_blocks, 0, j, sbi)
947 BASIC_BLOCK (j)->flags |= BB_DIRTY;
948 new_direct_jump_p |= purge_all_dead_edges ();
949 delete_noop_moves ();
951 update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
952 PROP_DEATH_NOTES | PROP_SCAN_DEAD_CODE
953 | PROP_KILL_DEAD_CODE);
955 /* Clean up. */
956 sbitmap_free (refresh_blocks);
957 free (uid_insn_cost);
958 free (reg_stat);
959 free (uid_cuid);
962 struct undo *undo, *next;
963 for (undo = undobuf.frees; undo; undo = next)
965 next = undo->next;
966 free (undo);
968 undobuf.frees = 0;
971 total_attempts += combine_attempts;
972 total_merges += combine_merges;
973 total_extras += combine_extras;
974 total_successes += combine_successes;
976 nonzero_sign_valid = 0;
977 rtl_hooks = general_rtl_hooks;
979 /* Make recognizer allow volatile MEMs again. */
980 init_recog ();
982 return new_direct_jump_p;
985 /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */
987 static void
988 init_reg_last (void)
990 unsigned int i;
991 for (i = 0; i < combine_max_regno; i++)
992 memset (reg_stat + i, 0, offsetof (struct reg_stat, sign_bit_copies));
995 /* Set up any promoted values for incoming argument registers. */
997 static void
998 setup_incoming_promotions (void)
1000 unsigned int regno;
1001 rtx reg;
1002 enum machine_mode mode;
1003 int unsignedp;
1004 rtx first = get_insns ();
1006 if (targetm.calls.promote_function_args (TREE_TYPE (cfun->decl)))
1008 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1009 /* Check whether this register can hold an incoming pointer
1010 argument. FUNCTION_ARG_REGNO_P tests outgoing register
1011 numbers, so translate if necessary due to register windows. */
1012 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (regno))
1013 && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0)
1015 record_value_for_reg
1016 (reg, first, gen_rtx_fmt_e ((unsignedp ? ZERO_EXTEND
1017 : SIGN_EXTEND),
1018 GET_MODE (reg),
1019 gen_rtx_CLOBBER (mode, const0_rtx)));
1024 /* Called via note_stores. If X is a pseudo that is narrower than
1025 HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero.
1027 If we are setting only a portion of X and we can't figure out what
1028 portion, assume all bits will be used since we don't know what will
1029 be happening.
1031 Similarly, set how many bits of X are known to be copies of the sign bit
1032 at all locations in the function. This is the smallest number implied
1033 by any set of X. */
1035 static void
1036 set_nonzero_bits_and_sign_copies (rtx x, rtx set,
1037 void *data ATTRIBUTE_UNUSED)
1039 unsigned int num;
1041 if (REG_P (x)
1042 && REGNO (x) >= FIRST_PSEUDO_REGISTER
1043 /* If this register is undefined at the start of the file, we can't
1044 say what its contents were. */
1045 && ! REGNO_REG_SET_P
1046 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start, REGNO (x))
1047 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
1049 if (set == 0 || GET_CODE (set) == CLOBBER)
1051 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1052 reg_stat[REGNO (x)].sign_bit_copies = 1;
1053 return;
1056 /* If this is a complex assignment, see if we can convert it into a
1057 simple assignment. */
1058 set = expand_field_assignment (set);
1060 /* If this is a simple assignment, or we have a paradoxical SUBREG,
1061 set what we know about X. */
1063 if (SET_DEST (set) == x
1064 || (GET_CODE (SET_DEST (set)) == SUBREG
1065 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
1066 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set)))))
1067 && SUBREG_REG (SET_DEST (set)) == x))
1069 rtx src = SET_SRC (set);
1071 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
1072 /* If X is narrower than a word and SRC is a non-negative
1073 constant that would appear negative in the mode of X,
1074 sign-extend it for use in reg_stat[].nonzero_bits because some
1075 machines (maybe most) will actually do the sign-extension
1076 and this is the conservative approach.
1078 ??? For 2.5, try to tighten up the MD files in this regard
1079 instead of this kludge. */
1081 if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD
1082 && GET_CODE (src) == CONST_INT
1083 && INTVAL (src) > 0
1084 && 0 != (INTVAL (src)
1085 & ((HOST_WIDE_INT) 1
1086 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
1087 src = GEN_INT (INTVAL (src)
1088 | ((HOST_WIDE_INT) (-1)
1089 << GET_MODE_BITSIZE (GET_MODE (x))));
1090 #endif
1092 /* Don't call nonzero_bits if it cannot change anything. */
1093 if (reg_stat[REGNO (x)].nonzero_bits != ~(unsigned HOST_WIDE_INT) 0)
1094 reg_stat[REGNO (x)].nonzero_bits
1095 |= nonzero_bits (src, nonzero_bits_mode);
1096 num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x));
1097 if (reg_stat[REGNO (x)].sign_bit_copies == 0
1098 || reg_stat[REGNO (x)].sign_bit_copies > num)
1099 reg_stat[REGNO (x)].sign_bit_copies = num;
1101 else
1103 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1104 reg_stat[REGNO (x)].sign_bit_copies = 1;
1109 /* See if INSN can be combined into I3. PRED and SUCC are optionally
1110 insns that were previously combined into I3 or that will be combined
1111 into the merger of INSN and I3.
1113 Return 0 if the combination is not allowed for any reason.
1115 If the combination is allowed, *PDEST will be set to the single
1116 destination of INSN and *PSRC to the single source, and this function
1117 will return 1. */
1119 static int
1120 can_combine_p (rtx insn, rtx i3, rtx pred ATTRIBUTE_UNUSED, rtx succ,
1121 rtx *pdest, rtx *psrc)
1123 int i;
1124 rtx set = 0, src, dest;
1125 rtx p;
1126 #ifdef AUTO_INC_DEC
1127 rtx link;
1128 #endif
1129 int all_adjacent = (succ ? (next_active_insn (insn) == succ
1130 && next_active_insn (succ) == i3)
1131 : next_active_insn (insn) == i3);
1133 /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0.
1134 or a PARALLEL consisting of such a SET and CLOBBERs.
1136 If INSN has CLOBBER parallel parts, ignore them for our processing.
1137 By definition, these happen during the execution of the insn. When it
1138 is merged with another insn, all bets are off. If they are, in fact,
1139 needed and aren't also supplied in I3, they may be added by
1140 recog_for_combine. Otherwise, it won't match.
1142 We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED
1143 note.
1145 Get the source and destination of INSN. If more than one, can't
1146 combine. */
1148 if (GET_CODE (PATTERN (insn)) == SET)
1149 set = PATTERN (insn);
1150 else if (GET_CODE (PATTERN (insn)) == PARALLEL
1151 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET)
1153 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
1155 rtx elt = XVECEXP (PATTERN (insn), 0, i);
1156 rtx note;
1158 switch (GET_CODE (elt))
1160 /* This is important to combine floating point insns
1161 for the SH4 port. */
1162 case USE:
1163 /* Combining an isolated USE doesn't make sense.
1164 We depend here on combinable_i3pat to reject them. */
1165 /* The code below this loop only verifies that the inputs of
1166 the SET in INSN do not change. We call reg_set_between_p
1167 to verify that the REG in the USE does not change between
1168 I3 and INSN.
1169 If the USE in INSN was for a pseudo register, the matching
1170 insn pattern will likely match any register; combining this
1171 with any other USE would only be safe if we knew that the
1172 used registers have identical values, or if there was
1173 something to tell them apart, e.g. different modes. For
1174 now, we forgo such complicated tests and simply disallow
1175 combining of USES of pseudo registers with any other USE. */
1176 if (REG_P (XEXP (elt, 0))
1177 && GET_CODE (PATTERN (i3)) == PARALLEL)
1179 rtx i3pat = PATTERN (i3);
1180 int i = XVECLEN (i3pat, 0) - 1;
1181 unsigned int regno = REGNO (XEXP (elt, 0));
1185 rtx i3elt = XVECEXP (i3pat, 0, i);
1187 if (GET_CODE (i3elt) == USE
1188 && REG_P (XEXP (i3elt, 0))
1189 && (REGNO (XEXP (i3elt, 0)) == regno
1190 ? reg_set_between_p (XEXP (elt, 0),
1191 PREV_INSN (insn), i3)
1192 : regno >= FIRST_PSEUDO_REGISTER))
1193 return 0;
1195 while (--i >= 0);
1197 break;
1199 /* We can ignore CLOBBERs. */
1200 case CLOBBER:
1201 break;
1203 case SET:
1204 /* Ignore SETs whose result isn't used but not those that
1205 have side-effects. */
1206 if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt))
1207 && (!(note = find_reg_note (insn, REG_EH_REGION, NULL_RTX))
1208 || INTVAL (XEXP (note, 0)) <= 0)
1209 && ! side_effects_p (elt))
1210 break;
1212 /* If we have already found a SET, this is a second one and
1213 so we cannot combine with this insn. */
1214 if (set)
1215 return 0;
1217 set = elt;
1218 break;
1220 default:
1221 /* Anything else means we can't combine. */
1222 return 0;
1226 if (set == 0
1227 /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs,
1228 so don't do anything with it. */
1229 || GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1230 return 0;
1232 else
1233 return 0;
1235 if (set == 0)
1236 return 0;
1238 set = expand_field_assignment (set);
1239 src = SET_SRC (set), dest = SET_DEST (set);
1241 /* Don't eliminate a store in the stack pointer. */
1242 if (dest == stack_pointer_rtx
1243 /* Don't combine with an insn that sets a register to itself if it has
1244 a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */
1245 || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX))
1246 /* Can't merge an ASM_OPERANDS. */
1247 || GET_CODE (src) == ASM_OPERANDS
1248 /* Can't merge a function call. */
1249 || GET_CODE (src) == CALL
1250 /* Don't eliminate a function call argument. */
1251 || (CALL_P (i3)
1252 && (find_reg_fusage (i3, USE, dest)
1253 || (REG_P (dest)
1254 && REGNO (dest) < FIRST_PSEUDO_REGISTER
1255 && global_regs[REGNO (dest)])))
1256 /* Don't substitute into an incremented register. */
1257 || FIND_REG_INC_NOTE (i3, dest)
1258 || (succ && FIND_REG_INC_NOTE (succ, dest))
1259 /* Don't substitute into a non-local goto, this confuses CFG. */
1260 || (JUMP_P (i3) && find_reg_note (i3, REG_NON_LOCAL_GOTO, NULL_RTX))
1261 #if 0
1262 /* Don't combine the end of a libcall into anything. */
1263 /* ??? This gives worse code, and appears to be unnecessary, since no
1264 pass after flow uses REG_LIBCALL/REG_RETVAL notes. Local-alloc does
1265 use REG_RETVAL notes for noconflict blocks, but other code here
1266 makes sure that those insns don't disappear. */
1267 || find_reg_note (insn, REG_RETVAL, NULL_RTX)
1268 #endif
1269 /* Make sure that DEST is not used after SUCC but before I3. */
1270 || (succ && ! all_adjacent
1271 && reg_used_between_p (dest, succ, i3))
1272 /* Make sure that the value that is to be substituted for the register
1273 does not use any registers whose values alter in between. However,
1274 If the insns are adjacent, a use can't cross a set even though we
1275 think it might (this can happen for a sequence of insns each setting
1276 the same destination; last_set of that register might point to
1277 a NOTE). If INSN has a REG_EQUIV note, the register is always
1278 equivalent to the memory so the substitution is valid even if there
1279 are intervening stores. Also, don't move a volatile asm or
1280 UNSPEC_VOLATILE across any other insns. */
1281 || (! all_adjacent
1282 && (((!MEM_P (src)
1283 || ! find_reg_note (insn, REG_EQUIV, src))
1284 && use_crosses_set_p (src, INSN_CUID (insn)))
1285 || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src))
1286 || GET_CODE (src) == UNSPEC_VOLATILE))
1287 /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get
1288 better register allocation by not doing the combine. */
1289 || find_reg_note (i3, REG_NO_CONFLICT, dest)
1290 || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest))
1291 /* Don't combine across a CALL_INSN, because that would possibly
1292 change whether the life span of some REGs crosses calls or not,
1293 and it is a pain to update that information.
1294 Exception: if source is a constant, moving it later can't hurt.
1295 Accept that special case, because it helps -fforce-addr a lot. */
1296 || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src)))
1297 return 0;
1299 /* DEST must either be a REG or CC0. */
1300 if (REG_P (dest))
1302 /* If register alignment is being enforced for multi-word items in all
1303 cases except for parameters, it is possible to have a register copy
1304 insn referencing a hard register that is not allowed to contain the
1305 mode being copied and which would not be valid as an operand of most
1306 insns. Eliminate this problem by not combining with such an insn.
1308 Also, on some machines we don't want to extend the life of a hard
1309 register. */
1311 if (REG_P (src)
1312 && ((REGNO (dest) < FIRST_PSEUDO_REGISTER
1313 && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest)))
1314 /* Don't extend the life of a hard register unless it is
1315 user variable (if we have few registers) or it can't
1316 fit into the desired register (meaning something special
1317 is going on).
1318 Also avoid substituting a return register into I3, because
1319 reload can't handle a conflict with constraints of other
1320 inputs. */
1321 || (REGNO (src) < FIRST_PSEUDO_REGISTER
1322 && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)))))
1323 return 0;
1325 else if (GET_CODE (dest) != CC0)
1326 return 0;
1329 if (GET_CODE (PATTERN (i3)) == PARALLEL)
1330 for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--)
1331 if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER)
1333 /* Don't substitute for a register intended as a clobberable
1334 operand. */
1335 rtx reg = XEXP (XVECEXP (PATTERN (i3), 0, i), 0);
1336 if (rtx_equal_p (reg, dest))
1337 return 0;
1339 /* If the clobber represents an earlyclobber operand, we must not
1340 substitute an expression containing the clobbered register.
1341 As we do not analyze the constraint strings here, we have to
1342 make the conservative assumption. However, if the register is
1343 a fixed hard reg, the clobber cannot represent any operand;
1344 we leave it up to the machine description to either accept or
1345 reject use-and-clobber patterns. */
1346 if (!REG_P (reg)
1347 || REGNO (reg) >= FIRST_PSEUDO_REGISTER
1348 || !fixed_regs[REGNO (reg)])
1349 if (reg_overlap_mentioned_p (reg, src))
1350 return 0;
1353 /* If INSN contains anything volatile, or is an `asm' (whether volatile
1354 or not), reject, unless nothing volatile comes between it and I3 */
1356 if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src))
1358 /* Make sure succ doesn't contain a volatile reference. */
1359 if (succ != 0 && volatile_refs_p (PATTERN (succ)))
1360 return 0;
1362 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1363 if (INSN_P (p) && p != succ && volatile_refs_p (PATTERN (p)))
1364 return 0;
1367 /* If INSN is an asm, and DEST is a hard register, reject, since it has
1368 to be an explicit register variable, and was chosen for a reason. */
1370 if (GET_CODE (src) == ASM_OPERANDS
1371 && REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER)
1372 return 0;
1374 /* If there are any volatile insns between INSN and I3, reject, because
1375 they might affect machine state. */
1377 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1378 if (INSN_P (p) && p != succ && volatile_insn_p (PATTERN (p)))
1379 return 0;
1381 /* If INSN contains an autoincrement or autodecrement, make sure that
1382 register is not used between there and I3, and not already used in
1383 I3 either. Neither must it be used in PRED or SUCC, if they exist.
1384 Also insist that I3 not be a jump; if it were one
1385 and the incremented register were spilled, we would lose. */
1387 #ifdef AUTO_INC_DEC
1388 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
1389 if (REG_NOTE_KIND (link) == REG_INC
1390 && (JUMP_P (i3)
1391 || reg_used_between_p (XEXP (link, 0), insn, i3)
1392 || (pred != NULL_RTX
1393 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (pred)))
1394 || (succ != NULL_RTX
1395 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (succ)))
1396 || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3))))
1397 return 0;
1398 #endif
1400 #ifdef HAVE_cc0
1401 /* Don't combine an insn that follows a CC0-setting insn.
1402 An insn that uses CC0 must not be separated from the one that sets it.
1403 We do, however, allow I2 to follow a CC0-setting insn if that insn
1404 is passed as I1; in that case it will be deleted also.
1405 We also allow combining in this case if all the insns are adjacent
1406 because that would leave the two CC0 insns adjacent as well.
1407 It would be more logical to test whether CC0 occurs inside I1 or I2,
1408 but that would be much slower, and this ought to be equivalent. */
1410 p = prev_nonnote_insn (insn);
1411 if (p && p != pred && NONJUMP_INSN_P (p) && sets_cc0_p (PATTERN (p))
1412 && ! all_adjacent)
1413 return 0;
1414 #endif
1416 /* If we get here, we have passed all the tests and the combination is
1417 to be allowed. */
1419 *pdest = dest;
1420 *psrc = src;
1422 return 1;
1425 /* LOC is the location within I3 that contains its pattern or the component
1426 of a PARALLEL of the pattern. We validate that it is valid for combining.
1428 One problem is if I3 modifies its output, as opposed to replacing it
1429 entirely, we can't allow the output to contain I2DEST or I1DEST as doing
1430 so would produce an insn that is not equivalent to the original insns.
1432 Consider:
1434 (set (reg:DI 101) (reg:DI 100))
1435 (set (subreg:SI (reg:DI 101) 0) <foo>)
1437 This is NOT equivalent to:
1439 (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>)
1440 (set (reg:DI 101) (reg:DI 100))])
1442 Not only does this modify 100 (in which case it might still be valid
1443 if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100.
1445 We can also run into a problem if I2 sets a register that I1
1446 uses and I1 gets directly substituted into I3 (not via I2). In that
1447 case, we would be getting the wrong value of I2DEST into I3, so we
1448 must reject the combination. This case occurs when I2 and I1 both
1449 feed into I3, rather than when I1 feeds into I2, which feeds into I3.
1450 If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source
1451 of a SET must prevent combination from occurring.
1453 Before doing the above check, we first try to expand a field assignment
1454 into a set of logical operations.
1456 If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which
1457 we place a register that is both set and used within I3. If more than one
1458 such register is detected, we fail.
1460 Return 1 if the combination is valid, zero otherwise. */
1462 static int
1463 combinable_i3pat (rtx i3, rtx *loc, rtx i2dest, rtx i1dest,
1464 int i1_not_in_src, rtx *pi3dest_killed)
1466 rtx x = *loc;
1468 if (GET_CODE (x) == SET)
1470 rtx set = x ;
1471 rtx dest = SET_DEST (set);
1472 rtx src = SET_SRC (set);
1473 rtx inner_dest = dest;
1474 rtx subdest;
1476 while (GET_CODE (inner_dest) == STRICT_LOW_PART
1477 || GET_CODE (inner_dest) == SUBREG
1478 || GET_CODE (inner_dest) == ZERO_EXTRACT)
1479 inner_dest = XEXP (inner_dest, 0);
1481 /* Check for the case where I3 modifies its output, as discussed
1482 above. We don't want to prevent pseudos from being combined
1483 into the address of a MEM, so only prevent the combination if
1484 i1 or i2 set the same MEM. */
1485 if ((inner_dest != dest &&
1486 (!MEM_P (inner_dest)
1487 || rtx_equal_p (i2dest, inner_dest)
1488 || (i1dest && rtx_equal_p (i1dest, inner_dest)))
1489 && (reg_overlap_mentioned_p (i2dest, inner_dest)
1490 || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest))))
1492 /* This is the same test done in can_combine_p except we can't test
1493 all_adjacent; we don't have to, since this instruction will stay
1494 in place, thus we are not considering increasing the lifetime of
1495 INNER_DEST.
1497 Also, if this insn sets a function argument, combining it with
1498 something that might need a spill could clobber a previous
1499 function argument; the all_adjacent test in can_combine_p also
1500 checks this; here, we do a more specific test for this case. */
1502 || (REG_P (inner_dest)
1503 && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER
1504 && (! HARD_REGNO_MODE_OK (REGNO (inner_dest),
1505 GET_MODE (inner_dest))))
1506 || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src)))
1507 return 0;
1509 /* If DEST is used in I3, it is being killed in this insn, so
1510 record that for later. We have to consider paradoxical
1511 subregs here, since they kill the whole register, but we
1512 ignore partial subregs, STRICT_LOW_PART, etc.
1513 Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the
1514 STACK_POINTER_REGNUM, since these are always considered to be
1515 live. Similarly for ARG_POINTER_REGNUM if it is fixed. */
1516 subdest = dest;
1517 if (GET_CODE (subdest) == SUBREG
1518 && (GET_MODE_SIZE (GET_MODE (subdest))
1519 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (subdest)))))
1520 subdest = SUBREG_REG (subdest);
1521 if (pi3dest_killed
1522 && REG_P (subdest)
1523 && reg_referenced_p (subdest, PATTERN (i3))
1524 && REGNO (subdest) != FRAME_POINTER_REGNUM
1525 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1526 && REGNO (subdest) != HARD_FRAME_POINTER_REGNUM
1527 #endif
1528 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
1529 && (REGNO (subdest) != ARG_POINTER_REGNUM
1530 || ! fixed_regs [REGNO (subdest)])
1531 #endif
1532 && REGNO (subdest) != STACK_POINTER_REGNUM)
1534 if (*pi3dest_killed)
1535 return 0;
1537 *pi3dest_killed = subdest;
1541 else if (GET_CODE (x) == PARALLEL)
1543 int i;
1545 for (i = 0; i < XVECLEN (x, 0); i++)
1546 if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest,
1547 i1_not_in_src, pi3dest_killed))
1548 return 0;
1551 return 1;
1554 /* Return 1 if X is an arithmetic expression that contains a multiplication
1555 and division. We don't count multiplications by powers of two here. */
1557 static int
1558 contains_muldiv (rtx x)
1560 switch (GET_CODE (x))
1562 case MOD: case DIV: case UMOD: case UDIV:
1563 return 1;
1565 case MULT:
1566 return ! (GET_CODE (XEXP (x, 1)) == CONST_INT
1567 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0);
1568 default:
1569 if (BINARY_P (x))
1570 return contains_muldiv (XEXP (x, 0))
1571 || contains_muldiv (XEXP (x, 1));
1573 if (UNARY_P (x))
1574 return contains_muldiv (XEXP (x, 0));
1576 return 0;
1580 /* Determine whether INSN can be used in a combination. Return nonzero if
1581 not. This is used in try_combine to detect early some cases where we
1582 can't perform combinations. */
1584 static int
1585 cant_combine_insn_p (rtx insn)
1587 rtx set;
1588 rtx src, dest;
1590 /* If this isn't really an insn, we can't do anything.
1591 This can occur when flow deletes an insn that it has merged into an
1592 auto-increment address. */
1593 if (! INSN_P (insn))
1594 return 1;
1596 /* Never combine loads and stores involving hard regs that are likely
1597 to be spilled. The register allocator can usually handle such
1598 reg-reg moves by tying. If we allow the combiner to make
1599 substitutions of likely-spilled regs, reload might die.
1600 As an exception, we allow combinations involving fixed regs; these are
1601 not available to the register allocator so there's no risk involved. */
1603 set = single_set (insn);
1604 if (! set)
1605 return 0;
1606 src = SET_SRC (set);
1607 dest = SET_DEST (set);
1608 if (GET_CODE (src) == SUBREG)
1609 src = SUBREG_REG (src);
1610 if (GET_CODE (dest) == SUBREG)
1611 dest = SUBREG_REG (dest);
1612 if (REG_P (src) && REG_P (dest)
1613 && ((REGNO (src) < FIRST_PSEUDO_REGISTER
1614 && ! fixed_regs[REGNO (src)]
1615 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (src))))
1616 || (REGNO (dest) < FIRST_PSEUDO_REGISTER
1617 && ! fixed_regs[REGNO (dest)]
1618 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (dest))))))
1619 return 1;
1621 return 0;
1624 struct likely_spilled_retval_info
1626 unsigned regno, nregs;
1627 unsigned mask;
1630 /* Called via note_stores by likely_spilled_retval_p. Remove from info->mask
1631 hard registers that are known to be written to / clobbered in full. */
1632 static void
1633 likely_spilled_retval_1 (rtx x, rtx set, void *data)
1635 struct likely_spilled_retval_info *info = data;
1636 unsigned regno, nregs;
1637 unsigned new_mask;
1639 if (!REG_P (XEXP (set, 0)))
1640 return;
1641 regno = REGNO (x);
1642 if (regno >= info->regno + info->nregs)
1643 return;
1644 nregs = hard_regno_nregs[regno][GET_MODE (x)];
1645 if (regno + nregs <= info->regno)
1646 return;
1647 new_mask = (2U << (nregs - 1)) - 1;
1648 if (regno < info->regno)
1649 new_mask >>= info->regno - regno;
1650 else
1651 new_mask <<= regno - info->regno;
1652 info->mask &= new_mask;
1655 /* Return nonzero iff part of the return value is live during INSN, and
1656 it is likely spilled. This can happen when more than one insn is needed
1657 to copy the return value, e.g. when we consider to combine into the
1658 second copy insn for a complex value. */
1660 static int
1661 likely_spilled_retval_p (rtx insn)
1663 rtx use = BB_END (this_basic_block);
1664 rtx reg, p;
1665 unsigned regno, nregs;
1666 /* We assume here that no machine mode needs more than
1667 32 hard registers when the value overlaps with a register
1668 for which FUNCTION_VALUE_REGNO_P is true. */
1669 unsigned mask;
1670 struct likely_spilled_retval_info info;
1672 if (!NONJUMP_INSN_P (use) || GET_CODE (PATTERN (use)) != USE || insn == use)
1673 return 0;
1674 reg = XEXP (PATTERN (use), 0);
1675 if (!REG_P (reg) || !FUNCTION_VALUE_REGNO_P (REGNO (reg)))
1676 return 0;
1677 regno = REGNO (reg);
1678 nregs = hard_regno_nregs[regno][GET_MODE (reg)];
1679 if (nregs == 1)
1680 return 0;
1681 mask = (2U << (nregs - 1)) - 1;
1683 /* Disregard parts of the return value that are set later. */
1684 info.regno = regno;
1685 info.nregs = nregs;
1686 info.mask = mask;
1687 for (p = PREV_INSN (use); info.mask && p != insn; p = PREV_INSN (p))
1688 note_stores (PATTERN (insn), likely_spilled_retval_1, &info);
1689 mask = info.mask;
1691 /* Check if any of the (probably) live return value registers is
1692 likely spilled. */
1693 nregs --;
1696 if ((mask & 1 << nregs)
1697 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno + nregs)))
1698 return 1;
1699 } while (nregs--);
1700 return 0;
1703 /* Adjust INSN after we made a change to its destination.
1705 Changing the destination can invalidate notes that say something about
1706 the results of the insn and a LOG_LINK pointing to the insn. */
1708 static void
1709 adjust_for_new_dest (rtx insn)
1711 rtx *loc;
1713 /* For notes, be conservative and simply remove them. */
1714 loc = &REG_NOTES (insn);
1715 while (*loc)
1717 enum reg_note kind = REG_NOTE_KIND (*loc);
1718 if (kind == REG_EQUAL || kind == REG_EQUIV)
1719 *loc = XEXP (*loc, 1);
1720 else
1721 loc = &XEXP (*loc, 1);
1724 /* The new insn will have a destination that was previously the destination
1725 of an insn just above it. Call distribute_links to make a LOG_LINK from
1726 the next use of that destination. */
1727 distribute_links (gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX));
1730 /* Return TRUE if combine can reuse reg X in mode MODE.
1731 ADDED_SETS is nonzero if the original set is still required. */
1732 static bool
1733 can_change_dest_mode (rtx x, int added_sets, enum machine_mode mode)
1735 unsigned int regno;
1737 if (!REG_P(x))
1738 return false;
1740 regno = REGNO (x);
1741 /* Allow hard registers if the new mode is legal, and occupies no more
1742 registers than the old mode. */
1743 if (regno < FIRST_PSEUDO_REGISTER)
1744 return (HARD_REGNO_MODE_OK (regno, mode)
1745 && (hard_regno_nregs[regno][GET_MODE (x)]
1746 >= hard_regno_nregs[regno][mode]));
1748 /* Or a pseudo that is only used once. */
1749 return (REG_N_SETS (regno) == 1 && !added_sets
1750 && !REG_USERVAR_P (x));
1754 /* Check whether X, the destination of a set, refers to part of
1755 the register specified by REG. */
1757 static bool
1758 reg_subword_p (rtx x, rtx reg)
1760 /* Check that reg is an integer mode register. */
1761 if (!REG_P (reg) || GET_MODE_CLASS (GET_MODE (reg)) != MODE_INT)
1762 return false;
1764 if (GET_CODE (x) == STRICT_LOW_PART
1765 || GET_CODE (x) == ZERO_EXTRACT)
1766 x = XEXP (x, 0);
1768 return GET_CODE (x) == SUBREG
1769 && SUBREG_REG (x) == reg
1770 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT;
1774 /* Try to combine the insns I1 and I2 into I3.
1775 Here I1 and I2 appear earlier than I3.
1776 I1 can be zero; then we combine just I2 into I3.
1778 If we are combining three insns and the resulting insn is not recognized,
1779 try splitting it into two insns. If that happens, I2 and I3 are retained
1780 and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2
1781 are pseudo-deleted.
1783 Return 0 if the combination does not work. Then nothing is changed.
1784 If we did the combination, return the insn at which combine should
1785 resume scanning.
1787 Set NEW_DIRECT_JUMP_P to a nonzero value if try_combine creates a
1788 new direct jump instruction. */
1790 static rtx
1791 try_combine (rtx i3, rtx i2, rtx i1, int *new_direct_jump_p)
1793 /* New patterns for I3 and I2, respectively. */
1794 rtx newpat, newi2pat = 0;
1795 rtvec newpat_vec_with_clobbers = 0;
1796 int substed_i2 = 0, substed_i1 = 0;
1797 /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */
1798 int added_sets_1, added_sets_2;
1799 /* Total number of SETs to put into I3. */
1800 int total_sets;
1801 /* Nonzero if I2's body now appears in I3. */
1802 int i2_is_used;
1803 /* INSN_CODEs for new I3, new I2, and user of condition code. */
1804 int insn_code_number, i2_code_number = 0, other_code_number = 0;
1805 /* Contains I3 if the destination of I3 is used in its source, which means
1806 that the old life of I3 is being killed. If that usage is placed into
1807 I2 and not in I3, a REG_DEAD note must be made. */
1808 rtx i3dest_killed = 0;
1809 /* SET_DEST and SET_SRC of I2 and I1. */
1810 rtx i2dest, i2src, i1dest = 0, i1src = 0;
1811 /* PATTERN (I2), or a copy of it in certain cases. */
1812 rtx i2pat;
1813 /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */
1814 int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0;
1815 int i2dest_killed = 0, i1dest_killed = 0;
1816 int i1_feeds_i3 = 0;
1817 /* Notes that must be added to REG_NOTES in I3 and I2. */
1818 rtx new_i3_notes, new_i2_notes;
1819 /* Notes that we substituted I3 into I2 instead of the normal case. */
1820 int i3_subst_into_i2 = 0;
1821 /* Notes that I1, I2 or I3 is a MULT operation. */
1822 int have_mult = 0;
1823 int swap_i2i3 = 0;
1825 int maxreg;
1826 rtx temp;
1827 rtx link;
1828 int i;
1830 /* Exit early if one of the insns involved can't be used for
1831 combinations. */
1832 if (cant_combine_insn_p (i3)
1833 || cant_combine_insn_p (i2)
1834 || (i1 && cant_combine_insn_p (i1))
1835 || likely_spilled_retval_p (i3)
1836 /* We also can't do anything if I3 has a
1837 REG_LIBCALL note since we don't want to disrupt the contiguity of a
1838 libcall. */
1839 #if 0
1840 /* ??? This gives worse code, and appears to be unnecessary, since no
1841 pass after flow uses REG_LIBCALL/REG_RETVAL notes. */
1842 || find_reg_note (i3, REG_LIBCALL, NULL_RTX)
1843 #endif
1845 return 0;
1847 combine_attempts++;
1848 undobuf.other_insn = 0;
1850 /* Reset the hard register usage information. */
1851 CLEAR_HARD_REG_SET (newpat_used_regs);
1853 /* If I1 and I2 both feed I3, they can be in any order. To simplify the
1854 code below, set I1 to be the earlier of the two insns. */
1855 if (i1 && INSN_CUID (i1) > INSN_CUID (i2))
1856 temp = i1, i1 = i2, i2 = temp;
1858 added_links_insn = 0;
1860 /* First check for one important special-case that the code below will
1861 not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
1862 and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case,
1863 we may be able to replace that destination with the destination of I3.
1864 This occurs in the common code where we compute both a quotient and
1865 remainder into a structure, in which case we want to do the computation
1866 directly into the structure to avoid register-register copies.
1868 Note that this case handles both multiple sets in I2 and also
1869 cases where I2 has a number of CLOBBER or PARALLELs.
1871 We make very conservative checks below and only try to handle the
1872 most common cases of this. For example, we only handle the case
1873 where I2 and I3 are adjacent to avoid making difficult register
1874 usage tests. */
1876 if (i1 == 0 && NONJUMP_INSN_P (i3) && GET_CODE (PATTERN (i3)) == SET
1877 && REG_P (SET_SRC (PATTERN (i3)))
1878 && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER
1879 && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3)))
1880 && GET_CODE (PATTERN (i2)) == PARALLEL
1881 && ! side_effects_p (SET_DEST (PATTERN (i3)))
1882 /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code
1883 below would need to check what is inside (and reg_overlap_mentioned_p
1884 doesn't support those codes anyway). Don't allow those destinations;
1885 the resulting insn isn't likely to be recognized anyway. */
1886 && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT
1887 && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART
1888 && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)),
1889 SET_DEST (PATTERN (i3)))
1890 && next_real_insn (i2) == i3)
1892 rtx p2 = PATTERN (i2);
1894 /* Make sure that the destination of I3,
1895 which we are going to substitute into one output of I2,
1896 is not used within another output of I2. We must avoid making this:
1897 (parallel [(set (mem (reg 69)) ...)
1898 (set (reg 69) ...)])
1899 which is not well-defined as to order of actions.
1900 (Besides, reload can't handle output reloads for this.)
1902 The problem can also happen if the dest of I3 is a memory ref,
1903 if another dest in I2 is an indirect memory ref. */
1904 for (i = 0; i < XVECLEN (p2, 0); i++)
1905 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1906 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1907 && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)),
1908 SET_DEST (XVECEXP (p2, 0, i))))
1909 break;
1911 if (i == XVECLEN (p2, 0))
1912 for (i = 0; i < XVECLEN (p2, 0); i++)
1913 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1914 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1915 && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3)))
1917 combine_merges++;
1919 subst_insn = i3;
1920 subst_low_cuid = INSN_CUID (i2);
1922 added_sets_2 = added_sets_1 = 0;
1923 i2dest = SET_SRC (PATTERN (i3));
1924 i2dest_killed = dead_or_set_p (i2, i2dest);
1926 /* Replace the dest in I2 with our dest and make the resulting
1927 insn the new pattern for I3. Then skip to where we
1928 validate the pattern. Everything was set up above. */
1929 SUBST (SET_DEST (XVECEXP (p2, 0, i)),
1930 SET_DEST (PATTERN (i3)));
1932 newpat = p2;
1933 i3_subst_into_i2 = 1;
1934 goto validate_replacement;
1938 /* If I2 is setting a pseudo to a constant and I3 is setting some
1939 sub-part of it to another constant, merge them by making a new
1940 constant. */
1941 if (i1 == 0
1942 && (temp = single_set (i2)) != 0
1943 && (GET_CODE (SET_SRC (temp)) == CONST_INT
1944 || GET_CODE (SET_SRC (temp)) == CONST_DOUBLE)
1945 && GET_CODE (PATTERN (i3)) == SET
1946 && (GET_CODE (SET_SRC (PATTERN (i3))) == CONST_INT
1947 || GET_CODE (SET_SRC (PATTERN (i3))) == CONST_DOUBLE)
1948 && reg_subword_p (SET_DEST (PATTERN (i3)), SET_DEST (temp)))
1950 rtx dest = SET_DEST (PATTERN (i3));
1951 int offset = -1;
1952 int width = 0;
1954 if (GET_CODE (dest) == ZERO_EXTRACT)
1956 if (GET_CODE (XEXP (dest, 1)) == CONST_INT
1957 && GET_CODE (XEXP (dest, 2)) == CONST_INT)
1959 width = INTVAL (XEXP (dest, 1));
1960 offset = INTVAL (XEXP (dest, 2));
1961 dest = XEXP (dest, 0);
1962 if (BITS_BIG_ENDIAN)
1963 offset = GET_MODE_BITSIZE (GET_MODE (dest)) - width - offset;
1966 else
1968 if (GET_CODE (dest) == STRICT_LOW_PART)
1969 dest = XEXP (dest, 0);
1970 width = GET_MODE_BITSIZE (GET_MODE (dest));
1971 offset = 0;
1974 if (offset >= 0)
1976 /* If this is the low part, we're done. */
1977 if (subreg_lowpart_p (dest))
1979 /* Handle the case where inner is twice the size of outer. */
1980 else if (GET_MODE_BITSIZE (GET_MODE (SET_DEST (temp)))
1981 == 2 * GET_MODE_BITSIZE (GET_MODE (dest)))
1982 offset += GET_MODE_BITSIZE (GET_MODE (dest));
1983 /* Otherwise give up for now. */
1984 else
1985 offset = -1;
1988 if (offset >= 0)
1990 HOST_WIDE_INT mhi, ohi, ihi;
1991 HOST_WIDE_INT mlo, olo, ilo;
1992 rtx inner = SET_SRC (PATTERN (i3));
1993 rtx outer = SET_SRC (temp);
1995 if (GET_CODE (outer) == CONST_INT)
1997 olo = INTVAL (outer);
1998 ohi = olo < 0 ? -1 : 0;
2000 else
2002 olo = CONST_DOUBLE_LOW (outer);
2003 ohi = CONST_DOUBLE_HIGH (outer);
2006 if (GET_CODE (inner) == CONST_INT)
2008 ilo = INTVAL (inner);
2009 ihi = ilo < 0 ? -1 : 0;
2011 else
2013 ilo = CONST_DOUBLE_LOW (inner);
2014 ihi = CONST_DOUBLE_HIGH (inner);
2017 if (width < HOST_BITS_PER_WIDE_INT)
2019 mlo = ((unsigned HOST_WIDE_INT) 1 << width) - 1;
2020 mhi = 0;
2022 else if (width < HOST_BITS_PER_WIDE_INT * 2)
2024 mhi = ((unsigned HOST_WIDE_INT) 1
2025 << (width - HOST_BITS_PER_WIDE_INT)) - 1;
2026 mlo = -1;
2028 else
2030 mlo = -1;
2031 mhi = -1;
2034 ilo &= mlo;
2035 ihi &= mhi;
2037 if (offset >= HOST_BITS_PER_WIDE_INT)
2039 mhi = mlo << (offset - HOST_BITS_PER_WIDE_INT);
2040 mlo = 0;
2041 ihi = ilo << (offset - HOST_BITS_PER_WIDE_INT);
2042 ilo = 0;
2044 else if (offset > 0)
2046 mhi = (mhi << offset) | ((unsigned HOST_WIDE_INT) mlo
2047 >> (HOST_BITS_PER_WIDE_INT - offset));
2048 mlo = mlo << offset;
2049 ihi = (ihi << offset) | ((unsigned HOST_WIDE_INT) ilo
2050 >> (HOST_BITS_PER_WIDE_INT - offset));
2051 ilo = ilo << offset;
2054 olo = (olo & ~mlo) | ilo;
2055 ohi = (ohi & ~mhi) | ihi;
2057 combine_merges++;
2058 subst_insn = i3;
2059 subst_low_cuid = INSN_CUID (i2);
2060 added_sets_2 = added_sets_1 = 0;
2061 i2dest = SET_DEST (temp);
2062 i2dest_killed = dead_or_set_p (i2, i2dest);
2064 SUBST (SET_SRC (temp),
2065 immed_double_const (olo, ohi, GET_MODE (SET_DEST (temp))));
2067 newpat = PATTERN (i2);
2068 goto validate_replacement;
2072 #ifndef HAVE_cc0
2073 /* If we have no I1 and I2 looks like:
2074 (parallel [(set (reg:CC X) (compare:CC OP (const_int 0)))
2075 (set Y OP)])
2076 make up a dummy I1 that is
2077 (set Y OP)
2078 and change I2 to be
2079 (set (reg:CC X) (compare:CC Y (const_int 0)))
2081 (We can ignore any trailing CLOBBERs.)
2083 This undoes a previous combination and allows us to match a branch-and-
2084 decrement insn. */
2086 if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL
2087 && XVECLEN (PATTERN (i2), 0) >= 2
2088 && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET
2089 && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0))))
2090 == MODE_CC)
2091 && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE
2092 && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx
2093 && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET
2094 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)))
2095 && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0),
2096 SET_SRC (XVECEXP (PATTERN (i2), 0, 1))))
2098 for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--)
2099 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER)
2100 break;
2102 if (i == 1)
2104 /* We make I1 with the same INSN_UID as I2. This gives it
2105 the same INSN_CUID for value tracking. Our fake I1 will
2106 never appear in the insn stream so giving it the same INSN_UID
2107 as I2 will not cause a problem. */
2109 i1 = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2,
2110 BLOCK_FOR_INSN (i2), INSN_LOCATOR (i2),
2111 XVECEXP (PATTERN (i2), 0, 1), -1, NULL_RTX,
2112 NULL_RTX);
2114 SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0));
2115 SUBST (XEXP (SET_SRC (PATTERN (i2)), 0),
2116 SET_DEST (PATTERN (i1)));
2119 #endif
2121 /* Verify that I2 and I1 are valid for combining. */
2122 if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src)
2123 || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src)))
2125 undo_all ();
2126 return 0;
2129 /* Record whether I2DEST is used in I2SRC and similarly for the other
2130 cases. Knowing this will help in register status updating below. */
2131 i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src);
2132 i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src);
2133 i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src);
2134 i2dest_killed = dead_or_set_p (i2, i2dest);
2135 i1dest_killed = i1 && dead_or_set_p (i1, i1dest);
2137 /* See if I1 directly feeds into I3. It does if I1DEST is not used
2138 in I2SRC. */
2139 i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src);
2141 /* Ensure that I3's pattern can be the destination of combines. */
2142 if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest,
2143 i1 && i2dest_in_i1src && i1_feeds_i3,
2144 &i3dest_killed))
2146 undo_all ();
2147 return 0;
2150 /* See if any of the insns is a MULT operation. Unless one is, we will
2151 reject a combination that is, since it must be slower. Be conservative
2152 here. */
2153 if (GET_CODE (i2src) == MULT
2154 || (i1 != 0 && GET_CODE (i1src) == MULT)
2155 || (GET_CODE (PATTERN (i3)) == SET
2156 && GET_CODE (SET_SRC (PATTERN (i3))) == MULT))
2157 have_mult = 1;
2159 /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd.
2160 We used to do this EXCEPT in one case: I3 has a post-inc in an
2161 output operand. However, that exception can give rise to insns like
2162 mov r3,(r3)+
2163 which is a famous insn on the PDP-11 where the value of r3 used as the
2164 source was model-dependent. Avoid this sort of thing. */
2166 #if 0
2167 if (!(GET_CODE (PATTERN (i3)) == SET
2168 && REG_P (SET_SRC (PATTERN (i3)))
2169 && MEM_P (SET_DEST (PATTERN (i3)))
2170 && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC
2171 || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC)))
2172 /* It's not the exception. */
2173 #endif
2174 #ifdef AUTO_INC_DEC
2175 for (link = REG_NOTES (i3); link; link = XEXP (link, 1))
2176 if (REG_NOTE_KIND (link) == REG_INC
2177 && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2))
2178 || (i1 != 0
2179 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1)))))
2181 undo_all ();
2182 return 0;
2184 #endif
2186 /* See if the SETs in I1 or I2 need to be kept around in the merged
2187 instruction: whenever the value set there is still needed past I3.
2188 For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3.
2190 For the SET in I1, we have two cases: If I1 and I2 independently
2191 feed into I3, the set in I1 needs to be kept around if I1DEST dies
2192 or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set
2193 in I1 needs to be kept around unless I1DEST dies or is set in either
2194 I2 or I3. We can distinguish these cases by seeing if I2SRC mentions
2195 I1DEST. If so, we know I1 feeds into I2. */
2197 added_sets_2 = ! dead_or_set_p (i3, i2dest);
2199 added_sets_1
2200 = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest)
2201 : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest)));
2203 /* If the set in I2 needs to be kept around, we must make a copy of
2204 PATTERN (I2), so that when we substitute I1SRC for I1DEST in
2205 PATTERN (I2), we are only substituting for the original I1DEST, not into
2206 an already-substituted copy. This also prevents making self-referential
2207 rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to
2208 I2DEST. */
2210 i2pat = (GET_CODE (PATTERN (i2)) == PARALLEL
2211 ? gen_rtx_SET (VOIDmode, i2dest, i2src)
2212 : PATTERN (i2));
2214 if (added_sets_2)
2215 i2pat = copy_rtx (i2pat);
2217 combine_merges++;
2219 /* Substitute in the latest insn for the regs set by the earlier ones. */
2221 maxreg = max_reg_num ();
2223 subst_insn = i3;
2225 #ifndef HAVE_cc0
2226 /* Many machines that don't use CC0 have insns that can both perform an
2227 arithmetic operation and set the condition code. These operations will
2228 be represented as a PARALLEL with the first element of the vector
2229 being a COMPARE of an arithmetic operation with the constant zero.
2230 The second element of the vector will set some pseudo to the result
2231 of the same arithmetic operation. If we simplify the COMPARE, we won't
2232 match such a pattern and so will generate an extra insn. Here we test
2233 for this case, where both the comparison and the operation result are
2234 needed, and make the PARALLEL by just replacing I2DEST in I3SRC with
2235 I2SRC. Later we will make the PARALLEL that contains I2. */
2237 if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET
2238 && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE
2239 && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx
2240 && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest))
2242 #ifdef SELECT_CC_MODE
2243 rtx *cc_use;
2244 enum machine_mode compare_mode;
2245 #endif
2247 newpat = PATTERN (i3);
2248 SUBST (XEXP (SET_SRC (newpat), 0), i2src);
2250 i2_is_used = 1;
2252 #ifdef SELECT_CC_MODE
2253 /* See if a COMPARE with the operand we substituted in should be done
2254 with the mode that is currently being used. If not, do the same
2255 processing we do in `subst' for a SET; namely, if the destination
2256 is used only once, try to replace it with a register of the proper
2257 mode and also replace the COMPARE. */
2258 if (undobuf.other_insn == 0
2259 && (cc_use = find_single_use (SET_DEST (newpat), i3,
2260 &undobuf.other_insn))
2261 && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use),
2262 i2src, const0_rtx))
2263 != GET_MODE (SET_DEST (newpat))))
2265 if (can_change_dest_mode(SET_DEST (newpat), added_sets_2,
2266 compare_mode))
2268 unsigned int regno = REGNO (SET_DEST (newpat));
2269 rtx new_dest;
2271 if (regno < FIRST_PSEUDO_REGISTER)
2272 new_dest = gen_rtx_REG (compare_mode, regno);
2273 else
2275 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
2276 new_dest = regno_reg_rtx[regno];
2279 SUBST (SET_DEST (newpat), new_dest);
2280 SUBST (XEXP (*cc_use, 0), new_dest);
2281 SUBST (SET_SRC (newpat),
2282 gen_rtx_COMPARE (compare_mode, i2src, const0_rtx));
2284 else
2285 undobuf.other_insn = 0;
2287 #endif
2289 else
2290 #endif
2292 /* It is possible that the source of I2 or I1 may be performing
2293 an unneeded operation, such as a ZERO_EXTEND of something
2294 that is known to have the high part zero. Handle that case
2295 by letting subst look at the innermost one of them.
2297 Another way to do this would be to have a function that tries
2298 to simplify a single insn instead of merging two or more
2299 insns. We don't do this because of the potential of infinite
2300 loops and because of the potential extra memory required.
2301 However, doing it the way we are is a bit of a kludge and
2302 doesn't catch all cases.
2304 But only do this if -fexpensive-optimizations since it slows
2305 things down and doesn't usually win.
