2006-03-15 Paul Brook <paul@codesourcery.com>
[official-gcc.git] / gcc / combine.c
blob2ff106d684cd5d3085a044ccf867ecc8f456873f
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 /* Maximum register number, which is the size of the tables below. */
144 static unsigned int combine_max_regno;
146 struct reg_stat {
147 /* Record last point of death of (hard or pseudo) register n. */
148 rtx last_death;
150 /* Record last point of modification of (hard or pseudo) register n. */
151 rtx last_set;
153 /* The next group of fields allows the recording of the last value assigned
154 to (hard or pseudo) register n. We use this information to see if an
155 operation being processed is redundant given a prior operation performed
156 on the register. For example, an `and' with a constant is redundant if
157 all the zero bits are already known to be turned off.
159 We use an approach similar to that used by cse, but change it in the
160 following ways:
162 (1) We do not want to reinitialize at each label.
163 (2) It is useful, but not critical, to know the actual value assigned
164 to a register. Often just its form is helpful.
166 Therefore, we maintain the following fields:
168 last_set_value the last value assigned
169 last_set_label records the value of label_tick when the
170 register was assigned
171 last_set_table_tick records the value of label_tick when a
172 value using the register is assigned
173 last_set_invalid set to nonzero when it is not valid
174 to use the value of this register in some
175 register's value
177 To understand the usage of these tables, it is important to understand
178 the distinction between the value in last_set_value being valid and
179 the register being validly contained in some other expression in the
180 table.
182 (The next two parameters are out of date).
184 reg_stat[i].last_set_value is valid if it is nonzero, and either
185 reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick.
187 Register I may validly appear in any expression returned for the value
188 of another register if reg_n_sets[i] is 1. It may also appear in the
189 value for register J if reg_stat[j].last_set_invalid is zero, or
190 reg_stat[i].last_set_label < reg_stat[j].last_set_label.
192 If an expression is found in the table containing a register which may
193 not validly appear in an expression, the register is replaced by
194 something that won't match, (clobber (const_int 0)). */
196 /* Record last value assigned to (hard or pseudo) register n. */
198 rtx last_set_value;
200 /* Record the value of label_tick when an expression involving register n
201 is placed in last_set_value. */
203 int last_set_table_tick;
205 /* Record the value of label_tick when the value for register n is placed in
206 last_set_value. */
208 int last_set_label;
210 /* These fields are maintained in parallel with last_set_value and are
211 used to store the mode in which the register was last set, the bits
212 that were known to be zero when it was last set, and the number of
213 sign bits copies it was known to have when it was last set. */
215 unsigned HOST_WIDE_INT last_set_nonzero_bits;
216 char last_set_sign_bit_copies;
217 ENUM_BITFIELD(machine_mode) last_set_mode : 8;
219 /* Set nonzero if references to register n in expressions should not be
220 used. last_set_invalid is set nonzero when this register is being
221 assigned to and last_set_table_tick == label_tick. */
223 char last_set_invalid;
225 /* Some registers that are set more than once and used in more than one
226 basic block are nevertheless always set in similar ways. For example,
227 a QImode register may be loaded from memory in two places on a machine
228 where byte loads zero extend.
230 We record in the following fields if a register has some leading bits
231 that are always equal to the sign bit, and what we know about the
232 nonzero bits of a register, specifically which bits are known to be
233 zero.
235 If an entry is zero, it means that we don't know anything special. */
237 unsigned char sign_bit_copies;
239 unsigned HOST_WIDE_INT nonzero_bits;
241 /* Record the value of the label_tick when the last truncation
242 happened. The field truncated_to_mode is only valid if
243 truncation_label == label_tick. */
245 int truncation_label;
247 /* Record the last truncation seen for this register. If truncation
248 is not a nop to this mode we might be able to save an explicit
249 truncation if we know that value already contains a truncated
250 value. */
252 ENUM_BITFIELD(machine_mode) truncated_to_mode : 8;
255 static struct reg_stat *reg_stat;
257 /* Record the cuid of the last insn that invalidated memory
258 (anything that writes memory, and subroutine calls, but not pushes). */
260 static int mem_last_set;
262 /* Record the cuid of the last CALL_INSN
263 so we can tell whether a potential combination crosses any calls. */
265 static int last_call_cuid;
267 /* When `subst' is called, this is the insn that is being modified
268 (by combining in a previous insn). The PATTERN of this insn
269 is still the old pattern partially modified and it should not be
270 looked at, but this may be used to examine the successors of the insn
271 to judge whether a simplification is valid. */
273 static rtx subst_insn;
275 /* This is the lowest CUID that `subst' is currently dealing with.
276 get_last_value will not return a value if the register was set at or
277 after this CUID. If not for this mechanism, we could get confused if
278 I2 or I1 in try_combine were an insn that used the old value of a register
279 to obtain a new value. In that case, we might erroneously get the
280 new value of the register when we wanted the old one. */
282 static int subst_low_cuid;
284 /* This contains any hard registers that are used in newpat; reg_dead_at_p
285 must consider all these registers to be always live. */
287 static HARD_REG_SET newpat_used_regs;
289 /* This is an insn to which a LOG_LINKS entry has been added. If this
290 insn is the earlier than I2 or I3, combine should rescan starting at
291 that location. */
293 static rtx added_links_insn;
295 /* Basic block in which we are performing combines. */
296 static basic_block this_basic_block;
298 /* A bitmap indicating which blocks had registers go dead at entry.
299 After combine, we'll need to re-do global life analysis with
300 those blocks as starting points. */
301 static sbitmap refresh_blocks;
303 /* The following array records the insn_rtx_cost for every insn
304 in the instruction stream. */
306 static int *uid_insn_cost;
308 /* Length of the currently allocated uid_insn_cost array. */
310 static int last_insn_cost;
312 /* Incremented for each label. */
314 static int label_tick;
316 /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the
317 largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */
319 static enum machine_mode nonzero_bits_mode;
321 /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can
322 be safely used. It is zero while computing them and after combine has
323 completed. This former test prevents propagating values based on
324 previously set values, which can be incorrect if a variable is modified
325 in a loop. */
327 static int nonzero_sign_valid;
330 /* Record one modification to rtl structure
331 to be undone by storing old_contents into *where. */
333 struct undo
335 struct undo *next;
336 enum { UNDO_RTX, UNDO_INT, UNDO_MODE } kind;
337 union { rtx r; int i; enum machine_mode m; } old_contents;
338 union { rtx *r; int *i; } where;
341 /* Record a bunch of changes to be undone, up to MAX_UNDO of them.
342 num_undo says how many are currently recorded.
344 other_insn is nonzero if we have modified some other insn in the process
345 of working on subst_insn. It must be verified too. */
347 struct undobuf
349 struct undo *undos;
350 struct undo *frees;
351 rtx other_insn;
354 static struct undobuf undobuf;
356 /* Number of times the pseudo being substituted for
357 was found and replaced. */
359 static int n_occurrences;
361 static rtx reg_nonzero_bits_for_combine (rtx, enum machine_mode, rtx,
362 enum machine_mode,
363 unsigned HOST_WIDE_INT,
364 unsigned HOST_WIDE_INT *);
365 static rtx reg_num_sign_bit_copies_for_combine (rtx, enum machine_mode, rtx,
366 enum machine_mode,
367 unsigned int, unsigned int *);
368 static void do_SUBST (rtx *, rtx);
369 static void do_SUBST_INT (int *, int);
370 static void init_reg_last (void);
371 static void setup_incoming_promotions (void);
372 static void set_nonzero_bits_and_sign_copies (rtx, rtx, void *);
373 static int cant_combine_insn_p (rtx);
374 static int can_combine_p (rtx, rtx, rtx, rtx, rtx *, rtx *);
375 static int combinable_i3pat (rtx, rtx *, rtx, rtx, int, rtx *);
376 static int contains_muldiv (rtx);
377 static rtx try_combine (rtx, rtx, rtx, int *);
378 static void undo_all (void);
379 static void undo_commit (void);
380 static rtx *find_split_point (rtx *, rtx);
381 static rtx subst (rtx, rtx, rtx, int, int);
382 static rtx combine_simplify_rtx (rtx, enum machine_mode, int);
383 static rtx simplify_if_then_else (rtx);
384 static rtx simplify_set (rtx);
385 static rtx simplify_logical (rtx);
386 static rtx expand_compound_operation (rtx);
387 static rtx expand_field_assignment (rtx);
388 static rtx make_extraction (enum machine_mode, rtx, HOST_WIDE_INT,
389 rtx, unsigned HOST_WIDE_INT, int, int, int);
390 static rtx extract_left_shift (rtx, int);
391 static rtx make_compound_operation (rtx, enum rtx_code);
392 static int get_pos_from_mask (unsigned HOST_WIDE_INT,
393 unsigned HOST_WIDE_INT *);
394 static rtx canon_reg_for_combine (rtx, rtx);
395 static rtx force_to_mode (rtx, enum machine_mode,
396 unsigned HOST_WIDE_INT, int);
397 static rtx if_then_else_cond (rtx, rtx *, rtx *);
398 static rtx known_cond (rtx, enum rtx_code, rtx, rtx);
399 static int rtx_equal_for_field_assignment_p (rtx, rtx);
400 static rtx make_field_assignment (rtx);
401 static rtx apply_distributive_law (rtx);
402 static rtx distribute_and_simplify_rtx (rtx, int);
403 static rtx simplify_and_const_int_1 (enum machine_mode, rtx,
404 unsigned HOST_WIDE_INT);
405 static rtx simplify_and_const_int (rtx, enum machine_mode, rtx,
406 unsigned HOST_WIDE_INT);
407 static int merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code,
408 HOST_WIDE_INT, enum machine_mode, int *);
409 static rtx simplify_shift_const_1 (enum rtx_code, enum machine_mode, rtx, int);
410 static rtx simplify_shift_const (rtx, enum rtx_code, enum machine_mode, rtx,
411 int);
412 static int recog_for_combine (rtx *, rtx, rtx *);
413 static rtx gen_lowpart_for_combine (enum machine_mode, rtx);
414 static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *);
415 static void update_table_tick (rtx);
416 static void record_value_for_reg (rtx, rtx, rtx);
417 static void check_conversions (rtx, rtx);
418 static void record_dead_and_set_regs_1 (rtx, rtx, void *);
419 static void record_dead_and_set_regs (rtx);
420 static int get_last_value_validate (rtx *, rtx, int, int);
421 static rtx get_last_value (rtx);
422 static int use_crosses_set_p (rtx, int);
423 static void reg_dead_at_p_1 (rtx, rtx, void *);
424 static int reg_dead_at_p (rtx, rtx);
425 static void move_deaths (rtx, rtx, int, rtx, rtx *);
426 static int reg_bitfield_target_p (rtx, rtx);
427 static void distribute_notes (rtx, rtx, rtx, rtx, rtx, rtx);
428 static void distribute_links (rtx);
429 static void mark_used_regs_combine (rtx);
430 static int insn_cuid (rtx);
431 static void record_promoted_value (rtx, rtx);
432 static int unmentioned_reg_p_1 (rtx *, void *);
433 static bool unmentioned_reg_p (rtx, rtx);
434 static void record_truncated_value (rtx);
435 static bool reg_truncated_to_mode (enum machine_mode, rtx);
436 static rtx gen_lowpart_or_truncate (enum machine_mode, rtx);
439 /* It is not safe to use ordinary gen_lowpart in combine.
440 See comments in gen_lowpart_for_combine. */
441 #undef RTL_HOOKS_GEN_LOWPART
442 #define RTL_HOOKS_GEN_LOWPART gen_lowpart_for_combine
444 /* Our implementation of gen_lowpart never emits a new pseudo. */
445 #undef RTL_HOOKS_GEN_LOWPART_NO_EMIT
446 #define RTL_HOOKS_GEN_LOWPART_NO_EMIT gen_lowpart_for_combine
448 #undef RTL_HOOKS_REG_NONZERO_REG_BITS
449 #define RTL_HOOKS_REG_NONZERO_REG_BITS reg_nonzero_bits_for_combine
451 #undef RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES
452 #define RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES reg_num_sign_bit_copies_for_combine
454 #undef RTL_HOOKS_REG_TRUNCATED_TO_MODE
455 #define RTL_HOOKS_REG_TRUNCATED_TO_MODE reg_truncated_to_mode
457 static const struct rtl_hooks combine_rtl_hooks = RTL_HOOKS_INITIALIZER;
460 /* Substitute NEWVAL, an rtx expression, into INTO, a place in some
461 insn. The substitution can be undone by undo_all. If INTO is already
462 set to NEWVAL, do not record this change. Because computing NEWVAL might
463 also call SUBST, we have to compute it before we put anything into
464 the undo table. */
466 static void
467 do_SUBST (rtx *into, rtx newval)
469 struct undo *buf;
470 rtx oldval = *into;
472 if (oldval == newval)
473 return;
475 /* We'd like to catch as many invalid transformations here as
476 possible. Unfortunately, there are way too many mode changes
477 that are perfectly valid, so we'd waste too much effort for
478 little gain doing the checks here. Focus on catching invalid
479 transformations involving integer constants. */
480 if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT
481 && GET_CODE (newval) == CONST_INT)
483 /* Sanity check that we're replacing oldval with a CONST_INT
484 that is a valid sign-extension for the original mode. */
485 gcc_assert (INTVAL (newval)
486 == trunc_int_for_mode (INTVAL (newval), GET_MODE (oldval)));
488 /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
489 CONST_INT is not valid, because after the replacement, the
490 original mode would be gone. Unfortunately, we can't tell
491 when do_SUBST is called to replace the operand thereof, so we
492 perform this test on oldval instead, checking whether an
493 invalid replacement took place before we got here. */
494 gcc_assert (!(GET_CODE (oldval) == SUBREG
495 && GET_CODE (SUBREG_REG (oldval)) == CONST_INT));
496 gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND
497 && GET_CODE (XEXP (oldval, 0)) == CONST_INT));
500 if (undobuf.frees)
501 buf = undobuf.frees, undobuf.frees = buf->next;
502 else
503 buf = XNEW (struct undo);
505 buf->kind = UNDO_RTX;
506 buf->where.r = into;
507 buf->old_contents.r = oldval;
508 *into = newval;
510 buf->next = undobuf.undos, undobuf.undos = buf;
513 #define SUBST(INTO, NEWVAL) do_SUBST(&(INTO), (NEWVAL))
515 /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution
516 for the value of a HOST_WIDE_INT value (including CONST_INT) is
517 not safe. */
519 static void
520 do_SUBST_INT (int *into, int newval)
522 struct undo *buf;
523 int oldval = *into;
525 if (oldval == newval)
526 return;
528 if (undobuf.frees)
529 buf = undobuf.frees, undobuf.frees = buf->next;
530 else
531 buf = XNEW (struct undo);
533 buf->kind = UNDO_INT;
534 buf->where.i = into;
535 buf->old_contents.i = oldval;
536 *into = newval;
538 buf->next = undobuf.undos, undobuf.undos = buf;
541 #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT(&(INTO), (NEWVAL))
543 /* Similar to SUBST, but just substitute the mode. This is used when
544 changing the mode of a pseudo-register, so that any other
545 references to the entry in the regno_reg_rtx array will change as
546 well. */
548 static void
549 do_SUBST_MODE (rtx *into, enum machine_mode newval)
551 struct undo *buf;
552 enum machine_mode oldval = GET_MODE (*into);
554 if (oldval == newval)
555 return;
557 if (undobuf.frees)
558 buf = undobuf.frees, undobuf.frees = buf->next;
559 else
560 buf = XNEW (struct undo);
562 buf->kind = UNDO_MODE;
563 buf->where.r = into;
564 buf->old_contents.m = oldval;
565 PUT_MODE (*into, newval);
567 buf->next = undobuf.undos, undobuf.undos = buf;
570 #define SUBST_MODE(INTO, NEWVAL) do_SUBST_MODE(&(INTO), (NEWVAL))
572 /* Subroutine of try_combine. Determine whether the combine replacement
573 patterns NEWPAT and NEWI2PAT are cheaper according to insn_rtx_cost
574 that the original instruction sequence I1, I2 and I3. Note that I1
575 and/or NEWI2PAT may be NULL_RTX. This function returns false, if the
576 costs of all instructions can be estimated, and the replacements are
577 more expensive than the original sequence. */
579 static bool
580 combine_validate_cost (rtx i1, rtx i2, rtx i3, rtx newpat, rtx newi2pat)
582 int i1_cost, i2_cost, i3_cost;
583 int new_i2_cost, new_i3_cost;
584 int old_cost, new_cost;
586 /* Lookup the original insn_rtx_costs. */
587 i2_cost = INSN_UID (i2) <= last_insn_cost
588 ? uid_insn_cost[INSN_UID (i2)] : 0;
589 i3_cost = INSN_UID (i3) <= last_insn_cost
590 ? uid_insn_cost[INSN_UID (i3)] : 0;
592 if (i1)
594 i1_cost = INSN_UID (i1) <= last_insn_cost
595 ? uid_insn_cost[INSN_UID (i1)] : 0;
596 old_cost = (i1_cost > 0 && i2_cost > 0 && i3_cost > 0)
597 ? i1_cost + i2_cost + i3_cost : 0;
599 else
601 old_cost = (i2_cost > 0 && i3_cost > 0) ? i2_cost + i3_cost : 0;
602 i1_cost = 0;
605 /* Calculate the replacement insn_rtx_costs. */
606 new_i3_cost = insn_rtx_cost (newpat);
607 if (newi2pat)
609 new_i2_cost = insn_rtx_cost (newi2pat);
610 new_cost = (new_i2_cost > 0 && new_i3_cost > 0)
611 ? new_i2_cost + new_i3_cost : 0;
613 else
615 new_cost = new_i3_cost;
616 new_i2_cost = 0;
619 if (undobuf.other_insn)
621 int old_other_cost, new_other_cost;
623 old_other_cost = (INSN_UID (undobuf.other_insn) <= last_insn_cost
624 ? uid_insn_cost[INSN_UID (undobuf.other_insn)] : 0);
625 new_other_cost = insn_rtx_cost (PATTERN (undobuf.other_insn));
626 if (old_other_cost > 0 && new_other_cost > 0)
628 old_cost += old_other_cost;
629 new_cost += new_other_cost;
631 else
632 old_cost = 0;
635 /* Disallow this recombination if both new_cost and old_cost are
636 greater than zero, and new_cost is greater than old cost. */
637 if (old_cost > 0
638 && new_cost > old_cost)
640 if (dump_file)
642 if (i1)
644 fprintf (dump_file,
645 "rejecting combination of insns %d, %d and %d\n",
646 INSN_UID (i1), INSN_UID (i2), INSN_UID (i3));
647 fprintf (dump_file, "original costs %d + %d + %d = %d\n",
648 i1_cost, i2_cost, i3_cost, old_cost);
650 else
652 fprintf (dump_file,
653 "rejecting combination of insns %d and %d\n",
654 INSN_UID (i2), INSN_UID (i3));
655 fprintf (dump_file, "original costs %d + %d = %d\n",
656 i2_cost, i3_cost, old_cost);
659 if (newi2pat)
661 fprintf (dump_file, "replacement costs %d + %d = %d\n",
662 new_i2_cost, new_i3_cost, new_cost);
664 else
665 fprintf (dump_file, "replacement cost %d\n", new_cost);
668 return false;
671 /* Update the uid_insn_cost array with the replacement costs. */
672 uid_insn_cost[INSN_UID (i2)] = new_i2_cost;
673 uid_insn_cost[INSN_UID (i3)] = new_i3_cost;
674 if (i1)
675 uid_insn_cost[INSN_UID (i1)] = 0;
677 return true;
680 /* Main entry point for combiner. F is the first insn of the function.
681 NREGS is the first unused pseudo-reg number.
683 Return nonzero if the combiner has turned an indirect jump
684 instruction into a direct jump. */
685 static int
686 combine_instructions (rtx f, unsigned int nregs)
688 rtx insn, next;
689 #ifdef HAVE_cc0
690 rtx prev;
691 #endif
692 int i;
693 unsigned int j = 0;
694 rtx links, nextlinks;
695 sbitmap_iterator sbi;
697 int new_direct_jump_p = 0;
699 combine_attempts = 0;
700 combine_merges = 0;
701 combine_extras = 0;
702 combine_successes = 0;
704 combine_max_regno = nregs;
706 rtl_hooks = combine_rtl_hooks;
708 reg_stat = XCNEWVEC (struct reg_stat, nregs);
710 init_recog_no_volatile ();
712 /* Compute maximum uid value so uid_cuid can be allocated. */
714 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
715 if (INSN_UID (insn) > i)
716 i = INSN_UID (insn);
718 uid_cuid = XNEWVEC (int, i + 1);
719 max_uid_cuid = i;
721 nonzero_bits_mode = mode_for_size (HOST_BITS_PER_WIDE_INT, MODE_INT, 0);
723 /* Don't use reg_stat[].nonzero_bits when computing it. This can cause
724 problems when, for example, we have j <<= 1 in a loop. */
726 nonzero_sign_valid = 0;
728 /* Compute the mapping from uids to cuids.
729 Cuids are numbers assigned to insns, like uids,
730 except that cuids increase monotonically through the code.
732 Scan all SETs and see if we can deduce anything about what
733 bits are known to be zero for some registers and how many copies
734 of the sign bit are known to exist for those registers.
736 Also set any known values so that we can use it while searching
737 for what bits are known to be set. */
739 label_tick = 1;
741 setup_incoming_promotions ();
743 refresh_blocks = sbitmap_alloc (last_basic_block);
744 sbitmap_zero (refresh_blocks);
746 /* Allocate array of current insn_rtx_costs. */
747 uid_insn_cost = XCNEWVEC (int, max_uid_cuid + 1);
748 last_insn_cost = max_uid_cuid;
750 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
752 uid_cuid[INSN_UID (insn)] = ++i;
753 subst_low_cuid = i;
754 subst_insn = insn;
756 if (INSN_P (insn))
758 note_stores (PATTERN (insn), set_nonzero_bits_and_sign_copies,
759 NULL);
760 record_dead_and_set_regs (insn);
762 #ifdef AUTO_INC_DEC
763 for (links = REG_NOTES (insn); links; links = XEXP (links, 1))
764 if (REG_NOTE_KIND (links) == REG_INC)
765 set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX,
766 NULL);
767 #endif
769 /* Record the current insn_rtx_cost of this instruction. */
770 if (NONJUMP_INSN_P (insn))
771 uid_insn_cost[INSN_UID (insn)] = insn_rtx_cost (PATTERN (insn));
772 if (dump_file)
773 fprintf(dump_file, "insn_cost %d: %d\n",
774 INSN_UID (insn), uid_insn_cost[INSN_UID (insn)]);
777 if (LABEL_P (insn))
778 label_tick++;
781 nonzero_sign_valid = 1;
783 /* Now scan all the insns in forward order. */
785 label_tick = 1;
786 last_call_cuid = 0;
787 mem_last_set = 0;
788 init_reg_last ();
789 setup_incoming_promotions ();
791 FOR_EACH_BB (this_basic_block)
793 for (insn = BB_HEAD (this_basic_block);
794 insn != NEXT_INSN (BB_END (this_basic_block));
795 insn = next ? next : NEXT_INSN (insn))
797 next = 0;
799 if (LABEL_P (insn))
800 label_tick++;
802 else if (INSN_P (insn))
804 /* See if we know about function return values before this
805 insn based upon SUBREG flags. */
806 check_conversions (insn, PATTERN (insn));
808 /* Try this insn with each insn it links back to. */
810 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
811 if ((next = try_combine (insn, XEXP (links, 0),
812 NULL_RTX, &new_direct_jump_p)) != 0)
813 goto retry;
815 /* Try each sequence of three linked insns ending with this one. */
817 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
819 rtx link = XEXP (links, 0);
821 /* If the linked insn has been replaced by a note, then there
822 is no point in pursuing this chain any further. */
823 if (NOTE_P (link))
824 continue;
826 for (nextlinks = LOG_LINKS (link);
827 nextlinks;
828 nextlinks = XEXP (nextlinks, 1))
829 if ((next = try_combine (insn, link,
830 XEXP (nextlinks, 0),
831 &new_direct_jump_p)) != 0)
832 goto retry;
835 #ifdef HAVE_cc0
836 /* Try to combine a jump insn that uses CC0
837 with a preceding insn that sets CC0, and maybe with its
838 logical predecessor as well.
839 This is how we make decrement-and-branch insns.
840 We need this special code because data flow connections
841 via CC0 do not get entered in LOG_LINKS. */
843 if (JUMP_P (insn)
844 && (prev = prev_nonnote_insn (insn)) != 0
845 && NONJUMP_INSN_P (prev)
846 && sets_cc0_p (PATTERN (prev)))
848 if ((next = try_combine (insn, prev,
849 NULL_RTX, &new_direct_jump_p)) != 0)
850 goto retry;
852 for (nextlinks = LOG_LINKS (prev); nextlinks;
853 nextlinks = XEXP (nextlinks, 1))
854 if ((next = try_combine (insn, prev,
855 XEXP (nextlinks, 0),
856 &new_direct_jump_p)) != 0)
857 goto retry;
860 /* Do the same for an insn that explicitly references CC0. */
861 if (NONJUMP_INSN_P (insn)
862 && (prev = prev_nonnote_insn (insn)) != 0
863 && NONJUMP_INSN_P (prev)
864 && sets_cc0_p (PATTERN (prev))
865 && GET_CODE (PATTERN (insn)) == SET
866 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (insn))))
868 if ((next = try_combine (insn, prev,
869 NULL_RTX, &new_direct_jump_p)) != 0)
870 goto retry;
872 for (nextlinks = LOG_LINKS (prev); nextlinks;
873 nextlinks = XEXP (nextlinks, 1))
874 if ((next = try_combine (insn, prev,
875 XEXP (nextlinks, 0),
876 &new_direct_jump_p)) != 0)
877 goto retry;
880 /* Finally, see if any of the insns that this insn links to
881 explicitly references CC0. If so, try this insn, that insn,
882 and its predecessor if it sets CC0. */
883 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
884 if (NONJUMP_INSN_P (XEXP (links, 0))
885 && GET_CODE (PATTERN (XEXP (links, 0))) == SET
886 && reg_mentioned_p (cc0_rtx, SET_SRC (PATTERN (XEXP (links, 0))))
887 && (prev = prev_nonnote_insn (XEXP (links, 0))) != 0
888 && NONJUMP_INSN_P (prev)
889 && sets_cc0_p (PATTERN (prev))
890 && (next = try_combine (insn, XEXP (links, 0),
891 prev, &new_direct_jump_p)) != 0)
892 goto retry;
893 #endif
895 /* Try combining an insn with two different insns whose results it
896 uses. */
897 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
898 for (nextlinks = XEXP (links, 1); nextlinks;
899 nextlinks = XEXP (nextlinks, 1))
900 if ((next = try_combine (insn, XEXP (links, 0),
901 XEXP (nextlinks, 0),
902 &new_direct_jump_p)) != 0)
903 goto retry;
905 /* Try this insn with each REG_EQUAL note it links back to. */
906 for (links = LOG_LINKS (insn); links; links = XEXP (links, 1))
908 rtx set, note;
909 rtx temp = XEXP (links, 0);
910 if ((set = single_set (temp)) != 0
911 && (note = find_reg_equal_equiv_note (temp)) != 0
912 && (note = XEXP (note, 0), GET_CODE (note)) != EXPR_LIST
913 /* Avoid using a register that may already been marked
914 dead by an earlier instruction. */
915 && ! unmentioned_reg_p (note, SET_SRC (set))
916 && (GET_MODE (note) == VOIDmode
917 ? SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set)))
918 : GET_MODE (SET_DEST (set)) == GET_MODE (note)))
920 /* Temporarily replace the set's source with the
921 contents of the REG_EQUAL note. The insn will
922 be deleted or recognized by try_combine. */
923 rtx orig = SET_SRC (set);
924 SET_SRC (set) = note;
925 next = try_combine (insn, temp, NULL_RTX,
926 &new_direct_jump_p);
927 if (next)
928 goto retry;
929 SET_SRC (set) = orig;
933 if (!NOTE_P (insn))
934 record_dead_and_set_regs (insn);
936 retry:
941 clear_bb_flags ();
943 EXECUTE_IF_SET_IN_SBITMAP (refresh_blocks, 0, j, sbi)
944 BASIC_BLOCK (j)->flags |= BB_DIRTY;
945 new_direct_jump_p |= purge_all_dead_edges ();
946 delete_noop_moves ();
948 update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
949 PROP_DEATH_NOTES | PROP_SCAN_DEAD_CODE
950 | PROP_KILL_DEAD_CODE);
952 /* Clean up. */
953 sbitmap_free (refresh_blocks);
954 free (uid_insn_cost);
955 free (reg_stat);
956 free (uid_cuid);
959 struct undo *undo, *next;
960 for (undo = undobuf.frees; undo; undo = next)
962 next = undo->next;
963 free (undo);
965 undobuf.frees = 0;
968 total_attempts += combine_attempts;
969 total_merges += combine_merges;
970 total_extras += combine_extras;
971 total_successes += combine_successes;
973 nonzero_sign_valid = 0;
974 rtl_hooks = general_rtl_hooks;
976 /* Make recognizer allow volatile MEMs again. */
977 init_recog ();
979 return new_direct_jump_p;
982 /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */
984 static void
985 init_reg_last (void)
987 unsigned int i;
988 for (i = 0; i < combine_max_regno; i++)
989 memset (reg_stat + i, 0, offsetof (struct reg_stat, sign_bit_copies));
992 /* Set up any promoted values for incoming argument registers. */
994 static void
995 setup_incoming_promotions (void)
997 unsigned int regno;
998 rtx reg;
999 enum machine_mode mode;
1000 int unsignedp;
1001 rtx first = get_insns ();
1003 if (targetm.calls.promote_function_args (TREE_TYPE (cfun->decl)))
1005 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1006 /* Check whether this register can hold an incoming pointer
1007 argument. FUNCTION_ARG_REGNO_P tests outgoing register
1008 numbers, so translate if necessary due to register windows. */
1009 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (regno))
1010 && (reg = promoted_input_arg (regno, &mode, &unsignedp)) != 0)
1012 record_value_for_reg
1013 (reg, first, gen_rtx_fmt_e ((unsignedp ? ZERO_EXTEND
1014 : SIGN_EXTEND),
1015 GET_MODE (reg),
1016 gen_rtx_CLOBBER (mode, const0_rtx)));
1021 /* Called via note_stores. If X is a pseudo that is narrower than
1022 HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero.
1024 If we are setting only a portion of X and we can't figure out what
1025 portion, assume all bits will be used since we don't know what will
1026 be happening.
1028 Similarly, set how many bits of X are known to be copies of the sign bit
1029 at all locations in the function. This is the smallest number implied
1030 by any set of X. */
1032 static void
1033 set_nonzero_bits_and_sign_copies (rtx x, rtx set,
1034 void *data ATTRIBUTE_UNUSED)
1036 unsigned int num;
1038 if (REG_P (x)
1039 && REGNO (x) >= FIRST_PSEUDO_REGISTER
1040 /* If this register is undefined at the start of the file, we can't
1041 say what its contents were. */
1042 && ! REGNO_REG_SET_P
1043 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start, REGNO (x))
1044 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
1046 if (set == 0 || GET_CODE (set) == CLOBBER)
1048 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1049 reg_stat[REGNO (x)].sign_bit_copies = 1;
1050 return;
1053 /* If this is a complex assignment, see if we can convert it into a
1054 simple assignment. */
1055 set = expand_field_assignment (set);
1057 /* If this is a simple assignment, or we have a paradoxical SUBREG,
1058 set what we know about X. */
1060 if (SET_DEST (set) == x
1061 || (GET_CODE (SET_DEST (set)) == SUBREG
1062 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
1063 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (set)))))
1064 && SUBREG_REG (SET_DEST (set)) == x))
1066 rtx src = SET_SRC (set);
1068 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
1069 /* If X is narrower than a word and SRC is a non-negative
1070 constant that would appear negative in the mode of X,
1071 sign-extend it for use in reg_stat[].nonzero_bits because some
1072 machines (maybe most) will actually do the sign-extension
1073 and this is the conservative approach.
1075 ??? For 2.5, try to tighten up the MD files in this regard
1076 instead of this kludge. */
1078 if (GET_MODE_BITSIZE (GET_MODE (x)) < BITS_PER_WORD
1079 && GET_CODE (src) == CONST_INT
1080 && INTVAL (src) > 0
1081 && 0 != (INTVAL (src)
1082 & ((HOST_WIDE_INT) 1
1083 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
1084 src = GEN_INT (INTVAL (src)
1085 | ((HOST_WIDE_INT) (-1)
1086 << GET_MODE_BITSIZE (GET_MODE (x))));
1087 #endif
1089 /* Don't call nonzero_bits if it cannot change anything. */
1090 if (reg_stat[REGNO (x)].nonzero_bits != ~(unsigned HOST_WIDE_INT) 0)
1091 reg_stat[REGNO (x)].nonzero_bits
1092 |= nonzero_bits (src, nonzero_bits_mode);
1093 num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x));
1094 if (reg_stat[REGNO (x)].sign_bit_copies == 0
1095 || reg_stat[REGNO (x)].sign_bit_copies > num)
1096 reg_stat[REGNO (x)].sign_bit_copies = num;
1098 else
1100 reg_stat[REGNO (x)].nonzero_bits = GET_MODE_MASK (GET_MODE (x));
1101 reg_stat[REGNO (x)].sign_bit_copies = 1;
1106 /* See if INSN can be combined into I3. PRED and SUCC are optionally
1107 insns that were previously combined into I3 or that will be combined
1108 into the merger of INSN and I3.
1110 Return 0 if the combination is not allowed for any reason.
1112 If the combination is allowed, *PDEST will be set to the single
1113 destination of INSN and *PSRC to the single source, and this function
1114 will return 1. */
1116 static int
1117 can_combine_p (rtx insn, rtx i3, rtx pred ATTRIBUTE_UNUSED, rtx succ,
1118 rtx *pdest, rtx *psrc)
1120 int i;
1121 rtx set = 0, src, dest;
1122 rtx p;
1123 #ifdef AUTO_INC_DEC
1124 rtx link;
1125 #endif
1126 int all_adjacent = (succ ? (next_active_insn (insn) == succ
1127 && next_active_insn (succ) == i3)
1128 : next_active_insn (insn) == i3);
1130 /* Can combine only if previous insn is a SET of a REG, a SUBREG or CC0.
1131 or a PARALLEL consisting of such a SET and CLOBBERs.
1133 If INSN has CLOBBER parallel parts, ignore them for our processing.
1134 By definition, these happen during the execution of the insn. When it
1135 is merged with another insn, all bets are off. If they are, in fact,
1136 needed and aren't also supplied in I3, they may be added by
1137 recog_for_combine. Otherwise, it won't match.
1139 We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED
1140 note.
1142 Get the source and destination of INSN. If more than one, can't
1143 combine. */
1145 if (GET_CODE (PATTERN (insn)) == SET)
1146 set = PATTERN (insn);
1147 else if (GET_CODE (PATTERN (insn)) == PARALLEL
1148 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET)
1150 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
1152 rtx elt = XVECEXP (PATTERN (insn), 0, i);
1153 rtx note;
1155 switch (GET_CODE (elt))
1157 /* This is important to combine floating point insns
1158 for the SH4 port. */
1159 case USE:
1160 /* Combining an isolated USE doesn't make sense.
1161 We depend here on combinable_i3pat to reject them. */
1162 /* The code below this loop only verifies that the inputs of
1163 the SET in INSN do not change. We call reg_set_between_p
1164 to verify that the REG in the USE does not change between
1165 I3 and INSN.
1166 If the USE in INSN was for a pseudo register, the matching
1167 insn pattern will likely match any register; combining this
1168 with any other USE would only be safe if we knew that the
1169 used registers have identical values, or if there was
1170 something to tell them apart, e.g. different modes. For
1171 now, we forgo such complicated tests and simply disallow
1172 combining of USES of pseudo registers with any other USE. */
1173 if (REG_P (XEXP (elt, 0))
1174 && GET_CODE (PATTERN (i3)) == PARALLEL)
1176 rtx i3pat = PATTERN (i3);
1177 int i = XVECLEN (i3pat, 0) - 1;
1178 unsigned int regno = REGNO (XEXP (elt, 0));
1182 rtx i3elt = XVECEXP (i3pat, 0, i);
1184 if (GET_CODE (i3elt) == USE
1185 && REG_P (XEXP (i3elt, 0))
1186 && (REGNO (XEXP (i3elt, 0)) == regno
1187 ? reg_set_between_p (XEXP (elt, 0),
1188 PREV_INSN (insn), i3)
1189 : regno >= FIRST_PSEUDO_REGISTER))
1190 return 0;
1192 while (--i >= 0);
1194 break;
1196 /* We can ignore CLOBBERs. */
1197 case CLOBBER:
1198 break;
1200 case SET:
1201 /* Ignore SETs whose result isn't used but not those that
1202 have side-effects. */
1203 if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt))
1204 && (!(note = find_reg_note (insn, REG_EH_REGION, NULL_RTX))
1205 || INTVAL (XEXP (note, 0)) <= 0)
1206 && ! side_effects_p (elt))
1207 break;
1209 /* If we have already found a SET, this is a second one and
1210 so we cannot combine with this insn. */
1211 if (set)
1212 return 0;
1214 set = elt;
1215 break;
1217 default:
1218 /* Anything else means we can't combine. */
1219 return 0;
1223 if (set == 0
1224 /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs,
1225 so don't do anything with it. */
1226 || GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1227 return 0;
1229 else
1230 return 0;
1232 if (set == 0)
1233 return 0;
1235 set = expand_field_assignment (set);
1236 src = SET_SRC (set), dest = SET_DEST (set);
1238 /* Don't eliminate a store in the stack pointer. */
1239 if (dest == stack_pointer_rtx
1240 /* Don't combine with an insn that sets a register to itself if it has
1241 a REG_EQUAL note. This may be part of a REG_NO_CONFLICT sequence. */
1242 || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX))
1243 /* Can't merge an ASM_OPERANDS. */
1244 || GET_CODE (src) == ASM_OPERANDS
1245 /* Can't merge a function call. */
1246 || GET_CODE (src) == CALL
1247 /* Don't eliminate a function call argument. */
1248 || (CALL_P (i3)
1249 && (find_reg_fusage (i3, USE, dest)
1250 || (REG_P (dest)
1251 && REGNO (dest) < FIRST_PSEUDO_REGISTER
1252 && global_regs[REGNO (dest)])))
1253 /* Don't substitute into an incremented register. */
1254 || FIND_REG_INC_NOTE (i3, dest)
1255 || (succ && FIND_REG_INC_NOTE (succ, dest))
1256 /* Don't substitute into a non-local goto, this confuses CFG. */
1257 || (JUMP_P (i3) && find_reg_note (i3, REG_NON_LOCAL_GOTO, NULL_RTX))
1258 #if 0
1259 /* Don't combine the end of a libcall into anything. */
1260 /* ??? This gives worse code, and appears to be unnecessary, since no
1261 pass after flow uses REG_LIBCALL/REG_RETVAL notes. Local-alloc does
1262 use REG_RETVAL notes for noconflict blocks, but other code here
1263 makes sure that those insns don't disappear. */
1264 || find_reg_note (insn, REG_RETVAL, NULL_RTX)
1265 #endif
1266 /* Make sure that DEST is not used after SUCC but before I3. */
1267 || (succ && ! all_adjacent
1268 && reg_used_between_p (dest, succ, i3))
1269 /* Make sure that the value that is to be substituted for the register
1270 does not use any registers whose values alter in between. However,
1271 If the insns are adjacent, a use can't cross a set even though we
1272 think it might (this can happen for a sequence of insns each setting
1273 the same destination; last_set of that register might point to
1274 a NOTE). If INSN has a REG_EQUIV note, the register is always
1275 equivalent to the memory so the substitution is valid even if there
1276 are intervening stores. Also, don't move a volatile asm or
1277 UNSPEC_VOLATILE across any other insns. */
1278 || (! all_adjacent
1279 && (((!MEM_P (src)
1280 || ! find_reg_note (insn, REG_EQUIV, src))
1281 && use_crosses_set_p (src, INSN_CUID (insn)))
1282 || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src))
1283 || GET_CODE (src) == UNSPEC_VOLATILE))
1284 /* If there is a REG_NO_CONFLICT note for DEST in I3 or SUCC, we get
1285 better register allocation by not doing the combine. */
1286 || find_reg_note (i3, REG_NO_CONFLICT, dest)
1287 || (succ && find_reg_note (succ, REG_NO_CONFLICT, dest))
1288 /* Don't combine across a CALL_INSN, because that would possibly
1289 change whether the life span of some REGs crosses calls or not,
1290 and it is a pain to update that information.
1291 Exception: if source is a constant, moving it later can't hurt.
1292 Accept that special case, because it helps -fforce-addr a lot. */
1293 || (INSN_CUID (insn) < last_call_cuid && ! CONSTANT_P (src)))
1294 return 0;
1296 /* DEST must either be a REG or CC0. */
1297 if (REG_P (dest))
1299 /* If register alignment is being enforced for multi-word items in all
1300 cases except for parameters, it is possible to have a register copy
1301 insn referencing a hard register that is not allowed to contain the
1302 mode being copied and which would not be valid as an operand of most
1303 insns. Eliminate this problem by not combining with such an insn.
1305 Also, on some machines we don't want to extend the life of a hard
1306 register. */
1308 if (REG_P (src)
1309 && ((REGNO (dest) < FIRST_PSEUDO_REGISTER
1310 && ! HARD_REGNO_MODE_OK (REGNO (dest), GET_MODE (dest)))
1311 /* Don't extend the life of a hard register unless it is
1312 user variable (if we have few registers) or it can't
1313 fit into the desired register (meaning something special
1314 is going on).
1315 Also avoid substituting a return register into I3, because
1316 reload can't handle a conflict with constraints of other
1317 inputs. */
1318 || (REGNO (src) < FIRST_PSEUDO_REGISTER
1319 && ! HARD_REGNO_MODE_OK (REGNO (src), GET_MODE (src)))))
1320 return 0;
1322 else if (GET_CODE (dest) != CC0)
1323 return 0;
1326 if (GET_CODE (PATTERN (i3)) == PARALLEL)
1327 for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--)
1328 if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER)
1330 /* Don't substitute for a register intended as a clobberable
1331 operand. */
1332 rtx reg = XEXP (XVECEXP (PATTERN (i3), 0, i), 0);
1333 if (rtx_equal_p (reg, dest))
1334 return 0;
1336 /* If the clobber represents an earlyclobber operand, we must not
1337 substitute an expression containing the clobbered register.
1338 As we do not analyze the constraint strings here, we have to
1339 make the conservative assumption. However, if the register is
1340 a fixed hard reg, the clobber cannot represent any operand;
1341 we leave it up to the machine description to either accept or
1342 reject use-and-clobber patterns. */
1343 if (!REG_P (reg)
1344 || REGNO (reg) >= FIRST_PSEUDO_REGISTER
1345 || !fixed_regs[REGNO (reg)])
1346 if (reg_overlap_mentioned_p (reg, src))
1347 return 0;
1350 /* If INSN contains anything volatile, or is an `asm' (whether volatile
1351 or not), reject, unless nothing volatile comes between it and I3 */
1353 if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src))
1355 /* Make sure succ doesn't contain a volatile reference. */
1356 if (succ != 0 && volatile_refs_p (PATTERN (succ)))
1357 return 0;
1359 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1360 if (INSN_P (p) && p != succ && volatile_refs_p (PATTERN (p)))
1361 return 0;
1364 /* If INSN is an asm, and DEST is a hard register, reject, since it has
1365 to be an explicit register variable, and was chosen for a reason. */
1367 if (GET_CODE (src) == ASM_OPERANDS
1368 && REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER)
1369 return 0;
1371 /* If there are any volatile insns between INSN and I3, reject, because
1372 they might affect machine state. */
1374 for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (p))
1375 if (INSN_P (p) && p != succ && volatile_insn_p (PATTERN (p)))
1376 return 0;
1378 /* If INSN contains an autoincrement or autodecrement, make sure that
1379 register is not used between there and I3, and not already used in
1380 I3 either. Neither must it be used in PRED or SUCC, if they exist.
1381 Also insist that I3 not be a jump; if it were one
1382 and the incremented register were spilled, we would lose. */
1384 #ifdef AUTO_INC_DEC
1385 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
1386 if (REG_NOTE_KIND (link) == REG_INC
1387 && (JUMP_P (i3)
1388 || reg_used_between_p (XEXP (link, 0), insn, i3)
1389 || (pred != NULL_RTX
1390 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (pred)))
1391 || (succ != NULL_RTX
1392 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (succ)))
1393 || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i3))))
1394 return 0;
1395 #endif
1397 #ifdef HAVE_cc0
1398 /* Don't combine an insn that follows a CC0-setting insn.
1399 An insn that uses CC0 must not be separated from the one that sets it.
1400 We do, however, allow I2 to follow a CC0-setting insn if that insn
1401 is passed as I1; in that case it will be deleted also.
1402 We also allow combining in this case if all the insns are adjacent
1403 because that would leave the two CC0 insns adjacent as well.
1404 It would be more logical to test whether CC0 occurs inside I1 or I2,
1405 but that would be much slower, and this ought to be equivalent. */
1407 p = prev_nonnote_insn (insn);
1408 if (p && p != pred && NONJUMP_INSN_P (p) && sets_cc0_p (PATTERN (p))
1409 && ! all_adjacent)
1410 return 0;
1411 #endif
1413 /* If we get here, we have passed all the tests and the combination is
1414 to be allowed. */
1416 *pdest = dest;
1417 *psrc = src;
1419 return 1;
1422 /* LOC is the location within I3 that contains its pattern or the component
1423 of a PARALLEL of the pattern. We validate that it is valid for combining.
1425 One problem is if I3 modifies its output, as opposed to replacing it
1426 entirely, we can't allow the output to contain I2DEST or I1DEST as doing
1427 so would produce an insn that is not equivalent to the original insns.
1429 Consider:
1431 (set (reg:DI 101) (reg:DI 100))
1432 (set (subreg:SI (reg:DI 101) 0) <foo>)
1434 This is NOT equivalent to:
1436 (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>)
1437 (set (reg:DI 101) (reg:DI 100))])
1439 Not only does this modify 100 (in which case it might still be valid
1440 if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100.
1442 We can also run into a problem if I2 sets a register that I1
1443 uses and I1 gets directly substituted into I3 (not via I2). In that
1444 case, we would be getting the wrong value of I2DEST into I3, so we
1445 must reject the combination. This case occurs when I2 and I1 both
1446 feed into I3, rather than when I1 feeds into I2, which feeds into I3.
1447 If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source
1448 of a SET must prevent combination from occurring.
1450 Before doing the above check, we first try to expand a field assignment
1451 into a set of logical operations.
1453 If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which
1454 we place a register that is both set and used within I3. If more than one
1455 such register is detected, we fail.
1457 Return 1 if the combination is valid, zero otherwise. */
1459 static int
1460 combinable_i3pat (rtx i3, rtx *loc, rtx i2dest, rtx i1dest,
1461 int i1_not_in_src, rtx *pi3dest_killed)
1463 rtx x = *loc;
1465 if (GET_CODE (x) == SET)
1467 rtx set = x ;
1468 rtx dest = SET_DEST (set);
1469 rtx src = SET_SRC (set);
1470 rtx inner_dest = dest;
1471 rtx subdest;
1473 while (GET_CODE (inner_dest) == STRICT_LOW_PART
1474 || GET_CODE (inner_dest) == SUBREG
1475 || GET_CODE (inner_dest) == ZERO_EXTRACT)
1476 inner_dest = XEXP (inner_dest, 0);
1478 /* Check for the case where I3 modifies its output, as discussed
1479 above. We don't want to prevent pseudos from being combined
1480 into the address of a MEM, so only prevent the combination if
1481 i1 or i2 set the same MEM. */
1482 if ((inner_dest != dest &&
1483 (!MEM_P (inner_dest)
1484 || rtx_equal_p (i2dest, inner_dest)
1485 || (i1dest && rtx_equal_p (i1dest, inner_dest)))
1486 && (reg_overlap_mentioned_p (i2dest, inner_dest)
1487 || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest))))
1489 /* This is the same test done in can_combine_p except we can't test
1490 all_adjacent; we don't have to, since this instruction will stay
1491 in place, thus we are not considering increasing the lifetime of
1492 INNER_DEST.
1494 Also, if this insn sets a function argument, combining it with
1495 something that might need a spill could clobber a previous
1496 function argument; the all_adjacent test in can_combine_p also
1497 checks this; here, we do a more specific test for this case. */
1499 || (REG_P (inner_dest)
1500 && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER
1501 && (! HARD_REGNO_MODE_OK (REGNO (inner_dest),
1502 GET_MODE (inner_dest))))
1503 || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src)))
1504 return 0;
1506 /* If DEST is used in I3, it is being killed in this insn, so
1507 record that for later. We have to consider paradoxical
1508 subregs here, since they kill the whole register, but we
1509 ignore partial subregs, STRICT_LOW_PART, etc.
1510 Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the
1511 STACK_POINTER_REGNUM, since these are always considered to be
1512 live. Similarly for ARG_POINTER_REGNUM if it is fixed. */
1513 subdest = dest;
1514 if (GET_CODE (subdest) == SUBREG
1515 && (GET_MODE_SIZE (GET_MODE (subdest))
1516 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (subdest)))))
1517 subdest = SUBREG_REG (subdest);
1518 if (pi3dest_killed
1519 && REG_P (subdest)
1520 && reg_referenced_p (subdest, PATTERN (i3))
1521 && REGNO (subdest) != FRAME_POINTER_REGNUM
1522 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1523 && REGNO (subdest) != HARD_FRAME_POINTER_REGNUM
1524 #endif
1525 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
1526 && (REGNO (subdest) != ARG_POINTER_REGNUM
1527 || ! fixed_regs [REGNO (subdest)])
1528 #endif
1529 && REGNO (subdest) != STACK_POINTER_REGNUM)
1531 if (*pi3dest_killed)
1532 return 0;
1534 *pi3dest_killed = subdest;
1538 else if (GET_CODE (x) == PARALLEL)
1540 int i;
1542 for (i = 0; i < XVECLEN (x, 0); i++)
1543 if (! combinable_i3pat (i3, &XVECEXP (x, 0, i), i2dest, i1dest,
1544 i1_not_in_src, pi3dest_killed))
1545 return 0;
1548 return 1;
1551 /* Return 1 if X is an arithmetic expression that contains a multiplication
1552 and division. We don't count multiplications by powers of two here. */
1554 static int
1555 contains_muldiv (rtx x)
1557 switch (GET_CODE (x))
1559 case MOD: case DIV: case UMOD: case UDIV:
1560 return 1;
1562 case MULT:
1563 return ! (GET_CODE (XEXP (x, 1)) == CONST_INT
1564 && exact_log2 (INTVAL (XEXP (x, 1))) >= 0);
1565 default:
1566 if (BINARY_P (x))
1567 return contains_muldiv (XEXP (x, 0))
1568 || contains_muldiv (XEXP (x, 1));
1570 if (UNARY_P (x))
1571 return contains_muldiv (XEXP (x, 0));
1573 return 0;
1577 /* Determine whether INSN can be used in a combination. Return nonzero if
1578 not. This is used in try_combine to detect early some cases where we
1579 can't perform combinations. */
1581 static int
1582 cant_combine_insn_p (rtx insn)
1584 rtx set;
1585 rtx src, dest;
1587 /* If this isn't really an insn, we can't do anything.
1588 This can occur when flow deletes an insn that it has merged into an
1589 auto-increment address. */
1590 if (! INSN_P (insn))
1591 return 1;
1593 /* Never combine loads and stores involving hard regs that are likely
1594 to be spilled. The register allocator can usually handle such
1595 reg-reg moves by tying. If we allow the combiner to make
1596 substitutions of likely-spilled regs, reload might die.
1597 As an exception, we allow combinations involving fixed regs; these are
1598 not available to the register allocator so there's no risk involved. */
1600 set = single_set (insn);
1601 if (! set)
1602 return 0;
1603 src = SET_SRC (set);
1604 dest = SET_DEST (set);
1605 if (GET_CODE (src) == SUBREG)
1606 src = SUBREG_REG (src);
1607 if (GET_CODE (dest) == SUBREG)
1608 dest = SUBREG_REG (dest);
1609 if (REG_P (src) && REG_P (dest)
1610 && ((REGNO (src) < FIRST_PSEUDO_REGISTER
1611 && ! fixed_regs[REGNO (src)]
1612 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (src))))
1613 || (REGNO (dest) < FIRST_PSEUDO_REGISTER
1614 && ! fixed_regs[REGNO (dest)]
1615 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (dest))))))
1616 return 1;
1618 return 0;
1621 struct likely_spilled_retval_info
1623 unsigned regno, nregs;
1624 unsigned mask;
1627 /* Called via note_stores by likely_spilled_retval_p. Remove from info->mask
1628 hard registers that are known to be written to / clobbered in full. */
1629 static void
1630 likely_spilled_retval_1 (rtx x, rtx set, void *data)
1632 struct likely_spilled_retval_info *info = data;
1633 unsigned regno, nregs;
1634 unsigned new_mask;
1636 if (!REG_P (XEXP (set, 0)))
1637 return;
1638 regno = REGNO (x);
1639 if (regno >= info->regno + info->nregs)
1640 return;
1641 nregs = hard_regno_nregs[regno][GET_MODE (x)];
1642 if (regno + nregs <= info->regno)
1643 return;
1644 new_mask = (2U << (nregs - 1)) - 1;
1645 if (regno < info->regno)
1646 new_mask >>= info->regno - regno;
1647 else
1648 new_mask <<= regno - info->regno;
1649 info->mask &= new_mask;
1652 /* Return nonzero iff part of the return value is live during INSN, and
1653 it is likely spilled. This can happen when more than one insn is needed
1654 to copy the return value, e.g. when we consider to combine into the
1655 second copy insn for a complex value. */
1657 static int
1658 likely_spilled_retval_p (rtx insn)
1660 rtx use = BB_END (this_basic_block);
1661 rtx reg, p;
1662 unsigned regno, nregs;
1663 /* We assume here that no machine mode needs more than
1664 32 hard registers when the value overlaps with a register
1665 for which FUNCTION_VALUE_REGNO_P is true. */
1666 unsigned mask;
1667 struct likely_spilled_retval_info info;
1669 if (!NONJUMP_INSN_P (use) || GET_CODE (PATTERN (use)) != USE || insn == use)
1670 return 0;
1671 reg = XEXP (PATTERN (use), 0);
1672 if (!REG_P (reg) || !FUNCTION_VALUE_REGNO_P (REGNO (reg)))
1673 return 0;
1674 regno = REGNO (reg);
1675 nregs = hard_regno_nregs[regno][GET_MODE (reg)];
1676 if (nregs == 1)
1677 return 0;
1678 mask = (2U << (nregs - 1)) - 1;
1680 /* Disregard parts of the return value that are set later. */
1681 info.regno = regno;
1682 info.nregs = nregs;
1683 info.mask = mask;
1684 for (p = PREV_INSN (use); info.mask && p != insn; p = PREV_INSN (p))
1685 note_stores (PATTERN (insn), likely_spilled_retval_1, &info);
1686 mask = info.mask;
1688 /* Check if any of the (probably) live return value registers is
1689 likely spilled. */
1690 nregs --;
1693 if ((mask & 1 << nregs)
1694 && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (regno + nregs)))
1695 return 1;
1696 } while (nregs--);
1697 return 0;
1700 /* Adjust INSN after we made a change to its destination.