2307 This is not done in the COMPARE case above because the
2308 unmodified I2PAT is used in the PARALLEL and so a pattern
2309 with a modified I2SRC would not match. */
2311 if (flag_expensive_optimizations)
2313 /* Pass pc_rtx so no substitutions are done, just
2314 simplifications. */
2315 if (i1)
2317 subst_low_cuid = INSN_CUID (i1);
2318 i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0);
2320 else
2322 subst_low_cuid = INSN_CUID (i2);
2323 i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0);
2327 n_occurrences = 0; /* `subst' counts here */
2329 /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we
2330 need to make a unique copy of I2SRC each time we substitute it
2331 to avoid self-referential rtl. */
2333 subst_low_cuid = INSN_CUID (i2);
2334 newpat = subst (PATTERN (i3), i2dest, i2src, 0,
2335 ! i1_feeds_i3 && i1dest_in_i1src);
2336 substed_i2 = 1;
2338 /* Record whether i2's body now appears within i3's body. */
2339 i2_is_used = n_occurrences;
2342 /* If we already got a failure, don't try to do more. Otherwise,
2343 try to substitute in I1 if we have it. */
2345 if (i1 && GET_CODE (newpat) != CLOBBER)
2347 /* Before we can do this substitution, we must redo the test done
2348 above (see detailed comments there) that ensures that I1DEST
2349 isn't mentioned in any SETs in NEWPAT that are field assignments. */
2351 if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX,
2352 0, (rtx*) 0))
2354 undo_all ();
2355 return 0;
2358 n_occurrences = 0;
2359 subst_low_cuid = INSN_CUID (i1);
2360 newpat = subst (newpat, i1dest, i1src, 0, 0);
2361 substed_i1 = 1;
2364 /* Fail if an autoincrement side-effect has been duplicated. Be careful
2365 to count all the ways that I2SRC and I1SRC can be used. */
2366 if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0
2367 && i2_is_used + added_sets_2 > 1)
2368 || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
2369 && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3)
2370 > 1))
2371 /* Fail if we tried to make a new register. */
2372 || max_reg_num () != maxreg
2373 /* Fail if we couldn't do something and have a CLOBBER. */
2374 || GET_CODE (newpat) == CLOBBER
2375 /* Fail if this new pattern is a MULT and we didn't have one before
2376 at the outer level. */
2377 || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT
2378 && ! have_mult))
2380 undo_all ();
2381 return 0;
2384 /* If the actions of the earlier insns must be kept
2385 in addition to substituting them into the latest one,
2386 we must make a new PARALLEL for the latest insn
2387 to hold additional the SETs. */
2389 if (added_sets_1 || added_sets_2)
2391 combine_extras++;
2393 if (GET_CODE (newpat) == PARALLEL)
2395 rtvec old = XVEC (newpat, 0);
2396 total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2;
2397 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2398 memcpy (XVEC (newpat, 0)->elem, &old->elem[0],
2399 sizeof (old->elem[0]) * old->num_elem);
2401 else
2403 rtx old = newpat;
2404 total_sets = 1 + added_sets_1 + added_sets_2;
2405 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2406 XVECEXP (newpat, 0, 0) = old;
2409 if (added_sets_1)
2410 XVECEXP (newpat, 0, --total_sets)
2411 = (GET_CODE (PATTERN (i1)) == PARALLEL
2412 ? gen_rtx_SET (VOIDmode, i1dest, i1src) : PATTERN (i1));
2414 if (added_sets_2)
2416 /* If there is no I1, use I2's body as is. We used to also not do
2417 the subst call below if I2 was substituted into I3,
2418 but that could lose a simplification. */
2419 if (i1 == 0)
2420 XVECEXP (newpat, 0, --total_sets) = i2pat;
2421 else
2422 /* See comment where i2pat is assigned. */
2423 XVECEXP (newpat, 0, --total_sets)
2424 = subst (i2pat, i1dest, i1src, 0, 0);
2428 /* We come here when we are replacing a destination in I2 with the
2429 destination of I3. */
2430 validate_replacement:
2432 /* Note which hard regs this insn has as inputs. */
2433 mark_used_regs_combine (newpat);
2435 /* If recog_for_combine fails, it strips existing clobbers. If we'll
2436 consider splitting this pattern, we might need these clobbers. */
2437 if (i1 && GET_CODE (newpat) == PARALLEL
2438 && GET_CODE (XVECEXP (newpat, 0, XVECLEN (newpat, 0) - 1)) == CLOBBER)
2440 int len = XVECLEN (newpat, 0);
2442 newpat_vec_with_clobbers = rtvec_alloc (len);
2443 for (i = 0; i < len; i++)
2444 RTVEC_ELT (newpat_vec_with_clobbers, i) = XVECEXP (newpat, 0, i);
2447 /* Is the result of combination a valid instruction? */
2448 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2450 /* If the result isn't valid, see if it is a PARALLEL of two SETs where
2451 the second SET's destination is a register that is unused and isn't
2452 marked as an instruction that might trap in an EH region. In that case,
2453 we just need the first SET. This can occur when simplifying a divmod
2454 insn. We *must* test for this case here because the code below that
2455 splits two independent SETs doesn't handle this case correctly when it
2456 updates the register status.
2458 It's pointless doing this if we originally had two sets, one from
2459 i3, and one from i2. Combining then splitting the parallel results
2460 in the original i2 again plus an invalid insn (which we delete).
2461 The net effect is only to move instructions around, which makes
2462 debug info less accurate.
2464 Also check the case where the first SET's destination is unused.
2465 That would not cause incorrect code, but does cause an unneeded
2466 insn to remain. */
2468 if (insn_code_number < 0
2469 && !(added_sets_2 && i1 == 0)
2470 && GET_CODE (newpat) == PARALLEL
2471 && XVECLEN (newpat, 0) == 2
2472 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2473 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2474 && asm_noperands (newpat) < 0)
2476 rtx set0 = XVECEXP (newpat, 0, 0);
2477 rtx set1 = XVECEXP (newpat, 0, 1);
2478 rtx note;
2480 if (((REG_P (SET_DEST (set1))
2481 && find_reg_note (i3, REG_UNUSED, SET_DEST (set1)))
2482 || (GET_CODE (SET_DEST (set1)) == SUBREG
2483 && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1)))))
2484 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2485 || INTVAL (XEXP (note, 0)) <= 0)
2486 && ! side_effects_p (SET_SRC (set1)))
2488 newpat = set0;
2489 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2492 else if (((REG_P (SET_DEST (set0))
2493 && find_reg_note (i3, REG_UNUSED, SET_DEST (set0)))
2494 || (GET_CODE (SET_DEST (set0)) == SUBREG
2495 && find_reg_note (i3, REG_UNUSED,
2496 SUBREG_REG (SET_DEST (set0)))))
2497 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2498 || INTVAL (XEXP (note, 0)) <= 0)
2499 && ! side_effects_p (SET_SRC (set0)))
2501 newpat = set1;
2502 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2504 if (insn_code_number >= 0)
2506 /* If we will be able to accept this, we have made a
2507 change to the destination of I3. This requires us to
2508 do a few adjustments. */
2510 PATTERN (i3) = newpat;
2511 adjust_for_new_dest (i3);
2516 /* If we were combining three insns and the result is a simple SET
2517 with no ASM_OPERANDS that wasn't recognized, try to split it into two
2518 insns. There are two ways to do this. It can be split using a
2519 machine-specific method (like when you have an addition of a large
2520 constant) or by combine in the function find_split_point. */
2522 if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET
2523 && asm_noperands (newpat) < 0)
2525 rtx m_split, *split;
2527 /* See if the MD file can split NEWPAT. If it can't, see if letting it
2528 use I2DEST as a scratch register will help. In the latter case,
2529 convert I2DEST to the mode of the source of NEWPAT if we can. */
2531 m_split = split_insns (newpat, i3);
2533 /* We can only use I2DEST as a scratch reg if it doesn't overlap any
2534 inputs of NEWPAT. */
2536 /* ??? If I2DEST is not safe, and I1DEST exists, then it would be
2537 possible to try that as a scratch reg. This would require adding
2538 more code to make it work though. */
2540 if (m_split == 0 && ! reg_overlap_mentioned_p (i2dest, newpat))
2542 enum machine_mode new_mode = GET_MODE (SET_DEST (newpat));
2544 /* First try to split using the original register as a
2545 scratch register. */
2546 m_split = split_insns (gen_rtx_PARALLEL
2547 (VOIDmode,
2548 gen_rtvec (2, newpat,
2549 gen_rtx_CLOBBER (VOIDmode,
2550 i2dest))),
2551 i3);
2553 /* If that didn't work, try changing the mode of I2DEST if
2554 we can. */
2555 if (m_split == 0
2556 && new_mode != GET_MODE (i2dest)
2557 && new_mode != VOIDmode
2558 && can_change_dest_mode (i2dest, added_sets_2, new_mode))
2560 enum machine_mode old_mode = GET_MODE (i2dest);
2561 rtx ni2dest;
2563 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
2564 ni2dest = gen_rtx_REG (new_mode, REGNO (i2dest));
2565 else
2567 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], new_mode);
2568 ni2dest = regno_reg_rtx[REGNO (i2dest)];
2571 m_split = split_insns (gen_rtx_PARALLEL
2572 (VOIDmode,
2573 gen_rtvec (2, newpat,
2574 gen_rtx_CLOBBER (VOIDmode,
2575 ni2dest))),
2576 i3);
2578 if (m_split == 0
2579 && REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2581 struct undo *buf;
2583 PUT_MODE (regno_reg_rtx[REGNO (i2dest)], old_mode);
2584 buf = undobuf.undos;
2585 undobuf.undos = buf->next;
2586 buf->next = undobuf.frees;
2587 undobuf.frees = buf;
2592 /* If recog_for_combine has discarded clobbers, try to use them
2593 again for the split. */
2594 if (m_split == 0 && newpat_vec_with_clobbers)
2595 m_split
2596 = split_insns (gen_rtx_PARALLEL (VOIDmode,
2597 newpat_vec_with_clobbers), i3);
2599 if (m_split && NEXT_INSN (m_split) == NULL_RTX)
2601 m_split = PATTERN (m_split);
2602 insn_code_number = recog_for_combine (&m_split, i3, &new_i3_notes);
2603 if (insn_code_number >= 0)
2604 newpat = m_split;
2606 else if (m_split && NEXT_INSN (NEXT_INSN (m_split)) == NULL_RTX
2607 && (next_real_insn (i2) == i3
2608 || ! use_crosses_set_p (PATTERN (m_split), INSN_CUID (i2))))
2610 rtx i2set, i3set;
2611 rtx newi3pat = PATTERN (NEXT_INSN (m_split));
2612 newi2pat = PATTERN (m_split);
2614 i3set = single_set (NEXT_INSN (m_split));
2615 i2set = single_set (m_split);
2617 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2619 /* If I2 or I3 has multiple SETs, we won't know how to track
2620 register status, so don't use these insns. If I2's destination
2621 is used between I2 and I3, we also can't use these insns. */
2623 if (i2_code_number >= 0 && i2set && i3set
2624 && (next_real_insn (i2) == i3
2625 || ! reg_used_between_p (SET_DEST (i2set), i2, i3)))
2626 insn_code_number = recog_for_combine (&newi3pat, i3,
2627 &new_i3_notes);
2628 if (insn_code_number >= 0)
2629 newpat = newi3pat;
2631 /* It is possible that both insns now set the destination of I3.
2632 If so, we must show an extra use of it. */
2634 if (insn_code_number >= 0)
2636 rtx new_i3_dest = SET_DEST (i3set);
2637 rtx new_i2_dest = SET_DEST (i2set);
2639 while (GET_CODE (new_i3_dest) == ZERO_EXTRACT
2640 || GET_CODE (new_i3_dest) == STRICT_LOW_PART
2641 || GET_CODE (new_i3_dest) == SUBREG)
2642 new_i3_dest = XEXP (new_i3_dest, 0);
2644 while (GET_CODE (new_i2_dest) == ZERO_EXTRACT
2645 || GET_CODE (new_i2_dest) == STRICT_LOW_PART
2646 || GET_CODE (new_i2_dest) == SUBREG)
2647 new_i2_dest = XEXP (new_i2_dest, 0);
2649 if (REG_P (new_i3_dest)
2650 && REG_P (new_i2_dest)
2651 && REGNO (new_i3_dest) == REGNO (new_i2_dest))
2652 REG_N_SETS (REGNO (new_i2_dest))++;
2656 /* If we can split it and use I2DEST, go ahead and see if that
2657 helps things be recognized. Verify that none of the registers
2658 are set between I2 and I3. */
2659 if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0
2660 #ifdef HAVE_cc0
2661 && REG_P (i2dest)
2662 #endif
2663 /* We need I2DEST in the proper mode. If it is a hard register
2664 or the only use of a pseudo, we can change its mode.
2665 Make sure we don't change a hard register to have a mode that
2666 isn't valid for it, or change the number of registers. */
2667 && (GET_MODE (*split) == GET_MODE (i2dest)
2668 || GET_MODE (*split) == VOIDmode
2669 || can_change_dest_mode (i2dest, added_sets_2,
2670 GET_MODE (*split)))
2671 && (next_real_insn (i2) == i3
2672 || ! use_crosses_set_p (*split, INSN_CUID (i2)))
2673 /* We can't overwrite I2DEST if its value is still used by
2674 NEWPAT. */
2675 && ! reg_referenced_p (i2dest, newpat))
2677 rtx newdest = i2dest;
2678 enum rtx_code split_code = GET_CODE (*split);
2679 enum machine_mode split_mode = GET_MODE (*split);
2680 bool subst_done = false;
2681 newi2pat = NULL_RTX;
2683 /* Get NEWDEST as a register in the proper mode. We have already
2684 validated that we can do this. */
2685 if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode)
2687 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
2688 newdest = gen_rtx_REG (split_mode, REGNO (i2dest));
2689 else
2691 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], split_mode);
2692 newdest = regno_reg_rtx[REGNO (i2dest)];
2696 /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to
2697 an ASHIFT. This can occur if it was inside a PLUS and hence
2698 appeared to be a memory address. This is a kludge. */
2699 if (split_code == MULT
2700 && GET_CODE (XEXP (*split, 1)) == CONST_INT
2701 && INTVAL (XEXP (*split, 1)) > 0
2702 && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0)
2704 SUBST (*split, gen_rtx_ASHIFT (split_mode,
2705 XEXP (*split, 0), GEN_INT (i)));
2706 /* Update split_code because we may not have a multiply
2707 anymore. */
2708 split_code = GET_CODE (*split);
2711 #ifdef INSN_SCHEDULING
2712 /* If *SPLIT is a paradoxical SUBREG, when we split it, it should
2713 be written as a ZERO_EXTEND. */
2714 if (split_code == SUBREG && MEM_P (SUBREG_REG (*split)))
2716 #ifdef LOAD_EXTEND_OP
2717 /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's
2718 what it really is. */
2719 if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (*split)))
2720 == SIGN_EXTEND)
2721 SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode,
2722 SUBREG_REG (*split)));
2723 else
2724 #endif
2725 SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode,
2726 SUBREG_REG (*split)));
2728 #endif
2730 /* Attempt to split binary operators using arithmetic identities. */
2731 if (BINARY_P (SET_SRC (newpat))
2732 && split_mode == GET_MODE (SET_SRC (newpat))
2733 && ! side_effects_p (SET_SRC (newpat)))
2735 rtx setsrc = SET_SRC (newpat);
2736 enum machine_mode mode = GET_MODE (setsrc);
2737 enum rtx_code code = GET_CODE (setsrc);
2738 rtx src_op0 = XEXP (setsrc, 0);
2739 rtx src_op1 = XEXP (setsrc, 1);
2741 /* Split "X = Y op Y" as "Z = Y; X = Z op Z". */
2742 if (rtx_equal_p (src_op0, src_op1))
2744 newi2pat = gen_rtx_SET (VOIDmode, newdest, src_op0);
2745 SUBST (XEXP (setsrc, 0), newdest);
2746 SUBST (XEXP (setsrc, 1), newdest);
2747 subst_done = true;
2749 /* Split "((P op Q) op R) op S" where op is PLUS or MULT. */
2750 else if ((code == PLUS || code == MULT)
2751 && GET_CODE (src_op0) == code
2752 && GET_CODE (XEXP (src_op0, 0)) == code
2753 && (INTEGRAL_MODE_P (mode)
2754 || (FLOAT_MODE_P (mode)
2755 && flag_unsafe_math_optimizations)))
2757 rtx p = XEXP (XEXP (src_op0, 0), 0);
2758 rtx q = XEXP (XEXP (src_op0, 0), 1);
2759 rtx r = XEXP (src_op0, 1);
2760 rtx s = src_op1;
2762 /* Split both "((X op Y) op X) op Y" and
2763 "((X op Y) op Y) op X" as "T op T" where T is
2764 "X op Y". */
2765 if ((rtx_equal_p (p,r) && rtx_equal_p (q,s))
2766 || (rtx_equal_p (p,s) && rtx_equal_p (q,r)))
2768 newi2pat = gen_rtx_SET (VOIDmode, newdest,
2769 XEXP (src_op0, 0));
2770 SUBST (XEXP (setsrc, 0), newdest);
2771 SUBST (XEXP (setsrc, 1), newdest);
2772 subst_done = true;
2774 /* Split "((X op X) op Y) op Y)" as "T op T" where
2775 T is "X op Y". */
2776 else if (rtx_equal_p (p,q) && rtx_equal_p (r,s))
2778 rtx tmp = simplify_gen_binary (code, mode, p, r);
2779 newi2pat = gen_rtx_SET (VOIDmode, newdest, tmp);
2780 SUBST (XEXP (setsrc, 0), newdest);
2781 SUBST (XEXP (setsrc, 1), newdest);
2782 subst_done = true;
2787 if (!subst_done)
2789 newi2pat = gen_rtx_SET (VOIDmode, newdest, *split);
2790 SUBST (*split, newdest);
2793 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2795 /* recog_for_combine might have added CLOBBERs to newi2pat.
2796 Make sure NEWPAT does not depend on the clobbered regs. */
2797 if (GET_CODE (newi2pat) == PARALLEL)
2798 for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--)
2799 if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER)
2801 rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0);
2802 if (reg_overlap_mentioned_p (reg, newpat))
2804 undo_all ();
2805 return 0;
2809 /* If the split point was a MULT and we didn't have one before,
2810 don't use one now. */
2811 if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult))
2812 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2816 /* Check for a case where we loaded from memory in a narrow mode and
2817 then sign extended it, but we need both registers. In that case,
2818 we have a PARALLEL with both loads from the same memory location.
2819 We can split this into a load from memory followed by a register-register
2820 copy. This saves at least one insn, more if register allocation can
2821 eliminate the copy.
2823 We cannot do this if the destination of the first assignment is a
2824 condition code register or cc0. We eliminate this case by making sure
2825 the SET_DEST and SET_SRC have the same mode.
2827 We cannot do this if the destination of the second assignment is
2828 a register that we have already assumed is zero-extended. Similarly
2829 for a SUBREG of such a register. */
2831 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2832 && GET_CODE (newpat) == PARALLEL
2833 && XVECLEN (newpat, 0) == 2
2834 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2835 && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND
2836 && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0)))
2837 == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0))))
2838 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2839 && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2840 XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0))
2841 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2842 INSN_CUID (i2))
2843 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2844 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2845 && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)),
2846 (REG_P (temp)
2847 && reg_stat[REGNO (temp)].nonzero_bits != 0
2848 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2849 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2850 && (reg_stat[REGNO (temp)].nonzero_bits
2851 != GET_MODE_MASK (word_mode))))
2852 && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG
2853 && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))),
2854 (REG_P (temp)
2855 && reg_stat[REGNO (temp)].nonzero_bits != 0
2856 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2857 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2858 && (reg_stat[REGNO (temp)].nonzero_bits
2859 != GET_MODE_MASK (word_mode)))))
2860 && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2861 SET_SRC (XVECEXP (newpat, 0, 1)))
2862 && ! find_reg_note (i3, REG_UNUSED,
2863 SET_DEST (XVECEXP (newpat, 0, 0))))
2865 rtx ni2dest;
2867 newi2pat = XVECEXP (newpat, 0, 0);
2868 ni2dest = SET_DEST (XVECEXP (newpat, 0, 0));
2869 newpat = XVECEXP (newpat, 0, 1);
2870 SUBST (SET_SRC (newpat),
2871 gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest));
2872 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2874 if (i2_code_number >= 0)
2875 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2877 if (insn_code_number >= 0)
2878 swap_i2i3 = 1;
2881 /* Similarly, check for a case where we have a PARALLEL of two independent
2882 SETs but we started with three insns. In this case, we can do the sets
2883 as two separate insns. This case occurs when some SET allows two
2884 other insns to combine, but the destination of that SET is still live. */
2886 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2887 && GET_CODE (newpat) == PARALLEL
2888 && XVECLEN (newpat, 0) == 2
2889 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2890 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT
2891 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART
2892 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2893 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2894 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2895 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2896 INSN_CUID (i2))
2897 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2898 XVECEXP (newpat, 0, 0))
2899 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)),
2900 XVECEXP (newpat, 0, 1))
2901 && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0)))
2902 && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1)))))
2904 /* Normally, it doesn't matter which of the two is done first,
2905 but it does if one references cc0. In that case, it has to
2906 be first. */
2907 #ifdef HAVE_cc0
2908 if (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0)))
2910 newi2pat = XVECEXP (newpat, 0, 0);
2911 newpat = XVECEXP (newpat, 0, 1);
2913 else
2914 #endif
2916 newi2pat = XVECEXP (newpat, 0, 1);
2917 newpat = XVECEXP (newpat, 0, 0);
2920 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2922 if (i2_code_number >= 0)
2923 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2926 /* If it still isn't recognized, fail and change things back the way they
2927 were. */
2928 if ((insn_code_number < 0
2929 /* Is the result a reasonable ASM_OPERANDS? */
2930 && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2)))
2932 undo_all ();
2933 return 0;
2936 /* If we had to change another insn, make sure it is valid also. */
2937 if (undobuf.other_insn)
2939 rtx other_pat = PATTERN (undobuf.other_insn);
2940 rtx new_other_notes;
2941 rtx note, next;
2943 CLEAR_HARD_REG_SET (newpat_used_regs);
2945 other_code_number = recog_for_combine (&other_pat, undobuf.other_insn,
2946 &new_other_notes);
2948 if (other_code_number < 0 && ! check_asm_operands (other_pat))
2950 undo_all ();
2951 return 0;
2954 PATTERN (undobuf.other_insn) = other_pat;
2956 /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they
2957 are still valid. Then add any non-duplicate notes added by
2958 recog_for_combine. */
2959 for (note = REG_NOTES (undobuf.other_insn); note; note = next)
2961 next = XEXP (note, 1);
2963 if (REG_NOTE_KIND (note) == REG_UNUSED
2964 && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn)))
2966 if (REG_P (XEXP (note, 0)))
2967 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
2969 remove_note (undobuf.other_insn, note);
2973 for (note = new_other_notes; note; note = XEXP (note, 1))
2974 if (REG_P (XEXP (note, 0)))
2975 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
2977 distribute_notes (new_other_notes, undobuf.other_insn,
2978 undobuf.other_insn, NULL_RTX, NULL_RTX, NULL_RTX);
2980 #ifdef HAVE_cc0
2981 /* If I2 is the CC0 setter and I3 is the CC0 user then check whether
2982 they are adjacent to each other or not. */
2984 rtx p = prev_nonnote_insn (i3);
2985 if (p && p != i2 && NONJUMP_INSN_P (p) && newi2pat
2986 && sets_cc0_p (newi2pat))
2988 undo_all ();
2989 return 0;
2992 #endif
2994 /* Only allow this combination if insn_rtx_costs reports that the
2995 replacement instructions are cheaper than the originals. */
2996 if (!combine_validate_cost (i1, i2, i3, newpat, newi2pat))
2998 undo_all ();
2999 return 0;
3002 /* We now know that we can do this combination. Merge the insns and
3003 update the status of registers and LOG_LINKS. */
3005 if (swap_i2i3)
3007 rtx insn;
3008 rtx link;
3009 rtx ni2dest;
3011 /* I3 now uses what used to be its destination and which is now
3012 I2's destination. This requires us to do a few adjustments. */
3013 PATTERN (i3) = newpat;
3014 adjust_for_new_dest (i3);
3016 /* We need a LOG_LINK from I3 to I2. But we used to have one,
3017 so we still will.
3019 However, some later insn might be using I2's dest and have
3020 a LOG_LINK pointing at I3. We must remove this link.
3021 The simplest way to remove the link is to point it at I1,
3022 which we know will be a NOTE. */
3024 /* newi2pat is usually a SET here; however, recog_for_combine might
3025 have added some clobbers. */
3026 if (GET_CODE (newi2pat) == PARALLEL)
3027 ni2dest = SET_DEST (XVECEXP (newi2pat, 0, 0));
3028 else
3029 ni2dest = SET_DEST (newi2pat);
3031 for (insn = NEXT_INSN (i3);
3032 insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3033 || insn != BB_HEAD (this_basic_block->next_bb));
3034 insn = NEXT_INSN (insn))
3036 if (INSN_P (insn) && reg_referenced_p (ni2dest, PATTERN (insn)))
3038 for (link = LOG_LINKS (insn); link;
3039 link = XEXP (link, 1))
3040 if (XEXP (link, 0) == i3)
3041 XEXP (link, 0) = i1;
3043 break;
3049 rtx i3notes, i2notes, i1notes = 0;
3050 rtx i3links, i2links, i1links = 0;
3051 rtx midnotes = 0;
3052 unsigned int regno;
3053 /* Compute which registers we expect to eliminate. newi2pat may be setting
3054 either i3dest or i2dest, so we must check it. Also, i1dest may be the
3055 same as i3dest, in which case newi2pat may be setting i1dest. */
3056 rtx elim_i2 = ((newi2pat && reg_set_p (i2dest, newi2pat))
3057 || i2dest_in_i2src || i2dest_in_i1src
3058 || !i2dest_killed
3059 ? 0 : i2dest);
3060 rtx elim_i1 = (i1 == 0 || i1dest_in_i1src
3061 || (newi2pat && reg_set_p (i1dest, newi2pat))
3062 || !i1dest_killed
3063 ? 0 : i1dest);
3065 /* Get the old REG_NOTES and LOG_LINKS from all our insns and
3066 clear them. */
3067 i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3);
3068 i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2);
3069 if (i1)
3070 i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1);
3072 /* Ensure that we do not have something that should not be shared but
3073 occurs multiple times in the new insns. Check this by first
3074 resetting all the `used' flags and then copying anything is shared. */
3076 reset_used_flags (i3notes);
3077 reset_used_flags (i2notes);
3078 reset_used_flags (i1notes);
3079 reset_used_flags (newpat);
3080 reset_used_flags (newi2pat);
3081 if (undobuf.other_insn)
3082 reset_used_flags (PATTERN (undobuf.other_insn));
3084 i3notes = copy_rtx_if_shared (i3notes);
3085 i2notes = copy_rtx_if_shared (i2notes);
3086 i1notes = copy_rtx_if_shared (i1notes);
3087 newpat = copy_rtx_if_shared (newpat);
3088 newi2pat = copy_rtx_if_shared (newi2pat);
3089 if (undobuf.other_insn)
3090 reset_used_flags (PATTERN (undobuf.other_insn));
3092 INSN_CODE (i3) = insn_code_number;
3093 PATTERN (i3) = newpat;
3095 if (CALL_P (i3) && CALL_INSN_FUNCTION_USAGE (i3))
3097 rtx call_usage = CALL_INSN_FUNCTION_USAGE (i3);
3099 reset_used_flags (call_usage);
3100 call_usage = copy_rtx (call_usage);
3102 if (substed_i2)
3103 replace_rtx (call_usage, i2dest, i2src);
3105 if (substed_i1)
3106 replace_rtx (call_usage, i1dest, i1src);
3108 CALL_INSN_FUNCTION_USAGE (i3) = call_usage;
3111 if (undobuf.other_insn)
3112 INSN_CODE (undobuf.other_insn) = other_code_number;
3114 /* We had one special case above where I2 had more than one set and
3115 we replaced a destination of one of those sets with the destination
3116 of I3. In that case, we have to update LOG_LINKS of insns later
3117 in this basic block. Note that this (expensive) case is rare.
3119 Also, in this case, we must pretend that all REG_NOTEs for I2
3120 actually came from I3, so that REG_UNUSED notes from I2 will be
3121 properly handled. */
3123 if (i3_subst_into_i2)
3125 for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++)
3126 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != USE
3127 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, i)))
3128 && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest
3129 && ! find_reg_note (i2, REG_UNUSED,
3130 SET_DEST (XVECEXP (PATTERN (i2), 0, i))))
3131 for (temp = NEXT_INSN (i2);
3132 temp && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3133 || BB_HEAD (this_basic_block) != temp);
3134 temp = NEXT_INSN (temp))
3135 if (temp != i3 && INSN_P (temp))
3136 for (link = LOG_LINKS (temp); link; link = XEXP (link, 1))
3137 if (XEXP (link, 0) == i2)
3138 XEXP (link, 0) = i3;
3140 if (i3notes)
3142 rtx link = i3notes;
3143 while (XEXP (link, 1))
3144 link = XEXP (link, 1);
3145 XEXP (link, 1) = i2notes;
3147 else
3148 i3notes = i2notes;
3149 i2notes = 0;
3152 LOG_LINKS (i3) = 0;
3153 REG_NOTES (i3) = 0;
3154 LOG_LINKS (i2) = 0;
3155 REG_NOTES (i2) = 0;
3157 if (newi2pat)
3159 INSN_CODE (i2) = i2_code_number;
3160 PATTERN (i2) = newi2pat;
3162 else
3163 SET_INSN_DELETED (i2);
3165 if (i1)
3167 LOG_LINKS (i1) = 0;
3168 REG_NOTES (i1) = 0;
3169 SET_INSN_DELETED (i1);
3172 /* Get death notes for everything that is now used in either I3 or
3173 I2 and used to die in a previous insn. If we built two new
3174 patterns, move from I1 to I2 then I2 to I3 so that we get the
3175 proper movement on registers that I2 modifies. */
3177 if (newi2pat)
3179 move_deaths (newi2pat, NULL_RTX, INSN_CUID (i1), i2, &midnotes);
3180 move_deaths (newpat, newi2pat, INSN_CUID (i1), i3, &midnotes);
3182 else
3183 move_deaths (newpat, NULL_RTX, i1 ? INSN_CUID (i1) : INSN_CUID (i2),
3184 i3, &midnotes);
3186 /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */
3187 if (i3notes)
3188 distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX,
3189 elim_i2, elim_i1);
3190 if (i2notes)
3191 distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX,
3192 elim_i2, elim_i1);
3193 if (i1notes)
3194 distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX,
3195 elim_i2, elim_i1);
3196 if (midnotes)
3197 distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3198 elim_i2, elim_i1);
3200 /* Distribute any notes added to I2 or I3 by recog_for_combine. We
3201 know these are REG_UNUSED and want them to go to the desired insn,
3202 so we always pass it as i3. We have not counted the notes in
3203 reg_n_deaths yet, so we need to do so now. */
3205 if (newi2pat && new_i2_notes)
3207 for (temp = new_i2_notes; temp; temp = XEXP (temp, 1))
3208 if (REG_P (XEXP (temp, 0)))
3209 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
3211 distribute_notes (new_i2_notes, i2, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3214 if (new_i3_notes)
3216 for (temp = new_i3_notes; temp; temp = XEXP (temp, 1))
3217 if (REG_P (XEXP (temp, 0)))
3218 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
3220 distribute_notes (new_i3_notes, i3, i3, NULL_RTX, NULL_RTX, NULL_RTX);
3223 /* If I3DEST was used in I3SRC, it really died in I3. We may need to
3224 put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets
3225 I3DEST, the death must be somewhere before I2, not I3. If we passed I3
3226 in that case, it might delete I2. Similarly for I2 and I1.
3227 Show an additional death due to the REG_DEAD note we make here. If
3228 we discard it in distribute_notes, we will decrement it again. */
3230 if (i3dest_killed)
3232 if (REG_P (i3dest_killed))
3233 REG_N_DEATHS (REGNO (i3dest_killed))++;
3235 if (newi2pat && reg_set_p (i3dest_killed, newi2pat))
3236 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
3237 NULL_RTX),
3238 NULL_RTX, i2, NULL_RTX, elim_i2, elim_i1);
3239 else
3240 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
3241 NULL_RTX),
3242 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3243 elim_i2, elim_i1);
3246 if (i2dest_in_i2src)
3248 if (REG_P (i2dest))
3249 REG_N_DEATHS (REGNO (i2dest))++;
3251 if (newi2pat && reg_set_p (i2dest, newi2pat))
3252 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
3253 NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3254 else
3255 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
3256 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3257 NULL_RTX, NULL_RTX);
3260 if (i1dest_in_i1src)
3262 if (REG_P (i1dest))
3263 REG_N_DEATHS (REGNO (i1dest))++;
3265 if (newi2pat && reg_set_p (i1dest, newi2pat))
3266 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
3267 NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3268 else
3269 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
3270 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3271 NULL_RTX, NULL_RTX);
3274 distribute_links (i3links);
3275 distribute_links (i2links);
3276 distribute_links (i1links);
3278 if (REG_P (i2dest))
3280 rtx link;
3281 rtx i2_insn = 0, i2_val = 0, set;
3283 /* The insn that used to set this register doesn't exist, and
3284 this life of the register may not exist either. See if one of
3285 I3's links points to an insn that sets I2DEST. If it does,
3286 that is now the last known value for I2DEST. If we don't update
3287 this and I2 set the register to a value that depended on its old
3288 contents, we will get confused. If this insn is used, thing
3289 will be set correctly in combine_instructions. */
3291 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
3292 if ((set = single_set (XEXP (link, 0))) != 0
3293 && rtx_equal_p (i2dest, SET_DEST (set)))
3294 i2_insn = XEXP (link, 0), i2_val = SET_SRC (set);
3296 record_value_for_reg (i2dest, i2_insn, i2_val);
3298 /* If the reg formerly set in I2 died only once and that was in I3,
3299 zero its use count so it won't make `reload' do any work. */
3300 if (! added_sets_2
3301 && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat))
3302 && ! i2dest_in_i2src)
3304 regno = REGNO (i2dest);
3305 REG_N_SETS (regno)--;
3309 if (i1 && REG_P (i1dest))
3311 rtx link;
3312 rtx i1_insn = 0, i1_val = 0, set;
3314 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
3315 if ((set = single_set (XEXP (link, 0))) != 0
3316 && rtx_equal_p (i1dest, SET_DEST (set)))
3317 i1_insn = XEXP (link, 0), i1_val = SET_SRC (set);
3319 record_value_for_reg (i1dest, i1_insn, i1_val);
3321 regno = REGNO (i1dest);
3322 if (! added_sets_1 && ! i1dest_in_i1src)
3323 REG_N_SETS (regno)--;
3326 /* Update reg_stat[].nonzero_bits et al for any changes that may have
3327 been made to this insn. The order of
3328 set_nonzero_bits_and_sign_copies() is important. Because newi2pat
3329 can affect nonzero_bits of newpat */
3330 if (newi2pat)
3331 note_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL);
3332 note_stores (newpat, set_nonzero_bits_and_sign_copies, NULL);
3334 /* Set new_direct_jump_p if a new return or simple jump instruction
3335 has been created.
3337 If I3 is now an unconditional jump, ensure that it has a
3338 BARRIER following it since it may have initially been a
3339 conditional jump. It may also be the last nonnote insn. */
3341 if (returnjump_p (i3) || any_uncondjump_p (i3))
3343 *new_direct_jump_p = 1;
3344 mark_jump_label (PATTERN (i3), i3, 0);
3346 if ((temp = next_nonnote_insn (i3)) == NULL_RTX
3347 || !BARRIER_P (temp))
3348 emit_barrier_after (i3);
3351 if (undobuf.other_insn != NULL_RTX
3352 && (returnjump_p (undobuf.other_insn)
3353 || any_uncondjump_p (undobuf.other_insn)))
3355 *new_direct_jump_p = 1;
3357 if ((temp = next_nonnote_insn (undobuf.other_insn)) == NULL_RTX
3358 || !BARRIER_P (temp))
3359 emit_barrier_after (undobuf.other_insn);
3362 /* An NOOP jump does not need barrier, but it does need cleaning up
3363 of CFG. */
3364 if (GET_CODE (newpat) == SET
3365 && SET_SRC (newpat) == pc_rtx
3366 && SET_DEST (newpat) == pc_rtx)
3367 *new_direct_jump_p = 1;
3370 combine_successes++;
3371 undo_commit ();
3373 if (added_links_insn
3374 && (newi2pat == 0 || INSN_CUID (added_links_insn) < INSN_CUID (i2))
3375 && INSN_CUID (added_links_insn) < INSN_CUID (i3))
3376 return added_links_insn;
3377 else
3378 return newi2pat ? i2 : i3;
3381 /* Undo all the modifications recorded in undobuf. */
3383 static void
3384 undo_all (void)
3386 struct undo *undo, *next;
3388 for (undo = undobuf.undos; undo; undo = next)
3390 next = undo->next;
3391 switch (undo->kind)
3393 case UNDO_RTX:
3394 *undo->where.r = undo->old_contents.r;
3395 break;
3396 case UNDO_INT:
3397 *undo->where.i = undo->old_contents.i;
3398 break;
3399 case UNDO_MODE:
3400 PUT_MODE (*undo->where.r, undo->old_contents.m);
3401 break;
3402 default:
3403 gcc_unreachable ();
3406 undo->next = undobuf.frees;
3407 undobuf.frees = undo;
3410 undobuf.undos = 0;
3413 /* We've committed to accepting the changes we made. Move all
3414 of the undos to the free list. */
3416 static void
3417 undo_commit (void)
3419 struct undo *undo, *next;
3421 for (undo = undobuf.undos; undo; undo = next)
3423 next = undo->next;
3424 undo->next = undobuf.frees;
3425 undobuf.frees = undo;
3427 undobuf.undos = 0;
3431 /* Find the innermost point within the rtx at LOC, possibly LOC itself,
3432 where we have an arithmetic expression and return that point. LOC will
3433 be inside INSN.
3435 try_combine will call this function to see if an insn can be split into
3436 two insns. */
3438 static rtx *
3439 find_split_point (rtx *loc, rtx insn)
3441 rtx x = *loc;
3442 enum rtx_code code = GET_CODE (x);
3443 rtx *split;
3444 unsigned HOST_WIDE_INT len = 0;
3445 HOST_WIDE_INT pos = 0;
3446 int unsignedp = 0;
3447 rtx inner = NULL_RTX;
3449 /* First special-case some codes. */
3450 switch (code)
3452 case SUBREG:
3453 #ifdef INSN_SCHEDULING
3454 /* If we are making a paradoxical SUBREG invalid, it becomes a split
3455 point. */
3456 if (MEM_P (SUBREG_REG (x)))
3457 return loc;
3458 #endif
3459 return find_split_point (&SUBREG_REG (x), insn);
3461 case MEM:
3462 #ifdef HAVE_lo_sum
3463 /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it
3464 using LO_SUM and HIGH. */
3465 if (GET_CODE (XEXP (x, 0)) == CONST
3466 || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
3468 SUBST (XEXP (x, 0),
3469 gen_rtx_LO_SUM (Pmode,
3470 gen_rtx_HIGH (Pmode, XEXP (x, 0)),
3471 XEXP (x, 0)));
3472 return &XEXP (XEXP (x, 0), 0);
3474 #endif
3476 /* If we have a PLUS whose second operand is a constant and the
3477 address is not valid, perhaps will can split it up using
3478 the machine-specific way to split large constants. We use
3479 the first pseudo-reg (one of the virtual regs) as a placeholder;
3480 it will not remain in the result. */
3481 if (GET_CODE (XEXP (x, 0)) == PLUS
3482 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3483 && ! memory_address_p (GET_MODE (x), XEXP (x, 0)))
3485 rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER];
3486 rtx seq = split_insns (gen_rtx_SET (VOIDmode, reg, XEXP (x, 0)),
3487 subst_insn);
3489 /* This should have produced two insns, each of which sets our
3490 placeholder. If the source of the second is a valid address,
3491 we can make put both sources together and make a split point
3492 in the middle. */
3494 if (seq
3495 && NEXT_INSN (seq) != NULL_RTX
3496 && NEXT_INSN (NEXT_INSN (seq)) == NULL_RTX
3497 && NONJUMP_INSN_P (seq)
3498 && GET_CODE (PATTERN (seq)) == SET
3499 && SET_DEST (PATTERN (seq)) == reg
3500 && ! reg_mentioned_p (reg,
3501 SET_SRC (PATTERN (seq)))
3502 && NONJUMP_INSN_P (NEXT_INSN (seq))
3503 && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET
3504 && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg
3505 && memory_address_p (GET_MODE (x),
3506 SET_SRC (PATTERN (NEXT_INSN (seq)))))
3508 rtx src1 = SET_SRC (PATTERN (seq));
3509 rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq)));
3511 /* Replace the placeholder in SRC2 with SRC1. If we can
3512 find where in SRC2 it was placed, that can become our
3513 split point and we can replace this address with SRC2.
3514 Just try two obvious places. */
3516 src2 = replace_rtx (src2, reg, src1);
3517 split = 0;
3518 if (XEXP (src2, 0) == src1)
3519 split = &XEXP (src2, 0);
3520 else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e'
3521 && XEXP (XEXP (src2, 0), 0) == src1)
3522 split = &XEXP (XEXP (src2, 0), 0);
3524 if (split)
3526 SUBST (XEXP (x, 0), src2);
3527 return split;
3531 /* If that didn't work, perhaps the first operand is complex and
3532 needs to be computed separately, so make a split point there.
3533 This will occur on machines that just support REG + CONST
3534 and have a constant moved through some previous computation. */
3536 else if (!OBJECT_P (XEXP (XEXP (x, 0), 0))
3537 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
3538 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
3539 return &XEXP (XEXP (x, 0), 0);
3541 break;
3543 case SET:
3544 #ifdef HAVE_cc0
3545 /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a
3546 ZERO_EXTRACT, the most likely reason why this doesn't match is that
3547 we need to put the operand into a register. So split at that
3548 point. */
3550 if (SET_DEST (x) == cc0_rtx
3551 && GET_CODE (SET_SRC (x)) != COMPARE
3552 && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT
3553 && !OBJECT_P (SET_SRC (x))
3554 && ! (GET_CODE (SET_SRC (x)) == SUBREG
3555 && OBJECT_P (SUBREG_REG (SET_SRC (x)))))
3556 return &SET_SRC (x);
3557 #endif
3559 /* See if we can split SET_SRC as it stands. */
3560 split = find_split_point (&SET_SRC (x), insn);
3561 if (split && split != &SET_SRC (x))
3562 return split;
3564 /* See if we can split SET_DEST as it stands. */
3565 split = find_split_point (&SET_DEST (x), insn);
3566 if (split && split != &SET_DEST (x))
3567 return split;
3569 /* See if this is a bitfield assignment with everything constant. If
3570 so, this is an IOR of an AND, so split it into that. */
3571 if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
3572 && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))
3573 <= HOST_BITS_PER_WIDE_INT)
3574 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT
3575 && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT
3576 && GET_CODE (SET_SRC (x)) == CONST_INT
3577 && ((INTVAL (XEXP (SET_DEST (x), 1))
3578 + INTVAL (XEXP (SET_DEST (x), 2)))
3579 <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))))
3580 && ! side_effects_p (XEXP (SET_DEST (x), 0)))
3582 HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2));
3583 unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1));
3584 unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x));
3585 rtx dest = XEXP (SET_DEST (x), 0);
3586 enum machine_mode mode = GET_MODE (dest);
3587 unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1;
3588 rtx or_mask;
3590 if (BITS_BIG_ENDIAN)
3591 pos = GET_MODE_BITSIZE (mode) - len - pos;
3593 or_mask = gen_int_mode (src << pos, mode);
3594 if (src == mask)
3595 SUBST (SET_SRC (x),
3596 simplify_gen_binary (IOR, mode, dest, or_mask));
3597 else
3599 rtx negmask = gen_int_mode (~(mask << pos), mode);
3600 SUBST (SET_SRC (x),
3601 simplify_gen_binary (IOR, mode,
3602 simplify_gen_binary (AND, mode,
3603 dest, negmask),
3604 or_mask));
3607 SUBST (SET_DEST (x), dest);
3609 split = find_split_point (&SET_SRC (x), insn);
3610 if (split && split != &SET_SRC (x))
3611 return split;
3614 /* Otherwise, see if this is an operation that we can split into two.
3615 If so, try to split that. */
3616 code = GET_CODE (SET_SRC (x));
3618 switch (code)
3620 case AND:
3621 /* If we are AND'ing with a large constant that is only a single
3622 bit and the result is only being used in a context where we
3623 need to know if it is zero or nonzero, replace it with a bit
3624 extraction. This will avoid the large constant, which might
3625 have taken more than one insn to make. If the constant were
3626 not a valid argument to the AND but took only one insn to make,
3627 this is no worse, but if it took more than one insn, it will
3628 be better. */
3630 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3631 && REG_P (XEXP (SET_SRC (x), 0))
3632 && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7
3633 && REG_P (SET_DEST (x))
3634 && (split = find_single_use (SET_DEST (x), insn, (rtx*) 0)) != 0
3635 && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE)
3636 && XEXP (*split, 0) == SET_DEST (x)
3637 && XEXP (*split, 1) == const0_rtx)
3639 rtx extraction = make_extraction (GET_MODE (SET_DEST (x)),
3640 XEXP (SET_SRC (x), 0),
3641 pos, NULL_RTX, 1, 1, 0, 0);
3642 if (extraction != 0)
3644 SUBST (SET_SRC (x), extraction);
3645 return find_split_point (loc, insn);
3648 break;
3650 case NE:
3651 /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X
3652 is known to be on, this can be converted into a NEG of a shift. */
3653 if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx
3654 && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0))
3655 && 1 <= (pos = exact_log2
3656 (nonzero_bits (XEXP (SET_SRC (x), 0),
3657 GET_MODE (XEXP (SET_SRC (x), 0))))))
3659 enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0));
3661 SUBST (SET_SRC (x),
3662 gen_rtx_NEG (mode,
3663 gen_rtx_LSHIFTRT (mode,
3664 XEXP (SET_SRC (x), 0),
3665 GEN_INT (pos))));
3667 split = find_split_point (&SET_SRC (x), insn);
3668 if (split && split != &SET_SRC (x))
3669 return split;
3671 break;
3673 case SIGN_EXTEND:
3674 inner = XEXP (SET_SRC (x), 0);
3676 /* We can't optimize if either mode is a partial integer
3677 mode as we don't know how many bits are significant
3678 in those modes. */
3679 if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT
3680 || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT)
3681 break;
3683 pos = 0;
3684 len = GET_MODE_BITSIZE (GET_MODE (inner));
3685 unsignedp = 0;
3686 break;
3688 case SIGN_EXTRACT:
3689 case ZERO_EXTRACT:
3690 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3691 && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT)
3693 inner = XEXP (SET_SRC (x), 0);
3694 len = INTVAL (XEXP (SET_SRC (x), 1));
3695 pos = INTVAL (XEXP (SET_SRC (x), 2));
3697 if (BITS_BIG_ENDIAN)
3698 pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos;
3699 unsignedp = (code == ZERO_EXTRACT);
3701 break;
3703 default:
3704 break;
3707 if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner)))
3709 enum machine_mode mode = GET_MODE (SET_SRC (x));
3711 /* For unsigned, we have a choice of a shift followed by an
3712 AND or two shifts. Use two shifts for field sizes where the
3713 constant might be too large. We assume here that we can
3714 always at least get 8-bit constants in an AND insn, which is
3715 true for every current RISC. */
3717 if (unsignedp && len <= 8)
3719 SUBST (SET_SRC (x),
3720 gen_rtx_AND (mode,
3721 gen_rtx_LSHIFTRT
3722 (mode, gen_lowpart (mode, inner),
3723 GEN_INT (pos)),
3724 GEN_INT (((HOST_WIDE_INT) 1 << len) - 1)));
3726 split = find_split_point (&SET_SRC (x), insn);
3727 if (split && split != &SET_SRC (x))
3728 return split;
3730 else
3732 SUBST (SET_SRC (x),
3733 gen_rtx_fmt_ee
3734 (unsignedp ? LSHIFTRT : ASHIFTRT, mode,
3735 gen_rtx_ASHIFT (mode,
3736 gen_lowpart (mode, inner),
3737 GEN_INT (GET_MODE_BITSIZE (mode)
3738 - len - pos)),
3739 GEN_INT (GET_MODE_BITSIZE (mode) - len)));
3741 split = find_split_point (&SET_SRC (x), insn);
3742 if (split && split != &SET_SRC (x))
3743 return split;
3747 /* See if this is a simple operation with a constant as the second
3748 operand. It might be that this constant is out of range and hence
3749 could be used as a split point. */
3750 if (BINARY_P (SET_SRC (x))
3751 && CONSTANT_P (XEXP (SET_SRC (x), 1))
3752 && (OBJECT_P (XEXP (SET_SRC (x), 0))
3753 || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG
3754 && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0))))))
3755 return &XEXP (SET_SRC (x), 1);
3757 /* Finally, see if this is a simple operation with its first operand
3758 not in a register. The operation might require this operand in a
3759 register, so return it as a split point. We can always do this
3760 because if the first operand were another operation, we would have
3761 already found it as a split point. */
3762 if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x)))
3763 && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode))
3764 return &XEXP (SET_SRC (x), 0);
3766 return 0;
3768 case AND:
3769 case IOR:
3770 /* We write NOR as (and (not A) (not B)), but if we don't have a NOR,
3771 it is better to write this as (not (ior A B)) so we can split it.
3772 Similarly for IOR. */
3773 if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT)
3775 SUBST (*loc,
3776 gen_rtx_NOT (GET_MODE (x),
3777 gen_rtx_fmt_ee (code == IOR ? AND : IOR,
3778 GET_MODE (x),
3779 XEXP (XEXP (x, 0), 0),
3780 XEXP (XEXP (x, 1), 0))));
3781 return find_split_point (loc, insn);
3784 /* Many RISC machines have a large set of logical insns. If the
3785 second operand is a NOT, put it first so we will try to split the
3786 other operand first. */
3787 if (GET_CODE (XEXP (x, 1)) == NOT)
3789 rtx tem = XEXP (x, 0);
3790 SUBST (XEXP (x, 0), XEXP (x, 1));
3791 SUBST (XEXP (x, 1), tem);
3793 break;
3795 default:
3796 break;
3799 /* Otherwise, select our actions depending on our rtx class. */
3800 switch (GET_RTX_CLASS (code))
3802 case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */
3803 case RTX_TERNARY:
3804 split = find_split_point (&XEXP (x, 2), insn);
3805 if (split)
3806 return split;
3807 /* ... fall through ... */
3808 case RTX_BIN_ARITH:
3809 case RTX_COMM_ARITH:
3810 case RTX_COMPARE:
3811 case RTX_COMM_COMPARE:
3812 split = find_split_point (&XEXP (x, 1), insn);
3813 if (split)
3814 return split;
3815 /* ... fall through ... */
3816 case RTX_UNARY:
3817 /* Some machines have (and (shift ...) ...) insns. If X is not
3818 an AND, but XEXP (X, 0) is, use it as our split point. */
3819 if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND)
3820 return &XEXP (x, 0);
3822 split = find_split_point (&XEXP (x, 0), insn);
3823 if (split)
3824 return split;
3825 return loc;
3827 default:
3828 /* Otherwise, we don't have a split point. */
3829 return 0;
3833 /* Throughout X, replace FROM with TO, and return the result.
3834 The result is TO if X is FROM;
3835 otherwise the result is X, but its contents may have been modified.
3836 If they were modified, a record was made in undobuf so that
3837 undo_all will (among other things) return X to its original state.
3839 If the number of changes necessary is too much to record to undo,
3840 the excess changes are not made, so the result is invalid.
3841 The changes already made can still be undone.
3842 undobuf.num_undo is incremented for such changes, so by testing that
3843 the caller can tell whether the result is valid.
3845 `n_occurrences' is incremented each time FROM is replaced.
3847 IN_DEST is nonzero if we are processing the SET_DEST of a SET.
3849 UNIQUE_COPY is nonzero if each substitution must be unique. We do this
3850 by copying if `n_occurrences' is nonzero. */
3852 static rtx
3853 subst (rtx x, rtx from, rtx to, int in_dest, int unique_copy)
3855 enum rtx_code code = GET_CODE (x);
3856 enum machine_mode op0_mode = VOIDmode;
3857 const char *fmt;
3858 int len, i;
3859 rtx new;
3861 /* Two expressions are equal if they are identical copies of a shared
3862 RTX or if they are both registers with the same register number
3863 and mode. */
3865 #define COMBINE_RTX_EQUAL_P(X,Y) \
3866 ((X) == (Y) \
3867 || (REG_P (X) && REG_P (Y) \
3868 && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y)))
3870 if (! in_dest && COMBINE_RTX_EQUAL_P (x, from))
3872 n_occurrences++;
3873 return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to);
3876 /* If X and FROM are the same register but different modes, they will
3877 not have been seen as equal above. However, flow.c will make a
3878 LOG_LINKS entry for that case. If we do nothing, we will try to
3879 rerecognize our original insn and, when it succeeds, we will
3880 delete the feeding insn, which is incorrect.
3882 So force this insn not to match in this (rare) case. */
3883 if (! in_dest && code == REG && REG_P (from)
3884 && REGNO (x) == REGNO (from))
3885 return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
3887 /* If this is an object, we are done unless it is a MEM or LO_SUM, both
3888 of which may contain things that can be combined. */
3889 if (code != MEM && code != LO_SUM && OBJECT_P (x))
3890 return x;
3892 /* It is possible to have a subexpression appear twice in the insn.
3893 Suppose that FROM is a register that appears within TO.
3894 Then, after that subexpression has been scanned once by `subst',
3895 the second time it is scanned, TO may be found. If we were
3896 to scan TO here, we would find FROM within it and create a
3897 self-referent rtl structure which is completely wrong. */
3898 if (COMBINE_RTX_EQUAL_P (x, to))
3899 return to;
3901 /* Parallel asm_operands need special attention because all of the
3902 inputs are shared across the arms. Furthermore, unsharing the
3903 rtl results in recognition failures. Failure to handle this case
3904 specially can result in circular rtl.
3906 Solve this by doing a normal pass across the first entry of the
3907 parallel, and only processing the SET_DESTs of the subsequent
3908 entries. Ug. */
3910 if (code == PARALLEL
3911 && GET_CODE (XVECEXP (x, 0, 0)) == SET
3912 && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS)
3914 new = subst (XVECEXP (x, 0, 0), from, to, 0, unique_copy);
3916 /* If this substitution failed, this whole thing fails. */
3917 if (GET_CODE (new) == CLOBBER
3918 && XEXP (new, 0) == const0_rtx)
3919 return new;
3921 SUBST (XVECEXP (x, 0, 0), new);
3923 for (i = XVECLEN (x, 0) - 1; i >= 1; i--)
3925 rtx dest = SET_DEST (XVECEXP (x, 0, i));
3927 if (!REG_P (dest)
3928 && GET_CODE (dest) != CC0
3929 && GET_CODE (dest) != PC)
3931 new = subst (dest, from, to, 0, unique_copy);
3933 /* If this substitution failed, this whole thing fails. */
3934 if (GET_CODE (new) == CLOBBER
3935 && XEXP (new, 0) == const0_rtx)
3936 return new;
3938 SUBST (SET_DEST (XVECEXP (x, 0, i)), new);
3942 else
3944 len = GET_RTX_LENGTH (code);
3945 fmt = GET_RTX_FORMAT (code);
3947 /* We don't need to process a SET_DEST that is a register, CC0,
3948 or PC, so set up to skip this common case. All other cases
3949 where we want to suppress replacing something inside a
3950 SET_SRC are handled via the IN_DEST operand. */
3951 if (code == SET
3952 && (REG_P (SET_DEST (x))
3953 || GET_CODE (SET_DEST (x)) == CC0
3954 || GET_CODE (SET_DEST (x)) == PC))
3955 fmt = "ie";
3957 /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a
3958 constant. */
3959 if (fmt[0] == 'e')
3960 op0_mode = GET_MODE (XEXP (x, 0));
3962 for (i = 0; i < len; i++)
3964 if (fmt[i] == 'E')
3966 int j;
3967 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3969 if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from))
3971 new = (unique_copy && n_occurrences
3972 ? copy_rtx (to) : to);
3973 n_occurrences++;
3975 else
3977 new = subst (XVECEXP (x, i, j), from, to, 0,
3978 unique_copy);
3980 /* If this substitution failed, this whole thing
3981 fails. */
3982 if (GET_CODE (new) == CLOBBER
3983 && XEXP (new, 0) == const0_rtx)
3984 return new;
3987 SUBST (XVECEXP (x, i, j), new);
3990 else if (fmt[i] == 'e')
3992 /* If this is a register being set, ignore it. */
3993 new = XEXP (x, i);
3994 if (in_dest
3995 && i == 0
3996 && (((code == SUBREG || code == ZERO_EXTRACT)
3997 && REG_P (new))
3998 || code == STRICT_LOW_PART))
4001 else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from))
4003 /* In general, don't install a subreg involving two
4004 modes not tieable. It can worsen register
4005 allocation, and can even make invalid reload
4006 insns, since the reg inside may need to be copied
4007 from in the outside mode, and that may be invalid
4008 if it is an fp reg copied in integer mode.
4010 We allow two exceptions to this: It is valid if
4011 it is inside another SUBREG and the mode of that
4012 SUBREG and the mode of the inside of TO is
4013 tieable and it is valid if X is a SET that copies
4014 FROM to CC0. */
4016 if (GET_CODE (to) == SUBREG
4017 && ! MODES_TIEABLE_P (GET_MODE (to),
4018 GET_MODE (SUBREG_REG (to)))
4019 && ! (code == SUBREG
4020 && MODES_TIEABLE_P (GET_MODE (x),
4021 GET_MODE (SUBREG_REG (to))))
4022 #ifdef HAVE_cc0
4023 && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx)
4024 #endif
4026 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
4028 #ifdef CANNOT_CHANGE_MODE_CLASS
4029 if (code == SUBREG
4030 && REG_P (to)
4031 && REGNO (to) < FIRST_PSEUDO_REGISTER
4032 && REG_CANNOT_CHANGE_MODE_P (REGNO (to),
4033 GET_MODE (to),
4034 GET_MODE (x)))
4035 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
4036 #endif
4038 new = (unique_copy && n_occurrences ? copy_rtx (to) : to);
4039 n_occurrences++;
4041 else
4042 /* If we are in a SET_DEST, suppress most cases unless we
4043 have gone inside a MEM, in which case we want to
4044 simplify the address. We assume here that things that
4045 are actually part of the destination have their inner
4046 parts in the first expression. This is true for SUBREG,
4047 STRICT_LOW_PART, and ZERO_EXTRACT, which are the only
4048 things aside from REG and MEM that should appear in a
4049 SET_DEST. */
4050 new = subst (XEXP (x, i), from, to,
4051 (((in_dest
4052 && (code == SUBREG || code == STRICT_LOW_PART
4053 || code == ZERO_EXTRACT))
4054 || code == SET)
4055 && i == 0), unique_copy);
4057 /* If we found that we will have to reject this combination,
4058 indicate that by returning the CLOBBER ourselves, rather than
4059 an expression containing it. This will speed things up as
4060 well as prevent accidents where two CLOBBERs are considered
4061 to be equal, thus producing an incorrect simplification. */
4063 if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx)
4064 return new;
4066 if (GET_CODE (x) == SUBREG
4067 && (GET_CODE (new) == CONST_INT
4068 || GET_CODE (new) == CONST_DOUBLE))
4070 enum machine_mode mode = GET_MODE (x);
4072 x = simplify_subreg (GET_MODE (x), new,
4073 GET_MODE (SUBREG_REG (x)),
4074 SUBREG_BYTE (x));
4075 if (! x)
4076 x = gen_rtx_CLOBBER (mode, const0_rtx);
4078 else if (GET_CODE (new) == CONST_INT
4079 && GET_CODE (x) == ZERO_EXTEND)
4081 x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
4082 new, GET_MODE (XEXP (x, 0)));
4083 gcc_assert (x);
4085 else
4086 SUBST (XEXP (x, i), new);
4091 /* Try to simplify X. If the simplification changed the code, it is likely
4092 that further simplification will help, so loop, but limit the number
4093 of repetitions that will be performed. */
4095 for (i = 0; i < 4; i++)
4097 /* If X is sufficiently simple, don't bother trying to do anything
4098 with it. */
4099 if (code != CONST_INT && code != REG && code != CLOBBER)
4100 x = combine_simplify_rtx (x, op0_mode, in_dest);
4102 if (GET_CODE (x) == code)
4103 break;
4105 code = GET_CODE (x);
4107 /* We no longer know the original mode of operand 0 since we
4108 have changed the form of X) */
4109 op0_mode = VOIDmode;
4112 return x;
4115 /* Simplify X, a piece of RTL. We just operate on the expression at the
4116 outer level; call `subst' to simplify recursively. Return the new
4117 expression.
4119 OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero
4120 if we are inside a SET_DEST. */
4122 static rtx
4123 combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest)
4125 enum rtx_code code = GET_CODE (x);
4126 enum machine_mode mode = GET_MODE (x);
4127 rtx temp;
4128 int i;
4130 /* If this is a commutative operation, put a constant last and a complex
4131 expression first. We don't need to do this for comparisons here. */
4132 if (COMMUTATIVE_ARITH_P (x)
4133 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
4135 temp = XEXP (x, 0);
4136 SUBST (XEXP (x, 0), XEXP (x, 1));
4137 SUBST (XEXP (x, 1), temp);
4140 /* If this is a simple operation applied to an IF_THEN_ELSE, try
4141 applying it to the arms of the IF_THEN_ELSE. This often simplifies
4142 things. Check for cases where both arms are testing the same
4143 condition.
4145 Don't do anything if all operands are very simple. */
4147 if ((BINARY_P (x)
4148 && ((!OBJECT_P (XEXP (x, 0))
4149 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4150 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))
4151 || (!OBJECT_P (XEXP (x, 1))
4152 && ! (GET_CODE (XEXP (x, 1)) == SUBREG
4153 && OBJECT_P (SUBREG_REG (XEXP (x, 1)))))))
4154 || (UNARY_P (x)
4155 && (!OBJECT_P (XEXP (x, 0))
4156 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4157 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))))
4159 rtx cond, true_rtx, false_rtx;
4161 cond = if_then_else_cond (x, &true_rtx, &false_rtx);
4162 if (cond != 0
4163 /* If everything is a comparison, what we have is highly unlikely
4164 to be simpler, so don't use it. */
4165 && ! (COMPARISON_P (x)
4166 && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))))
4168 rtx cop1 = const0_rtx;
4169 enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1);
4171 if (cond_code == NE && COMPARISON_P (cond))
4172 return x;
4174 /* Simplify the alternative arms; this may collapse the true and
4175 false arms to store-flag values. Be careful to use copy_rtx
4176 here since true_rtx or false_rtx might share RTL with x as a
4177 result of the if_then_else_cond call above. */
4178 true_rtx = subst (copy_rtx (true_rtx), pc_rtx, pc_rtx, 0, 0);
4179 false_rtx = subst (copy_rtx (false_rtx), pc_rtx, pc_rtx, 0, 0);
4181 /* If true_rtx and false_rtx are not general_operands, an if_then_else
4182 is unlikely to be simpler. */
4183 if (general_operand (true_rtx, VOIDmode)
4184 && general_operand (false_rtx, VOIDmode))
4186 enum rtx_code reversed;
4188 /* Restarting if we generate a store-flag expression will cause
4189 us to loop. Just drop through in this case. */
4191 /* If the result values are STORE_FLAG_VALUE and zero, we can
4192 just make the comparison operation. */
4193 if (true_rtx == const_true_rtx && false_rtx == const0_rtx)
4194 x = simplify_gen_relational (cond_code, mode, VOIDmode,
4195 cond, cop1);
4196 else if (true_rtx == const0_rtx && false_rtx == const_true_rtx
4197 && ((reversed = reversed_comparison_code_parts
4198 (cond_code, cond, cop1, NULL))
4199 != UNKNOWN))
4200 x = simplify_gen_relational (reversed, mode, VOIDmode,
4201 cond, cop1);
4203 /* Likewise, we can make the negate of a comparison operation
4204 if the result values are - STORE_FLAG_VALUE and zero. */
4205 else if (GET_CODE (true_rtx) == CONST_INT
4206 && INTVAL (true_rtx) == - STORE_FLAG_VALUE
4207 && false_rtx == const0_rtx)
4208 x = simplify_gen_unary (NEG, mode,
4209 simplify_gen_relational (cond_code,
4210 mode, VOIDmode,
4211 cond, cop1),
4212 mode);
4213 else if (GET_CODE (false_rtx) == CONST_INT
4214 && INTVAL (false_rtx) == - STORE_FLAG_VALUE
4215 && true_rtx == const0_rtx
4216 && ((reversed = reversed_comparison_code_parts
4217 (cond_code, cond, cop1, NULL))
4218 != UNKNOWN))
4219 x = simplify_gen_unary (NEG, mode,
4220 simplify_gen_relational (reversed,
4221 mode, VOIDmode,
4222 cond, cop1),
4223 mode);
4224 else
4225 return gen_rtx_IF_THEN_ELSE (mode,
4226 simplify_gen_relational (cond_code,
4227 mode,
4228 VOIDmode,
4229 cond,
4230 cop1),
4231 true_rtx, false_rtx);
4233 code = GET_CODE (x);
4234 op0_mode = VOIDmode;
4239 /* Try to fold this expression in case we have constants that weren't
4240 present before. */
4241 temp = 0;
4242 switch (GET_RTX_CLASS (code))
4244 case RTX_UNARY:
4245 if (op0_mode == VOIDmode)
4246 op0_mode = GET_MODE (XEXP (x, 0));
4247 temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode);
4248 break;
4249 case RTX_COMPARE:
4250 case RTX_COMM_COMPARE:
4252 enum machine_mode cmp_mode = GET_MODE (XEXP (x, 0));
4253 if (cmp_mode == VOIDmode)
4255 cmp_mode = GET_MODE (XEXP (x, 1));
4256 if (cmp_mode == VOIDmode)
4257 cmp_mode = op0_mode;
4259 temp = simplify_relational_operation (code, mode, cmp_mode,
4260 XEXP (x, 0), XEXP (x, 1));
4262 break;
4263 case RTX_COMM_ARITH:
4264 case RTX_BIN_ARITH:
4265 temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1));
4266 break;
4267 case RTX_BITFIELD_OPS:
4268 case RTX_TERNARY:
4269 temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0),
4270 XEXP (x, 1), XEXP (x, 2));
4271 break;
4272 default:
4273 break;
4276 if (temp)
4278 x = temp;
4279 code = GET_CODE (temp);
4280 op0_mode = VOIDmode;
4281 mode = GET_MODE (temp);
4284 /* First see if we can apply the inverse distributive law. */
4285 if (code == PLUS || code == MINUS
4286 || code == AND || code == IOR || code == XOR)
4288 x = apply_distributive_law (x);
4289 code = GET_CODE (x);
4290 op0_mode = VOIDmode;
4293 /* If CODE is an associative operation not otherwise handled, see if we
4294 can associate some operands. This can win if they are constants or
4295 if they are logically related (i.e. (a & b) & a). */
4296 if ((code == PLUS || code == MINUS || code == MULT || code == DIV
4297 || code == AND || code == IOR || code == XOR
4298 || code == SMAX || code == SMIN || code == UMAX || code == UMIN)
4299 && ((INTEGRAL_MODE_P (mode) && code != DIV)
4300 || (flag_unsafe_math_optimizations && FLOAT_MODE_P (mode))))
4302 if (GET_CODE (XEXP (x, 0)) == code)
4304 rtx other = XEXP (XEXP (x, 0), 0);
4305 rtx inner_op0 = XEXP (XEXP (x, 0), 1);
4306 rtx inner_op1 = XEXP (x, 1);
4307 rtx inner;
4309 /* Make sure we pass the constant operand if any as the second
4310 one if this is a commutative operation. */
4311 if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x))
4313 rtx tem = inner_op0;
4314 inner_op0 = inner_op1;
4315 inner_op1 = tem;
4317 inner = simplify_binary_operation (code == MINUS ? PLUS
4318 : code == DIV ? MULT
4319 : code,
4320 mode, inner_op0, inner_op1);
4322 /* For commutative operations, try the other pair if that one
4323 didn't simplify. */
4324 if (inner == 0 && COMMUTATIVE_ARITH_P (x))
4326 other = XEXP (XEXP (x, 0), 1);
4327 inner = simplify_binary_operation (code, mode,
4328 XEXP (XEXP (x, 0), 0),
4329 XEXP (x, 1));
4332 if (inner)
4333 return simplify_gen_binary (code, mode, other, inner);
4337 /* A little bit of algebraic simplification here. */
4338 switch (code)
4340 case MEM:
4341 /* Ensure that our address has any ASHIFTs converted to MULT in case
4342 address-recognizing predicates are called later. */
4343 temp = make_compound_operation (XEXP (x, 0), MEM);
4344 SUBST (XEXP (x, 0), temp);
4345 break;
4347 case SUBREG:
4348 if (op0_mode == VOIDmode)
4349 op0_mode = GET_MODE (SUBREG_REG (x));
4351 /* See if this can be moved to simplify_subreg. */
4352 if (CONSTANT_P (SUBREG_REG (x))
4353 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
4354 /* Don't call gen_lowpart if the inner mode
4355 is VOIDmode and we cannot simplify it, as SUBREG without
4356 inner mode is invalid. */
4357 && (GET_MODE (SUBREG_REG (x)) != VOIDmode
4358 || gen_lowpart_common (mode, SUBREG_REG (x))))
4359 return gen_lowpart (mode, SUBREG_REG (x));
4361 if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC)
4362 break;
4364 rtx temp;
4365 temp = simplify_subreg (mode, SUBREG_REG (x), op0_mode,
4366 SUBREG_BYTE (x));
4367 if (temp)
4368 return temp;
4371 /* Don't change the mode of the MEM if that would change the meaning
4372 of the address. */
4373 if (MEM_P (SUBREG_REG (x))
4374 && (MEM_VOLATILE_P (SUBREG_REG (x))
4375 || mode_dependent_address_p (XEXP (SUBREG_REG (x), 0))))
4376 return gen_rtx_CLOBBER (mode, const0_rtx);
4378 /* Note that we cannot do any narrowing for non-constants since
4379 we might have been counting on using the fact that some bits were
4380 zero. We now do this in the SET. */
4382 break;
4384 case NEG:
4385 temp = expand_compound_operation (XEXP (x, 0));
4387 /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be
4388 replaced by (lshiftrt X C). This will convert
4389 (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */
4391 if (GET_CODE (temp) == ASHIFTRT
4392 && GET_CODE (XEXP (temp, 1)) == CONST_INT
4393 && INTVAL (XEXP (temp, 1)) == GET_MODE_BITSIZE (mode) - 1)
4394 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (temp, 0),
4395 INTVAL (XEXP (temp, 1)));
4397 /* If X has only a single bit that might be nonzero, say, bit I, convert
4398 (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of
4399 MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to
4400 (sign_extract X 1 Y). But only do this if TEMP isn't a register
4401 or a SUBREG of one since we'd be making the expression more
4402 complex if it was just a register. */
4404 if (!REG_P (temp)
4405 && ! (GET_CODE (temp) == SUBREG
4406 && REG_P (SUBREG_REG (temp)))
4407 && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0)
4409 rtx temp1 = simplify_shift_const
4410 (NULL_RTX, ASHIFTRT, mode,
4411 simplify_shift_const (NULL_RTX, ASHIFT, mode, temp,
4412 GET_MODE_BITSIZE (mode) - 1 - i),
4413 GET_MODE_BITSIZE (mode) - 1 - i);
4415 /* If all we did was surround TEMP with the two shifts, we
4416 haven't improved anything, so don't use it. Otherwise,
4417 we are better off with TEMP1. */
4418 if (GET_CODE (temp1) != ASHIFTRT
4419 || GET_CODE (XEXP (temp1, 0)) != ASHIFT
4420 || XEXP (XEXP (temp1, 0), 0) != temp)
4421 return temp1;
4423 break;
4425 case TRUNCATE:
4426 /* We can't handle truncation to a partial integer mode here
4427 because we don't know the real bitsize of the partial
4428 integer mode. */
4429 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
4430 break;
4432 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4433 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
4434 GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))))
4435 SUBST (XEXP (x, 0),
4436 force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
4437 GET_MODE_MASK (mode), 0));
4439 /* Similarly to what we do in simplify-rtx.c, a truncate of a register
4440 whose value is a comparison can be replaced with a subreg if
4441 STORE_FLAG_VALUE permits. */
4442 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4443 && ((HOST_WIDE_INT) STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0
4444 && (temp = get_last_value (XEXP (x, 0)))
4445 && COMPARISON_P (temp))
4446 return gen_lowpart (mode, XEXP (x, 0));
4447 break;
4449 #ifdef HAVE_cc0
4450 case COMPARE:
4451 /* Convert (compare FOO (const_int 0)) to FOO unless we aren't
4452 using cc0, in which case we want to leave it as a COMPARE
4453 so we can distinguish it from a register-register-copy. */
4454 if (XEXP (x, 1) == const0_rtx)
4455 return XEXP (x, 0);
4457 /* x - 0 is the same as x unless x's mode has signed zeros and
4458 allows rounding towards -infinity. Under those conditions,
4459 0 - 0 is -0. */
4460 if (!(HONOR_SIGNED_ZEROS (GET_MODE (XEXP (x, 0)))
4461 && HONOR_SIGN_DEPENDENT_ROUNDING (GET_MODE (XEXP (x, 0))))
4462 && XEXP (x, 1) == CONST0_RTX (GET_MODE (XEXP (x, 0))))
4463 return XEXP (x, 0);
4464 break;
4465 #endif
4467 case CONST:
4468 /* (const (const X)) can become (const X). Do it this way rather than
4469 returning the inner CONST since CONST can be shared with a
4470 REG_EQUAL note. */
4471 if (GET_CODE (XEXP (x, 0)) == CONST)
4472 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4473 break;
4475 #ifdef HAVE_lo_sum
4476 case LO_SUM:
4477 /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we
4478 can add in an offset. find_split_point will split this address up
4479 again if it doesn't match. */
4480 if (GET_CODE (XEXP (x, 0)) == HIGH
4481 && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)))
4482 return XEXP (x, 1);
4483 break;
4484 #endif
4486 case PLUS:
4487 /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>)
4488 when c is (const_int (pow2 + 1) / 2) is a sign extension of a
4489 bit-field and can be replaced by either a sign_extend or a
4490 sign_extract. The `and' may be a zero_extend and the two
4491 <c>, -<c> constants may be reversed. */
4492 if (GET_CODE (XEXP (x, 0)) == XOR
4493 && GET_CODE (XEXP (x, 1)) == CONST_INT
4494 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
4495 && INTVAL (XEXP (x, 1)) == -INTVAL (XEXP (XEXP (x, 0), 1))
4496 && ((i = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) >= 0
4497 || (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
4498 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4499 && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND
4500 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
4501 && (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))
4502 == ((HOST_WIDE_INT) 1 << (i + 1)) - 1))
4503 || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND
4504 && (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)))
4505 == (unsigned int) i + 1))))
4506 return simplify_shift_const
4507 (NULL_RTX, ASHIFTRT, mode,
4508 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4509 XEXP (XEXP (XEXP (x, 0), 0), 0),
4510 GET_MODE_BITSIZE (mode) - (i + 1)),
4511 GET_MODE_BITSIZE (mode) - (i + 1));
4513 /* If only the low-order bit of X is possibly nonzero, (plus x -1)
4514 can become (ashiftrt (ashift (xor x 1) C) C) where C is
4515 the bitsize of the mode - 1. This allows simplification of
4516 "a = (b & 8) == 0;" */
4517 if (XEXP (x, 1) == constm1_rtx
4518 && !REG_P (XEXP (x, 0))
4519 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4520 && REG_P (SUBREG_REG (XEXP (x, 0))))
4521 && nonzero_bits (XEXP (x, 0), mode) == 1)
4522 return simplify_shift_const (NULL_RTX, ASHIFTRT, mode,
4523 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4524 gen_rtx_XOR (mode, XEXP (x, 0), const1_rtx),
4525 GET_MODE_BITSIZE (mode) - 1),
4526 GET_MODE_BITSIZE (mode) - 1);
4528 /* If we are adding two things that have no bits in common, convert
4529 the addition into an IOR. This will often be further simplified,
4530 for example in cases like ((a & 1) + (a & 2)), which can
4531 become a & 3. */
4533 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4534 && (nonzero_bits (XEXP (x, 0), mode)
4535 & nonzero_bits (XEXP (x, 1), mode)) == 0)
4537 /* Try to simplify the expression further. */
4538 rtx tor = simplify_gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1));
4539 temp = combine_simplify_rtx (tor, mode, in_dest);
4541 /* If we could, great. If not, do not go ahead with the IOR
4542 replacement, since PLUS appears in many special purpose
4543 address arithmetic instructions. */
4544 if (GET_CODE (temp) != CLOBBER && temp != tor)
4545 return temp;
4547 break;
4549 case MINUS:
4550 /* (minus <foo> (and <foo> (const_int -pow2))) becomes
4551 (and <foo> (const_int pow2-1)) */
4552 if (GET_CODE (XEXP (x, 1)) == AND
4553 && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT
4554 && exact_log2 (-INTVAL (XEXP (XEXP (x, 1), 1))) >= 0
4555 && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0)))
4556 return simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0),
4557 -INTVAL (XEXP (XEXP (x, 1), 1)) - 1);
4558 break;
4560 case MULT:
4561 /* If we have (mult (plus A B) C), apply the distributive law and then
4562 the inverse distributive law to see if things simplify. This
4563 occurs mostly in addresses, often when unrolling loops. */
4565 if (GET_CODE (XEXP (x, 0)) == PLUS)
4567 rtx result = distribute_and_simplify_rtx (x, 0);
4568 if (result)
4569 return result;
4572 /* Try simplify a*(b/c) as (a*b)/c. */
4573 if (FLOAT_MODE_P (mode) && flag_unsafe_math_optimizations
4574 && GET_CODE (XEXP (x, 0)) == DIV)
4576 rtx tem = simplify_binary_operation (MULT, mode,
4577 XEXP (XEXP (x, 0), 0),
4578 XEXP (x, 1));
4579 if (tem)
4580 return simplify_gen_binary (DIV, mode, tem, XEXP (XEXP (x, 0), 1));
4582 break;
4584 case UDIV:
4585 /* If this is a divide by a power of two, treat it as a shift if
4586 its first operand is a shift. */
4587 if (GET_CODE (XEXP (x, 1)) == CONST_INT
4588 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0
4589 && (GET_CODE (XEXP (x, 0)) == ASHIFT
4590 || GET_CODE (XEXP (x, 0)) == LSHIFTRT
4591 || GET_CODE (XEXP (x, 0)) == ASHIFTRT
4592 || GET_CODE (XEXP (x, 0)) == ROTATE
4593 || GET_CODE (XEXP (x, 0)) == ROTATERT))
4594 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i);
4595 break;
4597 case EQ: case NE:
4598 case GT: case GTU: case GE: case GEU:
4599 case LT: case LTU: case LE: case LEU:
4600 case UNEQ: case LTGT:
4601 case UNGT: case UNGE:
4602 case UNLT: case UNLE:
4603 case UNORDERED: case ORDERED:
4604 /* If the first operand is a condition code, we can't do anything
4605 with it. */
4606 if (GET_CODE (XEXP (x, 0)) == COMPARE
4607 || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC
4608 && ! CC0_P (XEXP (x, 0))))
4610 rtx op0 = XEXP (x, 0);
4611 rtx op1 = XEXP (x, 1);
4612 enum rtx_code new_code;
4614 if (GET_CODE (op0) == COMPARE)
4615 op1 = XEXP (op0, 1), op0 = XEXP (op0, 0);
4617 /* Simplify our comparison, if possible. */
4618 new_code = simplify_comparison (code, &op0, &op1);
4620 /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X
4621 if only the low-order bit is possibly nonzero in X (such as when
4622 X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to
4623 (xor X 1) or (minus 1 X); we use the former. Finally, if X is
4624 known to be either 0 or -1, NE becomes a NEG and EQ becomes
4625 (plus X 1).
4627 Remove any ZERO_EXTRACT we made when thinking this was a
4628 comparison. It may now be simpler to use, e.g., an AND. If a
4629 ZERO_EXTRACT is indeed appropriate, it will be placed back by
4630 the call to make_compound_operation in the SET case. */
4632 if (STORE_FLAG_VALUE == 1
4633 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4634 && op1 == const0_rtx
4635 && mode == GET_MODE (op0)
4636 && nonzero_bits (op0, mode) == 1)
4637 return gen_lowpart (mode,
4638 expand_compound_operation (op0));
4640 else if (STORE_FLAG_VALUE == 1
4641 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4642 && op1 == const0_rtx
4643 && mode == GET_MODE (op0)
4644 && (num_sign_bit_copies (op0, mode)
4645 == GET_MODE_BITSIZE (mode)))
4647 op0 = expand_compound_operation (op0);
4648 return simplify_gen_unary (NEG, mode,
4649 gen_lowpart (mode, op0),
4650 mode);
4653 else if (STORE_FLAG_VALUE == 1
4654 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4655 && op1 == const0_rtx
4656 && mode == GET_MODE (op0)
4657 && nonzero_bits (op0, mode) == 1)
4659 op0 = expand_compound_operation (op0);
4660 return simplify_gen_binary (XOR, mode,
4661 gen_lowpart (mode, op0),
4662 const1_rtx);
4665 else if (STORE_FLAG_VALUE == 1
4666 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4667 && op1 == const0_rtx
4668 && mode == GET_MODE (op0)
4669 && (num_sign_bit_copies (op0, mode)
4670 == GET_MODE_BITSIZE (mode)))
4672 op0 = expand_compound_operation (op0);
4673 return plus_constant (gen_lowpart (mode, op0), 1);
4676 /* If STORE_FLAG_VALUE is -1, we have cases similar to
4677 those above. */
4678 if (STORE_FLAG_VALUE == -1
4679 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4680 && op1 == const0_rtx
4681 && (num_sign_bit_copies (op0, mode)
4682 == GET_MODE_BITSIZE (mode)))
4683 return gen_lowpart (mode,
4684 expand_compound_operation (op0));
4686 else if (STORE_FLAG_VALUE == -1
4687 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4688 && op1 == const0_rtx
4689 && mode == GET_MODE (op0)
4690 && nonzero_bits (op0, mode) == 1)
4692 op0 = expand_compound_operation (op0);
4693 return simplify_gen_unary (NEG, mode,
4694 gen_lowpart (mode, op0),
4695 mode);
4698 else if (STORE_FLAG_VALUE == -1
4699 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4700 && op1 == const0_rtx
4701 && mode == GET_MODE (op0)
4702 && (num_sign_bit_copies (op0, mode)
4703 == GET_MODE_BITSIZE (mode)))
4705 op0 = expand_compound_operation (op0);
4706 return simplify_gen_unary (NOT, mode,
4707 gen_lowpart (mode, op0),
4708 mode);
4711 /* If X is 0/1, (eq X 0) is X-1. */
4712 else if (STORE_FLAG_VALUE == -1
4713 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4714 && op1 == const0_rtx
4715 && mode == GET_MODE (op0)
4716 && nonzero_bits (op0, mode) == 1)
4718 op0 = expand_compound_operation (op0);
4719 return plus_constant (gen_lowpart (mode, op0), -1);
4722 /* If STORE_FLAG_VALUE says to just test the sign bit and X has just
4723 one bit that might be nonzero, we can convert (ne x 0) to
4724 (ashift x c) where C puts the bit in the sign bit. Remove any
4725 AND with STORE_FLAG_VALUE when we are done, since we are only
4726 going to test the sign bit. */
4727 if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4728 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4729 && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode))
4730 == (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1))
4731 && op1 == const0_rtx
4732 && mode == GET_MODE (op0)
4733 && (i = exact_log2 (nonzero_bits (op0, mode))) >= 0)
4735 x = simplify_shift_const (NULL_RTX, ASHIFT, mode,
4736 expand_compound_operation (op0),
4737 GET_MODE_BITSIZE (mode) - 1 - i);
4738 if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx)
4739 return XEXP (x, 0);
4740 else
4741 return x;
4744 /* If the code changed, return a whole new comparison. */
4745 if (new_code != code)
4746 return gen_rtx_fmt_ee (new_code, mode, op0, op1);
4748 /* Otherwise, keep this operation, but maybe change its operands.
4749 This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */
4750 SUBST (XEXP (x, 0), op0);
4751 SUBST (XEXP (x, 1), op1);
4753 break;
4755 case IF_THEN_ELSE:
4756 return simplify_if_then_else (x);
4758 case ZERO_EXTRACT:
4759 case SIGN_EXTRACT:
4760 case ZERO_EXTEND:
4761 case SIGN_EXTEND:
4762 /* If we are processing SET_DEST, we are done. */
4763 if (in_dest)
4764 return x;
4766 return expand_compound_operation (x);
4768 case SET:
4769 return simplify_set (x);
4771 case AND:
4772 case IOR:
4773 return simplify_logical (x);
4775 case ASHIFT:
4776 case LSHIFTRT:
4777 case ASHIFTRT:
4778 case ROTATE:
4779 case ROTATERT:
4780 /* If this is a shift by a constant amount, simplify it. */
4781 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
4782 return simplify_shift_const (x, code, mode, XEXP (x, 0),
4783 INTVAL (XEXP (x, 1)));
4785 else if (SHIFT_COUNT_TRUNCATED && !REG_P (XEXP (x, 1)))
4786 SUBST (XEXP (x, 1),
4787 force_to_mode (XEXP (x, 1), GET_MODE (XEXP (x, 1)),
4788 ((HOST_WIDE_INT) 1
4789 << exact_log2 (GET_MODE_BITSIZE (GET_MODE (x))))
4790 - 1,
4791 0));
4792 break;
4794 default:
4795 break;
4798 return x;
4801 /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */
4803 static rtx
4804 simplify_if_then_else (rtx x)
4806 enum machine_mode mode = GET_MODE (x);
4807 rtx cond = XEXP (x, 0);
4808 rtx true_rtx = XEXP (x, 1);
4809 rtx false_rtx = XEXP (x, 2);
4810 enum rtx_code true_code = GET_CODE (cond);
4811 int comparison_p = COMPARISON_P (cond);
4812 rtx temp;
4813 int i;
4814 enum rtx_code false_code;
4815 rtx reversed;
4817 /* Simplify storing of the truth value. */
4818 if (comparison_p && true_rtx == const_true_rtx && false_rtx == const0_rtx)
4819 return simplify_gen_relational (true_code, mode, VOIDmode,
4820 XEXP (cond, 0), XEXP (cond, 1));
4822 /* Also when the truth value has to be reversed. */
4823 if (comparison_p
4824 && true_rtx == const0_rtx && false_rtx == const_true_rtx
4825 && (reversed = reversed_comparison (cond, mode)))
4826 return reversed;
4828 /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used
4829 in it is being compared against certain values. Get the true and false
4830 comparisons and see if that says anything about the value of each arm. */
4832 if (comparison_p
4833 && ((false_code = reversed_comparison_code (cond, NULL))
4834 != UNKNOWN)
4835 && REG_P (XEXP (cond, 0)))
4837 HOST_WIDE_INT nzb;
4838 rtx from = XEXP (cond, 0);
4839 rtx true_val = XEXP (cond, 1);
4840 rtx false_val = true_val;
4841 int swapped = 0;
4843 /* If FALSE_CODE is EQ, swap the codes and arms. */
4845 if (false_code == EQ)
4847 swapped = 1, true_code = EQ, false_code = NE;
4848 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4851 /* If we are comparing against zero and the expression being tested has
4852 only a single bit that might be nonzero, that is its value when it is
4853 not equal to zero. Similarly if it is known to be -1 or 0. */
4855 if (true_code == EQ && true_val == const0_rtx
4856 && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0)
4857 false_code = EQ, false_val = GEN_INT (nzb);
4858 else if (true_code == EQ && true_val == const0_rtx
4859 && (num_sign_bit_copies (from, GET_MODE (from))
4860 == GET_MODE_BITSIZE (GET_MODE (from))))
4861 false_code = EQ, false_val = constm1_rtx;
4863 /* Now simplify an arm if we know the value of the register in the
4864 branch and it is used in the arm. Be careful due to the potential
4865 of locally-shared RTL. */
4867 if (reg_mentioned_p (from, true_rtx))
4868 true_rtx = subst (known_cond (copy_rtx (true_rtx), true_code,
4869 from, true_val),
4870 pc_rtx, pc_rtx, 0, 0);
4871 if (reg_mentioned_p (from, false_rtx))
4872 false_rtx = subst (known_cond (copy_rtx (false_rtx), false_code,
4873 from, false_val),
4874 pc_rtx, pc_rtx, 0, 0);
4876 SUBST (XEXP (x, 1), swapped ? false_rtx : true_rtx);
4877 SUBST (XEXP (x, 2), swapped ? true_rtx : false_rtx);
4879 true_rtx = XEXP (x, 1);
4880 false_rtx = XEXP (x, 2);
4881 true_code = GET_CODE (cond);
4884 /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be
4885 reversed, do so to avoid needing two sets of patterns for
4886 subtract-and-branch insns. Similarly if we have a constant in the true
4887 arm, the false arm is the same as the first operand of the comparison, or
4888 the false arm is more complicated than the true arm. */
4890 if (comparison_p
4891 && reversed_comparison_code (cond, NULL) != UNKNOWN
4892 && (true_rtx == pc_rtx
4893 || (CONSTANT_P (true_rtx)
4894 && GET_CODE (false_rtx) != CONST_INT && false_rtx != pc_rtx)
4895 || true_rtx == const0_rtx
4896 || (OBJECT_P (true_rtx) && !OBJECT_P (false_rtx))
4897 || (GET_CODE (true_rtx) == SUBREG && OBJECT_P (SUBREG_REG (true_rtx))
4898 && !OBJECT_P (false_rtx))
4899 || reg_mentioned_p (true_rtx, false_rtx)
4900 || rtx_equal_p (false_rtx, XEXP (cond, 0))))
4902 true_code = reversed_comparison_code (cond, NULL);
4903 SUBST (XEXP (x, 0), reversed_comparison (cond, GET_MODE (cond)));
4904 SUBST (XEXP (x, 1), false_rtx);
4905 SUBST (XEXP (x, 2), true_rtx);
4907 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4908 cond = XEXP (x, 0);
4910 /* It is possible that the conditional has been simplified out. */
4911 true_code = GET_CODE (cond);
4912 comparison_p = COMPARISON_P (cond);
4915 /* If the two arms are identical, we don't need the comparison. */
4917 if (rtx_equal_p (true_rtx, false_rtx) && ! side_effects_p (cond))
4918 return true_rtx;
4920 /* Convert a == b ? b : a to "a". */
4921 if (true_code == EQ && ! side_effects_p (cond)
4922 && !HONOR_NANS (mode)
4923 && rtx_equal_p (XEXP (cond, 0), false_rtx)
4924 && rtx_equal_p (XEXP (cond, 1), true_rtx))
4925 return false_rtx;
4926 else if (true_code == NE && ! side_effects_p (cond)
4927 && !HONOR_NANS (mode)
4928 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4929 && rtx_equal_p (XEXP (cond, 1), false_rtx))
4930 return true_rtx;
4932 /* Look for cases where we have (abs x) or (neg (abs X)). */
4934 if (GET_MODE_CLASS (mode) == MODE_INT
4935 && GET_CODE (false_rtx) == NEG
4936 && rtx_equal_p (true_rtx, XEXP (false_rtx, 0))
4937 && comparison_p
4938 && rtx_equal_p (true_rtx, XEXP (cond, 0))
4939 && ! side_effects_p (true_rtx))
4940 switch (true_code)
4942 case GT:
4943 case GE:
4944 return simplify_gen_unary (ABS, mode, true_rtx, mode);
4945 case LT:
4946 case LE:
4947 return
4948 simplify_gen_unary (NEG, mode,
4949 simplify_gen_unary (ABS, mode, true_rtx, mode),
4950 mode);
4951 default:
4952 break;
4955 /* Look for MIN or MAX. */
4957 if ((! FLOAT_MODE_P (mode) || flag_unsafe_math_optimizations)
4958 && comparison_p
4959 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4960 && rtx_equal_p (XEXP (cond, 1), false_rtx)
4961 && ! side_effects_p (cond))
4962 switch (true_code)
4964 case GE:
4965 case GT:
4966 return simplify_gen_binary (SMAX, mode, true_rtx, false_rtx);
4967 case LE:
4968 case LT:
4969 return simplify_gen_binary (SMIN, mode, true_rtx, false_rtx);
4970 case GEU:
4971 case GTU:
4972 return simplify_gen_binary (UMAX, mode, true_rtx, false_rtx);
4973 case LEU:
4974 case LTU:
4975 return simplify_gen_binary (UMIN, mode, true_rtx, false_rtx);
4976 default:
4977 break;
4980 /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its
4981 second operand is zero, this can be done as (OP Z (mult COND C2)) where
4982 C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or
4983 SIGN_EXTEND as long as Z is already extended (so we don't destroy it).
4984 We can do this kind of thing in some cases when STORE_FLAG_VALUE is
4985 neither 1 or -1, but it isn't worth checking for. */
4987 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
4988 && comparison_p
4989 && GET_MODE_CLASS (mode) == MODE_INT
4990 && ! side_effects_p (x))
4992 rtx t = make_compound_operation (true_rtx, SET);
4993 rtx f = make_compound_operation (false_rtx, SET);
4994 rtx cond_op0 = XEXP (cond, 0);
4995 rtx cond_op1 = XEXP (cond, 1);
4996 enum rtx_code op = UNKNOWN, extend_op = UNKNOWN;
4997 enum machine_mode m = mode;
4998 rtx z = 0, c1 = NULL_RTX;
5000 if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS
5001 || GET_CODE (t) == IOR || GET_CODE (t) == XOR
5002 || GET_CODE (t) == ASHIFT
5003 || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT)
5004 && rtx_equal_p (XEXP (t, 0), f))
5005 c1 = XEXP (t, 1), op = GET_CODE (t), z = f;
5007 /* If an identity-zero op is commutative, check whether there
5008 would be a match if we swapped the operands. */
5009 else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR
5010 || GET_CODE (t) == XOR)
5011 && rtx_equal_p (XEXP (t, 1), f))
5012 c1 = XEXP (t, 0), op = GET_CODE (t), z = f;
5013 else if (GET_CODE (t) == SIGN_EXTEND
5014 && (GET_CODE (XEXP (t, 0)) == PLUS
5015 || GET_CODE (XEXP (t, 0)) == MINUS
5016 || GET_CODE (XEXP (t, 0)) == IOR
5017 || GET_CODE (XEXP (t, 0)) == XOR
5018 || GET_CODE (XEXP (t, 0)) == ASHIFT
5019 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5020 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5021 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5022 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5023 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5024 && (num_sign_bit_copies (f, GET_MODE (f))
5025 > (unsigned int)
5026 (GET_MODE_BITSIZE (mode)
5027 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 0))))))
5029 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5030 extend_op = SIGN_EXTEND;
5031 m = GET_MODE (XEXP (t, 0));
5033 else if (GET_CODE (t) == SIGN_EXTEND
5034 && (GET_CODE (XEXP (t, 0)) == PLUS
5035 || GET_CODE (XEXP (t, 0)) == IOR
5036 || GET_CODE (XEXP (t, 0)) == XOR)
5037 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5038 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5039 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5040 && (num_sign_bit_copies (f, GET_MODE (f))
5041 > (unsigned int)
5042 (GET_MODE_BITSIZE (mode)
5043 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 1))))))
5045 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5046 extend_op = SIGN_EXTEND;
5047 m = GET_MODE (XEXP (t, 0));
5049 else if (GET_CODE (t) == ZERO_EXTEND
5050 && (GET_CODE (XEXP (t, 0)) == PLUS
5051 || GET_CODE (XEXP (t, 0)) == MINUS
5052 || GET_CODE (XEXP (t, 0)) == IOR
5053 || GET_CODE (XEXP (t, 0)) == XOR
5054 || GET_CODE (XEXP (t, 0)) == ASHIFT
5055 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5056 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5057 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5058 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5059 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5060 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5061 && ((nonzero_bits (f, GET_MODE (f))
5062 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 0))))
5063 == 0))
5065 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5066 extend_op = ZERO_EXTEND;
5067 m = GET_MODE (XEXP (t, 0));
5069 else if (GET_CODE (t) == ZERO_EXTEND
5070 && (GET_CODE (XEXP (t, 0)) == PLUS
5071 || GET_CODE (XEXP (t, 0)) == IOR
5072 || GET_CODE (XEXP (t, 0)) == XOR)
5073 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5074 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5075 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5076 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5077 && ((nonzero_bits (f, GET_MODE (f))
5078 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 1))))
5079 == 0))
5081 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5082 extend_op = ZERO_EXTEND;
5083 m = GET_MODE (XEXP (t, 0));
5086 if (z)
5088 temp = subst (simplify_gen_relational (true_code, m, VOIDmode,
5089 cond_op0, cond_op1),
5090 pc_rtx, pc_rtx, 0, 0);
5091 temp = simplify_gen_binary (MULT, m, temp,
5092 simplify_gen_binary (MULT, m, c1,
5093 const_true_rtx));
5094 temp = subst (temp, pc_rtx, pc_rtx, 0, 0);
5095 temp = simplify_gen_binary (op, m, gen_lowpart (m, z), temp);
5097 if (extend_op != UNKNOWN)
5098 temp = simplify_gen_unary (extend_op, mode, temp, m);
5100 return temp;
5104 /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or
5105 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the
5106 negation of a single bit, we can convert this operation to a shift. We
5107 can actually do this more generally, but it doesn't seem worth it. */
5109 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5110 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5111 && ((1 == nonzero_bits (XEXP (cond, 0), mode)
5112 && (i = exact_log2 (INTVAL (true_rtx))) >= 0)
5113 || ((num_sign_bit_copies (XEXP (cond, 0), mode)
5114 == GET_MODE_BITSIZE (mode))
5115 && (i = exact_log2 (-INTVAL (true_rtx))) >= 0)))
5116 return
5117 simplify_shift_const (NULL_RTX, ASHIFT, mode,
5118 gen_lowpart (mode, XEXP (cond, 0)), i);
5120 /* (IF_THEN_ELSE (NE REG 0) (0) (8)) is REG for nonzero_bits (REG) == 8. */
5121 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5122 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5123 && GET_MODE (XEXP (cond, 0)) == mode
5124 && (INTVAL (true_rtx) & GET_MODE_MASK (mode))
5125 == nonzero_bits (XEXP (cond, 0), mode)
5126 && (i = exact_log2 (INTVAL (true_rtx) & GET_MODE_MASK (mode))) >= 0)
5127 return XEXP (cond, 0);
5129 return x;
5132 /* Simplify X, a SET expression. Return the new expression. */
5134 static rtx
5135 simplify_set (rtx x)
5137 rtx src = SET_SRC (x);
5138 rtx dest = SET_DEST (x);
5139 enum machine_mode mode
5140 = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest);
5141 rtx other_insn;
5142 rtx *cc_use;
5144 /* (set (pc) (return)) gets written as (return). */
5145 if (GET_CODE (dest) == PC && GET_CODE (src) == RETURN)
5146 return src;
5148 /* Now that we know for sure which bits of SRC we are using, see if we can
5149 simplify the expression for the object knowing that we only need the
5150 low-order bits. */
5152 if (GET_MODE_CLASS (mode) == MODE_INT
5153 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
5155 src = force_to_mode (src, mode, ~(HOST_WIDE_INT) 0, 0);
5156 SUBST (SET_SRC (x), src);
5159 /* If we are setting CC0 or if the source is a COMPARE, look for the use of
5160 the comparison result and try to simplify it unless we already have used
5161 undobuf.other_insn. */
5162 if ((GET_MODE_CLASS (mode) == MODE_CC
5163 || GET_CODE (src) == COMPARE
5164 || CC0_P (dest))
5165 && (cc_use = find_single_use (dest, subst_insn, &other_insn)) != 0
5166 && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn)
5167 && COMPARISON_P (*cc_use)
5168 && rtx_equal_p (XEXP (*cc_use, 0), dest))
5170 enum rtx_code old_code = GET_CODE (*cc_use);
5171 enum rtx_code new_code;
5172 rtx op0, op1, tmp;
5173 int other_changed = 0;
5174 enum machine_mode compare_mode = GET_MODE (dest);
5176 if (GET_CODE (src) == COMPARE)
5177 op0 = XEXP (src, 0), op1 = XEXP (src, 1);
5178 else
5179 op0 = src, op1 = CONST0_RTX (GET_MODE (src));
5181 tmp = simplify_relational_operation (old_code, compare_mode, VOIDmode,
5182 op0, op1);
5183 if (!tmp)
5184 new_code = old_code;
5185 else if (!CONSTANT_P (tmp))
5187 new_code = GET_CODE (tmp);
5188 op0 = XEXP (tmp, 0);
5189 op1 = XEXP (tmp, 1);
5191 else
5193 rtx pat = PATTERN (other_insn);
5194 undobuf.other_insn = other_insn;
5195 SUBST (*cc_use, tmp);
5197 /* Attempt to simplify CC user. */
5198 if (GET_CODE (pat) == SET)
5200 rtx new = simplify_rtx (SET_SRC (pat));
5201 if (new != NULL_RTX)
5202 SUBST (SET_SRC (pat), new);
5205 /* Convert X into a no-op move. */
5206 SUBST (SET_DEST (x), pc_rtx);
5207 SUBST (SET_SRC (x), pc_rtx);
5208 return x;
5211 /* Simplify our comparison, if possible. */
5212 new_code = simplify_comparison (new_code, &op0, &op1);
5214 #ifdef SELECT_CC_MODE
5215 /* If this machine has CC modes other than CCmode, check to see if we
5216 need to use a different CC mode here. */
5217 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
5218 compare_mode = GET_MODE (op0);
5219 else
5220 compare_mode = SELECT_CC_MODE (new_code, op0, op1);
5222 #ifndef HAVE_cc0
5223 /* If the mode changed, we have to change SET_DEST, the mode in the
5224 compare, and the mode in the place SET_DEST is used. If SET_DEST is
5225 a hard register, just build new versions with the proper mode. If it
5226 is a pseudo, we lose unless it is only time we set the pseudo, in
5227 which case we can safely change its mode. */
5228 if (compare_mode != GET_MODE (dest))
5230 if (can_change_dest_mode (dest, 0, compare_mode))
5232 unsigned int regno = REGNO (dest);
5233 rtx new_dest;
5235 if (regno < FIRST_PSEUDO_REGISTER)
5236 new_dest = gen_rtx_REG (compare_mode, regno);
5237 else
5239 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
5240 new_dest = regno_reg_rtx[regno];
5243 SUBST (SET_DEST (x), new_dest);
5244 SUBST (XEXP (*cc_use, 0), new_dest);
5245 other_changed = 1;
5247 dest = new_dest;
5250 #endif /* cc0 */
5251 #endif /* SELECT_CC_MODE */
5253 /* If the code changed, we have to build a new comparison in
5254 undobuf.other_insn. */
5255 if (new_code != old_code)
5257 int other_changed_previously = other_changed;
5258 unsigned HOST_WIDE_INT mask;
5260 SUBST (*cc_use, gen_rtx_fmt_ee (new_code, GET_MODE (*cc_use),
5261 dest, const0_rtx));
5262 other_changed = 1;
5264 /* If the only change we made was to change an EQ into an NE or
5265 vice versa, OP0 has only one bit that might be nonzero, and OP1
5266 is zero, check if changing the user of the condition code will
5267 produce a valid insn. If it won't, we can keep the original code
5268 in that insn by surrounding our operation with an XOR. */
5270 if (((old_code == NE && new_code == EQ)
5271 || (old_code == EQ && new_code == NE))
5272 && ! other_changed_previously && op1 == const0_rtx
5273 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
5274 && exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) >= 0)
5276 rtx pat = PATTERN (other_insn), note = 0;
5278 if ((recog_for_combine (&pat, other_insn, &note) < 0
5279 && ! check_asm_operands (pat)))
5281 PUT_CODE (*cc_use, old_code);
5282 other_changed = 0;
5284 op0 = simplify_gen_binary (XOR, GET_MODE (op0),
5285 op0, GEN_INT (mask));
5290 if (other_changed)
5291 undobuf.other_insn = other_insn;
5293 #ifdef HAVE_cc0
5294 /* If we are now comparing against zero, change our source if
5295 needed. If we do not use cc0, we always have a COMPARE. */
5296 if (op1 == const0_rtx && dest == cc0_rtx)
5298 SUBST (SET_SRC (x), op0);
5299 src = op0;
5301 else
5302 #endif
5304 /* Otherwise, if we didn't previously have a COMPARE in the
5305 correct mode, we need one. */
5306 if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode)
5308 SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1));
5309 src = SET_SRC (x);
5311 else if (GET_MODE (op0) == compare_mode && op1 == const0_rtx)
5313 SUBST(SET_SRC (x), op0);
5314 src = SET_SRC (x);
5316 else
5318 /* Otherwise, update the COMPARE if needed. */
5319 SUBST (XEXP (src, 0), op0);
5320 SUBST (XEXP (src, 1), op1);
5323 else
5325 /* Get SET_SRC in a form where we have placed back any
5326 compound expressions. Then do the checks below. */
5327 src = make_compound_operation (src, SET);
5328 SUBST (SET_SRC (x), src);
5331 /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation,
5332 and X being a REG or (subreg (reg)), we may be able to convert this to
5333 (set (subreg:m2 x) (op)).