1702 Changing the destination can invalidate notes that say something about
1703 the results of the insn and a LOG_LINK pointing to the insn. */
1705 static void
1706 adjust_for_new_dest (rtx insn)
1708 rtx *loc;
1710 /* For notes, be conservative and simply remove them. */
1711 loc = &REG_NOTES (insn);
1712 while (*loc)
1714 enum reg_note kind = REG_NOTE_KIND (*loc);
1715 if (kind == REG_EQUAL || kind == REG_EQUIV)
1716 *loc = XEXP (*loc, 1);
1717 else
1718 loc = &XEXP (*loc, 1);
1721 /* The new insn will have a destination that was previously the destination
1722 of an insn just above it. Call distribute_links to make a LOG_LINK from
1723 the next use of that destination. */
1724 distribute_links (gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX));
1727 /* Return TRUE if combine can reuse reg X in mode MODE.
1728 ADDED_SETS is nonzero if the original set is still required. */
1729 static bool
1730 can_change_dest_mode (rtx x, int added_sets, enum machine_mode mode)
1732 unsigned int regno;
1734 if (!REG_P(x))
1735 return false;
1737 regno = REGNO (x);
1738 /* Allow hard registers if the new mode is legal, and occupies no more
1739 registers than the old mode. */
1740 if (regno < FIRST_PSEUDO_REGISTER)
1741 return (HARD_REGNO_MODE_OK (regno, mode)
1742 && (hard_regno_nregs[regno][GET_MODE (x)]
1743 >= hard_regno_nregs[regno][mode]));
1745 /* Or a pseudo that is only used once. */
1746 return (REG_N_SETS (regno) == 1 && !added_sets
1747 && !REG_USERVAR_P (x));
1751 /* Check whether X, the destination of a set, refers to part of
1752 the register specified by REG. */
1754 static bool
1755 reg_subword_p (rtx x, rtx reg)
1757 /* Check that reg is an integer mode register. */
1758 if (!REG_P (reg) || GET_MODE_CLASS (GET_MODE (reg)) != MODE_INT)
1759 return false;
1761 if (GET_CODE (x) == STRICT_LOW_PART
1762 || GET_CODE (x) == ZERO_EXTRACT)
1763 x = XEXP (x, 0);
1765 return GET_CODE (x) == SUBREG
1766 && SUBREG_REG (x) == reg
1767 && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT;
1771 /* Try to combine the insns I1 and I2 into I3.
1772 Here I1 and I2 appear earlier than I3.
1773 I1 can be zero; then we combine just I2 into I3.
1775 If we are combining three insns and the resulting insn is not recognized,
1776 try splitting it into two insns. If that happens, I2 and I3 are retained
1777 and I1 is pseudo-deleted by turning it into a NOTE. Otherwise, I1 and I2
1778 are pseudo-deleted.
1780 Return 0 if the combination does not work. Then nothing is changed.
1781 If we did the combination, return the insn at which combine should
1782 resume scanning.
1784 Set NEW_DIRECT_JUMP_P to a nonzero value if try_combine creates a
1785 new direct jump instruction. */
1787 static rtx
1788 try_combine (rtx i3, rtx i2, rtx i1, int *new_direct_jump_p)
1790 /* New patterns for I3 and I2, respectively. */
1791 rtx newpat, newi2pat = 0;
1792 rtvec newpat_vec_with_clobbers = 0;
1793 int substed_i2 = 0, substed_i1 = 0;
1794 /* Indicates need to preserve SET in I1 or I2 in I3 if it is not dead. */
1795 int added_sets_1, added_sets_2;
1796 /* Total number of SETs to put into I3. */
1797 int total_sets;
1798 /* Nonzero if I2's body now appears in I3. */
1799 int i2_is_used;
1800 /* INSN_CODEs for new I3, new I2, and user of condition code. */
1801 int insn_code_number, i2_code_number = 0, other_code_number = 0;
1802 /* Contains I3 if the destination of I3 is used in its source, which means
1803 that the old life of I3 is being killed. If that usage is placed into
1804 I2 and not in I3, a REG_DEAD note must be made. */
1805 rtx i3dest_killed = 0;
1806 /* SET_DEST and SET_SRC of I2 and I1. */
1807 rtx i2dest, i2src, i1dest = 0, i1src = 0;
1808 /* PATTERN (I2), or a copy of it in certain cases. */
1809 rtx i2pat;
1810 /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */
1811 int i2dest_in_i2src = 0, i1dest_in_i1src = 0, i2dest_in_i1src = 0;
1812 int i2dest_killed = 0, i1dest_killed = 0;
1813 int i1_feeds_i3 = 0;
1814 /* Notes that must be added to REG_NOTES in I3 and I2. */
1815 rtx new_i3_notes, new_i2_notes;
1816 /* Notes that we substituted I3 into I2 instead of the normal case. */
1817 int i3_subst_into_i2 = 0;
1818 /* Notes that I1, I2 or I3 is a MULT operation. */
1819 int have_mult = 0;
1820 int swap_i2i3 = 0;
1822 int maxreg;
1823 rtx temp;
1824 rtx link;
1825 int i;
1827 /* Exit early if one of the insns involved can't be used for
1828 combinations. */
1829 if (cant_combine_insn_p (i3)
1830 || cant_combine_insn_p (i2)
1831 || (i1 && cant_combine_insn_p (i1))
1832 || likely_spilled_retval_p (i3)
1833 /* We also can't do anything if I3 has a
1834 REG_LIBCALL note since we don't want to disrupt the contiguity of a
1835 libcall. */
1836 #if 0
1837 /* ??? This gives worse code, and appears to be unnecessary, since no
1838 pass after flow uses REG_LIBCALL/REG_RETVAL notes. */
1839 || find_reg_note (i3, REG_LIBCALL, NULL_RTX)
1840 #endif
1842 return 0;
1844 combine_attempts++;
1845 undobuf.other_insn = 0;
1847 /* Reset the hard register usage information. */
1848 CLEAR_HARD_REG_SET (newpat_used_regs);
1850 /* If I1 and I2 both feed I3, they can be in any order. To simplify the
1851 code below, set I1 to be the earlier of the two insns. */
1852 if (i1 && INSN_CUID (i1) > INSN_CUID (i2))
1853 temp = i1, i1 = i2, i2 = temp;
1855 added_links_insn = 0;
1857 /* First check for one important special-case that the code below will
1858 not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
1859 and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case,
1860 we may be able to replace that destination with the destination of I3.
1861 This occurs in the common code where we compute both a quotient and
1862 remainder into a structure, in which case we want to do the computation
1863 directly into the structure to avoid register-register copies.
1865 Note that this case handles both multiple sets in I2 and also
1866 cases where I2 has a number of CLOBBER or PARALLELs.
1868 We make very conservative checks below and only try to handle the
1869 most common cases of this. For example, we only handle the case
1870 where I2 and I3 are adjacent to avoid making difficult register
1871 usage tests. */
1873 if (i1 == 0 && NONJUMP_INSN_P (i3) && GET_CODE (PATTERN (i3)) == SET
1874 && REG_P (SET_SRC (PATTERN (i3)))
1875 && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER
1876 && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3)))
1877 && GET_CODE (PATTERN (i2)) == PARALLEL
1878 && ! side_effects_p (SET_DEST (PATTERN (i3)))
1879 /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code
1880 below would need to check what is inside (and reg_overlap_mentioned_p
1881 doesn't support those codes anyway). Don't allow those destinations;
1882 the resulting insn isn't likely to be recognized anyway. */
1883 && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT
1884 && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART
1885 && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)),
1886 SET_DEST (PATTERN (i3)))
1887 && next_real_insn (i2) == i3)
1889 rtx p2 = PATTERN (i2);
1891 /* Make sure that the destination of I3,
1892 which we are going to substitute into one output of I2,
1893 is not used within another output of I2. We must avoid making this:
1894 (parallel [(set (mem (reg 69)) ...)
1895 (set (reg 69) ...)])
1896 which is not well-defined as to order of actions.
1897 (Besides, reload can't handle output reloads for this.)
1899 The problem can also happen if the dest of I3 is a memory ref,
1900 if another dest in I2 is an indirect memory ref. */
1901 for (i = 0; i < XVECLEN (p2, 0); i++)
1902 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1903 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1904 && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)),
1905 SET_DEST (XVECEXP (p2, 0, i))))
1906 break;
1908 if (i == XVECLEN (p2, 0))
1909 for (i = 0; i < XVECLEN (p2, 0); i++)
1910 if ((GET_CODE (XVECEXP (p2, 0, i)) == SET
1911 || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER)
1912 && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3)))
1914 combine_merges++;
1916 subst_insn = i3;
1917 subst_low_cuid = INSN_CUID (i2);
1919 added_sets_2 = added_sets_1 = 0;
1920 i2dest = SET_SRC (PATTERN (i3));
1921 i2dest_killed = dead_or_set_p (i2, i2dest);
1923 /* Replace the dest in I2 with our dest and make the resulting
1924 insn the new pattern for I3. Then skip to where we
1925 validate the pattern. Everything was set up above. */
1926 SUBST (SET_DEST (XVECEXP (p2, 0, i)),
1927 SET_DEST (PATTERN (i3)));
1929 newpat = p2;
1930 i3_subst_into_i2 = 1;
1931 goto validate_replacement;
1935 /* If I2 is setting a pseudo to a constant and I3 is setting some
1936 sub-part of it to another constant, merge them by making a new
1937 constant. */
1938 if (i1 == 0
1939 && (temp = single_set (i2)) != 0
1940 && (GET_CODE (SET_SRC (temp)) == CONST_INT
1941 || GET_CODE (SET_SRC (temp)) == CONST_DOUBLE)
1942 && GET_CODE (PATTERN (i3)) == SET
1943 && (GET_CODE (SET_SRC (PATTERN (i3))) == CONST_INT
1944 || GET_CODE (SET_SRC (PATTERN (i3))) == CONST_DOUBLE)
1945 && reg_subword_p (SET_DEST (PATTERN (i3)), SET_DEST (temp)))
1947 rtx dest = SET_DEST (PATTERN (i3));
1948 int offset = -1;
1949 int width = 0;
1951 if (GET_CODE (dest) == ZERO_EXTRACT)
1953 if (GET_CODE (XEXP (dest, 1)) == CONST_INT
1954 && GET_CODE (XEXP (dest, 2)) == CONST_INT)
1956 width = INTVAL (XEXP (dest, 1));
1957 offset = INTVAL (XEXP (dest, 2));
1958 dest = XEXP (dest, 0);
1959 if (BITS_BIG_ENDIAN)
1960 offset = GET_MODE_BITSIZE (GET_MODE (dest)) - width - offset;
1963 else
1965 if (GET_CODE (dest) == STRICT_LOW_PART)
1966 dest = XEXP (dest, 0);
1967 width = GET_MODE_BITSIZE (GET_MODE (dest));
1968 offset = 0;
1971 if (offset >= 0)
1973 /* If this is the low part, we're done. */
1974 if (subreg_lowpart_p (dest))
1976 /* Handle the case where inner is twice the size of outer. */
1977 else if (GET_MODE_BITSIZE (GET_MODE (SET_DEST (temp)))
1978 == 2 * GET_MODE_BITSIZE (GET_MODE (dest)))
1979 offset += GET_MODE_BITSIZE (GET_MODE (dest));
1980 /* Otherwise give up for now. */
1981 else
1982 offset = -1;
1985 if (offset >= 0)
1987 HOST_WIDE_INT mhi, ohi, ihi;
1988 HOST_WIDE_INT mlo, olo, ilo;
1989 rtx inner = SET_SRC (PATTERN (i3));
1990 rtx outer = SET_SRC (temp);
1992 if (GET_CODE (outer) == CONST_INT)
1994 olo = INTVAL (outer);
1995 ohi = olo < 0 ? -1 : 0;
1997 else
1999 olo = CONST_DOUBLE_LOW (outer);
2000 ohi = CONST_DOUBLE_HIGH (outer);
2003 if (GET_CODE (inner) == CONST_INT)
2005 ilo = INTVAL (inner);
2006 ihi = ilo < 0 ? -1 : 0;
2008 else
2010 ilo = CONST_DOUBLE_LOW (inner);
2011 ihi = CONST_DOUBLE_HIGH (inner);
2014 if (width < HOST_BITS_PER_WIDE_INT)
2016 mlo = ((unsigned HOST_WIDE_INT) 1 << width) - 1;
2017 mhi = 0;
2019 else if (width < HOST_BITS_PER_WIDE_INT * 2)
2021 mhi = ((unsigned HOST_WIDE_INT) 1
2022 << (width - HOST_BITS_PER_WIDE_INT)) - 1;
2023 mlo = -1;
2025 else
2027 mlo = -1;
2028 mhi = -1;
2031 ilo &= mlo;
2032 ihi &= mhi;
2034 if (offset >= HOST_BITS_PER_WIDE_INT)
2036 mhi = mlo << (offset - HOST_BITS_PER_WIDE_INT);
2037 mlo = 0;
2038 ihi = ilo << (offset - HOST_BITS_PER_WIDE_INT);
2039 ilo = 0;
2041 else if (offset > 0)
2043 mhi = (mhi << offset) | ((unsigned HOST_WIDE_INT) mlo
2044 >> (HOST_BITS_PER_WIDE_INT - offset));
2045 mlo = mlo << offset;
2046 ihi = (ihi << offset) | ((unsigned HOST_WIDE_INT) ilo
2047 >> (HOST_BITS_PER_WIDE_INT - offset));
2048 ilo = ilo << offset;
2051 olo = (olo & ~mlo) | ilo;
2052 ohi = (ohi & ~mhi) | ihi;
2054 combine_merges++;
2055 subst_insn = i3;
2056 subst_low_cuid = INSN_CUID (i2);
2057 added_sets_2 = added_sets_1 = 0;
2058 i2dest = SET_DEST (temp);
2059 i2dest_killed = dead_or_set_p (i2, i2dest);
2061 SUBST (SET_SRC (temp),
2062 immed_double_const (olo, ohi, GET_MODE (SET_DEST (temp))));
2064 newpat = PATTERN (i2);
2065 goto validate_replacement;
2069 #ifndef HAVE_cc0
2070 /* If we have no I1 and I2 looks like:
2071 (parallel [(set (reg:CC X) (compare:CC OP (const_int 0)))
2072 (set Y OP)])
2073 make up a dummy I1 that is
2074 (set Y OP)
2075 and change I2 to be
2076 (set (reg:CC X) (compare:CC Y (const_int 0)))
2078 (We can ignore any trailing CLOBBERs.)
2080 This undoes a previous combination and allows us to match a branch-and-
2081 decrement insn. */
2083 if (i1 == 0 && GET_CODE (PATTERN (i2)) == PARALLEL
2084 && XVECLEN (PATTERN (i2), 0) >= 2
2085 && GET_CODE (XVECEXP (PATTERN (i2), 0, 0)) == SET
2086 && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0))))
2087 == MODE_CC)
2088 && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE
2089 && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx
2090 && GET_CODE (XVECEXP (PATTERN (i2), 0, 1)) == SET
2091 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)))
2092 && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0),
2093 SET_SRC (XVECEXP (PATTERN (i2), 0, 1))))
2095 for (i = XVECLEN (PATTERN (i2), 0) - 1; i >= 2; i--)
2096 if (GET_CODE (XVECEXP (PATTERN (i2), 0, i)) != CLOBBER)
2097 break;
2099 if (i == 1)
2101 /* We make I1 with the same INSN_UID as I2. This gives it
2102 the same INSN_CUID for value tracking. Our fake I1 will
2103 never appear in the insn stream so giving it the same INSN_UID
2104 as I2 will not cause a problem. */
2106 i1 = gen_rtx_INSN (VOIDmode, INSN_UID (i2), NULL_RTX, i2,
2107 BLOCK_FOR_INSN (i2), INSN_LOCATOR (i2),
2108 XVECEXP (PATTERN (i2), 0, 1), -1, NULL_RTX,
2109 NULL_RTX);
2111 SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0));
2112 SUBST (XEXP (SET_SRC (PATTERN (i2)), 0),
2113 SET_DEST (PATTERN (i1)));
2116 #endif
2118 /* Verify that I2 and I1 are valid for combining. */
2119 if (! can_combine_p (i2, i3, i1, NULL_RTX, &i2dest, &i2src)
2120 || (i1 && ! can_combine_p (i1, i3, NULL_RTX, i2, &i1dest, &i1src)))
2122 undo_all ();
2123 return 0;
2126 /* Record whether I2DEST is used in I2SRC and similarly for the other
2127 cases. Knowing this will help in register status updating below. */
2128 i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src);
2129 i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src);
2130 i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src);
2131 i2dest_killed = dead_or_set_p (i2, i2dest);
2132 i1dest_killed = i1 && dead_or_set_p (i1, i1dest);
2134 /* See if I1 directly feeds into I3. It does if I1DEST is not used
2135 in I2SRC. */
2136 i1_feeds_i3 = i1 && ! reg_overlap_mentioned_p (i1dest, i2src);
2138 /* Ensure that I3's pattern can be the destination of combines. */
2139 if (! combinable_i3pat (i3, &PATTERN (i3), i2dest, i1dest,
2140 i1 && i2dest_in_i1src && i1_feeds_i3,
2141 &i3dest_killed))
2143 undo_all ();
2144 return 0;
2147 /* See if any of the insns is a MULT operation. Unless one is, we will
2148 reject a combination that is, since it must be slower. Be conservative
2149 here. */
2150 if (GET_CODE (i2src) == MULT
2151 || (i1 != 0 && GET_CODE (i1src) == MULT)
2152 || (GET_CODE (PATTERN (i3)) == SET
2153 && GET_CODE (SET_SRC (PATTERN (i3))) == MULT))
2154 have_mult = 1;
2156 /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd.
2157 We used to do this EXCEPT in one case: I3 has a post-inc in an
2158 output operand. However, that exception can give rise to insns like
2159 mov r3,(r3)+
2160 which is a famous insn on the PDP-11 where the value of r3 used as the
2161 source was model-dependent. Avoid this sort of thing. */
2163 #if 0
2164 if (!(GET_CODE (PATTERN (i3)) == SET
2165 && REG_P (SET_SRC (PATTERN (i3)))
2166 && MEM_P (SET_DEST (PATTERN (i3)))
2167 && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC
2168 || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC)))
2169 /* It's not the exception. */
2170 #endif
2171 #ifdef AUTO_INC_DEC
2172 for (link = REG_NOTES (i3); link; link = XEXP (link, 1))
2173 if (REG_NOTE_KIND (link) == REG_INC
2174 && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i2))
2175 || (i1 != 0
2176 && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (i1)))))
2178 undo_all ();
2179 return 0;
2181 #endif
2183 /* See if the SETs in I1 or I2 need to be kept around in the merged
2184 instruction: whenever the value set there is still needed past I3.
2185 For the SETs in I2, this is easy: we see if I2DEST dies or is set in I3.
2187 For the SET in I1, we have two cases: If I1 and I2 independently
2188 feed into I3, the set in I1 needs to be kept around if I1DEST dies
2189 or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set
2190 in I1 needs to be kept around unless I1DEST dies or is set in either
2191 I2 or I3. We can distinguish these cases by seeing if I2SRC mentions
2192 I1DEST. If so, we know I1 feeds into I2. */
2194 added_sets_2 = ! dead_or_set_p (i3, i2dest);
2196 added_sets_1
2197 = i1 && ! (i1_feeds_i3 ? dead_or_set_p (i3, i1dest)
2198 : (dead_or_set_p (i3, i1dest) || dead_or_set_p (i2, i1dest)));
2200 /* If the set in I2 needs to be kept around, we must make a copy of
2201 PATTERN (I2), so that when we substitute I1SRC for I1DEST in
2202 PATTERN (I2), we are only substituting for the original I1DEST, not into
2203 an already-substituted copy. This also prevents making self-referential
2204 rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to
2205 I2DEST. */
2207 i2pat = (GET_CODE (PATTERN (i2)) == PARALLEL
2208 ? gen_rtx_SET (VOIDmode, i2dest, i2src)
2209 : PATTERN (i2));
2211 if (added_sets_2)
2212 i2pat = copy_rtx (i2pat);
2214 combine_merges++;
2216 /* Substitute in the latest insn for the regs set by the earlier ones. */
2218 maxreg = max_reg_num ();
2220 subst_insn = i3;
2222 #ifndef HAVE_cc0
2223 /* Many machines that don't use CC0 have insns that can both perform an
2224 arithmetic operation and set the condition code. These operations will
2225 be represented as a PARALLEL with the first element of the vector
2226 being a COMPARE of an arithmetic operation with the constant zero.
2227 The second element of the vector will set some pseudo to the result
2228 of the same arithmetic operation. If we simplify the COMPARE, we won't
2229 match such a pattern and so will generate an extra insn. Here we test
2230 for this case, where both the comparison and the operation result are
2231 needed, and make the PARALLEL by just replacing I2DEST in I3SRC with
2232 I2SRC. Later we will make the PARALLEL that contains I2. */
2234 if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET
2235 && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE
2236 && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx
2237 && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest))
2239 #ifdef SELECT_CC_MODE
2240 rtx *cc_use;
2241 enum machine_mode compare_mode;
2242 #endif
2244 newpat = PATTERN (i3);
2245 SUBST (XEXP (SET_SRC (newpat), 0), i2src);
2247 i2_is_used = 1;
2249 #ifdef SELECT_CC_MODE
2250 /* See if a COMPARE with the operand we substituted in should be done
2251 with the mode that is currently being used. If not, do the same
2252 processing we do in `subst' for a SET; namely, if the destination
2253 is used only once, try to replace it with a register of the proper
2254 mode and also replace the COMPARE. */
2255 if (undobuf.other_insn == 0
2256 && (cc_use = find_single_use (SET_DEST (newpat), i3,
2257 &undobuf.other_insn))
2258 && ((compare_mode = SELECT_CC_MODE (GET_CODE (*cc_use),
2259 i2src, const0_rtx))
2260 != GET_MODE (SET_DEST (newpat))))
2262 if (can_change_dest_mode(SET_DEST (newpat), added_sets_2,
2263 compare_mode))
2265 unsigned int regno = REGNO (SET_DEST (newpat));
2266 rtx new_dest;
2268 if (regno < FIRST_PSEUDO_REGISTER)
2269 new_dest = gen_rtx_REG (compare_mode, regno);
2270 else
2272 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
2273 new_dest = regno_reg_rtx[regno];
2276 SUBST (SET_DEST (newpat), new_dest);
2277 SUBST (XEXP (*cc_use, 0), new_dest);
2278 SUBST (SET_SRC (newpat),
2279 gen_rtx_COMPARE (compare_mode, i2src, const0_rtx));
2281 else
2282 undobuf.other_insn = 0;
2284 #endif
2286 else
2287 #endif
2289 /* It is possible that the source of I2 or I1 may be performing
2290 an unneeded operation, such as a ZERO_EXTEND of something
2291 that is known to have the high part zero. Handle that case
2292 by letting subst look at the innermost one of them.
2294 Another way to do this would be to have a function that tries
2295 to simplify a single insn instead of merging two or more
2296 insns. We don't do this because of the potential of infinite
2297 loops and because of the potential extra memory required.
2298 However, doing it the way we are is a bit of a kludge and
2299 doesn't catch all cases.
2301 But only do this if -fexpensive-optimizations since it slows
2302 things down and doesn't usually win.
2304 This is not done in the COMPARE case above because the
2305 unmodified I2PAT is used in the PARALLEL and so a pattern
2306 with a modified I2SRC would not match. */
2308 if (flag_expensive_optimizations)
2310 /* Pass pc_rtx so no substitutions are done, just
2311 simplifications. */
2312 if (i1)
2314 subst_low_cuid = INSN_CUID (i1);
2315 i1src = subst (i1src, pc_rtx, pc_rtx, 0, 0);
2317 else
2319 subst_low_cuid = INSN_CUID (i2);
2320 i2src = subst (i2src, pc_rtx, pc_rtx, 0, 0);
2324 n_occurrences = 0; /* `subst' counts here */
2326 /* If I1 feeds into I2 (not into I3) and I1DEST is in I1SRC, we
2327 need to make a unique copy of I2SRC each time we substitute it
2328 to avoid self-referential rtl. */
2330 subst_low_cuid = INSN_CUID (i2);
2331 newpat = subst (PATTERN (i3), i2dest, i2src, 0,
2332 ! i1_feeds_i3 && i1dest_in_i1src);
2333 substed_i2 = 1;
2335 /* Record whether i2's body now appears within i3's body. */
2336 i2_is_used = n_occurrences;
2339 /* If we already got a failure, don't try to do more. Otherwise,
2340 try to substitute in I1 if we have it. */
2342 if (i1 && GET_CODE (newpat) != CLOBBER)
2344 /* Before we can do this substitution, we must redo the test done
2345 above (see detailed comments there) that ensures that I1DEST
2346 isn't mentioned in any SETs in NEWPAT that are field assignments. */
2348 if (! combinable_i3pat (NULL_RTX, &newpat, i1dest, NULL_RTX,
2349 0, (rtx*) 0))
2351 undo_all ();
2352 return 0;
2355 n_occurrences = 0;
2356 subst_low_cuid = INSN_CUID (i1);
2357 newpat = subst (newpat, i1dest, i1src, 0, 0);
2358 substed_i1 = 1;
2361 /* Fail if an autoincrement side-effect has been duplicated. Be careful
2362 to count all the ways that I2SRC and I1SRC can be used. */
2363 if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0
2364 && i2_is_used + added_sets_2 > 1)
2365 || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0
2366 && (n_occurrences + added_sets_1 + (added_sets_2 && ! i1_feeds_i3)
2367 > 1))
2368 /* Fail if we tried to make a new register. */
2369 || max_reg_num () != maxreg
2370 /* Fail if we couldn't do something and have a CLOBBER. */
2371 || GET_CODE (newpat) == CLOBBER
2372 /* Fail if this new pattern is a MULT and we didn't have one before
2373 at the outer level. */
2374 || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT
2375 && ! have_mult))
2377 undo_all ();
2378 return 0;
2381 /* If the actions of the earlier insns must be kept
2382 in addition to substituting them into the latest one,
2383 we must make a new PARALLEL for the latest insn
2384 to hold additional the SETs. */
2386 if (added_sets_1 || added_sets_2)
2388 combine_extras++;
2390 if (GET_CODE (newpat) == PARALLEL)
2392 rtvec old = XVEC (newpat, 0);
2393 total_sets = XVECLEN (newpat, 0) + added_sets_1 + added_sets_2;
2394 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2395 memcpy (XVEC (newpat, 0)->elem, &old->elem[0],
2396 sizeof (old->elem[0]) * old->num_elem);
2398 else
2400 rtx old = newpat;
2401 total_sets = 1 + added_sets_1 + added_sets_2;
2402 newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets));
2403 XVECEXP (newpat, 0, 0) = old;
2406 if (added_sets_1)
2407 XVECEXP (newpat, 0, --total_sets)
2408 = (GET_CODE (PATTERN (i1)) == PARALLEL
2409 ? gen_rtx_SET (VOIDmode, i1dest, i1src) : PATTERN (i1));
2411 if (added_sets_2)
2413 /* If there is no I1, use I2's body as is. We used to also not do
2414 the subst call below if I2 was substituted into I3,
2415 but that could lose a simplification. */
2416 if (i1 == 0)
2417 XVECEXP (newpat, 0, --total_sets) = i2pat;
2418 else
2419 /* See comment where i2pat is assigned. */
2420 XVECEXP (newpat, 0, --total_sets)
2421 = subst (i2pat, i1dest, i1src, 0, 0);
2425 /* We come here when we are replacing a destination in I2 with the
2426 destination of I3. */
2427 validate_replacement:
2429 /* Note which hard regs this insn has as inputs. */
2430 mark_used_regs_combine (newpat);
2432 /* If recog_for_combine fails, it strips existing clobbers. If we'll
2433 consider splitting this pattern, we might need these clobbers. */
2434 if (i1 && GET_CODE (newpat) == PARALLEL
2435 && GET_CODE (XVECEXP (newpat, 0, XVECLEN (newpat, 0) - 1)) == CLOBBER)
2437 int len = XVECLEN (newpat, 0);
2439 newpat_vec_with_clobbers = rtvec_alloc (len);
2440 for (i = 0; i < len; i++)
2441 RTVEC_ELT (newpat_vec_with_clobbers, i) = XVECEXP (newpat, 0, i);
2444 /* Is the result of combination a valid instruction? */
2445 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2447 /* If the result isn't valid, see if it is a PARALLEL of two SETs where
2448 the second SET's destination is a register that is unused and isn't
2449 marked as an instruction that might trap in an EH region. In that case,
2450 we just need the first SET. This can occur when simplifying a divmod
2451 insn. We *must* test for this case here because the code below that
2452 splits two independent SETs doesn't handle this case correctly when it
2453 updates the register status.
2455 It's pointless doing this if we originally had two sets, one from
2456 i3, and one from i2. Combining then splitting the parallel results
2457 in the original i2 again plus an invalid insn (which we delete).
2458 The net effect is only to move instructions around, which makes
2459 debug info less accurate.
2461 Also check the case where the first SET's destination is unused.
2462 That would not cause incorrect code, but does cause an unneeded
2463 insn to remain. */
2465 if (insn_code_number < 0
2466 && !(added_sets_2 && i1 == 0)
2467 && GET_CODE (newpat) == PARALLEL
2468 && XVECLEN (newpat, 0) == 2
2469 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2470 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2471 && asm_noperands (newpat) < 0)
2473 rtx set0 = XVECEXP (newpat, 0, 0);
2474 rtx set1 = XVECEXP (newpat, 0, 1);
2475 rtx note;
2477 if (((REG_P (SET_DEST (set1))
2478 && find_reg_note (i3, REG_UNUSED, SET_DEST (set1)))
2479 || (GET_CODE (SET_DEST (set1)) == SUBREG
2480 && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1)))))
2481 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2482 || INTVAL (XEXP (note, 0)) <= 0)
2483 && ! side_effects_p (SET_SRC (set1)))
2485 newpat = set0;
2486 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2489 else if (((REG_P (SET_DEST (set0))
2490 && find_reg_note (i3, REG_UNUSED, SET_DEST (set0)))
2491 || (GET_CODE (SET_DEST (set0)) == SUBREG
2492 && find_reg_note (i3, REG_UNUSED,
2493 SUBREG_REG (SET_DEST (set0)))))
2494 && (!(note = find_reg_note (i3, REG_EH_REGION, NULL_RTX))
2495 || INTVAL (XEXP (note, 0)) <= 0)
2496 && ! side_effects_p (SET_SRC (set0)))
2498 newpat = set1;
2499 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2501 if (insn_code_number >= 0)
2503 /* If we will be able to accept this, we have made a
2504 change to the destination of I3. This requires us to
2505 do a few adjustments. */
2507 PATTERN (i3) = newpat;
2508 adjust_for_new_dest (i3);
2513 /* If we were combining three insns and the result is a simple SET
2514 with no ASM_OPERANDS that wasn't recognized, try to split it into two
2515 insns. There are two ways to do this. It can be split using a
2516 machine-specific method (like when you have an addition of a large
2517 constant) or by combine in the function find_split_point. */
2519 if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET
2520 && asm_noperands (newpat) < 0)
2522 rtx m_split, *split;
2524 /* See if the MD file can split NEWPAT. If it can't, see if letting it
2525 use I2DEST as a scratch register will help. In the latter case,
2526 convert I2DEST to the mode of the source of NEWPAT if we can. */
2528 m_split = split_insns (newpat, i3);
2530 /* We can only use I2DEST as a scratch reg if it doesn't overlap any
2531 inputs of NEWPAT. */
2533 /* ??? If I2DEST is not safe, and I1DEST exists, then it would be
2534 possible to try that as a scratch reg. This would require adding
2535 more code to make it work though. */
2537 if (m_split == 0 && ! reg_overlap_mentioned_p (i2dest, newpat))
2539 enum machine_mode new_mode = GET_MODE (SET_DEST (newpat));
2541 /* First try to split using the original register as a
2542 scratch register. */
2543 m_split = split_insns (gen_rtx_PARALLEL
2544 (VOIDmode,
2545 gen_rtvec (2, newpat,
2546 gen_rtx_CLOBBER (VOIDmode,
2547 i2dest))),
2548 i3);
2550 /* If that didn't work, try changing the mode of I2DEST if
2551 we can. */
2552 if (m_split == 0
2553 && new_mode != GET_MODE (i2dest)
2554 && new_mode != VOIDmode
2555 && can_change_dest_mode (i2dest, added_sets_2, new_mode))
2557 enum machine_mode old_mode = GET_MODE (i2dest);
2558 rtx ni2dest;
2560 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
2561 ni2dest = gen_rtx_REG (new_mode, REGNO (i2dest));
2562 else
2564 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], new_mode);
2565 ni2dest = regno_reg_rtx[REGNO (i2dest)];
2568 m_split = split_insns (gen_rtx_PARALLEL
2569 (VOIDmode,
2570 gen_rtvec (2, newpat,
2571 gen_rtx_CLOBBER (VOIDmode,
2572 ni2dest))),
2573 i3);
2575 if (m_split == 0
2576 && REGNO (i2dest) >= FIRST_PSEUDO_REGISTER)
2578 struct undo *buf;
2580 PUT_MODE (regno_reg_rtx[REGNO (i2dest)], old_mode);
2581 buf = undobuf.undos;
2582 undobuf.undos = buf->next;
2583 buf->next = undobuf.frees;
2584 undobuf.frees = buf;
2589 /* If recog_for_combine has discarded clobbers, try to use them
2590 again for the split. */
2591 if (m_split == 0 && newpat_vec_with_clobbers)
2592 m_split
2593 = split_insns (gen_rtx_PARALLEL (VOIDmode,
2594 newpat_vec_with_clobbers), i3);
2596 if (m_split && NEXT_INSN (m_split) == NULL_RTX)
2598 m_split = PATTERN (m_split);
2599 insn_code_number = recog_for_combine (&m_split, i3, &new_i3_notes);
2600 if (insn_code_number >= 0)
2601 newpat = m_split;
2603 else if (m_split && NEXT_INSN (NEXT_INSN (m_split)) == NULL_RTX
2604 && (next_real_insn (i2) == i3
2605 || ! use_crosses_set_p (PATTERN (m_split), INSN_CUID (i2))))
2607 rtx i2set, i3set;
2608 rtx newi3pat = PATTERN (NEXT_INSN (m_split));
2609 newi2pat = PATTERN (m_split);
2611 i3set = single_set (NEXT_INSN (m_split));
2612 i2set = single_set (m_split);
2614 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2616 /* If I2 or I3 has multiple SETs, we won't know how to track
2617 register status, so don't use these insns. If I2's destination
2618 is used between I2 and I3, we also can't use these insns. */
2620 if (i2_code_number >= 0 && i2set && i3set
2621 && (next_real_insn (i2) == i3
2622 || ! reg_used_between_p (SET_DEST (i2set), i2, i3)))
2623 insn_code_number = recog_for_combine (&newi3pat, i3,
2624 &new_i3_notes);
2625 if (insn_code_number >= 0)
2626 newpat = newi3pat;
2628 /* It is possible that both insns now set the destination of I3.
2629 If so, we must show an extra use of it. */
2631 if (insn_code_number >= 0)
2633 rtx new_i3_dest = SET_DEST (i3set);
2634 rtx new_i2_dest = SET_DEST (i2set);
2636 while (GET_CODE (new_i3_dest) == ZERO_EXTRACT
2637 || GET_CODE (new_i3_dest) == STRICT_LOW_PART
2638 || GET_CODE (new_i3_dest) == SUBREG)
2639 new_i3_dest = XEXP (new_i3_dest, 0);
2641 while (GET_CODE (new_i2_dest) == ZERO_EXTRACT
2642 || GET_CODE (new_i2_dest) == STRICT_LOW_PART
2643 || GET_CODE (new_i2_dest) == SUBREG)
2644 new_i2_dest = XEXP (new_i2_dest, 0);
2646 if (REG_P (new_i3_dest)
2647 && REG_P (new_i2_dest)
2648 && REGNO (new_i3_dest) == REGNO (new_i2_dest))
2649 REG_N_SETS (REGNO (new_i2_dest))++;
2653 /* If we can split it and use I2DEST, go ahead and see if that
2654 helps things be recognized. Verify that none of the registers
2655 are set between I2 and I3. */
2656 if (insn_code_number < 0 && (split = find_split_point (&newpat, i3)) != 0
2657 #ifdef HAVE_cc0
2658 && REG_P (i2dest)
2659 #endif
2660 /* We need I2DEST in the proper mode. If it is a hard register
2661 or the only use of a pseudo, we can change its mode.
2662 Make sure we don't change a hard register to have a mode that
2663 isn't valid for it, or change the number of registers. */
2664 && (GET_MODE (*split) == GET_MODE (i2dest)
2665 || GET_MODE (*split) == VOIDmode
2666 || can_change_dest_mode (i2dest, added_sets_2,
2667 GET_MODE (*split)))
2668 && (next_real_insn (i2) == i3
2669 || ! use_crosses_set_p (*split, INSN_CUID (i2)))
2670 /* We can't overwrite I2DEST if its value is still used by
2671 NEWPAT. */
2672 && ! reg_referenced_p (i2dest, newpat))
2674 rtx newdest = i2dest;
2675 enum rtx_code split_code = GET_CODE (*split);
2676 enum machine_mode split_mode = GET_MODE (*split);
2677 bool subst_done = false;
2678 newi2pat = NULL_RTX;
2680 /* Get NEWDEST as a register in the proper mode. We have already
2681 validated that we can do this. */
2682 if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode)
2684 if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER)
2685 newdest = gen_rtx_REG (split_mode, REGNO (i2dest));
2686 else
2688 SUBST_MODE (regno_reg_rtx[REGNO (i2dest)], split_mode);
2689 newdest = regno_reg_rtx[REGNO (i2dest)];
2693 /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to
2694 an ASHIFT. This can occur if it was inside a PLUS and hence
2695 appeared to be a memory address. This is a kludge. */
2696 if (split_code == MULT
2697 && GET_CODE (XEXP (*split, 1)) == CONST_INT
2698 && INTVAL (XEXP (*split, 1)) > 0
2699 && (i = exact_log2 (INTVAL (XEXP (*split, 1)))) >= 0)
2701 SUBST (*split, gen_rtx_ASHIFT (split_mode,
2702 XEXP (*split, 0), GEN_INT (i)));
2703 /* Update split_code because we may not have a multiply
2704 anymore. */
2705 split_code = GET_CODE (*split);
2708 #ifdef INSN_SCHEDULING
2709 /* If *SPLIT is a paradoxical SUBREG, when we split it, it should
2710 be written as a ZERO_EXTEND. */
2711 if (split_code == SUBREG && MEM_P (SUBREG_REG (*split)))
2713 #ifdef LOAD_EXTEND_OP
2714 /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's
2715 what it really is. */
2716 if (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (*split)))
2717 == SIGN_EXTEND)
2718 SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode,
2719 SUBREG_REG (*split)));
2720 else
2721 #endif
2722 SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode,
2723 SUBREG_REG (*split)));
2725 #endif
2727 /* Attempt to split binary operators using arithmetic identities. */
2728 if (BINARY_P (SET_SRC (newpat))
2729 && split_mode == GET_MODE (SET_SRC (newpat))
2730 && ! side_effects_p (SET_SRC (newpat)))
2732 rtx setsrc = SET_SRC (newpat);
2733 enum machine_mode mode = GET_MODE (setsrc);
2734 enum rtx_code code = GET_CODE (setsrc);
2735 rtx src_op0 = XEXP (setsrc, 0);
2736 rtx src_op1 = XEXP (setsrc, 1);
2738 /* Split "X = Y op Y" as "Z = Y; X = Z op Z". */
2739 if (rtx_equal_p (src_op0, src_op1))
2741 newi2pat = gen_rtx_SET (VOIDmode, newdest, src_op0);
2742 SUBST (XEXP (setsrc, 0), newdest);
2743 SUBST (XEXP (setsrc, 1), newdest);
2744 subst_done = true;
2746 /* Split "((P op Q) op R) op S" where op is PLUS or MULT. */
2747 else if ((code == PLUS || code == MULT)
2748 && GET_CODE (src_op0) == code
2749 && GET_CODE (XEXP (src_op0, 0)) == code
2750 && (INTEGRAL_MODE_P (mode)
2751 || (FLOAT_MODE_P (mode)
2752 && flag_unsafe_math_optimizations)))
2754 rtx p = XEXP (XEXP (src_op0, 0), 0);
2755 rtx q = XEXP (XEXP (src_op0, 0), 1);
2756 rtx r = XEXP (src_op0, 1);
2757 rtx s = src_op1;
2759 /* Split both "((X op Y) op X) op Y" and
2760 "((X op Y) op Y) op X" as "T op T" where T is
2761 "X op Y". */
2762 if ((rtx_equal_p (p,r) && rtx_equal_p (q,s))
2763 || (rtx_equal_p (p,s) && rtx_equal_p (q,r)))
2765 newi2pat = gen_rtx_SET (VOIDmode, newdest,
2766 XEXP (src_op0, 0));
2767 SUBST (XEXP (setsrc, 0), newdest);
2768 SUBST (XEXP (setsrc, 1), newdest);
2769 subst_done = true;
2771 /* Split "((X op X) op Y) op Y)" as "T op T" where
2772 T is "X op Y". */
2773 else if (rtx_equal_p (p,q) && rtx_equal_p (r,s))
2775 rtx tmp = simplify_gen_binary (code, mode, p, r);
2776 newi2pat = gen_rtx_SET (VOIDmode, newdest, tmp);
2777 SUBST (XEXP (setsrc, 0), newdest);
2778 SUBST (XEXP (setsrc, 1), newdest);
2779 subst_done = true;
2784 if (!subst_done)
2786 newi2pat = gen_rtx_SET (VOIDmode, newdest, *split);
2787 SUBST (*split, newdest);
2790 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2792 /* recog_for_combine might have added CLOBBERs to newi2pat.
2793 Make sure NEWPAT does not depend on the clobbered regs. */
2794 if (GET_CODE (newi2pat) == PARALLEL)
2795 for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--)
2796 if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER)
2798 rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0);
2799 if (reg_overlap_mentioned_p (reg, newpat))
2801 undo_all ();
2802 return 0;
2806 /* If the split point was a MULT and we didn't have one before,
2807 don't use one now. */
2808 if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult))
2809 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2813 /* Check for a case where we loaded from memory in a narrow mode and
2814 then sign extended it, but we need both registers. In that case,
2815 we have a PARALLEL with both loads from the same memory location.
2816 We can split this into a load from memory followed by a register-register
2817 copy. This saves at least one insn, more if register allocation can
2818 eliminate the copy.
2820 We cannot do this if the destination of the first assignment is a
2821 condition code register or cc0. We eliminate this case by making sure
2822 the SET_DEST and SET_SRC have the same mode.
2824 We cannot do this if the destination of the second assignment is
2825 a register that we have already assumed is zero-extended. Similarly
2826 for a SUBREG of such a register. */
2828 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2829 && GET_CODE (newpat) == PARALLEL
2830 && XVECLEN (newpat, 0) == 2
2831 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2832 && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND
2833 && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0)))
2834 == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0))))
2835 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2836 && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2837 XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0))
2838 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2839 INSN_CUID (i2))
2840 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2841 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2842 && ! (temp = SET_DEST (XVECEXP (newpat, 0, 1)),
2843 (REG_P (temp)
2844 && reg_stat[REGNO (temp)].nonzero_bits != 0
2845 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2846 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2847 && (reg_stat[REGNO (temp)].nonzero_bits
2848 != GET_MODE_MASK (word_mode))))
2849 && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG
2850 && (temp = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))),
2851 (REG_P (temp)
2852 && reg_stat[REGNO (temp)].nonzero_bits != 0
2853 && GET_MODE_BITSIZE (GET_MODE (temp)) < BITS_PER_WORD
2854 && GET_MODE_BITSIZE (GET_MODE (temp)) < HOST_BITS_PER_INT
2855 && (reg_stat[REGNO (temp)].nonzero_bits
2856 != GET_MODE_MASK (word_mode)))))
2857 && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2858 SET_SRC (XVECEXP (newpat, 0, 1)))
2859 && ! find_reg_note (i3, REG_UNUSED,
2860 SET_DEST (XVECEXP (newpat, 0, 0))))
2862 rtx ni2dest;
2864 newi2pat = XVECEXP (newpat, 0, 0);
2865 ni2dest = SET_DEST (XVECEXP (newpat, 0, 0));
2866 newpat = XVECEXP (newpat, 0, 1);
2867 SUBST (SET_SRC (newpat),
2868 gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest));
2869 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2871 if (i2_code_number >= 0)
2872 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2874 if (insn_code_number >= 0)
2875 swap_i2i3 = 1;
2878 /* Similarly, check for a case where we have a PARALLEL of two independent
2879 SETs but we started with three insns. In this case, we can do the sets
2880 as two separate insns. This case occurs when some SET allows two
2881 other insns to combine, but the destination of that SET is still live. */
2883 else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0
2884 && GET_CODE (newpat) == PARALLEL
2885 && XVECLEN (newpat, 0) == 2
2886 && GET_CODE (XVECEXP (newpat, 0, 0)) == SET
2887 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT
2888 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART
2889 && GET_CODE (XVECEXP (newpat, 0, 1)) == SET
2890 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT
2891 && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART
2892 && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
2893 INSN_CUID (i2))
2894 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)),
2895 XVECEXP (newpat, 0, 0))
2896 && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)),
2897 XVECEXP (newpat, 0, 1))
2898 && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0)))
2899 && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1)))))
2901 /* Normally, it doesn't matter which of the two is done first,
2902 but it does if one references cc0. In that case, it has to
2903 be first. */
2904 #ifdef HAVE_cc0
2905 if (reg_referenced_p (cc0_rtx, XVECEXP (newpat, 0, 0)))
2907 newi2pat = XVECEXP (newpat, 0, 0);
2908 newpat = XVECEXP (newpat, 0, 1);
2910 else
2911 #endif
2913 newi2pat = XVECEXP (newpat, 0, 1);
2914 newpat = XVECEXP (newpat, 0, 0);
2917 i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes);
2919 if (i2_code_number >= 0)
2920 insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
2923 /* If it still isn't recognized, fail and change things back the way they
2924 were. */
2925 if ((insn_code_number < 0
2926 /* Is the result a reasonable ASM_OPERANDS? */
2927 && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2)))
2929 undo_all ();
2930 return 0;
2933 /* If we had to change another insn, make sure it is valid also. */
2934 if (undobuf.other_insn)
2936 rtx other_pat = PATTERN (undobuf.other_insn);
2937 rtx new_other_notes;
2938 rtx note, next;
2940 CLEAR_HARD_REG_SET (newpat_used_regs);
2942 other_code_number = recog_for_combine (&other_pat, undobuf.other_insn,
2943 &new_other_notes);
2945 if (other_code_number < 0 && ! check_asm_operands (other_pat))
2947 undo_all ();
2948 return 0;
2951 PATTERN (undobuf.other_insn) = other_pat;
2953 /* If any of the notes in OTHER_INSN were REG_UNUSED, ensure that they
2954 are still valid. Then add any non-duplicate notes added by
2955 recog_for_combine. */
2956 for (note = REG_NOTES (undobuf.other_insn); note; note = next)
2958 next = XEXP (note, 1);
2960 if (REG_NOTE_KIND (note) == REG_UNUSED
2961 && ! reg_set_p (XEXP (note, 0), PATTERN (undobuf.other_insn)))
2963 if (REG_P (XEXP (note, 0)))
2964 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
2966 remove_note (undobuf.other_insn, note);
2970 for (note = new_other_notes; note; note = XEXP (note, 1))
2971 if (REG_P (XEXP (note, 0)))
2972 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
2974 distribute_notes (new_other_notes, undobuf.other_insn,
2975 undobuf.other_insn, NULL_RTX, NULL_RTX, NULL_RTX);
2977 #ifdef HAVE_cc0
2978 /* If I2 is the CC0 setter and I3 is the CC0 user then check whether
2979 they are adjacent to each other or not. */
2981 rtx p = prev_nonnote_insn (i3);
2982 if (p && p != i2 && NONJUMP_INSN_P (p) && newi2pat
2983 && sets_cc0_p (newi2pat))
2985 undo_all ();
2986 return 0;
2989 #endif
2991 /* Only allow this combination if insn_rtx_costs reports that the
2992 replacement instructions are cheaper than the originals. */
2993 if (!combine_validate_cost (i1, i2, i3, newpat, newi2pat))
2995 undo_all ();
2996 return 0;
2999 /* We now know that we can do this combination. Merge the insns and
3000 update the status of registers and LOG_LINKS. */
3002 if (swap_i2i3)
3004 rtx insn;
3005 rtx link;
3006 rtx ni2dest;
3008 /* I3 now uses what used to be its destination and which is now
3009 I2's destination. This requires us to do a few adjustments. */
3010 PATTERN (i3) = newpat;
3011 adjust_for_new_dest (i3);
3013 /* We need a LOG_LINK from I3 to I2. But we used to have one,
3014 so we still will.
3016 However, some later insn might be using I2's dest and have
3017 a LOG_LINK pointing at I3. We must remove this link.
3018 The simplest way to remove the link is to point it at I1,
3019 which we know will be a NOTE. */
3021 /* newi2pat is usually a SET here; however, recog_for_combine might
3022 have added some clobbers. */
3023 if (GET_CODE (newi2pat) == PARALLEL)
3024 ni2dest = SET_DEST (XVECEXP (newi2pat, 0, 0));
3025 else
3026 ni2dest = SET_DEST (newi2pat);
3028 for (insn = NEXT_INSN (i3);
3029 insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3030 || insn != BB_HEAD (this_basic_block->next_bb));
3031 insn = NEXT_INSN (insn))
3033 if (INSN_P (insn) && reg_referenced_p (ni2dest, PATTERN (insn)))
3035 for (link = LOG_LINKS (insn); link;
3036 link = XEXP (link, 1))
3037 if (XEXP (link, 0) == i3)
3038 XEXP (link, 0) = i1;
3040 break;
3046 rtx i3notes, i2notes, i1notes = 0;
3047 rtx i3links, i2links, i1links = 0;
3048 rtx midnotes = 0;
3049 unsigned int regno;
3050 /* Compute which registers we expect to eliminate. newi2pat may be setting
3051 either i3dest or i2dest, so we must check it. Also, i1dest may be the
3052 same as i3dest, in which case newi2pat may be setting i1dest. */
3053 rtx elim_i2 = ((newi2pat && reg_set_p (i2dest, newi2pat))
3054 || i2dest_in_i2src || i2dest_in_i1src
3055 || !i2dest_killed
3056 ? 0 : i2dest);
3057 rtx elim_i1 = (i1 == 0 || i1dest_in_i1src
3058 || (newi2pat && reg_set_p (i1dest, newi2pat))
3059 || !i1dest_killed
3060 ? 0 : i1dest);
3062 /* Get the old REG_NOTES and LOG_LINKS from all our insns and
3063 clear them. */
3064 i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3);
3065 i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2);
3066 if (i1)
3067 i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1);
3069 /* Ensure that we do not have something that should not be shared but
3070 occurs multiple times in the new insns. Check this by first
3071 resetting all the `used' flags and then copying anything is shared. */
3073 reset_used_flags (i3notes);
3074 reset_used_flags (i2notes);
3075 reset_used_flags (i1notes);
3076 reset_used_flags (newpat);
3077 reset_used_flags (newi2pat);
3078 if (undobuf.other_insn)
3079 reset_used_flags (PATTERN (undobuf.other_insn));
3081 i3notes = copy_rtx_if_shared (i3notes);
3082 i2notes = copy_rtx_if_shared (i2notes);
3083 i1notes = copy_rtx_if_shared (i1notes);
3084 newpat = copy_rtx_if_shared (newpat);
3085 newi2pat = copy_rtx_if_shared (newi2pat);
3086 if (undobuf.other_insn)
3087 reset_used_flags (PATTERN (undobuf.other_insn));
3089 INSN_CODE (i3) = insn_code_number;
3090 PATTERN (i3) = newpat;
3092 if (CALL_P (i3) && CALL_INSN_FUNCTION_USAGE (i3))
3094 rtx call_usage = CALL_INSN_FUNCTION_USAGE (i3);
3096 reset_used_flags (call_usage);
3097 call_usage = copy_rtx (call_usage);
3099 if (substed_i2)
3100 replace_rtx (call_usage, i2dest, i2src);
3102 if (substed_i1)
3103 replace_rtx (call_usage, i1dest, i1src);
3105 CALL_INSN_FUNCTION_USAGE (i3) = call_usage;
3108 if (undobuf.other_insn)
3109 INSN_CODE (undobuf.other_insn) = other_code_number;
3111 /* We had one special case above where I2 had more than one set and
3112 we replaced a destination of one of those sets with the destination
3113 of I3. In that case, we have to update LOG_LINKS of insns later
3114 in this basic block. Note that this (expensive) case is rare.
3116 Also, in this case, we must pretend that all REG_NOTEs for I2
3117 actually came from I3, so that REG_UNUSED notes from I2 will be
3118 properly handled. */
3120 if (i3_subst_into_i2)
3122 for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++)
3123 if ((GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == SET
3124 || GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == CLOBBER)
3125 && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, i)))
3126 && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest
3127 && ! find_reg_note (i2, REG_UNUSED,
3128 SET_DEST (XVECEXP (PATTERN (i2), 0, i))))
3129 for (temp = NEXT_INSN (i2);
3130 temp && (this_basic_block->next_bb == EXIT_BLOCK_PTR
3131 || BB_HEAD (this_basic_block) != temp);
3132 temp = NEXT_INSN (temp))
3133 if (temp != i3 && INSN_P (temp))
3134 for (link = LOG_LINKS (temp); link; link = XEXP (link, 1))
3135 if (XEXP (link, 0) == i2)
3136 XEXP (link, 0) = i3;
3138 if (i3notes)
3140 rtx link = i3notes;
3141 while (XEXP (link, 1))
3142 link = XEXP (link, 1);
3143 XEXP (link, 1) = i2notes;
3145 else
3146 i3notes = i2notes;
3147 i2notes = 0;
3150 LOG_LINKS (i3) = 0;
3151 REG_NOTES (i3) = 0;
3152 LOG_LINKS (i2) = 0;
3153 REG_NOTES (i2) = 0;
3155 if (newi2pat)
3157 INSN_CODE (i2) = i2_code_number;
3158 PATTERN (i2) = newi2pat;
3160 else
3161 SET_INSN_DELETED (i2);
3163 if (i1)
3165 LOG_LINKS (i1) = 0;
3166 REG_NOTES (i1) = 0;
3167 SET_INSN_DELETED (i1);
3170 /* Get death notes for everything that is now used in either I3 or
3171 I2 and used to die in a previous insn. If we built two new
3172 patterns, move from I1 to I2 then I2 to I3 so that we get the
3173 proper movement on registers that I2 modifies. */
3175 if (newi2pat)
3177 move_deaths (newi2pat, NULL_RTX, INSN_CUID (i1), i2, &midnotes);
3178 move_deaths (newpat, newi2pat, INSN_CUID (i1), i3, &midnotes);
3180 else
3181 move_deaths (newpat, NULL_RTX, i1 ? INSN_CUID (i1) : INSN_CUID (i2),
3182 i3, &midnotes);
3184 /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */
3185 if (i3notes)
3186 distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL_RTX,
3187 elim_i2, elim_i1);
3188 if (i2notes)
3189 distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL_RTX,
3190 elim_i2, elim_i1);
3191 if (i1notes)
3192 distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL_RTX,
3193 elim_i2, elim_i1);
3194 if (midnotes)
3195 distribute_notes (midnotes, NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3196 elim_i2, elim_i1);
3198 /* Distribute any notes added to I2 or I3 by recog_for_combine. We
3199 know these are REG_UNUSED and want them to go to the desired insn,
3200 so we always pass it as i3. We have not counted the notes in
3201 reg_n_deaths yet, so we need to do so now. */
3203 if (newi2pat && new_i2_notes)
3205 for (temp = new_i2_notes; temp; temp = XEXP (temp, 1))
3206 if (REG_P (XEXP (temp, 0)))
3207 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
3209 distribute_notes (new_i2_notes, i2, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3212 if (new_i3_notes)
3214 for (temp = new_i3_notes; temp; temp = XEXP (temp, 1))
3215 if (REG_P (XEXP (temp, 0)))
3216 REG_N_DEATHS (REGNO (XEXP (temp, 0)))++;
3218 distribute_notes (new_i3_notes, i3, i3, NULL_RTX, NULL_RTX, NULL_RTX);
3221 /* If I3DEST was used in I3SRC, it really died in I3. We may need to
3222 put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets
3223 I3DEST, the death must be somewhere before I2, not I3. If we passed I3
3224 in that case, it might delete I2. Similarly for I2 and I1.
3225 Show an additional death due to the REG_DEAD note we make here. If
3226 we discard it in distribute_notes, we will decrement it again. */
3228 if (i3dest_killed)
3230 if (REG_P (i3dest_killed))
3231 REG_N_DEATHS (REGNO (i3dest_killed))++;
3233 if (newi2pat && reg_set_p (i3dest_killed, newi2pat))
3234 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
3235 NULL_RTX),
3236 NULL_RTX, i2, NULL_RTX, elim_i2, elim_i1);
3237 else
3238 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i3dest_killed,
3239 NULL_RTX),
3240 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3241 elim_i2, elim_i1);
3244 if (i2dest_in_i2src)
3246 if (REG_P (i2dest))
3247 REG_N_DEATHS (REGNO (i2dest))++;
3249 if (newi2pat && reg_set_p (i2dest, newi2pat))
3250 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
3251 NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3252 else
3253 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i2dest, NULL_RTX),
3254 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3255 NULL_RTX, NULL_RTX);
3258 if (i1dest_in_i1src)
3260 if (REG_P (i1dest))
3261 REG_N_DEATHS (REGNO (i1dest))++;
3263 if (newi2pat && reg_set_p (i1dest, newi2pat))
3264 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
3265 NULL_RTX, i2, NULL_RTX, NULL_RTX, NULL_RTX);
3266 else
3267 distribute_notes (gen_rtx_EXPR_LIST (REG_DEAD, i1dest, NULL_RTX),
3268 NULL_RTX, i3, newi2pat ? i2 : NULL_RTX,
3269 NULL_RTX, NULL_RTX);
3272 distribute_links (i3links);
3273 distribute_links (i2links);
3274 distribute_links (i1links);
3276 if (REG_P (i2dest))
3278 rtx link;
3279 rtx i2_insn = 0, i2_val = 0, set;
3281 /* The insn that used to set this register doesn't exist, and
3282 this life of the register may not exist either. See if one of
3283 I3's links points to an insn that sets I2DEST. If it does,
3284 that is now the last known value for I2DEST. If we don't update
3285 this and I2 set the register to a value that depended on its old
3286 contents, we will get confused. If this insn is used, thing
3287 will be set correctly in combine_instructions. */
3289 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
3290 if ((set = single_set (XEXP (link, 0))) != 0
3291 && rtx_equal_p (i2dest, SET_DEST (set)))
3292 i2_insn = XEXP (link, 0), i2_val = SET_SRC (set);
3294 record_value_for_reg (i2dest, i2_insn, i2_val);
3296 /* If the reg formerly set in I2 died only once and that was in I3,
3297 zero its use count so it won't make `reload' do any work. */
3298 if (! added_sets_2
3299 && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat))
3300 && ! i2dest_in_i2src)
3302 regno = REGNO (i2dest);
3303 REG_N_SETS (regno)--;
3307 if (i1 && REG_P (i1dest))
3309 rtx link;
3310 rtx i1_insn = 0, i1_val = 0, set;
3312 for (link = LOG_LINKS (i3); link; link = XEXP (link, 1))
3313 if ((set = single_set (XEXP (link, 0))) != 0
3314 && rtx_equal_p (i1dest, SET_DEST (set)))
3315 i1_insn = XEXP (link, 0), i1_val = SET_SRC (set);
3317 record_value_for_reg (i1dest, i1_insn, i1_val);
3319 regno = REGNO (i1dest);
3320 if (! added_sets_1 && ! i1dest_in_i1src)
3321 REG_N_SETS (regno)--;
3324 /* Update reg_stat[].nonzero_bits et al for any changes that may have
3325 been made to this insn. The order of
3326 set_nonzero_bits_and_sign_copies() is important. Because newi2pat
3327 can affect nonzero_bits of newpat */
3328 if (newi2pat)
3329 note_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL);
3330 note_stores (newpat, set_nonzero_bits_and_sign_copies, NULL);
3332 /* Set new_direct_jump_p if a new return or simple jump instruction
3333 has been created.
3335 If I3 is now an unconditional jump, ensure that it has a
3336 BARRIER following it since it may have initially been a
3337 conditional jump. It may also be the last nonnote insn. */
3339 if (returnjump_p (i3) || any_uncondjump_p (i3))
3341 *new_direct_jump_p = 1;
3342 mark_jump_label (PATTERN (i3), i3, 0);
3344 if ((temp = next_nonnote_insn (i3)) == NULL_RTX
3345 || !BARRIER_P (temp))
3346 emit_barrier_after (i3);
3349 if (undobuf.other_insn != NULL_RTX
3350 && (returnjump_p (undobuf.other_insn)
3351 || any_uncondjump_p (undobuf.other_insn)))
3353 *new_direct_jump_p = 1;
3355 if ((temp = next_nonnote_insn (undobuf.other_insn)) == NULL_RTX
3356 || !BARRIER_P (temp))
3357 emit_barrier_after (undobuf.other_insn);
3360 /* An NOOP jump does not need barrier, but it does need cleaning up
3361 of CFG. */
3362 if (GET_CODE (newpat) == SET
3363 && SET_SRC (newpat) == pc_rtx
3364 && SET_DEST (newpat) == pc_rtx)
3365 *new_direct_jump_p = 1;
3368 combine_successes++;
3369 undo_commit ();
3371 if (added_links_insn
3372 && (newi2pat == 0 || INSN_CUID (added_links_insn) < INSN_CUID (i2))
3373 && INSN_CUID (added_links_insn) < INSN_CUID (i3))
3374 return added_links_insn;
3375 else
3376 return newi2pat ? i2 : i3;
3379 /* Undo all the modifications recorded in undobuf. */
3381 static void
3382 undo_all (void)
3384 struct undo *undo, *next;
3386 for (undo = undobuf.undos; undo; undo = next)
3388 next = undo->next;
3389 switch (undo->kind)
3391 case UNDO_RTX:
3392 *undo->where.r = undo->old_contents.r;
3393 break;
3394 case UNDO_INT:
3395 *undo->where.i = undo->old_contents.i;
3396 break;
3397 case UNDO_MODE:
3398 PUT_MODE (*undo->where.r, undo->old_contents.m);
3399 break;
3400 default:
3401 gcc_unreachable ();
3404 undo->next = undobuf.frees;
3405 undobuf.frees = undo;
3408 undobuf.undos = 0;
3411 /* We've committed to accepting the changes we made. Move all
3412 of the undos to the free list. */
3414 static void
3415 undo_commit (void)
3417 struct undo *undo, *next;
3419 for (undo = undobuf.undos; undo; undo = next)
3421 next = undo->next;
3422 undo->next = undobuf.frees;
3423 undobuf.frees = undo;
3425 undobuf.undos = 0;
3429 /* Find the innermost point within the rtx at LOC, possibly LOC itself,
3430 where we have an arithmetic expression and return that point. LOC will
3431 be inside INSN.
3433 try_combine will call this function to see if an insn can be split into
3434 two insns. */
3436 static rtx *
3437 find_split_point (rtx *loc, rtx insn)
3439 rtx x = *loc;
3440 enum rtx_code code = GET_CODE (x);
3441 rtx *split;
3442 unsigned HOST_WIDE_INT len = 0;
3443 HOST_WIDE_INT pos = 0;
3444 int unsignedp = 0;
3445 rtx inner = NULL_RTX;
3447 /* First special-case some codes. */
3448 switch (code)
3450 case SUBREG:
3451 #ifdef INSN_SCHEDULING
3452 /* If we are making a paradoxical SUBREG invalid, it becomes a split
3453 point. */
3454 if (MEM_P (SUBREG_REG (x)))
3455 return loc;
3456 #endif
3457 return find_split_point (&SUBREG_REG (x), insn);
3459 case MEM:
3460 #ifdef HAVE_lo_sum
3461 /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it
3462 using LO_SUM and HIGH. */
3463 if (GET_CODE (XEXP (x, 0)) == CONST
3464 || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)
3466 SUBST (XEXP (x, 0),
3467 gen_rtx_LO_SUM (Pmode,
3468 gen_rtx_HIGH (Pmode, XEXP (x, 0)),
3469 XEXP (x, 0)));
3470 return &XEXP (XEXP (x, 0), 0);
3472 #endif
3474 /* If we have a PLUS whose second operand is a constant and the
3475 address is not valid, perhaps will can split it up using
3476 the machine-specific way to split large constants. We use
3477 the first pseudo-reg (one of the virtual regs) as a placeholder;
3478 it will not remain in the result. */
3479 if (GET_CODE (XEXP (x, 0)) == PLUS
3480 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
3481 && ! memory_address_p (GET_MODE (x), XEXP (x, 0)))
3483 rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER];
3484 rtx seq = split_insns (gen_rtx_SET (VOIDmode, reg, XEXP (x, 0)),
3485 subst_insn);
3487 /* This should have produced two insns, each of which sets our
3488 placeholder. If the source of the second is a valid address,
3489 we can make put both sources together and make a split point
3490 in the middle. */
3492 if (seq
3493 && NEXT_INSN (seq) != NULL_RTX
3494 && NEXT_INSN (NEXT_INSN (seq)) == NULL_RTX
3495 && NONJUMP_INSN_P (seq)
3496 && GET_CODE (PATTERN (seq)) == SET
3497 && SET_DEST (PATTERN (seq)) == reg
3498 && ! reg_mentioned_p (reg,
3499 SET_SRC (PATTERN (seq)))
3500 && NONJUMP_INSN_P (NEXT_INSN (seq))
3501 && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET
3502 && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg
3503 && memory_address_p (GET_MODE (x),
3504 SET_SRC (PATTERN (NEXT_INSN (seq)))))
3506 rtx src1 = SET_SRC (PATTERN (seq));
3507 rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq)));
3509 /* Replace the placeholder in SRC2 with SRC1. If we can
3510 find where in SRC2 it was placed, that can become our
3511 split point and we can replace this address with SRC2.
3512 Just try two obvious places. */
3514 src2 = replace_rtx (src2, reg, src1);
3515 split = 0;
3516 if (XEXP (src2, 0) == src1)
3517 split = &XEXP (src2, 0);
3518 else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e'
3519 && XEXP (XEXP (src2, 0), 0) == src1)
3520 split = &XEXP (XEXP (src2, 0), 0);
3522 if (split)
3524 SUBST (XEXP (x, 0), src2);
3525 return split;
3529 /* If that didn't work, perhaps the first operand is complex and
3530 needs to be computed separately, so make a split point there.
3531 This will occur on machines that just support REG + CONST
3532 and have a constant moved through some previous computation. */
3534 else if (!OBJECT_P (XEXP (XEXP (x, 0), 0))
3535 && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG
3536 && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0)))))
3537 return &XEXP (XEXP (x, 0), 0);
3539 break;
3541 case SET:
3542 #ifdef HAVE_cc0
3543 /* If SET_DEST is CC0 and SET_SRC is not an operand, a COMPARE, or a
3544 ZERO_EXTRACT, the most likely reason why this doesn't match is that
3545 we need to put the operand into a register. So split at that
3546 point. */
3548 if (SET_DEST (x) == cc0_rtx
3549 && GET_CODE (SET_SRC (x)) != COMPARE
3550 && GET_CODE (SET_SRC (x)) != ZERO_EXTRACT
3551 && !OBJECT_P (SET_SRC (x))
3552 && ! (GET_CODE (SET_SRC (x)) == SUBREG
3553 && OBJECT_P (SUBREG_REG (SET_SRC (x)))))
3554 return &SET_SRC (x);
3555 #endif
3557 /* See if we can split SET_SRC as it stands. */
3558 split = find_split_point (&SET_SRC (x), insn);
3559 if (split && split != &SET_SRC (x))
3560 return split;
3562 /* See if we can split SET_DEST as it stands. */
3563 split = find_split_point (&SET_DEST (x), insn);
3564 if (split && split != &SET_DEST (x))
3565 return split;
3567 /* See if this is a bitfield assignment with everything constant. If
3568 so, this is an IOR of an AND, so split it into that. */
3569 if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
3570 && (GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)))
3571 <= HOST_BITS_PER_WIDE_INT)
3572 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT
3573 && GET_CODE (XEXP (SET_DEST (x), 2)) == CONST_INT
3574 && GET_CODE (SET_SRC (x)) == CONST_INT
3575 && ((INTVAL (XEXP (SET_DEST (x), 1))
3576 + INTVAL (XEXP (SET_DEST (x), 2)))
3577 <= GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0))))
3578 && ! side_effects_p (XEXP (SET_DEST (x), 0)))
3580 HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2));
3581 unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1));
3582 unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x));
3583 rtx dest = XEXP (SET_DEST (x), 0);
3584 enum machine_mode mode = GET_MODE (dest);
3585 unsigned HOST_WIDE_INT mask = ((HOST_WIDE_INT) 1 << len) - 1;
3586 rtx or_mask;
3588 if (BITS_BIG_ENDIAN)
3589 pos = GET_MODE_BITSIZE (mode) - len - pos;
3591 or_mask = gen_int_mode (src << pos, mode);
3592 if (src == mask)
3593 SUBST (SET_SRC (x),
3594 simplify_gen_binary (IOR, mode, dest, or_mask));
3595 else
3597 rtx negmask = gen_int_mode (~(mask << pos), mode);
3598 SUBST (SET_SRC (x),
3599 simplify_gen_binary (IOR, mode,
3600 simplify_gen_binary (AND, mode,
3601 dest, negmask),
3602 or_mask));
3605 SUBST (SET_DEST (x), dest);
3607 split = find_split_point (&SET_SRC (x), insn);
3608 if (split && split != &SET_SRC (x))
3609 return split;
3612 /* Otherwise, see if this is an operation that we can split into two.
3613 If so, try to split that. */
3614 code = GET_CODE (SET_SRC (x));
3616 switch (code)
3618 case AND:
3619 /* If we are AND'ing with a large constant that is only a single
3620 bit and the result is only being used in a context where we
3621 need to know if it is zero or nonzero, replace it with a bit
3622 extraction. This will avoid the large constant, which might
3623 have taken more than one insn to make. If the constant were
3624 not a valid argument to the AND but took only one insn to make,
3625 this is no worse, but if it took more than one insn, it will
3626 be better. */
3628 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3629 && REG_P (XEXP (SET_SRC (x), 0))
3630 && (pos = exact_log2 (INTVAL (XEXP (SET_SRC (x), 1)))) >= 7
3631 && REG_P (SET_DEST (x))
3632 && (split = find_single_use (SET_DEST (x), insn, (rtx*) 0)) != 0
3633 && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE)
3634 && XEXP (*split, 0) == SET_DEST (x)
3635 && XEXP (*split, 1) == const0_rtx)
3637 rtx extraction = make_extraction (GET_MODE (SET_DEST (x)),
3638 XEXP (SET_SRC (x), 0),
3639 pos, NULL_RTX, 1, 1, 0, 0);
3640 if (extraction != 0)
3642 SUBST (SET_SRC (x), extraction);
3643 return find_split_point (loc, insn);
3646 break;
3648 case NE:
3649 /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X
3650 is known to be on, this can be converted into a NEG of a shift. */
3651 if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx
3652 && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0))
3653 && 1 <= (pos = exact_log2
3654 (nonzero_bits (XEXP (SET_SRC (x), 0),
3655 GET_MODE (XEXP (SET_SRC (x), 0))))))
3657 enum machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0));
3659 SUBST (SET_SRC (x),
3660 gen_rtx_NEG (mode,
3661 gen_rtx_LSHIFTRT (mode,
3662 XEXP (SET_SRC (x), 0),
3663 GEN_INT (pos))));
3665 split = find_split_point (&SET_SRC (x), insn);
3666 if (split && split != &SET_SRC (x))
3667 return split;
3669 break;
3671 case SIGN_EXTEND:
3672 inner = XEXP (SET_SRC (x), 0);
3674 /* We can't optimize if either mode is a partial integer
3675 mode as we don't know how many bits are significant
3676 in those modes. */
3677 if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_PARTIAL_INT
3678 || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT)
3679 break;
3681 pos = 0;
3682 len = GET_MODE_BITSIZE (GET_MODE (inner));
3683 unsignedp = 0;
3684 break;
3686 case SIGN_EXTRACT:
3687 case ZERO_EXTRACT:
3688 if (GET_CODE (XEXP (SET_SRC (x), 1)) == CONST_INT
3689 && GET_CODE (XEXP (SET_SRC (x), 2)) == CONST_INT)
3691 inner = XEXP (SET_SRC (x), 0);
3692 len = INTVAL (XEXP (SET_SRC (x), 1));
3693 pos = INTVAL (XEXP (SET_SRC (x), 2));
3695 if (BITS_BIG_ENDIAN)
3696 pos = GET_MODE_BITSIZE (GET_MODE (inner)) - len - pos;
3697 unsignedp = (code == ZERO_EXTRACT);
3699 break;
3701 default:
3702 break;
3705 if (len && pos >= 0 && pos + len <= GET_MODE_BITSIZE (GET_MODE (inner)))
3707 enum machine_mode mode = GET_MODE (SET_SRC (x));
3709 /* For unsigned, we have a choice of a shift followed by an
3710 AND or two shifts. Use two shifts for field sizes where the
3711 constant might be too large. We assume here that we can
3712 always at least get 8-bit constants in an AND insn, which is
3713 true for every current RISC. */
3715 if (unsignedp && len <= 8)
3717 SUBST (SET_SRC (x),
3718 gen_rtx_AND (mode,
3719 gen_rtx_LSHIFTRT
3720 (mode, gen_lowpart (mode, inner),
3721 GEN_INT (pos)),
3722 GEN_INT (((HOST_WIDE_INT) 1 << len) - 1)));
3724 split = find_split_point (&SET_SRC (x), insn);
3725 if (split && split != &SET_SRC (x))
3726 return split;
3728 else
3730 SUBST (SET_SRC (x),
3731 gen_rtx_fmt_ee
3732 (unsignedp ? LSHIFTRT : ASHIFTRT, mode,
3733 gen_rtx_ASHIFT (mode,
3734 gen_lowpart (mode, inner),
3735 GEN_INT (GET_MODE_BITSIZE (mode)
3736 - len - pos)),
3737 GEN_INT (GET_MODE_BITSIZE (mode) - len)));
3739 split = find_split_point (&SET_SRC (x), insn);
3740 if (split && split != &SET_SRC (x))
3741 return split;
3745 /* See if this is a simple operation with a constant as the second
3746 operand. It might be that this constant is out of range and hence
3747 could be used as a split point. */
3748 if (BINARY_P (SET_SRC (x))
3749 && CONSTANT_P (XEXP (SET_SRC (x), 1))
3750 && (OBJECT_P (XEXP (SET_SRC (x), 0))
3751 || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG
3752 && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0))))))
3753 return &XEXP (SET_SRC (x), 1);
3755 /* Finally, see if this is a simple operation with its first operand
3756 not in a register. The operation might require this operand in a
3757 register, so return it as a split point. We can always do this
3758 because if the first operand were another operation, we would have
3759 already found it as a split point. */
3760 if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x)))
3761 && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode))
3762 return &XEXP (SET_SRC (x), 0);
3764 return 0;
3766 case AND:
3767 case IOR:
3768 /* We write NOR as (and (not A) (not B)), but if we don't have a NOR,
3769 it is better to write this as (not (ior A B)) so we can split it.
3770 Similarly for IOR. */
3771 if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT)
3773 SUBST (*loc,
3774 gen_rtx_NOT (GET_MODE (x),
3775 gen_rtx_fmt_ee (code == IOR ? AND : IOR,
3776 GET_MODE (x),
3777 XEXP (XEXP (x, 0), 0),
3778 XEXP (XEXP (x, 1), 0))));
3779 return find_split_point (loc, insn);
3782 /* Many RISC machines have a large set of logical insns. If the
3783 second operand is a NOT, put it first so we will try to split the
3784 other operand first. */
3785 if (GET_CODE (XEXP (x, 1)) == NOT)
3787 rtx tem = XEXP (x, 0);
3788 SUBST (XEXP (x, 0), XEXP (x, 1));
3789 SUBST (XEXP (x, 1), tem);
3791 break;
3793 default:
3794 break;
3797 /* Otherwise, select our actions depending on our rtx class. */
3798 switch (GET_RTX_CLASS (code))
3800 case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */
3801 case RTX_TERNARY:
3802 split = find_split_point (&XEXP (x, 2), insn);
3803 if (split)
3804 return split;
3805 /* ... fall through ... */
3806 case RTX_BIN_ARITH:
3807 case RTX_COMM_ARITH:
3808 case RTX_COMPARE:
3809 case RTX_COMM_COMPARE:
3810 split = find_split_point (&XEXP (x, 1), insn);
3811 if (split)
3812 return split;
3813 /* ... fall through ... */
3814 case RTX_UNARY:
3815 /* Some machines have (and (shift ...) ...) insns. If X is not
3816 an AND, but XEXP (X, 0) is, use it as our split point. */
3817 if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND)
3818 return &XEXP (x, 0);
3820 split = find_split_point (&XEXP (x, 0), insn);
3821 if (split)
3822 return split;
3823 return loc;
3825 default:
3826 /* Otherwise, we don't have a split point. */
3827 return 0;
3831 /* Throughout X, replace FROM with TO, and return the result.
3832 The result is TO if X is FROM;
3833 otherwise the result is X, but its contents may have been modified.
3834 If they were modified, a record was made in undobuf so that
3835 undo_all will (among other things) return X to its original state.
3837 If the number of changes necessary is too much to record to undo,
3838 the excess changes are not made, so the result is invalid.
3839 The changes already made can still be undone.
3840 undobuf.num_undo is incremented for such changes, so by testing that
3841 the caller can tell whether the result is valid.
3843 `n_occurrences' is incremented each time FROM is replaced.
3845 IN_DEST is nonzero if we are processing the SET_DEST of a SET.
3847 UNIQUE_COPY is nonzero if each substitution must be unique. We do this
3848 by copying if `n_occurrences' is nonzero. */
3850 static rtx
3851 subst (rtx x, rtx from, rtx to, int in_dest, int unique_copy)
3853 enum rtx_code code = GET_CODE (x);
3854 enum machine_mode op0_mode = VOIDmode;
3855 const char *fmt;
3856 int len, i;
3857 rtx new;
3859 /* Two expressions are equal if they are identical copies of a shared
3860 RTX or if they are both registers with the same register number
3861 and mode. */
3863 #define COMBINE_RTX_EQUAL_P(X,Y) \
3864 ((X) == (Y) \
3865 || (REG_P (X) && REG_P (Y) \
3866 && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y)))
3868 if (! in_dest && COMBINE_RTX_EQUAL_P (x, from))
3870 n_occurrences++;
3871 return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to);
3874 /* If X and FROM are the same register but different modes, they will
3875 not have been seen as equal above. However, flow.c will make a
3876 LOG_LINKS entry for that case. If we do nothing, we will try to
3877 rerecognize our original insn and, when it succeeds, we will
3878 delete the feeding insn, which is incorrect.
3880 So force this insn not to match in this (rare) case. */
3881 if (! in_dest && code == REG && REG_P (from)
3882 && REGNO (x) == REGNO (from))
3883 return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
3885 /* If this is an object, we are done unless it is a MEM or LO_SUM, both
3886 of which may contain things that can be combined. */
3887 if (code != MEM && code != LO_SUM && OBJECT_P (x))
3888 return x;
3890 /* It is possible to have a subexpression appear twice in the insn.
3891 Suppose that FROM is a register that appears within TO.
3892 Then, after that subexpression has been scanned once by `subst',
3893 the second time it is scanned, TO may be found. If we were
3894 to scan TO here, we would find FROM within it and create a
3895 self-referent rtl structure which is completely wrong. */
3896 if (COMBINE_RTX_EQUAL_P (x, to))
3897 return to;
3899 /* Parallel asm_operands need special attention because all of the
3900 inputs are shared across the arms. Furthermore, unsharing the
3901 rtl results in recognition failures. Failure to handle this case
3902 specially can result in circular rtl.
3904 Solve this by doing a normal pass across the first entry of the
3905 parallel, and only processing the SET_DESTs of the subsequent
3906 entries. Ug. */
3908 if (code == PARALLEL
3909 && GET_CODE (XVECEXP (x, 0, 0)) == SET
3910 && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS)
3912 new = subst (XVECEXP (x, 0, 0), from, to, 0, unique_copy);
3914 /* If this substitution failed, this whole thing fails. */
3915 if (GET_CODE (new) == CLOBBER
3916 && XEXP (new, 0) == const0_rtx)
3917 return new;
3919 SUBST (XVECEXP (x, 0, 0), new);
3921 for (i = XVECLEN (x, 0) - 1; i >= 1; i--)
3923 rtx dest = SET_DEST (XVECEXP (x, 0, i));
3925 if (!REG_P (dest)
3926 && GET_CODE (dest) != CC0
3927 && GET_CODE (dest) != PC)
3929 new = subst (dest, from, to, 0, unique_copy);
3931 /* If this substitution failed, this whole thing fails. */
3932 if (GET_CODE (new) == CLOBBER
3933 && XEXP (new, 0) == const0_rtx)
3934 return new;
3936 SUBST (SET_DEST (XVECEXP (x, 0, i)), new);
3940 else
3942 len = GET_RTX_LENGTH (code);
3943 fmt = GET_RTX_FORMAT (code);
3945 /* We don't need to process a SET_DEST that is a register, CC0,
3946 or PC, so set up to skip this common case. All other cases
3947 where we want to suppress replacing something inside a
3948 SET_SRC are handled via the IN_DEST operand. */
3949 if (code == SET
3950 && (REG_P (SET_DEST (x))
3951 || GET_CODE (SET_DEST (x)) == CC0
3952 || GET_CODE (SET_DEST (x)) == PC))
3953 fmt = "ie";
3955 /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a
3956 constant. */
3957 if (fmt[0] == 'e')
3958 op0_mode = GET_MODE (XEXP (x, 0));
3960 for (i = 0; i < len; i++)
3962 if (fmt[i] == 'E')
3964 int j;
3965 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3967 if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from))
3969 new = (unique_copy && n_occurrences
3970 ? copy_rtx (to) : to);
3971 n_occurrences++;
3973 else
3975 new = subst (XVECEXP (x, i, j), from, to, 0,
3976 unique_copy);
3978 /* If this substitution failed, this whole thing
3979 fails. */
3980 if (GET_CODE (new) == CLOBBER
3981 && XEXP (new, 0) == const0_rtx)
3982 return new;
3985 SUBST (XVECEXP (x, i, j), new);
3988 else if (fmt[i] == 'e')
3990 /* If this is a register being set, ignore it. */
3991 new = XEXP (x, i);
3992 if (in_dest
3993 && i == 0
3994 && (((code == SUBREG || code == ZERO_EXTRACT)
3995 && REG_P (new))
3996 || code == STRICT_LOW_PART))
3999 else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from))
4001 /* In general, don't install a subreg involving two
4002 modes not tieable. It can worsen register
4003 allocation, and can even make invalid reload
4004 insns, since the reg inside may need to be copied
4005 from in the outside mode, and that may be invalid
4006 if it is an fp reg copied in integer mode.
4008 We allow two exceptions to this: It is valid if
4009 it is inside another SUBREG and the mode of that
4010 SUBREG and the mode of the inside of TO is
4011 tieable and it is valid if X is a SET that copies
4012 FROM to CC0. */
4014 if (GET_CODE (to) == SUBREG
4015 && ! MODES_TIEABLE_P (GET_MODE (to),
4016 GET_MODE (SUBREG_REG (to)))
4017 && ! (code == SUBREG
4018 && MODES_TIEABLE_P (GET_MODE (x),
4019 GET_MODE (SUBREG_REG (to))))
4020 #ifdef HAVE_cc0
4021 && ! (code == SET && i == 1 && XEXP (x, 0) == cc0_rtx)
4022 #endif
4024 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
4026 #ifdef CANNOT_CHANGE_MODE_CLASS
4027 if (code == SUBREG
4028 && REG_P (to)
4029 && REGNO (to) < FIRST_PSEUDO_REGISTER
4030 && REG_CANNOT_CHANGE_MODE_P (REGNO (to),
4031 GET_MODE (to),
4032 GET_MODE (x)))
4033 return gen_rtx_CLOBBER (VOIDmode, const0_rtx);
4034 #endif
4036 new = (unique_copy && n_occurrences ? copy_rtx (to) : to);
4037 n_occurrences++;
4039 else
4040 /* If we are in a SET_DEST, suppress most cases unless we
4041 have gone inside a MEM, in which case we want to
4042 simplify the address. We assume here that things that
4043 are actually part of the destination have their inner
4044 parts in the first expression. This is true for SUBREG,
4045 STRICT_LOW_PART, and ZERO_EXTRACT, which are the only
4046 things aside from REG and MEM that should appear in a
4047 SET_DEST. */
4048 new = subst (XEXP (x, i), from, to,
4049 (((in_dest
4050 && (code == SUBREG || code == STRICT_LOW_PART
4051 || code == ZERO_EXTRACT))
4052 || code == SET)
4053 && i == 0), unique_copy);
4055 /* If we found that we will have to reject this combination,
4056 indicate that by returning the CLOBBER ourselves, rather than
4057 an expression containing it. This will speed things up as
4058 well as prevent accidents where two CLOBBERs are considered
4059 to be equal, thus producing an incorrect simplification. */
4061 if (GET_CODE (new) == CLOBBER && XEXP (new, 0) == const0_rtx)
4062 return new;
4064 if (GET_CODE (x) == SUBREG
4065 && (GET_CODE (new) == CONST_INT
4066 || GET_CODE (new) == CONST_DOUBLE))
4068 enum machine_mode mode = GET_MODE (x);
4070 x = simplify_subreg (GET_MODE (x), new,
4071 GET_MODE (SUBREG_REG (x)),
4072 SUBREG_BYTE (x));
4073 if (! x)
4074 x = gen_rtx_CLOBBER (mode, const0_rtx);
4076 else if (GET_CODE (new) == CONST_INT
4077 && GET_CODE (x) == ZERO_EXTEND)
4079 x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
4080 new, GET_MODE (XEXP (x, 0)));
4081 gcc_assert (x);
4083 else
4084 SUBST (XEXP (x, i), new);
4089 /* Try to simplify X. If the simplification changed the code, it is likely
4090 that further simplification will help, so loop, but limit the number
4091 of repetitions that will be performed. */
4093 for (i = 0; i < 4; i++)
4095 /* If X is sufficiently simple, don't bother trying to do anything
4096 with it. */
4097 if (code != CONST_INT && code != REG && code != CLOBBER)
4098 x = combine_simplify_rtx (x, op0_mode, in_dest);
4100 if (GET_CODE (x) == code)
4101 break;
4103 code = GET_CODE (x);
4105 /* We no longer know the original mode of operand 0 since we
4106 have changed the form of X) */
4107 op0_mode = VOIDmode;
4110 return x;
4113 /* Simplify X, a piece of RTL. We just operate on the expression at the
4114 outer level; call `subst' to simplify recursively. Return the new
4115 expression.
4117 OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero
4118 if we are inside a SET_DEST. */
4120 static rtx
4121 combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest)
4123 enum rtx_code code = GET_CODE (x);
4124 enum machine_mode mode = GET_MODE (x);
4125 rtx temp;
4126 int i;
4128 /* If this is a commutative operation, put a constant last and a complex
4129 expression first. We don't need to do this for comparisons here. */
4130 if (COMMUTATIVE_ARITH_P (x)
4131 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
4133 temp = XEXP (x, 0);
4134 SUBST (XEXP (x, 0), XEXP (x, 1));
4135 SUBST (XEXP (x, 1), temp);
4138 /* If this is a simple operation applied to an IF_THEN_ELSE, try
4139 applying it to the arms of the IF_THEN_ELSE. This often simplifies
4140 things. Check for cases where both arms are testing the same
4141 condition.
4143 Don't do anything if all operands are very simple. */
4145 if ((BINARY_P (x)
4146 && ((!OBJECT_P (XEXP (x, 0))
4147 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4148 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))
4149 || (!OBJECT_P (XEXP (x, 1))
4150 && ! (GET_CODE (XEXP (x, 1)) == SUBREG
4151 && OBJECT_P (SUBREG_REG (XEXP (x, 1)))))))
4152 || (UNARY_P (x)
4153 && (!OBJECT_P (XEXP (x, 0))
4154 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4155 && OBJECT_P (SUBREG_REG (XEXP (x, 0)))))))
4157 rtx cond, true_rtx, false_rtx;
4159 cond = if_then_else_cond (x, &true_rtx, &false_rtx);
4160 if (cond != 0
4161 /* If everything is a comparison, what we have is highly unlikely
4162 to be simpler, so don't use it. */
4163 && ! (COMPARISON_P (x)
4164 && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))))
4166 rtx cop1 = const0_rtx;
4167 enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1);
4169 if (cond_code == NE && COMPARISON_P (cond))
4170 return x;
4172 /* Simplify the alternative arms; this may collapse the true and
4173 false arms to store-flag values. Be careful to use copy_rtx
4174 here since true_rtx or false_rtx might share RTL with x as a
4175 result of the if_then_else_cond call above. */
4176 true_rtx = subst (copy_rtx (true_rtx), pc_rtx, pc_rtx, 0, 0);
4177 false_rtx = subst (copy_rtx (false_rtx), pc_rtx, pc_rtx, 0, 0);
4179 /* If true_rtx and false_rtx are not general_operands, an if_then_else
4180 is unlikely to be simpler. */
4181 if (general_operand (true_rtx, VOIDmode)
4182 && general_operand (false_rtx, VOIDmode))
4184 enum rtx_code reversed;
4186 /* Restarting if we generate a store-flag expression will cause
4187 us to loop. Just drop through in this case. */
4189 /* If the result values are STORE_FLAG_VALUE and zero, we can
4190 just make the comparison operation. */
4191 if (true_rtx == const_true_rtx && false_rtx == const0_rtx)
4192 x = simplify_gen_relational (cond_code, mode, VOIDmode,
4193 cond, cop1);
4194 else if (true_rtx == const0_rtx && false_rtx == const_true_rtx
4195 && ((reversed = reversed_comparison_code_parts
4196 (cond_code, cond, cop1, NULL))
4197 != UNKNOWN))
4198 x = simplify_gen_relational (reversed, mode, VOIDmode,
4199 cond, cop1);
4201 /* Likewise, we can make the negate of a comparison operation
4202 if the result values are - STORE_FLAG_VALUE and zero. */
4203 else if (GET_CODE (true_rtx) == CONST_INT
4204 && INTVAL (true_rtx) == - STORE_FLAG_VALUE
4205 && false_rtx == const0_rtx)
4206 x = simplify_gen_unary (NEG, mode,
4207 simplify_gen_relational (cond_code,
4208 mode, VOIDmode,
4209 cond, cop1),
4210 mode);
4211 else if (GET_CODE (false_rtx) == CONST_INT
4212 && INTVAL (false_rtx) == - STORE_FLAG_VALUE
4213 && true_rtx == const0_rtx
4214 && ((reversed = reversed_comparison_code_parts
4215 (cond_code, cond, cop1, NULL))
4216 != UNKNOWN))
4217 x = simplify_gen_unary (NEG, mode,
4218 simplify_gen_relational (reversed,
4219 mode, VOIDmode,
4220 cond, cop1),
4221 mode);
4222 else
4223 return gen_rtx_IF_THEN_ELSE (mode,
4224 simplify_gen_relational (cond_code,
4225 mode,
4226 VOIDmode,
4227 cond,
4228 cop1),
4229 true_rtx, false_rtx);
4231 code = GET_CODE (x);
4232 op0_mode = VOIDmode;
4237 /* Try to fold this expression in case we have constants that weren't
4238 present before. */
4239 temp = 0;
4240 switch (GET_RTX_CLASS (code))
4242 case RTX_UNARY:
4243 if (op0_mode == VOIDmode)
4244 op0_mode = GET_MODE (XEXP (x, 0));
4245 temp = simplify_unary_operation (code, mode, XEXP (x, 0), op0_mode);
4246 break;
4247 case RTX_COMPARE:
4248 case RTX_COMM_COMPARE:
4250 enum machine_mode cmp_mode = GET_MODE (XEXP (x, 0));
4251 if (cmp_mode == VOIDmode)
4253 cmp_mode = GET_MODE (XEXP (x, 1));
4254 if (cmp_mode == VOIDmode)
4255 cmp_mode = op0_mode;
4257 temp = simplify_relational_operation (code, mode, cmp_mode,
4258 XEXP (x, 0), XEXP (x, 1));
4260 break;
4261 case RTX_COMM_ARITH:
4262 case RTX_BIN_ARITH:
4263 temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1));
4264 break;
4265 case RTX_BITFIELD_OPS:
4266 case RTX_TERNARY:
4267 temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0),
4268 XEXP (x, 1), XEXP (x, 2));
4269 break;
4270 default:
4271 break;
4274 if (temp)
4276 x = temp;
4277 code = GET_CODE (temp);
4278 op0_mode = VOIDmode;
4279 mode = GET_MODE (temp);
4282 /* First see if we can apply the inverse distributive law. */
4283 if (code == PLUS || code == MINUS
4284 || code == AND || code == IOR || code == XOR)
4286 x = apply_distributive_law (x);
4287 code = GET_CODE (x);
4288 op0_mode = VOIDmode;
4291 /* If CODE is an associative operation not otherwise handled, see if we
4292 can associate some operands. This can win if they are constants or
4293 if they are logically related (i.e. (a & b) & a). */
4294 if ((code == PLUS || code == MINUS || code == MULT || code == DIV
4295 || code == AND || code == IOR || code == XOR
4296 || code == SMAX || code == SMIN || code == UMAX || code == UMIN)
4297 && ((INTEGRAL_MODE_P (mode) && code != DIV)
4298 || (flag_unsafe_math_optimizations && FLOAT_MODE_P (mode))))
4300 if (GET_CODE (XEXP (x, 0)) == code)
4302 rtx other = XEXP (XEXP (x, 0), 0);
4303 rtx inner_op0 = XEXP (XEXP (x, 0), 1);
4304 rtx inner_op1 = XEXP (x, 1);
4305 rtx inner;
4307 /* Make sure we pass the constant operand if any as the second
4308 one if this is a commutative operation. */
4309 if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x))
4311 rtx tem = inner_op0;
4312 inner_op0 = inner_op1;
4313 inner_op1 = tem;
4315 inner = simplify_binary_operation (code == MINUS ? PLUS
4316 : code == DIV ? MULT
4317 : code,
4318 mode, inner_op0, inner_op1);
4320 /* For commutative operations, try the other pair if that one
4321 didn't simplify. */
4322 if (inner == 0 && COMMUTATIVE_ARITH_P (x))
4324 other = XEXP (XEXP (x, 0), 1);
4325 inner = simplify_binary_operation (code, mode,
4326 XEXP (XEXP (x, 0), 0),
4327 XEXP (x, 1));
4330 if (inner)
4331 return simplify_gen_binary (code, mode, other, inner);
4335 /* A little bit of algebraic simplification here. */
4336 switch (code)
4338 case MEM:
4339 /* Ensure that our address has any ASHIFTs converted to MULT in case
4340 address-recognizing predicates are called later. */
4341 temp = make_compound_operation (XEXP (x, 0), MEM);
4342 SUBST (XEXP (x, 0), temp);
4343 break;
4345 case SUBREG:
4346 if (op0_mode == VOIDmode)
4347 op0_mode = GET_MODE (SUBREG_REG (x));
4349 /* See if this can be moved to simplify_subreg. */
4350 if (CONSTANT_P (SUBREG_REG (x))
4351 && subreg_lowpart_offset (mode, op0_mode) == SUBREG_BYTE (x)
4352 /* Don't call gen_lowpart if the inner mode
4353 is VOIDmode and we cannot simplify it, as SUBREG without
4354 inner mode is invalid. */
4355 && (GET_MODE (SUBREG_REG (x)) != VOIDmode
4356 || gen_lowpart_common (mode, SUBREG_REG (x))))
4357 return gen_lowpart (mode, SUBREG_REG (x));
4359 if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC)
4360 break;
4362 rtx temp;
4363 temp = simplify_subreg (mode, SUBREG_REG (x), op0_mode,
4364 SUBREG_BYTE (x));
4365 if (temp)
4366 return temp;
4369 /* Don't change the mode of the MEM if that would change the meaning
4370 of the address. */
4371 if (MEM_P (SUBREG_REG (x))
4372 && (MEM_VOLATILE_P (SUBREG_REG (x))
4373 || mode_dependent_address_p (XEXP (SUBREG_REG (x), 0))))
4374 return gen_rtx_CLOBBER (mode, const0_rtx);
4376 /* Note that we cannot do any narrowing for non-constants since
4377 we might have been counting on using the fact that some bits were
4378 zero. We now do this in the SET. */
4380 break;
4382 case NEG:
4383 temp = expand_compound_operation (XEXP (x, 0));
4385 /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be
4386 replaced by (lshiftrt X C). This will convert
4387 (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */
4389 if (GET_CODE (temp) == ASHIFTRT
4390 && GET_CODE (XEXP (temp, 1)) == CONST_INT
4391 && INTVAL (XEXP (temp, 1)) == GET_MODE_BITSIZE (mode) - 1)
4392 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (temp, 0),
4393 INTVAL (XEXP (temp, 1)));
4395 /* If X has only a single bit that might be nonzero, say, bit I, convert
4396 (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of
4397 MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to
4398 (sign_extract X 1 Y). But only do this if TEMP isn't a register
4399 or a SUBREG of one since we'd be making the expression more
4400 complex if it was just a register. */
4402 if (!REG_P (temp)
4403 && ! (GET_CODE (temp) == SUBREG
4404 && REG_P (SUBREG_REG (temp)))
4405 && (i = exact_log2 (nonzero_bits (temp, mode))) >= 0)
4407 rtx temp1 = simplify_shift_const
4408 (NULL_RTX, ASHIFTRT, mode,
4409 simplify_shift_const (NULL_RTX, ASHIFT, mode, temp,
4410 GET_MODE_BITSIZE (mode) - 1 - i),
4411 GET_MODE_BITSIZE (mode) - 1 - i);
4413 /* If all we did was surround TEMP with the two shifts, we
4414 haven't improved anything, so don't use it. Otherwise,
4415 we are better off with TEMP1. */
4416 if (GET_CODE (temp1) != ASHIFTRT
4417 || GET_CODE (XEXP (temp1, 0)) != ASHIFT
4418 || XEXP (XEXP (temp1, 0), 0) != temp)
4419 return temp1;
4421 break;
4423 case TRUNCATE:
4424 /* We can't handle truncation to a partial integer mode here
4425 because we don't know the real bitsize of the partial
4426 integer mode. */
4427 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
4428 break;
4430 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4431 && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
4432 GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))))
4433 SUBST (XEXP (x, 0),
4434 force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
4435 GET_MODE_MASK (mode), 0));
4437 /* Similarly to what we do in simplify-rtx.c, a truncate of a register
4438 whose value is a comparison can be replaced with a subreg if
4439 STORE_FLAG_VALUE permits. */
4440 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4441 && ((HOST_WIDE_INT) STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0
4442 && (temp = get_last_value (XEXP (x, 0)))
4443 && COMPARISON_P (temp))
4444 return gen_lowpart (mode, XEXP (x, 0));
4445 break;
4447 #ifdef HAVE_cc0
4448 case COMPARE:
4449 /* Convert (compare FOO (const_int 0)) to FOO unless we aren't
4450 using cc0, in which case we want to leave it as a COMPARE
4451 so we can distinguish it from a register-register-copy. */
4452 if (XEXP (x, 1) == const0_rtx)
4453 return XEXP (x, 0);
4455 /* x - 0 is the same as x unless x's mode has signed zeros and
4456 allows rounding towards -infinity. Under those conditions,
4457 0 - 0 is -0. */
4458 if (!(HONOR_SIGNED_ZEROS (GET_MODE (XEXP (x, 0)))
4459 && HONOR_SIGN_DEPENDENT_ROUNDING (GET_MODE (XEXP (x, 0))))
4460 && XEXP (x, 1) == CONST0_RTX (GET_MODE (XEXP (x, 0))))
4461 return XEXP (x, 0);
4462 break;
4463 #endif
4465 case CONST:
4466 /* (const (const X)) can become (const X). Do it this way rather than
4467 returning the inner CONST since CONST can be shared with a
4468 REG_EQUAL note. */
4469 if (GET_CODE (XEXP (x, 0)) == CONST)
4470 SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
4471 break;
4473 #ifdef HAVE_lo_sum
4474 case LO_SUM:
4475 /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we
4476 can add in an offset. find_split_point will split this address up
4477 again if it doesn't match. */
4478 if (GET_CODE (XEXP (x, 0)) == HIGH
4479 && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)))
4480 return XEXP (x, 1);
4481 break;
4482 #endif
4484 case PLUS:
4485 /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>)
4486 when c is (const_int (pow2 + 1) / 2) is a sign extension of a
4487 bit-field and can be replaced by either a sign_extend or a
4488 sign_extract. The `and' may be a zero_extend and the two
4489 <c>, -<c> constants may be reversed. */
4490 if (GET_CODE (XEXP (x, 0)) == XOR
4491 && GET_CODE (XEXP (x, 1)) == CONST_INT
4492 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
4493 && INTVAL (XEXP (x, 1)) == -INTVAL (XEXP (XEXP (x, 0), 1))
4494 && ((i = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) >= 0
4495 || (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
4496 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4497 && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND
4498 && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == CONST_INT
4499 && (INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1))
4500 == ((HOST_WIDE_INT) 1 << (i + 1)) - 1))
4501 || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND
4502 && (GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)))
4503 == (unsigned int) i + 1))))
4504 return simplify_shift_const
4505 (NULL_RTX, ASHIFTRT, mode,
4506 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4507 XEXP (XEXP (XEXP (x, 0), 0), 0),
4508 GET_MODE_BITSIZE (mode) - (i + 1)),
4509 GET_MODE_BITSIZE (mode) - (i + 1));
4511 /* If only the low-order bit of X is possibly nonzero, (plus x -1)
4512 can become (ashiftrt (ashift (xor x 1) C) C) where C is
4513 the bitsize of the mode - 1. This allows simplification of
4514 "a = (b & 8) == 0;" */
4515 if (XEXP (x, 1) == constm1_rtx
4516 && !REG_P (XEXP (x, 0))
4517 && ! (GET_CODE (XEXP (x, 0)) == SUBREG
4518 && REG_P (SUBREG_REG (XEXP (x, 0))))
4519 && nonzero_bits (XEXP (x, 0), mode) == 1)
4520 return simplify_shift_const (NULL_RTX, ASHIFTRT, mode,
4521 simplify_shift_const (NULL_RTX, ASHIFT, mode,
4522 gen_rtx_XOR (mode, XEXP (x, 0), const1_rtx),
4523 GET_MODE_BITSIZE (mode) - 1),
4524 GET_MODE_BITSIZE (mode) - 1);
4526 /* If we are adding two things that have no bits in common, convert
4527 the addition into an IOR. This will often be further simplified,
4528 for example in cases like ((a & 1) + (a & 2)), which can
4529 become a & 3. */
4531 if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4532 && (nonzero_bits (XEXP (x, 0), mode)
4533 & nonzero_bits (XEXP (x, 1), mode)) == 0)
4535 /* Try to simplify the expression further. */
4536 rtx tor = simplify_gen_binary (IOR, mode, XEXP (x, 0), XEXP (x, 1));
4537 temp = combine_simplify_rtx (tor, mode, in_dest);
4539 /* If we could, great. If not, do not go ahead with the IOR
4540 replacement, since PLUS appears in many special purpose
4541 address arithmetic instructions. */
4542 if (GET_CODE (temp) != CLOBBER && temp != tor)
4543 return temp;
4545 break;
4547 case MINUS:
4548 /* (minus <foo> (and <foo> (const_int -pow2))) becomes
4549 (and <foo> (const_int pow2-1)) */
4550 if (GET_CODE (XEXP (x, 1)) == AND
4551 && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT
4552 && exact_log2 (-INTVAL (XEXP (XEXP (x, 1), 1))) >= 0
4553 && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0)))
4554 return simplify_and_const_int (NULL_RTX, mode, XEXP (x, 0),
4555 -INTVAL (XEXP (XEXP (x, 1), 1)) - 1);
4556 break;
4558 case MULT:
4559 /* If we have (mult (plus A B) C), apply the distributive law and then
4560 the inverse distributive law to see if things simplify. This
4561 occurs mostly in addresses, often when unrolling loops. */
4563 if (GET_CODE (XEXP (x, 0)) == PLUS)
4565 rtx result = distribute_and_simplify_rtx (x, 0);
4566 if (result)
4567 return result;
4570 /* Try simplify a*(b/c) as (a*b)/c. */
4571 if (FLOAT_MODE_P (mode) && flag_unsafe_math_optimizations
4572 && GET_CODE (XEXP (x, 0)) == DIV)
4574 rtx tem = simplify_binary_operation (MULT, mode,
4575 XEXP (XEXP (x, 0), 0),
4576 XEXP (x, 1));
4577 if (tem)
4578 return simplify_gen_binary (DIV, mode, tem, XEXP (XEXP (x, 0), 1));
4580 break;
4582 case UDIV:
4583 /* If this is a divide by a power of two, treat it as a shift if
4584 its first operand is a shift. */
4585 if (GET_CODE (XEXP (x, 1)) == CONST_INT
4586 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0
4587 && (GET_CODE (XEXP (x, 0)) == ASHIFT
4588 || GET_CODE (XEXP (x, 0)) == LSHIFTRT
4589 || GET_CODE (XEXP (x, 0)) == ASHIFTRT
4590 || GET_CODE (XEXP (x, 0)) == ROTATE
4591 || GET_CODE (XEXP (x, 0)) == ROTATERT))
4592 return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (x, 0), i);
4593 break;
4595 case EQ: case NE:
4596 case GT: case GTU: case GE: case GEU:
4597 case LT: case LTU: case LE: case LEU:
4598 case UNEQ: case LTGT:
4599 case UNGT: case UNGE:
4600 case UNLT: case UNLE:
4601 case UNORDERED: case ORDERED:
4602 /* If the first operand is a condition code, we can't do anything
4603 with it. */
4604 if (GET_CODE (XEXP (x, 0)) == COMPARE
4605 || (GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC
4606 && ! CC0_P (XEXP (x, 0))))
4608 rtx op0 = XEXP (x, 0);
4609 rtx op1 = XEXP (x, 1);
4610 enum rtx_code new_code;
4612 if (GET_CODE (op0) == COMPARE)
4613 op1 = XEXP (op0, 1), op0 = XEXP (op0, 0);
4615 /* Simplify our comparison, if possible. */
4616 new_code = simplify_comparison (code, &op0, &op1);
4618 /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X
4619 if only the low-order bit is possibly nonzero in X (such as when
4620 X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to
4621 (xor X 1) or (minus 1 X); we use the former. Finally, if X is
4622 known to be either 0 or -1, NE becomes a NEG and EQ becomes
4623 (plus X 1).
4625 Remove any ZERO_EXTRACT we made when thinking this was a
4626 comparison. It may now be simpler to use, e.g., an AND. If a
4627 ZERO_EXTRACT is indeed appropriate, it will be placed back by
4628 the call to make_compound_operation in the SET case. */
4630 if (STORE_FLAG_VALUE == 1
4631 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4632 && op1 == const0_rtx
4633 && mode == GET_MODE (op0)
4634 && nonzero_bits (op0, mode) == 1)
4635 return gen_lowpart (mode,
4636 expand_compound_operation (op0));
4638 else if (STORE_FLAG_VALUE == 1
4639 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4640 && op1 == const0_rtx
4641 && mode == GET_MODE (op0)
4642 && (num_sign_bit_copies (op0, mode)
4643 == GET_MODE_BITSIZE (mode)))
4645 op0 = expand_compound_operation (op0);
4646 return simplify_gen_unary (NEG, mode,
4647 gen_lowpart (mode, op0),
4648 mode);
4651 else if (STORE_FLAG_VALUE == 1
4652 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4653 && op1 == const0_rtx
4654 && mode == GET_MODE (op0)
4655 && nonzero_bits (op0, mode) == 1)
4657 op0 = expand_compound_operation (op0);
4658 return simplify_gen_binary (XOR, mode,
4659 gen_lowpart (mode, op0),
4660 const1_rtx);
4663 else if (STORE_FLAG_VALUE == 1
4664 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4665 && op1 == const0_rtx
4666 && mode == GET_MODE (op0)
4667 && (num_sign_bit_copies (op0, mode)
4668 == GET_MODE_BITSIZE (mode)))
4670 op0 = expand_compound_operation (op0);
4671 return plus_constant (gen_lowpart (mode, op0), 1);
4674 /* If STORE_FLAG_VALUE is -1, we have cases similar to
4675 those above. */
4676 if (STORE_FLAG_VALUE == -1
4677 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4678 && op1 == const0_rtx
4679 && (num_sign_bit_copies (op0, mode)
4680 == GET_MODE_BITSIZE (mode)))
4681 return gen_lowpart (mode,
4682 expand_compound_operation (op0));
4684 else if (STORE_FLAG_VALUE == -1
4685 && new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4686 && op1 == const0_rtx
4687 && mode == GET_MODE (op0)
4688 && nonzero_bits (op0, mode) == 1)
4690 op0 = expand_compound_operation (op0);
4691 return simplify_gen_unary (NEG, mode,
4692 gen_lowpart (mode, op0),
4693 mode);
4696 else if (STORE_FLAG_VALUE == -1
4697 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4698 && op1 == const0_rtx
4699 && mode == GET_MODE (op0)
4700 && (num_sign_bit_copies (op0, mode)
4701 == GET_MODE_BITSIZE (mode)))
4703 op0 = expand_compound_operation (op0);
4704 return simplify_gen_unary (NOT, mode,
4705 gen_lowpart (mode, op0),
4706 mode);
4709 /* If X is 0/1, (eq X 0) is X-1. */
4710 else if (STORE_FLAG_VALUE == -1
4711 && new_code == EQ && GET_MODE_CLASS (mode) == MODE_INT
4712 && op1 == const0_rtx
4713 && mode == GET_MODE (op0)
4714 && nonzero_bits (op0, mode) == 1)
4716 op0 = expand_compound_operation (op0);
4717 return plus_constant (gen_lowpart (mode, op0), -1);
4720 /* If STORE_FLAG_VALUE says to just test the sign bit and X has just
4721 one bit that might be nonzero, we can convert (ne x 0) to
4722 (ashift x c) where C puts the bit in the sign bit. Remove any
4723 AND with STORE_FLAG_VALUE when we are done, since we are only
4724 going to test the sign bit. */
4725 if (new_code == NE && GET_MODE_CLASS (mode) == MODE_INT
4726 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
4727 && ((STORE_FLAG_VALUE & GET_MODE_MASK (mode))
4728 == (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1))
4729 && op1 == const0_rtx
4730 && mode == GET_MODE (op0)
4731 && (i = exact_log2 (nonzero_bits (op0, mode))) >= 0)
4733 x = simplify_shift_const (NULL_RTX, ASHIFT, mode,
4734 expand_compound_operation (op0),
4735 GET_MODE_BITSIZE (mode) - 1 - i);
4736 if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx)
4737 return XEXP (x, 0);
4738 else
4739 return x;
4742 /* If the code changed, return a whole new comparison. */
4743 if (new_code != code)
4744 return gen_rtx_fmt_ee (new_code, mode, op0, op1);
4746 /* Otherwise, keep this operation, but maybe change its operands.