5335 We can always do this if M1 is narrower than M2 because that means that
5336 we only care about the low bits of the result.
5338 However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot
5339 perform a narrower operation than requested since the high-order bits will
5340 be undefined. On machine where it is defined, this transformation is safe
5341 as long as M1 and M2 have the same number of words. */
5343 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5344 && !OBJECT_P (SUBREG_REG (src))
5345 && (((GET_MODE_SIZE (GET_MODE (src)) + (UNITS_PER_WORD - 1))
5346 / UNITS_PER_WORD)
5347 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5348 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))
5349 #ifndef WORD_REGISTER_OPERATIONS
5350 && (GET_MODE_SIZE (GET_MODE (src))
5351 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5352 #endif
5353 #ifdef CANNOT_CHANGE_MODE_CLASS
5354 && ! (REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER
5355 && REG_CANNOT_CHANGE_MODE_P (REGNO (dest),
5356 GET_MODE (SUBREG_REG (src)),
5357 GET_MODE (src)))
5358 #endif
5359 && (REG_P (dest)
5360 || (GET_CODE (dest) == SUBREG
5361 && REG_P (SUBREG_REG (dest)))))
5363 SUBST (SET_DEST (x),
5364 gen_lowpart (GET_MODE (SUBREG_REG (src)),
5365 dest));
5366 SUBST (SET_SRC (x), SUBREG_REG (src));
5368 src = SET_SRC (x), dest = SET_DEST (x);
5371 #ifdef HAVE_cc0
5372 /* If we have (set (cc0) (subreg ...)), we try to remove the subreg
5373 in SRC. */
5374 if (dest == cc0_rtx
5375 && GET_CODE (src) == SUBREG
5376 && subreg_lowpart_p (src)
5377 && (GET_MODE_BITSIZE (GET_MODE (src))
5378 < GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (src)))))
5380 rtx inner = SUBREG_REG (src);
5381 enum machine_mode inner_mode = GET_MODE (inner);
5383 /* Here we make sure that we don't have a sign bit on. */
5384 if (GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_WIDE_INT
5385 && (nonzero_bits (inner, inner_mode)
5386 < ((unsigned HOST_WIDE_INT) 1
5387 << (GET_MODE_BITSIZE (GET_MODE (src)) - 1))))
5389 SUBST (SET_SRC (x), inner);
5390 src = SET_SRC (x);
5393 #endif
5395 #ifdef LOAD_EXTEND_OP
5396 /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this
5397 would require a paradoxical subreg. Replace the subreg with a
5398 zero_extend to avoid the reload that would otherwise be required. */
5400 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5401 && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))) != UNKNOWN
5402 && SUBREG_BYTE (src) == 0
5403 && (GET_MODE_SIZE (GET_MODE (src))
5404 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5405 && MEM_P (SUBREG_REG (src)))
5407 SUBST (SET_SRC (x),
5408 gen_rtx_fmt_e (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))),
5409 GET_MODE (src), SUBREG_REG (src)));
5411 src = SET_SRC (x);
5413 #endif
5415 /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we
5416 are comparing an item known to be 0 or -1 against 0, use a logical
5417 operation instead. Check for one of the arms being an IOR of the other
5418 arm with some value. We compute three terms to be IOR'ed together. In
5419 practice, at most two will be nonzero. Then we do the IOR's. */
5421 if (GET_CODE (dest) != PC
5422 && GET_CODE (src) == IF_THEN_ELSE
5423 && GET_MODE_CLASS (GET_MODE (src)) == MODE_INT
5424 && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE)
5425 && XEXP (XEXP (src, 0), 1) == const0_rtx
5426 && GET_MODE (src) == GET_MODE (XEXP (XEXP (src, 0), 0))
5427 #ifdef HAVE_conditional_move
5428 && ! can_conditionally_move_p (GET_MODE (src))
5429 #endif
5430 && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0),
5431 GET_MODE (XEXP (XEXP (src, 0), 0)))
5432 == GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (src, 0), 0))))
5433 && ! side_effects_p (src))
5435 rtx true_rtx = (GET_CODE (XEXP (src, 0)) == NE
5436 ? XEXP (src, 1) : XEXP (src, 2));
5437 rtx false_rtx = (GET_CODE (XEXP (src, 0)) == NE
5438 ? XEXP (src, 2) : XEXP (src, 1));
5439 rtx term1 = const0_rtx, term2, term3;
5441 if (GET_CODE (true_rtx) == IOR
5442 && rtx_equal_p (XEXP (true_rtx, 0), false_rtx))
5443 term1 = false_rtx, true_rtx = XEXP (true_rtx, 1), false_rtx = const0_rtx;
5444 else if (GET_CODE (true_rtx) == IOR
5445 && rtx_equal_p (XEXP (true_rtx, 1), false_rtx))
5446 term1 = false_rtx, true_rtx = XEXP (true_rtx, 0), false_rtx = const0_rtx;
5447 else if (GET_CODE (false_rtx) == IOR
5448 && rtx_equal_p (XEXP (false_rtx, 0), true_rtx))
5449 term1 = true_rtx, false_rtx = XEXP (false_rtx, 1), true_rtx = const0_rtx;
5450 else if (GET_CODE (false_rtx) == IOR
5451 && rtx_equal_p (XEXP (false_rtx, 1), true_rtx))
5452 term1 = true_rtx, false_rtx = XEXP (false_rtx, 0), true_rtx = const0_rtx;
5454 term2 = simplify_gen_binary (AND, GET_MODE (src),
5455 XEXP (XEXP (src, 0), 0), true_rtx);
5456 term3 = simplify_gen_binary (AND, GET_MODE (src),
5457 simplify_gen_unary (NOT, GET_MODE (src),
5458 XEXP (XEXP (src, 0), 0),
5459 GET_MODE (src)),
5460 false_rtx);
5462 SUBST (SET_SRC (x),
5463 simplify_gen_binary (IOR, GET_MODE (src),
5464 simplify_gen_binary (IOR, GET_MODE (src),
5465 term1, term2),
5466 term3));
5468 src = SET_SRC (x);
5471 /* If either SRC or DEST is a CLOBBER of (const_int 0), make this
5472 whole thing fail. */
5473 if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx)
5474 return src;
5475 else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx)
5476 return dest;
5477 else
5478 /* Convert this into a field assignment operation, if possible. */
5479 return make_field_assignment (x);
5482 /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified
5483 result. */
5485 static rtx
5486 simplify_logical (rtx x)
5488 enum machine_mode mode = GET_MODE (x);
5489 rtx op0 = XEXP (x, 0);
5490 rtx op1 = XEXP (x, 1);
5492 switch (GET_CODE (x))
5494 case AND:
5495 /* We can call simplify_and_const_int only if we don't lose
5496 any (sign) bits when converting INTVAL (op1) to
5497 "unsigned HOST_WIDE_INT". */
5498 if (GET_CODE (op1) == CONST_INT
5499 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5500 || INTVAL (op1) > 0))
5502 x = simplify_and_const_int (x, mode, op0, INTVAL (op1));
5503 if (GET_CODE (x) != AND)
5504 return x;
5506 op0 = XEXP (x, 0);
5507 op1 = XEXP (x, 1);
5510 /* If we have any of (and (ior A B) C) or (and (xor A B) C),
5511 apply the distributive law and then the inverse distributive
5512 law to see if things simplify. */
5513 if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR)
5515 rtx result = distribute_and_simplify_rtx (x, 0);
5516 if (result)
5517 return result;
5519 if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR)
5521 rtx result = distribute_and_simplify_rtx (x, 1);
5522 if (result)
5523 return result;
5525 break;
5527 case IOR:
5528 /* If we have (ior (and A B) C), apply the distributive law and then
5529 the inverse distributive law to see if things simplify. */
5531 if (GET_CODE (op0) == AND)
5533 rtx result = distribute_and_simplify_rtx (x, 0);
5534 if (result)
5535 return result;
5538 if (GET_CODE (op1) == AND)
5540 rtx result = distribute_and_simplify_rtx (x, 1);
5541 if (result)
5542 return result;
5544 break;
5546 default:
5547 gcc_unreachable ();
5550 return x;
5553 /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound
5554 operations" because they can be replaced with two more basic operations.
5555 ZERO_EXTEND is also considered "compound" because it can be replaced with
5556 an AND operation, which is simpler, though only one operation.
5558 The function expand_compound_operation is called with an rtx expression
5559 and will convert it to the appropriate shifts and AND operations,
5560 simplifying at each stage.
5562 The function make_compound_operation is called to convert an expression
5563 consisting of shifts and ANDs into the equivalent compound expression.
5564 It is the inverse of this function, loosely speaking. */
5566 static rtx
5567 expand_compound_operation (rtx x)
5569 unsigned HOST_WIDE_INT pos = 0, len;
5570 int unsignedp = 0;
5571 unsigned int modewidth;
5572 rtx tem;
5574 switch (GET_CODE (x))
5576 case ZERO_EXTEND:
5577 unsignedp = 1;
5578 case SIGN_EXTEND:
5579 /* We can't necessarily use a const_int for a multiword mode;
5580 it depends on implicitly extending the value.
5581 Since we don't know the right way to extend it,
5582 we can't tell whether the implicit way is right.
5584 Even for a mode that is no wider than a const_int,
5585 we can't win, because we need to sign extend one of its bits through
5586 the rest of it, and we don't know which bit. */
5587 if (GET_CODE (XEXP (x, 0)) == CONST_INT)
5588 return x;
5590 /* Return if (subreg:MODE FROM 0) is not a safe replacement for
5591 (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM
5592 because (SUBREG (MEM...)) is guaranteed to cause the MEM to be
5593 reloaded. If not for that, MEM's would very rarely be safe.
5595 Reject MODEs bigger than a word, because we might not be able
5596 to reference a two-register group starting with an arbitrary register
5597 (and currently gen_lowpart might crash for a SUBREG). */
5599 if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) > UNITS_PER_WORD)
5600 return x;
5602 /* Reject MODEs that aren't scalar integers because turning vector
5603 or complex modes into shifts causes problems. */
5605 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5606 return x;
5608 len = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)));
5609 /* If the inner object has VOIDmode (the only way this can happen
5610 is if it is an ASM_OPERANDS), we can't do anything since we don't
5611 know how much masking to do. */
5612 if (len == 0)
5613 return x;
5615 break;
5617 case ZERO_EXTRACT:
5618 unsignedp = 1;
5620 /* ... fall through ... */
5622 case SIGN_EXTRACT:
5623 /* If the operand is a CLOBBER, just return it. */
5624 if (GET_CODE (XEXP (x, 0)) == CLOBBER)
5625 return XEXP (x, 0);
5627 if (GET_CODE (XEXP (x, 1)) != CONST_INT
5628 || GET_CODE (XEXP (x, 2)) != CONST_INT
5629 || GET_MODE (XEXP (x, 0)) == VOIDmode)
5630 return x;
5632 /* Reject MODEs that aren't scalar integers because turning vector
5633 or complex modes into shifts causes problems. */
5635 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5636 return x;
5638 len = INTVAL (XEXP (x, 1));
5639 pos = INTVAL (XEXP (x, 2));
5641 /* This should stay within the object being extracted, fail otherwise. */
5642 if (len + pos > GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))
5643 return x;
5645 if (BITS_BIG_ENDIAN)
5646 pos = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - len - pos;
5648 break;
5650 default:
5651 return x;
5653 /* Convert sign extension to zero extension, if we know that the high
5654 bit is not set, as this is easier to optimize. It will be converted
5655 back to cheaper alternative in make_extraction. */
5656 if (GET_CODE (x) == SIGN_EXTEND
5657 && (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5658 && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
5659 & ~(((unsigned HOST_WIDE_INT)
5660 GET_MODE_MASK (GET_MODE (XEXP (x, 0))))
5661 >> 1))
5662 == 0)))
5664 rtx temp = gen_rtx_ZERO_EXTEND (GET_MODE (x), XEXP (x, 0));
5665 rtx temp2 = expand_compound_operation (temp);
5667 /* Make sure this is a profitable operation. */
5668 if (rtx_cost (x, SET) > rtx_cost (temp2, SET))
5669 return temp2;
5670 else if (rtx_cost (x, SET) > rtx_cost (temp, SET))
5671 return temp;
5672 else
5673 return x;
5676 /* We can optimize some special cases of ZERO_EXTEND. */
5677 if (GET_CODE (x) == ZERO_EXTEND)
5679 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we
5680 know that the last value didn't have any inappropriate bits
5681 set. */
5682 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5683 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5684 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5685 && (nonzero_bits (XEXP (XEXP (x, 0), 0), GET_MODE (x))
5686 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5687 return XEXP (XEXP (x, 0), 0);
5689 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5690 if (GET_CODE (XEXP (x, 0)) == SUBREG
5691 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5692 && subreg_lowpart_p (XEXP (x, 0))
5693 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5694 && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), GET_MODE (x))
5695 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5696 return SUBREG_REG (XEXP (x, 0));
5698 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo
5699 is a comparison and STORE_FLAG_VALUE permits. This is like
5700 the first case, but it works even when GET_MODE (x) is larger
5701 than HOST_WIDE_INT. */
5702 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5703 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5704 && COMPARISON_P (XEXP (XEXP (x, 0), 0))
5705 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5706 <= HOST_BITS_PER_WIDE_INT)
5707 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5708 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5709 return XEXP (XEXP (x, 0), 0);
5711 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5712 if (GET_CODE (XEXP (x, 0)) == SUBREG
5713 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5714 && subreg_lowpart_p (XEXP (x, 0))
5715 && COMPARISON_P (SUBREG_REG (XEXP (x, 0)))
5716 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5717 <= HOST_BITS_PER_WIDE_INT)
5718 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5719 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5720 return SUBREG_REG (XEXP (x, 0));
5724 /* If we reach here, we want to return a pair of shifts. The inner
5725 shift is a left shift of BITSIZE - POS - LEN bits. The outer
5726 shift is a right shift of BITSIZE - LEN bits. It is arithmetic or
5727 logical depending on the value of UNSIGNEDP.
5729 If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be
5730 converted into an AND of a shift.
5732 We must check for the case where the left shift would have a negative
5733 count. This can happen in a case like (x >> 31) & 255 on machines
5734 that can't shift by a constant. On those machines, we would first
5735 combine the shift with the AND to produce a variable-position
5736 extraction. Then the constant of 31 would be substituted in to produce
5737 a such a position. */
5739 modewidth = GET_MODE_BITSIZE (GET_MODE (x));
5740 if (modewidth + len >= pos)
5741 tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT,
5742 GET_MODE (x),
5743 simplify_shift_const (NULL_RTX, ASHIFT,
5744 GET_MODE (x),
5745 XEXP (x, 0),
5746 modewidth - pos - len),
5747 modewidth - len);
5749 else if (unsignedp && len < HOST_BITS_PER_WIDE_INT)
5750 tem = simplify_and_const_int (NULL_RTX, GET_MODE (x),
5751 simplify_shift_const (NULL_RTX, LSHIFTRT,
5752 GET_MODE (x),
5753 XEXP (x, 0), pos),
5754 ((HOST_WIDE_INT) 1 << len) - 1);
5755 else
5756 /* Any other cases we can't handle. */
5757 return x;
5759 /* If we couldn't do this for some reason, return the original
5760 expression. */
5761 if (GET_CODE (tem) == CLOBBER)
5762 return x;
5764 return tem;
5767 /* X is a SET which contains an assignment of one object into
5768 a part of another (such as a bit-field assignment, STRICT_LOW_PART,
5769 or certain SUBREGS). If possible, convert it into a series of
5770 logical operations.
5772 We half-heartedly support variable positions, but do not at all
5773 support variable lengths. */
5775 static rtx
5776 expand_field_assignment (rtx x)
5778 rtx inner;
5779 rtx pos; /* Always counts from low bit. */
5780 int len;
5781 rtx mask, cleared, masked;
5782 enum machine_mode compute_mode;
5784 /* Loop until we find something we can't simplify. */
5785 while (1)
5787 if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
5788 && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG)
5790 inner = SUBREG_REG (XEXP (SET_DEST (x), 0));
5791 len = GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)));
5792 pos = GEN_INT (subreg_lsb (XEXP (SET_DEST (x), 0)));
5794 else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
5795 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT)
5797 inner = XEXP (SET_DEST (x), 0);
5798 len = INTVAL (XEXP (SET_DEST (x), 1));
5799 pos = XEXP (SET_DEST (x), 2);
5801 /* A constant position should stay within the width of INNER. */
5802 if (GET_CODE (pos) == CONST_INT
5803 && INTVAL (pos) + len > GET_MODE_BITSIZE (GET_MODE (inner)))
5804 break;
5806 if (BITS_BIG_ENDIAN)
5808 if (GET_CODE (pos) == CONST_INT)
5809 pos = GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) - len
5810 - INTVAL (pos));
5811 else if (GET_CODE (pos) == MINUS
5812 && GET_CODE (XEXP (pos, 1)) == CONST_INT
5813 && (INTVAL (XEXP (pos, 1))
5814 == GET_MODE_BITSIZE (GET_MODE (inner)) - len))
5815 /* If position is ADJUST - X, new position is X. */
5816 pos = XEXP (pos, 0);
5817 else
5818 pos = simplify_gen_binary (MINUS, GET_MODE (pos),
5819 GEN_INT (GET_MODE_BITSIZE (
5820 GET_MODE (inner))
5821 - len),
5822 pos);
5826 /* A SUBREG between two modes that occupy the same numbers of words
5827 can be done by moving the SUBREG to the source. */
5828 else if (GET_CODE (SET_DEST (x)) == SUBREG
5829 /* We need SUBREGs to compute nonzero_bits properly. */
5830 && nonzero_sign_valid
5831 && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
5832 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
5833 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x))))
5834 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)))
5836 x = gen_rtx_SET (VOIDmode, SUBREG_REG (SET_DEST (x)),
5837 gen_lowpart
5838 (GET_MODE (SUBREG_REG (SET_DEST (x))),
5839 SET_SRC (x)));
5840 continue;
5842 else
5843 break;
5845 while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
5846 inner = SUBREG_REG (inner);
5848 compute_mode = GET_MODE (inner);
5850 /* Don't attempt bitwise arithmetic on non scalar integer modes. */
5851 if (! SCALAR_INT_MODE_P (compute_mode))
5853 enum machine_mode imode;
5855 /* Don't do anything for vector or complex integral types. */
5856 if (! FLOAT_MODE_P (compute_mode))
5857 break;
5859 /* Try to find an integral mode to pun with. */
5860 imode = mode_for_size (GET_MODE_BITSIZE (compute_mode), MODE_INT, 0);
5861 if (imode == BLKmode)
5862 break;
5864 compute_mode = imode;
5865 inner = gen_lowpart (imode, inner);
5868 /* Compute a mask of LEN bits, if we can do this on the host machine. */
5869 if (len >= HOST_BITS_PER_WIDE_INT)
5870 break;
5872 /* Now compute the equivalent expression. Make a copy of INNER
5873 for the SET_DEST in case it is a MEM into which we will substitute;
5874 we don't want shared RTL in that case. */
5875 mask = GEN_INT (((HOST_WIDE_INT) 1 << len) - 1);
5876 cleared = simplify_gen_binary (AND, compute_mode,
5877 simplify_gen_unary (NOT, compute_mode,
5878 simplify_gen_binary (ASHIFT,
5879 compute_mode,
5880 mask, pos),
5881 compute_mode),
5882 inner);
5883 masked = simplify_gen_binary (ASHIFT, compute_mode,
5884 simplify_gen_binary (
5885 AND, compute_mode,
5886 gen_lowpart (compute_mode, SET_SRC (x)),
5887 mask),
5888 pos);
5890 x = gen_rtx_SET (VOIDmode, copy_rtx (inner),
5891 simplify_gen_binary (IOR, compute_mode,
5892 cleared, masked));
5895 return x;
5898 /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero,
5899 it is an RTX that represents a variable starting position; otherwise,
5900 POS is the (constant) starting bit position (counted from the LSB).
5902 UNSIGNEDP is nonzero for an unsigned reference and zero for a
5903 signed reference.
5905 IN_DEST is nonzero if this is a reference in the destination of a
5906 SET. This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If nonzero,
5907 a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will
5908 be used.
5910 IN_COMPARE is nonzero if we are in a COMPARE. This means that a
5911 ZERO_EXTRACT should be built even for bits starting at bit 0.
5913 MODE is the desired mode of the result (if IN_DEST == 0).
5915 The result is an RTX for the extraction or NULL_RTX if the target
5916 can't handle it. */
5918 static rtx
5919 make_extraction (enum machine_mode mode, rtx inner, HOST_WIDE_INT pos,
5920 rtx pos_rtx, unsigned HOST_WIDE_INT len, int unsignedp,
5921 int in_dest, int in_compare)
5923 /* This mode describes the size of the storage area
5924 to fetch the overall value from. Within that, we
5925 ignore the POS lowest bits, etc. */
5926 enum machine_mode is_mode = GET_MODE (inner);
5927 enum machine_mode inner_mode;
5928 enum machine_mode wanted_inner_mode;
5929 enum machine_mode wanted_inner_reg_mode = word_mode;
5930 enum machine_mode pos_mode = word_mode;
5931 enum machine_mode extraction_mode = word_mode;
5932 enum machine_mode tmode = mode_for_size (len, MODE_INT, 1);
5933 rtx new = 0;
5934 rtx orig_pos_rtx = pos_rtx;
5935 HOST_WIDE_INT orig_pos;
5937 if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
5939 /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...),
5940 consider just the QI as the memory to extract from.
5941 The subreg adds or removes high bits; its mode is
5942 irrelevant to the meaning of this extraction,
5943 since POS and LEN count from the lsb. */
5944 if (MEM_P (SUBREG_REG (inner)))
5945 is_mode = GET_MODE (SUBREG_REG (inner));
5946 inner = SUBREG_REG (inner);
5948 else if (GET_CODE (inner) == ASHIFT
5949 && GET_CODE (XEXP (inner, 1)) == CONST_INT
5950 && pos_rtx == 0 && pos == 0
5951 && len > (unsigned HOST_WIDE_INT) INTVAL (XEXP (inner, 1)))
5953 /* We're extracting the least significant bits of an rtx
5954 (ashift X (const_int C)), where LEN > C. Extract the
5955 least significant (LEN - C) bits of X, giving an rtx
5956 whose mode is MODE, then shift it left C times. */
5957 new = make_extraction (mode, XEXP (inner, 0),
5958 0, 0, len - INTVAL (XEXP (inner, 1)),
5959 unsignedp, in_dest, in_compare);
5960 if (new != 0)
5961 return gen_rtx_ASHIFT (mode, new, XEXP (inner, 1));
5964 inner_mode = GET_MODE (inner);
5966 if (pos_rtx && GET_CODE (pos_rtx) == CONST_INT)
5967 pos = INTVAL (pos_rtx), pos_rtx = 0;
5969 /* See if this can be done without an extraction. We never can if the
5970 width of the field is not the same as that of some integer mode. For
5971 registers, we can only avoid the extraction if the position is at the
5972 low-order bit and this is either not in the destination or we have the
5973 appropriate STRICT_LOW_PART operation available.
5975 For MEM, we can avoid an extract if the field starts on an appropriate
5976 boundary and we can change the mode of the memory reference. */
5978 if (tmode != BLKmode
5979 && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0
5980 && !MEM_P (inner)
5981 && (inner_mode == tmode
5982 || !REG_P (inner)
5983 || TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (tmode),
5984 GET_MODE_BITSIZE (inner_mode))
5985 || reg_truncated_to_mode (tmode, inner))
5986 && (! in_dest
5987 || (REG_P (inner)
5988 && have_insn_for (STRICT_LOW_PART, tmode))))
5989 || (MEM_P (inner) && pos_rtx == 0
5990 && (pos
5991 % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode)
5992 : BITS_PER_UNIT)) == 0
5993 /* We can't do this if we are widening INNER_MODE (it
5994 may not be aligned, for one thing). */
5995 && GET_MODE_BITSIZE (inner_mode) >= GET_MODE_BITSIZE (tmode)
5996 && (inner_mode == tmode
5997 || (! mode_dependent_address_p (XEXP (inner, 0))
5998 && ! MEM_VOLATILE_P (inner))))))
6000 /* If INNER is a MEM, make a new MEM that encompasses just the desired
6001 field. If the original and current mode are the same, we need not
6002 adjust the offset. Otherwise, we do if bytes big endian.
6004 If INNER is not a MEM, get a piece consisting of just the field
6005 of interest (in this case POS % BITS_PER_WORD must be 0). */
6007 if (MEM_P (inner))
6009 HOST_WIDE_INT offset;
6011 /* POS counts from lsb, but make OFFSET count in memory order. */
6012 if (BYTES_BIG_ENDIAN)
6013 offset = (GET_MODE_BITSIZE (is_mode) - len - pos) / BITS_PER_UNIT;
6014 else
6015 offset = pos / BITS_PER_UNIT;
6017 new = adjust_address_nv (inner, tmode, offset);
6019 else if (REG_P (inner))
6021 if (tmode != inner_mode)
6023 /* We can't call gen_lowpart in a DEST since we
6024 always want a SUBREG (see below) and it would sometimes
6025 return a new hard register. */
6026 if (pos || in_dest)
6028 HOST_WIDE_INT final_word = pos / BITS_PER_WORD;
6030 if (WORDS_BIG_ENDIAN
6031 && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD)
6032 final_word = ((GET_MODE_SIZE (inner_mode)
6033 - GET_MODE_SIZE (tmode))
6034 / UNITS_PER_WORD) - final_word;
6036 final_word *= UNITS_PER_WORD;
6037 if (BYTES_BIG_ENDIAN &&
6038 GET_MODE_SIZE (inner_mode) > GET_MODE_SIZE (tmode))
6039 final_word += (GET_MODE_SIZE (inner_mode)
6040 - GET_MODE_SIZE (tmode)) % UNITS_PER_WORD;
6042 /* Avoid creating invalid subregs, for example when
6043 simplifying (x>>32)&255. */
6044 if (!validate_subreg (tmode, inner_mode, inner, final_word))
6045 return NULL_RTX;
6047 new = gen_rtx_SUBREG (tmode, inner, final_word);
6049 else
6050 new = gen_lowpart (tmode, inner);
6052 else
6053 new = inner;
6055 else
6056 new = force_to_mode (inner, tmode,
6057 len >= HOST_BITS_PER_WIDE_INT
6058 ? ~(unsigned HOST_WIDE_INT) 0
6059 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
6062 /* If this extraction is going into the destination of a SET,
6063 make a STRICT_LOW_PART unless we made a MEM. */
6065 if (in_dest)
6066 return (MEM_P (new) ? new
6067 : (GET_CODE (new) != SUBREG
6068 ? gen_rtx_CLOBBER (tmode, const0_rtx)
6069 : gen_rtx_STRICT_LOW_PART (VOIDmode, new)));
6071 if (mode == tmode)
6072 return new;
6074 if (GET_CODE (new) == CONST_INT)
6075 return gen_int_mode (INTVAL (new), mode);
6077 /* If we know that no extraneous bits are set, and that the high
6078 bit is not set, convert the extraction to the cheaper of
6079 sign and zero extension, that are equivalent in these cases. */
6080 if (flag_expensive_optimizations
6081 && (GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
6082 && ((nonzero_bits (new, tmode)
6083 & ~(((unsigned HOST_WIDE_INT)
6084 GET_MODE_MASK (tmode))
6085 >> 1))
6086 == 0)))
6088 rtx temp = gen_rtx_ZERO_EXTEND (mode, new);
6089 rtx temp1 = gen_rtx_SIGN_EXTEND (mode, new);
6091 /* Prefer ZERO_EXTENSION, since it gives more information to
6092 backends. */
6093 if (rtx_cost (temp, SET) <= rtx_cost (temp1, SET))
6094 return temp;
6095 return temp1;
6098 /* Otherwise, sign- or zero-extend unless we already are in the
6099 proper mode. */
6101 return (gen_rtx_fmt_e (unsignedp ? ZERO_EXTEND : SIGN_EXTEND,
6102 mode, new));
6105 /* Unless this is a COMPARE or we have a funny memory reference,
6106 don't do anything with zero-extending field extracts starting at
6107 the low-order bit since they are simple AND operations. */
6108 if (pos_rtx == 0 && pos == 0 && ! in_dest
6109 && ! in_compare && unsignedp)
6110 return 0;
6112 /* Unless INNER is not MEM, reject this if we would be spanning bytes or
6113 if the position is not a constant and the length is not 1. In all
6114 other cases, we would only be going outside our object in cases when
6115 an original shift would have been undefined. */
6116 if (MEM_P (inner)
6117 && ((pos_rtx == 0 && pos + len > GET_MODE_BITSIZE (is_mode))
6118 || (pos_rtx != 0 && len != 1)))
6119 return 0;
6121 /* Get the mode to use should INNER not be a MEM, the mode for the position,
6122 and the mode for the result. */
6123 if (in_dest && mode_for_extraction (EP_insv, -1) != MAX_MACHINE_MODE)
6125 wanted_inner_reg_mode = mode_for_extraction (EP_insv, 0);
6126 pos_mode = mode_for_extraction (EP_insv, 2);
6127 extraction_mode = mode_for_extraction (EP_insv, 3);
6130 if (! in_dest && unsignedp
6131 && mode_for_extraction (EP_extzv, -1) != MAX_MACHINE_MODE)
6133 wanted_inner_reg_mode = mode_for_extraction (EP_extzv, 1);
6134 pos_mode = mode_for_extraction (EP_extzv, 3);
6135 extraction_mode = mode_for_extraction (EP_extzv, 0);
6138 if (! in_dest && ! unsignedp
6139 && mode_for_extraction (EP_extv, -1) != MAX_MACHINE_MODE)
6141 wanted_inner_reg_mode = mode_for_extraction (EP_extv, 1);
6142 pos_mode = mode_for_extraction (EP_extv, 3);
6143 extraction_mode = mode_for_extraction (EP_extv, 0);
6146 /* Never narrow an object, since that might not be safe. */
6148 if (mode != VOIDmode
6149 && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode))
6150 extraction_mode = mode;
6152 if (pos_rtx && GET_MODE (pos_rtx) != VOIDmode
6153 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6154 pos_mode = GET_MODE (pos_rtx);
6156 /* If this is not from memory, the desired mode is the preferred mode
6157 for an extraction pattern's first input operand, or word_mode if there
6158 is none. */
6159 if (!MEM_P (inner))
6160 wanted_inner_mode = wanted_inner_reg_mode;
6161 else
6163 /* Be careful not to go beyond the extracted object and maintain the
6164 natural alignment of the memory. */
6165 wanted_inner_mode = smallest_mode_for_size (len, MODE_INT);
6166 while (pos % GET_MODE_BITSIZE (wanted_inner_mode) + len
6167 > GET_MODE_BITSIZE (wanted_inner_mode))
6169 wanted_inner_mode = GET_MODE_WIDER_MODE (wanted_inner_mode);
6170 gcc_assert (wanted_inner_mode != VOIDmode);
6173 /* If we have to change the mode of memory and cannot, the desired mode
6174 is EXTRACTION_MODE. */
6175 if (inner_mode != wanted_inner_mode
6176 && (mode_dependent_address_p (XEXP (inner, 0))
6177 || MEM_VOLATILE_P (inner)
6178 || pos_rtx))
6179 wanted_inner_mode = extraction_mode;
6182 orig_pos = pos;
6184 if (BITS_BIG_ENDIAN)
6186 /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to
6187 BITS_BIG_ENDIAN style. If position is constant, compute new
6188 position. Otherwise, build subtraction.
6189 Note that POS is relative to the mode of the original argument.
6190 If it's a MEM we need to recompute POS relative to that.
6191 However, if we're extracting from (or inserting into) a register,
6192 we want to recompute POS relative to wanted_inner_mode. */
6193 int width = (MEM_P (inner)
6194 ? GET_MODE_BITSIZE (is_mode)
6195 : GET_MODE_BITSIZE (wanted_inner_mode));
6197 if (pos_rtx == 0)
6198 pos = width - len - pos;
6199 else
6200 pos_rtx
6201 = gen_rtx_MINUS (GET_MODE (pos_rtx), GEN_INT (width - len), pos_rtx);
6202 /* POS may be less than 0 now, but we check for that below.
6203 Note that it can only be less than 0 if !MEM_P (inner). */
6206 /* If INNER has a wider mode, and this is a constant extraction, try to
6207 make it smaller and adjust the byte to point to the byte containing
6208 the value. */
6209 if (wanted_inner_mode != VOIDmode
6210 && inner_mode != wanted_inner_mode
6211 && ! pos_rtx
6212 && GET_MODE_SIZE (wanted_inner_mode) < GET_MODE_SIZE (is_mode)
6213 && MEM_P (inner)
6214 && ! mode_dependent_address_p (XEXP (inner, 0))
6215 && ! MEM_VOLATILE_P (inner))
6217 int offset = 0;
6219 /* The computations below will be correct if the machine is big
6220 endian in both bits and bytes or little endian in bits and bytes.
6221 If it is mixed, we must adjust. */
6223 /* If bytes are big endian and we had a paradoxical SUBREG, we must
6224 adjust OFFSET to compensate. */
6225 if (BYTES_BIG_ENDIAN
6226 && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode))
6227 offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode);
6229 /* We can now move to the desired byte. */
6230 offset += (pos / GET_MODE_BITSIZE (wanted_inner_mode))
6231 * GET_MODE_SIZE (wanted_inner_mode);
6232 pos %= GET_MODE_BITSIZE (wanted_inner_mode);
6234 if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN
6235 && is_mode != wanted_inner_mode)
6236 offset = (GET_MODE_SIZE (is_mode)
6237 - GET_MODE_SIZE (wanted_inner_mode) - offset);
6239 inner = adjust_address_nv (inner, wanted_inner_mode, offset);
6242 /* If INNER is not memory, we can always get it into the proper mode. If we
6243 are changing its mode, POS must be a constant and smaller than the size
6244 of the new mode. */
6245 else if (!MEM_P (inner))
6247 if (GET_MODE (inner) != wanted_inner_mode
6248 && (pos_rtx != 0
6249 || orig_pos + len > GET_MODE_BITSIZE (wanted_inner_mode)))
6250 return 0;
6252 if (orig_pos < 0)
6253 return 0;
6255 inner = force_to_mode (inner, wanted_inner_mode,
6256 pos_rtx
6257 || len + orig_pos >= HOST_BITS_PER_WIDE_INT
6258 ? ~(unsigned HOST_WIDE_INT) 0
6259 : ((((unsigned HOST_WIDE_INT) 1 << len) - 1)
6260 << orig_pos),
6264 /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we
6265 have to zero extend. Otherwise, we can just use a SUBREG. */
6266 if (pos_rtx != 0
6267 && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx)))
6269 rtx temp = gen_rtx_ZERO_EXTEND (pos_mode, pos_rtx);
6271 /* If we know that no extraneous bits are set, and that the high
6272 bit is not set, convert extraction to cheaper one - either
6273 SIGN_EXTENSION or ZERO_EXTENSION, that are equivalent in these
6274 cases. */
6275 if (flag_expensive_optimizations
6276 && (GET_MODE_BITSIZE (GET_MODE (pos_rtx)) <= HOST_BITS_PER_WIDE_INT
6277 && ((nonzero_bits (pos_rtx, GET_MODE (pos_rtx))
6278 & ~(((unsigned HOST_WIDE_INT)
6279 GET_MODE_MASK (GET_MODE (pos_rtx)))
6280 >> 1))
6281 == 0)))
6283 rtx temp1 = gen_rtx_SIGN_EXTEND (pos_mode, pos_rtx);
6285 /* Prefer ZERO_EXTENSION, since it gives more information to
6286 backends. */
6287 if (rtx_cost (temp1, SET) < rtx_cost (temp, SET))
6288 temp = temp1;
6290 pos_rtx = temp;
6292 else if (pos_rtx != 0
6293 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6294 pos_rtx = gen_lowpart (pos_mode, pos_rtx);
6296 /* Make POS_RTX unless we already have it and it is correct. If we don't
6297 have a POS_RTX but we do have an ORIG_POS_RTX, the latter must
6298 be a CONST_INT. */
6299 if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos)
6300 pos_rtx = orig_pos_rtx;
6302 else if (pos_rtx == 0)
6303 pos_rtx = GEN_INT (pos);
6305 /* Make the required operation. See if we can use existing rtx. */
6306 new = gen_rtx_fmt_eee (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT,
6307 extraction_mode, inner, GEN_INT (len), pos_rtx);
6308 if (! in_dest)
6309 new = gen_lowpart (mode, new);
6311 return new;
6314 /* See if X contains an ASHIFT of COUNT or more bits that can be commuted
6315 with any other operations in X. Return X without that shift if so. */
6317 static rtx
6318 extract_left_shift (rtx x, int count)
6320 enum rtx_code code = GET_CODE (x);
6321 enum machine_mode mode = GET_MODE (x);
6322 rtx tem;
6324 switch (code)
6326 case ASHIFT:
6327 /* This is the shift itself. If it is wide enough, we will return
6328 either the value being shifted if the shift count is equal to
6329 COUNT or a shift for the difference. */
6330 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6331 && INTVAL (XEXP (x, 1)) >= count)
6332 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0),
6333 INTVAL (XEXP (x, 1)) - count);
6334 break;
6336 case NEG: case NOT:
6337 if ((tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6338 return simplify_gen_unary (code, mode, tem, mode);
6340 break;
6342 case PLUS: case IOR: case XOR: case AND:
6343 /* If we can safely shift this constant and we find the inner shift,
6344 make a new operation. */
6345 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6346 && (INTVAL (XEXP (x, 1)) & ((((HOST_WIDE_INT) 1 << count)) - 1)) == 0
6347 && (tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6348 return simplify_gen_binary (code, mode, tem,
6349 GEN_INT (INTVAL (XEXP (x, 1)) >> count));
6351 break;
6353 default:
6354 break;
6357 return 0;
6360 /* Look at the expression rooted at X. Look for expressions
6361 equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND.
6362 Form these expressions.
6364 Return the new rtx, usually just X.
6366 Also, for machines like the VAX that don't have logical shift insns,
6367 try to convert logical to arithmetic shift operations in cases where
6368 they are equivalent. This undoes the canonicalizations to logical
6369 shifts done elsewhere.
6371 We try, as much as possible, to re-use rtl expressions to save memory.
6373 IN_CODE says what kind of expression we are processing. Normally, it is
6374 SET. In a memory address (inside a MEM, PLUS or minus, the latter two
6375 being kludges), it is MEM. When processing the arguments of a comparison
6376 or a COMPARE against zero, it is COMPARE. */
6378 static rtx
6379 make_compound_operation (rtx x, enum rtx_code in_code)
6381 enum rtx_code code = GET_CODE (x);
6382 enum machine_mode mode = GET_MODE (x);
6383 int mode_width = GET_MODE_BITSIZE (mode);
6384 rtx rhs, lhs;
6385 enum rtx_code next_code;
6386 int i;
6387 rtx new = 0;
6388 rtx tem;
6389 const char *fmt;
6391 /* Select the code to be used in recursive calls. Once we are inside an
6392 address, we stay there. If we have a comparison, set to COMPARE,
6393 but once inside, go back to our default of SET. */
6395 next_code = (code == MEM || code == PLUS || code == MINUS ? MEM
6396 : ((code == COMPARE || COMPARISON_P (x))
6397 && XEXP (x, 1) == const0_rtx) ? COMPARE
6398 : in_code == COMPARE ? SET : in_code);
6400 /* Process depending on the code of this operation. If NEW is set
6401 nonzero, it will be returned. */
6403 switch (code)
6405 case ASHIFT:
6406 /* Convert shifts by constants into multiplications if inside
6407 an address. */
6408 if (in_code == MEM && GET_CODE (XEXP (x, 1)) == CONST_INT
6409 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
6410 && INTVAL (XEXP (x, 1)) >= 0)
6412 new = make_compound_operation (XEXP (x, 0), next_code);
6413 new = gen_rtx_MULT (mode, new,
6414 GEN_INT ((HOST_WIDE_INT) 1
6415 << INTVAL (XEXP (x, 1))));
6417 break;
6419 case AND:
6420 /* If the second operand is not a constant, we can't do anything
6421 with it. */
6422 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6423 break;
6425 /* If the constant is a power of two minus one and the first operand
6426 is a logical right shift, make an extraction. */
6427 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6428 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6430 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6431 new = make_extraction (mode, new, 0, XEXP (XEXP (x, 0), 1), i, 1,
6432 0, in_code == COMPARE);
6435 /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */
6436 else if (GET_CODE (XEXP (x, 0)) == SUBREG
6437 && subreg_lowpart_p (XEXP (x, 0))
6438 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT
6439 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6441 new = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0),
6442 next_code);
6443 new = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new, 0,
6444 XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1,
6445 0, in_code == COMPARE);
6447 /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */
6448 else if ((GET_CODE (XEXP (x, 0)) == XOR
6449 || GET_CODE (XEXP (x, 0)) == IOR)
6450 && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT
6451 && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT
6452 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6454 /* Apply the distributive law, and then try to make extractions. */
6455 new = gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)), mode,
6456 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0),
6457 XEXP (x, 1)),
6458 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1),
6459 XEXP (x, 1)));
6460 new = make_compound_operation (new, in_code);
6463 /* If we are have (and (rotate X C) M) and C is larger than the number
6464 of bits in M, this is an extraction. */
6466 else if (GET_CODE (XEXP (x, 0)) == ROTATE
6467 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6468 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0
6469 && i <= INTVAL (XEXP (XEXP (x, 0), 1)))
6471 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6472 new = make_extraction (mode, new,
6473 (GET_MODE_BITSIZE (mode)
6474 - INTVAL (XEXP (XEXP (x, 0), 1))),
6475 NULL_RTX, i, 1, 0, in_code == COMPARE);
6478 /* On machines without logical shifts, if the operand of the AND is
6479 a logical shift and our mask turns off all the propagated sign
6480 bits, we can replace the logical shift with an arithmetic shift. */
6481 else if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6482 && !have_insn_for (LSHIFTRT, mode)
6483 && have_insn_for (ASHIFTRT, mode)
6484 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6485 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
6486 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
6487 && mode_width <= HOST_BITS_PER_WIDE_INT)
6489 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
6491 mask >>= INTVAL (XEXP (XEXP (x, 0), 1));
6492 if ((INTVAL (XEXP (x, 1)) & ~mask) == 0)
6493 SUBST (XEXP (x, 0),
6494 gen_rtx_ASHIFTRT (mode,
6495 make_compound_operation
6496 (XEXP (XEXP (x, 0), 0), next_code),
6497 XEXP (XEXP (x, 0), 1)));
6500 /* If the constant is one less than a power of two, this might be
6501 representable by an extraction even if no shift is present.
6502 If it doesn't end up being a ZERO_EXTEND, we will ignore it unless
6503 we are in a COMPARE. */
6504 else if ((i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6505 new = make_extraction (mode,
6506 make_compound_operation (XEXP (x, 0),
6507 next_code),
6508 0, NULL_RTX, i, 1, 0, in_code == COMPARE);
6510 /* If we are in a comparison and this is an AND with a power of two,
6511 convert this into the appropriate bit extract. */
6512 else if (in_code == COMPARE
6513 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
6514 new = make_extraction (mode,
6515 make_compound_operation (XEXP (x, 0),
6516 next_code),
6517 i, NULL_RTX, 1, 1, 0, 1);
6519 break;
6521 case LSHIFTRT:
6522 /* If the sign bit is known to be zero, replace this with an
6523 arithmetic shift. */
6524 if (have_insn_for (ASHIFTRT, mode)
6525 && ! have_insn_for (LSHIFTRT, mode)
6526 && mode_width <= HOST_BITS_PER_WIDE_INT
6527 && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0)
6529 new = gen_rtx_ASHIFTRT (mode,
6530 make_compound_operation (XEXP (x, 0),
6531 next_code),
6532 XEXP (x, 1));
6533 break;
6536 /* ... fall through ... */
6538 case ASHIFTRT:
6539 lhs = XEXP (x, 0);
6540 rhs = XEXP (x, 1);
6542 /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1,
6543 this is a SIGN_EXTRACT. */
6544 if (GET_CODE (rhs) == CONST_INT
6545 && GET_CODE (lhs) == ASHIFT
6546 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
6547 && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1)))
6549 new = make_compound_operation (XEXP (lhs, 0), next_code);
6550 new = make_extraction (mode, new,
6551 INTVAL (rhs) - INTVAL (XEXP (lhs, 1)),
6552 NULL_RTX, mode_width - INTVAL (rhs),
6553 code == LSHIFTRT, 0, in_code == COMPARE);
6554 break;
6557 /* See if we have operations between an ASHIFTRT and an ASHIFT.
6558 If so, try to merge the shifts into a SIGN_EXTEND. We could
6559 also do this for some cases of SIGN_EXTRACT, but it doesn't
6560 seem worth the effort; the case checked for occurs on Alpha. */
6562 if (!OBJECT_P (lhs)
6563 && ! (GET_CODE (lhs) == SUBREG
6564 && (OBJECT_P (SUBREG_REG (lhs))))
6565 && GET_CODE (rhs) == CONST_INT
6566 && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT
6567 && (new = extract_left_shift (lhs, INTVAL (rhs))) != 0)
6568 new = make_extraction (mode, make_compound_operation (new, next_code),
6569 0, NULL_RTX, mode_width - INTVAL (rhs),
6570 code == LSHIFTRT, 0, in_code == COMPARE);
6572 break;
6574 case SUBREG:
6575 /* Call ourselves recursively on the inner expression. If we are
6576 narrowing the object and it has a different RTL code from
6577 what it originally did, do this SUBREG as a force_to_mode. */
6579 tem = make_compound_operation (SUBREG_REG (x), in_code);
6582 rtx simplified;
6583 simplified = simplify_subreg (GET_MODE (x), tem, GET_MODE (tem),
6584 SUBREG_BYTE (x));
6586 if (simplified)
6587 tem = simplified;
6589 if (GET_CODE (tem) != GET_CODE (SUBREG_REG (x))
6590 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (tem))
6591 && subreg_lowpart_p (x))
6593 rtx newer = force_to_mode (tem, mode, ~(HOST_WIDE_INT) 0,
6596 /* If we have something other than a SUBREG, we might have
6597 done an expansion, so rerun ourselves. */
6598 if (GET_CODE (newer) != SUBREG)
6599 newer = make_compound_operation (newer, in_code);
6601 return newer;
6604 if (simplified)
6605 return tem;
6607 break;
6609 default:
6610 break;
6613 if (new)
6615 x = gen_lowpart (mode, new);
6616 code = GET_CODE (x);
6619 /* Now recursively process each operand of this operation. */
6620 fmt = GET_RTX_FORMAT (code);
6621 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6622 if (fmt[i] == 'e')
6624 new = make_compound_operation (XEXP (x, i), next_code);
6625 SUBST (XEXP (x, i), new);
6628 /* If this is a commutative operation, the changes to the operands
6629 may have made it noncanonical. */
6630 if (COMMUTATIVE_ARITH_P (x)
6631 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
6633 tem = XEXP (x, 0);
6634 SUBST (XEXP (x, 0), XEXP (x, 1));
6635 SUBST (XEXP (x, 1), tem);
6638 return x;
6641 /* Given M see if it is a value that would select a field of bits
6642 within an item, but not the entire word. Return -1 if not.
6643 Otherwise, return the starting position of the field, where 0 is the
6644 low-order bit.
6646 *PLEN is set to the length of the field. */
6648 static int
6649 get_pos_from_mask (unsigned HOST_WIDE_INT m, unsigned HOST_WIDE_INT *plen)
6651 /* Get the bit number of the first 1 bit from the right, -1 if none. */
6652 int pos = exact_log2 (m & -m);
6653 int len = 0;
6655 if (pos >= 0)
6656 /* Now shift off the low-order zero bits and see if we have a
6657 power of two minus 1. */
6658 len = exact_log2 ((m >> pos) + 1);
6660 if (len <= 0)
6661 pos = -1;
6663 *plen = len;
6664 return pos;
6667 /* If X refers to a register that equals REG in value, replace these
6668 references with REG. */
6669 static rtx
6670 canon_reg_for_combine (rtx x, rtx reg)
6672 rtx op0, op1, op2;
6673 const char *fmt;
6674 int i;
6675 bool copied;
6677 enum rtx_code code = GET_CODE (x);
6678 switch (GET_RTX_CLASS (code))
6680 case RTX_UNARY:
6681 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6682 if (op0 != XEXP (x, 0))
6683 return simplify_gen_unary (GET_CODE (x), GET_MODE (x), op0,
6684 GET_MODE (reg));
6685 break;
6687 case RTX_BIN_ARITH:
6688 case RTX_COMM_ARITH:
6689 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6690 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6691 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
6692 return simplify_gen_binary (GET_CODE (x), GET_MODE (x), op0, op1);
6693 break;
6695 case RTX_COMPARE:
6696 case RTX_COMM_COMPARE:
6697 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6698 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6699 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
6700 return simplify_gen_relational (GET_CODE (x), GET_MODE (x),
6701 GET_MODE (op0), op0, op1);
6702 break;
6704 case RTX_TERNARY:
6705 case RTX_BITFIELD_OPS:
6706 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6707 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6708 op2 = canon_reg_for_combine (XEXP (x, 2), reg);
6709 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1) || op2 != XEXP (x, 2))
6710 return simplify_gen_ternary (GET_CODE (x), GET_MODE (x),
6711 GET_MODE (op0), op0, op1, op2);
6713 case RTX_OBJ:
6714 if (REG_P (x))
6716 if (rtx_equal_p (get_last_value (reg), x)
6717 || rtx_equal_p (reg, get_last_value (x)))
6718 return reg;
6719 else
6720 break;
6723 /* fall through */
6725 default:
6726 fmt = GET_RTX_FORMAT (code);
6727 copied = false;
6728 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6729 if (fmt[i] == 'e')
6731 rtx op = canon_reg_for_combine (XEXP (x, i), reg);
6732 if (op != XEXP (x, i))
6734 if (!copied)
6736 copied = true;
6737 x = copy_rtx (x);
6739 XEXP (x, i) = op;
6742 else if (fmt[i] == 'E')
6744 int j;
6745 for (j = 0; j < XVECLEN (x, i); j++)
6747 rtx op = canon_reg_for_combine (XVECEXP (x, i, j), reg);
6748 if (op != XVECEXP (x, i, j))
6750 if (!copied)
6752 copied = true;
6753 x = copy_rtx (x);
6755 XVECEXP (x, i, j) = op;
6760 break;
6763 return x;
6766 /* Return X converted to MODE. If the value is already truncated to
6767 MODE we can just return a subreg even though in the general case we
6768 would need an explicit truncation. */
6770 static rtx
6771 gen_lowpart_or_truncate (enum machine_mode mode, rtx x)
6773 if (GET_MODE_SIZE (GET_MODE (x)) <= GET_MODE_SIZE (mode)
6774 || TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
6775 GET_MODE_BITSIZE (GET_MODE (x)))
6776 || (REG_P (x) && reg_truncated_to_mode (mode, x)))
6777 return gen_lowpart (mode, x);
6778 else
6779 return gen_rtx_TRUNCATE (mode, x);
6782 /* See if X can be simplified knowing that we will only refer to it in
6783 MODE and will only refer to those bits that are nonzero in MASK.
6784 If other bits are being computed or if masking operations are done
6785 that select a superset of the bits in MASK, they can sometimes be
6786 ignored.
6788 Return a possibly simplified expression, but always convert X to
6789 MODE. If X is a CONST_INT, AND the CONST_INT with MASK.
6791 If JUST_SELECT is nonzero, don't optimize by noticing that bits in MASK
6792 are all off in X. This is used when X will be complemented, by either
6793 NOT, NEG, or XOR. */
6795 static rtx
6796 force_to_mode (rtx x, enum machine_mode mode, unsigned HOST_WIDE_INT mask,
6797 int just_select)
6799 enum rtx_code code = GET_CODE (x);
6800 int next_select = just_select || code == XOR || code == NOT || code == NEG;
6801 enum machine_mode op_mode;
6802 unsigned HOST_WIDE_INT fuller_mask, nonzero;
6803 rtx op0, op1, temp;
6805 /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the
6806 code below will do the wrong thing since the mode of such an
6807 expression is VOIDmode.
6809 Also do nothing if X is a CLOBBER; this can happen if X was
6810 the return value from a call to gen_lowpart. */
6811 if (code == CALL || code == ASM_OPERANDS || code == CLOBBER)
6812 return x;
6814 /* We want to perform the operation is its present mode unless we know
6815 that the operation is valid in MODE, in which case we do the operation
6816 in MODE. */
6817 op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x))
6818 && have_insn_for (code, mode))
6819 ? mode : GET_MODE (x));
6821 /* It is not valid to do a right-shift in a narrower mode
6822 than the one it came in with. */
6823 if ((code == LSHIFTRT || code == ASHIFTRT)
6824 && GET_MODE_BITSIZE (mode) < GET_MODE_BITSIZE (GET_MODE (x)))
6825 op_mode = GET_MODE (x);
6827 /* Truncate MASK to fit OP_MODE. */
6828 if (op_mode)
6829 mask &= GET_MODE_MASK (op_mode);
6831 /* When we have an arithmetic operation, or a shift whose count we
6832 do not know, we need to assume that all bits up to the highest-order
6833 bit in MASK will be needed. This is how we form such a mask. */
6834 if (mask & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1)))
6835 fuller_mask = ~(unsigned HOST_WIDE_INT) 0;
6836 else
6837 fuller_mask = (((unsigned HOST_WIDE_INT) 1 << (floor_log2 (mask) + 1))
6838 - 1);
6840 /* Determine what bits of X are guaranteed to be (non)zero. */
6841 nonzero = nonzero_bits (x, mode);
6843 /* If none of the bits in X are needed, return a zero. */
6844 if (! just_select && (nonzero & mask) == 0)
6845 x = const0_rtx;
6847 /* If X is a CONST_INT, return a new one. Do this here since the
6848 test below will fail. */
6849 if (GET_CODE (x) == CONST_INT)
6851 if (SCALAR_INT_MODE_P (mode))
6852 return gen_int_mode (INTVAL (x) & mask, mode);
6853 else
6855 x = GEN_INT (INTVAL (x) & mask);
6856 return gen_lowpart_common (mode, x);
6860 /* If X is narrower than MODE and we want all the bits in X's mode, just
6861 get X in the proper mode. */
6862 if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)
6863 && (GET_MODE_MASK (GET_MODE (x)) & ~mask) == 0)
6864 return gen_lowpart (mode, x);
6866 switch (code)
6868 case CLOBBER:
6869 /* If X is a (clobber (const_int)), return it since we know we are
6870 generating something that won't match. */
6871 return x;
6873 case SIGN_EXTEND:
6874 case ZERO_EXTEND:
6875 case ZERO_EXTRACT:
6876 case SIGN_EXTRACT:
6877 x = expand_compound_operation (x);
6878 if (GET_CODE (x) != code)
6879 return force_to_mode (x, mode, mask, next_select);
6880 break;
6882 case SUBREG:
6883 if (subreg_lowpart_p (x)
6884 /* We can ignore the effect of this SUBREG if it narrows the mode or
6885 if the constant masks to zero all the bits the mode doesn't
6886 have. */
6887 && ((GET_MODE_SIZE (GET_MODE (x))
6888 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
6889 || (0 == (mask
6890 & GET_MODE_MASK (GET_MODE (x))
6891 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))))))
6892 return force_to_mode (SUBREG_REG (x), mode, mask, next_select);
6893 break;
6895 case AND:
6896 /* If this is an AND with a constant, convert it into an AND
6897 whose constant is the AND of that constant with MASK. If it
6898 remains an AND of MASK, delete it since it is redundant. */
6900 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6902 x = simplify_and_const_int (x, op_mode, XEXP (x, 0),
6903 mask & INTVAL (XEXP (x, 1)));
6905 /* If X is still an AND, see if it is an AND with a mask that
6906 is just some low-order bits. If so, and it is MASK, we don't
6907 need it. */
6909 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
6910 && ((INTVAL (XEXP (x, 1)) & GET_MODE_MASK (GET_MODE (x)))
6911 == mask))
6912 x = XEXP (x, 0);
6914 /* If it remains an AND, try making another AND with the bits
6915 in the mode mask that aren't in MASK turned on. If the
6916 constant in the AND is wide enough, this might make a
6917 cheaper constant. */
6919 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
6920 && GET_MODE_MASK (GET_MODE (x)) != mask
6921 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
6923 HOST_WIDE_INT cval = (INTVAL (XEXP (x, 1))
6924 | (GET_MODE_MASK (GET_MODE (x)) & ~mask));
6925 int width = GET_MODE_BITSIZE (GET_MODE (x));
6926 rtx y;
6928 /* If MODE is narrower than HOST_WIDE_INT and CVAL is a negative
6929 number, sign extend it. */
6930 if (width > 0 && width < HOST_BITS_PER_WIDE_INT
6931 && (cval & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
6932 cval |= (HOST_WIDE_INT) -1 << width;
6934 y = simplify_gen_binary (AND, GET_MODE (x),
6935 XEXP (x, 0), GEN_INT (cval));
6936 if (rtx_cost (y, SET) < rtx_cost (x, SET))
6937 x = y;
6940 break;
6943 goto binop;
6945 case PLUS:
6946 /* In (and (plus FOO C1) M), if M is a mask that just turns off
6947 low-order bits (as in an alignment operation) and FOO is already
6948 aligned to that boundary, mask C1 to that boundary as well.
6949 This may eliminate that PLUS and, later, the AND. */
6952 unsigned int width = GET_MODE_BITSIZE (mode);
6953 unsigned HOST_WIDE_INT smask = mask;
6955 /* If MODE is narrower than HOST_WIDE_INT and mask is a negative
6956 number, sign extend it. */
6958 if (width < HOST_BITS_PER_WIDE_INT
6959 && (smask & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
6960 smask |= (HOST_WIDE_INT) -1 << width;
6962 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6963 && exact_log2 (- smask) >= 0
6964 && (nonzero_bits (XEXP (x, 0), mode) & ~smask) == 0
6965 && (INTVAL (XEXP (x, 1)) & ~smask) != 0)
6966 return force_to_mode (plus_constant (XEXP (x, 0),
6967 (INTVAL (XEXP (x, 1)) & smask)),
6968 mode, smask, next_select);
6971 /* ... fall through ... */
6973 case MULT:
6974 /* For PLUS, MINUS and MULT, we need any bits less significant than the
6975 most significant bit in MASK since carries from those bits will
6976 affect the bits we are interested in. */
6977 mask = fuller_mask;
6978 goto binop;
6980 case MINUS:
6981 /* If X is (minus C Y) where C's least set bit is larger than any bit
6982 in the mask, then we may replace with (neg Y). */
6983 if (GET_CODE (XEXP (x, 0)) == CONST_INT
6984 && (((unsigned HOST_WIDE_INT) (INTVAL (XEXP (x, 0))
6985 & -INTVAL (XEXP (x, 0))))
6986 > mask))
6988 x = simplify_gen_unary (NEG, GET_MODE (x), XEXP (x, 1),
6989 GET_MODE (x));
6990 return force_to_mode (x, mode, mask, next_select);
6993 /* Similarly, if C contains every bit in the fuller_mask, then we may
6994 replace with (not Y). */
6995 if (GET_CODE (XEXP (x, 0)) == CONST_INT
6996 && ((INTVAL (XEXP (x, 0)) | (HOST_WIDE_INT) fuller_mask)
6997 == INTVAL (XEXP (x, 0))))
6999 x = simplify_gen_unary (NOT, GET_MODE (x),
7000 XEXP (x, 1), GET_MODE (x));
7001 return force_to_mode (x, mode, mask, next_select);
7004 mask = fuller_mask;
7005 goto binop;
7007 case IOR:
7008 case XOR:
7009 /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and
7010 LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...)
7011 operation which may be a bitfield extraction. Ensure that the
7012 constant we form is not wider than the mode of X. */
7014 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7015 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7016 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7017 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
7018 && GET_CODE (XEXP (x, 1)) == CONST_INT
7019 && ((INTVAL (XEXP (XEXP (x, 0), 1))
7020 + floor_log2 (INTVAL (XEXP (x, 1))))
7021 < GET_MODE_BITSIZE (GET_MODE (x)))
7022 && (INTVAL (XEXP (x, 1))
7023 & ~nonzero_bits (XEXP (x, 0), GET_MODE (x))) == 0)
7025 temp = GEN_INT ((INTVAL (XEXP (x, 1)) & mask)
7026 << INTVAL (XEXP (XEXP (x, 0), 1)));
7027 temp = simplify_gen_binary (GET_CODE (x), GET_MODE (x),
7028 XEXP (XEXP (x, 0), 0), temp);
7029 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), temp,
7030 XEXP (XEXP (x, 0), 1));
7031 return force_to_mode (x, mode, mask, next_select);
7034 binop:
7035 /* For most binary operations, just propagate into the operation and
7036 change the mode if we have an operation of that mode. */
7038 op0 = gen_lowpart_or_truncate (op_mode,
7039 force_to_mode (XEXP (x, 0), mode, mask,
7040 next_select));
7041 op1 = gen_lowpart_or_truncate (op_mode,
7042 force_to_mode (XEXP (x, 1), mode, mask,
7043 next_select));
7045 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
7046 x = simplify_gen_binary (code, op_mode, op0, op1);
7047 break;
7049 case ASHIFT:
7050 /* For left shifts, do the same, but just for the first operand.
7051 However, we cannot do anything with shifts where we cannot
7052 guarantee that the counts are smaller than the size of the mode
7053 because such a count will have a different meaning in a
7054 wider mode. */
7056 if (! (GET_CODE (XEXP (x, 1)) == CONST_INT
7057 && INTVAL (XEXP (x, 1)) >= 0
7058 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode))
7059 && ! (GET_MODE (XEXP (x, 1)) != VOIDmode
7060 && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1)))
7061 < (unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode))))
7062 break;
7064 /* If the shift count is a constant and we can do arithmetic in
7065 the mode of the shift, refine which bits we need. Otherwise, use the
7066 conservative form of the mask. */
7067 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7068 && INTVAL (XEXP (x, 1)) >= 0
7069 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (op_mode)
7070 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7071 mask >>= INTVAL (XEXP (x, 1));
7072 else
7073 mask = fuller_mask;
7075 op0 = gen_lowpart_or_truncate (op_mode,
7076 force_to_mode (XEXP (x, 0), op_mode,
7077 mask, next_select));
7079 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7080 x = simplify_gen_binary (code, op_mode, op0, XEXP (x, 1));
7081 break;
7083 case LSHIFTRT:
7084 /* Here we can only do something if the shift count is a constant,
7085 this shift constant is valid for the host, and we can do arithmetic
7086 in OP_MODE. */
7088 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7089 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
7090 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7092 rtx inner = XEXP (x, 0);
7093 unsigned HOST_WIDE_INT inner_mask;
7095 /* Select the mask of the bits we need for the shift operand. */
7096 inner_mask = mask << INTVAL (XEXP (x, 1));
7098 /* We can only change the mode of the shift if we can do arithmetic
7099 in the mode of the shift and INNER_MASK is no wider than the
7100 width of X's mode. */
7101 if ((inner_mask & ~GET_MODE_MASK (GET_MODE (x))) != 0)
7102 op_mode = GET_MODE (x);
7104 inner = force_to_mode (inner, op_mode, inner_mask, next_select);
7106 if (GET_MODE (x) != op_mode || inner != XEXP (x, 0))
7107 x = simplify_gen_binary (LSHIFTRT, op_mode, inner, XEXP (x, 1));
7110 /* If we have (and (lshiftrt FOO C1) C2) where the combination of the
7111 shift and AND produces only copies of the sign bit (C2 is one less
7112 than a power of two), we can do this with just a shift. */
7114 if (GET_CODE (x) == LSHIFTRT
7115 && GET_CODE (XEXP (x, 1)) == CONST_INT
7116 /* The shift puts one of the sign bit copies in the least significant
7117 bit. */
7118 && ((INTVAL (XEXP (x, 1))
7119 + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))))
7120 >= GET_MODE_BITSIZE (GET_MODE (x)))
7121 && exact_log2 (mask + 1) >= 0
7122 /* Number of bits left after the shift must be more than the mask
7123 needs. */
7124 && ((INTVAL (XEXP (x, 1)) + exact_log2 (mask + 1))
7125 <= GET_MODE_BITSIZE (GET_MODE (x)))
7126 /* Must be more sign bit copies than the mask needs. */
7127 && ((int) num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
7128 >= exact_log2 (mask + 1)))
7129 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7130 GEN_INT (GET_MODE_BITSIZE (GET_MODE (x))
7131 - exact_log2 (mask + 1)));
7133 goto shiftrt;
7135 case ASHIFTRT:
7136 /* If we are just looking for the sign bit, we don't need this shift at
7137 all, even if it has a variable count. */
7138 if (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
7139 && (mask == ((unsigned HOST_WIDE_INT) 1
7140 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
7141 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
7143 /* If this is a shift by a constant, get a mask that contains those bits
7144 that are not copies of the sign bit. We then have two cases: If
7145 MASK only includes those bits, this can be a logical shift, which may
7146 allow simplifications. If MASK is a single-bit field not within
7147 those bits, we are requesting a copy of the sign bit and hence can
7148 shift the sign bit to the appropriate location. */
7150 if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) >= 0
7151 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT)
7153 int i;
7155 /* If the considered data is wider than HOST_WIDE_INT, we can't
7156 represent a mask for all its bits in a single scalar.
7157 But we only care about the lower bits, so calculate these. */
7159 if (GET_MODE_BITSIZE (GET_MODE (x)) > HOST_BITS_PER_WIDE_INT)
7161 nonzero = ~(HOST_WIDE_INT) 0;
7163 /* GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7164 is the number of bits a full-width mask would have set.
7165 We need only shift if these are fewer than nonzero can
7166 hold. If not, we must keep all bits set in nonzero. */
7168 if (GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7169 < HOST_BITS_PER_WIDE_INT)
7170 nonzero >>= INTVAL (XEXP (x, 1))
7171 + HOST_BITS_PER_WIDE_INT
7172 - GET_MODE_BITSIZE (GET_MODE (x)) ;
7174 else
7176 nonzero = GET_MODE_MASK (GET_MODE (x));
7177 nonzero >>= INTVAL (XEXP (x, 1));
7180 if ((mask & ~nonzero) == 0)
7182 x = simplify_shift_const (NULL_RTX, LSHIFTRT, GET_MODE (x),
7183 XEXP (x, 0), INTVAL (XEXP (x, 1)));
7184 if (GET_CODE (x) != ASHIFTRT)
7185 return force_to_mode (x, mode, mask, next_select);
7188 else if ((i = exact_log2 (mask)) >= 0)
7190 x = simplify_shift_const
7191 (NULL_RTX, LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7192 GET_MODE_BITSIZE (GET_MODE (x)) - 1 - i);
7194 if (GET_CODE (x) != ASHIFTRT)
7195 return force_to_mode (x, mode, mask, next_select);
7199 /* If MASK is 1, convert this to an LSHIFTRT. This can be done
7200 even if the shift count isn't a constant. */
7201 if (mask == 1)
7202 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7203 XEXP (x, 0), XEXP (x, 1));
7205 shiftrt:
7207 /* If this is a zero- or sign-extension operation that just affects bits
7208 we don't care about, remove it. Be sure the call above returned
7209 something that is still a shift. */
7211 if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT)
7212 && GET_CODE (XEXP (x, 1)) == CONST_INT
7213 && INTVAL (XEXP (x, 1)) >= 0
7214 && (INTVAL (XEXP (x, 1))
7215 <= GET_MODE_BITSIZE (GET_MODE (x)) - (floor_log2 (mask) + 1))
7216 && GET_CODE (XEXP (x, 0)) == ASHIFT
7217 && XEXP (XEXP (x, 0), 1) == XEXP (x, 1))
7218 return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask,
7219 next_select);
7221 break;
7223 case ROTATE:
7224 case ROTATERT:
7225 /* If the shift count is constant and we can do computations
7226 in the mode of X, compute where the bits we care about are.
7227 Otherwise, we can't do anything. Don't change the mode of
7228 the shift or propagate MODE into the shift, though. */
7229 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7230 && INTVAL (XEXP (x, 1)) >= 0)
7232 temp = simplify_binary_operation (code == ROTATE ? ROTATERT : ROTATE,
7233 GET_MODE (x), GEN_INT (mask),
7234 XEXP (x, 1));
7235 if (temp && GET_CODE (temp) == CONST_INT)
7236 SUBST (XEXP (x, 0),
7237 force_to_mode (XEXP (x, 0), GET_MODE (x),
7238 INTVAL (temp), next_select));
7240 break;
7242 case NEG:
7243 /* If we just want the low-order bit, the NEG isn't needed since it
7244 won't change the low-order bit. */
7245 if (mask == 1)
7246 return force_to_mode (XEXP (x, 0), mode, mask, just_select);
7248 /* We need any bits less significant than the most significant bit in
7249 MASK since carries from those bits will affect the bits we are
7250 interested in. */
7251 mask = fuller_mask;
7252 goto unop;
7254 case NOT:
7255 /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the
7256 same as the XOR case above. Ensure that the constant we form is not
7257 wider than the mode of X. */
7259 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7260 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7261 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7262 && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (mask)
7263 < GET_MODE_BITSIZE (GET_MODE (x)))
7264 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT)
7266 temp = gen_int_mode (mask << INTVAL (XEXP (XEXP (x, 0), 1)),
7267 GET_MODE (x));
7268 temp = simplify_gen_binary (XOR, GET_MODE (x),
7269 XEXP (XEXP (x, 0), 0), temp);
7270 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7271 temp, XEXP (XEXP (x, 0), 1));
7273 return force_to_mode (x, mode, mask, next_select);
7276 /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must
7277 use the full mask inside the NOT. */
7278 mask = fuller_mask;
7280 unop:
7281 op0 = gen_lowpart_or_truncate (op_mode,
7282 force_to_mode (XEXP (x, 0), mode, mask,
7283 next_select));
7284 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7285 x = simplify_gen_unary (code, op_mode, op0, op_mode);
7286 break;
7288 case NE:
7289 /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included
7290 in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero,
7291 which is equal to STORE_FLAG_VALUE. */
7292 if ((mask & ~STORE_FLAG_VALUE) == 0 && XEXP (x, 1) == const0_rtx
7293 && GET_MODE (XEXP (x, 0)) == mode
7294 && exact_log2 (nonzero_bits (XEXP (x, 0), mode)) >= 0
7295 && (nonzero_bits (XEXP (x, 0), mode)
7296 == (unsigned HOST_WIDE_INT) STORE_FLAG_VALUE))
7297 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
7299 break;
7301 case IF_THEN_ELSE:
7302 /* We have no way of knowing if the IF_THEN_ELSE can itself be
7303 written in a narrower mode. We play it safe and do not do so. */
7305 SUBST (XEXP (x, 1),
7306 gen_lowpart_or_truncate (GET_MODE (x),
7307 force_to_mode (XEXP (x, 1), mode,
7308 mask, next_select)));
7309 SUBST (XEXP (x, 2),
7310 gen_lowpart_or_truncate (GET_MODE (x),
7311 force_to_mode (XEXP (x, 2), mode,
7312 mask, next_select)));
7313 break;
7315 default:
7316 break;
7319 /* Ensure we return a value of the proper mode. */
7320 return gen_lowpart_or_truncate (mode, x);
7323 /* Return nonzero if X is an expression that has one of two values depending on
7324 whether some other value is zero or nonzero. In that case, we return the
7325 value that is being tested, *PTRUE is set to the value if the rtx being
7326 returned has a nonzero value, and *PFALSE is set to the other alternative.
7328 If we return zero, we set *PTRUE and *PFALSE to X. */
7330 static rtx
7331 if_then_else_cond (rtx x, rtx *ptrue, rtx *pfalse)
7333 enum machine_mode mode = GET_MODE (x);
7334 enum rtx_code code = GET_CODE (x);
7335 rtx cond0, cond1, true0, true1, false0, false1;
7336 unsigned HOST_WIDE_INT nz;
7338 /* If we are comparing a value against zero, we are done. */
7339 if ((code == NE || code == EQ)
7340 && XEXP (x, 1) == const0_rtx)
7342 *ptrue = (code == NE) ? const_true_rtx : const0_rtx;
7343 *pfalse = (code == NE) ? const0_rtx : const_true_rtx;
7344 return XEXP (x, 0);
7347 /* If this is a unary operation whose operand has one of two values, apply
7348 our opcode to compute those values. */
7349 else if (UNARY_P (x)
7350 && (cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0)) != 0)
7352 *ptrue = simplify_gen_unary (code, mode, true0, GET_MODE (XEXP (x, 0)));
7353 *pfalse = simplify_gen_unary (code, mode, false0,
7354 GET_MODE (XEXP (x, 0)));
7355 return cond0;
7358 /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would
7359 make can't possibly match and would suppress other optimizations. */
7360 else if (code == COMPARE)
7363 /* If this is a binary operation, see if either side has only one of two
7364 values. If either one does or if both do and they are conditional on
7365 the same value, compute the new true and false values. */
7366 else if (BINARY_P (x))
7368 cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0);
7369 cond1 = if_then_else_cond (XEXP (x, 1), &true1, &false1);
7371 if ((cond0 != 0 || cond1 != 0)
7372 && ! (cond0 != 0 && cond1 != 0 && ! rtx_equal_p (cond0, cond1)))
7374 /* If if_then_else_cond returned zero, then true/false are the
7375 same rtl. We must copy one of them to prevent invalid rtl
7376 sharing. */
7377 if (cond0 == 0)
7378 true0 = copy_rtx (true0);
7379 else if (cond1 == 0)
7380 true1 = copy_rtx (true1);
7382 if (COMPARISON_P (x))
7384 *ptrue = simplify_gen_relational (code, mode, VOIDmode,
7385 true0, true1);
7386 *pfalse = simplify_gen_relational (code, mode, VOIDmode,
7387 false0, false1);
7389 else
7391 *ptrue = simplify_gen_binary (code, mode, true0, true1);
7392 *pfalse = simplify_gen_binary (code, mode, false0, false1);
7395 return cond0 ? cond0 : cond1;
7398 /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the
7399 operands is zero when the other is nonzero, and vice-versa,
7400 and STORE_FLAG_VALUE is 1 or -1. */
7402 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7403 && (code == PLUS || code == IOR || code == XOR || code == MINUS
7404 || code == UMAX)
7405 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7407 rtx op0 = XEXP (XEXP (x, 0), 1);
7408 rtx op1 = XEXP (XEXP (x, 1), 1);
7410 cond0 = XEXP (XEXP (x, 0), 0);
7411 cond1 = XEXP (XEXP (x, 1), 0);
7413 if (COMPARISON_P (cond0)
7414 && COMPARISON_P (cond1)
7415 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7416 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7417 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7418 || ((swap_condition (GET_CODE (cond0))
7419 == reversed_comparison_code (cond1, NULL))
7420 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7421 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7422 && ! side_effects_p (x))
7424 *ptrue = simplify_gen_binary (MULT, mode, op0, const_true_rtx);
7425 *pfalse = simplify_gen_binary (MULT, mode,
7426 (code == MINUS
7427 ? simplify_gen_unary (NEG, mode,
7428 op1, mode)
7429 : op1),
7430 const_true_rtx);
7431 return cond0;
7435 /* Similarly for MULT, AND and UMIN, except that for these the result
7436 is always zero. */
7437 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7438 && (code == MULT || code == AND || code == UMIN)
7439 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7441 cond0 = XEXP (XEXP (x, 0), 0);
7442 cond1 = XEXP (XEXP (x, 1), 0);
7444 if (COMPARISON_P (cond0)
7445 && COMPARISON_P (cond1)
7446 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7447 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7448 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7449 || ((swap_condition (GET_CODE (cond0))
7450 == reversed_comparison_code (cond1, NULL))
7451 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7452 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7453 && ! side_effects_p (x))
7455 *ptrue = *pfalse = const0_rtx;
7456 return cond0;
7461 else if (code == IF_THEN_ELSE)
7463 /* If we have IF_THEN_ELSE already, extract the condition and
7464 canonicalize it if it is NE or EQ. */
7465 cond0 = XEXP (x, 0);
7466 *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2);
7467 if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx)
7468 return XEXP (cond0, 0);
7469 else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx)
7471 *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1);
7472 return XEXP (cond0, 0);
7474 else
7475 return cond0;
7478 /* If X is a SUBREG, we can narrow both the true and false values
7479 if the inner expression, if there is a condition. */
7480 else if (code == SUBREG
7481 && 0 != (cond0 = if_then_else_cond (SUBREG_REG (x),
7482 &true0, &false0)))
7484 true0 = simplify_gen_subreg (mode, true0,
7485 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7486 false0 = simplify_gen_subreg (mode, false0,
7487 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7488 if (true0 && false0)
7490 *ptrue = true0;
7491 *pfalse = false0;
7492 return cond0;
7496 /* If X is a constant, this isn't special and will cause confusions
7497 if we treat it as such. Likewise if it is equivalent to a constant. */
7498 else if (CONSTANT_P (x)
7499 || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0)))
7502 /* If we're in BImode, canonicalize on 0 and STORE_FLAG_VALUE, as that
7503 will be least confusing to the rest of the compiler. */
7504 else if (mode == BImode)
7506 *ptrue = GEN_INT (STORE_FLAG_VALUE), *pfalse = const0_rtx;
7507 return x;
7510 /* If X is known to be either 0 or -1, those are the true and
7511 false values when testing X. */
7512 else if (x == constm1_rtx || x == const0_rtx
7513 || (mode != VOIDmode
7514 && num_sign_bit_copies (x, mode) == GET_MODE_BITSIZE (mode)))
7516 *ptrue = constm1_rtx, *pfalse = const0_rtx;
7517 return x;
7520 /* Likewise for 0 or a single bit. */
7521 else if (SCALAR_INT_MODE_P (mode)
7522 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
7523 && exact_log2 (nz = nonzero_bits (x, mode)) >= 0)
7525 *ptrue = gen_int_mode (nz, mode), *pfalse = const0_rtx;
7526 return x;
7529 /* Otherwise fail; show no condition with true and false values the same. */
7530 *ptrue = *pfalse = x;
7531 return 0;
7534 /* Return the value of expression X given the fact that condition COND
7535 is known to be true when applied to REG as its first operand and VAL
7536 as its second. X is known to not be shared and so can be modified in
7537 place.
7539 We only handle the simplest cases, and specifically those cases that
7540 arise with IF_THEN_ELSE expressions. */
7542 static rtx
7543 known_cond (rtx x, enum rtx_code cond, rtx reg, rtx val)
7545 enum rtx_code code = GET_CODE (x);
7546 rtx temp;
7547 const char *fmt;
7548 int i, j;
7550 if (side_effects_p (x))
7551 return x;
7553 /* If either operand of the condition is a floating point value,
7554 then we have to avoid collapsing an EQ comparison. */
7555 if (cond == EQ
7556 && rtx_equal_p (x, reg)
7557 && ! FLOAT_MODE_P (GET_MODE (x))
7558 && ! FLOAT_MODE_P (GET_MODE (val)))
7559 return val;
7561 if (cond == UNEQ && rtx_equal_p (x, reg))
7562 return val;
7564 /* If X is (abs REG) and we know something about REG's relationship
7565 with zero, we may be able to simplify this. */
7567 if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx)
7568 switch (cond)
7570 case GE: case GT: case EQ:
7571 return XEXP (x, 0);
7572 case LT: case LE:
7573 return simplify_gen_unary (NEG, GET_MODE (XEXP (x, 0)),
7574 XEXP (x, 0),
7575 GET_MODE (XEXP (x, 0)));
7576 default:
7577 break;
7580 /* The only other cases we handle are MIN, MAX, and comparisons if the
7581 operands are the same as REG and VAL. */
7583 else if (COMPARISON_P (x) || COMMUTATIVE_ARITH_P (x))
7585 if (rtx_equal_p (XEXP (x, 0), val))
7586 cond = swap_condition (cond), temp = val, val = reg, reg = temp;
7588 if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val))
7590 if (COMPARISON_P (x))
7592 if (comparison_dominates_p (cond, code))
7593 return const_true_rtx;
7595 code = reversed_comparison_code (x, NULL);
7596 if (code != UNKNOWN
7597 && comparison_dominates_p (cond, code))
7598 return const0_rtx;
7599 else
7600 return x;
7602 else if (code == SMAX || code == SMIN
7603 || code == UMIN || code == UMAX)
7605 int unsignedp = (code == UMIN || code == UMAX);
7607 /* Do not reverse the condition when it is NE or EQ.
7608 This is because we cannot conclude anything about
7609 the value of 'SMAX (x, y)' when x is not equal to y,
7610 but we can when x equals y. */
7611 if ((code == SMAX || code == UMAX)
7612 && ! (cond == EQ || cond == NE))
7613 cond = reverse_condition (cond);
7615 switch (cond)
7617 case GE: case GT:
7618 return unsignedp ? x : XEXP (x, 1);
7619 case LE: case LT:
7620 return unsignedp ? x : XEXP (x, 0);
7621 case GEU: case GTU:
7622 return unsignedp ? XEXP (x, 1) : x;
7623 case LEU: case LTU:
7624 return unsignedp ? XEXP (x, 0) : x;
7625 default:
7626 break;
7631 else if (code == SUBREG)
7633 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (x));
7634 rtx new, r = known_cond (SUBREG_REG (x), cond, reg, val);
7636 if (SUBREG_REG (x) != r)
7638 /* We must simplify subreg here, before we lose track of the
7639 original inner_mode. */
7640 new = simplify_subreg (GET_MODE (x), r,
7641 inner_mode, SUBREG_BYTE (x));
7642 if (new)
7643 return new;
7644 else
7645 SUBST (SUBREG_REG (x), r);
7648 return x;
7650 /* We don't have to handle SIGN_EXTEND here, because even in the
7651 case of replacing something with a modeless CONST_INT, a
7652 CONST_INT is already (supposed to be) a valid sign extension for
7653 its narrower mode, which implies it's already properly
7654 sign-extended for the wider mode. Now, for ZERO_EXTEND, the
7655 story is different. */
7656 else if (code == ZERO_EXTEND)
7658 enum machine_mode inner_mode = GET_MODE (XEXP (x, 0));
7659 rtx new, r = known_cond (XEXP (x, 0), cond, reg, val);
7661 if (XEXP (x, 0) != r)
7663 /* We must simplify the zero_extend here, before we lose
7664 track of the original inner_mode. */
7665 new = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
7666 r, inner_mode);
7667 if (new)
7668 return new;
7669 else
7670 SUBST (XEXP (x, 0), r);
7673 return x;
7676 fmt = GET_RTX_FORMAT (code);
7677 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7679 if (fmt[i] == 'e')
7680 SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val));
7681 else if (fmt[i] == 'E')
7682 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7683 SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j),
7684 cond, reg, val));
7687 return x;
7690 /* See if X and Y are equal for the purposes of seeing if we can rewrite an
7691 assignment as a field assignment. */
7693 static int
7694 rtx_equal_for_field_assignment_p (rtx x, rtx y)
7696 if (x == y || rtx_equal_p (x, y))
7697 return 1;
7699 if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y))
7700 return 0;
7702 /* Check for a paradoxical SUBREG of a MEM compared with the MEM.
7703 Note that all SUBREGs of MEM are paradoxical; otherwise they
7704 would have been rewritten. */
7705 if (MEM_P (x) && GET_CODE (y) == SUBREG
7706 && MEM_P (SUBREG_REG (y))
7707 && rtx_equal_p (SUBREG_REG (y),
7708 gen_lowpart (GET_MODE (SUBREG_REG (y)), x)))
7709 return 1;
7711 if (MEM_P (y) && GET_CODE (x) == SUBREG
7712 && MEM_P (SUBREG_REG (x))
7713 && rtx_equal_p (SUBREG_REG (x),
7714 gen_lowpart (GET_MODE (SUBREG_REG (x)), y)))
7715 return 1;
7717 /* We used to see if get_last_value of X and Y were the same but that's
7718 not correct. In one direction, we'll cause the assignment to have
7719 the wrong destination and in the case, we'll import a register into this
7720 insn that might have already have been dead. So fail if none of the
7721 above cases are true. */
7722 return 0;
7725 /* See if X, a SET operation, can be rewritten as a bit-field assignment.
7726 Return that assignment if so.
7728 We only handle the most common cases. */
7730 static rtx
7731 make_field_assignment (rtx x)
7733 rtx dest = SET_DEST (x);
7734 rtx src = SET_SRC (x);
7735 rtx assign;
7736 rtx rhs, lhs;
7737 HOST_WIDE_INT c1;
7738 HOST_WIDE_INT pos;
7739 unsigned HOST_WIDE_INT len;
7740 rtx other;
7741 enum machine_mode mode;
7743 /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is
7744 a clear of a one-bit field. We will have changed it to
7745 (and (rotate (const_int -2) POS) DEST), so check for that. Also check
7746 for a SUBREG. */
7748 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE
7749 && GET_CODE (XEXP (XEXP (src, 0), 0)) == CONST_INT
7750 && INTVAL (XEXP (XEXP (src, 0), 0)) == -2
7751 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7753 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7754 1, 1, 1, 0);
7755 if (assign != 0)
7756 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7757 return x;
7760 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG
7761 && subreg_lowpart_p (XEXP (src, 0))
7762 && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0)))
7763 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0)))))
7764 && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE
7765 && GET_CODE (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == CONST_INT
7766 && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2
7767 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7769 assign = make_extraction (VOIDmode, dest, 0,
7770 XEXP (SUBREG_REG (XEXP (src, 0)), 1),
7771 1, 1, 1, 0);
7772 if (assign != 0)
7773 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7774 return x;
7777 /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a
7778 one-bit field. */
7779 if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT
7780 && XEXP (XEXP (src, 0), 0) == const1_rtx
7781 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7783 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7784 1, 1, 1, 0);
7785 if (assign != 0)
7786 return gen_rtx_SET (VOIDmode, assign, const1_rtx);
7787 return x;
7790 /* If DEST is already a field assignment, i.e. ZERO_EXTRACT, and the
7791 SRC is an AND with all bits of that field set, then we can discard
7792 the AND. */
7793 if (GET_CODE (dest) == ZERO_EXTRACT
7794 && GET_CODE (XEXP (dest, 1)) == CONST_INT
7795 && GET_CODE (src) == AND
7796 && GET_CODE (XEXP (src, 1)) == CONST_INT)
7798 HOST_WIDE_INT width = INTVAL (XEXP (dest, 1));
7799 unsigned HOST_WIDE_INT and_mask = INTVAL (XEXP (src, 1));
7800 unsigned HOST_WIDE_INT ze_mask;
7802 if (width >= HOST_BITS_PER_WIDE_INT)
7803 ze_mask = -1;
7804 else
7805 ze_mask = ((unsigned HOST_WIDE_INT)1 << width) - 1;
7807 /* Complete overlap. We can remove the source AND. */
7808 if ((and_mask & ze_mask) == ze_mask)
7809 return gen_rtx_SET (VOIDmode, dest, XEXP (src, 0));
7811 /* Partial overlap. We can reduce the source AND. */
7812 if ((and_mask & ze_mask) != and_mask)
7814 mode = GET_MODE (src);
7815 src = gen_rtx_AND (mode, XEXP (src, 0),
7816 gen_int_mode (and_mask & ze_mask, mode));
7817 return gen_rtx_SET (VOIDmode, dest, src);
7821 /* The other case we handle is assignments into a constant-position
7822 field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents
7823 a mask that has all one bits except for a group of zero bits and
7824 OTHER is known to have zeros where C1 has ones, this is such an
7825 assignment. Compute the position and length from C1. Shift OTHER
7826 to the appropriate position, force it to the required mode, and
7827 make the extraction. Check for the AND in both operands. */
7829 if (GET_CODE (src) != IOR && GET_CODE (src) != XOR)
7830 return x;
7832 rhs = expand_compound_operation (XEXP (src, 0));
7833 lhs = expand_compound_operation (XEXP (src, 1));
7835 if (GET_CODE (rhs) == AND
7836 && GET_CODE (XEXP (rhs, 1)) == CONST_INT
7837 && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), dest))
7838 c1 = INTVAL (XEXP (rhs, 1)), other = lhs;
7839 else if (GET_CODE (lhs) == AND
7840 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
7841 && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), dest))
7842 c1 = INTVAL (XEXP (lhs, 1)), other = rhs;
7843 else
7844 return x;
7846 pos = get_pos_from_mask ((~c1) & GET_MODE_MASK (GET_MODE (dest)), &len);
7847 if (pos < 0 || pos + len > GET_MODE_BITSIZE (GET_MODE (dest))
7848 || GET_MODE_BITSIZE (GET_MODE (dest)) > HOST_BITS_PER_WIDE_INT
7849 || (c1 & nonzero_bits (other, GET_MODE (dest))) != 0)
7850 return x;
7852 assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0);
7853 if (assign == 0)
7854 return x;
7856 /* The mode to use for the source is the mode of the assignment, or of
7857 what is inside a possible STRICT_LOW_PART. */
7858 mode = (GET_CODE (assign) == STRICT_LOW_PART
7859 ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign));
7861 /* Shift OTHER right POS places and make it the source, restricting it
7862 to the proper length and mode. */
7864 src = canon_reg_for_combine (simplify_shift_const (NULL_RTX, LSHIFTRT,
7865 GET_MODE (src),
7866 other, pos),
7867 dest);
7868 src = force_to_mode (src, mode,
7869 GET_MODE_BITSIZE (mode) >= HOST_BITS_PER_WIDE_INT
7870 ? ~(unsigned HOST_WIDE_INT) 0
7871 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
7874 /* If SRC is masked by an AND that does not make a difference in
7875 the value being stored, strip it. */
7876 if (GET_CODE (assign) == ZERO_EXTRACT
7877 && GET_CODE (XEXP (assign, 1)) == CONST_INT
7878 && INTVAL (XEXP (assign, 1)) < HOST_BITS_PER_WIDE_INT
7879 && GET_CODE (src) == AND
7880 && GET_CODE (XEXP (src, 1)) == CONST_INT
7881 && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (src, 1))
7882 == ((unsigned HOST_WIDE_INT) 1 << INTVAL (XEXP (assign, 1))) - 1))
7883 src = XEXP (src, 0);
7885 return gen_rtx_SET (VOIDmode, assign, src);
7888 /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c)
7889 if so. */
7891 static rtx
7892 apply_distributive_law (rtx x)
7894 enum rtx_code code = GET_CODE (x);
7895 enum rtx_code inner_code;
7896 rtx lhs, rhs, other;
7897 rtx tem;
7899 /* Distributivity is not true for floating point as it can change the
7900 value. So we don't do it unless -funsafe-math-optimizations. */
7901 if (FLOAT_MODE_P (GET_MODE (x))
7902 && ! flag_unsafe_math_optimizations)
7903 return x;
7905 /* The outer operation can only be one of the following: */
7906 if (code != IOR && code != AND && code != XOR
7907 && code != PLUS && code != MINUS)
7908 return x;
7910 lhs = XEXP (x, 0);
7911 rhs = XEXP (x, 1);
7913 /* If either operand is a primitive we can't do anything, so get out
7914 fast. */
7915 if (OBJECT_P (lhs) || OBJECT_P (rhs))
7916 return x;
7918 lhs = expand_compound_operation (lhs);
7919 rhs = expand_compound_operation (rhs);
7920 inner_code = GET_CODE (lhs);
7921 if (inner_code != GET_CODE (rhs))
7922 return x;
7924 /* See if the inner and outer operations distribute. */
7925 switch (inner_code)
7927 case LSHIFTRT:
7928 case ASHIFTRT:
7929 case AND:
7930 case IOR:
7931 /* These all distribute except over PLUS. */
7932 if (code == PLUS || code == MINUS)
7933 return x;
7934 break;
7936 case MULT:
7937 if (code != PLUS && code != MINUS)
7938 return x;
7939 break;
7941 case ASHIFT:
7942 /* This is also a multiply, so it distributes over everything. */
7943 break;
7945 case SUBREG:
7946 /* Non-paradoxical SUBREGs distributes over all operations,
7947 provided the inner modes and byte offsets are the same, this
7948 is an extraction of a low-order part, we don't convert an fp
7949 operation to int or vice versa, this is not a vector mode,
7950 and we would not be converting a single-word operation into a
7951 multi-word operation. The latter test is not required, but
7952 it prevents generating unneeded multi-word operations. Some
7953 of the previous tests are redundant given the latter test,
7954 but are retained because they are required for correctness.
7956 We produce the result slightly differently in this case. */
7958 if (GET_MODE (SUBREG_REG (lhs)) != GET_MODE (SUBREG_REG (rhs))
7959 || SUBREG_BYTE (lhs) != SUBREG_BYTE (rhs)
7960 || ! subreg_lowpart_p (lhs)
7961 || (GET_MODE_CLASS (GET_MODE (lhs))
7962 != GET_MODE_CLASS (GET_MODE (SUBREG_REG (lhs))))
7963 || (GET_MODE_SIZE (GET_MODE (lhs))
7964 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))))
7965 || VECTOR_MODE_P (GET_MODE (lhs))
7966 || GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))) > UNITS_PER_WORD
7967 /* Result might need to be truncated. Don't change mode if
7968 explicit truncation is needed. */
7969 || !TRULY_NOOP_TRUNCATION
7970 (GET_MODE_BITSIZE (GET_MODE (x)),
7971 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (lhs)))))
7972 return x;
7974 tem = simplify_gen_binary (code, GET_MODE (SUBREG_REG (lhs)),
7975 SUBREG_REG (lhs), SUBREG_REG (rhs));
7976 return gen_lowpart (GET_MODE (x), tem);
7978 default:
7979 return x;
7982 /* Set LHS and RHS to the inner operands (A and B in the example
7983 above) and set OTHER to the common operand (C in the example).
7984 There is only one way to do this unless the inner operation is
7985 commutative. */
7986 if (COMMUTATIVE_ARITH_P (lhs)
7987 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0)))
7988 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1);
7989 else if (COMMUTATIVE_ARITH_P (lhs)
7990 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1)))
7991 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0);
7992 else if (COMMUTATIVE_ARITH_P (lhs)
7993 && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0)))
7994 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1);
7995 else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1)))
7996 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0);
7997 else
7998 return x;
8000 /* Form the new inner operation, seeing if it simplifies first. */
8001 tem = simplify_gen_binary (code, GET_MODE (x), lhs, rhs);
8003 /* There is one exception to the general way of distributing:
8004 (a | c) ^ (b | c) -> (a ^ b) & ~c */
8005 if (code == XOR && inner_code == IOR)
8007 inner_code = AND;
8008 other = simplify_gen_unary (NOT, GET_MODE (x), other, GET_MODE (x));
8011 /* We may be able to continuing distributing the result, so call
8012 ourselves recursively on the inner operation before forming the
8013 outer operation, which we return. */
8014 return simplify_gen_binary (inner_code, GET_MODE (x),
8015 apply_distributive_law (tem), other);
8018 /* See if X is of the form (* (+ A B) C), and if so convert to
8019 (+ (* A C) (* B C)) and try to simplify.
8021 Most of the time, this results in no change. However, if some of
8022 the operands are the same or inverses of each other, simplifications
8023 will result.
8025 For example, (and (ior A B) (not B)) can occur as the result of
8026 expanding a bit field assignment. When we apply the distributive
8027 law to this, we get (ior (and (A (not B))) (and (B (not B)))),
8028 which then simplifies to (and (A (not B))).
8030 Note that no checks happen on the validity of applying the inverse
8031 distributive law. This is pointless since we can do it in the
8032 few places where this routine is called.