4747 This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */
4748 SUBST (XEXP (x, 0), op0);
4749 SUBST (XEXP (x, 1), op1);
4751 break;
4753 case IF_THEN_ELSE:
4754 return simplify_if_then_else (x);
4756 case ZERO_EXTRACT:
4757 case SIGN_EXTRACT:
4758 case ZERO_EXTEND:
4759 case SIGN_EXTEND:
4760 /* If we are processing SET_DEST, we are done. */
4761 if (in_dest)
4762 return x;
4764 return expand_compound_operation (x);
4766 case SET:
4767 return simplify_set (x);
4769 case AND:
4770 case IOR:
4771 return simplify_logical (x);
4773 case ASHIFT:
4774 case LSHIFTRT:
4775 case ASHIFTRT:
4776 case ROTATE:
4777 case ROTATERT:
4778 /* If this is a shift by a constant amount, simplify it. */
4779 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
4780 return simplify_shift_const (x, code, mode, XEXP (x, 0),
4781 INTVAL (XEXP (x, 1)));
4783 else if (SHIFT_COUNT_TRUNCATED && !REG_P (XEXP (x, 1)))
4784 SUBST (XEXP (x, 1),
4785 force_to_mode (XEXP (x, 1), GET_MODE (XEXP (x, 1)),
4786 ((HOST_WIDE_INT) 1
4787 << exact_log2 (GET_MODE_BITSIZE (GET_MODE (x))))
4788 - 1,
4789 0));
4790 break;
4792 default:
4793 break;
4796 return x;
4799 /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */
4801 static rtx
4802 simplify_if_then_else (rtx x)
4804 enum machine_mode mode = GET_MODE (x);
4805 rtx cond = XEXP (x, 0);
4806 rtx true_rtx = XEXP (x, 1);
4807 rtx false_rtx = XEXP (x, 2);
4808 enum rtx_code true_code = GET_CODE (cond);
4809 int comparison_p = COMPARISON_P (cond);
4810 rtx temp;
4811 int i;
4812 enum rtx_code false_code;
4813 rtx reversed;
4815 /* Simplify storing of the truth value. */
4816 if (comparison_p && true_rtx == const_true_rtx && false_rtx == const0_rtx)
4817 return simplify_gen_relational (true_code, mode, VOIDmode,
4818 XEXP (cond, 0), XEXP (cond, 1));
4820 /* Also when the truth value has to be reversed. */
4821 if (comparison_p
4822 && true_rtx == const0_rtx && false_rtx == const_true_rtx
4823 && (reversed = reversed_comparison (cond, mode)))
4824 return reversed;
4826 /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used
4827 in it is being compared against certain values. Get the true and false
4828 comparisons and see if that says anything about the value of each arm. */
4830 if (comparison_p
4831 && ((false_code = reversed_comparison_code (cond, NULL))
4832 != UNKNOWN)
4833 && REG_P (XEXP (cond, 0)))
4835 HOST_WIDE_INT nzb;
4836 rtx from = XEXP (cond, 0);
4837 rtx true_val = XEXP (cond, 1);
4838 rtx false_val = true_val;
4839 int swapped = 0;
4841 /* If FALSE_CODE is EQ, swap the codes and arms. */
4843 if (false_code == EQ)
4845 swapped = 1, true_code = EQ, false_code = NE;
4846 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4849 /* If we are comparing against zero and the expression being tested has
4850 only a single bit that might be nonzero, that is its value when it is
4851 not equal to zero. Similarly if it is known to be -1 or 0. */
4853 if (true_code == EQ && true_val == const0_rtx
4854 && exact_log2 (nzb = nonzero_bits (from, GET_MODE (from))) >= 0)
4855 false_code = EQ, false_val = GEN_INT (nzb);
4856 else if (true_code == EQ && true_val == const0_rtx
4857 && (num_sign_bit_copies (from, GET_MODE (from))
4858 == GET_MODE_BITSIZE (GET_MODE (from))))
4859 false_code = EQ, false_val = constm1_rtx;
4861 /* Now simplify an arm if we know the value of the register in the
4862 branch and it is used in the arm. Be careful due to the potential
4863 of locally-shared RTL. */
4865 if (reg_mentioned_p (from, true_rtx))
4866 true_rtx = subst (known_cond (copy_rtx (true_rtx), true_code,
4867 from, true_val),
4868 pc_rtx, pc_rtx, 0, 0);
4869 if (reg_mentioned_p (from, false_rtx))
4870 false_rtx = subst (known_cond (copy_rtx (false_rtx), false_code,
4871 from, false_val),
4872 pc_rtx, pc_rtx, 0, 0);
4874 SUBST (XEXP (x, 1), swapped ? false_rtx : true_rtx);
4875 SUBST (XEXP (x, 2), swapped ? true_rtx : false_rtx);
4877 true_rtx = XEXP (x, 1);
4878 false_rtx = XEXP (x, 2);
4879 true_code = GET_CODE (cond);
4882 /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be
4883 reversed, do so to avoid needing two sets of patterns for
4884 subtract-and-branch insns. Similarly if we have a constant in the true
4885 arm, the false arm is the same as the first operand of the comparison, or
4886 the false arm is more complicated than the true arm. */
4888 if (comparison_p
4889 && reversed_comparison_code (cond, NULL) != UNKNOWN
4890 && (true_rtx == pc_rtx
4891 || (CONSTANT_P (true_rtx)
4892 && GET_CODE (false_rtx) != CONST_INT && false_rtx != pc_rtx)
4893 || true_rtx == const0_rtx
4894 || (OBJECT_P (true_rtx) && !OBJECT_P (false_rtx))
4895 || (GET_CODE (true_rtx) == SUBREG && OBJECT_P (SUBREG_REG (true_rtx))
4896 && !OBJECT_P (false_rtx))
4897 || reg_mentioned_p (true_rtx, false_rtx)
4898 || rtx_equal_p (false_rtx, XEXP (cond, 0))))
4900 true_code = reversed_comparison_code (cond, NULL);
4901 SUBST (XEXP (x, 0), reversed_comparison (cond, GET_MODE (cond)));
4902 SUBST (XEXP (x, 1), false_rtx);
4903 SUBST (XEXP (x, 2), true_rtx);
4905 temp = true_rtx, true_rtx = false_rtx, false_rtx = temp;
4906 cond = XEXP (x, 0);
4908 /* It is possible that the conditional has been simplified out. */
4909 true_code = GET_CODE (cond);
4910 comparison_p = COMPARISON_P (cond);
4913 /* If the two arms are identical, we don't need the comparison. */
4915 if (rtx_equal_p (true_rtx, false_rtx) && ! side_effects_p (cond))
4916 return true_rtx;
4918 /* Convert a == b ? b : a to "a". */
4919 if (true_code == EQ && ! side_effects_p (cond)
4920 && !HONOR_NANS (mode)
4921 && rtx_equal_p (XEXP (cond, 0), false_rtx)
4922 && rtx_equal_p (XEXP (cond, 1), true_rtx))
4923 return false_rtx;
4924 else if (true_code == NE && ! side_effects_p (cond)
4925 && !HONOR_NANS (mode)
4926 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4927 && rtx_equal_p (XEXP (cond, 1), false_rtx))
4928 return true_rtx;
4930 /* Look for cases where we have (abs x) or (neg (abs X)). */
4932 if (GET_MODE_CLASS (mode) == MODE_INT
4933 && GET_CODE (false_rtx) == NEG
4934 && rtx_equal_p (true_rtx, XEXP (false_rtx, 0))
4935 && comparison_p
4936 && rtx_equal_p (true_rtx, XEXP (cond, 0))
4937 && ! side_effects_p (true_rtx))
4938 switch (true_code)
4940 case GT:
4941 case GE:
4942 return simplify_gen_unary (ABS, mode, true_rtx, mode);
4943 case LT:
4944 case LE:
4945 return
4946 simplify_gen_unary (NEG, mode,
4947 simplify_gen_unary (ABS, mode, true_rtx, mode),
4948 mode);
4949 default:
4950 break;
4953 /* Look for MIN or MAX. */
4955 if ((! FLOAT_MODE_P (mode) || flag_unsafe_math_optimizations)
4956 && comparison_p
4957 && rtx_equal_p (XEXP (cond, 0), true_rtx)
4958 && rtx_equal_p (XEXP (cond, 1), false_rtx)
4959 && ! side_effects_p (cond))
4960 switch (true_code)
4962 case GE:
4963 case GT:
4964 return simplify_gen_binary (SMAX, mode, true_rtx, false_rtx);
4965 case LE:
4966 case LT:
4967 return simplify_gen_binary (SMIN, mode, true_rtx, false_rtx);
4968 case GEU:
4969 case GTU:
4970 return simplify_gen_binary (UMAX, mode, true_rtx, false_rtx);
4971 case LEU:
4972 case LTU:
4973 return simplify_gen_binary (UMIN, mode, true_rtx, false_rtx);
4974 default:
4975 break;
4978 /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its
4979 second operand is zero, this can be done as (OP Z (mult COND C2)) where
4980 C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or
4981 SIGN_EXTEND as long as Z is already extended (so we don't destroy it).
4982 We can do this kind of thing in some cases when STORE_FLAG_VALUE is
4983 neither 1 or -1, but it isn't worth checking for. */
4985 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
4986 && comparison_p
4987 && GET_MODE_CLASS (mode) == MODE_INT
4988 && ! side_effects_p (x))
4990 rtx t = make_compound_operation (true_rtx, SET);
4991 rtx f = make_compound_operation (false_rtx, SET);
4992 rtx cond_op0 = XEXP (cond, 0);
4993 rtx cond_op1 = XEXP (cond, 1);
4994 enum rtx_code op = UNKNOWN, extend_op = UNKNOWN;
4995 enum machine_mode m = mode;
4996 rtx z = 0, c1 = NULL_RTX;
4998 if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS
4999 || GET_CODE (t) == IOR || GET_CODE (t) == XOR
5000 || GET_CODE (t) == ASHIFT
5001 || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT)
5002 && rtx_equal_p (XEXP (t, 0), f))
5003 c1 = XEXP (t, 1), op = GET_CODE (t), z = f;
5005 /* If an identity-zero op is commutative, check whether there
5006 would be a match if we swapped the operands. */
5007 else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR
5008 || GET_CODE (t) == XOR)
5009 && rtx_equal_p (XEXP (t, 1), f))
5010 c1 = XEXP (t, 0), op = GET_CODE (t), z = f;
5011 else if (GET_CODE (t) == SIGN_EXTEND
5012 && (GET_CODE (XEXP (t, 0)) == PLUS
5013 || GET_CODE (XEXP (t, 0)) == MINUS
5014 || GET_CODE (XEXP (t, 0)) == IOR
5015 || GET_CODE (XEXP (t, 0)) == XOR
5016 || GET_CODE (XEXP (t, 0)) == ASHIFT
5017 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5018 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5019 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5020 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5021 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5022 && (num_sign_bit_copies (f, GET_MODE (f))
5023 > (unsigned int)
5024 (GET_MODE_BITSIZE (mode)
5025 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 0))))))
5027 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5028 extend_op = SIGN_EXTEND;
5029 m = GET_MODE (XEXP (t, 0));
5031 else if (GET_CODE (t) == SIGN_EXTEND
5032 && (GET_CODE (XEXP (t, 0)) == PLUS
5033 || GET_CODE (XEXP (t, 0)) == IOR
5034 || GET_CODE (XEXP (t, 0)) == XOR)
5035 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5036 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5037 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5038 && (num_sign_bit_copies (f, GET_MODE (f))
5039 > (unsigned int)
5040 (GET_MODE_BITSIZE (mode)
5041 - GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (t, 0), 1))))))
5043 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5044 extend_op = SIGN_EXTEND;
5045 m = GET_MODE (XEXP (t, 0));
5047 else if (GET_CODE (t) == ZERO_EXTEND
5048 && (GET_CODE (XEXP (t, 0)) == PLUS
5049 || GET_CODE (XEXP (t, 0)) == MINUS
5050 || GET_CODE (XEXP (t, 0)) == IOR
5051 || GET_CODE (XEXP (t, 0)) == XOR
5052 || GET_CODE (XEXP (t, 0)) == ASHIFT
5053 || GET_CODE (XEXP (t, 0)) == LSHIFTRT
5054 || GET_CODE (XEXP (t, 0)) == ASHIFTRT)
5055 && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG
5056 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5057 && subreg_lowpart_p (XEXP (XEXP (t, 0), 0))
5058 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f)
5059 && ((nonzero_bits (f, GET_MODE (f))
5060 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 0))))
5061 == 0))
5063 c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0));
5064 extend_op = ZERO_EXTEND;
5065 m = GET_MODE (XEXP (t, 0));
5067 else if (GET_CODE (t) == ZERO_EXTEND
5068 && (GET_CODE (XEXP (t, 0)) == PLUS
5069 || GET_CODE (XEXP (t, 0)) == IOR
5070 || GET_CODE (XEXP (t, 0)) == XOR)
5071 && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG
5072 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5073 && subreg_lowpart_p (XEXP (XEXP (t, 0), 1))
5074 && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f)
5075 && ((nonzero_bits (f, GET_MODE (f))
5076 & ~GET_MODE_MASK (GET_MODE (XEXP (XEXP (t, 0), 1))))
5077 == 0))
5079 c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0));
5080 extend_op = ZERO_EXTEND;
5081 m = GET_MODE (XEXP (t, 0));
5084 if (z)
5086 temp = subst (simplify_gen_relational (true_code, m, VOIDmode,
5087 cond_op0, cond_op1),
5088 pc_rtx, pc_rtx, 0, 0);
5089 temp = simplify_gen_binary (MULT, m, temp,
5090 simplify_gen_binary (MULT, m, c1,
5091 const_true_rtx));
5092 temp = subst (temp, pc_rtx, pc_rtx, 0, 0);
5093 temp = simplify_gen_binary (op, m, gen_lowpart (m, z), temp);
5095 if (extend_op != UNKNOWN)
5096 temp = simplify_gen_unary (extend_op, mode, temp, m);
5098 return temp;
5102 /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or
5103 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the
5104 negation of a single bit, we can convert this operation to a shift. We
5105 can actually do this more generally, but it doesn't seem worth it. */
5107 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5108 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5109 && ((1 == nonzero_bits (XEXP (cond, 0), mode)
5110 && (i = exact_log2 (INTVAL (true_rtx))) >= 0)
5111 || ((num_sign_bit_copies (XEXP (cond, 0), mode)
5112 == GET_MODE_BITSIZE (mode))
5113 && (i = exact_log2 (-INTVAL (true_rtx))) >= 0)))
5114 return
5115 simplify_shift_const (NULL_RTX, ASHIFT, mode,
5116 gen_lowpart (mode, XEXP (cond, 0)), i);
5118 /* (IF_THEN_ELSE (NE REG 0) (0) (8)) is REG for nonzero_bits (REG) == 8. */
5119 if (true_code == NE && XEXP (cond, 1) == const0_rtx
5120 && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT
5121 && GET_MODE (XEXP (cond, 0)) == mode
5122 && (INTVAL (true_rtx) & GET_MODE_MASK (mode))
5123 == nonzero_bits (XEXP (cond, 0), mode)
5124 && (i = exact_log2 (INTVAL (true_rtx) & GET_MODE_MASK (mode))) >= 0)
5125 return XEXP (cond, 0);
5127 return x;
5130 /* Simplify X, a SET expression. Return the new expression. */
5132 static rtx
5133 simplify_set (rtx x)
5135 rtx src = SET_SRC (x);
5136 rtx dest = SET_DEST (x);
5137 enum machine_mode mode
5138 = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest);
5139 rtx other_insn;
5140 rtx *cc_use;
5142 /* (set (pc) (return)) gets written as (return). */
5143 if (GET_CODE (dest) == PC && GET_CODE (src) == RETURN)
5144 return src;
5146 /* Now that we know for sure which bits of SRC we are using, see if we can
5147 simplify the expression for the object knowing that we only need the
5148 low-order bits. */
5150 if (GET_MODE_CLASS (mode) == MODE_INT
5151 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
5153 src = force_to_mode (src, mode, ~(HOST_WIDE_INT) 0, 0);
5154 SUBST (SET_SRC (x), src);
5157 /* If we are setting CC0 or if the source is a COMPARE, look for the use of
5158 the comparison result and try to simplify it unless we already have used
5159 undobuf.other_insn. */
5160 if ((GET_MODE_CLASS (mode) == MODE_CC
5161 || GET_CODE (src) == COMPARE
5162 || CC0_P (dest))
5163 && (cc_use = find_single_use (dest, subst_insn, &other_insn)) != 0
5164 && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn)
5165 && COMPARISON_P (*cc_use)
5166 && rtx_equal_p (XEXP (*cc_use, 0), dest))
5168 enum rtx_code old_code = GET_CODE (*cc_use);
5169 enum rtx_code new_code;
5170 rtx op0, op1, tmp;
5171 int other_changed = 0;
5172 enum machine_mode compare_mode = GET_MODE (dest);
5174 if (GET_CODE (src) == COMPARE)
5175 op0 = XEXP (src, 0), op1 = XEXP (src, 1);
5176 else
5177 op0 = src, op1 = CONST0_RTX (GET_MODE (src));
5179 tmp = simplify_relational_operation (old_code, compare_mode, VOIDmode,
5180 op0, op1);
5181 if (!tmp)
5182 new_code = old_code;
5183 else if (!CONSTANT_P (tmp))
5185 new_code = GET_CODE (tmp);
5186 op0 = XEXP (tmp, 0);
5187 op1 = XEXP (tmp, 1);
5189 else
5191 rtx pat = PATTERN (other_insn);
5192 undobuf.other_insn = other_insn;
5193 SUBST (*cc_use, tmp);
5195 /* Attempt to simplify CC user. */
5196 if (GET_CODE (pat) == SET)
5198 rtx new = simplify_rtx (SET_SRC (pat));
5199 if (new != NULL_RTX)
5200 SUBST (SET_SRC (pat), new);
5203 /* Convert X into a no-op move. */
5204 SUBST (SET_DEST (x), pc_rtx);
5205 SUBST (SET_SRC (x), pc_rtx);
5206 return x;
5209 /* Simplify our comparison, if possible. */
5210 new_code = simplify_comparison (new_code, &op0, &op1);
5212 #ifdef SELECT_CC_MODE
5213 /* If this machine has CC modes other than CCmode, check to see if we
5214 need to use a different CC mode here. */
5215 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
5216 compare_mode = GET_MODE (op0);
5217 else
5218 compare_mode = SELECT_CC_MODE (new_code, op0, op1);
5220 #ifndef HAVE_cc0
5221 /* If the mode changed, we have to change SET_DEST, the mode in the
5222 compare, and the mode in the place SET_DEST is used. If SET_DEST is
5223 a hard register, just build new versions with the proper mode. If it
5224 is a pseudo, we lose unless it is only time we set the pseudo, in
5225 which case we can safely change its mode. */
5226 if (compare_mode != GET_MODE (dest))
5228 if (can_change_dest_mode (dest, 0, compare_mode))
5230 unsigned int regno = REGNO (dest);
5231 rtx new_dest;
5233 if (regno < FIRST_PSEUDO_REGISTER)
5234 new_dest = gen_rtx_REG (compare_mode, regno);
5235 else
5237 SUBST_MODE (regno_reg_rtx[regno], compare_mode);
5238 new_dest = regno_reg_rtx[regno];
5241 SUBST (SET_DEST (x), new_dest);
5242 SUBST (XEXP (*cc_use, 0), new_dest);
5243 other_changed = 1;
5245 dest = new_dest;
5248 #endif /* cc0 */
5249 #endif /* SELECT_CC_MODE */
5251 /* If the code changed, we have to build a new comparison in
5252 undobuf.other_insn. */
5253 if (new_code != old_code)
5255 int other_changed_previously = other_changed;
5256 unsigned HOST_WIDE_INT mask;
5258 SUBST (*cc_use, gen_rtx_fmt_ee (new_code, GET_MODE (*cc_use),
5259 dest, const0_rtx));
5260 other_changed = 1;
5262 /* If the only change we made was to change an EQ into an NE or
5263 vice versa, OP0 has only one bit that might be nonzero, and OP1
5264 is zero, check if changing the user of the condition code will
5265 produce a valid insn. If it won't, we can keep the original code
5266 in that insn by surrounding our operation with an XOR. */
5268 if (((old_code == NE && new_code == EQ)
5269 || (old_code == EQ && new_code == NE))
5270 && ! other_changed_previously && op1 == const0_rtx
5271 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
5272 && exact_log2 (mask = nonzero_bits (op0, GET_MODE (op0))) >= 0)
5274 rtx pat = PATTERN (other_insn), note = 0;
5276 if ((recog_for_combine (&pat, other_insn, &note) < 0
5277 && ! check_asm_operands (pat)))
5279 PUT_CODE (*cc_use, old_code);
5280 other_changed = 0;
5282 op0 = simplify_gen_binary (XOR, GET_MODE (op0),
5283 op0, GEN_INT (mask));
5288 if (other_changed)
5289 undobuf.other_insn = other_insn;
5291 #ifdef HAVE_cc0
5292 /* If we are now comparing against zero, change our source if
5293 needed. If we do not use cc0, we always have a COMPARE. */
5294 if (op1 == const0_rtx && dest == cc0_rtx)
5296 SUBST (SET_SRC (x), op0);
5297 src = op0;
5299 else
5300 #endif
5302 /* Otherwise, if we didn't previously have a COMPARE in the
5303 correct mode, we need one. */
5304 if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode)
5306 SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1));
5307 src = SET_SRC (x);
5309 else if (GET_MODE (op0) == compare_mode && op1 == const0_rtx)
5311 SUBST(SET_SRC (x), op0);
5312 src = SET_SRC (x);
5314 else
5316 /* Otherwise, update the COMPARE if needed. */
5317 SUBST (XEXP (src, 0), op0);
5318 SUBST (XEXP (src, 1), op1);
5321 else
5323 /* Get SET_SRC in a form where we have placed back any
5324 compound expressions. Then do the checks below. */
5325 src = make_compound_operation (src, SET);
5326 SUBST (SET_SRC (x), src);
5329 /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation,
5330 and X being a REG or (subreg (reg)), we may be able to convert this to
5331 (set (subreg:m2 x) (op)).
5333 We can always do this if M1 is narrower than M2 because that means that
5334 we only care about the low bits of the result.
5336 However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot
5337 perform a narrower operation than requested since the high-order bits will
5338 be undefined. On machine where it is defined, this transformation is safe
5339 as long as M1 and M2 have the same number of words. */
5341 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5342 && !OBJECT_P (SUBREG_REG (src))
5343 && (((GET_MODE_SIZE (GET_MODE (src)) + (UNITS_PER_WORD - 1))
5344 / UNITS_PER_WORD)
5345 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))
5346 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD))
5347 #ifndef WORD_REGISTER_OPERATIONS
5348 && (GET_MODE_SIZE (GET_MODE (src))
5349 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5350 #endif
5351 #ifdef CANNOT_CHANGE_MODE_CLASS
5352 && ! (REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER
5353 && REG_CANNOT_CHANGE_MODE_P (REGNO (dest),
5354 GET_MODE (SUBREG_REG (src)),
5355 GET_MODE (src)))
5356 #endif
5357 && (REG_P (dest)
5358 || (GET_CODE (dest) == SUBREG
5359 && REG_P (SUBREG_REG (dest)))))
5361 SUBST (SET_DEST (x),
5362 gen_lowpart (GET_MODE (SUBREG_REG (src)),
5363 dest));
5364 SUBST (SET_SRC (x), SUBREG_REG (src));
5366 src = SET_SRC (x), dest = SET_DEST (x);
5369 #ifdef HAVE_cc0
5370 /* If we have (set (cc0) (subreg ...)), we try to remove the subreg
5371 in SRC. */
5372 if (dest == cc0_rtx
5373 && GET_CODE (src) == SUBREG
5374 && subreg_lowpart_p (src)
5375 && (GET_MODE_BITSIZE (GET_MODE (src))
5376 < GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (src)))))
5378 rtx inner = SUBREG_REG (src);
5379 enum machine_mode inner_mode = GET_MODE (inner);
5381 /* Here we make sure that we don't have a sign bit on. */
5382 if (GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_WIDE_INT
5383 && (nonzero_bits (inner, inner_mode)
5384 < ((unsigned HOST_WIDE_INT) 1
5385 << (GET_MODE_BITSIZE (GET_MODE (src)) - 1))))
5387 SUBST (SET_SRC (x), inner);
5388 src = SET_SRC (x);
5391 #endif
5393 #ifdef LOAD_EXTEND_OP
5394 /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this
5395 would require a paradoxical subreg. Replace the subreg with a
5396 zero_extend to avoid the reload that would otherwise be required. */
5398 if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src)
5399 && LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))) != UNKNOWN
5400 && SUBREG_BYTE (src) == 0
5401 && (GET_MODE_SIZE (GET_MODE (src))
5402 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))
5403 && MEM_P (SUBREG_REG (src)))
5405 SUBST (SET_SRC (x),
5406 gen_rtx_fmt_e (LOAD_EXTEND_OP (GET_MODE (SUBREG_REG (src))),
5407 GET_MODE (src), SUBREG_REG (src)));
5409 src = SET_SRC (x);
5411 #endif
5413 /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we
5414 are comparing an item known to be 0 or -1 against 0, use a logical
5415 operation instead. Check for one of the arms being an IOR of the other
5416 arm with some value. We compute three terms to be IOR'ed together. In
5417 practice, at most two will be nonzero. Then we do the IOR's. */
5419 if (GET_CODE (dest) != PC
5420 && GET_CODE (src) == IF_THEN_ELSE
5421 && GET_MODE_CLASS (GET_MODE (src)) == MODE_INT
5422 && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE)
5423 && XEXP (XEXP (src, 0), 1) == const0_rtx
5424 && GET_MODE (src) == GET_MODE (XEXP (XEXP (src, 0), 0))
5425 #ifdef HAVE_conditional_move
5426 && ! can_conditionally_move_p (GET_MODE (src))
5427 #endif
5428 && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0),
5429 GET_MODE (XEXP (XEXP (src, 0), 0)))
5430 == GET_MODE_BITSIZE (GET_MODE (XEXP (XEXP (src, 0), 0))))
5431 && ! side_effects_p (src))
5433 rtx true_rtx = (GET_CODE (XEXP (src, 0)) == NE
5434 ? XEXP (src, 1) : XEXP (src, 2));
5435 rtx false_rtx = (GET_CODE (XEXP (src, 0)) == NE
5436 ? XEXP (src, 2) : XEXP (src, 1));
5437 rtx term1 = const0_rtx, term2, term3;
5439 if (GET_CODE (true_rtx) == IOR
5440 && rtx_equal_p (XEXP (true_rtx, 0), false_rtx))
5441 term1 = false_rtx, true_rtx = XEXP (true_rtx, 1), false_rtx = const0_rtx;
5442 else if (GET_CODE (true_rtx) == IOR
5443 && rtx_equal_p (XEXP (true_rtx, 1), false_rtx))
5444 term1 = false_rtx, true_rtx = XEXP (true_rtx, 0), false_rtx = const0_rtx;
5445 else if (GET_CODE (false_rtx) == IOR
5446 && rtx_equal_p (XEXP (false_rtx, 0), true_rtx))
5447 term1 = true_rtx, false_rtx = XEXP (false_rtx, 1), true_rtx = const0_rtx;
5448 else if (GET_CODE (false_rtx) == IOR
5449 && rtx_equal_p (XEXP (false_rtx, 1), true_rtx))
5450 term1 = true_rtx, false_rtx = XEXP (false_rtx, 0), true_rtx = const0_rtx;
5452 term2 = simplify_gen_binary (AND, GET_MODE (src),
5453 XEXP (XEXP (src, 0), 0), true_rtx);
5454 term3 = simplify_gen_binary (AND, GET_MODE (src),
5455 simplify_gen_unary (NOT, GET_MODE (src),
5456 XEXP (XEXP (src, 0), 0),
5457 GET_MODE (src)),
5458 false_rtx);
5460 SUBST (SET_SRC (x),
5461 simplify_gen_binary (IOR, GET_MODE (src),
5462 simplify_gen_binary (IOR, GET_MODE (src),
5463 term1, term2),
5464 term3));
5466 src = SET_SRC (x);
5469 /* If either SRC or DEST is a CLOBBER of (const_int 0), make this
5470 whole thing fail. */
5471 if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx)
5472 return src;
5473 else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx)
5474 return dest;
5475 else
5476 /* Convert this into a field assignment operation, if possible. */
5477 return make_field_assignment (x);
5480 /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified
5481 result. */
5483 static rtx
5484 simplify_logical (rtx x)
5486 enum machine_mode mode = GET_MODE (x);
5487 rtx op0 = XEXP (x, 0);
5488 rtx op1 = XEXP (x, 1);
5490 switch (GET_CODE (x))
5492 case AND:
5493 /* We can call simplify_and_const_int only if we don't lose
5494 any (sign) bits when converting INTVAL (op1) to
5495 "unsigned HOST_WIDE_INT". */
5496 if (GET_CODE (op1) == CONST_INT
5497 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
5498 || INTVAL (op1) > 0))
5500 x = simplify_and_const_int (x, mode, op0, INTVAL (op1));
5501 if (GET_CODE (x) != AND)
5502 return x;
5504 op0 = XEXP (x, 0);
5505 op1 = XEXP (x, 1);
5508 /* If we have any of (and (ior A B) C) or (and (xor A B) C),
5509 apply the distributive law and then the inverse distributive
5510 law to see if things simplify. */
5511 if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR)
5513 rtx result = distribute_and_simplify_rtx (x, 0);
5514 if (result)
5515 return result;
5517 if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR)
5519 rtx result = distribute_and_simplify_rtx (x, 1);
5520 if (result)
5521 return result;
5523 break;
5525 case IOR:
5526 /* If we have (ior (and A B) C), apply the distributive law and then
5527 the inverse distributive law to see if things simplify. */
5529 if (GET_CODE (op0) == AND)
5531 rtx result = distribute_and_simplify_rtx (x, 0);
5532 if (result)
5533 return result;
5536 if (GET_CODE (op1) == AND)
5538 rtx result = distribute_and_simplify_rtx (x, 1);
5539 if (result)
5540 return result;
5542 break;
5544 default:
5545 gcc_unreachable ();
5548 return x;
5551 /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound
5552 operations" because they can be replaced with two more basic operations.
5553 ZERO_EXTEND is also considered "compound" because it can be replaced with
5554 an AND operation, which is simpler, though only one operation.
5556 The function expand_compound_operation is called with an rtx expression
5557 and will convert it to the appropriate shifts and AND operations,
5558 simplifying at each stage.
5560 The function make_compound_operation is called to convert an expression
5561 consisting of shifts and ANDs into the equivalent compound expression.
5562 It is the inverse of this function, loosely speaking. */
5564 static rtx
5565 expand_compound_operation (rtx x)
5567 unsigned HOST_WIDE_INT pos = 0, len;
5568 int unsignedp = 0;
5569 unsigned int modewidth;
5570 rtx tem;
5572 switch (GET_CODE (x))
5574 case ZERO_EXTEND:
5575 unsignedp = 1;
5576 case SIGN_EXTEND:
5577 /* We can't necessarily use a const_int for a multiword mode;
5578 it depends on implicitly extending the value.
5579 Since we don't know the right way to extend it,
5580 we can't tell whether the implicit way is right.
5582 Even for a mode that is no wider than a const_int,
5583 we can't win, because we need to sign extend one of its bits through
5584 the rest of it, and we don't know which bit. */
5585 if (GET_CODE (XEXP (x, 0)) == CONST_INT)
5586 return x;
5588 /* Return if (subreg:MODE FROM 0) is not a safe replacement for
5589 (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM
5590 because (SUBREG (MEM...)) is guaranteed to cause the MEM to be
5591 reloaded. If not for that, MEM's would very rarely be safe.
5593 Reject MODEs bigger than a word, because we might not be able
5594 to reference a two-register group starting with an arbitrary register
5595 (and currently gen_lowpart might crash for a SUBREG). */
5597 if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) > UNITS_PER_WORD)
5598 return x;
5600 /* Reject MODEs that aren't scalar integers because turning vector
5601 or complex modes into shifts causes problems. */
5603 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5604 return x;
5606 len = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)));
5607 /* If the inner object has VOIDmode (the only way this can happen
5608 is if it is an ASM_OPERANDS), we can't do anything since we don't
5609 know how much masking to do. */
5610 if (len == 0)
5611 return x;
5613 break;
5615 case ZERO_EXTRACT:
5616 unsignedp = 1;
5618 /* ... fall through ... */
5620 case SIGN_EXTRACT:
5621 /* If the operand is a CLOBBER, just return it. */
5622 if (GET_CODE (XEXP (x, 0)) == CLOBBER)
5623 return XEXP (x, 0);
5625 if (GET_CODE (XEXP (x, 1)) != CONST_INT
5626 || GET_CODE (XEXP (x, 2)) != CONST_INT
5627 || GET_MODE (XEXP (x, 0)) == VOIDmode)
5628 return x;
5630 /* Reject MODEs that aren't scalar integers because turning vector
5631 or complex modes into shifts causes problems. */
5633 if (! SCALAR_INT_MODE_P (GET_MODE (XEXP (x, 0))))
5634 return x;
5636 len = INTVAL (XEXP (x, 1));
5637 pos = INTVAL (XEXP (x, 2));
5639 /* This should stay within the object being extracted, fail otherwise. */
5640 if (len + pos > GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))))
5641 return x;
5643 if (BITS_BIG_ENDIAN)
5644 pos = GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))) - len - pos;
5646 break;
5648 default:
5649 return x;
5651 /* Convert sign extension to zero extension, if we know that the high
5652 bit is not set, as this is easier to optimize. It will be converted
5653 back to cheaper alternative in make_extraction. */
5654 if (GET_CODE (x) == SIGN_EXTEND
5655 && (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5656 && ((nonzero_bits (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
5657 & ~(((unsigned HOST_WIDE_INT)
5658 GET_MODE_MASK (GET_MODE (XEXP (x, 0))))
5659 >> 1))
5660 == 0)))
5662 rtx temp = gen_rtx_ZERO_EXTEND (GET_MODE (x), XEXP (x, 0));
5663 rtx temp2 = expand_compound_operation (temp);
5665 /* Make sure this is a profitable operation. */
5666 if (rtx_cost (x, SET) > rtx_cost (temp2, SET))
5667 return temp2;
5668 else if (rtx_cost (x, SET) > rtx_cost (temp, SET))
5669 return temp;
5670 else
5671 return x;
5674 /* We can optimize some special cases of ZERO_EXTEND. */
5675 if (GET_CODE (x) == ZERO_EXTEND)
5677 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we
5678 know that the last value didn't have any inappropriate bits
5679 set. */
5680 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5681 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5682 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5683 && (nonzero_bits (XEXP (XEXP (x, 0), 0), GET_MODE (x))
5684 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5685 return XEXP (XEXP (x, 0), 0);
5687 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5688 if (GET_CODE (XEXP (x, 0)) == SUBREG
5689 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5690 && subreg_lowpart_p (XEXP (x, 0))
5691 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
5692 && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), GET_MODE (x))
5693 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5694 return SUBREG_REG (XEXP (x, 0));
5696 /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo
5697 is a comparison and STORE_FLAG_VALUE permits. This is like
5698 the first case, but it works even when GET_MODE (x) is larger
5699 than HOST_WIDE_INT. */
5700 if (GET_CODE (XEXP (x, 0)) == TRUNCATE
5701 && GET_MODE (XEXP (XEXP (x, 0), 0)) == GET_MODE (x)
5702 && COMPARISON_P (XEXP (XEXP (x, 0), 0))
5703 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5704 <= HOST_BITS_PER_WIDE_INT)
5705 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5706 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5707 return XEXP (XEXP (x, 0), 0);
5709 /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */
5710 if (GET_CODE (XEXP (x, 0)) == SUBREG
5711 && GET_MODE (SUBREG_REG (XEXP (x, 0))) == GET_MODE (x)
5712 && subreg_lowpart_p (XEXP (x, 0))
5713 && COMPARISON_P (SUBREG_REG (XEXP (x, 0)))
5714 && (GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))
5715 <= HOST_BITS_PER_WIDE_INT)
5716 && ((HOST_WIDE_INT) STORE_FLAG_VALUE
5717 & ~GET_MODE_MASK (GET_MODE (XEXP (x, 0)))) == 0)
5718 return SUBREG_REG (XEXP (x, 0));
5722 /* If we reach here, we want to return a pair of shifts. The inner
5723 shift is a left shift of BITSIZE - POS - LEN bits. The outer
5724 shift is a right shift of BITSIZE - LEN bits. It is arithmetic or
5725 logical depending on the value of UNSIGNEDP.
5727 If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be
5728 converted into an AND of a shift.
5730 We must check for the case where the left shift would have a negative
5731 count. This can happen in a case like (x >> 31) & 255 on machines
5732 that can't shift by a constant. On those machines, we would first
5733 combine the shift with the AND to produce a variable-position
5734 extraction. Then the constant of 31 would be substituted in to produce
5735 a such a position. */
5737 modewidth = GET_MODE_BITSIZE (GET_MODE (x));
5738 if (modewidth + len >= pos)
5739 tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT,
5740 GET_MODE (x),
5741 simplify_shift_const (NULL_RTX, ASHIFT,
5742 GET_MODE (x),
5743 XEXP (x, 0),
5744 modewidth - pos - len),
5745 modewidth - len);
5747 else if (unsignedp && len < HOST_BITS_PER_WIDE_INT)
5748 tem = simplify_and_const_int (NULL_RTX, GET_MODE (x),
5749 simplify_shift_const (NULL_RTX, LSHIFTRT,
5750 GET_MODE (x),
5751 XEXP (x, 0), pos),
5752 ((HOST_WIDE_INT) 1 << len) - 1);
5753 else
5754 /* Any other cases we can't handle. */
5755 return x;
5757 /* If we couldn't do this for some reason, return the original
5758 expression. */
5759 if (GET_CODE (tem) == CLOBBER)
5760 return x;
5762 return tem;
5765 /* X is a SET which contains an assignment of one object into
5766 a part of another (such as a bit-field assignment, STRICT_LOW_PART,
5767 or certain SUBREGS). If possible, convert it into a series of
5768 logical operations.
5770 We half-heartedly support variable positions, but do not at all
5771 support variable lengths. */
5773 static rtx
5774 expand_field_assignment (rtx x)
5776 rtx inner;
5777 rtx pos; /* Always counts from low bit. */
5778 int len;
5779 rtx mask, cleared, masked;
5780 enum machine_mode compute_mode;
5782 /* Loop until we find something we can't simplify. */
5783 while (1)
5785 if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART
5786 && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG)
5788 inner = SUBREG_REG (XEXP (SET_DEST (x), 0));
5789 len = GET_MODE_BITSIZE (GET_MODE (XEXP (SET_DEST (x), 0)));
5790 pos = GEN_INT (subreg_lsb (XEXP (SET_DEST (x), 0)));
5792 else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT
5793 && GET_CODE (XEXP (SET_DEST (x), 1)) == CONST_INT)
5795 inner = XEXP (SET_DEST (x), 0);
5796 len = INTVAL (XEXP (SET_DEST (x), 1));
5797 pos = XEXP (SET_DEST (x), 2);
5799 /* A constant position should stay within the width of INNER. */
5800 if (GET_CODE (pos) == CONST_INT
5801 && INTVAL (pos) + len > GET_MODE_BITSIZE (GET_MODE (inner)))
5802 break;
5804 if (BITS_BIG_ENDIAN)
5806 if (GET_CODE (pos) == CONST_INT)
5807 pos = GEN_INT (GET_MODE_BITSIZE (GET_MODE (inner)) - len
5808 - INTVAL (pos));
5809 else if (GET_CODE (pos) == MINUS
5810 && GET_CODE (XEXP (pos, 1)) == CONST_INT
5811 && (INTVAL (XEXP (pos, 1))
5812 == GET_MODE_BITSIZE (GET_MODE (inner)) - len))
5813 /* If position is ADJUST - X, new position is X. */
5814 pos = XEXP (pos, 0);
5815 else
5816 pos = simplify_gen_binary (MINUS, GET_MODE (pos),
5817 GEN_INT (GET_MODE_BITSIZE (
5818 GET_MODE (inner))
5819 - len),
5820 pos);
5824 /* A SUBREG between two modes that occupy the same numbers of words
5825 can be done by moving the SUBREG to the source. */
5826 else if (GET_CODE (SET_DEST (x)) == SUBREG
5827 /* We need SUBREGs to compute nonzero_bits properly. */
5828 && nonzero_sign_valid
5829 && (((GET_MODE_SIZE (GET_MODE (SET_DEST (x)))
5830 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
5831 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x))))
5832 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)))
5834 x = gen_rtx_SET (VOIDmode, SUBREG_REG (SET_DEST (x)),
5835 gen_lowpart
5836 (GET_MODE (SUBREG_REG (SET_DEST (x))),
5837 SET_SRC (x)));
5838 continue;
5840 else
5841 break;
5843 while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
5844 inner = SUBREG_REG (inner);
5846 compute_mode = GET_MODE (inner);
5848 /* Don't attempt bitwise arithmetic on non scalar integer modes. */
5849 if (! SCALAR_INT_MODE_P (compute_mode))
5851 enum machine_mode imode;
5853 /* Don't do anything for vector or complex integral types. */
5854 if (! FLOAT_MODE_P (compute_mode))
5855 break;
5857 /* Try to find an integral mode to pun with. */
5858 imode = mode_for_size (GET_MODE_BITSIZE (compute_mode), MODE_INT, 0);
5859 if (imode == BLKmode)
5860 break;
5862 compute_mode = imode;
5863 inner = gen_lowpart (imode, inner);
5866 /* Compute a mask of LEN bits, if we can do this on the host machine. */
5867 if (len >= HOST_BITS_PER_WIDE_INT)
5868 break;
5870 /* Now compute the equivalent expression. Make a copy of INNER
5871 for the SET_DEST in case it is a MEM into which we will substitute;
5872 we don't want shared RTL in that case. */
5873 mask = GEN_INT (((HOST_WIDE_INT) 1 << len) - 1);
5874 cleared = simplify_gen_binary (AND, compute_mode,
5875 simplify_gen_unary (NOT, compute_mode,
5876 simplify_gen_binary (ASHIFT,
5877 compute_mode,
5878 mask, pos),
5879 compute_mode),
5880 inner);
5881 masked = simplify_gen_binary (ASHIFT, compute_mode,
5882 simplify_gen_binary (
5883 AND, compute_mode,
5884 gen_lowpart (compute_mode, SET_SRC (x)),
5885 mask),
5886 pos);
5888 x = gen_rtx_SET (VOIDmode, copy_rtx (inner),
5889 simplify_gen_binary (IOR, compute_mode,
5890 cleared, masked));
5893 return x;
5896 /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero,
5897 it is an RTX that represents a variable starting position; otherwise,
5898 POS is the (constant) starting bit position (counted from the LSB).
5900 UNSIGNEDP is nonzero for an unsigned reference and zero for a
5901 signed reference.
5903 IN_DEST is nonzero if this is a reference in the destination of a
5904 SET. This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If nonzero,
5905 a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will
5906 be used.
5908 IN_COMPARE is nonzero if we are in a COMPARE. This means that a
5909 ZERO_EXTRACT should be built even for bits starting at bit 0.
5911 MODE is the desired mode of the result (if IN_DEST == 0).
5913 The result is an RTX for the extraction or NULL_RTX if the target
5914 can't handle it. */
5916 static rtx
5917 make_extraction (enum machine_mode mode, rtx inner, HOST_WIDE_INT pos,
5918 rtx pos_rtx, unsigned HOST_WIDE_INT len, int unsignedp,
5919 int in_dest, int in_compare)
5921 /* This mode describes the size of the storage area
5922 to fetch the overall value from. Within that, we
5923 ignore the POS lowest bits, etc. */
5924 enum machine_mode is_mode = GET_MODE (inner);
5925 enum machine_mode inner_mode;
5926 enum machine_mode wanted_inner_mode;
5927 enum machine_mode wanted_inner_reg_mode = word_mode;
5928 enum machine_mode pos_mode = word_mode;
5929 enum machine_mode extraction_mode = word_mode;
5930 enum machine_mode tmode = mode_for_size (len, MODE_INT, 1);
5931 rtx new = 0;
5932 rtx orig_pos_rtx = pos_rtx;
5933 HOST_WIDE_INT orig_pos;
5935 if (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner))
5937 /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...),
5938 consider just the QI as the memory to extract from.
5939 The subreg adds or removes high bits; its mode is
5940 irrelevant to the meaning of this extraction,
5941 since POS and LEN count from the lsb. */
5942 if (MEM_P (SUBREG_REG (inner)))
5943 is_mode = GET_MODE (SUBREG_REG (inner));
5944 inner = SUBREG_REG (inner);
5946 else if (GET_CODE (inner) == ASHIFT
5947 && GET_CODE (XEXP (inner, 1)) == CONST_INT
5948 && pos_rtx == 0 && pos == 0
5949 && len > (unsigned HOST_WIDE_INT) INTVAL (XEXP (inner, 1)))
5951 /* We're extracting the least significant bits of an rtx
5952 (ashift X (const_int C)), where LEN > C. Extract the
5953 least significant (LEN - C) bits of X, giving an rtx
5954 whose mode is MODE, then shift it left C times. */
5955 new = make_extraction (mode, XEXP (inner, 0),
5956 0, 0, len - INTVAL (XEXP (inner, 1)),
5957 unsignedp, in_dest, in_compare);
5958 if (new != 0)
5959 return gen_rtx_ASHIFT (mode, new, XEXP (inner, 1));
5962 inner_mode = GET_MODE (inner);
5964 if (pos_rtx && GET_CODE (pos_rtx) == CONST_INT)
5965 pos = INTVAL (pos_rtx), pos_rtx = 0;
5967 /* See if this can be done without an extraction. We never can if the
5968 width of the field is not the same as that of some integer mode. For
5969 registers, we can only avoid the extraction if the position is at the
5970 low-order bit and this is either not in the destination or we have the
5971 appropriate STRICT_LOW_PART operation available.
5973 For MEM, we can avoid an extract if the field starts on an appropriate
5974 boundary and we can change the mode of the memory reference. */
5976 if (tmode != BLKmode
5977 && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0
5978 && !MEM_P (inner)
5979 && (inner_mode == tmode
5980 || !REG_P (inner)
5981 || TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (tmode),
5982 GET_MODE_BITSIZE (inner_mode))
5983 || reg_truncated_to_mode (tmode, inner))
5984 && (! in_dest
5985 || (REG_P (inner)
5986 && have_insn_for (STRICT_LOW_PART, tmode))))
5987 || (MEM_P (inner) && pos_rtx == 0
5988 && (pos
5989 % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode)
5990 : BITS_PER_UNIT)) == 0
5991 /* We can't do this if we are widening INNER_MODE (it
5992 may not be aligned, for one thing). */
5993 && GET_MODE_BITSIZE (inner_mode) >= GET_MODE_BITSIZE (tmode)
5994 && (inner_mode == tmode
5995 || (! mode_dependent_address_p (XEXP (inner, 0))
5996 && ! MEM_VOLATILE_P (inner))))))
5998 /* If INNER is a MEM, make a new MEM that encompasses just the desired
5999 field. If the original and current mode are the same, we need not
6000 adjust the offset. Otherwise, we do if bytes big endian.
6002 If INNER is not a MEM, get a piece consisting of just the field
6003 of interest (in this case POS % BITS_PER_WORD must be 0). */
6005 if (MEM_P (inner))
6007 HOST_WIDE_INT offset;
6009 /* POS counts from lsb, but make OFFSET count in memory order. */
6010 if (BYTES_BIG_ENDIAN)
6011 offset = (GET_MODE_BITSIZE (is_mode) - len - pos) / BITS_PER_UNIT;
6012 else
6013 offset = pos / BITS_PER_UNIT;
6015 new = adjust_address_nv (inner, tmode, offset);
6017 else if (REG_P (inner))
6019 if (tmode != inner_mode)
6021 /* We can't call gen_lowpart in a DEST since we
6022 always want a SUBREG (see below) and it would sometimes
6023 return a new hard register. */
6024 if (pos || in_dest)
6026 HOST_WIDE_INT final_word = pos / BITS_PER_WORD;
6028 if (WORDS_BIG_ENDIAN
6029 && GET_MODE_SIZE (inner_mode) > UNITS_PER_WORD)
6030 final_word = ((GET_MODE_SIZE (inner_mode)
6031 - GET_MODE_SIZE (tmode))
6032 / UNITS_PER_WORD) - final_word;
6034 final_word *= UNITS_PER_WORD;
6035 if (BYTES_BIG_ENDIAN &&
6036 GET_MODE_SIZE (inner_mode) > GET_MODE_SIZE (tmode))
6037 final_word += (GET_MODE_SIZE (inner_mode)
6038 - GET_MODE_SIZE (tmode)) % UNITS_PER_WORD;
6040 /* Avoid creating invalid subregs, for example when
6041 simplifying (x>>32)&255. */
6042 if (!validate_subreg (tmode, inner_mode, inner, final_word))
6043 return NULL_RTX;
6045 new = gen_rtx_SUBREG (tmode, inner, final_word);
6047 else
6048 new = gen_lowpart (tmode, inner);
6050 else
6051 new = inner;
6053 else
6054 new = force_to_mode (inner, tmode,
6055 len >= HOST_BITS_PER_WIDE_INT
6056 ? ~(unsigned HOST_WIDE_INT) 0
6057 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
6060 /* If this extraction is going into the destination of a SET,
6061 make a STRICT_LOW_PART unless we made a MEM. */
6063 if (in_dest)
6064 return (MEM_P (new) ? new
6065 : (GET_CODE (new) != SUBREG
6066 ? gen_rtx_CLOBBER (tmode, const0_rtx)
6067 : gen_rtx_STRICT_LOW_PART (VOIDmode, new)));
6069 if (mode == tmode)
6070 return new;
6072 if (GET_CODE (new) == CONST_INT)
6073 return gen_int_mode (INTVAL (new), mode);
6075 /* If we know that no extraneous bits are set, and that the high
6076 bit is not set, convert the extraction to the cheaper of
6077 sign and zero extension, that are equivalent in these cases. */
6078 if (flag_expensive_optimizations
6079 && (GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
6080 && ((nonzero_bits (new, tmode)
6081 & ~(((unsigned HOST_WIDE_INT)
6082 GET_MODE_MASK (tmode))
6083 >> 1))
6084 == 0)))
6086 rtx temp = gen_rtx_ZERO_EXTEND (mode, new);
6087 rtx temp1 = gen_rtx_SIGN_EXTEND (mode, new);
6089 /* Prefer ZERO_EXTENSION, since it gives more information to
6090 backends. */
6091 if (rtx_cost (temp, SET) <= rtx_cost (temp1, SET))
6092 return temp;
6093 return temp1;
6096 /* Otherwise, sign- or zero-extend unless we already are in the
6097 proper mode. */
6099 return (gen_rtx_fmt_e (unsignedp ? ZERO_EXTEND : SIGN_EXTEND,
6100 mode, new));
6103 /* Unless this is a COMPARE or we have a funny memory reference,
6104 don't do anything with zero-extending field extracts starting at
6105 the low-order bit since they are simple AND operations. */
6106 if (pos_rtx == 0 && pos == 0 && ! in_dest
6107 && ! in_compare && unsignedp)
6108 return 0;
6110 /* Unless INNER is not MEM, reject this if we would be spanning bytes or
6111 if the position is not a constant and the length is not 1. In all
6112 other cases, we would only be going outside our object in cases when
6113 an original shift would have been undefined. */
6114 if (MEM_P (inner)
6115 && ((pos_rtx == 0 && pos + len > GET_MODE_BITSIZE (is_mode))
6116 || (pos_rtx != 0 && len != 1)))
6117 return 0;
6119 /* Get the mode to use should INNER not be a MEM, the mode for the position,
6120 and the mode for the result. */
6121 if (in_dest && mode_for_extraction (EP_insv, -1) != MAX_MACHINE_MODE)
6123 wanted_inner_reg_mode = mode_for_extraction (EP_insv, 0);
6124 pos_mode = mode_for_extraction (EP_insv, 2);
6125 extraction_mode = mode_for_extraction (EP_insv, 3);
6128 if (! in_dest && unsignedp
6129 && mode_for_extraction (EP_extzv, -1) != MAX_MACHINE_MODE)
6131 wanted_inner_reg_mode = mode_for_extraction (EP_extzv, 1);
6132 pos_mode = mode_for_extraction (EP_extzv, 3);
6133 extraction_mode = mode_for_extraction (EP_extzv, 0);
6136 if (! in_dest && ! unsignedp
6137 && mode_for_extraction (EP_extv, -1) != MAX_MACHINE_MODE)
6139 wanted_inner_reg_mode = mode_for_extraction (EP_extv, 1);
6140 pos_mode = mode_for_extraction (EP_extv, 3);
6141 extraction_mode = mode_for_extraction (EP_extv, 0);
6144 /* Never narrow an object, since that might not be safe. */
6146 if (mode != VOIDmode
6147 && GET_MODE_SIZE (extraction_mode) < GET_MODE_SIZE (mode))
6148 extraction_mode = mode;
6150 if (pos_rtx && GET_MODE (pos_rtx) != VOIDmode
6151 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6152 pos_mode = GET_MODE (pos_rtx);
6154 /* If this is not from memory, the desired mode is the preferred mode
6155 for an extraction pattern's first input operand, or word_mode if there
6156 is none. */
6157 if (!MEM_P (inner))
6158 wanted_inner_mode = wanted_inner_reg_mode;
6159 else
6161 /* Be careful not to go beyond the extracted object and maintain the
6162 natural alignment of the memory. */
6163 wanted_inner_mode = smallest_mode_for_size (len, MODE_INT);
6164 while (pos % GET_MODE_BITSIZE (wanted_inner_mode) + len
6165 > GET_MODE_BITSIZE (wanted_inner_mode))
6167 wanted_inner_mode = GET_MODE_WIDER_MODE (wanted_inner_mode);
6168 gcc_assert (wanted_inner_mode != VOIDmode);
6171 /* If we have to change the mode of memory and cannot, the desired mode
6172 is EXTRACTION_MODE. */
6173 if (inner_mode != wanted_inner_mode
6174 && (mode_dependent_address_p (XEXP (inner, 0))
6175 || MEM_VOLATILE_P (inner)
6176 || pos_rtx))
6177 wanted_inner_mode = extraction_mode;
6180 orig_pos = pos;
6182 if (BITS_BIG_ENDIAN)
6184 /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to
6185 BITS_BIG_ENDIAN style. If position is constant, compute new
6186 position. Otherwise, build subtraction.
6187 Note that POS is relative to the mode of the original argument.
6188 If it's a MEM we need to recompute POS relative to that.
6189 However, if we're extracting from (or inserting into) a register,
6190 we want to recompute POS relative to wanted_inner_mode. */
6191 int width = (MEM_P (inner)
6192 ? GET_MODE_BITSIZE (is_mode)
6193 : GET_MODE_BITSIZE (wanted_inner_mode));
6195 if (pos_rtx == 0)
6196 pos = width - len - pos;
6197 else
6198 pos_rtx
6199 = gen_rtx_MINUS (GET_MODE (pos_rtx), GEN_INT (width - len), pos_rtx);
6200 /* POS may be less than 0 now, but we check for that below.