8034 N is the index of the term that is decomposed (the arithmetic operation,
8035 i.e. (+ A B) in the first example above). !N is the index of the term that
8036 is distributed, i.e. of C in the first example above. */
8037 static rtx
8038 distribute_and_simplify_rtx (rtx x, int n)
8040 enum machine_mode mode;
8041 enum rtx_code outer_code, inner_code;
8042 rtx decomposed, distributed, inner_op0, inner_op1, new_op0, new_op1, tmp;
8044 decomposed = XEXP (x, n);
8045 if (!ARITHMETIC_P (decomposed))
8046 return NULL_RTX;
8048 mode = GET_MODE (x);
8049 outer_code = GET_CODE (x);
8050 distributed = XEXP (x, !n);
8052 inner_code = GET_CODE (decomposed);
8053 inner_op0 = XEXP (decomposed, 0);
8054 inner_op1 = XEXP (decomposed, 1);
8056 /* Special case (and (xor B C) (not A)), which is equivalent to
8057 (xor (ior A B) (ior A C)) */
8058 if (outer_code == AND && inner_code == XOR && GET_CODE (distributed) == NOT)
8060 distributed = XEXP (distributed, 0);
8061 outer_code = IOR;
8064 if (n == 0)
8066 /* Distribute the second term. */
8067 new_op0 = simplify_gen_binary (outer_code, mode, inner_op0, distributed);
8068 new_op1 = simplify_gen_binary (outer_code, mode, inner_op1, distributed);
8070 else
8072 /* Distribute the first term. */
8073 new_op0 = simplify_gen_binary (outer_code, mode, distributed, inner_op0);
8074 new_op1 = simplify_gen_binary (outer_code, mode, distributed, inner_op1);
8077 tmp = apply_distributive_law (simplify_gen_binary (inner_code, mode,
8078 new_op0, new_op1));
8079 if (GET_CODE (tmp) != outer_code
8080 && rtx_cost (tmp, SET) < rtx_cost (x, SET))
8081 return tmp;
8083 return NULL_RTX;
8086 /* Simplify a logical `and' of VAROP with the constant CONSTOP, to be done
8087 in MODE. Return an equivalent form, if different from (and VAROP
8088 (const_int CONSTOP)). Otherwise, return NULL_RTX. */
8090 static rtx
8091 simplify_and_const_int_1 (enum machine_mode mode, rtx varop,
8092 unsigned HOST_WIDE_INT constop)
8094 unsigned HOST_WIDE_INT nonzero;
8095 unsigned HOST_WIDE_INT orig_constop;
8096 rtx orig_varop;
8097 int i;
8099 orig_varop = varop;
8100 orig_constop = constop;
8101 if (GET_CODE (varop) == CLOBBER)
8102 return NULL_RTX;
8104 /* Simplify VAROP knowing that we will be only looking at some of the
8105 bits in it.
8107 Note by passing in CONSTOP, we guarantee that the bits not set in
8108 CONSTOP are not significant and will never be examined. We must
8109 ensure that is the case by explicitly masking out those bits
8110 before returning. */
8111 varop = force_to_mode (varop, mode, constop, 0);
8113 /* If VAROP is a CLOBBER, we will fail so return it. */
8114 if (GET_CODE (varop) == CLOBBER)
8115 return varop;
8117 /* If VAROP is a CONST_INT, then we need to apply the mask in CONSTOP
8118 to VAROP and return the new constant. */
8119 if (GET_CODE (varop) == CONST_INT)
8120 return gen_int_mode (INTVAL (varop) & constop, mode);
8122 /* See what bits may be nonzero in VAROP. Unlike the general case of
8123 a call to nonzero_bits, here we don't care about bits outside
8124 MODE. */
8126 nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
8128 /* Turn off all bits in the constant that are known to already be zero.
8129 Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS
8130 which is tested below. */
8132 constop &= nonzero;
8134 /* If we don't have any bits left, return zero. */
8135 if (constop == 0)
8136 return const0_rtx;
8138 /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is
8139 a power of two, we can replace this with an ASHIFT. */
8140 if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), mode) == 1
8141 && (i = exact_log2 (constop)) >= 0)
8142 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i);
8144 /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR
8145 or XOR, then try to apply the distributive law. This may eliminate
8146 operations if either branch can be simplified because of the AND.
8147 It may also make some cases more complex, but those cases probably
8148 won't match a pattern either with or without this. */
8150 if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR)
8151 return
8152 gen_lowpart
8153 (mode,
8154 apply_distributive_law
8155 (simplify_gen_binary (GET_CODE (varop), GET_MODE (varop),
8156 simplify_and_const_int (NULL_RTX,
8157 GET_MODE (varop),
8158 XEXP (varop, 0),
8159 constop),
8160 simplify_and_const_int (NULL_RTX,
8161 GET_MODE (varop),
8162 XEXP (varop, 1),
8163 constop))));
8165 /* If VAROP is PLUS, and the constant is a mask of low bits, distribute
8166 the AND and see if one of the operands simplifies to zero. If so, we
8167 may eliminate it. */
8169 if (GET_CODE (varop) == PLUS
8170 && exact_log2 (constop + 1) >= 0)
8172 rtx o0, o1;
8174 o0 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 0), constop);
8175 o1 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 1), constop);
8176 if (o0 == const0_rtx)
8177 return o1;
8178 if (o1 == const0_rtx)
8179 return o0;
8182 /* Make a SUBREG if necessary. If we can't make it, fail. */
8183 varop = gen_lowpart (mode, varop);
8184 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
8185 return NULL_RTX;
8187 /* If we are only masking insignificant bits, return VAROP. */
8188 if (constop == nonzero)
8189 return varop;
8191 if (varop == orig_varop && constop == orig_constop)
8192 return NULL_RTX;
8194 /* Otherwise, return an AND. */
8195 constop = trunc_int_for_mode (constop, mode);
8196 return simplify_gen_binary (AND, mode, varop, GEN_INT (constop));
8200 /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done
8201 in MODE.
8203 Return an equivalent form, if different from X. Otherwise, return X. If
8204 X is zero, we are to always construct the equivalent form. */
8206 static rtx
8207 simplify_and_const_int (rtx x, enum machine_mode mode, rtx varop,
8208 unsigned HOST_WIDE_INT constop)
8210 rtx tem = simplify_and_const_int_1 (mode, varop, constop);
8211 if (tem)
8212 return tem;
8214 if (!x)
8215 x = simplify_gen_binary (AND, GET_MODE (varop), varop, GEN_INT (constop));
8216 if (GET_MODE (x) != mode)
8217 x = gen_lowpart (mode, x);
8218 return x;
8221 /* Given a REG, X, compute which bits in X can be nonzero.
8222 We don't care about bits outside of those defined in MODE.
8224 For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is
8225 a shift, AND, or zero_extract, we can do better. */
8227 static rtx
8228 reg_nonzero_bits_for_combine (rtx x, enum machine_mode mode,
8229 rtx known_x ATTRIBUTE_UNUSED,
8230 enum machine_mode known_mode ATTRIBUTE_UNUSED,
8231 unsigned HOST_WIDE_INT known_ret ATTRIBUTE_UNUSED,
8232 unsigned HOST_WIDE_INT *nonzero)
8234 rtx tem;
8236 /* If X is a register whose nonzero bits value is current, use it.
8237 Otherwise, if X is a register whose value we can find, use that
8238 value. Otherwise, use the previously-computed global nonzero bits
8239 for this register. */
8241 if (reg_stat[REGNO (x)].last_set_value != 0
8242 && (reg_stat[REGNO (x)].last_set_mode == mode
8243 || (GET_MODE_CLASS (reg_stat[REGNO (x)].last_set_mode) == MODE_INT
8244 && GET_MODE_CLASS (mode) == MODE_INT))
8245 && (reg_stat[REGNO (x)].last_set_label == label_tick
8246 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8247 && REG_N_SETS (REGNO (x)) == 1
8248 && ! REGNO_REG_SET_P
8249 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
8250 REGNO (x))))
8251 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8253 *nonzero &= reg_stat[REGNO (x)].last_set_nonzero_bits;
8254 return NULL;
8257 tem = get_last_value (x);
8259 if (tem)
8261 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
8262 /* If X is narrower than MODE and TEM is a non-negative
8263 constant that would appear negative in the mode of X,
8264 sign-extend it for use in reg_nonzero_bits because some
8265 machines (maybe most) will actually do the sign-extension
8266 and this is the conservative approach.
8268 ??? For 2.5, try to tighten up the MD files in this regard
8269 instead of this kludge. */
8271 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode)
8272 && GET_CODE (tem) == CONST_INT
8273 && INTVAL (tem) > 0
8274 && 0 != (INTVAL (tem)
8275 & ((HOST_WIDE_INT) 1
8276 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
8277 tem = GEN_INT (INTVAL (tem)
8278 | ((HOST_WIDE_INT) (-1)
8279 << GET_MODE_BITSIZE (GET_MODE (x))));
8280 #endif
8281 return tem;
8283 else if (nonzero_sign_valid && reg_stat[REGNO (x)].nonzero_bits)
8285 unsigned HOST_WIDE_INT mask = reg_stat[REGNO (x)].nonzero_bits;
8287 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode))
8288 /* We don't know anything about the upper bits. */
8289 mask |= GET_MODE_MASK (mode) ^ GET_MODE_MASK (GET_MODE (x));
8290 *nonzero &= mask;
8293 return NULL;
8296 /* Return the number of bits at the high-order end of X that are known to
8297 be equal to the sign bit. X will be used in mode MODE; if MODE is
8298 VOIDmode, X will be used in its own mode. The returned value will always
8299 be between 1 and the number of bits in MODE. */
8301 static rtx
8302 reg_num_sign_bit_copies_for_combine (rtx x, enum machine_mode mode,
8303 rtx known_x ATTRIBUTE_UNUSED,
8304 enum machine_mode known_mode
8305 ATTRIBUTE_UNUSED,
8306 unsigned int known_ret ATTRIBUTE_UNUSED,
8307 unsigned int *result)
8309 rtx tem;
8311 if (reg_stat[REGNO (x)].last_set_value != 0
8312 && reg_stat[REGNO (x)].last_set_mode == mode
8313 && (reg_stat[REGNO (x)].last_set_label == label_tick
8314 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8315 && REG_N_SETS (REGNO (x)) == 1
8316 && ! REGNO_REG_SET_P
8317 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
8318 REGNO (x))))
8319 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8321 *result = reg_stat[REGNO (x)].last_set_sign_bit_copies;
8322 return NULL;
8325 tem = get_last_value (x);
8326 if (tem != 0)
8327 return tem;
8329 if (nonzero_sign_valid && reg_stat[REGNO (x)].sign_bit_copies != 0
8330 && GET_MODE_BITSIZE (GET_MODE (x)) == GET_MODE_BITSIZE (mode))
8331 *result = reg_stat[REGNO (x)].sign_bit_copies;
8333 return NULL;
8336 /* Return the number of "extended" bits there are in X, when interpreted
8337 as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For
8338 unsigned quantities, this is the number of high-order zero bits.
8339 For signed quantities, this is the number of copies of the sign bit
8340 minus 1. In both case, this function returns the number of "spare"
8341 bits. For example, if two quantities for which this function returns
8342 at least 1 are added, the addition is known not to overflow.
8344 This function will always return 0 unless called during combine, which
8345 implies that it must be called from a define_split. */
8347 unsigned int
8348 extended_count (rtx x, enum machine_mode mode, int unsignedp)
8350 if (nonzero_sign_valid == 0)
8351 return 0;
8353 return (unsignedp
8354 ? (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
8355 ? (unsigned int) (GET_MODE_BITSIZE (mode) - 1
8356 - floor_log2 (nonzero_bits (x, mode)))
8357 : 0)
8358 : num_sign_bit_copies (x, mode) - 1);
8361 /* This function is called from `simplify_shift_const' to merge two
8362 outer operations. Specifically, we have already found that we need
8363 to perform operation *POP0 with constant *PCONST0 at the outermost
8364 position. We would now like to also perform OP1 with constant CONST1
8365 (with *POP0 being done last).
8367 Return 1 if we can do the operation and update *POP0 and *PCONST0 with
8368 the resulting operation. *PCOMP_P is set to 1 if we would need to
8369 complement the innermost operand, otherwise it is unchanged.
8371 MODE is the mode in which the operation will be done. No bits outside
8372 the width of this mode matter. It is assumed that the width of this mode
8373 is smaller than or equal to HOST_BITS_PER_WIDE_INT.
8375 If *POP0 or OP1 are UNKNOWN, it means no operation is required. Only NEG, PLUS,
8376 IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper
8377 result is simply *PCONST0.
8379 If the resulting operation cannot be expressed as one operation, we
8380 return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */
8382 static int
8383 merge_outer_ops (enum rtx_code *pop0, HOST_WIDE_INT *pconst0, enum rtx_code op1, HOST_WIDE_INT const1, enum machine_mode mode, int *pcomp_p)
8385 enum rtx_code op0 = *pop0;
8386 HOST_WIDE_INT const0 = *pconst0;
8388 const0 &= GET_MODE_MASK (mode);
8389 const1 &= GET_MODE_MASK (mode);
8391 /* If OP0 is an AND, clear unimportant bits in CONST1. */
8392 if (op0 == AND)
8393 const1 &= const0;
8395 /* If OP0 or OP1 is UNKNOWN, this is easy. Similarly if they are the same or
8396 if OP0 is SET. */
8398 if (op1 == UNKNOWN || op0 == SET)
8399 return 1;
8401 else if (op0 == UNKNOWN)
8402 op0 = op1, const0 = const1;
8404 else if (op0 == op1)
8406 switch (op0)
8408 case AND:
8409 const0 &= const1;
8410 break;
8411 case IOR:
8412 const0 |= const1;
8413 break;
8414 case XOR:
8415 const0 ^= const1;
8416 break;
8417 case PLUS:
8418 const0 += const1;
8419 break;
8420 case NEG:
8421 op0 = UNKNOWN;
8422 break;
8423 default:
8424 break;
8428 /* Otherwise, if either is a PLUS or NEG, we can't do anything. */
8429 else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG)
8430 return 0;
8432 /* If the two constants aren't the same, we can't do anything. The
8433 remaining six cases can all be done. */
8434 else if (const0 != const1)
8435 return 0;
8437 else
8438 switch (op0)
8440 case IOR:
8441 if (op1 == AND)
8442 /* (a & b) | b == b */
8443 op0 = SET;
8444 else /* op1 == XOR */
8445 /* (a ^ b) | b == a | b */
8447 break;
8449 case XOR:
8450 if (op1 == AND)
8451 /* (a & b) ^ b == (~a) & b */
8452 op0 = AND, *pcomp_p = 1;
8453 else /* op1 == IOR */
8454 /* (a | b) ^ b == a & ~b */
8455 op0 = AND, const0 = ~const0;
8456 break;
8458 case AND:
8459 if (op1 == IOR)
8460 /* (a | b) & b == b */
8461 op0 = SET;
8462 else /* op1 == XOR */
8463 /* (a ^ b) & b) == (~a) & b */
8464 *pcomp_p = 1;
8465 break;
8466 default:
8467 break;
8470 /* Check for NO-OP cases. */
8471 const0 &= GET_MODE_MASK (mode);
8472 if (const0 == 0
8473 && (op0 == IOR || op0 == XOR || op0 == PLUS))
8474 op0 = UNKNOWN;
8475 else if (const0 == 0 && op0 == AND)
8476 op0 = SET;
8477 else if ((unsigned HOST_WIDE_INT) const0 == GET_MODE_MASK (mode)
8478 && op0 == AND)
8479 op0 = UNKNOWN;
8481 /* ??? Slightly redundant with the above mask, but not entirely.
8482 Moving this above means we'd have to sign-extend the mode mask
8483 for the final test. */
8484 const0 = trunc_int_for_mode (const0, mode);
8486 *pop0 = op0;
8487 *pconst0 = const0;
8489 return 1;
8492 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
8493 The result of the shift is RESULT_MODE. Return NULL_RTX if we cannot
8494 simplify it. Otherwise, return a simplified value.
8496 The shift is normally computed in the widest mode we find in VAROP, as
8497 long as it isn't a different number of words than RESULT_MODE. Exceptions
8498 are ASHIFTRT and ROTATE, which are always done in their original mode. */
8500 static rtx
8501 simplify_shift_const_1 (enum rtx_code code, enum machine_mode result_mode,
8502 rtx varop, int orig_count)
8504 enum rtx_code orig_code = code;
8505 rtx orig_varop = varop;
8506 int count;
8507 enum machine_mode mode = result_mode;
8508 enum machine_mode shift_mode, tmode;
8509 unsigned int mode_words
8510 = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
8511 /* We form (outer_op (code varop count) (outer_const)). */
8512 enum rtx_code outer_op = UNKNOWN;
8513 HOST_WIDE_INT outer_const = 0;
8514 int complement_p = 0;
8515 rtx new, x;
8517 /* Make sure and truncate the "natural" shift on the way in. We don't
8518 want to do this inside the loop as it makes it more difficult to
8519 combine shifts. */
8520 if (SHIFT_COUNT_TRUNCATED)
8521 orig_count &= GET_MODE_BITSIZE (mode) - 1;
8523 /* If we were given an invalid count, don't do anything except exactly
8524 what was requested. */
8526 if (orig_count < 0 || orig_count >= (int) GET_MODE_BITSIZE (mode))
8527 return NULL_RTX;
8529 count = orig_count;
8531 /* Unless one of the branches of the `if' in this loop does a `continue',
8532 we will `break' the loop after the `if'. */
8534 while (count != 0)
8536 /* If we have an operand of (clobber (const_int 0)), fail. */
8537 if (GET_CODE (varop) == CLOBBER)
8538 return NULL_RTX;
8540 /* If we discovered we had to complement VAROP, leave. Making a NOT
8541 here would cause an infinite loop. */
8542 if (complement_p)
8543 break;
8545 /* Convert ROTATERT to ROTATE. */
8546 if (code == ROTATERT)
8548 unsigned int bitsize = GET_MODE_BITSIZE (result_mode);;
8549 code = ROTATE;
8550 if (VECTOR_MODE_P (result_mode))
8551 count = bitsize / GET_MODE_NUNITS (result_mode) - count;
8552 else
8553 count = bitsize - count;
8556 /* We need to determine what mode we will do the shift in. If the
8557 shift is a right shift or a ROTATE, we must always do it in the mode
8558 it was originally done in. Otherwise, we can do it in MODE, the
8559 widest mode encountered. */
8560 shift_mode
8561 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
8562 ? result_mode : mode);
8564 /* Handle cases where the count is greater than the size of the mode
8565 minus 1. For ASHIFT, use the size minus one as the count (this can
8566 occur when simplifying (lshiftrt (ashiftrt ..))). For rotates,
8567 take the count modulo the size. For other shifts, the result is
8568 zero.
8570 Since these shifts are being produced by the compiler by combining
8571 multiple operations, each of which are defined, we know what the
8572 result is supposed to be. */
8574 if (count > (GET_MODE_BITSIZE (shift_mode) - 1))
8576 if (code == ASHIFTRT)
8577 count = GET_MODE_BITSIZE (shift_mode) - 1;
8578 else if (code == ROTATE || code == ROTATERT)
8579 count %= GET_MODE_BITSIZE (shift_mode);
8580 else
8582 /* We can't simply return zero because there may be an
8583 outer op. */
8584 varop = const0_rtx;
8585 count = 0;
8586 break;
8590 /* An arithmetic right shift of a quantity known to be -1 or 0
8591 is a no-op. */
8592 if (code == ASHIFTRT
8593 && (num_sign_bit_copies (varop, shift_mode)
8594 == GET_MODE_BITSIZE (shift_mode)))
8596 count = 0;
8597 break;
8600 /* If we are doing an arithmetic right shift and discarding all but
8601 the sign bit copies, this is equivalent to doing a shift by the
8602 bitsize minus one. Convert it into that shift because it will often
8603 allow other simplifications. */
8605 if (code == ASHIFTRT
8606 && (count + num_sign_bit_copies (varop, shift_mode)
8607 >= GET_MODE_BITSIZE (shift_mode)))
8608 count = GET_MODE_BITSIZE (shift_mode) - 1;
8610 /* We simplify the tests below and elsewhere by converting
8611 ASHIFTRT to LSHIFTRT if we know the sign bit is clear.
8612 `make_compound_operation' will convert it to an ASHIFTRT for
8613 those machines (such as VAX) that don't have an LSHIFTRT. */
8614 if (GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8615 && code == ASHIFTRT
8616 && ((nonzero_bits (varop, shift_mode)
8617 & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (shift_mode) - 1)))
8618 == 0))
8619 code = LSHIFTRT;
8621 if (code == LSHIFTRT
8622 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8623 && !(nonzero_bits (varop, shift_mode) >> count))
8624 varop = const0_rtx;
8625 if (code == ASHIFT
8626 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8627 && !((nonzero_bits (varop, shift_mode) << count)
8628 & GET_MODE_MASK (shift_mode)))
8629 varop = const0_rtx;
8631 switch (GET_CODE (varop))
8633 case SIGN_EXTEND:
8634 case ZERO_EXTEND:
8635 case SIGN_EXTRACT:
8636 case ZERO_EXTRACT:
8637 new = expand_compound_operation (varop);
8638 if (new != varop)
8640 varop = new;
8641 continue;
8643 break;
8645 case MEM:
8646 /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH
8647 minus the width of a smaller mode, we can do this with a
8648 SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */
8649 if ((code == ASHIFTRT || code == LSHIFTRT)
8650 && ! mode_dependent_address_p (XEXP (varop, 0))
8651 && ! MEM_VOLATILE_P (varop)
8652 && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count,
8653 MODE_INT, 1)) != BLKmode)
8655 new = adjust_address_nv (varop, tmode,
8656 BYTES_BIG_ENDIAN ? 0
8657 : count / BITS_PER_UNIT);
8659 varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND
8660 : ZERO_EXTEND, mode, new);
8661 count = 0;
8662 continue;
8664 break;
8666 case SUBREG:
8667 /* If VAROP is a SUBREG, strip it as long as the inner operand has
8668 the same number of words as what we've seen so far. Then store
8669 the widest mode in MODE. */
8670 if (subreg_lowpart_p (varop)
8671 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8672 > GET_MODE_SIZE (GET_MODE (varop)))
8673 && (unsigned int) ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8674 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
8675 == mode_words)
8677 varop = SUBREG_REG (varop);
8678 if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode))
8679 mode = GET_MODE (varop);
8680 continue;
8682 break;
8684 case MULT:
8685 /* Some machines use MULT instead of ASHIFT because MULT
8686 is cheaper. But it is still better on those machines to
8687 merge two shifts into one. */
8688 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8689 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8691 varop
8692 = simplify_gen_binary (ASHIFT, GET_MODE (varop),
8693 XEXP (varop, 0),
8694 GEN_INT (exact_log2 (
8695 INTVAL (XEXP (varop, 1)))));
8696 continue;
8698 break;
8700 case UDIV:
8701 /* Similar, for when divides are cheaper. */
8702 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8703 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8705 varop
8706 = simplify_gen_binary (LSHIFTRT, GET_MODE (varop),
8707 XEXP (varop, 0),
8708 GEN_INT (exact_log2 (
8709 INTVAL (XEXP (varop, 1)))));
8710 continue;
8712 break;
8714 case ASHIFTRT:
8715 /* If we are extracting just the sign bit of an arithmetic
8716 right shift, that shift is not needed. However, the sign
8717 bit of a wider mode may be different from what would be
8718 interpreted as the sign bit in a narrower mode, so, if
8719 the result is narrower, don't discard the shift. */
8720 if (code == LSHIFTRT
8721 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8722 && (GET_MODE_BITSIZE (result_mode)
8723 >= GET_MODE_BITSIZE (GET_MODE (varop))))
8725 varop = XEXP (varop, 0);
8726 continue;
8729 /* ... fall through ... */
8731 case LSHIFTRT:
8732 case ASHIFT:
8733 case ROTATE:
8734 /* Here we have two nested shifts. The result is usually the
8735 AND of a new shift with a mask. We compute the result below. */
8736 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8737 && INTVAL (XEXP (varop, 1)) >= 0
8738 && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop))
8739 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
8740 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
8742 enum rtx_code first_code = GET_CODE (varop);
8743 unsigned int first_count = INTVAL (XEXP (varop, 1));
8744 unsigned HOST_WIDE_INT mask;
8745 rtx mask_rtx;
8747 /* We have one common special case. We can't do any merging if
8748 the inner code is an ASHIFTRT of a smaller mode. However, if
8749 we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2)
8750 with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2),
8751 we can convert it to
8752 (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0 C2) C3) C1).
8753 This simplifies certain SIGN_EXTEND operations. */
8754 if (code == ASHIFT && first_code == ASHIFTRT
8755 && count == (GET_MODE_BITSIZE (result_mode)
8756 - GET_MODE_BITSIZE (GET_MODE (varop))))
8758 /* C3 has the low-order C1 bits zero. */
8760 mask = (GET_MODE_MASK (mode)
8761 & ~(((HOST_WIDE_INT) 1 << first_count) - 1));
8763 varop = simplify_and_const_int (NULL_RTX, result_mode,
8764 XEXP (varop, 0), mask);
8765 varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode,
8766 varop, count);
8767 count = first_count;
8768 code = ASHIFTRT;
8769 continue;
8772 /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more
8773 than C1 high-order bits equal to the sign bit, we can convert
8774 this to either an ASHIFT or an ASHIFTRT depending on the
8775 two counts.
8777 We cannot do this if VAROP's mode is not SHIFT_MODE. */
8779 if (code == ASHIFTRT && first_code == ASHIFT
8780 && GET_MODE (varop) == shift_mode
8781 && (num_sign_bit_copies (XEXP (varop, 0), shift_mode)
8782 > first_count))
8784 varop = XEXP (varop, 0);
8785 count -= first_count;
8786 if (count < 0)
8788 count = -count;
8789 code = ASHIFT;
8792 continue;
8795 /* There are some cases we can't do. If CODE is ASHIFTRT,
8796 we can only do this if FIRST_CODE is also ASHIFTRT.
8798 We can't do the case when CODE is ROTATE and FIRST_CODE is
8799 ASHIFTRT.
8801 If the mode of this shift is not the mode of the outer shift,
8802 we can't do this if either shift is a right shift or ROTATE.
8804 Finally, we can't do any of these if the mode is too wide
8805 unless the codes are the same.
8807 Handle the case where the shift codes are the same
8808 first. */
8810 if (code == first_code)
8812 if (GET_MODE (varop) != result_mode
8813 && (code == ASHIFTRT || code == LSHIFTRT
8814 || code == ROTATE))
8815 break;
8817 count += first_count;
8818 varop = XEXP (varop, 0);
8819 continue;
8822 if (code == ASHIFTRT
8823 || (code == ROTATE && first_code == ASHIFTRT)
8824 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT
8825 || (GET_MODE (varop) != result_mode
8826 && (first_code == ASHIFTRT || first_code == LSHIFTRT
8827 || first_code == ROTATE
8828 || code == ROTATE)))
8829 break;
8831 /* To compute the mask to apply after the shift, shift the
8832 nonzero bits of the inner shift the same way the
8833 outer shift will. */
8835 mask_rtx = GEN_INT (nonzero_bits (varop, GET_MODE (varop)));
8837 mask_rtx
8838 = simplify_const_binary_operation (code, result_mode, mask_rtx,
8839 GEN_INT (count));
8841 /* Give up if we can't compute an outer operation to use. */
8842 if (mask_rtx == 0
8843 || GET_CODE (mask_rtx) != CONST_INT
8844 || ! merge_outer_ops (&outer_op, &outer_const, AND,
8845 INTVAL (mask_rtx),
8846 result_mode, &complement_p))
8847 break;
8849 /* If the shifts are in the same direction, we add the
8850 counts. Otherwise, we subtract them. */
8851 if ((code == ASHIFTRT || code == LSHIFTRT)
8852 == (first_code == ASHIFTRT || first_code == LSHIFTRT))
8853 count += first_count;
8854 else
8855 count -= first_count;
8857 /* If COUNT is positive, the new shift is usually CODE,
8858 except for the two exceptions below, in which case it is
8859 FIRST_CODE. If the count is negative, FIRST_CODE should
8860 always be used */
8861 if (count > 0
8862 && ((first_code == ROTATE && code == ASHIFT)
8863 || (first_code == ASHIFTRT && code == LSHIFTRT)))
8864 code = first_code;
8865 else if (count < 0)
8866 code = first_code, count = -count;
8868 varop = XEXP (varop, 0);
8869 continue;
8872 /* If we have (A << B << C) for any shift, we can convert this to
8873 (A << C << B). This wins if A is a constant. Only try this if
8874 B is not a constant. */
8876 else if (GET_CODE (varop) == code
8877 && GET_CODE (XEXP (varop, 0)) == CONST_INT
8878 && GET_CODE (XEXP (varop, 1)) != CONST_INT)
8880 rtx new = simplify_const_binary_operation (code, mode,
8881 XEXP (varop, 0),
8882 GEN_INT (count));
8883 varop = gen_rtx_fmt_ee (code, mode, new, XEXP (varop, 1));
8884 count = 0;
8885 continue;
8887 break;
8889 case NOT:
8890 /* Make this fit the case below. */
8891 varop = gen_rtx_XOR (mode, XEXP (varop, 0),
8892 GEN_INT (GET_MODE_MASK (mode)));
8893 continue;
8895 case IOR:
8896 case AND:
8897 case XOR:
8898 /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C)
8899 with C the size of VAROP - 1 and the shift is logical if
8900 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
8901 we have an (le X 0) operation. If we have an arithmetic shift
8902 and STORE_FLAG_VALUE is 1 or we have a logical shift with
8903 STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */
8905 if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS
8906 && XEXP (XEXP (varop, 0), 1) == constm1_rtx
8907 && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
8908 && (code == LSHIFTRT || code == ASHIFTRT)
8909 && count == (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
8910 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
8912 count = 0;
8913 varop = gen_rtx_LE (GET_MODE (varop), XEXP (varop, 1),
8914 const0_rtx);
8916 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
8917 varop = gen_rtx_NEG (GET_MODE (varop), varop);
8919 continue;
8922 /* If we have (shift (logical)), move the logical to the outside
8923 to allow it to possibly combine with another logical and the
8924 shift to combine with another shift. This also canonicalizes to
8925 what a ZERO_EXTRACT looks like. Also, some machines have
8926 (and (shift)) insns. */
8928 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8929 /* We can't do this if we have (ashiftrt (xor)) and the
8930 constant has its sign bit set in shift_mode. */
8931 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
8932 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
8933 shift_mode))
8934 && (new = simplify_const_binary_operation (code, result_mode,
8935 XEXP (varop, 1),
8936 GEN_INT (count))) != 0
8937 && GET_CODE (new) == CONST_INT
8938 && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop),
8939 INTVAL (new), result_mode, &complement_p))
8941 varop = XEXP (varop, 0);
8942 continue;
8945 /* If we can't do that, try to simplify the shift in each arm of the
8946 logical expression, make a new logical expression, and apply
8947 the inverse distributive law. This also can't be done
8948 for some (ashiftrt (xor)). */
8949 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8950 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
8951 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
8952 shift_mode)))
8954 rtx lhs = simplify_shift_const (NULL_RTX, code, shift_mode,
8955 XEXP (varop, 0), count);
8956 rtx rhs = simplify_shift_const (NULL_RTX, code, shift_mode,
8957 XEXP (varop, 1), count);
8959 varop = simplify_gen_binary (GET_CODE (varop), shift_mode,
8960 lhs, rhs);
8961 varop = apply_distributive_law (varop);
8963 count = 0;
8964 continue;
8966 break;
8968 case EQ:
8969 /* Convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE
8970 says that the sign bit can be tested, FOO has mode MODE, C is
8971 GET_MODE_BITSIZE (MODE) - 1, and FOO has only its low-order bit
8972 that may be nonzero. */
8973 if (code == LSHIFTRT
8974 && XEXP (varop, 1) == const0_rtx
8975 && GET_MODE (XEXP (varop, 0)) == result_mode
8976 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8977 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
8978 && STORE_FLAG_VALUE == -1
8979 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
8980 && merge_outer_ops (&outer_op, &outer_const, XOR,
8981 (HOST_WIDE_INT) 1, result_mode,
8982 &complement_p))
8984 varop = XEXP (varop, 0);
8985 count = 0;
8986 continue;
8988 break;
8990 case NEG:
8991 /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less
8992 than the number of bits in the mode is equivalent to A. */
8993 if (code == LSHIFTRT
8994 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8995 && nonzero_bits (XEXP (varop, 0), result_mode) == 1)
8997 varop = XEXP (varop, 0);
8998 count = 0;
8999 continue;
9002 /* NEG commutes with ASHIFT since it is multiplication. Move the
9003 NEG outside to allow shifts to combine. */
9004 if (code == ASHIFT
9005 && merge_outer_ops (&outer_op, &outer_const, NEG,
9006 (HOST_WIDE_INT) 0, result_mode,
9007 &complement_p))
9009 varop = XEXP (varop, 0);
9010 continue;
9012 break;
9014 case PLUS:
9015 /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C
9016 is one less than the number of bits in the mode is
9017 equivalent to (xor A 1). */
9018 if (code == LSHIFTRT
9019 && count == (GET_MODE_BITSIZE (result_mode) - 1)
9020 && XEXP (varop, 1) == constm1_rtx
9021 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
9022 && merge_outer_ops (&outer_op, &outer_const, XOR,
9023 (HOST_WIDE_INT) 1, result_mode,
9024 &complement_p))
9026 count = 0;
9027 varop = XEXP (varop, 0);
9028 continue;
9031 /* If we have (xshiftrt (plus FOO BAR) C), and the only bits
9032 that might be nonzero in BAR are those being shifted out and those
9033 bits are known zero in FOO, we can replace the PLUS with FOO.
9034 Similarly in the other operand order. This code occurs when
9035 we are computing the size of a variable-size array. */
9037 if ((code == ASHIFTRT || code == LSHIFTRT)
9038 && count < HOST_BITS_PER_WIDE_INT
9039 && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0
9040 && (nonzero_bits (XEXP (varop, 1), result_mode)
9041 & nonzero_bits (XEXP (varop, 0), result_mode)) == 0)
9043 varop = XEXP (varop, 0);
9044 continue;
9046 else if ((code == ASHIFTRT || code == LSHIFTRT)
9047 && count < HOST_BITS_PER_WIDE_INT
9048 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
9049 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9050 >> count)
9051 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9052 & nonzero_bits (XEXP (varop, 1),
9053 result_mode)))
9055 varop = XEXP (varop, 1);
9056 continue;
9059 /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */
9060 if (code == ASHIFT
9061 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9062 && (new = simplify_const_binary_operation (ASHIFT, result_mode,
9063 XEXP (varop, 1),
9064 GEN_INT (count))) != 0
9065 && GET_CODE (new) == CONST_INT
9066 && merge_outer_ops (&outer_op, &outer_const, PLUS,
9067 INTVAL (new), result_mode, &complement_p))
9069 varop = XEXP (varop, 0);
9070 continue;
9073 /* Check for 'PLUS signbit', which is the canonical form of 'XOR
9074 signbit', and attempt to change the PLUS to an XOR and move it to
9075 the outer operation as is done above in the AND/IOR/XOR case
9076 leg for shift(logical). See details in logical handling above
9077 for reasoning in doing so. */
9078 if (code == LSHIFTRT
9079 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9080 && mode_signbit_p (result_mode, XEXP (varop, 1))
9081 && (new = simplify_const_binary_operation (code, result_mode,
9082 XEXP (varop, 1),
9083 GEN_INT (count))) != 0
9084 && GET_CODE (new) == CONST_INT
9085 && merge_outer_ops (&outer_op, &outer_const, XOR,
9086 INTVAL (new), result_mode, &complement_p))
9088 varop = XEXP (varop, 0);
9089 continue;
9092 break;
9094 case MINUS:
9095 /* If we have (xshiftrt (minus (ashiftrt X C)) X) C)
9096 with C the size of VAROP - 1 and the shift is logical if
9097 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
9098 we have a (gt X 0) operation. If the shift is arithmetic with
9099 STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1,
9100 we have a (neg (gt X 0)) operation. */
9102 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
9103 && GET_CODE (XEXP (varop, 0)) == ASHIFTRT
9104 && count == (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
9105 && (code == LSHIFTRT || code == ASHIFTRT)
9106 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9107 && INTVAL (XEXP (XEXP (varop, 0), 1)) == count
9108 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
9110 count = 0;
9111 varop = gen_rtx_GT (GET_MODE (varop), XEXP (varop, 1),
9112 const0_rtx);
9114 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
9115 varop = gen_rtx_NEG (GET_MODE (varop), varop);
9117 continue;
9119 break;
9121 case TRUNCATE:
9122 /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt))
9123 if the truncate does not affect the value. */
9124 if (code == LSHIFTRT
9125 && GET_CODE (XEXP (varop, 0)) == LSHIFTRT
9126 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9127 && (INTVAL (XEXP (XEXP (varop, 0), 1))
9128 >= (GET_MODE_BITSIZE (GET_MODE (XEXP (varop, 0)))
9129 - GET_MODE_BITSIZE (GET_MODE (varop)))))
9131 rtx varop_inner = XEXP (varop, 0);
9133 varop_inner
9134 = gen_rtx_LSHIFTRT (GET_MODE (varop_inner),
9135 XEXP (varop_inner, 0),
9136 GEN_INT
9137 (count + INTVAL (XEXP (varop_inner, 1))));
9138 varop = gen_rtx_TRUNCATE (GET_MODE (varop), varop_inner);
9139 count = 0;
9140 continue;
9142 break;
9144 default:
9145 break;
9148 break;
9151 /* We need to determine what mode to do the shift in. If the shift is
9152 a right shift or ROTATE, we must always do it in the mode it was
9153 originally done in. Otherwise, we can do it in MODE, the widest mode
9154 encountered. The code we care about is that of the shift that will
9155 actually be done, not the shift that was originally requested. */
9156 shift_mode
9157 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
9158 ? result_mode : mode);
9160 /* We have now finished analyzing the shift. The result should be
9161 a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If
9162 OUTER_OP is non-UNKNOWN, it is an operation that needs to be applied
9163 to the result of the shift. OUTER_CONST is the relevant constant,
9164 but we must turn off all bits turned off in the shift. */
9166 if (outer_op == UNKNOWN
9167 && orig_code == code && orig_count == count
9168 && varop == orig_varop
9169 && shift_mode == GET_MODE (varop))
9170 return NULL_RTX;
9172 /* Make a SUBREG if necessary. If we can't make it, fail. */
9173 varop = gen_lowpart (shift_mode, varop);
9174 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
9175 return NULL_RTX;
9177 /* If we have an outer operation and we just made a shift, it is
9178 possible that we could have simplified the shift were it not
9179 for the outer operation. So try to do the simplification
9180 recursively. */
9182 if (outer_op != UNKNOWN)
9183 x = simplify_shift_const_1 (code, shift_mode, varop, count);
9184 else
9185 x = NULL_RTX;
9187 if (x == NULL_RTX)
9188 x = simplify_gen_binary (code, shift_mode, varop, GEN_INT (count));
9190 /* If we were doing an LSHIFTRT in a wider mode than it was originally,
9191 turn off all the bits that the shift would have turned off. */
9192 if (orig_code == LSHIFTRT && result_mode != shift_mode)
9193 x = simplify_and_const_int (NULL_RTX, shift_mode, x,
9194 GET_MODE_MASK (result_mode) >> orig_count);
9196 /* Do the remainder of the processing in RESULT_MODE. */
9197 x = gen_lowpart (result_mode, x);
9199 /* If COMPLEMENT_P is set, we have to complement X before doing the outer
9200 operation. */
9201 if (complement_p)
9202 x = simplify_gen_unary (NOT, result_mode, x, result_mode);
9204 if (outer_op != UNKNOWN)
9206 if (GET_MODE_BITSIZE (result_mode) < HOST_BITS_PER_WIDE_INT)
9207 outer_const = trunc_int_for_mode (outer_const, result_mode);
9209 if (outer_op == AND)
9210 x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const);
9211 else if (outer_op == SET)
9212 /* This means that we have determined that the result is
9213 equivalent to a constant. This should be rare. */
9214 x = GEN_INT (outer_const);
9215 else if (GET_RTX_CLASS (outer_op) == RTX_UNARY)
9216 x = simplify_gen_unary (outer_op, result_mode, x, result_mode);
9217 else
9218 x = simplify_gen_binary (outer_op, result_mode, x,
9219 GEN_INT (outer_const));
9222 return x;
9225 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
9226 The result of the shift is RESULT_MODE. If we cannot simplify it,
9227 return X or, if it is NULL, synthesize the expression with
9228 simplify_gen_binary. Otherwise, return a simplified value.
9230 The shift is normally computed in the widest mode we find in VAROP, as
9231 long as it isn't a different number of words than RESULT_MODE. Exceptions
9232 are ASHIFTRT and ROTATE, which are always done in their original mode. */
9234 static rtx
9235 simplify_shift_const (rtx x, enum rtx_code code, enum machine_mode result_mode,
9236 rtx varop, int count)
9238 rtx tem = simplify_shift_const_1 (code, result_mode, varop, count);
9239 if (tem)
9240 return tem;
9242 if (!x)
9243 x = simplify_gen_binary (code, GET_MODE (varop), varop, GEN_INT (count));
9244 if (GET_MODE (x) != result_mode)
9245 x = gen_lowpart (result_mode, x);
9246 return x;
9250 /* Like recog, but we receive the address of a pointer to a new pattern.
9251 We try to match the rtx that the pointer points to.
9252 If that fails, we may try to modify or replace the pattern,
9253 storing the replacement into the same pointer object.
9255 Modifications include deletion or addition of CLOBBERs.
9257 PNOTES is a pointer to a location where any REG_UNUSED notes added for
9258 the CLOBBERs are placed.
9260 The value is the final insn code from the pattern ultimately matched,
9261 or -1. */
9263 static int
9264 recog_for_combine (rtx *pnewpat, rtx insn, rtx *pnotes)
9266 rtx pat = *pnewpat;
9267 int insn_code_number;
9268 int num_clobbers_to_add = 0;
9269 int i;
9270 rtx notes = 0;
9271 rtx old_notes, old_pat;
9273 /* If PAT is a PARALLEL, check to see if it contains the CLOBBER
9274 we use to indicate that something didn't match. If we find such a
9275 thing, force rejection. */
9276 if (GET_CODE (pat) == PARALLEL)
9277 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
9278 if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER
9279 && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx)
9280 return -1;
9282 old_pat = PATTERN (insn);
9283 old_notes = REG_NOTES (insn);
9284 PATTERN (insn) = pat;
9285 REG_NOTES (insn) = 0;
9287 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9289 /* If it isn't, there is the possibility that we previously had an insn
9290 that clobbered some register as a side effect, but the combined
9291 insn doesn't need to do that. So try once more without the clobbers
9292 unless this represents an ASM insn. */
9294 if (insn_code_number < 0 && ! check_asm_operands (pat)
9295 && GET_CODE (pat) == PARALLEL)
9297 int pos;
9299 for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++)
9300 if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER)
9302 if (i != pos)
9303 SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i));
9304 pos++;
9307 SUBST_INT (XVECLEN (pat, 0), pos);
9309 if (pos == 1)
9310 pat = XVECEXP (pat, 0, 0);
9312 PATTERN (insn) = pat;
9313 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9315 PATTERN (insn) = old_pat;
9316 REG_NOTES (insn) = old_notes;
9318 /* Recognize all noop sets, these will be killed by followup pass. */
9319 if (insn_code_number < 0 && GET_CODE (pat) == SET && set_noop_p (pat))
9320 insn_code_number = NOOP_MOVE_INSN_CODE, num_clobbers_to_add = 0;
9322 /* If we had any clobbers to add, make a new pattern than contains
9323 them. Then check to make sure that all of them are dead. */
9324 if (num_clobbers_to_add)
9326 rtx newpat = gen_rtx_PARALLEL (VOIDmode,
9327 rtvec_alloc (GET_CODE (pat) == PARALLEL
9328 ? (XVECLEN (pat, 0)
9329 + num_clobbers_to_add)
9330 : num_clobbers_to_add + 1));
9332 if (GET_CODE (pat) == PARALLEL)
9333 for (i = 0; i < XVECLEN (pat, 0); i++)
9334 XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i);
9335 else
9336 XVECEXP (newpat, 0, 0) = pat;
9338 add_clobbers (newpat, insn_code_number);
9340 for (i = XVECLEN (newpat, 0) - num_clobbers_to_add;
9341 i < XVECLEN (newpat, 0); i++)
9343 if (REG_P (XEXP (XVECEXP (newpat, 0, i), 0))
9344 && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn))
9345 return -1;
9346 notes = gen_rtx_EXPR_LIST (REG_UNUSED,
9347 XEXP (XVECEXP (newpat, 0, i), 0), notes);
9349 pat = newpat;
9352 *pnewpat = pat;
9353 *pnotes = notes;
9355 return insn_code_number;
9358 /* Like gen_lowpart_general but for use by combine. In combine it
9359 is not possible to create any new pseudoregs. However, it is
9360 safe to create invalid memory addresses, because combine will
9361 try to recognize them and all they will do is make the combine
9362 attempt fail.
9364 If for some reason this cannot do its job, an rtx
9365 (clobber (const_int 0)) is returned.
9366 An insn containing that will not be recognized. */
9368 static rtx
9369 gen_lowpart_for_combine (enum machine_mode omode, rtx x)
9371 enum machine_mode imode = GET_MODE (x);
9372 unsigned int osize = GET_MODE_SIZE (omode);
9373 unsigned int isize = GET_MODE_SIZE (imode);
9374 rtx result;
9376 if (omode == imode)
9377 return x;
9379 /* Return identity if this is a CONST or symbolic reference. */
9380 if (omode == Pmode
9381 && (GET_CODE (x) == CONST
9382 || GET_CODE (x) == SYMBOL_REF
9383 || GET_CODE (x) == LABEL_REF))
9384 return x;
9386 /* We can only support MODE being wider than a word if X is a
9387 constant integer or has a mode the same size. */
9388 if (GET_MODE_SIZE (omode) > UNITS_PER_WORD
9389 && ! ((imode == VOIDmode
9390 && (GET_CODE (x) == CONST_INT
9391 || GET_CODE (x) == CONST_DOUBLE))
9392 || isize == osize))
9393 goto fail;
9395 /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart
9396 won't know what to do. So we will strip off the SUBREG here and
9397 process normally. */
9398 if (GET_CODE (x) == SUBREG && MEM_P (SUBREG_REG (x)))
9400 x = SUBREG_REG (x);
9402 /* For use in case we fall down into the address adjustments
9403 further below, we need to adjust the known mode and size of
9404 x; imode and isize, since we just adjusted x. */
9405 imode = GET_MODE (x);
9407 if (imode == omode)
9408 return x;
9410 isize = GET_MODE_SIZE (imode);
9413 result = gen_lowpart_common (omode, x);
9415 #ifdef CANNOT_CHANGE_MODE_CLASS
9416 if (result != 0 && GET_CODE (result) == SUBREG)
9417 record_subregs_of_mode (result);
9418 #endif
9420 if (result)
9421 return result;
9423 if (MEM_P (x))
9425 int offset = 0;
9427 /* Refuse to work on a volatile memory ref or one with a mode-dependent
9428 address. */
9429 if (MEM_VOLATILE_P (x) || mode_dependent_address_p (XEXP (x, 0)))
9430 goto fail;
9432 /* If we want to refer to something bigger than the original memref,
9433 generate a paradoxical subreg instead. That will force a reload
9434 of the original memref X. */
9435 if (isize < osize)
9436 return gen_rtx_SUBREG (omode, x, 0);
9438 if (WORDS_BIG_ENDIAN)
9439 offset = MAX (isize, UNITS_PER_WORD) - MAX (osize, UNITS_PER_WORD);
9441 /* Adjust the address so that the address-after-the-data is
9442 unchanged. */
9443 if (BYTES_BIG_ENDIAN)
9444 offset -= MIN (UNITS_PER_WORD, osize) - MIN (UNITS_PER_WORD, isize);
9446 return adjust_address_nv (x, omode, offset);
9449 /* If X is a comparison operator, rewrite it in a new mode. This
9450 probably won't match, but may allow further simplifications. */
9451 else if (COMPARISON_P (x))
9452 return gen_rtx_fmt_ee (GET_CODE (x), omode, XEXP (x, 0), XEXP (x, 1));
9454 /* If we couldn't simplify X any other way, just enclose it in a
9455 SUBREG. Normally, this SUBREG won't match, but some patterns may
9456 include an explicit SUBREG or we may simplify it further in combine. */
9457 else
9459 int offset = 0;
9460 rtx res;
9462 offset = subreg_lowpart_offset (omode, imode);
9463 if (imode == VOIDmode)
9465 imode = int_mode_for_mode (omode);
9466 x = gen_lowpart_common (imode, x);
9467 if (x == NULL)
9468 goto fail;
9470 res = simplify_gen_subreg (omode, x, imode, offset);
9471 if (res)
9472 return res;
9475 fail:
9476 return gen_rtx_CLOBBER (imode, const0_rtx);
9479 /* Simplify a comparison between *POP0 and *POP1 where CODE is the
9480 comparison code that will be tested.
9482 The result is a possibly different comparison code to use. *POP0 and
9483 *POP1 may be updated.