6201 Note that it can only be less than 0 if !MEM_P (inner). */
6204 /* If INNER has a wider mode, and this is a constant extraction, try to
6205 make it smaller and adjust the byte to point to the byte containing
6206 the value. */
6207 if (wanted_inner_mode != VOIDmode
6208 && inner_mode != wanted_inner_mode
6209 && ! pos_rtx
6210 && GET_MODE_SIZE (wanted_inner_mode) < GET_MODE_SIZE (is_mode)
6211 && MEM_P (inner)
6212 && ! mode_dependent_address_p (XEXP (inner, 0))
6213 && ! MEM_VOLATILE_P (inner))
6215 int offset = 0;
6217 /* The computations below will be correct if the machine is big
6218 endian in both bits and bytes or little endian in bits and bytes.
6219 If it is mixed, we must adjust. */
6221 /* If bytes are big endian and we had a paradoxical SUBREG, we must
6222 adjust OFFSET to compensate. */
6223 if (BYTES_BIG_ENDIAN
6224 && GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (is_mode))
6225 offset -= GET_MODE_SIZE (is_mode) - GET_MODE_SIZE (inner_mode);
6227 /* We can now move to the desired byte. */
6228 offset += (pos / GET_MODE_BITSIZE (wanted_inner_mode))
6229 * GET_MODE_SIZE (wanted_inner_mode);
6230 pos %= GET_MODE_BITSIZE (wanted_inner_mode);
6232 if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN
6233 && is_mode != wanted_inner_mode)
6234 offset = (GET_MODE_SIZE (is_mode)
6235 - GET_MODE_SIZE (wanted_inner_mode) - offset);
6237 inner = adjust_address_nv (inner, wanted_inner_mode, offset);
6240 /* If INNER is not memory, we can always get it into the proper mode. If we
6241 are changing its mode, POS must be a constant and smaller than the size
6242 of the new mode. */
6243 else if (!MEM_P (inner))
6245 if (GET_MODE (inner) != wanted_inner_mode
6246 && (pos_rtx != 0
6247 || orig_pos + len > GET_MODE_BITSIZE (wanted_inner_mode)))
6248 return 0;
6250 if (orig_pos < 0)
6251 return 0;
6253 inner = force_to_mode (inner, wanted_inner_mode,
6254 pos_rtx
6255 || len + orig_pos >= HOST_BITS_PER_WIDE_INT
6256 ? ~(unsigned HOST_WIDE_INT) 0
6257 : ((((unsigned HOST_WIDE_INT) 1 << len) - 1)
6258 << orig_pos),
6262 /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we
6263 have to zero extend. Otherwise, we can just use a SUBREG. */
6264 if (pos_rtx != 0
6265 && GET_MODE_SIZE (pos_mode) > GET_MODE_SIZE (GET_MODE (pos_rtx)))
6267 rtx temp = gen_rtx_ZERO_EXTEND (pos_mode, pos_rtx);
6269 /* If we know that no extraneous bits are set, and that the high
6270 bit is not set, convert extraction to cheaper one - either
6271 SIGN_EXTENSION or ZERO_EXTENSION, that are equivalent in these
6272 cases. */
6273 if (flag_expensive_optimizations
6274 && (GET_MODE_BITSIZE (GET_MODE (pos_rtx)) <= HOST_BITS_PER_WIDE_INT
6275 && ((nonzero_bits (pos_rtx, GET_MODE (pos_rtx))
6276 & ~(((unsigned HOST_WIDE_INT)
6277 GET_MODE_MASK (GET_MODE (pos_rtx)))
6278 >> 1))
6279 == 0)))
6281 rtx temp1 = gen_rtx_SIGN_EXTEND (pos_mode, pos_rtx);
6283 /* Prefer ZERO_EXTENSION, since it gives more information to
6284 backends. */
6285 if (rtx_cost (temp1, SET) < rtx_cost (temp, SET))
6286 temp = temp1;
6288 pos_rtx = temp;
6290 else if (pos_rtx != 0
6291 && GET_MODE_SIZE (pos_mode) < GET_MODE_SIZE (GET_MODE (pos_rtx)))
6292 pos_rtx = gen_lowpart (pos_mode, pos_rtx);
6294 /* Make POS_RTX unless we already have it and it is correct. If we don't
6295 have a POS_RTX but we do have an ORIG_POS_RTX, the latter must
6296 be a CONST_INT. */
6297 if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos)
6298 pos_rtx = orig_pos_rtx;
6300 else if (pos_rtx == 0)
6301 pos_rtx = GEN_INT (pos);
6303 /* Make the required operation. See if we can use existing rtx. */
6304 new = gen_rtx_fmt_eee (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT,
6305 extraction_mode, inner, GEN_INT (len), pos_rtx);
6306 if (! in_dest)
6307 new = gen_lowpart (mode, new);
6309 return new;
6312 /* See if X contains an ASHIFT of COUNT or more bits that can be commuted
6313 with any other operations in X. Return X without that shift if so. */
6315 static rtx
6316 extract_left_shift (rtx x, int count)
6318 enum rtx_code code = GET_CODE (x);
6319 enum machine_mode mode = GET_MODE (x);
6320 rtx tem;
6322 switch (code)
6324 case ASHIFT:
6325 /* This is the shift itself. If it is wide enough, we will return
6326 either the value being shifted if the shift count is equal to
6327 COUNT or a shift for the difference. */
6328 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6329 && INTVAL (XEXP (x, 1)) >= count)
6330 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0),
6331 INTVAL (XEXP (x, 1)) - count);
6332 break;
6334 case NEG: case NOT:
6335 if ((tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6336 return simplify_gen_unary (code, mode, tem, mode);
6338 break;
6340 case PLUS: case IOR: case XOR: case AND:
6341 /* If we can safely shift this constant and we find the inner shift,
6342 make a new operation. */
6343 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6344 && (INTVAL (XEXP (x, 1)) & ((((HOST_WIDE_INT) 1 << count)) - 1)) == 0
6345 && (tem = extract_left_shift (XEXP (x, 0), count)) != 0)
6346 return simplify_gen_binary (code, mode, tem,
6347 GEN_INT (INTVAL (XEXP (x, 1)) >> count));
6349 break;
6351 default:
6352 break;
6355 return 0;
6358 /* Look at the expression rooted at X. Look for expressions
6359 equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND.
6360 Form these expressions.
6362 Return the new rtx, usually just X.
6364 Also, for machines like the VAX that don't have logical shift insns,
6365 try to convert logical to arithmetic shift operations in cases where
6366 they are equivalent. This undoes the canonicalizations to logical
6367 shifts done elsewhere.
6369 We try, as much as possible, to re-use rtl expressions to save memory.
6371 IN_CODE says what kind of expression we are processing. Normally, it is
6372 SET. In a memory address (inside a MEM, PLUS or minus, the latter two
6373 being kludges), it is MEM. When processing the arguments of a comparison
6374 or a COMPARE against zero, it is COMPARE. */
6376 static rtx
6377 make_compound_operation (rtx x, enum rtx_code in_code)
6379 enum rtx_code code = GET_CODE (x);
6380 enum machine_mode mode = GET_MODE (x);
6381 int mode_width = GET_MODE_BITSIZE (mode);
6382 rtx rhs, lhs;
6383 enum rtx_code next_code;
6384 int i;
6385 rtx new = 0;
6386 rtx tem;
6387 const char *fmt;
6389 /* Select the code to be used in recursive calls. Once we are inside an
6390 address, we stay there. If we have a comparison, set to COMPARE,
6391 but once inside, go back to our default of SET. */
6393 next_code = (code == MEM || code == PLUS || code == MINUS ? MEM
6394 : ((code == COMPARE || COMPARISON_P (x))
6395 && XEXP (x, 1) == const0_rtx) ? COMPARE
6396 : in_code == COMPARE ? SET : in_code);
6398 /* Process depending on the code of this operation. If NEW is set
6399 nonzero, it will be returned. */
6401 switch (code)
6403 case ASHIFT:
6404 /* Convert shifts by constants into multiplications if inside
6405 an address. */
6406 if (in_code == MEM && GET_CODE (XEXP (x, 1)) == CONST_INT
6407 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
6408 && INTVAL (XEXP (x, 1)) >= 0)
6410 new = make_compound_operation (XEXP (x, 0), next_code);
6411 new = gen_rtx_MULT (mode, new,
6412 GEN_INT ((HOST_WIDE_INT) 1
6413 << INTVAL (XEXP (x, 1))));
6415 break;
6417 case AND:
6418 /* If the second operand is not a constant, we can't do anything
6419 with it. */
6420 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6421 break;
6423 /* If the constant is a power of two minus one and the first operand
6424 is a logical right shift, make an extraction. */
6425 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6426 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6428 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6429 new = make_extraction (mode, new, 0, XEXP (XEXP (x, 0), 1), i, 1,
6430 0, in_code == COMPARE);
6433 /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */
6434 else if (GET_CODE (XEXP (x, 0)) == SUBREG
6435 && subreg_lowpart_p (XEXP (x, 0))
6436 && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT
6437 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6439 new = make_compound_operation (XEXP (SUBREG_REG (XEXP (x, 0)), 0),
6440 next_code);
6441 new = make_extraction (GET_MODE (SUBREG_REG (XEXP (x, 0))), new, 0,
6442 XEXP (SUBREG_REG (XEXP (x, 0)), 1), i, 1,
6443 0, in_code == COMPARE);
6445 /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */
6446 else if ((GET_CODE (XEXP (x, 0)) == XOR
6447 || GET_CODE (XEXP (x, 0)) == IOR)
6448 && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT
6449 && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT
6450 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6452 /* Apply the distributive law, and then try to make extractions. */
6453 new = gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)), mode,
6454 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0),
6455 XEXP (x, 1)),
6456 gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1),
6457 XEXP (x, 1)));
6458 new = make_compound_operation (new, in_code);
6461 /* If we are have (and (rotate X C) M) and C is larger than the number
6462 of bits in M, this is an extraction. */
6464 else if (GET_CODE (XEXP (x, 0)) == ROTATE
6465 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6466 && (i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0
6467 && i <= INTVAL (XEXP (XEXP (x, 0), 1)))
6469 new = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code);
6470 new = make_extraction (mode, new,
6471 (GET_MODE_BITSIZE (mode)
6472 - INTVAL (XEXP (XEXP (x, 0), 1))),
6473 NULL_RTX, i, 1, 0, in_code == COMPARE);
6476 /* On machines without logical shifts, if the operand of the AND is
6477 a logical shift and our mask turns off all the propagated sign
6478 bits, we can replace the logical shift with an arithmetic shift. */
6479 else if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
6480 && !have_insn_for (LSHIFTRT, mode)
6481 && have_insn_for (ASHIFTRT, mode)
6482 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
6483 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
6484 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
6485 && mode_width <= HOST_BITS_PER_WIDE_INT)
6487 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
6489 mask >>= INTVAL (XEXP (XEXP (x, 0), 1));
6490 if ((INTVAL (XEXP (x, 1)) & ~mask) == 0)
6491 SUBST (XEXP (x, 0),
6492 gen_rtx_ASHIFTRT (mode,
6493 make_compound_operation
6494 (XEXP (XEXP (x, 0), 0), next_code),
6495 XEXP (XEXP (x, 0), 1)));
6498 /* If the constant is one less than a power of two, this might be
6499 representable by an extraction even if no shift is present.
6500 If it doesn't end up being a ZERO_EXTEND, we will ignore it unless
6501 we are in a COMPARE. */
6502 else if ((i = exact_log2 (INTVAL (XEXP (x, 1)) + 1)) >= 0)
6503 new = make_extraction (mode,
6504 make_compound_operation (XEXP (x, 0),
6505 next_code),
6506 0, NULL_RTX, i, 1, 0, in_code == COMPARE);
6508 /* If we are in a comparison and this is an AND with a power of two,
6509 convert this into the appropriate bit extract. */
6510 else if (in_code == COMPARE
6511 && (i = exact_log2 (INTVAL (XEXP (x, 1)))) >= 0)
6512 new = make_extraction (mode,
6513 make_compound_operation (XEXP (x, 0),
6514 next_code),
6515 i, NULL_RTX, 1, 1, 0, 1);
6517 break;
6519 case LSHIFTRT:
6520 /* If the sign bit is known to be zero, replace this with an
6521 arithmetic shift. */
6522 if (have_insn_for (ASHIFTRT, mode)
6523 && ! have_insn_for (LSHIFTRT, mode)
6524 && mode_width <= HOST_BITS_PER_WIDE_INT
6525 && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0)
6527 new = gen_rtx_ASHIFTRT (mode,
6528 make_compound_operation (XEXP (x, 0),
6529 next_code),
6530 XEXP (x, 1));
6531 break;
6534 /* ... fall through ... */
6536 case ASHIFTRT:
6537 lhs = XEXP (x, 0);
6538 rhs = XEXP (x, 1);
6540 /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1,
6541 this is a SIGN_EXTRACT. */
6542 if (GET_CODE (rhs) == CONST_INT
6543 && GET_CODE (lhs) == ASHIFT
6544 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
6545 && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1)))
6547 new = make_compound_operation (XEXP (lhs, 0), next_code);
6548 new = make_extraction (mode, new,
6549 INTVAL (rhs) - INTVAL (XEXP (lhs, 1)),
6550 NULL_RTX, mode_width - INTVAL (rhs),
6551 code == LSHIFTRT, 0, in_code == COMPARE);
6552 break;
6555 /* See if we have operations between an ASHIFTRT and an ASHIFT.
6556 If so, try to merge the shifts into a SIGN_EXTEND. We could
6557 also do this for some cases of SIGN_EXTRACT, but it doesn't
6558 seem worth the effort; the case checked for occurs on Alpha. */
6560 if (!OBJECT_P (lhs)
6561 && ! (GET_CODE (lhs) == SUBREG
6562 && (OBJECT_P (SUBREG_REG (lhs))))
6563 && GET_CODE (rhs) == CONST_INT
6564 && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT
6565 && (new = extract_left_shift (lhs, INTVAL (rhs))) != 0)
6566 new = make_extraction (mode, make_compound_operation (new, next_code),
6567 0, NULL_RTX, mode_width - INTVAL (rhs),
6568 code == LSHIFTRT, 0, in_code == COMPARE);
6570 break;
6572 case SUBREG:
6573 /* Call ourselves recursively on the inner expression. If we are
6574 narrowing the object and it has a different RTL code from
6575 what it originally did, do this SUBREG as a force_to_mode. */
6577 tem = make_compound_operation (SUBREG_REG (x), in_code);
6580 rtx simplified;
6581 simplified = simplify_subreg (GET_MODE (x), tem, GET_MODE (tem),
6582 SUBREG_BYTE (x));
6584 if (simplified)
6585 tem = simplified;
6587 if (GET_CODE (tem) != GET_CODE (SUBREG_REG (x))
6588 && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (tem))
6589 && subreg_lowpart_p (x))
6591 rtx newer = force_to_mode (tem, mode, ~(HOST_WIDE_INT) 0,
6594 /* If we have something other than a SUBREG, we might have
6595 done an expansion, so rerun ourselves. */
6596 if (GET_CODE (newer) != SUBREG)
6597 newer = make_compound_operation (newer, in_code);
6599 return newer;
6602 if (simplified)
6603 return tem;
6605 break;
6607 default:
6608 break;
6611 if (new)
6613 x = gen_lowpart (mode, new);
6614 code = GET_CODE (x);
6617 /* Now recursively process each operand of this operation. */
6618 fmt = GET_RTX_FORMAT (code);
6619 for (i = 0; i < GET_RTX_LENGTH (code); i++)
6620 if (fmt[i] == 'e')
6622 new = make_compound_operation (XEXP (x, i), next_code);
6623 SUBST (XEXP (x, i), new);
6626 /* If this is a commutative operation, the changes to the operands
6627 may have made it noncanonical. */
6628 if (COMMUTATIVE_ARITH_P (x)
6629 && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)))
6631 tem = XEXP (x, 0);
6632 SUBST (XEXP (x, 0), XEXP (x, 1));
6633 SUBST (XEXP (x, 1), tem);
6636 return x;
6639 /* Given M see if it is a value that would select a field of bits
6640 within an item, but not the entire word. Return -1 if not.
6641 Otherwise, return the starting position of the field, where 0 is the
6642 low-order bit.
6644 *PLEN is set to the length of the field. */
6646 static int
6647 get_pos_from_mask (unsigned HOST_WIDE_INT m, unsigned HOST_WIDE_INT *plen)
6649 /* Get the bit number of the first 1 bit from the right, -1 if none. */
6650 int pos = exact_log2 (m & -m);
6651 int len = 0;
6653 if (pos >= 0)
6654 /* Now shift off the low-order zero bits and see if we have a
6655 power of two minus 1. */
6656 len = exact_log2 ((m >> pos) + 1);
6658 if (len <= 0)
6659 pos = -1;
6661 *plen = len;
6662 return pos;
6665 /* If X refers to a register that equals REG in value, replace these
6666 references with REG. */
6667 static rtx
6668 canon_reg_for_combine (rtx x, rtx reg)
6670 rtx op0, op1, op2;
6671 const char *fmt;
6672 int i;
6673 bool copied;
6675 enum rtx_code code = GET_CODE (x);
6676 switch (GET_RTX_CLASS (code))
6678 case RTX_UNARY:
6679 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6680 if (op0 != XEXP (x, 0))
6681 return simplify_gen_unary (GET_CODE (x), GET_MODE (x), op0,
6682 GET_MODE (reg));
6683 break;
6685 case RTX_BIN_ARITH:
6686 case RTX_COMM_ARITH:
6687 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6688 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6689 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
6690 return simplify_gen_binary (GET_CODE (x), GET_MODE (x), op0, op1);
6691 break;
6693 case RTX_COMPARE:
6694 case RTX_COMM_COMPARE:
6695 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6696 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6697 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
6698 return simplify_gen_relational (GET_CODE (x), GET_MODE (x),
6699 GET_MODE (op0), op0, op1);
6700 break;
6702 case RTX_TERNARY:
6703 case RTX_BITFIELD_OPS:
6704 op0 = canon_reg_for_combine (XEXP (x, 0), reg);
6705 op1 = canon_reg_for_combine (XEXP (x, 1), reg);
6706 op2 = canon_reg_for_combine (XEXP (x, 2), reg);
6707 if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1) || op2 != XEXP (x, 2))
6708 return simplify_gen_ternary (GET_CODE (x), GET_MODE (x),
6709 GET_MODE (op0), op0, op1, op2);
6711 case RTX_OBJ:
6712 if (REG_P (x))
6714 if (rtx_equal_p (get_last_value (reg), x)
6715 || rtx_equal_p (reg, get_last_value (x)))
6716 return reg;
6717 else
6718 break;
6721 /* fall through */
6723 default:
6724 fmt = GET_RTX_FORMAT (code);
6725 copied = false;
6726 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6727 if (fmt[i] == 'e')
6729 rtx op = canon_reg_for_combine (XEXP (x, i), reg);
6730 if (op != XEXP (x, i))
6732 if (!copied)
6734 copied = true;
6735 x = copy_rtx (x);
6737 XEXP (x, i) = op;
6740 else if (fmt[i] == 'E')
6742 int j;
6743 for (j = 0; j < XVECLEN (x, i); j++)
6745 rtx op = canon_reg_for_combine (XVECEXP (x, i, j), reg);
6746 if (op != XVECEXP (x, i, j))
6748 if (!copied)
6750 copied = true;
6751 x = copy_rtx (x);
6753 XVECEXP (x, i, j) = op;
6758 break;
6761 return x;
6764 /* Return X converted to MODE. If the value is already truncated to
6765 MODE we can just return a subreg even though in the general case we
6766 would need an explicit truncation. */
6768 static rtx
6769 gen_lowpart_or_truncate (enum machine_mode mode, rtx x)
6771 if (GET_MODE_SIZE (GET_MODE (x)) <= GET_MODE_SIZE (mode)
6772 || TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
6773 GET_MODE_BITSIZE (GET_MODE (x)))
6774 || (REG_P (x) && reg_truncated_to_mode (mode, x)))
6775 return gen_lowpart (mode, x);
6776 else
6777 return simplify_gen_unary (TRUNCATE, mode, x, GET_MODE (x));
6780 /* See if X can be simplified knowing that we will only refer to it in
6781 MODE and will only refer to those bits that are nonzero in MASK.
6782 If other bits are being computed or if masking operations are done
6783 that select a superset of the bits in MASK, they can sometimes be
6784 ignored.
6786 Return a possibly simplified expression, but always convert X to
6787 MODE. If X is a CONST_INT, AND the CONST_INT with MASK.
6789 If JUST_SELECT is nonzero, don't optimize by noticing that bits in MASK
6790 are all off in X. This is used when X will be complemented, by either
6791 NOT, NEG, or XOR. */
6793 static rtx
6794 force_to_mode (rtx x, enum machine_mode mode, unsigned HOST_WIDE_INT mask,
6795 int just_select)
6797 enum rtx_code code = GET_CODE (x);
6798 int next_select = just_select || code == XOR || code == NOT || code == NEG;
6799 enum machine_mode op_mode;
6800 unsigned HOST_WIDE_INT fuller_mask, nonzero;
6801 rtx op0, op1, temp;
6803 /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the
6804 code below will do the wrong thing since the mode of such an
6805 expression is VOIDmode.
6807 Also do nothing if X is a CLOBBER; this can happen if X was
6808 the return value from a call to gen_lowpart. */
6809 if (code == CALL || code == ASM_OPERANDS || code == CLOBBER)
6810 return x;
6812 /* We want to perform the operation is its present mode unless we know
6813 that the operation is valid in MODE, in which case we do the operation
6814 in MODE. */
6815 op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x))
6816 && have_insn_for (code, mode))
6817 ? mode : GET_MODE (x));
6819 /* It is not valid to do a right-shift in a narrower mode
6820 than the one it came in with. */
6821 if ((code == LSHIFTRT || code == ASHIFTRT)
6822 && GET_MODE_BITSIZE (mode) < GET_MODE_BITSIZE (GET_MODE (x)))
6823 op_mode = GET_MODE (x);
6825 /* Truncate MASK to fit OP_MODE. */
6826 if (op_mode)
6827 mask &= GET_MODE_MASK (op_mode);
6829 /* When we have an arithmetic operation, or a shift whose count we
6830 do not know, we need to assume that all bits up to the highest-order
6831 bit in MASK will be needed. This is how we form such a mask. */
6832 if (mask & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1)))
6833 fuller_mask = ~(unsigned HOST_WIDE_INT) 0;
6834 else
6835 fuller_mask = (((unsigned HOST_WIDE_INT) 1 << (floor_log2 (mask) + 1))
6836 - 1);
6838 /* Determine what bits of X are guaranteed to be (non)zero. */
6839 nonzero = nonzero_bits (x, mode);
6841 /* If none of the bits in X are needed, return a zero. */
6842 if (! just_select && (nonzero & mask) == 0)
6843 x = const0_rtx;
6845 /* If X is a CONST_INT, return a new one. Do this here since the
6846 test below will fail. */
6847 if (GET_CODE (x) == CONST_INT)
6849 if (SCALAR_INT_MODE_P (mode))
6850 return gen_int_mode (INTVAL (x) & mask, mode);
6851 else
6853 x = GEN_INT (INTVAL (x) & mask);
6854 return gen_lowpart_common (mode, x);
6858 /* If X is narrower than MODE and we want all the bits in X's mode, just
6859 get X in the proper mode. */
6860 if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (mode)
6861 && (GET_MODE_MASK (GET_MODE (x)) & ~mask) == 0)
6862 return gen_lowpart (mode, x);
6864 switch (code)
6866 case CLOBBER:
6867 /* If X is a (clobber (const_int)), return it since we know we are
6868 generating something that won't match. */
6869 return x;
6871 case SIGN_EXTEND:
6872 case ZERO_EXTEND:
6873 case ZERO_EXTRACT:
6874 case SIGN_EXTRACT:
6875 x = expand_compound_operation (x);
6876 if (GET_CODE (x) != code)
6877 return force_to_mode (x, mode, mask, next_select);
6878 break;
6880 case SUBREG:
6881 if (subreg_lowpart_p (x)
6882 /* We can ignore the effect of this SUBREG if it narrows the mode or
6883 if the constant masks to zero all the bits the mode doesn't
6884 have. */
6885 && ((GET_MODE_SIZE (GET_MODE (x))
6886 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
6887 || (0 == (mask
6888 & GET_MODE_MASK (GET_MODE (x))
6889 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))))))
6890 return force_to_mode (SUBREG_REG (x), mode, mask, next_select);
6891 break;
6893 case AND:
6894 /* If this is an AND with a constant, convert it into an AND
6895 whose constant is the AND of that constant with MASK. If it
6896 remains an AND of MASK, delete it since it is redundant. */
6898 if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6900 x = simplify_and_const_int (x, op_mode, XEXP (x, 0),
6901 mask & INTVAL (XEXP (x, 1)));
6903 /* If X is still an AND, see if it is an AND with a mask that
6904 is just some low-order bits. If so, and it is MASK, we don't
6905 need it. */
6907 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
6908 && ((INTVAL (XEXP (x, 1)) & GET_MODE_MASK (GET_MODE (x)))
6909 == mask))
6910 x = XEXP (x, 0);
6912 /* If it remains an AND, try making another AND with the bits
6913 in the mode mask that aren't in MASK turned on. If the
6914 constant in the AND is wide enough, this might make a
6915 cheaper constant. */
6917 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT
6918 && GET_MODE_MASK (GET_MODE (x)) != mask
6919 && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT)
6921 HOST_WIDE_INT cval = (INTVAL (XEXP (x, 1))
6922 | (GET_MODE_MASK (GET_MODE (x)) & ~mask));
6923 int width = GET_MODE_BITSIZE (GET_MODE (x));
6924 rtx y;
6926 /* If MODE is narrower than HOST_WIDE_INT and CVAL is a negative
6927 number, sign extend it. */
6928 if (width > 0 && width < HOST_BITS_PER_WIDE_INT
6929 && (cval & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
6930 cval |= (HOST_WIDE_INT) -1 << width;
6932 y = simplify_gen_binary (AND, GET_MODE (x),
6933 XEXP (x, 0), GEN_INT (cval));
6934 if (rtx_cost (y, SET) < rtx_cost (x, SET))
6935 x = y;
6938 break;
6941 goto binop;
6943 case PLUS:
6944 /* In (and (plus FOO C1) M), if M is a mask that just turns off
6945 low-order bits (as in an alignment operation) and FOO is already
6946 aligned to that boundary, mask C1 to that boundary as well.
6947 This may eliminate that PLUS and, later, the AND. */
6950 unsigned int width = GET_MODE_BITSIZE (mode);
6951 unsigned HOST_WIDE_INT smask = mask;
6953 /* If MODE is narrower than HOST_WIDE_INT and mask is a negative
6954 number, sign extend it. */
6956 if (width < HOST_BITS_PER_WIDE_INT
6957 && (smask & ((HOST_WIDE_INT) 1 << (width - 1))) != 0)
6958 smask |= (HOST_WIDE_INT) -1 << width;
6960 if (GET_CODE (XEXP (x, 1)) == CONST_INT
6961 && exact_log2 (- smask) >= 0
6962 && (nonzero_bits (XEXP (x, 0), mode) & ~smask) == 0
6963 && (INTVAL (XEXP (x, 1)) & ~smask) != 0)
6964 return force_to_mode (plus_constant (XEXP (x, 0),
6965 (INTVAL (XEXP (x, 1)) & smask)),
6966 mode, smask, next_select);
6969 /* ... fall through ... */
6971 case MULT:
6972 /* For PLUS, MINUS and MULT, we need any bits less significant than the
6973 most significant bit in MASK since carries from those bits will
6974 affect the bits we are interested in. */
6975 mask = fuller_mask;
6976 goto binop;
6978 case MINUS:
6979 /* If X is (minus C Y) where C's least set bit is larger than any bit
6980 in the mask, then we may replace with (neg Y). */
6981 if (GET_CODE (XEXP (x, 0)) == CONST_INT
6982 && (((unsigned HOST_WIDE_INT) (INTVAL (XEXP (x, 0))
6983 & -INTVAL (XEXP (x, 0))))
6984 > mask))
6986 x = simplify_gen_unary (NEG, GET_MODE (x), XEXP (x, 1),
6987 GET_MODE (x));
6988 return force_to_mode (x, mode, mask, next_select);
6991 /* Similarly, if C contains every bit in the fuller_mask, then we may
6992 replace with (not Y). */
6993 if (GET_CODE (XEXP (x, 0)) == CONST_INT
6994 && ((INTVAL (XEXP (x, 0)) | (HOST_WIDE_INT) fuller_mask)
6995 == INTVAL (XEXP (x, 0))))
6997 x = simplify_gen_unary (NOT, GET_MODE (x),
6998 XEXP (x, 1), GET_MODE (x));
6999 return force_to_mode (x, mode, mask, next_select);
7002 mask = fuller_mask;
7003 goto binop;
7005 case IOR:
7006 case XOR:
7007 /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and
7008 LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...)
7009 operation which may be a bitfield extraction. Ensure that the
7010 constant we form is not wider than the mode of X. */
7012 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7013 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7014 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7015 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT
7016 && GET_CODE (XEXP (x, 1)) == CONST_INT
7017 && ((INTVAL (XEXP (XEXP (x, 0), 1))
7018 + floor_log2 (INTVAL (XEXP (x, 1))))
7019 < GET_MODE_BITSIZE (GET_MODE (x)))
7020 && (INTVAL (XEXP (x, 1))
7021 & ~nonzero_bits (XEXP (x, 0), GET_MODE (x))) == 0)
7023 temp = GEN_INT ((INTVAL (XEXP (x, 1)) & mask)
7024 << INTVAL (XEXP (XEXP (x, 0), 1)));
7025 temp = simplify_gen_binary (GET_CODE (x), GET_MODE (x),
7026 XEXP (XEXP (x, 0), 0), temp);
7027 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), temp,
7028 XEXP (XEXP (x, 0), 1));
7029 return force_to_mode (x, mode, mask, next_select);
7032 binop:
7033 /* For most binary operations, just propagate into the operation and
7034 change the mode if we have an operation of that mode. */
7036 op0 = gen_lowpart_or_truncate (op_mode,
7037 force_to_mode (XEXP (x, 0), mode, mask,
7038 next_select));
7039 op1 = gen_lowpart_or_truncate (op_mode,
7040 force_to_mode (XEXP (x, 1), mode, mask,
7041 next_select));
7043 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0) || op1 != XEXP (x, 1))
7044 x = simplify_gen_binary (code, op_mode, op0, op1);
7045 break;
7047 case ASHIFT:
7048 /* For left shifts, do the same, but just for the first operand.
7049 However, we cannot do anything with shifts where we cannot
7050 guarantee that the counts are smaller than the size of the mode
7051 because such a count will have a different meaning in a
7052 wider mode. */
7054 if (! (GET_CODE (XEXP (x, 1)) == CONST_INT
7055 && INTVAL (XEXP (x, 1)) >= 0
7056 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (mode))
7057 && ! (GET_MODE (XEXP (x, 1)) != VOIDmode
7058 && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1)))
7059 < (unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode))))
7060 break;
7062 /* If the shift count is a constant and we can do arithmetic in
7063 the mode of the shift, refine which bits we need. Otherwise, use the
7064 conservative form of the mask. */
7065 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7066 && INTVAL (XEXP (x, 1)) >= 0
7067 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (op_mode)
7068 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7069 mask >>= INTVAL (XEXP (x, 1));
7070 else
7071 mask = fuller_mask;
7073 op0 = gen_lowpart_or_truncate (op_mode,
7074 force_to_mode (XEXP (x, 0), op_mode,
7075 mask, next_select));
7077 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7078 x = simplify_gen_binary (code, op_mode, op0, XEXP (x, 1));
7079 break;
7081 case LSHIFTRT:
7082 /* Here we can only do something if the shift count is a constant,
7083 this shift constant is valid for the host, and we can do arithmetic
7084 in OP_MODE. */
7086 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7087 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT
7088 && GET_MODE_BITSIZE (op_mode) <= HOST_BITS_PER_WIDE_INT)
7090 rtx inner = XEXP (x, 0);
7091 unsigned HOST_WIDE_INT inner_mask;
7093 /* Select the mask of the bits we need for the shift operand. */
7094 inner_mask = mask << INTVAL (XEXP (x, 1));
7096 /* We can only change the mode of the shift if we can do arithmetic
7097 in the mode of the shift and INNER_MASK is no wider than the
7098 width of X's mode. */
7099 if ((inner_mask & ~GET_MODE_MASK (GET_MODE (x))) != 0)
7100 op_mode = GET_MODE (x);
7102 inner = force_to_mode (inner, op_mode, inner_mask, next_select);
7104 if (GET_MODE (x) != op_mode || inner != XEXP (x, 0))
7105 x = simplify_gen_binary (LSHIFTRT, op_mode, inner, XEXP (x, 1));
7108 /* If we have (and (lshiftrt FOO C1) C2) where the combination of the
7109 shift and AND produces only copies of the sign bit (C2 is one less
7110 than a power of two), we can do this with just a shift. */
7112 if (GET_CODE (x) == LSHIFTRT
7113 && GET_CODE (XEXP (x, 1)) == CONST_INT
7114 /* The shift puts one of the sign bit copies in the least significant
7115 bit. */
7116 && ((INTVAL (XEXP (x, 1))
7117 + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))))
7118 >= GET_MODE_BITSIZE (GET_MODE (x)))
7119 && exact_log2 (mask + 1) >= 0
7120 /* Number of bits left after the shift must be more than the mask
7121 needs. */
7122 && ((INTVAL (XEXP (x, 1)) + exact_log2 (mask + 1))
7123 <= GET_MODE_BITSIZE (GET_MODE (x)))
7124 /* Must be more sign bit copies than the mask needs. */
7125 && ((int) num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
7126 >= exact_log2 (mask + 1)))
7127 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7128 GEN_INT (GET_MODE_BITSIZE (GET_MODE (x))
7129 - exact_log2 (mask + 1)));
7131 goto shiftrt;
7133 case ASHIFTRT:
7134 /* If we are just looking for the sign bit, we don't need this shift at
7135 all, even if it has a variable count. */
7136 if (GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
7137 && (mask == ((unsigned HOST_WIDE_INT) 1
7138 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
7139 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
7141 /* If this is a shift by a constant, get a mask that contains those bits
7142 that are not copies of the sign bit. We then have two cases: If
7143 MASK only includes those bits, this can be a logical shift, which may
7144 allow simplifications. If MASK is a single-bit field not within
7145 those bits, we are requesting a copy of the sign bit and hence can
7146 shift the sign bit to the appropriate location. */
7148 if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) >= 0
7149 && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT)
7151 int i;
7153 /* If the considered data is wider than HOST_WIDE_INT, we can't
7154 represent a mask for all its bits in a single scalar.
7155 But we only care about the lower bits, so calculate these. */
7157 if (GET_MODE_BITSIZE (GET_MODE (x)) > HOST_BITS_PER_WIDE_INT)
7159 nonzero = ~(HOST_WIDE_INT) 0;
7161 /* GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7162 is the number of bits a full-width mask would have set.
7163 We need only shift if these are fewer than nonzero can
7164 hold. If not, we must keep all bits set in nonzero. */
7166 if (GET_MODE_BITSIZE (GET_MODE (x)) - INTVAL (XEXP (x, 1))
7167 < HOST_BITS_PER_WIDE_INT)
7168 nonzero >>= INTVAL (XEXP (x, 1))
7169 + HOST_BITS_PER_WIDE_INT
7170 - GET_MODE_BITSIZE (GET_MODE (x)) ;
7172 else
7174 nonzero = GET_MODE_MASK (GET_MODE (x));
7175 nonzero >>= INTVAL (XEXP (x, 1));
7178 if ((mask & ~nonzero) == 0)
7180 x = simplify_shift_const (NULL_RTX, LSHIFTRT, GET_MODE (x),
7181 XEXP (x, 0), INTVAL (XEXP (x, 1)));
7182 if (GET_CODE (x) != ASHIFTRT)
7183 return force_to_mode (x, mode, mask, next_select);
7186 else if ((i = exact_log2 (mask)) >= 0)
7188 x = simplify_shift_const
7189 (NULL_RTX, LSHIFTRT, GET_MODE (x), XEXP (x, 0),
7190 GET_MODE_BITSIZE (GET_MODE (x)) - 1 - i);
7192 if (GET_CODE (x) != ASHIFTRT)
7193 return force_to_mode (x, mode, mask, next_select);
7197 /* If MASK is 1, convert this to an LSHIFTRT. This can be done
7198 even if the shift count isn't a constant. */
7199 if (mask == 1)
7200 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7201 XEXP (x, 0), XEXP (x, 1));
7203 shiftrt:
7205 /* If this is a zero- or sign-extension operation that just affects bits
7206 we don't care about, remove it. Be sure the call above returned
7207 something that is still a shift. */
7209 if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT)
7210 && GET_CODE (XEXP (x, 1)) == CONST_INT
7211 && INTVAL (XEXP (x, 1)) >= 0
7212 && (INTVAL (XEXP (x, 1))
7213 <= GET_MODE_BITSIZE (GET_MODE (x)) - (floor_log2 (mask) + 1))
7214 && GET_CODE (XEXP (x, 0)) == ASHIFT
7215 && XEXP (XEXP (x, 0), 1) == XEXP (x, 1))
7216 return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask,
7217 next_select);
7219 break;
7221 case ROTATE:
7222 case ROTATERT:
7223 /* If the shift count is constant and we can do computations
7224 in the mode of X, compute where the bits we care about are.
7225 Otherwise, we can't do anything. Don't change the mode of
7226 the shift or propagate MODE into the shift, though. */
7227 if (GET_CODE (XEXP (x, 1)) == CONST_INT
7228 && INTVAL (XEXP (x, 1)) >= 0)
7230 temp = simplify_binary_operation (code == ROTATE ? ROTATERT : ROTATE,
7231 GET_MODE (x), GEN_INT (mask),
7232 XEXP (x, 1));
7233 if (temp && GET_CODE (temp) == CONST_INT)
7234 SUBST (XEXP (x, 0),
7235 force_to_mode (XEXP (x, 0), GET_MODE (x),
7236 INTVAL (temp), next_select));
7238 break;
7240 case NEG:
7241 /* If we just want the low-order bit, the NEG isn't needed since it
7242 won't change the low-order bit. */
7243 if (mask == 1)
7244 return force_to_mode (XEXP (x, 0), mode, mask, just_select);
7246 /* We need any bits less significant than the most significant bit in
7247 MASK since carries from those bits will affect the bits we are
7248 interested in. */
7249 mask = fuller_mask;
7250 goto unop;
7252 case NOT:
7253 /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the
7254 same as the XOR case above. Ensure that the constant we form is not
7255 wider than the mode of X. */
7257 if (GET_CODE (XEXP (x, 0)) == LSHIFTRT
7258 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
7259 && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0
7260 && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (mask)
7261 < GET_MODE_BITSIZE (GET_MODE (x)))
7262 && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT)
7264 temp = gen_int_mode (mask << INTVAL (XEXP (XEXP (x, 0), 1)),
7265 GET_MODE (x));
7266 temp = simplify_gen_binary (XOR, GET_MODE (x),
7267 XEXP (XEXP (x, 0), 0), temp);
7268 x = simplify_gen_binary (LSHIFTRT, GET_MODE (x),
7269 temp, XEXP (XEXP (x, 0), 1));
7271 return force_to_mode (x, mode, mask, next_select);
7274 /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must
7275 use the full mask inside the NOT. */
7276 mask = fuller_mask;
7278 unop:
7279 op0 = gen_lowpart_or_truncate (op_mode,
7280 force_to_mode (XEXP (x, 0), mode, mask,
7281 next_select));
7282 if (op_mode != GET_MODE (x) || op0 != XEXP (x, 0))
7283 x = simplify_gen_unary (code, op_mode, op0, op_mode);
7284 break;
7286 case NE:
7287 /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included
7288 in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero,
7289 which is equal to STORE_FLAG_VALUE. */
7290 if ((mask & ~STORE_FLAG_VALUE) == 0 && XEXP (x, 1) == const0_rtx
7291 && GET_MODE (XEXP (x, 0)) == mode
7292 && exact_log2 (nonzero_bits (XEXP (x, 0), mode)) >= 0
7293 && (nonzero_bits (XEXP (x, 0), mode)
7294 == (unsigned HOST_WIDE_INT) STORE_FLAG_VALUE))
7295 return force_to_mode (XEXP (x, 0), mode, mask, next_select);
7297 break;
7299 case IF_THEN_ELSE:
7300 /* We have no way of knowing if the IF_THEN_ELSE can itself be
7301 written in a narrower mode. We play it safe and do not do so. */
7303 SUBST (XEXP (x, 1),
7304 gen_lowpart_or_truncate (GET_MODE (x),
7305 force_to_mode (XEXP (x, 1), mode,
7306 mask, next_select)));
7307 SUBST (XEXP (x, 2),
7308 gen_lowpart_or_truncate (GET_MODE (x),
7309 force_to_mode (XEXP (x, 2), mode,
7310 mask, next_select)));
7311 break;
7313 default:
7314 break;
7317 /* Ensure we return a value of the proper mode. */
7318 return gen_lowpart_or_truncate (mode, x);
7321 /* Return nonzero if X is an expression that has one of two values depending on
7322 whether some other value is zero or nonzero. In that case, we return the
7323 value that is being tested, *PTRUE is set to the value if the rtx being
7324 returned has a nonzero value, and *PFALSE is set to the other alternative.
7326 If we return zero, we set *PTRUE and *PFALSE to X. */
7328 static rtx
7329 if_then_else_cond (rtx x, rtx *ptrue, rtx *pfalse)
7331 enum machine_mode mode = GET_MODE (x);
7332 enum rtx_code code = GET_CODE (x);
7333 rtx cond0, cond1, true0, true1, false0, false1;
7334 unsigned HOST_WIDE_INT nz;
7336 /* If we are comparing a value against zero, we are done. */
7337 if ((code == NE || code == EQ)
7338 && XEXP (x, 1) == const0_rtx)
7340 *ptrue = (code == NE) ? const_true_rtx : const0_rtx;
7341 *pfalse = (code == NE) ? const0_rtx : const_true_rtx;
7342 return XEXP (x, 0);
7345 /* If this is a unary operation whose operand has one of two values, apply
7346 our opcode to compute those values. */
7347 else if (UNARY_P (x)
7348 && (cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0)) != 0)
7350 *ptrue = simplify_gen_unary (code, mode, true0, GET_MODE (XEXP (x, 0)));
7351 *pfalse = simplify_gen_unary (code, mode, false0,
7352 GET_MODE (XEXP (x, 0)));
7353 return cond0;
7356 /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would
7357 make can't possibly match and would suppress other optimizations. */
7358 else if (code == COMPARE)
7361 /* If this is a binary operation, see if either side has only one of two
7362 values. If either one does or if both do and they are conditional on
7363 the same value, compute the new true and false values. */
7364 else if (BINARY_P (x))
7366 cond0 = if_then_else_cond (XEXP (x, 0), &true0, &false0);
7367 cond1 = if_then_else_cond (XEXP (x, 1), &true1, &false1);
7369 if ((cond0 != 0 || cond1 != 0)
7370 && ! (cond0 != 0 && cond1 != 0 && ! rtx_equal_p (cond0, cond1)))
7372 /* If if_then_else_cond returned zero, then true/false are the
7373 same rtl. We must copy one of them to prevent invalid rtl
7374 sharing. */
7375 if (cond0 == 0)
7376 true0 = copy_rtx (true0);
7377 else if (cond1 == 0)
7378 true1 = copy_rtx (true1);
7380 if (COMPARISON_P (x))
7382 *ptrue = simplify_gen_relational (code, mode, VOIDmode,
7383 true0, true1);
7384 *pfalse = simplify_gen_relational (code, mode, VOIDmode,
7385 false0, false1);
7387 else
7389 *ptrue = simplify_gen_binary (code, mode, true0, true1);
7390 *pfalse = simplify_gen_binary (code, mode, false0, false1);
7393 return cond0 ? cond0 : cond1;
7396 /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the
7397 operands is zero when the other is nonzero, and vice-versa,
7398 and STORE_FLAG_VALUE is 1 or -1. */
7400 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7401 && (code == PLUS || code == IOR || code == XOR || code == MINUS
7402 || code == UMAX)
7403 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7405 rtx op0 = XEXP (XEXP (x, 0), 1);
7406 rtx op1 = XEXP (XEXP (x, 1), 1);
7408 cond0 = XEXP (XEXP (x, 0), 0);
7409 cond1 = XEXP (XEXP (x, 1), 0);
7411 if (COMPARISON_P (cond0)
7412 && COMPARISON_P (cond1)
7413 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7414 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7415 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7416 || ((swap_condition (GET_CODE (cond0))
7417 == reversed_comparison_code (cond1, NULL))
7418 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7419 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7420 && ! side_effects_p (x))
7422 *ptrue = simplify_gen_binary (MULT, mode, op0, const_true_rtx);
7423 *pfalse = simplify_gen_binary (MULT, mode,
7424 (code == MINUS
7425 ? simplify_gen_unary (NEG, mode,
7426 op1, mode)
7427 : op1),
7428 const_true_rtx);
7429 return cond0;
7433 /* Similarly for MULT, AND and UMIN, except that for these the result
7434 is always zero. */
7435 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
7436 && (code == MULT || code == AND || code == UMIN)
7437 && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT)
7439 cond0 = XEXP (XEXP (x, 0), 0);
7440 cond1 = XEXP (XEXP (x, 1), 0);
7442 if (COMPARISON_P (cond0)
7443 && COMPARISON_P (cond1)
7444 && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL)
7445 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0))
7446 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1)))
7447 || ((swap_condition (GET_CODE (cond0))
7448 == reversed_comparison_code (cond1, NULL))
7449 && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1))
7450 && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0))))
7451 && ! side_effects_p (x))
7453 *ptrue = *pfalse = const0_rtx;
7454 return cond0;
7459 else if (code == IF_THEN_ELSE)
7461 /* If we have IF_THEN_ELSE already, extract the condition and
7462 canonicalize it if it is NE or EQ. */
7463 cond0 = XEXP (x, 0);
7464 *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2);
7465 if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx)
7466 return XEXP (cond0, 0);
7467 else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx)
7469 *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1);
7470 return XEXP (cond0, 0);
7472 else
7473 return cond0;
7476 /* If X is a SUBREG, we can narrow both the true and false values
7477 if the inner expression, if there is a condition. */
7478 else if (code == SUBREG
7479 && 0 != (cond0 = if_then_else_cond (SUBREG_REG (x),
7480 &true0, &false0)))
7482 true0 = simplify_gen_subreg (mode, true0,
7483 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7484 false0 = simplify_gen_subreg (mode, false0,
7485 GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x));
7486 if (true0 && false0)
7488 *ptrue = true0;
7489 *pfalse = false0;
7490 return cond0;
7494 /* If X is a constant, this isn't special and will cause confusions
7495 if we treat it as such. Likewise if it is equivalent to a constant. */
7496 else if (CONSTANT_P (x)
7497 || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0)))
7500 /* If we're in BImode, canonicalize on 0 and STORE_FLAG_VALUE, as that
7501 will be least confusing to the rest of the compiler. */
7502 else if (mode == BImode)
7504 *ptrue = GEN_INT (STORE_FLAG_VALUE), *pfalse = const0_rtx;
7505 return x;
7508 /* If X is known to be either 0 or -1, those are the true and
7509 false values when testing X. */
7510 else if (x == constm1_rtx || x == const0_rtx
7511 || (mode != VOIDmode
7512 && num_sign_bit_copies (x, mode) == GET_MODE_BITSIZE (mode)))
7514 *ptrue = constm1_rtx, *pfalse = const0_rtx;
7515 return x;
7518 /* Likewise for 0 or a single bit. */
7519 else if (SCALAR_INT_MODE_P (mode)
7520 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
7521 && exact_log2 (nz = nonzero_bits (x, mode)) >= 0)
7523 *ptrue = gen_int_mode (nz, mode), *pfalse = const0_rtx;
7524 return x;
7527 /* Otherwise fail; show no condition with true and false values the same. */
7528 *ptrue = *pfalse = x;
7529 return 0;
7532 /* Return the value of expression X given the fact that condition COND
7533 is known to be true when applied to REG as its first operand and VAL
7534 as its second. X is known to not be shared and so can be modified in
7535 place.
7537 We only handle the simplest cases, and specifically those cases that
7538 arise with IF_THEN_ELSE expressions. */
7540 static rtx
7541 known_cond (rtx x, enum rtx_code cond, rtx reg, rtx val)
7543 enum rtx_code code = GET_CODE (x);
7544 rtx temp;
7545 const char *fmt;
7546 int i, j;
7548 if (side_effects_p (x))
7549 return x;
7551 /* If either operand of the condition is a floating point value,
7552 then we have to avoid collapsing an EQ comparison. */
7553 if (cond == EQ
7554 && rtx_equal_p (x, reg)
7555 && ! FLOAT_MODE_P (GET_MODE (x))
7556 && ! FLOAT_MODE_P (GET_MODE (val)))
7557 return val;
7559 if (cond == UNEQ && rtx_equal_p (x, reg))
7560 return val;
7562 /* If X is (abs REG) and we know something about REG's relationship
7563 with zero, we may be able to simplify this. */
7565 if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx)
7566 switch (cond)
7568 case GE: case GT: case EQ:
7569 return XEXP (x, 0);
7570 case LT: case LE:
7571 return simplify_gen_unary (NEG, GET_MODE (XEXP (x, 0)),
7572 XEXP (x, 0),
7573 GET_MODE (XEXP (x, 0)));
7574 default:
7575 break;
7578 /* The only other cases we handle are MIN, MAX, and comparisons if the
7579 operands are the same as REG and VAL. */
7581 else if (COMPARISON_P (x) || COMMUTATIVE_ARITH_P (x))
7583 if (rtx_equal_p (XEXP (x, 0), val))
7584 cond = swap_condition (cond), temp = val, val = reg, reg = temp;
7586 if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val))
7588 if (COMPARISON_P (x))
7590 if (comparison_dominates_p (cond, code))
7591 return const_true_rtx;
7593 code = reversed_comparison_code (x, NULL);
7594 if (code != UNKNOWN
7595 && comparison_dominates_p (cond, code))
7596 return const0_rtx;
7597 else
7598 return x;
7600 else if (code == SMAX || code == SMIN
7601 || code == UMIN || code == UMAX)
7603 int unsignedp = (code == UMIN || code == UMAX);
7605 /* Do not reverse the condition when it is NE or EQ.
7606 This is because we cannot conclude anything about
7607 the value of 'SMAX (x, y)' when x is not equal to y,
7608 but we can when x equals y. */
7609 if ((code == SMAX || code == UMAX)
7610 && ! (cond == EQ || cond == NE))
7611 cond = reverse_condition (cond);
7613 switch (cond)
7615 case GE: case GT:
7616 return unsignedp ? x : XEXP (x, 1);
7617 case LE: case LT:
7618 return unsignedp ? x : XEXP (x, 0);
7619 case GEU: case GTU:
7620 return unsignedp ? XEXP (x, 1) : x;
7621 case LEU: case LTU:
7622 return unsignedp ? XEXP (x, 0) : x;
7623 default:
7624 break;
7629 else if (code == SUBREG)
7631 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (x));
7632 rtx new, r = known_cond (SUBREG_REG (x), cond, reg, val);
7634 if (SUBREG_REG (x) != r)
7636 /* We must simplify subreg here, before we lose track of the
7637 original inner_mode. */
7638 new = simplify_subreg (GET_MODE (x), r,
7639 inner_mode, SUBREG_BYTE (x));
7640 if (new)
7641 return new;
7642 else
7643 SUBST (SUBREG_REG (x), r);
7646 return x;
7648 /* We don't have to handle SIGN_EXTEND here, because even in the
7649 case of replacing something with a modeless CONST_INT, a
7650 CONST_INT is already (supposed to be) a valid sign extension for
7651 its narrower mode, which implies it's already properly
7652 sign-extended for the wider mode. Now, for ZERO_EXTEND, the
7653 story is different. */
7654 else if (code == ZERO_EXTEND)
7656 enum machine_mode inner_mode = GET_MODE (XEXP (x, 0));
7657 rtx new, r = known_cond (XEXP (x, 0), cond, reg, val);
7659 if (XEXP (x, 0) != r)
7661 /* We must simplify the zero_extend here, before we lose
7662 track of the original inner_mode. */
7663 new = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x),
7664 r, inner_mode);
7665 if (new)
7666 return new;
7667 else
7668 SUBST (XEXP (x, 0), r);
7671 return x;
7674 fmt = GET_RTX_FORMAT (code);
7675 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7677 if (fmt[i] == 'e')
7678 SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val));
7679 else if (fmt[i] == 'E')
7680 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7681 SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j),
7682 cond, reg, val));
7685 return x;
7688 /* See if X and Y are equal for the purposes of seeing if we can rewrite an
7689 assignment as a field assignment. */
7691 static int
7692 rtx_equal_for_field_assignment_p (rtx x, rtx y)
7694 if (x == y || rtx_equal_p (x, y))
7695 return 1;
7697 if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y))
7698 return 0;
7700 /* Check for a paradoxical SUBREG of a MEM compared with the MEM.
7701 Note that all SUBREGs of MEM are paradoxical; otherwise they
7702 would have been rewritten. */
7703 if (MEM_P (x) && GET_CODE (y) == SUBREG
7704 && MEM_P (SUBREG_REG (y))
7705 && rtx_equal_p (SUBREG_REG (y),
7706 gen_lowpart (GET_MODE (SUBREG_REG (y)), x)))
7707 return 1;
7709 if (MEM_P (y) && GET_CODE (x) == SUBREG
7710 && MEM_P (SUBREG_REG (x))
7711 && rtx_equal_p (SUBREG_REG (x),
7712 gen_lowpart (GET_MODE (SUBREG_REG (x)), y)))
7713 return 1;
7715 /* We used to see if get_last_value of X and Y were the same but that's
7716 not correct. In one direction, we'll cause the assignment to have
7717 the wrong destination and in the case, we'll import a register into this
7718 insn that might have already have been dead. So fail if none of the
7719 above cases are true. */
7720 return 0;
7723 /* See if X, a SET operation, can be rewritten as a bit-field assignment.
7724 Return that assignment if so.
7726 We only handle the most common cases. */
7728 static rtx
7729 make_field_assignment (rtx x)
7731 rtx dest = SET_DEST (x);
7732 rtx src = SET_SRC (x);
7733 rtx assign;
7734 rtx rhs, lhs;
7735 HOST_WIDE_INT c1;
7736 HOST_WIDE_INT pos;
7737 unsigned HOST_WIDE_INT len;
7738 rtx other;
7739 enum machine_mode mode;
7741 /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is
7742 a clear of a one-bit field. We will have changed it to
7743 (and (rotate (const_int -2) POS) DEST), so check for that. Also check
7744 for a SUBREG. */
7746 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE
7747 && GET_CODE (XEXP (XEXP (src, 0), 0)) == CONST_INT
7748 && INTVAL (XEXP (XEXP (src, 0), 0)) == -2
7749 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7751 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7752 1, 1, 1, 0);
7753 if (assign != 0)
7754 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7755 return x;
7758 if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG
7759 && subreg_lowpart_p (XEXP (src, 0))
7760 && (GET_MODE_SIZE (GET_MODE (XEXP (src, 0)))
7761 < GET_MODE_SIZE (GET_MODE (SUBREG_REG (XEXP (src, 0)))))
7762 && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE
7763 && GET_CODE (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == CONST_INT
7764 && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2
7765 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7767 assign = make_extraction (VOIDmode, dest, 0,
7768 XEXP (SUBREG_REG (XEXP (src, 0)), 1),
7769 1, 1, 1, 0);
7770 if (assign != 0)
7771 return gen_rtx_SET (VOIDmode, assign, const0_rtx);
7772 return x;
7775 /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a
7776 one-bit field. */
7777 if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT
7778 && XEXP (XEXP (src, 0), 0) == const1_rtx
7779 && rtx_equal_for_field_assignment_p (dest, XEXP (src, 1)))
7781 assign = make_extraction (VOIDmode, dest, 0, XEXP (XEXP (src, 0), 1),
7782 1, 1, 1, 0);
7783 if (assign != 0)
7784 return gen_rtx_SET (VOIDmode, assign, const1_rtx);
7785 return x;
7788 /* If DEST is already a field assignment, i.e. ZERO_EXTRACT, and the
7789 SRC is an AND with all bits of that field set, then we can discard
7790 the AND. */
7791 if (GET_CODE (dest) == ZERO_EXTRACT
7792 && GET_CODE (XEXP (dest, 1)) == CONST_INT
7793 && GET_CODE (src) == AND
7794 && GET_CODE (XEXP (src, 1)) == CONST_INT)
7796 HOST_WIDE_INT width = INTVAL (XEXP (dest, 1));
7797 unsigned HOST_WIDE_INT and_mask = INTVAL (XEXP (src, 1));
7798 unsigned HOST_WIDE_INT ze_mask;
7800 if (width >= HOST_BITS_PER_WIDE_INT)
7801 ze_mask = -1;
7802 else
7803 ze_mask = ((unsigned HOST_WIDE_INT)1 << width) - 1;
7805 /* Complete overlap. We can remove the source AND. */
7806 if ((and_mask & ze_mask) == ze_mask)
7807 return gen_rtx_SET (VOIDmode, dest, XEXP (src, 0));
7809 /* Partial overlap. We can reduce the source AND. */
7810 if ((and_mask & ze_mask) != and_mask)
7812 mode = GET_MODE (src);
7813 src = gen_rtx_AND (mode, XEXP (src, 0),
7814 gen_int_mode (and_mask & ze_mask, mode));
7815 return gen_rtx_SET (VOIDmode, dest, src);
7819 /* The other case we handle is assignments into a constant-position
7820 field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents
7821 a mask that has all one bits except for a group of zero bits and
7822 OTHER is known to have zeros where C1 has ones, this is such an
7823 assignment. Compute the position and length from C1. Shift OTHER
7824 to the appropriate position, force it to the required mode, and
7825 make the extraction. Check for the AND in both operands. */
7827 if (GET_CODE (src) != IOR && GET_CODE (src) != XOR)
7828 return x;
7830 rhs = expand_compound_operation (XEXP (src, 0));
7831 lhs = expand_compound_operation (XEXP (src, 1));
7833 if (GET_CODE (rhs) == AND
7834 && GET_CODE (XEXP (rhs, 1)) == CONST_INT
7835 && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), dest))
7836 c1 = INTVAL (XEXP (rhs, 1)), other = lhs;
7837 else if (GET_CODE (lhs) == AND
7838 && GET_CODE (XEXP (lhs, 1)) == CONST_INT
7839 && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), dest))
7840 c1 = INTVAL (XEXP (lhs, 1)), other = rhs;
7841 else
7842 return x;
7844 pos = get_pos_from_mask ((~c1) & GET_MODE_MASK (GET_MODE (dest)), &len);
7845 if (pos < 0 || pos + len > GET_MODE_BITSIZE (GET_MODE (dest))
7846 || GET_MODE_BITSIZE (GET_MODE (dest)) > HOST_BITS_PER_WIDE_INT
7847 || (c1 & nonzero_bits (other, GET_MODE (dest))) != 0)
7848 return x;
7850 assign = make_extraction (VOIDmode, dest, pos, NULL_RTX, len, 1, 1, 0);
7851 if (assign == 0)
7852 return x;
7854 /* The mode to use for the source is the mode of the assignment, or of
7855 what is inside a possible STRICT_LOW_PART. */
7856 mode = (GET_CODE (assign) == STRICT_LOW_PART
7857 ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign));
7859 /* Shift OTHER right POS places and make it the source, restricting it
7860 to the proper length and mode. */
7862 src = canon_reg_for_combine (simplify_shift_const (NULL_RTX, LSHIFTRT,
7863 GET_MODE (src),
7864 other, pos),
7865 dest);
7866 src = force_to_mode (src, mode,
7867 GET_MODE_BITSIZE (mode) >= HOST_BITS_PER_WIDE_INT
7868 ? ~(unsigned HOST_WIDE_INT) 0
7869 : ((unsigned HOST_WIDE_INT) 1 << len) - 1,
7872 /* If SRC is masked by an AND that does not make a difference in
7873 the value being stored, strip it. */
7874 if (GET_CODE (assign) == ZERO_EXTRACT
7875 && GET_CODE (XEXP (assign, 1)) == CONST_INT
7876 && INTVAL (XEXP (assign, 1)) < HOST_BITS_PER_WIDE_INT
7877 && GET_CODE (src) == AND
7878 && GET_CODE (XEXP (src, 1)) == CONST_INT
7879 && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (src, 1))
7880 == ((unsigned HOST_WIDE_INT) 1 << INTVAL (XEXP (assign, 1))) - 1))
7881 src = XEXP (src, 0);
7883 return gen_rtx_SET (VOIDmode, assign, src);
7886 /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c)
7887 if so. */
7889 static rtx
7890 apply_distributive_law (rtx x)
7892 enum rtx_code code = GET_CODE (x);
7893 enum rtx_code inner_code;
7894 rtx lhs, rhs, other;
7895 rtx tem;
7897 /* Distributivity is not true for floating point as it can change the
7898 value. So we don't do it unless -funsafe-math-optimizations. */
7899 if (FLOAT_MODE_P (GET_MODE (x))
7900 && ! flag_unsafe_math_optimizations)
7901 return x;
7903 /* The outer operation can only be one of the following: */
7904 if (code != IOR && code != AND && code != XOR
7905 && code != PLUS && code != MINUS)
7906 return x;
7908 lhs = XEXP (x, 0);
7909 rhs = XEXP (x, 1);
7911 /* If either operand is a primitive we can't do anything, so get out
7912 fast. */
7913 if (OBJECT_P (lhs) || OBJECT_P (rhs))
7914 return x;
7916 lhs = expand_compound_operation (lhs);
7917 rhs = expand_compound_operation (rhs);
7918 inner_code = GET_CODE (lhs);
7919 if (inner_code != GET_CODE (rhs))
7920 return x;
7922 /* See if the inner and outer operations distribute. */
7923 switch (inner_code)
7925 case LSHIFTRT:
7926 case ASHIFTRT:
7927 case AND:
7928 case IOR:
7929 /* These all distribute except over PLUS. */
7930 if (code == PLUS || code == MINUS)
7931 return x;
7932 break;
7934 case MULT:
7935 if (code != PLUS && code != MINUS)
7936 return x;
7937 break;
7939 case ASHIFT:
7940 /* This is also a multiply, so it distributes over everything. */
7941 break;
7943 case SUBREG:
7944 /* Non-paradoxical SUBREGs distributes over all operations,
7945 provided the inner modes and byte offsets are the same, this
7946 is an extraction of a low-order part, we don't convert an fp
7947 operation to int or vice versa, this is not a vector mode,
7948 and we would not be converting a single-word operation into a
7949 multi-word operation. The latter test is not required, but
7950 it prevents generating unneeded multi-word operations. Some
7951 of the previous tests are redundant given the latter test,
7952 but are retained because they are required for correctness.
7954 We produce the result slightly differently in this case. */
7956 if (GET_MODE (SUBREG_REG (lhs)) != GET_MODE (SUBREG_REG (rhs))
7957 || SUBREG_BYTE (lhs) != SUBREG_BYTE (rhs)
7958 || ! subreg_lowpart_p (lhs)
7959 || (GET_MODE_CLASS (GET_MODE (lhs))
7960 != GET_MODE_CLASS (GET_MODE (SUBREG_REG (lhs))))
7961 || (GET_MODE_SIZE (GET_MODE (lhs))
7962 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))))
7963 || VECTOR_MODE_P (GET_MODE (lhs))
7964 || GET_MODE_SIZE (GET_MODE (SUBREG_REG (lhs))) > UNITS_PER_WORD
7965 /* Result might need to be truncated. Don't change mode if
7966 explicit truncation is needed. */
7967 || !TRULY_NOOP_TRUNCATION
7968 (GET_MODE_BITSIZE (GET_MODE (x)),
7969 GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (lhs)))))
7970 return x;
7972 tem = simplify_gen_binary (code, GET_MODE (SUBREG_REG (lhs)),
7973 SUBREG_REG (lhs), SUBREG_REG (rhs));
7974 return gen_lowpart (GET_MODE (x), tem);
7976 default:
7977 return x;
7980 /* Set LHS and RHS to the inner operands (A and B in the example
7981 above) and set OTHER to the common operand (C in the example).
7982 There is only one way to do this unless the inner operation is
7983 commutative. */
7984 if (COMMUTATIVE_ARITH_P (lhs)
7985 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0)))
7986 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1);
7987 else if (COMMUTATIVE_ARITH_P (lhs)
7988 && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1)))
7989 other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0);
7990 else if (COMMUTATIVE_ARITH_P (lhs)
7991 && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0)))
7992 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1);
7993 else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1)))
7994 other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0);
7995 else
7996 return x;
7998 /* Form the new inner operation, seeing if it simplifies first. */
7999 tem = simplify_gen_binary (code, GET_MODE (x), lhs, rhs);
8001 /* There is one exception to the general way of distributing:
8002 (a | c) ^ (b | c) -> (a ^ b) & ~c */
8003 if (code == XOR && inner_code == IOR)
8005 inner_code = AND;
8006 other = simplify_gen_unary (NOT, GET_MODE (x), other, GET_MODE (x));
8009 /* We may be able to continuing distributing the result, so call
8010 ourselves recursively on the inner operation before forming the
8011 outer operation, which we return. */
8012 return simplify_gen_binary (inner_code, GET_MODE (x),
8013 apply_distributive_law (tem), other);
8016 /* See if X is of the form (* (+ A B) C), and if so convert to
8017 (+ (* A C) (* B C)) and try to simplify.
8019 Most of the time, this results in no change. However, if some of
8020 the operands are the same or inverses of each other, simplifications
8021 will result.
8023 For example, (and (ior A B) (not B)) can occur as the result of
8024 expanding a bit field assignment. When we apply the distributive
8025 law to this, we get (ior (and (A (not B))) (and (B (not B)))),
8026 which then simplifies to (and (A (not B))).
8028 Note that no checks happen on the validity of applying the inverse
8029 distributive law. This is pointless since we can do it in the
8030 few places where this routine is called.
8032 N is the index of the term that is decomposed (the arithmetic operation,
8033 i.e. (+ A B) in the first example above). !N is the index of the term that
8034 is distributed, i.e. of C in the first example above. */
8035 static rtx
8036 distribute_and_simplify_rtx (rtx x, int n)
8038 enum machine_mode mode;
8039 enum rtx_code outer_code, inner_code;
8040 rtx decomposed, distributed, inner_op0, inner_op1, new_op0, new_op1, tmp;
8042 decomposed = XEXP (x, n);
8043 if (!ARITHMETIC_P (decomposed))
8044 return NULL_RTX;
8046 mode = GET_MODE (x);
8047 outer_code = GET_CODE (x);
8048 distributed = XEXP (x, !n);
8050 inner_code = GET_CODE (decomposed);
8051 inner_op0 = XEXP (decomposed, 0);
8052 inner_op1 = XEXP (decomposed, 1);
8054 /* Special case (and (xor B C) (not A)), which is equivalent to
8055 (xor (ior A B) (ior A C)) */
8056 if (outer_code == AND && inner_code == XOR && GET_CODE (distributed) == NOT)
8058 distributed = XEXP (distributed, 0);
8059 outer_code = IOR;
8062 if (n == 0)
8064 /* Distribute the second term. */
8065 new_op0 = simplify_gen_binary (outer_code, mode, inner_op0, distributed);
8066 new_op1 = simplify_gen_binary (outer_code, mode, inner_op1, distributed);
8068 else
8070 /* Distribute the first term. */
8071 new_op0 = simplify_gen_binary (outer_code, mode, distributed, inner_op0);
8072 new_op1 = simplify_gen_binary (outer_code, mode, distributed, inner_op1);
8075 tmp = apply_distributive_law (simplify_gen_binary (inner_code, mode,
8076 new_op0, new_op1));
8077 if (GET_CODE (tmp) != outer_code
8078 && rtx_cost (tmp, SET) < rtx_cost (x, SET))
8079 return tmp;
8081 return NULL_RTX;
8084 /* Simplify a logical `and' of VAROP with the constant CONSTOP, to be done
8085 in MODE. Return an equivalent form, if different from (and VAROP
8086 (const_int CONSTOP)). Otherwise, return NULL_RTX. */
8088 static rtx
8089 simplify_and_const_int_1 (enum machine_mode mode, rtx varop,
8090 unsigned HOST_WIDE_INT constop)
8092 unsigned HOST_WIDE_INT nonzero;
8093 unsigned HOST_WIDE_INT orig_constop;
8094 rtx orig_varop;
8095 int i;
8097 orig_varop = varop;
8098 orig_constop = constop;
8099 if (GET_CODE (varop) == CLOBBER)
8100 return NULL_RTX;
8102 /* Simplify VAROP knowing that we will be only looking at some of the
8103 bits in it.
8105 Note by passing in CONSTOP, we guarantee that the bits not set in
8106 CONSTOP are not significant and will never be examined. We must
8107 ensure that is the case by explicitly masking out those bits
8108 before returning. */
8109 varop = force_to_mode (varop, mode, constop, 0);
8111 /* If VAROP is a CLOBBER, we will fail so return it. */
8112 if (GET_CODE (varop) == CLOBBER)
8113 return varop;
8115 /* If VAROP is a CONST_INT, then we need to apply the mask in CONSTOP
8116 to VAROP and return the new constant. */
8117 if (GET_CODE (varop) == CONST_INT)
8118 return gen_int_mode (INTVAL (varop) & constop, mode);
8120 /* See what bits may be nonzero in VAROP. Unlike the general case of
8121 a call to nonzero_bits, here we don't care about bits outside
8122 MODE. */
8124 nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
8126 /* Turn off all bits in the constant that are known to already be zero.
8127 Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS
8128 which is tested below. */
8130 constop &= nonzero;
8132 /* If we don't have any bits left, return zero. */
8133 if (constop == 0)
8134 return const0_rtx;
8136 /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is
8137 a power of two, we can replace this with an ASHIFT. */
8138 if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), mode) == 1
8139 && (i = exact_log2 (constop)) >= 0)
8140 return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i);
8142 /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR
8143 or XOR, then try to apply the distributive law. This may eliminate
8144 operations if either branch can be simplified because of the AND.
8145 It may also make some cases more complex, but those cases probably
8146 won't match a pattern either with or without this. */
8148 if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR)
8149 return
8150 gen_lowpart
8151 (mode,
8152 apply_distributive_law
8153 (simplify_gen_binary (GET_CODE (varop), GET_MODE (varop),
8154 simplify_and_const_int (NULL_RTX,
8155 GET_MODE (varop),
8156 XEXP (varop, 0),
8157 constop),
8158 simplify_and_const_int (NULL_RTX,
8159 GET_MODE (varop),
8160 XEXP (varop, 1),
8161 constop))));
8163 /* If VAROP is PLUS, and the constant is a mask of low bits, distribute
8164 the AND and see if one of the operands simplifies to zero. If so, we
8165 may eliminate it. */
8167 if (GET_CODE (varop) == PLUS
8168 && exact_log2 (constop + 1) >= 0)
8170 rtx o0, o1;
8172 o0 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 0), constop);
8173 o1 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 1), constop);
8174 if (o0 == const0_rtx)
8175 return o1;
8176 if (o1 == const0_rtx)
8177 return o0;
8180 /* Make a SUBREG if necessary. If we can't make it, fail. */
8181 varop = gen_lowpart (mode, varop);
8182 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
8183 return NULL_RTX;
8185 /* If we are only masking insignificant bits, return VAROP. */
8186 if (constop == nonzero)
8187 return varop;
8189 if (varop == orig_varop && constop == orig_constop)
8190 return NULL_RTX;
8192 /* Otherwise, return an AND. */
8193 constop = trunc_int_for_mode (constop, mode);
8194 return simplify_gen_binary (AND, mode, varop, GEN_INT (constop));
8198 /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done
8199 in MODE.
8201 Return an equivalent form, if different from X. Otherwise, return X. If
8202 X is zero, we are to always construct the equivalent form. */
8204 static rtx
8205 simplify_and_const_int (rtx x, enum machine_mode mode, rtx varop,
8206 unsigned HOST_WIDE_INT constop)
8208 rtx tem = simplify_and_const_int_1 (mode, varop, constop);
8209 if (tem)
8210 return tem;
8212 if (!x)
8213 x = simplify_gen_binary (AND, GET_MODE (varop), varop, GEN_INT (constop));
8214 if (GET_MODE (x) != mode)
8215 x = gen_lowpart (mode, x);
8216 return x;
8219 /* Given a REG, X, compute which bits in X can be nonzero.
8220 We don't care about bits outside of those defined in MODE.
8222 For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is
8223 a shift, AND, or zero_extract, we can do better. */
8225 static rtx
8226 reg_nonzero_bits_for_combine (rtx x, enum machine_mode mode,
8227 rtx known_x ATTRIBUTE_UNUSED,
8228 enum machine_mode known_mode ATTRIBUTE_UNUSED,
8229 unsigned HOST_WIDE_INT known_ret ATTRIBUTE_UNUSED,
8230 unsigned HOST_WIDE_INT *nonzero)
8232 rtx tem;
8234 /* If X is a register whose nonzero bits value is current, use it.
8235 Otherwise, if X is a register whose value we can find, use that
8236 value. Otherwise, use the previously-computed global nonzero bits
8237 for this register. */
8239 if (reg_stat[REGNO (x)].last_set_value != 0
8240 && (reg_stat[REGNO (x)].last_set_mode == mode
8241 || (GET_MODE_CLASS (reg_stat[REGNO (x)].last_set_mode) == MODE_INT
8242 && GET_MODE_CLASS (mode) == MODE_INT))
8243 && (reg_stat[REGNO (x)].last_set_label == label_tick
8244 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8245 && REG_N_SETS (REGNO (x)) == 1
8246 && ! REGNO_REG_SET_P
8247 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
8248 REGNO (x))))
8249 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8251 *nonzero &= reg_stat[REGNO (x)].last_set_nonzero_bits;
8252 return NULL;
8255 tem = get_last_value (x);
8257 if (tem)
8259 #ifdef SHORT_IMMEDIATES_SIGN_EXTEND
8260 /* If X is narrower than MODE and TEM is a non-negative
8261 constant that would appear negative in the mode of X,
8262 sign-extend it for use in reg_nonzero_bits because some
8263 machines (maybe most) will actually do the sign-extension
8264 and this is the conservative approach.
8266 ??? For 2.5, try to tighten up the MD files in this regard
8267 instead of this kludge. */
8269 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode)
8270 && GET_CODE (tem) == CONST_INT
8271 && INTVAL (tem) > 0
8272 && 0 != (INTVAL (tem)
8273 & ((HOST_WIDE_INT) 1
8274 << (GET_MODE_BITSIZE (GET_MODE (x)) - 1))))
8275 tem = GEN_INT (INTVAL (tem)
8276 | ((HOST_WIDE_INT) (-1)
8277 << GET_MODE_BITSIZE (GET_MODE (x))));
8278 #endif
8279 return tem;
8281 else if (nonzero_sign_valid && reg_stat[REGNO (x)].nonzero_bits)
8283 unsigned HOST_WIDE_INT mask = reg_stat[REGNO (x)].nonzero_bits;
8285 if (GET_MODE_BITSIZE (GET_MODE (x)) < GET_MODE_BITSIZE (mode))
8286 /* We don't know anything about the upper bits. */
8287 mask |= GET_MODE_MASK (mode) ^ GET_MODE_MASK (GET_MODE (x));
8288 *nonzero &= mask;
8291 return NULL;
8294 /* Return the number of bits at the high-order end of X that are known to
8295 be equal to the sign bit. X will be used in mode MODE; if MODE is
8296 VOIDmode, X will be used in its own mode. The returned value will always
8297 be between 1 and the number of bits in MODE. */
8299 static rtx
8300 reg_num_sign_bit_copies_for_combine (rtx x, enum machine_mode mode,
8301 rtx known_x ATTRIBUTE_UNUSED,
8302 enum machine_mode known_mode
8303 ATTRIBUTE_UNUSED,
8304 unsigned int known_ret ATTRIBUTE_UNUSED,
8305 unsigned int *result)
8307 rtx tem;
8309 if (reg_stat[REGNO (x)].last_set_value != 0
8310 && reg_stat[REGNO (x)].last_set_mode == mode
8311 && (reg_stat[REGNO (x)].last_set_label == label_tick
8312 || (REGNO (x) >= FIRST_PSEUDO_REGISTER
8313 && REG_N_SETS (REGNO (x)) == 1
8314 && ! REGNO_REG_SET_P
8315 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
8316 REGNO (x))))
8317 && INSN_CUID (reg_stat[REGNO (x)].last_set) < subst_low_cuid)
8319 *result = reg_stat[REGNO (x)].last_set_sign_bit_copies;
8320 return NULL;
8323 tem = get_last_value (x);
8324 if (tem != 0)
8325 return tem;
8327 if (nonzero_sign_valid && reg_stat[REGNO (x)].sign_bit_copies != 0
8328 && GET_MODE_BITSIZE (GET_MODE (x)) == GET_MODE_BITSIZE (mode))
8329 *result = reg_stat[REGNO (x)].sign_bit_copies;
8331 return NULL;
8334 /* Return the number of "extended" bits there are in X, when interpreted
8335 as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For
8336 unsigned quantities, this is the number of high-order zero bits.
8337 For signed quantities, this is the number of copies of the sign bit
8338 minus 1. In both case, this function returns the number of "spare"
8339 bits. For example, if two quantities for which this function returns
8340 at least 1 are added, the addition is known not to overflow.
8342 This function will always return 0 unless called during combine, which
8343 implies that it must be called from a define_split. */
8345 unsigned int
8346 extended_count (rtx x, enum machine_mode mode, int unsignedp)
8348 if (nonzero_sign_valid == 0)
8349 return 0;
8351 return (unsignedp
8352 ? (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
8353 ? (unsigned int) (GET_MODE_BITSIZE (mode) - 1
8354 - floor_log2 (nonzero_bits (x, mode)))
8355 : 0)
8356 : num_sign_bit_copies (x, mode) - 1);
8359 /* This function is called from `simplify_shift_const' to merge two
8360 outer operations. Specifically, we have already found that we need
8361 to perform operation *POP0 with constant *PCONST0 at the outermost
8362 position. We would now like to also perform OP1 with constant CONST1
8363 (with *POP0 being done last).
8365 Return 1 if we can do the operation and update *POP0 and *PCONST0 with
8366 the resulting operation. *PCOMP_P is set to 1 if we would need to
8367 complement the innermost operand, otherwise it is unchanged.
8369 MODE is the mode in which the operation will be done. No bits outside
8370 the width of this mode matter. It is assumed that the width of this mode
8371 is smaller than or equal to HOST_BITS_PER_WIDE_INT.
8373 If *POP0 or OP1 are UNKNOWN, it means no operation is required. Only NEG, PLUS,
8374 IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper
8375 result is simply *PCONST0.
8377 If the resulting operation cannot be expressed as one operation, we
8378 return 0 and do not change *POP0, *PCONST0, and *PCOMP_P. */
8380 static int
8381 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)
8383 enum rtx_code op0 = *pop0;
8384 HOST_WIDE_INT const0 = *pconst0;
8386 const0 &= GET_MODE_MASK (mode);
8387 const1 &= GET_MODE_MASK (mode);
8389 /* If OP0 is an AND, clear unimportant bits in CONST1. */
8390 if (op0 == AND)
8391 const1 &= const0;
8393 /* If OP0 or OP1 is UNKNOWN, this is easy. Similarly if they are the same or
8394 if OP0 is SET. */
8396 if (op1 == UNKNOWN || op0 == SET)
8397 return 1;
8399 else if (op0 == UNKNOWN)
8400 op0 = op1, const0 = const1;
8402 else if (op0 == op1)
8404 switch (op0)
8406 case AND:
8407 const0 &= const1;
8408 break;
8409 case IOR:
8410 const0 |= const1;
8411 break;
8412 case XOR:
8413 const0 ^= const1;
8414 break;
8415 case PLUS:
8416 const0 += const1;
8417 break;
8418 case NEG:
8419 op0 = UNKNOWN;
8420 break;
8421 default:
8422 break;
8426 /* Otherwise, if either is a PLUS or NEG, we can't do anything. */
8427 else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG)
8428 return 0;
8430 /* If the two constants aren't the same, we can't do anything. The
8431 remaining six cases can all be done. */
8432 else if (const0 != const1)
8433 return 0;
8435 else
8436 switch (op0)
8438 case IOR:
8439 if (op1 == AND)
8440 /* (a & b) | b == b */
8441 op0 = SET;
8442 else /* op1 == XOR */
8443 /* (a ^ b) | b == a | b */
8445 break;
8447 case XOR:
8448 if (op1 == AND)
8449 /* (a & b) ^ b == (~a) & b */
8450 op0 = AND, *pcomp_p = 1;
8451 else /* op1 == IOR */
8452 /* (a | b) ^ b == a & ~b */
8453 op0 = AND, const0 = ~const0;
8454 break;
8456 case AND:
8457 if (op1 == IOR)
8458 /* (a | b) & b == b */
8459 op0 = SET;
8460 else /* op1 == XOR */
8461 /* (a ^ b) & b) == (~a) & b */
8462 *pcomp_p = 1;
8463 break;
8464 default:
8465 break;
8468 /* Check for NO-OP cases. */
8469 const0 &= GET_MODE_MASK (mode);
8470 if (const0 == 0
8471 && (op0 == IOR || op0 == XOR || op0 == PLUS))
8472 op0 = UNKNOWN;
8473 else if (const0 == 0 && op0 == AND)
8474 op0 = SET;
8475 else if ((unsigned HOST_WIDE_INT) const0 == GET_MODE_MASK (mode)
8476 && op0 == AND)
8477 op0 = UNKNOWN;
8479 /* ??? Slightly redundant with the above mask, but not entirely.
8480 Moving this above means we'd have to sign-extend the mode mask
8481 for the final test. */
8482 const0 = trunc_int_for_mode (const0, mode);
8484 *pop0 = op0;
8485 *pconst0 = const0;
8487 return 1;
8490 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
8491 The result of the shift is RESULT_MODE. Return NULL_RTX if we cannot
8492 simplify it. Otherwise, return a simplified value.
8494 The shift is normally computed in the widest mode we find in VAROP, as
8495 long as it isn't a different number of words than RESULT_MODE. Exceptions
8496 are ASHIFTRT and ROTATE, which are always done in their original mode. */
8498 static rtx
8499 simplify_shift_const_1 (enum rtx_code code, enum machine_mode result_mode,
8500 rtx varop, int orig_count)
8502 enum rtx_code orig_code = code;
8503 rtx orig_varop = varop;
8504 int count;
8505 enum machine_mode mode = result_mode;
8506 enum machine_mode shift_mode, tmode;
8507 unsigned int mode_words
8508 = (GET_MODE_SIZE (mode) + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
8509 /* We form (outer_op (code varop count) (outer_const)). */
8510 enum rtx_code outer_op = UNKNOWN;
8511 HOST_WIDE_INT outer_const = 0;
8512 int complement_p = 0;
8513 rtx new, x;
8515 /* Make sure and truncate the "natural" shift on the way in. We don't
8516 want to do this inside the loop as it makes it more difficult to
8517 combine shifts. */
8518 if (SHIFT_COUNT_TRUNCATED)
8519 orig_count &= GET_MODE_BITSIZE (mode) - 1;
8521 /* If we were given an invalid count, don't do anything except exactly
8522 what was requested. */
8524 if (orig_count < 0 || orig_count >= (int) GET_MODE_BITSIZE (mode))
8525 return NULL_RTX;
8527 count = orig_count;
8529 /* Unless one of the branches of the `if' in this loop does a `continue',
8530 we will `break' the loop after the `if'. */
8532 while (count != 0)
8534 /* If we have an operand of (clobber (const_int 0)), fail. */
8535 if (GET_CODE (varop) == CLOBBER)
8536 return NULL_RTX;
8538 /* If we discovered we had to complement VAROP, leave. Making a NOT
8539 here would cause an infinite loop. */
8540 if (complement_p)
8541 break;
8543 /* Convert ROTATERT to ROTATE. */
8544 if (code == ROTATERT)
8546 unsigned int bitsize = GET_MODE_BITSIZE (result_mode);;
8547 code = ROTATE;
8548 if (VECTOR_MODE_P (result_mode))
8549 count = bitsize / GET_MODE_NUNITS (result_mode) - count;
8550 else
8551 count = bitsize - count;
8554 /* We need to determine what mode we will do the shift in. If the
8555 shift is a right shift or a ROTATE, we must always do it in the mode
8556 it was originally done in. Otherwise, we can do it in MODE, the
8557 widest mode encountered. */
8558 shift_mode
8559 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
8560 ? result_mode : mode);
8562 /* Handle cases where the count is greater than the size of the mode
8563 minus 1. For ASHIFT, use the size minus one as the count (this can
8564 occur when simplifying (lshiftrt (ashiftrt ..))). For rotates,
8565 take the count modulo the size. For other shifts, the result is
8566 zero.
8568 Since these shifts are being produced by the compiler by combining
8569 multiple operations, each of which are defined, we know what the
8570 result is supposed to be. */
8572 if (count > (GET_MODE_BITSIZE (shift_mode) - 1))
8574 if (code == ASHIFTRT)
8575 count = GET_MODE_BITSIZE (shift_mode) - 1;
8576 else if (code == ROTATE || code == ROTATERT)
8577 count %= GET_MODE_BITSIZE (shift_mode);
8578 else
8580 /* We can't simply return zero because there may be an
8581 outer op. */
8582 varop = const0_rtx;
8583 count = 0;
8584 break;
8588 /* An arithmetic right shift of a quantity known to be -1 or 0
8589 is a no-op. */
8590 if (code == ASHIFTRT
8591 && (num_sign_bit_copies (varop, shift_mode)
8592 == GET_MODE_BITSIZE (shift_mode)))
8594 count = 0;
8595 break;
8598 /* If we are doing an arithmetic right shift and discarding all but
8599 the sign bit copies, this is equivalent to doing a shift by the
8600 bitsize minus one. Convert it into that shift because it will often
8601 allow other simplifications. */
8603 if (code == ASHIFTRT
8604 && (count + num_sign_bit_copies (varop, shift_mode)
8605 >= GET_MODE_BITSIZE (shift_mode)))
8606 count = GET_MODE_BITSIZE (shift_mode) - 1;
8608 /* We simplify the tests below and elsewhere by converting
8609 ASHIFTRT to LSHIFTRT if we know the sign bit is clear.
8610 `make_compound_operation' will convert it to an ASHIFTRT for
8611 those machines (such as VAX) that don't have an LSHIFTRT. */
8612 if (GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8613 && code == ASHIFTRT
8614 && ((nonzero_bits (varop, shift_mode)
8615 & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (shift_mode) - 1)))
8616 == 0))
8617 code = LSHIFTRT;
8619 if (code == LSHIFTRT
8620 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8621 && !(nonzero_bits (varop, shift_mode) >> count))
8622 varop = const0_rtx;
8623 if (code == ASHIFT
8624 && GET_MODE_BITSIZE (shift_mode) <= HOST_BITS_PER_WIDE_INT
8625 && !((nonzero_bits (varop, shift_mode) << count)
8626 & GET_MODE_MASK (shift_mode)))
8627 varop = const0_rtx;
8629 switch (GET_CODE (varop))
8631 case SIGN_EXTEND:
8632 case ZERO_EXTEND:
8633 case SIGN_EXTRACT:
8634 case ZERO_EXTRACT:
8635 new = expand_compound_operation (varop);
8636 if (new != varop)
8638 varop = new;
8639 continue;
8641 break;
8643 case MEM:
8644 /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH
8645 minus the width of a smaller mode, we can do this with a
8646 SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */
8647 if ((code == ASHIFTRT || code == LSHIFTRT)
8648 && ! mode_dependent_address_p (XEXP (varop, 0))
8649 && ! MEM_VOLATILE_P (varop)
8650 && (tmode = mode_for_size (GET_MODE_BITSIZE (mode) - count,
8651 MODE_INT, 1)) != BLKmode)
8653 new = adjust_address_nv (varop, tmode,
8654 BYTES_BIG_ENDIAN ? 0
8655 : count / BITS_PER_UNIT);
8657 varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND
8658 : ZERO_EXTEND, mode, new);
8659 count = 0;
8660 continue;
8662 break;
8664 case SUBREG:
8665 /* If VAROP is a SUBREG, strip it as long as the inner operand has
8666 the same number of words as what we've seen so far. Then store
8667 the widest mode in MODE. */
8668 if (subreg_lowpart_p (varop)
8669 && (GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8670 > GET_MODE_SIZE (GET_MODE (varop)))
8671 && (unsigned int) ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (varop)))
8672 + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)
8673 == mode_words)
8675 varop = SUBREG_REG (varop);
8676 if (GET_MODE_SIZE (GET_MODE (varop)) > GET_MODE_SIZE (mode))
8677 mode = GET_MODE (varop);
8678 continue;
8680 break;
8682 case MULT:
8683 /* Some machines use MULT instead of ASHIFT because MULT
8684 is cheaper. But it is still better on those machines to
8685 merge two shifts into one. */
8686 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8687 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8689 varop
8690 = simplify_gen_binary (ASHIFT, GET_MODE (varop),
8691 XEXP (varop, 0),
8692 GEN_INT (exact_log2 (
8693 INTVAL (XEXP (varop, 1)))));
8694 continue;
8696 break;
8698 case UDIV:
8699 /* Similar, for when divides are cheaper. */
8700 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8701 && exact_log2 (INTVAL (XEXP (varop, 1))) >= 0)
8703 varop
8704 = simplify_gen_binary (LSHIFTRT, GET_MODE (varop),
8705 XEXP (varop, 0),
8706 GEN_INT (exact_log2 (
8707 INTVAL (XEXP (varop, 1)))));
8708 continue;
8710 break;
8712 case ASHIFTRT:
8713 /* If we are extracting just the sign bit of an arithmetic
8714 right shift, that shift is not needed. However, the sign
8715 bit of a wider mode may be different from what would be
8716 interpreted as the sign bit in a narrower mode, so, if
8717 the result is narrower, don't discard the shift. */
8718 if (code == LSHIFTRT
8719 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8720 && (GET_MODE_BITSIZE (result_mode)
8721 >= GET_MODE_BITSIZE (GET_MODE (varop))))
8723 varop = XEXP (varop, 0);
8724 continue;
8727 /* ... fall through ... */
8729 case LSHIFTRT:
8730 case ASHIFT:
8731 case ROTATE:
8732 /* Here we have two nested shifts. The result is usually the
8733 AND of a new shift with a mask. We compute the result below. */
8734 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8735 && INTVAL (XEXP (varop, 1)) >= 0
8736 && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop))
8737 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
8738 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
8739 && !VECTOR_MODE_P (result_mode))
8741 enum rtx_code first_code = GET_CODE (varop);
8742 unsigned int first_count = INTVAL (XEXP (varop, 1));
8743 unsigned HOST_WIDE_INT mask;
8744 rtx mask_rtx;
8746 /* We have one common special case. We can't do any merging if
8747 the inner code is an ASHIFTRT of a smaller mode. However, if
8748 we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2)
8749 with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2),
8750 we can convert it to
8751 (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0 C2) C3) C1).
8752 This simplifies certain SIGN_EXTEND operations. */
8753 if (code == ASHIFT && first_code == ASHIFTRT
8754 && count == (GET_MODE_BITSIZE (result_mode)
8755 - GET_MODE_BITSIZE (GET_MODE (varop))))
8757 /* C3 has the low-order C1 bits zero. */
8759 mask = (GET_MODE_MASK (mode)
8760 & ~(((HOST_WIDE_INT) 1 << first_count) - 1));
8762 varop = simplify_and_const_int (NULL_RTX, result_mode,
8763 XEXP (varop, 0), mask);
8764 varop = simplify_shift_const (NULL_RTX, ASHIFT, result_mode,
8765 varop, count);
8766 count = first_count;
8767 code = ASHIFTRT;
8768 continue;
8771 /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more
8772 than C1 high-order bits equal to the sign bit, we can convert
8773 this to either an ASHIFT or an ASHIFTRT depending on the
8774 two counts.
8776 We cannot do this if VAROP's mode is not SHIFT_MODE. */
8778 if (code == ASHIFTRT && first_code == ASHIFT
8779 && GET_MODE (varop) == shift_mode
8780 && (num_sign_bit_copies (XEXP (varop, 0), shift_mode)
8781 > first_count))
8783 varop = XEXP (varop, 0);
8784 count -= first_count;
8785 if (count < 0)
8787 count = -count;
8788 code = ASHIFT;
8791 continue;
8794 /* There are some cases we can't do. If CODE is ASHIFTRT,
8795 we can only do this if FIRST_CODE is also ASHIFTRT.
8797 We can't do the case when CODE is ROTATE and FIRST_CODE is
8798 ASHIFTRT.
8800 If the mode of this shift is not the mode of the outer shift,
8801 we can't do this if either shift is a right shift or ROTATE.
8803 Finally, we can't do any of these if the mode is too wide
8804 unless the codes are the same.
8806 Handle the case where the shift codes are the same
8807 first. */
8809 if (code == first_code)
8811 if (GET_MODE (varop) != result_mode
8812 && (code == ASHIFTRT || code == LSHIFTRT
8813 || code == ROTATE))
8814 break;
8816 count += first_count;
8817 varop = XEXP (varop, 0);
8818 continue;
8821 if (code == ASHIFTRT
8822 || (code == ROTATE && first_code == ASHIFTRT)
8823 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT
8824 || (GET_MODE (varop) != result_mode
8825 && (first_code == ASHIFTRT || first_code == LSHIFTRT
8826 || first_code == ROTATE
8827 || code == ROTATE)))
8828 break;
8830 /* To compute the mask to apply after the shift, shift the
8831 nonzero bits of the inner shift the same way the
8832 outer shift will. */
8834 mask_rtx = GEN_INT (nonzero_bits (varop, GET_MODE (varop)));
8836 mask_rtx
8837 = simplify_const_binary_operation (code, result_mode, mask_rtx,
8838 GEN_INT (count));
8840 /* Give up if we can't compute an outer operation to use. */
8841 if (mask_rtx == 0
8842 || GET_CODE (mask_rtx) != CONST_INT
8843 || ! merge_outer_ops (&outer_op, &outer_const, AND,
8844 INTVAL (mask_rtx),
8845 result_mode, &complement_p))
8846 break;
8848 /* If the shifts are in the same direction, we add the
8849 counts. Otherwise, we subtract them. */
8850 if ((code == ASHIFTRT || code == LSHIFTRT)
8851 == (first_code == ASHIFTRT || first_code == LSHIFTRT))
8852 count += first_count;
8853 else
8854 count -= first_count;
8856 /* If COUNT is positive, the new shift is usually CODE,
8857 except for the two exceptions below, in which case it is
8858 FIRST_CODE. If the count is negative, FIRST_CODE should
8859 always be used */
8860 if (count > 0
8861 && ((first_code == ROTATE && code == ASHIFT)
8862 || (first_code == ASHIFTRT && code == LSHIFTRT)))
8863 code = first_code;
8864 else if (count < 0)
8865 code = first_code, count = -count;
8867 varop = XEXP (varop, 0);
8868 continue;
8871 /* If we have (A << B << C) for any shift, we can convert this to
8872 (A << C << B). This wins if A is a constant. Only try this if
8873 B is not a constant. */
8875 else if (GET_CODE (varop) == code
8876 && GET_CODE (XEXP (varop, 0)) == CONST_INT
8877 && GET_CODE (XEXP (varop, 1)) != CONST_INT)
8879 rtx new = simplify_const_binary_operation (code, mode,
8880 XEXP (varop, 0),
8881 GEN_INT (count));
8882 varop = gen_rtx_fmt_ee (code, mode, new, XEXP (varop, 1));
8883 count = 0;
8884 continue;
8886 break;
8888 case NOT:
8889 /* Make this fit the case below. */
8890 varop = gen_rtx_XOR (mode, XEXP (varop, 0),
8891 GEN_INT (GET_MODE_MASK (mode)));
8892 continue;
8894 case IOR:
8895 case AND:
8896 case XOR:
8897 /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C)
8898 with C the size of VAROP - 1 and the shift is logical if
8899 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
8900 we have an (le X 0) operation. If we have an arithmetic shift
8901 and STORE_FLAG_VALUE is 1 or we have a logical shift with
8902 STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */
8904 if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS
8905 && XEXP (XEXP (varop, 0), 1) == constm1_rtx
8906 && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
8907 && (code == LSHIFTRT || code == ASHIFTRT)
8908 && count == (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
8909 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
8911 count = 0;
8912 varop = gen_rtx_LE (GET_MODE (varop), XEXP (varop, 1),
8913 const0_rtx);
8915 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
8916 varop = gen_rtx_NEG (GET_MODE (varop), varop);
8918 continue;
8921 /* If we have (shift (logical)), move the logical to the outside
8922 to allow it to possibly combine with another logical and the
8923 shift to combine with another shift. This also canonicalizes to
8924 what a ZERO_EXTRACT looks like. Also, some machines have
8925 (and (shift)) insns. */
8927 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8928 /* We can't do this if we have (ashiftrt (xor)) and the
8929 constant has its sign bit set in shift_mode. */
8930 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
8931 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
8932 shift_mode))
8933 && (new = simplify_const_binary_operation (code, result_mode,
8934 XEXP (varop, 1),
8935 GEN_INT (count))) != 0
8936 && GET_CODE (new) == CONST_INT
8937 && merge_outer_ops (&outer_op, &outer_const, GET_CODE (varop),
8938 INTVAL (new), result_mode, &complement_p))
8940 varop = XEXP (varop, 0);
8941 continue;
8944 /* If we can't do that, try to simplify the shift in each arm of the
8945 logical expression, make a new logical expression, and apply
8946 the inverse distributive law. This also can't be done
8947 for some (ashiftrt (xor)). */
8948 if (GET_CODE (XEXP (varop, 1)) == CONST_INT
8949 && !(code == ASHIFTRT && GET_CODE (varop) == XOR
8950 && 0 > trunc_int_for_mode (INTVAL (XEXP (varop, 1)),
8951 shift_mode)))
8953 rtx lhs = simplify_shift_const (NULL_RTX, code, shift_mode,
8954 XEXP (varop, 0), count);
8955 rtx rhs = simplify_shift_const (NULL_RTX, code, shift_mode,
8956 XEXP (varop, 1), count);
8958 varop = simplify_gen_binary (GET_CODE (varop), shift_mode,
8959 lhs, rhs);
8960 varop = apply_distributive_law (varop);
8962 count = 0;
8963 continue;
8965 break;
8967 case EQ:
8968 /* Convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE
8969 says that the sign bit can be tested, FOO has mode MODE, C is
8970 GET_MODE_BITSIZE (MODE) - 1, and FOO has only its low-order bit
8971 that may be nonzero. */
8972 if (code == LSHIFTRT
8973 && XEXP (varop, 1) == const0_rtx
8974 && GET_MODE (XEXP (varop, 0)) == result_mode
8975 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8976 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
8977 && STORE_FLAG_VALUE == -1
8978 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
8979 && merge_outer_ops (&outer_op, &outer_const, XOR,
8980 (HOST_WIDE_INT) 1, result_mode,
8981 &complement_p))
8983 varop = XEXP (varop, 0);
8984 count = 0;
8985 continue;
8987 break;
8989 case NEG:
8990 /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less
8991 than the number of bits in the mode is equivalent to A. */
8992 if (code == LSHIFTRT
8993 && count == (GET_MODE_BITSIZE (result_mode) - 1)
8994 && nonzero_bits (XEXP (varop, 0), result_mode) == 1)
8996 varop = XEXP (varop, 0);
8997 count = 0;
8998 continue;
9001 /* NEG commutes with ASHIFT since it is multiplication. Move the
9002 NEG outside to allow shifts to combine. */
9003 if (code == ASHIFT
9004 && merge_outer_ops (&outer_op, &outer_const, NEG,
9005 (HOST_WIDE_INT) 0, result_mode,
9006 &complement_p))
9008 varop = XEXP (varop, 0);
9009 continue;
9011 break;
9013 case PLUS:
9014 /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C
9015 is one less than the number of bits in the mode is
9016 equivalent to (xor A 1). */
9017 if (code == LSHIFTRT
9018 && count == (GET_MODE_BITSIZE (result_mode) - 1)
9019 && XEXP (varop, 1) == constm1_rtx
9020 && nonzero_bits (XEXP (varop, 0), result_mode) == 1
9021 && merge_outer_ops (&outer_op, &outer_const, XOR,
9022 (HOST_WIDE_INT) 1, result_mode,
9023 &complement_p))
9025 count = 0;
9026 varop = XEXP (varop, 0);
9027 continue;
9030 /* If we have (xshiftrt (plus FOO BAR) C), and the only bits
9031 that might be nonzero in BAR are those being shifted out and those
9032 bits are known zero in FOO, we can replace the PLUS with FOO.
9033 Similarly in the other operand order. This code occurs when
9034 we are computing the size of a variable-size array. */
9036 if ((code == ASHIFTRT || code == LSHIFTRT)
9037 && count < HOST_BITS_PER_WIDE_INT
9038 && nonzero_bits (XEXP (varop, 1), result_mode) >> count == 0
9039 && (nonzero_bits (XEXP (varop, 1), result_mode)
9040 & nonzero_bits (XEXP (varop, 0), result_mode)) == 0)
9042 varop = XEXP (varop, 0);
9043 continue;
9045 else if ((code == ASHIFTRT || code == LSHIFTRT)
9046 && count < HOST_BITS_PER_WIDE_INT
9047 && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
9048 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9049 >> count)
9050 && 0 == (nonzero_bits (XEXP (varop, 0), result_mode)
9051 & nonzero_bits (XEXP (varop, 1),
9052 result_mode)))
9054 varop = XEXP (varop, 1);
9055 continue;
9058 /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */
9059 if (code == ASHIFT
9060 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9061 && (new = simplify_const_binary_operation (ASHIFT, result_mode,
9062 XEXP (varop, 1),
9063 GEN_INT (count))) != 0
9064 && GET_CODE (new) == CONST_INT
9065 && merge_outer_ops (&outer_op, &outer_const, PLUS,
9066 INTVAL (new), result_mode, &complement_p))
9068 varop = XEXP (varop, 0);
9069 continue;
9072 /* Check for 'PLUS signbit', which is the canonical form of 'XOR
9073 signbit', and attempt to change the PLUS to an XOR and move it to
9074 the outer operation as is done above in the AND/IOR/XOR case
9075 leg for shift(logical). See details in logical handling above
9076 for reasoning in doing so. */
9077 if (code == LSHIFTRT
9078 && GET_CODE (XEXP (varop, 1)) == CONST_INT
9079 && mode_signbit_p (result_mode, XEXP (varop, 1))
9080 && (new = simplify_const_binary_operation (code, result_mode,
9081 XEXP (varop, 1),
9082 GEN_INT (count))) != 0
9083 && GET_CODE (new) == CONST_INT
9084 && merge_outer_ops (&outer_op, &outer_const, XOR,
9085 INTVAL (new), result_mode, &complement_p))
9087 varop = XEXP (varop, 0);
9088 continue;
9091 break;
9093 case MINUS:
9094 /* If we have (xshiftrt (minus (ashiftrt X C)) X) C)
9095 with C the size of VAROP - 1 and the shift is logical if
9096 STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1,
9097 we have a (gt X 0) operation. If the shift is arithmetic with
9098 STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1,
9099 we have a (neg (gt X 0)) operation. */
9101 if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1)
9102 && GET_CODE (XEXP (varop, 0)) == ASHIFTRT
9103 && count == (GET_MODE_BITSIZE (GET_MODE (varop)) - 1)
9104 && (code == LSHIFTRT || code == ASHIFTRT)
9105 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9106 && INTVAL (XEXP (XEXP (varop, 0), 1)) == count
9107 && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1)))
9109 count = 0;
9110 varop = gen_rtx_GT (GET_MODE (varop), XEXP (varop, 1),
9111 const0_rtx);
9113 if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT)
9114 varop = gen_rtx_NEG (GET_MODE (varop), varop);
9116 continue;
9118 break;
9120 case TRUNCATE:
9121 /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt))
9122 if the truncate does not affect the value. */
9123 if (code == LSHIFTRT
9124 && GET_CODE (XEXP (varop, 0)) == LSHIFTRT
9125 && GET_CODE (XEXP (XEXP (varop, 0), 1)) == CONST_INT
9126 && (INTVAL (XEXP (XEXP (varop, 0), 1))
9127 >= (GET_MODE_BITSIZE (GET_MODE (XEXP (varop, 0)))
9128 - GET_MODE_BITSIZE (GET_MODE (varop)))))
9130 rtx varop_inner = XEXP (varop, 0);
9132 varop_inner
9133 = gen_rtx_LSHIFTRT (GET_MODE (varop_inner),
9134 XEXP (varop_inner, 0),
9135 GEN_INT
9136 (count + INTVAL (XEXP (varop_inner, 1))));
9137 varop = gen_rtx_TRUNCATE (GET_MODE (varop), varop_inner);
9138 count = 0;
9139 continue;
9141 break;
9143 default:
9144 break;
9147 break;
9150 /* We need to determine what mode to do the shift in. If the shift is
9151 a right shift or ROTATE, we must always do it in the mode it was
9152 originally done in. Otherwise, we can do it in MODE, the widest mode
9153 encountered. The code we care about is that of the shift that will
9154 actually be done, not the shift that was originally requested. */
9155 shift_mode
9156 = (code == ASHIFTRT || code == LSHIFTRT || code == ROTATE
9157 ? result_mode : mode);
9159 /* We have now finished analyzing the shift. The result should be
9160 a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If
9161 OUTER_OP is non-UNKNOWN, it is an operation that needs to be applied
9162 to the result of the shift. OUTER_CONST is the relevant constant,
9163 but we must turn off all bits turned off in the shift. */
9165 if (outer_op == UNKNOWN
9166 && orig_code == code && orig_count == count
9167 && varop == orig_varop
9168 && shift_mode == GET_MODE (varop))
9169 return NULL_RTX;
9171 /* Make a SUBREG if necessary. If we can't make it, fail. */
9172 varop = gen_lowpart (shift_mode, varop);
9173 if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER)
9174 return NULL_RTX;
9176 /* If we have an outer operation and we just made a shift, it is
9177 possible that we could have simplified the shift were it not
9178 for the outer operation. So try to do the simplification
9179 recursively. */
9181 if (outer_op != UNKNOWN)
9182 x = simplify_shift_const_1 (code, shift_mode, varop, count);
9183 else
9184 x = NULL_RTX;
9186 if (x == NULL_RTX)
9187 x = simplify_gen_binary (code, shift_mode, varop, GEN_INT (count));
9189 /* If we were doing an LSHIFTRT in a wider mode than it was originally,
9190 turn off all the bits that the shift would have turned off. */
9191 if (orig_code == LSHIFTRT && result_mode != shift_mode)
9192 x = simplify_and_const_int (NULL_RTX, shift_mode, x,
9193 GET_MODE_MASK (result_mode) >> orig_count);
9195 /* Do the remainder of the processing in RESULT_MODE. */
9196 x = gen_lowpart_or_truncate (result_mode, x);
9198 /* If COMPLEMENT_P is set, we have to complement X before doing the outer
9199 operation. */
9200 if (complement_p)
9201 x = simplify_gen_unary (NOT, result_mode, x, result_mode);
9203 if (outer_op != UNKNOWN)
9205 if (GET_MODE_BITSIZE (result_mode) < HOST_BITS_PER_WIDE_INT)
9206 outer_const = trunc_int_for_mode (outer_const, result_mode);
9208 if (outer_op == AND)
9209 x = simplify_and_const_int (NULL_RTX, result_mode, x, outer_const);
9210 else if (outer_op == SET)
9211 /* This means that we have determined that the result is
9212 equivalent to a constant. This should be rare. */
9213 x = GEN_INT (outer_const);
9214 else if (GET_RTX_CLASS (outer_op) == RTX_UNARY)
9215 x = simplify_gen_unary (outer_op, result_mode, x, result_mode);
9216 else
9217 x = simplify_gen_binary (outer_op, result_mode, x,
9218 GEN_INT (outer_const));
9221 return x;
9224 /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift.
9225 The result of the shift is RESULT_MODE. If we cannot simplify it,
9226 return X or, if it is NULL, synthesize the expression with
9227 simplify_gen_binary. Otherwise, return a simplified value.
9229 The shift is normally computed in the widest mode we find in VAROP, as
9230 long as it isn't a different number of words than RESULT_MODE. Exceptions
9231 are ASHIFTRT and ROTATE, which are always done in their original mode. */
9233 static rtx
9234 simplify_shift_const (rtx x, enum rtx_code code, enum machine_mode result_mode,
9235 rtx varop, int count)
9237 rtx tem = simplify_shift_const_1 (code, result_mode, varop, count);
9238 if (tem)
9239 return tem;
9241 if (!x)
9242 x = simplify_gen_binary (code, GET_MODE (varop), varop, GEN_INT (count));
9243 if (GET_MODE (x) != result_mode)
9244 x = gen_lowpart (result_mode, x);
9245 return x;
9249 /* Like recog, but we receive the address of a pointer to a new pattern.
9250 We try to match the rtx that the pointer points to.
9251 If that fails, we may try to modify or replace the pattern,
9252 storing the replacement into the same pointer object.
9254 Modifications include deletion or addition of CLOBBERs.
9256 PNOTES is a pointer to a location where any REG_UNUSED notes added for
9257 the CLOBBERs are placed.
9259 The value is the final insn code from the pattern ultimately matched,
9260 or -1. */
9262 static int
9263 recog_for_combine (rtx *pnewpat, rtx insn, rtx *pnotes)
9265 rtx pat = *pnewpat;
9266 int insn_code_number;
9267 int num_clobbers_to_add = 0;
9268 int i;
9269 rtx notes = 0;
9270 rtx old_notes, old_pat;
9272 /* If PAT is a PARALLEL, check to see if it contains the CLOBBER
9273 we use to indicate that something didn't match. If we find such a
9274 thing, force rejection. */
9275 if (GET_CODE (pat) == PARALLEL)
9276 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
9277 if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER
9278 && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx)
9279 return -1;
9281 old_pat = PATTERN (insn);
9282 old_notes = REG_NOTES (insn);
9283 PATTERN (insn) = pat;
9284 REG_NOTES (insn) = 0;
9286 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9288 /* If it isn't, there is the possibility that we previously had an insn
9289 that clobbered some register as a side effect, but the combined
9290 insn doesn't need to do that. So try once more without the clobbers
9291 unless this represents an ASM insn. */
9293 if (insn_code_number < 0 && ! check_asm_operands (pat)
9294 && GET_CODE (pat) == PARALLEL)
9296 int pos;
9298 for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++)
9299 if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER)
9301 if (i != pos)
9302 SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i));
9303 pos++;
9306 SUBST_INT (XVECLEN (pat, 0), pos);
9308 if (pos == 1)
9309 pat = XVECEXP (pat, 0, 0);
9311 PATTERN (insn) = pat;
9312 insn_code_number = recog (pat, insn, &num_clobbers_to_add);
9314 PATTERN (insn) = old_pat;
9315 REG_NOTES (insn) = old_notes;
9317 /* Recognize all noop sets, these will be killed by followup pass. */
9318 if (insn_code_number < 0 && GET_CODE (pat) == SET && set_noop_p (pat))
9319 insn_code_number = NOOP_MOVE_INSN_CODE, num_clobbers_to_add = 0;
9321 /* If we had any clobbers to add, make a new pattern than contains
9322 them. Then check to make sure that all of them are dead. */
9323 if (num_clobbers_to_add)
9325 rtx newpat = gen_rtx_PARALLEL (VOIDmode,
9326 rtvec_alloc (GET_CODE (pat) == PARALLEL
9327 ? (XVECLEN (pat, 0)
9328 + num_clobbers_to_add)
9329 : num_clobbers_to_add + 1));
9331 if (GET_CODE (pat) == PARALLEL)
9332 for (i = 0; i < XVECLEN (pat, 0); i++)
9333 XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i);
9334 else
9335 XVECEXP (newpat, 0, 0) = pat;
9337 add_clobbers (newpat, insn_code_number);
9339 for (i = XVECLEN (newpat, 0) - num_clobbers_to_add;
9340 i < XVECLEN (newpat, 0); i++)
9342 if (REG_P (XEXP (XVECEXP (newpat, 0, i), 0))
9343 && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn))
9344 return -1;
9345 notes = gen_rtx_EXPR_LIST (REG_UNUSED,
9346 XEXP (XVECEXP (newpat, 0, i), 0), notes);
9348 pat = newpat;
9351 *pnewpat = pat;
9352 *pnotes = notes;
9354 return insn_code_number;
9357 /* Like gen_lowpart_general but for use by combine. In combine it
9358 is not possible to create any new pseudoregs. However, it is
9359 safe to create invalid memory addresses, because combine will
9360 try to recognize them and all they will do is make the combine
9361 attempt fail.