9485 It is possible that we might detect that a comparison is either always
9486 true or always false. However, we do not perform general constant
9487 folding in combine, so this knowledge isn't useful. Such tautologies
9488 should have been detected earlier. Hence we ignore all such cases. */
9490 static enum rtx_code
9491 simplify_comparison (enum rtx_code code, rtx *pop0, rtx *pop1)
9493 rtx op0 = *pop0;
9494 rtx op1 = *pop1;
9495 rtx tem, tem1;
9496 int i;
9497 enum machine_mode mode, tmode;
9499 /* Try a few ways of applying the same transformation to both operands. */
9500 while (1)
9502 #ifndef WORD_REGISTER_OPERATIONS
9503 /* The test below this one won't handle SIGN_EXTENDs on these machines,
9504 so check specially. */
9505 if (code != GTU && code != GEU && code != LTU && code != LEU
9506 && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT
9507 && GET_CODE (XEXP (op0, 0)) == ASHIFT
9508 && GET_CODE (XEXP (op1, 0)) == ASHIFT
9509 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG
9510 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG
9511 && (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0)))
9512 == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0))))
9513 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9514 && XEXP (op0, 1) == XEXP (op1, 1)
9515 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
9516 && XEXP (op0, 1) == XEXP (XEXP (op1, 0), 1)
9517 && (INTVAL (XEXP (op0, 1))
9518 == (GET_MODE_BITSIZE (GET_MODE (op0))
9519 - (GET_MODE_BITSIZE
9520 (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))))))))
9522 op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0));
9523 op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0));
9525 #endif
9527 /* If both operands are the same constant shift, see if we can ignore the
9528 shift. We can if the shift is a rotate or if the bits shifted out of
9529 this shift are known to be zero for both inputs and if the type of
9530 comparison is compatible with the shift. */
9531 if (GET_CODE (op0) == GET_CODE (op1)
9532 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
9533 && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ))
9534 || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT)
9535 && (code != GT && code != LT && code != GE && code != LE))
9536 || (GET_CODE (op0) == ASHIFTRT
9537 && (code != GTU && code != LTU
9538 && code != GEU && code != LEU)))
9539 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9540 && INTVAL (XEXP (op0, 1)) >= 0
9541 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
9542 && XEXP (op0, 1) == XEXP (op1, 1))
9544 enum machine_mode mode = GET_MODE (op0);
9545 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9546 int shift_count = INTVAL (XEXP (op0, 1));
9548 if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT)
9549 mask &= (mask >> shift_count) << shift_count;
9550 else if (GET_CODE (op0) == ASHIFT)
9551 mask = (mask & (mask << shift_count)) >> shift_count;
9553 if ((nonzero_bits (XEXP (op0, 0), mode) & ~mask) == 0
9554 && (nonzero_bits (XEXP (op1, 0), mode) & ~mask) == 0)
9555 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0);
9556 else
9557 break;
9560 /* If both operands are AND's of a paradoxical SUBREG by constant, the
9561 SUBREGs are of the same mode, and, in both cases, the AND would
9562 be redundant if the comparison was done in the narrower mode,
9563 do the comparison in the narrower mode (e.g., we are AND'ing with 1
9564 and the operand's possibly nonzero bits are 0xffffff01; in that case
9565 if we only care about QImode, we don't need the AND). This case
9566 occurs if the output mode of an scc insn is not SImode and
9567 STORE_FLAG_VALUE == 1 (e.g., the 386).
9569 Similarly, check for a case where the AND's are ZERO_EXTEND
9570 operations from some narrower mode even though a SUBREG is not
9571 present. */
9573 else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND
9574 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9575 && GET_CODE (XEXP (op1, 1)) == CONST_INT)
9577 rtx inner_op0 = XEXP (op0, 0);
9578 rtx inner_op1 = XEXP (op1, 0);
9579 HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1));
9580 HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1));
9581 int changed = 0;
9583 if (GET_CODE (inner_op0) == SUBREG && GET_CODE (inner_op1) == SUBREG
9584 && (GET_MODE_SIZE (GET_MODE (inner_op0))
9585 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner_op0))))
9586 && (GET_MODE (SUBREG_REG (inner_op0))
9587 == GET_MODE (SUBREG_REG (inner_op1)))
9588 && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (inner_op0)))
9589 <= HOST_BITS_PER_WIDE_INT)
9590 && (0 == ((~c0) & nonzero_bits (SUBREG_REG (inner_op0),
9591 GET_MODE (SUBREG_REG (inner_op0)))))
9592 && (0 == ((~c1) & nonzero_bits (SUBREG_REG (inner_op1),
9593 GET_MODE (SUBREG_REG (inner_op1))))))
9595 op0 = SUBREG_REG (inner_op0);
9596 op1 = SUBREG_REG (inner_op1);
9598 /* The resulting comparison is always unsigned since we masked
9599 off the original sign bit. */
9600 code = unsigned_condition (code);
9602 changed = 1;
9605 else if (c0 == c1)
9606 for (tmode = GET_CLASS_NARROWEST_MODE
9607 (GET_MODE_CLASS (GET_MODE (op0)));
9608 tmode != GET_MODE (op0); tmode = GET_MODE_WIDER_MODE (tmode))
9609 if ((unsigned HOST_WIDE_INT) c0 == GET_MODE_MASK (tmode))
9611 op0 = gen_lowpart (tmode, inner_op0);
9612 op1 = gen_lowpart (tmode, inner_op1);
9613 code = unsigned_condition (code);
9614 changed = 1;
9615 break;
9618 if (! changed)
9619 break;
9622 /* If both operands are NOT, we can strip off the outer operation
9623 and adjust the comparison code for swapped operands; similarly for
9624 NEG, except that this must be an equality comparison. */
9625 else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT)
9626 || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG
9627 && (code == EQ || code == NE)))
9628 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code);
9630 else
9631 break;
9634 /* If the first operand is a constant, swap the operands and adjust the
9635 comparison code appropriately, but don't do this if the second operand
9636 is already a constant integer. */
9637 if (swap_commutative_operands_p (op0, op1))
9639 tem = op0, op0 = op1, op1 = tem;
9640 code = swap_condition (code);
9643 /* We now enter a loop during which we will try to simplify the comparison.
9644 For the most part, we only are concerned with comparisons with zero,
9645 but some things may really be comparisons with zero but not start
9646 out looking that way. */
9648 while (GET_CODE (op1) == CONST_INT)
9650 enum machine_mode mode = GET_MODE (op0);
9651 unsigned int mode_width = GET_MODE_BITSIZE (mode);
9652 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9653 int equality_comparison_p;
9654 int sign_bit_comparison_p;
9655 int unsigned_comparison_p;
9656 HOST_WIDE_INT const_op;
9658 /* We only want to handle integral modes. This catches VOIDmode,
9659 CCmode, and the floating-point modes. An exception is that we
9660 can handle VOIDmode if OP0 is a COMPARE or a comparison
9661 operation. */
9663 if (GET_MODE_CLASS (mode) != MODE_INT
9664 && ! (mode == VOIDmode
9665 && (GET_CODE (op0) == COMPARE || COMPARISON_P (op0))))
9666 break;
9668 /* Get the constant we are comparing against and turn off all bits
9669 not on in our mode. */
9670 const_op = INTVAL (op1);
9671 if (mode != VOIDmode)
9672 const_op = trunc_int_for_mode (const_op, mode);
9673 op1 = GEN_INT (const_op);
9675 /* If we are comparing against a constant power of two and the value
9676 being compared can only have that single bit nonzero (e.g., it was
9677 `and'ed with that bit), we can replace this with a comparison
9678 with zero. */
9679 if (const_op
9680 && (code == EQ || code == NE || code == GE || code == GEU
9681 || code == LT || code == LTU)
9682 && mode_width <= HOST_BITS_PER_WIDE_INT
9683 && exact_log2 (const_op) >= 0
9684 && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) const_op)
9686 code = (code == EQ || code == GE || code == GEU ? NE : EQ);
9687 op1 = const0_rtx, const_op = 0;
9690 /* Similarly, if we are comparing a value known to be either -1 or
9691 0 with -1, change it to the opposite comparison against zero. */
9693 if (const_op == -1
9694 && (code == EQ || code == NE || code == GT || code == LE
9695 || code == GEU || code == LTU)
9696 && num_sign_bit_copies (op0, mode) == mode_width)
9698 code = (code == EQ || code == LE || code == GEU ? NE : EQ);
9699 op1 = const0_rtx, const_op = 0;
9702 /* Do some canonicalizations based on the comparison code. We prefer
9703 comparisons against zero and then prefer equality comparisons.
9704 If we can reduce the size of a constant, we will do that too. */
9706 switch (code)
9708 case LT:
9709 /* < C is equivalent to <= (C - 1) */
9710 if (const_op > 0)
9712 const_op -= 1;
9713 op1 = GEN_INT (const_op);
9714 code = LE;
9715 /* ... fall through to LE case below. */
9717 else
9718 break;
9720 case LE:
9721 /* <= C is equivalent to < (C + 1); we do this for C < 0 */
9722 if (const_op < 0)
9724 const_op += 1;
9725 op1 = GEN_INT (const_op);
9726 code = LT;
9729 /* If we are doing a <= 0 comparison on a value known to have
9730 a zero sign bit, we can replace this with == 0. */
9731 else if (const_op == 0
9732 && mode_width <= HOST_BITS_PER_WIDE_INT
9733 && (nonzero_bits (op0, mode)
9734 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9735 code = EQ;
9736 break;
9738 case GE:
9739 /* >= C is equivalent to > (C - 1). */
9740 if (const_op > 0)
9742 const_op -= 1;
9743 op1 = GEN_INT (const_op);
9744 code = GT;
9745 /* ... fall through to GT below. */
9747 else
9748 break;
9750 case GT:
9751 /* > C is equivalent to >= (C + 1); we do this for C < 0. */
9752 if (const_op < 0)
9754 const_op += 1;
9755 op1 = GEN_INT (const_op);
9756 code = GE;
9759 /* If we are doing a > 0 comparison on a value known to have
9760 a zero sign bit, we can replace this with != 0. */
9761 else if (const_op == 0
9762 && mode_width <= HOST_BITS_PER_WIDE_INT
9763 && (nonzero_bits (op0, mode)
9764 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9765 code = NE;
9766 break;
9768 case LTU:
9769 /* < C is equivalent to <= (C - 1). */
9770 if (const_op > 0)
9772 const_op -= 1;
9773 op1 = GEN_INT (const_op);
9774 code = LEU;
9775 /* ... fall through ... */
9778 /* (unsigned) < 0x80000000 is equivalent to >= 0. */
9779 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9780 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9782 const_op = 0, op1 = const0_rtx;
9783 code = GE;
9784 break;
9786 else
9787 break;
9789 case LEU:
9790 /* unsigned <= 0 is equivalent to == 0 */
9791 if (const_op == 0)
9792 code = EQ;
9794 /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */
9795 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9796 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9798 const_op = 0, op1 = const0_rtx;
9799 code = GE;
9801 break;
9803 case GEU:
9804 /* >= C is equivalent to > (C - 1). */
9805 if (const_op > 1)
9807 const_op -= 1;
9808 op1 = GEN_INT (const_op);
9809 code = GTU;
9810 /* ... fall through ... */
9813 /* (unsigned) >= 0x80000000 is equivalent to < 0. */
9814 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9815 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9817 const_op = 0, op1 = const0_rtx;
9818 code = LT;
9819 break;
9821 else
9822 break;
9824 case GTU:
9825 /* unsigned > 0 is equivalent to != 0 */
9826 if (const_op == 0)
9827 code = NE;
9829 /* (unsigned) > 0x7fffffff is equivalent to < 0. */
9830 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9831 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9833 const_op = 0, op1 = const0_rtx;
9834 code = LT;
9836 break;
9838 default:
9839 break;
9842 /* Compute some predicates to simplify code below. */
9844 equality_comparison_p = (code == EQ || code == NE);
9845 sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0);
9846 unsigned_comparison_p = (code == LTU || code == LEU || code == GTU
9847 || code == GEU);
9849 /* If this is a sign bit comparison and we can do arithmetic in
9850 MODE, say that we will only be needing the sign bit of OP0. */
9851 if (sign_bit_comparison_p
9852 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
9853 op0 = force_to_mode (op0, mode,
9854 ((HOST_WIDE_INT) 1
9855 << (GET_MODE_BITSIZE (mode) - 1)),
9858 /* Now try cases based on the opcode of OP0. If none of the cases
9859 does a "continue", we exit this loop immediately after the
9860 switch. */
9862 switch (GET_CODE (op0))
9864 case ZERO_EXTRACT:
9865 /* If we are extracting a single bit from a variable position in
9866 a constant that has only a single bit set and are comparing it
9867 with zero, we can convert this into an equality comparison
9868 between the position and the location of the single bit. */
9869 /* Except we can't if SHIFT_COUNT_TRUNCATED is set, since we might
9870 have already reduced the shift count modulo the word size. */
9871 if (!SHIFT_COUNT_TRUNCATED
9872 && GET_CODE (XEXP (op0, 0)) == CONST_INT
9873 && XEXP (op0, 1) == const1_rtx
9874 && equality_comparison_p && const_op == 0
9875 && (i = exact_log2 (INTVAL (XEXP (op0, 0)))) >= 0)
9877 if (BITS_BIG_ENDIAN)
9879 enum machine_mode new_mode
9880 = mode_for_extraction (EP_extzv, 1);
9881 if (new_mode == MAX_MACHINE_MODE)
9882 i = BITS_PER_WORD - 1 - i;
9883 else
9885 mode = new_mode;
9886 i = (GET_MODE_BITSIZE (mode) - 1 - i);
9890 op0 = XEXP (op0, 2);
9891 op1 = GEN_INT (i);
9892 const_op = i;
9894 /* Result is nonzero iff shift count is equal to I. */
9895 code = reverse_condition (code);
9896 continue;
9899 /* ... fall through ... */
9901 case SIGN_EXTRACT:
9902 tem = expand_compound_operation (op0);
9903 if (tem != op0)
9905 op0 = tem;
9906 continue;
9908 break;
9910 case NOT:
9911 /* If testing for equality, we can take the NOT of the constant. */
9912 if (equality_comparison_p
9913 && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0)
9915 op0 = XEXP (op0, 0);
9916 op1 = tem;
9917 continue;
9920 /* If just looking at the sign bit, reverse the sense of the
9921 comparison. */
9922 if (sign_bit_comparison_p)
9924 op0 = XEXP (op0, 0);
9925 code = (code == GE ? LT : GE);
9926 continue;
9928 break;
9930 case NEG:
9931 /* If testing for equality, we can take the NEG of the constant. */
9932 if (equality_comparison_p
9933 && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0)
9935 op0 = XEXP (op0, 0);
9936 op1 = tem;
9937 continue;
9940 /* The remaining cases only apply to comparisons with zero. */
9941 if (const_op != 0)
9942 break;
9944 /* When X is ABS or is known positive,
9945 (neg X) is < 0 if and only if X != 0. */
9947 if (sign_bit_comparison_p
9948 && (GET_CODE (XEXP (op0, 0)) == ABS
9949 || (mode_width <= HOST_BITS_PER_WIDE_INT
9950 && (nonzero_bits (XEXP (op0, 0), mode)
9951 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)))
9953 op0 = XEXP (op0, 0);
9954 code = (code == LT ? NE : EQ);
9955 continue;
9958 /* If we have NEG of something whose two high-order bits are the
9959 same, we know that "(-a) < 0" is equivalent to "a > 0". */
9960 if (num_sign_bit_copies (op0, mode) >= 2)
9962 op0 = XEXP (op0, 0);
9963 code = swap_condition (code);
9964 continue;
9966 break;
9968 case ROTATE:
9969 /* If we are testing equality and our count is a constant, we
9970 can perform the inverse operation on our RHS. */
9971 if (equality_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
9972 && (tem = simplify_binary_operation (ROTATERT, mode,
9973 op1, XEXP (op0, 1))) != 0)
9975 op0 = XEXP (op0, 0);
9976 op1 = tem;
9977 continue;
9980 /* If we are doing a < 0 or >= 0 comparison, it means we are testing
9981 a particular bit. Convert it to an AND of a constant of that
9982 bit. This will be converted into a ZERO_EXTRACT. */
9983 if (const_op == 0 && sign_bit_comparison_p
9984 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9985 && mode_width <= HOST_BITS_PER_WIDE_INT)
9987 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
9988 ((HOST_WIDE_INT) 1
9989 << (mode_width - 1
9990 - INTVAL (XEXP (op0, 1)))));
9991 code = (code == LT ? NE : EQ);
9992 continue;
9995 /* Fall through. */
9997 case ABS:
9998 /* ABS is ignorable inside an equality comparison with zero. */
9999 if (const_op == 0 && equality_comparison_p)
10001 op0 = XEXP (op0, 0);
10002 continue;
10004 break;
10006 case SIGN_EXTEND:
10007 /* Can simplify (compare (zero/sign_extend FOO) CONST) to
10008 (compare FOO CONST) if CONST fits in FOO's mode and we
10009 are either testing inequality or have an unsigned
10010 comparison with ZERO_EXTEND or a signed comparison with
10011 SIGN_EXTEND. But don't do it if we don't have a compare
10012 insn of the given mode, since we'd have to revert it
10013 later on, and then we wouldn't know whether to sign- or
10014 zero-extend. */
10015 mode = GET_MODE (XEXP (op0, 0));
10016 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10017 && ! unsigned_comparison_p
10018 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10019 && ((unsigned HOST_WIDE_INT) const_op
10020 < (((unsigned HOST_WIDE_INT) 1
10021 << (GET_MODE_BITSIZE (mode) - 1))))
10022 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10024 op0 = XEXP (op0, 0);
10025 continue;
10027 break;
10029 case SUBREG:
10030 /* Check for the case where we are comparing A - C1 with C2, that is
10032 (subreg:MODE (plus (A) (-C1))) op (C2)
10034 with C1 a constant, and try to lift the SUBREG, i.e. to do the
10035 comparison in the wider mode. One of the following two conditions
10036 must be true in order for this to be valid:
10038 1. The mode extension results in the same bit pattern being added
10039 on both sides and the comparison is equality or unsigned. As
10040 C2 has been truncated to fit in MODE, the pattern can only be
10041 all 0s or all 1s.
10043 2. The mode extension results in the sign bit being copied on
10044 each side.
10046 The difficulty here is that we have predicates for A but not for
10047 (A - C1) so we need to check that C1 is within proper bounds so
10048 as to perturbate A as little as possible. */
10050 if (mode_width <= HOST_BITS_PER_WIDE_INT
10051 && subreg_lowpart_p (op0)
10052 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) > mode_width
10053 && GET_CODE (SUBREG_REG (op0)) == PLUS
10054 && GET_CODE (XEXP (SUBREG_REG (op0), 1)) == CONST_INT)
10056 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
10057 rtx a = XEXP (SUBREG_REG (op0), 0);
10058 HOST_WIDE_INT c1 = -INTVAL (XEXP (SUBREG_REG (op0), 1));
10060 if ((c1 > 0
10061 && (unsigned HOST_WIDE_INT) c1
10062 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)
10063 && (equality_comparison_p || unsigned_comparison_p)
10064 /* (A - C1) zero-extends if it is positive and sign-extends
10065 if it is negative, C2 both zero- and sign-extends. */
10066 && ((0 == (nonzero_bits (a, inner_mode)
10067 & ~GET_MODE_MASK (mode))
10068 && const_op >= 0)
10069 /* (A - C1) sign-extends if it is positive and 1-extends
10070 if it is negative, C2 both sign- and 1-extends. */
10071 || (num_sign_bit_copies (a, inner_mode)
10072 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10073 - mode_width)
10074 && const_op < 0)))
10075 || ((unsigned HOST_WIDE_INT) c1
10076 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 2)
10077 /* (A - C1) always sign-extends, like C2. */
10078 && num_sign_bit_copies (a, inner_mode)
10079 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10080 - (mode_width - 1))))
10082 op0 = SUBREG_REG (op0);
10083 continue;
10087 /* If the inner mode is narrower and we are extracting the low part,
10088 we can treat the SUBREG as if it were a ZERO_EXTEND. */
10089 if (subreg_lowpart_p (op0)
10090 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) < mode_width)
10091 /* Fall through */ ;
10092 else
10093 break;
10095 /* ... fall through ... */
10097 case ZERO_EXTEND:
10098 mode = GET_MODE (XEXP (op0, 0));
10099 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10100 && (unsigned_comparison_p || equality_comparison_p)
10101 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10102 && ((unsigned HOST_WIDE_INT) const_op < GET_MODE_MASK (mode))
10103 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10105 op0 = XEXP (op0, 0);
10106 continue;
10108 break;
10110 case PLUS:
10111 /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do
10112 this for equality comparisons due to pathological cases involving
10113 overflows. */
10114 if (equality_comparison_p
10115 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10116 op1, XEXP (op0, 1))))
10118 op0 = XEXP (op0, 0);
10119 op1 = tem;
10120 continue;
10123 /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */
10124 if (const_op == 0 && XEXP (op0, 1) == constm1_rtx
10125 && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p)
10127 op0 = XEXP (XEXP (op0, 0), 0);
10128 code = (code == LT ? EQ : NE);
10129 continue;
10131 break;
10133 case MINUS:
10134 /* We used to optimize signed comparisons against zero, but that
10135 was incorrect. Unsigned comparisons against zero (GTU, LEU)
10136 arrive here as equality comparisons, or (GEU, LTU) are
10137 optimized away. No need to special-case them. */
10139 /* (eq (minus A B) C) -> (eq A (plus B C)) or
10140 (eq B (minus A C)), whichever simplifies. We can only do
10141 this for equality comparisons due to pathological cases involving
10142 overflows. */
10143 if (equality_comparison_p
10144 && 0 != (tem = simplify_binary_operation (PLUS, mode,
10145 XEXP (op0, 1), op1)))
10147 op0 = XEXP (op0, 0);
10148 op1 = tem;
10149 continue;
10152 if (equality_comparison_p
10153 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10154 XEXP (op0, 0), op1)))
10156 op0 = XEXP (op0, 1);
10157 op1 = tem;
10158 continue;
10161 /* The sign bit of (minus (ashiftrt X C) X), where C is the number
10162 of bits in X minus 1, is one iff X > 0. */
10163 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT
10164 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10165 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (op0, 0), 1))
10166 == mode_width - 1
10167 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10169 op0 = XEXP (op0, 1);
10170 code = (code == GE ? LE : GT);
10171 continue;
10173 break;
10175 case XOR:
10176 /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification
10177 if C is zero or B is a constant. */
10178 if (equality_comparison_p
10179 && 0 != (tem = simplify_binary_operation (XOR, mode,
10180 XEXP (op0, 1), op1)))
10182 op0 = XEXP (op0, 0);
10183 op1 = tem;
10184 continue;
10186 break;
10188 case EQ: case NE:
10189 case UNEQ: case LTGT:
10190 case LT: case LTU: case UNLT: case LE: case LEU: case UNLE:
10191 case GT: case GTU: case UNGT: case GE: case GEU: case UNGE:
10192 case UNORDERED: case ORDERED:
10193 /* We can't do anything if OP0 is a condition code value, rather
10194 than an actual data value. */
10195 if (const_op != 0
10196 || CC0_P (XEXP (op0, 0))
10197 || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC)
10198 break;
10200 /* Get the two operands being compared. */
10201 if (GET_CODE (XEXP (op0, 0)) == COMPARE)
10202 tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1);
10203 else
10204 tem = XEXP (op0, 0), tem1 = XEXP (op0, 1);
10206 /* Check for the cases where we simply want the result of the
10207 earlier test or the opposite of that result. */
10208 if (code == NE || code == EQ
10209 || (GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
10210 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10211 && (STORE_FLAG_VALUE
10212 & (((HOST_WIDE_INT) 1
10213 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
10214 && (code == LT || code == GE)))
10216 enum rtx_code new_code;
10217 if (code == LT || code == NE)
10218 new_code = GET_CODE (op0);
10219 else
10220 new_code = reversed_comparison_code (op0, NULL);
10222 if (new_code != UNKNOWN)
10224 code = new_code;
10225 op0 = tem;
10226 op1 = tem1;
10227 continue;
10230 break;
10232 case IOR:
10233 /* The sign bit of (ior (plus X (const_int -1)) X) is nonzero
10234 iff X <= 0. */
10235 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS
10236 && XEXP (XEXP (op0, 0), 1) == constm1_rtx
10237 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10239 op0 = XEXP (op0, 1);
10240 code = (code == GE ? GT : LE);
10241 continue;
10243 break;
10245 case AND:
10246 /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This
10247 will be converted to a ZERO_EXTRACT later. */
10248 if (const_op == 0 && equality_comparison_p
10249 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10250 && XEXP (XEXP (op0, 0), 0) == const1_rtx)
10252 op0 = simplify_and_const_int
10253 (NULL_RTX, mode, gen_rtx_LSHIFTRT (mode,
10254 XEXP (op0, 1),
10255 XEXP (XEXP (op0, 0), 1)),
10256 (HOST_WIDE_INT) 1);
10257 continue;
10260 /* If we are comparing (and (lshiftrt X C1) C2) for equality with
10261 zero and X is a comparison and C1 and C2 describe only bits set
10262 in STORE_FLAG_VALUE, we can compare with X. */
10263 if (const_op == 0 && equality_comparison_p
10264 && mode_width <= HOST_BITS_PER_WIDE_INT
10265 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10266 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT
10267 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10268 && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0
10269 && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT)
10271 mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10272 << INTVAL (XEXP (XEXP (op0, 0), 1)));
10273 if ((~STORE_FLAG_VALUE & mask) == 0
10274 && (COMPARISON_P (XEXP (XEXP (op0, 0), 0))
10275 || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0
10276 && COMPARISON_P (tem))))
10278 op0 = XEXP (XEXP (op0, 0), 0);
10279 continue;
10283 /* If we are doing an equality comparison of an AND of a bit equal
10284 to the sign bit, replace this with a LT or GE comparison of
10285 the underlying value. */
10286 if (equality_comparison_p
10287 && const_op == 0
10288 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10289 && mode_width <= HOST_BITS_PER_WIDE_INT
10290 && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10291 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
10293 op0 = XEXP (op0, 0);
10294 code = (code == EQ ? GE : LT);
10295 continue;
10298 /* If this AND operation is really a ZERO_EXTEND from a narrower
10299 mode, the constant fits within that mode, and this is either an
10300 equality or unsigned comparison, try to do this comparison in
10301 the narrower mode.
10303 Note that in:
10305 (ne:DI (and:DI (reg:DI 4) (const_int 0xffffffff)) (const_int 0))
10306 -> (ne:DI (reg:SI 4) (const_int 0))
10308 unless TRULY_NOOP_TRUNCATION allows it or the register is
10309 known to hold a value of the required mode the
10310 transformation is invalid. */
10311 if ((equality_comparison_p || unsigned_comparison_p)
10312 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10313 && (i = exact_log2 ((INTVAL (XEXP (op0, 1))
10314 & GET_MODE_MASK (mode))
10315 + 1)) >= 0
10316 && const_op >> i == 0
10317 && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode
10318 && (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (tmode),
10319 GET_MODE_BITSIZE (GET_MODE (op0)))
10320 || (REG_P (XEXP (op0, 0))
10321 && reg_truncated_to_mode (tmode, XEXP (op0, 0)))))
10323 op0 = gen_lowpart (tmode, XEXP (op0, 0));
10324 continue;
10327 /* If this is (and:M1 (subreg:M2 X 0) (const_int C1)) where C1
10328 fits in both M1 and M2 and the SUBREG is either paradoxical
10329 or represents the low part, permute the SUBREG and the AND
10330 and try again. */
10331 if (GET_CODE (XEXP (op0, 0)) == SUBREG)
10333 unsigned HOST_WIDE_INT c1;
10334 tmode = GET_MODE (SUBREG_REG (XEXP (op0, 0)));
10335 /* Require an integral mode, to avoid creating something like
10336 (AND:SF ...). */
10337 if (SCALAR_INT_MODE_P (tmode)
10338 /* It is unsafe to commute the AND into the SUBREG if the
10339 SUBREG is paradoxical and WORD_REGISTER_OPERATIONS is
10340 not defined. As originally written the upper bits
10341 have a defined value due to the AND operation.
10342 However, if we commute the AND inside the SUBREG then
10343 they no longer have defined values and the meaning of
10344 the code has been changed. */
10345 && (0
10346 #ifdef WORD_REGISTER_OPERATIONS
10347 || (mode_width > GET_MODE_BITSIZE (tmode)
10348 && mode_width <= BITS_PER_WORD)
10349 #endif
10350 || (mode_width <= GET_MODE_BITSIZE (tmode)
10351 && subreg_lowpart_p (XEXP (op0, 0))))
10352 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10353 && mode_width <= HOST_BITS_PER_WIDE_INT
10354 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
10355 && ((c1 = INTVAL (XEXP (op0, 1))) & ~mask) == 0
10356 && (c1 & ~GET_MODE_MASK (tmode)) == 0
10357 && c1 != mask
10358 && c1 != GET_MODE_MASK (tmode))
10360 op0 = simplify_gen_binary (AND, tmode,
10361 SUBREG_REG (XEXP (op0, 0)),
10362 gen_int_mode (c1, tmode));
10363 op0 = gen_lowpart (mode, op0);
10364 continue;
10368 /* Convert (ne (and (not X) 1) 0) to (eq (and X 1) 0). */
10369 if (const_op == 0 && equality_comparison_p
10370 && XEXP (op0, 1) == const1_rtx
10371 && GET_CODE (XEXP (op0, 0)) == NOT)
10373 op0 = simplify_and_const_int
10374 (NULL_RTX, mode, XEXP (XEXP (op0, 0), 0), (HOST_WIDE_INT) 1);
10375 code = (code == NE ? EQ : NE);
10376 continue;
10379 /* Convert (ne (and (lshiftrt (not X)) 1) 0) to
10380 (eq (and (lshiftrt X) 1) 0).
10381 Also handle the case where (not X) is expressed using xor. */
10382 if (const_op == 0 && equality_comparison_p
10383 && XEXP (op0, 1) == const1_rtx
10384 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT)
10386 rtx shift_op = XEXP (XEXP (op0, 0), 0);
10387 rtx shift_count = XEXP (XEXP (op0, 0), 1);
10389 if (GET_CODE (shift_op) == NOT
10390 || (GET_CODE (shift_op) == XOR
10391 && GET_CODE (XEXP (shift_op, 1)) == CONST_INT
10392 && GET_CODE (shift_count) == CONST_INT
10393 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
10394 && (INTVAL (XEXP (shift_op, 1))
10395 == (HOST_WIDE_INT) 1 << INTVAL (shift_count))))
10397 op0 = simplify_and_const_int
10398 (NULL_RTX, mode,
10399 gen_rtx_LSHIFTRT (mode, XEXP (shift_op, 0), shift_count),
10400 (HOST_WIDE_INT) 1);
10401 code = (code == NE ? EQ : NE);
10402 continue;
10405 break;
10407 case ASHIFT:
10408 /* If we have (compare (ashift FOO N) (const_int C)) and
10409 the high order N bits of FOO (N+1 if an inequality comparison)
10410 are known to be zero, we can do this by comparing FOO with C
10411 shifted right N bits so long as the low-order N bits of C are
10412 zero. */
10413 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10414 && INTVAL (XEXP (op0, 1)) >= 0
10415 && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p)
10416 < HOST_BITS_PER_WIDE_INT)
10417 && ((const_op
10418 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0)
10419 && mode_width <= HOST_BITS_PER_WIDE_INT
10420 && (nonzero_bits (XEXP (op0, 0), mode)
10421 & ~(mask >> (INTVAL (XEXP (op0, 1))
10422 + ! equality_comparison_p))) == 0)
10424 /* We must perform a logical shift, not an arithmetic one,
10425 as we want the top N bits of C to be zero. */
10426 unsigned HOST_WIDE_INT temp = const_op & GET_MODE_MASK (mode);
10428 temp >>= INTVAL (XEXP (op0, 1));
10429 op1 = gen_int_mode (temp, mode);
10430 op0 = XEXP (op0, 0);
10431 continue;
10434 /* If we are doing a sign bit comparison, it means we are testing
10435 a particular bit. Convert it to the appropriate AND. */
10436 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
10437 && mode_width <= HOST_BITS_PER_WIDE_INT)
10439 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10440 ((HOST_WIDE_INT) 1
10441 << (mode_width - 1
10442 - INTVAL (XEXP (op0, 1)))));
10443 code = (code == LT ? NE : EQ);
10444 continue;
10447 /* If this an equality comparison with zero and we are shifting
10448 the low bit to the sign bit, we can convert this to an AND of the
10449 low-order bit. */
10450 if (const_op == 0 && equality_comparison_p
10451 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10452 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10453 == mode_width - 1)
10455 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10456 (HOST_WIDE_INT) 1);
10457 continue;
10459 break;
10461 case ASHIFTRT:
10462 /* If this is an equality comparison with zero, we can do this
10463 as a logical shift, which might be much simpler. */
10464 if (equality_comparison_p && const_op == 0
10465 && GET_CODE (XEXP (op0, 1)) == CONST_INT)
10467 op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode,
10468 XEXP (op0, 0),
10469 INTVAL (XEXP (op0, 1)));
10470 continue;
10473 /* If OP0 is a sign extension and CODE is not an unsigned comparison,
10474 do the comparison in a narrower mode. */
10475 if (! unsigned_comparison_p
10476 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10477 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10478 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
10479 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10480 MODE_INT, 1)) != BLKmode
10481 && (((unsigned HOST_WIDE_INT) const_op
10482 + (GET_MODE_MASK (tmode) >> 1) + 1)
10483 <= GET_MODE_MASK (tmode)))
10485 op0 = gen_lowpart (tmode, XEXP (XEXP (op0, 0), 0));
10486 continue;
10489 /* Likewise if OP0 is a PLUS of a sign extension with a
10490 constant, which is usually represented with the PLUS
10491 between the shifts. */
10492 if (! unsigned_comparison_p
10493 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10494 && GET_CODE (XEXP (op0, 0)) == PLUS
10495 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10496 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == ASHIFT
10497 && XEXP (op0, 1) == XEXP (XEXP (XEXP (op0, 0), 0), 1)
10498 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10499 MODE_INT, 1)) != BLKmode
10500 && (((unsigned HOST_WIDE_INT) const_op
10501 + (GET_MODE_MASK (tmode) >> 1) + 1)
10502 <= GET_MODE_MASK (tmode)))
10504 rtx inner = XEXP (XEXP (XEXP (op0, 0), 0), 0);
10505 rtx add_const = XEXP (XEXP (op0, 0), 1);
10506 rtx new_const = simplify_gen_binary (ASHIFTRT, GET_MODE (op0),
10507 add_const, XEXP (op0, 1));
10509 op0 = simplify_gen_binary (PLUS, tmode,
10510 gen_lowpart (tmode, inner),
10511 new_const);
10512 continue;
10515 /* ... fall through ... */
10516 case LSHIFTRT:
10517 /* If we have (compare (xshiftrt FOO N) (const_int C)) and
10518 the low order N bits of FOO are known to be zero, we can do this
10519 by comparing FOO with C shifted left N bits so long as no
10520 overflow occurs. */
10521 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10522 && INTVAL (XEXP (op0, 1)) >= 0
10523 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
10524 && mode_width <= HOST_BITS_PER_WIDE_INT
10525 && (nonzero_bits (XEXP (op0, 0), mode)
10526 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0
10527 && (((unsigned HOST_WIDE_INT) const_op
10528 + (GET_CODE (op0) != LSHIFTRT
10529 ? ((GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)) >> 1)
10530 + 1)
10531 : 0))
10532 <= GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1))))
10534 /* If the shift was logical, then we must make the condition
10535 unsigned. */
10536 if (GET_CODE (op0) == LSHIFTRT)
10537 code = unsigned_condition (code);
10539 const_op <<= INTVAL (XEXP (op0, 1));
10540 op1 = GEN_INT (const_op);
10541 op0 = XEXP (op0, 0);
10542 continue;
10545 /* If we are using this shift to extract just the sign bit, we
10546 can replace this with an LT or GE comparison. */
10547 if (const_op == 0
10548 && (equality_comparison_p || sign_bit_comparison_p)
10549 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10550 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10551 == mode_width - 1)
10553 op0 = XEXP (op0, 0);
10554 code = (code == NE || code == GT ? LT : GE);
10555 continue;
10557 break;
10559 default:
10560 break;
10563 break;
10566 /* Now make any compound operations involved in this comparison. Then,
10567 check for an outmost SUBREG on OP0 that is not doing anything or is
10568 paradoxical. The latter transformation must only be performed when
10569 it is known that the "extra" bits will be the same in op0 and op1 or
10570 that they don't matter. There are three cases to consider:
10572 1. SUBREG_REG (op0) is a register. In this case the bits are don't
10573 care bits and we can assume they have any convenient value. So
10574 making the transformation is safe.
10576 2. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is not defined.
10577 In this case the upper bits of op0 are undefined. We should not make
10578 the simplification in that case as we do not know the contents of
10579 those bits.
10581 3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is defined and not
10582 UNKNOWN. In that case we know those bits are zeros or ones. We must
10583 also be sure that they are the same as the upper bits of op1.
10585 We can never remove a SUBREG for a non-equality comparison because
10586 the sign bit is in a different place in the underlying object. */
10588 op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET);
10589 op1 = make_compound_operation (op1, SET);
10591 if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0)
10592 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10593 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (op0))) == MODE_INT
10594 && (code == NE || code == EQ))
10596 if (GET_MODE_SIZE (GET_MODE (op0))
10597 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))
10599 /* For paradoxical subregs, allow case 1 as above. Case 3 isn't
10600 implemented. */
10601 if (REG_P (SUBREG_REG (op0)))
10603 op0 = SUBREG_REG (op0);
10604 op1 = gen_lowpart (GET_MODE (op0), op1);
10607 else if ((GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0)))
10608 <= HOST_BITS_PER_WIDE_INT)
10609 && (nonzero_bits (SUBREG_REG (op0),
10610 GET_MODE (SUBREG_REG (op0)))
10611 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10613 tem = gen_lowpart (GET_MODE (SUBREG_REG (op0)), op1);
10615 if ((nonzero_bits (tem, GET_MODE (SUBREG_REG (op0)))
10616 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10617 op0 = SUBREG_REG (op0), op1 = tem;
10621 /* We now do the opposite procedure: Some machines don't have compare
10622 insns in all modes. If OP0's mode is an integer mode smaller than a
10623 word and we can't do a compare in that mode, see if there is a larger
10624 mode for which we can do the compare. There are a number of cases in
10625 which we can use the wider mode. */
10627 mode = GET_MODE (op0);
10628 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10629 && GET_MODE_SIZE (mode) < UNITS_PER_WORD
10630 && ! have_insn_for (COMPARE, mode))
10631 for (tmode = GET_MODE_WIDER_MODE (mode);
10632 (tmode != VOIDmode
10633 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT);
10634 tmode = GET_MODE_WIDER_MODE (tmode))
10635 if (have_insn_for (COMPARE, tmode))
10637 int zero_extended;
10639 /* If the only nonzero bits in OP0 and OP1 are those in the
10640 narrower mode and this is an equality or unsigned comparison,
10641 we can use the wider mode. Similarly for sign-extended
10642 values, in which case it is true for all comparisons. */
10643 zero_extended = ((code == EQ || code == NE
10644 || code == GEU || code == GTU
10645 || code == LEU || code == LTU)
10646 && (nonzero_bits (op0, tmode)
10647 & ~GET_MODE_MASK (mode)) == 0
10648 && ((GET_CODE (op1) == CONST_INT
10649 || (nonzero_bits (op1, tmode)
10650 & ~GET_MODE_MASK (mode)) == 0)));
10652 if (zero_extended
10653 || ((num_sign_bit_copies (op0, tmode)
10654 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10655 - GET_MODE_BITSIZE (mode)))
10656 && (num_sign_bit_copies (op1, tmode)
10657 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10658 - GET_MODE_BITSIZE (mode)))))
10660 /* If OP0 is an AND and we don't have an AND in MODE either,
10661 make a new AND in the proper mode. */
10662 if (GET_CODE (op0) == AND
10663 && !have_insn_for (AND, mode))
10664 op0 = simplify_gen_binary (AND, tmode,
10665 gen_lowpart (tmode,
10666 XEXP (op0, 0)),
10667 gen_lowpart (tmode,
10668 XEXP (op0, 1)));
10670 op0 = gen_lowpart (tmode, op0);
10671 if (zero_extended && GET_CODE (op1) == CONST_INT)
10672 op1 = GEN_INT (INTVAL (op1) & GET_MODE_MASK (mode));
10673 op1 = gen_lowpart (tmode, op1);
10674 break;
10677 /* If this is a test for negative, we can make an explicit
10678 test of the sign bit. */
10680 if (op1 == const0_rtx && (code == LT || code == GE)
10681 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10683 op0 = simplify_gen_binary (AND, tmode,
10684 gen_lowpart (tmode, op0),
10685 GEN_INT ((HOST_WIDE_INT) 1
10686 << (GET_MODE_BITSIZE (mode)
10687 - 1)));
10688 code = (code == LT) ? NE : EQ;
10689 break;
10693 #ifdef CANONICALIZE_COMPARISON
10694 /* If this machine only supports a subset of valid comparisons, see if we
10695 can convert an unsupported one into a supported one. */
10696 CANONICALIZE_COMPARISON (code, op0, op1);
10697 #endif
10699 *pop0 = op0;
10700 *pop1 = op1;
10702 return code;
10705 /* Utility function for record_value_for_reg. Count number of
10706 rtxs in X. */
10707 static int
10708 count_rtxs (rtx x)
10710 enum rtx_code code = GET_CODE (x);
10711 const char *fmt;
10712 int i, ret = 1;
10714 if (GET_RTX_CLASS (code) == '2'
10715 || GET_RTX_CLASS (code) == 'c')
10717 rtx x0 = XEXP (x, 0);
10718 rtx x1 = XEXP (x, 1);
10720 if (x0 == x1)
10721 return 1 + 2 * count_rtxs (x0);
10723 if ((GET_RTX_CLASS (GET_CODE (x1)) == '2'
10724 || GET_RTX_CLASS (GET_CODE (x1)) == 'c')
10725 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10726 return 2 + 2 * count_rtxs (x0)
10727 + count_rtxs (x == XEXP (x1, 0)
10728 ? XEXP (x1, 1) : XEXP (x1, 0));
10730 if ((GET_RTX_CLASS (GET_CODE (x0)) == '2'
10731 || GET_RTX_CLASS (GET_CODE (x0)) == 'c')
10732 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10733 return 2 + 2 * count_rtxs (x1)
10734 + count_rtxs (x == XEXP (x0, 0)
10735 ? XEXP (x0, 1) : XEXP (x0, 0));
10738 fmt = GET_RTX_FORMAT (code);
10739 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10740 if (fmt[i] == 'e')
10741 ret += count_rtxs (XEXP (x, i));
10743 return ret;
10746 /* Utility function for following routine. Called when X is part of a value
10747 being stored into last_set_value. Sets last_set_table_tick
10748 for each register mentioned. Similar to mention_regs in cse.c */
10750 static void
10751 update_table_tick (rtx x)
10753 enum rtx_code code = GET_CODE (x);
10754 const char *fmt = GET_RTX_FORMAT (code);
10755 int i;
10757 if (code == REG)
10759 unsigned int regno = REGNO (x);
10760 unsigned int endregno
10761 = regno + (regno < FIRST_PSEUDO_REGISTER
10762 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
10763 unsigned int r;
10765 for (r = regno; r < endregno; r++)
10766 reg_stat[r].last_set_table_tick = label_tick;
10768 return;
10771 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10772 /* Note that we can't have an "E" in values stored; see
10773 get_last_value_validate. */
10774 if (fmt[i] == 'e')
10776 /* Check for identical subexpressions. If x contains
10777 identical subexpression we only have to traverse one of
10778 them. */
10779 if (i == 0 && ARITHMETIC_P (x))
10781 /* Note that at this point x1 has already been
10782 processed. */
10783 rtx x0 = XEXP (x, 0);
10784 rtx x1 = XEXP (x, 1);
10786 /* If x0 and x1 are identical then there is no need to
10787 process x0. */
10788 if (x0 == x1)
10789 break;
10791 /* If x0 is identical to a subexpression of x1 then while
10792 processing x1, x0 has already been processed. Thus we
10793 are done with x. */
10794 if (ARITHMETIC_P (x1)
10795 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10796 break;
10798 /* If x1 is identical to a subexpression of x0 then we
10799 still have to process the rest of x0. */
10800 if (ARITHMETIC_P (x0)
10801 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10803 update_table_tick (XEXP (x0, x1 == XEXP (x0, 0) ? 1 : 0));
10804 break;
10808 update_table_tick (XEXP (x, i));
10812 /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we
10813 are saying that the register is clobbered and we no longer know its
10814 value. If INSN is zero, don't update reg_stat[].last_set; this is
10815 only permitted with VALUE also zero and is used to invalidate the
10816 register. */
10818 static void
10819 record_value_for_reg (rtx reg, rtx insn, rtx value)
10821 unsigned int regno = REGNO (reg);
10822 unsigned int endregno
10823 = regno + (regno < FIRST_PSEUDO_REGISTER
10824 ? hard_regno_nregs[regno][GET_MODE (reg)] : 1);
10825 unsigned int i;
10827 /* If VALUE contains REG and we have a previous value for REG, substitute
10828 the previous value. */
10829 if (value && insn && reg_overlap_mentioned_p (reg, value))
10831 rtx tem;
10833 /* Set things up so get_last_value is allowed to see anything set up to
10834 our insn. */
10835 subst_low_cuid = INSN_CUID (insn);
10836 tem = get_last_value (reg);
10838 /* If TEM is simply a binary operation with two CLOBBERs as operands,
10839 it isn't going to be useful and will take a lot of time to process,
10840 so just use the CLOBBER. */
10842 if (tem)
10844 if (ARITHMETIC_P (tem)
10845 && GET_CODE (XEXP (tem, 0)) == CLOBBER
10846 && GET_CODE (XEXP (tem, 1)) == CLOBBER)
10847 tem = XEXP (tem, 0);
10848 else if (count_occurrences (value, reg, 1) >= 2)
10850 /* If there are two or more occurrences of REG in VALUE,
10851 prevent the value from growing too much. */
10852 if (count_rtxs (tem) > MAX_LAST_VALUE_RTL)
10853 tem = gen_rtx_CLOBBER (GET_MODE (tem), const0_rtx);
10856 value = replace_rtx (copy_rtx (value), reg, tem);
10860 /* For each register modified, show we don't know its value, that
10861 we don't know about its bitwise content, that its value has been
10862 updated, and that we don't know the location of the death of the
10863 register. */
10864 for (i = regno; i < endregno; i++)
10866 if (insn)
10867 reg_stat[i].last_set = insn;
10869 reg_stat[i].last_set_value = 0;
10870 reg_stat[i].last_set_mode = 0;
10871 reg_stat[i].last_set_nonzero_bits = 0;
10872 reg_stat[i].last_set_sign_bit_copies = 0;
10873 reg_stat[i].last_death = 0;
10874 reg_stat[i].truncated_to_mode = 0;
10877 /* Mark registers that are being referenced in this value. */
10878 if (value)
10879 update_table_tick (value);
10881 /* Now update the status of each register being set.
10882 If someone is using this register in this block, set this register
10883 to invalid since we will get confused between the two lives in this
10884 basic block. This makes using this register always invalid. In cse, we
10885 scan the table to invalidate all entries using this register, but this
10886 is too much work for us. */
10888 for (i = regno; i < endregno; i++)
10890 reg_stat[i].last_set_label = label_tick;
10891 if (!insn || (value && reg_stat[i].last_set_table_tick == label_tick))
10892 reg_stat[i].last_set_invalid = 1;
10893 else
10894 reg_stat[i].last_set_invalid = 0;
10897 /* The value being assigned might refer to X (like in "x++;"). In that
10898 case, we must replace it with (clobber (const_int 0)) to prevent
10899 infinite loops. */
10900 if (value && ! get_last_value_validate (&value, insn,
10901 reg_stat[regno].last_set_label, 0))
10903 value = copy_rtx (value);
10904 if (! get_last_value_validate (&value, insn,
10905 reg_stat[regno].last_set_label, 1))
10906 value = 0;
10909 /* For the main register being modified, update the value, the mode, the
10910 nonzero bits, and the number of sign bit copies. */
10912 reg_stat[regno].last_set_value = value;
10914 if (value)
10916 enum machine_mode mode = GET_MODE (reg);
10917 subst_low_cuid = INSN_CUID (insn);
10918 reg_stat[regno].last_set_mode = mode;
10919 if (GET_MODE_CLASS (mode) == MODE_INT
10920 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10921 mode = nonzero_bits_mode;
10922 reg_stat[regno].last_set_nonzero_bits = nonzero_bits (value, mode);
10923 reg_stat[regno].last_set_sign_bit_copies
10924 = num_sign_bit_copies (value, GET_MODE (reg));
10928 /* Called via note_stores from record_dead_and_set_regs to handle one
10929 SET or CLOBBER in an insn. DATA is the instruction in which the
10930 set is occurring. */
10932 static void
10933 record_dead_and_set_regs_1 (rtx dest, rtx setter, void *data)
10935 rtx record_dead_insn = (rtx) data;
10937 if (GET_CODE (dest) == SUBREG)
10938 dest = SUBREG_REG (dest);
10940 if (!record_dead_insn)
10942 if (REG_P (dest))
10943 record_value_for_reg (dest, NULL_RTX, NULL_RTX);
10944 return;
10947 if (REG_P (dest))
10949 /* If we are setting the whole register, we know its value. Otherwise
10950 show that we don't know the value. We can handle SUBREG in
10951 some cases. */
10952 if (GET_CODE (setter) == SET && dest == SET_DEST (setter))
10953 record_value_for_reg (dest, record_dead_insn, SET_SRC (setter));
10954 else if (GET_CODE (setter) == SET
10955 && GET_CODE (SET_DEST (setter)) == SUBREG
10956 && SUBREG_REG (SET_DEST (setter)) == dest
10957 && GET_MODE_BITSIZE (GET_MODE (dest)) <= BITS_PER_WORD
10958 && subreg_lowpart_p (SET_DEST (setter)))
10959 record_value_for_reg (dest, record_dead_insn,
10960 gen_lowpart (GET_MODE (dest),
10961 SET_SRC (setter)));
10962 else
10963 record_value_for_reg (dest, record_dead_insn, NULL_RTX);
10965 else if (MEM_P (dest)
10966 /* Ignore pushes, they clobber nothing. */
10967 && ! push_operand (dest, GET_MODE (dest)))
10968 mem_last_set = INSN_CUID (record_dead_insn);
10971 /* Update the records of when each REG was most recently set or killed
10972 for the things done by INSN. This is the last thing done in processing
10973 INSN in the combiner loop.
10975 We update reg_stat[], in particular fields last_set, last_set_value,
10976 last_set_mode, last_set_nonzero_bits, last_set_sign_bit_copies,
10977 last_death, and also the similar information mem_last_set (which insn
10978 most recently modified memory) and last_call_cuid (which insn was the
10979 most recent subroutine call). */
10981 static void
10982 record_dead_and_set_regs (rtx insn)
10984 rtx link;
10985 unsigned int i;
10987 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
10989 if (REG_NOTE_KIND (link) == REG_DEAD
10990 && REG_P (XEXP (link, 0)))
10992 unsigned int regno = REGNO (XEXP (link, 0));
10993 unsigned int endregno
10994 = regno + (regno < FIRST_PSEUDO_REGISTER
10995 ? hard_regno_nregs[regno][GET_MODE (XEXP (link, 0))]
10996 : 1);
10998 for (i = regno; i < endregno; i++)
10999 reg_stat[i].last_death = insn;
11001 else if (REG_NOTE_KIND (link) == REG_INC)
11002 record_value_for_reg (XEXP (link, 0), insn, NULL_RTX);
11005 if (CALL_P (insn))
11007 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
11008 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
11010 reg_stat[i].last_set_value = 0;
11011 reg_stat[i].last_set_mode = 0;
11012 reg_stat[i].last_set_nonzero_bits = 0;
11013 reg_stat[i].last_set_sign_bit_copies = 0;
11014 reg_stat[i].last_death = 0;
11015 reg_stat[i].truncated_to_mode = 0;
11018 last_call_cuid = mem_last_set = INSN_CUID (insn);
11020 /* We can't combine into a call pattern. Remember, though, that
11021 the return value register is set at this CUID. We could
11022 still replace a register with the return value from the
11023 wrong subroutine call! */
11024 note_stores (PATTERN (insn), record_dead_and_set_regs_1, NULL_RTX);
11026 else
11027 note_stores (PATTERN (insn), record_dead_and_set_regs_1, insn);
11030 /* If a SUBREG has the promoted bit set, it is in fact a property of the
11031 register present in the SUBREG, so for each such SUBREG go back and
11032 adjust nonzero and sign bit information of the registers that are
11033 known to have some zero/sign bits set.