9363 If for some reason this cannot do its job, an rtx
9364 (clobber (const_int 0)) is returned.
9365 An insn containing that will not be recognized. */
9367 static rtx
9368 gen_lowpart_for_combine (enum machine_mode omode, rtx x)
9370 enum machine_mode imode = GET_MODE (x);
9371 unsigned int osize = GET_MODE_SIZE (omode);
9372 unsigned int isize = GET_MODE_SIZE (imode);
9373 rtx result;
9375 if (omode == imode)
9376 return x;
9378 /* Return identity if this is a CONST or symbolic reference. */
9379 if (omode == Pmode
9380 && (GET_CODE (x) == CONST
9381 || GET_CODE (x) == SYMBOL_REF
9382 || GET_CODE (x) == LABEL_REF))
9383 return x;
9385 /* We can only support MODE being wider than a word if X is a
9386 constant integer or has a mode the same size. */
9387 if (GET_MODE_SIZE (omode) > UNITS_PER_WORD
9388 && ! ((imode == VOIDmode
9389 && (GET_CODE (x) == CONST_INT
9390 || GET_CODE (x) == CONST_DOUBLE))
9391 || isize == osize))
9392 goto fail;
9394 /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart
9395 won't know what to do. So we will strip off the SUBREG here and
9396 process normally. */
9397 if (GET_CODE (x) == SUBREG && MEM_P (SUBREG_REG (x)))
9399 x = SUBREG_REG (x);
9401 /* For use in case we fall down into the address adjustments
9402 further below, we need to adjust the known mode and size of
9403 x; imode and isize, since we just adjusted x. */
9404 imode = GET_MODE (x);
9406 if (imode == omode)
9407 return x;
9409 isize = GET_MODE_SIZE (imode);
9412 result = gen_lowpart_common (omode, x);
9414 #ifdef CANNOT_CHANGE_MODE_CLASS
9415 if (result != 0 && GET_CODE (result) == SUBREG)
9416 record_subregs_of_mode (result);
9417 #endif
9419 if (result)
9420 return result;
9422 if (MEM_P (x))
9424 int offset = 0;
9426 /* Refuse to work on a volatile memory ref or one with a mode-dependent
9427 address. */
9428 if (MEM_VOLATILE_P (x) || mode_dependent_address_p (XEXP (x, 0)))
9429 goto fail;
9431 /* If we want to refer to something bigger than the original memref,
9432 generate a paradoxical subreg instead. That will force a reload
9433 of the original memref X. */
9434 if (isize < osize)
9435 return gen_rtx_SUBREG (omode, x, 0);
9437 if (WORDS_BIG_ENDIAN)
9438 offset = MAX (isize, UNITS_PER_WORD) - MAX (osize, UNITS_PER_WORD);
9440 /* Adjust the address so that the address-after-the-data is
9441 unchanged. */
9442 if (BYTES_BIG_ENDIAN)
9443 offset -= MIN (UNITS_PER_WORD, osize) - MIN (UNITS_PER_WORD, isize);
9445 return adjust_address_nv (x, omode, offset);
9448 /* If X is a comparison operator, rewrite it in a new mode. This
9449 probably won't match, but may allow further simplifications. */
9450 else if (COMPARISON_P (x))
9451 return gen_rtx_fmt_ee (GET_CODE (x), omode, XEXP (x, 0), XEXP (x, 1));
9453 /* If we couldn't simplify X any other way, just enclose it in a
9454 SUBREG. Normally, this SUBREG won't match, but some patterns may
9455 include an explicit SUBREG or we may simplify it further in combine. */
9456 else
9458 int offset = 0;
9459 rtx res;
9461 offset = subreg_lowpart_offset (omode, imode);
9462 if (imode == VOIDmode)
9464 imode = int_mode_for_mode (omode);
9465 x = gen_lowpart_common (imode, x);
9466 if (x == NULL)
9467 goto fail;
9469 res = simplify_gen_subreg (omode, x, imode, offset);
9470 if (res)
9471 return res;
9474 fail:
9475 return gen_rtx_CLOBBER (imode, const0_rtx);
9478 /* Simplify a comparison between *POP0 and *POP1 where CODE is the
9479 comparison code that will be tested.
9481 The result is a possibly different comparison code to use. *POP0 and
9482 *POP1 may be updated.
9484 It is possible that we might detect that a comparison is either always
9485 true or always false. However, we do not perform general constant
9486 folding in combine, so this knowledge isn't useful. Such tautologies
9487 should have been detected earlier. Hence we ignore all such cases. */
9489 static enum rtx_code
9490 simplify_comparison (enum rtx_code code, rtx *pop0, rtx *pop1)
9492 rtx op0 = *pop0;
9493 rtx op1 = *pop1;
9494 rtx tem, tem1;
9495 int i;
9496 enum machine_mode mode, tmode;
9498 /* Try a few ways of applying the same transformation to both operands. */
9499 while (1)
9501 #ifndef WORD_REGISTER_OPERATIONS
9502 /* The test below this one won't handle SIGN_EXTENDs on these machines,
9503 so check specially. */
9504 if (code != GTU && code != GEU && code != LTU && code != LEU
9505 && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT
9506 && GET_CODE (XEXP (op0, 0)) == ASHIFT
9507 && GET_CODE (XEXP (op1, 0)) == ASHIFT
9508 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG
9509 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG
9510 && (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0)))
9511 == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0))))
9512 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9513 && XEXP (op0, 1) == XEXP (op1, 1)
9514 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
9515 && XEXP (op0, 1) == XEXP (XEXP (op1, 0), 1)
9516 && (INTVAL (XEXP (op0, 1))
9517 == (GET_MODE_BITSIZE (GET_MODE (op0))
9518 - (GET_MODE_BITSIZE
9519 (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))))))))
9521 op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0));
9522 op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0));
9524 #endif
9526 /* If both operands are the same constant shift, see if we can ignore the
9527 shift. We can if the shift is a rotate or if the bits shifted out of
9528 this shift are known to be zero for both inputs and if the type of
9529 comparison is compatible with the shift. */
9530 if (GET_CODE (op0) == GET_CODE (op1)
9531 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
9532 && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ))
9533 || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT)
9534 && (code != GT && code != LT && code != GE && code != LE))
9535 || (GET_CODE (op0) == ASHIFTRT
9536 && (code != GTU && code != LTU
9537 && code != GEU && code != LEU)))
9538 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9539 && INTVAL (XEXP (op0, 1)) >= 0
9540 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
9541 && XEXP (op0, 1) == XEXP (op1, 1))
9543 enum machine_mode mode = GET_MODE (op0);
9544 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9545 int shift_count = INTVAL (XEXP (op0, 1));
9547 if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT)
9548 mask &= (mask >> shift_count) << shift_count;
9549 else if (GET_CODE (op0) == ASHIFT)
9550 mask = (mask & (mask << shift_count)) >> shift_count;
9552 if ((nonzero_bits (XEXP (op0, 0), mode) & ~mask) == 0
9553 && (nonzero_bits (XEXP (op1, 0), mode) & ~mask) == 0)
9554 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0);
9555 else
9556 break;
9559 /* If both operands are AND's of a paradoxical SUBREG by constant, the
9560 SUBREGs are of the same mode, and, in both cases, the AND would
9561 be redundant if the comparison was done in the narrower mode,
9562 do the comparison in the narrower mode (e.g., we are AND'ing with 1
9563 and the operand's possibly nonzero bits are 0xffffff01; in that case
9564 if we only care about QImode, we don't need the AND). This case
9565 occurs if the output mode of an scc insn is not SImode and
9566 STORE_FLAG_VALUE == 1 (e.g., the 386).
9568 Similarly, check for a case where the AND's are ZERO_EXTEND
9569 operations from some narrower mode even though a SUBREG is not
9570 present. */
9572 else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND
9573 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9574 && GET_CODE (XEXP (op1, 1)) == CONST_INT)
9576 rtx inner_op0 = XEXP (op0, 0);
9577 rtx inner_op1 = XEXP (op1, 0);
9578 HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1));
9579 HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1));
9580 int changed = 0;
9582 if (GET_CODE (inner_op0) == SUBREG && GET_CODE (inner_op1) == SUBREG
9583 && (GET_MODE_SIZE (GET_MODE (inner_op0))
9584 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner_op0))))
9585 && (GET_MODE (SUBREG_REG (inner_op0))
9586 == GET_MODE (SUBREG_REG (inner_op1)))
9587 && (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (inner_op0)))
9588 <= HOST_BITS_PER_WIDE_INT)
9589 && (0 == ((~c0) & nonzero_bits (SUBREG_REG (inner_op0),
9590 GET_MODE (SUBREG_REG (inner_op0)))))
9591 && (0 == ((~c1) & nonzero_bits (SUBREG_REG (inner_op1),
9592 GET_MODE (SUBREG_REG (inner_op1))))))
9594 op0 = SUBREG_REG (inner_op0);
9595 op1 = SUBREG_REG (inner_op1);
9597 /* The resulting comparison is always unsigned since we masked
9598 off the original sign bit. */
9599 code = unsigned_condition (code);
9601 changed = 1;
9604 else if (c0 == c1)
9605 for (tmode = GET_CLASS_NARROWEST_MODE
9606 (GET_MODE_CLASS (GET_MODE (op0)));
9607 tmode != GET_MODE (op0); tmode = GET_MODE_WIDER_MODE (tmode))
9608 if ((unsigned HOST_WIDE_INT) c0 == GET_MODE_MASK (tmode))
9610 op0 = gen_lowpart (tmode, inner_op0);
9611 op1 = gen_lowpart (tmode, inner_op1);
9612 code = unsigned_condition (code);
9613 changed = 1;
9614 break;
9617 if (! changed)
9618 break;
9621 /* If both operands are NOT, we can strip off the outer operation
9622 and adjust the comparison code for swapped operands; similarly for
9623 NEG, except that this must be an equality comparison. */
9624 else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT)
9625 || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG
9626 && (code == EQ || code == NE)))
9627 op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code);
9629 else
9630 break;
9633 /* If the first operand is a constant, swap the operands and adjust the
9634 comparison code appropriately, but don't do this if the second operand
9635 is already a constant integer. */
9636 if (swap_commutative_operands_p (op0, op1))
9638 tem = op0, op0 = op1, op1 = tem;
9639 code = swap_condition (code);
9642 /* We now enter a loop during which we will try to simplify the comparison.
9643 For the most part, we only are concerned with comparisons with zero,
9644 but some things may really be comparisons with zero but not start
9645 out looking that way. */
9647 while (GET_CODE (op1) == CONST_INT)
9649 enum machine_mode mode = GET_MODE (op0);
9650 unsigned int mode_width = GET_MODE_BITSIZE (mode);
9651 unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode);
9652 int equality_comparison_p;
9653 int sign_bit_comparison_p;
9654 int unsigned_comparison_p;
9655 HOST_WIDE_INT const_op;
9657 /* We only want to handle integral modes. This catches VOIDmode,
9658 CCmode, and the floating-point modes. An exception is that we
9659 can handle VOIDmode if OP0 is a COMPARE or a comparison
9660 operation. */
9662 if (GET_MODE_CLASS (mode) != MODE_INT
9663 && ! (mode == VOIDmode
9664 && (GET_CODE (op0) == COMPARE || COMPARISON_P (op0))))
9665 break;
9667 /* Get the constant we are comparing against and turn off all bits
9668 not on in our mode. */
9669 const_op = INTVAL (op1);
9670 if (mode != VOIDmode)
9671 const_op = trunc_int_for_mode (const_op, mode);
9672 op1 = GEN_INT (const_op);
9674 /* If we are comparing against a constant power of two and the value
9675 being compared can only have that single bit nonzero (e.g., it was
9676 `and'ed with that bit), we can replace this with a comparison
9677 with zero. */
9678 if (const_op
9679 && (code == EQ || code == NE || code == GE || code == GEU
9680 || code == LT || code == LTU)
9681 && mode_width <= HOST_BITS_PER_WIDE_INT
9682 && exact_log2 (const_op) >= 0
9683 && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) const_op)
9685 code = (code == EQ || code == GE || code == GEU ? NE : EQ);
9686 op1 = const0_rtx, const_op = 0;
9689 /* Similarly, if we are comparing a value known to be either -1 or
9690 0 with -1, change it to the opposite comparison against zero. */
9692 if (const_op == -1
9693 && (code == EQ || code == NE || code == GT || code == LE
9694 || code == GEU || code == LTU)
9695 && num_sign_bit_copies (op0, mode) == mode_width)
9697 code = (code == EQ || code == LE || code == GEU ? NE : EQ);
9698 op1 = const0_rtx, const_op = 0;
9701 /* Do some canonicalizations based on the comparison code. We prefer
9702 comparisons against zero and then prefer equality comparisons.
9703 If we can reduce the size of a constant, we will do that too. */
9705 switch (code)
9707 case LT:
9708 /* < C is equivalent to <= (C - 1) */
9709 if (const_op > 0)
9711 const_op -= 1;
9712 op1 = GEN_INT (const_op);
9713 code = LE;
9714 /* ... fall through to LE case below. */
9716 else
9717 break;
9719 case LE:
9720 /* <= C is equivalent to < (C + 1); we do this for C < 0 */
9721 if (const_op < 0)
9723 const_op += 1;
9724 op1 = GEN_INT (const_op);
9725 code = LT;
9728 /* If we are doing a <= 0 comparison on a value known to have
9729 a zero sign bit, we can replace this with == 0. */
9730 else if (const_op == 0
9731 && mode_width <= HOST_BITS_PER_WIDE_INT
9732 && (nonzero_bits (op0, mode)
9733 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9734 code = EQ;
9735 break;
9737 case GE:
9738 /* >= C is equivalent to > (C - 1). */
9739 if (const_op > 0)
9741 const_op -= 1;
9742 op1 = GEN_INT (const_op);
9743 code = GT;
9744 /* ... fall through to GT below. */
9746 else
9747 break;
9749 case GT:
9750 /* > C is equivalent to >= (C + 1); we do this for C < 0. */
9751 if (const_op < 0)
9753 const_op += 1;
9754 op1 = GEN_INT (const_op);
9755 code = GE;
9758 /* If we are doing a > 0 comparison on a value known to have
9759 a zero sign bit, we can replace this with != 0. */
9760 else if (const_op == 0
9761 && mode_width <= HOST_BITS_PER_WIDE_INT
9762 && (nonzero_bits (op0, mode)
9763 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)
9764 code = NE;
9765 break;
9767 case LTU:
9768 /* < C is equivalent to <= (C - 1). */
9769 if (const_op > 0)
9771 const_op -= 1;
9772 op1 = GEN_INT (const_op);
9773 code = LEU;
9774 /* ... fall through ... */
9777 /* (unsigned) < 0x80000000 is equivalent to >= 0. */
9778 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9779 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9781 const_op = 0, op1 = const0_rtx;
9782 code = GE;
9783 break;
9785 else
9786 break;
9788 case LEU:
9789 /* unsigned <= 0 is equivalent to == 0 */
9790 if (const_op == 0)
9791 code = EQ;
9793 /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */
9794 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9795 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9797 const_op = 0, op1 = const0_rtx;
9798 code = GE;
9800 break;
9802 case GEU:
9803 /* >= C is equivalent to > (C - 1). */
9804 if (const_op > 1)
9806 const_op -= 1;
9807 op1 = GEN_INT (const_op);
9808 code = GTU;
9809 /* ... fall through ... */
9812 /* (unsigned) >= 0x80000000 is equivalent to < 0. */
9813 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9814 && (const_op == (HOST_WIDE_INT) 1 << (mode_width - 1)))
9816 const_op = 0, op1 = const0_rtx;
9817 code = LT;
9818 break;
9820 else
9821 break;
9823 case GTU:
9824 /* unsigned > 0 is equivalent to != 0 */
9825 if (const_op == 0)
9826 code = NE;
9828 /* (unsigned) > 0x7fffffff is equivalent to < 0. */
9829 else if ((mode_width <= HOST_BITS_PER_WIDE_INT)
9830 && (const_op == ((HOST_WIDE_INT) 1 << (mode_width - 1)) - 1))
9832 const_op = 0, op1 = const0_rtx;
9833 code = LT;
9835 break;
9837 default:
9838 break;
9841 /* Compute some predicates to simplify code below. */
9843 equality_comparison_p = (code == EQ || code == NE);
9844 sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0);
9845 unsigned_comparison_p = (code == LTU || code == LEU || code == GTU
9846 || code == GEU);
9848 /* If this is a sign bit comparison and we can do arithmetic in
9849 MODE, say that we will only be needing the sign bit of OP0. */
9850 if (sign_bit_comparison_p
9851 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
9852 op0 = force_to_mode (op0, mode,
9853 ((HOST_WIDE_INT) 1
9854 << (GET_MODE_BITSIZE (mode) - 1)),
9857 /* Now try cases based on the opcode of OP0. If none of the cases
9858 does a "continue", we exit this loop immediately after the
9859 switch. */
9861 switch (GET_CODE (op0))
9863 case ZERO_EXTRACT:
9864 /* If we are extracting a single bit from a variable position in
9865 a constant that has only a single bit set and are comparing it
9866 with zero, we can convert this into an equality comparison
9867 between the position and the location of the single bit. */
9868 /* Except we can't if SHIFT_COUNT_TRUNCATED is set, since we might
9869 have already reduced the shift count modulo the word size. */
9870 if (!SHIFT_COUNT_TRUNCATED
9871 && GET_CODE (XEXP (op0, 0)) == CONST_INT
9872 && XEXP (op0, 1) == const1_rtx
9873 && equality_comparison_p && const_op == 0
9874 && (i = exact_log2 (INTVAL (XEXP (op0, 0)))) >= 0)
9876 if (BITS_BIG_ENDIAN)
9878 enum machine_mode new_mode
9879 = mode_for_extraction (EP_extzv, 1);
9880 if (new_mode == MAX_MACHINE_MODE)
9881 i = BITS_PER_WORD - 1 - i;
9882 else
9884 mode = new_mode;
9885 i = (GET_MODE_BITSIZE (mode) - 1 - i);
9889 op0 = XEXP (op0, 2);
9890 op1 = GEN_INT (i);
9891 const_op = i;
9893 /* Result is nonzero iff shift count is equal to I. */
9894 code = reverse_condition (code);
9895 continue;
9898 /* ... fall through ... */
9900 case SIGN_EXTRACT:
9901 tem = expand_compound_operation (op0);
9902 if (tem != op0)
9904 op0 = tem;
9905 continue;
9907 break;
9909 case NOT:
9910 /* If testing for equality, we can take the NOT of the constant. */
9911 if (equality_comparison_p
9912 && (tem = simplify_unary_operation (NOT, mode, op1, mode)) != 0)
9914 op0 = XEXP (op0, 0);
9915 op1 = tem;
9916 continue;
9919 /* If just looking at the sign bit, reverse the sense of the
9920 comparison. */
9921 if (sign_bit_comparison_p)
9923 op0 = XEXP (op0, 0);
9924 code = (code == GE ? LT : GE);
9925 continue;
9927 break;
9929 case NEG:
9930 /* If testing for equality, we can take the NEG of the constant. */
9931 if (equality_comparison_p
9932 && (tem = simplify_unary_operation (NEG, mode, op1, mode)) != 0)
9934 op0 = XEXP (op0, 0);
9935 op1 = tem;
9936 continue;
9939 /* The remaining cases only apply to comparisons with zero. */
9940 if (const_op != 0)
9941 break;
9943 /* When X is ABS or is known positive,
9944 (neg X) is < 0 if and only if X != 0. */
9946 if (sign_bit_comparison_p
9947 && (GET_CODE (XEXP (op0, 0)) == ABS
9948 || (mode_width <= HOST_BITS_PER_WIDE_INT
9949 && (nonzero_bits (XEXP (op0, 0), mode)
9950 & ((HOST_WIDE_INT) 1 << (mode_width - 1))) == 0)))
9952 op0 = XEXP (op0, 0);
9953 code = (code == LT ? NE : EQ);
9954 continue;
9957 /* If we have NEG of something whose two high-order bits are the
9958 same, we know that "(-a) < 0" is equivalent to "a > 0". */
9959 if (num_sign_bit_copies (op0, mode) >= 2)
9961 op0 = XEXP (op0, 0);
9962 code = swap_condition (code);
9963 continue;
9965 break;
9967 case ROTATE:
9968 /* If we are testing equality and our count is a constant, we
9969 can perform the inverse operation on our RHS. */
9970 if (equality_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
9971 && (tem = simplify_binary_operation (ROTATERT, mode,
9972 op1, XEXP (op0, 1))) != 0)
9974 op0 = XEXP (op0, 0);
9975 op1 = tem;
9976 continue;
9979 /* If we are doing a < 0 or >= 0 comparison, it means we are testing
9980 a particular bit. Convert it to an AND of a constant of that
9981 bit. This will be converted into a ZERO_EXTRACT. */
9982 if (const_op == 0 && sign_bit_comparison_p
9983 && GET_CODE (XEXP (op0, 1)) == CONST_INT
9984 && mode_width <= HOST_BITS_PER_WIDE_INT)
9986 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
9987 ((HOST_WIDE_INT) 1
9988 << (mode_width - 1
9989 - INTVAL (XEXP (op0, 1)))));
9990 code = (code == LT ? NE : EQ);
9991 continue;
9994 /* Fall through. */
9996 case ABS:
9997 /* ABS is ignorable inside an equality comparison with zero. */
9998 if (const_op == 0 && equality_comparison_p)
10000 op0 = XEXP (op0, 0);
10001 continue;
10003 break;
10005 case SIGN_EXTEND:
10006 /* Can simplify (compare (zero/sign_extend FOO) CONST) to
10007 (compare FOO CONST) if CONST fits in FOO's mode and we
10008 are either testing inequality or have an unsigned
10009 comparison with ZERO_EXTEND or a signed comparison with
10010 SIGN_EXTEND. But don't do it if we don't have a compare
10011 insn of the given mode, since we'd have to revert it
10012 later on, and then we wouldn't know whether to sign- or
10013 zero-extend. */
10014 mode = GET_MODE (XEXP (op0, 0));
10015 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10016 && ! unsigned_comparison_p
10017 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10018 && ((unsigned HOST_WIDE_INT) const_op
10019 < (((unsigned HOST_WIDE_INT) 1
10020 << (GET_MODE_BITSIZE (mode) - 1))))
10021 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10023 op0 = XEXP (op0, 0);
10024 continue;
10026 break;
10028 case SUBREG:
10029 /* Check for the case where we are comparing A - C1 with C2, that is
10031 (subreg:MODE (plus (A) (-C1))) op (C2)
10033 with C1 a constant, and try to lift the SUBREG, i.e. to do the
10034 comparison in the wider mode. One of the following two conditions
10035 must be true in order for this to be valid:
10037 1. The mode extension results in the same bit pattern being added
10038 on both sides and the comparison is equality or unsigned. As
10039 C2 has been truncated to fit in MODE, the pattern can only be
10040 all 0s or all 1s.
10042 2. The mode extension results in the sign bit being copied on
10043 each side.
10045 The difficulty here is that we have predicates for A but not for
10046 (A - C1) so we need to check that C1 is within proper bounds so
10047 as to perturbate A as little as possible. */
10049 if (mode_width <= HOST_BITS_PER_WIDE_INT
10050 && subreg_lowpart_p (op0)
10051 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) > mode_width
10052 && GET_CODE (SUBREG_REG (op0)) == PLUS
10053 && GET_CODE (XEXP (SUBREG_REG (op0), 1)) == CONST_INT)
10055 enum machine_mode inner_mode = GET_MODE (SUBREG_REG (op0));
10056 rtx a = XEXP (SUBREG_REG (op0), 0);
10057 HOST_WIDE_INT c1 = -INTVAL (XEXP (SUBREG_REG (op0), 1));
10059 if ((c1 > 0
10060 && (unsigned HOST_WIDE_INT) c1
10061 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)
10062 && (equality_comparison_p || unsigned_comparison_p)
10063 /* (A - C1) zero-extends if it is positive and sign-extends
10064 if it is negative, C2 both zero- and sign-extends. */
10065 && ((0 == (nonzero_bits (a, inner_mode)
10066 & ~GET_MODE_MASK (mode))
10067 && const_op >= 0)
10068 /* (A - C1) sign-extends if it is positive and 1-extends
10069 if it is negative, C2 both sign- and 1-extends. */
10070 || (num_sign_bit_copies (a, inner_mode)
10071 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10072 - mode_width)
10073 && const_op < 0)))
10074 || ((unsigned HOST_WIDE_INT) c1
10075 < (unsigned HOST_WIDE_INT) 1 << (mode_width - 2)
10076 /* (A - C1) always sign-extends, like C2. */
10077 && num_sign_bit_copies (a, inner_mode)
10078 > (unsigned int) (GET_MODE_BITSIZE (inner_mode)
10079 - (mode_width - 1))))
10081 op0 = SUBREG_REG (op0);
10082 continue;
10086 /* If the inner mode is narrower and we are extracting the low part,
10087 we can treat the SUBREG as if it were a ZERO_EXTEND. */
10088 if (subreg_lowpart_p (op0)
10089 && GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0))) < mode_width)
10090 /* Fall through */ ;
10091 else
10092 break;
10094 /* ... fall through ... */
10096 case ZERO_EXTEND:
10097 mode = GET_MODE (XEXP (op0, 0));
10098 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10099 && (unsigned_comparison_p || equality_comparison_p)
10100 && (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10101 && ((unsigned HOST_WIDE_INT) const_op < GET_MODE_MASK (mode))
10102 && cmp_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
10104 op0 = XEXP (op0, 0);
10105 continue;
10107 break;
10109 case PLUS:
10110 /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do
10111 this for equality comparisons due to pathological cases involving
10112 overflows. */
10113 if (equality_comparison_p
10114 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10115 op1, XEXP (op0, 1))))
10117 op0 = XEXP (op0, 0);
10118 op1 = tem;
10119 continue;
10122 /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */
10123 if (const_op == 0 && XEXP (op0, 1) == constm1_rtx
10124 && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p)
10126 op0 = XEXP (XEXP (op0, 0), 0);
10127 code = (code == LT ? EQ : NE);
10128 continue;
10130 break;
10132 case MINUS:
10133 /* We used to optimize signed comparisons against zero, but that
10134 was incorrect. Unsigned comparisons against zero (GTU, LEU)
10135 arrive here as equality comparisons, or (GEU, LTU) are
10136 optimized away. No need to special-case them. */
10138 /* (eq (minus A B) C) -> (eq A (plus B C)) or
10139 (eq B (minus A C)), whichever simplifies. We can only do
10140 this for equality comparisons due to pathological cases involving
10141 overflows. */
10142 if (equality_comparison_p
10143 && 0 != (tem = simplify_binary_operation (PLUS, mode,
10144 XEXP (op0, 1), op1)))
10146 op0 = XEXP (op0, 0);
10147 op1 = tem;
10148 continue;
10151 if (equality_comparison_p
10152 && 0 != (tem = simplify_binary_operation (MINUS, mode,
10153 XEXP (op0, 0), op1)))
10155 op0 = XEXP (op0, 1);
10156 op1 = tem;
10157 continue;
10160 /* The sign bit of (minus (ashiftrt X C) X), where C is the number
10161 of bits in X minus 1, is one iff X > 0. */
10162 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT
10163 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10164 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (op0, 0), 1))
10165 == mode_width - 1
10166 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10168 op0 = XEXP (op0, 1);
10169 code = (code == GE ? LE : GT);
10170 continue;
10172 break;
10174 case XOR:
10175 /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification
10176 if C is zero or B is a constant. */
10177 if (equality_comparison_p
10178 && 0 != (tem = simplify_binary_operation (XOR, mode,
10179 XEXP (op0, 1), op1)))
10181 op0 = XEXP (op0, 0);
10182 op1 = tem;
10183 continue;
10185 break;
10187 case EQ: case NE:
10188 case UNEQ: case LTGT:
10189 case LT: case LTU: case UNLT: case LE: case LEU: case UNLE:
10190 case GT: case GTU: case UNGT: case GE: case GEU: case UNGE:
10191 case UNORDERED: case ORDERED:
10192 /* We can't do anything if OP0 is a condition code value, rather
10193 than an actual data value. */
10194 if (const_op != 0
10195 || CC0_P (XEXP (op0, 0))
10196 || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC)
10197 break;
10199 /* Get the two operands being compared. */
10200 if (GET_CODE (XEXP (op0, 0)) == COMPARE)
10201 tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1);
10202 else
10203 tem = XEXP (op0, 0), tem1 = XEXP (op0, 1);
10205 /* Check for the cases where we simply want the result of the
10206 earlier test or the opposite of that result. */
10207 if (code == NE || code == EQ
10208 || (GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT
10209 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10210 && (STORE_FLAG_VALUE
10211 & (((HOST_WIDE_INT) 1
10212 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
10213 && (code == LT || code == GE)))
10215 enum rtx_code new_code;
10216 if (code == LT || code == NE)
10217 new_code = GET_CODE (op0);
10218 else
10219 new_code = reversed_comparison_code (op0, NULL);
10221 if (new_code != UNKNOWN)
10223 code = new_code;
10224 op0 = tem;
10225 op1 = tem1;
10226 continue;
10229 break;
10231 case IOR:
10232 /* The sign bit of (ior (plus X (const_int -1)) X) is nonzero
10233 iff X <= 0. */
10234 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS
10235 && XEXP (XEXP (op0, 0), 1) == constm1_rtx
10236 && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1)))
10238 op0 = XEXP (op0, 1);
10239 code = (code == GE ? GT : LE);
10240 continue;
10242 break;
10244 case AND:
10245 /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This
10246 will be converted to a ZERO_EXTRACT later. */
10247 if (const_op == 0 && equality_comparison_p
10248 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10249 && XEXP (XEXP (op0, 0), 0) == const1_rtx)
10251 op0 = simplify_and_const_int
10252 (NULL_RTX, mode, gen_rtx_LSHIFTRT (mode,
10253 XEXP (op0, 1),
10254 XEXP (XEXP (op0, 0), 1)),
10255 (HOST_WIDE_INT) 1);
10256 continue;
10259 /* If we are comparing (and (lshiftrt X C1) C2) for equality with
10260 zero and X is a comparison and C1 and C2 describe only bits set
10261 in STORE_FLAG_VALUE, we can compare with X. */
10262 if (const_op == 0 && equality_comparison_p
10263 && mode_width <= HOST_BITS_PER_WIDE_INT
10264 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10265 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT
10266 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10267 && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0
10268 && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT)
10270 mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10271 << INTVAL (XEXP (XEXP (op0, 0), 1)));
10272 if ((~STORE_FLAG_VALUE & mask) == 0
10273 && (COMPARISON_P (XEXP (XEXP (op0, 0), 0))
10274 || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0
10275 && COMPARISON_P (tem))))
10277 op0 = XEXP (XEXP (op0, 0), 0);
10278 continue;
10282 /* If we are doing an equality comparison of an AND of a bit equal
10283 to the sign bit, replace this with a LT or GE comparison of
10284 the underlying value. */
10285 if (equality_comparison_p
10286 && const_op == 0
10287 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10288 && mode_width <= HOST_BITS_PER_WIDE_INT
10289 && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode))
10290 == (unsigned HOST_WIDE_INT) 1 << (mode_width - 1)))
10292 op0 = XEXP (op0, 0);
10293 code = (code == EQ ? GE : LT);
10294 continue;
10297 /* If this AND operation is really a ZERO_EXTEND from a narrower
10298 mode, the constant fits within that mode, and this is either an
10299 equality or unsigned comparison, try to do this comparison in
10300 the narrower mode.
10302 Note that in:
10304 (ne:DI (and:DI (reg:DI 4) (const_int 0xffffffff)) (const_int 0))
10305 -> (ne:DI (reg:SI 4) (const_int 0))
10307 unless TRULY_NOOP_TRUNCATION allows it or the register is
10308 known to hold a value of the required mode the
10309 transformation is invalid. */
10310 if ((equality_comparison_p || unsigned_comparison_p)
10311 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10312 && (i = exact_log2 ((INTVAL (XEXP (op0, 1))
10313 & GET_MODE_MASK (mode))
10314 + 1)) >= 0
10315 && const_op >> i == 0
10316 && (tmode = mode_for_size (i, MODE_INT, 1)) != BLKmode
10317 && (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (tmode),
10318 GET_MODE_BITSIZE (GET_MODE (op0)))
10319 || (REG_P (XEXP (op0, 0))
10320 && reg_truncated_to_mode (tmode, XEXP (op0, 0)))))
10322 op0 = gen_lowpart (tmode, XEXP (op0, 0));
10323 continue;
10326 /* If this is (and:M1 (subreg:M2 X 0) (const_int C1)) where C1
10327 fits in both M1 and M2 and the SUBREG is either paradoxical
10328 or represents the low part, permute the SUBREG and the AND
10329 and try again. */
10330 if (GET_CODE (XEXP (op0, 0)) == SUBREG)
10332 unsigned HOST_WIDE_INT c1;
10333 tmode = GET_MODE (SUBREG_REG (XEXP (op0, 0)));
10334 /* Require an integral mode, to avoid creating something like
10335 (AND:SF ...). */
10336 if (SCALAR_INT_MODE_P (tmode)
10337 /* It is unsafe to commute the AND into the SUBREG if the
10338 SUBREG is paradoxical and WORD_REGISTER_OPERATIONS is
10339 not defined. As originally written the upper bits
10340 have a defined value due to the AND operation.
10341 However, if we commute the AND inside the SUBREG then
10342 they no longer have defined values and the meaning of
10343 the code has been changed. */
10344 && (0
10345 #ifdef WORD_REGISTER_OPERATIONS
10346 || (mode_width > GET_MODE_BITSIZE (tmode)
10347 && mode_width <= BITS_PER_WORD)
10348 #endif
10349 || (mode_width <= GET_MODE_BITSIZE (tmode)
10350 && subreg_lowpart_p (XEXP (op0, 0))))
10351 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10352 && mode_width <= HOST_BITS_PER_WIDE_INT
10353 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT
10354 && ((c1 = INTVAL (XEXP (op0, 1))) & ~mask) == 0
10355 && (c1 & ~GET_MODE_MASK (tmode)) == 0
10356 && c1 != mask
10357 && c1 != GET_MODE_MASK (tmode))
10359 op0 = simplify_gen_binary (AND, tmode,
10360 SUBREG_REG (XEXP (op0, 0)),
10361 gen_int_mode (c1, tmode));
10362 op0 = gen_lowpart (mode, op0);
10363 continue;
10367 /* Convert (ne (and (not X) 1) 0) to (eq (and X 1) 0). */
10368 if (const_op == 0 && equality_comparison_p
10369 && XEXP (op0, 1) == const1_rtx
10370 && GET_CODE (XEXP (op0, 0)) == NOT)
10372 op0 = simplify_and_const_int
10373 (NULL_RTX, mode, XEXP (XEXP (op0, 0), 0), (HOST_WIDE_INT) 1);
10374 code = (code == NE ? EQ : NE);
10375 continue;
10378 /* Convert (ne (and (lshiftrt (not X)) 1) 0) to
10379 (eq (and (lshiftrt X) 1) 0).
10380 Also handle the case where (not X) is expressed using xor. */
10381 if (const_op == 0 && equality_comparison_p
10382 && XEXP (op0, 1) == const1_rtx
10383 && GET_CODE (XEXP (op0, 0)) == LSHIFTRT)
10385 rtx shift_op = XEXP (XEXP (op0, 0), 0);
10386 rtx shift_count = XEXP (XEXP (op0, 0), 1);
10388 if (GET_CODE (shift_op) == NOT
10389 || (GET_CODE (shift_op) == XOR
10390 && GET_CODE (XEXP (shift_op, 1)) == CONST_INT
10391 && GET_CODE (shift_count) == CONST_INT
10392 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
10393 && (INTVAL (XEXP (shift_op, 1))
10394 == (HOST_WIDE_INT) 1 << INTVAL (shift_count))))
10396 op0 = simplify_and_const_int
10397 (NULL_RTX, mode,
10398 gen_rtx_LSHIFTRT (mode, XEXP (shift_op, 0), shift_count),
10399 (HOST_WIDE_INT) 1);
10400 code = (code == NE ? EQ : NE);
10401 continue;
10404 break;
10406 case ASHIFT:
10407 /* If we have (compare (ashift FOO N) (const_int C)) and
10408 the high order N bits of FOO (N+1 if an inequality comparison)
10409 are known to be zero, we can do this by comparing FOO with C
10410 shifted right N bits so long as the low-order N bits of C are
10411 zero. */
10412 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10413 && INTVAL (XEXP (op0, 1)) >= 0
10414 && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p)
10415 < HOST_BITS_PER_WIDE_INT)
10416 && ((const_op
10417 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0)
10418 && mode_width <= HOST_BITS_PER_WIDE_INT
10419 && (nonzero_bits (XEXP (op0, 0), mode)
10420 & ~(mask >> (INTVAL (XEXP (op0, 1))
10421 + ! equality_comparison_p))) == 0)
10423 /* We must perform a logical shift, not an arithmetic one,
10424 as we want the top N bits of C to be zero. */
10425 unsigned HOST_WIDE_INT temp = const_op & GET_MODE_MASK (mode);
10427 temp >>= INTVAL (XEXP (op0, 1));
10428 op1 = gen_int_mode (temp, mode);
10429 op0 = XEXP (op0, 0);
10430 continue;
10433 /* If we are doing a sign bit comparison, it means we are testing
10434 a particular bit. Convert it to the appropriate AND. */
10435 if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 1)) == CONST_INT
10436 && mode_width <= HOST_BITS_PER_WIDE_INT)
10438 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10439 ((HOST_WIDE_INT) 1
10440 << (mode_width - 1
10441 - INTVAL (XEXP (op0, 1)))));
10442 code = (code == LT ? NE : EQ);
10443 continue;
10446 /* If this an equality comparison with zero and we are shifting
10447 the low bit to the sign bit, we can convert this to an AND of the
10448 low-order bit. */
10449 if (const_op == 0 && equality_comparison_p
10450 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10451 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10452 == mode_width - 1)
10454 op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0),
10455 (HOST_WIDE_INT) 1);
10456 continue;
10458 break;
10460 case ASHIFTRT:
10461 /* If this is an equality comparison with zero, we can do this
10462 as a logical shift, which might be much simpler. */
10463 if (equality_comparison_p && const_op == 0
10464 && GET_CODE (XEXP (op0, 1)) == CONST_INT)
10466 op0 = simplify_shift_const (NULL_RTX, LSHIFTRT, mode,
10467 XEXP (op0, 0),
10468 INTVAL (XEXP (op0, 1)));
10469 continue;
10472 /* If OP0 is a sign extension and CODE is not an unsigned comparison,
10473 do the comparison in a narrower mode. */
10474 if (! unsigned_comparison_p
10475 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10476 && GET_CODE (XEXP (op0, 0)) == ASHIFT
10477 && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1)
10478 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10479 MODE_INT, 1)) != BLKmode
10480 && (((unsigned HOST_WIDE_INT) const_op
10481 + (GET_MODE_MASK (tmode) >> 1) + 1)
10482 <= GET_MODE_MASK (tmode)))
10484 op0 = gen_lowpart (tmode, XEXP (XEXP (op0, 0), 0));
10485 continue;
10488 /* Likewise if OP0 is a PLUS of a sign extension with a
10489 constant, which is usually represented with the PLUS
10490 between the shifts. */
10491 if (! unsigned_comparison_p
10492 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10493 && GET_CODE (XEXP (op0, 0)) == PLUS
10494 && GET_CODE (XEXP (XEXP (op0, 0), 1)) == CONST_INT
10495 && GET_CODE (XEXP (XEXP (op0, 0), 0)) == ASHIFT
10496 && XEXP (op0, 1) == XEXP (XEXP (XEXP (op0, 0), 0), 1)
10497 && (tmode = mode_for_size (mode_width - INTVAL (XEXP (op0, 1)),
10498 MODE_INT, 1)) != BLKmode
10499 && (((unsigned HOST_WIDE_INT) const_op
10500 + (GET_MODE_MASK (tmode) >> 1) + 1)
10501 <= GET_MODE_MASK (tmode)))
10503 rtx inner = XEXP (XEXP (XEXP (op0, 0), 0), 0);
10504 rtx add_const = XEXP (XEXP (op0, 0), 1);
10505 rtx new_const = simplify_gen_binary (ASHIFTRT, GET_MODE (op0),
10506 add_const, XEXP (op0, 1));
10508 op0 = simplify_gen_binary (PLUS, tmode,
10509 gen_lowpart (tmode, inner),
10510 new_const);
10511 continue;
10514 /* ... fall through ... */
10515 case LSHIFTRT:
10516 /* If we have (compare (xshiftrt FOO N) (const_int C)) and
10517 the low order N bits of FOO are known to be zero, we can do this
10518 by comparing FOO with C shifted left N bits so long as no
10519 overflow occurs. */
10520 if (GET_CODE (XEXP (op0, 1)) == CONST_INT
10521 && INTVAL (XEXP (op0, 1)) >= 0
10522 && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT
10523 && mode_width <= HOST_BITS_PER_WIDE_INT
10524 && (nonzero_bits (XEXP (op0, 0), mode)
10525 & (((HOST_WIDE_INT) 1 << INTVAL (XEXP (op0, 1))) - 1)) == 0
10526 && (((unsigned HOST_WIDE_INT) const_op
10527 + (GET_CODE (op0) != LSHIFTRT
10528 ? ((GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)) >> 1)
10529 + 1)
10530 : 0))
10531 <= GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1))))
10533 /* If the shift was logical, then we must make the condition
10534 unsigned. */
10535 if (GET_CODE (op0) == LSHIFTRT)
10536 code = unsigned_condition (code);
10538 const_op <<= INTVAL (XEXP (op0, 1));
10539 op1 = GEN_INT (const_op);
10540 op0 = XEXP (op0, 0);
10541 continue;
10544 /* If we are using this shift to extract just the sign bit, we
10545 can replace this with an LT or GE comparison. */
10546 if (const_op == 0
10547 && (equality_comparison_p || sign_bit_comparison_p)
10548 && GET_CODE (XEXP (op0, 1)) == CONST_INT
10549 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (op0, 1))
10550 == mode_width - 1)
10552 op0 = XEXP (op0, 0);
10553 code = (code == NE || code == GT ? LT : GE);
10554 continue;
10556 break;
10558 default:
10559 break;
10562 break;
10565 /* Now make any compound operations involved in this comparison. Then,
10566 check for an outmost SUBREG on OP0 that is not doing anything or is
10567 paradoxical. The latter transformation must only be performed when
10568 it is known that the "extra" bits will be the same in op0 and op1 or
10569 that they don't matter. There are three cases to consider:
10571 1. SUBREG_REG (op0) is a register. In this case the bits are don't
10572 care bits and we can assume they have any convenient value. So
10573 making the transformation is safe.
10575 2. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is not defined.
10576 In this case the upper bits of op0 are undefined. We should not make
10577 the simplification in that case as we do not know the contents of
10578 those bits.
10580 3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is defined and not
10581 UNKNOWN. In that case we know those bits are zeros or ones. We must
10582 also be sure that they are the same as the upper bits of op1.
10584 We can never remove a SUBREG for a non-equality comparison because
10585 the sign bit is in a different place in the underlying object. */
10587 op0 = make_compound_operation (op0, op1 == const0_rtx ? COMPARE : SET);
10588 op1 = make_compound_operation (op1, SET);
10590 if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0)
10591 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT
10592 && GET_MODE_CLASS (GET_MODE (SUBREG_REG (op0))) == MODE_INT
10593 && (code == NE || code == EQ))
10595 if (GET_MODE_SIZE (GET_MODE (op0))
10596 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0))))
10598 /* For paradoxical subregs, allow case 1 as above. Case 3 isn't
10599 implemented. */
10600 if (REG_P (SUBREG_REG (op0)))
10602 op0 = SUBREG_REG (op0);
10603 op1 = gen_lowpart (GET_MODE (op0), op1);
10606 else if ((GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op0)))
10607 <= HOST_BITS_PER_WIDE_INT)
10608 && (nonzero_bits (SUBREG_REG (op0),
10609 GET_MODE (SUBREG_REG (op0)))
10610 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10612 tem = gen_lowpart (GET_MODE (SUBREG_REG (op0)), op1);
10614 if ((nonzero_bits (tem, GET_MODE (SUBREG_REG (op0)))
10615 & ~GET_MODE_MASK (GET_MODE (op0))) == 0)
10616 op0 = SUBREG_REG (op0), op1 = tem;
10620 /* We now do the opposite procedure: Some machines don't have compare
10621 insns in all modes. If OP0's mode is an integer mode smaller than a
10622 word and we can't do a compare in that mode, see if there is a larger
10623 mode for which we can do the compare. There are a number of cases in
10624 which we can use the wider mode. */
10626 mode = GET_MODE (op0);
10627 if (mode != VOIDmode && GET_MODE_CLASS (mode) == MODE_INT
10628 && GET_MODE_SIZE (mode) < UNITS_PER_WORD
10629 && ! have_insn_for (COMPARE, mode))
10630 for (tmode = GET_MODE_WIDER_MODE (mode);
10631 (tmode != VOIDmode
10632 && GET_MODE_BITSIZE (tmode) <= HOST_BITS_PER_WIDE_INT);
10633 tmode = GET_MODE_WIDER_MODE (tmode))
10634 if (have_insn_for (COMPARE, tmode))
10636 int zero_extended;
10638 /* If the only nonzero bits in OP0 and OP1 are those in the
10639 narrower mode and this is an equality or unsigned comparison,
10640 we can use the wider mode. Similarly for sign-extended
10641 values, in which case it is true for all comparisons. */
10642 zero_extended = ((code == EQ || code == NE
10643 || code == GEU || code == GTU
10644 || code == LEU || code == LTU)
10645 && (nonzero_bits (op0, tmode)
10646 & ~GET_MODE_MASK (mode)) == 0
10647 && ((GET_CODE (op1) == CONST_INT
10648 || (nonzero_bits (op1, tmode)
10649 & ~GET_MODE_MASK (mode)) == 0)));
10651 if (zero_extended
10652 || ((num_sign_bit_copies (op0, tmode)
10653 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10654 - GET_MODE_BITSIZE (mode)))
10655 && (num_sign_bit_copies (op1, tmode)
10656 > (unsigned int) (GET_MODE_BITSIZE (tmode)
10657 - GET_MODE_BITSIZE (mode)))))
10659 /* If OP0 is an AND and we don't have an AND in MODE either,
10660 make a new AND in the proper mode. */
10661 if (GET_CODE (op0) == AND
10662 && !have_insn_for (AND, mode))
10663 op0 = simplify_gen_binary (AND, tmode,
10664 gen_lowpart (tmode,
10665 XEXP (op0, 0)),
10666 gen_lowpart (tmode,
10667 XEXP (op0, 1)));
10669 op0 = gen_lowpart (tmode, op0);
10670 if (zero_extended && GET_CODE (op1) == CONST_INT)
10671 op1 = GEN_INT (INTVAL (op1) & GET_MODE_MASK (mode));
10672 op1 = gen_lowpart (tmode, op1);
10673 break;
10676 /* If this is a test for negative, we can make an explicit
10677 test of the sign bit. */
10679 if (op1 == const0_rtx && (code == LT || code == GE)
10680 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10682 op0 = simplify_gen_binary (AND, tmode,
10683 gen_lowpart (tmode, op0),
10684 GEN_INT ((HOST_WIDE_INT) 1
10685 << (GET_MODE_BITSIZE (mode)
10686 - 1)));
10687 code = (code == LT) ? NE : EQ;
10688 break;
10692 #ifdef CANONICALIZE_COMPARISON
10693 /* If this machine only supports a subset of valid comparisons, see if we
10694 can convert an unsupported one into a supported one. */
10695 CANONICALIZE_COMPARISON (code, op0, op1);
10696 #endif
10698 *pop0 = op0;
10699 *pop1 = op1;
10701 return code;
10704 /* Utility function for record_value_for_reg. Count number of
10705 rtxs in X. */
10706 static int
10707 count_rtxs (rtx x)
10709 enum rtx_code code = GET_CODE (x);
10710 const char *fmt;
10711 int i, ret = 1;
10713 if (GET_RTX_CLASS (code) == '2'
10714 || GET_RTX_CLASS (code) == 'c')
10716 rtx x0 = XEXP (x, 0);
10717 rtx x1 = XEXP (x, 1);
10719 if (x0 == x1)
10720 return 1 + 2 * count_rtxs (x0);
10722 if ((GET_RTX_CLASS (GET_CODE (x1)) == '2'
10723 || GET_RTX_CLASS (GET_CODE (x1)) == 'c')
10724 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10725 return 2 + 2 * count_rtxs (x0)
10726 + count_rtxs (x == XEXP (x1, 0)
10727 ? XEXP (x1, 1) : XEXP (x1, 0));
10729 if ((GET_RTX_CLASS (GET_CODE (x0)) == '2'
10730 || GET_RTX_CLASS (GET_CODE (x0)) == 'c')
10731 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10732 return 2 + 2 * count_rtxs (x1)
10733 + count_rtxs (x == XEXP (x0, 0)
10734 ? XEXP (x0, 1) : XEXP (x0, 0));
10737 fmt = GET_RTX_FORMAT (code);
10738 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10739 if (fmt[i] == 'e')
10740 ret += count_rtxs (XEXP (x, i));
10742 return ret;
10745 /* Utility function for following routine. Called when X is part of a value
10746 being stored into last_set_value. Sets last_set_table_tick
10747 for each register mentioned. Similar to mention_regs in cse.c */
10749 static void
10750 update_table_tick (rtx x)
10752 enum rtx_code code = GET_CODE (x);
10753 const char *fmt = GET_RTX_FORMAT (code);
10754 int i;
10756 if (code == REG)
10758 unsigned int regno = REGNO (x);
10759 unsigned int endregno
10760 = regno + (regno < FIRST_PSEUDO_REGISTER
10761 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
10762 unsigned int r;
10764 for (r = regno; r < endregno; r++)
10765 reg_stat[r].last_set_table_tick = label_tick;
10767 return;
10770 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10771 /* Note that we can't have an "E" in values stored; see
10772 get_last_value_validate. */
10773 if (fmt[i] == 'e')
10775 /* Check for identical subexpressions. If x contains
10776 identical subexpression we only have to traverse one of
10777 them. */
10778 if (i == 0 && ARITHMETIC_P (x))
10780 /* Note that at this point x1 has already been
10781 processed. */
10782 rtx x0 = XEXP (x, 0);
10783 rtx x1 = XEXP (x, 1);
10785 /* If x0 and x1 are identical then there is no need to
10786 process x0. */
10787 if (x0 == x1)
10788 break;
10790 /* If x0 is identical to a subexpression of x1 then while
10791 processing x1, x0 has already been processed. Thus we
10792 are done with x. */
10793 if (ARITHMETIC_P (x1)
10794 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
10795 break;
10797 /* If x1 is identical to a subexpression of x0 then we
10798 still have to process the rest of x0. */
10799 if (ARITHMETIC_P (x0)
10800 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
10802 update_table_tick (XEXP (x0, x1 == XEXP (x0, 0) ? 1 : 0));
10803 break;
10807 update_table_tick (XEXP (x, i));
10811 /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we
10812 are saying that the register is clobbered and we no longer know its
10813 value. If INSN is zero, don't update reg_stat[].last_set; this is
10814 only permitted with VALUE also zero and is used to invalidate the
10815 register. */
10817 static void
10818 record_value_for_reg (rtx reg, rtx insn, rtx value)
10820 unsigned int regno = REGNO (reg);
10821 unsigned int endregno
10822 = regno + (regno < FIRST_PSEUDO_REGISTER
10823 ? hard_regno_nregs[regno][GET_MODE (reg)] : 1);
10824 unsigned int i;
10826 /* If VALUE contains REG and we have a previous value for REG, substitute
10827 the previous value. */
10828 if (value && insn && reg_overlap_mentioned_p (reg, value))
10830 rtx tem;
10832 /* Set things up so get_last_value is allowed to see anything set up to
10833 our insn. */
10834 subst_low_cuid = INSN_CUID (insn);
10835 tem = get_last_value (reg);
10837 /* If TEM is simply a binary operation with two CLOBBERs as operands,
10838 it isn't going to be useful and will take a lot of time to process,
10839 so just use the CLOBBER. */
10841 if (tem)
10843 if (ARITHMETIC_P (tem)
10844 && GET_CODE (XEXP (tem, 0)) == CLOBBER
10845 && GET_CODE (XEXP (tem, 1)) == CLOBBER)
10846 tem = XEXP (tem, 0);
10847 else if (count_occurrences (value, reg, 1) >= 2)
10849 /* If there are two or more occurrences of REG in VALUE,
10850 prevent the value from growing too much. */
10851 if (count_rtxs (tem) > MAX_LAST_VALUE_RTL)
10852 tem = gen_rtx_CLOBBER (GET_MODE (tem), const0_rtx);
10855 value = replace_rtx (copy_rtx (value), reg, tem);
10859 /* For each register modified, show we don't know its value, that
10860 we don't know about its bitwise content, that its value has been
10861 updated, and that we don't know the location of the death of the
10862 register. */
10863 for (i = regno; i < endregno; i++)
10865 if (insn)
10866 reg_stat[i].last_set = insn;
10868 reg_stat[i].last_set_value = 0;
10869 reg_stat[i].last_set_mode = 0;
10870 reg_stat[i].last_set_nonzero_bits = 0;
10871 reg_stat[i].last_set_sign_bit_copies = 0;
10872 reg_stat[i].last_death = 0;
10873 reg_stat[i].truncated_to_mode = 0;
10876 /* Mark registers that are being referenced in this value. */
10877 if (value)
10878 update_table_tick (value);
10880 /* Now update the status of each register being set.
10881 If someone is using this register in this block, set this register
10882 to invalid since we will get confused between the two lives in this
10883 basic block. This makes using this register always invalid. In cse, we
10884 scan the table to invalidate all entries using this register, but this
10885 is too much work for us. */
10887 for (i = regno; i < endregno; i++)
10889 reg_stat[i].last_set_label = label_tick;
10890 if (!insn || (value && reg_stat[i].last_set_table_tick == label_tick))
10891 reg_stat[i].last_set_invalid = 1;
10892 else
10893 reg_stat[i].last_set_invalid = 0;
10896 /* The value being assigned might refer to X (like in "x++;"). In that
10897 case, we must replace it with (clobber (const_int 0)) to prevent
10898 infinite loops. */
10899 if (value && ! get_last_value_validate (&value, insn,
10900 reg_stat[regno].last_set_label, 0))
10902 value = copy_rtx (value);
10903 if (! get_last_value_validate (&value, insn,
10904 reg_stat[regno].last_set_label, 1))
10905 value = 0;
10908 /* For the main register being modified, update the value, the mode, the
10909 nonzero bits, and the number of sign bit copies. */
10911 reg_stat[regno].last_set_value = value;
10913 if (value)
10915 enum machine_mode mode = GET_MODE (reg);
10916 subst_low_cuid = INSN_CUID (insn);
10917 reg_stat[regno].last_set_mode = mode;
10918 if (GET_MODE_CLASS (mode) == MODE_INT
10919 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
10920 mode = nonzero_bits_mode;
10921 reg_stat[regno].last_set_nonzero_bits = nonzero_bits (value, mode);
10922 reg_stat[regno].last_set_sign_bit_copies
10923 = num_sign_bit_copies (value, GET_MODE (reg));
10927 /* Called via note_stores from record_dead_and_set_regs to handle one
10928 SET or CLOBBER in an insn. DATA is the instruction in which the
10929 set is occurring. */
10931 static void
10932 record_dead_and_set_regs_1 (rtx dest, rtx setter, void *data)
10934 rtx record_dead_insn = (rtx) data;
10936 if (GET_CODE (dest) == SUBREG)
10937 dest = SUBREG_REG (dest);
10939 if (!record_dead_insn)
10941 if (REG_P (dest))
10942 record_value_for_reg (dest, NULL_RTX, NULL_RTX);
10943 return;
10946 if (REG_P (dest))
10948 /* If we are setting the whole register, we know its value. Otherwise
10949 show that we don't know the value. We can handle SUBREG in
10950 some cases. */
10951 if (GET_CODE (setter) == SET && dest == SET_DEST (setter))
10952 record_value_for_reg (dest, record_dead_insn, SET_SRC (setter));
10953 else if (GET_CODE (setter) == SET
10954 && GET_CODE (SET_DEST (setter)) == SUBREG
10955 && SUBREG_REG (SET_DEST (setter)) == dest
10956 && GET_MODE_BITSIZE (GET_MODE (dest)) <= BITS_PER_WORD
10957 && subreg_lowpart_p (SET_DEST (setter)))
10958 record_value_for_reg (dest, record_dead_insn,
10959 gen_lowpart (GET_MODE (dest),
10960 SET_SRC (setter)));
10961 else
10962 record_value_for_reg (dest, record_dead_insn, NULL_RTX);
10964 else if (MEM_P (dest)
10965 /* Ignore pushes, they clobber nothing. */
10966 && ! push_operand (dest, GET_MODE (dest)))
10967 mem_last_set = INSN_CUID (record_dead_insn);
10970 /* Update the records of when each REG was most recently set or killed
10971 for the things done by INSN. This is the last thing done in processing
10972 INSN in the combiner loop.
10974 We update reg_stat[], in particular fields last_set, last_set_value,
10975 last_set_mode, last_set_nonzero_bits, last_set_sign_bit_copies,
10976 last_death, and also the similar information mem_last_set (which insn
10977 most recently modified memory) and last_call_cuid (which insn was the
10978 most recent subroutine call). */
10980 static void
10981 record_dead_and_set_regs (rtx insn)
10983 rtx link;
10984 unsigned int i;
10986 for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
10988 if (REG_NOTE_KIND (link) == REG_DEAD
10989 && REG_P (XEXP (link, 0)))
10991 unsigned int regno = REGNO (XEXP (link, 0));
10992 unsigned int endregno
10993 = regno + (regno < FIRST_PSEUDO_REGISTER
10994 ? hard_regno_nregs[regno][GET_MODE (XEXP (link, 0))]
10995 : 1);
10997 for (i = regno; i < endregno; i++)
10998 reg_stat[i].last_death = insn;
11000 else if (REG_NOTE_KIND (link) == REG_INC)
11001 record_value_for_reg (XEXP (link, 0), insn, NULL_RTX);
11004 if (CALL_P (insn))
11006 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
11007 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
11009 reg_stat[i].last_set_value = 0;
11010 reg_stat[i].last_set_mode = 0;
11011 reg_stat[i].last_set_nonzero_bits = 0;
11012 reg_stat[i].last_set_sign_bit_copies = 0;
11013 reg_stat[i].last_death = 0;
11014 reg_stat[i].truncated_to_mode = 0;
11017 last_call_cuid = mem_last_set = INSN_CUID (insn);
11019 /* We can't combine into a call pattern. Remember, though, that
11020 the return value register is set at this CUID. We could
11021 still replace a register with the return value from the
11022 wrong subroutine call! */
11023 note_stores (PATTERN (insn), record_dead_and_set_regs_1, NULL_RTX);
11025 else
11026 note_stores (PATTERN (insn), record_dead_and_set_regs_1, insn);
11029 /* If a SUBREG has the promoted bit set, it is in fact a property of the
11030 register present in the SUBREG, so for each such SUBREG go back and
11031 adjust nonzero and sign bit information of the registers that are
11032 known to have some zero/sign bits set.
11034 This is needed because when combine blows the SUBREGs away, the
11035 information on zero/sign bits is lost and further combines can be
11036 missed because of that. */
11038 static void
11039 record_promoted_value (rtx insn, rtx subreg)
11041 rtx links, set;
11042 unsigned int regno = REGNO (SUBREG_REG (subreg));
11043 enum machine_mode mode = GET_MODE (subreg);
11045 if (GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)
11046 return;
11048 for (links = LOG_LINKS (insn); links;)
11050 insn = XEXP (links, 0);
11051 set = single_set (insn);
11053 if (! set || !REG_P (SET_DEST (set))
11054 || REGNO (SET_DEST (set)) != regno
11055 || GET_MODE (SET_DEST (set)) != GET_MODE (SUBREG_REG (subreg)))
11057 links = XEXP (links, 1);
11058 continue;
11061 if (reg_stat[regno].last_set == insn)
11063 if (SUBREG_PROMOTED_UNSIGNED_P (subreg) > 0)
11064 reg_stat[regno].last_set_nonzero_bits &= GET_MODE_MASK (mode);
11067 if (REG_P (SET_SRC (set)))
11069 regno = REGNO (SET_SRC (set));
11070 links = LOG_LINKS (insn);
11072 else
11073 break;
11077 /* Check if X, a register, is known to contain a value already
11078 truncated to MODE. In this case we can use a subreg to refer to
11079 the truncated value even though in the generic case we would need
11080 an explicit truncation. */
11082 static bool
11083 reg_truncated_to_mode (enum machine_mode mode, rtx x)
11085 enum machine_mode truncated = reg_stat[REGNO (x)].truncated_to_mode;
11087 if (truncated == 0 || reg_stat[REGNO (x)].truncation_label != label_tick)
11088 return false;
11089 if (GET_MODE_SIZE (truncated) <= GET_MODE_SIZE (mode))
11090 return true;
11091 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
11092 GET_MODE_BITSIZE (truncated)))
11093 return true;
11094 return false;
11097 /* X is a REG or a SUBREG. If X is some sort of a truncation record
11098 it. For non-TRULY_NOOP_TRUNCATION targets we might be able to turn
11099 a truncate into a subreg using this information. */
11101 static void
11102 record_truncated_value (rtx x)
11104 enum machine_mode truncated_mode;
11106 if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x)))
11108 enum machine_mode original_mode = GET_MODE (SUBREG_REG (x));
11109 truncated_mode = GET_MODE (x);
11111 if (GET_MODE_SIZE (original_mode) <= GET_MODE_SIZE (truncated_mode))
11112 return;
11114 if (TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (truncated_mode),
11115 GET_MODE_BITSIZE (original_mode)))
11116 return;
11118 x = SUBREG_REG (x);
11120 /* ??? For hard-regs we now record everything. We might be able to
11121 optimize this using last_set_mode. */
11122 else if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER)
11123 truncated_mode = GET_MODE (x);
11124 else
11125 return;
11127 if (reg_stat[REGNO (x)].truncated_to_mode == 0
11128 || reg_stat[REGNO (x)].truncation_label < label_tick
11129 || (GET_MODE_SIZE (truncated_mode)
11130 < GET_MODE_SIZE (reg_stat[REGNO (x)].truncated_to_mode)))
11132 reg_stat[REGNO (x)].truncated_to_mode = truncated_mode;
11133 reg_stat[REGNO (x)].truncation_label = label_tick;
11137 /* Scan X for promoted SUBREGs and truncated REGs. For each one
11138 found, note what it implies to the registers used in it. */
11140 static void
11141 check_conversions (rtx insn, rtx x)
11143 if (GET_CODE (x) == SUBREG || REG_P (x))
11145 if (GET_CODE (x) == SUBREG
11146 && SUBREG_PROMOTED_VAR_P (x)
11147 && REG_P (SUBREG_REG (x)))
11148 record_promoted_value (insn, x);
11150 record_truncated_value (x);
11152 else
11154 const char *format = GET_RTX_FORMAT (GET_CODE (x));
11155 int i, j;
11157 for (i = 0; i < GET_RTX_LENGTH (GET_CODE (x)); i++)
11158 switch (format[i])
11160 case 'e':
11161 check_conversions (insn, XEXP (x, i));
11162 break;
11163 case 'V':
11164 case 'E':
11165 if (XVEC (x, i) != 0)
11166 for (j = 0; j < XVECLEN (x, i); j++)
11167 check_conversions (insn, XVECEXP (x, i, j));
11168 break;
11173 /* Utility routine for the following function. Verify that all the registers
11174 mentioned in *LOC are valid when *LOC was part of a value set when
11175 label_tick == TICK. Return 0 if some are not.
11177 If REPLACE is nonzero, replace the invalid reference with
11178 (clobber (const_int 0)) and return 1. This replacement is useful because
11179 we often can get useful information about the form of a value (e.g., if
11180 it was produced by a shift that always produces -1 or 0) even though
11181 we don't know exactly what registers it was produced from. */
11183 static int
11184 get_last_value_validate (rtx *loc, rtx insn, int tick, int replace)
11186 rtx x = *loc;
11187 const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
11188 int len = GET_RTX_LENGTH (GET_CODE (x));
11189 int i;
11191 if (REG_P (x))
11193 unsigned int regno = REGNO (x);
11194 unsigned int endregno
11195 = regno + (regno < FIRST_PSEUDO_REGISTER
11196 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11197 unsigned int j;
11199 for (j = regno; j < endregno; j++)
11200 if (reg_stat[j].last_set_invalid
11201 /* If this is a pseudo-register that was only set once and not
11202 live at the beginning of the function, it is always valid. */
11203 || (! (regno >= FIRST_PSEUDO_REGISTER
11204 && REG_N_SETS (regno) == 1
11205 && (! REGNO_REG_SET_P
11206 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
11207 regno)))
11208 && reg_stat[j].last_set_label > tick))
11210 if (replace)
11211 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11212 return replace;
11215 return 1;
11217 /* If this is a memory reference, make sure that there were
11218 no stores after it that might have clobbered the value. We don't
11219 have alias info, so we assume any store invalidates it. */
11220 else if (MEM_P (x) && !MEM_READONLY_P (x)
11221 && INSN_CUID (insn) <= mem_last_set)
11223 if (replace)
11224 *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx);
11225 return replace;
11228 for (i = 0; i < len; i++)
11230 if (fmt[i] == 'e')
11232 /* Check for identical subexpressions. If x contains
11233 identical subexpression we only have to traverse one of
11234 them. */
11235 if (i == 1 && ARITHMETIC_P (x))
11237 /* Note that at this point x0 has already been checked
11238 and found valid. */
11239 rtx x0 = XEXP (x, 0);
11240 rtx x1 = XEXP (x, 1);
11242 /* If x0 and x1 are identical then x is also valid. */
11243 if (x0 == x1)
11244 return 1;
11246 /* If x1 is identical to a subexpression of x0 then
11247 while checking x0, x1 has already been checked. Thus
11248 it is valid and so as x. */
11249 if (ARITHMETIC_P (x0)
11250 && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))
11251 return 1;
11253 /* If x0 is identical to a subexpression of x1 then x is
11254 valid iff the rest of x1 is valid. */
11255 if (ARITHMETIC_P (x1)
11256 && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
11257 return
11258 get_last_value_validate (&XEXP (x1,
11259 x0 == XEXP (x1, 0) ? 1 : 0),
11260 insn, tick, replace);
11263 if (get_last_value_validate (&XEXP (x, i), insn, tick,
11264 replace) == 0)
11265 return 0;
11267 /* Don't bother with these. They shouldn't occur anyway. */
11268 else if (fmt[i] == 'E')
11269 return 0;
11272 /* If we haven't found a reason for it to be invalid, it is valid. */
11273 return 1;
11276 /* Get the last value assigned to X, if known. Some registers
11277 in the value may be replaced with (clobber (const_int 0)) if their value
11278 is known longer known reliably. */
11280 static rtx
11281 get_last_value (rtx x)
11283 unsigned int regno;
11284 rtx value;
11286 /* If this is a non-paradoxical SUBREG, get the value of its operand and
11287 then convert it to the desired mode. If this is a paradoxical SUBREG,
11288 we cannot predict what values the "extra" bits might have. */
11289 if (GET_CODE (x) == SUBREG
11290 && subreg_lowpart_p (x)
11291 && (GET_MODE_SIZE (GET_MODE (x))
11292 <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
11293 && (value = get_last_value (SUBREG_REG (x))) != 0)
11294 return gen_lowpart (GET_MODE (x), value);
11296 if (!REG_P (x))
11297 return 0;
11299 regno = REGNO (x);
11300 value = reg_stat[regno].last_set_value;
11302 /* If we don't have a value, or if it isn't for this basic block and
11303 it's either a hard register, set more than once, or it's a live
11304 at the beginning of the function, return 0.
11306 Because if it's not live at the beginning of the function then the reg
11307 is always set before being used (is never used without being set).
11308 And, if it's set only once, and it's always set before use, then all
11309 uses must have the same last value, even if it's not from this basic
11310 block. */
11312 if (value == 0
11313 || (reg_stat[regno].last_set_label != label_tick
11314 && (regno < FIRST_PSEUDO_REGISTER
11315 || REG_N_SETS (regno) != 1
11316 || (REGNO_REG_SET_P
11317 (ENTRY_BLOCK_PTR->next_bb->il.rtl->global_live_at_start,
11318 regno)))))
11319 return 0;
11321 /* If the value was set in a later insn than the ones we are processing,
11322 we can't use it even if the register was only set once. */
11323 if (INSN_CUID (reg_stat[regno].last_set) >= subst_low_cuid)
11324 return 0;
11326 /* If the value has all its registers valid, return it. */
11327 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11328 reg_stat[regno].last_set_label, 0))
11329 return value;
11331 /* Otherwise, make a copy and replace any invalid register with
11332 (clobber (const_int 0)). If that fails for some reason, return 0. */
11334 value = copy_rtx (value);
11335 if (get_last_value_validate (&value, reg_stat[regno].last_set,
11336 reg_stat[regno].last_set_label, 1))
11337 return value;
11339 return 0;
11342 /* Return nonzero if expression X refers to a REG or to memory
11343 that is set in an instruction more recent than FROM_CUID. */
11345 static int
11346 use_crosses_set_p (rtx x, int from_cuid)
11348 const char *fmt;
11349 int i;
11350 enum rtx_code code = GET_CODE (x);
11352 if (code == REG)
11354 unsigned int regno = REGNO (x);
11355 unsigned endreg = regno + (regno < FIRST_PSEUDO_REGISTER
11356 ? hard_regno_nregs[regno][GET_MODE (x)] : 1);
11358 #ifdef PUSH_ROUNDING
11359 /* Don't allow uses of the stack pointer to be moved,
11360 because we don't know whether the move crosses a push insn. */
11361 if (regno == STACK_POINTER_REGNUM && PUSH_ARGS)
11362 return 1;
11363 #endif
11364 for (; regno < endreg; regno++)
11365 if (reg_stat[regno].last_set
11366 && INSN_CUID (reg_stat[regno].last_set) > from_cuid)
11367 return 1;
11368 return 0;
11371 if (code == MEM && mem_last_set > from_cuid)
11372 return 1;
11374 fmt = GET_RTX_FORMAT (code);
11376 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11378 if (fmt[i] == 'E')
11380 int j;
11381 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11382 if (use_crosses_set_p (XVECEXP (x, i, j), from_cuid))
11383 return 1;
11385 else if (fmt[i] == 'e'
11386 && use_crosses_set_p (XEXP (x, i), from_cuid))
11387 return 1;
11389 return 0;
11392 /* Define three variables used for communication between the following
11393 routines. */
11395 static unsigned int reg_dead_regno, reg_dead_endregno;
11396 static int reg_dead_flag;
11398 /* Function called via note_stores from reg_dead_at_p.
11400 If DEST is within [reg_dead_regno, reg_dead_endregno), set
11401 reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */
11403 static void
11404 reg_dead_at_p_1 (rtx dest, rtx x, void *data ATTRIBUTE_UNUSED)
11406 unsigned int regno, endregno;
11408 if (!REG_P (dest))
11409 return;
11411 regno = REGNO (dest);
11412 endregno = regno + (regno < FIRST_PSEUDO_REGISTER
11413 ? hard_regno_nregs[regno][GET_MODE (dest)] : 1);
11415 if (reg_dead_endregno > regno && reg_dead_regno < endregno)
11416 reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1;
11419 /* Return nonzero if REG is known to be dead at INSN.
11421 We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER
11422 referencing REG, it is dead. If we hit a SET referencing REG, it is
11423 live. Otherwise, see if it is live or dead at the start of the basic
11424 block we are in. Hard regs marked as being live in NEWPAT_USED_REGS
11425 must be assumed to be always live. */
11427 static int
11428 reg_dead_at_p (rtx reg, rtx insn)
11430 basic_block block;
11431 unsigned int i;
11433 /* Set variables for reg_dead_at_p_1. */
11434 reg_dead_regno = REGNO (reg);
11435 reg_dead_endregno = reg_dead_regno + (reg_dead_regno < FIRST_PSEUDO_REGISTER
11436 ? hard_regno_nregs[reg_dead_regno]
11437 [GET_MODE (reg)]
11438 : 1);
11440 reg_dead_flag = 0;
11442 /* Check that reg isn't mentioned in NEWPAT_USED_REGS. For fixed registers
11443 we allow the machine description to decide whether use-and-clobber
11444 patterns are OK. */
11445 if (reg_dead_regno < FIRST_PSEUDO_REGISTER)
11447 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11448 if (!fixed_regs[i] && TEST_HARD_REG_BIT (newpat_used_regs, i))
11449 return 0;
11452 /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, label, or
11453 beginning of function. */
11454 for (; insn && !LABEL_P (insn) && !BARRIER_P (insn);
11455 insn = prev_nonnote_insn (insn))
11457 note_stores (PATTERN (insn), reg_dead_at_p_1, NULL);
11458 if (reg_dead_flag)
11459 return reg_dead_flag == 1 ? 1 : 0;
11461 if (find_regno_note (insn, REG_DEAD, reg_dead_regno))
11462 return 1;
11465 /* Get the basic block that we were in. */
11466 if (insn == 0)
11467 block = ENTRY_BLOCK_PTR->next_bb;
11468 else
11470 FOR_EACH_BB (block)
11471 if (insn == BB_HEAD (block))
11472 break;
11474 if (block == EXIT_BLOCK_PTR)
11475 return 0;
11478 for (i = reg_dead_regno; i < reg_dead_endregno; i++)
11479 if (REGNO_REG_SET_P (block->il.rtl->global_live_at_start, i))
11480 return 0;
11482 return 1;
11485 /* Note hard registers in X that are used. This code is similar to
11486 that in flow.c, but much simpler since we don't care about pseudos. */
11488 static void
11489 mark_used_regs_combine (rtx x)
11491 RTX_CODE code = GET_CODE (x);
11492 unsigned int regno;
11493 int i;
11495 switch (code)
11497 case LABEL_REF:
11498 case SYMBOL_REF:
11499 case CONST_INT:
11500 case CONST:
11501 case CONST_DOUBLE:
11502 case CONST_VECTOR:
11503 case PC:
11504 case ADDR_VEC:
11505 case ADDR_DIFF_VEC:
11506 case ASM_INPUT:
11507 #ifdef HAVE_cc0
11508 /* CC0 must die in the insn after it is set, so we don't need to take
11509 special note of it here. */
11510 case CC0:
11511 #endif
11512 return;
11514 case CLOBBER:
11515 /* If we are clobbering a MEM, mark any hard registers inside the
11516 address as used. */
11517 if (MEM_P (XEXP (x, 0)))
11518 mark_used_regs_combine (XEXP (XEXP (x, 0), 0));
11519 return;
11521 case REG:
11522 regno = REGNO (x);
11523 /* A hard reg in a wide mode may really be multiple registers.
11524 If so, mark all of them just like the first. */
11525 if (regno < FIRST_PSEUDO_REGISTER)
11527 unsigned int endregno, r;
11529 /* None of this applies to the stack, frame or arg pointers. */
11530 if (regno == STACK_POINTER_REGNUM
11531 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
11532 || regno == HARD_FRAME_POINTER_REGNUM
11533 #endif
11534 #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
11535 || (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
11536 #endif
11537 || regno == FRAME_POINTER_REGNUM)
11538 return;
11540 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11541 for (r = regno; r < endregno; r++)
11542 SET_HARD_REG_BIT (newpat_used_regs, r);
11544 return;
11546 case SET:
11548 /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in
11549 the address. */
11550 rtx testreg = SET_DEST (x);
11552 while (GET_CODE (testreg) == SUBREG
11553 || GET_CODE (testreg) == ZERO_EXTRACT
11554 || GET_CODE (testreg) == STRICT_LOW_PART)
11555 testreg = XEXP (testreg, 0);
11557 if (MEM_P (testreg))
11558 mark_used_regs_combine (XEXP (testreg, 0));
11560 mark_used_regs_combine (SET_SRC (x));
11562 return;
11564 default:
11565 break;
11568 /* Recursively scan the operands of this expression. */
11571 const char *fmt = GET_RTX_FORMAT (code);
11573 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
11575 if (fmt[i] == 'e')
11576 mark_used_regs_combine (XEXP (x, i));
11577 else if (fmt[i] == 'E')
11579 int j;
11581 for (j = 0; j < XVECLEN (x, i); j++)
11582 mark_used_regs_combine (XVECEXP (x, i, j));
11588 /* Remove register number REGNO from the dead registers list of INSN.
11590 Return the note used to record the death, if there was one. */
11593 remove_death (unsigned int regno, rtx insn)
11595 rtx note = find_regno_note (insn, REG_DEAD, regno);
11597 if (note)
11599 REG_N_DEATHS (regno)--;
11600 remove_note (insn, note);
11603 return note;
11606 /* For each register (hardware or pseudo) used within expression X, if its
11607 death is in an instruction with cuid between FROM_CUID (inclusive) and
11608 TO_INSN (exclusive), put a REG_DEAD note for that register in the
11609 list headed by PNOTES.
11611 That said, don't move registers killed by maybe_kill_insn.
11613 This is done when X is being merged by combination into TO_INSN. These
11614 notes will then be distributed as needed. */
11616 static void
11617 move_deaths (rtx x, rtx maybe_kill_insn, int from_cuid, rtx to_insn,
11618 rtx *pnotes)
11620 const char *fmt;
11621 int len, i;
11622 enum rtx_code code = GET_CODE (x);
11624 if (code == REG)
11626 unsigned int regno = REGNO (x);
11627 rtx where_dead = reg_stat[regno].last_death;
11628 rtx before_dead, after_dead;
11630 /* Don't move the register if it gets killed in between from and to. */
11631 if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn)
11632 && ! reg_referenced_p (x, maybe_kill_insn))
11633 return;
11635 /* WHERE_DEAD could be a USE insn made by combine, so first we
11636 make sure that we have insns with valid INSN_CUID values. */
11637 before_dead = where_dead;
11638 while (before_dead && INSN_UID (before_dead) > max_uid_cuid)
11639 before_dead = PREV_INSN (before_dead);
11641 after_dead = where_dead;
11642 while (after_dead && INSN_UID (after_dead) > max_uid_cuid)
11643 after_dead = NEXT_INSN (after_dead);
11645 if (before_dead && after_dead
11646 && INSN_CUID (before_dead) >= from_cuid
11647 && (INSN_CUID (after_dead) < INSN_CUID (to_insn)
11648 || (where_dead != after_dead
11649 && INSN_CUID (after_dead) == INSN_CUID (to_insn))))
11651 rtx note = remove_death (regno, where_dead);
11653 /* It is possible for the call above to return 0. This can occur
11654 when last_death points to I2 or I1 that we combined with.
11655 In that case make a new note.
11657 We must also check for the case where X is a hard register
11658 and NOTE is a death note for a range of hard registers
11659 including X. In that case, we must put REG_DEAD notes for
11660 the remaining registers in place of NOTE. */
11662 if (note != 0 && regno < FIRST_PSEUDO_REGISTER
11663 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11664 > GET_MODE_SIZE (GET_MODE (x))))
11666 unsigned int deadregno = REGNO (XEXP (note, 0));
11667 unsigned int deadend
11668 = (deadregno + hard_regno_nregs[deadregno]
11669 [GET_MODE (XEXP (note, 0))]);
11670 unsigned int ourend
11671 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11672 unsigned int i;
11674 for (i = deadregno; i < deadend; i++)
11675 if (i < regno || i >= ourend)
11676 REG_NOTES (where_dead)
11677 = gen_rtx_EXPR_LIST (REG_DEAD,
11678 regno_reg_rtx[i],
11679 REG_NOTES (where_dead));
11682 /* If we didn't find any note, or if we found a REG_DEAD note that
11683 covers only part of the given reg, and we have a multi-reg hard
11684 register, then to be safe we must check for REG_DEAD notes
11685 for each register other than the first. They could have
11686 their own REG_DEAD notes lying around. */
11687 else if ((note == 0
11688 || (note != 0
11689 && (GET_MODE_SIZE (GET_MODE (XEXP (note, 0)))
11690 < GET_MODE_SIZE (GET_MODE (x)))))
11691 && regno < FIRST_PSEUDO_REGISTER
11692 && hard_regno_nregs[regno][GET_MODE (x)] > 1)
11694 unsigned int ourend
11695 = regno + hard_regno_nregs[regno][GET_MODE (x)];
11696 unsigned int i, offset;
11697 rtx oldnotes = 0;
11699 if (note)
11700 offset = hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))];
11701 else
11702 offset = 1;
11704 for (i = regno + offset; i < ourend; i++)
11705 move_deaths (regno_reg_rtx[i],
11706 maybe_kill_insn, from_cuid, to_insn, &oldnotes);
11709 if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x))
11711 XEXP (note, 1) = *pnotes;
11712 *pnotes = note;
11714 else
11715 *pnotes = gen_rtx_EXPR_LIST (REG_DEAD, x, *pnotes);
11717 REG_N_DEATHS (regno)++;
11720 return;
11723 else if (GET_CODE (x) == SET)
11725 rtx dest = SET_DEST (x);
11727 move_deaths (SET_SRC (x), maybe_kill_insn, from_cuid, to_insn, pnotes);
11729 /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG
11730 that accesses one word of a multi-word item, some
11731 piece of everything register in the expression is used by
11732 this insn, so remove any old death. */
11733 /* ??? So why do we test for equality of the sizes? */
11735 if (GET_CODE (dest) == ZERO_EXTRACT
11736 || GET_CODE (dest) == STRICT_LOW_PART
11737 || (GET_CODE (dest) == SUBREG
11738 && (((GET_MODE_SIZE (GET_MODE (dest))
11739 + UNITS_PER_WORD - 1) / UNITS_PER_WORD)
11740 == ((GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))
11741 + UNITS_PER_WORD - 1) / UNITS_PER_WORD))))
11743 move_deaths (dest, maybe_kill_insn, from_cuid, to_insn, pnotes);
11744 return;
11747 /* If this is some other SUBREG, we know it replaces the entire
11748 value, so use that as the destination. */
11749 if (GET_CODE (dest) == SUBREG)
11750 dest = SUBREG_REG (dest);
11752 /* If this is a MEM, adjust deaths of anything used in the address.
11753 For a REG (the only other possibility), the entire value is
11754 being replaced so the old value is not used in this insn. */
11756 if (MEM_P (dest))
11757 move_deaths (XEXP (dest, 0), maybe_kill_insn, from_cuid,
11758 to_insn, pnotes);
11759 return;
11762 else if (GET_CODE (x) == CLOBBER)
11763 return;
11765 len = GET_RTX_LENGTH (code);
11766 fmt = GET_RTX_FORMAT (code);
11768 for (i = 0; i < len; i++)
11770 if (fmt[i] == 'E')
11772 int j;
11773 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
11774 move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_cuid,
11775 to_insn, pnotes);
11777 else if (fmt[i] == 'e')
11778 move_deaths (XEXP (x, i), maybe_kill_insn, from_cuid, to_insn, pnotes);
11782 /* Return 1 if X is the target of a bit-field assignment in BODY, the
11783 pattern of an insn. X must be a REG. */
11785 static int
11786 reg_bitfield_target_p (rtx x, rtx body)
11788 int i;
11790 if (GET_CODE (body) == SET)
11792 rtx dest = SET_DEST (body);
11793 rtx target;
11794 unsigned int regno, tregno, endregno, endtregno;
11796 if (GET_CODE (dest) == ZERO_EXTRACT)
11797 target = XEXP (dest, 0);
11798 else if (GET_CODE (dest) == STRICT_LOW_PART)
11799 target = SUBREG_REG (XEXP (dest, 0));
11800 else
11801 return 0;
11803 if (GET_CODE (target) == SUBREG)
11804 target = SUBREG_REG (target);
11806 if (!REG_P (target))
11807 return 0;
11809 tregno = REGNO (target), regno = REGNO (x);
11810 if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER)
11811 return target == x;
11813 endtregno = tregno + hard_regno_nregs[tregno][GET_MODE (target)];
11814 endregno = regno + hard_regno_nregs[regno][GET_MODE (x)];
11816 return endregno > tregno && regno < endtregno;
11819 else if (GET_CODE (body) == PARALLEL)
11820 for (i = XVECLEN (body, 0) - 1; i >= 0; i--)
11821 if (reg_bitfield_target_p (x, XVECEXP (body, 0, i)))
11822 return 1;
11824 return 0;
11827 /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them
11828 as appropriate. I3 and I2 are the insns resulting from the combination
11829 insns including FROM (I2 may be zero).
11831 ELIM_I2 and ELIM_I1 are either zero or registers that we know will
11832 not need REG_DEAD notes because they are being substituted for. This
11833 saves searching in the most common cases.
11835 Each note in the list is either ignored or placed on some insns, depending
11836 on the type of note. */
11838 static void
11839 distribute_notes (rtx notes, rtx from_insn, rtx i3, rtx i2, rtx elim_i2,
11840 rtx elim_i1)
11842 rtx note, next_note;
11843 rtx tem;
11845 for (note = notes; note; note = next_note)
11847 rtx place = 0, place2 = 0;
11849 next_note = XEXP (note, 1);
11850 switch (REG_NOTE_KIND (note))
11852 case REG_BR_PROB:
11853 case REG_BR_PRED:
11854 /* Doesn't matter much where we put this, as long as it's somewhere.
11855 It is preferable to keep these notes on branches, which is most
11856 likely to be i3. */
11857 place = i3;
11858 break;
11860 case REG_VALUE_PROFILE:
11861 /* Just get rid of this note, as it is unused later anyway. */
11862 break;
11864 case REG_NON_LOCAL_GOTO:
11865 if (JUMP_P (i3))
11866 place = i3;
11867 else
11869 gcc_assert (i2 && JUMP_P (i2));
11870 place = i2;
11872 break;
11874 case REG_EH_REGION:
11875 /* These notes must remain with the call or trapping instruction. */
11876 if (CALL_P (i3))
11877 place = i3;
11878 else if (i2 && CALL_P (i2))
11879 place = i2;
11880 else
11882 gcc_assert (flag_non_call_exceptions);
11883 if (may_trap_p (i3))
11884 place = i3;
11885 else if (i2 && may_trap_p (i2))
11886 place = i2;
11887 /* ??? Otherwise assume we've combined things such that we
11888 can now prove that the instructions can't trap. Drop the
11889 note in this case. */
11891 break;
11893 case REG_NORETURN:
11894 case REG_SETJMP:
11895 /* These notes must remain with the call. It should not be
11896 possible for both I2 and I3 to be a call. */
11897 if (CALL_P (i3))
11898 place = i3;
11899 else
11901 gcc_assert (i2 && CALL_P (i2));
11902 place = i2;
11904 break;
11906 case REG_UNUSED:
11907 /* Any clobbers for i3 may still exist, and so we must process
11908 REG_UNUSED notes from that insn.
11910 Any clobbers from i2 or i1 can only exist if they were added by
11911 recog_for_combine. In that case, recog_for_combine created the
11912 necessary REG_UNUSED notes. Trying to keep any original
11913 REG_UNUSED notes from these insns can cause incorrect output
11914 if it is for the same register as the original i3 dest.
11915 In that case, we will notice that the register is set in i3,
11916 and then add a REG_UNUSED note for the destination of i3, which
11917 is wrong. However, it is possible to have REG_UNUSED notes from
11918 i2 or i1 for register which were both used and clobbered, so
11919 we keep notes from i2 or i1 if they will turn into REG_DEAD
11920 notes. */
11922 /* If this register is set or clobbered in I3, put the note there
11923 unless there is one already. */
11924 if (reg_set_p (XEXP (note, 0), PATTERN (i3)))
11926 if (from_insn != i3)
11927 break;
11929 if (! (REG_P (XEXP (note, 0))
11930 ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0)))
11931 : find_reg_note (i3, REG_UNUSED, XEXP (note, 0))))
11932 place = i3;
11934 /* Otherwise, if this register is used by I3, then this register
11935 now dies here, so we must put a REG_DEAD note here unless there
11936 is one already. */
11937 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3))
11938 && ! (REG_P (XEXP (note, 0))
11939 ? find_regno_note (i3, REG_DEAD,
11940 REGNO (XEXP (note, 0)))
11941 : find_reg_note (i3, REG_DEAD, XEXP (note, 0))))
11943 PUT_REG_NOTE_KIND (note, REG_DEAD);
11944 place = i3;
11946 break;
11948 case REG_EQUAL:
11949 case REG_EQUIV:
11950 case REG_NOALIAS:
11951 /* These notes say something about results of an insn. We can
11952 only support them if they used to be on I3 in which case they
11953 remain on I3. Otherwise they are ignored.
11955 If the note refers to an expression that is not a constant, we
11956 must also ignore the note since we cannot tell whether the
11957 equivalence is still true. It might be possible to do
11958 slightly better than this (we only have a problem if I2DEST
11959 or I1DEST is present in the expression), but it doesn't
11960 seem worth the trouble. */
11962 if (from_insn == i3
11963 && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0))))
11964 place = i3;
11965 break;
11967 case REG_INC:
11968 case REG_NO_CONFLICT:
11969 /* These notes say something about how a register is used. They must
11970 be present on any use of the register in I2 or I3. */
11971 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3)))
11972 place = i3;
11974 if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (i2)))
11976 if (place)
11977 place2 = i2;
11978 else
11979 place = i2;
11981 break;
11983 case REG_LABEL:
11984 /* This can show up in several ways -- either directly in the
11985 pattern, or hidden off in the constant pool with (or without?)
11986 a REG_EQUAL note. */
11987 /* ??? Ignore the without-reg_equal-note problem for now. */
11988 if (reg_mentioned_p (XEXP (note, 0), PATTERN (i3))
11989 || ((tem = find_reg_note (i3, REG_EQUAL, NULL_RTX))
11990 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
11991 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0)))
11992 place = i3;
11994 if (i2
11995 && (reg_mentioned_p (XEXP (note, 0), PATTERN (i2))
11996 || ((tem = find_reg_note (i2, REG_EQUAL, NULL_RTX))
11997 && GET_CODE (XEXP (tem, 0)) == LABEL_REF
11998 && XEXP (XEXP (tem, 0), 0) == XEXP (note, 0))))
12000 if (place)
12001 place2 = i2;
12002 else
12003 place = i2;
12006 /* Don't attach REG_LABEL note to a JUMP_INSN. Add
12007 a JUMP_LABEL instead or decrement LABEL_NUSES. */
12008 if (place && JUMP_P (place))
12010 rtx label = JUMP_LABEL (place);
12012 if (!label)
12013 JUMP_LABEL (place) = XEXP (note, 0);
12014 else
12016 gcc_assert (label == XEXP (note, 0));
12017 if (LABEL_P (label))
12018 LABEL_NUSES (label)--;
12020 place = 0;
12022 if (place2 && JUMP_P (place2))
12024 rtx label = JUMP_LABEL (place2);
12026 if (!label)
12027 JUMP_LABEL (place2) = XEXP (note, 0);
12028 else
12030 gcc_assert (label == XEXP (note, 0));
12031 if (LABEL_P (label))
12032 LABEL_NUSES (label)--;
12034 place2 = 0;
12036 break;
12038 case REG_NONNEG:
12039 /* This note says something about the value of a register prior
12040 to the execution of an insn. It is too much trouble to see
12041 if the note is still correct in all situations. It is better
12042 to simply delete it. */
12043 break;
12045 case REG_RETVAL:
12046 /* If the insn previously containing this note still exists,
12047 put it back where it was. Otherwise move it to the previous
12048 insn. Adjust the corresponding REG_LIBCALL note. */
12049 if (!NOTE_P (from_insn))
12050 place = from_insn;
12051 else
12053 tem = find_reg_note (XEXP (note, 0), REG_LIBCALL, NULL_RTX);
12054 place = prev_real_insn (from_insn);
12055 if (tem && place)
12056 XEXP (tem, 0) = place;
12057 /* If we're deleting the last remaining instruction of a
12058 libcall sequence, don't add the notes. */
12059 else if (XEXP (note, 0) == from_insn)
12060 tem = place = 0;
12061 /* Don't add the dangling REG_RETVAL note. */
12062 else if (! tem)
12063 place = 0;
12065 break;
12067 case REG_LIBCALL:
12068 /* This is handled similarly to REG_RETVAL. */
12069 if (!NOTE_P (from_insn))
12070 place = from_insn;
12071 else
12073 tem = find_reg_note (XEXP (note, 0), REG_RETVAL, NULL_RTX);
12074 place = next_real_insn (from_insn);
12075 if (tem && place)
12076 XEXP (tem, 0) = place;
12077 /* If we're deleting the last remaining instruction of a
12078 libcall sequence, don't add the notes. */
12079 else if (XEXP (note, 0) == from_insn)
12080 tem = place = 0;
12081 /* Don't add the dangling REG_LIBCALL note. */
12082 else if (! tem)
12083 place = 0;
12085 break;
12087 case REG_DEAD:
12088 /* If the register is used as an input in I3, it dies there.
12089 Similarly for I2, if it is nonzero and adjacent to I3.
12091 If the register is not used as an input in either I3 or I2
12092 and it is not one of the registers we were supposed to eliminate,
12093 there are two possibilities. We might have a non-adjacent I2
12094 or we might have somehow eliminated an additional register
12095 from a computation. For example, we might have had A & B where
12096 we discover that B will always be zero. In this case we will
12097 eliminate the reference to A.
12099 In both cases, we must search to see if we can find a previous
12100 use of A and put the death note there. */
12102 if (from_insn
12103 && CALL_P (from_insn)
12104 && find_reg_fusage (from_insn, USE, XEXP (note, 0)))
12105 place = from_insn;
12106 else if (reg_referenced_p (XEXP (note, 0), PATTERN (i3)))
12107 place = i3;
12108 else if (i2 != 0 && next_nonnote_insn (i2) == i3
12109 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12110 place = i2;
12112 if (place == 0
12113 && (rtx_equal_p (XEXP (note, 0), elim_i2)
12114 || rtx_equal_p (XEXP (note, 0), elim_i1)))
12115 break;
12117 if (place == 0)
12119 basic_block bb = this_basic_block;
12121 /* You might think you could search back from FROM_INSN
12122 rather than from I3, but combine tries to split invalid
12123 combined instructions. This can result in the old I2
12124 or I1 moving later in the insn sequence. */
12125 for (tem = PREV_INSN (i3); place == 0; tem = PREV_INSN (tem))
12127 if (! INSN_P (tem))
12129 if (tem == BB_HEAD (bb))
12130 break;
12131 continue;
12134 /* If the register is being set at TEM, see if that is all
12135 TEM is doing. If so, delete TEM. Otherwise, make this
12136 into a REG_UNUSED note instead. Don't delete sets to
12137 global register vars. */
12138 if ((REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER
12139 || !global_regs[REGNO (XEXP (note, 0))])
12140 && reg_set_p (XEXP (note, 0), PATTERN (tem)))
12142 rtx set = single_set (tem);
12143 rtx inner_dest = 0;
12144 #ifdef HAVE_cc0
12145 rtx cc0_setter = NULL_RTX;
12146 #endif
12148 if (set != 0)
12149 for (inner_dest = SET_DEST (set);
12150 (GET_CODE (inner_dest) == STRICT_LOW_PART
12151 || GET_CODE (inner_dest) == SUBREG
12152 || GET_CODE (inner_dest) == ZERO_EXTRACT);
12153 inner_dest = XEXP (inner_dest, 0))
12156 /* Verify that it was the set, and not a clobber that
12157 modified the register.
12159 CC0 targets must be careful to maintain setter/user
12160 pairs. If we cannot delete the setter due to side
12161 effects, mark the user with an UNUSED note instead
12162 of deleting it. */
12164 if (set != 0 && ! side_effects_p (SET_SRC (set))
12165 && rtx_equal_p (XEXP (note, 0), inner_dest)
12166 #ifdef HAVE_cc0
12167 && (! reg_mentioned_p (cc0_rtx, SET_SRC (set))
12168 || ((cc0_setter = prev_cc0_setter (tem)) != NULL
12169 && sets_cc0_p (PATTERN (cc0_setter)) > 0))
12170 #endif
12173 /* Move the notes and links of TEM elsewhere.
12174 This might delete other dead insns recursively.
12175 First set the pattern to something that won't use
12176 any register. */
12177 rtx old_notes = REG_NOTES (tem);
12179 PATTERN (tem) = pc_rtx;
12180 REG_NOTES (tem) = NULL;
12182 distribute_notes (old_notes, tem, tem, NULL_RTX,
12183 NULL_RTX, NULL_RTX);
12184 distribute_links (LOG_LINKS (tem));
12186 SET_INSN_DELETED (tem);
12188 #ifdef HAVE_cc0
12189 /* Delete the setter too. */
12190 if (cc0_setter)
12192 PATTERN (cc0_setter) = pc_rtx;
12193 old_notes = REG_NOTES (cc0_setter);
12194 REG_NOTES (cc0_setter) = NULL;
12196 distribute_notes (old_notes, cc0_setter,
12197 cc0_setter, NULL_RTX,
12198 NULL_RTX, NULL_RTX);
12199 distribute_links (LOG_LINKS (cc0_setter));
12201 SET_INSN_DELETED (cc0_setter);
12203 #endif
12205 else
12207 PUT_REG_NOTE_KIND (note, REG_UNUSED);
12209 /* If there isn't already a REG_UNUSED note, put one
12210 here. Do not place a REG_DEAD note, even if
12211 the register is also used here; that would not
12212 match the algorithm used in lifetime analysis
12213 and can cause the consistency check in the
12214 scheduler to fail. */
12215 if (! find_regno_note (tem, REG_UNUSED,
12216 REGNO (XEXP (note, 0))))
12217 place = tem;
12218 break;
12221 else if (reg_referenced_p (XEXP (note, 0), PATTERN (tem))
12222 || (CALL_P (tem)
12223 && find_reg_fusage (tem, USE, XEXP (note, 0))))
12225 /* This may not be the correct place for the death
12226 note if FROM_INSN is before TEM, and the reg is
12227 set between FROM_INSN and TEM. The reg might
12228 die two or more times. An existing death note
12229 means we are looking at the wrong live range. */
12230 if (from_insn
12231 && INSN_CUID (from_insn) < INSN_CUID (tem)
12232 && find_regno_note (tem, REG_DEAD,
12233 REGNO (XEXP (note, 0))))
12235 tem = from_insn;
12236 if (tem == BB_HEAD (bb))
12237 break;
12238 continue;
12241 place = tem;
12243 /* If we are doing a 3->2 combination, and we have a
12244 register which formerly died in i3 and was not used
12245 by i2, which now no longer dies in i3 and is used in
12246 i2 but does not die in i2, and place is between i2
12247 and i3, then we may need to move a link from place to
12248 i2. */
12249 if (i2 && INSN_UID (place) <= max_uid_cuid
12250 && INSN_CUID (place) > INSN_CUID (i2)
12251 && from_insn
12252 && INSN_CUID (from_insn) > INSN_CUID (i2)
12253 && reg_referenced_p (XEXP (note, 0), PATTERN (i2)))
12255 rtx links = LOG_LINKS (place);
12256 LOG_LINKS (place) = 0;
12257 distribute_links (links);
12259 break;
12262 if (tem == BB_HEAD (bb))
12263 break;
12266 /* We haven't found an insn for the death note and it
12267 is still a REG_DEAD note, but we have hit the beginning
12268 of the block. If the existing life info says the reg
12269 was dead, there's nothing left to do. Otherwise, we'll
12270 need to do a global life update after combine. */
12271 if (REG_NOTE_KIND (note) == REG_DEAD && place == 0
12272 && REGNO_REG_SET_P (bb->il.rtl->global_live_at_start,
12273 REGNO (XEXP (note, 0))))
12274 SET_BIT (refresh_blocks, this_basic_block->index);
12277 /* If the register is set or already dead at PLACE, we needn't do
12278 anything with this note if it is still a REG_DEAD note.
12279 We check here if it is set at all, not if is it totally replaced,
12280 which is what `dead_or_set_p' checks, so also check for it being
12281 set partially. */
12283 if (place && REG_NOTE_KIND (note) == REG_DEAD)
12285 unsigned int regno = REGNO (XEXP (note, 0));
12287 /* Similarly, if the instruction on which we want to place
12288 the note is a noop, we'll need do a global live update
12289 after we remove them in delete_noop_moves. */
12290 if (noop_move_p (place))
12291 SET_BIT (refresh_blocks, this_basic_block->index);
12293 if (dead_or_set_p (place, XEXP (note, 0))
12294 || reg_bitfield_target_p (XEXP (note, 0), PATTERN (place)))
12296 /* Unless the register previously died in PLACE, clear
12297 last_death. [I no longer understand why this is
12298 being done.] */
12299 if (reg_stat[regno].last_death != place)
12300 reg_stat[regno].last_death = 0;
12301 place = 0;
12303 else
12304 reg_stat[regno].last_death = place;
12306 /* If this is a death note for a hard reg that is occupying
12307 multiple registers, ensure that we are still using all
12308 parts of the object. If we find a piece of the object
12309 that is unused, we must arrange for an appropriate REG_DEAD
12310 note to be added for it. However, we can't just emit a USE
12311 and tag the note to it, since the register might actually
12312 be dead; so we recourse, and the recursive call then finds
12313 the previous insn that used this register. */
12315 if (place && regno < FIRST_PSEUDO_REGISTER
12316 && hard_regno_nregs[regno][GET_MODE (XEXP (note, 0))] > 1)
12318 unsigned int endregno
12319 = regno + hard_regno_nregs[regno]
12320 [GET_MODE (XEXP (note, 0))];
12321 int all_used = 1;
12322 unsigned int i;
12324 for (i = regno; i < endregno; i++)
12325 if ((! refers_to_regno_p (i, i + 1, PATTERN (place), 0)
12326 && ! find_regno_fusage (place, USE, i))
12327 || dead_or_set_regno_p (place, i))
12328 all_used = 0;
12330 if (! all_used)
12332 /* Put only REG_DEAD notes for pieces that are
12333 not already dead or set. */
12335 for (i = regno; i < endregno;
12336 i += hard_regno_nregs[i][reg_raw_mode[i]])
12338 rtx piece = regno_reg_rtx[i];
12339 basic_block bb = this_basic_block;
12341 if (! dead_or_set_p (place, piece)
12342 && ! reg_bitfield_target_p (piece,
12343 PATTERN (place)))
12345 rtx new_note
12346 = gen_rtx_EXPR_LIST (REG_DEAD, piece, NULL_RTX);
12348 distribute_notes (new_note, place, place,
12349 NULL_RTX, NULL_RTX, NULL_RTX);
12351 else if (! refers_to_regno_p (i, i + 1,
12352 PATTERN (place), 0)
12353 && ! find_regno_fusage (place, USE, i))
12354 for (tem = PREV_INSN (place); ;
12355 tem = PREV_INSN (tem))
12357 if (! INSN_P (tem))
12359 if (tem == BB_HEAD (bb))
12361 SET_BIT (refresh_blocks,
12362 this_basic_block->index);
12363 break;
12365 continue;
12367 if (dead_or_set_p (tem, piece)
12368 || reg_bitfield_target_p (piece,
12369 PATTERN (tem)))
12371 REG_NOTES (tem)
12372 = gen_rtx_EXPR_LIST (REG_UNUSED, piece,
12373 REG_NOTES (tem));
12374 break;
12380 place = 0;
12384 break;
12386 default:
12387 /* Any other notes should not be present at this point in the
12388 compilation. */
12389 gcc_unreachable ();
12392 if (place)
12394 XEXP (note, 1) = REG_NOTES (place);
12395 REG_NOTES (place) = note;
12397 else if ((REG_NOTE_KIND (note) == REG_DEAD
12398 || REG_NOTE_KIND (note) == REG_UNUSED)
12399 && REG_P (XEXP (note, 0)))
12400 REG_N_DEATHS (REGNO (XEXP (note, 0)))--;
12402 if (place2)
12404 if ((REG_NOTE_KIND (note) == REG_DEAD
12405 || REG_NOTE_KIND (note) == REG_UNUSED)
12406 && REG_P (XEXP (note, 0)))
12407 REG_N_DEATHS (REGNO (XEXP (note, 0)))++;
12409 REG_NOTES (place2) = gen_rtx_fmt_ee (GET_CODE (note),
12410 REG_NOTE_KIND (note),
12411 XEXP (note, 0),
12412 REG_NOTES (place2));
12417 /* Similarly to above, distribute the LOG_LINKS that used to be present on
12418 I3, I2, and I1 to new locations. This is also called to add a link
12419 pointing at I3 when I3's destination is changed. */
12421 static void
12422 distribute_links (rtx links)
12424 rtx link, next_link;
12426 for (link = links; link; link = next_link)
12428 rtx place = 0;
12429 rtx insn;
12430 rtx set, reg;
12432 next_link = XEXP (link, 1);
12434 /* If the insn that this link points to is a NOTE or isn't a single
12435 set, ignore it. In the latter case, it isn't clear what we
12436 can do other than ignore the link, since we can't tell which
12437 register it was for. Such links wouldn't be used by combine
12438 anyway.
12440 It is not possible for the destination of the target of the link to
12441 have been changed by combine. The only potential of this is if we
12442 replace I3, I2, and I1 by I3 and I2. But in that case the
12443 destination of I2 also remains unchanged. */
12445 if (NOTE_P (XEXP (link, 0))
12446 || (set = single_set (XEXP (link, 0))) == 0)
12447 continue;
12449 reg = SET_DEST (set);
12450 while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT
12451 || GET_CODE (reg) == STRICT_LOW_PART)
12452 reg = XEXP (reg, 0);
12454 /* A LOG_LINK is defined as being placed on the first insn that uses
12455 a register and points to the insn that sets the register. Start
12456 searching at the next insn after the target of the link and stop
12457 when we reach a set of the register or the end of the basic block.
12459 Note that this correctly handles the link that used to point from
12460 I3 to I2. Also note that not much searching is typically done here
12461 since most links don't point very far away. */
12463 for (insn = NEXT_INSN (XEXP (link, 0));
12464 (insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR
12465 || BB_HEAD (this_basic_block->next_bb) != insn));
12466 insn = NEXT_INSN (insn))
12467 if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn)))
12469 if (reg_referenced_p (reg, PATTERN (insn)))
12470 place = insn;
12471 break;
12473 else if (CALL_P (insn)
12474 && find_reg_fusage (insn, USE, reg))
12476 place = insn;
12477 break;
12479 else if (INSN_P (insn) && reg_set_p (reg, insn))
12480 break;
12482 /* If we found a place to put the link, place it there unless there
12483 is already a link to the same insn as LINK at that point. */
12485 if (place)
12487 rtx link2;
12489 for (link2 = LOG_LINKS (place); link2; link2 = XEXP (link2, 1))
12490 if (XEXP (link2, 0) == XEXP (link, 0))
12491 break;
12493 if (link2 == 0)
12495 XEXP (link, 1) = LOG_LINKS (place);
12496 LOG_LINKS (place) = link;
12498 /* Set added_links_insn to the earliest insn we added a
12499 link to. */
12500 if (added_links_insn == 0
12501 || INSN_CUID (added_links_insn) > INSN_CUID (place))
12502 added_links_insn = place;
12508 /* Subroutine of unmentioned_reg_p and callback from for_each_rtx.
12509 Check whether the expression pointer to by LOC is a register or
12510 memory, and if so return 1 if it isn't mentioned in the rtx EXPR.
12511 Otherwise return zero. */
12513 static int
12514 unmentioned_reg_p_1 (rtx *loc, void *expr)
12516 rtx x = *loc;
12518 if (x != NULL_RTX
12519 && (REG_P (x) || MEM_P (x))
12520 && ! reg_mentioned_p (x, (rtx) expr))
12521 return 1;
12522 return 0;
12525 /* Check for any register or memory mentioned in EQUIV that is not
12526 mentioned in EXPR. This is used to restrict EQUIV to "specializations"
12527 of EXPR where some registers may have been replaced by constants. */
12529 static bool
12530 unmentioned_reg_p (rtx equiv, rtx expr)
12532 return for_each_rtx (&equiv, unmentioned_reg_p_1, expr);
12535 /* Compute INSN_CUID for INSN, which is an insn made by combine. */
12537 static int
12538 insn_cuid (rtx insn)
12540 while (insn != 0 && INSN_UID (insn) > max_uid_cuid
12541 && NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE)
12542 insn = NEXT_INSN (insn);
12544 gcc_assert (INSN_UID (insn) <= max_uid_cuid);
12546 return INSN_CUID (insn);
12549 void
12550 dump_combine_stats (FILE *file)
12552 fprintf
12553 (file,
12554 ";; Combiner statistics: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n\n",
12555 combine_attempts, combine_merges, combine_extras, combine_successes);
12558 void
12559 dump_combine_total_stats (FILE *file)
12561 fprintf
12562 (file,
12563 "\n;; Combiner totals: %d attempts, %d substitutions (%d requiring new space),\n;; %d successes.\n",
12564 total_attempts, total_merges, total_extras, total_successes);
12568 static bool
12569 gate_handle_combine (void)
12571 return (optimize > 0);
12574 /* Try combining insns through substitution. */
12575 static unsigned int
12576 rest_of_handle_combine (void)
12578 int rebuild_jump_labels_after_combine
12579 = combine_instructions (get_insns (), max_reg_num ());
12581 /* Combining insns may have turned an indirect jump into a
12582 direct jump. Rebuild the JUMP_LABEL fields of jumping
12583 instructions. */
12584 if (rebuild_jump_labels_after_combine)
12586 timevar_push (TV_JUMP);
12587 rebuild_jump_labels (get_insns ());
12588 timevar_pop (TV_JUMP);
12590 delete_dead_jumptables ();
12591 cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_UPDATE_LIFE);
12593 return 0;
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 */