11035 This is needed because when combine blows the SUBREGs away, the
11036 information on zero/sign bits is lost and further combines can be
11037 missed because of that. */
11039 static void
11040 record_promoted_value (rtx insn, rtx subreg)
11042 rtx links, set;
11043 unsigned int regno = REGNO (SUBREG_REG (subreg));
11044 enum machine_mode mode = GET_MODE (subreg);
11046 if (GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)
11047 return;
11049 for (links = LOG_LINKS (insn); links;)
11051 insn = XEXP (links, 0);
11052 set = single_set (insn);
11054 if (! set || !REG_P (SET_DEST (set))
11055 || REGNO (SET_DEST (set)) != regno
11056 || GET_MODE (SET_DEST (set)) != GET_MODE (SUBREG_REG (subreg)))
11058 links = XEXP (links, 1);
11059 continue;
11062 if (reg_stat[regno].last_set == insn)
11064 if (SUBREG_PROMOTED_UNSIGNED_P (subreg) > 0)
11065 reg_stat[regno].last_set_nonzero_bits &= GET_MODE_MASK (mode);
11068 if (REG_P (SET_SRC (set)))
11070 regno = REGNO (SET_SRC (set));
11071 links = LOG_LINKS (insn);
11073 else
11074 break;
11078 /* Check if X, a register, is known to contain a value already
11079 truncated to MODE. In this case we can use a subreg to refer to
11080 the truncated value even though in the generic case we would need
11081 an explicit truncation. */
11083 static bool
11084 reg_truncated_to_mode (enum machine_mode mode, rtx x)
11086 enum machine_mode truncated = reg_stat[REGNO (x)].truncated_to_mode;
11088 if (truncated == 0 || reg_stat[REGNO (x)].truncation_label != label_tick)
11089 return false;
11090 if (GET_MODE_SIZE (truncated) <= GET_MODE_SIZE (mode))
11091 return true;
11092 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
11093 GET_MODE_BITSIZE (truncated)))
11094 return true;
11095 return false;
11098 /* X is a REG or a SUBREG. If X is some sort of a truncation record
11099 it. For non-TRULY_NOOP_TRUNCATION targets we might be able to turn
11100 a truncate into a subreg using this information. */
11102 static void
11103 record_truncated_value (rtx x)
11105 enum machine_mode truncated_mode;
11107 if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x)))
11109 enum machine_mode original_mode = GET_MODE (SUBREG_REG (x));
11110 truncated_mode = GET_MODE (x);
11112 if (GET_MODE_SIZE (original_mode) <= GET_MODE_SIZE (truncated_mode))
11113 return;
11115 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (truncated_mode),
11116 GET_MODE_BITSIZE (original_mode)))
11117 return;
11119 x = SUBREG_REG (x);
11121 /* ??? For hard-regs we now record everthing. We might be able to
11122 optimize this using last_set_mode. */
11123 else if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
11124 truncated_mode = GET_MODE (x);
11125 else
11126 return;
11128 if (reg_stat[REGNO (x)].truncated_to_mode == 0
11129 || reg_stat[REGNO (x)].truncation_label < label_tick
11130 || (GET_MODE_SIZE (truncated_mode)
11131 < GET_MODE_SIZE (reg_stat[REGNO (x)].truncated_to_mode)))
11133 reg_stat[REGNO (x)].truncated_to_mode = truncated_mode;
11134 reg_stat[REGNO (x)].truncation_label = label_tick;
11138 /* Scan X for promoted SUBREGs and truncated REGs. For each one
11139 found, note what it implies to the registers used in it. */
11141 static void
11142 check_conversions (rtx insn, rtx x)
11144 if (GET_CODE (x) == SUBREG || REG_P (x))
11146 if (GET_CODE (x) == SUBREG
11147 && SUBREG_PROMOTED_VAR_P (x)
11148 && REG_P (SUBREG_REG (x)))
11149 record_promoted_value (insn, x);
11151 record_truncated_value (x);
11153 else
11155 const char *format = GET_RTX_FORMAT (GET_CODE (x));
11156 int i, j;
11158 for (i = 0; i < GET_RTX_LENGTH (GET_CODE (x)); i++)
11159 switch (format[i])
11161 case 'e':
11162 check_conversions (insn, XEXP (x, i));
11163 break;
11164 case 'V':
11165 case 'E':
11166 if (XVEC (x, i) != 0)
11167 for (j = 0; j < XVECLEN (x, i); j++)
11168 check_conversions (insn, XVECEXP (x, i, j));
11169 break;
11174 /* Utility routine for the following function. Verify that all the registers
11175 mentioned in *LOC are valid when *LOC was part of a value set when
11176 label_tick == TICK. Return 0 if some are not.
11178 If REPLACE is nonzero, replace the invalid reference with
11179 (clobber (const_int 0)) and return 1. This replacement is useful because
11180 we often can get useful information about the form of a value (e.g., if
11181 it was produced by a shift that always produces -1 or 0) even though
11182 we don't know exactly what registers it was produced from. */
11184 static int
11185 get_last_value_validate (rtx *loc, rtx insn, int tick, int replace)
11187 rtx x = *loc;
11188 const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
11189 int len = GET_RTX_LENGTH (GET_CODE (x));
11190 int i;
11192 if (REG_P (x))
11194 unsigned int regno = REGNO (x);
11195 unsigned int endregno
11196 = regno + (regno < FIRST_PSEUDO_REGISTER
11197 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11198 unsigned int j;
11200 for (j = regno; j < endregno; j++)
11201 if (reg_stat[j].last_set_invalid
11202 /* If this is a pseudo-register that was only set once and not
11203 live at the beginning of the function, it is always valid. */
11204 || (! (regno >= FIRST_PSEUDO_REGISTER
11205 && REG_N_SETS (regno) == 1
11206 && (! REGNO_REG_SET_P
11207 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
11208 regno)))
11209 && reg_stat[j].last_set_label > tick))
11211 if (replace)
11212 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11213 return replace;
11216 return 1;
11218 /* If this is a memory reference, make sure that there were
11219 no stores after it that might have clobbered the value. We don't
11220 have alias info, so we assume any store invalidates it. */
11221 else if (MEM_P (x) && !MEM_READONLY_P (x)
11222 && INSN_CUID (insn) <= mem_last_set)
11224 if (replace)
11225 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11226 return replace;
11229 for (i = 0; i < len; i++)
11231 if (fmt[i] == 'e')
11233 /* Check for identical subexpressions. If x contains
11234 identical subexpression we only have to traverse one of
11235 them. */
11236 if (i == 1 && ARITHMETIC_P (x))
11238 /* Note that at this point x0 has already been checked
11239 and found valid. */
11240 rtx x0 = XEXP (x, 0);
11241 rtx x1 = XEXP (x, 1);
11243 /* If x0 and x1 are identical then x is also valid. */
11244 if (x0 == x1)
11245 return 1;
11247 /* If x1 is identical to a subexpression of x0 then
11248 while checking x0, x1 has already been checked. Thus
11249 it is valid and so as x. */
11250 if (ARITHMETIC_P (x0)
11251 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
11252 return 1;
11254 /* If x0 is identical to a subexpression of x1 then x is
11255 valid iff the rest of x1 is valid. */
11256 if (ARITHMETIC_P (x1)
11257 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
11258 return
11259 get_last_value_validate (&XEXP (x1,
11260 x0 == XEXP (x1, 0) ? 1 : 0),
11261 insn, tick, replace);
11264 if (get_last_value_validate (&XEXP (x, i), insn, tick,
11265 replace) == 0)
11266 return 0;
11268 /* Don't bother with these. They shouldn't occur anyway. */
11269 else if (fmt[i] == 'E')
11270 return 0;
11273 /* If we haven't found a reason for it to be invalid, it is valid. */
11274 return 1;
11277 /* Get the last value assigned to X, if known. Some registers
11278 in the value may be replaced with (clobber (const_int 0)) if their value
11279 is known longer known reliably. */
11281 static rtx
11282 get_last_value (rtx x)
11284 unsigned int regno;
11285 rtx value;
11287 /* If this is a non-paradoxical SUBREG, get the value of its operand and
11288 then convert it to the desired mode. If this is a paradoxical SUBREG,
11289 we cannot predict what values the "extra" bits might have. */
11290 if (GET_CODE (x) == SUBREG
11291 && subreg_lowpart_p (x)
11292 && (GET_MODE_SIZE (GET_MODE (x))
11293 <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
11294 && (value = get_last_value (SUBREG_REG (x))) != 0)
11295 return gen_lowpart (GET_MODE (x), value);
11297 if (!REG_P (x))
11298 return 0;
11300 regno = REGNO (x);
11301 value = reg_stat[regno].last_set_value;
11303 /* If we don't have a value, or if it isn't for this basic block and
11304 it's either a hard register, set more than once, or it's a live
11305 at the beginning of the function, return 0.
11307 Because if it's not live at the beginning of the function then the reg
11308 is always set before being used (is never used without being set).
11309 And, if it's set only once, and it's always set before use, then all
11310 uses must have the same last value, even if it's not from this basic
11311 block. */
11313 if (value == 0
11314 || (reg_stat[regno].last_set_label != label_tick
11315 && (regno < FIRST_PSEUDO_REGISTER
11316 || REG_N_SETS (regno) != 1
11317 || (REGNO_REG_SET_P
11318 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
11319 regno)))))
11320 return 0;
11322 /* If the value was set in a later insn than the ones we are processing,
11323 we can't use it even if the register was only set once. */
11324 if (INSN_CUID (reg_stat[regno].last_set) >= subst_low_cuid)
11325 return 0;
11327 /* If the value has all its registers valid, return it. */
11328 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11329 reg_stat[regno].last_set_label, 0))
11330 return value;
11332 /* Otherwise, make a copy and replace any invalid register with
11333 (clobber (const_int 0)). If that fails for some reason, return 0. */
11335 value = copy_rtx (value);
11336 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11337 reg_stat[regno].last_set_label, 1))
11338 return value;
11340 return 0;
11343 /* Return nonzero if expression X refers to a REG or to memory
11344 that is set in an instruction more recent than FROM_CUID. */
11346 static int
11347 use_crosses_set_p (rtx x, int from_cuid)
11349 const char *fmt;
11350 int i;
11351 enum rtx_code code = GET_CODE (x);
11353 if (code == REG)
11355 unsigned int regno = REGNO (x);
11356 unsigned endreg = regno + (regno < FIRST_PSEUDO_REGISTER
11357 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11359 #ifdef PUSH_ROUNDING
11360 /* Don't allow uses of the stack pointer to be moved,
11361 because we don't know whether the move crosses a push insn. */
11362 if (regno == STACK_POINTER_REGNUM && PUSH_ARGS)
11363 return 1;
11364 #endif
11365 for (; regno < endreg; regno++)
11366 if (reg_stat[regno].last_set
11367 && INSN_CUID (reg_stat[regno].last_set) > from_cuid)
11368 return 1;
11369 return 0;
11372 if (code == MEM && mem_last_set > from_cuid)
11373 return 1;
11375 fmt = GET_RTX_FORMAT (code);
11377 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11379 if (fmt[i] == 'E')
11381 int j;
11382 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11383 if (use_crosses_set_p (XVECEXP (x, i, j), from_cuid))
11384 return 1;
11386 else if (fmt[i] == 'e'
11387 && use_crosses_set_p (XEXP (x, i), from_cuid))
11388 return 1;
11390 return 0;
11393 /* Define three variables used for communication between the following
11394 routines. */
11396 static unsigned int reg_dead_regno, reg_dead_endregno;
11397 static int reg_dead_flag;
11399 /* Function called via note_stores from reg_dead_at_p.
11401 If DEST is within [reg_dead_regno, reg_dead_endregno), set
11402 reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */
11404 static void
11405 reg_dead_at_p_1 (rtx dest, rtx x, void *data ATTRIBUTE_UNUSED)
11407 unsigned int regno, endregno;
11409 if (!REG_P (dest))
11410 return;
11412 regno = REGNO (dest);
11413 endregno = regno + (regno < FIRST_PSEUDO_REGISTER
11414 ? hard_regno_nregs[regno][GET_MODE (dest)] : 1);
11416 if (reg_dead_endregno > regno && reg_dead_regno < endregno)
11417 reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1;
11420 /* Return nonzero if REG is known to be dead at INSN.
11422 We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER
11423 referencing REG, it is dead. If we hit a SET referencing REG, it is
11424 live. Otherwise, see if it is live or dead at the start of the basic
11425 block we are in. Hard regs marked as being live in NEWPAT_USED_REGS
11426 must be assumed to be always live. */
11428 static int
11429 reg_dead_at_p (rtx reg, rtx insn)
11431 basic_block block;
11432 unsigned int i;
11434 /* Set variables for reg_dead_at_p_1. */
11435 reg_dead_regno = REGNO (reg);
11436 reg_dead_endregno = reg_dead_regno + (reg_dead_regno < FIRST_PSEUDO_REGISTER
11437 ? hard_regno_nregs[reg_dead_regno]
11438 [GET_MODE (reg)]
11439 : 1);
11441 reg_dead_flag = 0;
11443 /* Check that reg isn't mentioned in NEWPAT_USED_REGS. For fixed registers
11444 we allow the machine description to decide whether use-and-clobber
11445 patterns are OK. */
11446 if (reg_dead_regno < FIRST_PSEUDO_REGISTER)
11448 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11449 if (!fixed_regs[i] && TEST_HARD_REG_BIT (newpat_used_regs, i))
11450 return 0;
11453 /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, label, or
11454 beginning of function. */
11455 for (; insn && !LABEL_P (insn) && !BARRIER_P (insn);
11456 insn = prev_nonnote_insn (insn))
11458 note_stores (PATTERN (insn), reg_dead_at_p_1, NULL);
11459 if (reg_dead_flag)
11460 return reg_dead_flag == 1 ? 1 : 0;
11462 if (find_regno_note (insn, REG_DEAD, reg_dead_regno))
11463 return 1;
11466 /* Get the basic block that we were in. */
11467 if (insn == 0)
11468 block = ENTRY_BLOCK_PTR->next_bb;
11469 else
11471 FOR_EACH_BB (block)
11472 if (insn == BB_HEAD (block))
11473 break;
11475 if (block == EXIT_BLOCK_PTR)
11476 return 0;
11479 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11480 if (REGNO_REG_SET_P (block->il.rtl->global_live_at_start, i))
11481 return 0;
11483 return 1;
11486 /* Note hard registers in X that are used. This code is similar to
11487 that in flow.c, but much simpler since we don't care about pseudos. */
11489 static void
11490 mark_used_regs_combine (rtx x)
11492 RTX_CODE code = GET_CODE (x);
11493 unsigned int regno;
11494 int i;
11496 switch (code)
11498 case LABEL_REF:
11499 case SYMBOL_REF:
11500 case CONST_INT:
11501 case CONST:
11502 case CONST_DOUBLE:
11503 case CONST_VECTOR:
11504 case PC:
11505 case ADDR_VEC:
11506 case ADDR_DIFF_VEC:
11507 case ASM_INPUT:
11508 #ifdef HAVE_cc0
11509 /* CC0 must die in the insn after it is set, so we don't need to take
11510 special note of it here. */
11511 case CC0:
11512 #endif
11513 return;
11515 case CLOBBER:
11516 /* If we are clobbering a MEM, mark any hard registers inside the
11517 address as used. */
11518 if (MEM_P (XEXP (x, 0)))
11519 mark_used_regs_combine (XEXP (XEXP (x, 0), 0));
11520 return;
11522 case REG:
11523 regno = REGNO (x);
11524 /* A hard reg in a wide mode may really be multiple registers.
11525 If so, mark all of them just like the first. */
11526 if (regno < FIRST_PSEUDO_REGISTER)
11528 unsigned int endregno, r;
11530 /* None of this applies to the stack, frame or arg pointers. */
11531 if (regno == STACK_POINTER_REGNUM
11532 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
11533 || regno == HARD_FRAME_POINTER_REGNUM
11534 #endif
11535 #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
11536 || (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
11537 #endif
11538 || regno == FRAME_POINTER_REGNUM)
11539 return;
11541 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11542 for (r = regno; r < endregno; r++)
11543 SET_HARD_REG_BIT (newpat_used_regs, r);
11545 return;
11547 case SET:
11549 /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in
11550 the address. */
11551 rtx testreg = SET_DEST (x);
11553 while (GET_CODE (testreg) == SUBREG
11554 || GET_CODE (testreg) == ZERO_EXTRACT
11555 || GET_CODE (testreg) == STRICT_LOW_PART)
11556 testreg = XEXP (testreg, 0);
11558 if (MEM_P (testreg))
11559 mark_used_regs_combine (XEXP (testreg, 0));
11561 mark_used_regs_combine (SET_SRC (x));
11563 return;
11565 default:
11566 break;
11569 /* Recursively scan the operands of this expression. */
11572 const char *fmt = GET_RTX_FORMAT (code);
11574 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11576 if (fmt[i] == 'e')
11577 mark_used_regs_combine (XEXP (x, i));
11578 else if (fmt[i] == 'E')
11580 int j;
11582 for (j = 0; j < XVECLEN (x, i); j++)
11583 mark_used_regs_combine (XVECEXP (x, i, j));
11589 /* Remove register number REGNO from the dead registers list of INSN.
11591 Return the note used to record the death, if there was one. */
11594 remove_death (unsigned int regno, rtx insn)
11596 rtx note = find_regno_note (insn, REG_DEAD, regno);
11598 if (note)
11600 REG_N_DEATHS (regno)--;
11601 remove_note (insn, note);
11604 return note;
11607 /* For each register (hardware or pseudo) used within expression X, if its
11608 death is in an instruction with cuid between FROM_CUID (inclusive) and
11609 TO_INSN (exclusive), put a REG_DEAD note for that register in the
11610 list headed by PNOTES.
11612 That said, don't move registers killed by maybe_kill_insn.
11614 This is done when X is being merged by combination into TO_INSN. These
11615 notes will then be distributed as needed. */
11617 static void
11618 move_deaths (rtx x, rtx maybe_kill_insn, int from_cuid, rtx to_insn,
11619 rtx *pnotes)
11621 const char *fmt;
11622 int len, i;
11623 enum rtx_code code = GET_CODE (x);
11625 if (code == REG)
11627 unsigned int regno = REGNO (x);
11628 rtx where_dead = reg_stat[regno].last_death;
11629 rtx before_dead, after_dead;
11631 /* Don't move the register if it gets killed in between from and to. */
11632 if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn)
11633 && ! reg_referenced_p (x, maybe_kill_insn))
11634 return;
11636 /* WHERE_DEAD could be a USE insn made by combine, so first we
11637 make sure that we have insns with valid INSN_CUID values. */
11638 before_dead = where_dead;
11639 while (before_dead && INSN_UID (before_dead) > max_uid_cuid)
11640 before_dead = PREV_INSN (before_dead);
11642 after_dead = where_dead;
11643 while (after_dead && INSN_UID (after_dead) > max_uid_cuid)
11644 after_dead = NEXT_INSN (after_dead);
11646 if (before_dead && after_dead
11647 && INSN_CUID (before_dead) >= from_cuid
11648 && (INSN_CUID (after_dead) < INSN_CUID (to_insn)
11649 || (where_dead != after_dead
11650 && INSN_CUID (after_dead) == INSN_CUID (to_insn))))
11652 rtx note = remove_death (regno, where_dead);
11654 /* It is possible for the call above to return 0. This can occur
11655 when last_death points to I2 or I1 that we combined with.
11656 In that case make a new note.
11658 We must also check for the case where X is a hard register
11659 and NOTE is a death note for a range of hard registers
11660 including X. In that case, we must put REG_DEAD notes for
11661 the remaining registers in place of NOTE. */
11663 if (note != 0 && regno < FIRST_PSEUDO_REGISTER
11664 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11665 > GET_MODE_SIZE (GET_MODE (x))))
11667 unsigned int deadregno = REGNO (XEXP (note, 0));
11668 unsigned int deadend
11669 = (deadregno + hard_regno_nregs[deadregno]
11670 [GET_MODE (XEXP (note, 0))]);
11671 unsigned int ourend
11672 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11673 unsigned int i;
11675 for (i = deadregno; i < deadend; i++)
11676 if (i < regno || i >= ourend)
11677 REG_NOTES (where_dead)
11678 = gen_rtx_EXPR_LIST (REG_DEAD,
11679 regno_reg_rtx[i],
11680 REG_NOTES (where_dead));
11683 /* If we didn't find any note, or if we found a REG_DEAD note that
11684 covers only part of the given reg, and we have a multi-reg hard
11685 register, then to be safe we must check for REG_DEAD notes
11686 for each register other than the first. They could have
11687 their own REG_DEAD notes lying around. */
11688 else if ((note == 0
11689 || (note != 0
11690 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11691 < GET_MODE_SIZE (GET_MODE (x)))))
11692 && regno < FIRST_PSEUDO_REGISTER
11693 && hard_regno_nregs[regno][GET_MODE (x)] > 1)
11695 unsigned int ourend
11696 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11697 unsigned int i, offset;
11698 rtx oldnotes = 0;
11700 if (note)
11701 offset = hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))];
11702 else
11703 offset = 1;
11705 for (i = regno + offset; i < ourend; i++)
11706 move_deaths (regno_reg_rtx[i],
11707 maybe_kill_insn, from_cuid, to_insn, &oldnotes);
11710 if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x))
11712 XEXP (note, 1) = *pnotes;
11713 *pnotes = note;
11715 else
11716 *pnotes = gen_rtx_EXPR_LIST (REG_DEAD, x, *pnotes);
11718 REG_N_DEATHS (regno)++;
11721 return;
11724 else if (GET_CODE (x) == SET)
11726 rtx dest = SET_DEST (x);
11728 move_deaths (SET_SRC (x), maybe_kill_insn, from_cuid, to_insn, pnotes);
11730 /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG
11731 that accesses one word of a multi-word item, some
11732 piece of everything register in the expression is used by
11733 this insn, so remove any old death. */
11734 /* ??? So why do we test for equality of the sizes? */
11736 if (GET_CODE (dest) == ZERO_EXTRACT
11737 || GET_CODE (dest) == STRICT_LOW_PART
11738 || (GET_CODE (dest) == SUBREG
11739 && (((GET_MODE_SIZE (GET_MODE (dest))
11740 + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
11741 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))
11742 + UNITS_PER_WORD - 1) / UNITS_PER_WORD))))
11744 move_deaths (dest, maybe_kill_insn, from_cuid, to_insn, pnotes);
11745 return;
11748 /* If this is some other SUBREG, we know it replaces the entire
11749 value, so use that as the destination. */
11750 if (GET_CODE (dest) == SUBREG)
11751 dest = SUBREG_REG (dest);
11753 /* If this is a MEM, adjust deaths of anything used in the address.
11754 For a REG (the only other possibility), the entire value is
11755 being replaced so the old value is not used in this insn. */
11757 if (MEM_P (dest))
11758 move_deaths (XEXP (dest, 0), maybe_kill_insn, from_cuid,
11759 to_insn, pnotes);
11760 return;
11763 else if (GET_CODE (x) == CLOBBER)
11764 return;
11766 len = GET_RTX_LENGTH (code);
11767 fmt = GET_RTX_FORMAT (code);
11769 for (i = 0; i < len; i++)
11771 if (fmt[i] == 'E')
11773 int j;
11774 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11775 move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_cuid,
11776 to_insn, pnotes);
11778 else if (fmt[i] == 'e')
11779 move_deaths (XEXP (x, i), maybe_kill_insn, from_cuid, to_insn, pnotes);
11783 /* Return 1 if X is the target of a bit-field assignment in BODY, the
11784 pattern of an insn. X must be a REG. */
11786 static int
11787 reg_bitfield_target_p (rtx x, rtx body)
11789 int i;
11791 if (GET_CODE (body) == SET)
11793 rtx dest = SET_DEST (body);
11794 rtx target;
11795 unsigned int regno, tregno, endregno, endtregno;
11797 if (GET_CODE (dest) == ZERO_EXTRACT)
11798 target = XEXP (dest, 0);
11799 else if (GET_CODE (dest) == STRICT_LOW_PART)
11800 target = SUBREG_REG (XEXP (dest, 0));
11801 else
11802 return 0;
11804 if (GET_CODE (target) == SUBREG)
11805 target = SUBREG_REG (target);
11807 if (!REG_P (target))
11808 return 0;
11810 tregno = REGNO (target), regno = REGNO (x);
11811 if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER)
11812 return target == x;
11814 endtregno = tregno + hard_regno_nregs[tregno][GET_MODE (target)];
11815 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11817 return endregno > tregno && regno < endtregno;
11820 else if (GET_CODE (body) == PARALLEL)
11821 for (i = XVECLEN (body, 0) - 1; i >= 0; i--)
11822 if (reg_bitfield_target_p (x, XVECEXP (body, 0, i)))
11823 return 1;
11825 return 0;
11828 /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them
11829 as appropriate. I3 and I2 are the insns resulting from the combination
11830 insns including FROM (I2 may be zero).
11832 ELIM_I2 and ELIM_I1 are either zero or registers that we know will
11833 not need REG_DEAD notes because they are being substituted for. This
11834 saves searching in the most common cases.
11836 Each note in the list is either ignored or placed on some insns, depending
11837 on the type of note. */
11839 static void
11840 distribute_notes (rtx notes, rtx from_insn, rtx i3, rtx i2, rtx elim_i2,
11841 rtx elim_i1)
11843 rtx note, next_note;
11844 rtx tem;
11846 for (note = notes; note; note = next_note)
11848 rtx place = 0, place2 = 0;
11850 next_note = XEXP (note, 1);
11851 switch (REG_NOTE_KIND (note))
11853 case REG_BR_PROB:
11854 case REG_BR_PRED:
11855 /* Doesn't matter much where we put this, as long as it's somewhere.
11856 It is preferable to keep these notes on branches, which is most
11857 likely to be i3. */
11858 place = i3;
11859 break;
11861 case REG_VALUE_PROFILE:
11862 /* Just get rid of this note, as it is unused later anyway. */
11863 break;
11865 case REG_NON_LOCAL_GOTO:
11866 if (JUMP_P (i3))
11867 place = i3;
11868 else
11870 gcc_assert (i2 && JUMP_P (i2));
11871 place = i2;
11873 break;
11875 case REG_EH_REGION:
11876 /* These notes must remain with the call or trapping instruction. */
11877 if (CALL_P (i3))
11878 place = i3;
11879 else if (i2 && CALL_P (i2))
11880 place = i2;
11881 else
11883 gcc_assert (flag_non_call_exceptions);
11884 if (may_trap_p (i3))
11885 place = i3;
11886 else if (i2 && may_trap_p (i2))
11887 place = i2;
11888 /* ??? Otherwise assume we've combined things such that we
11889 can now prove that the instructions can't trap. Drop the
11890 note in this case. */
11892 break;
11894 case REG_NORETURN:
11895 case REG_SETJMP:
11896 /* These notes must remain with the call. It should not be
11897 possible for both I2 and I3 to be a call. */
11898 if (CALL_P (i3))
11899 place = i3;
11900 else
11902 gcc_assert (i2 && CALL_P (i2));
11903 place = i2;
11905 break;
11907 case REG_UNUSED:
11908 /* Any clobbers for i3 may still exist, and so we must process
11909 REG_UNUSED notes from that insn.
11911 Any clobbers from i2 or i1 can only exist if they were added by
11912 recog_for_combine. In that case, recog_for_combine created the
11913 necessary REG_UNUSED notes. Trying to keep any original
11914 REG_UNUSED notes from these insns can cause incorrect output
11915 if it is for the same register as the original i3 dest.
11916 In that case, we will notice that the register is set in i3,
11917 and then add a REG_UNUSED note for the destination of i3, which
11918 is wrong. However, it is possible to have REG_UNUSED notes from
11919 i2 or i1 for register which were both used and clobbered, so
11920 we keep notes from i2 or i1 if they will turn into REG_DEAD
11921 notes. */
11923 /* If this register is set or clobbered in I3, put the note there
11924 unless there is one already. */
11925 if (reg_set_p (XEXP (note, 0), PATTERN (i3)))
11927 if (from_insn != i3)
11928 break;
11930 if (! (REG_P (XEXP (note, 0))
11931 ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0)))
11932 : find_reg_note (i3, REG_UNUSED, XEXP (note, 0))))
11933 place = i3;
11935 /* Otherwise, if this register is used by I3, then this register
11936 now dies here, so we must put a REG_DEAD note here unless there
11937 is one already. */
11938 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))
11939 && ! (REG_P (XEXP (note, 0))
11940 ? find_regno_note (i3, REG_DEAD,
11941 REGNO (XEXP (note, 0)))
11942 : find_reg_note (i3, REG_DEAD, XEXP (note, 0))))
11944 PUT_REG_NOTE_KIND (note, REG_DEAD);
11945 place = i3;
11947 break;
11949 case REG_EQUAL:
11950 case REG_EQUIV:
11951 case REG_NOALIAS:
11952 /* These notes say something about results of an insn. We can
11953 only support them if they used to be on I3 in which case they
11954 remain on I3. Otherwise they are ignored.
11956 If the note refers to an expression that is not a constant, we
11957 must also ignore the note since we cannot tell whether the
11958 equivalence is still true. It might be possible to do
11959 slightly better than this (we only have a problem if I2DEST
11960 or I1DEST is present in the expression), but it doesn't
11961 seem worth the trouble. */
11963 if (from_insn == i3
11964 && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0))))
11965 place = i3;
11966 break;
11968 case REG_INC:
11969 case REG_NO_CONFLICT:
11970 /* These notes say something about how a register is used. They must
11971 be present on any use of the register in I2 or I3. */
11972 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3)))
11973 place = i3;
11975 if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2)))
11977 if (place)
11978 place2 = i2;
11979 else
11980 place = i2;
11982 break;
11984 case REG_LABEL:
11985 /* This can show up in several ways -- either directly in the
11986 pattern, or hidden off in the constant pool with (or without?)
11987 a REG_EQUAL note. */
11988 /* ??? Ignore the without-reg_equal-note problem for now. */
11989 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))
11990 || ((tem = find_reg_note (i3, REG_EQUAL, NULL_RTX))
11991 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
11992 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0)))
11993 place = i3;
11995 if (i2
11996 && (reg_mentioned_p (XEXP (note, 0), PATTERN (i2))
11997 || ((tem = find_reg_note (i2, REG_EQUAL, NULL_RTX))
11998 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
11999 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0))))
12001 if (place)
12002 place2 = i2;
12003 else
12004 place = i2;
12007 /* Don't attach REG_LABEL note to a JUMP_INSN. Add
12008 a JUMP_LABEL instead or decrement LABEL_NUSES. */
12009 if (place && JUMP_P (place))
12011 rtx label = JUMP_LABEL (place);
12013 if (!label)
12014 JUMP_LABEL (place) = XEXP (note, 0);
12015 else
12017 gcc_assert (label == XEXP (note, 0));
12018 if (LABEL_P (label))
12019 LABEL_NUSES (label)--;
12021 place = 0;
12023 if (place2 && JUMP_P (place2))
12025 rtx label = JUMP_LABEL (place2);
12027 if (!label)
12028 JUMP_LABEL (place2) = XEXP (note, 0);
12029 else
12031 gcc_assert (label == XEXP (note, 0));
12032 if (LABEL_P (label))
12033 LABEL_NUSES (label)--;
12035 place2 = 0;
12037 break;
12039 case REG_NONNEG:
12040 /* This note says something about the value of a register prior
12041 to the execution of an insn. It is too much trouble to see
12042 if the note is still correct in all situations. It is better
12043 to simply delete it. */
12044 break;
12046 case REG_RETVAL:
12047 /* If the insn previously containing this note still exists,
12048 put it back where it was. Otherwise move it to the previous
12049 insn. Adjust the corresponding REG_LIBCALL note. */
12050 if (!NOTE_P (from_insn))
12051 place = from_insn;
12052 else
12054 tem = find_reg_note (XEXP (note, 0), REG_LIBCALL, NULL_RTX);
12055 place = prev_real_insn (from_insn);
12056 if (tem && place)
12057 XEXP (tem, 0) = place;
12058 /* If we're deleting the last remaining instruction of a
12059 libcall sequence, don't add the notes. */
12060 else if (XEXP (note, 0) == from_insn)
12061 tem = place = 0;
12062 /* Don't add the dangling REG_RETVAL note. */
12063 else if (! tem)
12064 place = 0;
12066 break;
12068 case REG_LIBCALL:
12069 /* This is handled similarly to REG_RETVAL. */
12070 if (!NOTE_P (from_insn))
12071 place = from_insn;
12072 else
12074 tem = find_reg_note (XEXP (note, 0), REG_RETVAL, NULL_RTX);
12075 place = next_real_insn (from_insn);
12076 if (tem && place)
12077 XEXP (tem, 0) = place;
12078 /* If we're deleting the last remaining instruction of a
12079 libcall sequence, don't add the notes. */
12080 else if (XEXP (note, 0) == from_insn)
12081 tem = place = 0;
12082 /* Don't add the dangling REG_LIBCALL note. */
12083 else if (! tem)
12084 place = 0;
12086 break;
12088 case REG_DEAD:
12089 /* If the register is used as an input in I3, it dies there.
12090 Similarly for I2, if it is nonzero and adjacent to I3.
12092 If the register is not used as an input in either I3 or I2
12093 and it is not one of the registers we were supposed to eliminate,
12094 there are two possibilities. We might have a non-adjacent I2
12095 or we might have somehow eliminated an additional register
12096 from a computation. For example, we might have had A & B where
12097 we discover that B will always be zero. In this case we will
12098 eliminate the reference to A.
12100 In both cases, we must search to see if we can find a previous
12101 use of A and put the death note there. */
12103 if (from_insn
12104 && CALL_P (from_insn)
12105 && find_reg_fusage (from_insn, USE, XEXP (note, 0)))
12106 place = from_insn;
12107 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)))
12108 place = i3;
12109 else if (i2 != 0 && next_nonnote_insn (i2) == i3
12110 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12111 place = i2;
12113 if (place == 0
12114 && (rtx_equal_p (XEXP (note, 0), elim_i2)
12115 || rtx_equal_p (XEXP (note, 0), elim_i1)))
12116 break;
12118 if (place == 0)
12120 basic_block bb = this_basic_block;
12122 /* You might think you could search back from FROM_INSN
12123 rather than from I3, but combine tries to split invalid
12124 combined instructions. This can result in the old I2
12125 or I1 moving later in the insn sequence. */
12126 for (tem = PREV_INSN (i3); place == 0; tem = PREV_INSN (tem))
12128 if (! INSN_P (tem))
12130 if (tem == BB_HEAD (bb))
12131 break;
12132 continue;
12135 /* If the register is being set at TEM, see if that is all
12136 TEM is doing. If so, delete TEM. Otherwise, make this
12137 into a REG_UNUSED note instead. Don't delete sets to
12138 global register vars. */
12139 if ((REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER
12140 || !global_regs[REGNO (XEXP (note, 0))])
12141 && reg_set_p (XEXP (note, 0), PATTERN (tem)))
12143 rtx set = single_set (tem);
12144 rtx inner_dest = 0;
12145 #ifdef HAVE_cc0
12146 rtx cc0_setter = NULL_RTX;
12147 #endif
12149 if (set != 0)
12150 for (inner_dest = SET_DEST (set);
12151 (GET_CODE (inner_dest) == STRICT_LOW_PART
12152 || GET_CODE (inner_dest) == SUBREG
12153 || GET_CODE (inner_dest) == ZERO_EXTRACT);
12154 inner_dest = XEXP (inner_dest, 0))
12157 /* Verify that it was the set, and not a clobber that
12158 modified the register.
12160 CC0 targets must be careful to maintain setter/user
12161 pairs. If we cannot delete the setter due to side
12162 effects, mark the user with an UNUSED note instead
12163 of deleting it. */
12165 if (set != 0 && ! side_effects_p (SET_SRC (set))
12166 && rtx_equal_p (XEXP (note, 0), inner_dest)
12167 #ifdef HAVE_cc0
12168 && (! reg_mentioned_p (cc0_rtx, SET_SRC (set))
12169 || ((cc0_setter = prev_cc0_setter (tem)) != NULL
12170 && sets_cc0_p (PATTERN (cc0_setter)) > 0))
12171 #endif
12174 /* Move the notes and links of TEM elsewhere.
12175 This might delete other dead insns recursively.
12176 First set the pattern to something that won't use
12177 any register. */
12178 rtx old_notes = REG_NOTES (tem);
12180 PATTERN (tem) = pc_rtx;
12181 REG_NOTES (tem) = NULL;
12183 distribute_notes (old_notes, tem, tem, NULL_RTX,
12184 NULL_RTX, NULL_RTX);
12185 distribute_links (LOG_LINKS (tem));
12187 SET_INSN_DELETED (tem);
12189 #ifdef HAVE_cc0
12190 /* Delete the setter too. */
12191 if (cc0_setter)
12193 PATTERN (cc0_setter) = pc_rtx;
12194 old_notes = REG_NOTES (cc0_setter);
12195 REG_NOTES (cc0_setter) = NULL;
12197 distribute_notes (old_notes, cc0_setter,
12198 cc0_setter, NULL_RTX,
12199 NULL_RTX, NULL_RTX);
12200 distribute_links (LOG_LINKS (cc0_setter));
12202 SET_INSN_DELETED (cc0_setter);
12204 #endif
12206 else
12208 PUT_REG_NOTE_KIND (note, REG_UNUSED);
12210 /* If there isn't already a REG_UNUSED note, put one
12211 here. Do not place a REG_DEAD note, even if
12212 the register is also used here; that would not
12213 match the algorithm used in lifetime analysis
12214 and can cause the consistency check in the
12215 scheduler to fail. */
12216 if (! find_regno_note (tem, REG_UNUSED,
12217 REGNO (XEXP (note, 0))))
12218 place = tem;
12219 break;
12222 else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem))
12223 || (CALL_P (tem)
12224 && find_reg_fusage (tem, USE, XEXP (note, 0))))
12226 /* This may not be the correct place for the death
12227 note if FROM_INSN is before TEM, and the reg is
12228 set between FROM_INSN and TEM. The reg might
12229 die two or more times. An existing death note
12230 means we are looking at the wrong live range. */
12231 if (from_insn
12232 && INSN_CUID (from_insn) < INSN_CUID (tem)
12233 && find_regno_note (tem, REG_DEAD,
12234 REGNO (XEXP (note, 0))))
12236 tem = from_insn;
12237 if (tem == BB_HEAD (bb))
12238 break;
12239 continue;
12242 place = tem;
12244 /* If we are doing a 3->2 combination, and we have a
12245 register which formerly died in i3 and was not used
12246 by i2, which now no longer dies in i3 and is used in
12247 i2 but does not die in i2, and place is between i2
12248 and i3, then we may need to move a link from place to
12249 i2. */
12250 if (i2 && INSN_UID (place) <= max_uid_cuid
12251 && INSN_CUID (place) > INSN_CUID (i2)
12252 && from_insn
12253 && INSN_CUID (from_insn) > INSN_CUID (i2)
12254 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12256 rtx links = LOG_LINKS (place);
12257 LOG_LINKS (place) = 0;
12258 distribute_links (links);
12260 break;
12263 if (tem == BB_HEAD (bb))
12264 break;
12267 /* We haven't found an insn for the death note and it
12268 is still a REG_DEAD note, but we have hit the beginning
12269 of the block. If the existing life info says the reg
12270 was dead, there's nothing left to do. Otherwise, we'll
12271 need to do a global life update after combine. */
12272 if (REG_NOTE_KIND (note) == REG_DEAD && place == 0
12273 && REGNO_REG_SET_P (bb->il.rtl->global_live_at_start,
12274 REGNO (XEXP (note, 0))))
12275 SET_BIT (refresh_blocks, this_basic_block->index);
12278 /* If the register is set or already dead at PLACE, we needn't do
12279 anything with this note if it is still a REG_DEAD note.
12280 We check here if it is set at all, not if is it totally replaced,
12281 which is what `dead_or_set_p' checks, so also check for it being
12282 set partially. */
12284 if (place && REG_NOTE_KIND (note) == REG_DEAD)
12286 unsigned int regno = REGNO (XEXP (note, 0));
12288 /* Similarly, if the instruction on which we want to place
12289 the note is a noop, we'll need do a global live update
12290 after we remove them in delete_noop_moves. */
12291 if (noop_move_p (place))
12292 SET_BIT (refresh_blocks, this_basic_block->index);
12294 if (dead_or_set_p (place, XEXP (note, 0))
12295 || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place)))
12297 /* Unless the register previously died in PLACE, clear
12298 last_death. [I no longer understand why this is
12299 being done.] */
12300 if (reg_stat[regno].last_death != place)
12301 reg_stat[regno].last_death = 0;
12302 place = 0;
12304 else
12305 reg_stat[regno].last_death = place;
12307 /* If this is a death note for a hard reg that is occupying
12308 multiple registers, ensure that we are still using all
12309 parts of the object. If we find a piece of the object
12310 that is unused, we must arrange for an appropriate REG_DEAD
12311 note to be added for it. However, we can't just emit a USE
12312 and tag the note to it, since the register might actually
12313 be dead; so we recourse, and the recursive call then finds
12314 the previous insn that used this register. */
12316 if (place && regno < FIRST_PSEUDO_REGISTER
12317 && hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))] > 1)
12319 unsigned int endregno
12320 = regno + hard_regno_nregs[regno]
12321 [GET_MODE (XEXP (note, 0))];
12322 int all_used = 1;
12323 unsigned int i;
12325 for (i = regno; i < endregno; i++)
12326 if ((! refers_to_regno_p (i, i + 1, PATTERN (place), 0)
12327 && ! find_regno_fusage (place, USE, i))
12328 || dead_or_set_regno_p (place, i))
12329 all_used = 0;
12331 if (! all_used)
12333 /* Put only REG_DEAD notes for pieces that are
12334 not already dead or set. */
12336 for (i = regno; i < endregno;
12337 i += hard_regno_nregs[i][reg_raw_mode[i]])
12339 rtx piece = regno_reg_rtx[i];
12340 basic_block bb = this_basic_block;
12342 if (! dead_or_set_p (place, piece)
12343 && ! reg_bitfield_target_p (piece,
12344 PATTERN (place)))
12346 rtx new_note
12347 = gen_rtx_EXPR_LIST (REG_DEAD, piece, NULL_RTX);
12349 distribute_notes (new_note, place, place,
12350 NULL_RTX, NULL_RTX, NULL_RTX);
12352 else if (! refers_to_regno_p (i, i + 1,
12353 PATTERN (place), 0)
12354 && ! find_regno_fusage (place, USE, i))
12355 for (tem = PREV_INSN (place); ;
12356 tem = PREV_INSN (tem))
12358 if (! INSN_P (tem))
12360 if (tem == BB_HEAD (bb))
12362 SET_BIT (refresh_blocks,
12363 this_basic_block->index);
12364 break;
12366 continue;
12368 if (dead_or_set_p (tem, piece)
12369 || reg_bitfield_target_p (piece,
12370 PATTERN (tem)))
12372 REG_NOTES (tem)
12373 = gen_rtx_EXPR_LIST (REG_UNUSED, piece,
12374 REG_NOTES (tem));
12375 break;
12381 place = 0;
12385 break;
12387 default:
12388 /* Any other notes should not be present at this point in the
12389 compilation. */
12390 gcc_unreachable ();
12393 if (place)
12395 XEXP (note, 1) = REG_NOTES (place);
12396 REG_NOTES (place) = note;
12398 else if ((REG_NOTE_KIND (note) == REG_DEAD
12399 || REG_NOTE_KIND (note) == REG_UNUSED)
12400 && REG_P (XEXP (note, 0)))
12401 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
12403 if (place2)
12405 if ((REG_NOTE_KIND (note) == REG_DEAD
12406 || REG_NOTE_KIND (note) == REG_UNUSED)
12407 && REG_P (XEXP (note, 0)))
12408 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
12410 REG_NOTES (place2) = gen_rtx_fmt_ee (GET_CODE (note),
12411 REG_NOTE_KIND (note),
12412 XEXP (note, 0),
12413 REG_NOTES (place2));
12418 /* Similarly to above, distribute the LOG_LINKS that used to be present on
12419 I3, I2, and I1 to new locations. This is also called to add a link
12420 pointing at I3 when I3's destination is changed. */
12422 static void
12423 distribute_links (rtx links)
12425 rtx link, next_link;
12427 for (link = links; link; link = next_link)
12429 rtx place = 0;
12430 rtx insn;
12431 rtx set, reg;
12433 next_link = XEXP (link, 1);
12435 /* If the insn that this link points to is a NOTE or isn't a single
12436 set, ignore it. In the latter case, it isn't clear what we
12437 can do other than ignore the link, since we can't tell which
12438 register it was for. Such links wouldn't be used by combine
12439 anyway.
12441 It is not possible for the destination of the target of the link to
12442 have been changed by combine. The only potential of this is if we
12443 replace I3, I2, and I1 by I3 and I2. But in that case the
12444 destination of I2 also remains unchanged. */
12446 if (NOTE_P (XEXP (link, 0))
12447 || (set = single_set (XEXP (link, 0))) == 0)
12448 continue;
12450 reg = SET_DEST (set);
12451 while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT
12452 || GET_CODE (reg) == STRICT_LOW_PART)
12453 reg = XEXP (reg, 0);
12455 /* A LOG_LINK is defined as being placed on the first insn that uses
12456 a register and points to the insn that sets the register. Start
12457 searching at the next insn after the target of the link and stop
12458 when we reach a set of the register or the end of the basic block.
12460 Note that this correctly handles the link that used to point from
12461 I3 to I2. Also note that not much searching is typically done here
12462 since most links don't point very far away. */
12464 for (insn = NEXT_INSN (XEXP (link, 0));
12465 (insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
12466 || BB_HEAD (this_basic_block->next_bb) != insn));
12467 insn = NEXT_INSN (insn))
12468 if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn)))
12470 if (reg_referenced_p (reg, PATTERN (insn)))
12471 place = insn;
12472 break;
12474 else if (CALL_P (insn)
12475 && find_reg_fusage (insn, USE, reg))
12477 place = insn;
12478 break;
12480 else if (INSN_P (insn) && reg_set_p (reg, insn))
12481 break;
12483 /* If we found a place to put the link, place it there unless there
12484 is already a link to the same insn as LINK at that point. */
12486 if (place)
12488 rtx link2;
12490 for (link2 = LOG_LINKS (place); link2; link2 = XEXP (link2, 1))
12491 if (XEXP (link2, 0) == XEXP (link, 0))
12492 break;
12494 if (link2 == 0)
12496 XEXP (link, 1) = LOG_LINKS (place);
12497 LOG_LINKS (place) = link;
12499 /* Set added_links_insn to the earliest insn we added a
12500 link to. */
12501 if (added_links_insn == 0
12502 || INSN_CUID (added_links_insn) > INSN_CUID (place))
12503 added_links_insn = place;
12509 /* Subroutine of unmentioned_reg_p and callback from for_each_rtx.
12510 Check whether the expression pointer to by LOC is a register or
12511 memory, and if so return 1 if it isn't mentioned in the rtx EXPR.
12512 Otherwise return zero. */
12514 static int
12515 unmentioned_reg_p_1 (rtx *loc, void *expr)
12517 rtx x = *loc;
12519 if (x != NULL_RTX
12520 && (REG_P (x) || MEM_P (x))
12521 && ! reg_mentioned_p (x, (rtx) expr))
12522 return 1;
12523 return 0;
12526 /* Check for any register or memory mentioned in EQUIV that is not
12527 mentioned in EXPR. This is used to restrict EQUIV to "specializations"
12528 of EXPR where some registers may have been replaced by constants. */
12530 static bool
12531 unmentioned_reg_p (rtx equiv, rtx expr)
12533 return for_each_rtx (&equiv, unmentioned_reg_p_1, expr);
12536 /* Compute INSN_CUID for INSN, which is an insn made by combine. */
12538 static int
12539 insn_cuid (rtx insn)
12541 while (insn != 0 && INSN_UID (insn) > max_uid_cuid
12542 && NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE)
12543 insn = NEXT_INSN (insn);
12545 gcc_assert (INSN_UID (insn) <= max_uid_cuid);
12547 return INSN_CUID (insn);
12550 void
12551 dump_combine_stats (FILE *file)
12553 fprintf
12554 (file,
12555 ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n",
12556 combine_attempts, combine_merges, combine_extras, combine_successes);
12559 void
12560 dump_combine_total_stats (FILE *file)
12562 fprintf
12563 (file,
12564 "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n",
12565 total_attempts, total_merges, total_extras, total_successes);
12569 static bool
12570 gate_handle_combine (void)
12572 return (optimize > 0);
12575 /* Try combining insns through substitution. */
12576 static void
12577 rest_of_handle_combine (void)
12579 int rebuild_jump_labels_after_combine
12580 = combine_instructions (get_insns (), max_reg_num ());
12582 /* Combining insns may have turned an indirect jump into a
12583 direct jump. Rebuild the JUMP_LABEL fields of jumping
12584 instructions. */
12585 if (rebuild_jump_labels_after_combine)
12587 timevar_push (TV_JUMP);
12588 rebuild_jump_labels (get_insns ());
12589 timevar_pop (TV_JUMP);
12591 delete_dead_jumptables ();
12592 cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_UPDATE_LIFE);
12596 struct tree_opt_pass pass_combine =
12598 "combine", /* name */
12599 gate_handle_combine, /* gate */
12600 rest_of_handle_combine, /* execute */
12601 NULL, /* sub */
12602 NULL, /* next */
12603 0, /* static_pass_number */
12604 TV_COMBINE, /* tv_id */
12605 0, /* properties_required */
12606 0, /* properties_provided */
12607 0, /* properties_destroyed */
12608 0, /* todo_flags_start */
12609 TODO_dump_func |
12610 TODO_ggc_collect, /* todo_flags_finish */
12611 'c' /* letter */